University

LECTURE
Series
Volume 54

Conformal Dimension
Theory and Application
John M. Mackay
Jeremy T. Tyson

American Mathematical Society

Conformal Dimension
Theory and Application
 University
LECTURE
Series
Volume 54

Conformal Dimension
Theory and Application

John M. Mackay
Jeremy T. Tyson

HEMATIC
AT A
M
L

ΤΡΗΤΟΣ ΜΗ
AME ICAN

ΕΙΣΙΤΩ

SOCIETY
ΑΓΕΩΜΕ
R

FO
UN 8 88
DED 1

American Mathematical Society

Providence, Rhode Island
 EDITORIAL COMMITTEE
Eric M. Friedlander (Chair) Benjamin Sudakov
William P. Minicozzi II Tatiana Toro

2010 Mathematics Subject Classification. Primary 30L10, 28A78.

For additional information and updates on this book, visit

www.ams.org/bookpages/ulect-54

Library of Congress Cataloging-in-Publication Data

Mackay, John M., 1982–
Conformal dimension : theory and application / John M. Mackay, Jeremy T. Tyson.
p. cm. — (University lecture series ; v. 54)
Includes bibliographical references and index.
ISBN 978-0-8218-5229-3 (alk. paper)
1. Quasiconformal mappings. 2. Hausdorff measures. I. Tyson, Jeremy T., 1972– II. Title.

QA360.M285 2010
515.93—dc22
2010014667

Copying and reprinting. Individual readers of this publication, and nonprofit libraries
acting for them, are permitted to make fair use of the material, such as to copy a chapter for use
in teaching or research. Permission is granted to quote brief passages from this publication in
reviews, provided the customary acknowledgment of the source is given.
Republication, systematic copying, or multiple reproduction of any material in this publication
is permitted only under license from the American Mathematical Society. Requests for such
permission should be addressed to the Acquisitions Department, American Mathematical Society,
201 Charles Street, Providence, Rhode Island 02904-2294 USA. Requests can also be made by
e-mail to [email protected].

c 2010 by the American Mathematical Society. All rights reserved.
The American Mathematical Society retains all rights
except those granted to the United States Government.
Printed in the United States of America.

∞ The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at http://www.ams.org/
10 9 8 7 6 5 4 3 2 1 15 14 13 12 11 10
Dedicated to Pierre Pansu on the occasion of his fiftieth birthday
 Contents

Preface ix

List of Figures xiii

Chapter 1. Background material 1

1.1. Notation and conventions 1
1.2. Mappings 1
1.3. Quasisymmetric invariants 8
1.4. Dimensions and measures 9
1.5. Notes 17

Chapter 2. Conformal gauges and conformal dimension 21

2.1. Conformal gauges 21
2.2. Conformal dimension 22
2.3. Notes 24

Chapter 3. Gromov hyperbolic groups and spaces and their boundaries 25

3.1. The real hyperbolic plane and CAT(κ) spaces 25
3.2. Gromov hyperbolic metric spaces 26
3.3. Boundaries of rank one symmetric spaces 30
3.4. Gromov hyperbolic groups 40
3.5. Boundaries of hyperbolic buildings 43
3.6. Notes 45

Chapter 4. Lower bounds for conformal dimension 49

4.1. Rich curve families and lower bounds for conformal dimension 49
4.2. The Loewner condition and examples of spaces minimal for
conformal dimension 55
4.3. The conformal dimension of the Sierpiński carpet 58
4.4. Annular LLC implies conformal dimension bounded away from
one 61
4.5. Rich families of sets of conformal dimension one 62
4.6. Notes 64

Chapter 5. Sets and spaces of conformal dimension zero 67

5.1. Conformal dimension, porosity and uniform disconnectedness 67
5.2. Conformal dimension and unions 71
5.3. Conformal dimension and products 73
vii
viii CONTENTS

5.4. Tukia’s example 77

5.5. Notes 78
Chapter 6. Gromov–Hausdorff tangent spaces and conformal
dimension 79
6.1. Tangent spaces and lower bounds for conformal Assouad
dimension 79
6.2. Lowering the conformal Assouad dimension 82
6.3. Notes 89
Chapter 7. Ahlfors regular conformal dimension 91
7.1. David–Semmes deformation theory 91
7.2. Cannon’s conjecture and Ahlfors regular conformal dimension 92
7.3. p cohomology of graphs 93
7.4. Dynamics of post-critically finite rational maps on S2 97
7.5. Notes 104
Chapter 8. Global quasiconformal dimension 107
8.1. Distortion of dimension by quasiconformal maps 107
8.2. Minimality for global quasiconformal dimension and middle
interval Cantor sets 108
8.3. Global quasiconformal dimensions of self-similar sets 109
8.4. Conformal dimensions of self-affine sets 115
8.5. Sets with global quasiconformal dimension zero 122
8.6. Notes 125
Bibliography 131
Index 139
 Preface

Quasisymmetric maps are homeomorphisms that distort relative dis-

tance in a uniform, scale-invariant fashion. Their role in geometric function
theory is pervasive. Euclidean quasiconformal maps are locally quasisym-
metric. Quasisymmetry provides a metric substitute for quasiconformality
suitable for the study of maps defined on lower-dimensional submanifolds
or subsets of an ambient Euclidean space. On the other hand, quasiconfor-
mal maps of Euclidean spaces of sufficiently high dimension, or of sufficiently
nice domains therein, are quasisymmetric. It is by now well understood that
quasisymmetry is a remarkably ubiquitous condition which also arises in
many settings outside of classical function theory. For instance, the bound-
ary extensions of quasi-isometries of hyperbolic groups are quasisymmetric.
Quasisymmetric maps of the boundaries of strictly pseudoconvex domains
in Cn , equipped with the standard Carnot-Carathéodory metric adapted
to their CR structures, control the behavior of proper holomorphic maps
near the boundary. Julia sets of rational maps, parametrized by hyperbolic
components of the moduli space, are quasisymmetrically equivalent. Other
examples could be given.
Determining the quasisymmetry type of a metric space is one of the
premier open problems in geometric function theory. One of the most well-
studied quasisymmetric invariants is conformal dimension. This invariant
was introduced by Pansu in the late 1980’s as a tool for the quasi-isometric
classification of the classical rank one symmetric spaces and lattices in such
spaces. In recent years a significant body of mathematics has accumulated
concerning conformal dimensions of metric spaces which may not necessarily
arise as group boundaries. Other applications are beginning to be identified,
in areas as diverse as first-order analysis in metric spaces, dynamics of ratio-
nal maps, Gromov hyperbolic geometry, and self-affine fractal geometry. In
our view, the theory has developed and expanded to a point which justifies
a presentation emphasizing its scope and maturity. We have attempted to
provide such a presentation in the present survey.
We provide a comprehensive picture of the current state of the art re-
garding both computations and estimates of conformal dimension. Our pre-
sentation emphasizes both pure theory—computation of the conformal di-
mensions of metric spaces for its own sake, as an interesting exercise in
geometric function theory—as well as applications of such computations in

ix
x PREFACE

related fields. We discuss the relevance of conformal dimension for the quasi-
isometric classification of Gromov hyperbolic groups, for the investigation
of post-critically finite branched covers of the Riemann sphere arising from
the iteration of rational maps, and for the study of self-similar and self-affine
subsets of Euclidean space. These applications serve both as motivation for
the development of a rich theory of conformal dimensions of metric spaces
and as a source for new questions and problems.

Outline. We begin with a chapter of background material (Chapter 1)

reviewing basic terminology of analysis in metric spaces. We recall sev-
eral important classes of mappings between metric spaces (Lipschitz and
bi-Lipschitz, Hölder, and quasisymmetric maps) and several quasisymmetric
invariants of metric spaces. In Section 1.4 we recall several key metrically
defined notions of dimension for metric spaces: Hausdorff dimension, As-
souad dimension and Ahlfors regular dimension. We also review some of the
primary tools used to compute or estimate these dimensions.
In Chapter 2 we introduce the principal actors of this story: the confor-
mal gauge of a metric space and its conformal dimension. More precisely, we
introduce two distinct yet related notions of conformal dimension, defined
using two different metrically natural notions of dimension. The former,
conformal Hausdorff dimension, was Pansu’s original concept, while the lat-
ter, Ahlfors regular conformal dimension, arises in many of the most striking
applications.
In Chapter 3 we recall the definition of Gromov hyperbolicity for met-
ric spaces and groups. We illustrate with examples of group boundaries,
including spheres at infinity for rank one symmetric spaces and hyperbolic
buildings. The calculations by Pansu and Bourdon of the conformal dimen-
sions of these boundaries were an early impetus for the development of the
theory of conformal dimension.
The heart of this survey resides in Chapter 4, where we indicate the
main tools and techniques used for establishing lower bounds on conformal
dimension. The typical approach is to exhibit a large, diffuse family of
rectifiable curves. These techniques can be used to show that many spaces
(Carnot groups, boundaries of hyperbolic buildings, Ahlfors regular Loewner
spaces, etc.) are minimal for conformal dimension.
Sets and spaces of conformal dimension zero are well understood. Chap-
ter 5 describes what is known. The most important result is Kovalev’s
theorem (Theorem 5.1.12) on spaces of conformal dimension less than one.
We indicate some ideas from the proof of Kovalev’s theorem in Section 8.5.
Chapter 6 discusses the relationship between conformal dimension and
Gromov–Hausdorff weak tangents. The highlight is a theorem of Keith and
Laakso (Theorem 6.2.1) characterizing certain spaces which are minimal for
Ahlfors regular conformal dimension in terms of weak tangents.
In Chapter 7, we review various applications of Ahlfors regular conformal
dimension to recognition and uniformization problems for metric 2-spheres
 PREFACE xi

arising as boundaries of hyperbolic groups or through the dynamics of post-

critically finite rational and branched covering maps. We also describe the
connection between Ahlfors regular conformal dimension and p cohomology
of hyperbolic graphs.
Conformal dimension is an abstract concept, defined in terms of arbi-
trary quasisymmetric images of a given metric space. For subsets of a fixed
space, we introduce a related invariant, the global quasiconformal dimension.
Chapter 8 outlines known results on the global quasiconformal dimensions of
subsets of Euclidean space. Of particular interest are constructions which re-
duce dimension by global quasiconformal maps. Such constructions are fea-
sible for sets which are totally disconnected or almost totally disconnected:
Cantor sets and post-critically finite self-similar fractals. We illustrate the
techniques developed in Chapter 4 through a discussion of the conformal
dimensions of self-affine fractals which are not self-similar. We conclude
with Kovalev’s theorem (Theorem 8.5.1) on sets of global quasiconformal
dimension less than one.
We have oriented this survey towards a broad audience with interests in
analysis in metric spaces, geometric group theory, dynamics, rigidity, and
related subjects. No significant prior background in these areas is required.
We have included proofs for results of particular interest or importance.
For other results, we only provide a brief sketch of the proof or refer the
interested reader to the literature. Our bibliography is extensive but by no
means complete; it should be treated merely as an invitation to explore this
rich subject in greater detail. Likewise, the choice of specific results which
we present is greatly influenced by our own research interests and expertise.
Finally, a warning to the reader. The term conformal dimension is also
used in mathematical physics with an entirely different meaning, where it
refers to a feature of Hilbert space operators which describes their scaling
behavior under change of variables. This conflict in terminology is unfor-
tunate; however, both uses of the term are now firmly entrenched in the
literature.

Acknowledgements. Many individuals have contributed to the de-

velopment of the rich theory described here. We have benefited greatly
from discussions of these topics with numerous colleagues and collabora-
tors. We would especially like to recognize Zoltán Balogh, Mario Bonk,
Bruce Kleiner, Ilya Kapovich, Sergiy Merenkov, Pierre Pansu, Kevin Pil-
grim, Kevin Wildrick and Jang–Mei Wu. We are particularly grateful to
Hrant Hakobyan and Leonid Kovalev for providing us with specific and de-
tailed comments on an early draft of this manuscript.
The notion of conformal dimension was introduced by Pierre Pansu in
his 1989 Ann. Acad. Sci. Fenn. paper “Dimension conforme et sphère à
l’infini des variétés à courbure négative” [133]. The influence of this notion
on the fields of geometric group theory and analysis in metric spaces has
been extensive. We have attempted to do justice to the intricate theory
xii PREFACE

that has developed in response to Pansu’s original paper. We dedicate this

survey to him with great admiration on the occasion of his fiftieth birthday
and the twentieth anniversary of the publication of [133].
The second author’s research was supported in part by NSF Grants
DMS-0228807 and DMS-0555869.

J.M.M.
J.T.T.
Urbana, Illinois
October 2009
 List of Figures

1.1 The von Koch snowflake curve K 4

4.1 The Sierpiński carpet SC 59

4.2 Weights on the Sierpiński carpet SC 60

7.1 A Lattès map of S2 99

7.2 PCF branched covering map of S with a Thurston obstruction
2 100
7.3 Lattès-type PCF branched covering map of S 2 101
7.4 PCF branched covering map of S with a Thurston obstruction and
2

fractal conformal gauge of Rickman’s rug type 104

8.1 The Sierpiński gasket SG 111

8.2 Iterated function systems defining the Sierpiński gasket SG and a
quasisymmetrically equivalent invariant set 114
8.3 A Bedford–McMullen carpet 117
8.4 Polygaskets P G(N ) 128

xiii
 CHAPTER 1

Background material

1.1. Notation and conventions

Throughout this survey, (X, d) and (Z, d) will denote metric spaces. We
often abuse notation by dropping explicit reference to the metric. We write
B(z, r), resp. B(z, r), for the open, resp. closed, ball in Z with center z and
radius r. For B = B(z, r) we let tB = B(z, tr); we emphasize that the set tB
depends on the choice of center and radius for B (which are not necessarily
unique in metric spaces). We write diam A for the diameter of a set A ⊂ Z,
and dist(A, B) for the distance between sets A, B ⊂ Z. Note that
(1.1.1) diam B(z, r) ≤ 2r
for all metric balls B(z, r) in Z.
Unless otherwise specified, we equip the Euclidean space RN of dimen-
sion N with the Euclidean (2 ) norm and metric. We will occasionally denote
the Euclidean metric in RN by deucl . We denote Lebesgue measure in RN
by LN . The cardinality of a finite set S will be written #S.
As is customary, we let C, c, . . . denote various finite positive constants.
These constants may depend on auxiliary data a, b, etc.; we indicate this by
writing C(a, b) or c(a, b). We will write C if we wish to emphasize that a
certain constant is finite, and we will write c or 1/C if we wish to emphasize
that the constant is positive.

1.2. Mappings
In this section, we review a number of important classes of mappings
between metric spaces. The most important of these, from the perspec-
tive of this survey, is the class of quasisymmetric maps (Definition 1.2.9),
which underlies the notions of conformal gauge and conformal dimension.
The study of quasiconformal and quasi-Möbius maps has a long history in
geometric function theory: the theory of these classes of maps generalizes
classical aspects of complex analysis.
We begin with some simpler classes: Lipschitz and Hölder maps.

1.2.1. Lipschitz and Hölder maps. Lipschitz functions are arguably

the most important class of functions in analysis in metric spaces. They
play a role there analogous to that played by smooth functions in classical
Euclidean analysis.
1
2 1. BACKGROUND MATERIAL

Definition 1.2.1. Let (Z, d) and (Z  , d ) be metric spaces and let L > 0.
A map f : Z → Z  is said to be L-Lipschitz if
d (f (z1 ), f (z2 )) ≤ Ld(z1 , z2 )
for all z1 , z2 ∈ Z. We say that f is Lipschitz if it is L-Lipschitz for some
L < ∞.
When Z  = R we say that f is a Lipschitz function.
The Lipschitz condition is an absolute metric distortion requirement:
distances in the image should not exceed distances in the source by more
than a fixed, universal multiplicative constant. The importance of Lips-
chitz functions in analysis rests largely on results such as the following two
theorems.
Theorem 1.2.2 (Rademacher Differentiation Theorem). Every Lipschitz
function defined on a subset of RN is differentiable LN -almost everywhere.

Theorem 1.2.3 (McShane Extension Theorem). Let Z be any metric

space and let f : A → R be an L-Lipschitz function defined on a set A ⊂ Z.
Then there exists an L-Lipschitz extension of f to Z, i.e., there exists an
L-Lipschitz function g : Z → R so that g|A = f .
Note that, as a consequence of Theorem 1.2.3, we may without loss of
generality assume that the function in Theorem 1.2.2 is defined in all of
RN . Indeed, suppose that f : A → R, A ⊂ RN , is a Lipschitz function. By
McShane’s Theorem 1.2.3, there is a Lipschitz extension g : RN → R of f .
By Rademacher’s Theorem 1.2.2,
Z := {x ∈ RN : g is not differentiable at x}
is a set of measure zero in RN . Since
{x ∈ A : f is not differentiable at x}
is a subset of Z ∩A, it also has measure zero. Thus f is differentiable LN -a.e.
in A, as asserted.
We note the following elementary characterization of Lipschitz maps in
terms of metric balls. A generalization of this observation to quasisymmetric
maps will be useful later on.
Lemma 1.2.4. Let f : Z → Z  . Then f is L-Lipschitz if and only
f (B(z, r)) ⊂ B(f (z), Lr) for all balls B(z, r) in Z.
We recall some basic examples. The distance function d : Z → R to a
point z0 ∈ Z, d(z) := d(z, z0 ), is a 1-Lipschitz function; this follows directly
from the triangle inequality. More generally, for any closed set F ⊂ Z, the
function d(z) = dist(z, F ) is 1-Lipschitz. The class of Lipschitz functions
on a metric space Z is closed under addition and the operations of taking
maxima and minima. Compositions of Lipschitz maps are Lipschitz.
 1.2. MAPPINGS 3

The projection maps

πj : RN → R, j = 1, . . . , N,
or
πj : ∞ → R, j = 1, 2, . . . ,
given by πj (x) = xj are 1-Lipschitz maps. Thus πj ◦ f is Lipschitz whenever
f : Z → RN or f : Z√→ ∞ is Lipschitz. Conversely, if πj ◦ f is L-Lipschitz
for all j, then f is L N -Lipschitz (if the target is RN ) or is L-Lipschitz (if
the target is ∞ ).
Definition 1.2.5. A homeomorphism f : (Z, d) → (Z  , d ) is L-bi-
Lipschitz if both f and f −1 are L-Lipschitz. We say that f is bi-Lipschitz if
it is L-bi-Lipschitz for some finite L.
In many contexts, bi-Lipschitz equivalent spaces are indistinguishable.
For instance, all of the most well-studied notions of metric dimension (see
section 1.4) are invariant under bi-Lipschitz maps. The problem of rec-
ognizing when two metric spaces are bi-Lipschitz equivalent is in general
quite challenging. For instance, there is still no simple list of intrinsic met-
ric properties which characterize bi-Lipschitz images of RN , except in the
simple case N = 1.
Products of bi-Lipschitz maps are always bi-Lipschitz. Given two metric
spaces (Z1 , d1 ) and (Z2 , d2 ) we consider the product space Z3 = Z1 × Z2 .
We may equip Z3 with the 2 product metric

d3 ((z1 , z2 ), (z1 , z2 )) = d1 (z1 , z1 )2 + d2 (z2 , z2 )2 .
Alternatively, we may use the p product metric
((z1 , z2 ), (z1 , z2 ))
→ (d1 (z1 , z1 )p + d2 (z2 , z2 )p )1/p
for any 1 ≤ p < ∞ or the corresponding ∞ product metric
((z1 , z2 ), (z1 , z2 ))
→ max{d1 (z1 , z1 ), d2 (z2 , z2 )}.
All of these metrics are bi-Lipschitz equivalent.
Proposition 1.2.6. Let f1 : (Z1 , d1 ) → (Z1 , d1 ) and f2 : (Z2 , d2 ) →
(Z2 , d2 )
be bi-Lipschitz maps. Then (f1 , f2 ) : (Z3 , d3 ) → (Z3 , d3 ) is bi-
Lipschitz, where Z3 = Z1 × Z2 and Z3 = Z1 × Z2 .
The concept of a Lipschitz map admits the following natural generaliza-
tion.
Definition 1.2.7. Let Z and Z  be as in Definition 1.2.1 and let L, α >
0. A map f : Z → Z  is said to be (L, α)-Hölder if
d (f (z1 ), f (z2 )) ≤ Ld(z1 , z2 )α
for all z1 , z2 ∈ Z. We say that f is α-Hölder if it is (L, α)-Hölder for some
L < ∞, and that f is Hölder if it is α-Hölder for some α > 0.
Again, when Z  = R we say that f is a Hölder function.
4 1. BACKGROUND MATERIAL

The obvious analog of Lemma 1.2.4 holds for Hölder functions. We leave
the precise formulation to the reader.
Lipschitz and Hölder maps are closely related. To state this more pre-
cisely we recall the notion of a snowflake metric space.
Definition 1.2.8. Let (Z, d) be any metric space, let 0 <  < 1, and
define d on Z by
d (z1 , z2 ) := d(z1 , z2 ) .
Then d is a metric. The metric space (Z, d ) is called the snowflake of (Z, d)
with parameter .
The primary example (and the reason for the terminology) is the von
Koch snowflake curve K ⊂ R2 , see Figure 1.1. This self-similar fractal set
is homeomorphic to the unit interval [0, 1]. It has metric dimension log 4
log 3 > 1.
(Here we refer to any of the notions of dimension introduced in section 1.4—
all of them agree for this set.) Let α = log 3
log 4 . Then (K, deucl ) is bi-Lipschitz
equivalent to the snowflaked space ([0, 1], dαeucl ).

Figure 1.1. The von Koch snowflake curve K

It follows immediately from the definitions that a function f : (Z, d) →
(Z  , d )
is α-Hölder (for some α ≤ 1) if and only if f : (Z, dα ) → (Z  , d ) is
Lipschitz. Coupling this with the McShane Extension Theorem, we observe
that every real-valued Hölder function defined on a subset of a metric space
Z extends to a Hölder function on all of Z, with the same Hölder exponent
and coefficient.
Hölder functions with exponent larger than one are of little interest in
the classical setting. In fact, if f is a (1 + )-Hölder function defined on
an open subset Ω of RN , then f is constant. For the proof, we note that
f is Lipschitz on compact subdomains of Ω, hence differentiable a.e. in Ω.
It is easy to check that the derivative must vanish wherever it is defined.
Since f is also absolutely continuous along every line segment in Ω and Ω
is connected, we conclude that f is constant. However, by the remarks of
the previous paragraph, (1 + )-Hölder functions defined on more general
subsets of RN need not be constant.
Note that the product of the snowflakes of (Z1 , d1 ) and (Z2 , d2 ) with the
same parameter  is bi-Lipschitz equivalent to the snowflake of (Z1 × Z2 , d3 )
with parameter . Products of snowflakes with different parameters are
never bi-Lipschitz equivalent with each other, and are rarely even quasisym-
metrically equivalent (for the definition of quasisymmetric maps, see the
following section). See the notes to section 4 for more details.
 1.2. MAPPINGS 5

1.2.2. Quasisymmetric, quasiconformal and quasi-Möbius maps.

The bi-Lipschitz condition imposes a restriction on absolute metric distor-
tion: distances in the image are uniformly comparable to the corresponding
distances in the source. The notion of quasisymmetry, which we now intro-
duce, is more subtle: it imposes a restriction on relative metric distortion.
Quasisymmetric maps are at the core of this survey. In the Euclidean setting
(and in many other contexts), quasisymmetry turns out to be equivalent to
the more classical notion of quasiconformality, which arose first in the con-
text of complex analysis. We review the basic facts.
A distortion function is a homeomorphism of [0, ∞) onto itself.
Definition 1.2.9. Let η be a distortion function. Let f : (Z, d) →
(Z  , d )
be a homeomorphism of metric spaces. We say that f is η-quasi-
symmetric if
 
d (f (z1 ), f (z2 )) d(z1 , z2 )
(1.2.1) ≤η
d (f (z1 ), f (z3 )) d(z1 , z3 )
for all z1 , z2 , z3 ∈ Z, z1 = z3 . We say that f is quasisymmetric if it is
η-quasisymmetric for some η, and that (Z, d) and (Z  , d ) are quasisymmet-
rically equivalent if there exists a quasisymmetry from Z to Z  .
Thus f is quasisymmetric if ratios of distances of pairs of points are
distorted in a controlled fashion. This three-point condition is extremely
useful. We will see numerous applications in later chapters.
Examples 1.2.10. (1) Every L-bi-Lipschitz map f : Z → Z  is η-quasi-
symmetric with distortion function η(t) = L2 t.
(2) The identity map from (Z, d) to its snowflake (Z, d ) with parameter
 < 1 is η-quasisymmetric with distortion function η(t) = t .
(3) If f1 : Z → Z  is an η1 -quasisymmetric homeomorphism and f2 :
Z → Z  is an η2 -quasisymmetric homeomorphism, then f2 ◦ f1 : Z → Z  is


η2 ◦η1 -quasisymmetric. The inverse of an η-quasisymmetric homeomorphism

is η̃-quasisymmetric with distortion function η̃(t) = 1/η −1 (1/t). Hence the
collection of all quasisymmetric homeomorphisms of Z onto itself forms a
group.
Quasisymmetry is a global condition: it imposes a uniform requirement
on the relative metric distortion of any triple of points. In the Euclidean
case, this condition turns out to be equivalent to an infinitesimal version.
Definition 1.2.11. Let F : Ω → Ω be a homeomorphism between
domains in RN . The linear dilatation of F is the function HF : Ω → R given
by
sup{|F (x) − F (y)| : |x − y| = r}
(1.2.2) HF (x) := lim sup .
r→0 inf{|F (x) − F (z)| : |x − z| = r}
We say that F is H-quasiconformal, H ≥ 1, if HF (x) ≤ H for all x ∈ Ω,
and that F is quasiconformal if it is H-quasiconformal for some H < ∞.
6 1. BACKGROUND MATERIAL

It is easy to see from the definitions that if F is η-quasisymmetric, then

it is H-quasiconformal with H = η(1). In fact, the stronger condition
sup{|F (x) − F (y)| : |x − y| = r}
sup ≤ η(1)
r>0 inf{|F (x) − F (z)| : |x − z| = r}
holds true. It is a deep fact, central to modern geometric function theory,
that the converse also holds true, at least for globally defined mappings.
The same result holds true for homeomorphisms between sufficiently nice
domains.
Theorem 1.2.12. Let F : RN → RN be a H-quasiconformal homeomor-
phism, N ≥ 2. Then F is η-quasisymmetric for some distortion function η
which depends only on H and N .
Remark 1.2.13. There are many characterizations of quasiconformal
maps. Here is an analytic one. Assume that F is a homeomorphism in the
1,N
local Sobolev class Wloc (Ω : RN ) (the latter condition means that the first-
order partial derivatives of the components of F exist in the distributional
loc (Ω)), and that there exists K < ∞ so that the almost
sense as elements of LN
everywhere defined differential matrix DF satisfies the following reverse
Hadamard inequality
(1.2.3) ||DF (x)||N ≤ K det DF (x)
for a.e. x ∈ Ω. Geometrically, inequality (1.2.3) asserts that the volume of
the image of an infinitesimal ball centered at x is comparable with the volume
of the smallest circumscribed ball. By elementary geometry, this condition
implies and is implied by the corresponding condition that infinitesimal balls
are mapped onto infinitesimal ellipsoids with uniformly bounded eccentric-
ity. Indeed, F is quasiconformal (in the sense of Definition 1.2.11) if and
only if it satisfies this condition. The constants H and K depend only on
each other and on the dimension N .
Remark 1.2.14. In the setting of (1.2.3), the regularity assumption
1,N
F ∈ Wloc is natural, since it is equivalent to the local integrability of the
Jacobian det DF (which is relevant for change of variables arguments). How-
ever, it is a deep fact in Euclidean geometric function theory that weaker
regularity assumptions are sufficient, and that stronger regularity assump-
tions are in fact valid in the presence of a distortion condition such as (1.2.3).
A fundamental tool is Gehring’s higher integrability lemma, which in its sim-
1,N
plest form asserts that a Wloc mapping F satisfying (1.2.3) for some K < ∞
1,p
in fact belongs to Wloc for all p < p(N, K) for some specific p(N, K) > N .
See the notes to this section for further remarks.
Remark 1.2.15. Theorem 1.2.12 admits the following local version. Let
us say that F : Ω → Ω is locally quasisymmetric if it is quasisymmetric
on every compact set S ⊂ Ω, with distortion function η depending only
on S (and N ). Then every quasiconformal mapping F : Ω → Ω between
arbitrary domains Ω, Ω ⊂ RN , N ≥ 2, is locally quasisymmetric.
 1.2. MAPPINGS 7

Remark 1.2.16. The restriction to dimensions N ≥ 2 is essential. Every

C1 homeomorphism F : R → R is quasiconformal, but such a homeomor-
phism need not be quasisymmetric.
Example 1.2.17. Every 1-quasiconformal map defined between domains
in R2 is a conformal map (in the sense of complex analysis). Every 1-
quasiconformal map defined between domains in RN , N ≥ 3, is the restric-
tion of a globally defined Möbius transformation of RN : this is a consequence
of the Gehring–Liouville theorem.
We remark on an interesting corollary of Theorem 1.2.12: inverses of
quasiconformal maps of RN (N ≥ 2) are quasiconformal. This follows im-
mediately from Theorem 1.2.12 and Example 1.2.10(3). It is not at all easy
to see this fact directly from the definition of quasiconformality.
The following lemma recasts quasisymmetry in terms of the distortion
of metric annuli. It is the analog of Lemma 1.2.4 from the Lipschitz setting.
Lemma 1.2.18. A homeomorphism f : Z → Z  is η-quasisymmetric if
and only if whenever B1 = B(z, r1 ) ⊂ B2 = B(z, r2 ) are concentric balls in
Z, then there exist concentric balls B1 = B  (f (z), r1 ) ⊂ B2 = B  (f (z), r2 )
so that B1 ⊂ f (B1 ) ⊂ f (B2 ) ⊂ B2 and r2 /r1 ≤ η(r2 /r1 ).
We note however an important distinction between quasiconformal maps
and quasisymmetric maps. Quasiconformal maps need not preserve bound-
edness. (Indeed, even conformal maps of the plane can change bounded
domains into unbounded domains.) By way of contrast, quasisymmetries
always preserve boundedness. More generally, we have the following well-
known result on the distortion of relative diameter by quasisymmetric maps.
Lemma 1.2.19 (Tukia–Väisälä). Let f : Z → Z  be an η-quasisymmetric
homeomorphism and assume that A ⊂ B ⊂ Z are such that 0 < diam A ≤
diam B < ∞. Then diam f (B) is finite and
 
1 diam f (A) 2 diam A
≤ ≤ η .
2η( diam B
diam A ) diam f (B) diam B

It is for this reason sometimes advantageous to work with a different class

of mappings: the so-called quasi-Möbius maps. The quasi-Möbius condition
is a four-point condition. One can recover quasisymmetric maps on locally
compact metric spaces from quasi-Möbius maps on the one-point (Alexan-
drov) compactifications by specializing one of the points to the point at
infinity.
Definition 1.2.20. The (real) cross ratio of four points z1 , z2 , z3 , z4 in
a metric space (Z, d) is
d(z1 , z3 )d(z2 , z4 )
[z1 , z2 , z3 , z4 ] := .
d(z1 , z4 )d(z2 , z3 )
8 1. BACKGROUND MATERIAL

A map F : (Z, d) → (Z  , d ) is called θ-quasi-Möbius, where θ is a distortion

function, if
[f (z1 ), f (z2 ), f (z3 ), f (z4 )] ≤ θ([z1 , z2 , z3 , z4 ])
for all quadruples z1 , z2 , z3 , z4 ∈ Z, z1 = z4 , z2 = z3 .
Quasisymmetric maps are always quasi-Möbius, while quasi-Möbius maps
are quasisymmetric provided the source and target space are simultaneously
bounded or simultaneously unbounded.

1.3. Quasisymmetric invariants

In order to distinguish spaces up to some notion of equivalence, one re-
quires invariants. By Example 1.2.10(3), quasisymmetry is an equivalence
relation on the class of metric spaces. In this section, we indicate a few
important quasisymmetric invariants: the doubling property, uniform per-
fectness, and uniform disconnectedness.

1.3.1. Doubling spaces.

Definition 1.3.1. A metric space (Z, d) is doubling if there exists a
constant N < ∞ so that any ball B(z, r) in Z can be covered by at most N
balls of radius r/2. We say that (Z, d) is N -doubling.
The doubling condition is a finite-dimensionality hypothesis which con-
trols the growth of the cardinalities of separated subsets of balls at all scales
and locations. (This remark is made precise in Definition 1.4.14.) Examples
of doubling spaces include Euclidean space of any dimension and arbitrary
subsets thereof.

1.3.2. Uniformly perfect spaces.

Definition 1.3.2. A metric space (Z, d) is uniformly perfect if there
exists a constant c > 0 so that B(z, r) \ B(z, cr) = ∅ for all z ∈ Z and
0 < r < diam Z. We say that (Z, d) is c-uniformly perfect.
Uniform perfectness is a quantitative, scale-invariant version of topo-
logical perfectness: it asserts the nonexistence of separating annuli of large
modulus. Every connected metric space is uniformly perfect. The Cantor
set is uniformly perfect.

1.3.3. Uniformly disconnected spaces.

Definition 1.3.3. Let (Z, d) be a metric space. A finite sequence of
points z0 , z1 , . . . , zN in X is called a δ-chain if d(zi−1 , zi ) ≤ δd(z0 , zN ) for all
i = 1, . . . , N . We do not require that the points be distinct. Such a chain
is called trivial if all of its points coincide. We say that (Z, d) is uniformly
disconnected if there exists δ > 0 so that there are no nontrivial δ-chains in
Z. In this case we also say that (Z, d) is δ-uniformly disconnected.
 1.4. DIMENSIONS AND MEASURES 9

Chains should be viewed as finite analogs of curves. Indeed, in any path

connected space, every pair of points can be joined by δ-chains with arbi-
trarily small δ. Uniform disconnectedness is a quantitative, scale-invariant
version of total disconnectedness: it asserts the existence of separating an-
nuli of a definite modulus at all scales and locations. The Cantor set is
uniformly disconnected for some specific positive value of δ.
Theorem 1.3.4. Let (Z1 , d1 ) have property P, where P is one of the
following conditions:
• the doubling property,
• uniform perfectness,
• uniform disconnectedness.
Let f : Z1 → Z2 be a quasisymmetric homeomorphism of Z1 onto another
metric space (Z2 , d2 ). Then Z2 has property P. The statement is quantita-
tive.

1.4. Dimensions and measures

The subject of this survey is another quasisymmetric invariant: confor-
mal dimension. Versions of conformal dimension are defined (see section
2.2) by infimizing certain metrically defined notions of dimension over all
quasisymmetric images. In this section, we review several notions of dimen-
sion for metric spaces. We begin with the most well-known and useful of
these notions, the Hausdorff dimension.
1.4.1. Hausdorff measure and dimension. The Hausdorff measures
on a metric space arise as instances of a general construction due to Carathé-
odory. As we will make use of this construction in several different ways, we
begin with a general review.

Carathéodory’s construction. Let (Z, d) be a metric space. The data

associated to Carathéodory’s construction are a family F of subsets of Z and
a function φ : F → [0, +∞] (called a gauge function).
Typical examples of families F which we will use are the full power set
of Z, the family of all metric balls in Z, or (in the case Z = RN ) the family
of dyadic cubes. We usually consider gauge functions φ which depend only
on the diameter of the argument, although this is not essential and there
are situations where more general gauges are relevant. See, for example, the
proof of Lemma 6.2.8.
Associated to (Z, d), F and φ, and for each 0 < δ ≤ ∞, we define the
following set function:
∞ ∞

 
Φδ (A) := inf φ(Ai ) : A ⊂ Ai , diam Ai < δ, Ai ∈ F , A ⊂ Z.
i=1 i=1
Some comments are in order. We have imposed no further assumptions on
F , thus, it may happen that the infimum in question is taken over the empty
10 1. BACKGROUND MATERIAL

set for certain choices of δ and A. In this case, we assign Φδ (A) = +∞. In
case A = ∅ we allow the trivial (empty) covering (one way to do this is
to include the empty set as an element of F and assign φ(∅) = 0), thus
Φδ (∅) = 0 for all δ. With these remarks in mind, the value of Φδ (A) is
well-defined for all δ and A.
We call (Φδ )δ>0 the family of Carathéodory premeasures associated to F
and φ. We also call Φ∞ the Carathéodory content.
For each 0 < δ ≤ ∞, the function Φδ is an outer measure on Z. Indeed, it
is clearly monotone and countably subadditive, and by convention it vanishes
on the empty set. It is not necessarily a Borel measure, however. The
following definition overcomes this shortcoming.
Definition 1.4.1. The Carathéodory measure associated to the data F
and φ is
Φ(A) := lim Φδ (A) = sup Φδ (A),
δ→0 δ>0
where A ⊂ Z.
The existence of the limit and the equality of these two formulas is an
easy consequence of the monotonicity of the function δ
→ Φδ (A).
We use the following criterion, also due to Carathéodory, to verify that
Φ is a Borel measure.
Theorem 1.4.2. Let μ be an outer measure on a metric space (Z, d).
Then μ is a Borel measure if and only if μ(A ∪ B) = μ(A) + μ(B) for every
pair of sets A, B ⊂ Z with dist(A, B) > 0.

Corollary 1.4.3. Carathéodory measures are Borel.

Proof. It suffices to verify the condition in Theorem 1.4.2. For fixed
A, B ⊂ Z with dist(A, B) > 0, we have Φδ (A∪B) = Φδ (A)+Φδ (B) whenever
δ ≤ 12 dist(A, B). Taking the limit as δ → 0 completes the proof. 

Hausdorff measure and content. We now specialize to the most

important example. Let F be the power set of Z, let s ≥ 0, and let
hs : F → [0, ∞] be given by
hs (A) := (diam A)s .
Definition 1.4.4. The Hausdorff premeasures, resp. Hausdorff content,
resp. Hausdorff measure, in dimension s are the functions (Hδs )δ>0 , H∞
s , Hs ,

given by Carathéodory’s construction for the gauge function φ = hs on F .

By the preceding remarks, Hs is a Borel measure on Z for any s ≥ 0.
Examples 1.4.5. On any metric space, the Hausdorff measure in di-
mension zero coincides with the counting measure
H0 (A) = #A.
 1.4. DIMENSIONS AND MEASURES 11

When Z = RN and s = m ≤ N is an integer, Hs coincides with the usual

surface area measure for m-dimensional smooth submanifolds in RN , up to
a positive multiple. In particular, HN coincides with the Lebesgue measure
LN up to a multiple. For intervals I ⊂ R, L1 (I) = H1 (I) = H∞
1 (I) = diam I.

Remark 1.4.6. Let F be the family of all metric balls in (Z, d), let
φs (A) = (diam A)s for A ∈ F , and let Φsδ be the corresponding family of
Carathéodory premeasures. Then
(1.4.1) Hδs (A) ≤ Φsδ (A) ≤ 2s Hδs (A)
for all A ⊂ Z and all δ ≥ 0. The left hand inequality is obvious; for the right
hand inequality we note that any nonempty subset A of Z is contained inside
a ball of radius diam A and use (1.1.1). These measures Φs are sometimes
known as the spherical Hausdorff measures on Z.

Remark 1.4.7. Let Z = RN and let F be the family of dyadic cubes,

i.e., closed cubes with sides parallel to the coordinate axes and vertices at
consecutive dyadic rational coordinates. More precisely, A ∈ F if and only
if
A = [k1 2−m , (k1 + 1)2−m ] × · · · × [kN 2−m , (kN + 1)2−m ]
for integers m and k1 , k2 , . . . , kN . Let φs be as in Remark 1.4.6 and let Φsδ
be the corresponding family of Carathéodory premeasures. Then
√
(1.4.2) Hδs (A) ≤ Φsδ (A) ≤ (4 N )s Hδs (A)
for all A ⊂ Z and all δ ≥ 0. The argument is similar to the previous one.
These measures Φs are sometimes known as the dyadic Hausdorff measures
on RN .
Lemma 1.4.8. Let 0 ≤ s < t < ∞ and A ⊂ Z. If Hs (A) < ∞, then
Ht (A)= 0.
Proof. For each 0 < δ < ∞ it follows from the definitions that Hδt (A) ≤
δ t−s Hδs (A)
≤ δ t−s Hs (A). Let δ → 0. 
Definition 1.4.9. The Hausdorff dimension of a set A ⊂ Z is
dim A = inf{s ≥ 0 : Hs (A) = 0} = sup{s ≥ 0 : Hs (A) = +∞}.
Note that the equality of these two formulas follows immediately upon
considering the contrapositive in Lemma 1.4.8. Throughout this survey, we
will denote by dim the Hausdorff dimension. On occasion, if we wish to
distinguish it from other notions of dimension, we may write dimH for the
Hausdorff dimension.
Hausdorff dimension is monotonic:
dim A ≤ dim B if A ⊂ B
and countably stable:
∞


dim Ai = sup dim Ai for A1 , A2 , . . . in Z.
i=1 i
12 1. BACKGROUND MATERIAL

We will see many examples of dimensions of specific sets and metric spaces
throughout this text. For now we merely observe that the Hausdorff dimen-
sion of RN (or any open set in RN ) is equal to N . Similarly, the Hausdorff
dimension of any smooth m-dimensional submanifold of RN is equal to m.
Remark 1.4.10. By (1.4.1) and (1.4.2), the value of dim A is unchanged
if we replace the Hausdorff measures in its definition with either the spherical
Hausdorff measures or the dyadic Hausdorff measures (in RN ).
Estimates for the Hausdorff dimension of a set A in Z arise primarily via
measure-theoretic arguments. It suffices to find Borel measures supported
on A with suitable volume growth properties. Two primary results along
these lines are the Mass Distribution Principle and its converse, Frostman’s
lemma.
Theorem 1.4.11. (1) (Mass Distribution Principle) If there exists a
positive measure μ on A satisfying μ(B(z, r)) ≤ rs for all B(z, r) ⊂ Z, then
Hs (A) > 0.
(2) (Frostman’s lemma) If A is compact and Hs (A) > 0, then there
exists a positive Borel measure μ on A satisfying μ(B(z, r)) ≤ rs for all
B(z, r) ⊂ Z.
The proof of (1) is trivial: Hs (A) ≥ H∞ s (A) ≥ μ(A) > 0. The proof

of Frostman’s lemma is significantly more involved; see the notes to this

section for specific references.
Every uniformly perfect metric space has positive Hausdorff dimension.
Indeed, by iterating the uniform perfectness condition one can construct a
subset which is bi-Lipschitz equivalent to a Cantor set of positive dimension.

1.4.2. Topological dimension. We will denote by dimT Z the topo-

logical dimension (or small inductive dimension) of Z. This is defined as
follows:
dimT ∅ = −1
and dimT Z ≤ n if Z admits a neighborhood basis of sets {Ai } whose bound-
aries satisfy dimT ∂Ai ≤ n − 1. It is clear that RN has topological dimension
N . Any compact totally disconnected metric space has topological dimen-
sion zero. Note that the topological dimension is always an integer, unlike
Hausdorff dimension which may assume any nonnegative real value.
The main fact which we want to highlight is the following.
Theorem 1.4.12. For any metric space Z, the inequality
(1.4.3) dimT Z ≤ dim Z
holds.
There is a beautiful “Fubini-type” proof for Theorem 1.4.12 which re-
duces it to the following implication: if Hn+1 (Z) = 0, then Hn (∂B(z, r)) = 0
 1.4. DIMENSIONS AND MEASURES 13

for each z ∈ Z and for L1 -a.e. r > 0. Since the metric balls are a neigh-
borhood basis in Z, Theorem 1.4.12 follows by induction on the topological
dimension.

1.4.3. Ahlfors regularity. Many natural spaces which arise in geo-

metric function theory, dynamics and rigidity exhibit strong self-similarity
properties. This self-similarity implies that the space looks similar at all
locations and scales. Ahlfors regularity provides one quantitative measure-
ment of such self-similarity, from the perspective of measure theory. A
metric space is Ahlfors regular, roughly speaking, if the measures of balls of
a fixed radius are all comparable, and scale like an appropriate power of the
radius. Here is the precise definition.
Definition 1.4.13. A metric space (Z, d) is Ahlfors s-regular for some
s > 0 if there exists a Borel measure μ on Z and a constant C < ∞ so that
1 s
(1.4.4) r ≤ μ(B(z, r)) ≤ Crs
C
for all balls B(z, r) in Z with r < diam Z. We say that (Z, d) is Ahlfors
regular if it is Ahlfors s-regular for some s > 0. We also say that μ is an
Ahlfors regular measure on (Z, d).
If (Z, d) is Ahlfors s-regular with respect to some Borel measure μ, then
it is Ahlfors s-regular with respect to the Hausdorff measure Hs (possibly
with a different constant C) and μ and Hs are comparable. Indeed, for any
covering of a ball B(z, r) with sets Ai , we compute
1 s  
r ≤ μ(B(z, r)) ≤ μ(Ai ) ≤ (diam Ai )s
C
i i

(where we included each set Ai in a ball of radius equal to its diameter),

from which it follows that
1
Hs (B(z, r)) ≥ H∞s
(B(z, r)) ≥ rs .
C
The remaining direction is similar, but uses also a suitable covering theorem
(see Lemma 4.1.1).
For each s > 0 there exist Ahlfors s-regular metric spaces. Examples can
be selected from among the self-similar Cantor subsets of Euclidean spaces.
In fact, all self-similar subsets of Euclidean space generated by iterated func-
tion systems satisfying the so-called open set condition are Ahlfors regular.
See section 8.3 for more details.
Every Ahlfors regular space is uniformly perfect. Indeed, suppose that
(Z, d) satisfies (1.4.4) for some s > 0, some C and some measure μ. If
B(z, r) \ B(z, cr) is empty for some positive r, then rs /C ≤ μ(B(z, r)) =
μ(B(z, cr)) ≤ C(cr)s and c ≥ C −2/s . Thus Z is 12 C −2/s -uniformly perfect.
Similarly, every Ahlfors s-regular space with s < 1 is uniformly discon-
nected. For a more general statement, see section 1.4.4.
14 1. BACKGROUND MATERIAL

1.4.4. Assouad dimension. In a metric space (Z, d), call a subset S

-separated if d(z1 , z2 ) ≥  for all distinct z1 , z2 ∈ S. Let N (S, ) be the
maximal cardinality of an -separated subset of S. Then Z is doubling (as
in Definition 1.3.1) if and only if there exists C < ∞ so that
(1.4.5) N (B(z, r), r/2) ≤ C
for all balls B(z, r) in Z with r < diam Z. An easy inductive argument
shows that if (1.4.5) holds, then there exists C  and s depending only on C
so that
(1.4.6) N (B(z, r), r) ≤ C  −s
for all balls B(z, r) in Z with r < diam Z and all 0 <  < 1. For instance,
we may take s = log2 C.
Definition 1.4.14. The Assouad dimension, dimA Z, of a doubling met-
ric space (Z, d) is the infimum of those s > 0 for which (1.4.6) holds for some
C  < ∞, for all balls B(z, r) and 0 <  ≤ 2.
Thus (Z, d) is doubling precisely when it has finite Assouad dimension.
The inequality
(1.4.7) dim Z ≤ dimA Z
between Hausdorff and Assouad dimension is valid for all spaces (Z, d); this is
an easy exercise. Every nonempty open set in RN has Assouad dimension N .
Assouad dimension is monotonic: dimA S ≤ dimA Z whenever S ⊂ Z.
Let (Z, d) be an Ahlfors s-regular space. Then (1.4.6) holds for some C 
depending only on the constant C from (1.4.4) and on s. To see this, note
that if z1 , . . . , zN are points in a ball B(z, r) which are all at distance r from
each other ( ≤ 2), then the balls B(zi , 12 r) are all disjoint and contained
in B(z, 2r). The estimate N ≤ C 2 4s −s follows from volume considerations.
From this and the previous remarks we conclude that
(1.4.8) Z Ahlfors s-regular =⇒ dim Z = dimA Z = s.
In Chapter 5 we will prove that every metric space with Assouad dimen-
sion strictly less than one is uniformly disconnected. It follows that Ahlfors
s-regular spaces with s < 1 are uniformly disconnected.
Several warnings are in order. Assouad dimension is not countably stable
(Example 1.4.15). Also, the Assouad dimension of a metric space Z agrees
with the Assouad dimension of its completion, thus the rationals Q have
Assouad dimension one. These properties illustrate that Assouad dimension
behaves quite differently from its Hausdorff counterpart.
Example 1.4.15. For each p > 0, the set {1, 2−p , 3−p , 4−p , . . .} has As-
souad dimension equal to one.
In Chapter 6 we will give an alternate derivation of Example 1.4.15 by
showing that Assouad dimension does not increase upon passing to weak
 1.4. DIMENSIONS AND MEASURES 15

tangents. It is not hard to verify that {1, 2−p , 3−p , 4−p , . . .} has the set
[0, ∞) as a tangent space at the origin. Then
1 = dimA [0, ∞) ≤ dimA {1, 2−p , 3−p , 4−p , . . .} ≤ dimA R = 1
and equality holds throughout.
A fundamental property of doubling metric spaces is Assouad’s Embed-
ding Theorem.
Theorem 1.4.16 (Assouad). If (Z, d) is a doubling metric space and
 < 1, then there exists N < ∞ so that the snowflaked space (Z, d ) admits
a bi-Lipschitz embedding in RN .
The conclusion in Theorem 1.4.16 need not hold if  = 1. Non-abelian
Carnot groups, equipped with their standard Carnot-Carathéodory (sub-
Riemannian) metric, provide counterexamples. We discuss the sub-Riemann-
ian geometry of Carnot groups in more detail in chapter 3.
1.4.5. Behavior of dimensions under mappings. Lipschitz and
Hölder maps behave well with respect to Hausdorff dimensions and Ahlfors
regularity, but can behave rather pathologically with respect to Assouad
dimension.
Theorem 1.4.17. (1) Let f : Z → Z  be L-Lipschitz. Then HLδ
s (f (A)) ≤

Ls Hδs (A)for all A ⊂ Z and all 0 ≤ δ ≤ ∞.

(2) Let f : Z → Z  be (L, α)-Hölder. Then HLδα (f (A)) ≤ Ls/α Hδs (A)
s/α

for all A ⊂ Z and all 0 ≤ δ ≤ ∞.

The proof is obvious.
Corollary 1.4.18. (1) If f : Z → Z  is Lipschitz, then dim f (A) ≤
dim A for all A ⊂ Z. If f : Z → Z  is a bi-Lipschitz homeomorphism, then
dim f (A) = dim A
for all A ⊂ Z.
(2) If f : Z → Z  is α-Hölder, then dim f (A) ≤ α−1 dim A for all A ⊂ Z.
If f : Z → Z  is an (α, β)-bi-Hölder homeomorphism (i.e., f is α-Hölder
and f −1 is β-Hölder) for some α, β with αβ ≤ 1, then
β dim A ≤ dim f (A) ≤ α−1 dim A
for all A ⊂ Z.
(3) For any  < 1 and any metric space (Z, d), dim(Z, d ) = −1 dim Z.
It is also easy to see that the image of an Ahlfors s-regular metric space
under a bi-Lipschitz homeomorphism is again s-regular. The image of an
Ahlfors regular metric space under an (α, β)-bi-Hölder homeomorphism need
not be Ahlfors regular, but the snowflake of an s-regular metric space Z with
parameter  < 1 is s/-regular.
Bi-Hölder homeomorphisms can also be ill-behaved with respect to the
Assouad dimension.
16 1. BACKGROUND MATERIAL

Example 1.4.19. There exists a countable doubling metric space (Z, d)

and a Lipschitz homeomorphism f : Z → Z  onto a non-doubling space
(Z  , d ). In fact, for each β < 1 we construct a (1, β)-bi-Hölder homeomor-
phism from a space Z with dimA Z = 1 onto a space Z  with dimA Z  = ∞.
By Example 1.4.15 the Assouad dimension of {n−p : n ∈ N} is equal
to one, for each p > 0. By the finite stability of the Assouad dimension,
{n−p : n ≥ n0 } has Assouad dimension equal to one for any (fixed) n0 ∈ N.
Let us consider Z = {n ∈ N : n ≥ n0 } with the metric d(p) defined by
d(p) (m, n) = |m−p − n−p |; then (Z, d(p) ) is isometric with {n−p : n ≥ n0 }.
Let β < 1 and define d : N × N → ∞ by d (m, n) = max{m−1/β , n−1/β }
if m = n and d (m, n) = 0 if m = n. We leave it to the reader to check that
d is a metric on N.
If we choose p > 0 such that p < β1 − 1 and choose n0 = n0 (p) ∈ N
appropriately, then the identity map from (Z, d(p) ) to (Z, d ) satisfies
1
d (m, n)1/β ≤ d (m, n) ≤ Cd(p) (m, n)
C (p)
for all m, n ∈ Z and a suitable constant C = C(p) < ∞, i.e., the identity
map is a (1, β)-bi-Hölder homeomorphism. In particular, it is Lipschitz.
We show that (Z, d ) is not doubling. Again by the finite stability of the
Assouad dimension, it suffices to show that (N, d ) is not doubling. Since
lim diamd {m, m + 1, m + 2, . . .} = lim m−1/β = 0,
m→∞ m→∞
(m) is a Cauchy sequence in (N, d ).
There is a unique limit point for this
Cauchy sequence in the completion of (N, d ); we denote this point by 0.
The metric d extends to {0} ∪ N and d (0, m) = m−1/β . We will show that
lim N (B(0, r), 12 r) = ∞
r→0
which implies that (N, d ) is not doubling.
Observe that B(0, r) = {0} ∪ {m ∈ N : m > r−β } and that the points
m, m + 1, . . . , m + p ∈ B(0, r) are all 12 r-separated provided
 1/β
1 r
≥ ,
m+p 2
i.e., m + p ≤ (2/r)β . Choose p = [(2β − 1)r−β ] − 1 and m0 = [r−β + 1],
where [·] denotes the integer part function. Then m0 , . . . , m0 + p are all
( 12 r)-separated and
N (B(0, r), 12 r) ≥ p + 1 = [(2β − 1)r−β ]
which tends to infinity as r → 0.
However, Assouad dimension is invariant under bi-Lipschitz maps.
We conclude this chapter with an application of Theorem 1.4.17.
Proposition 1.4.20. For any connected metric space (Z, d), we have
diam Z ≤ H∞
1 (Z).
 1.5. NOTES 17

Proof. Suppose first that Z is unbounded. We claim that H∞ 1 (Z) =

+∞. Fix z0 ∈ Z and consider the map f : Z → R given by f (z) = d(z, z0 ).

This map is 1-Lipschitz, so Theorem 1.4.17 implies that
H∞
1
(Z) ≥ H∞
1
(f (Z)) = H∞
1
([0, +∞)) = L1 ([0, +∞)) = ∞.
Here we used the fact that f (Z) is an unbounded interval in the nonnegative
real axis containing 0, hence is equal to [0, ∞).
Now suppose that Z is bounded. Let  > 0 and fix z0 ∈ Z so that
sup d(z, z0 ) ≥ diam Z − .
z∈Z
Let f be as above. Again, f (Z) is an interval in the nonnegative real axis
containing 0, with
L1 (f (Z)) ≥ sup |f (z) − f (z0 )| ≥ diam Z − .
z∈Z
As before we find
H∞
1
(Z) ≥ H∞
1
(f (Z)) = L1 (f (Z)) ≥ diam Z − .
Now let  → 0. 
As a corollary, note that every space of Hausdorff dimension strictly less
than one is totally disconnected. (If such a space Z contained a nontrivial
connected component, we would conclude that H1 (Z) > 0.)

1.5. Notes
Rademacher’s theorem 1.2.2 is a remarkable fact, highly indicative of
the geometry of the underlying Euclidean space. It asserts that Lipschitz
functions are intrinsically equipped with a powerful infinitesimal rigidity
property (local linearity), at least away from a null set. Rademacher’s origi-
nal paper appeared in 1919 [144] and has stimulated a wealth of subsequent
generalizations, reformulations and applications which we cannot hope to
describe in detail here. We mention only the striking extension of The-
orem 1.2.2 by Cheeger [44] to the category of arbitrary doubling metric
measure spaces satisfying a Poincaré inequality. This generalization extends
the Rademacher theorem to a variety of new spaces with nonsmooth infin-
itesimal geometry, such as Carnot groups and other Carnot-Carathéodory
manifolds (section 3.3.5) and the boundaries of certain hyperbolic build-
ings (section 3.5). It is now clear, as a consequence of the results of [44]
that the validity of a Rademacher-type differentiation theorem has strong
implications for the geometry of the underlying space.
McShane’s Extension Theorem 1.2.3 has also opened new lines of inves-
tigation which have expanded dramatically in recent years. The existence of
a Lipschitz extension is quite easy to prove. In fact, given any L-Lipschitz
function f : A → R defined on a nonempty set A ⊂ Z, it is easy to check
that the formulas
g+ (z) := inf{f (a) + Ld(a, z) : a ∈ A}
18 1. BACKGROUND MATERIAL

and
g− (z) := sup{f (a) − Ld(a, z) : a ∈ A}
define L-Lipschitz extensions of f to all of Z. Furthermore, g− (z) ≤ g+ (z)
for all z ∈ Z. The existence of absolute minimizing Lipschitz extensions
(AMLE’s), i.e., extensions which minimize the Lipschitz constant locally,
was first shown by Aronsson [2] in the Euclidean setting. The study of such
extensions involves significant use of tools from nonlinear PDE, especially
the theory of the ∞-Laplace equation. The uniqueness of AMLE’s was later
established by Jensen [86]. See [3] for a comprehensive overview of the
theory of ∞-harmonic functions and absolutely minimizing Lipschitz exten-
sions in the Euclidean setting. Quite recently, Peres, Schramm, Sheffield and
Wilson [143] provided a game-theoretic reinterpretation of the ∞-Laplace
equation in terms of “tug-of-war” and thereby obtain a probabilistic ap-
proach to Jensen’s uniqueness theorem valid in extremely general settings.
This area of research is still developing rapidly.
Lipschitz extension theorems with other target spaces are also of some
interest. Lang, Pavlović and Schroeder [100], [99] consider the case of target
spaces with curvature bounds. In a series of papers [103], [104], Lee and
Naor consider the problem of Lipschitz extension where the subset A is
fixed and the superset Z is allowed to vary. They obtain striking results on
this absolute extension problem for a variety of spaces and targets, including
doubling metric spaces, graphs and Banach spaces.
Theorem 1.2.12 and Remark 1.2.15 are folklore, implicit in early work
of Gehring [57], [58] and Väisälä [174]. The local quasisymmetry of qua-
siconformal maps is sometimes known as the egg yolk principle; Heinonen
[76, Chapter 11] attributes the term to McMullen. The precise definition of
quasisymmetric maps in the metric space context was first given by Tukia
and Väisälä [166]. However, the notion was implicit in many earlier works.
In particular, quasisymmetric maps on the real line (as the boundary maps
of quasiconformal maps of the upper half plane) have been well studied since
at least the work of Ahlfors and Beurling [14]. The canonical reference for
quasiconformal maps in RN is Väisälä’s monograph [175]. For a recent
treatment, see Iwaniec–Martin [84].
Liouville’s theorem is a strong rigidity statement about conformal map-
pings in Euclidean spaces of sufficiently high dimension. It asserts that any
conformal mapping between domains in RN is the restriction of a Möbius
transformation, provided N ≥ 3. Liouville’s original proof required a suit-
able a priori regularity assumption. The version of Liouville’s theorem stated
in Example 1.2.17 is due to Gehring [58]. The result is false when N = 2;
every analytic map with nonvanishing derivative is conformal. Liouville’s
theorem also holds in more general spaces such as Carnot groups.
Gehring’s higher integrability lemma (Remark 1.2.14) is one of the most
important tools in geometric function theory; it instigated numerous in-
vestigations of self-improving integrability properties which continue to the
 1.5. NOTES 19

present day. The sharp integrability exponent p(N, K) in Remark 1.2.14 is

not known. A well-known conjecture asserts that
N K 1/(N −1)
(1.5.1) p(N, K) = ;
K 1/(N −1) − 1
this value, if correct, would be sharp as indicated by the radial stretch map
F (x) = |x|(1/K)−1 x. In a seminal work, Astala [5] established the conjecture
(1.5.1) in the case N = 2.
Theorem 1.5.1 (Astala). Let F : Ω → Ω be a Wloc1,2
homeomorphism

between domains Ω, Ω ⊂ C satisfying the distortion inequality ||DF ||2 ≤
1,p
K det DF a.e. Then F ∈ Wloc 2K
for each p < K−1 .
Gehring’s higher integrability lemma and Astala’s theorem 1.5.1 have
direct applications to the study of Hausdorff dimension distortion by qua-
siconformal maps. See Chapter 8 for further information. A comprehensive
account of the theory of elliptic partial differential equations and quasicon-
formal mappings in the plane can be found in the recent book of Astala,
Iwaniec and Martin [6].
An excellent reference for the theory of topological dimension is the book
of Hurewicz and Wallman [81]. Theorem 1.4.12 can be found there; see also
[76, Theorem 8.14]. For the theory of Hausdorff dimension, we recommend
the books of Rogers [149] and Falconer [52], [51]. A vast amount of in-
formation can be found in Federer’s monumental treatise [53]. For a more
recent discussion, see Mattila [116].
Assouad dimension was considered (under a different name) by Assouad
[4]. An extensive development of the history and theory of Assouad di-
mension was provided by Luukkainen [107]. Example 1.4.19 is a slight
elaboration on an example in the appendix of [107].
Assouad’s embedding theorem 1.4.16 [4] features in a surprisingly wide
spectrum of applications. Bonk and Schramm [26] use it to construct
coarse embeddings of Gromov hyperbolic spaces, Hajlasz [71] uses it to
build Whitney-style surjections of high regularity with large critical locus,
and Semmes [153] uses it to establish a bi-Lipschitz embedding theorem for
doubling spaces with target equipped with a metric defined via deformation
by an A∞ weight (compare section 7.1).
 CHAPTER 2

Conformal gauges and conformal dimension

2.1. Conformal gauges

The problem of classifying metric spaces up to quasisymmetry arises nat-
urally in many contexts. Conformal gauges provide a convenient language
for discussing quasisymmetric invariance. We recall the definition.
Definition 2.1.1. The conformal gauge G of a metric space (Z, d) is the
collection of all metric spaces (Z  , d ) which are quasisymmetrically equiva-
lent to (Z, d).
Any quasisymmetrically invariant property of metric spaces can be un-
derstood as a property of the associated conformal gauge. For example, by
Theorem 1.3.4, we can speak of a doubling conformal gauge, a uniformly
perfect conformal gauge, or a uniformly disconnected conformal gauge.
The following theorem characterizes the conformal gauge of the Cantor
set. It is a quantitative version of the well-known topological characteriza-
tion of the Cantor set as the unique compact, perfect, totally disconnected
metric space (up to homeomorphism).
Theorem 2.1.2. A metric space is quasisymmetrically equivalent to the
Cantor set if and only if it is compact, uniformly perfect and uniformly
disconnected.
In other words, the Euclidean conformal gauge on the Cantor set is the
unique conformal gauge which is compact, uniformly perfect and uniformly
disconnected.
The proof of Theorem 2.1.2 consists of encoding the Cantor set quasisym-
metrically via the abstract symbol space {0, 1}N (equipped with a suitable
self-similar metric), and then showing that any space satisfying the assump-
tions of the theorem admits a similar quasisymmetric coding by the same
symbol space.
We note the following consequence of Theorem 2.1.2: if a conformal
gauge G is compact, uniformly perfect and uniformly disconnected, then G
is doubling.
Other spaces whose conformal gauges can be explicitly characterized
include the real line R and the circle S1 . In general, however, it is extremely
difficult to write down a simple list of properties characterizing a given
conformal gauge. For example, characterizing the conformal gauge of the
sphere SN , N ≥ 2, is one of the most difficult and important problems in
21
22 2. CONFORMAL GAUGES AND CONFORMAL DIMENSION

the area. We will have more to say about this problem later in this survey
(see section 7.2).
In the language of conformal gauges, Assouad’s embedding theorem
1.4.16 states that a conformal gauge G embeds in a Euclidean space if and
only if it is doubling.

2.2. Conformal dimension

We now recall one of the most important quasisymmetry invariants, the
focus of this survey. The following concept was first introduced by Pansu
[133] in connection with the quasi-isometric classification problem for rank
one symmetric spaces.
Definition 2.2.1. The conformal dimension, Cdim G, of a conformal
gauge G is the infimum of the Hausdorff dimensions of all metric spaces
in G.
We will abuse notation by writing Cdim Z for the conformal dimension of
the conformal gauge associated to a given metric space (Z, d). Occasionally,
we will write CdimH Z in place of Cdim Z to emphasize that Hausdorff
dimension is under consideration.
Let (Z, d) be any metric space. Then
(2.2.1) dimT Z = inf dim Z 
where the infimum is taken over all metric spaces (Z  , d ) which are homeo-
morphic with Z. In other words, there is never any obstruction to lowering
the Hausdorff dimension of a metric space by a homeomorphism (apart from
the obvious a priori restriction (1.4.3)). On the other hand, bi-Lipschitz
homeomorphisms preserve Hausdorff dimension. Quasisymmetric homeo-
morphisms form a class which is intermediate between bi-Lipschitz home-
omorphisms and arbitrary homeomorphisms. By (1.4.3) and (2.2.1), the
following a priori inequalities hold true:
dimT Z ≤ Cdim Z ≤ dim Z.
It follows from Corollary 1.4.18(3) that the supremum of dim Z  over all
spaces (Z  , d ) in a conformal gauge G is equal to either 0 or +∞. Confor-
mal dimension, on the other hand, is a subtle and useful invariant which
reveals much about the structure of the gauge. The goal of this survey is to
document and justify this claim.
As we will see later, it is of critical importance to determine when the
Hausdorff dimension of a space cannot be decreased by quasisymmetries,
i.e., when it is minimal for conformal dimension.
Definition 2.2.2. A metric space (Z, d) is minimal for conformal di-
mension if
Cdim Z = dim Z.
 2.2. CONFORMAL DIMENSION 23

Examples 2.2.3. (1) For each n ≥ 1, Euclidean space RN is minimal

for conformal dimension. Indeed, Cdim RN = dim RN = N . More gener-
ally, any nonfractal metric space (i.e., a metric space whose Hausdorff and
topological dimensions agree) is minimal for conformal dimension, for purely
topological reasons.
(2) The Cantor set K is not minimal for conformal dimension. Indeed,
dim K = log 2
log 3 but Cdim K = 0. To prove the latter result, consider a
one-parameter family of Cantor sets Kλ ⊂ [0, 1], 0 < λ < 12 , where Kλ
consists of two subsets, one of which contains 0 and the other contains 1,
so that each subset is similar to K by scale factor λ. Note that K = K1/3 .
It is easy to verify that Kλ is doubling, uniformly perfect and uniformly
disconnected. By Theorem 2.1.2, Kλ is quasisymmetrically equivalent to K
for each λ > 0. (It is also straightforward to verify directly that the canonical
monotone increasing homeomorphism K → Kλ is quasisymmetric.) Thus
log 2
Cdim K ≤ log 1/λ for each λ > 0 and hence Cdim K = 0. Observe that
this infimum is not achieved; every quasisymmetric image of K is uniformly
perfect and hence has positive Hausdorff dimension.
In fact, there are no metric spaces with conformal dimension strictly
between zero and one. More precisely, the range of conformal dimension
is equal to {0} ∪ [1, ∞]. The nonexistence of metric spaces with conformal
dimension in the interval (0, 1) is due to Kovalev and is one of the major
recent advances in the subject. We will discuss Kovalev’s theorem further
in Chapters 6 and 8.
Conformal dimension is obviously monotonic, but (in contrast with Haus-
dorff dimension) it is not countably stable. In fact, it is not even finitely
stable: there exist sets A, B ⊂ [0, 1] of conformal dimension zero such that
A ∪ B = [0, 1]! See Example 5.2.1. In section 5.2 we prove that finite
unions of disjoint compact sets have conformal dimension no larger than the
maximum of the conformal dimensions of the pieces. See Proposition 5.2.3.
We observe the following modest corollary: the conformal dimension of a
compact space Z is not increased by adjoining a finite number of points to Z.
Variant notions are occasionally more useful. Similar invariants can be
built from any of the standard metric notions of dimension: Minkowski
dimension, packing dimension, etc. One particularly useful variant is con-
formal Assouad dimension.
Definition 2.2.4. The conformal Assouad dimension, CdimA G, of a
conformal gauge G is the infimum of the Assouad dimensions of all metric
spaces in G.
Note that CdimA G is finite if and only if G is doubling.
Definition 2.2.5. The Ahlfors regular conformal dimension, ARCdim G,
of a conformal gauge G is the infimum of the dimensions of all Ahlfors regular
metric spaces in G.
24 2. CONFORMAL GAUGES AND CONFORMAL DIMENSION

Note that a conformal gauge may contain both Ahlfors regular and non-
Ahlfors regular metrics. However, every complete, doubling, uniformly per-
fect gauge contains a large supply of Ahlfors regular metrics. We will justify
the latter assertion in chapter 7.
We record two elementary inequalities which follow from (1.4.7) and
(1.4.8):
CdimH G ≤ CdimA G ≤ ARCdim G
for every conformal gauge G. In fact, the second inequality is an equality
for a large class of gauges.
Proposition 2.2.6. Let G be a complete, uniformly perfect conformal
gauge. Then ARCdim G = CdimA G. Indeed, for any s > CdimA G there
exists a closed Ahlfors s-regular set Y contained in some Euclidean space
RN so that Y ∈ G.

2.3. Notes
Conformal dimension was introduced by Pansu [133] in connection with
the quasi-isometric classification of rank one symmetric spaces. Bourdon
[27], [29] extended the method to treat lattices in some hyperbolic buildings.
We will review the results of Pansu and Bourdon in chapter 3 and discuss
ideas from the proofs of those results in chapter 4.
The language of conformal gauges was suggested by Heinonen [76, Chap-
ter 15] as a formalism for the study of quasisymmetrically invariant prop-
erties. Theorem 2.1.2 is due to David and Semmes [47, Proposition 15.11].
The quasisymmetric characterization of S1 is well-known [166, Theorem
4.9]: A topological circle (Z, d) is quasisymmetrically equivalent to S1 if and
only if it is doubling and has bounded turning.
Definition 2.3.1. A connected, locally connected metric space (Z, d) is
said to have bounded turning if there exists C < ∞ so that any two points
z0 , z1 ∈ Z are included in a connected set A ⊂ Z with diam A ≤ Cd(z0 , z1 ).
A topological circle Z has bounded turning if and only if the following
condition holds for some constant C < ∞: for all a, b ∈ Z,
min{diam γ1 , diam γ2 } ≤ Cd(a, b)
where γ1 and γ2 are the two maximal subarcs of Z \ {a, b}. A planar topo-
logical circle has bounded turning if and only if it is a quasicircle, i.e., it is
the image of S1 under a quasiconformal map C → C [14].
Proposition 2.2.6 was proved by Juha Heinonen in 1996 during a grad-
uate topics course at the University of Michigan and first appeared in print
in [76, Chapter 14]. We will sketch the proof in chapter 7.
 CHAPTER 3

Gromov hyperbolic groups and spaces and their

boundaries

The original context for the introduction of conformal dimension was

rigidity theory for lattices in rank one symmetric spaces and hyperbolic
buildings. In this chapter, we review the basics of Gromov hyperbolic ge-
ometry and discuss various examples (rank one symmetric spaces, Gromov
hyperbolic groups, and hyperbolic buildings). The boundaries of rank one
symmetric spaces modeled on the complex or quaternionic hyperbolic spaces
are canonically equipped with sub-Riemannian structures. Sub-Riemannian
stratified Lie groups (Carnot groups) have played a critical role in the devel-
opment of analysis and geometry in metric measure spaces. In the following
chapter, we indicate the main tools used to prove lower bounds on confor-
mal dimension, and illustrate these tools by showing that the various spaces
discussed in this chapter are minimal for conformal dimension.

3.1. The real hyperbolic plane and CAT(κ) spaces

We denote by HR2 the real hyperbolic plane, i.e., the unique simply con-
nected Riemannian manifold of dimension two with constant sectional cur-
vatures equal to −1. For an explicit model of HR2 we can use the Poincaré
disc D ⊂ C equipped with the hyperbolic metric ds = 2(1 − |z|2 )−1 |dz|. Let
ρ be the associated distance function. An easy computation shows that the
hyperbolic geodesic joining the origin to a point z = r ∈ (0, 1) is the  Eu-
clidean line segment [0, r]; the length of this geodesic is ρ(0, r) = log 1+r
1−r .
The hyperbolic metric is invariant under the group of automorphisms of D,
i.e., the group of Möbius transformations z
→ eiθ 1−az
z−a
, a ∈ D. Thus
⎛ ⎞
  z−w
z−w 1 + 1−wz

ρ(z, w) = ρ 0, = log ⎝ ⎠ z, w ∈ D.
1 − wz z−w ,
1 − 1−wz

By scaling the metric, we construct manifolds with any prescribed constant

negative sectional curvature. We will denote by HR2 (κ) the unique simply
connected Riemannian 2-manifold with constant sectional curvatures equal
to κ < 0. We write ρκ for the corresponding distance function.
Definition 3.1.1. Let (X, d) be a geodesic metric space. A geodesic
triangle Δ in X is the union of three geodesics γab ∪ γac ∪ γbc , a, b, c ∈ X.
25
26 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

Let Δ be a geodesic triangle in X and let κ < 0. A comparison triangle

for Δ in HR2 (κ) is a geodesic triangle Δ = γab ∪ γac ∪ γbc ⊂ HR2 (κ) so that
length(γxy ) = length(γxy ) whenever x, y ∈ {a, b, c}.
Since any three-point metric space can be isometrically embedded in
HR2 (κ) for any κ < 0, comparison triangles always exist. Moreover, Δ and
Δ are isometric. If Δ is a comparison triangle for Δ and u ∈ Δ, we denote
by u the unique point in Δ which corresponds to u under the isometry.
We are now ready to state the definition of a CAT(κ) metric space. This
concept is due to Alexandrov.
Definition 3.1.2. A geodesic metric space (X, d) is said to be a CAT(κ)
metric space for some κ < 0 if the following condition holds: whenever u
and v are two points contained in a geodesic triangle Δ ⊂ X, then
d(u, v) ≤ ρκ (u, v).
Of course, (HR2 (κ), ρκ ) itself is a CAT(κ) space. A simply connected Rie-
mannian manifold M (of any dimension) is a CAT(κ) space if and only if all
of the sectional curvatures of M are bounded above by κ. Complete met-
ric spaces which are locally CAT(κ) are also known as spaces with negative
upper bounds on curvature in the sense of Alexandrov.

3.2. Gromov hyperbolic metric spaces

Gromov’s axiomatization of metric hyperbolicity generalizes and extends
the class of spaces with negative upper bounds on curvature in the sense of
Alexandrov. It has been integral to the modern development of synthetic
differential geometry, geometric group theory and analysis in metric spaces.

A metric space (X, d) is proper if it satisfies the conclusion of the Heine–

Borel theorem: all closed and bounded subsets are compact. Let (X, d) be a
proper geodesic metric space. We denote by γab a geodesic segment joining
a, b ∈ X.
Definition 3.2.1. Let δ > 0. A geodesic metric space (X, d) is δ-
hyperbolic if for every geodesic triangle γab ∪ γbc ∪ γac , we have that γab is
contained within the δ-neighborhood of γbc ∪ γac . We say that (X, d) is
(Gromov) hyperbolic if it is δ-hyperbolic for some δ > 0.
Roughly speaking, this definition asks that all triangles in (X, d) be uni-
formly thin. Put another way, the space (X, d) seen from far away behaves
like a metric tree (this philosophy can be made precise with the language of
asymptotic cones).
The basic example of a Gromov hyperbolic metric space is the √ real hy-
2
perbolic plane (HR , ρ), which is δ-hyperbolic with δ = log(1 + 2). More
generally, any (sufficiently smooth) simply connected manifold with all sec-
tional curvatures negative and bounded away from zero is hyperbolic. By
comparison with the model spaces HR2 (κ), one easily sees that the class of
 3.2. GROMOV HYPERBOLIC METRIC SPACES 27

hyperbolic spaces includes all CAT(κ) spaces with κ < 0. The 0-hyperbolic
metric spaces are precisely the metric R-trees: geodesic metric spaces in
which every geodesic triangle is degenerate.
Let us fix a base point o ∈ X and consider the collection R of all geodesic
rays in X starting from o. Thus γ ∈ R if and only if γ : [0, ∞) → X is
an isometry and γ(0) = o. We equip R with a (possibly infinite-valued)
distance function by declaring the distance between two geodesic rays γ, γ 
to be sup{d(γ(t), γ  (t)) : t ≥ 0}. Let ∼ be the equivalence relation on R
defined as follows: γ ∼ γ  if and only if the distance between γ and γ  is
finite.
Definition 3.2.2. The Gromov boundary, ∂∞ X, of a Gromov hyperbolic
space X is R/ ∼.
This boundary is equipped with a natural topology, T∞ , coinciding with
the quotient of the compact-open topology on the function space R. We
can equip X ∪ ∂∞ X with a compact topology which restricts to the metric
topology on X and the boundary topology T∞ on ∂∞ X. This exhibits
X ∪ ∂∞ X as a compactification of X. In this topology, limt→∞ γ(t) = ξ if
ξ ∈ ∂∞ X is represented by γ ∈ R. In fact, any geodesic ray (starting from
any point) in X limits to a unique point in ∂∞ X.
The boundary topology T∞ is metrizable; we next discuss certain canon-
ical metrics on ∂∞ X which recover this topology.
Definition 3.2.3. Let a > 1. We say that a is a visual parameter for
∂∞ X relative to o if there exists a metric da = da,o (a visual metric) on ∂∞ X
for which the following conditions hold:
(1) the topology of (∂∞ X, da ) coincides with T∞ , and
(2) there exists a constant C > 0 so that for any bi-infinite geodesic
ray γ : R → X with limt→∞ γ(t) = ξ+ ∈ ∂∞ X and limt→−∞ γ(t) =
ξ− ∈ ∂∞ X, we have
1 − dist(o,γ)
a ≤ da (ξ+ , ξ− ) ≤ Ca− dist(o,γ) .
C
Theorem 3.2.4. For each δ > 0, there exists a0 (δ) > 1 so that the
Gromov boundaries of δ-hyperbolic metric spaces admit visual metrics with
any parameter 1 < a < a0 (δ).
Any two visual metrics da,o and da,o on ∂∞ X are bi-Lipschitz equivalent.
Any two visual metrics da and da on ∂∞ X are snowflake equivalent with
parameter log a / log a.
For α > 0, we say that two metric spaces (X, d) and (X  , d ) are snowflake
equivalent with parameter α if there exists a homeomorphism f : X → X 
and a constant C > 0 so that
1
d(x1 , x2 )α ≤ d (f (x1 ), f (x2 )) ≤ Cd(x1 , x2 )α
C
for all x1 , x2 ∈ X. In other words, f is bi-Lipschitz as a map from the
snowflaked space (X, dα ) to (X  , d ) (if α ≤ 1) or f is bi-Lipschitz as a
28 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

map from (X, d) to the snowflaked space (X  , (d )1/α ) (if α ≥ 1). Note that
snowflake equivalence is a stronger condition than quasisymmetric equiva-
lence.
3.2.1. Construction of visual metrics. We briefly review the well-
known construction of visual metrics. The key tool is the Gromov product.
Definition 3.2.5. Let X be a metric space equipped with a fixed base-
point o. The Gromov product of x, y ∈ X with respect to o is
1
(x|y)o := (d(x, o) + d(y, o) − d(x, y)).
2
The Gromov product admits the following simple geometric interpre-
tation in the case X = R2 . Let x, y, o ∈ R2 and let C be the inscribed
circle in the triangle oxy. We write [ab] for the line segment joining a to
b. Then (x|y)o is the common length of the subsegments [oa] ⊂ [ox] and
[ob] ⊂ [oy], where a and b are the points of tangency between C and [ox]
and [oy] respectively.
We are primarily interested in the case when X is Gromov hyperbolic.
When X is δ-hyperbolic, the values of (x|y)o and dist(o, γxy ) are comparable:
|(x|y)o − dist(o, γxy )| ≤ 2δ.
In view of the definition of hyperbolicity, it follows that (x|y)o measures, in
rough terms, the length of time that γox and γoy remain close (within 2δ).
Remark 3.2.6. Hyperbolicity can be characterized in terms of the Gro-
mov product: X is Gromov hyperbolic (in the sense of Definition 3.2.1) if
and only if there exists δ > 0 so that
(3.2.1) (x|y)o ≥ min{(x|z)o , (z|y)o } − δ
for all x, y, z, o ∈ X. In particular, the notion of hyperbolicity (although not
the choice of the constant δ) is independent of the choice of the basepoint.
We extend the Gromov product to the Gromov boundary as follows. For
ξ, η ∈ ∂∞ X, define
(ξ|η)o = sup lim inf (γ(s)|γ  (t))o ,
s,t→∞

where the supremum is taken over all geodesic rays γ, resp. γ  , representing
ξ, resp. η. In case X is CAT (κ) for some κ, the limit exists and is indepen-
dent of the choice of geodesic rays: no supremum is needed. For arbitrary
hyperbolic X, the inequality in (3.2.1) extends to triples x, y, z ∈ X ∪ ∂∞ X,
at the cost of replacing δ with 2δ.
Definition 3.2.7. Let a > 1. The visual premetric on ∂∞ X with pa-
rameter a is Da (ξ, η) := a−(ξ|η)o .
Typically, Da is not a metric. However, there exists K < ∞ so that
(3.2.2) Da (ξ, η) ≤ K max{Da (ξ, ζ), Da (ζ, η)}
for all ξ, ζ, η ∈ ∂∞ X. In fact, we may take K = aδ by (3.2.1).
 3.2. GROMOV HYPERBOLIC METRIC SPACES 29

Definition 3.2.8. Let Z be a set. A symmetric function D : Z × Z →

[0, ∞) is called a quasimetric if it vanishes only on the diagonal of Z × Z
and satisfies the quasi-triangle inequality
(3.2.3) D(x, y) ≤ K(D(x, z) + D(z, y))
for all x, y, z ∈ Z. We call K a quasimetric constant for (Z, D).
Note that if D satisfies (3.2.3) for some constant K, then it satisfies
(3.2.2) with constant 2K. Conversely, any D satisfying (3.2.2) also satisfies
(3.2.3) for the same value of K.
Recall the snowflaking construction: for any metric space (Z, d), the
space (Z, d ) is also metric if 0 <  < 1. It is easy to see that (Z, d ) is
a quasimetric space for any  > 0. The following theorem of Macı́as and
Segovia provides a sort of converse to this assertion.
Theorem 3.2.9 (Macı́as–Segovia). Let (Z, D) be a quasimetric space
with quasimetric constant K. Then there exist  > 0, a metric d = d on
Z, and a finite constant C so that d(x, y)/C ≤ D(x, y) ≤ Cd(x, y) for all
x, y ∈ Z. The data  and C depend only on K.
The proof of Theorem 3.2.9 uses a standard chaining  construction. For
 > 0, define d (z, w) to be the infimum of sums M j=1 D(x 
j−1 , xj ) over
all finite chains of points {z = z0 , z1 , . . . , zM = w} in Z. When z = w we
admit the trivial chain (M = 0) for which the corresponding sum is zero.
It is clear that d is symmetric, satisfies the triangle inequality, vanishes on
the diagonal of Z × Z and satisfies d ≤ D  . An induction argument using
(3.2.2) shows that d ≥ (2K)−2 D  provided  > 0 satisfies (2K)2 ≤ 2. This
completes the proof.
Applying Theorem 3.2.9 to the Gromov boundary ∂∞ X equipped with
its canonical family of quasimetrics Da (see Definition 3.2.7) generates com-
parable metrics da provided a is sufficiently close to one. To summarize:
There exists a0 > 1 so that each 1 < a < a0 is a visual parameter for ∂∞ X
with associated visual metric da .
Here is an interesting case.
Theorem 3.2.10 (Bourdon). If X is CAT(κ) for some κ < 0, then
√
da (ξ, η) := exp{− −κ(ξ|η)o }
defines a metric on ∂∞ X.
√
In other words, we may choose a0 = e− −κ in the preceding paragraph
and no appeal to Theorem 3.2.9 is required.
3.2.2. Quasi-isometries of hyperbolic spaces and the conformal
gauge of the boundary. In light of Theorem 3.2.4, all of the visual metrics
on the boundary of a Gromov hyperbolic space lie in a single conformal gauge
G. We call this gauge the canonical conformal gauge of the boundary of X.
From here on, we always assume that ∂∞ X is equipped with a metric taken
30 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

from this canonical conformal gauge. We will consider hypotheses on ∂∞ X

which are quasisymmetrically invariant (such as the doubling or uniformly
perfect assumptions) without further reference to a specific metric.
In order to make use of the large scale geometry of the ambient hyper-
bolic space, we need the following class of maps.
Definition 3.2.11. Let (X, d) and (X  , d ) be metric spaces. A function
f : X → X  is a quasi-isometry if it is roughly bi-Lipschitz. More precisely,
f is a (λ, C)-quasi-isometry, λ ≥ 1, C ≥ 0, if f is C-roughly onto (the
C-neighborhood of f (X) is equal to X  ) and
(3.2.4) λd(x, y) − C ≤ d (f (x), f (y)) ≤ λd(x, y) + C
for all x, y ∈ X.
Note that quasi-isometries need not be continuous. On small scales, the
only restriction on f imposed by (3.2.4) is local boundedness. On large
scales, (3.2.4) asserts that the map f is bi-Lipschitz. We will identify two
quasi-isometries f : X → X  and g : X → X  if they lie at a bounded
distance from each other, i.e., sup{d (f (x), g(x)) : x ∈ X} is finite.
A key property of quasi-isometries is preservation of hyperbolicity.
Theorem 3.2.12. Suppose X and X  are geodesic metric spaces and
f : X → X  is a (λ, C)-quasi-isometry. If X is δ-hyperbolic, then X  is
δ  -hyperbolic, where δ  depends only on δ, λ, and C.
Quasi-isometric maps of Gromov hyperbolic spaces are in one-to-one
correspondence with quasisymmetric maps on the boundaries. This fact is at
the heart of the applications of geometric function theory in geometric group
theory. In particular, it implies that properties of the boundary conformal
gauge (such as conformal dimension) have geometric consequences for the
associated hyperbolic space.
Theorem 3.2.13. Let X and X  be Gromov hyperbolic spaces with uni-
formly perfect Gromov boundaries ∂∞ X and ∂∞ X  . Then each quasi-isometry
F : X → X  induces a quasisymmetric boundary map F∞ : ∂∞ X → ∂∞ X  .
Conversely, each quasisymmetric map f : ∂∞ X → ∂∞ X  extends to a quasi-
isometry F : X → X  .

Corollary 3.2.14. The conformal dimension of the canonical confor-

mal gauge of the Gromov boundary is a quasi-isometry invariant for Gromov
hyperbolic spaces with uniformly perfect boundaries.

3.3. Boundaries of rank one symmetric spaces

The study of Riemannian symmetric spaces illuminates beautiful and
striking connections between Riemannian geometry and the structure theory
of semi-simple Lie groups. The classical rank one symmetric spaces of non-
compact type provide a unified context for hyperbolic geometry irrespective
 3.3. BOUNDARIES OF RANK ONE SYMMETRIC SPACES 31

of the underlying division algebra. These spaces are Gromov hyperbolic;

their boundaries at infinity are locally modeled on certain stratified Lie
(Carnot) groups. Quasiconformal analysis on those boundaries features in
the original proof of Mostow’s Rigidity Theorem for lattices in the underly-
ing symmetric spaces. The quasi-isometric classification of these spaces and
the nonexistence of metrics with pinched sectional curvature on their (com-
pact) quotients were among the primary motivations for the introduction of
conformal dimension. Deeper rigidity phenomena for the quaternionic and
octonionic (Cayley) hyperbolic spaces arise from a more detailed study of
the tangent structure of quasiconformal maps of the Carnot groups associ-
ated to their boundaries. These results are all due to Pansu. See Theorems
3.3.3, 3.3.7 and 3.6.1.

3.3.1. Symmetric spaces. A Riemannian symmetric space is a con-

nected Riemannian manifold M which admits a local symmetry at each point
p ∈ M , i.e., an isometry σp : M → M satisfying σp (p) = p and dσp (v) = −v
for all v ∈ Tp M . The rank of M is the maximal dimension of an embedded
Euclidean flat (subspace). The rank one symmetric spaces of noncompact
type can be completely classified: they include, for each n ≥ 2, the hyper-
bolic space of dimension n modeled on the real line R, the complex plane
C or the quaternions H, as well as the exceptional Cayley hyperbolic plane.
We restrict our discussion to the real, complex and quaternionic categories
only since the Cayley hyperbolic plane requires a different treatment. (See
the notes to this chapter for references on the Cayley hyperbolic plane.)
Let K ∈ {R, C, H} be one of the classical division algebras. Recall that
the quaternions are the unique four-dimensional algebra over R spanned by
1 and the symbols i, j, k, subject to the rules i2 = j 2 = k 2 = ijk = −1.
Equip Kn with the sesquilinear scalar product

n
(3.3.1) x, y = xj yj
j=1

where x = (x1 , . . . , xn ) and x denotes the conjugate of x ∈ K. Denote by

||x|| = x, x1/2 the corresponding norm.
We review two standard models of hyperbolic space.
Ball model. In this model, we view hyperbolic space HKn as the unit
ball {x ∈ Kn : ||x|| < 1}, which we equip with the metric dh given by the
formula
|1 − x, y|
(3.3.2) cosh dh (x, y) :=   .
1 − ||x||2 · 1 − ||y||2
Parabolic model. In this model, we view HKn as the subset of Kn con-
sisting of points u = (u1 , . . . , un ) = (u , un ) such that

Re(un ) > ||u ||2 ,

32 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

where Re(u) = 12 (u + u). When K = C, this domain is known as the Siegel

upper half space. In this model, the metric takes the form
| 12 (un + vn ) − u , v  |
cosh dh (u, v) =   .
Re(un ) − ||u ||2 · Re(vn ) − ||v  ||2
The equivalence between these two models is effected by the map x
→ u
given by
(3.3.3) u = x (1 + xn )−1 and un = (1 − xn )(1 + xn )−1
known as the Cayley transform. (Note that we use right scalar multiplication
in Kn .)
3.3.2. Gromov hyperbolicity of HKn . All of the hyperbolic spaces
HKnare Gromov hyperbolic, in fact, they are CAT(−1). The standard visual
metric on the boundary can be computed explicitly.
Proposition 3.3.1. The metric space (HKn , dh ) is CAT(−1) and hence
Gromov hyperbolic. The maximal visual parameter a0 is equal to e. The
boundary at infinity, ∂∞ HKn , coincides (as a set) with the sphere Skn−1 ,
where k = dimR K. The visual metric on ∂∞ HKn corresponding to the visual
parameter a0 = e is

|1 − ξ, η|
(3.3.4) da0 (ξ, η) = ,
2
where ·, · denotes the scalar product (3.3.1) in Kn .
Proof of (3.3.4). The metric dh on the hyperbolic space HKn is the
global distance function associated to a smooth Riemannian metric all of
whose sectional curvatures are bounded between −4 and −1 (when K = R,
all sectional curvatures are equal to −1). By comparison with HR2 , it follows
that (HKn , dh ) is CAT(−1). In fact, every nondegenerate geodesic triangle
in HKn is contained in an isometrically embedded copy of either HR2 or its
scaled version with constant curvature −4.
Let o denote the origin in the ball model of HKn . Using √(3.3.2) for
points x, y√∈ HKn together with the identity arccosh(r) = log(r + r2 − 1) =
− log(r − r2 − 1), we compute
1
(dh (x, o) + dh (y, o) − dh (x, y))
2

1/2
(1 + ||x||)(1 + ||y||)
= log  ;
|1 − x, y| + |1 − x, y|2 − (1 − ||x||2 )(1 − ||y||2 )

taking the limit as x → ξ and y → η gives (ξ|η)o = log |1−ξ,η| 2
. Then
da0 (ξ, η) = exp(−(ξ|η)o ) has the desired form. 
The metric in (3.3.4) is known as the Cygan metric. We will denote it
by dCygan .
 3.3. BOUNDARIES OF RANK ONE SYMMETRIC SPACES 33

3.3.3. Sub-Riemannian structure on the boundary of hyper-

bolic space. We continue to work with the ball model for the hyperbolic
space HKn . In this subsection we assume that n ≥ 2. The (real) tangent
space Tξ Skn−1 to the boundary at a point ξ is given by the set of vectors
u ∈ Kn for which Reξ, u = 0. We compute the Cygan distance from ξ
to the point η obtained at time s by the Riemannian exponential map at ξ
with initial velocity u:
1
η= (ξ + su).
||ξ + su||
An easy computation gives

1 + s Imξ, u + o(|s|), if Imξ, u = 0,
ξ, η =
1 − 12 s2 ||u||2 + o(s2 ), else.
Thus
⎧
⎨ Imξ,u 1/2
2 |s| + o(|s|1/2 ), if Imξ, u = 0,
da (ξ, η) =
⎩1
2 ||u|||s| + o(|s|), else.
This elementary computation illustrates a fundamental feature of the Cy-
gan metric on the boundary of complex and quaternionic hyperbolic space:
anisotropy. Distance computations in the direction of tangent vectors lying
in the horizontal tangent space
Hξ Skn−1 = {u ∈ Tξ Skn−1 : ξ, u = 0}
differ from those in the direction of tangent vectors which lie outside the
horizontal tangent space. The horizontal tangent bundle HSkn−1 is a con-
stant rank subbundle of the tangent bundle T Skn−1 of real codimension
k − 1. Tensoring with K yields certain canonical integrable subbundles.
For instance, in the case K = C both the holomorphic tangent bundle
T 1,0 S2n−1 = T 1,0 Cn ∩ (T S2n−1 ⊗ C) and its conjugate bundle T 0,1 S2n−1 are
integrable. (Recall that T 1,0 Cn denotes the subbundle of T Cn ⊗C consisting
of complexified tangent vectors of type (1, 0).) However, HSkn−1 itself is not
integrable, in fact, it is completely nonintegrable: the Lie span of horizon-
tal vector fields coincides with the full tangent bundle. More precisely, any
vector field on Skn−1 can be obtained as a linear combination of horizontal
vector fields and their (single) Lie brackets. We say that the distribution
HSkn−1 defines a sub-Riemannian structure of step two on Skn−1 .
According to Chow’s theorem, any pair of points ξ, η ∈ Skn−1 can be
joined by a (piecewise) smooth curve whose tangent vectors lie in the hor-
izontal distribution. We introduce a sub-Riemannian metric on Skn−1 by
fixing a smoothly varying family of inner products ·, · on the horizontal
tangent spaces and defining the length of a curve γ : [a, b] → Skn−1 in the
usual fashion:  b
γ  (s), γ  (s)γ(s) ds.
1/2
lengthcc (γ) =
a
34 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

Then the Carnot-Carathéodory metric dcc on Skn−1 is defined by

dcc (ξ, η) := inf lengthcc (γ),
where the infimum is taken over all piecewise smooth horizontal curves γ
joining ξ and η. The CC metric dcc and the visual (Cygan) metric dCygan
are comparable.
Proposition 3.3.2. The Hausdorff dimension of (Skn−1 , dcc ) is equal to
kn + k − 2 = kn − k + 2(k − 1).
Essentially, this boils down to the observation that the k − 1 nonhori-
zontal directions behave, from a metric point of view, like two-dimensional
subspaces. Note that this metric dimension exceeds the topological dimen-
sion kn − 1 in the case K = R.
Are any of the spaces HKn quasi-isometric? By Theorem 3.2.13, it is
equivalent to ask: are any of the spaces (∂∞ HKn , dcc ) quasisymmetric? Let
us note that (Hausdorff) dimension is not sufficient to answer this question,
since quasisymmetric maps can change Hausdorff dimension. Instead, we
consider conformal dimension.
Theorem 3.3.3 (Pansu). For each n ≥ 2 and K, Z = (∂∞ HKn , dcc ) is
minimal for conformal dimension:
Cdim Z = dim Z = kn + k − 2.
It is clear that the pair of integers (dimT Z, Cdim Z) = (kn−1, kn+k−2)
consisting of the topological and conformal dimensions of Z completely char-
acterizes Z among the spaces (∂∞ HKn , dcc ), where n ≥ 2 and K is either the
real, complex or quaternionic field. Since both of these values are quasisym-
metry invariants, all of these spaces are quasisymmetrically distinct. Thus
all of the hyperbolic spaces HKn are quasi-isometrically distinct.

3.3.4. Sub-Riemannian structure on the boundary of the pa-

raboloid model. The sub-Riemannian structure described in the preceding
paragraph can be transferred to the paraboloid model of HKn via the Cayley
transform. The boundary of this model consists of points u ∈ Kn satisfying
Re(un ) = ||u ||2 . Via the Cayley transform, it is equivalent to the sphere
Skn−1 with a single point deleted (namely, (0, −1)).
We coordinatize the hypersurface Σ = {Re(un ) = ||u ||2 }, parameteriz-
ing Σ by Kn−1 × Im(K) via the map
(u , v)
→ (u , ||u ||2 + v).
Here Im(x) = x − Re(x) for x ∈ K and Im(K) = {Im(x) : x ∈ K}.
We can equip Kn−1 × Im(K) with a nilpotent group law which captures
the boundary behavior of the automorphisms of HKn . For p = (x, t) ∈
Kn−1 × Im(K), define τp : HKn → HKn by
τp (u) := (u + x, un + t + 2u , x + ||x||2 ), u = (u , un ).
 3.3. BOUNDARIES OF RANK ONE SYMMETRIC SPACES 35

This map extends naturally to a self-map of Σ. The collection of such maps

{τp : p ∈ Kn−1 × Im(K)} acts simply transitively and effectively on Σ. We
identify Kn−1 × Im(K) with this collection via the map p ↔ τp , thereby
defining a group law on Kn−1 × Im(K) which induces a group action. In
order to ensure that the sub-Riemannian structure on the resulting space
(which we will shortly define) is left-invariant, we require that this action
be a right group action. Thus, for p, q ∈ Kn−1 × Im(K), we define p ◦ q by
the identity τq ◦ τp = τp◦q . A simple computation yields the explicit formula
(3.3.5) p ◦ q = (x, t) ◦ (y, s) = (x + y, t + s + 2 Imx, y).
Example 3.3.4. Let K = R. The model space is Rn−1 , equipped with
the abelian (Euclidean) group law.
Example 3.3.5. Let K = C. The model space is Cn−1 ×Im(C) which we
identify with Cn−1 × R. Equipped with the group law in the preceding para-
graph, this space is the (n − 1)st Heisenberg group. It is a sub-Riemannian
stratified Lie (Carnot) group of step two, with a (2n − 2)-dimensional hori-
zontal space and one dimensional center.
Example 3.3.6. Let K be the quaternions. The model space is Kn−1 ×
Im(K) which we identify with Kn−1 × R3 . Equipped with the group law in
the preceding paragraph, this space is the (n − 1)st quaternionic Heisenberg
group. It is a sub-Riemannian stratified Lie (Carnot) group of step two, with
a (4n − 4)-dimensional horizontal space and a three-dimensional center.
The horizontal tangent subbundle HSkn−1 can be transported to the
boundary of the paraboloid model via the Cayley transform. It induces a
horizontal subbundle in the tangent bundle of the nilpotent stratified Lie
group G = Kn−1 × Im(K). We may equip each of these groups with a sub-
Riemannian (Carnot-Carathéodory) metric dcc by the procedure described
in section 3.3.3. Due to the left invariance of the subbundle HG, this metric
is left invariant.
The analog of Theorem 3.3.3 holds for the model spaces G = Kn−1 ×
Im(K) in Examples 3.3.5 and 3.3.6, i.e., the Heisenberg groups and their
quaternionic analogs, equipped with Carnot-Carathéodory metrics. In fact,
all nilpotent stratified Lie (Carnot) groups are minimal for conformal di-
mension. We next recall the definition of Carnot groups. We indicate why
such spaces are minimal for conformal dimension in Example 4.2.3.
3.3.5. Carnot groups. Carnot groups are simply connected nilpo-
tent Lie groups equipped with a stratified Lie algebra and a left invari-
ant sub-Riemannian metric. Examples include the Heisenberg and quater-
nionic Heisenberg groups with their Lie algebras and associated Carnot-
Carathéodory metrics.
Let G be a simply connected nilpotent Lie group of dimension N with
Lie algebra g. We fix a vector space decomposition
g = v 1 ⊕ · · · ⊕ vs .
36 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

The stratification assumption on g asserts that this decomposition interacts

well with the Lie bracket: [v1 , vj ] = vj+1 for all j = 1, . . . , s, where we make
the convention vs+1 = {0}. In the usual manner, we identify elements of
g as either tangent vectors at the identity element of G or as left invariant
vector fields on G.
We distinguish the first layer v1 as the horizontal subspace of g. Sec-
tions of v1 are horizontal vector fields. The set of values at a point p ∈ G
of all horizontal vector fields is the horizontal subspace Hp G of the tan-
gent space Tp G. The left invariant bundle HG, called the horizontal bundle
defines the infinitesimally accessible directions for motion in G. The Carnot-
Carathéodory metric (see the following paragraph) quantifies distance esti-
mates in the horizontal directions.
Fix an inner product in the layer v1 and define a Carnot-Carathéodory
metric dcc in G as was previously done in the case of ∂∞ HKn . The stratifi-
cation assumption implies that the distribution HG is completely noninte-
grable. Chow’s theorem, asserts that dcc (p, q) is finite for all p, q ∈ G. The
resulting space is a sub-Riemannian manifold of step s. Topologically, it
coincides with the Euclidean space RN . However, it is quite different from
a metric point of view. For example, the Hausdorff dimension of G in the
metric dcc is equal to the homogeneous dimension

s
Q= j dim vj .
j=1

We will show that (G, dcc ) is in fact Ahlfors Q-regular. To this end,
introduce the (anisotropic) dilations δr : G → G, r > 0, by the formula
(3.3.6) δr (exp(X1 + · · · + Xs ) = exp(rX1 + · · · + rs Xs ),
where we have written an arbitrary element of G using the exponential
map exp : g → G in terms of a graded decomposition in the Lie algebra
g = v1 ⊕· · ·⊕vs . These dilations are similarities of the Carnot-Carathéodory
metric dcc in the following sense: dcc (δr (p), δr (q)) = rdcc (p, q) for all p, q ∈ G.
Computing the Jacobian determinant reveals that
μ(Bcc (p, r)) = rQ μ(Bcc (p, 1))
for any p ∈ G and r > 0, where μ denotes the Haar measure in G (which
coincides with the image of the Lebesgue measure from the Lie algebra under
the exponential map). Furthermore, since the distribution and the Carnot-
Carathéodory metric are left invariant (by construction), left translations
are isometries of dcc . Since any metric ball Bcc (p, r) ⊂ G can be written as
the image of the unit ball Bcc (o, 1) centered at the identity element o ∈ G
by a left translation and a dilation with scale factor r, we see that
μ(Bcc (p, r)) = CrQ
with C = μ(Bcc (o, 1)).
 3.3. BOUNDARIES OF RANK ONE SYMMETRIC SPACES 37

Note that Q > N = j dim vj whenever G is nonabelian, i.e., s > 1. For
instance, the nth Heisenberg group has homogeneous dimension Q = 2n + 2
and topological dimension N = 2n+1, while the nth quaternionic Heisenberg
group has homogeneous dimension Q = 4n + 6 and topological dimension
N = 4n + 3. The Ahlfors regularity also implies that every nonempty
open subset of a Carnot group G has Hausdorff dimension equal to the
homogeneous dimension Q, when computed in the Carnot-Carathéodory
metric.
The Euclidean spaces, Heisenberg groups and quaternionic Heisenberg
groups (together with one exceptional space, the first octonionic Heisenberg
group; see the notes to this section) are the only Carnot groups which arise
as local models for the boundaries at infinity of rank one symmetric spaces.
Collectively, these spaces are known as Iwasawa groups.
3.3.6. Conformal dimension and volume entropy. In this section
we describe an application of Theorem 3.3.3 to the nonexistence of metrics of
pinched sectional curvature on rank one symmetric spaces and their compact
quotients. This result has been proved in several different ways, by Gromov
and also by Pansu. The argument which we show below, using conformal
dimension, is due to Pansu and was one of the primary motivations for the
introduction of the concept.
As mentioned in the proof of Proposition 3.3.1, the metric (3.3.2) on HKn
has constant sectional curvature −1 if K = R, and has sectional curvatures
bounded between −4 and −1 if K = R. How close can a metric on HKn ,
or quotients thereof, come to constant (negative) sectional curvature? We
say that a metric g on a Riemannian manifold M has (a, b)-pinched sec-
tional curvature if a ≤ Kg ≤ b, where Kg denotes the Riemannian sectional
curvature on (M, g).
Theorem 3.3.7 (Gromov, Pansu). Let M be a compact Riemannian
manifold given as the quotient of HKn , K = R, by a group of isometries Γ
acting discretely and properly discontinuously. If
nk + k − 2
1≤a< , k = dimR K,
nk − 1
then M admits no metric with (−a2 , −1)-pinched sectional curvature.
In order to explain the proof, we recall the notion of volume entropy,
which captures the asymptotic growth rate of the volumes of balls in an
open manifold.
Definition 3.3.8. Let (M, g) be an open Riemannian manifold, and let
dg be the associated distance function on M . The volume entropy of (M, g)
is the quantity
log Volg Bg (o, R)
h(M, g) := lim sup ,
R→∞ R
where o is a base point in M , Volg denotes the volume element on (M, g),
and Bg (p, r) is the metric ball in (M, dg ) with center p and radius r.
38 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

The quantity h(M, g) is independent of the choice of the base point.

Lower bounds on Ricci curvature lead to volume growth estimates and
consequently to estimates for the volume entropy. We recall a standard
volume comparison theorem from Riemannian geometry.
Theorem 3.3.9 (Volume Comparison Theorem). Let (M, g) be a Rie-
mannian N -manifold with Ricci curvature bounded below by −(N − 1)a2 for
some a > 0. Then the volume of any ball of radius R in M is bounded above
by the volume of a ball of equal radius in the constant curvature space form
(HR2 (−a2 ), ρ−a2 ) with constant Ricci curvature −(N − 1)a2 . More precisely,
  
sinh(a|x|) N −1
Volg Bg (o, R) ≤ Volg̃ Bg̃ (o, R) = dx,
B(0,R) a
where B(0, R) denotes the Euclidean ball of radius R in RN .
For the rest of this section, assume that M is a compact manifold given
as the quotient of a rank one symmetric space HKn by a group of isometries
Γ which acts discretely and properly discontinuously, and that g is a Rie-
mannian metric on M . Then g lifts to a Γ-equivariant metric g̃ on M̃ = HKn .
Since M is compact, g̃ induces a distance function dg̃ which is bi-Lipschitz
equivalent to the standard metric dh on HKn , see (3.3.2).
Corollary 3.3.10. Let M be a compact Riemannian N -manifold with
sectional curvatures K ≥ −a2 , and given as the quotient of HKn by a discrete
and properly discontinuous group of isometries. Then h(M̃ , g̃) ≤ (N − 1)a.
Proof. The lifted metric g̃ on M̃ admits the same sectional curvature
bounds. Theorem 3.3.9 applies, giving Volg̃ Bdg̃ (o, R) ≤ C(N, a)e(N −1)aR .

We now discuss Theorem 3.3.7. Suppose that g is a Riemannian metric
on a compact manifold M , with (−a2 , −1)-pinched sectional curvature. As
discussed above, the lifted metric g̃ is bi-Lipschitz equivalent to the standard
metric on M̃ = HKn , whence (M̃ , g̃) is also Gromov hyperbolic with Gromov
metric on the boundary in the canonical conformal gauge of the boundary
at infinity of the standard rank one symmetric space HKn . See, for instance,
Theorems 3.2.12 and 3.2.13.
Theorem 3.3.3 tells us that
Cdim ∂∞ (M̃ , dg̃ ) = nk + k − 2.
By Corollary 3.3.10, we have h(M̃ , g̃) ≤ (nk−1)a. We deduce Theorem 3.3.7
by applying the following proposition to the universal cover M̃ equipped with
the metric g̃.
Proposition 3.3.11 (Pansu). Let (M, g) be a Riemannian manifold with
sectional curvatures K ≤ −1. Then Cdim ∂∞ (M, dg ) ≤ h(M, g).
 3.3. BOUNDARIES OF RANK ONE SYMMETRIC SPACES 39

Proof. The metric space (M, dg ) is CAT(−1), hence by Bourdon’s the-

orem 3.2.10, d(ξ, η) = exp(−(ξ|η)o ) is a metric in the canonical conformal
gauge of ∂∞ M . Thus

Cdim ∂∞ M ≤ dim(∂∞ M, d).

Let s > h(M, g); we will show that Hds (∂∞ M ) = 0. Choose s > t > h(M, g).
By the definition of h(M, g), there exists a constant C so that

Volg Bg (o, R) ≤ CetR

for sufficiently large R. Cover Bg (o, R) with balls of radius one; the minimal
cardinality of such a covering (by volume estimates) is at most

C  etR

for some constant C  . Let B(x1 , 1), . . . , B(xN , 1), N ≤ C  etR , be elements
of this covering which cover S(o, R) = ∂Bg (o, R). For each i = 1, . . . , N
we consider the shadow Ai of B(xi , 1) on ∂∞ M . This is defined as follows:
ξ ∈ Ai if and only if the geodesic ray from ξ to o intersects B(xi , 1). We
claim that the diameter of Ai (in the metric d) is bounded above by e1−R .
To see this, note that if ξ, η ∈ Ai and if x, y ∈ B(xi , 1) ∩ S(o, R) are points
so that x lies on the geodesic ray from ξ to o and y lies on the geodesic ray
from η to o, then
(ξ|η)o ≥ lim (a|b)o
a→ξ,b→η
1
= lim (dg (a, o) + dg (b, o) − dg (a, b))
2
a→ξ,b→η
1
= lim (dg (a, x) + dg (x, o) + dg (b, y) + dg (y, o) − dg (a, b))
a→ξ,b→η 2
1
= R + lim (dg (a, x) + dg (b, y) − dg (a, b))
a→ξ,b→η 2
1
≥ R − dg (x, y)
2
≥ R − 1.

Thus d(ξ, η) = exp(−(ξ|η)o ) ≤ e1−R .

Now the sets A1 , . . . , AN cover ∂∞ M (since they are the shadows of the
balls B(x1 , 1), . . . , B(xN , 1), which cover S(o, R)). Thus


N
Hes1−R (∂∞ M ) ≤ (diamd Ai )s ≤ N e(1−R)s ≤ (C  es )e(t−s)R .
i=1

Since s > t, we obtain Hs (∂∞ M ) = 0 by letting R → ∞. This shows that

dim(∂∞ M, d) ≤ s. Letting s decrease to h(M, g) completes the proof. 
40 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

3.4. Gromov hyperbolic groups

Any finitely generated group may be turned into a geometric object by
considering its Cayley graph. The conformal dimension of the boundary
often encodes algebraic features of the group. See for example Theorems
3.4.6 and 3.4.7.
Definition 3.4.1. Suppose G is a finitely generated group with genera-
tors S. The Cayley graph Γ(G, S) of G with respect to S is the (undirected)
graph with vertex set G, and an edge between g and gs, for every g ∈ G
and s ∈ S.
The Cayley graph has a natural geodesic metric given by assigning each
edge to have length one. This induces a word metric dS on G. The choice of
word metric depends on the choice of the generator set S, however, the quasi-
isometry type of the resulting metric space (Γ(G, S), dS ) is independent of
S. It is traditional to denote again by G the Cayley graph equipped with a
word metric.
Definition 3.4.2. We say that a finitely generated group G is Gromov
hyperbolic if its Cayley graph Γ(G, S) is Gromov hyperbolic, for some finite
set of generators S.
In light of Theorem 3.2.12, Gromov hyperbolicity is independent of the
choice of generating set.
The fundamental lemma of geometric group theory is the following ob-
servation:
Lemma 3.4.3 (Švarc, Milnor). Suppose a group G acts properly discon-
tinuously, cocompactly and isometrically on a proper, geodesic metric space
X. Then G is finitely generated and (G, dS ) is quasi-isometric to X for any
finite generating set S.
In particular, if M is a compact Riemannian manifold, then its fun-
damental group π1 (M ) is quasi-isometric to its universal cover M̃ . For
example, if M is a closed hyperbolic manifold, i.e. its universal cover is
HRn , then π1 (M ) is Gromov hyperbolic. More generally, the fundamental
group π1 (M ) of any compact, negatively curved, Riemannian manifold M
is Gromov hyperbolic.
As a consequence of Theorem 3.2.13, the boundary ∂∞ G = ∂∞ Γ(G, S)
of a Gromov hyperbolic group G is canonically defined with a canonical
conformal gauge. Moreover, the conformal dimension of ∂∞ G is a quasi-
isometry invariant of G. The following theorem assures us that Cdim ∂∞ G
is finite.
Theorem 3.4.4 (Bonk–Schramm). Let G be a Gromov hyperbolic group.
Then the canonical conformal gauge on ∂∞ G is doubling.
The group G is elementary if it is finite or if ∂∞ G has at most two
points, otherwise, it is non-elementary. The canonical conformal gauge on
the boundary of any non-elementary group is uniformly perfect.
 3.4. GROMOV HYPERBOLIC GROUPS 41

Hyperbolic group boundaries admit a rough type of self-similarity which

allows one to deduce quantitative conclusions from qualitative topological
properties. The point is that elements of the group G act isometrically on
the Cayley graph by left translation; by Theorem 3.2.13 this action extends
to an action by quasisymmetries on ∂∞ G. However, the correct notion to
consider is that of quasi-Möbius maps (recall Definition 1.2.20).
We say that a family of maps {fα : Z → Z} is uniformly quasi-Möbius
if there exists a distortion function η so that each fα is η-quasi-Möbius.
Theorem 3.4.5 (Paulin, Bowditch). The induced action of G on ∂∞ G
is by uniformly quasi-Möbius maps, and is discrete and cocompact on the
space of distinct triples of points.
Cocompactness on triples means that there exists some δ > 0 so that
every triple (z1 , z2 , z3 ) of distinct points in ∂∞ G can be mapped by a group
element to a δ-separated triple (all pairwise distances exceed δ). Theo-
rem 3.4.5 implies that, for some choice of visual metric on ∂∞ G, there exists
a constant δ > 0 and a distortion function η so that any ball B in ∂∞ G
is η-quasi-Möbius equivalent to an open set U ⊂ ∂∞ G of diameter at least
δ. In fact, up to a rescaling, one can show that B and U are bi-Lipschitz
equivalent with uniform distortion function. This is the notion of rough
self-similarity alluded to above.
We would like to understand the relationship between the algebraic prop-
erties of G and the geometric properties of the canonical conformal gauge
of ∂∞ G, in particular, its conformal dimension.
Certain low dimensional cases are well understood. For example, when G
is a non-cyclic free group with basis S, its Cayley graph Γ(G, S) is a regular
tree of degree (2#S). Its boundary ∂∞ G is a Cantor set, and it is straight-
forward to find a family of visual metrics with arbitrarily small Hausdorff
dimension (compare Example 2.2.3(2)). Thus in this case the conformal
dimension of the boundary is equal to zero. It is much less obvious that
the converse is also true; this fact follows from Stallings’ theorem on groups
with infinitely many ends and Dunwoody’s theorem on the accessibility of
finitely presented groups.
Theorem 3.4.6 (Stallings, Dunwoody). Suppose G is an infinite hyper-
bolic group. Then the following are equivalent:
• Cdim ∂∞ G = 0,
• ∂∞ G is homeomorphic to a Cantor set, or pair of points,
• ∂∞ G is quasisymmetric to an (Ahlfors regular) Cantor set, or pair
of points,
• G is virtually free (i.e. contains a free group as a finite index sub-
group).
Moreover, the conformal dimension of zero is attained if and only if G is
virtually cyclic.
42 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

In fact, to show that G is virtually free, it would suffice to assume that

Cdim ∂∞ G < 1, since this implies that the boundary is totally disconnected.
Therefore as a corollary we can conclude that Cdim ∂∞ G ∈ {0} ∪ [1, ∞) for
any hyperbolic group G. We will see later on that no metric space can have
conformal dimension in the interval (0, 1).
For general hyperbolic groups, Stallings’ theorem allows us to reduce to
the case when the group is one-ended, and the boundary is connected.
Deep results by several authors characterize the case when the boundary
is a topological circle. This extends relatively directly to characterize the
case when the conformal dimension of the boundary is one and achieved by
an Ahlfors regular metric.
Theorem 3.4.7 (Tukia, Casson–Jungreis, Gabai). Suppose G is a non-
elementary hyperbolic group. Then the following are equivalent:
• Cdim ∂∞ G = 1 and is attained by an Ahlfors regular metric,
• the Gromov boundary ∂∞ G is homeomorphic to S1 ,
• the Gromov boundary ∂∞ G is quasisymmetric to S1 ,
• G acts discretely and cocompactly by isometries on real hyperbolic
space HR2 , i.e., G is virtually Fuchsian.
There are examples due to Pansu of hyperbolic groups whose boundaries
have conformal dimension one, but which are not virtually Fuchsian. These
arise by amalgamating hyperbolic surface groups along cyclic subgroups.
One can exhibit a family of quasi-symmetric deformations of the boundary
that lowers the Hausdorff dimension arbitrarily close to one.
It is a major result of Bowditch that such an algebraic operation corre-
sponds to the existence of local cut points in the boundary.
Theorem 3.4.8 (Bowditch). Suppose G is a hyperbolic group with one
end, so that ∂∞ G is not homeomorphic to S1 . Then ∂∞ G is locally con-
nected, has no global cut points, and has local cut points if and only if G
splits over a virtually cyclic subgroup.
There is not yet a complete understanding of the conformal dimension of
boundaries of topological dimension one that have local cut points. On the
other hand, in section 4.4 we will see that boundaries which are connected
and have no local cut points have conformal dimension greater than one.
The absence of local cut points means that the topological type of a one
dimensional boundary is highly restricted.
Theorem 3.4.9 (Kapovich–Kleiner). Suppose G is a hyperbolic group
with one end, and that ∂∞ G has topological dimension one and has no local
cut points. Then ∂∞ G is homeomorphic to either the Sierpiński carpet or
the Menger sponge.
See Example 4.3.1 for more details regarding analysis on the Sierpiński
carpet. For the Menger sponge, see Example 3.5.1.
 3.5. BOUNDARIES OF HYPERBOLIC BUILDINGS 43

The conformal dimension of the boundary of a group can provide infor-

mation about the quasi-isometry type of the group which is not detected
by the topology of the boundary. For example, let G = π1 (M ), where M
is a real hyperbolic manifold of dimension four, and let H = π1 (N ), where
N is a complex hyperbolic manifold of dimension two (i.e., of real dimen-
sion four). Then, by Proposition 3.3.1 both ∂∞ G and ∂∞ H are topological
spheres of real dimension three. However, by Theorem 3.3.3 Cdim ∂∞ G = 3
and Cdim ∂∞ H = 4, so G and H cannot be quasi-isometric.
In section 3.5 we describe a family of groups due to Bourdon and Pajot,
all of whose boundaries are homeomorphic to the Menger sponge. For this
family of examples the conformal dimensions of the boundaries take on a
dense set of values in (1, ∞).
We have seen that topological information suffices to clarify the setting
when the boundary is homeomorphic to S1 , and is inadequate for the set-
ting when the boundary is homeomorphic to S3 . The two dimensional case
remains a major open problem.
Conjecture 3.4.10 (Cannon). Let G be a Gromov hyperbolic group
whose boundary ∂∞ G is homeomorphic to S2 . Then ∂∞ G is quasisymmet-
rically equivalent to S2 and therefore acts discretely and cocompactly by
isometries on real hyperbolic space HR3 .
We will return to this well-known conjecture in Section 7.2.

3.5. Boundaries of hyperbolic buildings

Bourdon and Pajot consider another interesting class of Gromov hyper-
bolic spaces whose boundaries have good analytic properties. These are
hyperbolic (Fuchsian) buildings with prescribed vertex geometry. After the
boundaries of rank one symmetric spaces, this was the next class of spaces
for which the conformal dimension of the boundary was computed explicitly.

Example 3.5.1 (Bourdon, Bourdon–Pajot). Let p ≥ 5 and q ≥ 3 be

integers. We construct a simply connected hyperbolic 2-complex Ipq as fol-
lows. The 2-cells of Ipq will be right-angled hyperbolic p-gons. Each edge
is assumed to lie on the boundary of q distinct polygons. Finally, the link
at each vertex is isomorphic to the complete bipartite graph Kqq on q + q
vertices. There is a unique hyperbolic building Ipq satisfying these assump-
tions. Equipped with a path metric in which the polygons have the metric
induced from the hyperbolic plane, Ipq becomes a Gromov hyperbolic space.
Its boundary ∂∞ Ipq consists of uncountably many copies of S1 (correspond-
ing to copies of the hyperbolic plane isometrically embedded in Ipq ), glued
together at pairs of points (corresponding to bi-infinite geodesics contained
in pairs of such hyperbolic planes). Topologically, this boundary is home-
omorphic to the Menger sponge M . We recall that M is obtained from
the unit cube in R3 by iterating the following procedure: divide all current
44 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

cubes into 27 subcubes of equal size (with side length equal to 13 the side
length of the current cube) and remove those subcubes which do not touch
an edge of the larger cube. The limiting set M is the Menger sponge; it
can be described as the invariant set for a self-similar iterated function sys-
tem. See section 8.3 for the general theory of iterated function systems and
self-similar fractals.
Each of the defining 2-cells σ for Ipq , together with a portion of ∂σ, can
be identified as a fundamental domain for the action on Ipq of the generalized
reflection group
(3.5.1) Gpq = s1 , . . . , sp : sqi = 1, [si , si+1 ] = 1,
where the relators run over all indices i = 1, . . . , p and indices are taken
mod p. The geometric meaning of the relators in (3.5.1) is the following:
the relator sqi = 1 corresponds to a cyclic permutation of the q cells adja-
cent to a given edge via a rotation by angle 2π/q, while the commutator
relator [si , si+1 ] = 1 encodes the right-angled link geometry. Equipped with
the word metric, Gpq is quasi-isometric to Ipq and hence is also Gromov
hyperbolic, with ∂∞ Ipq = ∂∞ Gpq .
The maximal parameter for the visual metrics da on ∂∞ Ipq is a0 = e.
The Hausdorff dimension of (∂∞ Ipq , da0 ) can be computed in terms of the
growth rate of the volume of combinatorial balls in the Cayley graph of Gpq .
The result is dim ∂∞ Ipq = ττ (p,2)
(p,q)
, where
1
τ (p, q) = lim suplog #{g ∈ Gpq : |g| ≤ m}.
m→∞ m
When q = 2 the space Ipq reduces to the real hyperbolic plane and τ (p, 2) =
arccosh((p − 2)/2) by elementary hyperbolic geometry.
Lemma 3.5.2 (Bourdon). τ (p, q) = τ (p, 2) + log(q − 1).
Corollary 3.5.3. The Hausdorff dimension of the boundary of Ipq ,
dim(∂∞ Ipq , da ), is equal to
log(q − 1)
Qpq := 1 + .
arccosh( p−2
2 )

Theorem 3.5.4 (Bourdon). For each p ≥ 5 and q ≥ 3, Z = (∂∞ Ipq , da )

is minimal for conformal dimension.
Theorem 3.5.4 can be used to distinguish many of the buildings Ipq up
to quasi-isometry. We will observe reasons why (∂∞ Ipq , da ) is minimal for
conformal dimension in chapter 4.
The values Qpq , p ≥ 5, q ≥ 3, satisfy several intriguing number-theoretic
properties. All are irrational, and {Qpq : p ≥ 5, q ≥ 3} is dense in (1, +∞).
The map (p, q)
→ Qpq is not injective, for instance Q5,3 = Q9,5 = Q20,9 =

· · · . In fact, Qpq = Qp q whenever q  − 1 = (q − 1)N and p 2−2 = TN ( p−2
2 ),
where TN denotes the N th Chebyshev polynomial.
 3.6. NOTES 45


p −2
Conjecture 3.5.5. Qpq = Qp q if and only if TN ( p−2
2 ) = TN ( 2 ) and


(q − 1)N = (q  − 1)N for some N, N  ∈ N.

3.6. Notes
Synthetic differential geometry provides an effective language for recast-
ing the machinery of smooth differential geometry into a wide class of metric
spaces. The introduction of this language has been one of the seminal de-
velopments in modern geometry. Signature examples include Alexandrov’s
abstraction of metric curvature bounds and Gromov’s metric characteriza-
tion of hyperbolicity. Suggested references for this material include Burago–
Burago–Ivanov [39], Ballmann [8], Bridson–Haefliger [37] and Gromov [62].
Definition 3.2.1 is the well-known Rips thin triangles condition for hy-
perbolicity. Hyperbolicity of metric spaces admits many equivalent formula-
tions. We refer to [37, Chapter III.H] for an extensive discussion. Theorem
3.2.13 appears in numerous places in the literature. In particular, versions
of this result can be found in Pansu [131], Bourdon [28], Paulin [140] (for
hyperbolic groups), Ghys-de la Harpe [60], and Bonk–Heinonen–Koskela
[20].
For the theory of symmetric spaces, we suggest Helgason [78] and Wolf
[179]. Mostow’s rigidity theorem [127], [128] indicates strong restrictions
on the geometry of negatively curved manifolds in high dimensions. More
information about rigidity of lattices can be found in the survey [67] by
Gromov and Pansu.
An extensive introduction to complex hyperbolic geometry can be found
in Goldman [61], see also the forthcoming book by Parker [139]. For the
role of the Heisenberg group as the boundary of the Siegel upper half space
and the concomitant interactions with several complex variables, harmonic
analysis, and so on, see Stein [161, Chapters XII and XIII].
Our treatment of the geometry of the hyperbolic spaces modeled on
the classical division algebras follows closely that of Bridson and Haefliger
[37]. We have deliberately avoided discussion of the octonionic case, which
requires slightly different ideas. An excellent reference for the geometry and
algebra of the octonions is Baez [7]. For an elegant construction of the
octonionic hyperbolic plane, see Allcock [1].
Sub-Riemannian geometry is a rich subject with connections throughout
mathematics and numerous applications, both within pure mathematics and
to scientific problems. We will not say any more about this subject here, but
refer the reader to the literature for additional details. References include
Montgomery [125], Bellaı̈che [13], and Gromov [66], as well as the recent
survey [42]. A terse but readable account of the role of sub-Riemannian
geometry in Mostow’s rigidity theorem can be found in [74]. Theorem 3.3.3
is implicit in Pansu [133]. An even stronger type of rigidity holds in the
Carnot groups which model the boundaries of quaternionic hyperbolic space
and the octonionic hyperbolic plane.
46 3. GROMOV HYPERBOLIC SPACES AND BOUNDARIES

Theorem 3.6.1 (Pansu). Every quasi-isometry of quaternionic hyper-

bolic space or of the octonionic hyperbolic plane lies at a bounded distance
from an isometry. Equivalently, every quasiconformal map of the quater-
nionic Heisenberg group or the first octonionic Heisenberg group is in fact
conformal.

See [134]. Following this theorem, the question of rigidity for Carnot
groups has received much attention: see Reimann [147], [148], Warhurst
[176], [177], [178], Otazzi [129], Čap et al. [41] and the references therein.
The interplay between the asymptotic geometry of negatively curved
spaces and the conformal geometry of the boundary has been a major theme
in work of Gromov and Pansu. The proof of Theorem 3.3.7 which we have
given is modelled on that in [133], although we have modified the argument
to avoid discussion of Pansu’s generalized (combinatorial) modulus. Other
derivations using capacity estimates or Lp cohomology can be found in [63]
and [131]. We note particularly the following recent result of Pansu [130]
which provides the sharp formulation of Theorem 3.3.7 in the case of the
complex hyperbolic plane: If M is any simply connected Riemannian man-
ifold which is quasi-isometric to HC2 , then the sectional curvatures of M fail
to be (−a2 , −1)-pinched for any a < 2.
The Volume Comparison Theorem 3.3.9 is a well-known result from com-
parative Riemannian geometry which can be found, for instance, in the book
of Bishop and Crittenden [18].
Numerous surveys, monographs and textbooks review the theory of Gro-
mov hyperbolic groups. We especially highlight those of Gromov [64], [65],
Coornaert–Delzant–Papadopoulos [45], Ghys–de la Harpe [60], Kapovich–
Benakli [89] and Kleiner [92]. The Milnor–Švarc Lemma 3.4.3 is a founda-
tional principle of geometric group theory. It guarantees that the large-scale
geometry of a proper geodesic space Z can be identified, within the frame-
work of quasi-isometric equivalence, with that of any (finitely generated)
group acting on Z by isometries.
Theorem 3.4.4 is a consequence of a result of Bonk and Schramm [26,
Theorem 9.2]. By combining this observation with Assouad’s embedding
theorem 1.4.16, Bonk and Schramm prove that all of the classical rank one
symmetric spaces admit roughly similar embeddings into real hyperbolic
space of sufficiently high dimension. Theorem 3.4.5 combines results of
Paulin [140] and Bowditch [36]. Theorem 3.4.6 is due to Stallings [159]
and Dunwoody [49]. See also Theorem 8.1 in [89]. Theorem 3.4.7 combines
work of Tukia [163], [164], Casson–Jungreis [43], and Gabai [55], [56].
Bowditch’s theorem 3.4.8 appears in [35]. Theorem 3.4.9 and the follow-
ing conjecture, which complements Cannon’s conjecture 3.4.10, come from
Kapovich–Kleiner [90].

Conjecture 3.6.2 (Kapovich–Kleiner). Let G be a Gromov hyperbolic

group whose boundary ∂∞ G is homeomorphic to the Sierpiński carpet SC.
 3.6. NOTES 47

Then G acts discretely and cocompactly by isometries on a convex subset

of HR3 with nonempty totally geodesic boundary.
Mineyev and Yu [124] construct a different invariant metric on the Cay-
ley graph of a hyperbolic group, in the quasi-isometry class of the word
metric, so that the group acts by conformal (not quasi-Möbius) maps on the
boundary equipped with the corresponding visual metric. Mineyev [123]
then considers a variant on conformal dimension for boundaries of groups
obtained by infimizing Hausdorff dimension over visual metrics of this type.
This invariant is also able to detect algebraic features of the group.
Bourdon [29], [27], [30] studied the quasi-isometric classification prob-
lem for lattices in several different classes of hyperbolic buildings. We have
presented only one of his examples (Example 3.5.1), which is the most well-
behaved from the perspective of analysis in metric spaces. We will have
more to say about this example in chapter 4 and the notes to chapter 7.
The details of the construction and proofs of the stated results can be found
in [29]. Bourdon also proves a Mostow rigidity-type result for lattices act-
ing on the buildings Ipq , q ≥ 3 [29, Thèoreme 1.3]. A more general class of
hyperbolic buildings in which the same conclusions hold was later studied
by Bourdon and Pajot [32]. Conjecture 3.5.5 was suggested in [168].
For another interesting perspective on the interplay of hyperbolic geom-
etry, rigidity and conformal analysis on the boundary, see the survey article
of Bourdon and Pajot [33].
 CHAPTER 4

Lower bounds for conformal dimension

In this chapter we indicate the primary tools for establishing lower

bounds on conformal dimension and in particular, for identifying spaces
which are minimal for conformal dimension. The basic philosophy is
existence of sufficiently rich curve families in Z  lower
bounds on Cdim Z
The notion of “sufficiently rich curve family Γ” can be formalized in various
ways, using modulus, Poincaré inequalities or, more directly, through the
existence of a sufficiently diffuse measure on Γ. Complications occur due
to the fact that the curve family may only appear “infinitesimally” in Z,
i.e., in some weak tangent space of Z. We discuss tangents of metric spaces
in Chapter 6. With this caveat in mind, all known sufficient conditions
for lower bounds on conformal dimension are instances of the previously
mentioned philosophy.

4.1. Rich curve families and lower bounds for conformal

dimension
In this section, we provide several propositions which formalize the pre-
ceding heuristic. See Propositions 4.1.3, 4.1.8 and 4.1.11. The proofs of
these three propositions are closely related and illustrate the most common
tools and techniques of the theory.
We recall a fundamental covering theorem in metric spaces.
Lemma 4.1.1. Let F be a family of balls in a metric space (Z, d) so
that Z = B∈F B. Assume that the diameters of the balls in F are uni-
formly bounded from above. Then there existsa subfamily G ⊂ F so that the
elements of G are pairwise disjoint and Z = B∈G 5B.
Recall that tB denotes the metric ball concentric with a given ball B
and with radius equal to t times the radius of B.
While quasisymmetric maps are defined solely in terms of the underlying
metrics, most of the methods used to study conformal dimension make use
of adapted measures. A metric measure space (Z, d, μ) is a metric space
(Z, d) equipped with a Borel regular measure μ which is positive and finite
on balls.
Definition 4.1.2. A metric measure space (Z, d, μ) is doubling if there
exists a constant C so that μ(2B) ≤ Cμ(B) for all balls B in Z.
49
50 4. LOWER BOUNDS FOR CONFORMAL DIMENSION

If (Z, d, μ) is a doubling metric measure space with doubling constant

C, then for each λ > 1 there exists C  = C  (λ, C) so that μ(λB) ≤ C  μ(B)
for all balls B in Z. On the other hand, if (Z, d) is Ahlfors Q-regular for
some Q > 0, then (Z, d, HQ ) is doubling. Every doubling metric measure
space (in the sense of Definition 4.1.2) is also a doubling metric space (in
the sense of Definition 1.3.1).
Here is our first criterion for nontrivial lower bounds on conformal di-
mension. In what follows, we denote by p = p/(p − 1) the Hölder conjugate
of an exponent p ∈ [1, ∞).
Proposition 4.1.3. Let (Z, d, μ) be a compact, doubling metric measure
space, and let 1 < p < ∞. Suppose that there exists a family E of connected
sets in Z and a probability measure ν on E with the following properties:
(i) there exists c > 0 so that diam E ≥ c for all E ∈ E, and
(ii) there exists C and r0 > 0 so that

ν{E ∈ E : E ∩ B = ∅} ≤ Cμ(B)1/p
for all balls B in Z with radius at most r0 .
Then Cdim Z ≥ p.
Condition (i) asks that the elements of E have diameters uniformly
bounded from below. In practice, E is typically a family of rectifiable curves.
Condition (ii) is the “diffusivity” condition: it forbids the curve family mea-
sure ν from concentrating on subsets of E which are supported in sets of
small μ-measure.
Remark 4.1.4. In case (Z, d) is Ahlfors Q-regular, criterion (ii) is equiv-
alent to the following:
(ii’) There exists C and r0 > 0 so that

(4.1.1) ν{E : E ∩ B(z, r) = ∅} ≤ CrQ/p
for all balls B(z, r) with r ≤ r0 .
The conclusion is p ≤ Cdim Z ≤ dim Z = Q. When p = Q−1 Q
, p = Q and
Z is minimal for conformal dimension. In this case, assumption (4.1.1) is a
Fubini-type condition, implying that Z is fibered by a family of rectifiable
curves parameterized by a codimension one (dimension Q − 1) set. See
Examples 4.1.7, 4.2.3 and 4.2.5 for specific examples.
Proof of Proposition 4.1.3. Suppose that f : Z → Z  is a qua-
sisymmetric homeomorphism onto a space with dim Z  < p. Since Z and
Z  are compact, f is uniformly continuous and assumption (i) implies that
diam f (E) ≥ c for all E ∈ E and some c > 0.
Let  > 0. Using Lemma 4.1.1, select pairwise disjoint
  balls Bi ⊂ Z  of
  
radius ri so that the collection {5Bi } covers Z and i (ri ) < . We may
p

assume that  is sufficiently small relative to c so that none of the sets f (E),
E ∈ E, is entirely contained in any of the balls 10Bi .
 4.1. RICH CURVE FAMILIES AND LOWER BOUNDS FOR Cdim 51

Applying Lemma 1.2.18 to the inverse map f −1 : Z  → Z yields a

constant H and balls Bi ⊂ Z of radius ri so that Bi ⊂ f −1 (Bi ) ⊂ f −1 (5Bi ) ⊂
HBi . Note that the balls Bi are pairwise disjoint. For each Bi , define a
function hi : E → {0, 1} by

1 if f (E) ∩ 5Bi = ∅,
hi (E) =
0 otherwise.
For each E ∈ E,
  
10ri hi (E) = 10ri ≥ diam(5Bi )
i i i
f (E)∩5Bi =∅ f (E)∩5Bi =∅

≥ H∞
1
(f (E)) ≥ diam f (E) ≥ c
by Proposition 1.4.20. Thus
 
1 
c ≤ ri hi (E) dν(E)
10 E i

= ri ν{E ∈ E : f (E) ∩ 5Bi = ∅}
i

≤ ri ν{E ∈ E : E ∩ HBi = ∅}
i
 
≤C ri μ(HBi )1/p
i

1/p
1/p
 
≤C (ri )p μ(Bi )
i i

≤ C 1/p
μ(Z)1/p .
This yields a contradiction if  is chosen sufficiently small. 
Our next criterion for conformal dimension lower bounds starts from a
slightly different formulation of the “sufficiently rich curve family” idea. It
uses a fundamental tool of geometric function theory: the modulus of a curve
family.
Definition 4.1.5. Let (Z, d, μ) be a metric measure space and let p ≥ 1.
The p-modulus of a family of curves Γ is the quantity

Modp Γ = inf ρp dμ,
ρ Z
where the infimum is taken over all Γ-admissible functions ρ, i.e., Borel
functions ρ : Z → [0, ∞] for which γ ρ ds ≥ 1 for all locally rectifiable
curves γ ∈ Γ.
Basic properties of the modulus are summarized in the following propo-
sition. Parts (i), (ii) and (iii) imply that Modp is an outer measure on the
52 4. LOWER BOUNDS FOR CONFORMAL DIMENSION

collection of all rectifiable curves. The proof of part (iv) uses the uniform
convexity of the Lp spaces when p > 1.
Proposition 4.1.6. Let (Z, d, μ) be a metric measure space and let p ≥
1. Then
(i) Modp ∅ = 0,
(ii) if Γ1 ⊂ Γ2 , then
 Modp Γ1 ≤ Modp Γ2 ,
(iii) Modp ∪i Γi ≤ i Modp Γi for any countable set of collections Γi ,
(iv) if p > 1 and Γ1 ⊂ Γ2 ⊂ Γ3 ⊂ · · · , then Modp ∪i Γi = limi→∞ Modp Γi ,
(v) a curve family Γ has p-modulus zero if and only if there is ρ ∈ Lp (Z)
so that γ ρ ds = +∞ for all locally rectifiable curves γ ∈ Γ,
(vi) if Γ1 , Γ2 are curve families with the property that each curve in Γ1
contains a subcurve in Γ2 , then Modp Γ1 ≤ Modp Γ2 ,
(vii) if Z is compact and p < q, then Modp Γ ≤ μ(Z)1−p/q (Modq Γ)p/q
for every curve family Γ,
The inequality in Proposition 4.1.6(vii) is best possible, as our next
example demonstrates.
Example 4.1.7. Let (Y, dY , ν) be any metric measure space,
 let h > 0,
and let Z = Y × [0, h]. We equip Z with the  metric d = d2Y + d2E and
2

the product measure μ = ν × L1 . If E ⊂ Y and Γ = {γy : y ∈ E}, where

γy (t) = (y, t) for 0 ≤ t ≤ h, then
μ(E × [0, h]) ν(E)
Modp Γ = p
= p−1 .
h h
To see this, let ρ be any Γ-admissible function. Then
    1/p

ν(E) ≤ ρ ds dν(y) = ρ dμ ≤ ρp dμ (ν(E) · h)1/p
E γy E×[0,h] Z

whence Z ≥
ρp dμ On the other hand, ρ(x) = h−1 χE×[0,1] (z) is
h1−p ν(E).
 p
Γ-admissible and Z ρ dμ = h1−p ν(E).
Proposition 4.1.8 gives sufficient conditions for lower bounds on con-
formal dimension in terms of nontriviality of modulus. In contrast with
Proposition 4.1.3 we require here a mass bound assumption (4.1.2), since
we are comparing the arc length measure on the family of curves with the
ambient volume measure μ.
Proposition 4.1.8. Let (Z, d, μ) be a compact, doubling metric measure
space satisfying the upper mass bound
(4.1.2) μ(B(z, r)) ≤ CrQ
for all balls B(z, r) in Z, for some fixed constant C. Suppose that Z contains
a curve family Γ with Modp Γ > 0 for some 1 < p ≤ Q. Then Cdim Z ≥ Q.
Observe that the nonvanishing condition for the p-modulus weakens as
p increases, by Proposition 4.1.6(vii).
 4.1. RICH CURVE FAMILIES AND LOWER BOUNDS FOR Cdim 53

Proof. As before, suppose that f : Z → Z  is a quasisymmetric home-

omorphism onto a space Z  with dim Z  < Q. Since p > 1, we may use
Proposition 4.1.6(iv) to extract from Γ a subcollection with a uniform lower
bound on diameter which still has positive p-modulus. (Write Γ as the
countable increasing union of the subfamilies of Γ consisting of curves with
1
diameter at least m .) We continue to denote this subfamily by Γ. Again,
there exists c > 0 so that diam f (γ) ≥ c for all γ ∈ Γ.
Let  > 0 and the  collections {Bi } in Z and {Bi } in Y be as in the

previous proof, with i (ri ) < . We now define a Borel function ρ : Z →
Q

[0, ∞] by
 r
i
ρ(z) = χ2HBi (z).
ri
i
A calculation similar to that in the proof of Proposition 4.1.3 reveals that
Cρ is Γ-admissible for some large constant C. Using the estimates
   
p
p ri
ρ dμ = χ2HBi dμ
Z Z ri
i
  
p
ri
≤C χB dμ (see Exercise 2.11 in [76])
Z ri i
i
  r  p
i
=C μ(Bi )
ri
i

≤C (ri )p μ(Bi )1−p/Q
i

p/Q
1−p/Q
 
≤C (ri )Q μ(Bi )
i i
≤ C p/Q
μ(Z)1−p/Q ,
we obtain a contradiction by taking  sufficiently small. 
Example 4.1.9. Let Y be a compact, Ahlfors (Q − 1)-regular space and
let ν be the Hausdorff (Q − 1)-measure on Y . Let Z = Y × [0, 1] be as
in Example 4.1.7. Then μ = ν × L1 satisfies (4.1.2) for all balls B(z, r)
in Z for a (possibly different) constant C. Moreover, Z supports curve
families of positive p-modulus for any p. It follows from Proposition 4.1.8
that Cdim Z ≥ Q. In fact, μ is a Q-regular measure and Z is minimal for
conformal dimension.
Since compact Ahlfors regular spaces exist in every dimension, we have
shown
Proposition 4.1.10 (Tyson). For each Q ≥ 1, there exist metric spaces
Z with Cdim Z = dim Z = Q.
54 4. LOWER BOUNDS FOR CONFORMAL DIMENSION

In the next chapter, we will see that Proposition 4.1.10 is sharp in the
following sense: any metric space of dimension strictly less than one has
conformal dimension zero. Thus the range of conformal dimension is
{0} ∪ [1, ∞].
If the space Y in Example 4.1.9 is a subset of a Euclidean space, we can
even dispense entirely with the Ahlfors regularity assumption.
Proposition 4.1.11 (Bishop–Tyson). For every compact set Y ⊂ RN −1 ,
the space Z = Y × [0, 1] is minimal for conformal dimension.
The idea is to use the geometry of Euclidean space, substituting cover-
ings with dyadic cubes for coverings with balls. Dyadic cube coverings have
the advantage that they can be selected to be essentially disjoint, avoiding
the need for covering theorems (and the resulting use of the doubling prop-
erty). The relevant upper mass bound condition can be obtained by use of
Frostman’s lemma. We sketch the argument.
Sketch of the proof of Proposition 4.1.11. Let Q(N ) be the col-
lection of all dyadic cubes in RN , i.e., cubes with sides parallel to the coor-
dinate axes and vertices with appropriate dyadic rational coordinates.
For each d < dim Y , there exists a positive Borel measure ν on Y satis-
fying
(4.1.3) ν(Q) ≤ C(Q)d
for all Q ∈ Q(N − 1), where (Q) denotes the side length of Q. This follows
from Frostman’s lemma 1.4.11(2) and the comparability of balls and dyadic
cubes. The product measure μ = ν × L1 satisfies
(4.1.4) μ(Q) ≤ C(Q)d+1
for all Q ∈ Q(N ). Note that dim Z = dim Y + 1 by standard product
formulas for the Hausdorff dimension (which require at least one factor—in
this case the interval [0, 1]—to be reasonably nice).
Suppose that f : Z → Z  is a quasisymmetric homeomorphism onto a
space Z  of dimension d + 1 < dim Z. For each  > 0, choose a covering
{Qj } of Z by dyadic cubes with disjoint interiors so that

diam f (Qj )d+1 < .
j

Here we use the fact that Hausdorff dimension can be computed using dyadic
Hausdorff measures, or measures generated by the Carathéodory construc-
tion using the images of dyadic cubes under a fixed quasisymmetric map.
Define ρ : Z → [0, ∞) by ρ(z) = diam f (Qj )/(Qj ) if z ∈ Qj and ρ(z) = 0
otherwise. Then
  diam(f (Qj ))d+1
ρd+1 dμ = μ(Qj ) < C.
Z (Qj )d+1
j
 4.2. LOEWNER SPACES AND SPACES MINIMAL FOR Cdim 55

In view of Example 4.1.7 and since  is arbitrary, the proof will be finished
if we can show that ρ is comparable to an admissible function for the curve
family Γ = {γy }. To show this, we compute
  diam f (Qj )
ρ ds = (γy ∩ Qj )
γy (Qj )
j
γy ∩Qj =∅

= diam f (Qj )
j
γy ∩Qj =∅

≥ H∞
1
(f (γy )) ≥ diam f (γy ) ≥ c > 0
for each y ∈ Y . Note that (γy ∩ Qj ) = (Qj ) whenever γy ∩ Qj = ∅, by the
choice of the curves γy . 
Remark 4.1.12. The compactness assumption in Propositions 4.1.3,
4.1.8 and 4.1.11 is not strictly necessary. It suffices for Z to have a compact
subset containing a curve family with the relevant properties.

4.2. The Loewner condition and examples of spaces minimal for

conformal dimension
Proposition 4.1.8 asserts that a qualitative condition on the nontriviality
of the p-modulus, coupled with a quantitative volume growth assumption,
suffices to deduce lower bounds on conformal dimension. In many self-similar
(or quasi-self-similar) spaces, this qualitative condition can be promoted to
a quantitative form. The following notion was introduced by Heinonen and
Koskela.
Definition 4.2.1 (Heinonen–Koskela). A metric measure space (Z, d, μ)
is a p-Loewner space if there exists a decreasing function ϕ : (0, ∞) → (0, ∞)
so that
Modp Γ(E, F ) ≥ ϕ(Δ(E, F ))
for all disjoint compact sets E, F ⊂ Z, where Γ(E, F ) denotes the collec-
dist(E,F )
tion of curves joining E to F and Δ(E, F ) = min{diam E,diam F } denotes the
relative distance between E and F .
In case (Z, d, μ) is Ahlfors regular of dimension Q, we abbreviate Q-
Loewner as Loewner. From Proposition 4.1.8 we conclude
Corollary 4.2.2 (Tyson). Every Ahlfors regular Loewner space is min-
imal for conformal dimension. More precisely, if (Z, d, μ) is Q-regular and
Q-Loewner, then Cdim Z = Q.
We illustrate with a number of examples.
Example 4.2.3. Carnot groups G equipped with their standard Carnot-
Carathéodory metric are minimal for conformal dimension. In fact, dim G =
Cdim G = Q where Q is the homogeneous dimension of G. Let g = v1 ⊕· · ·⊕
56 4. LOWER BOUNDS FOR CONFORMAL DIMENSION

vs be the Lie algebra of G, let W = span{X} ⊂ v1 be a one-dimensional

horizontal subspace with complementary subspace W ⊥ and consider the
family Γ of horizontal curves in G given by
γy (t) = y · exp(tX), 0 ≤ t ≤ h,
where y ranges over a compact subset K of S := exp(W ⊥ ). We may equip K
with a probability measure (a multiple of the Carnot-Carathéodory (Q − 1)-
dimensional Hausdorff measure). The hypotheses of Proposition 4.1.3 are
verified with p = Q.
The same curve family Γ has positive Q-modulus, and so the minimality
for conformal dimension can equivalently be obtained from Proposition 4.1.8.
Q
In fact, (G, dcc , Hcc ) is a Loewner space, although this is more difficult to
establish.
As a consequence of the conformal dimension minimality, observe that
(G, dcc ) is never quasisymmetrically equivalent to the underlying Euclidean
space RN when G is non-abelian (i.e., Q > N ).
Example 4.2.4. Any nonempty open set in a Carnot group G is minimal
for conformal dimension. It suffices to prove that every open ball Bcc (p, r)
in the Carnot-Carathéodory metric is minimal. Let us recall from (3.3.6)
that the dilations δr , r > 0, on G are given by δr (exp(X1 + · · · + Xs )) =
exp(rX1 + · · · + rs Xs ). Similarities of the Carnot-Carathéodory metric dcc
are compositions of dilations and left translations. These are all bi-Lipschitz
maps on (G, dcc ). Thus Cdim Bcc (p, r) is independent of p and r. By choos-
ing r sufficiently large, we can ensure that Bcc (p, r) contains the family Γ
from Example 4.2.3. Hence Cdim Bcc (p, r) ≥ Q for all p and r. The equality
follows since G itself has (conformal) dimension Q.
Example 4.2.5. The boundaries ∂∞ Ipq of Bourdon–Pajot hyperbolic
buildings Ipq , equipped with the visual metric da0 with maximal parame-
ter a0 = e, are minimal for conformal dimension. In fact, dim ∂∞ Ipq =
Cdim ∂∞ Ipq = Qpq . Fix ξ, η ∈ ∂∞ Ipq and consider the curve family Γ of
curves joining ξ to η which are boundaries of hyperbolic planes in Ipq con-
taining bi-infinite geodesics terminating at ξ and η. This curve family can be
equipped with a probability measure with respect to which the hypotheses
of Proposition 4.1.3 are satisfied with p = Q. As in the previous example,
Q
(∂∞ Ipq , da0 , Hdapq ) is also Loewner.
0
Theorem 3.5.4 implies that ∂∞ Ipq and ∂∞ Ip q are not quasisymmetri-
cally equivalent (and hence the corresponding hyperbolic spaces Ipq and Ip q
are not quasi-isometrically equivalent) if Qpq = Qp q . Recall that the value
of Qpq does not uniquely determine Ipq .

Example 4.2.6. For n ≥ 2 and K ∈ {R, C, H}, the boundary ∂∞ HKn

equipped with its Carnot-Carathéodory metric is minimal for conformal di-
mension. This is Pansu’s theorem 3.3.3. We are now in a position to give a
proof. The point is that this space is locally bi-Lipschitz equivalent to the
 4.2. LOEWNER SPACES AND SPACES MINIMAL FOR Cdim 57

corresponding Iwasawa group G = Kn−1 × Im(K), which as we have already

discussed is minimal for conformal dimension. We give a detailed proof of
this fact since it illustrates some beautiful explicit formulas of conformal
geometric analysis in Iwasawa groups.
Let us recall that the groups under consideration are given by G =
K n−1 × Im(K), with group law (3.3.5). In place of the Carnot-Carathéodory
metric dcc we introduce the bi-Lipschitz equivalent Korányi metric dKor ,
defined by
dKor (p, q) = ||p−1 ◦ q||Kor ,
where the Korányi norm || · ||Kor is defined by
 1/4
(4.2.1) ||p||Kor := ||x||4 + |t|2 , p = (x, t).
The bi-Lipschitz equivalence of these two metrics is a consequence of the fact
that each of them is left invariant and scales linearly under the dilations
δr (x, t) = (rx, r2 t). Thus the comparability reduces to the fact that the
(continuous) Korányi norm is bounded away from zero and infinity on the
(compact) unit sphere in the Carnot-Carathéodory metric.
The following lemma is a simple computation using (3.3.3) and (4.2.1).
Lemma 4.2.7. The Cayley transform C : ∂∞ HKn \{(0, −1)} → G satisfies

1 − 2(Re xn )2 + 2| Im xn |2 + |xn |4
||C(x)||4Kor = , x ∈ ∂∞ HKn , xn = −1.
|1 + xn |4
In particular, if Re xn ≥ 0, then ||C(x)||Kor ≤ 1.
Since open subsets of Carnot groups are minimal for conformal dimen-
sion, the minimality of ∂∞ HKn for conformal dimension will follow once we
establish the next proposition. Indeed, we have
kn + k − 2 = Q = Cdim{p : ||p||Kor < 1}
= Cdim{x : Re xn > 0}
≤ Cdim ∂∞ HKn ≤ dim ∂∞ HKn = kn + k − 2
and equality holds throughout.
Proposition 4.2.8. The Cayley transform C is bi-Lipschitz as a map
from {x ∈ ∂∞ HKn : Re xn ≥ 0} to {p ∈ G : ||p||Kor ≤ 1}.
Proof. We equip ∂∞ HKn with the Cygan metric dCygan , and we equip
G = Kn−1 × Im(K) with the Korányi metric dKor . We will prove that
dCygan (x, y) ≤ dKor (C(x), C(y)) ≤ 2dCygan (x, y)
for all x, y ∈ ∂∞ HKn with Re xn ≥ 0 and Re yn ≥ 0.
To this end, we use the group law (3.3.5) to compute
2
dKor (p, q)4 = ||p−1 ◦ q||4Kor = ||p1 ||2 − p2 − 2p1 , q1  + ||q1 ||2 + q2
58 4. LOWER BOUNDS FOR CONFORMAL DIMENSION

for p = (p1 , p2 ) and q = (q1 , q2 ) in G = Kn−1 × Im(K). If p = C(x) and

q = C(y) for x = (x , xn ) and y = (y  , yn ) in ∂∞ HKn , we find
||p1 ||2 − p2 = ||x ||2 |1 + xn |−2 + 2|1 + xn |−2 Im xn
= |1 + xn |−2 (1 − |xn |2 + 2 Im xn )
= (1 + xn )−1 (1 − xn )
and similarly
||q1 ||2 + q2 = (1 + yn )−1 (1 − yn ).
Furthermore,
p1 , q1  = (1 + xn )−1 x , y  (1 + yn )−1 .
Putting this all together, we find
||p1 ||2 − p2 − 2p1 , q1  + ||q1 ||2 + q2
 
= (1 + xn )−1 (1 + xn )(1 − yn ) − 2x , y   + (1 − xn )(1 + yn ) (1 + yn )−1
= 2(1 + xn )−1 (1 − x, y)(1 + yn )−1 .
Hence
4|1 − x, y|2 16dCygan (x, y)
(4.2.2) dKor (C(x), C(y))4 = = .
|1 + xn | |1 + yn |
2 2 |1 + xn |2 |1 + yn |2
Since 0 ≤ Re(xn ) ≤ 1 and 0 ≤ Re(yn ) ≤ 1, we have 1 ≤ |1 + xn |2 ≤ 4 and
1 ≤ |1 + yn |2 ≤ 4, which completes the proof. 
Remark 4.2.9. When K = R the Cayley transform is just the standard
stereographic projection C : Sn−1 → Rn−1 , C(x) = C(x , xn ) = (1+xn )−1 x .
In this case, the Korányi metric on Rn−1 coincides with the Euclidean metric,
while the Cygan metric on Sn−1 is equal to 12 times the chordal metric q on
Sn−1 (i.e., the metric induced by the inclusion Sn−1 ⊂ Rn ). In this case, the
explicit formula (4.2.2) takes the form
q(x, y)
|C(x) − C(y)| = √ √ , x = (x , xn ), y = (y  , yn ) ∈ Sn−1 ,
1 + xn 1 + yn
which is a well-known identity relating stereographic projection to the chordal
metric.

4.3. The conformal dimension of the Sierpiński carpet

In this section, we illustrate the theory developed in section 4.1 through
discussion of a well known open problem: the conformal dimension of the
classical Sierpiński carpet.
Example 4.3.1. Let SC denote the standard Sierpiński carpet; see Fig-
ure 4.1. We recall that SC is obtained from the unit square in R2 by iterating
the following procedure: divide all current squares into nine subsquares of
equal size (with side length equal to 13 the side length of the current square)
and remove the central square. The set SC can be described as the invari-
ant set for a self-similar iterated function system. See section 8.3 for the
 4.3. THE CONFORMAL DIMENSION OF THE SIERPIŃSKI CARPET 59

general theory of iterated function systems and self-similar fractals. The

dimension of SC is equal to its similarity dimension log 8
log 3 ≈ 1.892 · · · . This
is a consequence, for instance, of the Moran–Hutchinson theorem 8.3.2. We
note that both Hausdorff and Assouad dimensions agree for this set, indeed,
SC is Ahlfors regular of dimension log 8
log 3 .

Figure 4.1. The Sierpiński carpet SC

Note that SC contains a subset of the form K × [0, 1], where K denotes
the standard 13 Cantor set. Since K is α-Ahlfors regular (α = log 2
log 3 ), the
conditions in Example 4.1.9 are satisfied with Q = 1 + α. Thus
log 2
Cdim SC ≥ Cdim K × [0, 1] ≥ 1 + ≈ 1.631 · · · .
log 3
The exact value of Cdim SC is unknown. A result of Keith and Laakso
(which we will discuss in detail in chapter 6) implies that
(4.3.1) Cdim SC < dim SC.
Since SC is log 8
log 3 -Ahlfors regular, the inequality in (4.3.1) implies (by
Proposition 4.1.8) that Modp Γ = 0 for every curve family Γ in SC and
every 1 < p ≤ log 8
log 3 . We now give a direct proof of the triviality of the
p-modulus on SC for all p ≥ 1. This result in turn will play a role in the
proof of (4.3.1).
We begin with a simple lemma. A map f : Z → Z of a metric space
(Z, d) is called a λ-similarity if d(f (z1 ), f (z2 )) = λd(z1 , z2 ) for all z1 , z2 ∈ Z.
For instance, the dilations δr on a Carnot group G are similarities of the
Carnot-Carathéodory metric.
On a general metric space (Z, d) equipped with an Ahlfors Q-regular
measure μ, consider the volume derivative (Jacobian) of a map f : Z → Z,
defined as
d(f# μ) f# μ(B(z, r)) μ(B(z, λr))
Jf (z) = (z) := lim = lim .
dμ r→0 μ(B(z, r)) r→0 μ(B(z, r))

This quantity exists and is finite for μ-a.e. z ∈ Z. This is a differentiation

theorem for measures which can be proved using Vitali-type covering lemmas
which hold for doubling (in particular, for Ahlfors regular) measures. See
[116, Chapter 2] or [76, Chapter 1] for details.
60 4. LOWER BOUNDS FOR CONFORMAL DIMENSION

Lemma 4.3.2. If f is a λ-similarity of a Q-regular metric measure space

(Z, d, μ), then λQ /C ≤ Jf ≤ CλQ for a suitable constant C > 0. Further-
more,
1 Q−p
λ Modp Γ ≤ Modp f (Γ) ≤ C  λQ−p Modp Γ
C
for every curve family Γ and every p ≥ 1, for a suitable constant C  > 0.
Proof. This is a simple calculation using the effect of λ-similarities on
arc length measure and on the Ahlfors regular volume measure. 
Proposition 4.3.3. Let μ be the log 8
log 3 -Ahlfors regular Hausdorff measure
on the Sierpiński carpet SC. For each p ≥ 1, Modp Γ = 0 for every curve
family Γ in SC.
1
We normalize μ so that μ(SC) = 1. Then μ(Q) = 8 for each first level
subsquare Q in SC.
Lemma 4.3.4. Let Γ0 denote the family of curves joining the left edge of
SC to the right edge of SC. Then Modp Γ0 = 0 for each p ≥ 1.
Proof. Consider the piecewise constant weight ρ : SC → (0, ∞) defined
as follows: ρ(x) = a if x is contained in one of the six first level subsquares
of SC adjacent to either the left or right edge, and ρ(x) = b if x is contained
in one of the remaining two first level subsquares of SC. See Figure 4.2.
Here a and b are positive constants whose values will be determined shortly.
This function ρ will be Γ0 -admissible provided
2a + b
(4.3.2) ≥ 1.
3

a a

Figure 4.2. Weights on the Sierpiński carpet SC

Then the p-modulus of Γ0 is bounded above by


6 2 3ap + bp
ρp dμ = ap + bp = .
SC 8 8 4
By a simple iterative procedure, we define a family of admissible functions ρm
on SC. The values of ρm on the level m subsquares are given by monomials
in a and b according to a binomial distribution. The single admissibility
 4.4. ANNULAR LLC IMPLIES Cdim GREATER THAN ONE 61

condition (4.3.2) guarantees that each ρm is Γ0 -admissible. The p-modulus

of Γ0 is bounded above by
  p 
p 3a + bp m
ρm dμ = .
SC 4
To complete the proof, it suffices to show that there exist positive constants
a and b satisfying
2a + b
(4.3.3) =1
3
and
3ap + bp
< 1.
4
For p = 1, any choice of a < 1 and b > 1 satisfying (4.3.3) will work. For
p > 1 we solve the corresponding optimization problem to find appropriate
values for a and b. By the method of Lagrange multipliers, we obtain the
optimal choices a = (8λ/3p)1/(p−1) and b = (4λ/p)1/(p−1) , where
p3p
λ= .
4(2p/(p−1) + 31/(p−1) )p−1
We compute
3ap + bp 3p+1
=
4 4(2p/(p−1) + 31/(p−1) )p−1
which is strictly less than one for all 1 < p < ∞ (exercise). 
Proof of Proposition 4.3.3. By Proposition 4.1.6, it is sufficient to
prove the result for curve √families Γ satisfying inf γ∈Γ diam γ ≥  for fixed
 > 0. Choose m so that 2 · 31−m <  and write Γ as the union of curve
families Γw = {γ ∈ Γ : γ ∩ Qw = ∅}, where Qw ranges over all level m
subsquares on SC. Again by Proposition 4.1.6, it suffices to prove that
Modp Γw = 0 for all such Qw . The choice of m guarantees that each γ ∈ Γw
is not entirely contained in the set consisting of Qw and all other level m
squares which touch Qw . Then each such γ contains a subcurve inside some
square, or possibly inside a pair of adjacent squares, in Qw which joins the
left edge to the right edge (or the top edge to the bottom edge). There are a
countable number of such subsquares. Applying Proposition 4.1.6 one more
time, together with Lemmas 4.3.2 and 4.3.4, we finish the proof. 

4.4. Annular LLC implies conformal dimension bounded away

from one
The usual Sierpiński carpet SC contains the product of a Cantor set
and the unit interval. This pleasant situation gives us a lower bound on
conformal dimension greater than one, see Example 4.3.1.
In this section, we describe a class of spaces which may not be as strongly
connected as the Loewner spaces of Definition 4.2.1 — for example, they
need not contain any rectifiable curves — but nevertheless support a family
62 4. LOWER BOUNDS FOR CONFORMAL DIMENSION

of curves sufficiently rich to bound the conformal dimension away from one
via Lemma 4.1.3.
Let X be a metric space, and fix 0 < r < R and p ∈ X. We define the
annulus A(p, r, R) to be the set B(p, R) \ B(p, r).
Definition 4.4.1. A metric space X is called L-annularly linearly con-
nected (or L-annular LLC), if whenever x, y ∈ A(p, r, 2r) for some p ∈ X
and r > 0, then there exists an arc connecting x and y in A(p, r/L, 2rL).
We will always assume that X is complete and connected.
This property is a quantitative metric version of the topological prop-
erties of being locally connected and having no local cut points. The usual
Sierpiński carpet is annularly linearly connected.
Example 4.4.2. If ∂∞ G is the boundary of a Gromov hyperbolic group,
and ∂∞ G is connected with no local cut points, then ∂∞ G is annularly
linearly connected. The self-similarity of ∂∞ G allows us to promote the
topological condition to a quantitative metric version. See section 3.4 for
more details.
For our purposes, the interest in this property is the following result.
Theorem 4.4.3 (Mackay). If X is an N -doubling, L-annularly linearly
connected, complete metric space, then Cdim X ≥ C(L, N ) > 1.
The rough idea of the proof is as follows. Given an arc in the metric
space X, the annular linear connectivity condition permits one to “unzip”
it into two disjoint arcs nearby. One repeats this construction to get four
disjoint arcs, then eight, and so on, limiting to a “Cantor set of curves” in
X. Once this is done, Proposition 4.1.3 applied to this collection of curves
gives a lower bound for the conformal dimension that is strictly greater than
one.
The difficulty in the proof arises from trying to satisfy the upper mass
bound (ii) in Proposition 4.1.3. This requires that the curves must spread
out in a uniform way on every scale. To do this requires delicate compactness
arguments which we do not describe here.
All steps in the proof are quantitative, but there is no explicit estimate
for the lower bound C(L, N ).
Recall that the existence of local cut points in the boundary of a hy-
perbolic group implies that the group splits over a virtually cyclic subgroup
(Theorem 3.4.8). Therefore we can deduce the following corollary.
Corollary 4.4.4. If G is a hyperbolic group, and Cdim ∂∞ G = 1, then
either G is virtually Fuchsian, or G splits over an elementary subgroup.

4.5. Rich families of sets of conformal dimension one

Proposition 4.1.8 tells us that the existence of curve families of positive
modulus guarantees lower bounds on conformal modulus. That is, the space
 4.5. RICH FAMILIES OF SETS OF CONFORMAL DIMENSION ONE 63

must carry a sufficiently rich family of sets, each of which is a rectifiable

curve. It is natural to ask whether we can relax the assumption that these
sets be connected. More directly, motivated by Proposition 4.1.11 we ask
the following
Question 4.5.1. Does there exist a totally disconnected space K so that
Y × K is minimal for conformal dimension for every compact metric space
Y?
Progress on this question was made by Hakobyan.
Theorem 4.5.2 (Hakobyan). Let K ⊂ [0, 1] be a set which has conformal
dimension one and which supports a measure λ with the following property:
for each  > 0 there exists C > 0 so that
(4.5.1) C−1 r1+ ≤ λ(K ∩ (x − r, x + r)) ≤ C r1−
for all x ∈ K and r ≤ 1. Then K × Y is minimal for conformal dimension
for every compact metric space Y .
Examples of totally disconnected sets K ⊂ R satisfying the assumptions
of the theorem will be given in section 8. The sets in question are non-self-
similar middle interval Cantor sets of dimension one.
An important tool in the proof of Theorem 4.5.2 is Fuglede’s notion of
the modulus of a family of measures.
Definition 4.5.3 (Fuglede). Let (Z, μ) be a measure space. Let E be
a family of measures on Z, so that the μ-measurable sets are λ-measurable
for each λ ∈ E. We say that a μ-measurable function ρ : Z → [0, ∞] is
E-admissible if Z ρ dλ ≥ 1 for each λ ∈ E. For 1 ≤ p < ∞, the p-modulus
of E is the quantity 
Modp E = inf ρp dμ,
Z
where the infimum is taken over all E-admissible functions ρ.
We illustrate with some examples.
Example 4.5.4. Let EA be the family of Dirac measures δa , a ∈ A,
associated to a μ-measurable set A ⊂ Z. Then Modp EA = μ(A) for each
p ≥ 1.
Example 4.5.5. Let (Z, d, μ) be a metric measure space and let EΓ be
the family of arc length measures ds|γ , γ ∈ Γ, associated to a family Γ of
locally rectifiable curves. Then Modp EΓ = Modp (Γ), the classical p-modulus
defined in Definition 4.1.5.
The primary technical step in the proof of Theorem 4.5.2 is the following
proposition.
Proposition 4.5.6. Let (Z, d, μ) be a doubling metric measure space
satisfying the upper mass bound inequality (4.1.2) for some Q > 1. Let E be
a collection of measures on Z such that
64 4. LOWER BOUNDS FOR CONFORMAL DIMENSION

(i) Cdim Eλ = 1 for each λ ∈ E, where Eλ = supp(λ),

(ii) Modq E > 0 for some 1 ≤ q < Q,
(iii) there exists r0 > 0 and for each  > 0 there exists C > 0 so that
λ(Eλ ∩ B(z, r)) ≥ C−1 r1+ for each z ∈ supp(λ) and r ≤ r0 .
Then Cdim Z ≥ Q.
Proof of Theorem 4.5.2. Let Y be a compact metric space of pos-
itive Hausdorff dimension d, let s < d, and let  < d − s. By Frostman’s
lemma, there exists a measure ν on Y so that ν(Y ) > 0 and ν(B(y, r)) ≤
Crs+ for all y ∈ Y and r > 0. For each y ∈ Y , let Ey be the set K ×{y} and
let λy be the restriction of λ×L1 to Ey . In other words, λy (A×{y}) := λ(A)
for all A ⊂ K. Let E be the family of measures {λy }y∈Y . A simple adapta-
tion of the argument in Example 4.1.7 yields that Mods+1 E > 0.
Apply (4.5.1) with  replaced by /2. The product measure μ = λ × ν
on Z = K × Y satisfies the upper mass bound inequality
μ(B(z, r)) ≤ Cr1−/2 rs+ = Crs+1+/2
for all z ∈ Z and r > 0, where C depends on the constant C/2 from (4.5.1).
By Proposition 4.5.6, Cdim Z ≥ s+1+/2 > s+1. Letting s → d completes
the proof, since dim Z ≤ dim Y + 1 = d + 1. 
Remark 4.5.7. Let us reconcile Proposition 4.5.6 with the general com-
ments at the beginning of this chapter. Suppose that {Eλ } is a family of sets
as in the statement of the proposition. While it may be the case that some
or all of the sets Eλ are totally disconnected, their minimality for conformal
dimension ensures that there exist weak tangents of the space (Z, d, μ) in
which the limits of the sets Eλ are connected. (We discuss tangent and
weak tangent spaces in Chapter 6.) Thus it is still the case that some weak
tangent space of (Z, d, μ) contains a large curve family.

4.6. Notes
The principal message of this chapter is that lower bounds on conformal
dimension reflect strong connectivity properties of the underlying space.
Proposition 4.1.3 is a variation on results of Pansu [133, Proposition 2.9],
[134, Lemma 6.3] and Bourdon [27, Lemma 1.6]. The main novelty is our
use of the underlying measure of a ball (instead of the radius) to control
the diffusivity of the curve family measure ν; this eliminates the need for
any volume growth assumption. Indeed, Proposition 4.1.3 holds in arbitrary
compact doubling spaces. Propositions 4.1.8 and 4.1.10 were first proved via
a different method by Tyson [169]. The proof of Proposition 4.1.8 which
we gave here was found by Bonk and Tyson and appears in section 15 of
Heinonen’s monograph [76]. Proposition 4.1.11 was observed in [17]. The
triviality of the modulus on the Sierpiński carpet SC can be found in Lupo-
Krebs–Pajot [106]; see also Bourdon and Pajot [33] for a variant proof of
the nonexistence of Poincaré inequalities on SC.
 4.6. NOTES 65

To illustrate Example 4.1.9, let Y ⊂ R2 be the von Koch snowflake curve.

This space is Ahlfors Q-regular with Q = log 4
log 3 > 1, hence (4.1.3) holds when
ν is the Hausdorff Q-measure on Y . The product space Y × [0, 1] is known
as Rickman’s rug. The fact that it is not quasisymmetrically equivalent to
[0, 1]2 is well-known, dating back to Tukia [162]. Note that Y is bi-Lipschitz
equivalent to a snowflaked metric space ([0, 1], dE ), where dE denotes the
Euclidean metric on R and  = Q 1
= log 3
log 4 . A generalization of this example
can be found in [168, section 15]. For  = (1 , . . . , N ) ∈ (0, 1)N , let X()

denote the product of N snowflake spaces ([0, 1], dEj ), j = 1, . . . , N . Then
• X() and X(δ) are quasisymmetrically equivalent if and only if the
vector (1 , . . . , N ) is a positive multiple of the vector (δ1 , . . . , δN )
(up to permutation of the indices), and
• X() and X(δ) are bi-Lipschitz equivalent if and only if the vectors
 and δ coincide (up to permutation of the indices).
Put another way, X() and X(δ) are quasisymmetrically equivalent if and
only if they are conformally equivalent, and are bi-Lipschitz equivalent if
and only if they are isometric.
The observation that every doubling metric measure space is a doubling
metric space admits the following converse: Every complete doubling metric
space carries a doubling measure. See Wu [180] and Luukkainen–Saksman
[108]. More precisely, for each s > dimA Z there exists a Borel measure μ
on (Z, d) and a constant C so that

(4.6.1) μ(B(z, λr)) ≤ Cλs μ(B(z, r))

for all B(z, r) and all λ ≥ 1. Conversely, if Z supports a measure μ satisfying

(4.6.1) for some s, then dimA Z ≤ s. Thus the Assouad dimension of a
complete doubling metric space may be characterized as the infimum of
those s > 0 for which there exists a measure μ satisfying (4.6.1). The
following terminology is taken from Luukkainen [107].

Definition 4.6.1. A Borel measure μ on a metric space (Z, d) which

satisfies (4.6.1) for all balls B(z, r) and all λ ≥ 1 is called an s-homogeneous
measure.

The Loewner condition was introduced by Heinonen and Koskela [77].

It is nowadays a standard axiom in first-order analysis in metric spaces. In
sufficiently nice metric spaces, it is equivalent with a Poincaré inequality.
The Loewner property of R2 is an easy consequence of known identities
for the conformal 2-modulus (extremal width) arising in univalent function
theory. That RN is Loewner for N ≥ 3 was proved (essentially) by Loewner
in 1959 [105]. Example 4.2.3 is due to Heinonen [75], while Example 4.2.5
is due to Bourdon and Pajot [31].
Theorem 4.4.3 was proved in the Ph.D. thesis of Mackay [111], see also
[110].
66 4. LOWER BOUNDS FOR CONFORMAL DIMENSION

The material in section 4.5 comes from the recent paper of Hakobyan
[73]. For an application of Proposition 4.5.6 to the computation of the
conformal dimension of self-affine fractals, see section 8.4. Question 4.5.1
was posed by Bishop and Tyson in [17]. The notion of modulus of a family of
measures was introduced by Fuglede [54] in 1957. Recently it has featured
in a number of applications in geometric function theory. In addition to
the cited work of Hakobyan, it plays a role in Rajala’s work on moduli of
surface families, Fatou’s theorem and the local homeomorphism property
for quasiregular maps [145], [146].
 CHAPTER 5

Sets and spaces of conformal dimension zero

Sets of conformal dimension zero are well understood. In this chapter

we review what is known about such sets. The crucial result is Kovalev’s
Theorem 5.1.12: metric spaces of Hausdorff dimension strictly less than one
have conformal (Hausdorff ) dimension zero. Conformal gauges of conformal
Assouad dimension zero can be explicitly characterized (Theorem 5.1.9). We
discuss how unions and products affect the collection of spaces of conformal
dimension zero, and more generally, how such operations behave with respect
to conformal dimension.

5.1. Conformal dimension, porosity and uniform

disconnectedness
Observe first that any gauge of conformal dimension zero is totally
disconnected. Indeed, if Cdim G = 0 then there exists a metric space
(Z  , d ) ∈ G with dim Z  < 1. All metric spaces of Hausdorff dimension
strictly less than one are totally disconnected. Hence Z  , and also G, are
totally disconnected.
As indicated in chapter 1, the Cantor set has conformal dimension zero.
We now generalize this observation.
Definition 5.1.1. A set S in a metric space Z is porous in Z if there
exists a constant c > 0 so that for every ball B(z, r) in Z, there exists a ball
B(z1 , cr) ⊂ B(z, r) so that B(z1 , cr) ∩ S = ∅.
Porous sets have holes of a definite (proportional) size at every location
and scale. For porous sets in the real line, we can reduce the dimension
quasisymmetrically by expanding the size of these omitted intervals.
Proposition 5.1.2. Let S be porous in R. Then Cdim S = 0.
Proof. Assume for simplicity that S is bounded, say, S ⊂ [0, 1]. The
porosity condition on S implies the following: there exists c > 0 (depending
only on c) so that for every interval I ⊂ [0, 1], there exists an interval J ⊂ I
so that dist(J, ∂I) ≥ 14 |I|, J ∩ S = ∅ and |J| ≥ c |I|. (We write |E| for
the length of an interval E.) Fix  > 0 and δ < 2−1/ . We define an
increasing piecewise linear homeomorphism f : I → I as follows: f |∂I is the
identity map, the midpoint of f (J) coincides with the midpoint of f (I), and
|f (J)| = (1 − 2δ)|f (I)|. The map f is η-quasisymmetric with η depending
only on c.
67
68 5. SETS AND SPACES OF CONFORMAL DIMENSION ZERO

We now repeat this construction inductively, beginning with I = [0, 1]

and repeating on all complementary intervals. Thus in the first stage we
construct J ⊂ [ 14 , 34 ], J ∩ S = ∅, with complementary intervals I1 , I2 (i.e.,
[0, 1] = I1 ∪J ∪I2 ), together with an η-quasisymmetric map f1 : [0, 1] → [0, 1]
so that |f1 (J)| = 1 − 2δ. In the next stage we construct Jk ⊂ Ik , k = 1, 2,
Jk ∩ S = ∅, with complementary intervals I11 , I12 , I21 , I22 , together with an
η-quasisymmetric map f2 : [0, 1] → [0, 1] so that |f2 (Jk )| = (1 − 2δ)2 . After
m steps, we obtain an η-quasisymmetric map fm : [0, 1] → [0, 1], where η
depends only on c and not on m, so that |fm (Jw )| = (1 − 2δ)m for every
word w of length m in the letters 1 and 2. The maps fm converge uniformly
as m → ∞ to an increasing homeomorphism f : [0, 1] → [0, 1] which is
η-quasisymmetric. We will show that f (S) has dimension  less than .
Observe that, for each m, f (S) is contained in f (Iw ) where the union
is taken over all complementary intervals Iw at the mth level. Thus

Hδ m (f (S)) ≤ |f (Iw )| ≤ 2m δ m .
w
Since 2δ 
< 1, we obtain H (f (S)) =
 0 upon taking the limit as m → ∞.
Thus Cdim S ≤ dim f (S) < . Letting  → 0 completes the proof. 
Remark 5.1.3. The proof shows that S can be mapped onto sets of
arbitrarily small dimension via global quasisymmetries of the real line. In
the language of chapter 8, S has global quasiconformal dimension zero. A
stronger statement will appear in Theorem 8.5.1.
We amplify the preceding proposition with some further relations be-
tween porosity, Assouad dimension and uniform disconnectedness. The
proof of Proposition 5.1.4 uses the dyadic decomposition of RN and volume
comparison estimates. The corresponding result (with N replaced by Q)
holds if RN is replaced by any Ahlfors Q-regular metric space. On the other
hand, Proposition 5.1.5 holds if RN is replaced by any Ahlfors Q-regular
Loewner space.
Proposition 5.1.4. A set S ⊂ RN is porous in RN if and only if
dimA S < N .

Proposition 5.1.5. Let S ⊂ RN be porous in RN and let f : RN → RN

be quasiconformal (quasisymmetric if N = 1). Then f (S) is porous in RN .

Corollary 5.1.6. For any N ≥ 1, the image of a set of Assouad di-

mension N under a quasisymmetric map of RN has dimension N .
The analogous result for Hausdorff dimension is true, but only in the
case N ≥ 2. See Corollary 8.1.2 and the notes to chapter 8.
The following result was alluded to in Chapter 1.
Proposition 5.1.7. Every metric space with Assouad dimension strictly
less than one is uniformly disconnected.
 5.1. Cdim, POROSITY AND UNIFORM DISCONNECTEDNESS 69

Proof. Let Z be a metric space with dimA Z < s < 1. Then

N (B(z, r), r) ≤ C−s
for some constant C < ∞ and all balls B(z, r) in Z and  > 0.
Suppose that Z is not uniformly disconnected. Then for each δ > 0
there exist nontrivial δ-chains in Z. Fix  > 0 and let δ = 2 . Select a
nontrivial δ-chain {z0 , z1 , . . . , zn } in Z. Thus d(zi−1 , zi ) < δd(z0 , zn ) for all
i = 1, . . . , n.
Consider the ball B(z, r) where z = z0 and r = d(z0 , zn ). We select in-
ductively a subset of {z0 , z1 , . . . , zn } as follows: zk0 = z0 and for each j, zkj+1
is the first point in {zkj +1 , zkj +2 , . . . , zn } ∩ B(z, r) whose distance to each of
the preceding points zk0 , zk1 , . . . , zkj is at least r. By the choice of  and δ,
we conclude that each zkj lies at distance ( + 2 )r from one of the preceding
points. Continue this procedure, generating points zk0 , . . . , zkT until there
are no remaining points of the chain inside B(z, r). By construction, all of
these points are in B(z, r) and they are r-separated. Hence
(5.1.1) T + 1 ≤ C−s .
However,
1
r ≤ d(zk0 , zkT ) ≤ d(zk0 , zk1 ) + · · · + d(zkT −1 , zkT ) ≤ T ( + 2 )r
2
which implies that
1 1
(5.1.2) T ≥ .
2  + 2
Since s < 1 we obtain a contradiction between (5.1.1) and (5.1.2) in the
limit as  → 0. 
We collect the preceding results to deduce the following characterization
of subsets of R of Assouad dimension strictly less than one.
Theorem 5.1.8. For a set S ⊂ R the following are equivalent:
(1) S is porous in R,
(2) dimA S < 1,
(3) S is uniformly disconnected.

Proof. The implications (1) ⇔ (2) and (2) ⇒ (3) are Propositions 5.1.4
and 5.1.7 respectively. To complete the circle we show that (3) ⇒ (1). Let
S ⊂ R be uniformly disconnected with parameter  > 0. We will show
that S is porous in R with parameter c = 2−m−1 , where m is the least
integer so that 2m ≥ 1 + −1 . On the real line, the porosity condition can be
written as follows: for all intervals I, there exists I  ⊂ I with |I  | ≥ c|I| and
I ∩ S = ∅. Suppose this is false for some I. Decompose I into subintervals
Ik , k = 1, . . . , 2m+1 , with disjoint interiors. Since |Ik | = 2−m−1 |I| = c|I|,
each Ik meets S at some point xk . Let x = x0 and y = x2m+1 . We estimate
|x − y| ≥ (1 − 2−m )|I| and |xi − xi−1 | ≤ 2−m |I| for all i = 2, . . . , 2m+1 . Thus
70 5. SETS AND SPACES OF CONFORMAL DIMENSION ZERO

−m
|xi − xi−1 | ≤ 1−2
2
−m |x − y| ≤ |x − y| and we have a nontrivial -chain in S

(Definition 1.3.3). This contradicts the uniform disconnectedness. 

We now characterize gauges of conformal Assouad dimension zero.
Theorem 5.1.9. A conformal gauge G has CdimA G = 0 if and only if
G is doubling and uniformly disconnected.
The proof makes use of the following well known observation.
Lemma 5.1.10. A metric space (Z, d) is uniformly disconnected if and
only if there exists an ultrametric d on Z so that (Z, d) and (Z, d ) are
bi-Lipschitz equivalent.
We recall that an ultrametric on a space Z is a symmetric nonnegative
function d defined on Z × Z which vanishes only along the diagonal and
satisfies the improved triangle inequality
(5.1.3) d(z1 , z2 ) ≤ max{d(z1 , z3 ), d(z3 , z2 )}
for all z1 , z2 , z3 ∈ Z. The basic (and most important) example is the class
of metrics da , a > 1, on the symbolic Cantor set {0, 1}∞ defined by
da (v, w) := a− min{i:vi =wi } , v = (v1 , v2 , . . .), w = (w1 , w2 , . . .) ∈ {0, 1}∞ .
The usual 13 Cantor set K, equipped with its Euclidean metric, is uniformly
disconnected. It is bi-Lipschitz equivalent to ({0, 1}∞ , d1/3 ).
One direction in Lemma 5.1.10 is obvious. Ultrametric spaces are uni-
formly disconnected, and uniform disconnectedness is bi-Lipschitz invari-
ant. For the other direction, consider a uniformly disconnected metric space
(Z, d). Define d : Z × Z → [0, ∞) by
d (z1 , z2 ) = d(z1 , z2 ) · inf{δ : there exists an δ-chain from z1 to z2 }.
Then d is a symmetric function. The equivalent formulation
 
there exists a sequence z1 = w0 , w1 , . . . , wM = z2
d (z1 , z2 ) = inf  :
so that d(wi−1 , wi ) ≤  for all i = 1, . . . , M
shows that d verifies (5.1.3). Obviously d (z1 , z2 ) ≤ d(z1 , z2 ). To see that d
and d are comparable, we use the fact that (X, d) is uniformly disconnected.
This completes the proof of Lemma 5.1.10.
Proof of Theorem 5.1.9. If CdimA G = 0, then there exists Y ∈ G
with dimA Y < 1. By Proposition 5.1.7, Y , hence G, is uniformly discon-
nected. Since dimA Y is finite, Y , hence G, is doubling.
For the converse, suppose that G is doubling and uniformly disconnected.
Choose (Z, d) ∈ G and let d be an ultrametric on Z which is bi-Lipschitz
equivalent to d. When applied to an ultrametric, the snowflaking operation
d
→ (d ) produces a new (ultra)metric for all  > 0, not just for 0 <
 < 1 (cf. Example 1.2.10(2)). Since G is doubling, dimA Z is finite. Then
dimA (Z, d ) = dimA Z → 0 as  → ∞. We conclude that CdimA G = 0. 
 5.2. CONFORMAL DIMENSION AND UNIONS 71

Corollary 5.1.11 (Tyson). If G contains a metric space of Assouad

dimension less than one, then CdimA G = 0. In other words, for any con-
formal gauge G, CdimA G ∈ {0} ∪ [1, ∞].
The analogous result for Hausdorff dimension is a theorem of Leonid
Kovalev.
Theorem 5.1.12 (Kovalev). If G contains a metric space of Hausdorff
dimension less than one, then Cdim G = 0. In other words, for any confor-
mal gauge G, Cdim G ∈ {0} ∪ [1, ∞].
More concretely, this says: if dim Z < 1, then Cdim Z = 0.
We will have more to say about Kovalev’s theorem in chapter 8. In
particular we will discuss some of the influential new techniques and methods
which he developed for the proof of this theorem.

5.2. Conformal dimension and unions

As mentioned in chapter 2, we do not yet have a complete picture of how
conformal dimension changes under unions, products, etc. In this section
and the next we present some tentative results along these lines.
Since all notions of dimension which we consider in this survey (including
conformal dimension) are monotone, the dimension of a union is always
at least as large as the dimensions of the pieces. We recall that Assouad
dimension is finitely stable:

dimA (S1 ∪ · · · ∪ SM ) = max dimA S1 , . . . , dimA SM ,
and that Hausdorff dimension is countably stable:

dim(S1 ∪ S2 ∪ · · · ) = sup dim S1 , dim S2 , . . . ,
for any sets S1 , S2 , . . . contained in a fixed metric space (Z, d).
Our first observation is that conformal Hausdorff dimension is not even
finitely stable.
Example 5.2.1. There exist sets S1 , S2 ⊂ [0, 1] of conformal dimension
zero so that S1 ∪ S2 = [0, 1] has conformal dimension one.
To construct such sets, we make use of a classical example due to Tukia.
Example 5.2.2 (Tukia). There exists a quasisymmetric map f of [0, 1]
to itself together with an Fσ set S∞ ⊂ [0, 1] such that both f (S∞ ) and
[0, 1] \ S∞ have Hausdorff dimension strictly less than one.
For the reader’s convenience, we review the construction of such a map
in subsection 5.4.
We now describe the construction of sets S1 , S2 as in Example 5.2.1. Let
S∞ be as in Example 5.2.2 and set S1 = S∞ and S2 = [0, 1] \ S∞ . The fact
that S1 and S2 both have conformal dimension zero follows from Kovalev’s
theorem 5.1.12.
Example 5.2.2 should be contrasted with the following result.
72 5. SETS AND SPACES OF CONFORMAL DIMENSION ZERO

Proposition 5.2.3. Let S1 and S2 be disjoint compact subsets of a met-

ric space (Z, d). Then Cdim(S1 ∪ S2 ) = max{Cdim S1 , Cdim S2 }.
Corollary 5.2.4. The conformal dimension of a finite union of disjoint
compact sets is equal to the maximum of the conformal dimensions of the
pieces.
The situation for conformal Assouad dimension is quite different. No
example such as Example 5.2.2 can exist, by the following proposition.
Proposition 5.2.5. Every finite union of sets of conformal Assouad
dimension zero again has conformal Assouad dimension zero.
Proposition 5.2.5 is an easy consequence of the characterization of spaces
of conformal Assouad dimension zero (Theorem 5.1.9) and the preservation
of the doubling and uniform disconnectedness properties under finite unions.

It is easy to see that conformal Assouad dimension is not countably

stable. By Proposition 6.1.7 in the following chapter, the rationals Q have
conformal Assouad dimension one. (We leave as an exercise for the interested
reader to verify this fact directly.) However, all finite sets have conformal
Assouad dimension zero. For a compact example, consider the set Z =
{n−p : n ∈ N}, p > 0, from Example 1.4.15. This set also has conformal
Assouad dimension one, for similar reasons.
Proof of Proposition 5.2.3. It suffices to prove that Cdim(S1 ∪S2 ) ≤
max{Cdim S1 , Cdim S2 }.
Fix  > 0 and quasisymmetric maps fj from Sj to metric spaces (Yj , dj )
with dim Yj < Cdim Sj + , for each j = 1, 2. Then Y1 and Y2 are compact.
Define δ on the disjoint union of Y1 and Y2 as follows:
(1) the restriction of δ to Yj is dj , and
(2) δ(y1 , y2 ) = D1 + D2 (where Dj := diamdj Yj ) if yj ∈ Yj , j = 1, 2.
It is clear that δ is a metric and that
dim(Y1 ∪ Y2 , δ) < max{Cdim S1 , Cdim S2 } + .
We will show that f : S1 ∪ S2 → Y1 ∪ Y2 defined by f |Sj = fj is quasisym-
metric. More precisely, let b = dist(S1 , S2 ) > 0 and let η be a common
quasisymmetry distortion function for f1 and f2 . We will show that f is
η̃-quasisymmetric for some η̃ = η̃(η, b, diam Sj , Dj ).
Let x, y, z ∈ S1 ∪S2 ; we must show that δ(f (x),f (y)) d(x,y)
δ(f (x),f (z)) ≤ η̃( d(x,z) ). Switching
the roles of S1 and S2 if necessary, it suffices to verify quasisymmetry in three
cases: (i) x, y ∈ S1 , z ∈ S2 , (ii) x, z ∈ S1 , y ∈ S2 , (iii) y, z ∈ S1 , x ∈ S2 . Let
L = diam S1 + b + diam S2 .
In case (i), we have
δ(f (x), f (y)) d1 (f1 (x), f1 (y)) d(x, y) d(x, y)
= and ≥ .
δ(f (x), f (z)) D1 + D2 d(x, z) L
 5.3. CONFORMAL DIMENSION AND PRODUCTS 73

Using Lemma 1.2.19, we estimate

   
d1 (f1 (x), f1 (y)) 2d(x, y) D1 d(x, y) 2L D1
≤η ≤η · .
D1 + D2 diam S1 D1 + D2 d(x, z) diam S1 D1 + D2
In case (ii), we have
δ(f (x), f (y)) D1 + D2 d(x, y) b
= and ≥ .
δ(f (x), f (z)) d1 (f1 (x), f1 (z)) d(x, z) d(x, z)
A similar computation using Lemma 1.2.19 yields
 
D1 + D2 2(D1 + D2 ) diam S1 d(x, y)
≤ η · .
d1 (f1 (x), f1 (z)) D1 b d(x, z)
Finally, in case (iii), we have
δ(f (x), f (y)) d(x, y) b
=1 and ≥ .
δ(f (x), f (z)) d(x, z) L
We must choose η̃ so that the following two conditions hold:
    
D1 2L 2(D1 + D2 ) diam S1
(5.2.1) η̃(t) ≥ max η t , η t
D1 + D2 diam S1 D1 b
for all t > 0, and
(5.2.2) η̃( Lb ) ≥ 1.
Inequalities (5.2.1) and (5.2.2) are satisfied if we set η̃(t) = C1 η(C2 t) for
sufficiently large C1 and C2 depending only on the specified data. This
completes the proof. 

5.3. Conformal dimension and products

Product spaces provide another interesting source of questions regarding
conformal dimension. The inequality
(5.3.1) dim Z1 + dim Z2 ≤ dim(Z1 × Z2 )
relating the Hausdorff dimensions of spaces Z1 , Z2 and their product is
well known. (Recall that all of the p product metrics on Z1 × Z2 are
bi-Lipschitz equivalent.) Equality holds in (5.3.1) if one of the two factors
is sufficiently regular (for instance, if its Hausdorff and Assouad dimensions
agree), but need not hold in general, even for compact subsets of Euclidean
space. Falconer [52, Example 7.8] provides an example of two compact sets
S1 , S2 ⊂ [0, 1], each of Hausdorff dimension zero, with the property that
S1 × S2 admits a Lipschitz map onto [0, 1]. Thus dim(S1 × S2 ) ≥ 1.
The situation regarding conformal dimension is unclear. The preceding
discussion tells us that dim(Z1 × Z2 ) = 0 whenever Z1 and Z2 are spaces
satisfying dim Z1 = 0 and dimA Z2 = 0.
74 5. SETS AND SPACES OF CONFORMAL DIMENSION ZERO

Question 5.3.1. For which metric spaces Z1 , Z2 of conformal dimension

zero do we have
(5.3.2) Cdim(Z1 × Z2 ) = 0?
For instance, does (5.3.2) hold if one of the two spaces has conformal Assouad
dimension equal to zero?
There is a subtle issue regarding Question 5.3.1 which is worth elaborat-
ing. Product constructions rarely interact well with quasisymmetric maps.
Put more simply, the conformal gauge of the product space is not well de-
fined (in terms of the original conformal gauges on the factor spaces). It
is for this reason that we stated Question 5.3.1 for individual metric spaces
and not for conformal gauges.
Let Gj be conformal gauges on the factor spaces Zj , j = 1, 2. For
each choice of metrics dj ∈ Gj , the product metric d1 × d2 lies in some
conformal gauge on Z1 × Z2 . Let {G3α }α∈A be the collection of all possible
gauges on Z1 × Z2 which contain 2 products of metrics in G1 and G2 . We
emphasize that the cardinality of the index set A is arbitrary (and may even
be uncountable). We restate Question 5.3.1 more precisely as follows.
Question 5.3.2. For which conformal gauges G1 , G2 of conformal di-
mension zero is it true that
(1) Cdim G3α = 0 for each α?
(2) Cdim G3α = 0 for some α?
The answer to the corresponding question for conformal Assouad dimen-
sion is known. If G1 and G2 have conformal Assouad dimension zero, then
G3α has conformal Assouad dimension zero for every α. Again, we use the
characterization of spaces of conformal Assouad dimension zero in Theorem
5.1.9. It suffices to prove that the doubling property (finiteness of the As-
souad dimension) and uniform disconnectedness are preserved by product
constructions. For the former, we note that
(5.3.3) dimA (Z1 × Z2 ) ≤ dimA Z1 + dimA Z2
for any metric spaces Z1 , Z2 (this is an easy exercise using the definition).
For the latter we use Lemma 5.1.10, noting that ∞ products of ultrametrics
are ultrametric and that the 2 and ∞ product metrics are bi-Lipschitz
equivalent.
In view of (5.3.1), we also ask the following questions: For which metric
spaces (Z1 , d1 ), (Z2 , d2 ) do we have
Cdim(Z1 × Z2 ) ≥ Cdim Z1 + Cdim Z2 ?
For which spaces does equality hold? More precisely, we restate these ques-
tions as follows.
Question 5.3.3. For which conformal gauges G1 , G2 is it true that
(1) Cdim G1 + Cdim G2 ≤ Cdim G3α for each α?
 5.3. CONFORMAL DIMENSION AND PRODUCTS 75

(2) Cdim G1 + Cdim G2 ≤ Cdim G3α for some α?

(3) Cdim G1 + Cdim G2 = Cdim G3α for each α?
(4) Cdim G1 + Cdim G2 = Cdim G3α for some α?
The inequality in (5.3.3) suggests that the appopriate version of these
questions for conformal Assouad dimension is the following.
Question 5.3.4. For which conformal gauges G1 , G2 is it true that
(1) CdimA G1 + CdimA G2 ≥ CdimA G3α for each α?
(2) CdimA G1 + CdimA G2 ≥ CdimA G3α for some α?
(3) CdimA G1 + CdimA G2 = CdimA G3α for each α?
(4) CdimA G1 + CdimA G2 = CdimA G3α for some α?
It is easy to see that the answer to Question 5.3.4(2) is always yes, for
all gauges G1 , G2 . This comes from (5.3.3) and the definition of conformal
dimension as an infimum.
Finally, observe that it follows from Proposition 4.1.11 that
Cdim(Z × [0, 1]) ≥ Cdim Z + 1 = Cdim Z + Cdim[0, 1]
for every compact set Z contained in a Euclidean space. Moreover, equality
holds if Z itself is minimal for conformal dimension.
We conclude this section with an example showing that equality need
not hold in (5.3.3), even for subsets of Euclidean space.
Example 5.3.5 (Larman). For each δ > 0, there exist countable com-
pact sets S1 , S2 ⊂ R, each of which has Assouad dimension one, so that
dimA (S1 × S2 ) < 1 + δ.
We sketch the construction of such sets. Choose a sequence of positive
integers nm = am , m = 1, 2, . . ., for some large positive integer a. Let
Im = [2−nm+1 , 2−nm ] and let Km = (2−nm+1 Z) ∩ Im be a 2−nm+1 -dense set
of points in Im . Observe that the cardinality of Km is equal to 2nm+1 −nm ,
and tends to infinity with m.
Now set
 
S1 = K4p and S2 = K4q+2 .
p≥1 q≥0
To see that S1 and S2 have Assouad dimension one, we consider balls
B(xm , rm ) = (xm − rm , xm + rm ) ∩ Sj in Sj with center xm = 2−nm+1 and
radius rm = 2−nm for suitable m (equal to 4p or 4q + 2, respectively). Let
m = 2nm −nm+1 ; then m rm = 2−nm+1 is the separation between points in
Km . It follows that
1
N (B(xm , rm ), m rm ) ≥ 2nm+1 −nm = .
m
Since this holds for all m, we conclude that dimA Sj ≥ 1; since Sj ⊂ R we
have dimA Sj = 1. (This conclusion can also be established by computing
weak tangents, as in the following chapter.)
76 5. SETS AND SPACES OF CONFORMAL DIMENSION ZERO

We now indicate the reason why S1 ×S2 has Assouad dimension bounded
above by 1 + δ. We must show that

N (B(x, r), r) ≤ C−1−δ

for all balls B(x, r) ⊂ S1 × S2 and all 0 <  < 2, for some finite constant
C independent of x and r. Without loss of generality we may use the ∞
product metric, in which case

B(x, r) = (J1 ∩ S1 ) × (J2 ∩ S2 )

is the product of balls in the factor spaces (here J1 and J2 are intervals). We
assert that the worst case (i.e., the largest value for N (B(x, r), r)) occurs
when r = 2−n2k and r = 2−n2+1 for some integers k and . We have

N ((J1 ∩ S1 ) × (J2 ∩ S2 ), r) ≤ N (J1 ∩ K4p , r)N (J2 ∩ K4q+2 , r).
p,q

We let m = 2p or m = 2q + 1. Again the worst case (for the terms N (J1 ∩

K4p , r) and N (J2 ∩ K4q+2 , r)) occurs when K4p or K4q+2 is contained in
the corresponding interval J1 or J2 (in this case, 2−n2m ≤ 2−n2k so k ≤ m)
and 2−n2m+1 ≥ 2−n2+1 (so m ≤ ). We obtain N (K2m , r) ≤ 2n2m+1 −n2m .
Collecting all of this information, we find

N ((J1 ∩ S1 ) × (J2 ∩ S2 ), r) ≤ 2n4p+1 +n4q+3 −n4p −n4q+2
p,q
k≤2p≤,k≤2q+1≤

Since 2p ≤  and 2q + 1 ≤  we have 4p + 1 ≤ 2 + 1 and 4q + 3 ≤ 2 + 1.

However, 4p + 1 is congruent to 1 mod 4 while 4q + 3 is congruent to 3
mod 4, thus, one of these two values is in fact bounded above by 2 − 1. In
conclusion, n4p+1 + n4q+3 ≤ n2+1 + n2−1 .
In a similar fashion, we show that n4p + n4q+2 ≥ n2k + n2k+2 .
Then

N ((J1 ∩ S1 ) × (J2 ∩ S2 ), r) ≤ C( − k)2 2n2+1 +n2−1 −n2k −n2k+2 .

By the choice of nm , we have

1
n2+1 + n2−1 − n2k − n2k+2 = (1 + )(a2+1 − a2k )
a2
and
(1 + δ)(n2+1 − n2k ) = (1 + δ)(a2+1 − a2k ).
Choose a so that a−2 < δ. Then
 
( − k)2 2n2+1 +n2−1 −n2k −n2k+2 = O 2(1+δ)(n2+1 −n2k ) .

Since −1 = 2n2+1 −n2k , this finishes the proof of Example 5.3.5.
 5.4. TUKIA’S EXAMPLE 77

5.4. Tukia’s example

We sketch the construction of singular quasisymmetric maps of the real
line as in Tukia’s example 5.2.2. Our goal is to describe a quasisymmetric
map f : [0, 1] → [0, 1] and an Fσ set S∞ ⊂ [0, 1] such that both f (S∞ ) and
[0, 1] \ S∞ have Hausdorff dimension strictly less than one.
The construction is reminiscent
 of the proof of Proposition 5.1.2. Fix t,
0 < t < 1, and let W∗ = m≥0 {0, 1}m be the symbol space of all finite words
in the letters 0 and 1. We define closed subintervals Jw (t) of [0, 1] indexed
by the elements w of W∗ and depending on t as follows: J∅ (t) = [0, 1] and for
any word w ∈ W∗ of length m, Jw0 (t) is the closed subinterval of Jw (t) so
that the left endpoints coincide and |Jw0 (t)| is equal to either t|Jw (t)| (if m
is even) or (1 − t)|Jw (t)| (if m is odd), while Jw1 (t) = Jw (t) \ Jw0 (t). Thus
Jw (t) is decomposed into two intervals of lengths t|Jw (t)| and (1 − t)|Jw (t)|
in a pattern which alternates with the parity of the length of the word w.
Note that the collection of intervals {Jw ( 12 )}w∈W∗ is precisely the collection
of dyadic subintervals [j2−m , (j + 1)2−m ], m ≥ 0, j = 0, 1, . . . , 2m − 1.
Define an increasing homeomorphism g (t) from [0, 1] to itself by speci-
fying that g (t) (Jw ( 12 )) = Jw (t) for all w ∈ W∗ . It can be shown that g (t) is
quasisymmetric for some η = η (t) depending only on t. In fact, it suffices to
verify that the lengths of Jw (t) and Jw (t) are comparable whenever w and
w are words of equal length and Jw ( 12 ) ∩ Jw ( 12 ) = ∅. Note that
 2  2 
1−t t
|Jw (t)| ≤ max , |Jw (t)|
t 1−t

for all such w and w .

Fix t = 12 and consider f = g (t) ◦ (g (1−t) )−1 . By selecting subcollections
of the intervals {Jw ( 12 )} of each length m, we generate a Cantor set C in
the usual fashion. It is possible to arrange such a selection so that there
exists s = s(t) < 1 for which (i) dim g (t) (C) ≤ s < 1 and (ii) the Hausdorff
s-content of g (1−t) ([0, 1] \ C) is arbitrarily small. We omit the details of this
selection. Denote by Cs,n a Cantor set satisfying dim g (t) (Cs,n ) ≤ s and

H∞
s (1−t)
g ([0, 1] \ Cs,n ) < 2−n
 !
and let C∞ = n≥1 m≥n Cs,m and S∞ = g (1−t) (C∞ ). Then

dim f (S∞ ) = dim g (t) (C∞ )

 "
= dim g (t) (Cs,m )
n≥1 m≥n

≤ max dim g (t) (Cs,n ) ≤ s < 1

n≥1
78 5. SETS AND SPACES OF CONFORMAL DIMENSION ZERO

and for each n ≥ 1,

H∞
s
([0, 1] \ S∞ ) = H∞
s
(g (1−t) ([0, 1] \ C∞ ))
⎛ ⎞

s ⎝
≤ H∞ g (1−t) ([0, 1] \ Cs,m )⎠
m≥n

≤ H∞
s
(g (1−t) ([0, 1] \ Cs,m )) < 21−n ;
m≥n

letting n → ∞ yields H∞
s ([0, 1] \ S∞ ) = 0 so dim([0, 1] \ S∞ ) ≤ s < 1.

5.5. Notes
Porosity is an important geometric notion which arises in numerous con-
texts: dynamical systems, geometric measure theory, PDE, etc. A thorough
survey of various notions of porosity and their interrelations can be found
in [181]. Koskela and Rohde [93] introduced mean porosity which has been
useful in complex dynamics. Martio and Vuorinen [115] discuss the interplay
between porosity, Whitney decompositions and dimension of the boundary.
The equivalence between porosity and positive Assouad codimension
(Proposition 5.1.4) can be found in Luukkainen’s comprehensive study of
Assouad dimension [107]. Proposition 5.1.2 and Theorem 5.1.8 are taken
from [170]. Theorem 5.1.9 is a strengthened form of another result from
[170]. Theorem 5.1.12 is due to Kovalev [95] and is a key result in the
theory of conformal dimension.
Uniform disconnectedness was introduced by David and Semmes [47,
Chapter 15], where also Proposition 5.1.7 appears. Lemma 1.2.19 origi-
nates in Tukia and Väisälä’s article [166], which inaugurated the study of
quasisymmetric maps on metric spaces.
Tukia’s example 5.2.2 originates in [165]. Examples of this type arise
naturally in the study of Fuchsian groups and quasisymmetric welding. The
application of Example 5.2.2 to the problem of conformal dimensions of
unions was pointed out to the authors by Hakobyan. Theorem 5.2.4 is
original. Note that the corresponding result for countable unions is not
true. Examples were given by Hakobyan [73, Corollary 1.4] and Bishop–
Tyson [17, Corollary 2]. The product formula (5.3.1) was shown first by
Marstrand [114] for subsets of Euclidean space and later extended to metric
spaces by Howroyd [79], [80]. For examples of sets where strict inequality
in (5.3.1) occurs, see Marstrand [114] or Falconer [52, Example 7.8]. The
product formula (5.3.3) is an easy exercise. The fact that strict inequality
can occur has been a point of some confusion in the literature. The example
which we give (Example 5.3.5) comes from a paper of Larman [101, Theorem
9].
 CHAPTER 6

Gromov–Hausdorff tangent spaces and conformal

dimension

In this chapter we consider the connection between conformal Assouad

dimension and the moduli of curve families in tangent spaces of a given
metric space. We show that the conformal Assouad dimension of a tangent
space gives a lower bound for the conformal Assouad dimension of the space
itself; in particular, the existence of a positive modulus family of curves in
a weak tangent provides an obstruction to lowering dimension.
Amazingly, Keith and Laakso showed that a converse to this statement
is true (Theorem 6.2.1): if every weak tangent has few curves (in a suit-
able sense), then the conformal Assouad dimension can be lowered. This
leads to nearly optimal necessary and sufficient conditions for minimality
for conformal Assouad dimension.

6.1. Tangent spaces and lower bounds for conformal Assouad

dimension
A pointed metric space is a triple (Z, d, p), where (Z, d) is a metric space
and p ∈ Z is a base point. A pointed map between two pointed metric spaces
is a map that sends base point to base point.
In order to talk about tangent spaces for general metric spaces, we need
a notion of convergence for abstract metric spaces. An easy way to do this
is to isometrically embed our spaces into a single ambient metric space. We
first define convergence for subsets of a metric space.
Definition 6.1.1. A sequence {Fm } of nonempty closed subsets of a
metric space (M, dM ) converges to a nonempty closed subset F of M if
lim sup dist(z, F ) = 0,
m→∞ z∈F ∩B(q,R)
m

and

lim sup dist(z, Fm ) = 0,

m→∞ z∈F ∩B(q,R)

for all q ∈ M and R > 0. We interpret the supremum over an empty set to
be zero.
We can now define convergence of metric (measure) spaces.
79
80 6. TANGENT SPACES AND CONFORMAL DIMENSION

Definition 6.1.2. A sequence of complete pointed metric measure spaces

{(Zm , dm , μm , pm )} converges to a complete pointed metric measure space
(W, dW , μW , pW ) if there exists a pointed metric space (M, dM , pM ) and
pointed isometric embeddings ι : W → M , ιm : Zm → M , m ∈ N, so that
{ιm (Zm )} converges to ι(W ) in the above sense and ((ιm )# μm ) converges
to ι# μW weakly.
If we ignore the measures in the previous definition, then we have de-
fined convergence for sequences of pointed metric spaces. Up to taking
subsequences, this definition agrees with the usual definition of Gromov–
Hausdorff convergence.
Definition 6.1.3. A complete, pointed metric measure space
(W, dW , μW , pW )
is a measured weak tangent of a metric measure space (Z, d, μ) if there exist
sequences of positive real numbers (rm ), 0 < rm < diam Z, and (sm ), and
a sequence of points (pm ) in Z, so that (Z, rm −1 d, s μ, p ) converges to
m m
(W, dW , μW , pW ).
If we further require that (rm ) converges to zero, and pm = p for all m,
then we say that (W, dW , μW , pW ) is a tangent to Z at p.
As before, we can omit all mention of measures to define (unmeasured,
weak) tangents of a metric space. For the remainder of this section we will
consider only metric spaces. We will always assume that our metric spaces
are complete.
If a metric space (Z, d) is doubling, then any sequence of rescaled copies
of Z is uniformly doubling. Gromov’s compactness theorem then implies
that the sequence has a convergent subsequence. Hence any doubling metric
space (Z, d) has at least one tangent space at every point. All of the tangent
spaces are again doubling with the same doubling constant.
It is easy to find examples where some tangent spaces have smaller di-
mension, for any notion of dimension. For example, if a metric space Z has
an isolated point, then one of the tangents of Z will be the one point space.
The Hausdorff dimension of a tangent space can be larger than the am-
bient Hausdorff dimension.
−2m Z2 ∩ [0, 2−m ]2 ) ⊂ R2 , and let Z =
 Example 6.1.4. Let Zm = (2
m Zm . It is easy to see that the sequence {(Z, 2 d, (0, 0))} converges to a
m

tangent space W which contains an isometric copy of the unit square [0, 1]2 .
As W embeds isometrically into R2 , we see that dim W = 2. However,
dim Z = 0 since Z is countable. Moreover, Cdim W = 2 while Cdim Z = 0.
In view of these pathologies, we instead consider Assouad dimension.
This is much better, since the Assouad dimension of a tangent cannot in-
crease. The following proposition sharpens the earlier remark about preser-
vation of the doubling condition for (weak) tangents.
 6.1. TANGENT SPACES AND LOWER BOUNDS FOR CdimA 81

Proposition 6.1.5. If W is a (weak) tangent to Z, then dimA W ≤

dimA Z.
Proof. Denote by (δm ) the sequence of scales associated to the weak
tangent W .
Suppose that dimA Z < α < dimA W . Then for all C > 0, there exist
0 < r < R, and a set S ⊂ W , where S is an r-separated set in some ball of
radius R in W whose cardinality satisfies
#S ≥ C(R/r)α .
For sufficiently large m, we can use S to find a δm r/2-separated set Tm
in some ball of radius 2δm R in Z, whose cardinality is the same as the
cardinality of S. In particular,
 
−α 2δm R α
#Tm = #S ≥ C4 .
δm r/2
For sufficiently large C, this contradicts the inequality dimA Z < α. 
Exercise 6.1.6. Suppose that (Z, d) is a complete Ahlfors Q-regular
metric space. Then any weak tangent W is also Ahlfors Q-regular.
The convergence in Exercise 6.1.6 can be taken in either the measured
or the unmeasured sense.
Now suppose that Z1 and Z2 are doubling metric spaces and f : Z1 → Z2
is an η-quasisymmetric map. Suppose W1 is a weak tangent to Z1 . Then,
by an Arzelà-Ascoli argument, we can find a weak tangent W2 of Z2 and an
(η-)quasisymmetric map g : W1 → W2 . Hence CdimA W1 = CdimA g(W1 ) ≤
dimA W2 . Combining this construction with Proposition 6.1.5 yields
CdimA W1 ≤ dimA W2 ≤ dimA Z2 .
We have proved the following proposition.
Proposition 6.1.7. If W is a weak tangent to Z, then CdimA W ≤
CdimA Z. In particular, Cdim W ≤ CdimA Z.
The following corollary extends Proposition 4.1.8.
Corollary 6.1.8. Let (Z, d) be Ahlfors Q-regular. If Z has a weak
tangent W which contains a curve family of positive p-modulus for some
1 < p ≤ Q, then CdimA Z = ARCdim Z = Q.
Proof. By Exercise 6.1.6, W is Ahlfors Q-regular. By Proposition
4.1.8, Cdim W = Q. Then Q = Cdim W ≤ CdimA Z ≤ ARCdim Z ≤
dim Z = Q. 
The following example amplifies the discussion in Example 6.1.4.
Example 6.1.9. By considering the separated sets Zm , one sees that
dimA Z = 2. On the other hand, because Z has a tangent space W which
contains a copy of the unit square, Proposition 6.1.7 implies that 2 ≤
CdimA W ≤ CdimA Z ≤ dimA Z = 2, hence CdimA Z = 2.
82 6. TANGENT SPACES AND CONFORMAL DIMENSION

6.2. Lowering the conformal Assouad dimension

6.2.1. The Keith–Laakso theorem. Remarkably, there is a converse
to Corollary 6.1.8.
Theorem 6.2.1 (Keith–Laakso). Fix Q ≥ 1. Suppose that (Z, d, μ) is
a complete Ahlfors Q-regular metric measure space. Then Z is minimal for
conformal Assouad dimension (equivalently, for Ahlfors regular conformal
dimension) if and only if there exists a weak tangent (W, dW , μW ) of Z
which contains a curve family of positive p-modulus for some p ≥ 1. In
this case, there exists a further weak tangent (W  , dW , μW ) of Z which has
uniformly big 1-modulus.
Recall that in the context of the theorem, the conformal Assouad di-
mension and the Ahlfors regular conformal dimension agree.
Definition 6.2.2. Fix p ≥ 1.We say that a metric measure space (Z, d, μ)
has uniformly big p-modulus if μ is non-vanishing and there exists δ > 0 so
that
Modp Γ ≥ δr−p μ(B(z, r))
whenever z ∈ Z, 0 < r < diam Z and Γ is the set of all curves in B(z, r) of
diameter at least δr.
This condition is stronger than merely asserting positivity of the p-
modulus, but weaker than the Loewner condition. It becomes weaker as
the value of p increases; compare Proposition 4.1.6(vii).
We can use Theorem 6.2.1 to find better upper bounds for conformal
(Assouad) dimension. We illustrate by continuing Example 4.3.1.
Example 6.2.3. Let SC be the standard Sierpiński carpet (with metric
d induced from the plane) and let μ be the Hausdorff Q-measure on SC,
where Q = log 8
log 3 is the Hausdorff measure. Then (SC, d, μ) is complete and
Ahlfors Q-regular.
Since SC is self-similar, any weak tangent (W, dW , μW ) of SC will, at
every location and scale, consist of scaled copies of SC. By a similar ar-
gument as in the proof of Proposition 4.3.3, the p-modulus is zero on W
for every p > 1. Hence, by Theorem 6.2.1, (SC, d, μ) is not minimal for
conformal Assouad dimension:
log 8
Cdim SC ≤ CdimA SC = ARCdim SC < dim SC = .
log 3
6.2.2. Sketch of the proof of Theorem 6.2.1. Theorem 6.2.1 is
proven by showing the contrapositive statement: Suppose (Z, d, μ) is a com-
plete Ahlfors Q-regular metric measure space all of whose weak tangents have
trivial Q-modulus. Then for some Q < Q, there is an Ahlfors Q regular
metric θ and measure ν on Z so that (Z, d, μ) and (Z, θ, ν) are quasisym-
metric.
 6.2. LOWERING THE CONFORMAL ASSOUAD DIMENSION 83

In order to make the following outline of the proof more readable, we

will first restrict to the case
log(8)
(Z, d, μ) = (SC, d, HQ ), Q= ,
log(3)
where SC is the Sierpiński carpet with the Euclidean metric and the usual
Q-regular Hausdorff measure. (See Example 4.3.1.) The proof in the general
case is quite similar; see subsection 6.2.2.4 for additional remarks.
In the following we write A  B if A ≤ CB for some constant C, which
only depends on the relevant data, and A ≈ B if A  B and B  A.

6.2.2.1. The dyadic decomposition of Z = SC. The framework for build-

ing a new metric
 on Z is given by a generalized dyadic decomposition.
Let Δ = n≥0 Δn be the triadic decomposition of SC. Namely, for
n ≥ 0, we consider the collection of squares of the form
 
S = [(i − 1)3−n , i3−n ] × [(j − 1)3−n , j3−n ] ∩ SC,

where 1 ≤ i, j ≤ 3n , and μ(S) = 0. This collection of squares covers SC,

but is not pairwise disjoint as the squares may overlap along their edges.
We can adjust the sets here so that Δ satisfies:

(D1) for each n ≥ 0, S∈Δn S = Z, and
(D2) whenever S ∈ Δn , T ∈ Δm , and n ≤ m, then either T ∩ S = ∅ or
T ⊂ S.
For S ∈ Δn we denote the radius of S by rad(S) = 3−n .
Given τ ∈ (0, 1), the τ -boundary of a square S ∈ Δ is the set ∂τ (S)
defined as follows:
 
∂τ (S) := x ∈ S | d(x, Z\S) ≤ τ rad(S) ∪ x ∈ (Z\S) | d(x, S) ≤ τ rad(S) .
Suppose S ∈ Δn and T ∈ Δm are given. Let i ∈ N be maximal with the
property that S and T are contained in squares Ŝ and T̂ of Δi whose closures
intersect. We define the rough hyperbolic distance H : Δ × Δ → [0, ∞) by
H(S, T ) = |n − i| + |m − i| + 1.
Notice that H(S, T ) is roughly equal to
   
d(S, T ) rad(S)
log + 2 + log .
rad(S) + rad(T ) rad(T )
Later, we will need to space out the generations of our decomposition Δ,
to suitable subsequences of the sequence (Δn ). Given K ∈ N,
i.e., to pass 
we let ♦ = n≥0 ♦n , where ♦n = ΔnK . We call ♦ the K-refinement of Δ.
We set rad(♦n ) = rad(S), where S is any square in ♦n . For any S ∈ ♦n ,
let Children(S) denote the set of all T ∈ ♦n+1 with T ⊂ S.
We will construct a function ρ : ♦ → [0, ∞) whose value ρ(S) on any
square S ∈ ♦ describes how measure and distance near S are to be distorted
on scale rad(S).
84 6. TANGENT SPACES AND CONFORMAL DIMENSION

To any such function ρ it will be convenient to associate three other

functions. The accumulated distortion on S is given by the function
#
ρ∗ (S) = ρ(T ).
T ∈♦,S⊂T

Using ρ and ρ∗ we define two further functions ρk , ρ∗k : Z → [0, ∞), k ∈ N,

by minimizing ρ and ρ∗ respectively over all squares T ∈ ♦k near x. More
precisely, we set

ρk (x) = min ρ(T ) | T ∈ ♦k , T ∩ B(x, 10 rad(T )) = ∅
and 
ρ∗k (x) = min ρ∗ (T ) | T ∈ ♦k , T ∩ B(x, 10 rad(T )) = ∅

6.2.2.2. “A weight loss program” (after Keith and Laakso). In order to

reduce the dimension of Z, we need to find a suitable choice of ρ. The fol-
lowing theorem describes the conditions ρ must satisfy. In the remainder of
this section we will describe how to use ρ to construct a quasisymmetrically
equivalent Ahlfors regular metric θ on Z of dimension Q < Q.
Here by a (discrete) -path γ we understand a finite sequence
{γ(1), γ(2), . . . , γ(#γ)}
of points in Z, where #γ denotes the number of terms of the sequence, so
that d(γ(i), γ(i + 1)) ≤  for all 1 ≤ i < #γ.
Theorem 6.2.4. There exist τ > 0 and a > 1 so that for any sufficiently
large K ∈ N we can find Q < Q and a function ρ : ♦ → [0, ∞), where ♦ is
the K-refinement of Δ, so that:
(A1) For all S ∈ ♦,

  
μ(S)Q /Q = ρ(T )Q μ(T )Q /Q .
T ∈Children(S)

(A2) For any n ∈ N, and  any discrete 5 rad(♦n+1 )-path γ that begins
and ends in ∂τ ♦n := S∈♦n ∂τ S, we have

#γ
d(γ(1), γ(#γ)) ≤ d(γ(i), γ(i + 1))ρn+1 (γ(i)).
i=1
(A3) For every S, T ∈ ♦, we have
ρ∗ (S) ≤ aH(S,T )/K ρ∗ (T ).
Note that if Q = Q, and ρ ≡ 1, then condition (A1) follows from
additivity of measure for finite partitions, condition (A2) from the triangle
inequality, and (A3) is trivial.
Roughly speaking, (A1) implies that ρ will “de-snowflake” our metric,
(A2) that our new metric does not mix up scales too much locally, and (A3)
that the old and new metrics will be quasisymmetric, because the relative
 6.2. LOWERING THE CONFORMAL ASSOUAD DIMENSION 85

distortion of metric is controlled over both distances and scales. We will

return to the proof of Theorem 6.2.4 in the following subsection.
Before we use ρ to define a new metric, we translate (A3) from a con-
dition on squares to a condition on ρ∗k . This follows straightforwardly from
the definitions.
Lemma 6.2.5. There exists b = b(a) so that
 
1/K

d(x,y)
|k−l|+log +1
(6.2.1) ρ∗k (x) ≤ max b, b 3−k +3−l
ρ∗l (y)

for any x, y ∈ Z and k, l ∈ N.

We may choose b = a8 in Lemma 6.2.5.
We now build our new metric by constructing a sequence of metrics θn
that each consider scales larger than rad(♦n ).
Let G be the graph with vertex set V = Z × N and the following edges:
• an edge from (x, k) to (y, k) if d(x, y) ≤ 5 rad(♦k ), and
• an edge from (x, k) to (x, k + 1) for every x and k.
We will view G as a weighted graph, i.e., we will assign lengths to the
various edges. We declare edges of the first type (from (x, k) to (y, k)) to
have length d(x, y) min{ρ∗k (x), ρ∗k (y)}, and we declare edges of the second
type (from (x, k) to (x, k + 1)) to have length 5b rad(♦k )ρ∗k+1 (x).
This definition explains what we meant when we said “ρ distorts dis-
tances”. In fact, ρ∗k : Z → [0, ∞) will be the distortion factor for distances
of order rad(♦k ) = 3−kK . The presence of the (large) constant b in the
length of edges between (x, n) and (x, n + 1) ensures that different scales do
not interfere with each other too much.
For n ∈ N, let Gn be the subset of G with vertex set Vn = Z ×{1, 2, . . . , n}
and all allowable edges. Let θn : Vn × Vn → [0, ∞) be the distance function
on Vn induced by infimizing the lengths of all (finite) simplical paths between
vertices.
Lemma 6.2.6. For any fixed x, y ∈ Z, θn ((x, n), (y, n)) is a nondecreasing
function of n. If x, y ∈ Z, and d(x, y) ∈ [3−mK , 3−(m−1)K ], for some 1 ≤
m ≤ n, then
(6.2.2) θn ((x, n), (y, n)) ≈ ρ∗m (x)d(x, y).
Sketch of the proof. Provided that K is large enough depending
on τ , any 5 rad(♦n+1 )-discrete path in Z must meet ∂τ (♦n ) as it crosses
squares in ♦n .
For such a discrete path crossing an individual square of ♦n , condition
(A2) implies that there is no gain possible by travelling through Z × {n + 1}
rather than in one step across the square.
Thus, given a simplicial path from (x, n + 1) to (y, n + 1) in Gn+1 , we
can find a simplicial path from (x, n) to (x, n) in Gn of whose length is not
longer. This shows that θn (x, y) is a non-decreasing function of n.
86 6. TANGENT SPACES AND CONFORMAL DIMENSION

To get the upper bound in (6.2.2), we consider the candidate path be-
tween (x, n) and (y, n) given by
(x, n), (x, n − 1), . . . (x, m), (y, m), (y, m + 1), . . . , (y, n).
By the definition of G, there is an edge between (x, m) and (y, m) of length
bounded above by a constant times ρ∗m (x)d(x, y). Equation (6.2.1) provides
an upper bound for the sum of the remaining lengths in this path. This
upper bound is in the form of a geometric series, with ρ∗m (x)d(x, y) as the
leading term. As long as K is chosen sufficiently large, depending on b, the
geometric series converges and so the upper bound is independent of n.
For the lower bound, by the first statement of the lemma it suffices to
consider the case n = m + C, where C is chosen so that x and y are in
different cubes of ♦n .
Any simplicial path from (x, n) to (y, n) in Gn must do one of two things:
either it moves from level n to level n − 1 before leaving a ball of radius
comparable to d(x, y) around x, or it travels through such a ball on level n.
In the former case, the large cost of a jump between levels, controlled
by (6.2.1), gives the desired lower bound for the length of the path. In the
latter case, (6.2.1) and the triangle inequality gives the lower bound for the
length of the path. 
Applying (6.2.1) with l = 0 and k = m = [− log3K d(x, y)] + 1 to (6.2.2),
and assuming n ≥ m, we see that
θn ((x, n), (y, n))  b− log3K d(x,y) d(x, y) =: η(d(x, y)).
Notice that if K is chosen sufficiently large, η : [0, 1) → [0, ∞) will satisfy
η(t) → 0 as t → 0. This implies that θn distances are controlled locally,
independent of n, so we can use Arzéla-Ascoli to produce a subsequence of
the sequence (θn ) restricted to (Z, n)×(Z, n), for each n ∈ N, that converges
to a function θ : Z × Z → [0, ∞). Because of the lower bound in (6.2.2), θ
will be a metric on Z.
Lemma 6.2.7. The identity map from (Z, d) to (Z, θ) is a quasisymmetry.
Proof. Take distinct x, y, z ∈ Z. Suppose the distances d(x, y) and
d(x, z) lie in the intervals [3−kK , 3−(k−1)K ] and [3−lK , 3−(l−1)K ] respectively.
Then when d(x, y) ≤ d(x, z) it follows that
 
d(x, y)
(6.2.3) |k − l| ≤ − log3K + 1.
d(x, z)
By Lemma 6.2.6,
θ(x, y) ρ∗ (x)d(x, y)
(6.2.4) ≈ k∗ .
θ(x, z) ρl (x)d(x, z)
By (6.2.1) we have
ρ∗k (x)
(6.2.5)  b|k−l| ,
ρ∗l (x)
 6.2. LOWERING THE CONFORMAL ASSOUAD DIMENSION 87

When d(x, y) ≤ d(x, z), combining (6.2.4), (6.2.5) and (6.2.3) with the
definition of η gives the bound
 
θ(x, y) d(x, y)
η .
θ(x, z) d(x, z)
On the other hand, when d(x, y) ≥ d(x, z), it is enough to bound
   
ρ∗k (x) |k−l|
d(x,y)
log3K d(x,z) +1 d(x, y) C
b b  ,
ρ∗l (x) d(x, z)
for some constant C. 
It remains to show that (X, θ) is Ahlfors Q -regular. To do this, we use
ρ to build a new measure. We introduce a nonnegative function ν̃ defined
on ♦ by the equation
 
ν̃(S) = μ(S)Q /Q ρ∗ (S)Q .
Lemma 6.2.8. The function ν̃ extends to a measure on Z.
Proof. Condition (A1) implies that

(6.2.6) ν̃(S) = ν̃(T ), for all S ∈ ♦.
T ∈Children(S)

Given A ⊂ Z and δ > 0, cover A by squares in ♦ of diameter less than δ.

Infimize the sum of the ν̃ values of these squares, and then take the limit
as δ → 0. Call this limit ν(A). This (Carathéodory) construction always
produces a measure. By (6.2.6), ν(S) = ν̃(S) for all S ∈ ♦. 
Lemma 6.2.9. The measure ν is an Ahlfors Q -regular measure on (Z, θ).
Proof. By Lemma 6.2.7 balls in (Z, θ) can be well approximated by
balls in (Z, d), which themselves can be well approximated by squares in ♦,
so it suffices to check Ahlfors regularity for squares in ♦.
Take S ∈ ♦n for some n ∈ N. By Lemma 6.2.6,
diamθ (S) ≈ ρ∗ (S)3−nK .
Since μ is Ahlfors Q-regular, μ(S) ≈ (3−nK )Q , so
    
ν(S) = μ(S)Q /Q ρ∗ (S)Q ≈ (3−nK )Q ρ∗ (S)Q ≈ (diamθ (S))Q . 

6.2.2.3. Building ρ. Here we sketch the proof of Theorem 6.2.4 for the
case Z = SC. The basic idea is that since the modulus of all curve families
is zero, we can find an admissible function ρ that keeps a cost to travelling
across squares but whose integral is less than one.
We assume that τ is fixed. We will find a so that for all sufficiently large
K, we can find ρ satisfying (A1),(A2) and (A3).
In the proof of Lemma 4.3.4 we saw that, given any  > 0, for any
sufficiently large K1 ∈ N there is a function ρ1 : SC → [0, ∞) so that
• ρ1 is constant on squares of ΔK1 ,
88 6. TANGENT SPACES AND CONFORMAL DIMENSION

• ρ1 is admissible for the family of all paths joining the left and right
edges of SC, and
• The integral of ρQ1 with respect to the measure μ is less than .
1
Fix  = 100 , say, and the corresponding K1 and ρ1 . Essentially, ρ1 lets us
obstruct paths in SC at very little cost.
Suppose we are given a square T ∈ Δn , for some n ∈ N. Given a fixed
choice of τ > 0, we can find an annulus A(x, r, 2r) = B(x, 2r) \B(x, r) inside
T \ ∂τ (T ). We can suppose that r ≈ rad(T ), with comparability constant
depending only on τ . Choose K2 ∈ N, with 3−K2 rad(T ) ≈ r, so that all
paths from B(x, r) to T \ B(x, 2r) must pass through an entire square of
Δn+K2 .
We can use ρ1 suitably scaled and oriented to define a function ρ on T
so that:
• ρ is constant on squares of Δn+K2 +K1 ,
• ρ ≥ 1 outside B(x, r), and ρ ≡ 1 outside B(x, 2r),
• ρ =  inside B(x, r),
• the integral of ρ along any rectifiable path joining T \ B(x, 2r) to
B(x, r) must be at least 2r, and
• the integral of ρQ over T is strictly less than μ(T ).

We set K = K1 + K2 , and define ρ : ♦ → [0, ∞), where ♦ = n≥0 ΔnK ,
by repeating this construction for all T ∈ ♦, and setting ρ(S) = 1 where
S ∈ ♦0 . We now check the three conditions.
First we show that (A1) is satisfied. By construction, there is a δ > 0
so that for any S ∈ ♦, we have

ρ(T )Q μ(T ) ≤ (1 − δ)μ(S).
T ∈Children(S)

Therefore,
  
ρ(T )Q μ(T )Q /Q
T ∈Children(S)
  
Q −Q
Q −Q
≤ max ρ(T ) μ(T ) Q ρ(T )Q μ(T )
T ∈Children(S)
T ∈Children(S)
 
Q −Q Q−Q
Q −Q 
≤ max ρ(T ) μ(T ) Q μ(S) Q (1 − δ)μ(S)Q /Q .
T ∈Children(S)

Since ρ is bounded away from zero, by taking Q sufficiently close to Q, we


can ensure that the right hand side is at most μ(S)Q /Q . We then increase
ρ inside B(x, r) from  to some value no more than one so that we have an
equality in condition (A1).
Condition (A2) follows from the admissibility for ρ1 , and the fact that ρ1
is constant on squares. Any 5 rad(♦n+1 )-path that begins and ends in ∂τ ♦n
can do one of two things. Either it avoids the inside of the annuli where
we modified ρ, in which case ρ is at least one, so the triangle inequality
 6.3. NOTES 89

suffices. Alternatively, it may have some subpath that travels from outside
some B(x, 2r) into B(x, r) and back out again, but then our construction of
ρ implies that this subpath picks up at least 4r as required.
Finally, we consider condition (A3). Suppose we have S ∈ ♦n and
T ∈ ♦m , for 1 ≤ n ≤ m. Let U be the square in ♦n that contains T . Since
ρ has values in [ C1 , C], it follows that
ρ∗ (U ) ≈ ρ∗ (T ),
with approximation constant (C 2 )m−n .
Let l be maximal so that we have U ⊂ V ∈ ♦l , and S ⊂ W ∈ ♦l ,
and W and V are the same, or adjacent, squares. For all scales k < l, if
S ⊂ V  ∈ ♦k , and U ⊂ W  ∈ ♦k , then ρ(V  ) = ρ(W  ), since either V  = W  ,
or V  and W  are both near ∂τ (♦k−1 ), and so ρ(V  ) = 1 = ρ(W  ). Hence
ρ∗ (S) ≈ ρ∗ (V ) ≈ ρ∗ (W ) ≈ ρ∗ (U )
with approximation constant (C 2 )(n−l)+1+(n−l). Therefore
ρ∗ (S) ≈ ρ∗ (T )
with approximation constant C (n−l)+1+(m−l) ≤ (C  )H(S,T )/K .

6.2.2.4. Comments on the general case. The general case is not vastly
different to the above argument. We need to use a measurable dyadic de-
composition rather than a Euclidean one, but this does not substantially
affect the argument given in subsections 6.2.2.1 and 6.2.2.2.
However, we need a more general argument to prove the analogous state-
ment to Theorem 6.2.4. Since the modulus vanishes, one gets admissible
functions with small integral. However, condition (A2) may not be satisfied
for discrete paths. To do this requires showing that a suitable “discrete
modulus” of the family of curves joining the inside to the outside of certain
annuli is uniformly sufficiently small on enough scales and locations. This
follows from the weak tangent condition in Theorem 6.2.1: if the discrete
modulus stays large on scales tending to zero, one shows that some weak
tangent will have positive modulus.

6.3. Notes
Most of the material in this chapter is taken from Keith–Laakso [91].
In particular, our definitions for the various modes of Gromov–Hausdorff
convergence agree with theirs. Gromov’s compactness theorem is the fun-
damental tool for the analysis of metric tangent spaces. It can be found
in Gromov [62] or Burago–Burago–Ivanov [39]. The remaining material in
section 6.1 is original.
Theorem 6.2.1 was proved by Keith and Laakso in [91]. It plays a key
role in the work of Bonk and Kleiner on quasisymmetric uniformization
of 2-spheres and Cannon’s conjecture. We review this application in the
following chapter. Example 6.2.3 is mentioned in [91, Corollary 1.0.4], where
90 6. TANGENT SPACES AND CONFORMAL DIMENSION

a brief sketch of the proof is given. We have amplified their discussion. The
argument also works for the usual Menger sponge as well as other self-similar
fractals.
We mention another result in the spirit of Theorem 6.2.1, taken from
[172]. Let us call a map f : (Z1 , d1 ) → (Z2 , d2 ) a snowflake map if there
exists α > 0 and L < ∞ so that
1
d1 (z, z  )α ≤ d2 (f (z), f (z )) ≤ Ld1 (z, z  )α .
L
In other words, f is bi-Lipschitz as a map from the snowflake metric space
(Z1 , dα1 ) to (Z2 , d2 ) (if α ≤ 1), or as a map from (Z1 , d1 ) to the snowflake
1/α
metric space (Z2 , d2 ) (if α ≥ 1). See also Theorem 3.2.4. Recall that
dim Z2 = dim Z1 /α. The snowflake relation on metric spaces is an equiva-
lence relation.
Definition 6.3.1. The snowflake dimension, Sdim Z, of a metric space
(Z, d) is the infimum of the Hausdorff dimensions of all metric spaces which
are snowflake equivalent to (Z, d).
Example 1.2.10(2) shows that snowflake maps are quasisymmetric. Thus
Sdim Z ≥ Cdim Z for all Z. The following theorem from [172] characterizes
minimality for snowflake dimension within a certain class of metric spaces.
Theorem 6.3.2 (Tyson–Wu). Let (Z, d) be a compact doubling metric
space which admits a bi-Lipschitz embedding into a uniformly convex Banach
space. Then (Z, d) is minimal for snowflake dimension if and only if there
exists a weak tangent of (Z, d) which contains at least one rectifiable curve.
For examples illustrating Theorem 6.3.2, see the notes to chapter 8.
 CHAPTER 7

Ahlfors regular conformal dimension

In many applications of conformal dimension, the appropriate invariant

to consider is Ahlfors regular conformal dimension. In rough terms, this
assertion can be explained by the fact that typical constructions in the theory
of hyperbolic groups or the dynamics of rational maps generate metrics with
strong symmetries and scaling properties which are consequently Ahlfors
regular.
More to the point, many natural conformal gauges contain a plentiful
supply of Ahlfors regular metrics. We already hinted at this phenomenon
in chapter 3.4 where we showed that the canonical conformal gauge on the
boundary of any finitely generated non-elementary hyperbolic group is dou-
bling and uniformly perfect. The existence of Ahlfors regular metrics in such
a gauge then follows from Proposition 2.2.6.
In section 7.1 we review the David–Semmes quasisymmetric deformation
theory and sketch the proof of Proposition 2.2.6. Section 7.2 is devoted to
work of Bonk and Kleiner verifying Cannon’s conjecture 3.4.10 under an
additional hypothesis involving Ahlfors regular conformal dimension. In
section 7.3 we give a detailed presentation of work of Bourdon and Pajot
on the relationship between Ahlfors regular conformal dimension and the
critical p cohomology exponent. Finally, in section 7.4 we show how the
Ahlfors regular conformal dimension of certain dynamically defined metrics
on the sphere can be computed or estimated in terms of combinatorial data.

7.1. David–Semmes deformation theory

Any doubling measure on a doubling uniformly perfect metric space can
be used to deform the metric to a quasisymmetrically equivalent Ahlfors
regular metric. Since every doubling metric space supports doubling mea-
sures, this procedure constructs a variety of Ahlfors regular metrics in any
doubling uniformly perfect conformal gauge.
Let (Z, d, μ) be a complete doubling metric measure space which is uni-
formly perfect. Define δ : Z × Z → [0, ∞) by setting
δ(z1 , z2 ) = μ(Bz1 z2 ),
where Bxy := B(x, d(x, y)) ∪ B(y, d(x, y)). It is clear that δ is a symmetric,
finite-valued distance function which vanishes only along the diagonal in
Z × Z. It is typically not a metric since the triangle inequality may be
violated. However, from the doubling condition one quickly verifies that δ
91
92 7. AHLFORS REGULAR CONFORMAL DIMENSION

is a quasimetric (Definition 3.2.8). By Theorem 3.2.9, d := δ  is a metric

for sufficiently small .
The identity map from (Z, d) to (Z, d ) is quasisymmetric. Thus (Z, d ) is
also doubling and uniformly perfect. Moreover, the triple (Z, d , μ) is Ahlfors
regular of dimension 1/. In fact, for any z ∈ Z and 0 < r < diamd Z there
exists 0 < R < diamd Z so that Bd (z, R/C) ⊂ Bd (z, r) ⊂ Bd (z, R) and
μ(Bd (z, R)) is comparable to r1/ , for some universal constant C. This
follows from the uniform perfectness and doubling conditions on (Z, d) and
(Z, d ). We conclude that the conformal gauge of (Z, d) contains Ahlfors
regular metrics adapted to each doubling measure.
Suppose in particular that the measure μ is s-homogeneous for some
s > 0 (Definition 4.6.1). Recall that such measures can be constructed for
each s > dimA Z. In this case, a careful analysis of the proof of Theorem
3.2.9 reveals that δ(z1 , z2 )1/s = μ(Bz1 z2 )1/s is comparable to a metric ds . In
other words, we may select  = 1/s in Theorem 3.2.9.
We now sketch the proof of Proposition 2.2.6. Let G be a doubling
and uniformly perfect conformal gauge, and suppose that ARCdim G >
CdimA G. Choose a metric space (Z, d) ∈ G and choose s so that
(7.1.1) ARCdim G > s > dimA Z.
Using the David–Semmes deformation theory as in the previous paragraph,
we construct a new metric space (Z  , d ) ∈ G which is Ahlfors s-regular. This
violates the first inequality in (7.1.1).

7.2. Cannon’s conjecture and Ahlfors regular conformal

dimension
Recall that Cannon’s conjecture 3.4.10 predicts that hyperbolic group
boundaries which are topological 2-spheres are necessarily quasisymmetric
spheres. Bonk and Kleiner proved this conjecture under an additional hy-
pothesis phrased in terms of the Ahlfors regular conformal dimension.
Theorem 7.2.1 (Bonk–Kleiner). Let G be a Gromov hyperbolic group
whose boundary ∂∞ Γ is homeomorphic to S2 . Let G be the canonical confor-
mal gauge on ∂∞ G. Assume that ARCdim G is achieved. Then the Euclidean
metric on S2 lies in G. In particular, there exist Ahlfors regular Loewner
metrics in G. The group G acts discretely and cocompactly by isometries on
HR3 .
Sketch of the proof. Let G be the conformal gauge of the visual
metrics on ∂∞ G. By assumption, there exists a metric d in G so that
(∂∞ G, d) is Ahlfors Q-regular with Q = ARCdim ∂∞ G. The Keith–Laakso
theorem 6.2.1 guarantees the existence of a weak tangent W of (∂∞ G, d) on
which the Q-modulus is nontrivial. However, hyperbolic group boundaries
admit a sufficiently strong quasi-self-similar structure which permits one to
promote this conclusion from W to the original space (∂∞ G, d). Now re-
call that G acts on ∂∞ G via uniformly quasi-Möbius maps, discretely and
 7.3. p COHOMOLOGY OF GRAPHS 93

cocompactly on the space of distinct triples in ∂∞ G. Using the quasisym-

metric (equivalently, quasi-Möbius) quasi-invariance of the Q-modulus on
Ahlfors Q-regular spaces [167], one further promotes the nontriviality of
the Q-modulus on (∂∞ G, d) to the Loewner condition. The desired conclu-
sion now follows from another theorem of Bonk and Kleiner (Theorem 7.2.2)
on quasisymmetric uniformization of S2 . 
Theorem 7.2.2 (Bonk–Kleiner). Let (Z, d) be a metric space which is
homeomorphic to S2 . If (Z, d) is Ahlfors Q-regular and Loewner for some
Q ≥ 2, then Q = 2 and (Z, d) is quasisymmetrically equivalent to S2 .
Theorem 7.2.1 reduces Cannon’s conjecture to the problem of showing
that for hyperbolic group boundaries which are topological 2-spheres, the
Ahlfors regular conformal dimension of the canonical conformal gauge is
realized. The self-similarity of the situation makes this conclusion equivalent
to the existence of an Ahlfors regular Loewner metric. The prediction is that
the only gauge of this type is the standard Euclidean one.
We emphasize that there are metrics on S2 which are not in the standard
Euclidean gauge, yet for which the Ahlfors regular conformal dimension is
realized at a value strictly larger than 2. Examples of such metrics arise
from the dynamics of rational maps. See Example 7.4.11.

7.3. p cohomology of graphs

In this section we describe a striking connection between conformal
gauges and the p cohomology groups of hyperbolic graphs, which yields es-
timates for the Ahlfors regular conformal dimension in terms of the critical
p cohomology exponent. This circle of ideas goes back to work of Gromov
and Pansu. Theorems 7.3.1 and 7.3.2 are due to Bourdon and Pajot.
7.3.1. Overview. To each compact, doubling, uniformly perfect con-
formal gauge G there corresponds a quasi-isometry class of Gromov hyper-
bolic graphs G whose Gromov boundaries ∂∞ G lie in G. The first p coho-
mology group of G, denoted p H 1 (G), will be defined as a suitable quotient
of the space of functions f : VG → R defined on the vertex set VG whose
differential df : EG → R (defined on the edge set EG ) is p-summable. The
isomorphism type of p H 1 (G) turns out to be a quasi-isometry invariant of
the underlying graph G, and thus an invariant of G. We will write p H 1 (G)
for the isomorphism class of first p cohomology groups determined by G.
The spaces p H 1 (G) increase with increasing values of p, which leads us to
introduce the critical p cohomology exponent
(7.3.1) p(G) := inf{p ≥ 1 : p H 1 (G) = {0}}.
This value is closely related to the conformal dimension of the boundary
gauge, as evidenced in the following two theorems.
Theorem 7.3.1 (Bourdon–Pajot). If G is a compact, doubling, uni-
formly perfect conformal gauge, then p(G) ≤ ARCdim G.
94 7. AHLFORS REGULAR CONFORMAL DIMENSION

Theorem 7.3.2 (Bourdon–Pajot). If G is a compact, doubling, uni-

formly perfect conformal gauge which contains an Ahlfors regular Loewner
metric space (Z, d) of dimension Q, then p(G) = Q.
In the setting of Theorem 7.3.2, Corollary 4.2.2 implies that p(G) =
ARCdim G = Cdim G. Thus equality of critical p cohomology exponent
and conformal dimension is necessary for the existence of Ahlfors regular
Loewner metrics in a given conformal gauge. In Theorem 7.2.1 we saw
the relevance of the existence of such metrics for the characterization of
conformal gauges on the sphere. See the notes to this chapter for further
remarks and examples.

7.3.2. Construction of the Gromov hyperbolic graphs G. Let

(Z, d) be a compact, doubling, uniformly perfect metric space. We may
assume without loss of generality that diam Z < 1. For each m ∈ N, let Zm
be an e−m -net in Z, i.e., a maximal e−m -separated subset. The construction
of such nets is a standard application of Zorn’s lemma. We next define a
graph G as follows:

(i) the vertex set VG of G is equal to m≥0 Zm , and
(ii) the edge set EG consists of pairs (a, b) ∈ VG × VG so that a ∈ Zm
and b ∈ Zn with |m − n| ≤ 1 and B(a, e−m ) ∩ B(b, e−n ) = ∅.
We denote the edge joining a to b by ab. We equip G with a length metric
ρ by identifying each edge with a Euclidean line segment of length one.
The set Z0 contains exactly one element which we denote by o. Each in-
finite geodesic ray in G starting at o intersects VG in a sequence o, v1 , v2 , . . .,
where vm ∈ Zm and B(vm , e−m ) ∩ B(vm+1 , e−m−1 ) is nonempty. Consider-
ing these elements as points in G, we have ρ(vm , vm+1 ) = 1 for all m. In
(Z, d), (vm ) is Cauchy. We obtain a map from the collection R of infinite
geodesic rays out of o onto the original metric space (Z, d).
The doubling assumption on Z ensures that G has bounded valence. We
also note that the graph (G, ρ) is Gromov hyperbolic. In fact, for a, b ∈ VG ,
the quantities e−(a|b)o and d(a, b)+e−m +e−n are comparable (where a ∈ Zm
and b ∈ Zn ). The uniform perfectness assumption is used here to relate the
diameter of a ball with its radius. Note that diam B(z, r) ≥ cr in any
c-uniformly perfect metric space.
As described in chapter 3, the Gromov boundary ∂∞ G can be identified
with the collection R of geodesic rays out of o. Relative to this identifica-
tion, the map from ∂∞ G to (Z, d) described above becomes a bi-Lipschitz
equivalence, when ∂∞ G is equipped with the visual metric with parameter
a = e. Indeed, observe that d(vm x , v y ) + e−m + e−n → d(x, y) as m, n → +∞.
n
Theorem 3.2.13 tells us that the quasi-isometry type of G depends only on
the quasisymmetry type of (Z, d), i.e., on the associated conformal gauge G.

7.3.3. The first p cohomology group of G. Let p (VG ), resp. p (EG ),

denote the collection of p-summable functions f : VG → R, resp. g : EG → R.
 7.3. p COHOMOLOGY OF GRAPHS 95

The differential of f : VG → R, denoted df , is the map from EG to R defined

by
df (ab) = f (b) − f (a).

Since G has bounded valence, df ∈ p (EG ) whenever f ∈ p (VG ). Since G is

connected, df = 0 if and only if f is a constant.
The first p cohomology group of G, written p H 1 (G), is defined as the
quotient space
 
{f : VG → R|df ∈ p (EG )} / p (VG ) ⊕ R · 1 ,

where 1 denotes the constant function: 1(v) = 1 for all v ∈ G.

When p < q we have p (VG ) ⊂ q (VG ) and p (EG ) ⊂ q (EG ). It fol-
lows that p H 1 (G) ⊂ q H 1 (G) in the sense that every equivalence class
in p H 1 (G) contains a representative which defines an equivalence class in
q H 1 (G). We define the critical p cohomology exponent p(G) as in (7.3.1).

Proposition 7.3.3. Let G and G be hyperbolic graphs associated to

compact, doubling, uniformly perfect metric spaces as in section 7.3.2. Each
quasi-isometry G → G induces an isomorphisms p H 1 (G) → p H 1 (G ) of
topological vector spaces for each p ≥ 1. In particular, the isomorphism type
of p H 1 (G) depends only on the conformal gauge of (Z, d).

We emphasize this point by writing p H 1 (G) for the isomorphism class

and p(G) for the corresponding critical exponent defined in (7.3.1).

7.3.4. Identifying functions on G and Z = ∂∞ G. In the rest of

this section we assume that Z is equipped with a Borel measure μ which is
positive and finite on balls. Recall that the space R of geodesic rays in G
out of o surjects to Z as follows:

(7.3.2) γ ∈ R
→ z
z = lim vm , where γ ∩ VG = {o, v1z , v2z , . . .}.
m→∞

Define the boundary map ∂∞ f : Z → R for f : VG → R by the rule f∞ (z) =

limm→∞ f (vmz ), where (v z ) is as in (7.3.2). It is easy to see that this is
m
well-defined. The map ∂∞ descends to a map on the quotient {f : VG →
R|df ∈ p (EG )}/p (VG ) and thence to a map ∂∞ from p H 1 (G) to {u : Z →
R}/constants.

Proposition 7.3.4. Assume that (Z, d, μ) is Ahlfors Q-regular for some

Q > 0. Then ∂∞ maps p H 1 (G) to Lp (Z, μ)/constants. In fact, there exists
C > 0 so that
||∂∞ f − f (o)||Lp (Z,μ) ≤ C||df ||p(EG )

for all f ∈ p H 1 (G).

96 7. AHLFORS REGULAR CONFORMAL DIMENSION

The proof is straightforward:

   p

|∂∞ f (z) − f (o)| dμ(z) =
p
df (vm vm+1 ) dμ(z)
z z
Z Z m≥0
  
≤C |df (vm
z vz p mQ
m+1 )| e dμ(z)
Z m≥0

(insert emQ/p e−mQ/p and use Hölder’s inequality)

 
=C emQ |df (ab)|p μ{z : vm
z z
= a, vm+1 = b}.
m≥0 a∈Zm ,b∈Zm+1
z = a and v z
The set of points z such that vm −m )
m+1 = b is contained in B(a, e
and hence has measure at most Ce−mQ . Thus
  
|∂∞ f (x) − f (o)|p dμ(x) ≤ C |df (ab)|p ≤ C||df ||pp(EG ) .
Z m≥0 a∈Zm ,b∈Zm+1
 
For E ⊂ Z, μ(E) > 0, let E f dμ = μ(E)−1 E f dμ denote the average
value of an integrable function f on E. For each locally integrable u : Z → R,
we define a map fu : VG → R by the rule

fu (v) = u dμ,
Bv
where Bv = B(v, e−m )
if v ∈ Zm . The Lebesgue Differentiation Theorem
gives ∂∞ fu = u μ-a.e., which means that f
→ ∂∞ f can be viewed as a right
inverse for u
→ fu .
Proposition 7.3.5. If (Z, d, μ) is Q-regular and p > Q, then dfu ∈
p (EG ) if and only if
 
(7.3.3) |u(x) − u(y)|p dμ(x)dμ(y) < ∞.
Ba Bb
ab

One direction in this proposition is completely trivial:

p
   
p
||dfu ||p (EG ) = |fu (b)−fu (a)| =
p (u(x) − u(y)) dμ(x)dμ(y)

Ba Bb
ab ab
which is no more than the quantity in (7.3.3) by Hölder’s inequality. The
other direction is similar to the previous argument, using the triangle in-
equality and a chaining argument to bound the quantity in (7.3.3) from
above by a multiple of ||dfu ||p (EG ) .
Corollary 7.3.6. If u : Z → R is Lipschitz, then dfu ∈ p (EG ).
In fact, the map f
→ ∂∞ f is a two-sided inverse for the map u
→ fu
in the following sense: If f : VG → R satisfies df ∈ p (EG ) and u = ∂∞ f ,
then fu − f ∈ p (VG ). In particular, [fu ] = [f ] as elements of p H 1 (G). The
proof is similar to the previous ones; we omit it.
 7.4. DYNAMICS OF POST-CRITICALLY FINITE RATIONAL MAPS ON S2 97

Proof of Theorem 7.3.1. It suffices to prove the following: If (Z, d)

is an Ahlfors Q-regular space in G and p > Q, then p H 1 (G) = {0}.
We may select some hyperbolic graph G associated to (Z, d) and consider
p H 1 (G). Choose a nonconstant Lipschitz function u on Z. By Proposition
7.3.4 and Corollary 7.3.6,
0 < ||u − fu (o)||Lp (Z) = ||∂∞ fu − fu (o)||Lp (Z) ≤ C||dfu ||p (EG ) < ∞,
so fu defines a nontrivial element in p H 1 (G). 
Sketch of the proof of Theorem 7.3.2. It suffices to prove that
p(G) ≥ Q if G contains an Ahlfors Q-regular Loewner metric space (Z, d).
In fact, we claim that under this assumption, Q H 1 (G) = {0}, that is, every
function f : VG → R with df ∈ Q (EG ) is constant.
Assume f : VG → R satisfies df ∈ Q (EG ). As discussed in the proof of
Theorem 7.3.1, we have
 
(7.3.4) |∂∞ f (x) − ∂∞ f (y)|Q dμ(x)dμ(y) < ∞.
Ba Bb
ab
Lemma 7.3.7. u = ∂∞ f satisfies a local Hajlasz–Sobolev inequality: for
each compact set K in Z there exists a function ρ = ρu,K ∈ LQ (Z) so that
|u(z1 ) − u(z2 )| ≤ d(z1 , z2 )(ρ(z1 ) + ρ(z2 )) for all z1 , z2 ∈ K.

Lemma 7.3.8. There exists Z1 ⊂ Z, μ(Z1 ) = 0, so that u|Z\Z1 is constant

along ModQ -a.e. curve.
Lemma 7.3.7 follows from (7.3.4) and a simple integral estimate. Lemma
7.3.8 uses Lemma 7.3.7 and arguments reminiscent of Fuglede’s lemma.
It is known that Ahlfors Q-regular Loewner spaces satisfy the MECQ
(Main Equivalence Class) property: For each Borel ρ ∈ LQ (Z), there exists
Z2 ⊂ Z, μ(Z2 ) = 0, and any pair of points in Z \ Z2 can be joined by a
rectifiable curve γ with γ ρ ds < ∞. Let Γ0 be the family of rectifiable
curves γ in (Z, d) for which u|γ\Z1 is not constant, where Z1 is the set
from Lemma 7.3.8. This curve family has Q-modulus  zero. By Proposition
4.1.6(v), we may choose ρ ∈ LQ (Z) so that γ ρ ds = ∞ for all γ ∈ Γ0 .
We now use the MECQ property to find a μ-null set Z1 so that any points
z1 , z2 ∈ Z \ Z2 are joined by a rectifiable curve γ so that γ ρ ds < ∞. Then
γ ∈ Γ0 so u|γ\Z1 is constant. That is, u is constant almost everywhere. By
the definition of fu , we see that fu is constant everywhere. Finally, [fu ] = [f ]
as elements of p H 1 (G) and we conclude that [f ] = 0 as desired. 

7.4. Dynamics of post-critically finite rational maps on S2

The dynamics of branched covering maps of the Riemann sphere leads
to another natural field of application for the quasisymmetric classification
problem for abstract metric spaces. The Ahlfors regular conformal dimen-
sions of certain dynamically defined metrics on S2 can conjecturally be com-
puted in terms of a finite list of combinatorial information arising from the
98 7. AHLFORS REGULAR CONFORMAL DIMENSION

underlying dynamics (Conjecture 7.4.6). Half of this conjecture is known to

hold (Theorem 7.4.7). In this section we briefly review the relevant theory.
7.4.1. Post-critically finite rational maps and branched cover-
ing maps of S2 . We identify S2 with the Riemann sphere C. A rational
map R : C →  C is called post-critically finite if it has a finite post-critical
set P (R) = m≥1 R◦m (C(R)), where C(R) = {z ∈ C : R (z) = 0} is the set
of critical points of R and f ◦m denotes the m-fold iterate of f .
More generally, we consider branched covering maps f : S2 → S2 , i.e.,
discrete and open mappings.
Question 7.4.1. When is a branched covering map f : S2 → S2 topo-
logically conjugate to a rational map?
Thurston answered Question 7.4.1 for a large class of branched covering
maps. Details of the proof of Thurston’s characterization were provided by
Douady and Hubbard. See Theorem 7.4.3 for the precise statement.
We continue with definitions and notation. Let f be a branched covering
map of S2 . The (local) degree of f at a point z0 , denoted deg(f, z0 ), is the
unique integer m so that f is conjugate to z
→ z m near z0 . Such a map is
proper, has finite maximal degree deg(f ) = maxz0 ∈S2 deg(f, z0 ), and has a
discrete branch set Bf := {z0 ∈ S2 : deg(f, z0 ) > 1}. On the complement of
Bf , f acts as a covering map. As before, the post-critical set P (f ) is defined
as the iterated forward image of the critical set Bf . Again, f is called post-
critically finite (PCF) if P (f ) is a finite set. The associated orbifold is the
pair Of := (S2 , νf ), where the weight function νf : S2 → N is defined as
follows: νf (z0 ) is the least common multiple of the values deg(f, w) taken
over all iterated preimages
 w of z0 . The Euler characteristic of the orbifold
Of is χ(Of ) := 2 − z0 ∈S2 (1 − νf (z0 )−1 ). We say that Of is hyperbolic,
resp. parabolic, if χ(Of ) < 0, resp. χ(Of ) = 0.
The following classical example goes back to work of Lattès and Schröder.
Example 7.4.2. Consider the torus T = C/Z[i] obtained as the quotient
of C by the Gaussian integers Z[i]. The involution z
→ −z is a automor-
phism of this torus with four fixed points 0, 1/2, i/2, (1 + i)/2. Form the
quotient Riemann surface S = T /(z ∼ −z). The ramification index ep of
the degree d = 2 quotient map π : T → S at each of the fixed points p is
equal to two. We compute the Euler characteristic of S from the Riemann–
Hurwitz formula 
(ep − 1) = d · χ(S) − χ(T );
p∈T
the answer χ(S) = 2 reveals that S is a sphere. In fact, upon equipping S
with the standard conformal structure on S2 , the quotient map π becomes
a Weierstrass ℘-function.
Consider the degree four self-map z
→ 2z of T . This map is equivariant
with respect to π, hence descends to S2 where it defines a PCF branched
cover f with four postcritical points, each with weight νf = 2. There are no
 7.4. DYNAMICS OF POST-CRITICALLY FINITE RATIONAL MAPS ON S2 99

periodic critical points: f is chaotic on the entire sphere S2 . The associated

orbifold Of , the (2,2,2,2)-orbifold, has Euler characteristic χ(Of ) = 2 −
4(1 − 12 ) = 0.
There exist rational maps R : C → C in the topological conjugacy class
of f . For example,
(z 2 + 1)2
R(z) =
4z(z 2 − 1)
is ramified of degree two at the four points 0, ±1, ∞. This is an example of
a so-called (integral) Lattès map. See Figure 7.1.

x2

π π

Figure 7.1. A Lattès map of S2

In the remainder of this section, we restrict attention to the class of

PCF branched covering maps f with expanding dynamics: f has no peri-
odic critical points (in which case the Julia set J(f ) is all of S2 , i.e., f is
everywhere chaotic) and Of is hyperbolic. Note that Example 7.4.2 does
not have expanding dynamics.
Theorem 7.4.3 (Thurston, Douady–Hubbard). A PCF branched cov-
ering map f : S2 → S2 with expanding dynamics is conjugate to a rational
map (by homotopies relative to P (f )) if and only if f has no Thurston ob-
structions.
Thurston obstructions are combinatorial conditions phrased in terms of
the topological properties of the action of f on a finite collection of homo-
topically distinct, non-peripheral curves in S2 \ P (f ). We briefly review the
definition.
7.4.2. Thurston obstructions. Fix a PCF branched covering map f
with expanding dynamics. A multicurve for f is a finite collection Λ =
{γ1 , . . . , γm } of simple, closed, unoriented, disjoint, non-peripheral curves in
S2 \ P (f ) which are pairwise distinct up to free homotopy in S2 \ P (f ). A
100 7. AHLFORS REGULAR CONFORMAL DIMENSION

curve γ in S2 \ P (f ) is peripheral if some component of S2 \ γ contains at

most one post-critical point. Peripheral curves play no role in the definition
of Thurston obstructions.
A multicurve Λ is f -stable if for each j = 1, . . . , m, each component of
f −1 (γj ) is either peripheral or freely homotopic in S2 \ P (f ) to some γi ∈ Λ.
To each f -stable multicurve Λ we associate m × m matrices Ap (f, Λ) =
(aij,p ), p ≥ 1, as follows. For fixed i, j ∈ {1, . . . , m}, let γijk denote the
components of f −1 (γj ) which are freely homotopic to γi in S2 \ P (f ). The
map f |γijk : γijk → γj is a covering map; let dijk denote its degree. Then
define  1−p
aij,p = dijk .
k
We call Ap (f, Λ) the Thurston matrix of order p associated to Λ. By the
Perron–Frobenius theorem, Ap (f, Λ) has a largest eigenvalue λp (f, Λ) ≥ 0.
The multicurve Λ is called a Thurston obstruction if λ2 (f, Λ) ≥ 1. Then
Theorem 7.4.3 admits the following reformulation: f is conjugate to a ra-
tional map if and only if λ2 (f, Λ) < 1 for all f -stable multicurves Λ.
Example 7.4.4. Consider the pair of topological 2-spheres S and S 
shown in Figure 7.2. The sphere S is obtained by gluing together ten squares
along the indicated edges (there are four occluded squares on the lower face),
while S  is obtained by gluing together two squares (there is one occluded
square on the lower face). Note the “flap” on the upper square arising from
gluing a pair of squares along three edges.

γ3 f
γ2 γ
γ1

Figure 7.2. PCF branched covering map of S2 with a

Thurston obstruction

A map f : S → S  is defined by mapping black squares to black squares

and white squares to white squares via similarities, with coherent orienta-
tions obtained by reflection across boundary edges. This map is PCF; the
post-critical set P (f ) consists of the four vertices of the large square. In this
case the orbifold Of has negative Euler characteristic and Theorem 7.4.3
applies.
An f -stable multicurve Λ = {γ} is indicated on the right. The curve
γ has three preimages γi , i = 1, 2, 3, shown on the left. The curve γ3 is
peripheral, while γ1 and γ2 are freely homotopic to γ in the complement of
P (f ). Observe that f maps each of the curves γ1 and γ2 to γ by a degree
two map. Then A2 (f, Λ) is the 1 × 1 matrix (a11,2 ) where a11,2 = λ2 (f, Λ) =
2 · 21−2 = 1. Thus Λ is a Thurston obstruction and f is not conjugate to a
rational map.
 7.4. DYNAMICS OF POST-CRITICALLY FINITE RATIONAL MAPS ON S2 101

Example 7.4.5. To illustrate the necessity of the restriction to expand-

ing maps in Theorem 7.4.3, consider Example 7.4.2. An f -stable multicurve
Λ = {γ} is indicated in Figure 7.3. Up to relabelling of the post-critical
points via an isometry of the square and up to homotopy relative to P (f ),
this multicurve is unique. The curve γ has two preimages γ1 , γ2 , each of
which f maps to γ by a degree two map. Again A2 (f, Λ) is the 1 × 1 matrix
(a11,2 ) where a11,2 = λ2 (f, Λ) = 2 · 21−2 = 1. However, f is conjugate to a
rational map.

f
γ2 γ
γ1

Figure 7.3. Lattès-type PCF branched covering map of S2

7.4.3. The induced conformal gauge. To each PCF branched cover-

ing map f : S2 → S2 with expanding dynamics there corresponds a canonical
doubling conformal gauge Gf on S2 so that f acts on (S2 , d) (for any d ∈ Gf )
with uniformly bounded metric distortion (more precisely, f is “uniformly
quasiregular” for an appropriate metric space notion of quasiregularity).
The gauge in question is the canonical boundary gauge for a certain Gro-
mov hyperbolic graph G generated by the action of f . The details of the
construction are quite similar to that in section 7.3.2. In particular, the
vertices of G are taken to be the components of all preimages of elements of
a fixed finite cover U0 of S2 by iterates of f , while the edges are defined by
adjacency considerations similar to those in section 7.3.2. We metrize this
graph by declaring each edge to have length one. Since f has expanding
dynamics, G is Gromov hyperbolic. The induced Gromov boundary ∂∞ G
is homeomorphic to S2 , and the boundary gauge Gf depends only on the
conjugacy class of the original map f .
In the case when f is conjugate to a rational map, Gf contains the stan-
dard Euclidean metric on S2 and the Ahlfors regular conformal dimension
of G is attained at 2. One may ask to what extent the Ahlfors regular con-
formal dimension of the boundary gauge provides a quantitative measure of
the deviation of f from the rational map case.
The structure theory of nonnegative matrices implies that, apart from
certain degenerate multicurves Λ (so-called reducible multicurves), the func-
tion p
→ λp (f, Λ) is continuous and strictly decreasing on [1, ∞), with
λ1 (f, Λ) ≥ 1 and limp→∞ λp (f, Λ) = 0. Thus for each f -stable Λ, there
exists a unique p = p(f, Λ) ≥ 1 so that λp (f, Λ) = 1.
The following conjecture was proposed by Bonk, Geyer and Pilgrim.
102 7. AHLFORS REGULAR CONFORMAL DIMENSION

Conjecture 7.4.6 (Bonk–Geyer–Pilgrim). Let pf := supΛ p(f, Λ), where

the supremum is taken over all irreducible f -stable multicurves Λ. Then
ARCdim Gf = pf .
It is worth pointing out the following observation: if f has a Thurston
obstruction and satisfies ARCdim Gf = 2, then ARCdim Gf is not attained.
This follows from Theorem 7.4.3. For instance, see Example 7.4.4.
One direction in Conjecture 7.4.6 is known.
Theorem 7.4.7 (Haı̈ssinsky–Pilgrim). ARCdim Gf ≥ pf .
A key tool in the proof of Theorem 7.4.7 is a discrete, combinatorial
version of the modulus. Discretization plays an important role in many
contemporary approaches to conformal geometry in metric spaces. See the
notes for more details and references.
Definition 7.4.8. Let Z be a set and let S be a covering of Z, i.e., a
collection of subsets whose union is Z. We emphasize that Z is equipped
with no a priori metric. Let Γ be any family of subsets of Z. The discrete
p-modulus of Γ relative to S is the value

d-modp (Γ, S) = inf ρ(A)p ,
ρ
A∈S
where the infimum is taken over all ρ : S → [0, ∞) with

inf ρ(A) ≥ 1.
γ∈Γ
A∈S:A∩γ=∅

Lemma 7.4.9. Let (Z, d) be a compact Ahlfors Q-regular metric space,

Q > 1, and let Sm , m ≥ 1, be an approximating sequence of uniform quasi-
packings of Z. For fixed L > 0, let ΓL be the collection of curves in Z whose
diameter is at least L. Then exactly one of the following conclusions holds:
• ModQ ΓL = 0 and limm→∞ d-modQ (ΓL , Sm ) = 0, or
• ModQ ΓL > 0 and there exists M = M (L) so that for all m ≥ M
we have
1
ModQ ΓL ≤ d-modQ (ΓL , Sm ) ≤ C ModQ ΓL .
C
Here C denotes a constant which is independent of m and L.
A uniform quasipacking of (Z, d) is a collection S of subsets of Z for
which there exists a constant K < ∞ so that
• each A ∈ S is K-quasiround, i.e., there exists a ball BA in Z so
that BA ⊂ A ⊂ KBA , and
• the collection S is essentially disjoint in the following sense: the
balls BA , A ∈ S, defined in the previous item are pairwise disjoint.
An approximating sequence of uniform quasipackings is a sequence of uni-
form quasipackings Sm , m ≥ 1, with the property that supA∈Sm diam A
tends to zero.
 7.4. DYNAMICS OF POST-CRITICALLY FINITE RATIONAL MAPS ON S2 103

Earlier we observed that each PCF branched covering map f of S2 with

expanding dynamics induces a canonical conformal gauge Gf on S2 with
respect to which f acts uniformly with bounded metric distortion. In addi-
tion, f induces approximating sequences of uniform quasipackings. We may
choose S0 to be a finite cover of S2 by balls and consider the collections Sm ,
m ≥ 1, defined as the set of connected components of preimages of elements
of S0 by the iterates f (m) .
Now suppose that Λ is an f -stable multicurve with λp (f, Λ) ≥ 1 for
some p > 1. Denote by Γ(Λ) the family of all curves in S2 \ P (f ) which are
homotopic to some curve in Λ.
Lemma 7.4.10 (Haı̈ssinsky–Pilgrim). lim inf m→∞ d-modp (Γ(Λ), Sm ) >
0.
Sketch of the proof of Theorem 7.4.7. For simplicity, we only con-
sider the case pf > 2.
Choose an Ahlfors Q-regular metric d ∈ Gf , Q > 2, and an approximat-
ing sequence of uniform quasipackings Sm . Fix p, 2 < p < min{pf , Q}. By
the definition of pf , there exists an f -stable multicurve Λ so that λp (f, Λ) ≥
1. Note that the curves in Γ(Λ) have diameter uniformly bounded away from
zero by the non-peripherality condition. Thus Γ(Λ) ⊂ ΓL for some L > 0.
By Lemma 7.4.10, d-modp (ΓL , Sm ) remains bounded away from zero for m
sufficiently large. By Lemma 7.4.9, Modp ΓL > 0. Now Proposition 4.1.8
implies that
Cdim Z ≥ p.
Taking the supremum over the allowable values of p gives
Cdim Z ≥ min{pf , Q}
for any Q such that there exists an Ahlfors Q-regular metric d in Gf . Thus
ARCdim Z ≥ pf
which completes the proof. 
We conclude with an example where the conclusion of Conjecture 7.4.6
is satisfied with pf > 2.
Example 7.4.11. Consider the pair of topological 2-spheres S and S 
shown in Figure 7.4. The sphere S is obtained by gluing together 24 rectan-
gles along the indicated edges (there are twelve occluded rectangles on the
lower face), while S  is obtained by gluing together two squares (there is one
occluded square on the lower face). A map f : S → S  is defined by map-
ping black rectangles to black squares and white rectangles to white squares
via affine maps, with coherent orientations obtained by reflection across the
boundary edges. This map is PCF; the post-critical set P (f ) consists of the
four vertices of the large square.
The f -stable multicurve Λ = {γ} with maximal p(f, Λ) is indicated on
the right. (There is a second f -stable multicurve in the orthogonal direction
104 7. AHLFORS REGULAR CONFORMAL DIMENSION

f
γ4
γ γ
γ2 3
γ1

Figure 7.4. PCF branched covering map of S2 with a

Thurston obstruction and fractal conformal gauge of Rick-
man’s rug type

with a smaller value of p(f, Λ).) In the figure, γ has four preimages γi ,
i = 1, 2, 3, 4, shown on the left. Each of these curves is freely homotopic to
γ in the complement of P (f ), and f maps each of these curves to γ by a
degree three map. Then Ap = Ap (f, Λ) is the 1 × 1 matrix (a11,p ) where
a11,p = λp (f, Λ) = 4 · 31−p . In particular, a11,2 = λ2 = 43 > 1 and Λ is a
Thurston obstruction. Solving λp (f, Λ) = 1 gives
log 4
pf = 1 + .
log 3
In this setting, the optimal metric in the conformal gauge Gf is obtained by
snowflaking the horizontal direction with parameter  = log 3
log 4 , yielding the
product space Y × [0, 1], where Y = ([0, 1], d ). This is precisely Rickman’s
rug; see Example 4.1.9 and the notes to Chapter 4. As indicated in Example
4.1.9, this space is minimal for conformal dimension, and thus provides an
example where equality holds in the conclusion of the conjecture.
The associated orbifold Of for Example 7.4.11 is parabolic, hence this
example is not strictly speaking relevant for Conjecture 7.4.6. To generate
a hyperbolic orbifold, one can add flaps to the sphere as in Example 7.4.4.

7.5. Notes
David and Semmes introduced deformations of metric spaces by dou-
bling measures in [46]. The preservation of analytic machinery (such as the
Loewner condition or Poincaré inequalities) under such deformations makes
them a useful tool. See also [154] and [47].
Theorems 7.2.1 and 7.2.2 of Bonk and Kleiner are taken from [22] and
[21] respectively. A different line of attack on Cannon’s conjecture has
been suggested by Cannon, Floyd and Parry [40]. It is worth observing
that Theorem 7.2.2 is false in dimension three. The sphere S3 admits a
Carnot-Carathéodory metric induced by its standard CR structure (as a
real submanifold of C2 ) which is locally modelled by the Heisenberg group.
Equipped with this metric and the associated volume measure, S3 becomes
an Ahlfors 4-regular Loewner metric space. As described in section 3.3, the
Carnot-Carathéodory metric is a visual metric on the boundary of complex
hyperbolic space.
 7.5. NOTES 105

The connection between p cohomology and conformal geometry goes

back to work of Gromov [65] and has been studied extensively by Pansu
[132], [136], [137], [135], [138, §5]. Theorems 7.3.1 and 7.3.2 are due to
Bourdon and Pajot [34]. Our presentation is loosely based on theirs, al-
though we omit a number of details, including the precise identification of
p H 1 (G) as a certain Besov space on the boundary. (Implicitly, this identi-
fication is expressed in the criterion of Proposition 7.3.5 for membership of
dfu in p (EG ).)
A quasiconvex Ahlfors Q-regular metric space is Loewner if and only if
it satisfies a (1, Q)-Poincaré inequality in the sense of Heinonen and Koskela
[77]. The MECp property was introduced by Shanmugalingam as a qualita-
tive substitute for the (1, p)-Poincaré inequality which still provides geomet-
ric control on connectivity. The validity of the MECp property for spaces
satisfying a (1, p)-Poincaré inequality was proved by Shanmugalingam in
[155]; a minor technical issue in the proof was clarified in [85].
A key point of the proof of Theorem 7.3.2 is the fact that the cohomology
group in the critical exponent, Q H 1 (G), is trivial in the case when the
boundary contains an Ahlfors regular Loewner space. Bonk and Saksman
[25] show that the corresponding space Q,∞ H 1 (G) (obtained by replacing
Q-summability with a suitable weak-type Chebyshev condition) is nontrivial.
The interplay between conformal geometry and the dynamics of rational
maps has a long and rich history dating back to the early work of Fatou and
Julia. The Lattès map in Example 7.4.2 was presented by Lattès [102]
following work of Schröder. See also [121, Chapter 7]. An elegant survey
on the theory of Lattès maps can be found in [122].
The characterization in Theorem 7.4.3 of branched covering maps of
S2 which are conjugate to rational maps is proved by Douady–Hubbard
[48] following work of Thurston. See also McMullen [118]. The boundary
conformal gauge associated to a PCF branched covering maps of S2 appears
in various places in the literature. Our discussion follows Haı̈ssinsky–Pilgrim
[70, Chapter 3]. Conjecture 7.4.6 was proposed by Bonk, Geyer and Pilgrim
(unpublished), while Theorem 7.4.7 is due to Haı̈ssinsky–Pilgrim [69]. The
analogy between Theorem 7.4.7 and Theorem 7.3.1 is striking. See also the
related work of Bonk and Meyer, [120], [119], [24].
Discrete or combinatorial versions of the modulus arise repeatedly in
the theory of quasiconformal maps on metric spaces, see for example Pansu
[133], Cannon–Floyd–Parry [40], Schramm [152], Tyson [167], [168], Bonk–
Kleiner [21], Haı̈ssinsky [68], and Bonk–Merenkov [23]. The version de-
scribed here, and particularly Lemma 7.4.9, are taken from Haı̈ssinsky [68,
Appendix B], see also [69, Proposition 3.2]. Example 7.4.11 can be found
in [70, §4.5].
The expository article by Bonk [19] in the proceedings of the 2006 ICM
provides another overview of the material in sections 7.2 and 7.4, as well as
further context, motivation and history.
106 7. AHLFORS REGULAR CONFORMAL DIMENSION

The following example shows that Ahlfors regular Loewner metrics need
not exist, even in conformal gauges arising as hyperbolic group boundaries
and satisfying rather natural and mild restrictions.
Example 7.5.1 (Bourdon–Pajot). There exists a conformal gauge G,
containing a hyperbolic group boundary which is connected and locally con-
nected with no local cut points, yet there is no Ahlfors regular Loewner
metric in G.
Let Γ be the hyperbolic group obtained as an amalgamated product
of cocompact lattices L1 and L2 in Bourdon–Pajot hyperbolic buildings
Ip1 q1 and Ip2 q2 respectively, amalgamated over a free group Fn . The group
Γ contains quasiconvex copies of L1 and L2 . Consequently ∂∞ Γ contains
quasisymmetrically embedded copies of ∂∞ Ipν qν , ν = 1, 2, and
log(qν − 1)
Cdim ∂∞ Γ ≥ max Cdim ∂∞ Ipν qν ≥ max 1 + ;
ν=1,2 ν=1,2 arccosh( pν2−2 )
see Example 3.5.1. The space ∂∞ Γ is a compact, connected, locally con-
nected, doubling space without local cut points.
We analyze the critical p cohomology exponent of the hyperbolic 2-
complex Γ. Let pν = 4gν for some gν ≥ 2. Identification on the boundary of
a hyperbolic pν -gon which is a fundamental domain for the deck transfor-
mations of Ipν qν preserving the lattice Lν generates a finite connected CW
2-complex Kν of Euler characteristic χ(Kν ) = 2 − 2gν . We write p H • (Kν )
for the p cohomology groups of Kν . The Euler characteristic may be com-
puted via 2 Betti numbers:
χ(Kν ) = rank 2 H 0 (Kν ) − rank 2 H 1 (Kν ) + rank 2 H 2 (Kν ).
Every finite P D 2 complex with infinite fundamental group has trivial second
2 cohomology group. The complex Kν is of this type. Since Kν is connected,
we conclude that the first 2 Betti number of H 1 (Kν ) is −χ(Kν ) = 2gν −2 =
pν
2 − 2. On the other hand, the first  Betti number for Fn is n − 1. From
2

the Mayer–Vietoris exact sequence in 2 cohomology we obtain

rank 2 H 1 (Γ) ≥ rank 2 H 1 (K1 ) + rank 2 H 1 (K2 ) − rank 2 H 1 (Fn )
p1 + p2
=( − 4) − (n − 1).
2
Choose pν and qν so that 12 (p1 +p2 )−4 > n−1 and log(q2 −1) > arccosh((p2 −
2)/2). Then Cdim ∂∞ Γ > 2 but rank 2 H 1 (Γ) > 0 so p(∂∞ Γ) ≤ 2. In
particular, p(∂∞ Γ) < Cdim ∂∞ Γ. The conclusion now follows from Theorem
7.3.2.
 CHAPTER 8

Global quasiconformal dimension

8.1. Distortion of dimension by quasiconformal maps

How do quasiconformal maps of RN distort the dimensions of subsets?
This classical problem of Euclidean geometric function theory leads to a
variant on the notion of conformal dimension. Recall that quasisymmetry
and quasiconformality are equivalent for homeomorphisms of RN , N ≥ 2.
Our starting point is a theorem of Gehring and Väisälä.
Theorem 8.1.1 (Gehring–Väisälä). Let N ≥ 2, K < ∞, and 0 < s <
N . Then there exist α(N, K, s) > 0 and β(N, K, s) < N so that for any
set S ⊂ RN of (Hausdorff ) dimension s and any K-quasiconformal map
F : RN → RN we have α(N, K, s) ≤ dim F (S) ≤ β(N, K, s).
Recalling that the inverse of a quasiconformal map is quasiconformal,
we conclude:
Corollary 8.1.2. The image of a set of Hausdorff dimension N under
a quasiconformal map of RN , N ≥ 2, has dimension N .
Theorem 8.1.1 provides dilatation-dependent estimates on the distortion
of Hausdorff dimension by quasiconformal maps. What can be said about
dilatation-independent estimates? Nothing, if we allow completely general
sets. In fact, Theorem 8.1.1 is sharp in the following sense: for each 0 <
s < s < N there exist sets S, S  ⊂ RN with dim S = s, dim S  = s and a
quasiconformal map F : RN → RN with F (S) = S  . The sets S, S  can be
chosen to be self-similar Cantor sets.
However, the situation is quite different if we restrict our attention to a
fixed subset of RN .
Definition 8.1.3. Let S ⊂ RN for some N ≥ 1. The global quasicon-
formal dimension of S is
GQCdim S = inf dim F (S)
where the infimum is taken over all quasiconformal (quasisymmetric if N =
1) homeomorphisms F : RN → RN .
It is clear that
Cdim S ≤ GQCdim S ≤ dim S
for all S ⊂RN . A priori, the first inequality may be strict. This could hap-
pen for one of two reasons: first, the quasisymmetrically equivalent metric
107
108 8. GLOBAL QUASICONFORMAL DIMENSION

spaces (Z  , d ) which realize the conformal dimension of S may not admit bi-
Lipschitz embeddings in any Euclidean space (let alone the space RN where
S lies) and second, even if (Z  , d ) does admit a bi-Lipschitz embedding in
RN , the quasisymmetric map which demonstrates the equivalence may not
extend to a global map of the ambient space.
As discussed in Chapter 2, the problem of determining the supremum
of the dimensions of quasisymmetric images of a fixed metric space is com-
pletely understood. Indeed, the snowflaking operation shows that this supre-
mum is equal to either zero or infinity in all cases. The analogous problem
associated to global quasiconformal dimension (to determine the supremum
of the dimensions of the images of a fixed subset under quasisymmetric maps
of the ambient space) is also well understood, at least in the Euclidean case.
This is the content of the next theorem.
Theorem 8.1.4 (Bishop). If S ⊂ RN is a compact set of positive Haus-
dorff dimension, then for any  > 0 there exists a quasisymmetric map
F : RN → RN so that dim F (S) > N − .
In any dimension, quasiconformal maps preserve sets of dimension zero.
Since Hausdorff measures are Borel regular and Borel sets of positive Hs
measure have compact subsets of positive Hs measure for any s > 0 [116,
Theorem 8.13], we conclude
Corollary 8.1.5. If S ⊂ RN then sup dim F (S), where the supremum
is taken over all quasisymmetric maps F : RN → RN , is equal to either zero
or N .

8.2. Minimality for global quasiconformal dimension and middle

interval Cantor sets
Proposition 8.2.1. For each 1 ≤ α ≤ N there exist sets S ⊂ RN with
GQCdim S = dim S = α.
This follows from Example 4.1.9 and Proposition 4.1.10, after noting
that Y can be chosen to be a self-similar Cantor set in RN −1 .
Remark 8.2.2. There exist totally disconnected (but not uniformly dis-
connected!) examples as in Proposition 8.2.1.
For middle interval Cantor sets on the real line, the situation is well
understood.
Definition 8.2.3. Let a = (aj ) be a sequence of real numbers, each of
which lies in (0, 1). The middle interval Cantor set K(a) is defined to be
" 
K(a) := Iw ,
m≥1 w

where the union is taken over all subintervals Iw of I = [0, 1] indexed by

words w of length m in the letters 1 and 2. For each word w, the interval
 8.3. GLOBAL QUASICONFORMAL DIMENSIONS OF SELF-SIMILAR SETS 109

Iw satisfies
Iw = Iw1 ∪ Jw ∪ Iw2 ,
where the subinterval Jw has the same center as Iw and |Jw | = am |Iw |.
The standard 13 Cantor set K coincides with K(a) for a = ( 13 , 13 , . . .). It
is easy to verify that
dim K(a) = 1 ⇔ a ∈ c0
and
L1 (K(a)) > 0 ⇔ a ∈ 1 .
We restrict our attention to uniformly perfect middle interval Cantor sets.
An easy exercise shows that K(a) is uniformly perfect if and only if ||a||∞ <
1.
Theorem 8.2.4. The (global quasi)conformal dimension of any uni-
formly perfect middle interval Cantor set is either zero or one. The uni-
formly perfect middle interval Cantor set K(a) is minimal for (global quasi)-
conformal dimension if and only if it has Hausdorff dimension one (i.e., if
and only if a ∈ c0 ).
One direction follows from the theorem of Kovalev, while the other di-
rection is a result of Hakobyan.
Corollary 8.2.5. There exist subsets of the real line of length zero
which are minimal for (global quasi)conformal dimension.
It suffices to consider uniformly perfect middle interval Cantor sets K(a)
with a ∈ c0 \ 1 . Such sets also carry measures satisfying the growth condi-
tions (4.5.1) and thus serve as examples for Question 4.5.1. See section 8.4
and the notes to this chapter for additional discussion.

8.3. Global quasiconformal dimensions of self-similar sets

A general philosophy which is surely evident by now is that it is easier
to lower dimension quasisymmetrically when the set or space in question is
highly disconnected. After Cantor sets (the totally disconnected case), post-
critically finite spaces are the least connected spaces. Roughly speaking, a
space is post-critically finite if it admits finite cut sets at all scales and
locations. Post-critical finiteness already surfaced in our discussion of the
dynamics of rational maps (section 7.4). A related concept arises in self-
similar and self-affine fractal geometry. We begin with some definitions.
8.3.1. Self-similar iterated function systems and their invariant
sets. A map f : RN → RN is a contractive similarity if f (x) = a + rA · x
where a ∈ RN , r ∈ (0, 1) and A is an orthogonal matrix. We call r the
contraction ratio associated to f .
Definition 8.3.1. A self-similar iterated function system (IFS) is a
finite collection of contractive similarities.
110 8. GLOBAL QUASICONFORMAL DIMENSION

The invariant set for a self-similar IFS F = {f1 , . . . , fM } is the unique

nonempty compact set K = KF satisfying

M
(8.3.1) K= fi (K).
i=1
Existence and uniqueness of K follows from the completeness of the hy-
perspace of nonempty compact subsets of RN equipped with the Hausdorff
metric.
One of the most important questions about the invariant set of a self-
similar IFS is the value of its Hausdorff dimension. Let F = {f1 , . . . , fM }
be a self-similar IFS and let ri be the contraction ratio of fi . The similarity
dimension of F is the unique nonnegative number s = s(F ) satisfying the
equation

M
(8.3.2) ris = 1.
i=1
Note that s(F ) depends only on the contraction ratios {r1 , . . . , rM }. The
quantity s(F ) is always at least as large as, and often equal to, the Hausdorff
dimension of the invariant set of F . For instance, Falconer and Solomyak
have shown that
(8.3.3) dim K(F ) = s(F )
for almost every IFS F = {f1 , . . . , fM } with fj (x) = aj + rj Aj · x. Here the
orthogonal matrices Aj and the contraction ratios rj < 12 are assumed to
be fixed, and the almost sure statement is interpreted with respect to the
Ln×M measure on the translation parameters (a1 , . . . , aM ) ∈ Rn×M .
The Moran–Hutchinson theorem provides another setting in which (8.3.3)
holds.
Theorem 8.3.2 (Moran–Hutchinson). Let F be an iterated function sys-
tem as above which satisfies the open set condition. Then dim K(F ) = s(F ).
Moreover, K(F ) is Ahlfors s(F )-regular.
An IFS F = {f1 , . . . , fM } is said to satisfy the open set condition if
there exists a nonempty bounded open set O for which fj (O) ⊂ O for all
j and the sets fj (O), j = 1, . . . , M , are pairwise disjoint. This separation
condition ensures that the self-similar pieces of the invariant set K = K(F )
do not overlap, which in turn guarantees that there are no “accidentally
optimal” coverings of K which could reduce the Hausdorff dimension from
the expected value s(F ). In case the open set condition is violated, the value
of dim K(F ) can decrease from s(F ).
8.3.2. Symbolic dynamics. Let W ∞ := {1, . . . , M }∞ be the symbol
space for IFS(RN , M ). The symbolic representation map π = πF : W ∞ →
K assigns to an element w1 w2 · · · ∈ W ∞ the unique point
lim fw1 ◦ fw2 ◦ · · · ◦ fwm (0)
m→∞
 8.3. GLOBAL QUASICONFORMAL DIMENSIONS OF SELF-SIMILAR SETS 111

in K. Let C(F ) := i=j fi (K) ∩ fj (K) be the critical set of the IFS F . The
post-critical set P (F ) is the set of points in K which map into C(F ) under
some finite composition of elements of F , i.e., x ∈ P (F ) if and only if there
exists a finite word w1 · · · wm so that fw1 ◦ · · · ◦ fwm (x) ∈ C(F ).
Definition 8.3.3. A self-similar IFS F is post-critically finite (PCF) if
#π −1 (P (F )) < ∞.
Compare the definition in section 7.4.
Example 8.3.4. The premier example of a PCF self-similar fractal with
the open set condition is the Sierpiński gasket SG. This set is the invariant
set for a three element IFS
F 0 = {f0 , f1 , f2 }
where fj (z) = 12 (z + aj ) and a0 , a1 , a2 are non-collinear points in R2 . Here
C(F 0 ), P (F 0 ) and π −1 (P (F 0 )) each contain three elements. By the Moran–
Hutchinson theorem 8.3.2, dim SG = log 3/ log 2 and SG is Ahlfors regular
in this dimension. See Figure 8.1 for the case aj = ω j , ω = e2πi/3 .

Figure 8.1. The Sierpiński gasket SG

Two iterated function systems F and G in IFS(RN , M ) are said to be

isomorphic (or conjugate) if, up to a permutation of the letters 1, . . . , M ,
−1
the map F = πG ◦ πF : KF → KG is well defined. In this case F is a
homeomorphism which intertwines (commutes with) the defining similarities
for the two systems: F ◦ fk = gk ◦ F for all k = 1, . . . , M .
We are interested in the following questions.
Question 8.3.5. Let F and G be isomorphic IFS’s on RN .
(1) When is the canonical homeomorphism F : KF → KG quasisym-
metric?
(2) When does there exist a quasiconformal extension of F to RN ?
An answer to Question 8.3.5 has obvious consequences for the confor-
mal and global quasiconformal dimensions of invariant sets. In Example
8.3.11, we will show that the Sierpiński gasket SG has global quasiconfor-
mal dimension one by exhibiting a sequence of IFS’s whose invariant sets
112 8. GLOBAL QUASICONFORMAL DIMENSION

are equivalent with SG via quasiconformal maps of the plane. In Example

8.3.14 we provide an application to the dynamics of rational maps.
The following elementary example illustrates a necessary condition for
quasisymmetry of the canonical homeomorphism.
Example 8.3.6. Consider the following two iterated function systems
on the real line:
1 1 1
F = {f0 , f1 }, f0 (x) = x, f1 (x) = x + ,
2 2 2
and
1 2 1
G = {g0 , g1 }, g0 (x) = x, g1 (x) = x + .
3 3 3
Each of these IFS’s has [0, 1] as its invariant set. Both IFS’s are PCF, with
critical sets C(F ) = { 12 } and C(G) = { 13 }. However, the decomposition
of [0, 1] into self-similar pieces is quite different. Indeed, F decomposes
[0, 1] into two pieces each of length 12 , while G decomposes [0, 1] into one
piece of length 13 and another piece of length 23 . For large m, the size of
the level m pieces adjoining the critical points diverge at different rates.
−1
This fact precludes the canonical homeomorphism F = πG ◦ πF from be-
ing quasisymmetric. Check that F ( 2 + 2m ) = 3m+1 + 3 , F ( 2 ) = 13 and
1 1 2 1 1
2m
F ( 12 − 21m ) = 13 − 3m+1 . For any fixed distortion function η, the quasisym-
metry inequality (1.2.1) will be violated for this triple of points when m is
sufficiently large.
To avoid this issue we introduce the following notion.
Definition 8.3.7. The boundary, ∂F , of a PCF self-similar IFS F is
the set of maps in F whose fixed point lies in the post-critical set P (F ). We
say that F is boundary congruent if the contraction ratios λf , f ∈ ∂F , are
all equal.
Boundary congruence is a necessary condition for the canonical home-
−1
omorphism F = πG ◦ πF to be quasisymmetric. It is possible that the
invariant sets of two boundary incongruent IFS’s might be quasisymmetri-
cally invariant by a different map (which does not commute with the defining
iteration schemes).
8.3.3. Quasisymmetric and quasiconformal equivalence of self-
similar sets. The key result in this section is Theorem 8.3.9, which gives
conditions guaranteeing positive answers to Question 8.3.5. We introduce a
particular class of PCF self-similar iterated function systems which satisfy
the open set condition.
Definition 8.3.8. A post-critically finite self-similar IFS F in RN is of
gasket type if the following conditions hold:
(i) the convex hull Π of the invariant set K is a finite-sided polyhedron,
and each vertex of Π is the fixed point of an element of F ,
 8.3. GLOBAL QUASICONFORMAL DIMENSIONS OF SELF-SIMILAR SETS 113

(ii) fi (K) ∩ fj (K) = fi (Π) ∩ fj (Π) whenever i = j, and when this

intersection is nonempty, it contains exactly one point which is a
common vertex of fi (Π) and fj (Π), and
(iii) fi (K) ∩ fj (K) ∩ fk (K) = ∅ whenever i, j, k are distinct indices.
Gasket type IFS’s satisfy the open set condition; the interior of Π serves
as the required open set. The assumption that Π is a finite-sided polyhe-
dron ensures the existence of a positive lower bound on the angles between
adjacent child polyhedra fi (Π). The existence of such bounds is another
obvious necessary condition for quasisymmetry of the canonical homeomor-
phism. (Exercise: verify this claim.) Conditions (ii) and (iii) are technical
assumptions used in the proof of quasisymmetry; it is possible that they
could be relaxed.
We are now ready to state the main theorem.
Theorem 8.3.9 (Tyson–Wu). Let F and G be isomorphic boundary con-
−1
gruent, gasket type iterated function systems in RN . Then F = πG ◦ πF :
KF → KG is quasisymmetric. If in addition F and G can be joined by a
nondegenerate isotopy of boundary congruent, gasket type iterated function
systems, then F can be extended to a quasiconformal map of RN .
Nondegeneracy means that the dimension of the affine subspace of RN
spanned by the invariant sets of the systems F t is constant in t. The isotopy
in Theorem 8.3.9 refers to a continuous path of iterated function systems in
the topology of locally uniform convergence on IFS(RN , M ).
Corollary 8.3.10. Let K be the invariant set for a boundary congruent,
gasket type IFS F in RN . Then
Cdim K ≤ inf dim K  ,
where the infimum is taken over all boundary congruent, gasket type IFS’s
with invariant set K  which are isomorphic with F . Similarly,
GQCdim K ≤ inf dim K  ,
where the infimum is taken over all IFS’s which can be joined to F by a
nondegenerate isotopy of boundary congruent, gasket type IFS’s.
Example 8.3.11 (Tyson–Wu). The global quasiconformal dimension of
SG is equal to one. To show this, we exhibit a sequence of nondegenerate
isotopies [0, 1]  t
→ F t,ν , ν = 1, 2, . . ., of boundary congruent gasket type
systems for which F 0 is the IFS in Example 8.3.4. See Figure 8.2 for pictures
of the systems F 0 and F t,ν for some 0 < t < 1 and ν = 2. (The parameter
ν counts the number of triangles along each limb of the deformed gasket.)
Theorem 8.3.9 guarantees that the corresponding invariant sets K t,ν are
all equivalent via quasiconformal maps of R2 . A direct computation gives
dim K 1,ν → 1 as ν → +∞.
We observe an easier approach to the construction of the quasiconformal
extensions to R2 . The complementary regions for K 0 = SG are bounded
114 8. GLOBAL QUASICONFORMAL DIMENSION

by quasicircles (in fact, they are equilateral triangles). Since K 0 and K t,ν ,
t < 1, ν ∈ N, are quasisymmetrically equivalent, the complementary re-
gions for K t,ν are also bounded by quasicircles. Now the quasisymmetry
F : K 0 → K t,ν can be extended to these complementary regions by the
Beurling–Ahlfors extension theorem (see the notes to chapter 1). By self-
similarity, the extension is K-quasiconformal for some finite K.

Figure 8.2. Iterated function systems defining the

Sierpiński gasket SG and a quasisymmetrically equivalent
invariant set

Example 8.3.12 (Tyson–Wu). The n-dimensional Sierpiński gasket SGn

in RN , n ≥ 3, is defined as the invariant set for the IFS F 0 = {f0 , . . . , fn }
where fj (x) = 12 (x + aj ) and the points a0 , . . . , an are the vertices of a non-
degenerate simplex in RN . This system is post-critically finite and satisfies
the open set condition; the dimension of the invariant set SGn is equal to
log(n + 1)/ log 2.
For each n ≥ 2, SGn has global quasiconformal dimension equal to
one. When n ≥ 3 we require the second half of Theorem 8.3.9 to construct
the quasiconformal extensions. The main tool used in the construction of
such extensions is a type of quasisymmetric extension theorem due to Tukia
and Väisälä involving s-quasisymmetries, which are quasisymmetric maps
close (in a suitable sense) to similarities. Roughly speaking, the isotopy of
IFS’s induces a flow of quasisymmetric equivalences which (for short time)
are s-quasisymmetries. These s-quasisymmetries then extend to ambient
quasiconformal maps of RN .
We conclude this section with an application of Example 8.3.11 to the
dynamics of rational maps.
An IFS F = {f1 , . . . , fM } with invariant set K is invertible if there exists
a continuous map g : K → K with the property that the restriction of g
to each similarity image fj (K) coincides with the restriction of fj−1 . This
condition imposes a compatibility relation along the critical set C(F ). We
illustrate with an example. The IFS defining the Sierpiński gasket SG in
Example 8.3.4 is post-critically finite but not invertible. Consider the IFS

(8.3.4) $0 = {f˜0 , f˜1 , f˜2 }

F
 8.4. CONFORMAL DIMENSIONS OF SELF-AFFINE SETS 115

where f˜0 (z) = 12 (z + 1), f˜1 (z) = 12 ω(ωz + 1), and f˜2 (z) = 12 ω(z + ω). Then
$0 is both post-critically finite and invertible; its invariant set is still SG.
F
The following result of Eroğlu, Rohde and Solomyak provides general
criteria under which a topological conjugacy between a rational map dy-
namical system and a self-similar IFS can be promoted to quasisymmetric
conjugacy.
Theorem 8.3.13 (Eroğlu–Rohde–Solomyak). Let F be a planar invert-
ible PCF self-similar IFS all of whose contraction ratios are equal. Let g
be the associated expanding map on the invariant set K = K(F ). Assume
that K has bounded turning. Let R : C → C be a rational map with ex-
panding dynamics, with Julia set J(R). If the dynamical systems (K, g)
and (J(R), R) are topologically conjugate, then they are quasisymmetrically
conjugate. In fact, there exists a quasiconformal map of C conjugating the
actions of g : K → K and R : J(R) → J(R).
The post-critical finiteness assumption in Theorem 8.3.13 can be relaxed.
For the definition of bounded turning, see Definition 2.3.1.
Example 8.3.14 (Eroğlu–Rohde–Solomyak). Consider the invertible IFS
$
F in (8.3.4) with invariant set SG and associated inverse mapping g : SG →
0

SG. The self-similar dynamical system (SG, g) is topologically conjugate to

the rational map dynamical system (J(R), R), where
16
(8.3.5) R(z) = z 2 − .
27z
The criteria in Theorem 8.3.13 are satisfied, hence the two systems are
quasisymmetrically conjugate. Using Example 8.3.11 we conclude that
(8.3.6) GQCdim J(R) = 1.

8.4. Conformal dimensions of self-affine sets

In this section we present recent work of Binder–Hakobyan and Mackay
on the conformal dimensions of Bedford–McMullen self-affine carpets. It
may seem somewhat out of place to include this section in a chapter on
global quasiconformal dimension. Indeed, Theorem 8.4.11, which bounds
from below the conformal dimensions of self-affine carpets, is a consequence
of the abstract theory from chapter 4. However, to fully describe the result
we need the vocabulary of iterated function systems and symbolic dynamics.
A fortiori, Theorem 8.4.11 also bounds from below the global quasiconfor-
mal dimension. It is an interesting question whether analogs of Theorem
8.3.9 and Corollary 8.3.10 hold in this context, and if so, whether they can
be used in conjunction with explicit constructions to reduce the global qua-
siconformal dimensions of specific self-affine sets.
Conformal Assouad dimensions of self-affine sets turn out to be easier
to understand. Every non-self-similar, self-affine carpet is either minimal
116 8. GLOBAL QUASICONFORMAL DIMENSION

for conformal Assouad dimension or has conformal Assouad dimension zero.

See Theorem 8.4.14.

8.4.1. Self-similar measures. We start with a brief digression on self-

similar measures. We will use analogous ideas later on in the self-affine
setting.
Let F = {f1 , . . . , fM } be a self-similar IFS with associated contrac-
tion ratios r1 , . . . , rM . Let K be the invariant set for this IFS. Recall that
dim K ≤ s, where s denotes the similarity dimension for F , and that equality
holds if the open set condition is satisfied.
We construct measures on K with self-similar properties; these measures
are built from probability vectors p on the alphabet W = {1, . . . , M } which
defines the symbol space W ∞ . The dimensions of these measures can be
computed in terms of the entropy of p. These dimensions are all bounded
above by the similarity dimension; equality holds for the measure of maximal
entropy.
 Let p = (p1 , . . . , pM⊗∞ ) be a probability vector: pj ≥ 0 for all j and
j pj = 1. Let λp = p be the unique product probability measure on
W ∞ given by the Kolmogorov consistency theorem. Note that λp ({w ∈
W ∞ : w1 = j1 , . . . , wm = jm }) = pj1 · · · pjm for all j1 , . . . , jm ∈ W m .
The pushforward measure μp := (πF )# (λp ) via the symbolic representa-
tion map πF is a Borel probability measure on K. It is the unique probability
measure satisfying the self-similar equation

M
μp = pj (fj )# (μp ).
j=1

Definition 8.4.1. The dimension of a measure μ is

dim μ = inf{dim S : μ(S) > 0}.

Lemma 8.4.2. Assume that the open set condition is satisfied. Then
M
j=1 pj log 1/pj
(8.4.1) dim μp = M .
j=1 pj log 1/rj

The numerator on the right hand side of (8.4.1) is the entropy of the
probability vector p. It is easy to check that the right hand side of (8.4.1)
is maximized in the case when pj = rjs , where s denotes the similarity
dimension. Furthermore, in this case, dim μp = s.

8.4.2. Self-affine fractal geometry. We now turn to the study of

self-affine sets and their dimensions. First, some linear algebra.
Definition 8.4.3. The singular values of an n×n matrix A are the non-
negative square roots of the eigenvalues of AT A. We will write σ1 , . . . , σn
 8.4. CONFORMAL DIMENSIONS OF SELF-AFFINE SETS 117

for the singular values of A and always assume that these are in non-
increasing order. Geometrically, these quantities measure the lengths of
the principal semi-axes of the (possibly degenerate) ellipsoid A(B), where
B = {x ∈ RN : |x| ≤ 1}. The largest value σ1 coincides with the operator
norm ||A|| = maxx∈B |Ax| of A; this is the maximal stretching induced by
A. Similarly, σn is the minimal stretching. In the case when A is invertible,
σn = ||A−1 ||−1 . We say that A is contractive if ||A|| < 1. In the case when
A is orthogonal, all of the σj ’s are equal to one. An affine transformation
of RN is a map f : RN → RN of the form f (x) = Ax + b for some matrix A
and vector b ∈ RN .
Definition 8.4.4. A self-affine iterated function system is a finite col-
lection of contractive affine transformations.
As in the case of self-similar IFS, each self-affine IFS F = {f1 , . . . , fM }
generates a unique nonempty compact invariant set K. The set K satisfies
the same identity (8.3.1).
The following example is fundamental.
Example 8.4.5 (Bedford, McMullen). Let Q = [0, 1]2 be the unit square.
Fix integers 2 ≤ q ≤ n and divide Q into qn rectangles of width 1/q and
height 1/n. Select some number N ≤ qn of these rectangles Q1 , . . . , QN .
Consider the self-affine IFS F consisting of the contractions (without ro-
tation or reflection) of Q onto the rectangles Qj . Let K be the invariant
set. We call K the Bedford–McMullen carpet associated to the rectangles
Q1 , . . . , QM . For example, if q = n = 3 and Q1 , . . . , Q8 are the eight squares
which touch ∂Q, then K is the standard Sierpiński carpet. Figure 8.3 illus-
trates a typical Bedford–McMullen carpet (with parameters q = 2, n = 3,
N = 5).

Figure 8.3. A Bedford–McMullen carpet

Several pathologies arise in the study of the dimensions and conformal

dimensions of self-affine sets which have no counterpart for self-similar sets.
Let us begin with the self-similar case.
118 8. GLOBAL QUASICONFORMAL DIMENSION

(i) Self-similar sets which satisfy the open set condition have (Haus-
dorff) dimension equal to the similarity dimension. Moreover, both
Minkowski and Assouad dimensions also equal this value.
(ii) The dimension of the invariant set varies continuously when we
vary the IFS continuously in the space IFS(RN , M ) of self-similar
IFS’s, provided the open set condition remains in effect.
(iii) There exist self-similar Bedford–McMullen sets of Hausdorff dimen-
sion strictly greater than one but conformal dimension equal to
one. For instance, let q = n = 2 and select three of the four sub-
squares. The resulting invariant set is a Sierpiński gasket of dimen-
sion log 3/ log 2 and conformal dimension one (see the preceding
section).
Each of these properties may fail if the self-similarity assumption is
relaxed.
(i) There is no obvious quantity analogous to the similarity dimen-
sion for self-affine systems which are not self-similar. Non-self-
similar Bedford–McMullen carpets may have distinct Hausdorff,
Minkowski and Assouad dimensions. See Theorems 8.4.6, 8.4.7
and 8.4.8 for more details.
(ii) There exist examples of one-parameter families of self-affine IFS’s,
all of which satisfy the open set condition, yet where the dimensions
of the invariant sets change discontinuously.
(iii) All non-self-similar Bedford–McMullen carpets either have Haus-
dorff dimension equal to one or have conformal dimension strictly
greater than one. See Theorem 8.4.11.
We briefly elaborate on the reason why the distinction in (iii) holds.
For simplicity, let n = 2. Suppose that F is a self-affine IFS in R2 which
is not self-similar. Fix a large square Q which contains the invariant set
K and consider its iterated forward images Am := ∪w∈W m fw (Q), where
f!w = fw1 ◦ · · · ◦ fwm if w = w1 · · · wm . Then Q = A0 ⊃ A1 ⊃ A2 ⊃ · · · and
m≥0 Am = K. (In fact, (Am ) converges to K in the Hausdorff metric.)
Since F is not self-similar, σ2 < σ1 . For large m, Am contains many long,
thin rectangles with dimensions σ2m  σ1m . If all of the contracting affine
transformations have a common linear part (as is the case for Bedford–
McMullen carpets, for example), we obtain a structure similar to the product
spaces of Examples 4.1.7 and 4.1.9. Recalling the discussion in Chapters 4
and 6, this observation suggests the existence of nontrivial lower bounds on
conformal dimension. Theorems 8.4.11 and 8.4.14 confirm this suspicion.
We now discuss the dimensions of Bedford–McMullen carpets. Let K
be the Bedford–McMullen carpet associated to rectangles Q1 , . . . , QN . We
assume that we are in the non-self-similar case, i.e., q < n. For each i =
1, . . . , q, let Ni be the number of selected rectangles in the ith column.
 8.4. CONFORMAL DIMENSIONS OF SELF-AFFINE SETS 119

Theorem 8.4.6 (McMullen). The Hausdorff dimension dim K of K is


log ( qi=1 Niα )
(8.4.2) ,
log q
log q
where α = log n < 1.

In many cases, the (upper) Minkowski dimension dimM K of K will be

larger than this value. In particular, so will the Assouad dimension dimA K
(since dimA K ≥ dimM K always).
Theorem 8.4.7 (McMullen). The upper Minkowski dimension of K,
dimM K, is
log q  log(N/q  )
+ ,
log q log n
where q  = #{h ∈ {1, . . . , q} : Nh ≥ 1} denotes the number of columns which
contain at least one selected rectangle.
We omit the definition of upper Minkowski dimension.
Observe that
q
α
q  q 
Niα = Niα χ{h:Nh ≥1} (i) ≤ Ni (q  )1−α = N α (q  )1−α
i=1 i=1 i=1
from which it follows that

log( qi=1 Niα ) log q  log(N/q  ) log q  log(N/q  )
≤ +α = + .
log q log q log q log q log n
Strict inequality holds unless the vectors
(8.4.3) (N1α , . . . , Nqα )
and
(8.4.4) (χ{h:Nh ≥1} (1), . . . , χ{h:Nh ≥1} (q))
are proportional, i.e., there exists λ so that
(8.4.5) Niα = λχ{h:Nh ≥1} (i)
for all i = 1, . . . , q. Equation 8.4.5 implies that for each i ∈ {1, . . . , q},
either Ni = 0 or Ni = λ1/α . Thus the Hausdorff and Minkowski dimensions
agree only in the case of uniform vertical fibers, when the number of selected
rectangles in each nonempty column is the same.
Next, we discuss the Assouad dimension of such carpets. Let Nmax
denote the maximum of the quantities N1 , . . . , Nq .
Theorem 8.4.8 (Mackay). The Assouad dimension of K, dimA K, is
log q  log(Nmax )
+ ,
log q log n
where q  is as in Theorem 8.4.7.
120 8. GLOBAL QUASICONFORMAL DIMENSION
q
Since N = i=1 Ni ≤ #{h : Nh ≥ 1}Nmax ,
log q  log(N/q  ) log q  log(Nmax )
+ ≤ + .
log q log n log q log n
Strict inequality holds unless the vectors in (8.4.3) and (8.4.4) are pro-
portional, which implies that for each i ∈ {1, . . . , q}, either Ni = 0 or
Ni = Nmax . Thus Minkowski and Assouad dimensions also agree only in
the case of uniform vertical fibers.
Corollary 8.4.9. Assume that K is a Bedford–McMullen carpet with
uniform vertical fibers. Then dim K = dimM K = dimA K.
Note that the Hausdorff dimension of a general Bedford–McMullen car-
pet depends on not only the number of selected rectangles but also their loca-
tion relative to each other. For an example, consider the Bedford–McMullen
carpet K illustrated in Figure 8.3. We compute
log(2 + 2α )
dim K = ≈ 1.827 · · ·
log 2
log 5/2
< dimM K = 1 + ≈ 1.834 · · ·
log 3
< dimA K = 2,
where α = log 2
log 3 .
We briefly sketch the proof of Theorem 8.4.6, illustrating the use of self-
affine measures. Consider probability vectors p = (p1 , . . . , pN ) associated to
the rectangles Q1 , . . . , QN . Following the notation in the statement of the
theorem, we will reindex the entries in p, writing p = (p11 , . . . , pqNq ). As
before, let λp be the product measure on W ∞ and let μp = π# (λp ) be the
pushforward measure on K.
Lemma 8.4.10. Suppose that pik = pi is independent of k. Then
q  
1  1 α−1
dim μp = pi log + pi log(Ni ) Ni .
log q pi
i=1

The maximum value among all nonnegative vectors p satisfying qi=1 Ni pi =
1 is obtained when
N α−1
pi = q i α.
h=1 Nh
This maximum value is equal to the quantity in (8.4.2).
The upper bounds for (upper) Minkowski and Assouad dimension in
Theorems 8.4.7 and 8.4.8 follow from rather straightforward covering ar-
guments. The lower bounds are more difficult. In the Assouad dimension
case, one can appeal to the preservation of Assouad dimension lower bounds
under weak tangents (Proposition 6.1.5). The weak tangents of non-self-
similar carpets can be explicitly described; they include products of suitable
 8.4. CONFORMAL DIMENSIONS OF SELF-AFFINE SETS 121

self-similar Cantor sets or intervals. Applying (5.3.1) to such weak tan-

gents results in the equality of the upper and lower bounds for the Assouad
dimension as asserted in Theorem 8.4.8.
We conclude with estimates for the conformal dimensions of self-affine
Bedford–McMullen carpets.
Theorem 8.4.11 (Binder–Hakobyan). Let p be the probability vector
given by
(8.4.6) pi = 1/(qNi ) for i = 1, . . . , q.
Then
%q 1/q %q α 1/q
log ( i=1 Ni ) log ( i=1 Ni )
(8.4.7) Cdim K ≥ dim μp = 1 + =1+ .
log n log q
The significance of the probability vector in (8.4.6) is the following: the
corresponding measure μp on K projects (in the first coordinate) onto the
Lebesgue measure L1 . Moreover, for Lebesgue almost every x ∈ [0, 1], the
dimension of the vertical slice {(x, y) : y ∈ R} ∩ K is equal to dim μp − 1.
Again, we obtain the structure of a space crossed with an interval, which we
know to be minimal for conformal dimension.
Remarks 8.4.12. (1) The inequality
% 1/q 
log ( qi=1 Niα ) log qi=1 Niα
1+ <
log q log q
between conformal Hausdorff dimension and Hausdorff dimension is the
arithmetic-geometric-mean inequality for the values N1α , . . . , Nqα . Equal-
ity holds if and only if all of these values are equal; this means that we are
in the setting of uniform vertical fibers, with all columns containing at least
one selected rectangle.
(2) The lower bound in (8.4.7) is always strictly greater than one. (Recall
that we are in the non-self-similar case.)
(3) In the example (Figure 8.3), Theorem 8.4.11 gives
log 6
Cdim K ≥ 1 + ≈ 1.815 · · · .
log 9
We indicate how Theorem 8.4.11 reduces to Proposition 4.5.6. The idea
is to identify parameterized families of subsets of K consisting of sets of
conformal dimension one, together with natural transversal measures. As
discussed above, the lack of self-similarity suggests a mechanism for this
identification, arising from the discrepancy between the two singular values
of the defining contractive matrices.
More precisely, we introduce the following notion.
Definition 8.4.13. Let Q = {Q1 , . . . , QN } be the defining rectangles
for a Bedford–McMullen carpet K. A subcollection E ⊂ Q is called a graph
collection if E contains exactly one square in each column.
122 8. GLOBAL QUASICONFORMAL DIMENSION

More generally, let Q(m) be the collection of mth level iterates of the
initial square [0, 1]2 under the action of the associated iterated function
system. (Thus Q = Q(1) .) A collection E ⊂ Q(m) is called a graph collection
if E contains exactly one square in each column.
A sequence of graph collections E (m) ⊂ Q(m) is said to be nested if for
each R ∈ E (m) there exists a unique R ∈ E (m−1) satisfying R ⊃ R.

Each nested sequence of graph collections E (1) , E (2) , . . . generates a set

E ⊂ K, defined as the intersection over all m ≥ 1 of the union of the elements
of E (m) . The set E is a graph over the interval [0, 1] in the horizontal axis.
Typically E is the graph of a function which is discontinuous everywhere, and
E is a Cantor set. However, E could also be connected. We call E = EE (m)
the limit set for the sequence (E (m) ).
Each such set E is equipped with a natural measure λE = (proj |E )# (L1 ),
where proj denotes projection to the horizontal axis. Note that λE (R) =
L1 (proj(R)) if R ∈ ∪m E (m) . Let E be the set of all possible measures λE
associated to limit sets E of sequences of graph collections. Since q < n
it follows (by the argument sketched above) that each of these limit sets is
minimal for conformal dimension: Cdim E = 1. Modulo technical details,
one can verify the remaining conditions in Proposition 4.5.6 for this family
E. In particular, (K, μp ) is (locally) dim μp -regular and Mod1 E > 0. By
the proposition, Cdim K ≥ dim μp as asserted.
The exact value of the conformal Assouad dimension can always be com-
puted for non-self-similar carpets.

Theorem 8.4.14 (Mackay). Every non-self-similar Bedford–McMullen

carpet is either minimal for conformal Assouad dimension or has conformal
Assouad dimension zero. More precisely, if K denotes such a carpet, then
CdimA K = dimA K if either Nmax = n or q  = q, while CdimA K = 0 if
both Nmax < n and q  < q.

In brief, the idea for the proof is to consider weak tangents as before.
In case either Nmax = n or q  = q, the weak tangents previously identi-
fied include the product of a self-similar Cantor set with an interval. Such
spaces have already been shown to be minimal for conformal dimension.
Since conformal Assouad dimension is well-behaved under weak tangents
(Proposition 6.1.7), the first claim follows. For the second claim, one must
show by a direct geometric argument that the carpet in question is uniformly
disconnected, and then apply Theorem 5.1.9.

8.5. Sets with global quasiconformal dimension zero

Here we revisit the topic of sets and spaces of conformal dimension zero.
The full statement of Kovalev’s theorem is in fact much stronger than the
result stated in Theorem 5.1.12.
 8.5. SETS WITH GLOBAL QUASICONFORMAL DIMENSION ZERO 123

Theorem 8.5.1 (Kovalev). Let S be a subset of a separable Banach

space V , and assume that dim S < 1. Then for each  > 0 there exists a
quasisymmetric map F : V → V so that dim F (S) < .
Recall that any separable metric space (Z, d) embeds isometrically in
the separable Banach space C([0, 1]) (with the uniform norm); this is the
Banach–Mazur theorem. On the other hand, non-separable spaces necessar-
ily have infinite conformal dimension. Thus Theorem 5.1.12 is a consequence
of Theorem 8.5.1. On the other hand, choosing V = RN in Theorem 8.5.1
yields
Corollary 8.5.2. If S ⊂ RN satisfies dim S < 1, then GQCdim S = 0.
In other words, for any S ⊂ RN , GQCdim S ∈ {0} ∪ [1, n].
The importance of Theorem 8.5.1 lies especially in the powerful new tech-
niques developed for its proof, in particular, the notions of quasisymmetric
accretivity and δ-monotone mappings. These tools have found significant
use in a variety of contexts; see the notes to this section for more details.
To simplify the discussion, we consider the case when V is a Hilbert
space. The general case is only slightly more complicated, requiring the
notions of duality mapping and semi-inner product in general Banach spaces.
Definition 8.5.3 (Kovalev). Let V be a Hilbert space and let F : V →
V be a nonconstant mapping. We say that F is η-quasisymmetrically accre-
tive (η-QSA), where η is a distortion function, if the inequality
 
||x − y|| F (x) − F (z), x − z
(8.5.1) ||F (x) − F (y)|| ≤ η
||x − z|| ||x − z||
holds for all x, y, z ∈ V , x = z.
We collect a few easy consequences of the definition. Every quasisym-
metrically accretive mapping is injective, indeed, if F satisfies (8.5.1) and
F (x) = F (z) for some x = z then we quickly deduce that f is constant. Ev-
ery quasisymmetrically accretive mapping is also quasisymmetric, with the
same distortion function. Finally, every quasisymmetrically accretive map-
ping is accretive (or monotone): F (x) − F (z), x − z ≥ 0 for all x, z ∈ V .
Informally speaking, “F makes an angle of at most π2 with the identity
mapping”.
Example 8.5.4. Let V be a Hilbert space and let F : V → V be defined
by F (x) = ||x||α−1 x. Then F is quasisymmetrically accretive for every
α > 0. We check that F is both quasisymmetric and accretive. (Note that
quasisymmetric accretivity is in general stronger than the combination of
quasisymmetry and accretivity.)
To see that F is accretive, let x, z ∈ V . Then
||x||α−1 x − ||z||α−1 z, x − z = ||x||α+1 − (||x||α−1 + ||z||α−1 )x, z + ||z||α+1
≥ (||x||α − ||z||α )(||x|| − ||z||) ≥ 0.
124 8. GLOBAL QUASICONFORMAL DIMENSION

To see that F is quasisymmetric, we observe that in order to verify the

relevant three-point condition it suffices to verify it in finite-dimensional
(in fact, three-dimensional) subspaces W of V . By the definition of F , the
restriction of F to W maps W into W . Since V is Hilbert, W is a Euclidean
space. Now the quasisymmetry of this function on Euclidean space is a
classical fact; it is easy to prove by analytic techniques that the map is
quasiconformal.
One of the most important features of QSA mappings is the fact that
they form a large linear subclass of all quasisymmetric mappings. In fact,
the set of all η-QSA maps f : V → V forms a convex cone. The full class of
quasisymmetric mappings is highly nonlinear and difficult to investigate by
linear techniques. The linearity of QSA mappings allows us to form averages
and convolutions.
Let Lp ([0, 1] : V ) be the Lebesgue–Bochner space of p-integrable V -
valued functions on the interval [0, 1]. Each f ∈ Lp ([0, 1] : V ) and 0 < α < 1
produce a map Tα,f : V → V given by
 1
Tα,f (x) = x + ||x − f (t)||α−1 (x − f (t)) dt
0
=x+ ||x − z||α−1 (x − z) d(f# L1 )(z).
V
By Example 8.5.4, Tα,f is QSA. It is also surjective. The estimate
Tα,f (x) − Tα,f (y), x − y ≥ ||x − y||2 + c||x − y||α+1 (f# L1 )(B(x, ||x − y||))
for x, y ∈ V follows from Example 8.5.4, (8.5.1) and a simple calculation.
Hence
(8.5.2) ||Tα,f (x) − Tα,f (y)|| ≥ ||x − y|| + c||x − y||α (f# L1 )(B(x, ||x − y||)).
Now define
Dα f := Tα,f ◦ f.
Proposition 8.5.5 (Kovalev). For f and α as above, the following con-
clusions hold:
(a) Dα is a surjective map from Lp ([0, 1] : V ) to itself, and
(b) Dα is expansive: ||Dα f (t1 )−Dα f (t2 )|| ≥ ||f (t1 )−f (t2 )|| for all 0 ≤
t1 , t2 ≤ 1.
The most difficult point is the surjectivity of Dα . The proof involves a
density argument coupled with the Aubin–Lions compactness theorem and
some basic functional analysis.
We illustrate the usefulness of this theory by proving the following vari-
ant of Theorem 8.5.1, where Hausdorff dimension is replaced by Assouad
dimension.
Theorem 8.5.6. Let S be a bounded subset of a Hilbert space V and
assume that the Assouad dimension of S is strictly less than one. Then
 8.6. NOTES 125

for each  > 0 there exists a quasisymmetric map f : V → V so that

dimA f (S) < . In other words, for any bounded S ⊂ V , GQCdimA S ∈
{0} ∪ [1, ∞].
Proof of Theorem 8.5.6. Let dimA S < s < 1. Then there exists
an s-homogeneous measure μ (see Definition 4.6.1) supported on S. In
particular, μ(B(x, r)) ≥ crs for all x ∈ S and r ≤ r0 for some r0 > 0 and
c > 0. An appeal to the isomorphism theorem for uncountable complete
separable metric measure spaces yields a function f ∈ L1 ([0, 1] : V ) so that
μ = f# L1 . Define
α := (s−1 − 1).
By Proposition 8.5.5(a), there exists g ∈ L1 ([0, 1] : V ) so that Dα g = f .
−1 . Note that F is defined on all of V since T
Set F := Tα,g α,g is surjective.
As indicated above, F is QSA, hence quasisymmetric for some distortion
function η. We will show that ν := F# μ = g# L1 is -homogeneous; this will
imply that dimA F (S) ≤ .
For each ball B(x, r) in V , choose  ≤ L so that
(8.5.3) B(F (x), ) ⊂ F (B(x, r)) ⊂ B(F (x), L).
with L/ ≤ η(1). (See Lemma 1.2.18.) Choose y ∈ V so that ||x − y|| = r
and ||F (x) − F (y)|| = L. By (8.5.2),
r = ||x − y|| = ||Tα,g (F (x)) − Tα,g (F (y))||
≥ c||F (x) − F (y)||α (g# L1 )(B(F (x), ||F (x) − F (y)||))
≥ cLα (f# L1 )(B(x, r)) = cLα μ(B(x, r)) ≥ cLα rs .

Thus r1−s ≥ cLα or L ≤ Crs/ . From (8.5.3) we observe that ν is an

-homogeneous doubling measure supported on F (S). This completes the
proof. 

8.6. Notes
Distortion of Hausdorff dimension by quasiconformal maps is a classical
topic in Euclidean geometric function theory. Gehring and Väisälä’s theorem
8.1.1 comes from [59] and relies on Gehring’s higher integrability lemma (see
Remark 1.2.14). The sharp K-dependent bounds for Hausdorff dimension
distortion are unknown except in dimension two, where they follow from
Astala’s theorem 1.5.1. The precise exponents in the planar case are
2s 2Ks
α(2, K, s) = and β(2, K, s) = .
s + K(2 − s) Ks + (2 − s)
In particular, if S ⊂ C satisfies dim S = 1 and f : C → C is K-quasiconformal,
then
2 2K
≤ dim f (S) ≤ .
1+K 1+K
126 8. GLOBAL QUASICONFORMAL DIMENSION

Smirnov [156] has established a long-standing conjecture by verifying an

improved upper bound for the dimensions of quasicircles: if f : C → C is
K-quasiconformal, then
 
K−1 2
dim f (S ) ≤ 1 +
1
.
K+1
Corollary 8.1.2 sharpens the well-known fact that quasiconformal maps
of RN , n ≥ 2, preserve sets of positive Lebesgue measure (Lusin’s condition
N). Both Lusin’s condition N and Corollary 8.1.2 fail dramatically in the
case n = 1. See Tukia’s example 5.2.2.
Bishop’s theorem 8.1.4 provides further evidence for the flexibility of
quasiconformal maps in RN ; global dimension raising mappings exist in
abundance. This theorem appeared in [16]. Its analog for Assouad dimen-
sion remains an open problem.
Quasisymmetric deformations of Cantor sets have been studied exten-
sively. Examples illustrating Remark 8.2.2 were given by Bishop and Tyson
in [17]. The remaining direction in Theorem 8.2.4 (dim K(a) = 1 implies
K(a) is minimal for conformal dimension) is due to Hakobyan [72], as is
Corollary 8.2.5. A more restrictive condition, introduced by Staples and
Ward [160] is quasisymmetric thickness. A set K ⊂ R is called quasisym-
metrically thick if L1 (f (K)) > 0 for all quasisymmetric f : R → R. Note
that this notion is vacuous in higher dimensions in view of the Lusin prop-
erty of quasiconformal maps. The characterization of quasisymmetrically
thick middle interval Cantor sets is due to Staples–Ward [160] and Buckley–
Hanson–MacManus [38]: K(a) is quasisymmetrically thick if and only if
a ∈ ∩p>0 p .
Corollary 8.1.2 shows that sets in RN of dimension n are minimal for
global quasiconformal dimension. However, they need not be minimal for
conformal dimension. It is a subtle question to understand precisely which
subsets of RN of dimension n are minimal for conformal dimension. Theorem
8.2.4 answers this question completely for middle interval Cantor sets in the
real line. For a sequence a = (am ) of reciprocals of odd integers, consider
the middle square Sierpiński carpet
" 
SC(a) := Qw ,
m≥1 w

where the union is taken over all subcubes Qw of Q = [0, 1]2 , w = w1 w2 · · · wm ,

and wj is chosen from the letters 1, 2, . . . , a2j −1. For each word w, we assume
a2m −1
that Qw = j=1 Qwj ∪ Rw , where the subcube Rw has the same center
as Qw and has side length equal to am times the side length of Qw , while
the remaining cubes Qwj are Euclidean translates of Rw . The standard
Sierpiński carpet SC coincides with SC(a) for a = ( 13 , 13 , 13 , . . .). Similar to
the picture for middle interval Cantor sets, we have dim SC(a) = 2 ⇔ a ∈ c0
and L2 (SC(a)) > 0 ⇔ a ∈ 2 . The following observation from [112] is a
consequence of more general results on nontriviality of modulus and validity
 8.6. NOTES 127

of Poincaré inequalities for middle square Sierpiński carpets: If a ∈ 2 , then

SC(a) is minimal for conformal dimension.
The term “fractal” was introduced by Mandelbrot [113] to describe sets
and spaces with no a priori smooth structure but which exhibit some degree
of rough self-similarity. It is common to use the term to refer to spaces
whose (Hausdorff) dimension exceeds its topological dimension; this defini-
tion covers most of the examples described in this survey. Hutchinson [82]
was the first to coherently argue for the description of fractal sets in terms
of their defining equations, i.e., through the use of iterated function systems.
The Moran–Hutchinson theorem 8.3.2 appeared in [82] but dates back to
the earlier work of Moran [126]. Schief [150] made a major advance by
proving the following converse to Theorem 8.3.2.
Theorem 8.6.1 (Schief). If K is the invariant set for a self-similar
iterated function system F in RN and 0 < Hd (K) < ∞, where d is the
similarity dimension for F , then F satisfies the open set condition.
Further work on the Moran–Hutchinson theorem in the setting of more
general spaces can be found in Schief [151] and Balogh–Rohner [10]. The
approach to fractal geometry via iterated function systems is highly popular
in the computational and computer graphics communities.
Determining the conformal dimension of the standard 13 Sierpiński carpet
SC is a long standing open problem. Recall that SC is not minimal for
conformal dimension, by the work of Keith and Laakso. In the same vein,
we ask:
Question 8.6.2. Is SC minimal for global quasiconformal dimension?
In other words, does there exist a quasiconformal map F : C → C such that
dim F (SC) < log 8
log 3 = dim SC?

Theorem 8.3.9, Example 8.3.11 and Remark 8.3.12 are taken from Tyson–
Wu [173]. Laakso (unpublished) obtained similar results at approximately
the same time. For invariant sets of boundary congruent gasket type sys-
tems, Theorem 8.3.9 provides upper bounds on the conformal and global
quasiconformal dimensions in terms of arbitrary isomorphic systems. The
technique can be used to compute the conformal and global quasiconformal
dimensions of other invariant sets.
For example, the polygasket P G(N ), N ≥ 3, defined by replacing the
equilateral triangle in the definition of SG with an N -gon and adjusting the
contracting ratios until the self-similar pieces are “just-touching”, has global
quasiconformal dimension equal to one whenever N is not a multiple of 4.
See Figure 8.4.
As an application of Theorem 6.3.2, we observe that the polygasket
P G(N ) is not minimal for snowflake dimension if N = 3, 4, 6 [172]. Put
another way, for each N ∈ {5, 7, 8, 9, . . .} there exists p = p(N ) > 1 and a
metric d on P G(N ) so that d1/p is bi-Lipschitz equivalent to the Euclidean
metric on P G(N ). In particular, all nontrivial connected subsets of P G(N )
128 8. GLOBAL QUASICONFORMAL DIMENSION

Figure 8.4. Polygaskets P G(N )

have Hausdorff dimension at least p > 1. (The restriction on N is necessary

since P G(3), P G(4) and P G(6) are rectifiably connected.) To see why
Theorem 6.3.2 applies in this setting, we note that the self-similarity of
P G(N ) implies that it is enough to verify that P G(N ) itself contains no
rectifiable curves, or even that it contains no nontrivial line segments. The
latter fact can be observed directly from the construction of P G(N ), by
elementary geometry.
The relationship between the dynamics of rational maps and self-similar
planar iteration function systems has been explored by Kameyama [87],
[88], Bandt [11] and Solomyak [157], [158]. The definition of an invertible
IFS is taken from Kameyama [87], while Theorem 8.3.13 is proved in [50].
Example 8.3.14 appears as Example 4.1 in [50]; the topological equivalence
of the systems (SG, g) and (J(R), R) was shown previously by Kameyama
[88]. The conclusion in (8.3.6) is original to this survey. We do not know
whether the quantity
inf dim J(R̃),
where the infimum is taken over all rational maps R̃ whose Julia set J(R̃)
is quasisymmetrically equivalent to J(R), is equal to one.
McMullen [117] was the first to compute the Hausdorff dimensions of
self-affine carpets. The computation was later redone and extended by Bed-
ford [12], Peres [141], [142] and others. Theorem 8.4.11 comes from the
recent work [15], while Theorems 8.4.8 and 8.4.14 come from [109].
The material from section 8.5 comes from the work of Kovalev [95].
The concepts of quasisymmetric accretivity and monotonicity have proved
extremely fruitful in geometric function theory. A closely related class of
maps are the so-called δ-monotone maps, δ > 0, defined by the condition
F (x) − F (z), x − z ≥ δ||F (x) − F (z)|| ||x − z||, x, z ∈ V.
Informally speaking, such maps “form an angle strictly less than π2 with
the identity map”. They have been studied intensively by Kovalev and
his collaborators [94], [96], [98], [97], [83]. A key observation is that δ-
monotone mappings of RN are quasisymmetric. The theory of such maps
is closely allied with the study of quasiconformal gradient mappings, the
Monge–Ampere equation, existence and uniqueness theorems for first-order
 8.6. NOTES 129

ODE’s with weakly regular coefficients, the quasiconformal Jacobian prob-

lem and many other subjects. See the above references for more details.
The proof of Theorem 8.5.6 is somewhat simpler than that for the full
Theorem 8.5.1 of Kovalev. In the case of Hausdorff dimension distortion the
choice of the measure μ is more intricate. In case V = RN , a different proof
for Theorem 8.5.6 appears in [170].
Conformal and global quasiconformal dimensions of subsets of the Heisen-
berg group have been studied by Balogh [9] and Tyson [171].
 Bibliography

[1] Allcock, D. Reflection groups on the octave hyperbolic plane. J. Algebra 213, 2
(1999), 467–498.
[2] Aronsson, G. Extension of functions satisfying Lipschitz conditions. Ark. Mat. 6
(1967), 551–561.
[3] Aronsson, G., Crandall, M. G., and Juutinen, P. A tour of the theory of
absolutely minimizing functions. Bull. Amer. Math. Soc. (N.S.) 41, 4 (2004), 439–
505 (electronic).
[4] Assouad, P. Plongements lipschitziens dans Rn . Bull. Soc. Math. France 111
(1983), 429–448.
[5] Astala, K. Area distortion of quasiconformal mappings. Acta Math. 173, 1 (1994),
37–60.
[6] Astala, K., Iwaniec, T., and Martin, G. Elliptic partial differential equations
and quasiconformal mappings in the plane. Princeton Mathematical Series. Prince-
ton University Press, Princeton, NJ, 2008.
[7] Baez, J. C. The octonions. Bull. Amer. Math. Soc. (N.S.) 39, 2 (2002), 145–205
(electronic).
[8] Ballmann, W. Lectures on spaces of nonpositive curvature, vol. 25 of DMV Semi-
nar. Birkhäuser Verlag, Basel, 1995. With an appendix by Misha Brin.
[9] Balogh, Z. M. Hausdorff dimension distribution of quasiconformal mappings on
the Heisenberg group. J. Anal. Math. 83 (2001), 289–312.
[10] Balogh, Z. M., and Rohner, H. Self-similar sets in doubling spaces. Illinois J.
Math. 51, 4 (2007), 1275–1297.
[11] Bandt, C. On the Mandelbrot set for pairs of linear maps. Nonlinearity 15, 4 (2002),
1127–1147.
[12] Bedford, T. Crinkly curves, Markov partitions and dimension. PhD thesis, Uni-
versity of Warwick, 1984.
[13] Bellaı̈che, A. The tangent space in sub-Riemannian geometry. In Sub-Riemannian
geometry. Birkhäuser, Basel, 1996, pp. 1–78.
[14] Beurling, A., and Ahlfors, L. The boundary correspondence under quasiconfor-
mal mappings. Acta Math. 96 (1956), 125–142.
[15] Binder, I., and Hakobyan, H. Conformal dimension of self-affine and random
Bedford–McMullen carpets. In preparation.
[16] Bishop, C. J. Quasiconformal mappings which increase dimension. Ann. Acad. Sci.
Fenn. Ser. A I Math. 24 (1999), 397–407.
[17] Bishop, C. J., and Tyson, J. T. Locally minimal sets for conformal dimension.
Ann. Acad. Sci. Fenn. Ser. A I Math. 26 (2001), 361–373.
[18] Bishop, R. L., and Crittenden, R. J. Geometry of manifolds. AMS Chelsea
Publishing, Providence, RI, 2001. Reprint of the 1964 original.
[19] Bonk, M. Quasiconformal geometry of fractals. In International Congress of Math-
ematicians. Vol. II. Eur. Math. Soc., Zürich, 2006, pp. 1349–1373.
[20] Bonk, M., Heinonen, J., and Koskela, P. Uniformizing Gromov hyperbolic
spaces. Astérisque 270 (2001).

131
132 BIBLIOGRAPHY

[21] Bonk, M., and Kleiner, B. Quasisymmetric parametrizations of two-dimensional

metric spheres. Invent. Math. 150, 1 (2002), 127–183.
[22] Bonk, M., and Kleiner, B. Conformal dimension and Gromov hyperbolic groups
with 2-sphere boundary. Geom. Topol. 9 (2005), 219–246.
[23] Bonk, M., and Merenkov, S. Quasisymmetric rigidity of Sierpiński carpets.
Preprint, 2008.
[24] Bonk, M., and Meyer, D. In preparation.
[25] Bonk, M., and Saksman, E. In preparation.
[26] Bonk, M., and Schramm, O. Embeddings of Gromov hyperbolic spaces. Geom.
Funct. Anal. 10, 2 (2000), 266–306.
[27] Bourdon, M. Au bord de certains polyèdres hyperboliques. Ann. Inst. Fourier
(Grenoble) 45 (1995), 119–141.
[28] Bourdon, M. Structure conforme au bord et flot géodésique d’un CAT(−1)-espace.
Enseign. Math. (2) 41, 1-2 (1995), 63–102.
[29] Bourdon, M. Immeubles hyperboliques, dimension conforme et rigidité de Mostow.
Geom. Funct. Anal. 7 (1997), 245–268.
[30] Bourdon, M. Sur les immeubles fuchsiens et leur type de quasi-isométrie. Ergodic
Theory Dynam. Systems 20, 2 (2000), 343–364.
[31] Bourdon, M., and Pajot, H. Poincaré inequalities and quasiconformal structure
on the boundaries of some hyperbolic buildings. Proc. Amer. Math. Soc. 127, 8
(1999), 2315–2324.
[32] Bourdon, M., and Pajot, H. Rigidity of quasi-isometries for some hyperbolic
buildings. Comment. Math. Helv. 75, 4 (2000), 701–736.
[33] Bourdon, M., and Pajot, H. Quasi-conformal geometry and hyperbolic geometry.
In Rigidity in dynamics and geometry (Cambridge, 2000). Springer, Berlin, 2002,
pp. 1–17.
[34] Bourdon, M., and Pajot, H. Cohomologie lp et espaces de Besov. J. reine angew.
Math. 558 (2003), 85–108.
[35] Bowditch, B. H. Cut points and canonical splittings of hyperbolic groups. Acta
Math. 180, 2 (1998), 145–186.
[36] Bowditch, B. H. A topological characterisation of hyperbolic groups. J. Amer.
Math. Soc. 11, 3 (1998), 643–667.
[37] Bridson, M. R., and Haefliger, A. Metric spaces of non-positive curvature,
vol. 319 of Grundlehren der Mathematischen Wissenschaften [Fundamental Princi-
ples of Mathematical Sciences]. Springer-Verlag, Berlin, 1999.
[38] Buckley, S. M., Hanson, B., and MacManus, P. Doubling for general sets.
Math. Scand. 88, 2 (2001), 229–245.
[39] Burago, D., Burago, Y., and Ivanov, S. A course in metric geometry, vol. 33
of Graduate Studies in Mathematics. American Mathematical Society, Providence,
RI, 2001.
[40] Cannon, J. W., Floyd, W. J., and Parry, W. R. Squaring rectangles: the
finite Riemann mapping theorem. In The mathematical legacy of Wilhelm Magnus:
groups, geometry and special functions (Brooklyn, NY, 1992), no. 169 in Contemp.
Math. Amer. Math. Soc., Providence, RI, 1994, pp. 133–212.
ˇ
[41] Cap, A., Cowling, M. G., de Mari, F., Eastwood, M., and McCallum, R.
The Heisenberg group, SL(3, R), and rigidity. In Harmonic analysis, group represen-
tations, automorphic forms and invariant theory, vol. 12 of Lect. Notes Ser. Inst.
Math. Sci. Natl. Univ. Singap. World Sci. Publ., Hackensack, NJ, 2007, pp. 41–52.
[42] Capogna, L., Danielli, D., Pauls, S. D., and Tyson, J. T. An introduction
to the Heisenberg group and the sub-Riemannian isoperimetric problem, vol. 259 of
Progress in Mathematics. Birkhäuser, 2007.
[43] Casson, A., and Jungreis, D. Convergence groups and Seifert fibered 3-manifolds.
Invent. Math. 118, 3 (1994), 441–456.
 BIBLIOGRAPHY 133

[44] Cheeger, J. Differentiability of Lipschitz functions on metric measure spaces.

Geom. Funct. Anal. 9 (1999), 428–517.
[45] Coornaert, M., Delzant, T., and Papadopoulos, A. Géométrie et théorie des
groupes. No. 1441 in Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1990.
Les groupes hyperboliques de Gromov. [Gromov hyperbolic groups], With an English
summary.
[46] David, G., and Semmes, S. Strong A∞ weights, Sobolev inequalities and quasicon-
formal mappings. In Analysis and partial differential equations, vol. 122 of Lecture
Notes in Pure and Appl. Math. Marcel Dekker, 1990, pp. 101–111.
[47] David, G., and Semmes, S. Fractured fractals and broken dreams: self-similar
geometry through metric and measure, vol. 7 of Oxford Lecture Series in Mathematics
and its Applications. Oxford University Press, 1997.
[48] Douady, A., and Hubbard, J. H. A proof of Thurston’s topological characteriza-
tion of rational functions. Acta Math. 171, 2 (1993), 263–297.
[49] Dunwoody, M. J. The accessibility of finitely presented groups. Invent. Math. 81,
3 (1985), 449–457.
[50] Eroğlu, K. I., Rohde, S., and Solomyak, B. Quasisymmetric conjugacy between
quadratic dynamics and iterated function systems. Ergodic Theory Dynam. Systems.
to appear.
[51] Falconer, K. Techniques in fractal geometry. John Wiley & Sons Ltd., Chichester,
1997.
[52] Falconer, K. J. Fractal geometry. Mathematical Foundations and Applications.
John Wiley and Sons Ltd., Chichester, 1990.
[53] Federer, H. Geometric Measure Theory, vol. 153 of Die Grundlehren der mathe-
matischen Wissenschaften. Springer-Verlag, New York, 1969.
[54] Fuglede, B. Extremal length and functional completion. Acta Math. 98 (1957),
171–219.
[55] Gabai, D. Convergence groups are Fuchsian groups. Bull. Amer. Math. Soc. (N.S.)
25, 2 (1991), 395–402.
[56] Gabai, D. Convergence groups are Fuchsian groups. Ann. of Math. (2) 136, 3
(1992), 447–510.
[57] Gehring, F. W. Symmetrization of rings in space. Trans. Amer. Math. Soc. 101
(1961), 499–519.
[58] Gehring, F. W. Rings and quasiconformal mappings in space. Trans. Amer. Math.
Soc. 103 (1962), 353–393.
[59] Gehring, F. W., and Väisälä, J. Hausdorff dimension and quasiconformal map-
pings. J. London Math. Soc. (2) 6 (1973), 504–512.
[60] Ghys, E., and de la Harpe, P. Sur les Groupes Hyperboliques d’après Mikhael
Gromov. Progress in Mathematics. Birkhäuser, Boston-Basel-Berlin, 1990.
[61] Goldman, W. M. Complex hyperbolic geometry. Oxford Mathematical Monographs.
The Clarendon Press Oxford University Press, New York, 1999. Oxford Science
Publications.
[62] Gromov, M. Structures Métriques pour les Variétés Riemanniennes, (J. Lafontaine
and P. Pansu, ed.). Cedic/Fernand Nathan, Paris, 1981.
[63] Gromov, M. Asymptotic geometry of homogeneous spaces. Rend. Sem. Mat. Univ.
Politec. Torino, Special Issue (1983), 59–60 (1984). Conference on differential geom-
etry on homogeneous spaces (Turin, 1983).
[64] Gromov, M. Hyperbolic groups. In Essays in Group Theory, S. Gersten, Editor.
MSRI Publications, Springer-Verlag, 1987, pp. 75–265.
[65] Gromov, M. Asymptotic invariants of infinite groups. In Geometric group theory,
Vol. 2 (Sussex, 1991), vol. 182 of London Math. Soc. Lecture Note Ser. Cambridge
Univ. Press, Cambridge, 1993, pp. 1–295.
134 BIBLIOGRAPHY

[66] Gromov, M. Carnot-Carathéodory spaces seen from within. In Sub-Riemannian

Geometry, vol. 144 of Progress in Mathematics. Birkhäuser, Basel, 1996, pp. 79–
323.
[67] Gromov, M., and Pansu, P. Rigidity of lattices: an introduction. In Geometric
topology: recent developments (Montecatini Terme, 1990), vol. 1504 of Lecture Notes
in Math. Springer, Berlin, 1991, pp. 39–137.
[68] Haı̈ssinsky, P. Empilements de cercles et modules combinatoires. Ann. Inst.
Fourier (Grenoble) 59, 6 (2009), 2175–2222.
[69] Haı̈ssinsky, P., and Pilgrim, K. M. Thurston obstructions and Ahlfors regular
conformal dimension. J. Math. Pures Appl. (9) 90, 3 (2008), 229–241.
[70] Haı̈ssinsky, P., and Pilgrim, K. M. Coarse expanding conformal dynamics.
Astérisque 325 (2009).
[71] Hajlasz, P. Whitney’s example by way of Assouad’s embedding. Proc. Amer. Math.
Soc. 131, 11 (2003), 3463–3467.
[72] Hakobyan, H. Cantor sets which are minimal for quasisymmetric maps. J. Con-
temp. Math. Analysis 41 (2006), 5–13.
[73] Hakobyan, H. Conformal dimension: Cantor sets and Fuglede modulus. Int. Math.
Res. Notices 2010, 1 (2009), 87–111.
[74] Heinonen, J. Calculus on Carnot groups. In Fall School in Analysis (Jyväskylä,
1994), vol. 68. Ber. Univ. Jyväskylä Math. Inst., Jyväskylä, 1995, pp. 1–31.
[75] Heinonen, J. A capacity estimate on Carnot groups. Bull. Sci. Math. 119 (1995),
475–484.
[76] Heinonen, J. Lectures on analysis on metric spaces. Springer-Verlag, New York,
2001.
[77] Heinonen, J., and Koskela, P. Quasiconformal maps in metric spaces with con-
trolled geometry. Acta Math. 181 (1998), 1–61.
[78] Helgason, S. Differential geometry, Lie groups, and symmetric spaces, vol. 34 of
Graduate Studies in Mathematics. American Mathematical Society, Providence, RI,
2001. Corrected reprint of the 1978 original.
[79] Howroyd, J. D. On dimension and on the existence of sets of finite positive Haus-
dorff measure. Proc. London Math. Soc. (3) 70, 3 (1995), 581–604.
[80] Howroyd, J. D. On Hausdorff and packing dimension of product spaces. Math.
Proc. Cambridge Philos. Soc. 119, 4 (1996), 715–727.
[81] Hurewicz, W., and Wallman, H. Dimension Theory, vol. 4 of Princeton Mathe-
matical Series. Princeton University Press, Princeton, N. J., 1941.
[82] Hutchinson, J. E. Fractals and self-similarity. Indiana Univ. Math. J. 30 (1981),
713–747.
[83] Iwaniec, T., Kovalev, L. V., and Onninen, J. Dynamics of quasiconformal fields.
Preprint, 2008.
[84] Iwaniec, T., and Martin, G. Geometric function theory and non-linear analysis.
Oxford Mathematical Monographs. The Clarendon Press Oxford University Press,
New York, 2001.
[85] Järvenpää, E., Järvenpää, M., Rogovin, K., Rogovin, S., and Shanmu-
galingam, N. Measurability of equivalence classes and MECp -property in metric
spaces. Rev. Mat. Iberoam. 23, 3 (2007), 811–830.
[86] Jensen, R. Uniqueness of Lipschitz extensions: minimizing the sup norm of the
gradient. Arch. Rational Mech. Anal. 123, 1 (1993), 51–74.
[87] Kameyama, A. Julia sets and self-similar sets. Topology Appl. 54, 1-3 (1993), 241–
251.
[88] Kameyama, A. Julia sets of postcritically finite rational maps and topological self-
similar sets. Nonlinearity 13, 1 (2000), 165–188.
 BIBLIOGRAPHY 135

[89] Kapovich, I., and Benakli, N. Boundaries of hyperbolic groups. In Combinato-

rial and geometric group theory (New York, 2000/Hoboken, NJ, 2001), vol. 296 of
Contemp. Math. Amer. Math. Soc., Providence, RI, 2002, pp. 39–93.
[90] Kapovich, M., and Kleiner, B. Hyperbolic groups with low-dimensional bound-
ary. Ann. Sci. École Norm. Sup. (4) 33, 5 (2000), 647–669.
[91] Keith, S., and Laakso, T. Conformal Assouad dimension and modulus. Geom.
Funct. Anal. 14, 6 (2004), 1278–1321.
[92] Kleiner, B. The asymptotic geometry of negatively curved spaces: uniformization,
geometrization and rigidity. In International Congress of Mathematicians. Vol. II.
Eur. Math. Soc., Zürich, 2006, pp. 743–768.
[93] Koskela, P., and Rohde, S. Hausdorff dimension and mean porosity. Math. Ann.
309, 4 (1997), 593–609.
[94] Kovalev, L., Maldonado, D., and Wu, J.-M. Doubling measures, monotonicity,
and quasiconformality. Math. Z. 257, 3 (2007), 525–545.
[95] Kovalev, L. V. Conformal dimension does not assume values between zero and
one. Duke Math. J. 134, 1 (2006), 1–13.
[96] Kovalev, L. V. Quasiconformal geometry of monotone mappings. J. Lond. Math.
Soc. (2) 75, 2 (2007), 391–408.
[97] Kovalev, L. V., and Maldonado, D. Mappings with convex potentials and the
quasiconformal Jacobian problem. Illinois J. Math. 49, 4 (2005), 1039–1060 (elec-
tronic).
[98] Kovalev, L. V., and Maldonado, D. Convex functions and quasiconformal
mappings. In Harmonic analysis, partial differential equations, and related topics,
vol. 428 of Contemp. Math. Amer. Math. Soc., Providence, RI, 2007, pp. 93–103.
[99] Lang, U., Pavlović, B., and Schroeder, V. Extensions of Lipschitz maps into
Hadamard spaces. Geom. Funct. Anal. 10, 6 (2000), 1527–1553.
[100] Lang, U., and Schroeder, V. Kirszbraun’s theorem and metric spaces of bounded
curvature. Geom. Funct. Anal. 7, 3 (1997), 535–560.
[101] Larman, D. G. A new theory of dimension. Proc. London Math. Soc. 17 (1967),
178–192.
[102] Lattès, S. Sur l’itération des substitutions rationelles et les fonctions de Poincaré.
C. R. Acad. Sci. Paris 16 (1918), 26–28.
[103] Lee, J. R., and Naor, A. Absolute Lipschitz extendability. C. R. Math. Acad. Sci.
Paris 338, 11 (2004), 859–862.
[104] Lee, J. R., and Naor, A. Extending Lipschitz functions via random metric parti-
tions. Invent. Math. 160, 1 (2005), 59–95.
[105] Loewner, C. On the conformal capacity in space. J. Math. Mech. 8 (1959), 411–
414.
[106] Lupo-Krebs, G., and Pajot, H. Dimensions conformes, espaces Gromov-
hyperboliques et ensembles autosimilaires. In Séminaire de Théorie Spectrale et
Géométrie. Vol. 22. Année 2003–2004, vol. 22 of Sémin. Théor. Spectr. Géom.
Univ. Grenoble I, Saint, 2004, pp. 153–182.
[107] Luukkainen, J. Assouad dimension: antifractal metrization, porous sets, and ho-
mogeneous measures. J. Korean Math. Soc. 35 (1998), 23–76.
[108] Luukkainen, J., and Saksman, E. Every complete doubling metric space carries
a doubling measure. Proc. Amer. Math. Soc. 126 (1998), 531–534.
[109] Mackay, J. Assouad dimension of self-affine carpets. Preprint, 2010.
[110] Mackay, J. Spaces with conformal dimension greater than one. Duke Math J . to
appear.
[111] Mackay, J. Conformal dimension and the quasisymmetric geometry of metric
spaces. PhD thesis, University of Michigan, 2008.
[112] Mackay, J., Tyson, J. T., and Wildrick, K. Modulus and Poincaré inequalities
on non-self-similar Sierpiński carpets. In preparation.
136 BIBLIOGRAPHY

[113] Mandelbrot, B. B. The fractal geometry of nature. W. H. Freeman and Co., San
Francisco, Calif., 1982. Schriftenreihe für den Referenten. [Series for the Referee].
[114] Marstrand, J. M. The dimension of Cartesian product sets. Proc. Cambridge
Philos. Soc. 50 (1954), 198–202.
[115] Martio, O., and Vuorinen, M. Whitney cubes, p-capacity, and Minkowski con-
tent. Exposition. Math. 5, 1 (1987), 17–40.
[116] Mattila, P. Geometry of sets and measures in Euclidean spaces, vol. 44 of Cam-
bridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge,
1995.
[117] McMullen, C. The Hausdorff dimension of general Sierpiński carpets. Nagoya
Math. J. 96 (1984), 1–9.
[118] McMullen, C. T. Complex dynamics and renormalization, vol. 135 of Annals of
Mathematics Studies. Princeton University Press, Princeton, NJ, 1994.
[119] Meyer, D. Invariant Peano curves of expanding Thurston maps. Preprint, 2009.
[120] Meyer, D. Quasisymmetric embedding of self similar surfaces and origami with
rational maps. Ann. Acad. Sci. Fenn. Math. 27, 2 (2002), 461–484.
[121] Milnor, J. Dynamics in one complex variable, third ed., vol. 160 of Annals of
Mathematics Studies. Princeton University Press, Princeton, NJ, 2006.
[122] Milnor, J. On Lattès maps. In Dynamics on the Riemann sphere. Eur. Math. Soc.,
Zürich, 2006, pp. 9–43.
[123] Mineyev, I. Metric conformal structures and hyperbolic dimension. Conf. Geom.
Dynam. 11 (2007), 137–163.
[124] Mineyev, I., and Yu, G. The Baum-Connes conjecture for hyperbolic groups.
Invent. Math. 149, 1 (2002), 97–122.
[125] Montgomery, R. A tour of subriemannian geometries, their geodesics and appli-
cations, vol. 91 of Mathematical Surveys and Monographs. American Mathematical
Society, Providence, RI, 2002.
[126] Moran, P. A. P. Additive functions of intervals and Hausdorff measure. Proc.
Cambridge Philos. Soc. 42 (1946), 15–23.
[127] Mostow, G. D. Quasiconformal mappings in n-space and the rigidity of hyperbolic
space forms. Inst. Hautes Études Sci. Publ. Math. 34 (1968), 53–104.
[128] Mostow, G. D. Strong rigidity of locally symmetric spaces. Princeton University
Press, Princeton, N.J., 1973. Annals of Mathematics Studies, No. 78.
[129] Ottazzi, A. A sufficient condition for nonrigidity of Carnot groups. Math. Z. 259,
3 (2008), 617–629.
[130] Pansu, P. Pincement du plan hyperbolique complexe. Preprint, 2009.
[131] Pansu, P. Quasiconformal mappings and manifolds of negative curvature. In Cur-
vature and topology of Riemannian manifolds (Katata, 1985), vol. 1201 of Lecture
Notes in Math. Springer, Berlin, 1986, pp. 212–229.
[132] Pansu, P. Cohomologie Lp des variétés à courbure négative, cas du degré 1. Rend.
Sem. Mat. Univ. Politec. Torino, Special Issue (1989), 95–120 (1990). Conference
on Partial Differential Equations and Geometry (Torino, 1988).
[133] Pansu, P. Dimension conforme et sphère à l’infini des variétés à courbure négative.
Ann. Acad. Sci. Fenn. Ser. A I Math. 14 (1989), 177–212.
[134] Pansu, P. Métriques de Carnot-Carathéodory et quasiisométries des espaces
symétriques de rang un. Ann. of Math. (2) 129 (1989), 1–60.
[135] Pansu, P. Lp -cohomology and pinching. In Rigidity in dynamics and geometry
(Cambridge, 2000). Springer, Berlin, 2002, pp. 379–389.
[136] Pansu, P. Cohomologie Lp en degré 1 des espaces homogènes. Potential Anal. 27,
2 (2007), 151–165.
[137] Pansu, P. Cohomologie Lp et pincement. Comment. Math. Helv. 83, 2 (2008), 327–
357.
 BIBLIOGRAPHY 137

[138] Pansu, P. Lp -cohomology of symmetric spaces. In Geometry, analysis and topology

of discrete groups, vol. 6 of Adv. Lect. Math. (ALM). Int. Press, Somerville, MA,
2008, pp. 305–326.
[139] Parker, J. R. Complex Hyperbolic Kleinian Groups. Cambridge University Press.
to appear.
[140] Paulin, F. Un groupe hyperbolique est déterminé par son bord. J. London Math.
Soc. (2) 54, 1 (1996), 50–74.
[141] Peres, Y. The packing measure of self-affine carpets. Math. Proc. Cambridge Philos.
Soc. 115 (1994), 437–450.
[142] Peres, Y. The self-affine carpets of McMullen and Bedford have infinite Hausdorff
measure. Math. Proc. Cambridge Philos. Soc. 116, 3 (1994), 513–526.
[143] Peres, Y., Schramm, O., Sheffield, S., and Wilson, D. B. Tug-of-war and the
infinity Laplacian. J. Amer. Math. Soc. 22, 1 (2009), 167–210.
[144] Rademacher, H. Über partielle und totale differenzierbarkeit von Funktionen
mehrerer Variabeln und über die Transformation der Doppelintegrale. Math. Ann.
79, 4 (1919), 340–359.
[145] Rajala, K. The local homeomorphism property of spatial quasiregular mappings
with distortion close to one. Geom. Funct. Anal. 15, 5 (2005), 1100–1127.
[146] Rajala, K. Surface families and boundary behavior of quasiregular mappings. Illi-
nois J. Math. 49, 4 (2005), 1145–1153 (electronic).
[147] Reimann, H. M. Rigidity of H-type groups. Math. Z. 237, 4 (2001), 697–725.
[148] Reimann, H. M., and Ricci, F. The complexified Heisenberg group. In Proceedings
on Analysis and Geometry (Russian) (Novosibirsk Akademgorodok, 1999) (2000),
Izdat. Ross. Akad. Nauk Sib. Otd. Inst. Mat., Novosibirsk, pp. 465–480.
[149] Rogers, C. A. Hausdorff Measures. Cambridge University Press, London, 1970.
[150] Schief, A. Separation properties for self-similar sets. Proc. Amer. Math. Soc. 122,
1 (1994), 111–115.
[151] Schief, A. Self-similar sets in complete metric spaces. Proc. Amer. Math. Soc. 124,
2 (1996), 481–490.
[152] Schramm, O. Transboundary extremal length. J. Anal. Math. 66 (1995), 307–329.
[153] Semmes, S. On the nonexistence of bi-Lipschitz parameterizations and geometric
problems about A∞ -weights. Rev. Mat. Iberoamericana 12 (1996), 337–410.
[154] Semmes, S. Metric spaces and mappings seen at many scales (appendix). In Metric
structures for Riemannian and non-Riemannian spaces, by M. Gromov. Birkhäuser,
Boston, MA, 1999. Based on the 1981 French original, Translated from the French
by Sean Michael Bates.
[155] Shanmugalingam, N. Newtonian spaces: an extension of Sobolev spaces to metric
measure spaces. Rev. Mat. Iberoamericana 16, 2 (2000), 243–279.
[156] Smirnov, S. Dimension of quasicircles. Acta Math. to appear.
[157] Solomyak, B. “Mandelbrot set” for a pair of linear maps: the local geometry. Anal.
Theory Appl. 20, 2 (2004), 149–157.
[158] Solomyak, B. On the ‘Mandelbrot set’ for pairs of linear maps: asymptotic self-
similarity. Nonlinearity 18, 5 (2005), 1927–1943.
[159] Stallings, J. R. On torsion-free groups with infinitely many ends. Ann. of Math.
(2) 88 (1968), 312–334.
[160] Staples, S. G., and Ward, L. Quasisymmetrically thick sets. Ann. Acad. Sci.
Fenn. Ser. A I Math. 23 (1998), 151–168.
[161] Stein, E. M. Harmonic analysis: real-variable methods, orthogonality, and oscilla-
tory integrals. Princeton Univ. Press, Princeton, NJ, 1993. With the assistance of
Timothy S. Murphy, Monographs in Harmonic Analysis, III.
[162] Tukia, P. A quasiconformal group not isomorphic to a Möbius group. Ann. Acad.
Sci. Fenn. Ser. A I Math. 6 (1981), 149–160.
138 BIBLIOGRAPHY

[163] Tukia, P. Homeomorphic conjugates of Fuchsian groups. J. Reine Angew. Math.

391 (1988), 1–54.
[164] Tukia, P. Homeomorphic conjugates of Fuchsian groups: an outline. In Complex
analysis, Joensuu 1987, vol. 1351 of Lecture Notes in Math. Springer, Berlin, 1988,
pp. 344–353.
[165] Tukia, P. Hausdorff dimension and quasisymmetric mappings. Math. Scand. 65
(1989), 152–160.
[166] Tukia, P., and Väisälä, J. Quasisymmetric embeddings of metric spaces. Ann.
Acad. Sci. Fenn. Ser. A I Math. 5 (1980), 97–114.
[167] Tyson, J. T. Quasiconformality and quasisymmetry in metric measure spaces. Ann.
Acad. Sci. Fenn. Ser. A I Math. 23 (1998), 525–548.
[168] Tyson, J. T. Geometric and Analytic Applications of a Generalized Definition of
the Conformal Modulus. PhD thesis, University of Michigan, 1999.
[169] Tyson, J. T. Sets of minimal Hausdorff dimension for quasiconformal maps. Proc.
Amer. Math. Soc. 128, 11 (2000), 3361–3367.
[170] Tyson, J. T. Lowering the Assouad dimension by quasisymmetric mappings. Illinois
J. Math. 45 (2001), 641–656.
[171] Tyson, J. T. Global conformal Assouad dimension in the Heisenberg group. Con-
form. Geom. Dyn. 12 (2008), 32–57.
[172] Tyson, J. T., and Wu, J.-M. Characterizations of snowflake metric spaces. Ann.
Acad. Sci. Fenn. Ser. A I Math. 30 (2005), 313–336.
[173] Tyson, J. T., and Wu, J.-M. Quasiconformal dimensions of self-similar fractals.
Rev. Mat. Iberoamericana 22 (2006), 205–258.
[174] Väisälä, J. On quasiconformal mappings in space. Ann. Acad. Sci. Fenn. Ser. A I
Math. 298 (1961), 1–36.
[175] Väisälä, J. Lectures on n-dimensional quasiconformal mappings. No. 229 in Lecture
Notes in Mathematics. Springer-Verlag, Berlin, 1971.
[176] Warhurst, B. Contact and quasiconformal mappings on real model filiform groups.
Bull. Austral. Math. Soc. 68, 2 (2003), 329–343.
[177] Warhurst, B. Jet spaces as nonrigid Carnot groups. J. Lie Theory 15, 1 (2005),
341–356.
[178] Warhurst, B. Tanaka prolongation of free Lie algebras. Geom. Dedicata 130
(2007), 59–69.
[179] Wolf, J. A. Spaces of constant curvature. McGraw-Hill Book Co., New York, 1967.
[180] Wu, J.-M. Hausdorff dimension and doubling measures on metric spaces. Proc.
Amer. Math. Soc. 126 (1998), 1453–1459.
[181] Zajı́ček, L. Porosity and σ-porosity. Real Anal. Exchange 13, 2 (1987/88), 314–350.
 Index

(2, 2, 2, 2)-orbifold, 99 conformal dimension of the boundary

of, 44
A∞ weight, 19 Loewner property for, 56
absolutely minimizing Lipschitz branched covering map, xi, 98, 101
extension, 18 associated orbifold for, 98
accretive mapping, 123 expanding dynamics for, 99
accumulated distortion, 84 Thurston obstruction for, 100
admissible function, 51, 61
Cannon’s conjecture, 43, 89, 92, 104
affine transformation, 117
canonical boundary conformal gauge, 29
Ahlfors regular conformal dimension, x,
Cantor set, 12, 13, 59, 67
24, 91–93
middle interval, 108
Ahlfors regular measure, 13
quasisymmetric classification of, 21
Ahlfors regularity, x, 13, 106
Carathéodory construction, 9, 87
analysis in metric spaces, ix, x, 26
Carathéodory content, 10
annular linear connectivity, 62
Carathéodory criterion, 10
approximating sequence of uniform
Carathéodory measure, 10
quasipackings, 102
cardinality, 1, 10
associated orbifold, 98
Carnot group, x, 15, 17, 35
Assouad dimension, x, 14, 15
homogeneous dimension of, 36
and porosity, 68
Loewner property for, 56
Assouad’s embedding theorem, 15, 19,
Carnot-Carathéodory metric, ix, 15, 34,
46
36, 104
Aubin–Lions compactness theorem, 124
CAT(κ) space, 26, 28, 39
canonical metric on the boundary of,
Banach–Mazur theorem, 123 29
Bedford–McMullen carpet, 117, 128 examples of, 32
conformal dimensions of, 121 Cayley graph, 40
dimensions of, 119 Cayley hyperbolic plane, 31
uniform vertical fibers for, 119 Cayley transform, 32, 34, 57, 58
Besov norm, 96 chain, 9
Besov space, 105 Chebyshev polynommial, 44
Beurling–Ahlfors extension theorem, 24, Cheeger’s differentiation theorem, 17
114 chordal metric, 58
bi-Hölder homeomorphism, 15 Chow’s theorem, 33, 36
bi-Lipschitz map, x, 3, 15 coarse embedding, 19
binomial distribution, 60 cocompactness on triples, 41
boundary congruence, 112 cohomology group, 93
bounded turning, 24, 115 combinatorial modulus, 102, 105
Bourdon–Pajot space, 17, 43, 47, 106 comparison triangle, 26

139
140 INDEX

complete bipartite graph, 43 topological, 12

completely nonintegrable subbundle, 33 discrete modulus, 102, 105
complex analysis, 65 distance between sets, 1
complex hyperbolic space, 31, 45, 104 distance function, 2
conformal Assouad dimension, 23, 79 distortion function, 5
and products, 74, 75 division algebra, 31
and unions, 72 doubling measure, 49, 91, 104
minimality for, 82 doubling metric space, 8, 15, 21, 62, 91,
conformal Assouad dimension zero 93, 101
characterization of, 70 dyadic cube, 9, 11, 54
conformal dimension, ix–xi, 1, 22, 30, dyadic decomposition, 83
115 dyadic Hausdorff measure, 11
and volume entropy on negatively
curved manifolds, 38 egg yolk principle, 18
classification of Gromov hyperbolic elementary Gromov hyperbolic group,
groups with conformal dimension 40
one, 42 entropy, 116, 120
classification of Gromov hyperbolic Euclidean metric, 1
groups with conformal dimension Euclidean space, 1
zero, 41 Euler characteristic, 98
disambiguation, xi expanding dynamics, 99, 101, 115
minimality for, 22 f -stable multicurve, 100
of Bedford–McMullen carpets, 121 finitely generated group
of subsets of the Heisenberg group, Cayley graph of, 40
129 Gromov hyperbolicity of, 40
of the Sierpiński carpet, 59, 127 fractal, 127
of the boundary of a Gromov Frostman’s lemma, 12
hyperbolic space, 30 Fuchsian buildings, 43
zero, 67 Fuglede modulus, 63, 66, 122
conformal gauge, x, 1, 21, 30, 74, 91, Fuglede’s lemma, 97
93, 101, 103, 106
critical p cohomology exponent of, 93 gasket type iterated function system,
contraction ratio, 109 112
contractive, 117 gauge function, 9
contractive similarity, 109 Gehring’s higher integrability lemma, 6,
counting measure, 10 18, 125
covering theorem, 49 geodesic triangle, 25
critical p cohomology exponent, 93, 106 geometric function theory, ix, 30
cross ratio, 7 geometric group theory, 26, 30
CW complex, 106 global quasiconformal dimension, xi, 68,
Cygan metric, 32, 34 107, 115, 122
of n-dimensional Sierpiński gaskets,
David–Semmes deformation theory, 91, 114
104 of Julia sets of rational maps, 115
diameter of a set, 1 of middle interval Cantor sets, 109
dilation of middle space Sierpiński carpets,
on a Carnot group, 36, 56 127
dimension of subsets of the Heisenberg group,
Assouad, 14 129
Hausdorff, 11 of the Sierpiński gasket, 113
of a measure, 116 graph collection, 121
product formula for, 73 Gromov boundary, 27
small inductive, 12 topology on, 27
 INDEX 141

Gromov hyperbolic group, ix, x, 40, 46, infinity Laplacian, 18

106 invariant set, 110, 117
canonical boundary conformal gauge invertible iterated function system, 114,
of, 40 128
elementary, 40 isomorphism
non-elementary, 40 of iterated function systems, 111
Gromov hyperbolic metric space, x, 19, iterated function system, 13
26, 28, 93, 101 boundary of, 112
p cohomology groups of, 93 gasket-type, 112
canonical boundary conformal gauge invertibility of, 114
of, 29 isomorphism of, 111
examples of Bourdon–Pajot, 43 open set condition for, 110
quasi-isometric invariance of, 30 self-affine, 117
Gromov product, 28 self-similar, 109
Gromov–Hausdorff convergence, 80 symbolic representation map for, 110
Gromov–Hausdorff weak tangent, x Iwasawa groups, 37
Gromov–Pansu theorem, 37
Julia set, ix, 115, 128
HKn ,31 global quasiconformal dimension of,
sectional curvatures of, 32 115
Hölder’s inequality, 96
Hajlasz–Sobolev space, 97 Kapovich–Kleiner conjecture, 47
Kolmogorov consistency theorem, 116
harmonic analysis, 45
Korányi metric, 57
Hausdorff content, 10
Korányi norm, 57
Hausdorff convergence, 79
Hausdorff dimension, x, 11 2 Betti numbers, 106
Hausdorff measure, 10, 13 p cohomology, xi, 93, 105
Hausdorff metric, 110 Lagrange multipliers, 61
Heisenberg group, 35, 45, 104, 129 Lattès map, 99, 105
Hölder function, 4 Lebesgue Differentiation Theorem, 96
Hölder map, x, 4, 15 Lebesgue measure, 1
Hölder conjugate exponent, 50 Lebesgue–Bochner space, 124
Hölder’s inequality, 119 linear dilatation, 5
holomorphic tangent bundle, 33 Liouville’s theorem, 7, 18
homogeneous dimension, 36 Lipschitz function, 2
homogeneous measure, 65 operations on, 2
horizontal subspace, 36 Lipschitz map, x, 2, 15
horizontal tangent bundle, 33 local symmetry, 31
horizontal tangent space, 33 Loewner space, x, 55, 65, 68, 97,
horizontal vector field, 36 104–106
hyperbolic building, x, 24, 106 Lusin’s condition, 126
hyperbolic dimension, 47
hyperbolic metric space, 26 Main Equivalence Class property, 97
hyperbolic orbifold, 98 Mandelbrot, Benoit, 127
hyperbolic polygon, 43 Mass Distribution Principle, 12
hyperbolic space, 31 maximal stretching, 117
ball model for, 31 Mayer–Vietoris exact sequence, 106
Hausdorff dimension of the boundary McShane’s Extension Theorem, 2, 17
of, 34 measure
paraboloid model for, 31 Ahlfors regular, 13
quasi-isometric classification of, 34 dimension of, 116
sectional curvatures of, 32 doubling, 49
hyperspace, 110 MECp property, 97, 105
142 INDEX

Menger sponge, 42, 43, 90 post-critical finiteness, x, xi, 98, 101,

metric ball, 1, 9, 13 111, 114
metric measure space, 49 post-critical set, 98, 111
s-homogeneous, 65 probability vector, 116, 120
middle interval Cantor set, 108 entropy of, 116
global quasiconformal dimension of, projection map, 3
109 proper metric space, 26
middle square Sierpiński carpet, 126 pseudoconvexity, ix
middle square Sierpiński carpet
global quasiconformal dimension of, quasi-isometric rigidity, 46
127 quasi-isometry, ix, 30, 93
minimal stretching, 117 boundary quasisymmetric map for, 30
minimality for conformal dimension, 22 preservation of hyperbolicity under,
Möbius transformation, 25 30
modulus, 51, 59, 60 quasicircle, 24
examples of, 52 quasiconformal gradient mappings, 128
Fuglede, 63 quasiconformal Jacobian problem, 129
uniformly big, 82 quasiconformal map, ix, 6
Monge–Ampere equation, 128 quasimetric, 29
monotone mapping, 123, 128 snowflake of, 29
Moran–Hutchinson theorem, 110, 127 quasi-Möbius map, 8, 41
Mostow’s Rigidity Theorem, 31 quasiround set, 102
multicurve, 99 quasisymmetric accretivity, 123, 128
Thurston matrix for, 100 quasisymmetric classification
of S1 , 24
negatively curved manifolds of S2 , 93
volume entropy and conformal of the Cantor set, 21
dimension on, 38 quasisymmetric equivalence, 5
net, 94 quasisymmetric map, ix, x, 1, 5
non-elementary Gromov hyperbolic as the boundary map of a
group, 40 quasi-isometry, 30
octonionic Heisenberg group, 37 characterization via distortion of
quasisymmetric rigidity of, 46 metric balls, 7
octonionic hyperbolic plane, 45 operations on, 5
quasi-isometric rigidity of, 46 quasisymmetric thickness, 126
open set condition, 13, 110 quaternionic Heisenberg group, 35
orbifold, 98 quasisymmetric rigidity of, 46
Euler characteristic for, 98 quaternionic hyperbolic space, 31
quasi-isometric rigidity of, 46
Pansu, Pierre, ix, xi
parabolic orbifold, 98 Rademacher’s Differentiation Theorem,
peripheral curve, 100 2, 17
Perron–Frobenius theorem, 100 rank of a symmetric space, 31
pinched sectional curvature, 31, 37 rank one symmetric space, ix, x, 31
Poincaré inequality, 65 quasi-isometric classification of, 24
Poincaré disc, 25 rational map, ix, xi, 98, 115
Poincaré inequality, 104 post-critically finite, 98
polygasket, 127 real cross ratio, 7
porosity real hyperbolic plane, 25
and Assouad dimension, 68 real hyperbolic space, 31
and uniform disconnectedness, 69 refinement, 83
quasisymmetric invariance of, 68 reverse Hadamard inequality, 6
porous set, 67 Ricci curvature, 38
 INDEX 143

Rickman’s rug, 65, 104 Thurston–Douady–Hubbard theorem,

Riemann sphere, 97 99
Riemannian sectional curvature, 37 topological dimension, 12
Riemannian symmetric space, 31 totally disconnected, 67
rigidity, xi, 46, 47 tug-of-war
Rips condition, 45 and the infinity Laplacian, 18
rough hyperbolic distance, 83 Tukia’s example, 71
roughly onto map, 30
ultrametric, 70
uniform quasipacking, 102
s-homogeneous measure, 65, 92, 125 uniform vertical fibers, 119
s-quasisymmetry, 114 uniformly big modulus, 82
self-affine fractal, ix, xi, 66, 117 uniformly disconnected metric space, 9,
self-affine iterated function system, 117 13, 21
self-similar fractal, 110 uniformly perfect metric space, 8, 12,
self-similar iterated function system, 13, 21, 23, 40, 91, 93
109 uniformly quasi-Möbius family, 41
self-similar measure, 116 univalent function theory, 65
separated subset, 14
sesquilinear, 31 virtually Fuchsian group, 42
several complex variables, 45 visual metric, 27
shadow, 39 dependence on parameters, 27
Siegel upper half space, 32 visual parameter, 27
Sierpiński carpet, 42, 58, 82, 90 visual premetric, 28
conformal dimension of, 59, 127 Volume Comparison Theorem, 38
middle square, 126 volume derivative, 59
Sierpiński gasket, 111, 114 volume entropy, 37
n-dimensional, global quasiconformal and conformal dimension on
dimension of, 114 negatively curved manifolds, 38
global quasiconformal dimension of, von Koch snowflake curve, 4
113
weak tangent, 15, 64, 79, 92
similarity, 59, 109
word metric, 40
similarity dimension, 110, 116
singular values Zorn’s lemma, 94
of a matrix, 116
small inductive dimension, 12
snowflake curve, 4
snowflake dimension, 90, 127
minimality for, 90
snowflake equivalence, 27
snowflake map, 90
snowflake metric space, 4, 5, 15
sphere at infinity, x
spherical Hausdorff measures, 11
stereographic projection, 58
sub-Riemannian geometry, 15, 33, 45
sub-Riemannian metric, 33
surface area measure, 11
symbolic representation map, 110, 116
symmetric space, 31

Thurston matrix, 100

Thurston obstruction, 100
Titles in This Series
54 John M. Mackay and Jeremy T. Tyson, Conformal dimension: Theory and
application, 2010
53 John W. Morgan and Frederick Tsz-Ho Fong, Ricci flow and geometrization of
3-manifolds, 2010
52 Jan Nagel and Marian Aprodu, Koszul cohomology and algebraic geometry, 2010
51 J. Ben Hough, Manjunath Krishnapur, Yuval Peres, and Bálint Virág, Zeros of
Gaussian analytic functions and determinantal point processes, 2009
50 John T. Baldwin, Categoricity, 2009
49 József Beck, Inevitable randomness in discrete mathematics, 2009
48 Achill Schürmann, Computational geometry of positive definite quadratic forms, 2008
47 Ernst Kunz (with the assistance of and contributions by David A. Cox and
Alicia Dickenstein), Residues and duality for projective algebraic varieties, 2008
46 Lorenzo Sadun, Topology of tiling spaces, 2008
45 Matthew Baker, Brian Conrad, Samit Dasgupta, Kiran S. Kedlaya, and Jeremy
Teitelbaum (David Savitt and Dinesh S. Thakur, Editors), p-adic geometry:
Lectures from the 2007 Arizona Winter School, 2008
44 Vladimir Kanovei, Borel equivalence relations: structure and classification, 2008
43 Giuseppe Zampieri, Complex analysis and CR geometry, 2008
42 Holger Brenner, Jürgen Herzog, and Orlando Villamayor (Juan Elias, Teresa
Cortadellas Benı́tez, Gemma Colomé-Nin, and Santiago Zarzuela, Editors),
Three Lectures on Commutative Algebra, 2008
41 James Haglund, The q, t-Catalan numbers and the space of diagonal harmonics (with an
appendix on the combinatorics of Macdonald polynomials), 2008
40 Vladimir Pestov, Dynamics of infinite-dimensional groups. The Ramsey–Dvoretzky–
Milman phenomenon, 2006
39 Oscar Zariski, The moduli problem for plane branches (with an appendix by Bernard
Teissier), 2006
38 Lars V. Ahlfors, Lectures on Quasiconformal Mappings, Second Edition, 2006
37 Alexander Polishchuk and Leonid Positselski, Quadratic algebras, 2005
36 Matilde Marcolli, Arithmetic noncommutative geometry, 2005
35 Luca Capogna, Carlos E. Kenig, and Loredana Lanzani, Harmonic measure:
Geometric and analytic points of view, 2005
34 E. B. Dynkin, Superdiffusions and positive solutions of nonlinear partial differential
equations, 2004
33 Kristian Seip, Interpolation and sampling in spaces of analytic functions, 2004
32 Paul B. Larson, The stationary tower: Notes on a course by W. Hugh Woodin, 2004
31 John Roe, Lectures on coarse geometry, 2003
30 Anatole Katok, Combinatorial constructions in ergodic theory and dynamics, 2003
29 Thomas H. Wolff (Izabella L
 aba and Carol Shubin, editors), Lectures on harmonic
analysis, 2003
28 Skip Garibaldi, Alexander Merkurjev, and Jean-Pierre Serre, Cohomological
invariants in Galois cohomology, 2003
27 Sun-Yung A. Chang, Paul C. Yang, Karsten Grove, and Jon G. Wolfson,
Conformal, Riemannian and Lagrangian geometry, The 2000 Barrett Lectures, 2002
26 Susumu Ariki, Representations of quantum algebras and combinatorics of Young
tableaux, 2002
25 William T. Ross and Harold S. Shapiro, Generalized analytic continuation, 2002
24 Victor M. Buchstaber and Taras E. Panov, Torus actions and their applications in
topology and combinatorics, 2002
23 Luis Barreira and Yakov B. Pesin, Lyapunov exponents and smooth ergodic theory,
2002
 TITLES IN THIS SERIES

22 Yves Meyer, Oscillating patterns in image processing and nonlinear evolution equations,
2001
21 Bojko Bakalov and Alexander Kirillov, Jr., Lectures on tensor categories and
modular functors, 2001
20 Alison M. Etheridge, An introduction to superprocesses, 2000
19 R. A. Minlos, Introduction to mathematical statistical physics, 2000
18 Hiraku Nakajima, Lectures on Hilbert schemes of points on surfaces, 1999
17 Marcel Berger, Riemannian geometry during the second half of the twentieth century,
2000
16 Harish-Chandra, Admissible invariant distributions on reductive p-adic groups (with
notes by Stephen DeBacker and Paul J. Sally, Jr.), 1999
15 Andrew Mathas, Iwahori-Hecke algebras and Schur algebras of the symmetric group, 1999
14 Lars Kadison, New examples of Frobenius extensions, 1999
13 Yakov M. Eliashberg and William P. Thurston, Confoliations, 1998
12 I. G. Macdonald, Symmetric functions and orthogonal polynomials, 1998
11 Lars Gårding, Some points of analysis and their history, 1997
10 Victor Kac, Vertex algebras for beginners, Second Edition, 1998
9 Stephen Gelbart, Lectures on the Arthur-Selberg trace formula, 1996
8 Bernd Sturmfels, Gröbner bases and convex polytopes, 1996
7 Andy R. Magid, Lectures on differential Galois theory, 1994
6 Dusa McDuff and Dietmar Salamon, J-holomorphic curves and quantum cohomology,
1994
5 V. I. Arnold, Topological invariants of plane curves and caustics, 1994
4 David M. Goldschmidt, Group characters, symmetric functions, and the Hecke algebra,
1993
3 A. N. Varchenko and P. I. Etingof, Why the boundary of a round drop becomes a
curve of order four, 1992
2 Fritz John, Nonlinear wave equations, formation of singularities, 1990
1 Michael H. Freedman and Feng Luo, Selected applications of geometry to
low-dimensional topology, 1989
Conformal dimension measures the extent to which the Hausdorff dimension of a metric
space can be lowered by quasisymmetric deformations. Introduced by Pansu in 1989, this
concept has proved extremely fruitful in a diverse range of areas, including geometric func-
tion theory, conformal dynamics, and geometric group theory.
This survey leads the reader from the definitions and basic theory through to active research
applications in geometric function theory, Gromov hyperbolic geometry, and the dynamics
of rational maps, amongst other areas. It reviews the theory of dimension in metric
spaces and of deformations of metric spaces. It summarizes the basic tools for estimating
conformal dimension and illustrates their application to concrete problems of independent
interest. Numerous examples and proofs are provided.
Working from basic definitions through to current research areas, this book can be used
as a guide for graduate students interested in this field, or as a helpful survey for experts.
Background needed for a potential reader of the book consists of a working knowledge of
real and complex analysis on the level of first- and second-year graduate courses.

For additional information

and updates on this book, visit
www.ams.org/bookpages/ulect-54

AMS on the Web

ULECT/54 www.ams.org