Introduction to

Dynamical
Systems

Michael Brin
and Garrett Stuck
Introduction to Dynamical Systems

This book provides a broad introduction to the subject of dynamical systems,

suitable for a one- or two-semester graduate course. In the first chapter, the au-
thors introduce over a dozen examples, and then use these examples throughout
the book to motivate and clarify the development of the theory. Topics include
topological dynamics, symbolic dynamics, ergodic theory, hyperbolic dynamics,
one-dimensional dynamics, complex dynamics, and measure-theoretic entropy.
The authors top off the presentation with some beautiful and remarkable appli-
cations of dynamical systems to such areas as number theory, data storage, and
Internet search engines.
This book grew out of lecture notes from the graduate dynamical systems
course at the University of Maryland, College Park, and reflects not only the
tastes of the authors, but also to some extent the collective opinion of the Dy-
namics Group at the University of Maryland, which includes experts in virtually
every major area of dynamical systems.

Michael Brin is a professor of mathematics at the University of Maryland. He

is the author of over thirty papers, three of which appeared in the Annals of
Mathematics. Professor Brin is also an editor of Forum Mathematicum.

Garrett Stuck is a former professor of mathematics at the University of

Maryland and currently works in the structured finance industry. He has coau-
thored several textbooks, including The Mathematica Primer (Cambridge
University Press, 1998). Dr. Stuck is also a founder of Chalkfree, Inc., a soft-
ware company.
Introduction to Dynamical
Systems

MICHAEL BRIN
University of Maryland

GARRETT STUCK
University of Maryland
PUBLISHED BY CAMBRIDGE UNIVERSITY PRESS (VIRTUAL PUBLISHING)
FOR AND ON BEHALF OF THE PRESS SYNDICATE OF THE UNIVERSITY OF
CAMBRIDGE
The Pitt Building, Trumpington Street, Cambridge CB2 IRP
40 West 20th Street, New York, NY 10011-4211, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia

http://www.cambridge.org

© Michael Brin, Garrett Stuck 2002

This edition © Michael Brin, Garrett Stuck 2003

First published in printed format 2002

A catalogue record for the original printed book is available

from the British Library and from the Library of Congress
Original ISBN 0 521 80841 3 hardback

ISBN 0 511 02072 4 virtual (netLibrary Edition)

To Eugenia, Pamela, Sergey, Sam, Jonathan, and Catherine
for their patience and support.
Contents

Introduction page xi

1 Examples and Basic Concepts 1

1.1 The Notion of a Dynamical System 1
1.2 Circle Rotations 3
1.3 Expanding Endomorphisms of the Circle 5
1.4 Shifts and Subshifts 7
1.5 Quadratic Maps 9
1.6 The Gauss Transformation 11
1.7 Hyperbolic Toral Automorphisms 13
1.8 The Horseshoe 15
1.9 The Solenoid 17
1.10 Flows and Differential Equations 19
1.11 Suspension and Cross-Section 21
1.12 Chaos and Lyapunov Exponents 23
1.13 Attractors 25

2 Topological Dynamics 28
2.1 Limit Sets and Recurrence 28
2.2 Topological Transitivity 31
2.3 Topological Mixing 33
2.4 Expansiveness 35
2.5 Topological Entropy 36
2.6 Topological Entropy for Some Examples 41
2.7 Equicontinuity, Distality, and Proximality 45
2.8 Applications of Topological Recurrence to
Ramsey Theory 48

vii
viii Contents

3 Symbolic Dynamics 54
3.1 Subshifts and Codes 55
3.2 Subshifts of Finite Type 56
3.3 The Perron–Frobenius Theorem 57
3.4 Topological Entropy and the Zeta Function of an SFT 60
3.5 Strong Shift Equivalence and Shift Equivalence 62
3.6 Substitutions 64
3.7 Sofic Shifts 66
3.8 Data Storage 67

4 Ergodic Theory 69
4.1 Measure-Theory Preliminaries 69
4.2 Recurrence 71
4.3 Ergodicity and Mixing 73
4.4 Examples 77
4.5 Ergodic Theorems 80
4.6 Invariant Measures for Continuous Maps 85
4.7 Unique Ergodicity and Weyl’s Theorem 87
4.8 The Gauss Transformation Revisited 90
4.9 Discrete Spectrum 94
4.10 Weak Mixing 97
4.11 Applications of Measure-Theoretic Recurrence
to Number Theory 101
4.12 Internet Search 103

5 Hyperbolic Dynamics 106

5.1 Expanding Endomorphisms Revisited 107
5.2 Hyperbolic Sets 108
5.3 -Orbits 110
5.4 Invariant Cones 114
5.5 Stability of Hyperbolic Sets 117
5.6 Stable and Unstable Manifolds 118
5.7 Inclination Lemma 122
5.8 Horseshoes and Transverse Homoclinic Points 124
5.9 Local Product Structure and Locally Maximal
Hyperbolic Sets 128
5.10 Anosov Diffeomorphisms 130
5.11 Axiom A and Structural Stability 133
5.12 Markov Partitions 134
5.13 Appendix: Differentiable Manifolds 137
Contents ix

6 Ergodicity of Anosov Diffeomorphisms 141

6.1 Hölder Continuity of the Stable and Unstable
Distributions 141
6.2 Absolute Continuity of the Stable and Unstable
Foliations 144
6.3 Proof of Ergodicity 151

7 Low-Dimensional Dynamics 153

7.1 Circle Homeomorphisms 153
7.2 Circle Diffeomorphisms 160
7.3 The Sharkovsky Theorem 162
7.4 Combinatorial Theory of Piecewise-Monotone
Mappings 170
7.5 The Schwarzian Derivative 178
7.6 Real Quadratic Maps 181
7.7 Bifurcations of Periodic Points 183
7.8 The Feigenbaum Phenomenon 189

8 Complex Dynamics 191

8.1 Complex Analysis on the Riemann Sphere 191
8.2 Examples 194
8.3 Normal Families 197
8.4 Periodic Points 198
8.5 The Julia Set 200
8.6 The Mandelbrot Set 205

9 Measure-Theoretic Entropy 208

9.1 Entropy of a Partition 208
9.2 Conditional Entropy 211
9.3 Entropy of a Measure-Preserving Transformation 213
9.4 Examples of Entropy Calculation 218
9.5 Variational Principle 221

Bibliography 225
Index 231
Introduction

The purpose of this book is to provide a broad and general introduction

to the subject of dynamical systems, suitable for a one- or two-semester
graduate course. We introduce the principal themes of dynamical systems
both through examples and by explaining and proving fundamental and
accessible results. We make no attempt to be exhaustive in our treatment of
any particular area.
This book grew out of lecture notes from the graduate dynamical systems
course at the University of Maryland, College Park. The choice of topics
reflects not only the tastes of the authors, but also to a large extent the
collective opinion of the Dynamics Group at the University of Maryland,
which includes experts in virtually every major area of dynamical systems.
Early versions of this book have been used by several instructors at
Maryland, the University of Bonn, and Pennsylvania State University. Expe-
rience shows that with minor omissions the first five chapters of the book can
be covered in a one-semester course. Instructors who wish to cover a differ-
ent set of topics may safely omit some of the sections at the end of Chapter 1,
§§2.7–§2.8, §§3.5–3.8, and §§4.8–4.12, and then choose from topics in later
chapters. Examples from Chapter 1 are used throughout the book. Chapter 6
depends on Chapter 5, but the other chapters are essentially independent.
Every section ends with exercises (starred exercises are the most difficult).
The exposition of most of the concepts and results in this book has been
refined over the years by various authors. Since most of these ideas have
appeared so often and in so many variants in the literature, we have not at-
tempted to identify the original sources. In many cases, we followed the writ-
ten exposition from specific sources listed in the bibliography. These sources
cover particular topics in greater depth than we do here, and we recommend
them for further reading. We also benefited from the advice and guidance
of a number of specialists, including Joe Auslander, Werner Ballmann,
xi
xii Introduction

Ken Berg, Mike Boyle, Boris Hasselblatt, Michael Jakobson, Anatole Katok,
Michal Misiurewicz, and Dan Rudolph. We thank them for their contribu-
tions. We are especially grateful to Vitaly Bergelson for his contributions to
the treatment of applications of topological dynamics and ergodic theory to
combinatorial number theory. We thank the students who used early ver-
sions of this book in our classes, and who found many typos, errors, and
omissions.
CHAPTER ONE

Examples and Basic Concepts

Dynamical systems is the study of the long-term behavior of evolving

systems. The modern theory of dynamical systems originated at the end
of the 19th century with fundamental questions concerning the stability and
evolution of the solar system. Attempts to answer those questions led to
the development of a rich and powerful field with applications to physics,
biology, meteorology, astronomy, economics, and other areas.
By analogy with celestial mechanics, the evolution of a particular state of
a dynamical system is referred to as an orbit. A number of themes appear
repeatedly in the study of dynamical systems: properties of individual orbits;
periodic orbits; typical behavior of orbits; statistical properties of orbits;
randomness vs. determinism; entropy; chaotic behavior; and stability under
perturbation of individual orbits and patterns. We introduce some of these
themes through the examples in this chapter.
We use the following notation throughout the book: N is the set of
positive integers; N0 = N ∪ {0}; Z is the set of integers; Q is the set of rational
numbers; R is the set of real numbers; C is the set of complex numbers; R+
is the set of positive real numbers; R+ +
0 = R ∪ {0}.

1.1 The Notion of a Dynamical System

A discrete-time dynamical system consists of a non-empty set X and a map
f : X → X. For n ∈ N, the nth iterate of f is the n-fold composition f n =
f ◦ · · · ◦ f ; we define f 0 to be the identity map, denoted Id. If f is invertible,
then f −n = f −1 ◦ · · · ◦ f −1 (n times). Since f n+m = f n ◦ f m, these iterates
form a group if f is invertible, and a semigroup otherwise.
Although we have defined dynamical systems in a completely abstract
setting, where X is simply a set, in practice X usually has additional structure

1
2 1. Examples and Basic Concepts

that is preserved by the map f . For example, (X, f ) could be a measure space
and a measure-preserving map; a topological space and a continuous map;
a metric space and an isometry; or a smooth manifold and a differentiable
map.
A continuous-time dynamical system consists of a space X and a one-
parameter family of maps of { f t : X → X}, t ∈ R or t ∈ R+ 0 , that forms a one-
parameter group or semigroup, i.e., f t+s = f t ◦ f s and f 0 = Id. The dynam-
ical system is called a flow if the time t ranges over R, and a semiflow if t
ranges over R+ t
0 . For a flow, the time-t map f is invertible, since f
−t
= ( f t )−1 .
Note that for a fixed t0 , the iterates ( f ) = f form a discrete-time dynam-
t0 n t0 n

ical system.
We will use the term dynamical system to refer to either discrete-time
or continuous-time dynamical systems. Most concepts and results in dy-
namical systems have both discrete-time and continuous-time versions. The
continuous-time version can often be deduced from the discrete-time ver-
sion. In this book, we focus mainly on discrete-time dynamical systems, where
the results are usually easier to formulate and prove.
To avoid having to define basic terminology in four different cases, we
write the elements of a dynamical system as f t , where t ranges over Z, N0 , R,
or R+ , as appropriate. For x ∈ X, we define the positive semiorbit O+f (x) =

0 t
f (x). In the invertible case, we define the negative semiorbit O−f (x) =

t≥0 t + −
t
t≤0 f (x), and the orbit O f (x) = O f (x) ∪ O f (x) = t f (x) (we omit the
subscript “ f ” if the context is clear). A point x ∈ X is a periodic point of
period T > 0 if f T (x) = x. The orbit of a periodic point is called a periodic
orbit. If f t (x) = x for all t, then x is a fixed point. If x is periodic, but not
fixed, then the smallest positive T, such that f T (x) = x, is called the minimal
period of x. If f s (x) is periodic for some s > 0, we say that x is eventu-
ally periodic. In invertible dynamical systems, eventually periodic points are
periodic.
For a subset A ⊂ X and t > 0, let f t (A) be the image of Aunder f t , and let
f (A) be the preimage under f t , i.e., f −t (A) = ( f t )−1 (A) = {x ∈ X: f t (x) ∈
−t

A}. Note that f −t ( f t (A)) contains A, but, for a non-invertible dynamical

system, is generally not equal to A. A subset A⊂X is f -invariant if f t (A) ⊂A
for all t; forward f -invariant if f t (A) ⊂ A for all t ≥ 0; and backward
f -invariant if f −t (A) ⊂ A for all t ≥ 0.
In order to classify dynamical systems, we need a notion of equivalence.
Let f t : X → X and g t : Y → Y be dynamical systems. A semiconjugacy from
(Y, g) to (X, f ) (or, briefly, from g to f ) is a surjective map π : Y →X sat-
isfying f t ◦ π = π ◦ g t , for all t. We express this formula schematically by
1.2. Circle Rotations 3

saying that the following diagram commutes:

g
Y −→ Y
π| |
↓ ↓π
f
X −→ X
An invertible semiconjugacy is called a conjugacy. If there is a conju-
gacy from one dynamical system to another, the two systems are said to
be conjugate; conjugacy is an equivalence relation. To study a particular
dynamical system, we often look for a conjugacy or semiconjugacy with a
better-understood model. To classify dynamical systems, we study equiva-
lence classes determined by conjugacies preserving some specified structure.
Note that for some classes of dynamical systems (e.g., measure-preserving
transformations) the word isomorphism is used instead of “conjugacy.”
If there is a semiconjugacy π from g to f , then (X, f ) is a factor of (Y, g),
and (Y, g) is an extension of (X, f ). The map π: Y → X is also called a fac-
tor map or projection. The simplest example of an extension is the direct
product
( f1 × f2 )t : X1 × X2 → X1 × X2
of two dynamical systems fit : Xi →Xi , i = 1, 2, where ( f1 × f2 )t (x1 , x2 ) =
( f1t (x1 ), f2t (x2 )). Projection of X1 × X2 onto X1 or X2 is a semiconjugacy, so
(X1 , f1 ) and (X2 , f2 ) are factors of (X1 × X2 , f1 × f2 ).
An extension (Y, g) of (X, f ) with factor map π: Y →X is called a skew
product over (X, f ) if Y = X × F, and π is the projection onto the first factor
or, more generally, if Y is a fiber bundle over X with projection π.
Exercise 1.1.1. Show that the complement of a forward invariant set is
backward invariant, and vice versa. Show that if f is bijective, then an in-
variant set A satisfies f t (A) = A for all t. Show that this is false, in general,
if f is not bijective.
Exercise 1.1.2. Suppose (X, f ) is a factor of (Y, g) by a semiconjugacy
π : Y → X. Show that if y ∈ Y is a periodic point, then π (y) ∈ X is periodic.
Give an example to show that the preimage of a periodic point does not
necessarily contain a periodic point.

1.2 Circle Rotations

Consider the unit circle S1 = [0, 1] / ∼, where ∼ indicates that 0 and 1 are
identified. Addition mod 1 makes S1 an abelian group. The natural distance
4 1. Examples and Basic Concepts

on [0, 1] induces a distance on S1 ; specifically,

d(x, y) = min(|x − y|, 1 − |x − y|).

Lebesgue measure on [0, 1] gives a natural measure λ on S1 , also called

Lebesgue measure λ.
We can also describe the circle as the set S1 = {z ∈ C: |z| = 1}, with com-
plex multiplication as the group operation. The two notations are related by
z = e2πi x , which is an isometry if we divide arc length on the multiplicative
circle by 2π. We will generally use the additive notation for the circle.
For α ∈ R, let Rα be the rotation of S1 by angle 2π α, i.e.,

Rα x = x + α mod 1.

The collection {Rα : α ∈ [0, 1)} is a commutative group with composition as

group operation, Rα ◦ Rβ = Rγ , where γ = α + β mod 1. Note that Rα is an
isometry: It preserves the distance d. It also preserves Lebesgue measure λ,
i.e., the Lebesgue measure of a set is the same as the Lebesgue measure of
its preimage.
q
If α = p/q is rational, then Rα = Id, so every orbit is periodic. On the
other hand, if α is irrational, then every positive semiorbit is dense in S1 .
Indeed, the pigeon-hole principle implies that, for any
> 0, there are m, n <
1/
such that m < n and d(Rαm, Rαn ) <
. Thus Rn−m is rotation by an angle
less than
, so every positive semiorbit is
-dense in S1 (i.e., comes within
distance
of every point in S1 ). Since
is arbitrary, every positive semiorbit
is dense.
For α irrational, density of every orbit of Rα implies that S1 is the only
Rα -invariant closed non-empty subset. A dynamical system with no proper
closed non-empty invariant subsets is called minimal. In Chapter 4, we show
that any measurable Rα -invariant subset of S1 has either measure zero or
full measure. A measurable dynamical system with this property is called
ergodic.
Circle rotations are examples of an important class of dynamical systems
arising as group translations. Given a group G and an element h ∈ G, define
maps Lh : G →G and Rh : G →G by

Lh g = hg and Rh g = gh.

These maps are called left and right translation by h. If G is commutative,

Lh = Rh .
A topological group is a topological space G with a group structure
such that group multiplication (g, h) →gh, and the inverse g →g −1 are
1.3. Expanding Endomorphisms of the Circle 5

continuous maps. A continuous homomorphism of a topological group to

itself is called an endomorphism; an invertible endomorphism is an automor-
phism. Many important examples of dynamical systems arise as translations
or endomorphisms of topological groups.
Exercise 1.2.1. Show that for any k ∈ Z, there is a continuous semiconju-
gacy from Rα to Rkα .
Exercise 1.2.2. Prove that for any finite sequence of decimal digits there is
an integer n > 0 such that the decimal representation of 2n starts with that
sequence of digits.
Exercise 1.2.3. Let G be a topological group. Prove that for each g ∈ G, the
closure H(g) of the set {g n }∞
n=−∞ is a commutative subgroup of G. Thus, if
G has a minimal left translation, then G is abelian.
*Exercise 1.2.4. Show that Rα and Rβ are conjugate by a homeomorphism
if and only if α = ±β mod 1.

1.3 Expanding Endomorphisms of the Circle

For m ∈ Z, |m| > 1, define the times-m map Em: S1 →S1 by
Em x = mx mod 1.
This map is a non-invertible group endomorphism of S1 . Every point has
m preimages. In contrast to a circle rotation, Em expands arc length and
distances between nearby points by a factor of m: If d(x, y) ≤ 1/(2m), then
d(Em x, Em y) = md(x, y). A map (of a metric space) that expands distances
between nearby points by a factor of at least µ > 1 is called expanding.
The map Em preserves Lebesgue measure λ on S1 in the following sense:
−1
if A ⊂ S1 is measurable, then λ(Em (A)) = λ(A) (Exercise 1.3.1). Note, how-
ever, that for a sufficiently small interval I, λ(Em(I)) = mλ(I). We will show
later that Em is ergodic (Proposition 4.4.2).
Fix a positive integer m > 1. We will now construct a semiconjugacy from
another natural dynamical system to Em. Let = {0, . . . , m − 1}N be the set
of sequences of elements in {0, . . . , m − 1}. The shift σ : → discards the
first element of a sequence and shifts the remaining elements one place to
the left:
σ ((x1 , x2 , x3 , . . .)) = (x2 , x3 , x4 , . . .).
A base-m expansion of x ∈ [0, 1] is a sequence (xi )i∈N ∈ such that
∞
x = i=1 xi /mi . In analogy with decimal notation, we write x = 0.x1 x2 x3 . . . .
6 1. Examples and Basic Concepts

Base-m expansions are not always unique: A fraction whose denominator

is a power of m is represented both by a sequence with trailing m − 1s and
a sequence with trailing zeros. For example, in base 5, we have 0.144 . . . =
0.200 . . . = 2/5.
Define a map

∞
xi
φ: → [0, 1], φ((xi )i∈N ) = .
i=1
mi

We can consider φ as a map into S1 by identifying 0 and 1. This map is

surjective, and one-to-one except on the countable set of sequences with
trailing zeros or m − 1’s. If x = 0.x1 x2 x3 . . . ∈ [0, 1), then Em x = 0.x2 x3 . . . .
Thus, φ ◦ σ = Em ◦ φ, so φ is a semiconjugacy from σ to Em.
We can use the semiconjugacy of Em with the shift σ to deduce properties
of Em. For example, a sequence (xi ) ∈ is a periodic point for σ with period
k if and only if it is a periodic sequence with period k, i.e., xk+i = xi for all
i. It follows that the number of periodic points of σ of period k is mk. More
generally, (xi ) is eventually periodic for σ if and only if the sequence (xi ) is
eventually periodic. A point x ∈ S1 = [0, 1] / ∼ is periodic for Em with period
k if and only if x has a base-m expansion x = 0.x1 x2 . . . that is periodic with
period k. Therefore, the number of periodic points of Em of period k is mk − 1
(since 0 and 1 are identified).


Let Fm = ∞ k=1 {0, . . . , m − 1} be the set of all finite sequences of ele-
k

ments of the set {0, . . . , m − 1}. A subset A ⊂ [0, 1] is dense if and only if
every finite sequence w ∈ Fm occurs at the beginning of the base-m expan-
sion of some element of A. It follows that the set of periodic points is dense
in S1 . The orbit of a point x = 0.x1 x2 . . . is dense in S1 if and only if every
finite sequence from Fm appears in the sequence (xi ). Since Fm is countable,
we can construct such a point by concatenating all elements of Fm.
Although φ is not one-to-one, we can construct a right inverse to φ. Con-
sider the partition of S1 = [0, 1] / ∼ into intervals

Pk = [k/m, (k + 1)/m), 0 ≤ k ≤ m − 1.

For x ∈ [0, 1], define ψi (x) = k if Em i

x ∈ Pk. The map ψ: S1 → , given by
∞
x →(ψi (x))i=0 , is a right inverse for φ, i.e., φ ◦ ψ = Id: S1 →S1 . In particular,
x ∈ S1 is uniquely determined by the sequence (ψi (x)).
The use of partitions to code points by sequences is the principal motiva-
tion for symbolic dynamics, the study of shifts on sequence spaces, which is
the subject of the next section and Chapter 3.
1.4. Shifts and Subshifts 7

−1
Exercise 1.3.1. Prove that λ(Em ([a, b])) = λ([a, b]) for any interval
[a, b] ⊂ [0, 1].

Exercise 1.3.2. Prove that Ek ◦ El = El ◦ Ek = Ekl . When is Ek ◦ Rα =

Rα ◦ Ek?

Exercise 1.3.3. Show that the set of points with dense orbits is uncountable.

Exercise 1.3.4. Prove that the set

 
C = x ∈ [0, 1]: E3k x ∈
/ (1/3, 2/3) ∀ k ∈ N0

is the standard middle-thirds Cantor set.

*Exercise 1.3.5. Show that the set of points with dense orbits under Em has
Lebesgue measure 1.

1.4 Shifts and Subshifts

In this section, we generalize the notion of shift space introduced in the
previous section. For an integer m > 1 set Am = {1, . . . , m}. We refer to Am
as an alphabet and its elements as symbols. A finite sequence of symbols
is called a word. Let m = AZm be the set of infinite two-sided sequences of
+
symbols in Am, and m = ANm be the set of infinite one-sided sequences. We
say that a sequence x = (xi ) contains the word w = w1 w2 . . . wk (or that w
occurs in x) if there is some j such that wi = x j+i for i = 1, . . . , k.
Given a one-sided or two-sided sequence x = (xi ), let σ (x) = (σ (x)i ) be
the sequence obtained by shifting x one step to the left, i.e., σ (x)i = xi+1 .
+
This defines a self-map of both m and m called the shift. The pair ( m, σ ) is
+
called the full two-sided shift; ( m , σ ) is the full one-sided shift. The two-sided
shift is invertible. For a one-sided sequence, the leftmost symbol disappears,
so the one-sided shift is non-invertible, and every point has m preimages.
Both shifts have mn periodic points of period n.
+
The shift spaces m and m are compact topological spaces in the product
topology. This topology has a basis consisting of cylinders
,...,nk
C nj11,..., jk = {x = (xl ): xni = ji , i = 1, . . . , k},

where n1 < n2 < · · · < nk are indices in Z or N, and ji ∈ Am. Since the preim-
+
age of a cylinder is a cylinder, σ is continuous on m and is a homeomorphism
of m. The metric

d(x, x  ) = 2−l , where l = min{|i|: xi =xi }

8 1. Examples and Basic Concepts

1 1
1 2 1 1

1 1 0
1 0 1
1 2 3
0 0 1

Figure 1.1. Examples of directed graphs with labeled vertices and the corresponding
adjacency matrices.

+
generates the product topology on m and m (Exercise 1.4.3). In m,
the open ball B(x, 2 ) is the symmetric cylinder Cx−l,−l+1,...,l
−l
−l ,x−l+1 ,...,xl
, and in m +
,
−l + 
B(x, 2 ) = Cx1 ,...,xl . The shift is expanding on m ; if d(x, x ) < 1/2, then
1,...,l

d(σ (x), σ (x  )) = 2d(x, x  ).

In the product topology, periodic points are dense, and there are dense
orbits (Exercise 1.4.5).
Now we describe a natural class of closed shift-invariant subsets of the full
shift spaces. These subshifts can be described in terms of adjacency matrices
or their associated directed graphs. An adjacency matrix A = (ai j ) is an m ×
m matrix whose entries are zeros and ones. Associated to A is a directed
graph  A with m vertices such that ai j is the number of edges from the
ith vertex to the jth vertex. Conversely, if  is a finite directed graph with
vertices v1 , . . . , vm, then  determines an adjacency matrix B, and  =  B.
Figure 1.1 shows two adjacency matrices and the associated graphs.
Given an m × m adjacency matrix A = (ai j ), we say that a word or in-
finite sequence x (in the alphabet Am) is allowed if axi xi+1 > 0 for every i;
equivalently, if there is a directed edge from xi to xi+1 for every i. A word
or sequence that is not allowed is said to be forbidden. Let A ⊂ m be the
set of allowed two-sided sequences (xi ), and + +
A ⊂ m be the set of allowed
one-sided sequences. We can view a sequence (xi ) ∈ A (or + A ) as an infinite
walk along directed edges in the graph  A, where xi is the index of the vertex
visited at time i. The sets A and + A are closed shift-invariant subsets of m
and m +
, and inherit the subspace topology. The pairs ( A, σ ) and ( + A, σ )
are called the two-sided and one-sided vertex shifts determined by A.
A point (xi ) ∈ A (or + A ) is periodic of period n if and only if xi+n = xi
for every i. The number of periodic points of period n (in A or + A ) is equal
to the trace of An (Exercise 1.4.2).

Exercise 1.4.1. Let A be a matrix of zeros and ones. A vertex vi can be

reached (in n steps) from a vertex v j if there is a path (consisting of n edges)
from vi to v j along directed edges of  A. What properties of A correspond
to the following properties of  A?
1.5. Quadratic Maps 9

(a) Any vertex can be reached from some other vertex.

(b) There are no terminal vertices, i.e., there is at least one directed edge
starting at each vertex.
(c) Any vertex can be reached in one step from any other vertex .
(d) Any vertex can be reached from any other vertex in exactly n steps.

Exercise 1.4.2. Let Abe an m × m matrix of zeros and ones. Prove that:
(a) the number of fixed points in A (or +A ) is the trace of A;
(b) the number of allowed words of length n + 1 beginning with the sym-
bol i and ending with j is the i, jth entry of An ; and
(c) the number of periodic points of period n in A (or + A ) is the trace
of An .
+
Exercise 1.4.3. Verify that the metrics on m and m generate the product
topology.

Exercise 1.4.4. Show that the semiconjugacy φ: → [0, 1] of §1.3 is con-

tinuous with respect to the product topology on .

Exercise 1.4.5. Assume that all entries of some power of A are positive.
Show that in the product topology on A and +
A , periodic points are dense,
and there are dense orbits.

1.5 Quadratic Maps

The expanding maps of the circle introduced in §1.3 are linear maps in the
sense that they come from linear maps of the real line. The simplest non-
linear dynamical systems in dimension one are the quadratic maps

qµ (x) = µx(1 − x), µ > 0.

Figure 1.2 shows the graph of q3 and successive images xi = q3i (x0 ) of a point
x0 .
If µ > 1 and x ∈
/ [0, 1], then qµn (x) → −∞ as n → ∞. For this reason, we
focus our attention on the interval [0, 1]. For µ ∈ [0, 4], the interval
 [0, 1]
is forward
 invariant under qµ . For µ > 4, the interval (1/2 − 1/4 − 1/µ,
1/2 + 1/4 − 1/µ) maps outside [0, 1]; we show in Chapter 7 that the set of
points µ whose forward orbits stay in [0, 1] is a Cantor set, and (µ , qµ ) is
equivalent to the full one-sided shift on two symbols.
Let X be a locally compact metric space and f : X → X a continuous
map. A fixed point p of f is attracting if it has a neighborhood U such that Ū

is compact, f (Ū) ⊂ U, and n≥0 f n (U) = { p}. A fixed point p is repelling
10 1. Examples and Basic Concepts

0.8

0.6

0.4

0.2

2
x0 q3(x0) q 3(x0)

Figure 1.2. Quadratic map of q3 .


if it has a neighborhood U such that Ū ⊂ f (U), and n≥0 f −n (U) = { p}.
Note that if f is invertible, then p is attracting for f if and only if it is
repelling for f −1 , and vice versa. A fixed point p is called isolated if there is
a neighborhood of p that contains no other fixed points.
If x is a periodic point of f of period n, then we say that f is an attracting
(repelling) periodic point if x is an attracting (repelling) fixed point of f n . We
also say that the periodic orbit O(x) is attracting or repelling, respectively.
The fixed points of qµ are 0 and 1 − 1/µ. Note that qµ (0) = µ and that
qµ (1 − 1/µ) = 2 − µ. Thus, 0 is attracting for µ < 1 and repelling for µ > 1,
and 1 − 1/µ is attracting for µ ∈ (1, 3) and repelling for µ ∈ / [1, 3]
(Exercise 1.5.4).
The maps qµ , µ > 4, have interesting and complicated dynamical be-
havior. In particular, periodic points abound. For example,
qµ ([1/µ, 1/2]) ⊃ [1 − 1/µ, 1],
qµ ([1 − 1/µ, 1]) ⊃ [0, 1 − 1/µ] ⊃ [1/µ, 1/2].
Hence, qµ2 ([1/µ, 1/2]) ⊃ [1/µ, 1/2], so the Intermediate Value Theorem
implies that qµ2 has a fixed point p2 ∈ [1/µ, 1/2]. Thus, p2 and qµ ( p2 ) are
non-fixed periodic points of period 2. This approach to showing existence
of periodic points applies to many one-dimensional maps. We exploit this
technique in Chapter 7 to prove the Sharkovsky Theorem (Theorem 7.3.1),
which asserts, for example, that for continuous self-maps of the interval the
existence of an orbit of period three implies the existence of periodic orbits
of all orders.
/ [0, 1], qµn (x) → −∞ as n → ∞.
Exercise 1.5.1. Show that for any x ∈
Exercise 1.5.2. Show that a repelling fixed point is an isolated fixed point.
1.6. The Gauss Transformation 11

Exercise 1.5.3. Suppose p is an attracting fixed point for f . Show that there
is a neighborhood U of p such that the forward orbit of every point in U
converges to p.

Exercise 1.5.4. Let f : R → R be a C 1 map, and p be a fixed point. Show

that if | f  ( p)| < 1, then p is attracting, and if | f  ( p)| > 1, then p is repelling.

Exercise 1.5.5. Are 0 and 1 − 1/µ attracting or repelling for µ = 1? for

µ = 3?

Exercise 1.5.6. Show the existence of a non-fixed periodic point of qµ of

period 3, for µ > 4.

Exercise 1.5.7. Is the period-2 orbit { p2 , qµ ( p2 )} attracting or repelling for

µ > 4?

1.6 The Gauss Transformation

Let [x] denote the greatest integer less than or equal to x, for x ∈ R. The
map ϕ: [0, 1] → [0, 1] defined by

1/x − [1/x] if x ∈ (0, 1],
ϕ(x) =
0 if x = 0

was studied by C. Gauss, and is now called the Gauss transformation. Note
that ϕ maps each interval (1/(n + 1), 1/n] continuously and monotonically
onto [0, 1); it is discontinuous at 1/n for all n ∈ N. Figure 1.3 shows the graph
of ϕ.

0.8

0.6

0.4

0.2

1/4 1/3 1/2 1

Figure 1.3. Gauss transformation.
12 1. Examples and Basic Concepts

Gauss discovered a natural invariant measure µ for ϕ. The Gauss measure

of an interval A = (a, b) is
 b
1 dx 1+b
µ(A) = = (log 2)−1 log .
log 2 a 1 + x 1+a
This measure is ϕ-invariant in the sense that µ(ϕ −1 (A)) = µ(A) for any inter-
val A = (a, b). To prove invariance, note that the preimage of (a, b) consists
of infinitely many intervals: In the interval (1/(n + 1), 1/n), the preimage is
(1/(n + b), 1/(n + a)). Thus,
 

∞
1 1
µ(ϕ −1 ((a, b))) = µ ,
n=1
n+b n+a
 
1  ∞
n+a+1 n+b
= log · = µ((a, b)).
log 2 n=1 n+a n+b+1

Note that in general µ(ϕ(A)) =µ(A).

The Gauss transformation is closely related to continued fractions. The
expression
1
[a1 , a2 , . . . , an ] = , a1 , . . . , an ∈ N,
1
a1 +
1
a2 + · · ·
an
is called a finite continued fraction. For x ∈ (0, 1], we have x = 1/([ x1 ] + ϕ(x)).
More generally, if ϕ n−1 (x) =0, set ai = [1/ϕ i−1 (x)] ≥ 1 for i ≤ n. Then,
1
x=
1
a1 +
1
a2 +
1
··· +
an + ϕ n (x)
Note that x is rational if and only if ϕ m(x) = 0 for some m ∈ N (Exercise 1.6.2).
Thus any rational number is uniquely represented by a finite continued frac-
tion.
For an irrational number x ∈ (0, 1), the sequence of finite continued
fractions
1
[a1 , a2 , . . . , an ] =
1
a1 +
1
a2 +
1
··· +
an
1.7. Hyperbolic Toral Automorphisms 13

converges to x (where ai = [1/ϕ i−1 (x)]) (Exercise 1.6.4). This is expressed

concisely with the infinite continued fraction notation
1
x = [a1 , a2 , . . .] = .
1
a1 +
a2 + · · ·
Conversely, given a sequence (bi )i∈N , bi ∈ N, the sequence [b1 , b2 , . . . , bn ]
converges, as n → ∞, to a number y ∈ [0, 1], and the representation y =
[b1 , b2 , . . .] is unique (Exercise 1.6.4). Hence ϕ(y) = [b2 , b3 , . . .], because
bn = [1/ϕ n−1 (y)].
We summarize this discussion by saying that the continued fraction rep-
resentation conjugates the Gauss transformation and the shift on the space
ω
of finite or infinite integer-valued sequences (bi )i=1 , ω ∈ N ∪ {∞}, bi ∈ N. (By
convention, the shift of a finite sequence is obtained by deleting the first term;
the empty sequence represents 0.) As an immediate consequence, we obtain
a description of the eventually periodic points of ϕ (see Exercise 1.6.3).
Exercise 1.6.1. What are the fixed points of the Gauss transformation?
Exercise 1.6.2. Show that x ∈ [0, 1] is rational if and only if ϕ m(x) = 0 for
some m ∈ N.
Exercise 1.6.3. Show that:
(a) a number with periodic continued fraction expansion satisfies a
quadratic equation with integer coefficients; and
(b) a number with eventually periodic continued fraction expansion
satisfies a quadratic equation with integer coefficients.
The converse of the second statement is also true, but is more difficult to
prove [Arc70], [HW79].
*Exercise 1.6.4. Show that given any infinite sequence bk ∈ N, k = 1, 2, . . . ,
the sequence [b1 , . . . , bn ] of finite continued fractions converges. Show that
for any x ∈ R, the continued fraction [a1 , a2 , . . .], ai = [1/φ i−1 (x)], converges
to x, and that this continued fraction representation is unique.

1.7 Hyperbolic Toral Automorphisms

Consider the linear map of R2 given by the matrix
 
2 1
A= .
1 1
√
The eigenvalues are λ = (3 + 5)/2 > 1 and 1/λ. The√map expands by a fac-
tor of λ in the direction of the eigenvector vλ = ((1 + 5)/2, 1), and contracts
14 1. Examples and Basic Concepts

(1, 1)
(2, 1)

(0, 0)
(3, 0)

Figure 1.4. The image of the torus under A.

√
by 1/λ in the direction of v1/λ = ((1 − 5)/2, 1). The eigenvectors are per-
pendicular because A is symmetric.
Since A has integer entries, it preserves the integer lattice Z2 ⊂ R2 and
induces a map (which we also call A) of the torus T2 = R2 /Z2 . The torus
can be viewed as the unit square [0, 1] × [0, 1] with opposite sides identified:
(x1 , 0) ∼ (x1 , 1) and (0, x2 ) ∼ (1, x2 ), x1 , x2 ∈ [0, 1]. The map A is given in
coordinates by
 
x1 (2x1 + x2 ) mod 1
A =
x2 (x1 + x2 ) mod 1

(see Figure 1.4). Note that T2 is a commutative group and A is an automor-

phism, since A−1 is also an integer matrix.
The periodic points of A: T2 → T2 are the points with rational coordinates
(Exercise 1.7.1).
The lines in R2 parallel to the eigenvector vλ project to a family Wu of
parallel lines on T2 . For x ∈ T2 , the line Wu (x) through x is called the unstable
manifold of x. The family Wu partitions T2 and is called the unstable folia-
tion of A. This foliation is invariant in the sense that A(Wu (x)) = Wu (Ax).
Moreover, A expands each line in Wu by a factor of λ. Similarly, the stable
foliation Ws is obtained by projecting the family of lines in R2 parallel to
v1/λ . This foliation is also invariant under A, and A contracts each stable
manifold Ws (x) by 1/λ. Since the slopes of vλ and v1/λ are irrational, each of
the stable and unstable manifolds is dense in T2 (Exercise 1.11.1).
In a similar way, any n × n integer matrix B induces a group endomor-
phism of the n-torus Tn = Rn /Zn = [0, 1]n / ∼. The map is invertible (an
1.8. The Horseshoe 15

automorphism) if and only if B−1 is an integer matrix, which happens if and

only if | det B| = 1 (Exercise 1.7.2). If B is invertible and the eigenvalues do
not lie on the unit circle, then B: Tn → Tn has expanding and contracting
subspaces of complementary dimensions and is called a hyperbolic toral
automorphism. The stable and unstable manifolds of a hyperbolic toral
automorphism are dense in Tn (§5.10). This is easy to show in dimension
two (Exercise 1.7.3 and Exercise 1.11.1).
Hyperbolic toral automorphisms are prototypes of the more general class
of hyperbolic dynamical systems. These systems have uniform expansion and
contraction in complementary directions at every point. We discuss them in
detail in Chapter 5.

Exercise 1.7.1. Consider the automorphism of T2 corresponding to a non-

singular 2 × 2 integer matrix whose eigenvalues are not roots of 1.
(a) Prove that every point with rational coordinates is eventually periodic.
(b) Prove that every eventually periodic point has rational coordinates.

Exercise 1.7.2. Prove that the inverse of an n × n integer matrix B is also

an integer matrix if and only if | det B| = 1.

Exercise 1.7.3. Show that the eigenvalues of a two-dimensional hyperbolic

toral automorphism are irrational (so the stable and unstable manifolds are
dense by Exercise 1.11.1).

Exercise 1.7.4. Show that the number of fixed points of a hyperbolic toral
automorphism A is det(A − I) (hence the number of periodic points of
period n is det(An − I)).

1.8 The Horseshoe

Consider a region D ⊂ R2 consisting of two semicircular regions D1 and D5
together with a unit square R = D2 ∪ D3 ∪ D4 (see Figure 1.5).
Let f : D → D be a differentiable map that stretches and bends D into
a horseshoe as shown in Figure 1.5. Assume also that f stretches D2 ∪ D4
uniformly in the horizontal direction by a factor of µ > 2 and contracts

D1 D2 D3 D4 D5 p

Figure 1.5. The horseshoe map.

16 1. Examples and Basic Concepts

R01
R0
R00

R10
R1
R11

R
Figure 1.6. Horizontal rectangles.

uniformly in the vertical direction by λ < 1/2. Since f (D5 ) ⊂ D5 , the

Brouwer fixed point theorem implies the existence of a fixed point p ∈ D5 .
Set R0 = f (D2 ) ∩ R and R1 = f (D4 ) ∩ R. Note that f (R) ∩ R = R0 ∪ R1 .
The set f 2 (R) ∩ f (R) ∩ R = f 2 (R) ∩ R consists of four horizontal rectan-
gles Ri j , i, j ∈ {0, 1}, of height λ2 (see Figure 1.6). More generally, for any
finite sequence ω0 , . . . , ωn of zeros and ones,
   
Rω0 ω1 ... ωn = Rω0 ∩ f Rω1 ∩ · · · ∩ f n Rωn

is a horizontal rectangle of height λn , and f n (R) ∩ R is the union of

2n such rectangles. For an infinite sequence ω = (ωi ) ∈ {0, 1}N0 , let Rω =
∞ i +
∞ n

i=0 f (Rωi ). The set H = n=0 f (R) = ω Rω is the product of a hor-
izontal interval of length 1 and a vertical Cantor set C + (a Cantor set is a
compact, perfect, totally disconnected set). Note that f (H+ ) = H+ .
We now construct, in a similar way, a set H− using preimages. Observe that
f (R0 ) = f −1 (R) ∩ D2 , and f −1 (R1 ) = f −1 (R) ∩ D4 are vertical rectangles
−1

of width µ−1 . For any sequence ω−m, ω−m+1 , . . . , ω−1 of zeros and ones,
m −i −m
∞ −i
i=1 f (Rωi ) is a vertical rectangle of width µ , and H− = i=1 f (R)
is the product of a vertical interval (of length 1) and a horizontal Cantor
set C − .
∞
The horseshoe set H = H+ ∩ H− = i=−∞ f i (R) is the product of the
Cantor sets C − and C + and is closed and f -invariant. It is locally maximal,
i.e., there is an open set U containing H such that any f -invariant subset of U
containing H coincides with H (Exercise 1.8.2). The map φ: 2 = {0, 1}Z →
H that assigns to each infinite sequence ω = (ωi ) ∈ 2 the unique point

φ(ω) = ∞ i
−∞ f (Rωi ) is a bijection (Exercise 1.8.3). Note that


∞
 
f (φ(ω)) = f i+1 Rωi = φ(σr (ω)),
−∞
1.9. The Solenoid 17

where σr is the right shift in 2 , σr (ω)i+1 = ωi . Thus, φ conjugates f |H and

the full two-sided 2-shift.
The horseshoe was introduced by S. Smale in the 1960s as an example of
a hyperbolic set that “survives” small perturbations. We discuss hyperbolic
sets in Chapter 5.

Exercise 1.8.1. Draw a picture of f −1 (R) ∩ f (R) and f −2 (R) ∩ f 2 (R).

Exercise 1.8.2. Prove that H is a locally maximal f -invariant set.

Exercise 1.8.3. Prove that φ is a bijection, and that both φ and φ −1 are
continuous.

1.9 The Solenoid

Consider the solid torus T = S1 × D2 , where S1 = [0, 1] mod 1 and D2 =
{(x, y) ∈ R2 : x 2 + y2 ≤ 1}. Fix λ ∈ (0, 1/2), and define F: T → T by
 
F(φ, x, y) = 2φ, λx + 12 cos 2π φ, λy + 12 sin 2π φ .

The map F stretches by a factor of 2 in the S1 -direction, contracts by a

factor of λ in the D2 -direction, and wraps the image twice inside T (see
Figure 1.7).
The image F(T ) is contained in the interior int(T ) of T , and F n+1 (T ) ⊂
int(F n (T )). Note that F is one-to-one (Exercise 1.9.1). A slice F(T ) ∩ {φ =
const} consists of two disks of radius λ centered at diametrically opposite
points at distance 1/2 from the center of the slice. A slice F n (T ) ∩ {φ = const}

Figure 1.7. The solid torus and its image F(T ).

18 1. Examples and Basic Concepts

Figure 1.8. A cross-section of the solenoid.

consists of 2n disks of radius λn : two disks inside each of the 2n−1 disks of
F n−1 (T ) ∩ {φ = const}. Slices of F(T ), F 2 (T ), and F 3 (T ) for φ = 1/8 are
shown in Figure 1.8.

The set S = ∞ n
n=0 F (T ) is called a solenoid. It is a closed F-invariant
subset of T on which F is bijective (Exercise 1.9.1). It can be shown that S is
locally the product of an interval with a Cantor set in the two-dimensional
disk.
The solenoid is an attractor for F. In fact, any neighborhood of S con-
tains F n (T ) for n sufficiently large, so the forward orbit of every point in
T converges to S. Moreover, S is a hyperbolic set, and is therefore called a
hyperbolic attractor. We give a precise definition of attractors in §1.13.
∞
Let  denote the set of sequences (φi )i=0 , where φi ∈ S1 and φi =
2φi+1 mod 1 for all i. The product topology on (S1 )N0 induces the subspace
topology on . The space  is a commutative group under component-wise
addition (mod 1). The map (φ, ψ) →φ − ψ is continuous, so  is a topo-
logical group. The map α:  → , (φ0 , φ1 , . . .) →(2φ0 , φ0 , φ1 , . . .) is a group
automorphism and a homeomorphism (Exercise 1.9.3).
For s ∈ S, the first (angular) coordinates of the preimages F −n (s) =
−
(φn , xn , yn ) form a sequence h(s) = (φ0 , φ1 , . . .) ∈ . This defines a map

h: S → . The inverse of h is the map (φ0 , φ1 , . . .) → ∞ n=0 F ({φn } × D ),
n 2

and h is a homeomorphism (Exercise 1.9.2). Note that h: S →  conjugates

F and α, i.e., h ◦ F = α ◦ h. This conjugation allows one to study properties
of (S, F) by studying properties of the algebraic system (, α).

Exercise 1.9.1. Prove that (a) F: T → T is injective, and (b) F: S → S is

bijective.
1.10. Flows and Differential Equations 19

Exercise 1.9.2. Prove that for every (φ0 , φ1 , . . .) ∈  the intersection

∞
n=0 F ({φn } × D ) consists of a single point s, and h(s) = (φ0 , φ1 , . . .). Show
n 2

that h is a homeomorphism.
Exercise 1.9.3. Show that  is a topological group, and α is an automor-
phism and homeomorphism.
Exercise 1.9.4. Find the fixed point of F and all periodic points of period 2.

1.10 Flows and Differential Equations

Flows arise naturally from first-order autonomous differential equations.
Suppose ẋ = F(x) is a differential equation in Rn , where F: Rn → Rn is
continuously differentiable. For each point x ∈ Rn , there is a unique solution
f t (x) starting at x at time 0 and defined for all t in some neighborhood of
0. To simplify matters, we will assume that the solution is defined for all
t ∈ R; this will be the case, for example, if F is bounded, or is dominated
in norm by a linear function. For fixed t ∈ R, the time-t map x →f t (x) is a
C 1 diffeomorphism of Rn . Because the equation is autonomous, f t+s (x) =
f t ( f s (x)), i.e., f t is a flow.
Conversely, given a flow f t : Rn → Rn , if the map (t, x) →f t (x) is differ-
entiable, then f t is the time-t map of the differential equation

d 
ẋ =  f t (x).
dt t=0
Here are some examples. Consider the linear autonomous differential
equation ẋ = Ax in Rn , where A is a real n × n matrix. The flow of this
differential equation is f t (x) = e At x, where e At is the matrix exponential. If
Ais non-singular, the flow has exactly one fixed point at the origin. If all the
eigenvalues of A have negative real part, then every orbit approaches the
origin, and the origin is asymptotically stable. If some eigenvalue has positive
real part, then the origin is unstable.
Most differential equations that arise in applications are non-linear. The
differential equation governing an ideal frictionless pendulum is one of the
most familiar:
θ̈ + sin θ = 0.
This equation cannot be solved in closed form, but it can be studied by
qualitative methods. It is equivalent to the system
ẋ = y,
ẏ = −sin x.
20 1. Examples and Basic Concepts

The energy E of the system is the sum of the kinetic and potential energies,
E(x, y) = 1 − cos x + y2 /2. One can show (Exercise 1.10.2) by differentiating
E(x, y) with respect to t that E is constant along solutions of the differential
equation. Equivalently, if f t is the flow in R2 of this differential equation, then
E is invariant by the flow, i.e., E( f t (x, y)) = E(x, y), for all t ∈ R, (x, y) ∈ R2 .
A function that is constant on the orbits of a dynamical system is called a
first integral of the system.
The fixed points in the phase plane for the undamped pendulum are
(kπ, 0), k ∈ Z. The points (2kπ, 0) are local minima of the energy. The points
(2(k + 1)π, 0) are saddle points.
Now consider the damped pendulum θ̈ + γ θ̇ + sin θ = 0, or the equi-
valent system

ẋ = y,
ẏ = − sin x − γ y.

A simple calculation shows that Ė < 0 except at the fixed points (kπ, 0), k ∈
Z, which are the local extrema of the energy. Thus the energy is strictly
decreasing along every non-constant solution. In particular, every trajec-
tory approaches a critical point of the energy, and almost every trajectory
approaches a local minimum.
The energy of the pendulum is an example of a Lyapunov function, i.e.,
a continuous function that is non-increasing along the orbits of the flow.
Any strict local minimum of a Lyapunov function is an asymptotically sta-
ble equilibrium point of the differential equation. Moreover, any bounded
orbit must converge to the maximal invariant subset M of the set of points
satisfying Ė = 0. In the case of the damped pendulum, M = {(kπ, 0): k ∈ Z}.
Here is another class of examples that appears frequently in applications,
particularly optimization problems. Given a smooth function f : Rn → R,
the flow of the differential equation

ẋ = grad f (x)

is called the gradient flow of f . The function −f is a Lyapunov function

for the gradient flow. The trajectories are the projections to Rn of paths of
steepest ascent along the graph of f and are orthogonal to the level sets of
f (Exercise 1.10.3).
A Hamiltonian system is a flow in R2n given by a system of differential
equations of the form
∂H ∂H
q̇i = , ṗi = − , i = 1, . . . , n,
∂ pi ∂qi
1.11. Suspension and Cross-Section 21

where the Hamiltonian function H( p, q) is assumed to be smooth. Since

the divergence of the right-hand side is 0, the flow preserves volume. The
Hamiltonian function is a first integral, so that the level surfaces of H are
invariant under the flow. If for some C ∈ R the level surface H( p, q) = C
is compact, the restriction of the flow to the level surface preserves a finite
measure with smooth density. Hamiltonian flows have many applications
in physics and mathematics. For example, the flow associated with the un-
damped pendulum is a Hamiltonian flow, where the Hamiltonian function
is the total energy of the pendulum (Exercise 1.10.5).
Exercise 1.10.1. Show that the scalar differential equation ẋ = x log x in-
duces the flow f t (x) = x exp(t) on the line.
Exercise 1.10.2. Show that the energy is constant along solutions of the
undamped pendulum equation and strictly decreasing along non-constant
solutions of the damped pendulum equation.
Exercise 1.10.3. Show that −f is a Lyapunov function for the gradient flow
of f , and that the trajectories are orthogonal to the level sets of f .
Exercise 1.10.4. Prove that any differentiable one-parameter group of
linear maps of R is the flow of a differential equation ẋ = kx.
Exercise 1.10.5. Show that the flow of the undamped pendulum is a
Hamiltonian flow.

1.11 Suspension and Cross-Section

There are natural constructions for passing from a map to a flow, and vice
versa. Given a map f : X → X, and a function c: X → R+ bounded away
from 0, consider the quotient space
Xc = {(x, t) ∈ X × R+ : 0 ≤ t ≤ c(x)}/ ∼,
where ∼ is the equivalence relation (x, c(x)) ∼ ( f (x), 0). The suspension
of f with ceiling function c is the semiflow φ t : Xc → Xc given by φ t (x, s) =
( f n (x), s  ), where n and s  satisfy

n−1
c( f i (x)) + s  = t + s, 0 ≤ s  ≤ c( f n (x)).
i=0

In other words, flow along {x} × R+ to (x, c(x)), then jump to ( f (x), 0) and
continue along { f (x)} × R+ , and so on. See Figure 1.9. A suspension flow is
also called a flow under a function.
22 1. Examples and Basic Concepts

ψ t (a)
R+
(f (x), c(f (x)))
a
(x, c(x))
g(a)

A
(x, 0) (f (x), 0) X

Figure 1.9. Suspension and cross-section.

Conversely, a cross-section of a flow or semiflow ψ t : Y → Y is a subset A ⊂

Y with the following property: the set Ty = {t ∈ R+ : ψ t (y) ∈ A} is a non-
empty discrete subset of R+ for every y ∈ Y. For a ∈ A, let τ (a) = min Ta be
the return time to A. Define the first return map g: A → Aby g(a) = ψ τ (a) (a),
i.e., g(a) is the first point after a in Oψ+ (x) ∩ A (see Figure 1.9). The first
return map is often called the Poincaré map. Since the dimension of the
cross-section is less by 1, in many cases maps in dimension n present the
same level of difficulty as flows in dimension n + 1.
Suspension and cross-section are inverse constructions: the suspension of
g with ceiling function τ is ψ t , and X × {0} is a cross-section of φ with first
return map f . If φ is a suspension of f , then the dynamical properties of
f and φ are closely related, e.g., the periodic orbits of f correspond to the
periodic orbits of φ. Both of these constructions can be tailored to specific
settings (topological, measurable, smooth, etc.).
As an example, consider the 2-torus T2 = R2 /Z2 = S1 × S1 , with topo-
logy and metric induced from the topology and metric on R2 . Fix α ∈ R, and
define the linear flow φαt : T2 → T2 by

φαt (x, y) = (x + αt, y + t) mod 1.

Note that φαt is the suspension of the circle rotation Rα with ceiling function 1,
and S1 × {0} is a cross-section for φα with constant return time τ (y) = 1 and
first return map Rα . The flow φαt consists of left translations by the elements
g t = (αt, t) mod 1, which form a one-parameter subgroup of T2 .

Exercise 1.11.1. Show that if α is irrational, then every orbit of φα is dense

in T 2 , and if α is rational, then every orbit of φα is periodic.

Exercise 1.11.2. Let φ t be a suspension of f . Show that a periodic orbit

of φ t corresponds to a periodic orbit of f , and that a dense orbit of φ t
corresponds to a dense orbit of f .
1.12. Chaos and Lyapunov Exponents 23

∗
Exercise 1.11.3. Suppose 1, s, and αs are real numbers that are linearly
independent over Q. Show that every orbit of the time s map φαs is dense in
T 2.

1.12 Chaos and Lyapunov Exponents

A dynamical system is deterministic in the sense that the evolution of the sys-
tem is described by a specific map, so that the present (the initial state) com-
pletely determines the future (the forward orbit of the state). At the same
time, dynamical systems often appear to be chaotic in that they have sensitive
dependence on initial conditions, i.e., minor changes in the initial state lead to
dramatically different long-term behavior. Specifically, a dynamical system
(X, f ) has sensitive dependence on initial conditions on a subset X ⊂ X if
there is
> 0, such that for every x ∈ X and δ > 0 there are y ∈ X and n ∈ N
for which d(x, y) < δ and d( f n (x), f n (y)) >
. Although there is no univer-
sal agreement on a definition of chaos, it is generally agreed that a chaotic
dynamical system should exhibit sensitive dependence on initial conditions.
Chaotic systems are usually assumed to have some additional properties,
e.g., existence of a dense orbit.
The study of chaotic behavior has become one of the central issues in
dynamical systems during the last two decades. In practice, the term chaos
has been applied to a variety of systems that exhibit some type of random
behavior. This random behavior is observed experimentally in some situ-
ations, and in others follows from specific properties of the system. Often
a system is declared to be chaotic based on the observation that a typical
orbit appears to be randomly distributed, and different orbits appear to be
uncorrelated. The variety of views and approaches in this area precludes a
universal definition of the word “chaos.”
The simplest example of a chaotic system is the circle endomorphism
(S , Em), m > 1 (§1.3). Distances between points x and y are expanded by a
1

factor of m if d(x, y) ≤ 1/(2m), so any two points are moved at least 1/2m
apart by some iterate of Em, so Em has sensitive dependence on initial con-
ditions. A typical orbit is dense (§1.3) and is uniformly distributed on the
circle (Proposition 4.4.2).
The simplest non-linear chaotic dynamical systems in dimension one are
the quadratic maps qµ (x) = µx(1 − x), µ ≥ 4, restricted to the forward in-
variant set µ ⊂ [0, 1] (see §1.5 and Chapter 7).
Sensitive dependence on initial conditions is usually associated with
positive Lyapunov exponents. Let f be a differentiable map of an open
subset U ⊂ Rm into itself, and let d f (x) denote the derivative of f at x. For
24 1. Examples and Basic Concepts

x ∈ U and a non-zero vector v ∈ Rm define the Lyapunov exponent χ (x, v)

by
1
χ (x, v) = lim log d f n (x)v.
n→∞ n

If f has uniformly bounded first derivatives, then χ is well defined for every
x ∈ U and every non-zero vector v.
The Lyapunov exponent measures the exponential growth rate of tangent
vectors along orbits, and has the following properties:

χ (x, λv) = χ (x, v) for all real λ =0,

χ (x, v + w) ≤ max(χ (x, v), χ(x, w)), (1.1)
χ ( f (x), d f (x)v) = χ(x, v).

See Exercise 1.12.1.

If χ(x, v) = χ > 0 for some vector v, then there is a sequence n j → ∞
such that for every η > 0

d f n j (x)v ≥ e(χ −η)n j v.

This implies that, for a fixed j, there is a point y ∈ U such that

1 (χ −η)n j
d( f n j (x), f n j (y)) ≥ e d(x, y).
2
In general, this does not imply sensitive dependence on initial conditions,
since the distance between x and y cannot be controlled. However, most
dynamical systems with positive Lyapunov exponents have sensitive depen-
dence on initial conditions.
Conversely, if two close points are moved far apart by f n , by the interme-
diate value theorem, there must exist points x and directions v for which
d f n (x)v > v. Therefore, we expect f to have positive Lyapunov expo-
nents if it has sensitive dependence on initial conditions, though this is not
always the case.
The circle endomorphisms √ Em, m > 1, have positive exponents at all points.
A quadratic map qµ , µ > 2 + 5, has positive exponents at any point whose
forward orbit does not contain 0.

Exercise 1.12.1. Prove (1.1).

Exercise 1.12.2. Compute the Lyapunov exponents for Em.

Exercise 1.12.3. Compute the Lyapunov exponents for the solenoid, §1.9.
1.13. Attractors 25

Exercise
√ 1.12.4. Using a computer, calculate the first 100 points in the orbit
of 2 − 1 under the map E2 . What proportion of these points is contained
in each of the intervals [0, 14 ), [ 14 , 12 ), [ 12 , 34 ), and [ 34 , 1)?

1.13 Attractors
Let X be a compact topological space, and f : X → X be a continuous map.
Generalizing the notion of an attracting fixed point, we say that a compact
set C ⊂ X is an attractor if there is an open set U containing C, such that

f (Ū) ⊂ U and C = n≥0 f n (U). It follows that f (C) = C, since f (C) =
 
n≥1 f (U) ⊂ C; on the other hand, C = n≥1 f (U) = f (C), since f (U) ⊂
n n

U. Moreover, the forward orbit of any point x ∈ U converges to C, i.e., for

any open set V containing C, there is some N > 0 such that f n (x) ∈ V for
all n ≥ N. To see this, observe that X is covered by V together with the
open sets X\ f n (Ū), n ≥ 0. By compactness, there is a finite subcover, and
since f n (U) ⊂ f n−1 (U), we conclude that there is some N > 0 such that
X = V ∪ (X\ f n (Ū)) for all n ≥ N. Thus, f n (x) ∈ f n (U) ⊂ V for n ≥ N.


The basin of attraction of C is the set BA(C) = n≥0 f −n (U). The basin
BA(C) is precisely the set of points whose forward orbits converge to C
(Exercise 1.13.1).
An open set U ⊂ X such that Ū is compact and f (Ū) ⊂ U is called a

trapping region for f . If U is a trapping region, then n≥0 f n (U) is an at-
tractor. For flows generated by differential equations, any region with the
property that along the boundary the vector field points into the region is
a trapping region for the flow. In practice, the existence of an attractor is
proved by constructing a trapping region. An attractor can be studied ex-
perimentally by numerically approximating orbits that start in the trapping
region.
The simplest examples of attractors are: the intersection of the images
of the whole space (if the space is compact); attracting fixed points; and
attracting periodic orbits. For flows, the examples include asymptotically
stable fixed points and asymptotically stable periodic orbits.
Many dynamical systems have attractors of a more complicated nature.
For example, recall that the solenoid S (§1.9) is a (hyperbolic) attractor for
(T , F). Locally, S is the product of an interval with a Cantor set. The structure
of hyperbolic attractors is relatively well understood. However, some non-
linear systems have attractors that are chaotic (with sensitive dependence
on initial conditions) but not hyperbolic. These attractors are called strange
attractors. The best-known examples of strange attractors are the Hénon
attractor and the Lorenz attractor.
26 1. Examples and Basic Concepts

The study of strange attractors began with the publication by E. N. Lorenz

in 1963 of the paper “Deterministic non-periodic flow” [Lor63]. In the pro-
cess of investigating meteorological models, Lorenz studied the non-linear
system of differential equations

ẋ = σ (y − x),
ẏ = Rx − y − xz, (1.2)
ż = −bz + xy,

now called the Lorenz system. He observed that at parameter values σ =

10, b = 8/3, and R = 28, the solutions of (1.2) eventually√start revolving
√
alternately about two repelling equilibrium points at (± 72, ± 72, 27).
The number of times the solution revolves about one equilibrium before
switching to the other has no discernible pattern. There is a trapping region
U that contains 0 but not the two repelling equilibrium points. The attractor
contained in U is called the Lorenz attractor. It is an extremely complicated
set consisting of uncountably many orbits (including a saddle fixed point at
0), and non-fixed periodic orbits that are known to be knotted [Wil84]. The
attractor is not hyperbolic in the usual sense, though it has strong expansion
and contraction and sensitive dependence on initial conditions. The attractor
persists for small changes in the parameter values (see Figure 1.10).
The Hénon map H = ( f, g): R2 → R2 is defined by

f (x, y) = a − by − x 2 ,
g(x, y) = x,

where a and b are constants [Hén76]. The Jacobian of the derivative dH

50

40

30

25
20 20
15
10
5
10 0
-5 y
-20 -15 -10
-10 -5 -15
0 5 -20
10 15 -25
x 20

Figure 1.10. Lorenz attractor.

1.13. Attractors 27

-1

-2
-2 -1 0 1 2
Figure 1.11. Hénon attractor.

equals b. If b =0, the Hénon map is invertible; the inverse is

(x, y) →(y, (a − x − y2 )/b).
The map changes area by a factor of |b|, and is orientation reversing if b < 0.
For the specific parameter values a = 1.4 and b = −0.3, Hénon showed
that there is a trapping region U homeomorphic to a disk. His numerical
experiments suggested that the resulting attractor has a dense orbit and
sensitive dependence on initial conditions, though these properties have not
been rigorously proved. Figure 1.11 shows a long segment of an orbit starting
in the trapping region, which is believed to approximate the attractor. It is
known that for a large set of parameter values a ∈ [1, 2], b ∈ [−1, 0], the
attractor has a dense orbit and a positive Lyapunov exponent, but is not
hyperbolic [BC91].
Exercise 1.13.1. Let A be an attractor. Show that x ∈ B(A) if and only if
the forward orbit of x converges to A.
Exercise 1.13.2. Find a trapping region for the flow generated by the Lorenz
equations with parameter values σ = 10, b = 8/3, and R = 28.
Exercise 1.13.3. Find a trapping region for the Hénon map with parameter
values a = 1.4, b = −0.3.
Exercise 1.13.4. Using a computer, plot the first 1000 points in an orbit of
the Hénon map starting in a trapping region.
CHAPTER TWO

Topological Dynamics

A topological dynamical system is a topological space X and either a contin-

uous map f : X → X or a continuous (semi)flow f t on X, i.e., a (semi)flow f t
for which the map (t, x) →f t (x) is continuous. To simplify the exposition, we
usually assume that X is locally compact, metrizable, and second countable,
though many of the results in this chapter are true under weaker assumptions
on X. As we noted earlier, we will focus our attention on discrete-time sys-
tems, though all general results in this chapter are valid for continuous-time
systems as well.
Let X and Y be topological spaces. Recall that a continuous map f : X →Y
is a homeomorphism if it is one-to-one and the inverse is continuous.
Let f : X → X and g: Y → Y be topological dynamical systems. A topo-
logical semiconjugacy from g to f is a surjective continuous map h: Y → X
such that f ◦ h = h ◦ g. If h is a homeomorphism, it is called a topological
conjugacy, and f and g are said to be topologically conjugate or isomorphic.
Topologically conjugate dynamical systems have identical topological prop-
erties. Consequently, all the properties and invariants we introduce in this
chapter, including minimality, topological transitivity, topological mixing,
and topological entropy, are preserved by topological conjugacy.
Throughout this chapter, a metric space X with metric d is denoted (X, d).
If x ∈ X and r > 0, then B(x, r ) denotes the open ball of radius r centered
at x. If (X, d) and (Y, d ) are metric spaces, then f : X → Y is an isometry if
d ( f (x1 ), f (x2 )) = d(x1 , x2 ) for all x1 , x2 ∈ X.

2.1 Limit Sets and Recurrence

Let f : X → X be a topological dynamical system. Let x be a point in X.
A point y ∈ X is an ω-limit point of x if there is a sequence of natural num-
bers nk → ∞ (as k → ∞) such that f nk (x) → y. The ω-limit set of x is the
28
2.1. Limit Sets and Recurrence 29

set ω(x) = ω f (x) of all ω-limit points of x. Equivalently,


ω(x) = f i (x).
n∈N i≥n
 
If f is invertible, the α-limit set of x is α(x) = α f (x) = n∈N i≥n f −i (x).
A point in α(x) is an α-limit point of x. Both the α- and ω-limit sets are closed
and f -invariant (Exercise 2.1.1).
A point x is called (positively) recurrent if x ∈ ω(x); the set R( f ) of
recurrent points is f -invariant. Periodic points are recurrent.
A point x is non-wandering if for any neighborhood U of x there exists
n ∈ N such that f n (U) ∩ U = ∅. The set NW( f ) of non-wandering points is
closed, f -invariant, and contains ω(x) and α(x) for all x ∈ X (Exercise 2.1.2).
Every recurrent point is non-wandering, and in fact R( f ) ⊂ NW( f )
(Exercise 2.1.3); in general, however, NW( f ) ⊂ R( f ) (Exercise 2.1.11).

Recall the notation O(x) = n∈Z f (x) for an invertible mapping f , and
n

O+ (x) = n∈N0 f n (x).

PROPOSITION 2.1.1
1. Let f be a homeomorphism, y ∈ O(x), and z ∈ O(y). Then z ∈ O(x).
2. Let f be a continuous map, y ∈ O+ (x), and z ∈ O+ (y). Then z ∈ O+ (x).
Proof. Exercise 2.1.7. 

Let X be compact. A closed, non-empty, forward f -invariant subset Y ⊂
X is a minimal set for f if it contains no proper, closed, non-empty, forward
f -invariant subset. A compact invariant set Y is minimal if and only if the for-
ward orbit of every point in Y is dense in Y (Exercise 2.1.4). Note that a peri-
odic orbit is a minimal set. If X itself is a minimal set, we say that f is minimal.

PROPOSITION 2.1.2. Let f : X → X be a topological dynamical system. If

X is compact, then X contains a minimal set for f .
Proof. The proof is a straightforward application of Zorn’s lemma. Let
C be the collection of non-empty, closed, f -invariant subsets of X, with the
partial ordering given by inclusion. Then C is not empty, since X ∈ C. Suppose
K ⊂ C is a totally ordered subset. Then any finite intersection of elements
of K is non-empty, so by the finite intersection property for compact sets,

K∈K K = ∅. Thus, by Zorn’s lemma, C contains a minimal element, which
is a minimal set for f. 

In a compact topological space, every point in a minimal set is recurrent
(Exercise 2.1.4), so the existence of minimal sets implies the existence of
recurrent points.
30 2. Topological Dynamics

A subset A ⊂ N (or Z) is relatively dense (or syndetic) if there is k > 0 such

that {n, n + 1, . . . , n + k} ∩ A =for
∅ any n. A point x ∈ X is almost periodic
if for any neighborhood U of x, the set {i ∈ N: f i (x) ∈ U} is relatively dense
in N.

PROPOSITION 2.1.3. If X is a compact Hausdorff space and f : X → X is

continuous, then O+ (x) is minimal for f if and only if x is almost periodic.
Proof. Suppose x is almost periodic and y ∈ O+ (x). We need to show that
x ∈ O+ (y). Let U be a neighborhood of x. There is an open set U  ⊂ X, x ∈
U  ⊂ U, and an open set V ⊂ X ×X containing the diagonal, such that if
x1 ∈ U  and (x1 , x2 ) ∈ V, then x2 ∈ U. Since x is almost periodic, there is K ∈ N
such that for every j ∈ N we have that f j+k(x) ∈ U  for some 0 ≤ k ≤ K.
 K −i
Let V  = i=0 f (V). Note that V  is open and contains the diagonal of
X × X. There is a neighborhood W  y such that W × W ⊂ V  . Choose n
such that f n (x) ∈ W, and choose k such that f n+k(x) ∈ U  with 0 ≤ k ≤ K.
Then ( f n+k(x), f k(y)) ∈ V, and hence f k(y) ∈ U.
Conversely, suppose x is not almost periodic. Then there is a neigh-
borhood U of x such that A = {i: f i (x) ∈ U} is not relatively dense. Thus,
there are sequences ai ∈ N and ki ∈ N, ki → ∞, such that f ai + j (x) ∈ / U for
j = 0, . . . , ki . Let y be a limit point of { f ai (x)}. By passing to a subsequence,
we may assume that f ai (x) → y. Fix j ∈ N. Note that f ai + j (x) → f j (y),
and f ai + j (x) ∈ / U for i sufficiently large. Thus f j (y) ∈ / U for all j ∈ N, so
x∈ + +
/ O (y), which implies that O (x) is not minimal. 

Recall that an irrational circle rotation Rα is minimal (§1.2). Therefore
every point is non-wandering, recurrent, and almost periodic. An expanding
endomorphism Em of the circle has dense orbits (§1.3), but is not minimal
because it has periodic points. Every point is non-wandering, but not all
points are recurrent (Exercise 2.1.5).

Exercise 2.1.1. Show that the α- and ω-limit sets of a point are closed in-
variant sets.

Exercise 2.1.2. Show that the set of non-wandering points is closed, is f -

invariant, and contains ω(x) and α(x) for all x ∈ X.

Exercise 2.1.3. Show that R( f ) ⊂ NW( f ).

Exercise 2.1.4. Let X be compact, f : X → X continuous.

(a) Show that Y ⊂ X is minimal if and only if ω(y) = Y for every y ∈ Y.
(b) Show that Y is minimal if and only if the forward orbit of every point
in Y is dense in Y.
2.2. Topological Transitivity 31

Exercise 2.1.5. Show that there are points that are non-recurrent and not
eventually periodic for an expanding circle endomorphism Em.

Exercise 2.1.6. For a hyperbolic toral automorphism A: T2 → T2 , show

that:
(a) R(A) is dense, and hence NW(A) = T2 , but
(b) R(A) = T2 .

Exercise 2.1.7. Prove Proposition 2.1.1.

Exercise 2.1.8. Prove that a homeomorphism f : X → X is minimal if

and only if for each non-empty open set U ⊂ X there is n ∈ N, such that
n
k=−n f (U) = X.
k

Exercise 2.1.9. Prove that a homeomorphism f of a compact metric space

X is minimal if and only if for every  > 0 there is N = N() ∈ N, such that
for every x ∈ X the set {x, f (x), . . . , f N (x)} is -dense in X.

Exercise 2.1.10. Let f : X → X and g: Y → Y be continuous maps of com-

pact metric spaces. Prove that O+f ×g (x, y) = O+f (x) × Og+ (y) if and only if
(x, g(y)) ∈ O+f ×g (x, y).
Assume that f and g are minimal. Find necessary and sufficient conditions
for f × g to be minimal.

*Exercise 2.1.11. Give an example of a dynamical system where NW( f ) ⊂

R( f ).

2.2 Topological Transitivity

We assume throughout this section that X is second countable.
A topological dynamical system f : X → X is topologically transitive if
there is a point x ∈ X whose forward orbit is dense in X. If X has no isolated
points, this condition is equivalent to the existence of a point whose ω-limit
set is dense in X (Exercise 2.2.1).

PROPOSITION 2.2.1. Let f : X → X be a continuous map of a locally com-

pact Hausdorff space X. Suppose that for any two non-empty open sets U and
V there is n ∈ N such that f n (U) ∩ V = ∅. Then f is topologically transitive.
Proof. The hypothesis implies that given any open set V ⊂ X, the set
 −n
n∈N f (V) is dense in X, since it intersects every open set. Let {Vi } be
 
a countable basis for the topology of X. Then Y = i n∈N f −n (Vi ) is a
countable intersection of open, dense sets and is therefore non-empty by
32 2. Topological Dynamics

the Baire category theorem. The forward orbit of any point y ∈ Y enters
each Vi , hence is dense in X. 

In most topological spaces, existence of a dense full orbit for a home-
omorphism implies existence of a dense forward orbit, as we show in the
next proposition. Note, however, that density of a particular full orbit O(x)
does not imply density of the corresponding forward orbit O+ (x) (see
Exercise 2.2.2).
PROPOSITION 2.2.2. Let f : X → X be a homeomorphism of a compact
metric space, and suppose that X has no isolated points. If there is a dense full
orbit O(x), then there is a dense forward orbit O+ (y).
Proof. Since O(x) = X, the orbit O(x) visits every non-empty open set U
at least once, and therefore infinitely many times because X has no isolated
points. Hence there is a sequence nk, with |nk| → ∞, such that f nk (x) ∈
B(x, 1/k) for k ∈ N, i.e., f nk (x) → x as k → ∞. Thus, f nk+l (x) → f l (x) for
any l ∈ Z. There are either infinitely many positive or infinitely many neg-
ative indices nk, and it follows that either O(x) ⊂ O+ (x) or O(x) ⊂ O− (x).
In the former case, O+ (x) = X, and we are done. In the latter case, let
U, V be non-empty open sets. Since O− (x) = X, there are integers i <
j < 0 such that f i (x) ∈ U and f j (x) ∈ V, so f j−i (U) ∩ V =. ∅Hence, by
Proposition 2.2.1, f is topologically transitive. 

Exercise 2.2.1. Show that if X has no isolated points and O+ (x) is dense,
then ω(x) is dense. Give an example to show that this is not true if X has
isolated points.
Exercise 2.2.2. Give an example of a dynamical system with a dense full
orbit but no dense forward orbit.
Exercise 2.2.3. Is the product of two topologically transitive systems topo-
logically transitive? Is a factor of a topologically transitive system topologi-
cally transitive?
Exercise 2.2.4. Let f : X → X be a homeomorphism. Show that if f has
a non-constant first integral or Lyapunov function (§1.10), then it is not
topologically transitive.
Exercise 2.2.5. Let f : X → X be a topological dynamical system with at
least two orbits. Show that if f has an attracting periodic point, then it is not
topologically transitive.
Exercise 2.2.6. Let α be irrational and f : T 2 → T 2 be the homeomor-
phism of the 2-torus given by f (x, y) = (x + α, x + y).
2.3. Topological Mixing 33

(a) Prove that every non-empty, open, f -invariant set is dense, i.e., f is
topologically transitive.
(b) Suppose the forward orbit of (x0 , y0 ) is dense. Prove that for ev-
ery y ∈ S1 the forward orbit of (x0 , y) is dense. Moreover, if the set
n n
k=0 f (x0 , y0 ) is -dense, then k=0 f (x0 , y) is -dense.
k k

(c) Prove that every forward orbit is dense, i.e., f is minimal.

2.3 Topological Mixing

A topological dynamical system f : X → X is topologically mixing if for any
two non-empty open sets U, V ⊂ X, there is N > 0 such that f n (U) ∩ V = ∅
for n ≥ N. Topological mixing implies topological transitivity by Proposi-
tion 2.2.1, but not vice versa. For example, an irrational circle rotation is
minimal and therefore topologically transitive, but not topologically mixing
(Exercise 2.3.1).
The following propositions establish topological mixing for some of the
examples from Chapter 1.

PROPOSITION 2.3.1. Any hyperbolic toral automorphism A: T2 → T2 is

topologically mixing.
Proof. By Exercise 1.7.3, for each x ∈ T2 , the unstable manifold Wu (x) of A
is dense in T2 . Thus for every  > 0, the collection of balls of radius  centered
at points of Wu (x) covers T2 . By compactness, a finite subcollection of these
balls also covers T2 . Hence, there is a bounded segment S0 ⊂ Wu (x) whose
-neighborhood covers T2 . Since group translations of T2 are isometries, the
-neighborhood of any translate Lg S0 = g + S0 ⊂ Wu (g + x) covers T2 . To
summarize: For every  > 0, there is L() > 0 such that every segment S of
length L() in an unstable manifold is -dense in T2 , i.e., d(y, S) ≤  for every
y ∈ T2 .
Let U and V be non-empty open sets in T2 . Choose y ∈ V and  > 0 such
that B(y, ) ⊂ V. The open set U contains a segment of length δ > 0 in some
unstable manifold Wu (x). Let λ, |λ| > 1, be the expanding eigenvalue of A,
and choose N > 0 such that |λ| N δ ≥ L(). Then for any n ≥ N, the image
AnU contains a segment of length at least L() in some unstable manifold,
so AnU is -dense in T2 and therefore intersects V. 

PROPOSITION 2.3.2. The full two-sided shift (m, σ ) and the full one-sided
+
shift (m , σ ) are topologically mixing.
Proof. Recall from §1.4 that the topology on m has a basis consisting
of open metric balls B(ω, 2−l ) = {ω : ωi = ωi , |i| ≤ l}. Thus it suffices to
34 2. Topological Dynamics

show that for any two balls B(ω, 2−l1 ) and B(ω , 2−l2 ), there is N > 0 such
that σ n B(ω, 2−l1 ) ∩ B(ω , 2−l2 ) = for
∅ n ≥ N. Elements of σ n B(ω, 2−l1 ) are
sequences with specified values in the places −n − l1 , . . . , −n + l1 . Therefore
the intersection is non-empty when −n + l1 < −l2 , i.e., n ≥ N = l1 + l2 + 1.
+
This proves that (m, σ ) is topologically mixing; the proof for (m , σ ) is an
exercise (Exercise 2.3.4). 

COROLLARY 2.3.3. The horseshoe (H, f ) (§1.8) is topologically mixing.
Proof. The horseshoe (H, f ) is topologically conjugate to the full two-shift
(2 , σ ) (see Exercise 1.8.3). 

PROPOSITION 2.3.4. The solenoid (S, F) is topologically mixing.
Proof. Recall (Exercise 1.9.2) that (S, F) is topologically conjugate to
( , α), where


∞
= {(φi ): φi ∈ S1 , φi = 2φi+1 , ∀i} ⊂ S1 = T∞ ,
i=0

α
and (φ0 , φ1 , φ2 , . . .) →(2φ0 , φ0 , φ1 , . . .). Thus, it suffices to show that ( , α)
is topologically mixing. The topology in T∞ has a basis consisting of open
∞
sets i=0 Ik, where the I j s are open in S1 and all but finitely many are equal
to S1 . Let U = (I0 × I1 × · · · × Ik × S1 × S1 × · · ·) ∩ and V = (J0 × J1 ×
· · · × Jl × S1 × S1 × · · ·) ∩ be non-empty open sets from this basis. Choose
m > 0 so that 2m I0 = S1 . Then for n > m + l, the first n − m components of

α n (U) = (2n I0 × 2n−1 I0 × · · · × I0 × I1 × · · · × Ik × S1 × S1 × · · ·) ∩

are S1 , so α n (U) ∩ V =. ∅ 

Exercise 2.3.1. Show that a circle rotation is not topologically mixing. Show
that an isometry is not topologically mixing if there is more than one point
in the space.

Exercise 2.3.2. Show that expanding endomorphisms of S1 are topologi-

cally mixing (see §1.3).

Exercise 2.3.3. Show that a factor of a topologically mixing system is also

topologically mixing.

+
Exercise 2.3.4. Prove that (m , σ ) is topologically mixing.
2.4. Expansiveness 35

2.4 Expansiveness
A homeomorphism f : X → X is expansive if there is δ > 0 such that for any
two distinct points x, y ∈ X, there is some n ∈ Z such that d( f n (x), f n (y)) ≥
δ. A non-invertible continuous map f : X → X is positively expansive if there
is δ > 0 such that for any two distinct points x, y ∈ X, there is some n ≥ 0
such that d( f n (x), f n (y)) ≥ δ. Any number δ > 0 with this property is called
an expansiveness constant for f .
Among the examples from Chapter 1, the following are expansive (or
positively expansive): the circle endomorphisms Em, |m| ≥ 2; the full and
one-sided shifts; the hyperbolic toral automorphisms; the horseshoe; and
the solenoid (Exercise 2.4.2). For sufficiently large values of the parameter
µ, the quadratic map qµ is expansive on the invariant set µ . Circle rotations,
group translations, and other equicontinuous homeomorphisms (see §2.7)
are not expansive.
PROPOSITION 2.4.1. Let f be a homeomorphism of an infinite compact
metric space X. Then for every  > 0 there are distinct points x0 , y0 ∈ X such
that d( f n (x0 ), f n (y0 )) ≤  for all n ∈ N0 .
Proof [Kin90]. Fix  > 0. Let E be the set of natural numbers m for which
there is a pair x, y ∈ X such that
d(x, y) ≥  and d( f n (x), f n (y)) ≤  for n = 1, . . . , m. (2.1)
Let M = sup E if E = ∅, and M = 0 if E = ∅.
If M = ∞, then for every m ∈ N there is a pair xm, ym satisfying (2.1). By
compactness, there is a sequence mk → ∞ such that the limits
lim xmk = x  , lim ymk = y
k→∞ k→∞

exist. By (2.1), d(x  , y ) ≥  and, since f j is continuous,

    
d( f j (x  ), f j (y )) = lim d f j xmk , f j ymk ≤ 
k→∞

for every j ∈ N. Thus, x0 = f (x ), y0 = f (y ) are the desired points.



Suppose now that M is finite. Since any finite collection of iterates of f is

equicontinuous, there is δ > 0 such that if d(x, y) < δ, then d( f n (x), f n (y)) <
 for 0 ≤ n ≤ M; the definition of M then implies that d( f −1 (x), f −1 (y)) < .
By induction, we conclude that d( f − j (x), f − j (y)) <  for j ∈ N whenever
d(x, y) < δ. By compactness, there is a finite collection B of open δ/2-balls
that covers X. Let K be the cardinality of B. Since X is infinite, we can
choose a set W ⊂ X consisting of K + 1 distinct points. By the pigeon-hole
principle, for each j ∈ Z, there are distinct points a j , b j ∈ W such that f j (a j )
36 2. Topological Dynamics

and f j (b j ) belong to the same ball Bj ∈ B, so d( f j (a j ), f j (b j )) < δ. Thus,

d( f n (a j ), f n (b j )) <  for −∞ < n ≤ j. Since W is finite, there are distinct
x0 , y0 ∈ W such that

a j = x0 and b j = y0

for infinitely many positive j and hence d( f n (x0 ), f n (y0 )) <  for all n ≥ 0.


Proposition 2.4.1 is also true for non-invertible maps (Exercise 2.4.3).

COROLLARY 2.4.2. Let f be an expansive homeomorphism of an infinite

compact metric space X. Then there are x0 , y0 ∈ X such that d( f n (x0 ),
f n (y0 )) → 0 as n → ∞.
Proof. Let δ > 0 be an expansiveness constant for f . By Proposition 2.4.1,
there are x0 , y0 ∈ X such that d( f n (x0 ), f n (y0 )) < δ for all n ∈ N. Suppose
d( f n (x0 ), f n (y0 ))  0. Then by compactness, there is a sequence nk → ∞
such that f nk (x0 ) → x  and f nk (y0 ) → y with x  = y . Then f nk+m(x0 ) →
m  nk +m m 
f (x ) and f (y0 ) → f (y ) for any m ∈ Z. For k large, nk + m > 0 and
hence d( f m(x  ), f m(y )) ≤ δ for all m ∈ Z, which contradicts expansiveness.


Exercise 2.4.1. Prove that every isometry of a compact metric space to
itself is surjective and therefore is a homeomorphism.

Exercise 2.4.2. Show that the expanding circle endomorphisms Em, |m| ≥
2, the full one- and two-sided shifts, the hyperbolic toral automorphisms,
the horseshoe, and the solenoid are expansive, and compute expansiveness
constants for each.

Exercise 2.4.3. Show that Proposition 2.4.1 is true for non-invertible con-
tinuous maps of infinite metric spaces.

2.5 Topological Entropy

Topological entropy is the exponential growth rate of the number of es-
sentially different orbit segments of length n. It is a topological invariant
that measures the complexity of the orbit structure of a dynamical system.
Topological entropy is analogous to measure-theoretic entropy, which we
introduce in Chapter 9.
2.5. Topological Entropy 37

Let (X, d) be a compact metric space, and f : X → X a continuous map.

For each n ∈ N, the function

dn (x, y) = max d( f k(x), f k(y))

0≤k≤n−1

measures the maximum distance between the first n iterates of x and y. Each
dn is a metric on X, dn ≥ dn−1 , and d1 = d. Moreover, the di are all equivalent
metrics in the sense that they induce the same topology on X(Exercise 2.5.1).
Fix  > 0. A subset A ⊂ X is (n, )-spanning if for every x ∈ X there is
y ∈ A such that dn (x, y) < . By compactness, there are finite (n, )-spanning
sets. Let span(n, , f ) be the minimum cardinality of an (n, )-spanning set.
A subset A ⊂ X is (n, )-separated if any two distinct points in A are at
least  apart in the metric dn . Any (n, )-separated set is finite. Let sep(n, , f )
be the maximum cardinality of an (n, )-separated set.
Let cov(n, , f ) be the minimum cardinality of a covering of X by sets of
dn -diameter less than  (the diameter of a set is the supremum of distances
between pairs of points in the set). Again, by compactness, cov(n, , f ) is
finite.
The quantities span(n, , f ), sep(n, , f ), and cov(n, , f ) count the num-
ber of orbit segments of length n that are distinguishable at scale . These
quantities are related by the following lemma.

LEMMA 2.5.1. cov(n, 2, f ) ≤ span(n, , f ) ≤ sep(n, , f ) ≤ cov(n, , f ).

Proof. Suppose A is an (n, )-spanning set of minimum cardinality. Then the
open balls of radius  centered at the points of A cover X. By compactness,
there exists 1 <  such that the balls of radius 1 centered at the points of
Aalso cover X. Their diameter is 21 < 2, so cov(n, 2, f ) ≤ span(n, , f ).
The other inequalities are left as an exercise (Exercise 2.5.2). 

Let
1
h ( f ) = lim log(cov(n, , f )). (2.2)
n→∞ n

The quantity cov(n, , f ) increases monotonically as  decreases, so h ( f )

does as well. Thus the limit

htop = h( f ) = lim+ h ( f )
→0

exists; it is called the topological entropy of f . The inequalities in Lemma 2.5.1

imply that equivalent definitions of h( f ) can be given using span(n, , f ) or
38 2. Topological Dynamics

sep(n, , f ), i.e.,

1
h( f ) = lim+ lim log(span(n, , f )) (2.3)
→0 n
n→∞

1
= lim+ lim log(sep(n, , f )). (2.4)
→0 n→∞ n

LEMMA 2.5.2. The limit lim n→∞ 1

n
log(cov(n, , f )) = h ( f ) exists and is
finite.
Proof. Let U have dm-diameter less than , and V have dn -diameter less
than . Then U ∩ f −m(V) has dm+n -diameter less than . Hence

cov(m + n, , f ) ≤ cov(m, , f ) · cov(n, , f ),

so the sequence an = log(cov(n, , f )) ≥ 0 is subadditive. A standard

lemma from calculus implies that an /n converges to a finite limit as n → ∞
(Exercise 2.5.3). 

It follows from Lemmas 2.5.1 and 2.5.2 that the lim sups in Formulas (2.2),
(2.3), and (2.4) are finite. Moreover, the corresponding lim infs are finite, and

1
h( f ) = lim+ lim log(cov(n, , f )) (2.5)
→0 n→∞ n
1
= lim+ lim log(span(n, , f )) (2.6)
→0 n→∞ n
1
= lim+ lim log(sep(n, , f )). (2.7)
→0 n→∞ n

The topological entropy is either +∞ or a finite non-negative number.

There are dramatic differences between dynamical systems with positive
entropy and dynamical systems with zero entropy. Any isometry has zero
topological entropy (Exercise 2.5.4). In the next section, we show that topo-
logical entropy is positive for several of the examples from Chapter 1.

PROPOSITION 2.5.3. The topological entropy of a continuous map f : X →

X does not depend on the choice of a particular metric generating the topology
of X.
Proof. Suppose d and d are metrics generating the topology of X. For  > 0,
let δ() = sup{d (x, y): d(x, y) ≤ }. By compactness, δ() → 0 as  → 0. If
U is a set of dn -diameter less than , then U has dn -diameter at most δ().
Thus cov (n, δ(), f ) ≤ cov(n, , f ), where cov and cov correspond to the
2.5. Topological Entropy 39

metrics d and d , respectively. Hence,

1 1
lim lim log(cov (n, δ, f )) ≤ lim+ lim log(cov(n, , f )).
δ→0+ n→∞ n →0 n→∞ n

Interchanging d and d gives the opposite inequality. 


COROLLARY 2.5.4. Topological entropy is an invariant of topological con-
jugacy .
Proof. Suppose f : X → X and g: Y → Y are topologically conjugate dy-
namical systems, with conjugacy φ: Y → X. Let d be a metric on X. Then
d (y1 , y2 ) = d(φ(y1 ), φ(y2 )) is a metric on Y generating the topology of Y.
Since φ is an isometry of (X, d) and (Y, d ), and the entropy is independent
of the metric by Proposition 2.5.3, it follows that h( f ) = h(g). 

PROPOSITION 2.5.5. Let f : X → X be a continuous map of a compact met-
ric space X.
1. h( f m) = m · h( f ) for m ∈ N.
2. If f is invertible, then h( f −1 ) = h( f ). Thus h( f m) = |m| · h( f ) for all
m ∈ Z.
3. If Ai , i = 1, . . . , k are closed (not necessarily disjoint) forward f -
invariant subsets of X, whose union is X, then

h( f ) = max h( f |Ai ).
1≤i≤k

In particular, if A is a closed forward invariant subset of X, then

h( f |A) ≤ h( f ).

Proof. 1: Note that

max d( f mi (x), f mi (y)) ≤ max d( f j (x), f j (y)).

0≤i<n 0≤ j<mn

Thus, span(n, , f m) ≤ span(mn, , f ), so h( f m) ≤ m · h( f ). Conversely, for

 > 0, there is δ() > 0 such that d(x, y) < δ() implies that d( f i (x), f i (y)) <
 for i = 0, . . . , m. Then span(n, δ(), f m) ≥ span(mn, , f ), so h( f m) ≥ m ·
h( f ).
2: The nth image of an (n, )-separated set for f is (n, )-separated for
f −1 , and vice versa.
3: Any (n, )-separated set in Ai is (n, )-separated in X, so h( f |Ai ) ≤
h( f ). Conversely, the union of (n, )-spanning sets for the Ai s is an (n, )-
spanning set for X. Thus if spani (n, , f ) is the minimum cardinality of an
40 2. Topological Dynamics

(n, )-spanning subset of Ai , then

k
span(n, , f ) ≤ spani (n, , f ) ≤ k · max spani (n, , f ).
1≤i≤k
i=1

Therefore,


1 1 1
lim log (span(n, , f )) ≤ lim log k + lim log max spani (n, , f )
n→∞ n n→∞ n n→∞ n 1≤i≤k

1
= 0 + max lim log (spani (n, , f ))
1≤i≤k n→∞ n

The result follows by taking the limit as  → 0. 


PROPOSITION 2.5.6. Let (X, d X) and (Y, dY ) be compact metric spaces, and
f : X → X, g: Y → Y continuous maps. Then:
1. h( f × g) = h( f ) + h(g); and
2. if g is a factor of f (or equivalently, f is an extension of g), then h( f ) ≥
h(g).

Proof. To prove part 1, note that the metric d((x, y), (x  , y )) =

max{d X(x, x  ), dY (y, y )} generates the product topology on X × Y, and

dn ((x, y), (x  , y )) = max dnX(x, x  ), dnY (y, y ) .
If U ⊂ X and V ⊂ Y have diameters less than , then U × V has d-diameters
less than . Hence
cov(n, , f × g) ≤ cov(n, , f ) · cov(n, , g),
so h( f × g) ≤ h( f ) + h(g). On the other hand, if A ⊂ X and B ⊂ Y are
(n, )-separated, then A× B is (n, )-separated for d. Hence
sep(n, , f × g) ≥ sep(n, , f ) · sep(n, , g),
so, by (2.7), h( f × g) ≥ h( f ) + h(g).
The proof of part 2 is left as an exercise (Exercise 2.5.5). 

PROPOSITION 2.5.7. Let (X, d) be a compact metric space, and f : X → X
an expansive homeomorphism with expansiveness constant δ. Then h( f ) =
h ( f ) for any  < δ.

Proof. Fix γ and  with 0 < γ <  < δ. We will show that h2γ ( f ) = h ( f ).
By monotonicity, it suffices to show that h2γ ( f ) ≤ h ( f ).
By expansiveness, for distinct points x and y, there is some i ∈ Z such
that d( f i (x), f i (y)) ≥ δ > . Since the set {(x, y) ∈ X × X: d(x, y) ≥ γ }
is compact, there is k = k(γ , ) ∈ N such that if d(x, y) ≥ γ , then
2.6. Topological Entropy for Some Examples 41

d( f i (x), f i (y)) >  for some |i| ≤ k. Thus if A is an (n, γ )-separated set,
then f −k(A) is (n + 2k, )-separated. Hence, by Lemma 2.5.1, h ( f ) ≥
h2γ ( f ). 

REMARK 2.5.8. The topological entropy of a continuous (semi)flow can be
defined as the entropy of the time-1 map. Alternatively, it can be defined using
the analog dT , T > 0, of the metrics dn . The two definitions are equivalent
because of the equicontinuity of the family of time-t maps, t ∈ [0, 1].
Exercise 2.5.1. Let (X, d) be a compact metric space. Show that the metrics
di all induce the same topology on X.
Exercise 2.5.2. Prove the remaining inequalities in Lemma 2.5.1.
Exercise 2.5.3. Let {an } be a subadditive sequence of non-negative real
numbers, i.e., 0 ≤ am+n ≤ am + an for all m, n ≥ 0. Show that limn→∞ an /n =
infn≥0 an /n.
Exercise 2.5.4. Show that the topological entropy of an isometry is zero.
Exercise 2.5.5. Let g: Y → Y be a factor of f : X → X. Prove that h( f ) ≥
h(g).
Exercise 2.5.6. Let Y and Z be compact metric spaces, X = Y × Z, and
π be the projection to Y. Suppose f : X → X is an isometric extension
of a continuous map g: Y → Y, i.e., π ◦ f = g ◦ π and d( f (x1 ), f (x2 )) =
d((x1 ), (x2 )) for all x1 , x2 ∈ Y with π (x1 ) = π(x2 ). Prove that h( f ) = h(g).
Exercise 2.5.7. Prove that the topological entropy of a continuously differ-
entiable map of a compact manifold is finite.

2.6 Topological Entropy for Some Examples

In this section, we compute the topological entropy for some of the examples
from Chapter 1.
PROPOSITION 2.6.1. Let à be a 2 × 2 integer matrix with determinant 1
and eigenvalues λ, λ−1 , with |λ| > 1; and let A: T2 → T2 be the associated
hyperbolic toral automorphism. Then h(A) = log |λ|.
Proof. The natural projection π : R2 → R2 /Z2 = T2 is a local homeomor-
phism, and π Ã = Aπ. Any metric d˜ on R2 invariant under integer transla-
tions induces a metric d on T2 , where d(x, y) is the d-distance
˜ between the
sets π −1 (x) and π −1 (y). For these metrics, π is a local isometry.
Let v1 , v2 be eigenvectors of A with (Euclidean) length 1 correspond-
ing to the eigenvalues λ, λ−1 . For x, y ∈ R2 , write x − y = a1 v1 + a2 v2 and
42 2. Topological Dynamics

˜ y) = max(|a1 |, |a2 |). This is a translation-invariant metric on R2 .

define d(x,
˜
A d-ball of radius  is a parallelogram whose sides are of (Euclidean) length
2 and parallel to v1 and v2 . In the metric d˜n (defined for Ã), a ball of radius
 is a parallelogram with side length 2|λ|−n in the v1 -direction and 2 in the
v2 -direction. In particular, the Euclidean area of a d˜n -ball of radius  is not
greater than 4 2 |λ|−n . Since the induced metric d on T2 is locally isometric
to d,˜ we conclude that for sufficiently small , the Euclidean area of a dn -ball
of radius  in T2 is at most 4 2 |λ|−n . It follows that the minimal number of
balls of dn -radius  needed to cover T2 is at least

area(T2 )/4 2 |λ|−n = |λ|n /4 2 .

Since a set of diameter  is contained in an open ball of radius , we conclude

that cov(n, , A) ≥ |λ|n /4 2 . Thus, h(A) ≥ log |λ|.
Conversely, since the closed d˜n -balls are parallelograms, there is a tiling
of the plane by -balls whose interiors are disjoint. The Euclidean area of
such a ball is C 2 |λ|−n , where C depends on the angle between v1 and v2 .
For small enough , any -ball that intersects the unit square [0, 1] × [0, 1]
is entirely contained in the larger square [−1, 2] × [−1, 2]. Therefore the
number of the balls that intersect the unit square does not exceed the area
of the larger square divided by the area of a d˜n -ball of radius . Thus, the
torus can be covered by 9λn /C 2 closed dn -balls of radius . It follows that
cov(n, 2, A) ≤ 9λn /C 2 , so h(A) ≤ log |λ|. 

To establish the corresponding result in higher dimensions, we need some
results from linear algebra. Let B be a k × k complex matrix. If λ is an
eigenvalue of B, let

Vλ = {v ∈ Ck: (B − λI)i v = 0 for some i ∈ N}.

If B is real and γ is a real eigenvalue, let

VγR = Rk ∩ Vγ = {v ∈ Rk: (B − γ I)i v = 0 for some i ∈ N}.

If B is real and λ, λ̄ is a pair of complex eigenvalues, let

Vλ,Rλ̄ = Rk ∩ (Vλ ⊕ Vλ̄ ).

These spaces are called generalized eigenspaces.

LEMMA 2.6.2. Let B be a k × k complex matrix, and λ be an eigenvalue of

B. Then for every δ > 0 there is C(δ) > 0 such that

C(δ)−1 (|λ| − δ)n v ≤ Bn v ≤ C(δ)(|λ| + δ)n v

for every n ∈ N and every v ∈ Vλ .

2.6. Topological Entropy for Some Examples 43

Proof. It suffices to prove the lemma for a Jordan block. Thus without loss
of generality, we assume that B has λs on the diagonal, ones above and zeros
elsewhere. In this setting, Vλ = Ck and in the standard basis e1 , . . . , ek, we
have Be1 = λe1 and Bei = λei + ei−1 , for i = 2, . . . , k. For δ > 0, consider the
basis e1 , δe2 , δ 2 e3 , . . . , δ k−1 ek. In this basis, the linear map B is represented by
the matrix
 
λ δ
 λ δ 
 
 . . 
Bδ =  .. .. .
 
 λ δ
λ
Observe that Bδ = λI + δ A with A ≤ 1, where A = supv=0Av|/v.
Therefore
 
(|λ| − δ)n v ≤  Bδn v  ≤ (|λ| + δ)n v.
Since Bδ is conjugate to B, there is a constant C(δ) > 0 that bounds the
distortion of the change of basis. 

LEMMA 2.6.3. Let B be a k × k real matrix and λ an eigenvalue of B. Then
for every δ > 0 there is C(δ) > 0 such that
C(δ)−1 (|λ| − δ)n v ≤ Bn v ≤ C(δ)(|λ| + δ)n v
for every n ∈ N and every v ∈ Vλ (if λ ∈ R) or every v ∈ Vλ,λ̄ (if λ ∈
/ R).
Proof. If λ is real, then the result follows from Lemma 2.6.2. If λ is complex,
then the estimates for Vλ and Vλ̄ from Lemma 2.6.2 imply a similar estimate
for Vλ, λ̄ , with a new constant C(δ) depending on the angle between Vλ and
Vλ̄ and the constants in the estimates for Vλ and Vλ̄ (since |λ| = |λ̄|). 

PROPOSITION 2.6.4. Let à be a k × k integer matrix with determinant 1
and with eigenvalues α1 , . . . , αk, where
|α1 | ≥ |α2 | ≥ · · · ≥ |αm| > 1 > |αm+1 | ≥ · · · ≥ |αk|.
Let A: T → Tk be the associated hyperbolic toral automorphism. Then
k

m
h(A) = log |αi |.
i=1

Proof. Let γ1 , . . . , γ j be the distinct real eigenvalues of Ã, and λ1 , λ1 , . . . ,

λm, λm be the distinct complex eigenvalues of Ã. Then

j 
m
Rk = Vγi ⊕ Vλi ,λi ,
i=1 i=1
44 2. Topological Dynamics

any vector v ∈ Rk can be decomposed uniquely as v = v1 + · · · + v j+m with vi

in the corresponding generalized eigenspace. Given x, y ∈ Rk, let v = x − y,
˜ y) = max(|v1 |, . . . , |v j+m|). This is a translation-invariant
and define d(x,
metric on Rk, and therefore descends to a metric on Tk. Now, using
Lemma 2.6.3, the proposition follows by an argument similar to the one
in the proof of Proposition 2.6.1 (Exercise 2.6.3). 

The next example we consider is the solenoid from §1.9.

PROPOSITION 2.6.5. The topological entropy of the solenoid map F: S → S

is log 2.
Proof. Recall from §1.9 that F is topologically conjugate to the automor-
phism α: → , where
 ∞

= (φi )i=0 : φi ∈ [0, 1), φi = 2φi+1 mod 1 ,

and α is coordinatewise multiplication by 2 (mod 1). Thus, h(F) = h(α). Let

|x − y| denote the distance on S1 = [0, 1] mod 1. The distance function
∞
1
d(φ, φ  ) = |φ − φn |
n n
n=0
2

generates the topology in introduced in §1.9.

∞
The map π: → S1 , (φi )i=0 → φ0 , is a semiconjugacy from α to E2 . Hence,
h(α) ≥ h(E2 ) = log 2 (Exercise 2.6.1). We will establish the inequality h(α) ≤
log 2 by constructing an (n, )-spanning set.
Fix  > 0 and choose k ∈ N such that 2−k < /2. For n ∈ N, let An ⊂ con-
sist of the 2n+2k sequences ψ j = (ψi ), where ψi = j · 2−(n+k+i) mod 1, j =
j j

0, . . . , 2n+2k − 1. We claim that An is (n, )-spanning. Let φ = (φi ) be a point

in . Choose j ∈ {0, . . . , 2n+2k − 1} so that |φk − j · 2−(n+2k) | ≤ 2−(n+2k+1) .
Then |φi − ψi | ≤ 2k−i 2−(n+2k+1) , for 0 ≤ i ≤ k. It follows that for 0 ≤ m ≤ n,
j


∞  m j  j
2 φi − 2mψi  k
2mφi − ψi  1
d(α φ, α ψ ) =
m m j
i
< i
+ k
i=0
2 i=0
2 2

k
2k−i 2−(n+2k+1) 1 1
< 2m + < k−1 < .
i=0
2i 2k 2

Thus dn (φ, ψ j ) < , so An is (n, )-spanning. Hence,

1
h(α) ≤ lim log cardAn = log 2.
n→∞ n 

2.7. Equicontinuity, Distality, and Proximality 45

Note that α: → is expansive with expansiveness constant 1/3

(Exercise 2.6.4), so by Proposition 2.5.7, h (α) = h(α) for any  < 1/3.

Exercise 2.6.1. Compute the topological entropy of an expanding endo-

morphism Em: S1 → S1 .

Exercise 2.6.2. Compute the topological entropy of the full one- and two-
sided m-shifts.

Exercise 2.6.3. Finish the proof of Proposition 2.6.4.

Exercise 2.6.4. Prove that the solenoid map (§1.9) is expansive.

2.7 Equicontinuity, Distality, and Proximality1

In this section, we describe a number of properties related to the asymptotic
behavior of the distance between corresponding points on pairs of orbits.
Let f : X → X be a homeomorphism of a compact Hausdorff space. Points
x, y ∈ X are called proximal if the closure O((x, y)) of the orbit of (x, y)
under f × f intersects the diagonal  = {(z, z) ∈ X × X: z ∈ X}. Every point
is proximal to itself. If two points x and y are not proximal, i.e., if O((x, y)) ∩
 = ∅, they are called distal. A homeomorphism f : X → X is distal if
every pair of distinct points x, y ∈ X is distal. If (X, d) is a compact met-
ric space, then x, y ∈ X are proximal if there is a sequence nk ∈ Z such that
d( f nk (x), f nk (y)) → 0 as k → ∞; points x, y ∈ X are distal if there is  > 0
such that d( f n (x), f n (y)) >  for all n ∈ Z (Exercise 2.7.2)
A homeomorphism f of a compact metric space (X, d) is said to be
equicontinuous if the family of all iterates of f is an equicontinuous fam-
ily, i.e., for any  > 0, there exists δ > 0 such that d(x, y) < δ implies that
d( f n (x), f n (y)) <  for all n ∈ Z. An isometry preserves distances and is
therefore equicontinuous. Equicontinuous maps share many of the dynam-
ical properties of isometries. The only examples from Chapter 1 that are
equicontinuous are the group translations, including circle rotations.
We denote by f × f the induced action of f in X × X, defined by
f × f (x, y) = ( f (x), f (y)).

PROPOSITION 2.7.1. An expansive homeomorphism of an infinite compact

metric space is not distal.
Proof. Exercise 2.7.1. 


1 Several arguments in this section were conveyed to us by J. Auslander.

46 2. Topological Dynamics

PROPOSITION 2.7.2. Equicontinuous homeomorphisms are distal.

Proof. Suppose the equicontinuous homeomorphism f : X → X is not dis-
tal. Then there is a pair of proximal points x, y ∈ X, so d( f nk (x), f nk (y)) → 0
for some sequence nk ∈ Z. Let xk = f nk (x) and yk = f nk (y). Let  = d(x, y).
Then for any δ > 0, there is some k ∈ N such that d(xk, yk) < δ, but
d( f −nk (xk), f −nk (yk)) = , so f is not equicontinuous. 

Distal homeomorphisms are not necessarily equicontinuous. Consider the
map F: T2 → T2 defined by

x →x + α mod 1,
y →x + y mod 1.

We view T2 as the unit square with opposite sides identified and use the
metric inherited from the Euclidean metric. To see that this map is distal, let
(x, y), (x  , y ) be distinct points in T2 . If x = x  , then d(F n (x, y), F n (x  , y ))
is at least d((x, 0), (x , 0)), which is constant. If x = x  , then d(F n (x, y),


F n (x  , y )) = d((x, y), (x  , y )). Therefore, the pair (x, y), (x  , y ) is distal. To

see that F is not equicontinuous, let p = (0, 0) and q = (δ, 0). Then for all
n, the difference between the first coordinates of F n ( p) and F n (q) is δ. The
difference between the second coordinates of F n ( p) and F n (q) is nδ as long
as nδ < 1/2. Therefore there are points that are arbitrarily close together
that are moved at least 1/4 apart, so F is not equicontinuous.
The preceding map is an example of a distal extension. Suppose a home-
omorphism g: Y → Y is an extension of a homeomorphism f : X → X with
projection π: Y → X. We say that the extension is distal if any pair of dis-
tinct points y, y ∈ Y with π (y) = π(y ) is distal. The map F: T2 → T2 in the
preceding paragraph is a distal extension of a circle rotation, with projection
on the first factor as the factor map. A straightforward generalization of the
argument in the previous paragraph shows that a distal extension of a distal
homeomorphism is distal. Moreover, as we show later in this section, any
factor of a distal map is distal. Thus, (X1 , f1 ) and (X2 , f2 ) are distal if and
only if (X1 × X2 , f1 × f2 ) is distal.
Similarly, π: Y → X is an isometric extension if d(g(y), g(y )) = d(y, y )
whenever π(y) = π(y ). The extension π : Y → X is an equicontinuous ex-
tension if for any  > 0, there exists δ > 0 such that if π (y) = π(y ) and
d(y, y ) < δ, then d(g n (y), g n (y )) < , for all n. An isometric extension
is an equicontinuous extension; an equicontinuous extension is a distal
extension.
To prove Theorem 2.7.4, we need the following notion: For a subset A ⊂ X
and a homeomorphism f : X → X, denote by f A the induced action of f in
2.7. Equicontinuity, Distality, and Proximality 47

the product space XA (an element zof XA is a function z: A → X, and f A(z) =

f ◦ z). We say that A ⊂ X is almost periodic if every z ∈ XA with range A
is an almost periodic point of (XA, f A ). That is, A is almost periodic if for
every finite subset a1 , . . . , an ∈ A, and neighborhoods U1  a1 , . . . , Un  an ,
the set {k ∈ Z: f k(ai ) ∈ Ui , 1 ≤ i ≤ n} is syndetic in Z. Every subset of an
almost periodic set is an almost periodic set. Note that if x is an almost
periodic point of f , then {x} is an almost periodic set.

LEMMA 2.7.3. Every almost periodic set is contained in a maximal almost

periodic set.
Proof. Let A be an almost periodic set, and C be a collection, totally ordered

by inclusion, of almost periodic sets containing A. The set C∈C C is an
almost periodic set and a maximal element of C. By Zorn’s lemma there is
a maximal almost periodic set containing A. 

THEOREM 2.7.4. Let f be a homeomorphism of a compact Hausdorff space
X. Then every x ∈ X is proximal to an almost periodic point.
Proof. If x is an almost periodic point, then we are done, since x is proximal
to itself. Suppose x is not almost periodic, and let A be a maximal almost
periodic set. By definition, x ∈ / A. Let z ∈ XA have range A, and consider
(x, z) ∈ (X × XA ). Let (x0 , z0 ) be an almost periodic point (of ( f × f A )) in
O(x, z). Since z is almost periodic, z ∈ O(z0 ). Hence there is x  ∈ X such that
(x  , z) is almost periodic and (x  , z) ∈ O(x, z) (Proposition 2.1.1). Therefore
{x  } ∪ range(z) = {x  } ∪ Ais an almost periodic set. Since A is maximal, x  ∈
A, i.e., x  appears as one of the coordinates of z. It follows that (x  , x  ) ∈
O(x, x  ), and x is proximal to x  . 

A homeomorphism f of a compact Hausdorff space X is called pointwise
almost periodic if every point is almost periodic. By Proposition 2.1.3, this
happens if and only if X is a union of minimal sets.

PROPOSITION 2.7.5. Let f be a distal homeomorphism of a compact

Hausdorff space X. Then f is pointwise almost periodic.
Proof. Let x ∈ X. Then, by Theorem 2.7.4, x is proximal to an almost peri-
odic y ∈ X. Since f is distal, x = y and x is almost periodic. 

PROPOSITION 2.7.6. A homeomorphism of a compact Hausdorff space is
distal if and only if the product system (X × X, f × f ) is pointwise almost
periodic.
Proof. If f is distal, so is f × f , and hence f × f is pointwise almost
periodic. Conversely, assume that f × f is pointwise almost periodic, and
let x, y ∈ X be distinct points. If x and y are proximal, then there is z with
48 2. Topological Dynamics

(z, z) ∈ O(x, y). Recall that O(x, y) is minimal (Proposition 2.1.3). Since
(x, y) ∈/ O(z, z), we obtain a contradiction. 

COROLLARY 2.7.7. A factor of a distal homeomorphism f of a compact
Hausdorff space X is distal.
Proof. Let g: Y → Y be a factor of f . Then f × f is pointwise almost pe-
riodic by Proposition 2.7.6. Since (g × g) is a factor of f × f , it is pointwise
almost periodic (Exercise 2.7.5), and hence is distal. 

The class of distal dynamical systems is of special interest because it is
closed under factors and isometric extensions. The class of minimal distal sys-
tems is the smallest such class of minimal systems: According to Furstenberg’s
structure theorem [Fur63], every minimal distal homeomorphism (or flow)
can be obtained by a (possibly transfinite) sequence of isometric extensions
starting with the one-point dynamical system.

Exercise 2.7.1. Prove Proposition 2.7.1.

Exercise 2.7.2. Prove the equivalence of the topological and metric defini-
tions of distal and proximal points at the beginning of this section.

Exercise 2.7.3. Give an example of a homeomorphism f of a compact

metric space (X, d) such that d( f n (x), f n (y)) → 0 as n → ∞ for every pair
x, y ∈ X.

Exercise 2.7.4. Show that any infinite closed shift-invariant subset of m

contains a proximal pair of points.

Exercise 2.7.5. Prove that a factor of a pointwise almost periodic system is

pointwise almost periodic.

2.8 Applications of Topological Recurrence to Ramsey Theory2

In this section, we establish several Ramsey-type results to illustrate how
topological dynamics is applied in combinatorial number theory. One of the
main principles of the Ramsey theory is that a sufficiently rich structure is
indestructible by finite partitioning (see [Ber96] for more information on
Ramsey theory). An example of such a statement is van der Waerden’s
theorem, which we prove later in this section. We conclude this section by

2 The exposition in this section follows, to a large extent, [Ber00].

2.8. Applications of Topological Recurrence to Ramsey Theory 49

proving a result in Ramsey theory about infinite-dimensional vector spaces

over finite fields.
m
THEOREM 2.8.1 (van der Waerden). For each finite partition Z = k=1 Sk,
one of the sets Sk contains arbitrarily long (finite) arithmetic progressions.

We will obtain van der Waerden’s theorem as a consequence of a general

recurrence property in topological dynamics.
Recall from §1.4 that m = {1, 2, . . . , m}Z with metric d(ω, ω ) = 2−k,
where k = min{|i|: ωi = ωi }, is a compact metric space. The shift σ : m →

m, (σ ω)i = ωi+1 , is a homeomorphism. A finite partition Z = m k=1 Sk can
 ∞
be viewed as a sequence ξ ∈ m for which ξi = k if i ∈ Sk. Let X = i=−∞ σiξ
be the orbit closure of ξ under σ , and let Ak = {ω ∈ X: ω0 = k}. If ω ∈
Ak, ω ∈ X, and d(ω , ω) < 1, then ω ∈ Ak. Hence if there are integers p, q ∈
N and ω ∈ X such that d(σ i p ω, ω) < 1 for 0 ≤ i ≤ q − 1, then there is r ∈ Z
such that ξ j = ω0 for i = r, r + p, . . . , r + (q − 1) p. Therefore, Theorem 2.8.1
follows from the following multiple recurrence property (Exercise 2.8.1).

PROPOSITION 2.8.2. Let T be a homeomorphism of a compact metric space

X. Then for every  > 0 and q ∈ N there are p ∈ N and x ∈ X such that
d(T j p (x), x) <  for 0 ≤ j ≤ q.

We will obtain Proposition 2.8.2 as a consequence of a more general state-

ment (Theorem 2.8.3), which has other corollaries useful in combinatorial
number theory.
Let F be the collection of all finite non-empty subsets of N. For α, β ∈ F,
we write α < β if each element of α is less than each element of β. For a
commutative group G, a map T: F → G, α →Tα , defines an IP-system in G
if

T{i1 ,...,ik} = T{i1 } · . . . · T{ik}

for every {i 1 , . . . , i k} ∈ F; in particular, if α, β ∈ F and α ∩ β = ∅, then Tα∪β =

Tα Tβ . Every IP-system T is generated by the elements T{n} ∈ G, n ∈ N.
Let G be a group of homeomorphisms of a topological space X. For x ∈ X,
denote by Gx the orbit of x under G. We say that G acts minimally on X if
for each x ∈ X, the orbit Gx is dense in X.

THEOREM 2.8.3 (Furstenberg–Weiss [FW78]). Let G be a commutative

group acting minimally on a compact topological space X. Then for every
non-empty open set U ⊂ X, every n ∈ N, every α ∈ F, and any IP-systems
50 2. Topological Dynamics

T (1) , . . . , T (n) in G, there is β ∈ F such that α < β and

(1) (n)
U ∩ Tβ (U) ∩ · · · ∩ Tβ (U) =. ∅

Proof [Ber00]. Since G acts minimally, and X is compact, there are ele-
m
ments g1 , . . . , gm ∈ G such that i=1 gi (U) = X (Exercise 2.8.2).
We argue by induction on n. For n = 1, let T be an IP-system and U ⊂ X
be open and not empty. Set V0 = U. Define recursively Vk = T{k} (Vk−1 ) ∩
gik (U), where i k is chosen so that 1 ≤ i k ≤ m and T{k} (Vk−1 ) ∩ gik (U) =. ∅ By
−1
construction, T{k} (Vk) ⊂ Vk−1 and Vk ⊂ gik (U). In particular, by the pigeon-
hole principle, there are 1 ≤ i ≤ m and arbitrarily large p < q such that
Vp ∪ Vq ⊂ gi (U). Choose p so that β = { p + 1, p + 2, . . . , q} > α. Then the
set W = gi−1 (Vq ) ⊂ U is not empty, and
   −1 
Tβ−1 (W) = gi−1 T{−1 −1
p+1} · · · T{q} (Vq ) ⊂ gi
−1
T{ p+1} (Vp+1 ) ⊂ gi−1 (Vp ) ⊂ U.

Therefore, U ∩ Tβ (U) ⊃ W =. ∅
Assume that the theorem holds true for any n IP-systems in G. Let U ⊂ X
be open and not empty. Let T (1) , . . . , T (n+1) be IP-systems in G. We will
construct a sequence of non-empty open subsets Vk ⊂ X and an increasing
 ( j) −1
sequence αk ∈ F, αk > α, such that V0 = U, n+1 j=1 (Tαk ) (Vk) ⊂ Vk−1 , and
Vk ⊂ gik (U) for some 1 ≤ i k ≤ m.
By the inductive assumption applied to V0 = U and the n IP-systems
(T (n+1) )−1 T ( j) , j = 1, . . . , n, there is α1 > α such that
 −1 (1)  −1 (n)
V0 ∩ Tα(n+1)1
Tα1 (V0 ) ∩ · · · ∩ Tα(n+1)
1
Tα1 (V0 ) =. ∅
(n+1)
Apply Tα1 and, for an appropriate 1 ≤ i 1 ≤ m, set

V1 = gi1 (V0 ) ∩ Tα(1)

1
(V0 ) ∩ Tα(2)
1
(V0 ) ∩ · · · ∩ Tα(n+1)
1
(V0 ) =. ∅

If Vk−1 and αk−1 have been constructed, apply the inductive assumption to
Vk−1 and the IP-systems (T (n+1) )−1 T ( j) , j = 1, . . . , n, to get αk > αk−1 such
that
 −1 (1)  −1 (n)
Vk−1 ∩ Tα(n+1)
k
Tαk (Vk−1 ) ∩ · · · ∩ Tα(n+1)
k
Tαk (Vk−1 ) =. ∅
(n+1)
Apply Tαk and, for an appropriate 1 ≤ i k ≤ m, set

Vk = gik (V0 ) ∩ Tα(1)

k
(Vk−1 ) ∩ Tα(2)
k
(Vk−1 ) ∩ · · · ∩ Tα(n+1)
k
(Vk−1 ) =. ∅

By construction, the sequences αk and Vk have the desired properties. Since

Vk ⊂ gik (U), there is 1 ≤ i ≤ m such that Vk ⊂ gi (U) for infinitely many
k’s. Hence there are arbitrarily large p < q such that Vp ∪ Vq ⊂ gi (U). Let
2.8. Applications of Topological Recurrence to Ramsey Theory 51

W = gi−1 (Vq ) ⊂ U and β = α p+1 ∪ · · · ∪ αq . Then W = ∅, and for each

1 ≤ j ≤ n + 1,
 ( j) −1  j) −1
Tβ (W) = gi−1 Tα(s+1 (Vq )
 −1
⊂ gi−1 Tα(s+1
j)
(Vq−1 ) ⊂ · · · ⊂ gi−1 (Vp ) ⊂ U.
n+1 n+1
Therefore
( j) −1
j=1 (Tβ ) W ⊂ U, and hence
( j) −1
j=1 (Tβ ) U = ∅. 

COROLLARY 2.8.4. Let G be a commutative group of homeomorphisms
of a compact metric space X and let T (1) , . . . , T (n) be IP-systems in G. Then
for every α ∈ F and every  > 0 there are x ∈ X and β > α such that
(i)
d(x, Tβ (x)) <  for each 1 ≤ i ≤ n.
Proof. Similarly to Proposition 2.1.2, there is a non-empty closed G-
invariant subset X ⊂ X on which G acts minimally (Exercise 2.8.3). Thus
the corollary follows from Theorem 2.8.3. 

Proof of Proposition 2.8.2. Let G = {T k}k∈Z . For α ∈ F, denote by |α| the
sum of the elements in α. Apply Corollary 2.8.4 to G, X, and the IP-systems
T (αj) = T j|α| , 1 ≤ j ≤ q − 1. 

The following generalization of Theorem 2.8.1 also follows from Corol-
lary 2.8.4.

THEOREM 2.8.5. Let d ∈ N, and let A be a finite subset of Zd . Then for each

finite partition Zd = mk=1 Sk, there are k ∈ {1, . . . , m}, z0 ∈ Z , and n ∈ N such
d

that z0 + na ∈ Sk for each a ∈ A, i.e., z0 + nA ⊂ Sk.

Proof. Exercise 2.8.5. 

Let VF be an infinite vector space over a finite field F. A subset A ⊂ VF is
a d-dimensional affine subspace if there are v ∈ VF and linearly independent
x1 , . . . , xd ∈ VF such that A = v + Span(x1 , . . . , xd ).
m
THEOREM 2.8.6 [GLR72], [GLR73]. For each finite partition VF = k=1 Sk,
one of the sets Sk contains affine subspaces of arbitrarily large (finite) dimen-
sion.
Proof ([Ber00]; see Theorem 2.8.3). We say that a subset L ⊂ VF is mono-
chromatic of color j if L ⊂ S j .
Since VF is infinite, it contains a countable subspace isomorphic to the
abelian group
 ∞

F∞ = a = (ai )i=1 ∈ F N : ai = 0 for all but finitely many i ∈ N .
52 2. Topological Dynamics

Without loss of generality we assume that VF = F∞ . The set  = {1, . . . , m} F∞

of all functions F∞ → {1, . . . , m} is naturally identified with the set of all par-
titions of F∞ into m subsets. The discrete topology on {1, . . . , m} and product
topology on  make it a compact Hausdorff space.

Let ξ ∈  correspond to a partition F∞ = m k=1 Sk, i.e., ξ : F∞ →
{1, . . . , m}, ξ (a) = k if and only if a ∈ Sk. Each b ∈ F∞ induces a homeomor-
phism Tb :  → , (Tb η)(a) = η(a + b). Denote by X ⊂  the orbit closure

of ξ, X = { b∈F∞ Tb ξ }. Similarly to the argument in the proof of Proposi-
tion 2.1.2, Zorn’s lemma implies that there is a non-empty closed subset
X ⊂ X on which the group F∞ acts minimally.
Let g: F → F∞ be an IP-system in F∞ such that the elements gn , n ∈ N,
are linearly independent. Define an IP-system T of homeomorphisms of
(f )
X by setting Tα = Tgα . For each f ∈ F, set Tα = Tf gα to get |F| = card F
IP-systems of commuting homeomorphisms of X. Let 0 = (0, 0, . . .) be the
zero element of F∞ and Ai = {η ∈ : η(0) = i}. Then each Ai is open and
m 
i=1 Ai = . Therefore, there is j ∈ {1, . . . , m} such that U = Aj ∩ X =. ∅
 (f )
By Theorem 2.8.3, there is β1 ∈ F such that U1 = f ∈F Tβ1 (U) =. ∅ If η ∈
U1 , then η( f gβ1 ) = j for each f ∈ F. In other words, η contains a monochro-
matic affine line of color j. Since the orbit of ξ is dense in X , there is
b1 ∈ F∞ such that ξ ( f gβ1 + b1 ) = η( f gβ1 ) = j. Thus, S j contains an affine
line.
To obtain a two-dimensional affine subspace in S j apply Theorem 2.8.3 to
U1 , β1 and the same collection of IP-systems to get β2 > β1 such that U2 =
 (f )
f ∈F Tβ2 (U1 ) =. ∅ Since gβ2 is linearly independent with every gα , α < β2 ,
each η ∈ U2 contains a monochromatic two-dimensional affine subspace
of color j. Since η can be arbitrarily approximated by the shifts of ξ , the
latter also contains a monochromatic two-dimensional affine subspace of
color j.
Proceeding in this manner, we obtain a monochromatic subspace of
arbitrarily large dimension. 

Exercise 2.8.1. Prove Theorem 2.8.1 using Proposition 2.8.2.

Exercise 2.8.2. Prove that a group G acts minimally on a compact topo-

logical space X if and only if for every non-empty open set U ⊂ X there are
n
elements g1 , . . . , gn ∈ G such that i=1 gi (U) = X.

Exercise 2.8.3. Prove the following generalization of Proposition 2.1.2. If a

group G acts by homeomorphisms on a compact metric space X, then there
is a non-empty closed G-invariant subset X on which G acts minimally.
2.8. Applications of Topological Recurrence to Ramsey Theory 53

Exercise 2.8.4. Prove that van der Waerden’s Theorem 2.8.1 is equivalent
to the following finite version: For each m, n ∈ N there is k ∈ N such that if
the set {1, 2, . . . , k} is partitioned into m subsets, then one of them contains
an arithmetic progression of length n.
*Exercise 2.8.5. For z ∈ Zd , the translation by z in Zd induces a homeomor-
phism (shift) Tz in  = {1, . . . , m}Z . Prove Theorem 2.8.5 by considering the
d

orbit closure under the group of shifts of the element ξ ∈  corresponding

to the partition of Zd and the IP-systems in Zd generated by the translations
Tf , f ∈ A.
CHAPTER THREE

Symbolic Dynamics

+
In §1.4, we introduced the symbolic dynamical systems (
m, σ ) and (
m , σ ),
and we showed by example throughout Chapter 1 how these shift spaces arise
naturally in the study of other dynamical systems. In all of those examples,
we encoded an orbit of the dynamical system by its itinerary through a finite
collection of disjoint subsets. Specifically, following an idea that goes back
to J. Hadamard, suppose f : X → X is a discrete dynamical system. Con-
sider a partition P = {P1 , P2 , . . . , Pm} of X, i.e., P1 ∪ P2 ∪ · · · ∪ Pm = X and
Pi ∩ Pj = ∅ for i
= j. For each x ∈ X, let ψi (x) be the index of the element
of P containing f (x). The sequence (ψi (x))i∈N0 is called the itinerary of x.
i

This defines a map

+
ψ: X →
m = {1, 2, . . . , m}N0 , x → {ψ ∞
i (x)}i=0 ,

+
which satisfies ψ ◦ f = σ ◦ ψ. The space
m is totally disconnected, and the
map ψ usually is not continuous. If f is invertible, then positive and negative
iterates of f define a similar map X →
m = {1, 2, . . . , m}Z . The image of
+
ψ in
m or
m is shift-invariant, and ψ semiconjugates f to the shift on
the image of ψ. The indices ψi (x) are symbols – hence the name symbolic
dynamics. Any finite set can serve as the symbol set, or alphabet, of a symbolic
dynamical system. Throughout this chapter, we identify every finite alphabet
with {1, 2, . . . , m}.
Recall that the cylinder sets
,...,nk

C nj11,..., jk = ω = (ωl ) : ωni = ji , i = 1, . . . , k ,

+
form a basis for the product topology of
m and
m , and that the metric

d(ω, ω ) = 2−l , ωi }

where l = min{|i|: ωi
=

generates the product topology.

54
3.1. Subshifts and Codes 55

3.1 Subshifts and Codes1

In this section, we concentrate on two-sided shifts. The case of one-sided
shifts is similar.
A subshift is a closed subset X ⊂
m invariant under the shift σ and its
inverse. We refer to
m as the full m-shift.
Let Xi ⊂
mi , i = 1, 2, be two subshifts. A continuous map c: X1 → X2
is a code if it commutes with the shifts, i.e., σ ◦ c = c ◦ σ (here and later, σ
denotes the shift in any sequence space). Note that a surjective code is a
factor map. An injective code is called an embedding; a bijective code gives
a topological conjugacy of the subshifts and is called an isomorphism (since

m is compact, a bijective code is a homeomorphism).
For a subshift X ⊂
m, denote by Wn (X) the set of words of length n
that occur in X, and by |Wn (X)| its cardinality. Since different elements of
X differ in at least one position, the restriction σ | X is expansive. Therefore,
Proposition 2.5.7 allows us to compute the topological entropy of σ | X through
the asymptotic growth rate of |Wn (X)|.

PROPOSITION 3.1.1. Let X ⊂
m be a subshift. Then

1
h(σ | X) = lim log |Wn (X)|.
n→∞ n
Proof. Exercise 3.1.1. 

Let X be a subshift, k, l ∈ N0 , n = k + l + 1, and let α be a map from
Wn (X) to an alphabet Am  . The (k, l) block code cα from X to the full shift

m  assigns to a sequence x = (xi ) ∈ X the sequence cα (x) with cα (x)i =
α(xi−k, . . . , xi , . . . , xi+l ). Any block code is a code, since it is continuous and
commutes with the shift.

PROPOSITION 3.1.2 (Curtis–Lyndon–Hedlund). Every code c: X → Y is a

block code.
Proof. Let A be the symbol set of Y, and define α̃: X → A by α̃(x) = c(x)0 .
Since X is compact, α̃ is uniformly continuous, so there is a δ > 0 such that
α̃(x) = α̃(x  ) whenever d(x, x  ) < δ. Choose k ∈ N so that 2−k < δ. Then α̃(x)
depends only on x−k, . . . , x0 , . . . , xk, and therefore defines a map α: W2k+1 →
A satisfying c(x)0 = α(x−k . . . x0 . . . xk). Since c commutes with the shift, we
conclude that c = cα . 


1 The exposition of this section as well as §3.2, §3.4, and §3.5 follows in part the lectures of
M. Boyle [Boy93].
56 3. Symbolic Dynamics

There is a canonical class of block codes obtained by taking the alphabet

of the target shift to be the set of words of length n in the original shift.
Specifically, let k, l ∈ N, l < k, and let X be a subshift. For x ∈ X set
c(x)i = xi−k+l+1 . . . xi . . . xi+l , i ∈ Z.
This defines a block code c from X to the full shift on the alphabet Wk(X)
which is an isomorphism onto its image (Exercise 3.1.2). Such a code (or
sometimes its image) is called a higher block presentation of X.
Exercise 3.1.1. Prove Proposition 3.1.1.
Exercise 3.1.2. Prove that a higher block presentation of X is an isomor-
phism.
Exercise 3.1.3. Use a higher block presentation to prove that for any block
code c: X → Y, there is a subshift Z and an isomorphism f : Z → X such that
c ◦ f : Z → Y is a (0, 0) block code.
Exercise 3.1.4. Show that the full shift has points whose full orbit is dense
but whose forward orbit is nowhere dense.

3.2 Subshifts of Finite Type

The complement of a subshift X ⊂
m is open and hence is a union of at
most countably many cylinders. By shift invariance, if C is a cylinder and
C ⊂
m\X, then σ n (C) ⊂
m\X for all n ∈ Z, i.e., there is a countable list
of forbidden words such that no sequence in X contains a forbidden word
and each sequence in
m\X contains at least one forbidden word. If there
is a finite list of finite words such that X consists of precisely the sequences
in
m that do not contain any of these words, then X is called a subshift of
finite type (SFT); X is a k-step SFT if it is defined by a set of words of length
at most k + 1. A 1-step SFT is called a topological Markov chain.
In §1.4 we introduced a vertex shift
vA determined by an adjacency matrix
A of zeros and ones. A vertex shift is an example of an SFT. The forbidden
words have length 2 and are precisely those that are not allowed by A, i.e.,
a word uv is forbidden if there is no edge from u to v in the graph  A
determined by A. Since the list of forbidden words is finite,
vA is an SFT.
A sequence in
vA can be viewed as an infinite path in the directed graph
 A, labeled by the vertices.
An infinite path in the graph  A can also be specified by a sequence
of edges (rather than vertices). This gives a subshift
eA whose alphabet is
the set of edges in  A. More generally, a finite directed graph , possibly
3.3. The Perron–Frobenius Theorem 57

with multiple directed edges connecting pairs of vertices, corresponds to

an adjacency matrix B whose i, jth entry is a non-negative integer specify-
ing the number of directed edges in  =  B from the ith vertex to the jth
vertex. The set
eB of infinite directed paths in  B, labeled by the edges, is
closed and shift-invariant and is called the edge shift determined by B. Any
edge shift is a subshift of finite type (Exercise 3.2.3).
For any matrix A of zeros and ones, the map uv →e, where e is the edge
from u to v, defines a 2-block isomorphism from
vA to
eA. Conversely, any
edge shift is naturally isomorphic to a vertex shift (Exercise 3.2.4).
PROPOSITION 3.2.1. Every SFT is isomorphic to a vertex shift.
Proof. Let X be a k-step SFT with k > 0. Let Wk(X) be the set of words of
length k that occur in X. Let  be the directed graph whose set of vertices
is Wk(X); a vertex x1 . . . xk is connected to a vertex x1 . . . xk by a directed
edge if x1 . . . xk xk = x1 x1 . . . xk ∈ Wk+1 (X). Let A be the adjacency matrix of
. The code c(x)i = xi . . . xi+k−1 gives an isomorphism from X to
vA. 

COROLLARY 3.2.2. Every SFT is isomorphic to an edge shift.

The last proposition implies that “the future is independent of the past” in
an SFT; i.e., with appropriate one-step coding, if the sequences . . . x−2 x−1 x0
and x0 x1 x2 . . . are allowed, then . . . x−2 x−1 x0 x1 x2 . . . is allowed.
Exercise 3.2.1. Show that the collection of all isomorphism classes of sub-
shifts of finite type is countable.
∗
Exercise 3.2.2. Show that the collection of all subshifts of
2 is uncount-
able.
Exercise 3.2.3. Show that every edge shift is an SFT.
Exercise 3.2.4. Show that every edge shift is naturally isomorphic to a ver-
tex shift. What are the vertices?

3.3 The Perron–Frobenius Theorem

The Perron–Frobenius Theorem guarantees the existence of special invari-
ant measures, called Markov measures, for subshifts of finite type.
A vector or matrix all of whose coordinates are positive (non-negative)
is called positive (non-negative). Let A be a square non-negative matrix. If
for any i, j there is n ∈ N such that (An )i j > 0, then A is called irreducible;
otherwise A is called reducible. If some power of A is positive, A is called
primitive.
58 3. Symbolic Dynamics

An integer non-negative square matrix A is primitive if and only if the

directed graph  A has the property that there is n ∈ N such that, for every
pair of vertices u and v, there is a directed path from u to v of length n (see
Exercise 1.4.2). An integer non-negative square matrix A is irreducible if
and only if the directed graph  A has the property that, for every pair of
vertices u and v, there is a directed path from u to v (see Exercise 1.4.2).
A real non-negative m × m matrix is stochastic if the sum of the entries
in each row is 1 or, equivalently, the column vector with all entries 1 is an
eigenvector with eigenvalue 1.

THEOREM 3.3.1 (Perron). Let A be a primitive m × m matrix. Then A has

a positive eigenvalue λ with the following properties:
1. λ is a simple root of the characteristic polynomial of A,
2. λ has a positive eigenvector v,
3. any other eigenvalue of A has modulus strictly less than λ,
4. any non-negative eigenvector of A is a positive multiple of v.
Proof. Denote by int(W) the interior of a set W. We will need the following
lemma.

LEMMA 3.3.2. Let L: Rk → Rk be a linear operator, and assume that there

is a non-empty compact set P such that 0 ∈ int(P) and Li (P) ⊂ int(P) for
some i > 0. Then the modulus of any eigenvalue of L is strictly less than 1.
Proof. If the conclusion holds for Li with some i > 0, then it holds for L.
Hence we may assume that L(P) ⊂ int(P). It follows that Ln (P) ⊂ int(P)
for all n > 0. The matrix L cannot have an eigenvalue of modulus greater
than 1, since otherwise the iterates of L would move some vector in the open
set int(P) off to ∞.
Suppose that σ is an eigenvalue of L and |σ | = 1. If σ j = 1, then Lj has
a fixed point on ∂ P, a contradiction.
If σ is not a root of unity, there is a 2-dimensional subspace U on which
L acts as an irrational rotation and any point p ∈ ∂ P ∩ U is a limit point of
 n
n>0 L (P), a contradiction. 

Since A is non-negative, it induces a continuous map f from the unit

simplex S = {x ∈ Rm : x j = 1, x j ≥ 0, j = 1, . . . , m} into itself; f (x) is the
radial projection of Ax onto S. By the Brouwer fixed point theorem, there
is a fixed point v ∈ S of f , which is a non-negative eigenvector of A with
eigenvalue λ > 0. Since some power of A is positive, all coordinates of v are
positive.
Let V be the diagonal matrix that has the entries of v on the diagonal.
The matrix M = λ−1 V −1 AV is primitive, and the column vector 1 with all
3.3. The Perron–Frobenius Theorem 59

entries 1 is an eigenvector of M with eigenvalue 1, i.e., M is a stochastic

matrix. To prove parts 1 and 3, it suffices to show that 1 is a simple root of
the characteristic polynomial of M and that all other eigenvalues of M have
moduli strictly less than 1. Consider the action of M on row vectors. Since
M is stochastic and non-negative, the row action preserves the unit simplex
S. By the Brouwer fixed point theorem, there is a fixed row vector w ∈ S
all of whose coordinates are positive. Let P = S − w be the translation of S
by −w. Since for some j > 0 all entries of M j are positive, M j (P) ⊂ int(P)
and, by Lemma 3.3.2, the modulus of any eigenvalue of the row action of M
in the (m − 1)-dimensional invariant subspace spanned by P is strictly less
than 1.
The last statement of the theorem follows from the fact that the
codimension-one subspace spanned by P is M t -invariant and its intersec-
tion with the cone of non-negative vectors in Rn is {0}. 

COROLLARY 3.3.3. Let A be a primitive stochastic matrix. Then 1 is a simple
root of the characteristic polynomial of A, both A and the transpose of A
have positive eigenvectors with eigenvalue 1, and any other eigenvalue of A
has modulus strictly less than 1.

Frobenius extended Theorem 3.3.1 to irreducible matrices.

THEOREM 3.3.4 (Frobenius). Let A be a non-negative irreducible square

matrix. Then there exists an eigenvalue λ of A with the following properties:
(i) λ > 0, (ii) λ is a simple root of the characteristic polynomial, (iii) λ has a
positive eigenvector, (iv) if µ is any other eigenvalue of A, then |µ| ≤ λ, (v) if k
is the number of eigenvalues of modulus |λ|, then the spectrum of A(with mul-
tiplicity) is invariant under the rotation of the complex plane by angle 2π/k.

A proof of Theorem 3.3.4 is outlined in Exercise 3.3.3. A complete argu-

ment can be found in [Gan59] or [BP94].

Exercise 3.3.1. Show that if A is a primitive integral matrix, then the edge
shift
eA is topologically mixing.

Exercise 3.3.2. Show that if A is an irreducible integral matrix, then the

edge shift
eA is topologically transitive.

Exercise 3.3.3. This exercise outlines the main steps in the proof of
Theorem 3.3.4. Let A be a non-negative irreducible matrix, and let B be
the matrix with entries bi j = 0 if ai j = 0 and bi j = 1 if ai j > 0. Let  be the
graph whose adjacency matrix is B. For a vertex v in , let d = d(v) be the
greatest common divisor of the lengths of closed paths in  starting from v.
Let Vk, k = 0, 1, . . . , d − 1, be the set of vertices of  that can be connected
60 3. Symbolic Dynamics

to v by a path whose length is congruent to k mod d.

(a) Prove that d does not depend on v.
(b) Prove that any path of length l starting in Vk ends in Vm with m con-
gruent to k + l mod d.
(c) Prove that there is a permutation of the vertices that conjugates B d
to a block-diagonal matrix with square blocks Bk, k = 0, 1, . . . , d − 1,
along the diagonal and zeros elsewhere, each Bk being a primitive
matrix whose size equals the cardinality of Vk.
(d) What are the implications for the spectrum of A?
(e) Deduce Theorem 3.3.4.

3.4 Topological Entropy and the Zeta Function of an SFT

For an edge or vertex shift, dynamical invariants can be computed from
the adjacency matrix. In this section, we compute the topological entropy
of an edge shift and introduce the zeta function, an invariant that collects
combinatorial information about the periodic points.

PROPOSITION 3.4.1. Let A be a square non-negative integer matrix. Then

the topological entropy of the edge shift
eA and the vertex shift
vA equals the
logarithm of the largest eigenvalue of A.

Proof. We consider only the edge shift. By Proposition 3.1.1, it suffices to

compute the cardinality of Wn (
A) (the words of length n in
A), which is
the sum Sn of all entries of An (Exercise 1.4.2). The proposition now follows
from Exercise 3.4.1. 

For a discrete dynamical system f , denote by Fix( f ) the set of fixed points
of f and by |Fix( f )| its cardinality. If |Fix( f n )| is finite for every n, we define
the zeta function ζ f (z) of f to be the formal power series

∞
1
ζ f (z) = exp |Fix( f n )|zn .
n=1
n

The zeta function can also be expressed by the product formula:

 −1
ζ f (z) = 1 − z|γ | ,
γ

where the product is taken over all periodic orbits γ of f , and |γ | is the
number of points in γ (Exercise 3.4.4). The generating function g f (z) is
3.4. Topological Entropy and the Zeta Function of an SFT 61

another way to collect information about the periodic points of f :


∞
g f (z) = |Fix( f n )|zn .
n=1

The generating function is related to the zeta function by ζ f (z) = exp(zg f (z)).
The zeta function of the edge shift determined by an adjacency matrix A
is denoted by ζ A. A priori, the zeta function is merely a formal power series.
The next proposition shows that the zeta function of an SFT is a rational
function.

PROPOSITION 3.4.2. ζ A(z) = (det(I − zA))−1 .

Proof. Observe that

∞
xn 1
exp = exp(− log(1 − x)) = ,
n=1
n 1−x

and that |Fix(σ n |
A)| = tr(An ) = λ λn , where the sum is over the eigenval-
ues of A, repeated with the proper multiplicity (see Exercise 1.4.2). There-
fore, if A is N × N,

 ∞ 
(λz)n  ∞
(λz)n 
ζ A(z) = exp = exp = (1 − λz)−1
n=1 λ
n λ n=1
n λ

  −1   −1
1 1 1
= N −λ = zN det I−A = (det(I − zA))−1 .
z λ z z


The following theorem addresses the rationality of the zeta function for
a general subshift.

THEOREM 3.4.3 (Bowen–Lanford [BL70]). The zeta function of a subshift

X ⊂
m is rational if and only if there are matrices A and B such that
|Fix(σ n | X)| = trAn − trBn for all n ∈ N0 .

Exercise 3.4.1. Let A be a non-negative, non-zero, square matrix, Sn the

sum of entries of An , and λ the eigenvalue of A with largest modulus. Prove
that limn→∞ (log Sn )/n = log λ.

Exercise 3.4.2. Calculate the zeta and generating functions of the full
2-shift.

Exercise 3.4.3. Let A = ( 11 10 ). Calculate the zeta function of
eA.

62 3. Symbolic Dynamics

Exercise 3.4.4. Prove the product formula for the zeta function.

Exercise 3.4.5. Calculate the generating function of an edge shift with ad-
jacency matrix A.

Exercise 3.4.6. Calculate the zeta function of a hyperbolic toral automor-

phism (see Exercise 1.7.4).

Exercise 3.4.7. Prove that if the zeta function is rational, then so is the
generating function.

3.5 Strong Shift Equivalence and Shift Equivalence

We saw in §3.2 that any subshift of finite type is isomorphic to an edge shift
eA
for some adjacency matrix A. In this section, we give an algebraic condition
on pairs of adjacency matrices that is equivalent to topological conjugacy of
the corresponding edge shifts.
Square matrices A and B are elementary strong shift equivalent if there
are (not necessarily square) non-negative integer matrices U and V such
that A = UV and B = VU. Matrices A and B are strong shift equivalent if
there are (square) matrices A1 , . . . , An such that A1 = A, An = B, and the
matrices Ai and Ai+1 are elementary strong shift equivalent. For example,
the matrices
 
1 1 0
1 1 1 and (3)
2 2 1
are strong shift equivalent but not elementary strong shift equivalent
(Exercise 3.5.1).

THEOREM 3.5.1 (Williams [Wil73]). The edge shifts
eA and
eB are topolog-
ically conjugate if and only if the matrices Aand B are strong shift equivalent.

Proof. We show here only that strong shift equivalence gives an isomor-
phism of the edge shifts. The other direction is much more difficult (see
[LM95]).
It is sufficient to consider the case when A and B are elementary strong
shift equivalent. Let A= UV, B = VU, and  A,  B be the (disjoint) directed
graphs with adjacency matrices Aand B. If A is k × k and B is l × l, then U
is k × l and V is l × k. We interpret the entry Ui j as the number of (additional)
edges from vertex i of  A to vertex j of  B, and similarly we interpret
Vji as the number of edges from vertex j of  B to vertex i of  A. Since
3.5. Strong Shift Equivalence and Shift Equivalence 63

a−1 a0 a1

v−1 v0 v1
u−1 u0 u1

b−1 b0
Figure 3.1. A graph constructed from an elementary strong shift equivalence.


Apq = lj=1 Upj Vjq , the number of edges in  A from vertex p to vertex q is
the same as the number of paths of length 2 from vertex p to vertex q through
a vertex in  B. Therefore we can choose a one-to-one correspondence φ
between the edges a of  A and pairs uv of edges determined by U and V,
i.e., φ(a) = uv, so that the starting vertex of u is the starting vertex of a, the
terminal vertex of u is the starting vertex of v, and the terminal vertex of v
is the terminal vertex of a. Similarly, there is a bijection ψ from the edges
b of  B to pairs vu of edges determined by V and U. For each sequence
. . . a−1 a0 a1 . . . ∈
eA apply φ to get

. . . φ(a−1 )φ(a0 )φ(a1 ) . . . = . . . u−1 v−1 u0 v0 u1 v1 . . . ,

and then apply ψ −1 to get . . . b−1 b0 b1 . . . ∈
eB with bi = ψ −1 (vi ui+1 ) (see

Figure 3.1). This gives an isomorphism from
eA to
eB. 

Square matrices A and B are shift equivalent if there are (not necessarily
square) non-negative integer matrices U, V, and a positive integer k (called
the lag) such that

Ak = UV, B k = VU, AU = UB, BV = VA.

The notion of shift equivalence was introduced by R. Williams, who con-

jectured that if two primitive matrices are shift equivalent, then they are
strong shift equivalent, or, in view of Theorem 3.5.1, that shift equivalence
classifies subshifts of finite type. K. Kim and F. Roush [KR99] constructed a
counterexample to this conjecture.
For other notions of equivalence for SFTs see [Boy93].

Exercise 3.5.1. Show that the matrices

 
1 1 0
A= 1 1 1 and B = (3)
2 2 1

are strong shift equivalent but not elementary strong shift equivalent. Write
down an explicit isomorphism from (
A, σ ) to (
B, σ ).
64 3. Symbolic Dynamics

Exercise 3.5.2. Show that strong shift equivalence and shift equivalence
are equivalence relations and elementary strong shift equivalence is not.

3.6 Substitutions2
For an alphabet Am = {0, 1, . . . , m − 1}, denote by A∗m the collection of all
finite words in Am, and by |w| the length of w ∈ A∗m. A substitution s: Am →
A∗m assigns to every symbol a ∈ Am a finite word s(a) ∈ A∗m. We assume
throughout this section that |s(a)| > 1 for some a ∈ Am, and that |s n (b)| → ∞
for every b ∈ Am. Applying the substitution to each element of a sequence
or a word gives maps s: A∗m → A∗m and s:
m +
→
m+
,
s
x0 x1 . . . →s(x0 )s(x1 ) . . . .
These maps are continuous but not surjective. If s(a) has the same length
for all a ∈ Am, then s is said to have constant length.
Consider the example m = 2, s(0) = 01, s(1) = 10. We have: s 2 (0) =
0110, s 3 (0) = 01101001, s 4 (0) = 0110100110010110, . . . . If w̄ is the word
obtained from w by interchanging 0 and 1, then s n+1 (0) = s n (0)s n (0). The
sequence of finite words s n (0) stabilizes to an infinite sequence
M = 01101001100101101001011001101001 . . .
called the Morse sequence. The sequences M and M̄ are the only fixed
+
points of s in
m .
+
PROPOSITION 3.6.1. Every substitution s has a periodic point in
m .
Proof. Consider the map a →s(a)0 . Since Am contains m elements, there
are n ∈ {1, . . . , m} and a ∈ Am such that s n (a)0 = a. If |s n (a)| = 1, then the
sequence aaa . . . is a fixed point of s n . Otherwise, |s ni (a)| → ∞, and the
sequence of finite words s ni (a) stabilizes to a fixed point of s n in
m +
. 

+
If a substitution s has a fixed point x = x0 x1 . . . ∈
m and |s(x0 )| > 1, then
s(x0 )0 = x0 and the sequence s (x0 ) stabilizes to x; we write x = s ∞ (x0 ). If
n

|s(a)| > 1 for every a ∈ Am, then s has at most m fixed points in
m.
The closure
s (a) of the (forward) orbit of a fixed point s ∞ (a) under the
shift σ is a subshift.
We call a substitution s: Am → A∗m irreducible if for any a, b ∈ Am there is
n(a, b) ∈ N such that s n(a,b) (a) contains b; s is primitive if there is n ∈ N such
that s n (a) contains b for all a, b ∈ Am.
We assume from now on that |s n (b)| → ∞ for every b ∈ Am.

2 Several arguments in this section follow in part those of [Que87].

3.6. Substitutions 65

PROPOSITION 3.6.2. Let s be an irreducible substitution over Am. If s(a)0 =

a for some a ∈ Am, then s is primitive and the subshift (
s (a), σ ) is minimal.
Proof. Observe that s n (a)0 = a for all n ∈ N. Since s is irreducible, for
every b ∈ Am there is n(b) such that b appears in s n(b) (a), and therefore
appears in s n (a) for all n ≥ n(b). Hence, s n (a) contains all symbols from
Am if n ≥ N = max n(b). Since s is irreducible, for every b ∈ Am there is k(b)
such that a appears in s k(b) (b) and hence in s n (b) with n ≥ k(b). It follows that
for every c ∈ Am, s n (c) contains all symbols from Am if n ≥ 2(N + max k(b)),
so s is primitive.
Recall (Proposition 2.1.3) that (
s (a), σ ) is minimal if and only if s ∞ (a) is
almost periodic, i.e., for every n ∈ N the word s n (a) occurs in s ∞ (a) infinitely
often, and the gaps between successive occurrences are bounded. This hap-
pens if and only if a recurs in s ∞ (a) with bounded gaps, which holds true
because s is primitive (Exercise 3.6.1). 

For two words u, v ∈ A∗m denote by Nu (v) the number of times u occurs in
v. The composition matrix M = M(s) of a substitution s is the non-negative
integer matrix with entries Mi j = Ni (s( j)). The matrix M(s) is primitive
(respectively, irreducible) if and only if the substitution s is primitive (re-
spectively, irreducible). For a word w ∈ A∗m, the numbers Ni (w), i ∈ Am,
form a vector N(w) ∈ Rm. Observe that M(s n ) = (M(s))n for all n ∈ N and
N(s(w)) = M(s)N(w). If s has constant length l, then the sum of every col-
umn of M is l and the transpose of l −1 M is a stochastic matrix.

PROPOSITION 3.6.3. Let s: Am → A∗m be a primitive substitution, and let λ

be the largest in modulus eigenvalue of M(s). Then for every a ∈ Am
1. limn→∞ λ−n N(s n (a)) is an eigenvector of M(s) with eigenvalue λ,
|s n+1 (a)|
2. lim = λ,
n→∞ |s n (a)|
3. v = limn→∞ |s n (a)|−1 N(s n (a)) is an eigenvector of M(s) corresponding
m−1
to λ, and i=0 vi = 1.
Proof. The proposition follows directly from Theorem 3.3.1 (Exercise 3.6.2).


PROPOSITION 3.6.4. Let s be a primitive substitution, s ∞ (a) be a fixed point
of s, and ln be the number of different words of length n occurring in s ∞ (a).
Then there is a constant C such that ln ≤ C · n for all n ∈ N. Consequently, the
topological entropy of (
s (a), σ ) is 0.
Proof. Let ν k = mina∈Am |s k(a)| and ν̄k = maxa∈Am |s k(a)|, and note that
ν k, ν̄k → ∞ monotonically in k. Hence for every n ∈ N there is k = k(n) ∈ N
such that ν k−1 ≤ n ≤ ν k. Therefore, every word of length n occurring in x
66 3. Symbolic Dynamics

is contained in s k(ab) for a pair of consecutive symbols ab from x. Let λ

be the maximal-modulus eigenvalue λ of the primitive composition matrix
M = M(s). Then for every non-zero vector v with non-negative components
there are constants C1 (v) and C2 (v) such that for all k ∈ N,

C1 (v)λk ≤ Mkv ≤ C2 (v)λk,

where · is the Euclidean norm. Hence, by Proposition 3.6.3(1), there are
positive constants C1 and C2 such that for all k ∈ N

C1 · λk ≤ ν k ≤ ν̄k ≤ C2 · λk.

Since for every a ∈ Am there are at most ν̄k different words of length n in
s k(ab) with initial symbol in s k(a), we have
     
C2 2 C2 2 C2 2
ln ≤ m2 ν̄k ≤ C2 λkm2 = m λ C1 λk−1 ≤ m λ ν k−1 ≤ m λ n.
C1 C1 C1


Exercise 3.6.1. Prove that if s is primitive and s(a)0 = a, then each symbol
b ∈ Am appears in s ∞ (a) infinitely often and with bounded gaps.

Exercise 3.6.2. Prove Proposition 3.6.3.

3.7 Sofic Shifts

A subshift X ⊂
m is called sofic if it is a factor of a subshift of finite type, i.e.,
there is an adjacency matrix A and a code c:
eA → X such that c ◦ σ = σ ◦ c.
Sofic shifts have applications in finite-state automata and data transmission
and storage [MRS95].
A simple example of a sofic shift is the following subshift of (
2 , σ ), called
the even system of Weiss [Wei73]. Let A be the adjacency matrix of the graph
 A consisting of two vertices u and v, an edge from u to itself labeled 1,
an edge from u to v labeled 01 , and an edge from v to u labeled 02 (see
Figure 3.2). Let X be the set of sequences of 0s and 1s such that there is
an even number of 0s between every two 1s. The surjective code c:
A → X
replaces both 01 and 02 by 0.
As Proposition 3.7.1 shows, every sofic shift can be obtained by the fol-
lowing construction. Let  be a finite directed labeled graph, i.e., the edges
of  are labeled by an alphabet Am. Note that we do not assume that differ-
ent edges of  are labeled differently. The subset X ⊂
m consisting of all
infinite directed paths in  is closed and shift invariant.
3.8. Data Storage 67

01
1 u v

02

Figure 3.2. The directed graph used to construct the even system of Weiss.

If a subshift (X, σ ) is isomorphic to (X , σ ) for some directed labeled

graph , then we say that  is a presentation of X. For example, a presentation
for the even system of Weiss is obtained by replacing the labels 01 and 02
with 0 in Figure 3.2.

PROPOSITION 3.7.1. A subshift X ⊂
m is sofic if and only if it admits a

presentation by a finite directed labeled graph.
Proof. Since X is sofic, there is a matrix A and a code c:
eA → X (see
Corollary 3.2.2). By Proposition 3.1.2, c is a block code. By passing to a
higher block presentation we may assume that c is a 1-block code. Hence,
X admits a presentation by a finite directed labeled graph. The converse is
Exercise 3.7.2. 


Exercise 3.7.1. Prove that the even system of Weiss is not a subshift of finite
type.

Exercise 3.7.2. Prove that for any directed labeled graph , the set X is a
sofic shift.

Exercise 3.7.3. Show that there are only countably many non-isomorphic
sofic shifts. Conclude that there are subshifts that are not sofic.

3.8 Data Storage3

Most computer storage devices (floppy disk, hard drive, etc.) store data as a
chain of magnetized segments on tracks. A magnetic head can either change
or detect the polarity of a segment as it passes the head. Since it is technically
much easier to detect a change of polarity than to measure the polarity, a
common technique is to record a 1 as a change of polarity and a 0 as no
change in polarity. The two major problems that restrict the effectiveness
of this method are intersymbol interference and clock drift. Both of these

3 The presentation of this section follows in part [BP94].

68 3. Symbolic Dynamics

problems can be ameliorated by applying a block code to the data before it

is written to the storage device.
Intersymbol interference occurs when two polarity changes are adjacent
to each other on the track; the magnetic fields from the adjacent positions
partially cancel each other, and the magnetic head may not read the track
correctly. This effect can be minimized by requiring that in the encoded
sequence every two 1s are separated by at least one 0.
A sequence of n 0s with 1s on both ends is read off the track as two
pulses separated by n non-pulses. The length n is obtained by measuring the
time between the pulses. Every time a 1 is read, the clock is synchronized.
However, for a long sequence of 0s, clock error accumulates, which may
cause the data to read incorrectly. To counteract this effect the encoded
sequence is required to have no long stretches of 0s.
A common coding scheme called modified frequency modulation (MFM)
inserts a 0 between each two symbols unless they are both 0s, in which case
it inserts a 1. For example, the sequence

10100110001

is encoded for storage as

100010010010100101001.

This requires twice the length of the track, but results in fewer read/write
errors. The set of sequences produced by the MFM coding is a sofic system
(Exercise 3.8.3).
There are other considerations for storage devices that impose additional
conditions on the sequences used to encode data. For example, the total
magnetic charge of the device should not be too large. This restriction leads
to a subset of (
2 , σ ) that is not of finite type and not sofic.
Recall that the topological entropy of the factor does not exceed the
topological entropy of the extension (Exercise 2.5.5). Therefore in any one-
to-one coding scheme, which increases the length of the sequence by a factor
of n > 1, the topological entropy of the original subshift must be not more
than n times the topological entropy of the target subshift.

Exercise 3.8.1. Prove that the sequences produced by MFM have at least
one and at most three 0s between every two 1s.

Exercise 3.8.2. Describe an algorithm to reverse the MFM coding.

Exercise 3.8.3. Prove that the set of sequences produced by the MFM cod-
ing is a sofic system.
CHAPTER FOUR

Ergodic Theory

Ergodic1 theory is the study of statistical properties of dynamical systems

relative to a measure on the underlying space of the dynamical system. The
name comes from classical statistical mechanics, where the “ergodic hypoth-
esis” asserts that, asymptotically, the time average of an observable is equal
to the space average. Among the dynamical systems with natural invariant
measures that we have encountered before are circle rotations (§1.2) and
toral automorphisms (§1.7). Unlike topological dynamics, which studies the
behavior of individual orbits (e.g., periodic orbits), ergodic theory is con-
cerned with the behavior of the system on a set of full measure and with the
induced action in spaces of measurable functions such as Lp (especially L2 ).
The proper setting for ergodic theory is a dynamical system on a measure
space. Most natural (non-atomic) measure spaces are measure-theoretically
isomorphic to an interval [0, a] with Lebesgue measure, and the results in this
chapter are most important in that setting. The first section of this chapter
recalls some notation, definitions, and facts from measure theory. It is not
intended to serve as a complete exposition of measure theory (for a full
introduction see, for example, [Hal50] or [Rud87]).

4.1 Measure-Theory Preliminaries

A non-empty collection A of subsets of a set X is called a σ-algebra if A is
closed under complements and countable unions (and hence countable in-
tersections). A measure µ on A is a non-negative (possibly infinite) function


on A that is σ-additive, i.e., µ( i Ai ) = i µ(Ai ) for any countable collec-
tion of disjoint sets Ai ∈ A. A set of measure 0 is called a null set. A set
whose complement is a null set is said to have full measure. The σ-algebra

1 From the Greek words ργ 

´ oν, “work,” and óδoς , “path.”

69
70 4. Ergodic Theory

is complete (relative to µ) if it contains every subset of every null set. Given

a σ-algebra A and a measure µ, the completion Ā is the smallest σ-algebra
containing A and all subsets of null sets in A; the σ-algebra Ā is complete.
A measure space is a triple (X, A, µ), where X is a set, A is a σ-algebra
of subsets of X, and µ is a σ-additive measure. We always assume that A is
complete, and that µ is σ-finite, i.e., that X is a countable union of subsets
of finite measure. The elements of A are called measurable sets.
If µ(X) = 1, then (X, A, µ) is called a probability space and µ is a proba-
bility measure. If µ(X) is finite, then we can rescale µ by the factor 1/µ(X)
to obtain a probability measure.
Let (X, A, µ) and (Y, B, ν) be measure spaces. The product measure space
is the triple (X × Y, C, µ × ν), where C is the completion relative to µ × ν of
the σ-algebra generated by A × B.
Let (X, A, µ) and (Y, B, ν) be measure spaces. A map T: X → Y is called
measurable if the preimage of any measurable set is measurable. A measur-
able map T is non-singular if the preimage of every set of measure 0 has
measure 0, and is measure-preserving if µ(T −1 (B)) = ν(B) for every B ∈ B.
A non-singular map from a measure space into itself is called a non-singular
transformation (or simply a transformation). If a transformation T preserves
a measure µ, then µ is called T-invariant. If T is an invertible measurable
transformation, and its inverse is measurable and non-singular, then the iter-
ates T n , n ∈ Z, form a group of measurable transformations. Measure spaces
(X, A, µ) and (Y, B, ν) are isomorphic if there is a subset X of full measure
in X, a subset Y  of full measure in Y, and an invertible bijection T: X → Y 
such that T and T −1 are measurable and measure-preserving with respect
to (A, µ) and (B, ν). An isomorphism from a measure space into itself is an
automorphism.
Denote by λ the Lebesgue measure on R. A flow T t on a measure space
(X, A, µ) is measurable if the map T: X × R → X, (x, t) → T t (x ), is measur-
able with respect to the product measure on X × R, and T t : X → X is a
non-singular measurable transformation for each t ∈ R. A measurable flow
T t is a measure-preserving flow if each T t is a measure-preserving transfor-
mation.
Let T be a measure-preserving transformation of a measure space
(X, A, µ), and S a measure-preserving transformation of a measure space
(Y, B, ν). We say that T is an extension of S if there are sets X ⊂ X and
Y  ⊂ Y of full measure and a measure-preserving map ψ: X → Y  such that
ψ ◦ T = S ◦ ψ. A similar definition holds for measure-preserving flows. If ψ
is an isomorphism, then T and S are called isomorphic. The product T × S
is a measure-preserving transformation of (X × Y, C, µ × ν), where C is the
completion of the σ-algebra generated by A × B.
4.2. Recurrence 71

Let X be a topological space. The smallest σ-algebra containing all the

open subsets of X is called the Borel σ-algebra of X. If A is the Borel σ-
algebra, then a measure µ on A is a Borel measure if the measure of any
compact set is finite. A Borel measure is regular in the sense that the measure
of any set is the infimum of measures of open sets containing it, and the
supremum of measures of compact sets contained in it.
A one-point subset with positive measure is called an atom. A finite mea-
sure space is a Lebesgue space if it is isomorphic to the union of an interval
[0, a] (with Lebesgue measure) and at most countably many atoms. Most nat-
ural measure spaces are Lebesgue spaces. For example, if X is a complete
separable metric space, µ a finite Borel measure on X, and A the comple-
tion of the Borel σ-algebra with respect to µ, then (X, A, µ) is a Lebesgue
space. In particular, the unit square [0, 1] × [0, 1] with Lebesgue measure is
(measure-theoretically) isomorphic to the unit interval [0, 1] with Lebesgue
measure (Exercise 4.1.1).
A Lebesgue space without atoms is called non-atomic, and is isomorphic
to an interval [0, a] with Lebesgue measure.
A set has full measure if its complement has measure 0. We say a property
holds mod 0 in X, or holds for µ-almost every (a.e.) x, if it holds on a subset
of full µ-measure in X. We also use the word essentially to indicate that a
property holds mod 0.
Let (X, A, µ) be a measure space. Two measurable functions are equiv-
alent if they coincide on a set of full measure. For p ∈ (0, ∞), the space
L p (X, µ) consists of equivalence classes mod 0 of measurable functions
f : X → C such that | f | p dµ < ∞. As a rule, if there is no ambiguity, we
identify the function with its equivalence class. For p ≥ 1, the L p norm is de-
fined by f p = ( | f | dµ)1/ p . The space L2 (X, µ) is a Hilbert space with
p

inner product  f, g = f · g dµ. The space L∞ (X, µ) consists of equiva-

lence classes of essentially bounded measurable functions. If µ is finite, then
L∞ (X, µ) ⊂ L p (X, µ) for all p > 0. If X is a topological space and µ is a
Borel measure on X, then the space C0 (X, C) of continuous, complex-valued,
compactly supported functions on X is dense in L p (X, µ) for all p > 0.

Exercise 4.1.1. Prove that the unit square [0, 1] × [0, 1] with Lebesgue
measure is (measure-theoretically) isomorphic to the unit interval [0, 1] with
Lebesgue measure.

4.2 Recurrence
The following famous result of Poincaré implies that recurrence is a generic
property of orbits of measure-preserving dynamical systems.
72 4. Ergodic Theory

THEOREM 4.2.1 (Poincaré Recurrence Theorem). Let T be a measure-

preserving transformation of a probability space (X, A, µ). If A is a measur-
able set, then for a.e. x ∈ A, there is some n ∈ N such that T n (x) ∈ A. Conse-
quently, for a.e. x ∈ A, there are infinitely many k ∈ N for which T k(x) ∈ A.
Proof. Let

B = {x ∈ A: T (x) ∈
k
/ A for all k ∈ N} = A T −k(A).
k∈N
−k
Then B ∈ A, and all the preimages T (B) are disjoint, are measurable, and
have the same measure as B. Since X has finite total measure, it follows that
B has measure 0. Since every point in A\B returns to A, this proves the first
assertion. The proof of the second assertion is Exercise 4.2.1. 

For continuous maps of topological spaces, there is a connection between
measure-theoretic recurrence and the topological recurrence introduced in
Chapter 2. If X is a topological space, and µ is a Borel measure on X, then
supp µ (the support of µ) is the complement of the union of all open sets
with measure 0 or, equivalently, the intersection of all closed sets with full
measure. Recall from §2.1 that the set of recurrent points of a continuous
map T: X → X is R(T) = {x ∈ X : x ∈ ω(x)}.
PROPOSITION 4.2.2. Let X be a separable metric space, µ a Borel proba-
bility measure on X, and f : X → X a continuous measure-preserving trans-
formation. Then almost every point is recurrent, and hence supp µ ⊂ R( f ).
Proof. Since X is separable, there is a countable basis {Ui }i∈Z for the topol-
ogy of X. A point x ∈ X is recurrent if it returns (in the future) to every
basis element containing it. By the Poincaré recurrence theorem, for each
i, there is a subset Ũ i of full measure in Ui such that every point of Ũ i returns

to Ui . Then Xi = Ũ i ∪ (X\Ui ) has full measure in X, so X̃ = i∈Z Xi = R(T)
has full measure in X. 

We will discuss some applications of measure-theoretic recurrence in
§4.11.
Given a measure-preserving transformation T in a finite measure space
(X, A, µ) and a measurable subset A ∈ A of positive measure, the derivative
transformation TA: A → A is defined by TA(x) = T k(x), where k ∈ N is the
smallest natural number for which T k(x) ∈ A. The derivative transformation
is often called the first return map, or the Poincaré map. By Theorem 4.2.1,
TA is defined on a subset of full measure in A.
Let T be a transformation on a measure space (X, A, µ), and f : X → N a
measurable function. Let X f = {(x, k): x ∈ X, 1 ≤ k ≤ f (x)} ⊂ X × N. Let
A f be the σ-algebra generated by the sets A× {k}, A ∈ A, k ∈ N, and define
4.3. Ergodicity and Mixing 73

µ f (A× {k}) = µ(A). Define the primitive transformation Tf : X f → X f by

Tf (x, k) = (x, k + 1) if k < f (x) and Tf (x, f (x)) = (T(x), 1). If µ(X) < ∞
and f ∈ L1 (X, A, µ), then µ f (X f ) = X f (x) dµ . Note that the derivative
transformation of Tf on the set X × {1} is just the original transformation T.
Primitive and derivative transformations are both referred to as induced
transformations; we will encounter them later.

Exercise 4.2.1. Prove the second assertion of Theorem 4.2.1.

Exercise 4.2.2. Suppose T: X → X is a continuous transformation of a

topological space X, and µ is a finite T-invariant Borel measure on X with
supp µ = X. Show that every point is non-wandering and µ-a.e. point is
recurrent.

Exercise 4.2.3. Prove that if T is a measure-preserving transformation,

then so are the induced transformations.

4.3 Ergodicity and Mixing

A dynamical system induces an action on functions: T acts on a function f
by (T∗ f )(x) = f (T(x)). The ergodic properties of a dynamical system cor-
respond to the degree of statistical independence between f and T∗n f . The
strongest possible dependence happens for an invariant function f (T(x)) =
f (x). The strongest possible independence happens when a non-zero L2
function is orthogonal to its images.
Let T be a measure-preserving transformation (or flow) on a measure
space (X, A, µ). A measurable function f : X → R is essentially T-invariant
if µ({x ∈ X: f (T t x) = f (x)}) = 0 for every t. A measurable set A is essen-
tially T-invariant if its characteristic function 1 A is essentially T-invariant;
equivalently, if µ(T −1 (A)  A) = 0 (we denote by  the symmetric differ-
ence, A  B = (A\B) ∪ (B\A)).
A measure-preserving transformation (or flow) T is ergodic if any es-
sentially T-invariant measurable set has either measure 0 or full measure.
Equivalently (Exercise 4.3.1), T is ergodic if any essentially T-invariant
measurable function is constant mod 0.

PROPOSITION 4.3.1. Let T be a measure-preserving transformation or flow

on a finite measure space (X, A, µ), and let p ∈ (0, ∞]. Then T is ergodic if and
only if every essentially invariant function f ∈ L p (X, µ) is constant mod 0.
Proof. If T is ergodic, then every essentially invariant function is con-
stant mod 0.
74 4. Ergodic Theory

To prove the converse, let f be an essentially invariant measurable func-

tion on X. Then for every M > 0, the function

f (x) if f (x) ≤ M,
fM (x) =
0 if f (x) > M

is bounded, is essentially invariant, and belongs to L p (X, µ). Therefore it is

constant mod 0. It follows that f itself is constant mod 0. 

As the following proposition shows, any essentially invariant set or func-
tion is equal mod 0 to a strictly invariant set or function.

PROPOSITION 4.3.2. Let (X, A, µ) be a measure space, and suppose that

f : X → R is essentially invariant for a measurable transformation or flow
T on X. Then there is a strictly invariant measurable function f˜ such that
f (x) = f˜(x) mod 0.
Proof. We prove the proposition for a measurable flow. The case of a mea-
surable transformation follows by a similar but easier argument and is left
as an exercise.
Consider the measurable map : X × R → R, (x, t) = f (T t x) − f (x),
and the product measure ν = µ × λ in X × R, where λ is Lebesgue mea-
sure on R. The set A = −1 (0) is a measurable subset of X × R. Since f is
essentially T-invariant, for each t ∈ R the set

At = {(x, t) ∈ (X × R): f (T t x) = f (x)}

has full µ-measure in X × {t}. By the Fubini theorem, the set

Af = {x ∈ X: f (T t x) = f (x) for a.e. t ∈ R}

has full µ-measure in X. Set


˜f (x) = f (y) if T x = y ∈ Af for some t ∈ R,
t

0 otherwise.

If T t x = y ∈ Af and T s x = z ∈ Af , then y and z lie on the same orbit, and

the value of f along this orbit is equal λ-almost everywhere to f (y) and
to f (z), so f (y) = f (z). Therefore f˜ is well defined and strictly T-invariant.


A measure-preserving transformation (or flow) T on a probability space
(X, A, µ) is called (strong) mixing if

lim µ(T −t (A) ∩ B) = µ(A) · µ(B)

t→∞
4.3. Ergodicity and Mixing 75

for any two measurable sets A, B ∈ A. Equivalently (Exercise 4.3.3), T is

mixing if
  
lim f (T t (x)) · g(x) dµ = f (x) dµ · g(x) dµ
t→∞ X X X

for any bounded measurable functions f, g.

A measure-preserving transformation T of a probability space (X, A, µ)
is called weak mixing if for all A, B ∈ A,

1 n−1
lim |µ(T −i (A) ∩ B) − µ(A) · µ(B)| = 0
n→∞ n
i=0

or, equivalently (Exercise 4.3.3), if for all bounded measurable functions

f, g,
n−1

  

= 0.
lim
f (T i
(x))g(x) dµ − f dµ · g dµ

n→∞ n X X X
i=0

A measure-preserving flow T t on (X, A, µ) is weak mixing if for all A, B ∈ A


1 t
lim |µ(T −s (A) ∩ B) − µ(A) · µ(B)| ds = 0,
t→∞ t 0

or, equivalently (Exercise 4.3.3), if for all bounded measurable functions

f, g,

  

1 t

= 0.
lim
f (T s
(x))g(x) dµ ds − f dµ · g dµ

t→∞ t 0 X X X

In practice, the definitions of ergodicity and mixing in terms of L2 func-

tions are often easier to work with than the definitions in terms of measurable
sets. For example, to establish a certain property for each L2 function on a
separable topological space with Borel measure it suffices to do it for a count-
able set of continuous functions that is dense in L2 (Exercise 4.3.5). If the
property is “linear”, it is enough to check it for a basis in L2 , e.g., for the
exponential functions e2πi x on the circle [0, 1).

PROPOSITION 4.3.3. Mixing implies weak mixing, and weak mixing implies
ergodicity.
Proof. Suppose T is a measure-preserving transformation of the probability
space (X, A, µ). Let A and B be measurable subsets of X. If T is mixing,
then |µ(T −i (A) ∩ B) − µ(A) · µ(B)| converges to 0, so the averages do as
well; thus T is weak mixing.
76 4. Ergodic Theory

Let A be an invariant measurable set. Then applying the definition of

weak mixing with B = A, we conclude that µ(A) = µ(A)2 , so either µ(A) = 1
or µ(A) = 0. 

For continuous maps, ergodicity and mixing have the following topologi-
cal consequences.

PROPOSITION 4.3.4. Let X be a compact metric space, T: X → X a contin-

uous map, and µ a T-invariant Borel measure on X.
1. If T is ergodic, then the orbit of µ-almost every point is dense in supp µ.
2. If T is mixing, then T is topologically mixing on supp µ.
Proof. Suppose T is ergodic. Let U be a non-empty open set in supp µ.


Then µ(U) > 0. By ergodicity, the backward invariant set k∈N T −k(U) has
full measure. Thus the forward orbit of almost every point visits U. It follows
that the set of points whose forward orbit visits every element of a countable
open basis has full measure in X. This proves the first assertion.
The proof of the second assertion is Exercise 4.3.4. 

Exercise 4.3.1. Show that a measurable transformation is ergodic if and
only if every essentially invariant measurable function is constant mod 0
(see the remark after Corollary 4.5.7).

Exercise 4.3.2. Let T be an ergodic measure-preserving transformation

in a finite measure space (X, A, µ), A ∈ A, µ(A) > 0, and f ∈ L1 (X, A, µ),
f : X → N. Prove that the induced transformations TA and Tf are ergodic.

Exercise 4.3.3. Show that the two definitions of strong and weak mixing
given in terms of sets and bounded measurable functions are equivalent.

Exercise 4.3.4. Prove the second statement of Proposition 4.3.4.

Exercise 4.3.5. Let T be a measure-preserving transformation of (X, A, µ),

and let f ∈ L1 (X, µ) satisfy f (T(x)) ≤ f (x) for a.e. x. Prove that f (T(x)) =
f (x) for a.e. x.

Exercise 4.3.6. Let X be a compact topological space, µ a Borel measure,

and T: X → X a transformation preserving µ. Suppose that for every con-
tinuous f and g with 0 integrals,

f (T n (x)) · g(x) dµ → 0 as n → ∞.
X

Prove that T is mixing.

4.4. Examples 77

Exercise 4.3.7. Show that if T: X → X is mixing, then T × T: X × X →

X × X is mixing.

4.4 Examples
We now prove ergodicity or mixing for some of the examples from Chapter 1.

PROPOSITION 4.4.1. The circle rotation Rα is ergodic with respect to

Lebesgue measure if and only if α is irrational.
Proof. Suppose α is irrational. By Proposition 4.3.1, it is enough to prove
that any bounded Rα -invariant function f : S1 → R is constant mod 0. Since

f ∈ L2 (S1 , λ), the Fourier series ∞ n=−∞ an e
2nπi x
of f converges to f in
 ∞
2
the L norm. The series n=−∞ an e 2nπi(x+α)
converges to f ◦ Rα . Since f =
f ◦ Rα mod 0, uniqueness of Fourier coefficients implies that an = an e2nπiα
for all n ∈ Z. Since e2nπiα =1 for n =0, we conclude that an = 0 for n =0,
so f is constant mod 0.
The proof of the converse is left as an exercise. 

PROPOSITION 4.4.2. An expanding endomorphism Em: S1 → S1 is mixing
with respect to Lebesgue measure.
Proof. Since any measurable subset of S1 can be approximated by a
finite union of intervals, it is sufficient to consider two intervals
A = [ p/mi , ( p + 1)/mi ], p ∈ {0, . . . , mi − 1}, and B = [q/m j , (q + 1)/m j ],
−1
q ∈ {0, . . . , m j − 1}. Recall that Em (B) is the union of m uniformly spaced
j+1
intervals of length 1/m :

m−1
−1
Em (B) = [(km j + q)/m j+1 , (km j + q + 1)/m j+1 ].
k=0
−n
Similarly, Em (B) is the union of mn uniformly spaced intervals of length
−n
1/m j+n . Thus for n > i, the intersection A ∩ Em (B) consists of mn−i intervals
−(n+ j)
of length m . Thus
 
µ A ∩ Em −n
(B) = mn−i (1/mn+ j ) = m−i− j = µ(A) · µ(B). 


PROPOSITION 4.4.3. Any hyperbolic toral automorphism A: T n → T n is

ergodic with respect to Lebesgue measure.
Proof. We consider here only the case

2 1
A= : T2 → T2 ;
1 1
78 4. Ergodic Theory

the argument in the general case is similar. Let f : T2 → R be a bounded A-


invariant measurable function. The Fourier series ∞ m,n=−∞ amn e
2πi(mx+ny)
of
2
f converges to f in L . The series

∞
amn e2πi(m(2x+y)+n(x+y))
m,n=−∞

converges to f ◦ A. Since f is invariant, uniqueness of Fourier coefficients

implies that amn = a(2m+n)(m+n) for all m, n. Since A does not have eigenvalues
on the unit circle, if amn =0 for some (m, n) =(0, 0), then ai j = amn =0 with
arbitrarily large |i| + | j|, and the Fourier series diverges. 

A toral automorphism of T n corresponding to an integer matrix A
is ergodic if and only if no eigenvalue of A is a root of unity; for a proof
see, for example, [Pet89]. A hyperbolic toral automorphism is mixing
(Exercise 4.4.3).
Let A be an m × m stochastic matrix, i.e., A has non-negative entries, and
the sum of every row is 1. Suppose A has a non-negative left eigenvector q
with eigenvalue 1 and sum of entries equal to 1 (recall that if A is irreducible,
then by Corollary 3.3.3, q exists and is unique). We define a Borel probability
+
measure P = PA,q on m (and m ) as follows: for a cylinder C nj of length 1, we
+
define P(C nj ) = q j ; for a cylinder C n,n+1,...,n+k
j0 , j1 ,..., jk ⊂ m (or m ) with k + 1 > 1
consecutive indices,

  
k−1
P C n,n+1,...,n+k
j0 , j1 ,..., jk = q j0 Aji ji+1 .
i=0

In other words, we interpret q as an initial probability distribution on the

set {1, . . . , m}, and A as the matrix of transition probabilities. The number
P(C nj ) is the probability of observing symbol j in the nth place, and Ai j is
the probability of passing from i to j. The fact that q A = q means that the
probability distribution q is invariant under transition probabilities A, i.e.,

  
 m−1
q j = P C n+1
j = P Cin Ai j .
i=0

The pair (A, q) is called a Markov chain on the set {1, . . . , m}.
It can be shown that P extends uniquely to a shift-invariant σ-additive
measure defined on the completion C of the Borel σ-algebra generated by the
cylinders (Exercise 4.4.5); it is called the Markov measure corresponding to
A and q. The measure space (m, C, P) is a non-atomic Lebesgue probability
space. If A is irreducible, this measure is uniquely determined by A.
4.4. Examples 79

A very important particular case of this situation arises when the transi-
tion probabilities do not depend on the initial state. In this case each row of
A is the left eigenvector q, the shift-invariant measure P is called a Bernoulli
measure, and the shift is called a Bernoulli automorphism.
Let A be the adjacency matrix defined by Ai j = 0 if Ai j = 0 and Ai j = 1
if Ai j > 0. Then the support of P is precisely  vA ⊂ m (Exercise 4.4.6).

PROPOSITION 4.4.4. If A is a primitive stochastic m × m matrix, then the

shift σ is mixing in m with respect to the Markov measure P(A).
Proof. Exercise 4.4.7. 

Markov chains can be generalized to the class of stationary (discrete)
stochastic processes, dynamical systems with invariant measures on shift
spaces with a continuous alphabet. Let (, A, P) be a probability space.
A random variable on  is a measurable real-valued function on . A se-
∞
quence ( fi )i=−∞ of random variables is stationary if, for any i 1 , . . . , i k ∈ Z
and any Borel subsets B1 , . . . , Bk ⊂ R,
   
P ω ∈ : fi j (ω) ∈ Bj , j = 1, . . . , k = P ω ∈ : fi j +n (ω) ∈ Bj , j = 1, . . . , k .

Define the map :  → RZ by

(ω) = (. . . , f−1 (ω), f0 (ω), f1 (ω), . . .),

and the measure µ on the Borel subsets of RZ by µ(A) = P(−1 (A)). Since
the sequence ( fi ) is stationary, the shift σ : RZ → RZ defined by (σ x)n = xn+1
preserves µ (Exercise 4.4.8).

Exercise 4.4.1. Prove that the circle rotation Rα is not weak mixing.

Exercise 4.4.2. Let α ∈ R be irrational, and let F: T2 → T2 be the map

(x, y) →(x + α, x + y) mod 1 introduced in §2.4. Prove that F preserves
the Lebesgue measure and is ergodic but not weak mixing.

Exercise 4.4.3. Prove that any hyperbolic automorphism of T n is mixing.

Exercise 4.4.4. Show that an isometry of a compact metric space is not

mixing for any invariant Borel measure whose support is not a single point.
In particular, circle rotations are not mixing.

Exercise 4.4.5. Prove that any Markov measure is shift invariant.

Exercise 4.4.6. Prove that supp PA,q =  vA .

Exercise 4.4.7. Prove Proposition 4.4.4.

80 4. Ergodic Theory

Exercise 4.4.8. Prove that the measure µ on RZ constructed above for a

stationary sequence ( fi ) is invariant under the shift σ .

4.5 Ergodic Theorems2

The collection of all orbits represents a complete evolution of the dynam-
ical system T. The values f (T n (x)) of a (measurable) function f may
represent observations such as position or velocity. Long-term averages
1 n−1 k
n k=0 f (T (x)) of these quantities are important in statistical physics and
other areas. A central question in ergodic theory is whether these averages
converge as n → ∞ and, if so, whether the limit depends on x. In the context
of statistical physics, the ergodic hypothesis states that the asymptotic
 time
n−1
average limn→∞ (1/n) k=0 f (T k(x)) equals the space average X f dµ for
a.e. x. We show that this happens if T is ergodic.
Let (X, A, µ) be a measure space and T: X → X a measure-preserving
transformation. For a measurable function f : X → C set (UT f )(x) =
f (T(x)). The operator UT is linear and multiplicative: UT ( f · g) = UT f ·
UT g. Since T is measure-preserving, UT is an isometry of L p (X, A, µ) for
any p ≥ 1, i.e., UT f p = f p for any f ∈ L p (Exercise 4.5.3). If T is an
automorphism, then UT−1 = UT−1 is also an isometry, and hence UT is a uni-
tary operator on L2 (X, A, µ). We denote the scalar product on L2 (X, A, µ)
by  f, g, the norm by . , and the adjoint operator of U by U ∗ .
LEMMA 4.5.1. Let U be an isometry of a Hilbert space H. Then U f = f if
and only if U ∗ f = f .
Proof. For every f, g ∈ H we have U ∗ U f, g = U f, Ug =  f, g and hence
U ∗ U f = f . If U f = f , then (multiplying both sides by U ∗ )U ∗ f = f .
Conversely, if U ∗ f = f , then  f, U f  = U ∗ f, f  = f 2 and U f, f  =
 f, U ∗ f  = f 2 . Therefore U f − f, U f − f  = U f 2 −  f, U f  −
U f, f  + f 2 = 0. 

THEOREM 4.5.2 (von Neumann Ergodic Theorem). Let U be an isometry
of a separable Hilbert space H, and let P be orthogonal projection onto the
subspace I = { f ∈ H: U f = f } of U-invariant vectors in H. Then for every
f ∈H

1 n−1
lim U i f = P f.
n→∞ n
i=0

2 Several proofs in this section are due to F. Riesz; see [Hal60].

4.5. Ergodic Theorems 81

n−1 i
Proof. Let Un = n1 i=0 U and L = {g − Ug: g ∈ H}. Note that Land I are
n−1 i
U-invariant, and I is closed. If f = g − Ug ∈ L, then i=0 U f = g − Ung
and hence Un f → 0 as n → ∞. If f ∈ I, then Un f = f for all n ∈ N. We will
show that L ⊥ I and H = L̄ ⊕ I, where L̄ is the closure of L.
Let { fk} be a sequence in L, and suppose fk → f ∈ L̄. Then Un f ≤
Un ( f − fk) + Un fk ≤ Un · f − fk + Un fk , and hence Un f → 0
as n → ∞.
Let ⊥ denote the orthogonal complement, and note that L̄⊥ = L⊥ . If
h ∈ L⊥ , then 0 = h, g − Ug = h − U ∗ h, g for all g ∈ H so that h = U ∗ h,
and hence Uh = h, by Lemma 4.5.1. Conversely (again using Lemma 4.5.1),
if h ∈ I, then h, g − Ug = h, g − U ∗ h, g = 0 for every g ∈ H, and hence
h ∈ L⊥ .
Therefore, H = L̄ ⊕ I, and limn→∞ Un is the identity on I and 0 on L̄.


The following theorem is an immediate corollary of the von Neumann
ergodic theorem.

THEOREM 4.5.3. Let T be a measure-preserving transformation of a finite

measure space (X, A, µ). For f ∈ L2 (X, A, µ), set

1
N−1
f N+ (x) = f (T n (x)).
N n=0

Then f N+ converges in L2 (X, A, µ) to a T-invariant function f¯.

 N−1
If T is invertible, then fN− (x) = N1 n=0 f (T −n (x)) also converges in
L (X, A, µ) to f .
2 ¯
Similarly, let T be a measure-preserving flow in a finite measure space
(X, A, µ). For a function f ∈ L2 (X, A, µ) set
 
1 τ 1 τ
fτ+ (x) = f (T t (x)) dt and fτ− (x) = f (T −t (x)) dt.
τ 0 τ 0

Then fτ+ and fτ− converge in L2 (X, A, µ) to a T-invariant function f¯. 


Our next objective is to prove a pointwise version of the preceding theo-
rem. First, we need a combinatorial lemma. If a1 , . . . , am are real numbers
and 1 ≤ n ≤ m, we say that ak is an n-leader if ak + · · · + ak+ p−1 ≥ 0 for some
p, 1 ≤ p ≤ n.

LEMMA 4.5.4. For every n, 1 ≤ n ≤ m, the sum of all n-leaders is non-

negative.
82 4. Ergodic Theory

Proof. If there are no n-leaders, the lemma is true. Otherwise, let ak be the
first n-leader, and p ≥ 1 be the smallest integer for which ak + · · · + ak+ p−1 ≥
0. If k ≤ j ≤ k + p − 1, then a j + · · · + ak+ p−1 ≥ 0, by the choice of p, and
hence a j is an n-leader. The same argument can be applied to the sequence
ak+ p , . . . , am, which proves the lemma. 

THEOREM 4.5.5 (Birkhoff Ergodic Theorem). Let T be a measure-preserving
transformation in a finite measure space (X, A, µ), and let f ∈ L1 (X, A, µ).
Then the limit

1 n−1
f¯(x) = lim f (T k(x))
n→∞ n
k=0

exists for a.e. x ∈ X, is µ-integrable and T-invariant, and satisfies

 
¯f (x) dµ = f (x) dµ.
X X

If, in addition, f ∈ L2 (X, A, µ), then by Theorem 4.5.3, f¯ is the orthogonal

projection of f to the subspace of T-invariant functions.
n−1
If T is invertible, then n1 k=0 f (T −k(x)) also converges almost everywhere
to f¯.
Similarly, let T be a measure-preserving flow in a finite measure space
(X, A, µ). Then
 
1 τ 1 τ
fτ+ (x) = f (T t (x)) dt and fτ− (x) = f (T −t (x)) dt
τ 0 τ 0
converge almosteverywhere to the same µ-integrable and T-invariant limit
function f¯, and X f (x) dµ = X f¯(x) dµ.

Proof. We consider only the case of a transformation. We assume without

loss of generality that f is real-valued. Let

A = {x ∈ X: f (x) + f (T(x)) + · · · + f (T k(x)) ≥ 0 for some k ∈ N0 }.


LEMMA 4.5.6 (Maximal Ergodic Theorem). A f (x) dµ ≥ 0.
k
Proof. Let An = {x ∈ X: i=0 f (T i (x)) ≥ 0 for some k, 0 ≤ k ≤ n}. Then


An ⊂ An+1 , A = n∈N  An and, by the dominated convergence theorem, it
suffices to show that An f (x) dµ ≥ 0 for each n.
Fix an arbitrary m ∈ N. Let sn (x) be the sum of the n-leaders in the
sequence f (x), f (T(x)), . . . , f (T m+n−1 (x)). For k ≤ m − 1, let Bk ⊂ X be
the set of points for which f (T k(x)) is an n-leader of this sequence. By
4.5. Ergodic Theorems 83

Lemma 4.5.4,
 
m+n−1
0≤ sn (x) dµ = f (T k(x)) dµ. (4.1)
X k=0 Bk

Note that x ∈ Bk if and only if T(x) ∈ Bk−1 . Therefore, Bk = T −1 (Bk−1 ) and

Bk = T −k(B0 ) for 1 ≤ k ≤ m − 1, and hence
  
f (T (x)) dµ =
k
f (T (x)) dµ =
k
f (x) dµ.
Bk T −k (B0 ) B0

Thus the first m terms in (4.1) are equal, and since B0 = An ,

 
m f (x) dµ + n | f (x)| dµ ≥ 0.
An X

Since m is arbitrary, the lemma follows. 


Now we can finish the proof of the Birkhoff ergodic theorem. For any
a, b ∈ R, a < b, the set
 
1 n−1
1
n−1
X(a, b) = x ∈ X: lim f (T i (x)) < a < b < lim f (T i (x))
n→∞ n i=0 n→∞ n
i=0

is measurable and T-invariant. We claim that  µ(X(a, b)) = 0. Apply

Lemma 4.5.6 to T| X(a,b) and f − b to obtain that X(a,b) ( f (x) − b) dµ ≥ 0.
 
Similarly, X(a,b) (a − f (x)) dµ ≥ 0, and hence X(a,b) (a − b) dµ ≥ 0. There-
fore µ(X(a, b)) = 0. Since a and b are arbitrary, we conclude that the aver-
n−1
ages n1 i=0 f (T i (x)) converge for a.e. x ∈ X.
n−1
For n ∈ N, let fn (x) = n1 i=0 f (T i (x)). Define f¯: X → R by f¯(x) =
limn→∞ fn (x). Then f is measurable, and fn converges a.e. to f¯. By Fatou’s
¯
lemma and invariance of µ,
 
lim | fn (x)|q dµ ≤ lim | fn (x)| dµ
X n→∞ n→∞ X
 
1 n−1
≤ lim | f (T j (x))| dµ = | f (x)| dµ
n→∞ n j=0 X X
 
Thus X | f¯(x)| dµ = X lim| fn (x)|dµ is finite, so f¯ is integrable.
The proof that X f (x) dµ = X f¯(x) dµ is left as an exercise (Exer-
cise 4.5.2). 

The following facts are immediate corollaries of Theorem 4.5.5 (Exer-
cise 4.5.4, Exercise 4.5.5).
84 4. Ergodic Theory

COROLLARY 4.5.7. A measure-preserving transformation T in a finite mea-

sure space (X, A, µ) is ergodic if and only if for each f ∈ L1 (X, A, µ)

1 n−1
1
lim f (T (x)) =
i
f (x) dµ for a.e.x, (4.2)
n→∞ n
i=0
µ(X) X

i.e., if and only if the time average equals the space average for every L1
function. 

The preceding corollary implies that to check the ergodicity of a measure-
preserving transformation, it suffices to verify (4.2) for a dense subset of
L1 (X, A, µ), e.g., for all continuous functions if X is a compact topological
space and µ is a Borel measure. Moreover, due to linearity it suffices to check
the convergence for a countable collection of functions that form a basis.
COROLLARY 4.5.8. A measure-preserving transformation T of a finite mea-
sure space (X, A, µ) is ergodic if and only if for every A ∈ A, for a.e. x ∈ X,

1 n−1
µ(A)
lim χ A(T k(x)) = ,
n→∞ n
k=0
µ(X)

where χ A is the characteristic function of A. 


Exercise 4.5.1. Let T be a measure-preserving transformation of a finite
measure space (X, A, µ). Prove that T is ergodic if and only if

1 n−1
 
lim µ T −k(A) ∩ B = µ(A) · µ(B)
n→∞ n
k=0

for any A, B ∈ A.
Exercise 4.5.2. Using the dominated convergence theorem, finish the

proof of Theorem 4.5.5 by showing that the averages n1 n−1
j=0 f converge
to f¯ in L1 .
Exercise 4.5.3. Prove that if T is a measure-preserving transformation,
then UT is an isometry of L p (X, A, µ) for any p ≥ 1.
Exercise 4.5.4. Prove Corollary 4.5.7.
Exercise 4.5.5. Prove Corollary 4.5.8.
Exercise 4.5.6. A real number x is said to be normal in base n if for any
k ∈ N, every finite word of length k in the alphabet {0, . . . , n − 1} appears
with asymptotic frequency n−k in the base-n expansion of x. Prove that almost
every real number is normal with respect to every base n ∈ N.
4.6. Invariant Measures for Continuous Maps 85

4.6 Invariant Measures for Continuous Maps

In this section, we show that a continuous map T of a compact metric space X
into itself has at least one invariant Borel probability measure. Every
 finite
Borel measure µ on X defines a bounded linear functional Lµ ( f ) = X f dµ
on the space C(X) of continuous functions on X; moreover, Lµ is positive
in the sense that Lµ ( f ) ≥ 0 if f ≥ 0. The Riesz representation theorem
[Rud87] states that the converse is also true: for every positive bounded
linear functional
 L on C(X), there is a finite Borel measure µ on X such
that L = X f dµ.

THEOREM 4.6.1 (Krylov–Bogolubov). Let X be a compact metric space and

T: X → X a continuous map. Then there is a T-invariant Borel probability
measure µ on X.
n−1
Proof. Fix x ∈ X. For a function f : X → R set S nf (x) = n1 i=0 f (T i (x)).
Let F ⊂ C(X) be a dense countable collection of continuous functions on X.
For any f ∈ F the sequence S nf (x) is bounded, and hence has a convergent
subsequence. Since F is countable, there is a sequence n j → ∞ such that
the limit

S∞
n
f (x) = lim S f (x)
j

j→∞

exists for every f ∈ F. For any g ∈ C(X) and any  > 0 there is f ∈ F such
that max y∈X |g(y) − f (y)| < . Therefore, for a large enough j,

nj

Sg (x) − S∞ (x)
≤ S n j (x) +
S n j (x) − S∞ (x)
≤ 2,
f |g− f | f f

n
so Sg j (x) is a Cauchy sequence. Thus, the limit Sg∞ (x) exists for every g ∈
C(X) and defines a bounded positive linear functional Lx on C(X). By the
Riesz representation
 theorem, there is a Borel probability measure µ such
that Lx (g) = X g dµ. Note that

nj

Sg (T(x)) − Sgn j (x)
= 1 |g(T n j (x)) − g(x)|.
nj

Therefore, Sg∞ (T(x)) = Sg∞ (x) and µ is T-invariant. 


Let M = M(x) denote the set of all Borel probability measures on X. A
∗
 measures µ
sequence of  n ∈ M converges in the weak topology to a measure
µ ∈ M if X f dµn → X f dµ for every f ∈ C(X). If µn is any sequence in
M and F ⊂ C(X) is a dense countable  subset, then, by a diagonal process,
there is a subsequence µn j such that X f dµn j converges for every f ∈ F,
86 4. Ergodic Theory


and hence the sequence X g dµn j converges for every g ∈ C(X). Therefore,
M is compact in the weak∗ topology. It is also convex: tµ + (1 − t)ν ∈ M
for any t ∈ [0, 1] and µ, ν ∈ M. A point in a convex set is extreme if it cannot
be represented as a non-trivial convex combination of two other points. The
extreme points of M are the probability measures supported on points; they
are called Dirac measures.
Let MT ⊂ M denote the set of all T-invariant Borel probability measures
on X. Then MT is closed, and therefore compact, in the weak∗ topology, and
convex.
Recall that if µ and ν are finite measures on a space X with σ-algebra
A, then ν is absolutely continuous with respect to µ if ν(A) = 0 whenever
µ(A) = 0, for A ∈ A. If ν is absolutely continuous with respect to µ, then
the Radon–Nikodym theorem asserts that there is an L1 function dν/dµ,
called the Radon–Nikodym derivative, such that ν(A) = A(dν/dµ)(x) dµ
for every A ∈ A [Roy88].

PROPOSITION 4.6.2. Ergodic T-invariant measures are precisely the ex-

treme points of MT .
Proof. If µ is not ergodic, then there is a T-invariant measurable sub-
set A ⊂ X with 0 < µ(A) < 1. Let µ A(B) = µ(B ∩ A)/µ(A) and µ X\A(B) =
µ(B ∩ (X \ A))/µ(X \ A) for any measurable set B. Then µ A and µ X\A are
T-invariant and µ = µ(A)µ A + µ(X \ A)µ X\A, so µ is not an extreme point.
Conversely, assume that µ is ergodic and that µ = tν + (1 − t)κ with
ν, κ ∈ MT and t ∈ (0, 1). Then ν is absolutely continuous with respect to
µ and ν(A) = A r dµ, where r = dν/dµ ∈ L1 (X, µ) is the Radon–Nikodym
derivative. Observe that r ≤ 1t almost everywhere. Therefore r ∈ L2 (X, µ).
Let U be the isometry of L2 (X, µ) given by U f = f ◦ T. Invariance of ν
implies that for every f ∈ L2 (X, µ)
 
U f, r µ = ( f ◦ T)r dµ = f r dµ =  f, r µ .

It follows that  f, U ∗r µ = U f, r µ =  f, r µ , and hence U ∗r = r . By

Lemma 4.5.1 Ur = r . Since µ is ergodic, the function r is essentially con-
stant, so µ = ν = κ. 

By the Krein–Milman theorem [Roy88], [Rud91], MT is the closed con-
vex hull of its extreme points. Therefore, the set MeT of all T-invariant,
ergodic, Borel probability measures is not empty. However, MeT may be
rather complicated; for example, it may be dense in MT in the weak∗ topol-
ogy (Exercise 4.6.5).
4.7. Unique Ergodicity and Weyl’s Theorem 87

Exercise 4.6.1. Describe MT and MeT for the homeomorphism of the

circle T(x) = x + a sin 2π x mod 1, 0 < a ≤ 2π
1
.

Exercise 4.6.2. Describe MT and MeT for the homeomorphism of the torus
T(x, y) = (x, x + y) mod 1.

Exercise 4.6.3
(a) Give an example of a map of the circle that is discontinuous at ex-
actly one point and does not have non-trivial finite invariant Borel
measures.
(b) Give an example of a continuous map of the real line that does not
have non-trivial finite invariant Borel measures.

Exercise 4.6.4. Let X and Y be compact metric spaces and T: X → Y a

continuous map. Show that T induces a natural map M(X) → M(Y), and
that this map is continuous in the weak∗ topology.

*Exercise 4.6.5. Prove that if σ is the two-sided 2-shift, then Meσ is dense
in Mσ in the weak∗ topology.

4.7 Unique Ergodicity and Weyl’s Theorem3

In this section T is a continuous map of a compact metric space X. By §4.6,
there are T-invariant Borel probability measures. If there is only one such
measure, then T is said to be uniquely ergodic. Note that this unique invariant
measure is necessarily ergodic by Proposition 4.6.2.
An irrational circle rotation is uniquely ergodic (Exercise 4.7.1). More-
over, any topologically transitive translation on a compact abelian group is
uniquely ergodic (Exercise 4.7.2). On the other hand, unique ergodicity does
not imply topological transitivity (Exercise 4.7.3).

PROPOSITION 4.7.1. Let X be a compact metric space. A continuous map

n−1
T: X → X is uniquely ergodic if and only if Snf = n1 i=0 f ◦ T i converges
∞
uniformly to a constant function S f for any continuous function f ∈ C(X).
Proof. Suppose first that T is uniquely ergodic and µ is the unique T-
invariant Borel probability measure. We will show that



lim max

Sn f (x) − f dµ

→ 0.
n→∞ x∈X X

3 The arguments of this section follow in part those of [Fur81a] and [CFS82].
88 4. Ergodic Theory

Assume, for a contradiction, that there are f ∈ C(X) and sequences xk ∈

X and nk → ∞ such that limk→∞ S nf k (xk) = c = X f dµ. As in the proof
of Proposition 4.6.1, there is a subsequence nki → ∞ such that the limit
nk
L(g) = limi→∞ S f i (xki ) exists for any g ∈ C(X). As in Proposition 4.6.1,
L defines a T-invariant, positive, bounded linear functional on C(X). By
the Riesz representation
 theorem, L(g) = X g dν for some ν ∈ MT . Since
L( f ) = c = X f dµ, the measures µ and ν are different, which contradicts
unique ergodicity.
The proof of the converse is left as an exercise (Exercise 4.7.4). 

Uniform convergence of the time averages of continuous functions does
not, by itself, imply unique ergodicity. For example, if (X, T) is uniquely
ergodic and I = [0, 1], then (X × I, T × Id) is not uniquely ergodic, but the
time averages converge uniformly for all continuous functions.
PROPOSITION 4.7.2. Let T be a topologically transitive continuous map
of a compact metric space X. Suppose that the sequence of time averages
Snf converges uniformly for every continuous function f ∈ C(X). Then T is
uniquely ergodic.
Proof. Since the convergence is uniform, S∞ f = limn→∞ S f is a continuous
n

function. As in the proof of Proposition 4.6.1, S f (T(x)) = S∞

∞
f (x) for every x.
Since T is topologically transitive, S∞f is constant. As in previous
 arguments,
the linear functional f →S f defines a measure µ ∈ MT with X f dµ = S∞
∞
f .
∞
Let
 ν ∈ M T . By the Birkhoff ergodic theorem (Theorem 4.5.5), S f (x) =
X f dν for every f ∈ C(X) and ν a.e. x ∈ X. Therefore, ν = µ. 

Let X be a compact metric space with a Borel probability measure µ.
Let T: X → X be a homeomorphism preserving µ. A point x ∈ X is called
generic for (X, µ, T) if for every continuous function f

1 n−1
lim f (T k(x)) = f dµ.
n→∞ n X
k=0

If T is ergodic, then by Corollary 4.5.8, µ-a.e. x is generic.

For a compact topological group G, the Haar measure on G is the unique
Borel probability measure invariant under all left and right translations. Let
T: X → X be a homeomorphism of a compact metric space, G a compact
group, and φ: X → G a continuous function. The homeomorphism S: X ×
G → X × G given by S(x, g) = (T(x), φ(x)g) is a group extension (or G-
extension) of T. Observe that S commutes with the right translations
Rg (x, h) = (x, hg). If µ is a T-invariant measure on X and m is the Haar mea-
sure on G, then the product measure µ × m is S-invariant (Exercise 4.7.7).
4.7. Unique Ergodicity and Weyl’s Theorem 89

PROPOSITION 4.7.3 (Furstenberg). Let G be a compact group with Haar

measure m, X a compact metric space with a Borel probability measure µ,
T: X → X a homeomorphism preserving µ, Y = X × G, ν = µ × m, and
S: Y → Y a G-extension of T. If T is uniquely ergodic and S is ergodic, then
S is uniquely ergodic.
Proof. Since ν is Rg -invariant for every g ∈ G, if (x, h) is generic for ν, then
(x, hg) is generic for ν. Since S is ergodic, ν-a.e. (x, h) is ν-generic. Therefore
for µ-a.e. x ∈ X, the point (x, h) is ν-generic for every h. If a measure ν  =ν
is S-invariant and ergodic, then ν  -a.e. (x, h) is ν  -generic. The points that
are ν  -generic cannot be ν-generic. Hence there is a subset N ⊂ X such that
µ(N) = 0 and the first coordinate x of every ν  -generic point (x, h) lies in
N. However, the projection of ν  to X is T-invariant and therefore is µ. This
is a contradiction. 

PROPOSITION 4.7.4. Let α ∈ (0, 1) be irrational, and let T: Tk → Tk be
defined by
T(x1 , . . . , xk) = (x1 + α, x2 + a21 x1 , . . . , xk + ak1 x1 + · · · ak k−1 xk−1 ),
where the coefficients ai j are integers and ai i−1 =0, i = 2, . . . , k. Then T is
uniquely ergodic.
Proof. By Exercise 4.7.8, T is ergodic with respect to Lebesgue measure on
Tk. An inductive application of Proposition 4.7.3 yields the result. 

Let X be a compact topological space with a Borel probability measure
µ. A sequence (xi )i∈N in X is uniformly distributed if for any continuous
function f on X,

1 n
lim f (xk) = f dµ.
n→∞ n X
k=1

THEOREM 4.7.5 (Weyl). If P(x) = bk x k + · · · + b0 is a real polynomial such

that at least one of the coefficients bi , i > 0, is irrational, then the sequence
(P(n) mod 1)n∈N is uniformly distributed in [0, 1].
Proof [Fur81a]. Assume first that bk = α/k! with α irrational. Consider the
map T: Tk → Tk given by
T(x1 , . . . , xk) = (x1 + α, x2 + x1 , . . . , xk + xk−1 ).
Let π: Rk → Tk be the projection. Let Pk(x) = P(x) and Pi−1 (x) =
Pi (x + 1) − Pi (x), i = k, . . . , 1. Then P1 (x) = αx + β. Observe that
T n (π(P1 (0), . . . , Pk(0))) = π (P1 (n), . . . , Pk(n)). Since T is uniquely ergodic
by Proposition 4.7.4, this orbit (and any other orbit) is uniformly distributed
90 4. Ergodic Theory

on Tk. It follows that the last coordinate Pk(n) = P(n) is uniformly dis-
tributed on S1 .
Exercise 4.7.9 finishes the proof. 

Exercise 4.7.1. Prove that an irrational circle rotation is uniquely ergodic.
Exercise 4.7.2. Prove that any topologically transitive translation on a com-
pact abelian group is uniquely ergodic.
Exercise 4.7.3. Prove that the diffeomorphism T: S1 → S1 defined by
T(x) = x + a sin2 (π x), a < 1/π, is uniquely ergodic but not topologically
transitive.
Exercise 4.7.4. Prove the remaining statement of Proposition 4.7.1.
Exercise 4.7.5. Prove that the subshift defined by a fixed point a of a prim-
itive substitution s is uniquely ergodic.
Exercise 4.7.6. Let T be a uniquely ergodic continuous transformation of
a compact metric space X, and µ the unique invariant Borel probability
measure. Show that supp µ is a minimal set for T.
Exercise 4.7.7. Let S: X × G → X × G be a G-extension of T: (X, µ) →
(X, µ), and let m be the Haar measure on G. Prove that the product measure
µ × m is S-invariant.
Exercise 4.7.8. Use Fourier series on Tk to prove that T from Proposi-
tion 4.7.4 is ergodic with respect to Lebesgue measure.
Exercise 4.7.9. Reduce the general case of Theorem 4.7.5 to the case where
the leading coefficient is irrational.

4.8 The Gauss Transformation Revisited4

Recall that the Gauss transformation (§1.6) is the map of the unit interval
to itself defined by
 
1 1
φ(x) = − for x ∈ (0, 1], φ(0) = 0.
x x
The Gauss measure µ defined by

1 dx
µ(A) = (4.3)
log 2 A1+x
is a φ-invariant probability measure on [0, 1].

4 The arguments of this section follow in part those of [Bil65].

4.8. The Gauss Transformation Revisited 91

For an irrational x ∈ (0, 1], the nth entry an (x) = [1/φ n−1 (x)] of the con-
tinued fraction representing x is called the n-th quotient, and we write
x = [a1 (x), a2 (x), . . .]. The irreducible fraction pn (x)/qn (x) that is equal to
the truncated continued fraction [a1 (x), . . . , an (x)] is called the nth conver-
gent of x. The numerators and denominators of the convergents satisfy the
following relations:

p0 (x) = 0, p1 (x) = 1, pn (x) = an (x) pn−1 (x) + pn−2 (x), (4.4)

q0 (x) = 1, q1 (x) = a1 (x), qn (x) = an (x)qn−1 (x) + qn−2 (x) (4.5)

for n > 1. We have

pn (x) + (φ n (x)) pn−1 (x)
x= .
qn (x) + (φ n (x))qn−1 (x)
By an inductive argument

pn (x) ≥ 2(n−2)/2 and qn (x) ≥ 2(n−1)/2 for n ≥ 2,

and

pn−1 (x)qn (x) − pn (x)qn−1 (x) = (−1)n , n ≥ 1. (4.6)

For positive integers bk, k = 1, . . . , n, let

b1 ,...,bn = {x ∈ (0, 1]: ak(x) = bk, k = 1, . . . , n}.

The interval b1 ,...,bn is the image of the interval [0, 1) under the map ψb1 ,...,bn
defined by

ψb1 ,...,bn (t) = [b1 , . . . , bn−1 , bn + t].

If n is odd, ψb1 ,...,bn is decreasing; if n is even, it is increasing. For x ∈ b1 ,...,bn

pn + t pn−1
x = ψb1 ,...,bn (t) = , (4.7)
qn + tqn−1
where pn and qn are given by the recursive relations (4.4) and (4.5) with
an (x) replaced by bn . Therefore
 
pn pn + pn−1
b1 ,...,bn = , if n is even,
qn qn + qn−1
and

pn + pn−1 pn
b1 ,...,bn = , if n is odd.
qn + qn−1 qn
 
If λ is Lebesgue measure, then λ b1 ,...,bn = (qn (qn + qn−1 ))−1 .
92 4. Ergodic Theory

PROPOSITION 4.8.1. The Gauss transformation is ergodic for the Gauss

measure µ.
Proof. For a measure ν and measurable sets A and B with ν(B) =0, let
ν(A|B) = ν(A ∩ B)/ν(B) denote the conditional measure. Fix b1 , . . . , bn ,
and let n = b1 ,...,bn , ψn = ψb1 ,...,bn . The length of n is ±(ψn (1) − ψn (0)),
and for 0 ≤ x < y ≤ 1,
λ({z: x ≤ φ n (z) < y} ∩ n ) = ±(ψn (y) − ψn (x)),
where the sign depends on the parity of n. Therefore
ψn (y) − ψn (x)
λ(φ −n ([x, y)) | n ) = ,
ψn (1) − ψn (0)
and by (4.6) and (4.7),
qn (qn + qn−1 )
λ(φ −n ([x, y)) | n ) = (y − x) · .
(qn + xqn−1 )(qn + yqn−1 )
The second factor in the right-hand side is between 1/2 and 2. Hence
1
λ([x, y)) ≤ λ(φ −n ([x, y)) | n ) ≤ 2λ([x, y)).
2
Since the intervals [x, y) generate the σ-algebra,
1
λ(A) ≤ λ(φ −n (A) | n ) ≤ 2λ(A) (4.8)
2
for any measurable set A ⊂ [0, 1].
Because the density of the Gauss measure µ is between 1/(2 log 2) and
1/ log 2,
1 1
λ(A) ≤ µ(A) ≤ λ(A).
2 log 2 log 2
By (4.8),
1
µ(A) ≤ µ(φ −n (A) | n ) ≤ 4µ(A)
4
for any measurable A ⊂ [0, 1].
Let A be a measurable φ-invariant set with µ(A) > 0. Then 14 µ(A) ≤
µ(A| n ), or, equivalently, 14 µ(n ) ≤ µ(n | A). Since the intervals n gen-
erate the σ-algebra, 14 µ(B) ≤ µ(B |A) for any measurable set B. By choosing
B = [0, 1]\A we obtain that µ(A) = 1. 

The ergodicity of the Gauss transformation has the following number-
theoretic consequences.
4.8. The Gauss Transformation Revisited 93

PROPOSITION 4.8.2. For almost every x ∈ [0, 1] (with respect to µ measure

or Lebesgue measure), we have the following:
1. Each integer k ∈ N appears in the sequence a1 (x), a2 (x), . . . with asymp-
totic frequency

1 k+ 1
log .
log 2 k

1
2. lim (a1 (x) + · · · + an (x)) = ∞.
n
n→∞
 ∞
 log k/ log 2
1
3. lim a1 (x)a2 (x) · · · an (x) =
n 1+ .
n→∞
k=1
k2 + 2k
log qn (x) π2
4. lim =
n→∞ n 12 log 2
Proof. 1: Let f be the characteristic function of the semiopen interval
[1/k, 1/(k + 1)). Then an (x) = k if and only if f (φ n (x)) = 1. By the Birkhoff
ergodic theorem, for almost every x,

   
1 n−1 1
1 1 1 k+ 1
lim f (φ i (x)) = f dµ = µ , = log ,
n→∞ n
i=0 0 k k+ 1 log 2 k

which proves the first assertion. 1

2: Let f (x) = [1/x], i.e., f (x) = a1 (x). Note that 0 f (x)/(1 + x) dx = ∞,
 1 (1−x)
since f (x) > (1 − x)/x and 0 x(1+x) dx = ∞. For N > 0, define


f (x) if f (x) ≤ N,
f N (x) =
0 otherwise.

Then, for any N > 0, for almost every x,

1 n−1
1 n−1
lim f (φ k(x)) ≥ lim f N (φ k(x))
n→∞ n k=0 n→∞ n k=0

1
n−1
= lim f N (φ k(x))
n→∞ n
k=0
 1
1 f N (x)
= dx.
log 2 0 1 + x
1 f N (x)
Since lim N→∞ 0 1+x
dx → ∞, the conclusion follows.
94 4. Ergodic Theory

3: Let f (x) = log a1 (x) = log([ x1 ]). Then f ∈ L1 ([0, 1]) with respect to the
Gauss measure µ (Exercise 4.8.1). By the Birkhoff ergodic theorem,
 1
1 n
1 f (x)
lim log ak(x) = dx
n→∞ n
k=1
log 2 0 1 + x
∞  k1
1 log k
= dx
log 2 k=1 k+1 1 1+x
∞ 
log k 1
= · log 1 + 2 .
k=1
log 2 k + 2k

Exponentiating this expression gives part 3.

4: Note that pn (x) = qn−1 (φ(x)) (Exercise 4.8.2), so
1 pn (x) pn−1 (φ(x)) p1 (φ n−1 (x))
= ··· .
qn (x) qn (x) qn−1 (φ(x)) q1 (φ n−1 (x))
Thus

1 1 n−1
pn−k(φ k(x))
− log qn (x) = log
n n k=0 qn−k(φ k(x))
n−1 
1 n−1
1 pn−k(φ k(x))
= log(φ k(x)) + log − log(φ k
(x)) . (4.9)
n k=0 n k=0 qn−k(φ k(x))

It follows from the Birkhoff 1 Ergodic Theorem that the first term of (4.9)
converges a.e. to (1/ log 2) 0 log x/(1 + x) dx = −π 2 /12. The second term
converges to 0 (Exercise 4.8.2). 

Exercise 4.8.1. Show that log([1/x]) ∈ L1 ([0, 1]) with respect to the Gauss
measure µ.
Exercise 4.8.2. Show that pn (x) = qn (φ(x)) and that
n−1 
1 pn−k(φ k(x))
lim log(φ (x)) − log
k
= 0.
n→∞ n
k=0
qn−k(φ k(x))

4.9 Discrete Spectrum

Let T be an automorphism of a probability space (X, A, µ). The operator
UT : L2 (X, A, µ) → L2 (X, A, µ) is unitary, and each of its eigenvalues is a
complex number of absolute value 1. Denote by T the set of all eigenval-
ues of UT . Since constant functions are T-invariant, 1 is an eigenvalue of
UT . Any T-invariant function is an eigenfunction of UT with eigenvalue 1.
4.9. Discrete Spectrum 95

Therefore, T is ergodic if and only if 1 is a simple eigenvalue of UT . If f, g

are two eigenfunctions with different eigenvalues σ =κ, then  f, g = 0,
since  f, g = UT f, UT g = σ κ̄ f, g. Note that UT is a multiplicative oper-
ator, i.e., UT ( f · g) = UT ( f ) · UT (g), which has important implications for
its spectrum.

PROPOSITION 4.9.1. T is a subgroup of the unit circle S1 = {z ∈ C: |z| =

1}. If T is ergodic, then every eigenvalue of UT is simple.
Proof. If σ ∈ T and f (T(x)) = σ f (x), then f¯(T(x)) = σ̄ f¯(x), and hence
σ̄ = σ −1 ∈ T . If σ1 , σ2 ∈ T and f1 (T(x)) = σ1 f1 (x), f2 (T(x)) = σ2 f2 (x),
then f = f1 f2 has eigenvalue σ1 σ2 , and hence σ1 σ2 ∈ T . Therefore, T is a
subgroup of S1 .
If T is ergodic, the absolute value of any eigenfunction f is essentially
constant (and non-zero). Thus, if f and g are eigenfunctions with the same
eigenvalue σ , then f/g is in L2 and is an eigenfunction with eigenvalue
1, so it is essentially constant by ergodicity. Therefore every eigenvalue is
simple. 

An ergodic automorphism T has discrete spectrum if the eigenfunctions
of UT span L2 (X, A, µ). An automorphism T has continuous spectrum if 1 is
a simple eigenvalue of UT and UT has no other eigenvalues.
Consider a circle rotation Rα (x) = x + α mod 1, x ∈ [0, 1). For each n ∈
Z, the function fn (x) = exp(2πinx) is an eigenfunction of URα with eigen-
value 2π nα. If α is irrational, the eigenfunctions fn span L2 , and hence Rα
has discrete spectrum. On the other hand, every weak mixing transformation
has continuous spectrum (Exercise 4.9.1).
Let G be an abelian topological group. A character is a continuous ho-
momorphism χ : G → S 1 . The set of characters of G with the compact–open
topology forms a topological group Ĝ called the group of characters (or the
dual group). For every g ∈ G, the evaluation map χ → χ(g) is a character
ˆ ˆ
ιg ∈ Ĝ, the dual of Ĝ, and the map ι: G → Ĝ is a homomorphism. If ιg (χ) ≡ 1,
then χ(g) = 1 for every χ ∈ Ĝ, and hence ι is injective. By the Pontryagin
duality theorem [Hel95], ι is also surjective and Ĝˆ ∼ = G. Moreover, if G is
discrete, Ĝ is a compact abelian group, and conversely.
For example, each character χ ∈ Ẑ is completely determined by the value
χ (1) ∈ S1 . Therefore Ẑ ∼= S1 . On the other hand, if λ ∈ Ŝ1 , then λ: S1 → S1
is a homomorphism, so λ(z) = zn for some n ∈ Z. Therefore, Ŝ1 = Z.
On a compact abelian group G with Haar measure λ, every character is
in L∞ , and therefore in L2 . The integral of any non-trivial character with
respect to Haar measure is 0 (Exercise 4.9.3). If σ and σ  are characters of
96 4. Ergodic Theory

G, then σ σ  is also a character. If σ and σ  are different, then

 

σ, σ  = σ (g)σ (g) dλ(g) = (σ σ  )(g) dλ(g) = 0.

G G

Thus the characters of G are pairwise orthogonal in L2 (G, λ).

THEOREM 4.9.2. For every countable subgroup  ⊂ S1 there is an ergodic

automorphism T with discrete spectrum such that T = .
Proof. The identity character Id:  → S1 , Id(σ ) = σ , is a character of .
Let T: ˆ → ˆ be the translation χ → χ · Id. The normalized Haar measure
λ on ˆ is invariant under T. For σ ∈ , let fσ ∈ ˆˆ be the character of ˆ such
that fσ (χ ) = χ(σ ). Since

UT fσ (χ ) = fσ (χId) = fσ (χ ) fσ (Id) = σ fσ (χ ),

fσ is an eigenfunction with eigenvalue σ .

We claim that the linear span A of the set of characters { fσ : σ ∈ }, is
dense in L2 (,
ˆ λ), which will complete the proof. The set of characters sep-
arates points of ,
ˆ is closed under complex conjugation, and contains the
constant function 1. Since the set of characters is closed under multiplica-
tion, A is closed under multiplication, and is therefore an algebra. By the
Stone–Weierstrass theorem [Roy88], A is dense in C(, C), and therefore
in L2 (,
ˆ λ). 

The following theorem (which we do not prove) is a converse to
Theorem 4.9.2.

THEOREM 4.9.3 (Halmos–von Neumann). Let T be an ergodic automor-

phism with discrete spectrum, and let  ⊂ S1 be its spectrum. Then T is
isomorphic to the translation on ˆ by the identity character Id:  → S1 .

A measure-preserving transformation T: (X, A, µ) → (X, A, µ) is aperi-

odic if µ({x ∈ X: T n (x) = x}) = 0 for every n ∈ N.
Theorem 4.9.4 (which we do not prove) implies that every aperiodic trans-
formation can be approximated by a periodic transformation with an arbi-
trary period n. Many of the examples and counterexamples in abstract er-
godic theory are constructed using the method of cutting and stacking based
on this theorem.

THEOREM 4.9.4 (Rokhlin–Halmos [Hal60]). Let T be an aperiodic auto-

morphism of a Lebesgue probability space (X, A, µ). Then for every n ∈ N
and  > 0 there is a measurable subset A = A(n, ) ⊂ X such that the sets

n−1 i
T i (A), i = 0, . . . , n − 1, are pairwise disjoint and µ(X \ i=0 T (A)) < .
4.10. Weak Mixing 97

Exercise 4.9.1. Prove that every weak mixing measure-preserving transfor-

mation has continuous spectrum.

Exercise 4.9.2. Suppose that α, β ∈ (0, 1) are irrational and α/β is irra-
tional. Let T be the translation of T2 given by T(x, y) = (x + α, y + β). Prove
that T is topologically transitive and ergodic and has discrete spectrum.

Exercise 4.9.3. Show that on a compact topological group G, the integral

of any non-trivial character with respect to the Haar measure is 0.

4.10 Weak Mixing5

The property of weak mixing is typical in the following sense. Since each
non-atomic probability Lebesgue space is isomorphic to the unit interval
with Lebesgue measure λ, every measure-preserving transformation can be
viewed as a transformation of [0, 1] preserving λ. The weak topology on
the set of all measure-preserving transformations of [0, 1] is given by Tn →
T if λ(Tn (A) ! T(A)) → 0 for each measurable A ⊂ [0, 1]. Halmos showed
[Hal44] that a residual (in the weak topology) subset of transformations
are weak mixing. V. Rokhlin showed [Roh48] that the set of strong mixing
transformations is of first category (in the weak topology).
The weak mixing transformations, as Theorem 4.10.6 below shows, are
precisely those that have continuous spectrum. To show this we first prove
a splitting theorem for isometries in a Hilbert space.
We say that a sequence of complex numbers an , n ∈ Z is non-negative
definite if for each N ∈ N,

N
zk z̄mak−m ≥ 0
k,m=−N

for each finite sequence of complex numbers zk, −N ≤ k ≤ N.

For a (linear) isometry U in a separable Hilbert space H, denote by U ∗
the adjoint of U, and for n ≥ 0 set Un = U n and U−n = U ∗n .

LEMMA 4.10.1. For every v ∈ H, the sequence Un v, v is non-negative def-

inite.
Proof.
 2

N
N  N 
 
zk z̄mUk−mv, v = zk z̄mUkv, Umv =  zl Ul v  . 

k,m=−N k,m=−N
l=−N


5 The presentation of this section to a large extent follows §2.3 of [Kre85].

98 4. Ergodic Theory

4.10.2 (Wiener). For a finite measure ν on [0, 1) set ν̂k =

LEMMA
1 2πikx 
0 e ν(dx). Then limn→∞ n−1 n−1 k=0 |ν̂k| = 0 if and only if ν has no atoms.
 n−1
Proof. Observe that n−1 n−1 k=0 |ν̂k| → 0 if and only if n
−1
k=0 |ν̂k| → 0.
2

Now
n−1  1  1
1 n−1
1
|νk|2 = e2πikx ν(dx) e−2πiky ν(dy)
n k=0 n k=0 0 0
 1 1 
1 n−1 2πik(x−y)
= e ν(dx)ν(dy).
0 0 n k=0

The functions n−1 n−1k=0 exp(2πik(x − y)) are bounded in absolute value by
1 and converge to 1 for x = y and to 0 for x =y. Therefore the last integral
tends to the product measure ν × ν of the diagonal of [0, 1) × [0, 1). It fol-
lows that
1 n−1
lim |νk|2 = (ν({x}))2 .
n→∞ n
k=0 0≤x<1 


For a (linear) isometry U of a separable Hilbert space H, set

 
1 n−1
 
Hw (U) = v ∈ H: lim |U v, v | = 0 for each v ∈ H ,
k
n→∞ n
k=0

and denote by He (U) the closure of the subspace spanned by the eigenvectors
of U. Both Hw (U) and He (U) are closed and U-invariant.
PROPOSITION 4.10.3. Let U be a (linear) isometry of a separable Hilbert
space H. Then
1. For each v ∈ H, there is a unique finite measure νv on the interval [0, 1)
(called the spectral measure) such that for every n ∈ Z
 1
Un v, v = e2πinx νv (dx).
0

2. If v is an eigenvector of U with eigenvalue exp(2πiα), then νv consists

of a single atom at α of measure 1.
3. If v ⊥ He (U), then νv has no atoms and v ∈ Hw (U).

Proof. The first statement follows immediately from Lemma 4.10.1 and
the spectral theorem for isometries in a Hilbert space [Hel95], [Fol95]. The
second statement follows from the first (Exercise 4.10.3).
To prove the last statement let v ⊥ He and W = e−2πi x U. Applying the

von Neumann ergodic theorem 4.5.2, let u = limn→∞ n−1 n−1 k=0 W v. Then
k
4.10. Weak Mixing 99

Wu = u. By Proposition 4.10.3,

1 n−1
u, v = lim e−2πi xkU kv, v
n→∞ n
k=0
 1
1 n−1 −2πi(x−y)k
= lim e νv (dy) = νv (x).
n→∞ 0 n
k=0

If νv (x) > 0, then u is a non-zero eigenvector of U with eigenvalue e2πi x

and v ⊥u, which is a contradiction. Therefore νv (x) = 0 for each x, and
Lemma 4.10.2 completes the proof. 

For a finite subset B ⊂ N denote by |B| the cardinality of B. For a subset
A ⊂ N, define the upper density d(A)
¯ by
1
¯
d(A) = lim sup |A∩ [1, n]|.
n→∞ n

We say that a sequence bn converges in density to b and write d-limn bn = b

if there is a subset A ⊂ N such that d(A)
¯ = 0 and limn→∞, n/∈A bn = b.

LEMMA 4.10.4. If (bn ) is a bounded sequence, then d-limn bn = 0 if and only

n−1
if limn→∞ n1 k=0 |bn − b| = 0.

Proof. Exercise 4.10.1. 


The following splitting theorem is an immediate consequence of Propo-
sition 4.10.3.

THEOREM 4.10.5 (Koopman–von Neumann [KvN32]). Let U be an isom-

etry in a separable Hilbert space H. Then H = He ⊕ Hw . A vector v ∈ H
lies in Hw (U) if and only if d-limn U n v, v = 0, and if and only if d-limn
U n v, w = 0 for each w ∈ H.

Proof. The splitting follows from Proposition 4.10.3. To prove the remain-
ing statement that d-limU n v, v = 0 if and only if d-limU n v, w = 0, ob-
serve that U n v, w ≡ 0 if v ⊥ U kv for all k ∈ N. If w = U kv, then U n v, w =
U n v, U kv = U n−kv, v. 

Recall that if T and S are measure-preserving transformations in finite
measure spaces (X, A, µ) and (Y, B, ν), then T × S is a measure-preserving
transformation in the product space (X × Y, A × B, µ × ν). As in §4.9, we
denote by UT the isometry UT f (x) = f (T(x)) of L2 (X, A, µ).

THEOREM 4.10.6. Let T be a measure-preserving transformation of a prob-

ability space (X, A, µ). Then the following are equivalent:
100 4. Ergodic Theory

1. T is weak mixing.
2. T has continuous
 spectrum. 
3. d- limn X f (T n (x)) f (x) dµ = 0 if f ∈ L2 (X, A, µ) and X f dµ = 0.
  
4. d- limn X f (T n (x)) g(x) dµ = X f dµ · X g dµ for all functions
f, g ∈ L2 (X, A, µ).
5. T × T is ergodic.
6. T × S is weak mixing for each weak mixing S.
7. T × S is ergodic for each ergodic S.
Proof. The transformation T is weak mixing if and only if Hw (UT )
is the orthogonal complement of the constants in L2 (X, A, µ). Therefore,
by Proposition 4.10.3, 1 ⇔ 2. By Lemma 4.10.4, 1 ⇔ 3. Clearly 4 ⇒ 3. As-
sume that 3 holds. It is enough to show 4 for f with X f dµ = 0. Observe

that 4 holds for g satisfying X f (T k(x)) g(x) dµ = 0 for all k ∈ N. Hence it

suffices to consider g(x) = f (T k(x)). But X f (T n (x)) f (T k(x)) dµ =

X f (T
n−k
(x)) f (x) dµ → 0 as n → ∞ by 3. Therefore 3 ⇔ 4.
Assume 5. Observe that T is ergodic and if UT has an eigenfunction f ,
then | f | is T-invariant, and hence constant. Therefore f (x)/ f (y) is T × T-
invariant and 5 ⇒ 2. Clearly 6 ⇒ 2 and 7 ⇒ 5.
 (X × Y, A × B, µ × ν) is spanned
2
Assume 3. To prove 7 observe that L
by functions of the form f (x)g(y). Let X f dµ = Y g dν = 0. Then

f (T n (x))g(Sn (y)) f (x)g(y) dµ × ν
X×Y
 
= f (T n (x)) f (x) dµ · g(Sn (y))g(y) dν.
X Y

The first integral on the right-hand side converges in density to 0 by part 3,

while the second one is bounded. Therefore the product converges in density
to 0, and part 7 follows. The proof of 3 ⇒ 6 is similar (Exercise 4.10.4). 
Exercise 4.10.1. Let (bn ) be a bounded sequence. Prove that d-lim bn = b

if and only if limn→∞ n1 n−1
k=0 |bn − b| = 0.

Exercise 4.10.2. Prove that d-lim has the usual arithmetic properties of
limits.
Exercise 4.10.3. Prove the second statement of Proposition 4.10.3.
Exercise 4.10.4. Prove that 3 ⇒ 6 in Theorem 4.10.6.
Exercise 4.10.5. Let T be a weak mixing measure-preserving transforma-
tion, and let S be a measure-preserving transformation such that Sk = T for
some k ∈ N (S is called a kth root of T). Prove that S is weak mixing.
4.11. Measure-Theoretic Recurrence and Number Theory 101

4.11 Applications of Measure-Theoretic Recurrence

to Number Theory
In this section we give highlights of applications of measure-theoretic recur-
rence to number theory initiated by H. Furstenberg. As an illustration of
this approach we prove Sárközy’s Theorem (Theorem 4.11.5). Our exposi-
tion follows to a large extent [Fur77] and [Fur81a].
For a finite subset F ⊂ Z, denote by |F| the number of elements in F.
A subset D ⊂ Z has positive upper density if there are an , bn ∈ Z such that
bn − an → ∞ and for some δ > 0,
|D ∩ [an , bn ]|
>δ for all n ∈ N.
bn − an + 1
Let D ⊂ Z have positive upper density. Let ω D ∈ 2 = {0, 1}Z be the se-
quence for which (ω D)n = 1 if n ∈ A and (ω D)n = 0 if n ∈ / D, and let XD be
the closure of its orbit under the shift σ in 2 . Set YD = {ω ∈ XD : ω0 = 1}.

PROPOSITION 4.11.1 (Furstenberg). Let D ⊂ Z have positive upper den-

sity. Then there exists a shift-invariant Borel probability measure µ on XD
such that µ(YD) > 0.
Proof. By §4.6, every σ-invariant Borel probability measure on XD is a
linear functional L on the space C(XD) of continuous functions on XD such
that L( f ) ≥ 0 if f ≥ 0, L(1) = 1, and L( f ◦ σ ) = L( f ).
For a function f ∈ C(XD), set

1 bn
Ln ( f ) = f (σ i (ω D)),
bn − an + 1 i=an

where an , bn , and δ are associated with D as in the preceding paragraph.

Observe that Ln ( f ) ≤ max f for each n. Let ( f j ) j∈N be a countable dense
subset in C(XD). By a diagonal process, one can find a sequence nk → ∞
such that limk→∞ Lnk ( f j ) exists for each j. Since ( f j ) j∈N is dense in C(XD),
we have that

1
bnk
L( f ) = lim f (σ i (ω D))
k→∞ bnk − ank + 1
i=an k

exists for each f ∈ C(XD) and determines a σ-invariant Borel probability

measure µ.
Let χ ∈ C(XD) be the characteristic function of YD. Then

L(χ ) = χ dµ = µ(YD) > 0. 

102 4. Ergodic Theory

PROPOSITION 4.11.2. Let p(k) be a polynomial with integer coefficients

and p(0) = 0. Let U be an isometry of a separable Hilbert space H, and
Hrat ⊂ H be the closure of the subspace spanned by the eigenvectors of U
whose eigenvalues are roots of 1. Suppose v ∈ H is such that U p(k) v, v = 0
for all k ∈ N. Then v ⊥ Hrat .
Proof. Let v = vrat + w with vrat ∈ Hrat and w ⊥ Hrat . We use the following
lemma, whose proof is similar to the proof of Lemma 4.10.2 (Exercise 4.11.1).


lim n−1
LEMMA 4.11.3. n→∞ 1
n k=0 U p(k) w = 0 for all w ⊥ Hrat .
 
Fix  > 0, and let vrat ∈ Hrat and m be such that vrat − vrat <  and
 
U mvrat = vrat . Then U vrat − vrat < 2 for each k and, since p(mk) is di-
mk

visible by m,
 
1 n−1 
 
 U p(mk)
vrat − vrat  < 2.
 n k=0 
n−1
Since (1/n) k=0 U p(mk) w → 0 by Lemma 4.11.3, for n large enough we have
 
1 n−1 
 
 U p(mk)
v − vrat  < 2.
 n k=0 

By assumption, U p(mk) v, v = 0. Hence |vrat , v| < 2 v , so vrat , v = 0.



As a corollary of the preceding proposition we obtain Furstenberg’s poly-
nomial recurrence theorem.6

THEOREM 4.11.4 (Furstenberg). Let p(t) be a polynomial with integer co-

efficients and p(0) = 0. Let T be a measure-preserving transformation of a
finite measure space (X, A, µ), and A ∈ A be a set with positive measure. Then
there is n ∈ N such that µ(A∩ T p(n) A) > 0.
Proof. Let U be the isometry induced by T in H = L2 (X, A, µ), (Uh)(x) =
h(T −1 (x)). If µ(A ∩ T p(n) A) = 0 for each n ∈ N, then the characteristic func-
tion χ A of A satisfies U p(n) χ A, χ A = 0 for each n. By Proposition 4.11.2,
χ A is orthogonal to all eigenfunctions of U whose eigenvalues are roots
of 1. However 1(x) ≡ 1 is an eigenfunction of U with eigenvalue 1 and
1, χ A = µ(A) =0. 


6 A slight modification of the arguments above yields Proposition 4.11.2 and Theorem 4.11.4
for polynomials with integer values at integer points (rather than integer coefficients).
4.12. Internet Search 103

Theorem 4.11.4 and Proposition 4.11.1 imply the following result in com-
binatorial number theory.

THEOREM 4.11.5 (Sárközy [Sár78]). Let D ⊂ Z have positive upper density,

and let p be a polynomial with integer coefficients and p(0) = 0. Then there
are x, y ∈ D and n ∈ N such that x − y = p(n).

The following extension of the Poincaré recurrence theorem (whose

proof is beyond the scope of this book) was used by Furstenberg to give an
ergodic-theoretic proof of the Szemerédi theorem on arithmetic progres-
sions (Theorem 4.11.7).

THEOREM 4.11.6 (Furstenberg’s Multiple Recurrence Theorem [Fur77]).

Let T be an automorphism of a probability space (X, A, µ). Then for every
n ∈ N and every A ∈ A with µ(A) > 0 there is k ∈ N such that

µ(A ∩ T −k(A) ∩ T −2k(A) ∩ · · · ∩ T −nk(A)) > 0.

THEOREM 4.11.7 (Szemerédi [Sze69]). Every subset D ⊂ Z of positive up-

per density contains arbitrarily long arithmetic progressions.
Proof. Exercise 4.11.3. 

Exercise 4.11.1. Prove Lemma 4.11.3.

Exercise 4.11.2. Use Theorem 4.11.4 and Proposition 4.11.1 to prove

Theorem 4.11.5.

Exercise 4.11.3. Use Proposition 4.11.1 and Theorem 4.11.6 to prove

Theorem 4.11.7.

4.12 Internet Search7

In this section, we describe a surprising application of ergodic theory to the
problem of searching the Internet. This approach is is used by the Internet
search engine GoogleTM (www.google.com).
The Internet offers enormous amounts of information. Looking for infor-
mation on the Internet is analogous to looking for a book in a huge library
without a catalog. The task of locating information on the web is performed
by search engines. The first search engines appeared in the early 1990s. The
most popular engines handle tens of millions of searches per day.

7 The exposition in this section follows to a certain extent that of [BP98].

104 4. Ergodic Theory

The main tasks performed by a search engine are: gathering information

from web pages; processing and storing this information in a database; and
producing from this database a list of web pages relevant to a query consisting
of one or more words. The gathering of information is performed by robot
programs called crawlers that “crawl” the web by following links embedded
in web pages. Raw information collected by the crawlers is parsed and coded
by the indexer, which produces, for each web page, a set of word occurrences
(including word position, font type, and capitalization) and records all links
from this web page to other pages, thus creating the forward index. The sorter
rearranges information by words (rather than web pages), thus creating the
inverted index. The searcher uses the inverted index to answer the query, i.e.,
to compile a list of documents relevant to the keywords and phrases of the
query.
The order of the documents on the list is extremely important. A typical
list may contain tens of thousands of web pages, but at best only the first sev-
eral dozen may be reviewed by the user. Google uses two characteristics of
the web page to determine the order of the returned pages – the relevance of
the document to the query and the PageRank of the web page. The relevance
is based on the relative position, fontification, and frequency of the key-
word(s) in the document. This factor by itself often does not produce good
search results. For example, a query on the word “Internet” in one of the early
search engines returned a list whose first entry was a web page in Chinese
containing no English words other than “Internet.” Even now, many search
engines return barely relevant results when searching on common terms.
Google uses Markov chains to rank web pages. The collection of all web
pages and links between them is viewed as a directed graph G in which the
web pages serve as vertices and the links as directed edges (from the web
page on which they appear to the web page to which they point). At the
moment there are about 1.5 billion web pages with about 10 times as many
links. We number the vertices with positive integers i = 1, 2, . . . , N. Let G̃
be the graph obtained from G by adding a vertex 0 with edges to and from
all other vertices. Let bi j = 1 if there is an edge from vertex i to vertex j
in G̃, and let O(i) be the number of outgoing edges adjacent to vertex i
in G̃. Note that O(i) > 0 for all i. Fix a damping parameter p ∈ (0, 1) (for
example, p = .75). Set Bii = 0 for i ≥ 0. For i, j > 0 and i = j set
 
1 1 if O(i) = 1, 0 if bi j = 0,
B0i = , Bi0 = Bi j = p
N 1− p if O(i) =1, O(i)
if bi j = 1.

The matrix B is stochastic and primitive. Therefore, by Corollary 3.3.3,

it has a unique positive left eigenvector q with eigenvalue 1 whose entries
4.12. Internet Search 105

add up to 1. The pair (B, q) is a Markov chain on the vertices of G̃. Google
interprets qi as the PageRank of web page i and uses it together with the
relevance factor of the page to determine how high on the return list this
page should be.
For any initial probability distribution q on the vertices of G̃, the sequence
 n
q B converges exponentially to q. Thus one can find an approximation for
q by computing pBn , where q is the uniform distribution. This approach to
finding q is computationally much easier than trying to find an eigenvector
for a matrix with 1.5 billion rows and columns.
Exercise 4.12.1. Let A be an N × N stochastic matrix, and let Ainj be the
entries of An , i.e., Ainj is the probability of going from i to j in exactly n steps
(§4.4). Suppose q is an invariant probability distribution, q A = q.
(a) Suppose that for some j, we have Ai j = 0 for all i = j, and Anjk > 0
for some k = j and some n ∈ N. Show that q j = 0.
(b) Prove that if Ai j > 0 for some j =i and Anji = 0 for all n ∈ N, then
qi = 0.
CHAPTER FIVE

Hyperbolic Dynamics

In Chapter 1, we saw several examples of dynamical systems that were locally

linear and had complementary expanding and/or contracting directions: ex-
panding endomorphisms of S1 , hyperbolic toral automorphisms, the horse-
shoe, and the solenoid. In this chapter, we develop the theory of hyperbolic
differentiable dynamical systems, which include these examples. Locally, a
differentiable dynamical system is well approximated by a linear map –
its derivative. Hyperbolicity means that the derivative has complementary
expanding and contracting directions.
The proper setting for a differentiable dynamical system is a differen-
tiable manifold with a differentiable map, or flow. A detailed introduction
to the theory of differentiable manifolds is beyond the scope of this book.
For the convenience of the reader, we give a brief formal introduction to
differentiable manifolds in §5.13, and an even briefer informal introduction
here.
For the purposes of this book, and without loss of generality (see the em-
bedding theorems in [Hir94]), it suffices to think of a differentiable manifold
Mn as an n-dimensional differentiable surface, or submanifold, in R N , N > n.
The implicit function theorem implies that each point in M has a local coordi-
nate system that identifies a neighborhood of the point with a neighborhood
of 0 in Rn . For each point x on such a surface M ⊂ R N , the tangent space
Tx M ⊂ R N is the space of all vectors tangent to M at x. The standard inner
product on R N induces an inner product ·, ·x on each Tx M. The collection
of inner products is called a Riemannian metric, and a manifold M together
with a Riemannian metric is called a Riemannian manifold. The (intrinsic)
distance d between two points in M is the infimum of the lengths of differ-
entiable curves in M connecting the two points.
A one-to-one differentiable mapping with a differentiable inverse is called
a diffeomorphism.
106
5.1. Expanding Endomorphisms Revisited 107

A discrete-time differentiable dynamical system on a differentiable man-

ifold M is a differentiable map f : M → M. The derivative d fx is a linear map
from Tx M to Tf (x) M. In local coordinates d fx is given by the matrix of partial
derivatives of f . A continuous-time differentiable dynamical system on M is
a differentiable flow, i.e., a one-parameter group { f t }, t ∈ R, of differentiable
maps f t : M → M that depend differentiably on t. Since f −t ◦ f t = Id, each
map f t is a diffeomorphism. The derivative


d t


v(·) = f (·)

dt t=0

is a differentiable vector field tangent to M, and the flow { f t } is the one-

parameter group of time-t maps of the differential equation ẋ = v(x).
Differentiability, and even subtle differences in the degree of differen-
tiability, have important and sometimes surprising consequences. See, for
example, Exercise 2.5.7 and §7.2.

5.1 Expanding Endomorphisms Revisited

To illustrate and motivate some of the main ideas of this chapter we con-
sider again expanding endomorphisms of the circle Em x = mx mod 1, x ∈
[0, 1), m > 1, introduced in §1.3.
Fix  < 1/2. A finite or infinite sequence of points (xi ) in the circle is
called an -orbit of Em if d(xi+1 , Em xi ) <  for all i. The point xi has m
preimages under Em that are uniformly spread on the circle. Exactly one
of them, yii−1 , is closer than /m to xi−1 . Similarly, yii−1 has m preimages
under Em; exactly one of them, yii−2 , is closer than /m to xi−2 . Continuing
j
in this manner, we obtain a point yi0 with the property that d(Em yi0 , x j ) < 
for 0 ≤ j ≤ i. In other words, any finite -orbit of Em can be approximated
∞
or shadowed by a real orbit. If (xi )i=0 is an infinite -orbit, then the limit
y = limi→∞ yi exists (Exercise 5.1.1), and d(Em
0 i
y, xi ) ≤  for i ≥ 0. Since two
different orbits of Em diverge exponentially, there can be only one shadowing
orbit for a given infinite -orbit. By construction, y depends continuously on
(xi ) in the product topology (Exercise 5.1.2).
The above discussion of the -orbits of Em is based solely on the uniform
forward expansion of Em. Similar arguments show that if f is C 1 -close to
Em, then each infinite -orbit (xi ) of f is shadowed by a unique real orbit of
f that depends continuously on (xi ) (Exercise 5.1.3).
Consider now f that is C 1 -close enough to Em. View each orbit ( f i (x)) as
an -orbit of Em. Let y = φ(x) be the unique point whose orbit (Em i
y) shad-
ows ( f (x)). By the above discussion, the map φ is a homeomorphism and
i
108 5. Hyperbolic Dynamics

Emφ(x) = φ( f (x)) for each x (Exercise 5.1.4). This means that any differen-
tiable map that is C 1 -close enough to Em is topologically conjugate to Em.
In other words, Em is structurally stable; see §5.5 and §5.11.
Hyperbolicity is characterized by local expansion and contraction, in com-
plementary directions. This property, which causes local instability of orbits,
surprisingly leads to the global stability of the topological pattern of the
collection of all orbits.

Exercise 5.1.1. Prove that limi→∞ yi0 exists.

Exercise 5.1.2. Prove that limi→∞ yi0 depends continuously on (xi ) in the
product topology.

Exercise 5.1.3. Prove that if f is C 1 -close to Em, then each infinite -orbit
(xi ) of f is approximated by a unique real orbit of f that depends continu-
ously on (xi ).

Exercise 5.1.4. Prove that φ is a homeomorphism conjugating f and Em.

5.2 Hyperbolic Sets

In this section, M is a C 1 Riemannian manifold, U ⊂ M a non-empty open
subset, and f : U → f (U) ⊂ M a C 1 diffeomorphism. A compact, f -invariant
subset  ⊂ U is called hyperbolic if there are λ ∈ (0, 1), C > 0, and families
of subspaces Es (x) ⊂ Tx M and Eu (x) ⊂ Tx M, x ∈ , such that for every
x ∈ ,

1. Tx M = Es (x) ⊕ Eu (x),
2. d fxn v s  ≤ Cλn v s  for every v s ∈ Es (x) and n ≥ 0,
3. d fx−n v u  ≤ Cλn v u  for every v u ∈ Eu (x) and n ≥ 0,
4. d fx Es (x) = Es ( f (x)) and d fx Eu (x) = Eu ( f (x)).

The subspace Es (x) (respectively, Eu (x)) is called the stable (unstable)

subspace at x, and the family {Es (x)}x∈ ({Eu (x)}x∈ ) is called the stable (un-
stable) distribution of f | . The definition allows the two extreme cases
Es (x) = {0} or Eu (x) = {0}.
The horseshoe (§1.8) and the solenoid (§1.9) are examples of hyperbolic
sets. If  = M, then f is called an Anosov diffeomorphism. Hyperbolic toral
automorphisms (§1.7) are examples of Anosov diffeomorphisms. Any closed
invariant subset of a hyperbolic set is a hyperbolic set.

PROPOSITION 5.2.1. Let  be a hyperbolic set of f . Then the subspaces

Es (x) and Eu (x) depend continuously on x ∈ .
5.2. Hyperbolic Sets 109

Proof. Let xi be a sequence of points in  converging to x0 ∈ . Passing to

a subsequence, we may assume that dimEs (xi ) is constant. Let w1,i , . . . , wk,i
be an orthonormal basis in Es (xi ). The restriction of the unit tangent bundle
T 1 M to  is compact. Hence, by passing to a subsequence, w j,i converges
to w j,0 ∈ Tx10 M for each j = 1, . . . , k. Since condition 2 of the definition of a
hyperbolic set is a closed condition, each vector from the orthonormal frame
w1,0 , . . . , wk,0 satisfies condition 2 and, by the invariance (condition 4), lies
in Es (x0 ). It follows that dimEs (x0 ) ≥ k = dimEs (xi ). A similar argument
shows that dimEu (x0 ) ≥ dimEu (xi ). Hence, by (1), dimEs (x0 ) = dimEs (xi )
and dimEu (x0 ) = dimEu (xi ), and continuity follows. 

Any two Riemannian metrics on M are equivalent on a compact set, in the
sense that the ratios of the lengths of non-zero vectors are bounded above
and away from zero. Thus the notion of a hyperbolic set does not depend
on the choice of the Riemannian metric on M. The constant C depends on
the metric, but λ does not (Exercise 5.2.2). However, as the next proposition
shows, we can choose a particularly nice metric and C = 1 by using a slightly
larger λ.

PROPOSITION 5.2.2. If  is a hyperbolic set of f with constants C and λ,

then for every  > 0 there is a C 1 Riemannian metric ·, · in a neighborhood
of , called the Lyapunov, or adapted, metric (to f ), with respect to which f
satisfies the conditions of hyperbolicity with constants C  = 1 and λ = λ + ,
and the subspaces Es (x), Eu (x) are -orthogonal, i.e., v s , v u  <  for all unit
vectors v s ∈ Es (x), v u ∈ Eu (x), x ∈ .
Proof. For x ∈ , v s ∈ Es (x), and v u ∈ Eu (x), set

∞   
∞  
v s  = (λ + )−n d fxn v s , v u  = (λ + )−n d fx−n v u . (5.1)
n=0 n=0

Both series converge uniformly for v s , v u  ≤ 1 and x ∈ . We have


∞  
d fx v s  = (λ + )−n d fxn+1 v s  = (λ + )(v s  − v s ) < (λ + )v s  ,
n=0

and similarly for d fx−1 v u  . For v = v s + v u ∈ Tx M, x ∈ , define v =


(v s  )2 + (v u  )2 . The metric is recovered from the norm:
1   
v, w = (v + w 2 − v 2 − w 2 ).
2
With respect to this continuous metric, Es and Eu are orthogonal and f
satisfies the conditions of hyperbolicity with constant 1 and λ + . Now, by
110 5. Hyperbolic Dynamics

standard methods of differential topology [Hir94], ·, · can be uniformly

approximated on  by a smooth metric defined in a neighborhood of .


Observe that to construct an adapted metric it is enough to consider
sufficiently long finite sums instead of infinite sums in (5.1).
A fixed point x of a differentiable map f is called hyperbolic if no eigen-
value of d fx lies on the unit circle. A periodic point x of f of period k is
called hyperbolic if no eigenvalue of d fxk lies on the unit circle.

Exercise 5.2.1. Construct a diffeomorphism of the circle that satisfies the

first three conditions of hyperbolicity (with  being the whole circle) but
not the fourth condition.

Exercise 5.2.2. Prove that if  is a hyperbolic set of f : U → M for some

Riemannian metric on M, then  is a hyperbolic set of f for any other
Riemannian metric on M with the same constant λ.

Exercise 5.2.3. Let x be a fixed point of a diffeomorphism f . Prove that

{x} is a hyperbolic set if and only if x is a hyperbolic fixed point. Identify the
constants C and λ. Give an example when d fx has exactly two eigenvalues
µ ∈ (0, 1) and µ−1 , but λ =µ.

Exercise 5.2.4. Prove that the horseshoe (§1.8) is a hyperbolic set.

Exercise 5.2.5. Let i be a hyperbolic set of fi : Ui → Mi , i = 1, 2. Prove

that 1 × 2 is a hyperbolic set of f1 × f2 : U1 × U2 → M1 × M2 .

Exercise 5.2.6. Let M be a fiber bundle over N with projection π. Let U be

an open set in M, and suppose that  ⊂ U is a hyperbolic set of f : U → M
and that g: N → N is a factor of f . Prove that π () is a hyperbolic set of g.

Exercise 5.2.7. What are necessary and sufficient conditions for a periodic
orbit to be a hyperbolic set?

5.3
-Orbits
An -orbit of f : U → M is a finite or infinite sequence (xn ) ⊂ U such that
d( f (xn ), xn+1 ) ≤  for all n. Sometimes -orbits are referred to as pseudo-
orbits. For r ∈ {0, 1}, denote by distr the distance in the space of Cr -functions
(see §5.13).

THEOREM 5.3.1. Let  be a hyperbolic set of f : U → M. Then there is an

open set O ⊂ U containing  and positive 0 , δ0 with the following property:
5.3.
-Orbits 111

for every  > 0 there is δ > 0 such that for any g: O → M with dist1 (g, f ) < 0 ,
any homeomorphism h: X → X of a topological space X, and any continuous
map φ: X → O satisfying dist0 (φ ◦ h, g ◦ φ) < δ there is a continuous map
ψ: X → O with ψ ◦ h = g ◦ ψ and dist0 (φ, ψ) < . Moreover, ψ is unique in
the sense that if ψ  ◦ h = g ◦ ψ  for some ψ  : X → O with dist0 (φ, ψ  ) < δ0 ,
then ψ  = ψ.

Theorem 5.3.1 implies, in particular, that any collection of bi-infinite

pseudo-orbits near a hyperbolic set is close to a unique collection of genuine
orbits that shadow it (Corollary 5.3.2). Moreover, this property holds not
only for f itself but for any diffeomorphism C 1 -close to f . In the simplest
example, if X is a single point x (and h is the identity), Theorem 5.3.1 implies
the existence of a fixed point near h(x) for any diffeomorphism C 1 -close to f .
Proof.1 By the Whitney embedding theorem [Hir94], we may assume that
the manifold M is an m-dimensional submanifold in R N for some large N. For
y ∈ M, let Dα (y) be the disk of radius α centered at y in the (N − m)-plane
E⊥ (y) ⊂ R N that passes through y and is perpendicular to Ty M. Since  is
compact, by the tubular neighborhood theorem [Hir94], for any relatively
compact open neighborhood O of  in M there is α ∈ (0, 1) such that the α-
neighborhood Oα of O in R N is foliated by the disks Dα (y). For each z ∈ Oα
there is a unique point π(z) ∈ M closest to y, and the map π is the projection
to M along the disks Dα (y). Each map g: O → M can be extended to a map
g̃: Oα → M by

g̃(z) = g(π (z)).

Let C(X, Oα ) be the set of continuous maps from X to Oα with distance

dist0 . Note that Oα is bounded and φ ∈ C(X, Oα ). Let  be the Banach
space of bounded continuous vector fields v: X → R N with the norm v =
supx∈X v(x). The map φ  →φ  − φ is an isometry from the ball of radius
α centered at φ in C(X, Oα ) onto the ball Bα of radius α centered at 0 in .
Define : Bα →  by

((v))(x) = g̃(φ(h−1 (x)) + v(h−1 (x))) − φ(x), v ∈ Bα , x ∈ X.

If v is a fixed point of  and ψ(x) = φ(x) + v(x), then g̃(ψ(h−1 (x))) = ψ(x).
Observe that g̃(y) ∈ M and hence ψ(x) ∈ M for x ∈ X and g(ψ(h−1 (x))) =
ψ(x). Thus to prove the theorem it suffices to show that  has a unique fixed
point near φ, which depends continuously on g.

1 The main idea of this proof was communicated to us by A. Katok.

112 5. Hyperbolic Dynamics

The map  is differentiable as a map of Banach spaces, and the derivative

(dv w)(x) = d g̃ φ(h−1 (x))+v(h−1 (x)) w(h−1 (x))

is continuous in v. To establish the existence and uniqueness of a fixed point v

and its continuous dependence on g, we study the derivative of . By taking
the maximum of appropriate derivatives we obtain that (dv w)(x) ≤
Lw, where L depends on the first derivatives of g and on the embedding
but does not depend on X, h, and φ. For v = 0,

(d0 w)(x) = d g̃ φ(h−1 (x)) w(h−1 (x)).

Since  is a hyperbolic set, for some constants λ ∈ (0, 1) and C > 1, we have
for every y ∈  and n ∈ N
 n 
d f v  ≤ Cλn v if v ∈ Es (y), (5.2)
y
 −n 
d f v  ≤ Cλn v if v ∈ Eu (y). (5.3)
y

For z ∈ Oα , let T̃ z denote the m-dimensional plane through z that is or-

thogonal to the disk Dα (π(z)). The planes T̃ z form a differentiable distri-
bution on Oα . Note that T̃ z = Tz M for z ∈ O. Extend the splitting Ty M =
Es (y) ⊕ Eu (y) continuously from  to Oα (decreasing the neighborhood O
and α if necessary) so that Es (z) ⊕ Eu (z) = T̃ z and TzR N = Es (z) ⊕ Eu (z) ⊕
E⊥ (π(z)). Denote by Ps , Pu , and P⊥ the projections in each tangent space
TzR N onto Es (z), Eu (z), and E⊥ (π (z)), respectively.
Fix n ∈ N so that Cλn < 1/2. By (5.2)–(5.3) and continuity, for a small
enough α > 0 and small enough neighborhood O ⊃ , there is 0 > 0 such
that for every g with dist1 ( f, g) < 0 , every z ∈ Oα , and every v s ∈ Es (z), v u ∈
Eu (z), v ⊥ ∈ E⊥ (π(z)) we have

 s n s 1 s  u n s
 P d g̃ v  ≤ v ,  P d g̃ v  ≤ 1 v s , (5.4)
z z
2 100
 u n u  s n u
 P d g̃ v  ≥ 2v u ,  P d g̃ v  ≤ 1 v u , (5.5)
z z
100
d g̃ nz v ⊥ = 0. (5.6)

Denote

 ν = {v ∈ : v(x) ∈ Eν (φ(x)) for all x ∈ X}, ν = s, u, ⊥.

5.3.
-Orbits 113

The subspaces  s ,  u ,  ⊥ are closed and  =  s ⊕  u ⊕  ⊥ . By construction,

 ss 
A Asu 0
d0 =  Aus Auu 0  ,
0 0 0
where Ai j :  i →  j , i, j = s, u. By (5.4)–(5.6), there are positive 0 and δ
such that if dist1 ( f, g) < 0 and dist0 (φ ◦ h, g ◦ φ) < δ, then the spectrum of
d0 is separated from the unit circle. Therefore the operator d0 − Id is
invertible and
(d0 − Id)−1  < K,
where K depends only on f and φ.
As for maps of finite-dimensional linear spaces, (v) = (0) + d0 v +
H(v), where H(v1 ) − H(v2 ) ≤ C1 max{v1 , v2 } · v1 − v2  for some
C1 > 0 and small v1 , v2 . A fixed point v of  satisfies
F(v) = −(d0 − Id)−1 ((0) + H(v)) = v.
If ζ > 0 is small enough, then for any v1 , v2 ∈  with v1 , v2  < ζ ,
1
F(v1 ) − F(v2 ) < v1 − v2 .
2
Thus for an appropriate choice of constants and neighborhoods in the con-
struction, F:  →  is a contraction, and therefore has a unique fixed point,
which depends continuously on g. 

Theorem 5.3.1 implies that an -orbit lying in a small enough neighbor-
hood of a hyperbolic set can be globally (i.e., for all times) approximated by
a real f -orbit in the hyperbolic set. This property is called shadowing (the
real orbit shadows the pseudo-orbit). A continuous map f of a topological
space X has the shadowing property if for every  > 0 there is δ > 0 such
that every δ-orbit is -approximated by a real orbit.
For  > 0, denote by  the open -neighborhood of .
COROLLARY 5.3.2 (Anosov’s Shadowing Theorem). Let  be a hyperbolic
set of f : U → M. Then for every  > 0 there is δ > 0 such that if (xk) is a finite
or infinite δ-orbit of f and dist(xk, ) < δ for all k, then there is x ∈  with
dist( f k(x), xk) < .
Proof. Choose a neighborhood O satisfying the conclusion of Theorem
5.3.1, and choose δ > 0 such that δ ⊂ O. If (xk) is finite or semi-infinite,
add to (xk) the preimages of some y0 ∈  whose distance to the first point
of (xk) is < δ, and/or the images of some ym ∈  whose distance to the
114 5. Hyperbolic Dynamics

last point of (xk) is < δ, to obtain a doubly infinite δ-orbit lying in the δ-
neighborhood of . Let X = (xk) with discrete topology, g = f, h be the
shift xk →xk+1 , and φ: X → U be the inclusion, φ(xk) = xk. Since (xk) is a
δ-orbit, dist(φ(h(xk)), f (φ(xk)) < δ. Theorem 5.3.1 applies, and the corollary
follows. 

As in Chapter 2, denote by NW( f ) the set of non-wandering points, and
by Per( f ) the set of periodic points of f . If  is f -invariant, denote by
NW( f | ) the set of non-wandering points of f restricted to . In general,
NW( f | ) =NW( f ) ∩ .

PROPOSITION 5.3.3. Let  be a hyperbolic set of f : U → M. Then Per( f | )

= NW( f | ).
Proof. Fix  > 0 and let x ∈ NW( f | ). Choose δ from Theorem 5.3.1, and
let V = B(x, δ/2) ∩ . Since x ∈ NW( f | ), there is n ∈ N such that f n (V) ∩
V = ∅. Let z ∈ f −n ( f n (V) ∩ V) = V ∩ f −n (V). Then {z, f (z), . . . , f n−1 (z)}
is a δ-orbit, so by Theorem 5.3.1 there is a periodic point of period n within
2 of z. 

COROLLARY 5.3.4. If f : M → M is Anosov, then Per( f ) = NW( f ).

Exercise 5.3.1. Interpret Theorem 5.3.1 for X = Zm and h(z) = z + 1

mod m.

Exercise 5.3.2. Let  be a hyperbolic set of f : U → M. Prove that the

restriction f | is expansive.

Exercise 5.3.3. Let T: [0, 1] → [0, 1] be the tent map: T(x) = 2x for 0 ≤
x ≤ 1/2 and T(x) = 2(1 − x) for 1/2 ≤ x ≤ 1. Does T have the shadowing
property?

Exercise 5.3.4. Prove that a circle rotation does not have the shadowing
property. Prove that no isometry of a manifold has the shadowing property.

Exercise 5.3.5. Show that every minimal hyperbolic set consists of exactly
one periodic orbit.

5.4 Invariant Cones

Although hyperbolic sets are defined in terms of invariant families of linear
spaces, it is often convenient, and in more general settings even necessary,
to work with invariant families of linear cones instead of subspaces. In this
section, we characterize hyperbolicity in terms of families of invariant cones.
5.4. Invariant Cones 115

Let  be a hyperbolic set of f : U → M. Since the distributions Es and

u
E are continuous (Proposition 5.2.1), we extend them to continuous distri-
butions Ẽs and Ẽu defined in a neighborhood U() ⊃ . If x ∈ U() and
v ∈ Tx M, let v = v s + v u with v s ∈ Ẽs (x) and v u ∈ Ẽu (x). Assume that the
metric is adapted with constant λ. For α > 0, define the stable and unstable
cones of size α by

Kαs (x) = {v ∈ Tx M: v u  ≤ αv s },

Kαu (x) = {v ∈ Tx M: v s  ≤ αv u }.
◦
For a cone K, let K= int(K) ∪ {0}. Let  = {x ∈ U: dist(x, ) <  }.

PROPOSITION 5.4.1. For every α > 0 there is  = (α) > 0 such that f i ( )
⊂ U(), i = −1, 0, 1, and for every x ∈ 
◦ ◦
d fx Kαu (x) ⊂ K uα ( f (x)) and d f f−1
(x) Kα ( f (x)) ⊂ K α (x).
s s

Proof. The inclusions hold for x ∈ . The statement follows by continuity.



PROPOSITION 5.4.2. For every δ > 0 there are α > 0 and  > 0 such that
f i ( ) ⊂ U(), i = −1, 0, 1, and for every x ∈ 
 −1 
d f v  ≤ (λ + δ)v i f v ∈ Kαu (x),
x

and

d fx v ≤ (λ + δ)v if v ∈ Kαs (x).

Proof. The statement follows by continuity for a small enough α and  =

(α) from Proposition 5.4.1. 

The following proposition is the converse of Propositions 5.4.1 and 5.4.2.

PROPOSITION 5.4.3. Let  be a compact invariant set of f : U → M. Sup-

pose that there is α > 0 and for every x ∈  there are continuous subspaces
Ẽs (x) and Ẽu (x) such that Ẽs (x) ⊕ Ẽu (x) = Tx M, and the α-cones Kαs (x) and
Kαu (x) determined by the subspaces satisfy
1. d fx Kαu (x) ⊂ Kαu ( f (x)) and d f f−1
(x) Kα ( f (x)) ⊂ Kα (x), and
s s

2. d fx v < v for non-zero v ∈ Kαs (x), and d fx−1 v < v for non-
zero v ∈ Kαu (x).
Then  is a hyperbolic set of f .
116 5. Hyperbolic Dynamics

Proof. By compactness of  and of the unit tangent bundle of M, there is

a constant λ ∈ (0, 1) such that
 
d fx v ≤ λv for v ∈ Kαs (x) and d fx−1 v  ≤ λv for v ∈ Kαu (x).

For x ∈ , the subspaces

Es (x) = d f f−n s n
n (x) K ( f (x)) and Eu (x) = d f fn−n (x) Ku ( f −n (x))
n≥0 n≥0

satisfy the definition of hyperbolicity with constants λ and C = 1. 


Let

s = {x ∈ U: dist( f n (x), ) <  for all n ∈ N0 },

u = {x ∈ U: dist( f −n (x), ) <  for all n ∈ N0 }.

Note that both sets are contained in  and that f (s ) ⊂ s , f −1 (u ) ⊂ u .

PROPOSITION 5.4.4. Let  be a hyperbolic set of f with adapted metric.

Then for every δ > 0 there is  > 0 such that the distributions Es and Eu can
be extended to  so that
1. Es is continuous on s , and Eu is continuous on u ,
2. if x ∈  ∩ f ( ) then d fx Es (x) = Es ( f (x)) and d fx Eu (x) = Eu ( f (x)),
3. d fx v < (λ + δ)v for every x ∈  and v ∈ Es (x),
4. d fx−1 v < (λ + δ)v for every x ∈  and v ∈ Eu (x).
Proof. Choose  > 0 small enough that  ⊂ U(). For x ∈ s , let Es (x) =
limn→∞ d f f−n n (x) ( Ẽ ( f (x))). By Proposition 5.4.2, the limit exists if δ, α, and 
s n

are small enough. If x ∈  \s , let n(x) ∈ N be such that f n (x) ∈  for n =
−n(x)
0, 1, . . . , n(x) and f n(x)+1 (x) ∈ /  , and let Es (x) = d f f n (x) ( Ẽs ( f n(x) (x))).
The continuity of Es on s and the required properties follow from Proposi-
tion 5.4.2. A similar construction with f replaced by f −1 gives an extension
of Eu . 

Exercise 5.4.1. Prove that the solenoid (§1.9) is a hyperbolic set.

Exercise 5.4.2. Let  be a hyperbolic set of f . Prove that there is an open

set O ⊃  and  > 0 such that for every g with dist1 ( f, g) < , the invariant

set g = ∞ n
n=−∞ g (O) is a hyperbolic set of g.

Exercise 5.4.3. Prove that the topological entropy of an Anosov diffeo-

morphism is positive.
5.5. Stability of Hyperbolic Sets 117

Exercise 5.4.4. Let  be a hyperbolic set of f . Prove that if dim Eu (x) > 0
for each x ∈ , then f has sensitive dependence on initial conditions on 
(see §1.12).

5.5 Stability of Hyperbolic Sets

In this section, we use pseudo-orbits and invariant cones to obtain key prop-
erties of hyperbolic sets. The next two propositions imply that hyperbolicity
is “persistent.”

PROPOSITION 5.5.1. Let  be a hyperbolic set of f : U → M. There is an

open set U() ⊃  and 0 > 0 such that if K ⊂ U() is a compact invari-
ant subset of a diffeomorphism g: U → M with dist1 (g, f ) < 0 , then K is a
hyperbolic set of g.
Proof. Assume that the metric is adapted to f , and extend the distributions
Esf and Euf to continuous distributions Ẽsf and Ẽuf defined in an open neigh-
borhood U() of . For an appropriate choice of U(), 0 , and α, the stable
and unstable α-cones determined by Ẽsf and Ẽuf satisfy the assumptions of
Proposition 5.4.3 for the map g. 

Denote by Diff1 (M) the space of C 1 diffeomorphisms of M with the C 1
topology.

COROLLARY 5.5.2. The set of Anosov diffeomorphisms of a given compact

manifold is open in Diff1 (M).

PROPOSITION 5.5.3. Let  be a hyperbolic set of f : U → M. For every

open set V ⊂ U containing  and every  > 0, there is δ > 0 such that
for every g: V → M with dist1 (g, f ) < δ, there is a hyperbolic set K ⊂ V
of g and a homeomorphism χ: K →  such that χ ◦ g| K = f | ◦ χ and
dist0 (χ , Id) < .
Proof. Let X = , h = f | , and let φ:  → U be the inclusion. By
Theorem 5.3.1, there is a continuous map ψ:  → U such that ψ ◦ f | =
g ◦ ψ. Set K = ψ(). Now apply Theorem 5.3.1 to X = K, h = g| K , and the
inclusion φ: K → M to get ψ  : K → U with ψ  ◦ g| K = f | ◦ ψ. By unique-
ness, ψ −1 = ψ  . For a small enough δ, the map χ = ψ  is close to the identity,
and, by Proposition 5.5.1, K is hyperbolic. 

A C 1 diffeomorphism f of a C 1 manifold M is called structurally stable
if for every  > 0 there is δ > 0 such that if g ∈ Diff1 (M) and dist1 (g, f ) <
δ, then there is a homeomorphism h: M → M for which f ◦ h = h ◦ g and
118 5. Hyperbolic Dynamics

dist0 (h, Id) < . If one demands that the conjugacy h be C 1 , the definition
becomes vacuous. For example, if f has a hyperbolic fixed point x, then any
small enough perturbation g has a fixed point y nearby; if the conjugation
is differentiable, then the matrices dg y and d fx are similar. This condition
restricts g to lie in a proper submanifold of Diff1 (M).

COROLLARY 5.5.4. Anosov diffeomorphisms are structurally stable.

Exercise 5.5.1. Interpret Proposition 5.5.3 when  is a hyperbolic periodic

point of f .

5.6 Stable and Unstable Manifolds

Hyperbolicity is defined in terms of infinitesimal objects: a family of linear
subspaces invariant by the differential of a map. In this section, we construct
the corresponding integral objects, the stable and unstable manifolds.
For δ > 0, let Bδ = B(0, δ) ⊂ Rm be the ball of radius δ at 0.

PROPOSITION 5.6.1 (Hadamard–Perron). Let f = ( fn )n∈N0 , fn : Bδ → Rm,

be a sequence of C 1 diffeomorphisms onto their images such that fn (0) = 0.
Suppose that for each n there is a splitting Rm = Es (n) ⊕ Eu (n) and λ ∈ (0, 1)
such that
1. d fn (0)Es (n) = Es (n + 1) and d fn (0)Eu (n) = Eu (n + 1),
2. d fn (0)v s  < λv s  for every v s ∈ Es (n),
3. d fn (0)v u  > λ−1 v u  for every v u ∈ Eu (n),
4. the angles between Es (n) and Eu (n) are uniformly bounded away
from 0,
5. {d fn (·)}n∈N0 is an equicontinuous family of functions from Bδ to
GL(m, R).
Then there are  > 0 and a sequence φ = (φn )n∈N0 of uniformly Lipschitz
continuous maps φn : Bs = {v ∈ Es (n): v < } → Eu (n) such that
1. graph(φn ) ∩ B = Ws (n) := {x ∈ B :  fn+k−1 ◦ · · · ◦ fn+1 ◦ fn (x)
→k→∞ 0},
2. fn (graph(φn )) ⊂ graph(φn+1 ),
3. if x ∈ graph(φn ), then  fn (x) ≤ λx, so by (2), fnk(x) → 0 exponen-
tially as k → ∞,
4. for x ∈ B \graph(φn ),
 u     
 P fn (x) − φn+1 Ps fn (x)  > λ−1  Pu x − φn Ps x ,
n+1 n+1 n n

where Pns (Pnu ) denotes the projection onto Es (n)(Eu (n)) parallel to
Eu (n)(Es (n)),
5.6. Stable and Unstable Manifolds 119

5. φn is differentiable at 0 and dφn (0) = 0, i.e., the tangent plane to

graph(φn ) is Es (n).
6. φ depends continuously on f in the topologies induced by the following
distance functions:

d0 (φ, ψ) = sup 2−n |φn (x) − ψn (x)|,

n∈N0 , x∈B

d( f, g) = sup 2−n dist1 ( fn , gn ),

n∈N0

where dist1 is the C 1 distance.

Proof. For positive constants L and , let (L, ) be the space of se-
quences φ = (φn )n∈N0 , where φn : Bs → Eu (n) is a Lipschitz-continuous map
with Lipschitz constant L and φn (0) = 0. Define distance on (L, ) by
d(φ, ψ) = supn∈N0 , x∈B |φn (x) − ψn (x)|. This metric is complete.
We now define an operator F: (L, ) → (L, ) called the graph trans-
form. Suppose φ = (φn ) ∈ . We prove in the next lemma that for a small
enough , the projection of the set fn−1 (graph(φn+1 )) onto Es (n) covers
Es (n), and fn−1 (graph(φn+1 )) contains the graph of a continuous function
ψn : Bs → Eu (n) with Lipschitz constant L. We set F(φ)n = ψn .
Note that a map h: Rk → Rl is Lipschitz continuous at 0 with Lipschitz
constant L if and only if the graph of h lies in the L-cone about Rk, and is
Lipschitz continuous at x ∈ Rk if and only if its graph lies in the L-cone about
the translate of Rk by (x, h(x)).

LEMMA 5.6.2. For any L > 0, there exists  > 0 such that the graph transform
F is a well-defined operator on (L, ).
Proof. For L > 0 and x ∈ B , let KLs (n) denote the stable cone

KLs (n) = {v ∈ Rm: v = v s + v u , v s ∈ Es (n), v u ∈ Eu (n), |v u | ≤ L|v s |}.

Note that d fn−1 (0)KLs (n + 1) ⊂ KLs (n) for any L > 0. Therefore, by the uni-
form continuity of d fn , for any L > 0 there is  > 0 such that d fn−1 (x)KLs (n +
1) ⊂ KLs (n) for any n ∈ N0 and x ∈ B . Hence the preimage under fn of
the graph of a Lipschitz-continuous function is the graph of a Lipschitz-
continuous function. For φ ∈ (L, ), consider the following composition
β = Ps (n) ◦ fn−1 ◦ φn , where Ps (n) is the projection onto Es (n) parallel to
Eu (n). If  is small enough, then β is an expanding map and its image covers
Bs (n) (Exercise 5.6.1). Hence F(φ) ∈ (L, ). 

The next lemma shows that F is a contracting operator for an appropriate
choice of  and L.
120 5. Hyperbolic Dynamics

E u (n + 1) fn (cu )
E u (n) φn

fn
φn+1
cu
ψn+1
ψn
E s (n + 1)
E s (n)
0 y 0 z

Figure 5.1. Graph transform applied to φ and ψ.

LEMMA 5.6.3. There are  > 0 and L > 0 such that F is a contracting oper-
ator on (L, ).
Proof. For L ∈ (0, 0.1), let KLu (n) denote the unstable cone
KLu (n) = {v ∈ Rm: v = v s + v u , v s ∈ Es (n), v u ∈ Eu (n), |v u | ≥ L−1 |v s |},
and note that d fn (0)KLu (n) ⊂ KLu (n + 1). As in Lemma 5.6.2, by the uniform
continuity of d fn , for any L > 0 there is  > 0 such that the inclusion
d fn KLu (n) ⊂ KLu (n + 1) holds for every n ∈ N0 and x ∈ B .
Let φ, ψ ∈ (L, ), φ  = F(φ), ψ  = F(ψ) (see Figure 5.1). For any η > 0
there are n ∈ N0 and y ∈ Bs such that |φn (y) − ψn (y)| > d(φ  , ψ  ) − η. Let cu
be the straight line segment from (y, φn (y)) to (y, φn (y)). Since cu is parallel
to Eu (n), we have that length( fn (cu )) > λ−1 length(cu ). Let fn (y, ψn (y)) =
(z, ψn+1 (z)), and consider the curvilinear triangle formed by the straight line
segment from (z, φn+1 (z)) to (z, ψn+1 (z)), fn (cu ), and the shortest curve on
the graph of ψn+1 connecting the ends of these curves. For a small enough
 > 0 the tangent vectors to the image fn (cu ) lie in KLu (n + 1) and the tangent
vectors to the graph of φn+1 lie in KLs (n + 1). Therefore,
length( fn (cu ))
|φn+1 (z) − ψn+1 (z)| ≥ − L(1 + L) · length( fn (cu ))
1 + 2L
≥ (1 − 4L)length( fn (cu )),
and
d(φ, ψ) ≥ |φn+1 (z) − ψn+1 (z)| ≥ (1 − 4L) length ( fn (cu ))
> (1 − 4L)λ−1 length (cu ) = (1 − 4L)λ−1 (d(φ  , ψ  ) − η).
Since η is arbitrary, F is contracting for small enough L and . 

Since F is contracting (Lemma 5.6.3) and depends continuously on f ,
it has a unique fixed point φ ∈ (L, ), which depends continuously on f
(property 6) and automatically satisfies property 2. The invariance of the
5.6. Stable and Unstable Manifolds 121

stable and unstable cones (with a small enough ) implies that φ satisfies
properties 3 and 4. Property 1 follows immediately from 3 and 4. Since
property 1 gives a geometric characterization of graph(φn ), the fixed point
of F for a smaller  is a restriction of the fixed point of F for a larger  to
a smaller domain. As  → 0 and L → ∞, the stable cone KLs (0, n) (which
contains graph(φn )) tends to Es (n). Therefore Es (n) is the tangent plane to
graph(φn ) at 0 (property 5). 

The following theorem establishes the existence of local stable manifolds
for points in a hyperbolic set  and in sδ , and of local unstable manifolds
for points in  and in uδ (see §5.4); recall that sδ ⊃  and uδ ⊃ .

THEOREM 5.6.4. Let f : M → M be a C 1 diffeomorphism of a differentiable

manifold, and let  ⊂ M be a hyperbolic set of f with constant λ (the metric
is adapted).
Then there are , δ > 0 such that for every x s ∈ sδ and every x u ∈ uδ
(see §5.4)
1. the sets

Ws (x s ) = {y ∈ M: dist( f n (x s ), f n (y)) <  for all n ∈ N0 },

Wu (x u ) = {y ∈ M: dist( f −n (x u ), f −n (y)) <  for all n ∈ N0 },

called the local stable manifold of x s and the local unstable manifold
of x u , are C 1 embedded disks,
2. Tys Ws (x s ) = Es (ys ) for every ys ∈ Ws (x s ), and Tyu Wu (x u ) = Eu (yu )
for every yu ∈ Wu (x u ) (see Proposition 5.4.4),
3. f (Ws (x s )) ⊂ Ws ( f (x s )) and f −1 (Wu ( f (x u ))) ⊂ Wu (x u ),
4. if ys , zs ∈ Ws (x s ), then ds ( f (ys ), f (zs )) < λds (ys , zs ), where ds is the
distance along Ws (x s ),
if yu , zu ∈ Wu (x u ), then du ( f −1 (yu ), f −1 (zu )) < λdu (yu , zu ), where du
is the distance along Wu (x u ),
5. if 0 < dist(x s , y) <  and exp−1 x s (y) lies in the δ-cone Kδ (x ), then dist
u s
−1
( f (x ), f (y)) > λ dist(x , y),
s s

if 0 < dist(x u , y) <  and exp−1 x u (y) lies in the δ-cone Kδ (x ), then
s u

dist( f (x u ), f (y)) < λdist(x s , y),

6. if ys ∈ Ws (x s ), then Wαs (ys ) ⊂ Ws (x s ) for some α > 0,
if yu ∈ Wu (x u ), then Wβu (yu ) ⊂ Wu (x u ) for some β > 0,
Proof. Since sδ ⊃  is compact, for a small enough δ there is a collec-
tion U of coordinate charts (Ux , ψx ), x ∈ sδ , such that Ux covers the
δ-neighborhood of x and the changes of coordinates ψx ◦ ψ y−1 between
the charts have equicontinuous first derivatives. For any point x s ∈ sδ , let
122 5. Hyperbolic Dynamics

fn = ψ f n (xs ) ◦ f ◦ ψ −1
f n−1 (x s )
, Es (n) = dψ f n (xs ) (x s )Es ( f n (x s )), and Eu (n) =
dψ f n (x) (x)E ( f (x)), apply Proposition 5.6.1, and set Ws (x) = W0s (). Simi-
u n

larly, apply Proposition 5.6.1 to f −1 to construct the local unstable manifolds.

Properties 1–6 follow immediately from Proposition 5.6.1. 

Let  be a hyperbolic set of f : U → M and x ∈ . The (global) stable and
unstable manifolds of x are defined by

Ws (x) = {y ∈ M: d( f n (x), f n (y)) → 0 as n → ∞},

Wu (x) = {y ∈ M: d( f −n (x), f −n (y)) → 0 as n → ∞}.

PROPOSITION 5.6.5. There is 0 > 0 such that for every  ∈ (0, 0 ), for every
x ∈ ,

∞
 
∞

Ws (x) = f −n Ws ( f n (x) , Wu (x) = f n Wu ( f −n (x)).
n=0 n=0

Proof. Exercise 5.6.2. 


COROLLARY 5.6.6. The global stable and unstable manifolds are embed-
ded C 1 submanifolds of M homeomorphic to the unit balls in corresponding
dimensions.
Proof. Exercise 5.6.3. 

Exercise 5.6.1. Suppose f : Rm → Rm is a continuous map such that
| f (x) − f (y)| ≥ a|x − y| for some a > 1, for all x, y ∈ Rm. If f (0) = 0, show
that the image of a ball of radius r > 0 centered at 0 contains the ball of
radius ar centered at 0.

Exercise 5.6.2. Prove Proposition 5.6.5.

Exercise 5.6.3. Prove Corollary 5.6.6.

5.7 Inclination Lemma

Let M be a differentiable manifold. Recall that two submanifolds N1 , N2 ⊂
M of complementary dimensions intersect transversely (or are transverse) at
a point p ∈ N1 ∩ N2 if Tp M = Tp N1 ⊕ Tp N2 . We write N1 ∩| N2 if every point
of intersection of N1 and N2 is a point of transverse intersection.
Denote by Bi the open ball of radius  centered at 0 in Ri . For v ∈ Rm =
Rk × Rl , denote by v u ∈ Rk and v s ∈ Rl the components of v = v u + v s , and
by π u : Rm → Rk the projection to Rk. For δ > 0, let Kδu = {v ∈ Rm: v s  ≤
δv u } and Kδs = {v ∈ Rm: v u  ≤ δv s }.
5.7. Inclination Lemma 123

Bl
graph(φ)
φ(Dn )

In

Rk 0 Bk

Figure 5.2. The image of the graph of φ under f n .

LEMMA 5.7.1. Let λ ∈ (0, 1), , δ ∈ (0, 0.1). Suppose f : Bk × Bl → Rm and
φ: Bk → Bl are C 1 maps such that
1. 0 is a hyperbolic fixed point of f ,
2. Wu (0) = Bk × {0} and Ws (0) = {0} × Bl ,
3. d fx (v) ≥ λ−1 v for every v ∈ Kδu whenever both x, f (x) ∈ Bk × Bl ,
4. d fx (v) ≤ λv for every v ∈ Kδs whenever both x, f (x) ∈ Bk × Bl ,
5. d fx (Kδu ) ⊂ Kδu whenever both x, f (x) ∈ Bk × Bl ,
6. d( f −1 )x (Kδs ) ⊂ Kδs whenever both x, f −1 (x) ∈ Bk × Bl ,
7. T(y,φ(y)) graph(φ) ⊂ Kδu for every y ∈ Bk,
Then for every n ∈ N there is a subset Dn ⊂ Bk diffeomorphic to B k and such
that the image In under f n of the graph of the restriction φ| Dn has the following
properties: π u (In ) ⊃ B/2
k
and Tx In ⊂ Kδλ
u
2n for each x ∈ In .

Proof. The lemma follows from the invariance of the cones (Exercise 5.7.2).


The meaning of the lemma is that the tangent planes to the image of
the graph of φ under f n are exponentially (in n) close to the “horizontal”
space Rk, and the image spreads over Bk in the horizontal direction (see
Figure 5.2).
The following theorem, which is also sometimes called the Lambda
Lemma, implies that if f is Cr with r ≥ 1, and D is any C 1 -disk that in-
tersects transversely the stable manifold Ws (x) of a hyperbolic fixed point
x, then the forward images of D converge in the Cr topology to the unstable
manifold Wu (x) [PdM82]. We prove only C 1 convergence. Let BRu be the ball
of radius R centered at x in Wu (x) in the induced metric.

THEOREM 5.7.2 (Inclination Lemma). Let x be a hyperbolic fixed point

of a diffeomorphism f : U → M, dimWu (x) = k, and dimWs (x) = l. Let y ∈
Ws (x), and suppose that D  y is a C 1 submanifold of dimension k intersecting
Ws (x) transversely at y.
124 5. Hyperbolic Dynamics

Then for every R > 0 and β > 0 there are n0 ∈ N and, for each n ≥ n0 ,
a subset D̃ = D̃(R, β, n), y ∈ D̃ ⊂ D, diffeomorphic to an open k-disk and
such that the C 1 distance between f n ( D̃)) and BRu is less than β.
Proof. We will show that for some n1 ∈ N, an appropriate subset D1 ⊂
f n1 (D) satisfies the assumptions of Lemma 5.7.1. Since {x} is a hyperbolic
set of f , for any δ > 0 there is  > 0 such that Es (x) and Eu (x) can be ex-
tended to invariant distributions Ẽs and Ẽu in the -neighborhood B of
x and the hyperbolicity constant is at most λ + δ (Proposition 5.4.4). Since
f n (y) → x, there is n2 ∈ N such that z = f n2 (y) ∈ B . Since D intersects
Ws (x) transversely, so does f n2 (D). Therefore there is η > 0 such that if
v ∈ Tz f n2 (D), v = 1, v = v s + v u , v s ∈ Ẽs (z), v u ∈ Ẽu (z), and v u =0, then
v u  ≥ ηv s . By Proposition 5.4.4, for a small enough δ > 0, the norm
d f n v s  decays exponentially and d f n v u  grows exponentially. Therefore,
for an arbitrarily small cone size, there exists n2 ∈ N such that Tf n2 (y) f n2 (D)
lies inside the unstable cone at f n2 (y). 

Exercise 5.7.1. Prove that if x is a homoclinic point, then x is non-wandering
but not recurrent.

Exercise 5.7.2. Prove Lemma 5.7.1.

Exercise 5.7.3. Let p be a hyperbolic fixed point of f . Suppose Ws ( p) and

Wu ( p) intersect transversely at q. Prove that the union of p with the orbit
of q is a hyperbolic set of .

5.8 Horseshoes and Transverse Homoclinic Points

Let Rm = Rk × Rl . We will refer to Rk and Rl as the unstable and stable
subspaces, respectively, and denote by π u and π s the projections to those
subspaces. For v ∈ Rm, denote v u = π u (v) ∈ Rk and v s = π s (v) ∈ Rl . For
α ∈ (0, 1), call the sets Kαu = {v ∈ Rm: |v s | ≤ α|v u |} and Kαs = {v ∈ Rm: |v u | ≤
α|v s |} the unstable and stable cones, respectively. Let Ru = {x ∈ Rk: |x| ≤
1}, Rs = {x ∈ Rl : |x| ≤ 1}, and R = Ru × Rs . For z = (x, y), x ∈ R k, y ∈ R l ,
the sets F s (z) = {x} × Rs and F u (z) = Ru × {y} will be referred to as unsta-
ble and stable fibers, respectively. We say that a C 1 map f : R → Rm has a
horseshoe if there are λ, α ∈ (0, 1) such that

1. f is one-to-one on R;
2. f (R) ∩ R has at least two components 0 , . . . , q−1 ;
3. if z ∈ R and f (z) ∈ i , 0 ≤ i < q, then the sets Giu (z) = f (F u (z)) ∩
5.8. Horseshoes and Transverse Homoclinic Points 125

Rl
R

F u (z) z

∆i Gui (z) Gsi (z)

f (z)

Rk

Figure 5.3. A non-linear horseshoe.

i and Gis (z) = f −1 (F s ( f (z)) ∩ i ) are connected, and the restric-

tions of π u to Giu (z) and of π s to Gis (z) are onto and one-to-one;
4. if z, f (z) ∈ R, then the derivative d fz preserves the unstable cone Kαu
and λ|d fzv| ≥ |v| for every v ∈ Kαu , and the inverse d f f−1(z) preserves
the stable cone Kαs and λ|d f f−1
(z) v| ≥ |v| for every v ∈ Kα .
s

The intersection  = n∈Z f n (R) is called a horseshoe.

THEOREM 5.8.1. The horseshoe  = f n (R) is a hyperbolic set of f .

n∈Z
If f (R) ∩ R has q components, then the restriction of f to  is topologically
conjugate to the full two-sided shift σ in the space q of bi-infinite sequences
in the alphabet {0, 1, . . . , q − 1}.
Proof. The hyperbolicity of  follows from the invariance of the cones and
the stretching of vectors inside the cones. The topological conjugacy of f |
to the two-sided shift is left as an exercise (Exercise 5.8.2). 


COROLLARY 5.8.2. If a diffeomorphism f has a horseshoe, then the topo-

logical entropy of f is positive.

Let p be a hyperbolic periodic point of a diffeomorphism f : U → M.

A point q ∈ U is called homoclinic (for p) if q =p and q ∈ Ws ( p) ∩ Wu ( p);
it is called transverse homoclinic (for p) if in addition Ws ( p) and Wu ( p)
intersect transversely at q.
The next theorem shows that horseshoes, and hence hyperbolic sets in
general, are rather common.
126 5. Hyperbolic Dynamics

THEOREM 5.8.3. Let p be a hyperbolic periodic point of a diffeomorphism

f : U → M, and let q be a transverse homoclinic point of p. Then for every
 > 0 the union of the -neighborhoods of the orbits of p and q contains a
horseshoe of f .
Proof. We consider only the two-dimensional case; the argument for higher
dimensions is a routine generalization of the proof below. We assume without
loss of generality that f ( p) = p and f is orientation preserving. There is a
C 1 coordinate system in a neighborhood V = V u × V s of p such that p is the
origin and the stable and unstable manifolds of p coincide locally with the
coordinate axes (Figure 5.4). For a point x ∈ V and a vector v ∈ R2 , we write
x = (x u , x s ) and v = (v u , v s ), where s and u indicate the stable (vertical) and
unstable (horizontal) components, respectively. We also assume that there is
λ ∈ (0, 1) such that |d f p v s | < λ|v s | and |d f p−1 v u | < λ|v u | for every v =0. Fix
δ > 0, and let Kδ/2 s
and Kδ/2 u
be the stable and unstable δ/2-cones. Choose V
small enough so that for every x ∈ V
 u 
−1

d fx Kδ/2 ⊂ Kδ/2 u
,
d f v
< λ|v| if v ∈ Ku ,
x δ/2

d fx−1 Kδ/2
s
⊂ Kδ/2
s
, |d fx v| < λ|v| if v ∈ Kδ/2
s
.

Since q ∈ Ws ( p) ∩ Wu ( p), we have that f n (q) ∈ V and f −n (q) ∈ V for all

sufficiently large n. By invariance, Ws ( p) and Wu ( p) pass through all images
f n (q). Since Wu ( p) intersects Ws ( p) transversely at q, by Theorem 5.7.2
there is nu such that f n (q) ∈ V for n ≥ nu , and an appropriate neighborhood

W s (p)

f N (R)
f nu (q)

R
nu +1
f (q)

p W u (p)
f k (R) f −ns (q)

Figure 5.4. A horseshoe at a homoclinic point.

5.8. Horseshoes and Transverse Homoclinic Points 127

Du of f n (q) in Wu ( p) is a C 1 submanifold that “stretches across” V and

u
whose tangent planes lie in Kδ/2 , i.e., Du is the graph of a C 1 function
φ : V → V with dφ  < δ/2. Similarly since q ∈ Wu ( p), there is ns ∈ N
u u s u

such that f −n (q) ∈ U for n ≥ ns and a small neighborhood Ds of f −n (q) in

Ws ( p) is the graph of a C 1 function φ s : V s → V u with dφ s  < δ/2. Note that
since f preserves orientation, the point f nu +1 (q) is not the next intersection
of Wu ( p) with Ws ( p) after f nu (q); in Figure 5.4 it is shown as the second
intersection after f nu (q) along Ws ( p).
Consider a narrow “box” R shown in Figure 5.4, and let N = k + nu +
ns + 1. We will show that for an appropriate choice of the size and position
of R and of k ∈ N, the map f˜ = f N , the box R, and its image f˜(R) satisfy the
definition of a horseshoe. The smaller the width of the box and the closer it
lies to Ws ( p), the larger is k for which f k(R) reaches the vicinity of f −ns (q).
The number n̄ = nu + ns + 1 is fixed. If v u is a horizontal vector at f −ns (q),
its image w = d f fn̄−ns (q) v u is tangent to Wu ( p) at f nu +1 (q) and therefore lies in
u
Kδ/2 . Moreover, |w| ≥ 2β|v u | for some β > 0. For any sufficiently close vector
v at a close enough base point, the image will lie in Kδu and |d f n̄ v| ≥ β|v|.
The same holds for “almost horizontal” vectors at points close to f −ns −1 (q).
On the other hand, d fx (Kαu ) ⊂ Kλα u
for every small enough α > 0 and ev-
ery x ∈ V. Therefore, if x ∈ R, f (x), . . . , f k(x) ∈ V and v ∈ Kδu is a tangent
−k
vector at x, then d fxkv ∈ Kδλ u
k and |d fx v| > λ
k
|v|. Suppose now that x ∈ R is
such that f (x) is close to either f (q) or f −ns −1 (q). Let k be large enough
k −ns

so that β/λk > 10. There is λ ∈ (0, 1) such that if x ∈ R and f N (x) is close to
either f nu (q) or f nu +1 (q) (i.e., f k(x) is close to f −ns (q) or f −ns −1 (q)), then
Kδu is invariant under d fxN and λ |d fxN v| ≥ |v| for every v ∈ Kδu . Similarly,
for an appropriate choice of R and k, the stable δ-cones are invariant under
d f −N and vectors from Kδs are stretched by d f −N by a factor at least (λ )−1 .
To guarantee the correct intersection of f N (R) with R we must choose R
carefully. Choose the horizontal boundary segments of R to be straight line
segments, and let R stretch vertically so that it crosses Wu ( p) near f nu (q) and
f nu +1 (q). By Theorem 5.7.2, the images of these horizontal segments under
f k are almost horizontal line segments. To construct the vertical boundary
segments of R, take two vertical segments s1 and s2 to the left of f −ns −1 (q)
and to the right of f −ns (q), and truncate their preimages f −k(si ) by the
horizontal boundary segments. By Theorem 5.7.2, the preimages are almost
vertical line segments. This choice of R satisfies the definition of a horseshoe.


Exercise 5.8.1. Let f : U → M be a diffeomorphism, p a periodic point
of f , and q a (non-transverse) homoclinic point (for p). Prove that every
128 5. Hyperbolic Dynamics

arbitrarily small C 1 neighborhood of f contains a diffeomorphism g such

that p is a periodic point of g and q is a transverse homoclinic point (for p).

Exercise 5.8.2. Prove that if f (R) ∩ R in Theorem 5.8.1 has q connected

components, then the restriction of f to  is topologically conjugate to the
full two-sided shift in the space q of bi-infinite sequences in the alphabet
{1, . . . , q}.

Exercise 5.8.3. Let p1 , . . . , pk be periodic points (of possibly different pe-

riods) of f : U → M. Suppose Wu ( pi ) intersects Ws ( pi+1 ) transversely at
qi , i = 1, . . . , k, pk+1 = p1 (in particular, dimWs ( pi ) = dimWs ( p1 ) and
dimWu ( pi ) = dimWu ( p1 ), i = 2, . . . , k). The points qi are called transverse
heteroclinic points. Prove the following generalization of Theorem 5.8.3: Any
neighborhood of the union of the orbits of pi s and qi s contains a horseshoe.

5.9 Local Product Structure and Locally Maximal Hyperbolic Sets

A hyperbolic set  of f : U → M is called locally maximal if there is an

open set V such that  ⊂ V ⊂ U and  = ∞ n

n=−∞ f (V). The horseshoe
(§5.8) and the solenoid (§1.9) are examples of locally maximal hyperbolic
sets (Exercise 5.9.1).
Since every closed invariant subset of a hyperbolic set is also a hyperbolic
set, the geometric structure of a hyperbolic set may be very complicated
and difficult to describe. However, due to their special properties, locally
maximal hyperbolic sets allow a geometric characterization.
Since Es (x) ∩ Eu (x) = {0}, the local stable and unstable manifolds of x
intersect at x transversely. By continuity, this transversality extends to a
neighborhood of the diagonal in  × .

PROPOSITION 5.9.1. Let  be a hyperbolic set of f . For every small enough

 > 0 there is δ > 0 such that if x, y ∈  and d(x, y) < δ, then the intersec-
tion Ws (x) ∩ Wu (y) is transverse and consists of exactly one point [x, y],
which depends continuously on x and y. Furthermore, there is C p = C p (δ) >
0 such that if x, y ∈  and d(x, y) < δ, then ds (x, [x, y]) ≤ C p d(x, y) and
du (x, [x, y]) ≤ C p d(x, y), where ds and du denote distances along the stable
and unstable manifolds.
Proof. The proposition follows immediately from the uniform transversal-
ity of Es and Eu and Lemma 5.9.2. 

Let  > 0, k, l ∈ N, and let Bk ⊂ Rk, Bl ⊂ Rl be the -balls centered at the
origin.
5.9. Local Product Structure 129

LEMMA 5.9.2. For every  > 0 there is δ > 0 such that if φ: Bk → Rl and
ψ: Bl → Rk are differentiable maps and |φ(x)|, dφ(x), |ψ(y)|, dφ(y) < δ
for all x ∈ Bk and y ∈ Bl , then the intersection graph(φ) ∩ graph(ψ) ⊂ Rk+l
is transverse and consists of exactly one point, which depends continuously
on φ and ψ in the C 1 topology.
Proof. Exercise 5.9.3. 

The following property of hyperbolic sets plays a major role in their geo-
metric description and is equivalent to local maximality. A hyperbolic set 
has a local product structure if there are (small enough)  > 0 and δ > 0 such
that (i) for all x, y ∈  the intersection Ws (x) ∩ Wu (y) consists of at most one
point, which belongs to , and (ii) for x, y ∈  with d(x, y) < δ, the inter-
section consists of exactly one point of , denoted [x, y] = Ws (x) ∩ Wu (y),
and the intersection is transverse (Proposition 5.9.1). If a hyperbolic set 
has a local product structure, then for every x ∈  there is a neighborhood
U(x) such that
 
U(x) ∩  = [y, z]: y ∈ U(x) ∩ Ws (x), z ∈ U(x) ∩ Wu (x) .

PROPOSITION 5.9.3. A hyperbolic set  is locally maximal if and only if it

has a local product structure.
Proof. Suppose  is locally maximal. If x, y ∈  and dist(x, y) is small
enough, then by Proposition 5.9.1, Ws (x) ∩ Wu (y) = [x, y] =: z exists and,
by Theorem 5.6.4(4), the forward and backward semiorbits of z stay close to
. Since  is locally maximal, z ∈ .
Conversely, assume that  has a local product structure with constants
, δ, and C p from Proposition 5.9.1. We must show that if the whole orbit
of a point q lies close to , then the point lies in . Fix α ∈ (0, δ/3) such
that f ( p) ∈ Wδ/3
u
( f (x)) for each x ∈  and p ∈ Wαu (x). Assume first that q ∈
Wαu (x0 ) for some x0 ∈  and that there are yn ∈  such that d( f n (q), yn ) <
α/C p for all n > 0. Since f (x0 ), y1 ∈  and d( f (x0 ), y1 ) < d( f (x0 ), f (q)) +
d( f (q), y1 ) < δ/3 + α/C p < δ, we have that x1 = [y1 , f (x0 )] ∈  and, by
Proposition 5.9.1, f (q) ∈ Wαu (x1 ). Similarly, x2 = [y2 , f (x1 )] ∈  and f 2 (q) ∈
Wαu (x2 ). By repeating this argument we construct points xn = [yn , f n (q)] ∈ 
with f n (q) ∈ Wαu (xn ). Observe that qn = f −n (xn ) → q as n → ∞. Since  is
closed, q ∈ . Similarly, if q ∈ Wαs (x0 ) for some x0 ∈  and f n (q) stays close
enough to  for all n < 0, then q ∈ .
Assume now that f n (y) is close enough to xn ∈  for all n ∈ Z. Then
y ∈s and y ∈u . Hence, by Propositions 5.4.4 and 5.4.3, the union  ∪ O f (y)
is a hyperbolic set (with close constants), and the local stable and unstable
130 5. Hyperbolic Dynamics

manifolds of y are well defined. Observe that the forward semiorbit of

p = [y, x0 ] and the backward semiorbit of q = [x0 , y] stay close to . There-
fore, by the above argument, p, q ∈  and (by the local product structure)
y = [ p, q] ∈ . 

Exercise 5.9.1. Prove that horseshoes (§5.8) and the solenoid (§1.9) are
locally maximal hyperbolic sets.

Exercise 5.9.2. Let p be a hyperbolic fixed point of f and q ∈ Ws ( p) ∩

Wu ( p) a transverse homoclinic point. By Exercise 5.7.3, the union of p with
the orbit of q is a hyperbolic set of f . Is it locally maximal?

Exercise 5.9.3. Prove Lemma 5.9.2.

5.10 Anosov Diffeomorphisms

Recall that a C 1 diffeomorphism f of a connected differentiable manifold
M is called Anosov if M is a hyperbolic set for f ; it follows directly from the
definition that M is locally maximal and compact.
The simplest example of an Anosov diffeomorphism is the automorphism
of T2 given by the matrix (21 11). More generally, any linear hyperbolic au-
tomorphism of the n-torus Tn is Anosov. Such an automorphism is given
by an n × n integer matrix with determinant ±1 and with no eigenvalues of
modulus 1.
Toral automorphisms can be generalized as follows. Let N be a simply
connected nilpotent Lie group, and  a uniform discrete subgroup of N. The
quotient M = N/  is a compact nilmanifold. Let f¯ be an automorphism of
N that preserves  and whose derivative at the identity is hyperbolic. The
induced diffeomorphism f of M is Anosov. For specific examples of this type
see [Sma67]. Up to finite coverings, all known Anosov diffeomorphisms are
topologically conjugate to automorphisms of nilmanifolds.
The families of stable and unstable manifolds of an Anosov diffeomor-
phism form two foliations (see §5.13) called the stable foliation Ws and un-
stable foliation Wu (Exercise 5.10.1). These foliations are in general not C 1 ,
or even Lipschitz [Ano67], but they are Hölder continuous (Theorem 6.1.3).
In spite of the lack of Lipschitz continuity, the stable and unstable folia-
tions possess a uniqueness property similar to the uniqueness theorem for
ordinary differential equations (Exercise 5.10.2).
Proposition 5.10.1 states basic properties of the stable and unstable dis-
tributions Es and Eu , and the stable and unstable foliations Ws and Wu , of
an Anosov diffeomorphism f . These properties follow immediately from
the previous sections of this chapter. We assume that the metric is adapted
5.10. Anosov Diffeomorphisms 131

to f and denoted by ds and du , the distances along the stable and unstable
leaves.

PROPOSITION 5.10.1. Let f : M → M be an Anosov diffeomorphism. Then

there are λ ∈ (0, 1), C p > 0,  > 0, δ > 0, and, for every x ∈ M, a splitting
Tx M = Es (x) ⊕ Eu (x) such that
1. d fx (Es (x)) = Es ( f (x)) and d fx (Eu (x)) = Eu ( f (x));
2. d fx v s  ≤ λv s  and d fx−1 v u  ≤ λv u  for all v s ∈ Es (x), v u ∈ Eu (x);
3. Ws (x) = {y ∈ M: d( f n (x), f n (y)) → 0 as n → ∞}
and ds ( f (x), f (y)) ≤ λds (x, y) for every y ∈ Ws (x);
4. Wu (x) = {y ∈ M: d( f −n (x), f −n (y)) → 0 as n → ∞}
and du ( f −1 (x), f −1 (y)) ≤ λn du (x, y) for every y ∈ Wu (x);
5. f (Ws (x)) = Ws ( f (x)) and f (Wu (x)) = Wu ( f (x));
6. Tx Ws (x) = Es (x) and Tx Wu (x) = Eu (x);
7. if d(x, y) < δ, then the intersection Ws (x) ∩ Wu (y) is exactly one point
[x, y], which depends continuously on x and y, and ds ([x, y], x) ≤
C p d(x, y), du ([x, y], y) ≤ C p d(x, y).

For convenience we restate several properties of Anosov diffeomor-

phisms. Recall that a diffeomorphism f : M → M is structurally stable if
for every  > 0 there is a neighborhood U ⊂ Diff1 (M) of f such that for
every g ∈ U there is a homeomorphism h: M → M with h ◦ f = g ◦ h and
dist0 (h, Id) < .

PROPOSITION 5.10.2
1. Anosov diffeomorphisms form an open (possibly empty) subset in the
C 1 topology (Corollary 5.5.2).
2. Anosov diffeomorphisms are structurally stable (Corollary 5.5.4).
3. The set of periodic points of an Anosov diffeomorphism is dense in the
set of non-wandering points (Corollary 5.3.4).

Here is a more direct proof of the density of periodic points. Let  and
δ satisfy Proposition 5.10.1. If x ∈ M is non-wandering, then there is n ∈ N
and y ∈ M such that dist(x, y), dist( f n (y), y) < δ/(2C p ). Assume that λn <
1/(2C p ). Then the map z →[y, f n (z)] is well defined for z ∈ Wδs (y). It maps
Wδs (y) into itself and, by the Brouwer fixed point theorem, has a fixed point
y1 such that ds (y1 , y) < δ, f n (y1 ) ∈ Wu (y1 ) and du (y1 , f n (y1 )) < δ. The map
f −n sends Wδu ( f n (y1 )) to itself and therefore has a fixed point.

THEOREM 5.10.3. Let f : M → M be an Anosov diffeomorphism. Then the

following are equivalent:
1. NW( f ) = M,
2. every unstable manifold is dense in M,
132 5. Hyperbolic Dynamics

3. every stable manifold is dense in M,

4. f is topologically transitive,
5. f is topologically mixing.
Proof. We say that a set Ais -dense in a metric space (X, d) if d(x, A) < 
for every x ∈ X.
1 ⇒ 2: We will show that every unstable manifold is -dense in M for
an arbitrary  > 0. By Proposition 5.10.2(3), the periodic points are dense.
Assume that  > 0 satisfies Proposition 5.10.1(7) and that periodic points
xi , i = 1, . . . , N, form an /4-net in M. Let P be the product of the periods
of the xi s, and set g = f P . Note that the stable and unstable manifolds of g
are the same as those of f .
LEMMA 5.10.4. There is q ∈ N such that if dist(Wu (y), xi ) < /2 and
dist(xi , x j ) < /2 for some y ∈ M, i, j, then dist(g nq (Wu (y)), xi ) < /2 and
dist(g nq (Wu (y)), x j ) < /2 for every n ∈ N.
Proof. By Proposition 5.10.2(3), there is z ∈ Wu (y) ∩ WCs p  p (xi ). Therefore
dist(g t (z), xi ) < /2 for any t ≥ t0 , where t0 depends on  but not on z.
Since dist(g t (z), x j ) < , by Proposition 5.10.2(3) there exists a point
w ∈ Wu (g t (z)) ∩ WCs p  p (x j ). Hence dist(g τ (w), x j ) < /2 for any τ ≥ s0 which
depends only on  but not on w. The lemma follows with q = s0 + t0 . 

Since M is compact and connected, any xi can be connected to any x j by
a chain of not more than N periodic points xk with distance < /2 between
any two consecutive points. By Lemma 5.10.4, g Nq (Wu (y) is -dense in M
for any y ∈ M. Hence, Wu (x) is -dense for any x = g −Nq (y) ∈ M. Therefore,
Wu (x) is dense for each x. Reversing the time gives 1 ⇒ 3.
LEMMA 5.10.5. If every (un)stable manifold is dense in M, then for every
 > 0 there is R = R() > 0 such that every ball of radius R in every (un)stable
manifold is -dense in M.

Proof. Let x ∈ M. Since Wu (x) = R WRu (x) is dense, there is R(x) such that
u
WR(x) (x) is /2-dense. Since Wu is a continuous foliation, there is δ(x) > 0
u
such that WR(x) (y) is -dense for any y ∈ B(x, δ(x)). By the compactness of
M, a finite collection B of the δ(x)-balls covers M. The maximal R(x) for the
balls from B satisfies the lemma. 

2 ⇒ 5: Let U, V ⊂ M be non-empty open sets. Let x, y ∈ M and δ > 0 be
such that Wδu (x) ⊂ U and B(y, δ) ⊂ V, and let R = R(δ) (see Lemma 5.10.5).
Since f expands unstable manifolds exponentially and uniformly, there
is N ∈ N such that f n (Wδu (x)) ⊃ WRu ( f n (x)) for n ≥ N. By Lemma 5.10.5,
f n (U) ∩ V = ∅and hence f is topologically mixing. Similarly 3 ⇒ 5.
1 ⇒ 3 follows by reversing the time. Obviously 5 ⇒ 4 and 4 ⇒ 1. 

5.11. Axiom A and Structural Stability 133

Exercise 5.10.1. Prove that the stable and unstable manifolds of an Anosov
diffeomorphism form foliations (see §5.13).

Exercise 5.10.2. Although the stable and unstable distributions of an

Anosov diffeomorphism, in general, are not Lipschitz continuous, the fol-
lowing uniqueness property holds true. Let γ (·) be a differentiable curve
such that γ̇ (t) ∈ Es (γ (t)) for every t. Prove that γ lies in one stable mani-
fold.

5.11 Axiom A and Structural Stability

Some of the results of §5.10 extend to a natural wider class of hyperbolic
dynamical systems. Throughout this section we assume that f is a diffeo-
morphism of a compact manifold M. Recall that the set of non-wandering
points NW( f ) is closed and f -invariant, and that Per( f ) ⊂ NW( f ).
A diffeomorphism f satisfies Smale’s Axiom A if the set NW( f ) is hyper-
bolic and Per( f ) = NW( f ). The second condition does not follow from the
first. By Proposition 5.3.3, the set Per( f ) is dense in the set NW( f |NW( f ) ) of
non-wandering points of the restriction of f to NW( f ). However, in general
NW( f |NW( f ) ) =NW( f ) (Exercise 5.11.1, Exercise 5.11.2).
For a hyperbolic periodic point p of f , denote by Ws (O( p)) and Wu (O( p))
the unions of the stable and unstable manifolds of p and its images, re-
spectively. If p and q are hyperbolic periodic points, we write p ≤ q when
Ws (O( p)) and Wu (O(q)) have a point of transverse intersection. The re-
lation ≤ is reflexive. It follows from Theorem 5.7.2 that ≤ is transitive
(Exercise 5.11.3). If p ≤ q and q ≤ p, we write p ∼ q and say that p and
q are heteroclinically related. The relation ∼ is an equivalence relation.

THEOREM 5.11.1 (Smale’s Spectral Decomposition [Sma67]). If f satisfies

Axiom A, then there is a unique representation of NW( f ),

NW( f ) = 1 ∪ 2 ∪ · · · ∪ k,

as a disjoint union of closed f -invariant sets (called basic sets) such that
1. each i is a locally maximal hyperbolic set of f ;
2. f is topologically transitive on each i ; and
j
3. each i is a disjoint union of closed sets i , 1 ≤ j ≤ mi , the diffeomor-
j
phism f cyclically permutes the sets i , and f mi is topologically mixing
j
on each i .

The basic sets are precisely the closures of the equivalence classes of ∼.
For two basic sets, we write i ≤  j if there are periodic points p ∈ i and
q ∈  j such that p ≤ q.
134 5. Hyperbolic Dynamics

Let f satisfy Axiom A. We say that f satisfies the strong transversality

condition if Ws (x) intersects Wu (y) transversely (at all common points) for
all x, y ∈ NW( f ).
THEOREM 5.11.2 (Structural Stability Theorem). A C 1 diffeomorphism is
structurally stable if and only if it satisfies Axiom A and the strong transver-
sality condition.
J. Robbin [Rob71] showed that a C 2 diffeomorphism satisfying Axiom
A and the strong transversality condition is structurally stable. C. Robinson
[Rob76] weakened C 2 to C 1 . R. Mañé [Mañ88] proved that a structurally
stable C 1 diffeomorphism satisfies Axiom A and the strong transversality
condition.
Exercise 5.11.1. Give an example of a diffeomorphism f such that
NW( f |NW( f ) ) =NW( f ).
Exercise 5.11.2. Give an example of a diffeomorphism f for which NW( f )
is hyperbolic and NW( f |NW( f ) ) =NW( f ).
Exercise 5.11.3. Prove that ≤ is a transitive relation.
Exercise 5.11.4. Suppose that f satisfies Axiom A. Prove that NW( f ) is a
locally maximal hyperbolic set.

5.12 Markov Partitions

Recall (Chapter 1, Chapter 3) that a partition of the phase space of a dy-
namical system induces a coding of the orbits and hence a semiconjugacy
with a subshift. For hyperbolic dynamical systems, there is a special class of
partitions – Markov partitions – for which the target subshift is a subshift of
finite type. A Markov partition P for an invariant subset  of a diffeomor-
phism f of a compact manifold M is a collection of sets Ri called rectangles
such that for all i, j, k
1. each Ri is the closure of its interior,
2. int Ri ∩ int Rj = ∅ if i = j,

3.  ⊂ i Ri ,
4. if f m(int Ri ) ∩ int Rj ∩  = ∅for some m ∈ Z and f n (int Rj ) ∩ int Rk ∩
 = ∅for some n ∈ Z, then f m+n (int Ri ) ∩ int Rk ∩  = ∅.
The last condition guarantees the Markov property of the subshift cor-
responding to P, i.e., the independence of the future from the past. For
hyperbolic dynamical systems, each rectangle is closed under the local
5.12. Markov Partitions 135

p0 p0

A3
B2 B2
p2
p2

B1
p3

A2 A2
p1
p1
A1

p0 p0
A3

p3

Figure 5.5. Markov partition for the toral automorphism fM .

product structure “commutator” [x, y], i.e., if x, y ∈ Ri , then [x, y] ∈ Ri .

 
For x ∈ Ri let Ws (x, Ri ) = y∈Ri [x, y] and Wu (x, Ri ) = y∈Ri [y, x]. The last
condition means that if x ∈ int Ri and f (x) ∈ int Rj , then Wu ( f (x), Rj ) ⊂
f (Wu (x, Ri )) and Ws (x, Ri ) ⊂ f −1 (Ws ( f (x), Rj )).
The partition of the unit interval [0, 1] into m intervals [k/m, (k + 1)/m) is
a Markov partition for the expanding endomorphism Em. The target subshift
in this case is the full shift on m symbols.
We now describe a Markov partition for the hyperbolic toral automor-
phism f = f M given by the matrix
 
2 1
M= ,
1 1

which√was constructed by R. Adler and B. Weiss [AW67]. The eigenvalues are

(3 ± 5)/2. We begin by partitioning the unit square representing the torus
T2 in Figure 5.5 into two rectangles: A, consisting of three parts A1 , A2 , A3 ;
and B, consisting of two parts B1 , B2 . The longer sides of the√rectangles
are parallel to the eigendirection of the larger eigenvalue (3 + 5)/2, and
the shorter
√ sides are parallel to the eigendirection of the smaller eigenvalue
(3 − 5)/2. In Figure 5.5, the identified points and regions are marked by the
same symbols. The images of Aand B are shown in Figure 5.6. We subdivide
136 5. Hyperbolic Dynamics

p2 Delta

∆
4

p1
∆
2
p0
∆
3 p0

∆5
p3 ∆4 p2
p1 ∆1 p1
∆2
p0
p0
∆3

p3
Figure 5.6. The image of the Markov partition under fM .

A and B into five subrectangles 1 , 2 , 3 , 4 , 5 that are the connected

components of the intersections of Aand B with f (A) and f (B). The image
of A consists of 1 , 3 and 4 ; the image of B consists of 2 and 5 . The part
of the boundary of the i ’s that is parallel to the eigendirection of the larger
eigenvalue is called stable; the part that is parallel to the eigendirection of the
smaller eigenvalue is called unstable. By construction, the partition  of T2
into five rectangles i has the property that the image of the stable bound-
ary is contained in the stable boundary, and the preimage of the unstable
boundary is contained in the unstable boundary (Exercise 5.12.1). In other
words, for each i, j, the intersection i j = i ∩ f ( j ) consists of one or two
rectangles that stretch “all the way” through i , and the stable boundary
of i j is contained in the stable boundary of i ; similarly, the intersection
i−1
j = i ∩ f
−1
( j ) consists of one or two rectangles that stretch “all the
way” through i , and the unstable boundary of i−1 j is contained in the un-
stable boundary of i . Let ai j = 1 if the interior of f (i ) ∩  j is not empty,
and ai j = 0 otherwise, i, j = 1, . . . , 5. This defines the adjacency matrix
 
1 0 1 1 0
1 0 1 1 0
 
 
A = 1 0 1 1 0 .
 
0 1 0 0 1
0 1 0 0 1
5.13. Appendix: Differentiable Manifolds 137

If ω = (. . . , ω−1 , ω0 , ω1 , . . .) is an allowed infinite sequence for this adjacency

∞
matrix, then the intersection i=−∞ f −i (ωi ) consists of exactly one point
φ(ω); it follows that there is a continuous semiconjugacy φ:  A → T2 , i.e.,
f ◦ φ = φ ◦ σ , where σ is the shift in  A (Exercise 5.12.2). Conversely, let B0
∞
be the union of the boundaries of the i ’s, and let B = i=−∞ f i (B0 ). For
∞
x ∈ T \B, set ψi (x) = j if f (x) ∈  j . The itinerary sequence (ψi (x))i=−∞
2 i

is an element of  A, and φ ◦ ψ = Id (Exercise 5.12.3).

In higher dimensions, this direct geometric construction does not work.
Even for a hyperbolic toral automorphism, the boundary is nowhere differ-
entiable. Nevertheless, as R. Bowen showed [Bow70], any locally maximal
hyperbolic set  has a Markov partition [Bow70] which provides a semicon-
jugacy from a subshift of finite type to .

Exercise 5.12.1. Prove that the stable boundary is forward invariant and
the unstable boundary is backward invariant under fM .

Exercise 5.12.2. Prove that for the toral automorphism f M , the intersec-
tion of the preimages of rectangles i along an allowed infinite sequence
ω consists of exactly one point. Prove that there is a semiconjugacy φ from
σ | A to the toral automorphism fM .

Exercise 5.12.3. Prove that the map ψ defined in the text above satisfies
ψ(x) ∈  A and that φ ◦ ψ = Id.

Exercise 5.12.4. Construct Markov partitions for the linear horseshoe

(§1.8) and the solenoid (§1.9).

5.13 Appendix: Differentiable Manifolds

An m-dimensional C k manifold M is a second-countable Hausdorff topolog-
ical space together with a collection U of open sets in M and for each U ∈ U
a homeomorphism φU from U onto the unit ball B m ⊂ Rm such that:

1. U is a cover of M, and
2. for U, V ∈ U, if U ∩ V = ∅, the mapφU ◦ φV−1 : φV (U ∩ V) → φU (U ∩
V) is C k.

We may take k ∈ N ∪ {∞, ω}, where C ω denotes the class of real analytic
functions.
We write Mm to indicate that M has dimension m. If x ∈ M and U ∈ U
contains x, then the pair (U, φU ), U ∈ U, is called a coordinate chart at x,
and the n component functions x1 , x2 , . . . , xm of φU are called coordinates on
138 5. Hyperbolic Dynamics

U. The collection of coordinate charts {(U, φU )}U∈U is called an atlas on M.

Note that any open subset of Rm is a C k manifold, for any k ∈ N ∪ {∞, ω}.
If Mm and Nn are C k manifolds, then a continuous map f : M → N is C k if
for any coordinate chart (U, φU ) on M, and any coordinate chart (V, ψV ) on
N, the map ψV ◦ f ◦ φU−1 : φU (U ∩ f −1 (V)) → Rn is a C k map. For k ≥ 0, the
set of C k maps from M to N is denoted C k(M, N). We say that a sequence of
functions fn ∈ C k(M, N) converges if the functions and all their derivatives
up to order k converge uniformly on compact sets. This defines a topology
on C k(M, N) called the C k topology.
We set C k(M) = C k(M, R). The subset of C k(M, M) consisting of diffeo-
morphisms of M is denoted Diffk(M).
A C k curve in Mm is a C k map α: (−, ) → M. The tangent vector to α at
α(0) = p is the linear map v: C 1 (M) → R defined by


d


v( f ) = f (α(t))
dt
t=0

for f ∈ C 1 (M). The tangent space at p is the linear space Tp M of all tangent
vectors at p.
Suppose (U, φ) is a coordinate chart, with coordinate functions x1 , . . . , xm,
and let p ∈ U. For i = 1, . . . , m, consider the curves

αi (t) = φ −1 (x1 ( p), . . . , xi−1 ( p), xi ( p) + t, xi+1 ( p), . . . , xm( p)).

p

p
Define (∂/∂ xi ) p to be the tangent vector to αi at p, i.e., for g ∈ C 1 (M),
 
 
∂ d

 p  ∂
(g) = g α (t) = (g ◦ φ) .
∂ xi p dt
t=0 i
∂ xi φ( p)

The vectors ∂/∂ xi , i = 1, . . . , m, are linearly independent at p, and span Tp M.

In particular, Tp M is a vector space of dimension m.
Let f : M → N be a C k map, k ≥ 1. For p ∈ M, the tangent map d f p :
Tp M → Tf ( p) N is defined by d f p (v)(g) = v(g ◦ f ), for g ∈ C 1 (N). In terms
of curves, if v is tangent to α at p = α(0), then d f p (v) is tangent to f ◦ α at
f ( p).

The tangent bundle T M = x∈M Tx M of M is a C k−1 manifold of twice the
dimension of M with coordinate charts defined as follows. Let (U, φU ) be a
coordinate chart on M, φU = (x1 , . . . , xm): U → Rm. For each i, the deriva-

tive dxi is a function from TU = p∈U Tp M to R, defined by dxi (v) = v(xi ),
for v ∈ TU. The function (x1 , . . . , xm, dx1 , . . . , dxm): TU → R2m is a coordi-
nate chart on TU, which we denote dφU . Note that if y, w ∈ Rm, then
   
dφU ◦ dφV−1 (y, w) = φU ◦ φV−1 (y), d φU ◦ φV−1 y (w) .
5.13. Appendix: Differentiable Manifolds 139

Let π: T M → M be the projection map that sends a vector v ∈ Tp M to its

base point p. A Cr vector field X on M is a Cr map X: M → T M such that
π ◦ X is the identity on M. We write Xp = X( p).
Let Mm and Nn be C k manifolds. We say that M is a C k submanifold of N
if M is a subset of N and the inclusion map i: M → N is C k and has rank m for
each x ∈ M. If the topology of M coincides with the subspace topology, then
M is an embedded submanifold. For each x ∈ M, the tangent space Tx M is
naturally identified with a subspace of Tx N. Two submanifolds M1 , M2 ⊂ N
of complementary dimensions intersect transversely (or are transverse) at a
point p ∈ N1 ∩ N2 if Tp N = Tp M1 ⊕ Tp M2 .
A distribution E on a differentiable manifold M is a family of k-
dimensional subspaces E(x) ⊂ Tx M, x ∈ M. The distribution is Cl , l ≥ 0, if
locally it is spanned by k Cl vector fields.
Suppose W is a partition of a differentiable manifold M into C 1 subman-
ifolds of dimension k. For x ∈ M, let W(x) be the submanifold containing x.
We say that W is a k-dimensional continuous foliation with C 1 leaves (or sim-
ply a foliation) if every x ∈ M has a neighborhood U and a homeomorphism
h: Bk × Bm−k → U such that
1. for each z ∈ Bm−k, the set h(Bk × {z}) is the connected component of
W(h(0, z)) ∩ U containing h(0, z), and
2. h(·, z) is C 1 and depends continuously on z in the C 1 topology.
The pair (U, h) is called a foliation coordinate chart. The sets h(Bk × {z})
are called local leaves (or plaques), and the sets h({y} × Bm−k) are called
local transversals. For x ∈ U, we denote by WU (x) the local leaf containing
x. More generally, a differentiable submanifold Lm−k ⊂ M is a transversal if
L is transverse to the leaves of the foliation. Each submanifold W(x) of the
foliation is called a leaf of W.
A continuous foliation W is a C k foliation, k ≥ 1, if the maps h can be
chosen to be C k. For example, lines of constant slope on T2 form a C ∞
foliation.
A foliation W defines a distribution E = TW consisting of the tangent
spaces to the leaves. A distribution E is integrable if it is tangent to a foliation.
A C k Riemannian metric on a C k+1 manifold M consists of a positive
definite symmetric bilinear form  ,  p in each tangent space Tp M such that
for any C k vector fields X and Y, the function p → Xp , Yp  p is C k. For each
v ∈ Tp M, we write v = (v, v p )1/2 . If α: [a, b] → M is a differentiable
b
curve, we define the length of α to be a α̇(s) ds. The (intrinsic) distance
d between two points in M is defined to be the infimum of the lengths of
differentiable curves in M connecting the two points.
140 5. Hyperbolic Dynamics

A C k Riemannian manifold is a C k+1 manifold with a C k Riemannian met-

ric. We denote by T 1 M the set of tangent vectors of length 1 in a Riemannian
manifold M.
A Riemannian manifold carries a natural measure called the Riemannian
volume. Roughly speaking, the Riemannian metric allows one to compute
the Jacobian of a differentiable map, and therefore allows one to define
integration in a coordinate-free way.
If X is a topological space and (Y, d) is a metric space with metric, define
a metric dist0 on C(X, Y) by
 
dist0 ( f, g) = min 1, sup max{d( f (x), g(x)) .
x∈X

If X is compact, then this metric induces the topology of uniform conver-

gence on compact sets. If X is not compact, this metric induces a finer topol-
ogy. For example, the sequence of functions fn (x) = x n in C((0, 1), R) con-
verges to 0 in the topology of uniform convergence on compact sets, but not
in the metric dist0 . The topology of uniform convergence on compact sets is
metrizable even for non-compact sets, but we will not need this metric.
If M m and Nn are C 1 Riemannian manifolds, we define a distance function
dist1 on C 1 (M, N) as follows: The Riemannian metric on N induces a metric
(distance function) on the tangent bundle T N, making T N a metric space.
For f ∈ C 1 (M, N), the differential of f gives a map d f : T 1 M → T N on the
unit tangent bundle of M. We set dist1 ( f, g) = dist0 (d f, dg). If M is compact,
the topology induced by this metric is the C 1 topology.
A differentiable manifold M is a (differentiable) fiber bundle over a differ-
entiable manifold N with fiber F and (differentiable) projection π : M → N
if for every x ∈ N there is a neighborhood V  x such that π −1 (V) is dif-
feomorphic to V × F and π −1 (y) ∼ = y × F. A diffeomorphism f : M → M
is an extension of or a skew product over a diffeomorphism g: N → N if
π ◦ f = g ◦ π; in this case g is called a factor of f .
CHAPTER SIX

Ergodicity of Anosov Diffeomorphisms

The purpose of this chapter is to establish the ergodicity of volume-

preserving Anosov diffeomorphisms (Theorem 6.3.1). This result, which was
first obtained by D. Anosov [Ano69] (see also [AS67]), shows that hyper-
bolicity has strong implications for the ergodic properties of a dynamical
system. Moreover, since a small perturbation of an Anosov diffeomorphism
is also Anosov (Proposition 5.10.2), this gives an open set of ergodic diffeo-
morphisms.
Our proof is an improvement of the arguments in [Ano69] and [AS67].
It is based on the classical approach called Hopf’s argument. The first obser-
vation is that any f -invariant function is constant mod 0 on the stable and
unstable manifolds (Lemma 6.3.2). Since these manifolds have complemen-
tary dimensions, one would expect the Fubini theorem to imply that the
function is constant mod 0, and ergodicity would follow. The major difficulty
is that, although the stable and unstable manifolds are differentiable, they
need not depend differentiably on the point they pass through, even if f
is real analytic. Thus the local product structure defined by the stable and
unstable foliations does not yield a differentiable coordinate system, and
we cannot apply the usual Fubini theorem. So we establish a property of
the stable and unstable foliations called absolute continuity that implies the
Fubini theorem.
The reason the stable and unstable manifolds do not vary differentiably
is that they depend on the infinite future and past, respectively.

6.1 Hölder Continuity of the Stable and Unstable Distributions

For a subspace A ⊂ R N and a vector v ∈ R N , set

dist(v, A) = min v − w.

w∈A

141
142 6. Ergodicity of Anosov Diffeomorphisms

For subspaces A, B in R N , define


dist(A, B) = max max dist(v, B), max dist(w, A) .
v∈A,v=1 w∈B,w=1

The following lemmas can be used to prove the Hölder continuity of in-
variant distributions for a variety of dynamical systems. Our objective is the
Hölder continuity of the stable and unstable distributions of an Anosov dif-
feomorphism, which was first established by Anosov [Ano67]. We consider
only the stable distribution; Hölder continuity of the unstable distribution
follows by reversing the time.

LEMMA 6.1.1. Let Lin : R N → R N , i = 1, 2, n ∈ N, be two sequences of linear

maps. Assume that for some b > 0 and δ ∈ (0, 1),
 1 
 L − L2  ≤ δbn
n n

for each positive integer n.

Suppose that there are two subspaces E1 , E2 ⊂ R N and positive constants
C > 1 and λ < µ with λ < b such that
 i 
 L v  ≤ Cλn v if v ∈ Ei ,
n
 i 
 L w  ≥ C −1 µn w if w ⊥ Ei .
n

Then
µ
dist(E1 , E2 ) ≤ 3C 2 δ (log µ−log λ)/(log b−log λ) .
λ
Proof. Set Kn1 = {v ∈ R N : L1n v ≤ 2Cλn v}. Let v ∈ Kn1 . Write v = v 1 +
v⊥
1
, where v 1 ∈ E1 and v⊥1
⊥ E1 . Then
 1   1 1       
 L v  =  L v + v 1  ≥  L1 v 1  −  L1 v 1  ≥ C −1 µn v 1  − Cλn v 1 ,
n n ⊥ n ⊥ n ⊥

and hence

n
 1   
v  ≤ Cµ−n  L1 v  + Cλn v 1  ≤ 3C 2 λ v.
⊥ n
µ
It follows that

n
λ
dist(v, E ) ≤ 3C
1 2
v. (6.1)
µ
Set γ = λ/b < 1. There is a unique non-negative integer k such that
γ k+1 < δ ≤ γ k. Let v 2 ∈ E2 . Then
 1 2  2 2  1 
 L v  ≤  L v  +  L − L2  · v 2 
k k k k

≤ Cλkv 2  + bkδv 2 
≤ (Cλk + (bγ )k)v 2  ≤ 2Cλkv 2 .
6.1. Hölder Continuity of the Stable and Unstable Distributions 143

It follows that v 2 ∈ Kk1 and hence E2 ⊂ Kk1 . By symmetry, E1 ⊂ Kk2 . By (6.1)

and by the choice of k,

k
λ µ
dist(E , E ) ≤ 3C
1 2 2
≤ 3C 2 δ (log µ−log λ)/(log b−log λ) . 

µ λ

LEMMA 6.1.2. Let f be a C 2 diffeomorphism of a compact C 2 submanifold

M ⊂ R N . Then for each n ∈ N and all x, y ∈ M,
 n 
d f − d f n  ≤ bn · x − y,
x y

where b = maxz∈M d fz(1 + maxz∈M dz2 f ).

Proof. Let b1 = maxz∈M d fz ≥ 1 and b2 = maxz∈M dz2 f , so that b =
b1 (1 + b2 ). Observe that  f n (x) − f n (y) ≤ (b1 )n x − y for all x, y ∈ M.
The lemma obviously holds for n = 1. For the inductive step we have
 n+1         
d f − d fyn+1  ≤ d f f n (x)  · d fxn − d fyn  + d f f n (x) − d f f n (y)  · d fyn 
x

≤ b1 bn x − y + b2 b1n x − yb1 ≤ bn+1 x − y. 


Let M be a manifold embedded in R N , and suppose E is a distribution on
M. We say that E is Hölder continuous with Hölder exponent α ∈ (0, 1] and
Hölder constant L if

dist(E(x), E(y)) ≤ L · x − yα

for all x, y ∈ M with x − y ≤ 1.

One can define Hölder continuity for a distribution on an abstract
Riemannian manifold by using parallel transport along geodesics to iden-
tify tangent spaces at nearby points. However, for a compact manifold M it
suffices to consider Hölder continuity for some embedding of M in R N . This
is so because on a compact manifold M, the ratio of any two Riemannian
metrics is bounded above and below. So is the ratio between the intrinsic
distance function on M and the extrinsic distance on M obtained by restrict-
ing the distance in R N to M. Thus the Hölder exponent is independent of
both the Riemannian metric and the embedding, but the Hölder constant
does change. So, without loss of generality, and to simplify the arguments in
this section and the next one, we will deal only with manifolds embedded
in R N .

THEOREM 6.1.3. Let M be a compact C 2 manifold and f : M → M a C 2

Anosov diffeomorphism. Suppose that 0 < λ < 1 < µ and C > 0 are such that
d fxn v s  ≤ Cλn v s  and d fxn v u  ≥ Cµn v u  for all x ∈ M, vs ∈ Es (x), v u ∈
Eu (x), and n ∈ N. Set b = maxz∈M d fz(1 + maxz∈M dz2 f ). Then the
144 6. Ergodicity of Anosov Diffeomorphisms

stable distribution Es is Hölder continuous with exponent α = (log µ − log λ)/

(log b − log λ).
Proof. As indicated above, we may assume that M is embedded in R N . For
x ∈ M, let E⊥ (x) denote the orthogonal complement to the tangent plane
Tx M in R N . Since E⊥ is a smooth distribution, it is sufficient to prove the
Hölder continuity of Es ⊕ E⊥ on M.
Since M is compact, there is a constant C̄ > 1 such that for any x ∈ M, if
v ∈ Tx M is perpendicular to Es , then d fxn v ≥ C̄ −1 µn v.
For x ∈ M, extend d fx to a linear map L(x): R N → R N by setting
L(x)| E⊥ (x) = 0, and set Ln (x) = L( f n−1 (x)) ◦ · · · ◦ L( f (x)) ◦ L(x). Note
that Ln (x)|Tx M = d fxn .
Fix x1 , x2 ∈ M with x1 − x2  < 1. By Lemma 6.1.2, the conditions of
Lemma 6.1.1 are satisfied with Lin = Ln (xi ) and Ei = Es (xi ), i = 1, 2, and
the theorem follows. 

Exercise 6.1.1. Let β ∈ (0, 1], and M be a compact C 1+β manifold, i.e.,
the first derivatives of the coordinate functions are Hölder continuous with
exponent β. Let f : M → M be a C 1+β Anosov diffeomorphism. Prove that
the stable and unstable distributions of f are Hölder continuous.

6.2 Absolute Continuity of the Stable and Unstable Foliations

Let M be a smooth n-dimensional manifold. Recall (§5.13) that a contin-
uous k-dimensional foliation W with C 1 leaves is a partition of M into C 1
submanifolds W(x)  x which locally depend continuously in the C 1 topo-
logy on x ∈ M. Denote by m the Riemannian volume in M, and by mN the
induced Riemannian volume in a C 1 submanifold N. Note that every leaf
W(x) and every transversal carry an induced Riemannian volume.
Let (U, h) be a foliation coordinate chart on M (§5.13), and let L =
h({y} × Bn−k) be a C 1 local transversal. The foliation W is called absolutely
continuous if for any such L and U there is a measurable family of positive
measurable functions δx : WU (x) → R (called the conditional densities) such
that for any measurable subset A ⊂ U
 
m(A) = 1 A(x, y) δx (y) dmW(x) (y) dmL(x).
L WU (x)

Note that the conditional densities are automatically integrable.

PROPOSITION 6.2.1. Let W be an absolutely continuous foliation of a

Riemannian manifold M, and let f : M → R be a measurable function.
6.2. Absolute Continuity of the Stable and Unstable Foliations 145

W
U1
U2

p
Figure 6.1. Holonomy map p for a foliation W and transversals U1 and U2 .

Suppose there is a set A ⊂ M of measure 0 such that f is constant on W(x)\A

for every leaf W(x).
Then f is essentially constant on almost every leaf, i.e., for any transversal
L, the function f is mW(x) -essentially constant for mL-almost every x ∈ L.
Proof. Absolute continuity implies that mW(x) (A ∩ W(x)) = 0 for mL-
almost every x ∈ L. 

Absolute continuity of the stable and unstable foliations is the property we
need in order to prove the ergodicity of Anosov diffeomorphisms. However,
we will prove a stronger property, called transverse absolute continuity; see
Proposition 6.2.2.
Let W be a foliation of M, and (U, h) a foliation coordinate chart. Let
Li = h({yi } × Bm−k) for yi ∈ Bk, i = 1, 2. Define a homeomorphism p: L1 →
L2 by p(h(y1 , z)) = h(y2 , z), for z ∈ Bm−k; p is called the holonomy map
(see Figure 6.1). The foliation W is transversely absolutely continuous if the
holonomy map p is absolutely continuous for any foliation coordinate chart
and any transversals Li as above, i.e., if there is a positive measurable function
q: L1 → R (called the Jacobian of p) such that for any measurable subset
A ⊂ L1

mL2 ( p(A)) = 1 Aq(z) dmL1 (z).
L1

If the Jacobian q is bounded on compact subsets of L1 , then W is said to be

transversely absolutely continuous with bounded Jacobians.

PROPOSITION 6.2.2. If W is transversely absolutely continuous, then it is

absolutely continuous.
146 6. Ergodicity of Anosov Diffeomorphisms

L = FU (x)

WU (s)
z=r
s
WU (y) p̄s
x y

py

FU (y)

Figure 6.2. Holonomy maps for W and F.

Proof. Let L and U be as in the definition of an absolutely continuous

foliation, x ∈ L and let F be an (n − k)-dimensional C 1 -foliation such that

F(x) ⊃ L, FU (x) = L, and U = y∈WU (x) FU (y); see Figure 6.2. Obviously,
F is absolutely continuous and transversely absolutely continuous. Let δ̄ y (·)
denote the conditional densities for F. Since F is a C 1 foliation, δ̄ is contin-
uous and hence measurable. For any measurable set A ⊂ U, by the Fubini
theorem,
 
m(A) = 1 A(y, z) δ̄ y (z) dmF(y) (z) dmW(x) (y). (6.2)
WU (x) FU (y)

Let py denote the holonomy map along the leaves of W from FU (x) = L
to FU (y), and let qy (·) denote the Jacobian of py . We have
 
1 A(y, z) δ̄ y (z) dmF(y) (z) = 1 A( py (s))qy (s) δ̄ y ( py (s)) dmL(s),
FU (y) L

and by changing the order of integration in (6.2), which is an integral with

respect to the product measure, we get
 
m(A) = 1 A( py (s))qy (s) δ̄ y ( py (s)) dmW(x) (y) dmL(s). (6.3)
L WU (x)

Similarly, let p̄s denote the holonomy map along the leaves of F from WU (x)
to WU (s), s ∈ L, and let q̄s denote the Jacobian of p̄s . We transform the
integral over WU (x) into an integral over WU (s) using the change of variables
6.2. Absolute Continuity of the Stable and Unstable Foliations 147

r = py (s), y = p̄−1
s (r ):

1 A( py (s))qy (s) δ̄ y ( py (s)) dmW(x) (y)
WU (x)

= 1 A(r )qy (s) δ̄ y (r )q̄−1
s (r ) dmW(s) (r ).
WU (s)

The last formula together with (6.3) gives the absolute continuity of W. 

The converse of Proposition 6.2.2 is not true in general (Exercise 6.2.2).

LEMMA 6.2.3. Let (X, A, µ), (Y, B, ν) be two compact metric spaces
with Borel σ-algebras and σ-additive Borel measures, and let pn : X → Y,
n = 1, 2, . . . , and p: X → Y be continuous maps such that
1. each pn and p are homeomorphisms onto their images,
2. pn converges to p uniformly as n → ∞,
3. there is a constant J such that ν( pn (A)) ≤ J µ(A) for every A ∈ A.
Then ν( p(A)) ≤ J µ(A) for every A ∈ A.
Proof. It is sufficient to prove the statement for an arbitrary open ball
Br (x) in X. If δ < r then p(Br −δ (x)) ⊂ pn (Br (x)) for n large enough,
and hence ν( p(Br −δ (x))) ≤ ν( pn (Br (x))) ≤ J µ(Br (x)). Observe now that
ν( p(Br −δ (x)))  ν( p(Br (x))) as δ  0. 

For subspaces A, B ⊂ R N , set

(A, B) = min{v − w: v ∈ A, v = 1; w ∈ B, w = 1}.

√
For θ ∈ [0, 2], we say that a subspace A ⊂ R N is θ-transverse to a subspace
B ⊂ R N if (A, B) ≥ θ.

LEMMA 6.2.4. Let Ê be a smooth k-dimensional distribution on a compact

subset of R N . Then for every ξ > 0 and  > 0 there is δ > 0 with the follow-
ing property. Suppose Q1 , Q2 ⊂ R N are (N − k)-dimensional C 1 submani-
folds with a smooth holonomy map p̂: Q1 → Q2 such that p̂(x) ∈ Q2 , p̂(x) −
x ∈ Ê(x), (Tx Q1 , Ê(x)) ≥ ξ, (Tp̂(x) Q2 , Ê(x)) ≥ ξ, dist(Tx Q1 , Tp̂(x) Q2 ) ≤ δ,
and  p̂(x) − x ≤ δ for each x ∈ Q1 . Then the Jacobian of p̂ does not
exceed 1 + .
Proof. Since only the first derivatives of Q1 and Q2 affect the Jacobian
of p̂ at x ∈ Q1 , it equals the Jacobian at x of the holonomy map p̃: Tx Q1 →
Tp̂(x) Q2 along Ê. By applying an appropriate linear transformation L(whose
determinant depends only on ξ ), switching to new coordinates (u, v) in R N ,
and using the same notation for the images of all objects under L, we may
148 6. Ergodicity of Anosov Diffeomorphisms

assume that (a) x = (0, 0), (b) T(0,0) Q1 = {v = 0}, (c) p(x) = (0, v0 ), where
v0  =  p̂(x) − x, (d) T(0,v0 ) Q2 is given by the equation v = v0 + Bu, where
B is a k × (N − k) matrix whose norm depends only on δ, and (e) Ê(0, 0) =
{u = 0}, and Ê(w, 0) is given by the equation u = w + A(w)v, where A(w)
is an (N − k) × k matrix which is C 1 in w and A(0) = 0.
The image of (w, 0) under p̂ is the intersection point of the planes v =
v0 + Bu and u = w + A(w)v. Since the norm of B is bounded from above in
terms of ξ , it suffices to estimate the determinant of the derivative ∂u/∂w at
w = 0. We substitute the first equation into the second one,

u = w + A(w)v0 + A(w)Bu;

differentiate with respect to w,

∂u ∂ A(w) ∂ A(w) ∂u
=I+ v0 + Bu + A(w)B ;
∂w ∂w ∂w ∂w
and obtain for w = 0 (using u(0) = 0 and A(0) = 0)
 
∂u  ∂ A(w) 
=I+ v0 .
∂w w=0 ∂w w=0 

THEOREM 6.2.5. The stable and unstable foliations of a C 2 Anosov diffeo-
morphism are transversely absolutely continuous.
Proof. Let f : M → M be a C 2 Anosov diffeomorphism with stable and
unstable distributions Es and Eu , and hyperbolicity constants C and 0 <
λ < 1 < µ. We will prove the absolute continuity of the stable foliation Ws .
Absolute continuity of the unstable foliation Wu follows by reversing the
time. To prove the theorem, we are going to uniformly approximate the
holonomy map by continuous maps with uniformly bounded Jacobians.
As in the proof of Theorem 6.1.3, we assume that M is a compact subman-
ifold in R N [Hir94] and denote by Tx M⊥ the orthogonal complement of Tx M
in R N . Let Ês be a smooth distribution that approximates the continuous
distribution Ẽs (x) = Es (x) ⊕ Tx M⊥ . 

LEMMA 6.2.6. For every θ > 0 there is a constant C1 > 0 such that for every
x ∈ M, for every subspace H ⊂ Tx M of the same dimension as Eu (x) and
θ-transverse to Es (x), and for every k ∈ N,
1. d fxkv ≥ C1 µkv for every v ∈ H,
2. dist(d fxk H, d fxk Eu (x)) ≤ C1 ( µλ )k dist(H, Eu (x)).
Proof. Exercise 6.2.3. 

By compactness of M, there is θ0 > 0 such that (Es (x), Eu (x)) ≥ θ0
for every x ∈ M. Also by compactness, there is a covering of M by finitely
6.2. Absolute Continuity of the Stable and Unstable Foliations 149

f n (L1 )
p̂(y1 ) = f n (x2 )
s
W
x2 = pn (x1 )
pn f n (p(x1 ))
x1 p(x1 ) y1 = f n (x1 )
p
Ê(y1 )
s
W

L1 L2 f n (L2 )

Figure 6.3. Construction of approximating maps pn .

many foliation coordinate charts (Ui , hi ), i = 1, . . . , l, of the stable foliation

Ws . It follows that there are positive constants  and δ such that every y ∈ M
is contained in a coordinate chart U j with the following property: If L is a
compact connected submanifold of U j such that
1. L intersects transversely every local stable leaf of U j ,
2. (Tz L, Es ) > θ0 /3 for all z ∈ L, and
3. dist(y, L) < δ,
then for any subspace E ⊂ Rn with dist(E, Es (y) ⊕ Ty M⊥ ) < , the affine
plane y + E intersects L transversely in a unique point zy , and y − zy  <
6δ/θ0 .
Let (U, h) be a foliation coordinate chart, and L1 , L2 local transversals
in U with holonomy map p: L1 → L2 . Define a map p̂: f n (L1 ) → f n (L2 ) as
follows: For x ∈ L1 , let p̂( f n (x)) be the unique intersection point of the affine
plane f n (x) + Ê( f n (x)) with f n (L2 ) that is closest to f n ( p(x)) along f n (L2 )
(note that there may be several such intersection points). The map p̂ is well
defined by Lemma 6.2.6 and the remarks in the preceding paragraph.
For x ∈ L1 , set pn (x) = f −n ( p̂( f n (x))). Let x1 ∈ L1 , x2 = pn (x1 ) and set
yi = f n (xi ); see Figure 6.3. Observe that
dist( f k(x1 ), f k( p(x1 ))) ≤ Cλk dist(x1 , p(x1 )) for k = 0, 1, 2, . . . . (6.4)
Assuming that Ês is C 0 -close enough to Ẽs , it is, by Lemma 6.2.6, uniformly
transverse to f n (L1 ) and f n (L2 ). Therefore, there is C2 > 0 such that
dist( p̂( f n (x1 )), f n ( p(x1 ))) ≤ C2 dist( f n (x1 ), f n ( p(x1 )))
≤ C2 Cλn dist(x1 , p(x1 )).
Therefore, by (6.4) and Lemma 6.2.6,

n
C2 C λ
dist( pn (x1 ), p(x1 )) ≤ dist(x1 , p(x1 )), (6.5)
C1 µ
and hence pn converges uniformly to p as n → ∞.
150 6. Ergodicity of Anosov Diffeomorphisms

Combining (6.4) and (6.5), we get

dist( f k(x1 ), f k(x2 )) ≤ dist( f k(x1 ), f k( p(x1 ))) + dist( f k( p(x1 )), f k(x2 ))
≤ C3 λk. (6.6)

Let J ( f k(xi )) be the Jacobian of f˜ in the direction of the tangent plane

Ti (xi ) = Tf k(xi ) Li , i = 1, 2, k = 0, 1, 2, . . . . Also, denote by Jac pn the Jacobian
k

of pn , and by Jac p̂ the Jacobian of p̂: f n (L1 ) → f n (L2 ), which is uniformly

bounded by Lemma 6.2.4. Then

n−1
n−1
Jac pn (x1 ) = (J ( f k(x2 )))−1 · Jac p̂ ( f n (x1 )) · J ( f k(x1 )).
k=0 k=0

To obtain a uniform bound on Jac pn we need to estimate the quan-

n−1
tity P = k=0 (J ( f k(x1 ))/J ( f k(x2 ))) from above. By Theorem 6.1.3,
Lemma 6.2.6, and (6.6), for some C4 , C5 , C6 > 0 and ᾱ,
   
dist T1k(x1 ), T2k(x2 ) ≤ dist T1k(x1 ), Ẽu ( f k(x1 ))
+ dist( Ẽu ( f k(x1 )), Ẽu ( f k(x2 )))
 
+ dist T2k(x2 ), Ẽu ( f k(x2 ))

k
λ
≤ 2C1 + C4 (dist( f k(x1 ), f k(x2 )))α
µ

k
λ
≤ 2C1 + C5 λαk ≤ C6 λαk. (6.7)
µ
Since f is a C 2 diffeomorphism, its derivative is Lipschitz continuous, and
the Jacobians J ( f k(x1 )) and J ( f k(x2 )) are bounded away from 0 and ∞.
Therefore it follows from (6.7) that |J ( f k(x1 )) − J ( f k(x2 ))|/|J ( f k(x2 ))| <
C7 λαk. Hence the product P converges and is bounded. 

Exercise 6.2.1. Let W be a k-dimensional foliation of M, and let L be an
(n − k)-dimensional local transversal to W at x ∈ M, i.e., Tx M = Tx W(x) ⊕
Tx L. Prove that there is a neighborhood U  x and a C 1 coordinate chart
w: Bk × Bn−k → U such that the connected component of L ∩ U containing
x is w(0, Bn−k) and there are C 1 functions fy : Bk → Bn−k, y ∈ Bn−k, with the
following properties:
(i) fy depends continuously on y in the C 1 -topology;
(ii) w(graph ( fy )) = WU (w(0, y)).

Exercise 6.2.2. Give an example of an absolutely continuous foliation,

which is not transversely absolutely continuous.
6.3. Proof of Ergodicity 151

Exercise 6.2.3. Prove Lemma 6.2.6.

Exercise 6.2.4. Let Wi , i = 1, 2, be two transverse foliations of dimensions
ki on a smooth manifold M, i.e., Tx W1 (x) ∩ Tx W2 (x) = {0} for each x ∈
M. The foliations W1 and W2 are called integrable if there is a (k1 + k2 )-
dimensional foliation W (called the integral hull of W1 and W2 ) such that
 
W(x) = y∈W1 (x) W2 (y) = y∈W2 (x) W1 (y) for every x ∈ M.
Let W1 be a C 1 foliation and W2 be an absolutely continuous foliation,
and assume that W1 and W2 are integrable with integral hull W. Prove that
W is absolutely continuous.

6.3 Proof of Ergodicity

The proof of Theorem 6.3.1 below follows the main ideas of E. Hopf’s
argument for the ergodicity of the geodesic flow on a compact surface of
variable negative curvature.
We say that a measure µ on a differentiable Riemannian manifold M
is smooth if it has a continuous
 density q with respect to the Riemannian
volume m, i.e., µ(A) = A q(x) dm(x) for each bounded Borel set A ⊂ M.
THEOREM 6.3.1. A C 2 Anosov diffeomorphism preserving a smooth
measure is ergodic.
Proof. Let (X, A, µ) be a finite measure space such that X is a compact
metric space with distance d, µ is a Borel measure, and A is the µ-completion
of the Borel σ-algebra. Let f : X → X be a homeomorphism. For x ∈ X,
define the stable set V s (x) and unstable set V u (x) by the formulas
V s (x) = {y ∈ X: d( f n (x), f n (y)) → 0 as n → ∞},
V u (x) = {y ∈ X: d( f n (x), f n (y)) → 0 as n → −∞}.
LEMMA 6.3.2. Let φ: X → R be an f -invariant measurable function. Then
φ is constant mod 0 on stable and unstable sets, i.e., there is a null set N such
that φ is constant on V s (x)\N and on V u (x)\N for every x ∈ X\N.
Proof. We will only deal with the stable sets. Without loss of generality
assume that φ is non-negative. For a real C set φC (x) = min(φ(x), C). The
function φC is f -invariant, and it suffices to prove the lemma for φC with
arbitrary
 C. For k ∈ N, let ψk: X → R be a continuous function such that
X |φC − ψk| dµ(x) < k . By the Birkhoff ergodic theorem, the limit
1

1n−1
ψk+ (x) = lim ψk( f i (x))
n→∞ n
i=0
152 6. Ergodicity of Anosov Diffeomorphisms

exists for µ-a.e. x. By the invariance of µ and φC , for every j ∈ Z,

 
1
> |φC (x) − ψk(x)| dµ(x) = |φC ( f j (y)) − ψk( f j (y))| dµ(y)
k X X

= |φC (y) − ψk( f j (y))| dµ(y),
X

and hence
  

 1n−1

φC (y) − ψk( f (y)) dµ(y)
i
X n i=0 

1n−1
1
≤ |φC (y) − ψk( f i (y))| dµ(y) < .
n i=0 X k

Since ψk is uniformly continuous, ψk+ (y) = ψk+ (x) whenever y ∈ V s (x) and
ψk+ (x) is defined. Therefore, there is a null set Nk such that ψk+ exists and is
constant on the stable sets in X\Nk. It follows that φC+ (x) = limk→∞ ψk+ (x)

is constant on the stable sets in X\ Nk. Clearly φC (x) = φC+ (x) mod 0.


Let φ be a µ-measurable f -invariant function. By Lemma 6.3.2, there is
a µ-null set Ns such that φ is constant on the leaves of Ws in M\Ns and
another µ-null set Nu such that φ is constant on the leaves of Wu in M\Nu .
Let x ∈ M, and let U  x be a small neighborhood, as in the definition of
absolute continuity for Ws and Wu . Let Gs ⊂ U be the set of points z ∈ U for
which mWs (z) (Ns ∩ Ws (z)) = 0 and z ∈/ Ns . Let Gu ⊂ U be the set of points
z ∈ U for which mWu (z) (Nu ∩ Wu (z)) = 0 and z ∈ / Nu . By Proposition 6.2.1
and the absolute continuity of Wu and Ws (Theorem 6.2.5), both sets Gs
and Gu have full µ-measure in U, and hence so does Gs ∩ Gu . Again, by the
absolute continuity of Wu , there is a full-µ-measure subset of points z ∈ U
such that z ∈ Gs ∩ Gu and mWu (z) -a.e. point from Wu (z) also lies in Gs ∩ Gu .
It follows that φ(x) = φ(z) for µ-a.e. point x ∈ U. Since M is connected, φ is
constant mod 0 on M. 

Exercise 6.3.1. Prove that a C 2 Anosov diffeomorphism preserving a
smooth measure is weak mixing.
CHAPTER SEVEN

Low-Dimensional Dynamics

As we have seen in the previous chapters, general dynamical systems ex-

hibit a wide variety of behaviors and cannot be completely classified by
their invariants. The situation is considerably better in low-dimensional dy-
namics and especially in one-dimensional dynamics. The two crucial tools
for studying one-dimensional dynamical systems are the intermediate value
theorem (for continuous maps) and conformality (for non-singular differen-
tiable maps). A differentiable map f is conformal if the derivative at each
point is a non-zero scalar multiple of an orthogonal transformation, i.e., if the
derivative expands or contracts distances by the same amount in all direc-
tions. In dimension one, any non-singular differentiable map is conformal.
The same is true for complex analytic maps, which we study in Chapter 8.
But in higher dimensions, differentiable maps are rarely conformal.

7.1 Circle Homeomorphisms

The circle S1 = [0, 1] mod 1 can be considered as the quotient space R/Z.
The quotient map π: R → S1 is a covering map, i.e., each x ∈ S1 has a neigh-
borhood Ux such that π −1 (Ux ) is a disjoint union of connected open sets,
each of which is mapped homeomorphically onto Ux by π.
Let f : S1 → S1 be a homeomorphism. We will assume throughout this sec-
tion that f is orientation-preserving (see Exercise 7.1.3 for the orientation-
reversing case). Since π is a covering map, we can lift f to an increasing home-
omorphism F: R → R such that π ◦ F = f ◦ π . For each x0 ∈ π −1 ( f (0))
there is a unique lift F such that F(0) = x0 , and any two lifts differ by an
integer translation. For any lift F and any n ∈ Z, F(x + n) = F(x) + n for
any x ∈ R.

153
154 7. Low-Dimensional Dynamics

THEOREM 7.1.1. Let f : S1 → S1 be an orientation-preserving homeomor-

phism, and F: R → R a lift of f . Then for every x ∈ R, the limit
F n (x) − x
ρ(F) = lim
n→∞ n
exists, and is independent of the point x. The number ρ( f ) = π (ρ(F)) is
independent of the lift F, and is called the rotation number of f . If f has a
periodic point, then ρ( f ) is rational.
Proof. Suppose for the moment that the limit exists for some x ∈ [0, 1).
Since F maps any interval of length 1 to an interval of length 1, it follows
that |F n (x) − F n (y)| ≤ 1 for any y ∈ [0, 1). Thus
|(F n (x) − x) − (F n (y) − y)| ≤ |F n (x) − F n (y)| + |x − y| ≤ 2,
so
F n (x) − x F n (y) − y
lim = lim .
n→∞ n n→∞ n
Since F (y + k) = F n (y) + k, the same holds for any y ∈ R.
n

Suppose F q (x) = x + p for some x ∈ [0, 1) and some p, q ∈ N. This

is equivalent to asserting that π (x) is a periodic point for f with pe-
riod q. For n ∈ N, write n = kq + r, 0 ≤ r < q. Then F n (x) = F r (F kq x)) =
F r (x + kp) = F r (x) + kp, and since |F r (x) − x| is bounded for 0 ≤ r < q,
F n (x) − x p
lim = .
n→∞ n q
Thus the rotation number exists and is rational whenever f has a periodic
point.
Suppose now that F q (x)
=x + p for all x ∈ R and p, q ∈ N. By continuity,
for each pair p, q ∈ N, either F q (x) > x + p for all x ∈ R, or F q (x) < x + p
for all x ∈ R. For n ∈ N, choose pn ∈ N so that pn − 1 < F n (x) − x < pn for
all x ∈ R. Then for any m ∈ N,


m−1
m( pn − 1) < F mn (x) − x = F n (F kn (x)) − F kn (x) < mpn ,
k=0

which implies that

pn 1 F mn (x) − x pn
− < < .
n n mn n
Interchanging the roles of m and n, we also have
pm 1 F mn (x) − x pm
− < < .
m m mn m
7.1. Circle Homeomorphisms 155

Thus, | pm/m − pn /n| < |1/m + 1/n|, so { pn /n} is a Cauchy sequence. It fol-
lows that (F n (x) − x)/n converges as n → ∞.
If G = F + k is another lift of f , then ρ(G) = ρ(F) + k, so ρ( f ) is inde-
pendent of the lift F. Moreover, there is a unique lift F such that ρ(F) = ρ( f )
(Exercise 7.1.1). 

Since S1 = [0, 1] mod 1, we will often abuse notation by writing ρ( f ) = x
for some x ∈ [0, 1].

PROPOSITION 7.1.2. The rotation number depends continuously on the

map in the C 0 topology.
Proof. Let f be an orientation-preserving circle homeomorphism, and
choose p, q, p , q ∈ N such that p/q < ρ( f ) < p /q . Let F be the lift of
f such that p < F q (x) − x < p + q. Then for all x ∈ R, p < F q (x) − x <
p + q, since otherwise we would have ρ( f ) = p/q. If g is another circle
homeomorphism close to F, then there is a lift G close to F, and for g suffi-
ciently close to f , the same inequality p < Gq (x) − x < p + q holds for all
x ∈ R. Thus p/q < ρ(g). A similar argument involving p and q completes
the proof. 

PROPOSITION 7.1.3. Rotation number is an invariant of topological conju-
gacy.
Proof. Let f and h be orientation-preserving homeomorphisms of S1 , and
let F and H be lifts of f and h. Then H ◦ F ◦ H−1 is a lift of h ◦ f ◦ h−1 , and
for x ∈ R,
(HF H−1 )n (x) − x (HF n H−1 )(x) − x
=
n n
H(F H (x)) − F H (x) F n H−1 (x) − H−1 (x)
n −1 n −1
H−1 (x) − x
= + + .
n n n
Since the numerators in the first and third terms of the last expression are
bounded independent of n, we conclude that
(HF H−1 )n (x) − x F n (x) − x
ρ(hf h−1 ) = lim
n→∞ n
= lim
n→∞ n
= ρ( f ). 

PROPOSITION 7.1.4. If f : S1 → S1 is a homeomorphism, then ρ( f ) is ra-

tional if and only if f has a periodic point. Moreover, if ρ( f ) = p/q where p
and q are relatively prime non-negative integers, then every periodic point of
f has minimal period q, and if x ∈ R projects to a periodic point of f , then
F q (x) = x + p for the unique lift F with ρ(F) = p/q.
156 7. Low-Dimensional Dynamics

Proof. The “if” part of the first assertion is contained in Theorem 7.1.1.
Suppose ρ( f ) = p/q, where p, q ∈ N. If F and F̃ = F + l are two lifts
of f , then F̃ q = F q + lq. Thus we may choose F to be the unique lift with
p ≤ F q (0) < p + q. To show the existence of a periodic point of f , it suffices
to show the existence of a point x ∈ [0, 1] such that F q (x) = x + k for some
k ∈ N. We may assume that x + p < F q (x) < x + p + q for all x ∈ [0, 1],
since otherwise we have F q (x) = x + l for k = p or k = p + q, and we are
done. Choose  > 0 such that for any x ∈ [0, 1], x + p +  < F q (x) < x +
p + q − . The same inequality then holds for all x ∈ R, since F q (x + k) =
F q (x) + k for all k ∈ N. Thus

p+ k( p + ) F kq (x) − x k( p + q − ) p+1−

= < < =
q kq kq kq q

for all k ∈ N, contradicting ρ( f ) = p/q. We conclude that F q (x) = x + p or

F q (x) = x + p + q for some x, and x is periodic with period q.
Now assume ρ( f ) = p/q, with p and q relatively prime, and suppose
x ∈ [0, 1) is a periodic point of f . Then there are integers p , q ∈ N such

that F q (x) = x + p . By the proof of Theorem 7.1.1, ρ( f ) = p /q , so if
d is the greatest common divisor of p and q , then q = qd and p = pd.
We claim that F q (x) = x + p. If not, then either F q (x) > x + p or F q (x) <
x + p. Suppose the former holds (the other case is similar). Then by mono-
tonicity,

F dq (x) > F (d−1)q (x) + p > · · · > x + dp,


contradicting the fact that F q (x) = x + p . Thus, x is periodic with period q.


Suppose f is a homeomorphism of S1 . Given any subset A ⊂ S1 and a
distinguished point x ∈ A, we define an ordering on A by lifting A to the
interval [x̃, x̃ + 1) ⊂ R, where x̃ ∈ π −1 (x), and using the natural ordering on
R. In particular, if x ∈ S1 , then the orbit {x, f (x), f 2 (x), . . .} has a natural
order (using x as the distinguished point).

THEOREM 7.1.5. Let f : S1 → S1 be an orientation-preserving homeomor-

phism with rational rotation number ρ = p/q, where p and q are relatively
prime. Then for any periodic point x ∈ S1 , the ordering of the orbit {x, f (x),
f 2 (x), . . . , f q−1 (x)} is the same as the ordering of the set {0, p/q, 2 p/q, . . . ,
(q − 1) p/q}, which is the orbit of 0 under the rotation Rρ .
7.1. Circle Homeomorphisms 157

Proof. Let x be a periodic point of f , and let i ∈ {0, . . . , q − 1} be the

unique number such that f i (x) is the first point to the right of x in the
orbit of x. Then f 2i (x) must be the first point to the right of f i (x), since if
f l (x) ∈ ( f i (x), f 2i (x)) then l > i and f l−i (x) ∈ (x, f i (x)), contradicting the
choice of i. Thus the points of the orbit are ordered as x, f i (x), f 2i (x), . . . ,
f (q−1)i (x).
Let x̃ be a lift of x. Since f i carries each interval [ f ki (x), f (k+1)i (x)] to its
successor, and there are q of these intervals, there is a lift F̄ of f i such that
F̄ q x̃ = x̃ + 1. Let F be the lift of f with F q (x) = x + p. Then F i is a lift of
f i , so F i = F̄ + k for some k. We have

x + i p = F qi (x) = ( F̄ + k)q (x) = F̄ q (x) + qk = x + 1 + qk.

Thus i p = 1 + qk, so i is the unique number between 0 and q such that

i p = 1 mod q. Since the points of the set {0, p/q, 2 p/q, . . . , (q − 1) p/q} are
ordered as 0, i p/q, . . . , (q − 1)i p/q, the theorem follows. 

Now we turn to the study of orientation-preserving homeomorphisms
with irrational rotation number. If x and y are two points in S1 , then we
define the interval [x, y] ⊂ S1 to be π([x̃, ỹ]), where x̃ ∈ π −1 (x) and ỹ =
π −1 (y) ∩ [x̃, x̃ + 1). Open and half-open intervals are defined in a similar
way.

LEMMA 7.1.6. Suppose ρ( f ) is irrational. Then for any x ∈ S1 and any dis-
tinct integers m > n, every forward orbit of f intersects the interval
I = [ f m(x), f n (x)].

Proof. It suffices to show that S1 = ∞
k=0 f
−k
I. Suppose not. Then


∞ 
∞
 −(k−1)m+kn 
S1
⊂ f −k(m−n) I = f (x), f −km+(k+1)n (x) .
k=1 k=1

Since the intervals f −k(m−n) I abut at the endpoints, we conclude that

f −k(m−n) f n (x) converges monotonically to a point z ∈ S1 , which is a fixed
point for f m−n , contradicting the irrationality of ρ( f ). 

PROPOSITION 7.1.7. If ρ( f ) is irrational, then ω(x) = ω(y) for any x, y ∈
S1 , and either ω(x) = S1 or ω(x) is perfect and nowhere dense.
Proof. Fix x, y ∈ S1 . Suppose f an (x) → x0 ∈ ω(x) for some sequence
an  ∞. By Lemma 7.1.6, for each n ∈ N, we can choose bn such that
f bn (y) ∈ [ f an−1 (x), f an (x)]. Then f bn (y) → x0 , so ω(x) ⊂ ω(y). By sym-
metry, ω(x) = ω(y).
158 7. Low-Dimensional Dynamics

To show that ω(x) is perfect, we fix z ∈ ω(x). Since ω(x) is invariant, we

have that z ∈ ω(z) is a limit point of { f n (z)} ⊂ ω(x), so ω(x) is perfect.
To prove the last claim, we suppose that ω(x)
= S1 . Then ∂ω(x) is a
non-empty closed invariant set. If z ∈ ∂ω(x), then ω(z) = ω(x). Therefore,
ω(x) ⊂ ∂ω(x) and ω(x) is nowhere dense. 

LEMMA 7.1.8. Suppose ρ( f ) is irrational. Let F be a lift of f , and ρ =
ρ(F). Then for any x ∈ R, n1 ρ + m1 < n2 ρ + m2 if and only if F n1 (x) + m1 <
F n2 (x) + m2 , for any m1 , m2 , n1 , n2 ∈ Z,.
Proof. Suppose F n1 (x) + m1 < F n2 (x) + m2 or, equivalently,
F (n1 −n2 ) (x) < x + m2 − m1 .
This inequality holds for all x, since otherwise the rotation number would
be rational. In particular, for x = 0 we have F (n1 −n2 ) (0) < m2 − m1 . By an
inductive argument, F k(n1 −n2 ) (0) < k(m2 − m1 ). If n1 − n2 > 0, it follows that
F k(n1 −n2 ) (0) − 0 m2 − m1
< ,
k(n1 − n2 ) n1 − n2
so ρ = limk→∞ F k(n1 −n2 ) (0)/k(n1 − n2 ) ≤ (m2 − m1 )/(n1 − n2 ). Irrationality
of ρ implies strict inequality, so n1 ρ + m1 < n2 ρ + m2 . The same result holds
in the case n1 − n2 < 0 by a similar argument. The converse follows by re-
versing the inequality. 

THEOREM 7.1.9 (Poincaré Classification). Let f : S1 → S1 be an orientation-
preserving homeomorphism with irrational rotation number ρ.
1. If f is topologically transitive, then f is topologically conjugate to the
rotation Rρ .
2. If f is not topologically transitive, then Rρ is a factor of f , and the factor
map h: S1 → S1 can be chosen to be monotone.
Proof. Let F be a lift of f , and fix x ∈ R. Let A = {F n (x) + m: n, m ∈ Z} and
B = {nρ + m: n, m ∈ Z}. Then B is dense in R (§1.2). Define H: A → B by
H(F n (x) + m) = nρ + m. By the preceding lemma, H preserves order and
is bijective. Extend H to a map H: R → R by defining
H(y) = sup{nρ + m: F n (x) + m < y}.
Then H(y) = inf{nρ + m: F n (x) + m > y}, since otherwise R\B would con-
tain an interval.
We claim that H: R → R is continuous. If y ∈ Ā, then H(y) = sup{H(z):
z ∈ A, z < y} and H(y) = inf{H(z): z ∈ A, z > y} implies that H is continuous
7.1. Circle Homeomorphisms 159

on Ā. If I is an interval in R\ Ā, then H is constant on I and the constant

agrees with the values at the endpoints. Thus H: R → R is a continuous
extension of H: A → B.
Note that H is surjective, non-decreasing, and that

H(y + 1) = sup{nρ + m: F n (x) + m < y + 1}

= sup{nρ + m: F n (x) + (m − 1) < y} = H(y) + 1.

Moreover,

H(F(y)) = sup{nρ + m: F n (x) + m < F(y)}

= sup{nρ + m: F n−1 (x) + m < y}
= ρ + H(y).

We conclude that H descends to a map h: S1 → S1 and h ◦ f = Rρ ◦ h.

Finally, note that f is transitive if and only if {F n (x) + m: n, m ∈ Z} is
dense in R. Since H is constant on any interval in R\ Ā, we conclude that h is
injective if and only if f is transitive. (Note that by Proposition 7.1.7, either
every orbit is dense or no orbit is dense.) 

Exercise 7.1.1. Show that if F and G = F + kare two lifts of f , then ρ(F) =
ρ(G) + k, so ρ( f ) is independent of the choice of lift used in its definition.
Show that there is a unique lift F of f such that ρ(F) = ρ( f ).

Exercise 7.1.2. Show that ρ( f m) = mρ( f ).

Exercise 7.1.3. Show that if f is an orientation-reversing homeomorphism

of S1 , then ρ( f 2 ) = 0.

Exercise 7.1.4. Suppose f has rational rotation number. Show that:

(a) if f has exactly one periodic orbit, then every non-periodic point is
both forward and backward asymptotic to the periodic orbit; and
(b) if f has more than one periodic orbit, then every non-periodic orbit is
forward asymptotic to some periodic orbit and backward asymptotic
to a different periodic orbit.

Exercise 7.1.5. Show that Theorems 7.1.1 and 7.1.5 hold under the weaker
hypothesis that f : S1 → S1 is a continuous map such that any (and thus
every) lift F of f is non-decreasing.
160 7. Low-Dimensional Dynamics

7.2 Circle Diffeomorphisms

The total variation of a function f : S1 → R is


n
Var( f ) = sup | f (xk) − f (xk+1 )|,
k=1

where the supremum is taken over all partitions 0 ≤ x1 < · · · < xn ≤ 1, for
all n ∈ N. We say that g has bounded variation if Var(g) is finite. Note that
any Lipshitz function has bounded variation. In particular, any C 1 function
has bounded variation.
THEOREM 7.2.1 (Denjoy). Let f be an orientation-preserving C 1 diffeo-
morphism of the circle with irrational rotation number ρ = ρ( f ). If f  has
bounded variation, then f is topologically conjugate to the rigid rotation Rρ .
Proof. We know from Theorem 7.1.9 that if f is transitive, it is conjugate
to Rρ . Thus we assume that f is not transitive, and argue to obtain a contra-
diction. By Proposition 7.1.7, we may assume that ω(0) is a perfect, nowhere
dense set. Then S1 \ω(0) is a disjoint union of open intervals. Let I = (a, b)
be one of these intervals. Then the intervals { f n (I)}n∈Z are pairwise disjoint,

since otherwise f would have a periodic point. Thus n∈Z l ( f n (I)) ≤ 1,
b
where l ( f n (I)) = a ( f n ) (t) dt is the length of f n (I).
LEMMA 7.2.2. Let J be an interval in S1 , and suppose the interiors of the
intervals J, f (J ), . . . , f n−1 (J ) are pairwise disjoint. Let g = log f  , and fix
x, y ∈ J . Then for any n ∈ Z,
Var(g) ≥ | log( f n ) (x) − log( f n ) (y)|.

Proof. Using the fact that the intervals J, f (J ), . . . , f n (J ) are disjoint, we

get
 


n−1 
n−1 
 
Var(g) ≥ |g( f (y)) − g( f (x))| ≥ 
k k
g( f (y)) − g( f (x))
k k

k=0
 k=0

 

n−1
n−1 
 
= log f  ( f k(y)) − log f  ( f k(x))
 k=0 k=0


= | log( f n ) (y) − log( f n ) (x)|. 

Fix x ∈ S1 . We claim that there are infinitely many n ∈ N such that the
intervals (x, f −n (x)), ( f (x), f 1−n (x)), . . . , ( f n (x), x) are pairwise disjoint. It
suffices to show that there are infinitely many n such that f k(x) is not in the
interval (x, f n (x)) for 0 ≤ |k| ≤ n. Lemma 7.1.8 implies that the orbit of x is
7.2. Circle Diffeomorphisms 161

ordered in the same way as the orbit of a point under the irrational rotation
Rρ . Since the orbit of a point under an irrational rotation is dense, the claim
follows.
Choose n as in the preceding paragraph. Then by applying Lemma 7.2.2
with y = f −n (x), we obtain
 
 ( f n ) (x) 
Var(g) ≥ log n  = | log(( f n ) (x)( f −n ) (x))|.
( f ) (y) 
Thus for infinitely many n ∈ N, we have

l ( f n (I)) + l ( f −n (I)) = ( f n ) (x) dx + ( f −n ) (x) dx

I I

= [( f n ) (x) + ( f −n ) (x)] dx
I


≥ ( f n ) (x)( f −n ) (x) dx
I

 
1
≥ exp (−Var(g)) dx = exp − Var(g) l (I).
I 2

This contradicts the fact that n∈Z l ( f n (I)) ≤ ∞, so we conclude that f is
transitive, and therefore conjugate to Rρ . 

THEOREM 7.2.3 (Denjoy Example). For any irrational number ρ ∈ (0, 1),
there is a non-transitive C 1 orientation-preserving diffeomorphism f : S1 → S1
with rotation number ρ.
Proof. We know from Lemma 7.1.8 that if ρ( f ) = ρ, then for any x ∈ S1 ,
the orbit of x is ordered the same way as any orbit of Rρ , i.e., f k(x) < f l (x) <
f m(x) if and only if Rρk(x) < Rρl (x) < Rρm(x). Thus in constructing f , we have
no choice about the order of the orbit of any point. We do, however, have a
choice about the spacing between points in the orbit.

Let {ln }n∈Z be a sequence of positive real numbers such that n∈Z ln = 1
and ln is decreasing as n → ±∞ (we will impose additional constraints later).
Fix x0 ∈ S1 , and define


an = l k, bn = an + ln .
{k∈Z:Rρk (x0 )∈[x0 ,Rρn (x0 )}


The intervals [an , bn ] are pairwise disjoint. Since n∈Z ln = 1, the union of
these intervals covers a set of measure 1 in [0, 1], and is therefore dense.
To define a C 1 homeomorphism f : S1 → S1 it suffices to define a contin-
uous, positive function g on S1 with total integral 1. Then f will be defined
162 7. Low-Dimensional Dynamics

to be the
 b integral of g. The function g should satisfy:
1. ann g(t) dt = ln+1 .
To construct such a g it suffices to define g on each interval [an , bn ] so that it
also satisfies:
2. g(an ) = g(bn ) = 1.

3. For any sequence {xk} ⊂ n∈Z [an , bn ], if y = lim xk, then g(xk) → 1.

We then define g to be 1 on S1 \ n∈Z [an , bn ].
There are many such possibilities for g|[an , bn ]. We use the quadratic
polynomial
6(ln+1 − ln )
g(x) = 1 + (bn − x)(x − an ),
ln3
which clearly satisfies condition 1. For n ≥ 0, we have ln+1 − ln < 0, so

6(ln − ln+1 ) ln 2 3ln+1 − ln
1 ≥ g(x) ≥ 1 − 3
= .
ln 2 2ln
For n < 0, we have ln+1 − ln > 0, so
3ln+1 − ln
1 ≤ g(x) ≤ .
2ln
Thus if we choose ln such that (3ln+1 − ln )/2ln → 1 as n → ±∞, then condi-
tion 3 is satisfied. For example, we could choose ln = α(|n| + 2)−1 (|n| + 3)−1 ,

where α = 1/ n∈Z ((|n| + 2)−1 (|n| + 3)−1 ).
x
Now define f (x) = a1 + 0 g(t) dt. Using the results above, it follows
that f : S1 → S1 is a C 1 homeomorphism of S1 with rotation number ρ

(Exercise 7.2.1). Moreover, f n (0) = an , and ω(0) = S1 \ n∈Z (an , bn ) is a
closed, perfect, invariant set of measure zero. 

Exercise 7.2.1. Verify the statements in the last paragraph of the proof of
Theorem 7.2.3.
Exercise 7.2.2. Show directly that the example constructed in the proof of
Theorem 7.2.3 is not C 2 .

7.3 The Sharkovsky Theorem

We consider the set NSh = N ∪ {2∞ } obtained by adding the formal symbol
2∞ to the set of natural numbers. The Sharkovsky ordering of this set is
1 ≺ 2 ≺ · · · ≺ 2n ≺ · · · ≺ 2∞ ≺ · · ·
≺ 2m · (2n + 1) ≺ · · · ≺ 2m · 7 ≺ 2m · 5 ≺ 2m · 3 ≺ · · ·
≺ 2(2n + 1) ≺ · · · ≺ 14 ≺ 10 ≺ 6 ≺ · · ·
≺ 2n + 1 ≺ · · · ≺ 7 ≺ 5 ≺ 3.
7.3. The Sharkovsky Theorem 163

The symbol 2∞ is added so that NSh has the least-upper-bound property,

i.e., every subset of NSh has a supremum. The Sharkovsky ordering is pre-
served by multiplication by 2k, for any k ≥ 0 (where 2k · 2∞ = 2∞ , by
definition).
For α ∈ NSh , let S(α) = {k ∈ N: k  α} (note that S(α) is defined to be a
subset of N, not NSh ). For a map f : [0, 1] → [0, 1], we denote by MinPer( f )
the set of minimal periods of periodic points of f .

THEOREM 7.3.1 (Sharkovsky [Sha64]). For every continuous map f :

[0, 1] → [0, 1], there is α ∈ NSh such that MinPer( f ) = S(α). Conversely, for
every α ∈ NSh , there is a continuous map f : [0, 1] → [0, 1] with MinPer( f ) =
S(α).

The proof of the first assertion of the Sharkovsky theorem proceeds as

follows: We assume that f has a periodic point x of minimal period n > 1,
since otherwise there is nothing to show. The orbit of x partitions the interval
[0, 1] into a finite collection of subintervals whose endpoints are elements of
the orbit. The endpoints of these intervals are permuted by f . By examining
the combinatorial possibilities for the permutations of pairs of endpoints,
and using the intermediate value theorem, one establishes the existence of
periodic points of the desired periods.
The second assertion of the Sharkovsky theorem is proved as
Lemma 7.3.9.
If I and J are intervals in [0, 1] and f (I) ⊃ J , we say that I f-covers J ,
and we write I → J . If a, b ∈ [0, 1], then we will use [a, b] to represent the
closed interval between a and b, regardless of whether a ≥ b or a ≤ b.

LEMMA 7.3.2
1. If f (I) ⊃ I, then the closure of I contains a fixed point of f .
2. Fix m ∈ N ∪ {∞}, and suppose that {Ik}1≤k<m is a finite or infinite se-
quence of non-empty closed intervals in [0, 1] such that f (Ik) ⊃ Ik+1 for
1 ≤ k < m − 1. Then there is a point x ∈ I1 such that f k(x) ∈ Ik+1 for
1 ≤ k < m − 1. Moreover, if In = I1 for some n > 0, then I1 contains a
periodic point x of period n such that f k(x) ∈ Ik+1 for k = 1, . . . , n − 1.
Proof. The proof of part 1 is a simple application of the intermediate value
theorem.
To prove part 2, note that since f (I1 ) ⊃ I2 , there are points a0 , b0 ∈ I1
that map to the endpoints of I2 . Let J1 be the subinterval of I1 with
endpoints a0 , b0 . Then f (J1 ) = I2 . Suppose we have defined subintervals
J1 ⊃ J2 ⊃ · · · ⊃ Jn in I1 such that f k(Jk) = Ik+1 . Then f n+1 (Jn ) = f (In+1 ) ⊃
In+2 , so there is an interval Jn+1 ⊂ Jn such that f n+1 (Jn+1 ) = In+2 . Thus we
obtain a nested sequence {Jn } of non-empty closed intervals. The intersection
164 7. Low-Dimensional Dynamics

m−1
i=1 Ji is non-empty, and for any x in the intersection, f k(x) ∈ Ik+1 for
1 ≤ k < m − 1.
The last assertion follows from the preceding paragraph together with
part 1. 

A partition of an interval I is a (finite or infinite) collection of closed subin-
tervals {Ik}, with pairwise disjoint interiors, whose union is I. The Markov
graph of f associated to the partition {Ik} is the directed graph with ver-
tices Ik, and a directed edge from Ii to I j if and only if Ii f-covers I j . By
Lemma 7.3.2, any loop of length n in the Markov graph of f forces the
existence of a periodic point of (not necessarily minimal) period n.
As a warmup to the proof of the full Sharkovsky theorem, we prove
that the existence of a periodic point of minimal period three implies the
existence of periodic points of all periods. This result was rediscovered in
1975 by T. Y. Li and J. Yorke, and popularized in their paper “Period three
implies chaos” [LY75].
Let x be a point of period three. Replacing x with f (x) or f 2 (x) if nec-
essary, we may assume that x < f (x) and x < f 2 (x). Then there are two
cases: (1) x < f (x) < f 2 (x) or (2) x < f 2 (x) < f (x). In the first case, we let
I1 = [x, f (x)] and I2 = [ f (x), f 2 (x)]. The associated Markov graph is one
of the two graphs shown in Figure 7.1.
For k ≥ 2, the path I1 → I2 → I2 → · · · → I2 → I1 of length k implies the
existence of a periodic point y of period k with the itinerary I1 , I2 , I2 , . . . ,
I2 , I1 . If the minimal period of y is less than k, then y ∈ I1 ∩ I2 = { f (x)}. But
f (x) does not have the specified itinerary for k
= 3, so the minimal period of
y is k. A similar argument applies to case (2), and this proves the Sharkovsky
theorem for n = 3.
To prove the full Sharkovsky theorem it is convenient to use a sub-
graph of the Markov graph defined as follows. Let P = {x1 , x2 , . . . , xn } be
a periodic orbit of (minimal) period n > 1, where x1 < x2 < · · · < xn . Let

I1 I2

I1 I2

Figure 7.1. The two possible Markov graphs for period three.
7.3. The Sharkovsky Theorem 165

I j = [x j , x j+1 ]. The P-graph of f is the directed graph with vertices I j ,

and a directed edge from I j to Ik if and only if Ik ⊂ [ f (x j ), f (x j+1 )]. Since
f (I j ) ⊃ [ f (x j ), f (x j+1 )], it follows that the P-graph is a subgraph of the
Markov graph associated to the same partition. In particular, any loop in
the P-graph is also a loop in the Markov graph. The P-graph has the virtue
that it is completely determined by the ordering of the periodic orbit, and is
independent of the behavior of the map on the intervals I j . For example, in
Figure 7.1, the top graph is the unique P-graph for a periodic orbit of period
three with ordering x < f (x) < f 2 (x).

LEMMA 7.3.3. The P-graph of f contains a trivial loop, i.e., there is a vertex
I j with a directed edge from I j to itself.
Proof. Let j = max{i: f (xi ) > xi }. Then f (x j ) > x j and f (x j+1 ) ≤ x j+1 , so
f (x j ) ≥ x j+1 and f (x j+1 ) ≤ x j . Thus [ f (x j ), f (x j+1 )]) ⊃ [x j , x j+1 ]. 

We will renumber the vertices of the P-graph (but not the points of P)
so that I1 = [x j , x j+1 ], where j = max{i: f (xi ) > xi }. By the proof of the
preceding lemma, I1 is a vertex with a directed edge from itself to itself.
For any two points xi < xk in P, define


k−1
fˆ([xi , xk]) = [ f (xl ), f (xl+1 )].
l=i

In particular, fˆ(Ik) = [ f (xk), f (xk + 1)]. If fˆ(Ik) ⊃ Il , we say that Ik fˆ-covers

Il . Since we will only be using P-graphs throughout the remainder of this
section, we also redefine the notation Ik → Il to mean that Ik fˆ-covers Il .

PROPOSITION 7.3.4. Any vertex of the P-graph can be reached from I1 .

Proof. The nested sequence I1 ⊂ fˆ(I1 ) ⊂ fˆ2 (I1 ) ⊂ · · · must eventually sta-
bilize, since fˆk(I1 ) is an interval whose endpoints are in the orbit of x. Then
for k sufficiently large, O(x) ∩ fˆk(I1 ) is an invariant subset of O(x), and is
therefore equal to O(x). It follows that fˆk(I1 ) = [x1 , xn ], so any vertex of the
P-graph can be reached from I1 . 

LEMMA 7.3.5. Suppose the P-graph has no directed edge from any interval
Ik, k
=
1, to I1 . Then n is even, and f has a periodic point of period 2.
Proof. Let J0 = [x1 , x j ] and J1 = [x j+1 , xn−1 ], where j = max{i: f (xi ) > xi }
(the case j = 1 is not excluded a priori). Then fˆ(J0 ) ∈ / J0 (since f (x j ) > x j )
and f (J0 ) ∈
ˆ / I1 , so f (J0 ) ⊂ J1 , since f (J0 ) is connected. Likewise, fˆ(J1 ) ⊂ J0 .
ˆ ˆ
Now f (J0 ) ∪ fˆ(J1 ) ⊃ O(x), so fˆ(J0 ) = J1 and fˆ(J1 ) = J0 . Thus J0 f-covers
ˆ
166 7. Low-Dimensional Dynamics

I1 In−1

I2 I3 I4 I5 In−2

Figure 7.2. The P-graph for Lemmas 7.3.6 and 7.3.9.

J1 and J1 f -covers J0 , so f has a periodic point of minimal period 2, and

n = |O(x)| = 2|O(x) ∩ J0 | is even. 

LEMMA 7.3.6. Suppose n > 1 is odd and f has no non-fixed periodic points
of smaller odd period. Then there is a numbering of the vertices of the P-graph
so that the graph contains the following edges, and no others (see Figure 7.2):
1. I1 → I1 and In−1 → I1 ,
2. Ii → Ii+1 , for i = 1, . . . , n − 2 ,
3. In−1 → I2i+1 , for 0 ≤ i < (n − 1)/2.
Proof. By Lemma 7.3.5 and Lemma 7.3.4, there is a non-trivial loop in the
P-graph starting from I1 . By choosing a shortest such loop and renumbering
the vertices of the graph, we may assume that we have a loop

I1 → I2 → · · · → Ik → I1 (7.1)

in the P-graph, k ≤ n − 1. The existence of this loop implies that f has a

periodic point of minimal period k. The path

I1 → I1 → I2 → · · · → Ik → I1

implies the existence of a periodic point of minimal period k + 1. By the

minimality of n, we conclude that k = n − 1, which proves statement 1.
Let I1 = [x j , x j+1 ]. Note that fˆ(I1 ) contains I1 and I2 , but no other Ii ,
since otherwise we would have a shorter path than (7.1). Similarly, if 1 ≤
i < n − 2, then fˆ(Ii ) cannot contain Ik for k > i + 1. Thus fˆ(I1 ) = [x j , x j+2 ]
or fˆ(I1 ) = [x j−1 , x j+1 ]. Suppose the latter holds (the other case is similar).
Then I2 = [x j−1 , x j ], f (x j+1 ) = x j−1 , and f (x j ) = x j+1 . If 2 < n − 1, then
fˆ(I2 ) can contain at most I2 and I3 , so f (x j−1 ) = x j+2 . Continuing in this
way (see Figure 7.3), we find that the intervals of the partition are ordered

In−1 In−3 I2 I1 I3 In−4 In−2

x1 x2 x3 xj−1 xj xj+1 xn−1 xn

Figure 7.3. The action of f from Lemma 7.3.6 on xk is shown by arrows.

7.3. The Sharkovsky Theorem 167

on the interval I as follows:

In−1 , In−3 , . . . , I2 , I1 , I3 , . . . , In−2 .

Moreover, f (x2 ) = xn , f (xn ) = x1 , and f (x1 ) = x j , so fˆ(In−1 ) = [x j , xn−1 ],

and fˆ(In−1 ) contains all the odd-numbered intervals, which completes the
proof of the lemma. 

COROLLARY 7.3.7. If n is odd, then f has a periodic point of minimal period
q for any q > n and for any even integer q < n.
Proof. Let m > 1 be the minimal odd period of a non-fixed periodic point.
By the preceding lemma, there are paths of the form

I1 → I1 → · · · → I1 → I2 → · · · → Im−1 → I1

of any length q ≥ m. For q = 2i < m, the path

Im−1 → Im−2i → Im−2i+1 → · · · → Im−1

gives a periodic point of period q. The verification that these periodic points
have minimal period q is left as an exercise (Exercise 7.3.3). 

LEMMA 7.3.8. If n is even, then f has a periodic point of minimal period 2.
Proof. Let m be the smallest even period of a non-fixed periodic point, and
let I1 be an interval of the associated partition that fˆ-covers itself. If no other
interval fˆ-covers I1 , then Lemma 7.3.5 implies that m = 2.
Suppose then that some other interval fˆ-covers I1 . In the proof of
Lemma 7.3.6, we used the hypothesis that n is odd only to conclude the
existence of such an interval. Thus the same argument as in the proof of that
lemma implies that the P-graph contains the paths

I1 → I2 → · · · → In−1 → I1 and In−1 → I2i for 0 ≤ i < n/2.

Then In−1 → In−2 → In−1 implies the existence of a periodic point of minimal
period 2. 

Conclusion of the proof of the first assertion of the Sharkovsky
Theorem. There are two cases to consider:
1. n = 2k, k > 0. If q ≺ n, then q = 2l with 0 ≤ l < k. The case l = 0 is
l−1
trivial. If l > 0, then g = f q/2 = f 2 has a periodic point of period
2k−l+1 , so by Lemma 7.3.8, g has a non-fixed periodic point of period
2. This point is a fixed point for f q , i.e., it has period q for f . Since it
is not fixed by g, its minimal period is q.
168 7. Low-Dimensional Dynamics

k
2. n = p2k, p odd. The map f 2 has a periodic point of minimal period p,
k
so by Corollary 7.3.7, f 2 has periodic points of minimal period m for
all m ≥ p and all even m < p. Thus f has periodic points of minimal
period m2k for all m ≥ p and all even m < p. In particular, f has a
periodic point of minimal period 2k+1 , so by case 1, f has periodic
points of minimal period 2i for i = 0, . . . , k. 

The next lemma finishes the proof of the Sharkovsky theorem.

LEMMA 7.3.9. For any α ∈ NSh , there is a continuous map f : [0, 1] → [0, 1]
such that MinPer( f ) = S(α).
Proof. We distinguish three cases:
1. α ∈ N, α odd,
2. α ∈ N, α even, and
3. α = 2∞ .
Case 1. Suppose n ∈ N is odd, and α = n. Choose points x0 , . . . , xn−1 ∈
[0, 1] such that

0 = xn−1 < · · · < x4 < x2 < x0 < x1 < x3 < · · · < xn−2 = 1,

and let I1 = [x0 , x1 ], I2 = [x2 , x0 ], I3 = [x1 , x3 ], etc. Let f : [0, 1] → [0, 1] be

the unique map defined by:
1. f (xi ) = xi+1 , i = 0, . . . , n − 2, and f (xn−1 ) = x0 ,
2. f is linear (or affine, to be precise) on each interval I j , j = 1, . . . ,
n − 1.
Then x0 is periodic of period n, and the associated P-graph is shown in
Figure 7.2. Any path that avoids I1 has even length. Loops of length less
than n must be of the following form:
1. Ii → Ii+1 → · · · → In−1 → I2 j+1 → I2 j+2 → · · · → Ii for i > 1, or
2. In−1 → I2i+1 → · · · → In−1 , or
3. I1 → I1 → · · · → I1 → I1
Paths of type 1 or 2 have even length, so no point in int(I j ), j = 2, . . . , n − 1,
can have odd period k < n. Since f (I1 ) = I1 ∪ I2 , we have | f  | > 1 on I1 ,
so every non-fixed point in int(I1 ) must move away from the (unique) fixed
point in I1 , and therefore eventually enters I2 . Once a point enters I2 , it must
enter every I j before it returns to int(I1 ). Thus there is no non-fixed periodic
point in I1 of period less than n. It follows that no point has odd period less
than n. This finishes the proof of the theorem for n odd.
7.3. The Sharkovsky Theorem 169

f D(f ) D(D(f ))
Figure 7.4. Graphs of D ( f ) for f ≡ 1/2.
k

Case 2. Suppose n ∈ N is even, and α = n. For f : [0, 1] → [0, 1], define a

new function D: [0, 1] → [0, 1] by
2 1  

 + f (3x) x ∈ 0, 13 ,
3 3    
D( f )(x) = (2 + f (1)) 23 − x x ∈ 13 , 23 ,

  

x − 23 x ∈ 23 , 1 .

The operator D( f ) is sometimes called the doubling operator, because

MinPer(D( f )) = 2 MinPer( f ) ∪ {1}, i.e., D doubles the periods of a map. To
see this, let g = D( f ), and let I1 = [0, 1/3], I2 = [1/3, 2/3], and I3 = [2/3, 1].
For x ∈ I1 , we have g 2 (x) = f (3x)/3, so g 2k(x) = f k(3x)/3. Thus g 2k(x) = x
if and only if f k(3x) = 3x, so MinPer(g) ⊃ 2 MinPer( f ) (see Figure 7.4).
On the interval I2 , |g  | ≥ 2, so there is a unique repelling fixed point in
(1/3, 2/3), and every other point eventually leaves this interval and never
returns, since g(I1 ∪ I3 ) ∩ I2 = ∅. Thus no non-fixed point in I2 is periodic.
Finally, any periodic point in I3 enters I1 , so its period is in 2MinPer( f ),
which verifies our claim that MinPer(D( f )) = 2MinPer( f ) ∪ {1}.
Since n is even, we can write n = p2k, where p is odd and k > 0. Let f be
a map whose minimum odd period is p (see case 1). Then MinPer(Dk( f )) =
2kMinPer( f ) ∪ {2k−1 , 2k−2 , . . . , 1}, which settles case 2 of the lemma.
Case 3. Suppose α = 2∞ . Let gk = Dk(Id), where Id is the identity map.
Then, by the induction and the remarks in the proof of case 2, MinPer(gk) =
{2k−1 , 2k−2 , . . . , 1}. The sequence {gk}k∈N converges uniformly to a contin-
uous map g∞ : [0, 1] → [0, 1], and g∞ = gk on [2/3k, 1] (Exercise 7.3.4). It
follows that MinPer(g∞ ) ⊃ S(2∞ ).
Let x be a periodic point of g∞ . If 0 ∈ / O(x), then O(x) ⊂ [2/3k, 1] for k
sufficiently large, so x is a periodic point of gk and has even period. Suppose
then that 0 is periodic with period p. If p  2∞ , then there is q ∈ N such
170 7. Low-Dimensional Dynamics

that p  q  2∞ . By the first part of the Sharkovsky theorem, g∞ has a

periodic point y with minimal period q. Since 0 ∈ O(y), we conclude by
the preceding argument that q is even, which contradicts q  2∞ . Thus
MinPer(g∞ ) = S(2∞ ).
This concludes the proof of Lemma 7.3.9, and thus the proof of
Theorem 7.3.1. 

Exercise 7.3.1. Let σ be a permutation of {1, . . . , n − 1}. Show that there is
a continuous map f : [0, 1] → [0, 1] with a periodic point x of period n such
that x < f σ (1) < · · · < f σ (n−1) .

Exercise 7.3.2. Show that there are maps f, g: [0, 1] → [0, 1], each with a
periodic point of period n (for some n), such that the associated P-graphs
are not isomorphic. (Note that for n = 3, all P-graphs are isomorphic.)

Exercise 7.3.3. Verify that the periodic points in the proof of Corollary 7.3.7
have minimal period q.

Exercise 7.3.4. Show that the sequence {gk}k∈N defined near the end of the
proof of Lemma 7.3.9 converges uniformly, and the limit g∞ satisfies g∞ = gk
on [2/3k, 1].

7.4 Combinatorial Theory of Piecewise-Monotone Mappings1

Let I = [a, b] be a compact interval. A continuous map f : I → I is piecewise
monotone if there are points a = c0 < c1 < · · · < cl < cl+1 = b such that f
is strictly monotone on each interval Ii = [ci−1 , ci ], i = 1, . . . , l + 1. We al-
ways assume that each interval [ci−1 , ci ] is a maximal interval on which f is
monotone, so the orientation of f reverses at the turning points c1 , . . . , cl .
The intervals Ii are called laps of f .
Note that any piecewise-monotone map f : I → I can be extended to a
piecewise-monotone map of a larger interval J in such a way that f (∂ J ) ⊂
∂ J . Thus we assume (without losing much generality) that f (∂ I) ⊂ ∂ I. If f
has l turning points and f (∂ I) ⊂ ∂ I, then f is l-modal. If f has exactly one
turning point, then f is unimodal.
The address of a point x ∈ I is the symbol c j if x = c j for some j ∈
{1, . . . , l}, or the symbol I j if x ∈ I j and x ∈
/ {c1 , . . . , cl }. Note that c0 and
cl+1 are not included as addresses. The itinerary of x is the sequence i(x) =

1 Our arguments in this section follow in part those of [CE80] and [MT88].
7.4. Combinatorial Theory of Piecewise-Monotone Mappings 171

(i k(x))k∈N0 , where i k(x) is the address of f k(x). Let

= {I1 , . . . , Il+1 , c1 , . . . , cl }N0 .
Then i: I → , and i ◦ f = σ ◦ i, where σ is the one-sided shift on .
Example. Any quadratic map qµ (x) = µx(1 − x), 0 < µ ≤ 4, is a unimodal
map of I = [0, 1], with turning point c1 = 1/2, I1 = [0, 1/2], I2 = [1/2, 1]. If
0 < µ < 2, then f (I) ⊂ [0, 1/2), so the only possible itineraries are
(I1 , I1 , . . .), (c1 , I1 , I1 , . . .), and (I2 , I1 , I1 , . . .). Note that the map i: [0, 1] →
is not continuous at c1 .
If µ = 2, then the possible itineraries are (I1 , I1 , . . .), (c1 , c1 , . . .), and
(I2 , I1 , I1 , . . .). If 2 < µ < 3, there is an attracting fixed point (µ − 1)/µ ∈
(1/2, 2/3). Thus the possible itineraries are:
(I1 , I1 , . . .),
(c1 , I2 , I2 , . . .),
(I2 , I2 , . . .),
(I1 , . . . , I1 , I2 , I2 , . . .),
(I1 , . . . , I1 , C1 , I2 , I2 , . . .),
any of the above preceded by I2 .

LEMMA 7.4.1. The itinerary i(x) is eventually periodic if and only if the
iterates of x converge to a periodic orbit of f .
Proof. If i(x) is eventually periodic, then by replacing x by one of its forward
iterates, we may assume that i(x) is periodic, of period p. If i j (x) = c j for
some j, then c j is periodic, and we are done. Thus we may assume that f k(x)
is contained in the interior of a lap of f for each k. For j = 0, . . . , p − 1,
let J j be the smallest closed interval containing { f k(x): k = j mod p}. Since
the itinerary is periodic of period p, each Ji is contained in a single lap,
so f : J j → J j+1 is strictly monotone. It follows that f p : J0 → J0 is strictly
monotone.
Suppose f p : J0 → J0 is increasing. If f p (x) ≥ x, then by induction,
f kp (x) ≥ f (k−1) p (x) for all k > 0, so { f kp (x)} converges to a point y ∈ J0 ,
which is fixed for f p . A similar argument holds if f p (x) < x.
If f p : J0 → J0 is decreasing, then f 2 p : J0 → J0 is increasing, and by the
argument in the preceding paragraph, the sequence { f 2kp (x)} converges to
a fixed point of f 2 p .
Conversely, suppose that f kq (x) → y as k → ∞, where f q (y) = y. If the
orbit of y does not contain any turning points, then eventually x has the same
172 7. Low-Dimensional Dynamics

itinerary as y. The case where O(y) does contain a turning point is left as an
exercise (Exercise 7.4.1). 

Let  be a function defined on {I1 , . . . , Il , c1 , . . . , cl } such that (I1 ) =
±1, (Ik) = (−1)k+1 (I1 ), and (ck) = 1 for k = 0, . . . , l. Associated to  is a
signed lexicographic ordering ≺ on , defined as follows. For s ∈ , define

τn (s) = (sk).
0≤k<n

We order the symbols {±I j , ±ck} by

−Il+1 < −cl < −Il < · · · < −I1 < I1 < c1 < I2 < · · · < cl < Il+1 .

Given s = (si ), t = (ti ) ∈ , we say s ≺ t if and only if s0 < t0 , or there is

n > 0 such that si = ti for i = 0, . . . , n − 1, and τn (s)sn < τn (t)tn . The proof
that ≺ is an ordering is left as an exercise.
Associated to an l-modal map f is a natural signed lexicographic order-
ing with (Ik) = 1 if f is increasing on Ik and (Ik) = −1 otherwise, and
(ck) = 1, for k = 1, . . . , l. For x ∈ I, we define τn (x) = τn (i(x)). Note that
if {x, f (x), . . . , f n−1 (x)} contains no turning points, then τn (x) is the orien-
tation of f n at x: positive (i.e., increasing) if and only if τn (x) = 1.

LEMMA 7.4.2. For x, y ∈ I, if x < y, then i(x)  i(y). Conversely, if i(x) ≺

i(y), then x < y.
Proof. Suppose i(x)
= i (y), i k(x) = i k(y) for k = 0, . . . , n − 1, and i n (x)
=
i n (y). Then there is no turning point in the intervals [x, y], f ([x, y]), . . . ,
f n−1 ([x, y]), so f n is monotone on [x, y], and is increasing if and only if
τn (i(x)) = 1. Thus x < y if and only if τn (x) f n (x) < τn (y) f n (y), and the
latter holds if and only if τn (x)i n (x) < τn (y)i n (y) since i n (x)
= i n (y). 

LEMMA 7.4.3. Let I(x) = {y: i(y) = i(x)}. Then:
1. I(x) is an interval (which may consist of a single point).
2. If I(x)
=x},{ then f n (I(x)) does not contain any turning points for
n ≥ 0. In particular, every power of f is strictly monotone on I(x).
3. Either the intervals I(x), f (I(x)), f 2 (I(x)), . . . are pairwise disjoint, or
the iterates of every point in I(x) converge to a periodic orbit of f .
Proof. Lemma 7.4.2 implies immediately that I(x) is an interval. To prove
part 2, suppose there is y ∈ I(x) such that f n (y) is a turning point. If I(x) is
not a single point, then there is some point z ∈ I(x) such that f n (y)
= f n (z),
since f is not constant on any interval. Thus i n (z)
=
n
i n (y) = f (y), which
contradicts the fact that y, z ∈ I(x). Thus I(x) must be a single point.
7.4. Combinatorial Theory of Piecewise-Monotone Mappings 173

To prove part 3, suppose the intervals I(x), f (I(x)), f 2 (I(x)), . . . are not
pairwise disjoint. Then there are integers n ≥ 0, p > 0, such that f n (I(x)) ∩
f n+ p (I(x))
= ∅. Thenf n+kp (I(x)) ∩ f n+(k+1) p (I(x))
=for
∅ all k ≥ 1. It fol-

lows that L = k≥1 f (I(x)) is a non-empty interval that contains no turn-
kp

ing points and is invariant by f p . Since f p is strictly monotone on L, for any

y ∈ L, the sequence { f 2kp (y)} is monotone and converges to a fixed point of
f 2 p. 

An interval J ⊂ I is wandering if the intervals J, f (J ), f 2 (J ), . . . are
pairwise disjoint, and f n (J ) does not converge to a periodic orbit of f . Recall
that if x is an attracting periodic point, then the basin of attraction BA(x) of
x is the set of all points whose ω-limit set is O(x).

COROLLARY 7.4.4. Suppose f does not have wandering intervals, attracting

periodic points, or intervals of periodic points. Then i: I → is an injection,
and therefore a bijective order-preserving map onto its image.
Proof. To prove that i is injective we need only show that I(x) = {x} for
every x ∈ I. If not, then by the proof of Lemma 7.4.3, either I(x) is wandering
or there is an interval L with non-empty interior and p > 0 such that f p is
monotone on L, f p (L) ∈ L, and the iterates of any point in L converge to a
periodic orbit of f of period 2 p. The former case is excluded by hypothesis.
In the latter case, by Exercise 7.4.2, either L contains an interval of periodic
points, or some open interval in L converges to a single periodic point,
contrary to the hypothesis. So I(x) = {x}. 

Our next goal is to characterize the subset i(I) ⊂ . As we indicated
above, the map i: I → is not continuous. Nevertheless, for any x ∈ I and
k ∈ N0 , there is δ > 0 such that i k(y) is constant on (x, x + δ) and on (x − δ, x)
(but not necessarily the same on both intervals). Thus the limits i(x + ) =
lim y→x+ i(y) and i(x − ) = lim y→x− i(y) exist. Moreover, i(x + ) and i(x − ) are
both contained in {I1 , . . . , Il }N0 ⊂ . For j = 1, . . . , l, we define the jth
kneading invariant of f to be ν j = i(c+j ). For convenience we also define
sequences ν0 = i(c0 ) = i(c0+ ) and νl+1 = i(cl+1 ) = i(cl+1 −
). Note that ν0 and
νl+1 are eventually periodic of period 1 or 2, since by hypothesis the set
{c0 , cl+1 } is invariant. In fact, there are only four possibilities for the pair
ν0 , νl+1 , corresponding to the four possibilities for f |∂ I .

LEMMA 7.4.5. For any x ∈ I, i(x) satisfies the following:

1. σ n i(x) = i(ck) if f n (x) = ck.
2. σ νk  σ n+1 i(x)  σ νk+1 if f n (x) ∈ Ik+1 and f is increasing on Ik+1 .
3. σ νk  σ n+1 i(x)  σ νk+1 if f n (x) ∈ Ik+1 and f is decreasing on Ik+1 .
174 7. Low-Dimensional Dynamics

Moreover, if f has no wandering intervals, attracting periodic points, or inter-

vals of periodic points, then the inequalities in conditions 2 and 3 are strict.
Proof. The first assertion is obvious. To prove the second, suppose that
f n (x) ∈ Ik+1 , and f is increasing on Ik+1 . Then for y ∈ (ck, f n (x)), we have
f (ck) < f (y) < f n+1 (x), so
i( f (ck))  i( f (y))  i( f n+1 (x)) = σ n+1 i(x).
Since νk = lim y→ck+ i(y), we conclude that σ νk  σ n+1 i(x). The other inequal-
ities are proved in a similar way.
If f has no wandering intervals, attracting periodic points, or intervals of
periodic points, then Corollary 7.4.4 implies that i is injective, so  can be
replaced by ≺ everywhere in the preceding paragraph. 

The following immediate corollary of Lemma 7.4.5 gives an admissibility
criterion for kneading invariants.
COROLLARY 7.4.6. If σ n (ν j ) = (Ik+1 , . . .), then
1. σ νk  σ n+1 ν j  σ νk+1 if f is increasing on Ik+1 ,
2. σ νk  σ n+1 ν j  σ νk+1 if f is decreasing on Ik+1 .
Let f : I → I be an l-modal map with kneading invariants ν1 , . . . , νl , and
let ν0 , νl+1 be the itineraries of the endpoints of I. Define f to be the set of
all sequences t = (tn ) ∈ satisfying the following:
1. σ n t = i(ck) if tn = ck, k ∈ {0, . . . , l}.
2. σ νk ≺ σ n+1 t ≺ σ νk+1 if tn = Ik+1 and (Ik+1 ) = +1.
3. σ νk  σ n+1 t  σ νk+1 if tn = Ik+1 and (Ik+1 ) = −1.
Similarly, we define ˆ f to be the set of sequences in satisfying conditions
1–3 with ≺ replaced by .
THEOREM 7.4.7. Let f : I → I be an l-modal map with kneading invariants
ν1 , . . . , νl , and let ν0 , νl+1 be the itineraries of the endpoints. Then i(I) ⊂
ˆ f . Moreover, if f has no wandering intervals, attracting periodic points,
or intervals of periodic points, then i(I) = f , and i: I → f is an order-
preserving bijection.
Proof. Lemma 7.4.5 implies that i(I) ⊂ ˆ f , and i(I) ⊂ f if there are no
wandering intervals, attracting periodic points, or intervals of periodic points.
Suppose f has no wandering intervals, attracting periodic points or inter-
vals of periodic points. Let t = (tn ) ∈ f , and suppose t ∈ / i( f ). Then
Lt = {x ∈ I: i(x) ≺ t}, Rt = {x ∈ I: i(x)  t}

are disjoint intervals, and I = Lt ∪ Rt .

7.4. Combinatorial Theory of Piecewise-Monotone Mappings 175

We claim that Lt and Rt are non-empty. The proof of this claim breaks into
four cases according to the four possibilities for f |∂ I . We prove it in the case
f (c0 ) = f (cl+1 ) = c0 . Then ν0 = i(c0+ ) = (I1 , I1 , . . .), νl+1 = i(cl+1
−
) = (Il+1 ,
I1 , I1 , . . .), (I1 ) = 1, and (Il+1 ) = −1. Note that t
= i (c0 ) = ν0 and t
=
i(cl+1 ) = νl+1 , since t ∈ / i( f ). Thus ν0 ≺ t, so c0 ∈ Lt . If t0 < Il+1 , then t ≺
νl+1 , so cl+1 ∈ Rt , and we are done. So suppose t0 = Il+1 . If t1 > I1 , then
t ≺ νl+1 , and again we are done. If t1 = I1 , then condition 2 implies that
σ ν0 ≺ σ 2 t, which implies in turn that t ≺ νl+1 . Thus νl+1 ∈ Rt .
Let a = sup Lt . We will show that a ∈ / Lt . Suppose for a contradiction that
a ∈ Lt . Since x ∈ / Lt for all x > a, we conclude that i(a) ≺ t  i(a + ). This
implies that the orbit of a contains a turning point. Let n ≥ 0 be the smallest
integer such that i n (a) = ck for some k ∈ {1, . . . , l}. Then i j (a) = t j = i j (a + )
for j = 1, . . . , n − 1, and i n (a + ) = Ik or i n (a + ) = Ik+1 . Suppose the latter
holds. Then f n is increasing on a neighborhood of a. Since i(a) ≺ t  i(a + )
and i j (a) = t j = i j (a + ) for j = 0, . . . , n − 1, it follows that

i(ck) = σ n (i(a)) ≺ σ n (t) ≺ σ n (i(a + )) = νk,

and ck ≤ tn ≤ Ik+1 .
If tn = ck, then by condition 1, σ n (t) = i(ck), so t = i(a), contradicting
the fact that t ∈/ i( f ). Thus we may assume that tn = Ik+1 . If f is increasing
on Ik+1 , then condition 2 implies that σ n+1 (t)  σ νk. But σ n (t)  σ n (i(a + )),
τn (t) = +1 and tn = i n (a + ) imply that

σ n+1 (t)  σ n+1 (i(a + )) = σ (νk).

Similarly, if f is decreasing on Ik+1 , then condition 3 implies that σ n+1 ≺ σ νk,

which contradicts σ n (t)  σ n (i(a + )), τn+1 (t) = −1, and tn = i n (a + ).
We have shown that the case i n (a + ) = Ik+1 leads to a contradiction. Sim-
ilarly, the case i n (a + ) = Ik leads to a contradiction. Thus a ∈ / Lt . By similar
arguments, inf Rt ∈ / Rt , which contradicts the fact that I is the disjoint union
of Lt and Rt . Thus t ∈ i(I), so i(I) = f .
Lemma 7.4.2 now implies that i: I → f is an order-preserving bijection.


COROLLARY 7.4.8. Let f and g be l-modal maps of I with no wandering
intervals, no attracting periodic points, and no intervals of periodic points. If
f and g have the same kneading invariants and endpoint itineraries, then f
and g are topologically conjugate.
Proof. Let i f and i g be the itinerary maps of f and g, respectively. Then
i −1
f ◦ i g : I → (ν0 , ν1 , . . . , νl+1 ) → I is an order-preserving bijection, and
therefore a homeomorphism, which conjugates f and g. 

176 7. Low-Dimensional Dynamics

REMARK 7.4.9. One can show that the following extension of Corollary 7.4.8
is also true: Let f and g be l-modal maps of I, and suppose f has no wandering
intervals, no attracting periodic points, and no intervals of periodic points. If
f and g have the same kneading invariants and endpoint itineraries, then f
and g are topologically semiconjugate.

Example. Consider the unimodal quadratic map f : [−1, 1] → [−1, 1],

f (x) = −2x 2 + 1. This map is conjugate to the quadratic map q4 : [0, 1] →
[0, 1], q4 (x) = 4x(1 − x), via the homeomorphism h: [−1, 1] → [0, 1], h(x) =
1
2
(x + 1). The orbit of the turning point c = 0 of f is 0, 1, −1, −1, . . . , so the
kneading invariant is ν = (I2 , I2 , I1 , I1 , . . .).
Now let I = [−1, 1], and consider the tent map T: I → I defined by

2x + 1, x ≤ 0,
T(x) =
−2x + 1, x > 0.

The homeomorphism φ: I → I, φ(x) = (2/π) sin−1 (x) conjugates f to T.

For any n > 0, the map f n+1 maps each of the intervals [k/2n , (k + 1)/2n ],
k = −2n , . . . , 2n , homeomorphically onto I. Thus the forward iterates of
any open set cover I, or equivalently, the backward orbit of any point in
I is dense in I. It follows from the next lemma that T has no wandering
intervals, attracting periodic points, or intervals of periodic points, so any
unimodal map with the same kneading invariants as T is semiconjugate to
T. In particular, any unimodal map g: [a, b] → [a, b] with g(a) = g(b) = a
and g(c) = b is semiconjugate to T.

LEMMA 7.4.10. Let I = [a, b] be an interval, and f : I → I a continuous map

with f (∂ I) ⊂ ∂ I. Suppose that every backward orbit is dense in I, and that f
has a fixed point x0 not in ∂ I. Then f has no wandering intervals, no intervals
of periodic points, and no attracting periodic points.
 −n
Proof. Let U ∈ I be an open interval. Fix x ∈ U. By density of f (x),
there is n > 0 such that f −n (x) ∩ U
=. Then
∅ f n (U) ∩ U
=, so
∅ U is not a
wandering interval.
Suppose z ∈ I is an attracting periodic point. Then the basin of attraction
BA(z) is a forward-invariant set with non-empty interior. Since backward
orbits are dense, BA(z) is a dense open subset of I and therefore intersects
the backward orbit of x0 . Thus z = x0 . On the other hand, the backward orbits
of a and b are dense, and therefore intersect BA(z), which is a contradiction.
Thus there can be no attracting periodic point.
7.4. Combinatorial Theory of Piecewise-Monotone Mappings 177

Any point in Per( f ) has finitely many preimages in Per( f ), so if Per( f )

had non-empty interior, the backward orbit of a point in Per( f ) would not
be dense in Per( f ). Thus f has no intervals of periodic points. 

The final result of this section is a realization theorem, which asserts that
any “admissible” set of sequences in is the set of kneading invariants of
an l-modal map.
Note that for an l-modal map f , the endpoint itineraries are determined
completely by the orientation of f on the first and last laps of f . Thus, given
l and a function  as in the definition of signed lexicographic orderings, we
can define natural endpoint itineraries ν0 and νl+1 as sequences in the symbol
space {I1 , Il+1 }.

THEOREM 7.4.11. Let ν1 , . . . , νl ∈ {I1 , . . . , Il+1 }N0 , and (I j ) = 0 (−1) j ,

where 0 = ±1. Let ≺ be the signed lexicographic ordering on = {I1 , . . . ,
Il+1 , c1 , . . . , cl }N0 associated to . Let ν0 , νl+1 be the endpoint itineraries deter-
mined uniquely by  and l. If {ν0 , . . . , νl+1 } satisfies the admissibility criterion
of Corollary 7.4.6, then there is a continuous l-modal map f : [0, 1] → [0, 1]
with kneading invariants ν1 , . . . , νl+1 .
Proof. Define an equivalence relation ∼ on by the rule t ∼ s if and
only if t = s, or σ (t) = σ (s) and t0 = Ik, s0 = Ik±1 . To paraphrase: t and s are
equivalent if and only if they differ at most in the first position, and then only
if the first positions are adjacent intervals. (Thus, for example, i(ck− ) ∼ i(ck+ )
for a turning point of an l-modal map.)
We will define a sequence of l-modal maps f N , N ∈ N0 , whose kneading
invariants agree up to order N with ν1 , . . . , νl . The desired map f will be the
limit in the C 0 topology of these maps.
Let p0j = c j , j = 0, . . . , l + 1. Choose points p1j ∈ [0, 1], j = 0, . . . , l + 1,
such that
1. if σ m(νi ) ∼ σ n (ν j ) then pim = pnj ;
2. pim < pnj if and only if σ m(νi ) ≺ σ n (ν j ) and σ m(νi )
∼ σ n (ν j ); and
3. the new points are equidistributed in each of the intervals [ p0j , p0j+1 ],
j = 0, . . . , l + 1.
Define f1 : [0, 1] → [0, 1] to be the piecewise-linear map specified by f ( p0j ) =
p1j . Note that p1j < p1j+1 if and only if σ ν j < σ ν j+1 , which happens if and only
if (I j+1 ) = +1. Thus f1 is l-modal.
For N > 0 we define inductively points pN j ∈ [0, 1], j = 0, . . . , l + 1, sat-
isfying conditions 1 and 2 for all n, m ≤ N and j = 0, . . . , l + 1, and so that
in any subinterval defined by the points { pnj : 0 < n < N, 0 ≤ j ≤ l + 1}, the
178 7. Low-Dimensional Dynamics

new points { pN j } in that interval are equidistributed. Then we define the map
f N : I → I to be the piecewise-linear map connecting the points ( pnj , pn+1 j ),
j = 0, . . . , l + 1, n = 0, . . . , N − 1. It follows (Exercise 7.4.5) that:
1. f N is l-modal for each N > 0;
2. { f N } converges in the C 0 topology to an l-modal map f with turning
points c1 , . . . , cl ; and
3. the kneading invariants of f are ν1 , . . . , νl . 

Exercise 7.4.1. Finish the proof of Lemma 7.4.1.

Exercise 7.4.2. Let Lbe an interval and f : L → La strictly monotone map.

Show that either L contains an interval of periodic points, or some open
interval in L converges to a single periodic point.

Exercise 7.4.3. Work out the ordering on the set of itineraries of the
quadratic map qµ for 2 < µ < 3.

Exercise 7.4.4. Show that the tent map has exactly 2n periodic points of
period n, and the set of periodic points is dense in [−1, 1].

Exercise 7.4.5. Verify the last three assertions in the proof of Theorem
7.4.11.

7.5 The Schwarzian Derivative

Let f be a C 3 function defined on an interval I ⊂ R. If f  (x)
=
0, we define
the Schwarzian derivative of f at x to be

f  (x) 3 f  (x) 2
Sf (x) =  − .
f (x) 2 f  (x)
If x is an isolated critical point of f , we define Sf (x) = lim y→x Sf (y) if the
limit exists.
For the quadratic map qµ (x) = µx(1 − x), we have that
Sqµ (x) = −6/(1 − 2x)2 for x
= 1/2, and Sf (1/2) = −∞. We also have
S exp(x) = −1/2 and S log(x) = 1/2x 2 .

LEMMA 7.5.1. The Schwarzian derivative has the following properties:

1. S( f ◦ g) = (Sf ◦ g)(g  )2 + Sg.
n−1
2. S( f n ) = i=0 Sf ( f i (x)) · (( f i ) (x))2 .
3. If Sf < 0, then S( f n ) < 0 for all n > 0.

The proof is left as an exercise (Exercise 7.5.3).

7.5. The Schwarzian Derivative 179

A function with negative Schwarzian derivative satisfies the following

minimum principle.

LEMMA 7.5.2 (Minimum Principle). Let I be an interval and f : I → I a C 3

map with f  (x)
=
0 for all x ∈ I. If Sf < 0, then | f  (x)| does not attain a local
minimum in the interior of I.
Proof. Let z be a critical point of f  . Then f  (z) = 0, which implies that
f  (z)/ f  (z) < 0, since Sf < 0. Thus f  (z) and f  (z) have opposite signs.
If f  (z) < 0, then f  (z) > 0 and z is a local minimum of f  , so z is a local
maximum of | f  |. Similarly, if f  (z) > 0, then z is also a local maximum of
| f  |. Since f  is never zero on I, this implies that | f  | does not have a local
minimum on I. 

THEOREM 7.5.3 (Singer). Let I be a closed interval (possibly unbounded),
and f : I → I a C 3 map with negative Schwarzian derivative. If f has n critical
points, then f has at most n + 2 attracting periodic orbits.
Proof. Let z be an attracting periodic point of period m. Let W(z) be the
maximal interval about z such that f mn (y) → z as n → ∞ for all y ∈ U. Then
W(z) is open (in I), and f m(W(z)) ⊂ W(z).
Suppose that W(z) is bounded and does not contain a point in ∂ I, so
W(z) = (a, b) for some a < b ∈ R. We claim that f m has a critical point in
W(z). By maximality of W(z), f m must preserve the set of endpoints of W(z).
If f m(a) = f m(b), then f m must have a maximum or minimum in W(z), and
therefore a critical point in W(z). If f m(a)
= f m(b), then f m must permute
a and b. Suppose f (a) = a and f (b) = b. Then ( f m) ≥ 1 on ∂U, since
m m

otherwise a or b would be an attracting fixed point for f m whose basin

of attraction overlaps U. By the minimum principle, if f m has no critical
points in U, then ( f m) > 1 on U, which contradicts f m(W(z)) = W(z), so
f m has a critical point in W(z). If f m(a) = b and f m(b) = a, then applying the
preceding argument to f 2m, we conclude that f 2m has a critical point in W(z).
Since f m(W(z)) = W(z), it follows that f m also has a critical point in W(z).
By the chain rule, if p ∈ W(z) is a critical point of f m, then one of the
points p, f ( p), . . . , f m−1 ( p) is a critical point of f . Thus we have shown that
either W(z) is unbounded, or it meets ∂ I, or there is a critical point of f
whose orbit meets W(z). Since there are only n critical points, and there are
only two boundary points (or unbounded ends) of I, the theorem is proved.


COROLLARY 7.5.4. For any µ > 4, the quadratic map qµ : R → R has at
most one (finite) attracting periodic orbit.
180 7. Low-Dimensional Dynamics

Proof. The proof of Theorem 7.5.3 shows that if z is an attracting periodic

point, then W(z) either is unbounded or contains the critical point of qµ .
Since ∞ is an attracting periodic point, the basin of attraction of z must be
bounded, and therefore must contain the critical point. 

We now discuss a relation between the Schwarzian derivative and length
distortion that is used in producing absolutely continuous invariant measures
for maps of the interval with negative Schwarzian derivative.2
Let f be a piecewise-monotone real-valued function defined on a bounded
interval I. Suppose J ⊂ I is a subinterval such that I \ J consists of disjoint
non-empty intervals L and R. Denote by |F| the length of an interval F.
Define the cross-ratios
|I| · |J | |I| · |J |
C(I, J ) = , D(I, J ) = .
|J ∪ L| · |J ∪ R| |L| · |R|
If f is monotone on I, set
C( f (I), f (J )) D( f (I), f (J ))
A(I, J ) = , B(I, J ) = .
C(I, J ) D(I, J )
The group M of real Möbius transformations consists of maps of the
extended real line R ∪ {∞} of the form φ(x) = (ax + b)/(cx + d), where
a, b, c, d ∈ R and ad − bc
= 0. Möbius transformations have Schwarzian
derivative equal to 0 and preserve the cross-ratios C and D (Exercise 7.5.4).
The group of Möbius transformations is simply transitive on triples of points
in the extended real line, i.e., given any three distinct points a, b, c ∈ R ∪ {∞},
there is a unique Möbius transformation φ ∈ M such that φ(0) = a, φ(1) = b
and φ(∞) = c (Exercise 7.5.5). Möbius transformations are also called linear
fractional transformations.

PROPOSITION 7.5.5. Let f be a C 3 real-valued function defined on a com-

pact interval I such that f has negative Schwarzian derivative and f  (x)
=
0, x ∈ I. Let J ⊂ I be a closed subinterval that does not contain the endpoints
of I. Then A(I, J ) > 1 and B(I, J ) > 1.
Proof. Since every Möbius transformation has Schwarzian derivative 0 and
preserves C and D, we may assume, by composing f on the left and on the
right with appropriate Möbius transformations and using Lemma 7.5.1, that
I = [0, 1], J = [a, b] with 0 < a < b < 1, f (0) = 0, f (a) = a, and f (1) = 1.
By Lemma 7.5.2, | f  | does not have a local minimum in [0, 1], and hence f
cannot have fixed points except 0, a, and 1. Therefore f (x) < x if 0 < x < a

2 Our exposition here follows to a large extent [vS88] and [dMvS93]

7.6. Real Quadratic Maps 181

and f (x) > x if a < x < 1; in particular, f (b) > b. We have


| f (1) − f (0)| · | f (b) − f (a)| |1 − 0| · |b − a| −1
B(I, J ) = ·
| f (a) − f (0)| · | f (1) − f (b)| |a − 0| · |1 − b|
1 · ( f (b) − a) · a · (1 − b)
= > 1.
a · (1 − f (b)) · 1 · (b − a)
This proves the second inequality. The first one is left as an exercise
(Exercise 7.5.6). 

The following proposition, which we do not prove, describes bounded dis-
tortion properties of maps with negative Schwarzian derivative on intervals
without critical points.
PROPOSITION 7.5.6 [vS88], [dMvS93]. Let f : [a, b] → R be a C 3 map. As-
sume that Sf < 0 and f  (x)
= 0, for all x ∈ [a, b]. Then
 
1. | f (a)| · | f (b)| ≥ (| f (b) − f (a)|/(b − a))2 ;
| f  (x)| · | f (b) − f (a)| | f (x) − f (a)| | f (b) − f (x)|
2. ≥ · for every x ∈
b−a x−a b− x
(a, b).
Exercise 7.5.1. Prove that if f : I → R is a C 3 diffeomorphism onto its im-
age and g(x) = dx
d
log | f  (x)|, then

1 d2 1
Sf (x) = g  (x) − (g(x))2 = −2 | f  (x)| · 2  .
2 dx | f  (x)|

Exercise 7.5.2. Show that any polynomial with distinct real roots has neg-
ative Schwarzian derivative.
Exercise 7.5.3. Prove Lemma 7.5.1.
Exercise 7.5.4. Prove that each Möbius transformation has Schwarzian
derivative 0 and preserves the cross-ratios C and D.
Exercise 7.5.5. Prove that the action of the group of Möbius transforma-
tions on the extended real line is simply transitive on triples of points.
Exercise 7.5.6. Prove the remaining inequality of Proposition 7.5.5.

7.6 Real Quadratic Maps

In §1.5, we introduced the one-parameter family of real quadratic maps
qµ (x) = µx(1 − x), µ ∈ R. We showed that for µ > 1, the orbit of any point
182 7. Low-Dimensional Dynamics

1.2

0.8

0.6

0.4

0.2

0 I0 a b I1 1

Figure 7.5. Quadratic map.

outside I = [0, 1] converges monotonically to −∞. Thus the interesting dy-

namics is concentrated on the set
 
µ = x ∈ I | qµn (x) ∈ I ∀n ≥ 0 .

THEOREM 7.6.1. Let µ > 4. Then µ is a Cantor set, i.e., a perfect, nowhere
dense subset of [0, 1]. The restriction qµ |µ is topologically conjugate to the
one-sided shift σ : 2+ → 2+ .
 
Proof. Let a = 1/2 − 1/4 − 1/µ and b = 1/2 + 1/4 − 1/µ be the two
solutions of qµ (x) = 1, and let I0 = [0, a], I1 = [b, 1]. Then qµ (I0 ) = qµ (I1 ) =
I, and qµ ((a, b)) ∩ I = ∅ (see Figure 7.5). Observe that the images qµn (1/2)
of the critical point 1/2 lie outside I and tend to −∞. Therefore the two
inverse branches f0 : I → I0 and f1 : I → I1 and their compositions are well
defined. For k ∈ N, denote by Wk the set of all words of length k in the
alphabet {0, 1}. For w = ω1 ω2 . . . ωk ∈ Wk and j ∈ {0, 1}, set Iw j = f j (Iw ) and
gw = fωk ◦ · · · ◦ fω2 ◦ fω1 , so that Iw = gw (I).

LEMMA 7.6.2. lim max max |gw (x)| = 0.

k→∞ w∈Wk x∈I
√ 
Proof. If µ > 2 + 5, then 1 > | f j (1)| = µ 1 − 4/µ ≥ | f j (x)| for every
x ∈ I, j = 0, 1, and the
√ lemma follows.
For 4 < µ < 2 + 5, the lemma follows from Theorem 8.5.10 (see also
Theorem 8.5.11). 

Lemma 7.6.2 implies that the length of the interval Iw tends to 0 as
the length of w tends to infinity. Therefore, for each ω = ω1 ω2 . . . ∈ 2+
7.7. Bifurcations of Periodic Points 183


the intersection n∈N Iω1 ...ωn consists of exactly one point h(ω). The
map h: 2+ → µ is a homeomorphism conjugating the shift σ and qµ |µ
(Exercise 7.6.2). 


Exercise 7.6.1. Prove that if µ > 4 and 1/2 − 1/4 − 1/µ < x < 1/2 +

1/4 − 1/µ, then qµn (x) → −∞ as n → ∞.

Exercise 7.6.2. Prove that the map h: 2+ → µ in the proof of Theo-

rem 7.6.1 is a homeomorphism and that qµ ◦ h = h ◦ σ .

7.7 Bifurcations of Periodic Points3

The family of real quadratic maps qµ (x) = µx(1 − x) (§1.5, §7.6) is an ex-
ample of a (one-dimensional) parametrized family of dynamical systems.
Although the specific quantitative behavior of a dynamical system depends
on the parameter, it is often the case that the qualitative behavior remains
unchanged for certain ranges of the parameter. A parameter value where the
qualitative behavior changes is called a bifurcation value of the parameter.
For example, in the family of quadratic maps, the parameter value µ = 3 is
a bifurcation value because the stability of the fixed point 1 − 1/µ changes
from repelling to attracting. The parameter value µ = 1 is a bifurcation value
because for µ < 1, 0 is the only fixed point, and for µ > 1, qµ has two fixed
points.
A bifurcation is called generic if the same bifurcation occurs for all nearby
families of dynamical systems, where “nearby” is defined with respect to an
appropriate topology (usually the C 2 or C 3 topology). For example, the
bifurcation value µ = 3 is generic for the family of quadratic maps. To see
this, note that for µ close to the 3; the graph of qµ crosses the diagonal
transversely at the fixed point xµ = 1 − 1/µ, and the magnitude of qµ (xµ ) is
less than 1 for µ < 3 and greater than 1 for µ > 3. If fµ is another family of
maps C 1 -close to qµ , then the graph of fµ (x) must also cross the diagonal
at a point yµ near xµ , and the magnitude of fµ (yµ ) must cross 1 at some
parameter value close to 3. Thus fµ has the same kind of bifurcation as qµ .
Similar reasoning shows that the bifurcation value µ = 1 is also generic.
Generic bifurcations are the primary ones of interest. The notion of gener-
icity depends on the dimension of the parameter space (e.g., a bifurcation
may be generic for a one-parameter family, but not for a two-parameter

3 The exposition in this section follows to a certain extent that of [Rob95].

184 7. Low-Dimensional Dynamics

family). Bifurcations that are generic for one-parameter families of dy-

namical systems are called codimension-one bifurcations. In this section, we
describe codimension-one bifurcations of fixed and periodic points for one-
dimensional maps.
We begin with a non-bifurcation result. If the graph of a differentiable
map f intersects the diagonal transversely at a point x0 , then the fixed point
x0 persists under a small C 1 perturbation of f .

PROPOSITION 7.7.1. Let U ⊂ Rm and V ⊂ Rn be open subsets, and let

fµ : U → Rm, µ ∈ V, be a family of C 1 maps such that
1. the map (x, µ) "→ fµ (x) is a C 1 map,
2. fµ0 (x0 ) = x0 for some x0 ∈ U and µ0 ∈ V,
3. 1 is not an eigenvalue of d fµ0 (x0 ).
Then there are open sets U  ⊂ U, V  ⊂ V with x0 ∈ U  , µ0 ∈ V  and a C 1 func-
tion ξ : V  → U  such that for each µ ∈ V  , ξ (µ) is the only fixed point of fµ
in U  .
Proof. The proposition is an immediate consequence of the implicit func-
tion theorem applied to the map (x, µ) "→ fµ (x) − x (Exercise 7.7.1). 

Proposition 7.7.1 shows that if 1 is not an eigenvalue of the derivative,
then the fixed point does not bifurcate into multiple fixed points and does
not disappear. The next proposition shows that periodic points cannot appear
in a neighborhood of a hyperbolic fixed point.

PROPOSITION 7.7.2. Under the assumption (and notation) of Proposition

7.7.1, suppose in addition that x0 is a hyperbolic fixed point of fµ0 , i.e., no
eigenvalue of d fµ0 (x0 ) has absolute value 1. Then for each k ∈ N there are
neighborhoods Uk ⊂ U  of x0 and Vk ⊂ V  of µ0 such that ξ (µ) is the only
fixed point of fµk in Uk.
If, in addition, x0 is an attracting fixed point of fµ0 , i.e., all eigenvalues of
d fµ0 (x0 ) are strictly less than 1 in absolute value, then the neighborhoods Uk
and Vk can be chosen independent of k.
Proof. Since no eigenvalue of d fµ0 (x0 ) has absolute value 1, it follows
that 1 is not an eigenvalue of d fµk0 (x0 ), so the first statement follows from
Proposition 7.7.1.
The second statement is left as an exercise (Exercise 7.7.2). 

Propositions 7.7.1 and 7.7.2 show that, for differentiable one-dimensional
maps, bifurcations of fixed or periodic points can occur only if the abso-
lute value of the derivative is 1. For one-dimensional maps there are only
7.7. Bifurcations of Periodic Points 185

two types of generic bifurcations: The saddle–node bifurcation (or the fold
bifurcation) may occur if the derivative at a periodic point is 1, and the
period-doubling bifurcation (or flip bifurcation) may occur if the derivative
at a periodic point is −1. We describe these bifurcations in the next two
propositions. See [CH82] or [HK91] for a more extensive discussion of bi-
furcation theory, or [GG73] for a thorough exposition on the closely related
topic of singularities of differentiable maps.

PROPOSITION 7.7.3 (Saddle–Node Bifurcation). Let I, J ⊂ R be open in-

tervals and f : I × J → R be a C 2 map such that
∂f
1. f (x0 , µ0 ) = x0 and ∂ x (x0 , µ0 ) = 1 for some x0 ∈ I and µ0 ∈ J ,
∂2 f ∂f
2. ∂ x2 (x0 , µ0 ) < 0 and ∂µ (x0 , µ0 ) > 0.
Then there are , δ > 0 and a C 2 function α: (x0 − , x0 + ) → (µ0 − δ, µ0 +
δ) such that:
1. α(x0 ) = µ0 , α  (x0 ) = 0, α  (x0 ) = − ∂∂ x2f (x0 , µ0 )/ ∂µ
∂f
2
(x0 , µ0 ) > 0.
2. Each x ∈ (x0 − , x0 + ) is a fixed point of f (·, α(x)), i.e., f (x, α(x)) =
x, and α −1 (µ) is exactly the fixed point set of f (·, µ) in (x0 − , x0 + )
for µ ∈ (µ0 − δ, µ0 + δ).
3. For each µ ∈ (µ0 , µ0 + δ), there are exactly two fixed points x1 (µ) <
x2 (µ) of f (·, µ) in (x0 − , x0 + ) with
∂f ∂f
(x1 (µ), µ) > 1 and 0< (x2 (µ), µ) < 1;
∂x ∂x
α(xi (µ)) = µ for i = 1, 2.
4. f (·, µ) does not have fixed points in (x0 − , x0 + ) for each
µ ∈ (µ0 − δ, µ0 ).

REMARK 7.7.4. The inequalities in the second hypothesis of Proposition 7.7.3

correspond to one of the four possible generic cases when the two derivatives
do not vanish. The other three cases are similar (Exercise 7.7.3).
Proof. Consider the function g(x, µ) = f (x, µ) − x (see Figure 7.6). Ob-
serve that
∂g ∂f
(x0 , µ0 ) = (x0 , µ0 ) > 0.
∂µ ∂µ

Therefore, by the implicit function theorem, there are , δ > 0 and a C 2

function α: (x0 − , x0 + ) → J such that g(x, α(x)) = 0 for each x ∈
(x0 − , x0 + ) and there are no other zeros of g in (x0 − , x0 + ) × (µ0 −
, µ0 + ). A direct calculation shows that α satisfies statement 1. Since
186 7. Low-Dimensional Dynamics

y=x
y
µ = µ0

µ > µ0
µ < µ0

fµ

x1 (µ) x0 x2 (µ) x

Figure 7.6. Saddle-node bifurcation.

α  (x0 ) > 0, statements 3 and 4 are satisfied for  and δ sufficiently small
(Exercise 7.7.4). 

PROPOSITION 7.7.5 (Period-Doubling Bifurcation). Let I, J ⊂ R be open

intervals, and f : I × J → R be a C 3 map such that:
∂f
1. f (x0 , µ0 ) = x0 and ∂ x (x0 , µ0 ) = −1 for some x0 ∈ I and µ0 ∈ J , so that
by Proposition 7.7.1, there is a curve µ "→ξ (µ) of fixed points of f (·, µ)
for µ close to µ0 .
∂f
2. η = dµ d
|µ=µ0 ∂ x (ξ (µ), µ) < 0.
∂ 3 f ( f (x , µ ),µ ) ∂3 f ∂2 f
3. ζ = 0
∂ x3
0 0
= −2 ∂ x3 (x0 , µ0 ) − 3( ∂ x2 (x0 , µ0 ))2 < 0 .
Then there are , δ > 0 and C 3 functions ξ : (µ0 − δ, µ0 + δ) → R with
ξ (µ0 ) = x0 and α: (x0 − , x0 + ) → R with α(x0 ) = µ0 , α  (x0 ) = 0, and
α  (x0 ) = −2η/ζ > 0 such that:
1. f (ξ (µ), µ) = ξ (µ), and ξ (µ) is the only fixed point of f (·, µ) in (x0 −
, x0 + ) for µ ∈ (µ0 − δ, µ0 + δ).
2. ξ (µ) is an attracting fixed point of f (·, µ) for µ0 − δ < µ < µ0 and is
a repelling fixed point for µ0 < µ < µ0 + δ.
3. For each µ ∈ (µ0 , µ0 + δ), the map f (·, µ) has, in addition to the fixed
point ξ (µ), exactly two attracting period-2 points x1 (µ), x2 (µ) in the
interval (x0 − , x0 + ); moreover, α(xi (µ)) = µ and xi (µ) → x0 as
µ $ µ0 for i = 1, 2.
4. For each µ ∈ (µ0 − δ, µ0 ], the map f ( f (·, µ), µ) has exactly one fixed
point ξ (µ) in (x0 − , x0 + ).
7.7. Bifurcations of Periodic Points 187

REMARK 7.7.6. The stability of the fixed point ξ (µ) and of the periodic points
x1 (µ) and x2 (µ) depend on the signs of the derivatives in the third and fourth
hypotheses of Proposition 7.7.5. Proposition 7.7.5 deals with only one of the
four possible generic cases when the derivatives do not vanish. The other three
cases are similar, and we do not consider them here (Exercise 7.7.5).
Proof. Since
∂f
(x0 , µ0 ) = −1
=
1,
∂x
we can apply the implicit function theorem to f (x, µ) − x = 0 to obtain a
differentiable function ξ such that f (ξ (µ), µ) = ξ (µ) for µ close to µ0 and
ξ (µ0 ) = x0 . This proves statement 1.
Differentiating f (ξ (µ), µ) = ξ (µ) with respect to µ gives
d ∂f ∂f
f (ξ (µ), µ) = (ξ (µ), µ) + (ξ (µ), µ) · ξ  (µ) = ξ  (µ),
dµ ∂µ ∂x
and hence
∂f
∂µ
(ξ (µ), µ) 1 ∂f
ξ  (µ) = ∂f
, ξ  (µ0 ) = (x0 , µ0 ).
1 − ∂ x (ξ (µ), µ) 2 ∂µ
Therefore
d  ∂f ∂2 f 1 ∂f ∂2 f
 (ξ (µ), µ) = (x0 , µ0 ) + (x0 , µ0 ) · (x0 , µ0 ) = η,
dµ µ=µ0 ∂ x ∂µ ∂ x 2 ∂µ ∂ x2
and assumption 2 yields statement 2.
To prove statements 3 and 4 consider the change of variables y = x −
ξ (µ), 0 = x0 − ξ (µ0 ) and the function g(y, µ) = f ( f (y + ξ (µ), µ), µ) −
ξ (µ). Observe that fixed points of f ( f (·, µ), µ) correspond to solutions of
g(y, µ) = y. Moreover,
∂g ∂2g
g(0, µ) ≡ 0, (0, µ0 ) = 1, (0, µ0 ) = 0,
∂y ∂ y2
i.e., the graph of the second iterate of f (·, µ0 ) is tangent to the diagonal
at (x0 , µ0 ) with second derivative 0. (See Figure 7.7.) A direct calculation
shows that, by assumption 3, the third derivative does not vanish:
 2 2
∂3g ∂3 f ∂f
(0, µ0 ) = −2 3 (x0 , µ0 ) − 3 (x0 , µ0 ) = ζ < 0.
∂ y3 ∂x ∂ x2
Therefore
1 3
g(y, µ0 ) = y + ζ y + o(y3 ).
3!
Since ξ (µ) is a fixed point of f (·, µ), we have that g(0, µ) ≡ 0 in an interval
188 7. Low-Dimensional Dynamics

µ < µ0 µ = µ0 µ > µ0
Figure 7.7. Period-doubling bifurcation: the graph of the second iterate.

about µ0 . Therefore there is a differentiable function h such that g(y, µ) =

y · h(y, µ), and to find the period-2 points of f (·, µ) different from ξ (µ) we
must solve the equation h(y, µ) = 1. From (7.7.6) we obtain
1 2
h(y, µ0 ) = 1 + ζ y + o(y2 ),
3!
i.e.,
∂h ∂ 2h ζ
h(0, µ0 ) = 1, (0, µ0 ) = 0, and (0, µ0 ) = .
∂y ∂y 2 3
On the other hand,
∂h 1 ∂g ∂2g
(0, µ0 ) = lim (y, µ0 ) = (0, µ0 )
∂µ y→0 y ∂µ ∂µ ∂ y
 2 
d ∂f 
= (ξ (µ), µ)  = −2η > 0.
dµ ∂ x µ=µ0

By the implicit function theorem, there is  > 0 and a differentiable func-

tion β: (−, ) → R such that h(y, β(y)) = 1 for |y| <  and β(0) = µ0 . Dif-
ferentiating h(y, β(y)) = 1 with respect to y, we obtain that β  (0) = 0. The
second differentiation yields β  (0) = ζ /6η > 0. Therefore β(y) > 0 for
y
= 0, and the new period-2 orbit appears only for µ > µ0 .
Note that since g(·, µ) has three fixed points near x0 for µ close to µ0 ,
and the middle one, ξ (µ), is unstable, the other two must be stable. In fact,
a direct calculation shows that
∂g ∂g 1 ∂2g 1 ∂3g
(y, β(y)) = (0, µ0 ) + (0, µ0 )y + (0, µ0 )y2 + o(y2 )
∂y ∂y 2! ∂ y2 3! ∂ y3
ζ
= 1 + y2 + o(y2 ).
6
Since ζ < 0, the period-2 orbit is stable. 

Exercise 7.7.1. Prove Proposition 7.7.1.
Exercise 7.7.2. Prove the second statement of Proposition 7.7.2.
7.8. The Feigenbaum Phenomenon 189

Exercise 7.7.3. State the analog of Proposition 7.7.3 for the remaining
three generic cases when the derivatives from assumption 3 do not vanish.
Exercise 7.7.4. Prove statements 3 and 4 of Proposition 7.7.3.
Exercise 7.7.5. State the analog of Proposition 7.7.5 for the remaining three
generic cases when the derivatives from assumptions 3 and 4 do not vanish.
Exercise 7.7.6. Prove that a period-doubling bifurcation occurs for the
family fµ (x) = 1 − µx 2 at µ0 = 3/4, x0 = 2/3.

7.8 The Feigenbaum Phenomenon

M. Feigenbaum [Fei79] studied the family
fµ (x) = 1 − µx 2 , 0 < µ ≤ 2,
of unimodal maps of the interval [−1, 1]. For µ < 3/4, the unique attracting
fixed point of fµ is

1 + 4µ − 1
xµ = .
2µ

The derivative fµ (xµ ) = 1 − 1 + 4µ is greater than −1 for µ < 3/4, equals
−1 for µ = 3/4, and is less than −1 for µ > 3/4. A period-doubling bifur-
cation occurs at µ = 3/4 (Exercise 7.7.6). For µ > 3/4, the map fµ has an
attracting period-2 orbit. Numerical studies show that there is an increasing
sequence of bifurcation values µn at which an attracting periodic orbit of
period 2n for fµ loses stability and an attracting periodic orbit of period 2n+1
is born. The sequence µn converges, as n → ∞, to a limit µ∞ , and
µ∞ − µn−1
lim = δ = 4.669201609 . . . . (7.2)
n→∞ µ∞ − µn
The constant δ is called the Feigenbaum constant. Numerical experiments
show that the Feigenbaum constant appears for many other one-parameter
families.
The Feigenbaum phenomenon can be explained as follows. Consider the
infinite-dimensional space A of real analytic maps ψ: [−1, 1] → [−1, 1] with
ψ(0) = 1, and the map : A → A given by the formula
1
(ψ)(x) = ψ ◦ ψ(λx), λ = ψ(1). (7.3)
λ
A fixed point g of  (which Feigenbaum estimated numerically) is an even
function satisfying the Cvitanović–Feigenbaum equation
g ◦ g(λx) − λg(x) = 0. (7.4)
190 7. Low-Dimensional Dynamics

fµ1
B1

fµn
fµn+1 Bn

Bn+1
g W s (g)

W u (g) fµ

Figure 7.8. Fixed point and stable and unstable manifolds for the Feigenbaum
map .

The function g is a hyperbolic fixed point of . The stable manifold Ws (g)

has codimension one, and the unstable manifold Wu (g) has dimension one
and corresponds to a simple eigenvalue δ = 4.669201609 . . . of the derivative
dg . The codimension-one bifurcation set B1 , of maps ψ for which an at-
tracting fixed point loses stability and an attracting period two orbit is born,
intersects Wu (g) transversely. The preimage Bn = 1−n (B1 ) is the bifurca-
tion set of maps for which an attracting orbit of period 2n−1 is replaced by
an attracting orbit of period 2n (Exercise 7.8.1). Figure 7.8 is a graphical
depiction of the process underlying the Feigenbaum phenomenon.
By the infinite-dimensional version of the Inclination Lemma 5.7.2, the
codimension-one bifurcation sets Bn accumulate to Ws (g). Let fµ be a one-
parameter family of maps that intersects Ws (g) transversely, and let µn be
the sequence of period-doubling bifurcation parameters, fµn ∈ Bn . Using
the inclination lemma, one can show that the sequence µn satisfies (7.2).
O. E. Lanford established the correctness of this model through a computer-
assisted proof [Lan84].
Exercise 7.8.1. Prove that if ψ has an attracting periodic orbit of period 2k,
then (ψ) has an attracting periodic orbit of period k.
CHAPTER EIGHT

Complex Dynamics

In this chapter1 , we consider rational maps R(z) = P(z)/Q(z) of the Riemann

sphere C̄ = C ∪ {∞}, where P and Q are complex polynomials. These maps
exhibit many interesting dynamical properties, and lend themselves to the
computer-aided drawing of fractals and other fascinating pictures in the
complex plane. For a more thorough exposition of the dynamics of rational
maps see [Bea91] and [CG93].

8.1 Complex Analysis on the Riemann Sphere

We assume that the reader is familiar with the basic ideas of complex analysis
(see, for example, [BG91] or [Con95]).
Recall that a function f from a domain D ⊂ C to C̄ is said to be mero-
morphic if it is analytic except at a discrete set of singularities, all of which
are poles. In particular, rational functions are meromorphic.
The Riemann sphere is the one-point compactification of the complex
plane, C̄ = C ∪ {∞}. The space C̄ has the structure of a complex manifold,
given by the standard coordinate system on C and the coordinate z → z−1 on
C̄\{0}. If M and N are complex manifolds, then a map f : M → N is analytic
if for every point ζ ∈ M, there are complex coordinate neighborhoods U of
ζ and V of f (ζ ) such that f : U → V is analytic in the coordinates on U and
V. An analytic map into C̄ is said to be meromorphic. This terminology is
somewhat confusing, because in the modern sense (as maps of manifolds)
meromorphic functions are analytic, while in the classical sense (as functions
on C), meromorphic functions are generally not analytic. Nevertheless, the
terminology is so entrenched that it cannot be avoided.

1 Many of the proofs in this chapter follow the corresponding arguments from [CG93].

191
192 8. Complex Dynamics

It is easy to see that a map f : C̄ → C̄ is analytic (and meromorphic) if

and only if both f (z) and f (1/z) are meromorphic (in the classical sense)
on C. It is known that every analytic map from the Riemann sphere to itself
is a rational map. Note that the constant map f (z) = ∞ is considered to be
analytic.
The group of Möbius transformations


az + b
z→ : a, b, c, d ∈ C; ad − bc = 1
cz + d
acts on the Riemann sphere and is simply transitive on triples of points, i.e., for
any three distinct points x, y, z ∈ C̄, there is a unique Möbius transformation
that carries x, y, z to 0, 1, ∞, respectively (see §7.5).
Suppose f : C̄ → C̄ is a meromorphic map and ζ is a periodic point of
minimal period k. If ζ = ∞, the multiplier of ζ is the derivative λ(ζ ) =
( f k) (ζ ). If ζ = ∞, the multiplier of ζ is g  (0), where g(z) = 1/ f (1/z). The
periodic point ζ is attracting if 0 < |λ(ζ )| < 1, superattracting if λ(ζ ) = 0,
repelling if |λ(ζ )| > 1, rationally neutral if λ(ζ )m = 1 for some m ∈ N, and
irrationally neutral if |λ(ζ )| = 1 but λ(ζ )m =1 for every m ∈ N. One can
prove that a periodic point is attracting or superattracting if and only if it is
a topologically attracting periodic point in the sense of Chapter 1; similarly
for repelling periodic points. The orbit of an attracting or superattracting
periodic point is said to be an attracting or superattracting periodic orbit,
respectively.
For an attracting or superattracting fixed point ζ of a meromorphic map
f , we define the basin of attraction BA(ζ ) as the set of points z ∈ C̄ for
which f n (z) → ζ as n → ∞. Since the multiplier of ζ is less than 1, there is a

neighborhood U of ζ that is contained in BA(ζ ), and BA(ζ ) = n∈N f −n (U).
The set BA(ζ ) is open. The connected component of BA(ζ ) containing ζ is
called the immediate basin of attraction, and is denoted BA◦ (ζ ).
If ζ is an attracting or superattracting periodic point of period k, then the
basin of attraction of the periodic orbit is the set of all points z for which
f nk(z) → f j (ζ ) as n → ∞ for some j ∈ {0, 1, . . . , k} and is denoted BA(ζ ).
The union of the connected components of BA(ζ ) containing a point in the
orbit of ζ is called the immediate basin of attraction and is denoted BA◦ (ζ ).
A point ζ is a critical point (or branch point) of a meromorphic function
f if f is not 1-to-1 on a neighborhood of ζ . A critical point ζ has multiplicity
m if f is (m + 1)-to-1 on U\{ζ } for a sufficiently small neighborhood U of ζ
(this number is also called the branch number of f at ζ ). Equivalently, ζ is
a critical point of multiplicity m if ζ is a zero of f  (in local coordinates) of
multiplicity m. If ζ is a critical point, then f (ζ ) is called a critical value.
8.1. Complex Analysis on the Riemann Sphere 193

For a rational map R = P/Q, with P and Q relatively prime polynomials

of degree p and q, respectively, the degree of R is deg(R) = max( p, q). If R
has degree d, then the map R: C̄ → C̄ is a branched covering of degree d,
i.e., any ξ ∈ C̄ that is not a critical value has exactly d preimages; in fact, every
point has exactly d preimages if critical points are counted with multiplicity.
Since the number of preimages of a generic point is a topological invariant
of R, the degree is invariant under conjugation by a Möbius transformation.
The rational maps of degree 1 are the Möbius transformations. A rational
map is a polynomial if and only if the only preimage of ∞ is ∞.
PROPOSITION 8.1.1. Let R be a rational map of degree d. Then the number
of critical points, counted with multiplicity, is 2d − 2. If there are exactly two
distinct critical points, then R is conjugate by a Möbius transformation to zd
or z−d .
Proof. By composing with a Möbius transformation we may assume that
R(∞) = 0 and that ∞ is neither a critical point nor a critical value. Then
R(∞) = 0 and the fact that ∞ is not a critical point imply that
αzd−1 + · · ·
R(z) = ,
βzd + · · ·
where α =0 and β =0. Hence
αβz2d−2 + · · ·
R (z) = − ,
(βzd + · · ·)2
and the critical points of R are the zeros of the numerator (since ∞ is not a
critical value).
The proof of the second assertion is left as an exercise (Exercise 8.1.5).


A family F of meromorphic functions in a domain D ⊂ C̄ is normal if
every sequence from F contains a subsequence that converges uniformly on
compact subsets of D in the standard spherical metric on C̄ ≈ S 2 . A family
F is normal at a point z ∈ C̄ if it is normal in a neighborhood of z.
The Fatou set F(R) ⊂ C̄ of a rational map R: C̄ → C̄ is the set of points
z ∈ C̄ such that the family of forward iterates {R n }n∈N is normal at z. The
Julia set J (R) is the complement of the Fatou set. Both F(R) and J (R) are
completely invariant under R (see Proposition 8.5.1). Points belonging to the
same component of F(R) have the same asymptotic behavior. As we will
see later, the Fatou set contains all basins of attraction and the Julia set is the
closure of the set of all repelling periodic points. The “interesting” dynamics
is concentrated on the Julia set, which is often a fractal set. The case when
J (R) is a hyperbolic set is reasonably well understood (Theorem 8.5.10).
194 8. Complex Dynamics

Exercise 8.1.1. Prove that any Möbius transformation is conjugate by

another Möbius transformation to either z →
az or z →
z + a.
Exercise 8.1.2. Prove that a non-constant rational map R is conjugate to
a polynomial by a Möbius transformation if and only if R−1 (z0 ) = {z0 } for
some z0 ∈ C̄.
Exercise 8.1.3. Find all Möbius transformations that commute with q0 (z) =
z2 .
Exercise 8.1.4. Let R be a rational map such that R(∞) = ∞, and let f
be a Möbius transformation such that f (∞) is finite. Define the multiplier
λ R(∞) of R at ∞ to be the multiplier of f ◦ R ◦ f −1 at f (∞). Prove that
λ R(∞) does not depend on the choice of f .
Exercise 8.1.5. Prove the second assertion of Proposition 8.1.1.
Exercise 8.1.6. Let R be a non-constant rational map. Prove that
deg(R) − 1 ≤ deg(R ) ≤ 2 deg(R)
with equality on the left if and only if R is a polynomial and with equality on
the right if and only if all poles of R are simple and finite.

8.2 Examples
The global dynamics of a rational map R depends heavily on the behavior of
the critical points of R under its iterates. In most of the examples below, the
Fatou set consists of finitely many components, each of which is a basin of
attraction. Some of the assertions in the following examples will be proved
in later sections of this chapter. Proofs of most of the assertions that are not
proved here can be found in [CG93].
Let qa : C̄ → C̄ be the quadratic map qa (z) = z2 − a, and denote by S1
the unit circle {z ∈ C: |z| = 1}. The critical points of qa are 0 and ∞, and
the critical values are −a and ∞; if a =0, the only superattracting periodic
(fixed) point is ∞. In the examples below, we observe drastically different
global dynamics depending on whether the critical point lies in the basin of
a finite attracting periodic point, or in the basin of ∞, or in the Julia set.
1. q0 (z) = z2 . There is a superattracting fixed point at 0, whose basin of
attraction is the open disk 1 = {z ∈ C: |z| < 1}, and a superattract-
ing fixed point at ∞, whose basin of attraction is the exterior of S1 .
There is also a repelling fixed point at 1, and for each n ∈ N there
are 2n repelling periodic points of period n on S1 . The Julia set is
S1 ; the Fatou set is the complement of S1 . The map q0 acts on S1 by
8.2. Examples 195

Figure 8.1. The Julia set for a = 1.

φ → 2φ mod 2π (where φ is the angular coordinate of a point z ∈ S1 ).


If U is a neighborhood of ζ ∈ S1 = J (q0 ), then n∈N0 q0n (U) = C\{0}.
2. q (z) = z2 − , 0 <   1. There is an attracting fixed point near 0,
a superattracting fixed point at ∞, and, for each n ∈ N, 2n repelling
periodic points near S1 . The Julia set J (q ) is a closed continuous q -
invariant curve that is C 0 close to S1 and is not differentiable at a dense
set of points; it has a Hausdorff dimension greater than 1. The basins
of attraction of the fixed points near 0 and at ∞ are, respectively, the
interior and exterior of J (q ). The critical point and critical value lie in
the immediate basin of attraction of the attracting fixed point near 0.
The same properties hold true for maps of the form f (z) = z2 +  P(z),
where P is a polynomial and  is small enough.
3. q1 (z) = z2 − 1. Note that q1 (0) = −1, q1 (−1) = 0. Therefore, 0 and −1
are superattracting periodic √ points of period 2. On the real line the
repelling fixed point (1 − 5)/2 separates the basins of attraction of
0 and −1. The Julia set J (q1 ) contains two simple closed curves σ0
and σ−1 that surround 0 and −1 and bound their immediate basins of
attraction. The only preimage of −1 is 0; hence the only preimage of
σ−1 is σ0 . However, 0 has two preimages, +1 and −1. Therefore σ0 has
two preimages, σ−1 and a closed curve σ1 surrounding 1. Continuing in
this manner and using the complete invariance of the Julia set (Propo-
sition 8.5.1), we conclude that J (q1 ) contains infinitely many closed
curves. Their interiors are components of the Fatou set. Figure 8.1
shows the Julia set for q1 .2

2 The pictures in this chapter were produced with mandelspawn; see http://www.araneus.fi/
gson/mandelspawn/.
196 8. Complex Dynamics

Figure 8.2. The Julia set for a = −i.

4. q−i = z2 + i. The critical point 0 is eventually periodic: q−i

2
(0) = i − 1,
and i − 1 is a repelling periodic point of period 2. The only attracting
periodic fixed point is ∞. The Fatou set consists of one component
and coincides with BA(∞). The Julia set is a dendrite, i.e., a compact,
path-connected, locally connected, nowhere dense subset of C that
does not separate C. Figure 8.2 shows the Julia set for q−i .
5. q2 (z) = z2 − 2. The change of variables z = ζ + ζ −1 conjugates q2 on
C̄\[−2, 2] with ζ → ζ 2 on the exterior of S1 . Hence J (q2 ) = [−2, 2],
and F(q2 ) = C̄\[−2, 2] is the basin of attraction of ∞. The image of the
critical point 0 is −2 ∈ J (q2 ). The change of variables y = (2 − x)/4
conjugates the action of q2 on the real axis to y → 4y(1 − y). The only
attracting periodic point is ∞.
6. q4 (z) = z2 − 4. The only attracting periodic point is ∞; the critical
value −4 lies in the (immediate) basin of attraction of ∞; and J (q4 ) is
a Cantor set on the real axis; BA(∞) is the complement of J (q4 ).
7. This example illustrates the connection between the dynamics of ra-
tional maps and issues of convergence for the Newton method. Let
Q(z) = (z − a)(z − b) with a =b. To find the roots a and b using the
Newton method one iterates the map
Q(z) 1
f (z) = z − = z− .
Q (z) 1
z− a
+ 1
z− b

The change of variables ζ = (z − a)/(z − b) sends a to 0, b to ∞, ∞

to 1, and the line l = {(a + b)/2 + ti(a − b): t ∈ R} to the unit circle,
and conjugates f with ζ → ζ 2 . Therefore the Newton method for
Q converges to a or b if the initial point lies in the half plane of l
8.3. Normal Families 197

containing a or b, respectively; the Newton method diverges if the

initial point lies on l.
Exercise 8.2.1. Prove the properties of q0 described above.
Exercise 8.2.2. Let U be a neighborhood of a point z ∈ S1 . Prove that

n∈N q0 (U) = C\{0}.

Exercise 8.2.3. Check the above conjugacies for q2 .

Exercise 8.2.4. Prove that ∞ is the only attracting periodic point of q4 .
Exercise 8.2.5. Let |a| > 2 and |z| ≥ |a|. Prove that qan (z) → ∞ as n → ∞.
Exercise 8.2.6. Prove the statements in example 7.

8.3 Normal Families

The theory of normal families of meromorphic functions is a keystone in the
study of complex dynamics. The principal result, Theorem 8.3.2, is due to
P. Montel [Mon27].
PROPOSITION 8.3.1. Suppose F is a family of analytic functions in a domain
D, and suppose that for every compact subset K ⊂ D there is C(K) > 0 such
that | f (z)| < C(K) for all z ∈ K and f ∈ F. Then F is a normal family.
Proof. Let δ = 1
2
minz∈K dist(z, ∂ D). By the Cauchy formula,

 1 f (ξ )
f (z) = dξ
2π γ (ξ − z)2
for any smooth closed curve γ in D that contains z in its interior. Let K ⊂ D
be compact, Kδ be the closure of the δ-neighborhood of K, and γ be the
circle of radius δ centered at z. Then | f  (z)| < C(Kδ )/δ for every f ∈ F and
z ∈ K. Thus the family F is equicontinuous on K, and therefore normal by
the Arzela–Ascoli theorem. 

We say that a family F of functions on a domain D omits a point a if
f (z) =a for every f ∈ F and z ∈ D.
THEOREM 8.3.2 (Montel). Suppose that a family F of meromorphic func-
tions in a domain D ⊂ C̄ omits three distinct points a, b, c ∈ C̄. Then F is
normal in D.
Proof. Since Dis covered by disks, we may assume without loss of generality
that D is a disk. By applying a Möbius transformation, we may also assume
that a = 0, b = 1, and c = ∞. Let 1 be the unit disk. By the uniformization
198 8. Complex Dynamics

theorem [Ahl73], there is an analytic covering map φ: 1 → (C\{0, 1})(φ is

called the modular function). For every function f : D → (C\{0, 1}) there is a
lift f˜: D → 1 such that φ ◦ f˜ = f . The family F̃ = { f˜: f ∈ F} is bounded
and therefore, by Proposition 8.3.1, normal. The normality of F follows
immediately. 

Exercise 8.3.1. Let f be a meromorphic map defined on a domain D ⊂ C̄,
and let k > 1. Show that the family { f n }n∈N is normal on D if and only if the
family { f kn }n∈N is normal.

8.4 Periodic Points

THEOREM 8.4.1. Let ζ be an attracting fixed point of a meromorphic map

f : C̄ → C̄. Then there is a neighborhood U  ζ and an analytic map φ: U →
C that conjugates f and z → λ(ζ )z, i.e., φ( f (z)) = λ(ζ )φ(z) for all z ∈ U.
Proof. We abbreviate λ(ζ ) = λ. Conjugating by a translation (or by z → 1/z
if ζ = ∞), we replace ζ by 0. Then on any sufficiently small neighborhood
of 0, say 1/2 = {z: |z| < 1/2}, there is a C > 0 such that | f (z) − λz| ≤ C|z|2 .
Hence for every  > 0 there is a neighborhood U of 0 such that | f (z)| <
(|λ| + )|z|, for all z ∈ U, and, assuming that |λ| +  < 1,

| f n (z)| < (|λ| + )n |z|.

Set φn (z) = λ−n f n (z). Then, for z ∈ U,

 
 f ( f n (z)) − λ f n (z)  C(|λ| + )2n |z|2
|φn+1 (z) − φn (z)| =  ≤
 ,
λn+1 |λ|n+1

and hence the sequence φn converges uniformly in U if (|λ| + )2 < |λ|.
By construction, φn ( f (z)) = λφn+1 (z). Therefore the limit φ = limn→∞ φn
is the required conjugation. 

COROLLARY 8.4.2. Let ζ be a repelling fixed point of a meromorphic map
f . Then there is a neighborhood U of ζ and an analytic map φ: U → C̄ that
conjugates f and z → λ(ζ )z in U, i.e., φ( f (z)) = λ(ζ )φ(z) for z ∈ U.
Proof. Apply Theorem 8.4.1 to the branch g of f −1 with g(ζ ) = ζ . 

PROPOSITION 8.4.3. Let ζ be a fixed point of a meromorphic map f .
Assume that λ = f  (ζ ) is not 0 and is not a root of 1, and suppose that an
analytic map φ conjugates f and z → λz. Then φ is unique up to multiplica-
tion by a constant.
8.4. Periodic Points 199

Proof. Again, we assume ζ = 0. If there are two conjugating maps φ and

ψ, then η = φ −1 ◦ ψ conjugates z → λz with itself, i.e., η(λz) = λη(z). If
η = a1 z + a2 z2 + · · · , then an λn = λan and an = 0 for n > 1. 

LEMMA 8.4.4. Any rational map R of degree >1 has infinitely many periodic
points.
Proof. Observe that the number of solutions of R n (z) − z = 0 (counted
with multiplicity) tends to ∞ as n → ∞. Therefore, if R has only finitely
many periodic points, their multiplicities cannot be bounded in n.
On the other hand, if ζ is a multiple root of R n (z) − z = 0, then (R n ) (ζ ) =
1 and R n (z) = ζ + (z − ζ ) + a(z − ζ )m + · · · for some a =0 and m ≥ 2. By
induction, R nk(z) = ζ + (z − ζ ) + ka(z − ζ )m + · · · for k ∈ N. Therefore, ζ
has the same multiplicity as a fixed point of R n and as a fixed point of R nk.


PROPOSITION 8.4.5. Let f be a meromorphic map of C̄. If ζ is an attracting
or superattracting periodic point of f , then the family { f n }n≥0 is normal in
BA(ζ ).
If ζ is a repelling periodic point of f , then the family { f n } is not normal
at ζ .
Proof. Exercise 8.4.1. 

THEOREM 8.4.6. Let ζ be an attracting periodic point of a rational map R.
Then the immediate basin of attraction BA◦ (ζ ) contains a critical point of f .
Proof. Consider first the case when ζ is a fixed point. Suppose that BA◦ (ζ )
does not contain a critical point. For a small enough  > 0, there is a branch
g of R−1 that is defined in the open -disk D about ζ and satisfies g(ζ ) = ζ .
The map g: D → BA◦ (ζ ) is a diffeomorphism onto its image, and therefore
g(D ) is simply connected and does not contain a critical point. Thus g
extends uniquely to a map on g(D ). By induction, g extends uniquely to
g n (D ), which is a simply connected subset of BA◦ (ζ ). The sequence {g n } is
normal on D , since it omits infinitely many periodic points of R different
from ζ (Lemma 8.4.4). (Note that if R is a polynomial, then {g n } omits a
neighborhood of ∞, and Lemma 8.4.4 is not needed.) On the other hand,
|g  (ζ )| > 1 and hence (g n ) (ζ ) → ∞ as n → ∞, and therefore the family {g n }
is not normal (Proposition 8.4.5); a contradiction.
If ζ is a periodic point of period n, then the preceding argument shows
that the immediate basin of attraction of ζ for the map R n contains a critical
point of R n . Since the components of BA◦ (ζ ) are permuted by R, it follows
from the chain rule that one of the components contains a critical point
of R. 

200 8. Complex Dynamics

COROLLARY 8.4.7. A rational map has at most 2d − 2 attracting and super-

attracting periodic orbits.
Proof. The corollary follows immediately from Theorem 8.4.6 and Propo-
sition 8.1.1. 

More delicate analysis that is beyond the scope of this book leads to the
following theorem.

THEOREM 8.4.8 (Shishikura [Shi87]). The total number of attracting, super-

attracting, and neutral periodic orbits of a rational map of degree d is at most
2d − 2.

The upper bound 6d − 6 was obtained by P. Fatou.

Exercise 8.4.1. Prove Proposition 8.4.5.

Exercise 8.4.2. Let D ⊂ C̄ be a domain whose complement contains at

least three points, and let f : D → D be a meromorphic map with an attract-
ing fixed point z0 ∈ D. Prove that the sequence of iterates f n converges in
D to z0 uniformly on compact sets.

Exercise 8.4.3. Prove that every rational map R =Id of degree d ≥ 1 has
d + 1 fixed points in C̄ counted with multiplicity.

8.5 The Julia Set

Recall that the Fatou set F(R) of a rational map R is the set of points z ∈
C̄ such that the family of forward iterates R n , n ∈ N, is normal at z. The
Julia set J (R) is the complement of F(R). The Julia set of a rational map is
closed by definition, and non-empty by Lemma 8.4.4, Proposition 8.4.5, and
Theorem 8.4.8. If U is a connected component of F(R), then R(U) is also a
connected component of F(R) (Exercise 8.5.1).
Suppose V =BA(∞) is a component of BA(∞). Then R n (V) ⊂ BA◦ (∞)
for some n > 0. Moreover, R n (V) is both open and closed in BA◦ (∞), since
R n (V) = R n (V ∪ J (R))\J (R). It follows that R n (V) = BA◦ (∞).

PROPOSITION 8.5.1. Let R: C̄ → C̄ be a rational map. Then F(R) and J (R)

are completely invariant, i.e., R−1 (F(R)) = F(R) and R(J (R)) = J (R), and
similarly for J (R).
Proof. Let ζ = R(ξ ). Then R nk converges in a neighborhood of ζ if and
only if R nk+1 converges in a neighborhood of ξ . 

8.5. The Julia Set 201

PROPOSITION 8.5.2. Let R: C̄ → C̄ be a rational map. Then either J (R) =

C̄ or J (R) has no interior.
Proof. Suppose U ⊂ J (R) is non-empty and open in C̄. Then the family

{R n }n∈N is not normal on U and, in particular, by Theorem 8.3.2, n R n (U)
omits at most two points in C̄. Since J (R) is invariant and closed, J (R) = C̄.


Let R: C̄ → C̄ be a rational map, and U an open set such that U ∩ J (R) =
∅. The family of iterates {R n }n∈N0 is not normal in U, so it omits at most two
points in C̄. The set EU of omitted points is called the exceptional set of R

on U. The exceptional set of R is the set E = EU , where the union is over
all open sets U with U ∩ J (R) = ∅. A point inE is called an exceptional point
of R.

PROPOSITION 8.5.3. Let R be a rational map of degree greater than 1. Then

the exceptional set of R contains at most two points. If the exceptional set
consists of a single point, then R is conjugate by a Möbius transformation to
a polynomial. If it consists of two points, then R is conjugate by a Möbius
transformation to zm or 1/zm, for some m > 1. The exceptional set is disjoint
from J (R).
Proof. If EU is empty for every U with U ∩ J (R) = ∅, there is nothing to
show.
Suppose {R n }n∈N0 omits two points z0 , z1 on U for some open set U with
U ∩ J (R) = ∅. Then after conjugating by the rational map φ(z) =
(z − z1 )/(z − z0 ), R becomes a rational map whose family of iterates omits
only 0 and ∞ on the set φ(U). Thus there are no solutions of R(z) = ∞ except
possibly 0 or ∞. If R(0) = ∞, thenR has no poles, so it is a polynomial, and
is therefore equal to zm, m > 0, since R(z) = 0 has no non-zero solutions. If
R(0) = ∞, then R has a unique pole at 0; since there are no finite solutions
of R(z) = 0, it follows that R(z) = 1/zm. We have shown that R is conjugate
to zm, |m| > 1, if the exceptional set of some open set U has two points. In
this case the exceptional set is {0, ∞}.
Suppose that {R n }n∈N0 omits at most a single point on U for every open
set U with U ∩ J (R) = ∅. Fix such a setU with EU = ∅, and letz0 be the
omitted point. Replacing R with its conjugate by the rational map φ(z) =
1/(z − z0 ), we may take z0 = ∞. Since {∞} is omitted, R has no poles, and is
therefore a polynomial. Thus R omits ∞ on every open subset U ⊂ C, and
(by hypothesis) omits only a single point on U if U ∩ J (R) = ∅, so ∞is the
only exceptional point of R.
In either case, J (R) does not contain any exceptional points. 

202 8. Complex Dynamics

The following proposition shows that the Julia set possesses self-similarity,
a characteristic property of fractal sets.

PROPOSITION 8.5.4. Let R : C̄ → C̄ be a rational map of degree >1 with

exceptional set E, and let U be a neighborhood of a point ζ ∈ J (R). Then

n∈N R (U) = C̄\E, and J (R) ⊂ R (U) for some n ∈ N.
n n

Proof. If E contains two points, then by Proposition 8.5.3, R is conjugate

to zm, |m| > 1, and the proof is left as an exercise (Exercise 8.5.4).
Suppose E is empty or consists of a single point. If the latter, we may
and do assume that the omitted point is ∞ and R is a polynomial. Since
repelling periodic points are dense in J (R), we may choose a neighborhood
V ⊂ U and n > 0 such that R n (V) ⊃ V. The family {R nk}k∈N on V does not
omit any points in C, and ∞ is omitted if and only if R is a polynomial, in

which case ∞ ∈ / J (R). Hence J (R) ⊂ n R n (V). Since J (R) is compact and
R nk(V) ⊃ R n(k−1) (V), the proposition follows. 

COROLLARY 8.5.5. Let R: C̄ → C̄ be a rational map of degree >1. For any
point ζ ∈ / E, J (R) is contained in the closure of the set of backward iterates
of ζ . In particular, J (R) is the closure of the set of backward iterates of any
point in J (R).

PROPOSITION 8.5.6. The Julia set of a rational map of degree >1 is perfect,
i.e., it does not have isolated points.
Proof. Exercise 8.5.3. 

PROPOSITION 8.5.7. Let R: C̄ → C̄ be a rational map of degree >1. Then
J (R) is the closure of the set of repelling periodic points.
Proof. We will show that J (R) is contained in the closure of the set Per(R)
of the periodic points of R. The result will follow, since J (R) is perfect and
there are only finitely many non-repelling periodic points.
Suppose ζ ∈ J (R) has a neighborhood U that contains no periodic points,
no poles, and no critical values of R. Since the degree of R is >1, there are
distinct branches f and g of R−1 in U, and f (z) =g(z), f (z) =R n (z), and
g(z) =R n (z) for all n ≥ 0 and all z ∈ U. Hence the family

R n (z) − f (z) z − g(z)

hn (z) = · , n ∈ N,
R n (z) − g(z) z − f (z)

omits 0,1, and ∞ in U and therefore is normal by Theorem 8.3.2. Since

R n can be expressed in terms of hn , the family {R n } is also normal in U, a
contradiction. Therefore J (R) ⊂ Per(R). 

8.5. The Julia Set 203

Let P: C̄ → C̄ be a polynomial. Then P(∞) = ∞, and locally near ∞

there are deg P branches of P−1 . The complete preimage of any connected
domain containing ∞ is connected, since ∞ = P−1 (∞) must belong to every
connected component of the preimage. Therefore BA(∞) is connected, i.e.,
BA(∞) = BA◦ (∞).

LEMMA 8.5.8. Let f : C̄ → C̄ be a meromorphic function, and suppose ζ

is an attracting periodic point. Then every component of BA◦ (ζ ) is simply
connected.
Proof. Since f cyclically permutes the components of BA◦ (ζ ), we may
replace f by f n , where n is the minimal period of ζ , and assume that ζ is
fixed. After conjugating by a Möbius transformation, we may assume that ζ
is finite.
Let γ be a smooth simple closed curve in BA◦ (ζ ), and let D be the simply
connected region (in C) that it bounds. Suppose D ⊆BA◦ (ζ ). Let δ be the
distance from ζ to the boundary of BA◦ (ζ ), and let U be the disk of radius δ/2
around ζ . Because ζ is attracting, and γ is a compact subset of BA◦ (ζ ), there
is n > 0 such that f n (γ ) ⊂ U. Let g(z) = f n (z) − ζ . Then |g(z)| < δ/2 on γ ,
but |g(z)| > δ for some z ∈ D, since f n (D) ⊆BA◦ (ζ ). This contradicts the
maximum principle for analytic functions. Thus D ⊂ BA◦ (ζ ), and BA◦ (ζ ) is
simply connected. 

PROPOSITION 8.5.9. Let R : C̄ → C̄ be a rational map of degree >1. If
U is any completely invariant component of F(R), then J (R) = Ū\U, and
J (R) = ∂U if F(R) is not connected. Every other component of F(R) is simply
connected. There are at most two completely invariant components. If R is a
polynomial, then BA(∞) is completely invariant.
Proof. Suppose U is a completely invariant component of F(R). Then, by
Corollary 8.5.5, J (R) is contained in the closure of U, and also of F(R)\U if
the latter is non-empty. This proves the first assertion. Since J (R) ∪ U = Ū
is connected, every component of the complement in C̄ is simply connected
(by a basic result of homotopy theory).
Suppose there is more than one completely invariant component of F(R).
Then, by the preceding paragraph, each must be simply connected. Let U be
such a component. Then R: U → U is a branched covering of degree d, so
there must be d − 1 critical points, counted with multiplicity. Since the total
number of critical points is 2d − 2 (Proposition 8.1.1), this implies that there
are at most two completely invariant components.
If R is a polynomial, then BA(∞) is completely invariant (Exercise 8.5.1).


204 8. Complex Dynamics

The postcritical set of a rational map R is the union of the forward orbits
of all critical points of R, and is denoted CL(R).

THEOREM 8.5.10 (Fatou). Let R be a rational map of degree >1. Suppose

that all critical points of R tend to attracting periodic points of R under the
forward iterates of R. Then J (R) is a hyperbolic set for R, i.e., there are a > 1
and n ∈ N such that |(R n ) (z)| ≥ a for every z ∈ J (R).
Proof. If R has exactly two critical points, then it is conjugate to zd or z−d
(Proposition 8.1.1), and the theorem follows by a direct computation.
We assume then that there are at least three critical points. Let U =
C̄\CL(R); then R−1 (U) ⊂ U. By the uniformization theorem [Ahl73], there
is an analytic covering map φ: 1 → U. Let g: 1 → 1 be the lift of a locally
defined branch of R−1 , so R ◦ φ ◦ g = φ.
The family {φ ◦ g n } is normal, since it omits CL(R). Let f be the uniform
limit of a sequence φ ◦ g nk . Let z0 ∈ φ −1 (J (R)), and let O ⊂ 1 be a neigh-
borhood of z0 such that φ(O) does not contain any attracting periodic points
of R. Since J (R) is invariant (Proposition 8.5.1) and closed, f (z0 ) ∈ J (R).
If f  (z0 ) =0, then f (O) contains a neighborhood of f (z0 ), and hence (by
Proposition 8.5.9) contains a point z1 ∈ BA(ξ ), where ξ is an attracting peri-
odic point. Since φ ◦ g nk → f , the value z1 is taken on by every φ ◦ g nk with k
large enough. This implies that R nk (z1 ) ∈ φ(O) for k sufficiently large, which
contradicts the fact that z1 ∈ BA(ξ ) and ξ ∈ / φ(O). Therefore, f  (z0 ) = 0, so
f is constant on φ (J (R)). It follows that (R nk ) = 1/(g nk ) goes to infinity
−1

uniformly on J (P), which proves the theorem. 


THEOREM 8.5.11 (Fatou). Let P: C̄ → C̄ be a polynomial such that Pn (c) →
∞ as n → ∞ for every critical point c. Then the Julia set J (P) is totally
disconnected, i.e., J (P) is a Cantor set.
Proof. Let D be a disk centered at 0 that contains J (P), and choose N
large enough that P N carries all critical points outside of D̄. Then for n ≥ N,
branches of P−n are globally defined on D. Fix z0 ∈ J (P), and let gn be the
branch of P−n with gn (Pn (z0 )) = z0 , for n ≥ N. The family F = {gn }n≥N is
uniformly bounded on D̄, and is therefore normal on D̄. Let f be the uniform
limit of a sequence in F. Since P is hyperbolic on J (P) (Theorem 8.5.10), f
must be constant on J (P), and therefore constant on D̄, since f is analytic
and J (P) has no isolated points. If y =z0 is any other point of J (P), then
y∈/ gn (D) for n sufficiently large, since the diameter of gn (D) converges to
zero. The set gn (D) ∩ J (P) is both open and closed in J (P), because ∂ D
does not intersect J (P). Thus z0 and y are in different components of J (P),
so J (P) is totally disconnected. 

8.6. The Mandelbrot Set 205

PROPOSITION 8.5.12. Let P: C̄ → C̄ be a polynomial such that no critical

point lies in BA(∞). Then J (P) is connected.
Proof. BA(∞) is simply connected (Lemma 8.5.8) and completely invari-
ant. If F(P) has only one component, then J (P) is the complement in C̄
of BA(∞), and is therefore connected by a fundamental result of algebraic
topology.
We assume then that F(P) has at least two components. We conjugate by a
Möbius transformation that carries ∞ to O, and one of the other components
of F(P) to a neighborhood of ∞. We obtain a rational map R such that 0
is a superattracting fixed point and BA(0) is a bounded, simply connected,
completely invariant component of F(R) that contains no critical points. Let
gn be the branch of R n on BA(0) with gn (0) = 0. Let γ be the unit circle.
Then gn (γ ) converges to J (R), so J (R) is connected. 

There are many other results about the Fatou and Julia sets that are be-
yond the scope of this book. For example, results of Wolff–Denjoy [Wol26],
[Den26] and of Douady–Hubbard [Dou83] show that if a component of the
Fatou set is eventually mapped back to itself, then its closure contains either
an attracting periodic point or a neutral periodic point. A result of Sullivan
[Sul85] shows that the Fatou set has no wandering components, i.e., no orbit
in the set of components is infinite.
Exercise 8.5.1. Show that if U is a connected component of F(R), then
R(U) is also a connected component of F(R). Show that if P is a polynomial,
then BA(∞) is completely invariant.
Exercise 8.5.2. Show that, for m > 1, the Julia set of z → zm is the unit circle
S1 , BA(∞) is the exterior of S1 , and the α-limit set of every z =0 is S1 .
Exercise 8.5.3. Prove Proposition 8.5.6.
Exercise 8.5.4. Prove Proposition 8.5.4 for R(z) = zm, |m| > 1.
Exercise 8.5.5. Let P be a polynomial of degree at least 2. Prove that Pn →
∞ on the component of F(P) that contains ∞.
Exercise 8.5.6. Show that if R is a rational map of degree >1, and F(R)
has only finitely many components, then it has either 0, 1, or 2 components.

8.6 The Mandelbrot Set

For a general quadratic function q(z) = αz2 + βz + γ with α =0, the change
of variables ζ = z + β/2 maps the critical point to 0 and conjugates q with
206 8. Complex Dynamics

Figure 8.3. The Mandelbrot set.

qa (z) = z2 − a. Since the conjugation is unique, the maps qa , a ∈ C, are

in one-to-one correspondence with conjugacy classes of quadratic maps. If
qan (0) → ∞, then J (qa ) is totally disconnected (see Theorem 8.5.11).
Otherwise, the orbit {qan (0)}n∈N is bounded and J (qa ) is connected
(Proposition 8.5.12).
The Mandelbrot set M is the set of parameter values a for which the
/ BA(∞) for qa }. The
orbit of 0 is bounded, or equivalently, M = {a ∈ C: 0 ∈
Mandelbrot set is shown in Figure 8.3.

THEOREM 8.6.1 (Douady–Hubbard [DH82]). M = {a ∈ C: |qan (0)| ≤ 2 for

all n ∈ N}. M is closed and simply connected.
Proof. Let |a| > 2. We have |qa (0)| = |a| > 2, |qa2 (0)| = |qa (a)| ≥ |a 2 | − |a| =
|a|(|a| − 1), and |qan (0)| ≥ |a|(|a| − 1)n−1 for n ∈ N (Exercise 8.6.1). There-
fore a ∈/ M. If |a| ≤ 2 and |qan (0)| = 2 + α for some n ∈ N and α > 0, then
|qa (0)| ≥ (2 + α)2 − 2 > 2 + 4α and |qan+k(0)| ≥ 2 + 4kα → ∞ as k → ∞.
n+1

Therefore a ∈ / M. The first and second statements follow.

If D is a bounded component of C\M, then maxa∈ D̄ |qan (0)| > 2 for some
n ∈ N, and, by the maximum principle, |qan (0)| > 2 for some a ∈ ∂ D ⊂ M.
This contradicts the first assertion of the theorem. Thus C\M has no bounded
components, has only one unbounded component containing ∞, and is there-
fore connected. Hence M is simply connected. 

±
√ ±
√ The fixed points of qa are za √= (1 ± 1 + 4a)/2 with multipliers λ = 1 ±
1 + 4a. The set {a ∈ C: |1 ± 1 + 4a| < 1} is a subset of M (Exercise 8.6.3)
and is called the main cardioid of M.
8.6. The Mandelbrot Set 207

PROPOSITION 8.6.2. Every point in ∂ M is an accumulation point of the set

of values of a for which qa has a superattracting cycle.
Proof. Since 0 is the only critical point of qa , a periodic orbit is superattract-
ing if and only if it contains 0. Let D be a disk that intersects ∂ M and does
not contain 0, and suppose that 0 is not a periodic point of qa for any a ∈ D.
√
Then (qan (0))2 =a for all a ∈ D and n ∈ N. Let a be a branch of the inverse
√
of z → z2 defined on D, and define fn (a) = qan (0)/ a for n ∈ N and a ∈ D.
Then the family { fn }n∈N omits 0, 1, and ∞ on D, and is therefore normal
in D. On the other hand, since D intersects ∂ M, it contains both points a
for which fn (a) is bounded and points a for which fn (a) → ∞, and hence
the family { fn } is not normal on D. Thus 0 must be periodic for qa for some
a ∈ D. 

Exercise 8.6.1. Prove by induction that if |a| > 2, then |qan (0)| ≥
|a|(|a| − 1)n−1 for n ∈ N.
Exercise 8.6.2. Prove that the intersection of M with the real axis is
[−2, 1/4].
Exercise 8.6.3. Prove that the main cardioid is contained in M.
Exercise 8.6.4. Prove that the set of values a in C for which qa has an
attracting periodic point of period 2 is the disk of radius 1/4 centered at −1
(it is tangent to the main cardioid). Prove that this set is contained in M.
CHAPTER NINE

Measure-Theoretic Entropy

In this chapter, we give a short introduction to measure-theoretic entropy,

also called metric entropy, for measure-preserving transformations. This
invariant was introduced by A. Kolmogorov [Kol58], [Kol59] to classify
Bernoulli automorphisms and developed further by Ya. Sinai [Sin59] for
general measure-preserving dynamical systems. The measure-theoretic en-
tropy has deep roots in thermodynamics, statistical mechanics, and informa-
tion theory. We explain the interpretation of entropy from the perspective
of information theory at the end of the first section.

9.1 Entropy of a Partition

Throughout this chapter (X, A, µ) is a Lebesgue space with µ(X) = 1. We
use the notation of Chapter 4. A (finite) partition of X is a finite collection
ζ of essentially disjoint measurable sets Ci (called elements or atoms of ζ )
whose union covers X mod 0. We say that a partition ζ  is a refinement of
ζ and write ζ ≤ ζ  (or ζ  ≥ ζ ) if every element of ζ  is contained mod 0 in
an element of ζ . Partitions ζ and ζ  are equivalent if each is a refinement of
the other. We will deal with equivalence classes of partitions. The common
refinement ζ ∨ ζ  of partitions ζ and ζ  is the partition into intersections
Cα ∩ Cβ , where Cα ∈ ζ and Cβ ∈ ζ  ; it is the smallest partition which is ≥ ζ
and ζ  . The intersection ζ ∧ ζ  is the largest measurable partition which is
≤ζ and ζ  . The trivial partition consisting of a single element X is denoted
by ν.
Although many definitions and statements in this chapter hold for infinite
partitions, we discuss only finite partitions.
For A, B ⊂ X, let A
B = (A\B) ∪ (B\A). Let ξ = {Ci : 1 ≤ i ≤ m} and
η = {Dj : 1 ≤ j ≤ n} be finite partitions. By adding null sets if necessary, we
208
9.1. Entropy of a Partition 209

may assume that m = n. Define

m
 
d(ξ, η) = min µ Ci
Dσ (i) ,
σ ∈Sm
i=1

where the minimum is taken over all permutations of m elements. The ax-
ioms of distance are satisfied by d (Exercise 9.1.1).
Partitions ζ and ζ  are independent, and we write ζ ⊥ ζ  , if µ(C ∩ C  ) =
µ(C) · µ(C  ) for all C ∈ ζ and C  ∈ ζ  .
For a transformation T and partition ξ = {C1 , . . . , Cm}, let T −1 (ξ ) =
{T (C1 ), . . . , T −1 (Cm)}.
−1

To motivate the definition of entropy below, consider a Bernoulli auto-

morphism of
m with probabilities qi > 0, q1 + · · · + qm = 1 (see §4.4). Let
ξ be the partition of
m into m sets Ci = {ω ∈
m: ω0 = i}, µ(Ci ) = qi . Set
 −k
ηn = n−1k=0 σ (ξ ), and let ηn (ω) denote the element of ηn containing ω. For
ω ∈
m, let fin (ω) be the relative frequency of symbol i in the word ω1 . . . ωn .
Since σ is ergodic with respect to µ, by the Birkhoff ergodic theorem 4.5.5,
for every > 0 there are N ∈ N and a subset A ⊂
m with µ(A ) > 1 −
such that | fin (ω) − qi | < for each ω ∈ A and n ≥ N. Therefore, if ω ∈ A ,
then

m
(qi + i )n
m
µ(ηn (ω)) = qi = 2n i=1 (qi + i ) log qi ,
i=1

where | i | < , and from now on log denotes logarithm base 2 with 0 log 0 =
0. It follows that for µ-a.e. ω ∈
m
1
m
lim log µ(ηn (ω)) = qi log qi ,
n→∞ n
i=1

and hence the number of elements of ηn with approximately correct

frequency of symbols 1, . . . , m grows exponentially as 2nh , where h =
m
− i=1 qi log qi .
For a partition ζ = {C1 , . . . , Cn } define the entropy of ζ by


n
H(ζ ) = − µ(Ci ) log µ(Ci )
i=1

(recall that 0 log 0 = 0). Note that −x log x is a strictly concave continuous

function on [0, 1], i.e., if xi ≥ 0, λi ≥ 0, i = 1, . . . , n, and i λi = 1, then
 

n
n
n
− λi xi · log λi xi ≥ − λi xi log xi (9.1)
i=1 i=1 i=1
210 9. Measure-Theoretic Entropy

with equality if and only if all xi s are equal. For x ∈ X, let m(x, ζ ) denote
the measure of the element of ζ containing x. Then

H(ζ ) = − log m(x, ζ ) dµ.
X

PROPOSITION 9.1.1. Let ξ and η be finite partitions. Then

1. H(ξ ) ≥ 0, and H(ξ ) = 0 if and only if ξ = ν;
2. if ξ ≤ η, then H(ξ ) ≤ H(η), and equality holds if and only if ξ = η;
3. if ξ has n elements, then H(ξ ) ≤ log n, and equality holds if and only if
each element of ξ has measure 1/n;
4. H(ξ ∨ η) ≤ H(ξ ) + H(η) with equality if and only if ξ ⊥ η.
Proof. We leave the first three statements as exercises (Exercise 9.1.2). To
prove the last statement, let µi , ν j , and κi j be the measures of the elements
 
of ξ, η, and ξ ∨ η, respectively, so that j κi j = µi and i κi j = ν j . It follows
from (9.1) that

κi j κi j


−ν j log ν j ≥ − µi · log =− κi j log κi j + κi j log µi ,
i
µi µi i i

and summation over j finishes the proof of the inequality. The equality is
achieved if and only if xi = κi j /µi does not depend on i for each j, which is
equivalent to the independence of ξ and η. 

The entropy of a partition has a natural interpretation as the “average
information of the elements of the partition.” For example, suppose X
represents the set of all possible outcomes of an experiment, and µ is the
probability distribution of the outcomes. To extract information from the
experiment, we devise a measuring scheme that effectively partitions X into
finitely many observable subsets, or events, C1 , C2 , . . . , Cn . We define the
information of an event C to be I(C) = − log µ(C). This is a natural choice
given that the information should have the following properties:

1. The information is a non-negative and decreasing function of the prob-

ability of an event; the lower the probability of an event, the greater
the informational content of observing that event.
2. The information of the trivial event X is 0.
3. For independent events C and D, the information is additive, i.e.,
I(C ∩ D) = I(C) + I(D).

Up to a constant, − log µ(C) is the only such function.

9.2. Conditional Entropy 211

With this definition of information, the entropy of a partition is simply

the average information of the elements of the partition.

Exercise 9.1.1. Prove: (i) d(ξ, η) ≥ 0 with equality if and only if ξ = η mod 0
and (ii) d(ξ, ζ ) ≤ d(ξ, η) + d(η, ζ ).

Exercise 9.1.2. Prove the first three statements of Proposition 9.1.1.

Exercise 9.1.3. For n ∈ N, let Pn be the the space of equivalence classes of

finite partitions with n elements with metric d. Prove that the entropy is a
continuous function on Pn .

9.2 Conditional Entropy

For measurable subsets C, D ⊂ X with µ(D) > 0, set µ(C|D) = µ(C ∩ D)/
µ(D). Let ξ = {Ci : i ∈ I} and η = {Dj : j ∈ J } be partitions. The conditional
entropy of ξ with respect to η is defined by the formula



H(ξ |η) = − µ(Dj ) µ(Ci |Dj ) log µ(Ci |Dj ).
j∈J i∈I

The quantity H(ξ |η) is the average entropy of the partition induced by ξ
on an element of η. If C(x) ∈ ξ and D(x) ∈ η are the elements containing x,
then

H(ξ |η) = − log µ(C(x)|D(x)) dµ.
X

The following proposition gives several simple properties of conditional

entropy.

PROPOSITION 9.2.1. Let ξ, η, and ζ be finite partitions. Then

1. H(ξ |η) ≥ 0 with equality if and only if ξ ≤ η;
2. H(ξ |ν) = H(ξ );
3. if η ≤ ζ , then H(ξ |η) ≥ H(ξ |ζ );
4. if η ≤ ζ , then H(ξ ∨ η|ζ ) = H(ξ |ζ );
5. if ξ ≤ η, then H(ξ |ζ ) ≤ H(η|ζ ) with equality if and only if ξ ∨ ζ =
η ∨ ζ;
6. H(ξ ∨ η|ζ ) = H(ξ |ζ ) + H(η|ξ ∨ ζ ) and H(ξ ∨ η) = H(ξ ) + H(η|ξ );
7. H(ξ |η ∨ ζ ) ≤ H(ξ |ζ );
8. H(ξ |η) ≤ H(ξ ) with equality if and only if ξ ⊥ η.
212 9. Measure-Theoretic Entropy

Proof. To prove part 6, let ξ = {Ai }, η = {Bj }, ζ = {Ck}. Then

µ(Ai ∩ Bj ∩ Ck)
H(ξ ∨ η|ζ ) = − µ(Ai ∩ Bj ∩ Ck) · log
i, j,k
µ(Ck)

µ(Ai ∩ Ck)
=− µ(Ai ∩ Bj ∩ Ck) log
i, j,k
µ(Ck)

µ(Ai ∩ Bj ∩ Ck)
− µ(Ai ∩ Bj ∩ Ck) log
i, j,k
µ(Ai ∩ Ck)
= H(ξ |ζ ) + H(η|ξ ∨ ζ ),

and the first equality follows. The second equality follows from the first one
with ζ = ν.
The remaining statements of Proposition 9.2.1 are left as exercises
(Exercise 9.2.1). 


For finite partitions ξ and η, define

ρ(ξ, η) = H(ξ |η) + H(η|ξ ).

The function ρ, which is called the Rokhlin metric, defines a metric on the
space of equivalence classes of partitions (Exercise 9.2.2).

PROPOSITION 9.2.2. For every > 0 and m ∈ N there is δ > 0 such that if
ξ and η are finite partitions with at most m elements and d(ξ, η) < δ, then
ρ(ξ, η) < .
Proof ([KH95], Proposition 4.3.5). Let partitions, ξ = {Ci : 1 ≤ i ≤ m}, η =
m
{Di : 1 ≤ i ≤ m} satisfy d(ξ, η) = i=1 µ(Ci
Di ) = δ. We will estimate
H(η|ξ ) in terms of δ and m.
If µ(Ci ) > 0, set αi = µ(Ci \Di )/µ(Ci ). Then

µ(Ci ∩ Di )
−µ(Ci ∩ Di ) log ≤ −µ(Ci )(1 − αi ) log(1 − αi )
µ(Ci )

and, by Proposition 9.1.1(3) applied to the partition of Ci \Di induced by η,

µ(Ci ∩ Dj )
− µ(Ci ∩ Dj ) log ≤ −µ(Ci )αi (log αi − log(m − 1)).
j=i
µ(Ci )
9.3. Entropy of a Measure-Preserving Transformation 213

Therefore, since log x is concave,

µ(Ci ∩ Dj )
− µ(Ci ∩ Dj ) log
j
µ(Ci )


1 m− 1
≤ µ(Ci ) (1 − αi ) log + αi log ≤ µ(Ci ) log m.
1 − αi αi
It follows that


H(η|ξ ) ≤ µ(Ci ) log m
√
µ(Ci )< δ


+ µ(Ci )(−(1 − αi ) log(1 − αi ) − αi log αi + αi log(m − 1)).
√
µ(Ci )≥ δ
√
The first term does not exceed δ m log m. To √ estimate the√second term,
observe that αi µ(Ci ) ≤ δ. Hence, if µ(Ci ) ≥ δ, then αi ≤ δ. Since the
function f (x) = −x log x − (1 − x) log(1 − √x) is increasing
√ on (0, 1/2), for
small δ the second term does not exceed f ( δ) + δ log(m − 1), and
√ √
H(η|ξ ) ≤ f ( δ) + δ(m log m + log(m − 1)).
Since f (x) → 0 as x → 0, the proposition follows. 

Exercise 9.2.1. Prove the remaining statements of Proposition 9.2.1.
Exercise 9.2.2. Prove that (i) ρ(ξ, η) ≥ 0 with equality if and only if ξ =
η mod 0 and (ii) ρ(ξ, ζ ) ≤ ρ(ξ, η) + ρ(η, ζ ).

9.3 Entropy of a Measure-Preserving Transformation

Let T be a measure-preserving transformation of a measure space (X, A, µ)
and ζ = {Cα : α ∈ I} be a partition of X with finite entropy. For k, n ∈ N,
set T −k(ζ ) = {T −k(Cα ): α ∈ I} and
ζ n = ζ ∨ T −1 (ζ ) ∨ · · · ∨ T −n+1 (ζ ).
Since H(T −k(ζ )) = H(ζ ) and H(ξ ∨ η) ≤ H(ξ ) + H(η), we have that
H(ζ m+n ) ≤ H(ζ m) + H(ζ n ). By subadditivity (Exercise 2.5.3), the limit
1
h(T, ζ ) = lim H(ζ n )
n→∞ n

exists, and is called the metric (or measure-theoretic) entropy of T relative to

ζ . Note that h(T, ζ ) ≤ H(ζ ).
PROPOSITION 9.3.1. h(T, ζ ) = limn→∞ H(ζ |T −1 (ζ n )).
214 9. Measure-Theoretic Entropy

 
Proof. Since H(ξ |η) ≥ H(ξ |ζ ) for η ≤ ζ , the sequence H ζ |T −1 (ζ n ) is
non-increasing in n. Since H(T −1 ξ ) = H(ξ ) and H(ξ ∨ η) = H(ξ ) + H(η|ξ ),
we get

H(ζ n ) = H(T −1 (ζ n−1 ) ∨ ζ ) = H(ζ n−1 ) + H(ζ |T −1 (ζ n−1 ))

= H(ζ n−2 ) + H(ζ |T −1 (ζ n−2 )) + H(ζ |T −1 (ζ n−1 )) = · · ·


n−1
= H(ζ ) + H(ζ |T −1 (ζ k)).
k=1

Dividing by n and passing to the limit as n → ∞ finishes the proof. 


Proposition 9.3.1 means that h(T, ζ ) is the average information added by
the present state on condition that all past states are known.

PROPOSITION 9.3.2. Let ξ and η be finite partitions. Then

1. h(T, T −1 (ξ )) = h(T, ξ ); if T is invertible, then h(T, T(ξ )) = h(T, ξ );
n
2. h(T, ξ ) = h(T, i=0 T −i (ξ )) for n ∈ N; if T is invertible, then
n
h(T, ξ ) = h(T, i=−n T −i (ξ )) for n ∈ N;
3. h(T, ξ ) ≤ h(T, η) + H(ξ |η); if ξ ≤ η, then h(T, ξ ) ≤ h(T, η);
4. |h(T, ξ ) − h(T, η)| ≤ ρ(ξ, η) = H(ξ |η) + H(η|ξ ) (the Rokhlin inequal-
ity);
5. h(T, ξ ∨ η) ≤ h(T, ξ ) + h(T, η);
Proof. To prove statement 3 observe that, by the second statement of
Proposition 9.2.1(6), H(ξ n ) ≤ H(ξ n ∨ ηn ) = H(ηn ) + H(ξ n |ηn ). We apply
Proposition 9.2.1(6) n times to get

H(ξ n |ηn ) = H(ξ ∨ T −1 (ξ n−1 )|ηn ) = H(ξ |ηn ) + H(T −1 (ξ n−1 )|ξ ∨ ηn )
≤ H(ξ |η) + H(T −1 (ξ n−1 )|ηn )
≤ H(ξ |η) + H(T −1 (ξ )|T −1 (η)) + H(T −2 (ξ n−2 )|ηn )
..
.
≤ nH(ξ |η).

Therefore
1 1
H(ξ n ) ≤ H(ηn ) + H(ξ |η),
n n
and statement 3 follows.
The remaining statements of Proposition 9.3.2 are left as exercises
(Exercise 9.3.2). 

9.3. Entropy of a Measure-Preserving Transformation 215

The metric (or measure-theoretic) entropy is the supremum of the

entropies h(T, ζ ) over all finite measurable partitions ζ of X.
If two measure-preserving transformations are isomorphic (i.e., if there
exists a measure-preserving conjugacy), then their measure-theoretic en-
tropies are equal. If the entropies are different, the transformations are not
isomorphic.
We will need the following lemma.

LEMMA 9.3.3. Let η be a finite partition, and let ζn be a sequence of finite

partitions such that d(ζn , η) → 0. Then there are finite partitions ξn ≤ ζn such
that H(η|ξn ) → 0.
Proof. Let η = {Dj : 1 ≤ j ≤ m}. For each j choose a sequence C nj ∈ ζn
such that µ(Dj
C nj ) → 0. Let ξn consist of C nj , 1 ≤ j ≤ m, and Cm+1
n
=
m
X\ j+1 C j . Then µ(C j ) → µ(Dj ) and µ(Cm+1 ) → 0. We have
n n n

 

m
  µ Cin ∩ Di
H(η|ξn ) = − µ Cin ∩ Di · log  n
i=1
µ Ci
 n 

m
 n  µ Cm+1 ∩ Dj
− µ Cm+1 ∩ Dj · log  n 
j=1
µ Cm+1
 

m

 n  µ Cin ∩ Dj
− µ Ci ∩ Dj · log   .
i=1 j=i
µ Cin

The first sum tends to 0 because µ(Cin ∩ Di ) → µ(Cin ). The second and third
sums tend to 0 because µ(Cin ∩ Dj ) → 0 for j =i. 

A sequence (ζn ) of finite partitions is called refining if ζn+1 ≥ ζn for n ∈ N.
A sequence (ζn ) of finite partitions is called generating if for every finite
partition ξ and every δ > 0 there is n0 ∈ N such that for every n ≥ n0 there
n
is a partition ξn with ξn ≤ i=1 ζi and d(ξn , ξ ) < δ, or equivalently if every
n
measurable set can be approximated by a union of elements of i=1 ζi for a
large enough n.
Every Lebesgue space has a generating sequence of finite partitions
(Exercise 9.3.3). If X is a compact metric space with a non-atomic Borel
measure µ, then a sequence of finite partitions ζn is generating if the maxi-
mal diameter of elements of ζn tends to 0 as n → ∞ (Exercise 9.3.4).

PROPOSITION 9.3.4. If (ζn ) is a refining and generating sequence of finite

partitions, then h(T) = limn→∞ h(T, ζn ).
Proof. Let ξ be a partition of X with m elements. Fix > 0. Since (ζn )
is a refining and generating, for every δ > 0 there is n ∈ N and a partition ξn
216 9. Measure-Theoretic Entropy

n
with melements such that ξn ≤ i=1 ζi and d(ξn , ξ ) < δ. By Proposition 9.2.2,

ρ(ξ, ζn ) = H(ξ |ζn ) + H(ζn |ξ ) < .

By the Rokhlin inequality (Proposition 9.3.2(4)), h(T, ξ ) < h(T, ζn ) + .



A (one-sided) generator for a non-invertible measure-preserving trans-

formation T is a finite partition ξ such that the sequence ξ n = nk=0 T −k(ξ )
is generating. For an invertible T, a (two-sided) generator is a finite parti-

tion ξ such that the sequence nk=−n T k(ξ ) is generating. Equivalently, ξ is a

generator if for any finite partition η there are partitions ζn ≤ nk=0 T −k(ξ )
n
(or ζn ≤ k=−n T k(ξ )) such that d(ζn , η) → 0.
The following corollary of Proposition 9.3.4 allows one to calculate the
entropy of many measure-preserving transformations.

THEOREM 9.3.5 (Kolmogorov–Sinai). Let ξ be a generator for T. Then

h(T) = h(T, ξ ).
Proof. We consider only the non-invertible case. Let η be a finite parti-
n
tion. Since ξ is a generator, there are partitions ζn ≤ i=0 T −i (ξ ) such that
d(ζn , η) → 0. By Lemma 9.3.3 for any δ > 0 there is n ∈ N and a parti-
n
tion ξn ≤ ζn ≤ i=0 T −i (ξ ) with H(ξn |η) < δ. By statements 3, 5, and 2 of
Proposition 9.3.2,
 
n
h(T, η) ≤ h(T, ξn ) + H(η|ξn ) ≤ h T, T (ξ ) + δ = h(T, ξ ) + δ. 
−i

i=0

PROPOSITION 9.3.6. Let T and S be measure-preserving transformations

of measure spaces (X, A, µ) and (Y, B, ν), respectively.
1. h(T k) = kh(T) for every k ∈ N; if T is invertible, then h(T −1 ) = h(T)
and h(T k) = |k|h(T) for every k ∈ Z.
2. If T is a factor of S, then hµ (T) ≤ hν (S).
3. hµ×ν (T × S) = hµ (T) + hν (S).
Proof. To prove statement 3, consider refining and generating sequences
of partitions ξk and ηk in X and Y, respectively. Then

ζk = (ξk × ν) ∨ (µ × ηk)

is a refining and generating sequence in X × Y. Since

       
ζkn = ξkn × ν ∨ µ × ηkn and ξkn × ν ⊥ µ × ηkn ,
9.3. Entropy of a Measure-Preserving Transformation 217

we obtain, by Proposition 9.1.1 and Proposition 9.3.4, that

1  n 1   n  
h(T × S) = lim lim H ζk lim lim H ξk + H ηkn = h(T) + h(S).
k→∞ n→∞ n k→∞ n→∞ n
The first two statements are left as exercises (Exercise 9.3.6). 

Let T be a measure-preserving transformation of a probability space
(X, A, µ), and ζ a finite partition. As before, let m(x, ζ n ) be the measure
of the element of ζ n containing x ∈ X. The amount of information con-
veyed by the fact that x lies in a particular element of ζ n (or that the
points x, T(x), . . . , T n−1 (x) lie in particular elements of ζ ) is Iζ n (x) = − log
× m(x, ζ n ). A proof of the following theorem can be found in [Pet89] or
[Mañ88].

THEOREM 9.3.7 (Shannon–McMillan–Breiman). Let T be an ergodic

measure-preserving transformation of a probability space (X, A, µ), and ζ
a finite partition. Then
1
lim Iζ n (x) = h(T, ζ ) for a.e. x ∈ X and in L1 (X, A, µ).
n→∞ n
Theorem 9.3.7 implies that, for a typical point x ∈ X, the information
Iζ n (x) grows asymptotically as n · h(T, ζ ) and the measure m(x, ζ n ) decays
exponentially as e−nh(T,ζ ) . The proof of the following corollary is left as an
exercise (Exercise 9.3.8).

COROLLARY 9.3.8. Let T be an ergodic measure-preserving transformation

of a probability space (X, A, µ), and ζ a finite partition. Then for every > 0
there is n0 ∈ N and for every n ≥ n0 a subset Sn of the elements of ζ n such
that the total measure of the elements from Sn is ≥ 1 − and for each element
C ∈ Sn

−n(h(T, ζ ) + ) < log µ(C) < −n(h(T, ζ ) − ).

Exercise 9.3.1. Let T be a measure-preserving transformation of a non-

atomic measure space (X, A, µ). For a finite partition ξ and x ∈ X, let ξn (x)
be the element of ξ n containing x. Prove that µ(ξ n (x)) → 0 as n → ∞ for a.e.
x and every non-trivial finite partition ξ if and only if all powers T n , n ∈ N,
are ergodic.

Exercise 9.3.2. Prove the remaining statements of Proposition 9.3.2.

Exercise 9.3.3. Prove that every Lebesgue space has a generating sequence
of partitions.
218 9. Measure-Theoretic Entropy

Exercise 9.3.4. If ζ is a partition of a finite metric space, then we define

the diameter of ζ to be diam(ζ ) = supC∈ζ diam(C). Prove that a sequence
(ζn ) of finite partitions of a compact metric space X with a non-atomic Borel
measure µ is generating if the diameter of ζn tends to 0 as n → ∞.

Exercise 9.3.5. Suppose a measure-preserving transformation T has a gen-

erator with k elements. Prove that h(T) ≤ log k.

Exercise 9.3.6. Prove the first two statements of Proposition 9.3.6.

Exercise 9.3.7. Show that if an invertible transformation T has a one-sided

generator, then h(T) = 0.

Exercise 9.3.8. Prove Corollary 9.3.8.

9.4 Examples of Entropy Calculation

Let (X, d) be a compact metric space, and µ a non-atomic Borel measure on
X. By Exercise 9.3.4, any sequence of finite partitions whose diameter tends
to 0 is generating. We will use this fact repeatedly in computing the metric
entropy of some topological maps.

Rotations of S1 . Let λ be the Lebesgue measure on S1 . If α is rational,

then Rαn = Id for some n, so hλ (Rα ) = (1/n)hλ (Rnα ) = (1/n)hλ (Id) = 0. If α
is irrational, let ξ N be a partition of S1 into N intervals of equal length. Then
ξ Nn consists of nN intervals, so H(ξ Nn ) ≤ log nN. Thus h(Rα , ξ N ) ≤ limn−>∞
(log nN)/n = 0. The collection of partitions ξ N , N ∈ N, is clearly generating,
so h(Rα ) = 0.
This result can also be deduced from Exercise 9.3.7 by noting that ev-
ery forward semiorbit is dense, so any non-trivial partition is a one-sided
generator for Rα .

Expanding Maps. The partition

ξ = {[0, 1/k), [1/k, 2/k), . . . , [(k − 1)/k, 1)}

is a generator for the expanding map Ek: S1 → S1 , since the elements of ξ n

are of the form [i/kn , (i + 1)/kn ). We have

1

1
H(ξ n ) = − log = n log |k|,
|k|n |k|n
so hλ (Ek) = log |k|.
9.4. Examples of Entropy Calculation 219

Shifts. Let σ :
m →
m be the one or two-sided shift on m symbols, and
m
let p = ( p1 , . . . , pm) be a non-negative vector with i=1 pi = 1. The vector
p defines a measure on the alphabet {1, 2, . . . , m}. The associated product
measure µ p on
m is called a Bernoulli measure. For a cylinder set, we have

 ,...,nk   k
µ p C nj11,..., jk = p ji .
i=1

Let ξ = j = 1, . . . , m}. Then ξ is a (one- or two-sided) generator

{C 0j :
m i
for σ , since diam( i=0 σ ξ ) → 0 with respect to the metric d(ω, ω ) = 2−l ,
where l = min{|i|: ωi =ωi }. Thus
 
1 m
−i
hµ p (σ ) = hµ p (σ, ξ ) = lim H σ ξ
n→∞ n
i=0

For i = j, σ i ξ and σ j ξ are independent, so

 
m
−i
H σ ξ = nH(ξ ).
i=0

Thus hµ p (σ ) = H(ξ ) = − pi log pi .
Recall that the topological entropy of σ is log m. Thus the metric entropy
of σ with respect to any Bernoulli measure is less than or equal to the
topological entropy, and equality holds if and only if p = (1/n, . . . , 1/n).
We next calculate the metric entropy of σ with respect to the Markov
measures defined in §4.4. Let A be an irreducible m × m stochastic matrix,
and q the unique positive left eigenvector whose entries sum to 1. Recall
that for the measure P = PA,q , the measure of a cylinder set is

  
k−1
P C n,n+1,...,n+k
j0 , j1 ,..., jk = q j0 Aji ji+1 .
i=0

By Proposition 9.3.1, we have h P (σ, ξ ) = limn→∞ H(ξ |σ −1 (ξ n )). By defini-

tion,

P(C ∩ D)
H(ξ |σ −1 (ξ n )) = − P(C ∩ D) log .
C∈ξ,D∈σ −1 (ξ n )
P(D)

−1 n
For C = C 0j0 ∈ ξ and D = C 1,...,n
j1 ,..., jn ∈ σ (ξ ), we have


n−1 
n−1
P(C ∩ D) = q j0 Aji ji+1 and P(D) = q j1 Aji ji+1 .
i=0 i=1
220 9. Measure-Theoretic Entropy

Thus


m 
n−1

−1 q j0 Aj0 , j1
H(ξ |σ (ξ )) = −
n
q j0 Aji , ji+1 log
j0 , j1 ,..., jn =1 i=1
q j1


m 
n−1
= − q j0 Aji , ji+1 (log Aj0 , j1 + log q j0 − log q j1 ).
j0 , j1 ,..., jn =1 i=1

(9.2)
n n
Using the identities i=1 qi Ai,k = qk and k=1 Ai,k = 1, we find that


n−1

q j0 Aji , ji+1 log Aj0 , j1 = q j0 Aj0 , j1 log Aj0 , j1 , (9.3)
j0 , j1 ,..., jn i=0 j0 , j1


n−1

q j0 Aji , ji+1 log q j0 = q0 log q j0 , (9.4)
j0 , j1 ,..., jn i=0 j0

m 
n−1

q j0 Aji , ji+1 log q j1 = q j1 log q j1 . (9.5)
j0 , j1 ,..., jn =1 i=1 j1

It follows from (9.2)–(9.5) that

h P (σ ) = − q j0 Aj0 , j1 log Aj0 , j1 .
j0 , j1

There are many Markov measures for a given subshift. We now construct a
special Markov measure, called the Shannon–Parry measure, that maximizes
the entropy. By the results of the next section, a Markov measure maximizes
the entropy if and only if the metric entropy with respect to the measure is
the same as the topological entropy of the underlying subshift.
Let B be a primitive matrix of zeros and ones. Let λ be the largest positive
eigenvalue of B, and let q be a positive left eigenvector of B with eigenvalue
λ. Let v be a positive right eigenvector of B with eigenvalue λ normalized so
that q, v = 1. Let V be the diagonal matrix whose diagonal entries are the
coordinates of v, i.e., Vi j = δi j v j . Then A = λ−1 V −1 BV is a stochastic matrix:
all elements of Aare positive, and the rows sum to 1. The elements of A are
Ai j = λ−1 vi−1 Bi j v j . Let p = qV = (q1 v1 , . . . , qn vn ). Then p is a positive left
n
eigenvector of A with eigenvalue 1, and i=1 pi = q, v = 1.
The Markov measure P = PA, p is called the Shannon–Parry measure for
the subshift σ A. Recall that while P is defined on the full shift space
,
its support is the subspace
A. Thus h P (σ A) = h P (σ ). Using the properties
9.5. Variational Principle 221

qB = λq, q, v = 1, and Bi j log Bi j = 0, we have

h P (σ ) = − pi Ai j log Ai j
i, j

 
=− qi vi λ−1 vi−1 Bi j v j log λ−1 vi−1 Bi j v j
i, j

 
=− λ−1 qi v j Bi j log λ−1 vi−1 Bi j v j
i, j



= λ−1 qi v j Bi j log λ + λ−1 qi v j Bi j (log vi − log Bi j v j )
i, j i, j



= log λ + q j v j log v j − λ−1 qi v j Bi j log v j
j i, j


−1
− λ qi v j Bi j log Bi j
i, j



= log λ + v j q j log v j − vi qi log vi = log λ.
j i

Thus h P (σ A) = log λ, which is the topological entropy of σ A (Proposition

3.4.1).

Toral Automorphisms. We consider only the two-dimensional case. Let

A : T2 → T2 be a hyperbolic toral automorphism. The Markov partition con-
structed in §5.12 gives a (measurable) semiconjugacy φ:
A → T2 between
a subshift of finite type and A. Since the image of the Lebesgue measure
under φ ∗ is the Parry measure, the metric entropy of A (with respect to
the Lebesgue measure) is the logarithm of the largest eigenvalue of A
(Exercise 9.4.1).
Exercise 9.4.1. Let A be a hyperbolic toral automorphism. Prove that the
image of the Lebesgue measure on T2 under the semiconjugacy φ is the Parry
measure, and calculate the metric entropy of A.

9.5 Variational Principle1

In this section, we establish the variational principle for metric entropy
[Din71], [Goo69], which asserts that for a homeomorphism of a compact
metric space, the topological entropy is the supremum of the metric en-
tropies for all invariant probability measures.

1 The proof of the variational principle below follows the argument of M. Misiurewicz [Mis76];
see also [KH95] and [Pet89].
222 9. Measure-Theoretic Entropy

Let f be a homeomorphism of a compact metric space X, and M the

space of Borel probability measures on X.

LEMMA 9.5.1. Let µ, ν ∈ M and t ∈ (0, 1). Then for any measurable parti-
tion of ξ of X,

t Hµ (ξ ) + (1 − t)Hν (ξ ) ≤ Htµ+(1−t)ν (ξ ).

Proof. The proof is a straightforward consequence of the convexity of

x log x (Exercise 9.5.1). 

For a partition ξ = {A1 , . . . , Ak}, define the boundary of ξ to be the set
k
∂ξ = i=1 ∂ Ai , where ∂ A = Ā∩ X − A.

LEMMA 9.5.2. Let µ ∈ M. Then:

1. for any x ∈ X and δ > 0, there is δ  ∈ (0, δ) such that µ(∂ B(x, δ  )) = 0;
2. for any δ > 0, there is a finite measurable partition ξ = {C1 , . . . , Ck}
with diam(Ci ) < δ for all i and µ(∂ξ ) = 0;
3. if {µn } ⊂ M is a sequence of probability measures that converges to µ
in the weak∗ topology, and A is a measurable set with µ(∂ A) = 0, then
µ(A) = limn→∞ µn (A).

Proof. Let S(x, δ) = {y ∈ X: d(x, y) = δ}. Then B(x, δ) = 0≤δ <δ S(x, δ  ).
This is an uncountable union, so at least one of these must have measure 0.
Since ∂ B(x, δ) ⊂ S(x, δ), statement 1 follows.
To prove statement 2, let {B1 , . . . , Bk} be an open cover by balls of radius

less than δ/2 and µ(∂ Bi ) = 0. Let C1 = B̄1 , C2 = B2 \B1 , Ci = Bi \ i−1 j=1 B j .
 k
Then ξ = {C1 , . . . , Ck} is a partition, and ∂ξ = ∂Ci ⊂ i=1 ∂ Bi .
To prove statement 3, let A be a measurable set with µ(∂ A) = 0. Since
X is a normal topological space, there is a sequence { fk} of non-negative
continuous functions on X such that fk  χ Ā. Then, for fixed k,

lim µn (A) ≤ lim µn ( Ā) ≤ lim µn ( fk) = µ( fk).

n→∞ n→∞ n→∞

Taking the limit as k → ∞, we obtain

lim µn (A) ≤ lim µ( fk) = µ( Ā) = µ(A).

n→∞ k→∞

Similarly,

lim µn (X\A) ≤ µ(X\A),

n→∞

from which the result follows. 


Let |E| denote the cardinality of a finite set E.
9.5. Variational Principle 223


LEMMA 9.5.3. Let En be an (n, )-separated set, νn = (1/|En |) x∈En δx , and
n−1
µn = n1 i=0 f∗i νn . If µ is any weak∗ accumulation point of {µn }n∈N , then µ is
f -invariant and
1
lim log |En | ≤ hµ ( f ).
n→∞ n

Proof. Let µ be an accumulation point of {µn }n∈N . Then µ is clearly

f -invariant.
Let ξ be a measurable partition with elements of diameter less than and
µ(∂ξ ) = 0. If C ∈ ξ n , then νn (C) = 0 or 1/|En |, since C contains at most one
element of En . Thus Hνn (ξ n ) = log |En |.
Fix 0 < q < n, and 0 ≤ k < q. Let a(k) = [ n−k q
].
Let S = {k + rq + i : 0 ≤ r < a(k), 0 ≤ i < q}, and let T be the comple-
ment of S in {0, 1, . . . , n − 1}. The cardinality of T is at most k + q − 1 ≤ 2q.
Since
a(k)−1   

n−1
ξn = f −i ξ = f −rq−kξ q ∨ f −i ξ ,
i=0 r =0 i∈T

it follows that


a(k)−1
 
 
log |En | = Hνn (ξ n ) ≤ Hνn f −(rq+k) ξ q + Hνn f −i ξ
r =0 i∈T

a(k)−1
≤ Hf∗rq+kνn (ξ q ) + 2q log |ξ |.
r =0

Summing over k and using Lemma 9.5.1, we get

 
1


1
q−1 q−1 a(k)−1
q 2q2
log |En | = Hνn (ξ ) ≤
n
Hf∗rq+kνn (ξ ) +
q
log |ξ |
n n k=0 k=0 r =0
n n
2q2
≤ Hµn (ξ q ) + log |ξ |
n
Thus, by Lemma 9.5.2(3), for fixed q,
1 1 1
lim log |En | ≤ lim Hµn (ξ q ) = Hµ (ξ q ).
n→∞ n n→∞ q q

Letting q → ∞, we get limn→∞ 1

n
log |En | ≤ hµ ( f, ξ ). 

THEOREM 9.5.4 (Variational Principle). Let f be a homeomorphism of a
compact metric space (X, d). Then htop ( f ) = sup{hµ ( f )|µ ∈ M f }.
224 9. Measure-Theoretic Entropy

Proof. Lemma 9.5.3 shows that htop ( f ) ≤ supµ∈M f hµ ( f ), so we need only

demonstrate the opposite inequality.
Let µ ∈ M f be an f -invariant Borel probability measure on X, and
ξ = {C1 , . . . , Ck} a measurable partition of X. By the regularity of µ and
Lemma 9.3.3, we may choose compact sets Bi ⊂ Ci so that the partition
k
β = {B0 = X\ i=1 Bi , B1 , . . . , Bk} satisfies H(ξ |β) < 1. Thus
hµ ( f, ξ ) ≤ hµ ( f, β) + Hµ (ξ |β) ≤ hµ ( f, β) + 1.
The collection B = {B0 ∪ B1 , . . . , B0 ∪ Bk} is a covering of X by open
sets. Moreover, |β n | ≤ 2n |B n |, since each element of Bn intersects at most
two elements of β. Thus
Hµ (β n ) ≤ log |β n | ≤ n log 2 + log |Bn |
Let δ0 be the Lebesgue number of B, i.e., the supremum of all δ such that for
all x ∈ X, B(x, δ) is contained in some B0 ∪ Bi . Then δ0 is also the Lebesgue
number of B n with respect to the metric dn .
No subcollection of B covers X, and the same is true of B n . Thus each ele-
ment C ∈ B contains a point xC that is not contained in any other element, so
B(xC , δ0 , n) ⊂ C. If follows that the collection of all xC is an (n, δ0 )-separated
set. Thus sep(n, δ0 , f ) ≥ |B n |, from which it follows that
1 1
h( f, δ0 ) = lim log(sep(n, δ0 , f )) ≥ lim log |B n |
n→∞ n n→∞ n

1 1
≥ lim (log |β n | − n log 2) ≥ lim Hµ (β n ) − log 2
n→∞ n n→∞ n

= hµ ( f, β) − log 2 ≥ hµ ( f, ξ ) − log 2 − 1
We conclude that hµ ( f ) = hµ ( f n )/n ≤ n1 (htop ( f n ) + log 2 + 1) for all
n > 0. Letting n → ∞, we see that hµ ( f ) ≤ htop ( f ) for all µ ∈ M, which
proves the theorem. 

Exercise 9.5.1. Prove Lemma 9.5.1.
Exercise 9.5.2. Let f be an expansive map of a compact metric space with
expansiveness constant δ0 . Show that f has a measure of maximal entropy,
i.e., there is µ ∈ M f such that hµ ( f ) = htop ( f ). (Hint: Start with a measure
supported on an (n, )-separated set, where ≤ δ0 .)
Bibliography

[AF91] Roy Adler and Leopold Flatto. Geodesic flows, interval maps, and symbolic
dynamics. Bull. Amer. Math. Soc. (N.S.), 25(2):229–334, 1991.
[Ahl73] Lars V. Ahlfors. Conformal invariants: topics in geometric function theory.
McGraw-Hill Series in Higher Mathematics. McGraw-Hill Book Co., New
York, 1973.
[Ano67] D. V. Anosov. Tangential fields of transversal foliations in y-systems. Math.
Notes, 2:818–823, 1967.
[Ano69] D. V. Anosov. Geodesic flows on closed Riemann manifolds with negative cur-
vature. American Mathematical Society, Providence, RI, 1969.
[Arc70] Ralph G. Archibald. An introduction to the theory of numbers. Charles E.
Merrill Publishing Co., Columbus, OH, 1970.
[AS67] D. V. Anosov and Ya. G. Sinai. Some smooth ergodic systems. Russian Math.
Surveys, 22(5):103–168, 1967.
[AW67] R. L. Adler and B. Weiss. Entropy, a complete metric invariant for automor-
phisms of the torus. Proc. Nat. Acad. Sci. U.S.A., 57:1573–1576, 1967.
[BC91] M. Benedicks and L. Carleson. The dynamics of the Hénon map. Ann. of Math.
(2), 133:73–169, 1991.
[Bea91] Alan F. Beardon. Iteration of rational functions. Springer-Verlag, New York,
1991.
[Ber96] Vitaly Bergelson. Ergodic Ramsey theory – an update. In Ergodic theory of Zd
actions (Warwick, 1993–1994), pages 1–61. Cambridge Univ. Press, Cambridge,
1996.
[Ber00] Vitaly Bergelson. Ergodic theory and Diophantine problems. In Topics in sym-
bolic dynamics and applications (Temuco, 1997), pages 167–205. Cambridge
Univ. Press, Cambridge, 2000.
[BG91] Carlos A. Berenstein and Roger Gay. Complex variables. Springer-Verlag,
New York, 1991.
[Bil65] Patrick Billingsley. Ergodic theory and information. John Wiley & Sons Inc.,
New York, 1965.
[BL70] R. Bowen and O. E. Lanford, III. Zeta functions of restrictions of the shift
transformation. In Global Analysis (Proc. Sympos. Pure Math., Vol. XIV,
Berkeley, CA, 1968), pages 43–49. American Mathematical Society, Prov-
idence, RI, 1970.

225
226 Bibliography

[Bow70] Rufus Bowen. Markov partitions for Axiom a diffeomorphisms. Amer. J.

Math., 92:725–747, 1970.
[Boy93] Mike Boyle. Symbolic dynamics and matrices. In Combinatorial and graph-
theoretical problems in linear algebra (Minneapolis, MN, 1991), IMA Vol.
Math. Appl., volume 50, pages 1–38. Springer-Verlag, New York, 1993.
[BP94] Abraham Berman and Robert J. Plemmons. Nonnegative matrices in the math-
ematical sciences. Society for Industrial and Applied Mathematics (SIAM),
Philadelphia, PA, 1994.
[BP98] Sergey Brin and Lawrence Page. The anatomy of a large-scale hypertextual
web search engine. In Seventh International World Wide Web Conference
(Brisbane, Australia, 1998). http://www7.scu.edu.au/programme/fullpapers/
1921/com1921.htm, 1998.
[CE80] Pierre Collet and Jean-Pierre Eckmann. Iterated maps on the interval as dy-
namical systems. Birkhäuser, Boston, MA, 1980.
[CFS82] I. P. Cornfeld, S. V. Fomin, and Ya. G. Sinaı̆. Ergodic theory, volume 245 of
Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, New York,
1982.
[CG93] Lennart Carleson and Theodore W. Gamelin. Complex dynamics. Universi-
text: Tracts in Mathematics. Springer-Verlag, New York, 1993.
[CH82] Shui Nee Chow and Jack K. Hale. Methods of bifurcation theory. Springer-
Verlag, New York, 1982.
[Con95] John B. Conway. Functions of one complex variable. II. Springer-Verlag, New
York, 1995.
[Den26] Arnaud Denjoy. Sur l’itération des fonctions analytique. C. R. Acad. Sci. Paris
Sér. A-B, 182:255–257, 1926.
[Dev89] Robert L. Devaney. An introduction to chaotic dynamical systems. Addison-
Wesley, Redwood City, Calif., 1989.
[DH82] Adrien Douady and John Hamal Hubbard. Itération des polynômes quadra-
tiques complexes. C. R. Acad. Sci. Paris Sér. I Math., 294(3):123–126, 1982.
[Din71] Efim I. Dinaburg. On the relation among various entropy characterizatistics
of dynamical systems. Math. USSR, Izvestia, 5:337–378, 1971.
[dMvS93] Welington de Melo and Sebastian van Strien. One-dimensional dynamics.
Springer-Verlag, Berlin, 1993.
[Dou83] Adrien Douady. Systèmes dynamiques holomorphes. In Bourbaki seminar,
Vol. 1982/83, pages 39–63. Soc. Math. France, Paris, 1983.
[DS88] Nelson Dunford and Jacob T. Schwartz. Linear operators. Part II. John Wiley
& Sons Inc., New York, 1988.
[Fei79] Mitchell J. Feigenbaum. The universal metric properties of non-linear trans-
formations. J. Statist. Phys., 21(6):669–706, 1979.
[Fol95] Gerald B. Folland. A course in abstract harmonic analysis. CRC Press, Boca
Raton, FL, 1995.
[Fri70] Nathaniel A. Friedman. Introduction to ergodic theory. Van Nostrand
Reinhold Mathematical Studies, No. 29. Van Nostrand Reinhold Co., New
York, 1970.
[Fur63] H. Furstenberg. The structure of distal flows. Amer. J. Math., 85:477–515, 1963.
[Fur77] Harry Furstenberg. Ergodic behavior of diagonal measures and a theorem of
Szemerédi on arithmetic progressions. J. Analyse Math., 31:204–256, 1977.
Bibliography 227

[Fur81a] H. Furstenberg. Recurrence in ergodic theory and combinatorial number the-

ory. M. B. Porter Lectures. Princeton University Press, Princeton, NJ, 1981.
[Fur81b] Harry Furstenberg. Poincaré recurrence and number theory. Bull. Amer. Math.
Soc. (N.S.), 5(3):211–234, 1981.
[FW78] H. Furstenberg and B. Weiss. Topological dynamics and combinatorial number
theory. J. Analyse Math., 34:61–85 (1979), 1978.
[Gan59] F. R. Gantmacher. The theory of matrices. Vols. 1, 2. Chelsea Publishing Co.,
New York, 1959. Translated by K. A. Hirsch.
[GG73] M. Golubitsky and V. Guillemin. Stable mappings and their singularities. Grad-
uate Texts in Mathematics, Vol. 14. Springer-Verlag, New York, 1973.
[GH55] W. Gottschalk and G. Hedlund. Topological Dynamics. A.M.S. Colloquim
Publications, volume XXXVI. Amer. Math. Soc., Providence, RI, 1955.
[GLR72] R. L. Graham, K. Leeb, and B. L. Rothschild. Ramsey’s theorem for a class of
categories. Advances in Math., 8:417–433, 1972.
[GLR73] R. L. Graham, K. Leeb, and B. L. Rothschild. Errata: Ramsey’s theorem for
a class of categories. Advances in Math., 10:326–327, 1973.
[Goo69] L. Wayne Goodwyn. Topological entropy bounds measure-theoretic entropy.
Proc. Amer. Math. Soc., 23:679–688, 1969.
[Hal44] Paul R. Halmos. In general a measure preserving transformation is mixing.
Ann. of Math. (2), 45:786–792, 1944.
[Hal50] Paul R. Halmos. Measure theory. D. Van Nostrand Co., Princeton, NJ, 1950.
[Hal60] Paul R. Halmos. Lectures on ergodic theory. Chelsea Publishing Co., New
York, 1960.
[Hel95] Henry Helson. Harmonic analysis. Henry Helson, Berkeley, CA, second
edition, 1995.
[Hén76] M. Hénon. A two-dimensional mapping with a strange attractor. Comm. Math.
Phys., 50:69–77, 1976.
[Hir94] Morris W. Hirsch. Differential topology. Springer-Verlag, New York, 1994.
Corrected reprint of the 1976 original.
[HK91] Jack K. Hale and Hüseyin Koçak. Dynamics and bifurcations. Springer-Verlag,
New York, 1991.
[HW79] G. H. Hardy and E. M. Wright. An introduction to the theory of numbers. The
Clarendon Press, Oxford University Press, New York, fifth edition, 1979.
[Kat72] A. B. Katok. Dynamical systems with hyperbolic structure, pages 125–211,
1972. Three papers on smooth dynamical systems. Translations of the AMS
(series 2), volume 11b, AMS, Providence, RI, 1981.
[KH95] Anatole Katok and Boris Hasselblatt. Introduction to the modern theory of
dynamical systems. Encyclopedia of Mathematics and Its Applications, vol-
ume 54. Cambridge University Press, Cambridge, 1995. With a supplementary
chapter by Katok and Leonardo Mendoza.
[Kin90] J. L. King. A map with topological minimal self-joinings in the sense of Del
Junco. Ergodic Theory Dynamical Systems, 10(4):745–761, 1990.
[Kol58] A. N. Kolmogorov. A new metric invariant of transient dynamical systems and
automorphisms in Lebesgue spaces. Dokl. Akad. Nauk SSSR (N.S.), 119:861–
864, 1958.
[Kol59] A. N. Kolmogorov. Entropy per unit time as a metric invariant of automor-
phisms. Dokl. Akad. Nauk SSSR, 124:754–755, 1959.
228 Bibliography

[KR99] K. H. Kim and F. W. Roush. The Williams conjecture is false for irreducible
subshifts. Ann. of Math. (2), 149(2):545–558, 1999.
[Kre85] Ulrich Krengel. Ergodic theorems. Walter de Gruyter & Co., Berlin, 1985.
With a supplement by Antoine Brunel.
[KvN32] B. O. Koopman and J. von Neumann. Dynamical systems of continuous spectra.
Proc. Nat. Acad. Sci. U.S.A., 18:255–263, 1932.
[Lan84] Oscar E. Lanford, III. A shorter proof of the existence of the Feigenbaum
fixed point. Comm. Math. Phys., 96(4):521–538, 1984.
[LM95] Douglas Lind and Brian Marcus. An introduction to symbolic dynamics and
coding. Cambridge University Press, Cambridge, 1995.
[Lor63] E. N. Lorenz. Deterministic non-periodic flow. J. Atmos. Sci., 20:130–141, 1963.
[LY75] Tien Yien Li and James A. Yorke. Period three implies chaos. Amer. Math.
Monthly, 82(10):985–992, 1975.
[Mañ88] Ricardo Mañé. A proof of the C 1 stability conjecture. Inst. Hautes Études Sci.
Publ. Math., 66:161–210, 1988.
[Mat68] John N. Mather. Characterization of Anosov diffeomorphisms. Nederl. Akad.
Wetensch. Proc. Ser. A 71: Indag. Math., 30:479–483, 1968.
[Mil65] J. Milnor. Topology from the differentiable viewpoint. University Press of
Virginia, Charlottesville, VA, 1965.
[Mis76] Michal Misiurewicz. A short proof of the variational principle for a
n+ action
on a compact space. In International Conference on Dynamical Systems in
Mathematical Physics (Rennes, 1975), pages 147–157. Astérisque, No. 40. Soc.
Math. France, Paris, 1976.
[Mon27] P. Montel. Leçons sur les families normales de fonctions analytiques et leurs
applications. Gauthier-Villars, Paris, 1927.
[Moo66] C. C. Moore. Ergodicity of flows on homogeneous spaces. Amer. J. Math.,
88:154–178, 1966.
[MRS95] Brian Marcus, Ron M. Roth, and Paul H. Siegel. Modulation codes for digital
data storage. In Different aspects of coding theory (San Francisco, CA, 1995),
pages 41–94. American Mathematical Society, Providence, RI, 1995.
[MT88] John Milnor and William Thurston. On iterated maps of the interval. In Dy-
namical systems (College Park, MD, 1986–87), Lecture Notes in Mathematics,
volume 1342, pages 465–563. Springer-Verlag, Berlin, 1988.
[PdM82] Jacob Palis, Jr., and Welington de Melo. Geometric theory of dynamical sys-
tems. Springer-Verlag, New York, 1982. An introduction.
[Pes77] Ya. Pesin. Characteristic Lyapunov exponents and smooth ergodic theory.
Russian Math. Surveys, 32:4:55–114, 1977.
[Pet89] Karl Petersen. Ergodic theory. Cambridge University Press, Cambridge, 1989.
[Pon66] L. S. Pontryagin. Topological groups. Gordon and Breach Science Publishers,
Inc., New York, 1966. Translated from the second Russian edition by Arlen
Brown.
[Que87] Martine Queffélec. Substitution dynamical systems – spectral analysis. Lecture
Notes in Mathematics, volume 1294. Springer-Verlag, Berlin, 1987.
[Rob71] J. W. Robbin. A structural stability theorem. Ann. of Math. (2), 94:447–493,
1971.
[Rob76] Clark Robinson. Structural stability of C 1 diffeomorphisms. J. Differential
Equations, 22(1):28–73, 1976.
Bibliography 229

[Rob95] Clark Robinson. Dynamical systems. Studies in Advanced Mathematics. CRC

Press, Boca Raton, Fla., 1995.
[Roh48] V. A. Rohlin. A “general” measure-preserving transformation is not mixing.
Dokl. Akad. Nauk SSSR (N.S.), 60:349–351, 1948.
[Rok67] V. Rokhlin. Lectures on the entropy theory of measure preserving transfor-
mations. Russian Math. Surveys, 22:1–52, 1967.
[Roy88] H. L. Royden. Real analysis. Macmillan, third edition, 1988.
[Rud87] Walter Rudin. Real and complex analysis. McGraw-Hill Book Co., New York,
third edition, 1987.
[Rud91] Walter Rudin. Functional analysis. McGraw-Hill Book Co. Inc., New York,
second edition, 1991.
[Rue89] David Ruelle. Elements of differentiable dynamics and bifurcation theory.
Academic Press Inc., Boston, Mass., 1989.
[Sár78] A. Sárközy. On difference sets of sequences of integers. III. Acta Math. Acad.
Sci. Hungar., 31(3–4):355–386, 1978.
[Sha64] A. N. Sharkovskii. Co-existence of cycles of a continuous mapping of the line
into itself. Ukrainian Math. J., 16:61–71, 1964.
[Shi87] Mitsuhiro Shishikura. On the quasiconformal surgery of rational functions.
Ann. Sci. École Norm. Sup. (4), 20(1):1–29, 1987.
[Sin59] Ya. G. Sinai. On the concept of entropy of a dynamical system. Dokl. Akad.
Nauk SSSR (N.S.), 124:768–771, 1959.
[Sma67] S. Smale. Differentiable dynamical systems. Bull. Amer. Math. Soc., 73:747–
817, 1967.
[Sul85] Dennis Sullivan. Quasiconformal homeomorphisms and dynamics. I. Solu-
tion of the Fatou–Julia problem on wandering domains. Ann. of Math. (2),
122(3):401–418, 1985.
[SW49] Claude E. Shannon and Warren Weaver. The Mathematical Theory of Com-
munication. The University of Illinois Press, Urbana, IL, 1949.
[Sze69] E. Szemerédi. On sets of integers containing no four elements in arithmetic
progression. Acta Math. Acad. Sci. Hungar., 20:89–104, 1969.
[vS88] Sebastian van Strien. Smooth dynamics on the interval (with an emphasis on
quadratic-like maps). In New directions in dynamical systems, pages 57–119.
Cambridge Univ. Press, Cambridge, 1988.
[Wal75] Peter Walters. Ergodic theory – Introductory Lectures. Lecture Notes in Math-
ematics, volume 458. Springer-Verlag, Berlin, 1975.
[Wei73] Benjamin Weiss. Subshifts of finite type and sofic systems. Monatsh. Math.,
77:462–474, 1973.
[Wil73] R. F. Williams. Classification of subshifts of finite type. Ann. of Math. (2),
98:120–153, 1973. Errata, ibid., 99 (1974), 380–381.
[Wil84] R. F. Williams. Lorenz knots are prime. Ergodic Theory Dynam. Systems,
4(1):147–163, 1984.
[Wol26] Julius Wolff. Sur l’itération des fonctions bornées. C. R. Acad. Sci. Paris Sér.
A-B, 182:200–201, 1926.
This Page Intentionally Left Blank
Index

(X, A, µ), 70 C k curve, 138

[x, y], 128 C k topology, 138
1-step SFT, 56 C k(M, N), 138

, 73 C0 (X, C), 71
1 , 194 Diff1 (M), 117
⊥, 209 d(ξ, η), 209
≺, 162, 172 dn , 37
∼, 3 Es ,Eu , 108
∨, 208 Em, 5
∧, 208 F(R), 193
α-limit point, 29 f -covers, 163
α-limit set, 29 G-extension, 88
-dense, 4 H(ξ |η), 211
-orbit, 107, 110 H(ζ ), 209
ζ f (z), 60 He (U), 98
ζ n , 213 Hw (U), 98
θ-transversality, 147 h(T) = hµ (T), 215
s , u , 116 h(T, ζ ), 213
ν, 209 h( f ), 37
ν j , 173 I(C), 210
ρ(ξ, η), 212 I → J , 163
ρ( f ), 154 i(x), 170
eA, 56 J (R), 193
vA, 56 k-step SFT, 56
m , 7 Lp (X, µ), 71
m +, 7 l-modal map, 170
σ-algebra, 69 M, 85
σ , 5, 7 MT , 85
ω-limit point, 28 N, 1
ω-limit set, 28 N0 , 1
ω f (x), 29 NW( f ), 29, 114
BA(ζ ), 192 n-leader, 81
B(x, r ), 28 O f (x), 2
C, 1 P-graph, 165
C(X), 85 Q, 1
CL(R), 204 qµ , 9

231
232 Index

R( f ), 29 backward invariant, set, 2

Rα , 4 base-m expansion, 5
R, 1 basic set, 133
S1 , 4 basin of attraction, 25, 173, 192
S∞ f , 87 immediate, 192
S nf , 85, 87 Bernoulli
Sf (x), 178 automorphism, 79, 209
supp, 72, 90 measure, 79, 219
UT , 80 bifurcation, 183
Var( f ), 160 codimension-one, 184
Ws (x), Wu (x), 118, 122 flip, 185
Ws (x), Wu (x), 121 fold, 185
Z, 1 generic, 183
period-doubling, 185
a.e., 71 saddle–node, 185
abelian group, 87 Birkhoff ergodic theorem, 82, 93
absolutely continuous block code, 55, 67
foliation, 144 Borel
measure, 86 σ-algebra, 71
adapted metric, 109 measure, 71, 72
address of a point, 170 boundary of a partition, 222
adjacency matrix, 8, 56 bounded distortion, 181
Adler, 135 bounded Jacobian, 145
affine subspace, 51 bounded variation, 160
allowed word, 8 Bowen, 61, 137
almost every, 71 branch number, 192
almost periodic point, 30 branch point, 192
almost periodic set, 47 Breiman, 217
alphabet, 7, 54, 55
analytic function, 191 Cantor set, 7, 16, 18, 25, 182, 196,
Anosov diffeomorphism, 108, 130, 141, 142 204
topological entropy of, 116 ceiling function, 21
Anosov’s shadowing theorem, 113 chain, Markov, see Markov, chain
aperiodic transformation, 96 chaos, 23
arithmetic progression, 103 chaotic behavior, 23
atlas, 138 chaotic dynamical system, 23
atom, 71 character, 95
attracting periodic point, 192 circle endomorphism, 23, 35, 36
attracting point, 9, 25, 173 circle rotation, 4, 35, 45, 95
attractor, 18, 25 ergodicity of, 77
Hénon, 25 irrational, 4, 30, 77, 90
hyperbolic, 18 is uniquely ergodic, 87
Lorenz, 25 clock drift, 67
strange, 25 code, 55
automorphism color, 51
Bernoulli, 79, 209 combinatorial number theory, 48
group, 5 common refinement, 208
of a measure space, 70 completion of a σ-algebra, 70
average composition matrix, 65
space, 80, 84 conditional density, 144
time, 80, 84 conditional entropy, 211
Axiom A, 133 conditional measure, 92
Index 233

cone differentiable
stable, 115 flow, 107
unstable, 115 manifold, 106, 137
conformal map, 153 map, 107
conjugacy, 3 vector field, 107
measure-theoretic, 70 differential equation, 19
topological, 28, 39, 55 Diffk(M), 138
conjugate, 3 Dirac measure, 86
constant-length substitution, 64 direct product, 3
continued fraction, 12, 91 directed graph, 8, 56, 67
continuous (semi)flow, 28 discrete spectrum, 94–97
continuous spectrum, 95, 97 discrete-time dynamical system, 1
continuous-time dynamical system, 2 dist, 140
convergence in density, 99 distal
convergent, 91 extension, 46
coordinate chart, 137 homeomorphism, 45
coordinate system, local, 106 points, 45
covering map, 153 distribution, 139
cov(n, , f ), 37 Hölder-continuous, 143
critical point, 192 integrable, 139
critical value, 192 stable, 108
cross-ratio, 180 unstable, 108
cross-section, of a flow, 22 Douady, 205
Curtis, 55 Douady–Hubbard theorem,
cutting and stacking, 96 206
Cvitanović–Feigenbaum equation, doubling operator, 169
189 dual group, 95
cylinder, 54 dynamical system, 2
cylinder set, 7, 78 chaotic, 23
continuous-time, 2
d-lim, 99 deterministic, 23
deg(R), 193 differentiable, 106
degree, of a rational map, 193 discrete-time, 1
dendrite, 196 ergodic, 4
Denjoy, 205 hyperbolic, 22, 106
example, 161 minimal, 4, 29
theorem, 160 symbolic, 54
dense topological, 28
forward orbit, 32 topologically mixing, 33
full orbit, 32 topologically transitive, 31
orbit, 23, 32 dynamics
dense, -, 4 symbolic, 6, 54
derivative transformation, 72 topological, 28
deterministic dynamical system,
23 edge shift, 57
diameter, 37 elementary strong shift equivalence,
of a partition, 218 62
diffeomorphism, 106 embedded submanifold, 139
Anosov, 130, 141 embedding of a subshift, 55
Axiom A, 133 endomorphism
group, 138 expanding, 5
of the circle, 160 group, 5
234 Index

entropy Fatou, 200

conditional, 211 Fatou set, 193
horseshoe, 125 Fatou theorem, 204
measure-theoretic, 208, 213, 214 Feigenbaum constant, 189
metric, 208, 213, 214 Feigenbaum phenomenon, 189
of a partition, 209 fiber, 140
of a transformation, 214 fiber bundle, 140
topological, 36, 37, 116, 125 first integral, 20, 21
of an SFT, 60 first return map, 22, 72
equicontinuous fixed point, 2
extension, 46 Fix( f ), 60
homeomorphism, 45 flip bifurcation, 185
equivalent partitions, 208 flow, 2, 19
ergodic continuous, 28
dynamical system, 4 differentiable, 107
hypothesis, 69, 80 entropy of, 41
measure, 86 gradient, 20
theorem, von Neumann, 80 measurable, 70, 74
theory, 69 measure preserving, 70
transformation, 73 under a function, 21
ergodicity, 73, 75, 77 weak mixing, 75
essentially, 71 fold bifurcation, 185
essentially invariant foliation, 139
function, 73 absolutely continuous, 144
set, 73 coordinate chart, 139
even system of Weiss, 66 leaf of, 139
eventually periodic point, 2, 171 stable, 14, 130
exceptional point, 201 transversely absolutely continuous,
exceptional set, 201 145
expanding endomorphism, 5, 30, 34, 107 unstable, 14, 130
is mixing, 77 forbidden word, 8
expanding map, 5 forward invariant set, 2
entropy of, 218 Fourier series, 78
expansive fractal, 191, 193, 202
homeomorphism, 35, 40 Frobenius theorem, 59
map, 35 full m-shift, 55
positively, 35 full measure, 69
expansiveness constant, 35 full one-sided shift, 7, 33
exponent full shift, 35, 36
Hölder, 143 full two-sided shift, 7, 33
Lyapunov, 23 function
extension, 3, 140 analytic, 191
distal, 46 ceiling, 21
equicontinuous, 46 essentially invariant, 73
group, 88 generating, 60
isometric, 41, 46 Hamiltonian, 21
measure-theoretic, 70 Lyapunov, 20
extreme point, 86 meromorphic, 191
strictly invariant, 74
factor, 3, 55, 140 zeta, 60
factor map, 3 Furstenberg, 49, 101–103
Index 235

Gauss, 11 homoclinic point, 125

measure, 12–13, 90–94 Hopf, 151
transformation, 11–13, 90–94 Hopf’s argument, 141
generalized eigenspaces, 42 horseshoe, 34–36, 108, 124, 125, 128
generating function, 60 entropy of, 125
generating sequence of partitions, linear, 15, 16
215 non-linear, 124
generator Hubbard, 205
one-sided, 216 hyperbolic
two-sided, 216 attractor, 18
generic bifurcation, 183 dynamical system, 15, 106
generic point, 88 fixed point, 110
global stable manifold, 122 periodic point, 110
global unstable manifold, 122 set, 108
Google,TM 103 hyperbolic toral automorphism, 14, 22, 33,
gradient flow, 20 35, 36, 41, 108, 130, 135
graph transform, 119 is mixing, 77
group metric entropy of, 221
automorphism, 5 Hyperbolicity, 106, 108
dual, 95
endomorphism, 5, 14 immediate basin of attraction, 192
extension, 88 Inclination lemma, 123
of characters, 95 independent partitions, 209
topological, 4, 95 induced transformation, 73
translation, 4, 35, 45, 87 information, 210
integrable distribution, 139
Haar measure, 88 integrable foliations, 151
Hadamard, 54 integral hull, 151
Hadamard–Perron theorem, 118 Internet, 103
Halmos, 97 intersection of partitions, 208
Halmos–von Neumann theorem, 96 intersymbol interference, 67
Hamiltonian invariant
function, 21 kneading, 173
system, 20 measure, 70, 85
Hedlund, 55 set, 2
Hénon attractor, 25 IP-system, 49
heteroclinic point, 128 irrational circle rotation, 4, 33, 77, 87,
heteroclinically related, 133 90
higher block presentation, 56, 67 irrationally neutral periodic point, 192
Hölder irreducible
constant, 143 matrix, 57
continuity, 142, 143 substitution, 64
exponent, 143 isolated fixed point, 10
holonomy map, 145 isometric extension, 41, 46
homeomorphism, 28 isometry, 28, 34, 36, 114
distal, 45 entropy of, 38
equicontinuous, 45 isomorphism, 3
expansive, 35, 40 measure-theoretic, 70
minimal, 29 of subshifts, 55
of the circle, 153 of topological dynamical systems, 28
pointwise almost periodic, 47 itinerary, 54, 170
236 Index

Jacobian, 145 Mañé, 134

Julia set, 193, 200 manifold
differentiable, 106, 137
Katok, 111 Riemannian, 106, 140
Kim, 63 stable, 14, 118, 122
kneading invariant, 173 unstable, 14, 118, 122
Kolmogorov, 208, 216 map
Kolmogorov–Sinai theorem, 216 covering, 153
Koopman–von Neumann theorem, 99 expanding, 5
Krein–Milman theorem, 86 first return, 72
Krylov–Bogolubov theorem, 85 l-modal, 170
measurable, 70
lag, 63 measure preserving, 70
Lambda lemma, 123 non-singular, 70
Lanford, 61 piecewise-monotone, 170
lap, 170 Poincaré, 22, 72
leader, 81 quadratic, 9, 35, 171, 176, 178, 179, 182, 205
leaf, 139 rational, 191
leaf, local, 139 tent, 176
Lebesgue number, 224 unimodal, 170
Lebesgue space, 71 Markov
left translation, 4 chain, 78, 104, 105
lemma graph, 164, 165
inclination, 123 measure, 57, 78, 104, 105, 219
lambda, 123 partition, 134
Rokhlin–Halmos, 96 matrix
length distortion, 180 adjacency, 56
Li, 164 composition, 65
linear fractional transformation, 180 irreducible, 57
linear functional, 85 non-negative, 57
Lipshitz function, 160 positive, 57
local primitive, 57, 79
coordinate system, 106 reducible, 57
leaf, 139 shift-equivalent, 62
product structure, 129 stochastic, see stochastic matrix
stable manifold, 121 strong shift-equivalent, 62
transversal, 139 maximal almost periodic set, 47
unstable manifold, 121 maximal ergodic theorem, 82
locally maximal McMillan, 217
hyperbolic set, 128 measurable
invariant set, 16 flow, 70, 74
log, 209 map, 70
Lorenz, 26 set, 70
Lorenz attractor, 25 measure, 69
Lyapunov absolutely continuous, 86
exponent, 23 Bernoulli, 79, 219
function, 20 Borel, 72
metric, 109 conditional, 92
Lyndon, 55 Dirac, 86
ergodic, 86
main cardioid, 206 Gauss, see also Gauss measure
Mandelbrot set, 206 Haar, 88
Index 237

invariant, 70, 85 normal family, 193, 197

Markov, 57, 78, see also Markov measure at a point, 193
of maximal entropy, 224 normal number, 84
regular, 71 null set, 69
Shannon–Parry, 220 number theory, 49, 92, 103
smooth, 151
space, 69 omit a point, 197
spectral, 98 one-parameter group, 107
measure-preserving one-sided shift, 35
flow, 70 orbit, 2
transformation, 70 dense, 23, 32
measure-theoretic periodic, 2
entropy, 213, 215 uniformly distributed, 23, 89
extension, 70
isomorphism, 70 Parry, 220
meromorphic function, 191 partition, 208
metric boundary of, 222
adapted, 109 Markov, 134
entropy, 213, 215 of an interval, 164
Lyapunov, 109 partitions, independent, 209
Riemannian, 106, 139 pendulum, 19
space, 28 Per( f ), 114
minimal period, 2
action of a group, 49 period, minimal, 2
dynamical system, 4, 29 period-doubling bifurcation, 185
homeomorphism, 29 periodic
period, 2 orbit, 2
set, 29, 90 point, 2, 162, 198
minimum principle, 179 attracting, 10, 192
mixing, 74, 77 irrationally neutral, 192
transformation, 74 rationally neutral, 192
weak, 75, 100, 152 repelling, 10, 192
Möbius transformation, 180, 192 superattracting, 192
mod 0, 71 Perron theorem, 58
modified frequency modulation, 68 Perron–Frobenius theorem, 57
modular function, 198 piecewise-monotone map, 170
monochromatic, 51 plaque, 139
Montel theorem, 197 Poincaré, 71
Morse sequence, 64 map, 22, 72
multiple recurrence theorem, 103 Poincaré classification theorem,
multiple recurrence, topological, 49 158
multiplier, 192 Poincaré recurrence theorem, 72, 103
pointwise almost periodic homeomorphism,
negative semiorbit, 2 47
Newton method, 196 polynomial recurrence theorem, 102
nilmanifold, 130 Pontryagin duality theorem, 95
non-atomic Lebesgue space, 71 positive semiorbit, 2
non-linear horseshoe, 124 positive upper density, 101
non-negative definite sequence, 97 positively expansive, 35
non-singular map, 70 positively recurrent point, 29
non-singular transformation, 70 postcritical set, 204
non-wandering point, 29 presentation, 67
238 Index

primitive Robbin, 134

matrix, 57, 79, 104 Robinson, 134
substitution, 64, 90 Rokhlin, 97
transformation, 73 Rokhlin metric, 212
probability Rokhlin–Halmos lemma, 96
measure, 70 rotation number, 154
space, 70 rotation, entropy of, 218
product Roush, 63
direct, 3
formula, 60 saddle–node bifurcation, 184
skew, 3 Sárközy, 103
structure, local, 129 Schwarzian derivative, 178
projection, 3, 139, 140 search engine, 103
proximal points, 45 self-similarity, 202
pseudo-orbit, 110 semiconjugacy, 2
topological, 28
quadratic map, 9, 35, 171, 176, 178, 179, 182, semiflow, 2
205 semiorbit
quotient, 91 negative, 2
positive, 2
Radon–Nikodym sensitive dependence on initial conditions,
derivative, 86 23, 117
theorem, 86 sep(n, , f ), 37
Ramsey, 48 separated set, 37
Ramsey theory, 48 sequence of partitions
random behavior, 23 generating, 215
random variable, 79 refining, 215
rational map, 191 SFT, 56
degree of, 193 shadowing, 107, 111, 113
rationally neutral periodic point, 192 Shannon, 220
rectangle, 134 Shannon–McMillan–Breiman theorem,
recurrence 217
measure-theoretic, 71 Shannon–Parry measure, 220
topological, 29, 72 Sharkovsky ordering, 162
recurrent point, 29, 72 Sharkovsky theorem, 163
reducible matrix, 57 shift, 5, 7, 54, 60
refinement, 208 edge, 57
refining sequence of partitions, 215 entropy of, 219
regular measure, 71 equivalence, 62
relatively dense subset, 30 full, 7, 35, 36
repelling periodic point, 192 one-sided, 7
repelling point, 9 two-sided, 7
return time, 22 vertex, 8, 56
Riemann sphere, 191 shift equivalence, strong, 62
Riemannian Shishikura theorem, 200
manifold, 106, 140 signed lexicographic ordering, 172
metric, 106, 139 simply transitive, 180, 192
volume, 140 Sinai, 208, 216
Riesz, 80, 85 Singer theorem, 179
Riesz representation theorem, 85 skew product, 3, 140
right translation, 4 Smale, 133
Index 239

Smale’s spectral decomposition theorem, symmetric difference, 73

133 syndetic subset, 30
smooth measure, 151 Szemerédi, 103
sofic subshift, 66 Szemerédi theorem, 103
solenoid, 18, 35, 36, 44, 45, 108, 128
space average, 84 tangent
space, metric, 28 bundle, 138
span(n, , f ), 37 map, 138
spanning set, 37 space, 138
spectral measure, 98 vector, 138
spectrum tent map, 114, 176
continuous, see continuous spectrum terminal vertex, 9
discrete, see discrete spectrum theorem
stable Anosov’s shadowing, 113
cone, 115 Birkhoff ergodic, 82
distribution, 108 Bowen–Lanford, 61
foliation, 14, 130 Curtis–Lyndon–Hedlund, 55
manifold, 14, 118, 122 Denjoy, 160
local, 121 Douady–Hubbard, 206
set, 151 Fatou, 204
subspace, 108 Frobenius, 59
stationary sequence, 79 Furstenberg–Weiss, 49
stochastic matrix, 58, 78, 104 Hadamard–Perron, 118
stochastic process, 79 Halmos–von Neumann, 96
strange attractor, 25 Kolmogorov–Sinai, 216
strong mixing transformation, 74 Koopman–von Neumann, 99
strong transversality condition, 134 Krein–Milman, 86
structural stability, 108, 117, 131 Krylov–Bogolubov, 85
Structural Stability Theorem, 134 maximal ergodic, 82
structurally stable diffeomorphism, 117 Montel, 197
submanifold, 106, 139 multiple recurrence, 103
embedded, 139 Perron, 58
subshift, 8, 55, 134 Perron–Frobenius, 57
entropy of, 219 Poincaré classification, 158
generating function of, 60 Poincaré recurrence, 72
of finite type, 56 polynomial recurrence, 102
one-sided, 55 Pontryagin duality, 95
sofic, 66 Riesz representation, 85
topological entropy of, 60 Sárközy, 103
two-sided, 55 Shannon–McMillan–Breiman, 217
zeta function of, 61 Sharkovsky, 163
substitution, 64 Shishikura, 200
constant length, 64 Singer, 179
irreducible, 64 Smale’s spectral decomposition, 133
primitive, 64, 90 structural stability, 134
Sullivan, 205 Szemerédi, 103
superattracting periodic point, 192 van der Waerden, 49
support, 72 von Neumann ergodic, 80
suspension, 21 Weyl, 89
symbol, 7, 54 Williams, 62
symbolic dynamics, 6, 54 time average, 84, 88
240 Index

time-t map, 2 uniformly distributed, 89, 90

times-m map, 5 unimodal map, 170
topological uniquely ergodic, 87–90
conjugacy, 28, 39 unitary operator, 80
dynamical system, 28 unstable
dynamics, 28 cone, 115
entropy, 36, 37, 60, 116, 125 distribution, 108
group, 4, 95 foliation, 14, 130
Markov chain, 56 manifold, 14, 118, 122
mixing, 33, 76 local, 121
properties, 28 set, 151
recurrence, 29, 72 subspace, 108
semiconjugacy, 28 upper density, 99
transitivity, 31, 33, 76, 88
topology, C k, 138 van der Waerden’s theorem, 49
toral automorphism, see hyperbolic toral variational principle for metric entropy, 221
automorphism vector
transformation, 70 non-negative, 57
aperiodic, 96 positive, 57
derivative, 72 vector field, 107, 139
entropy of, 215 vertex shift, 8, 56
ergodic, 73 one-sided, 8
Gauss, see Gauss transformation two-sided, 8
induced, 73 vertex, terminal, 9
linear fractional, 180 von Neumann, 80
measure-preserving, 70 von Neumann ergodic theorem, 80
mixing, 74
Möbius, 180, 192
wandering domain, 205
non-singular, 70
wandering interval, 173
primitive, 73
weak mixing, 75, 97, 100, 152
strong mixing, 74
flow, 75
uniquely ergodic, 87
transformation, 75
weak mixing, 75
weak topology, 97
transition probability, 78
weak∗ topology, 85
translation, 90
Weiss, 49, 135
group, 4
even system of, 66
left, 4
Weyl theorem, 89
right, 4
Whitney embedding theorem, 111
translation-invariant metric, 44
Wiener lemma, 98
transversal, 139
Williams, 63
local, 139
Wolff, 205
transverse, 139
word, 7
transverse homoclinic point, 125
allowed, 8, 56
transverse submanifolds, 122
forbidden, 8, 56
transversely absolutely continuous foliation,
145
trapping region, 25 Yorke, 164
turning point, 170
zeta function, 60
uniform convergence, 88 rational, 61