Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
Editors:
J.M. Morel, Cachan
F. Takens, Groningen
B. Teissier, Paris
Rufus Bowen
Equilibrium States
and the Ergodic Theory
of Anosov Diffeomorphisms
Second revised edition
JeanRené Chazottes
Editor
Preface by David Ruelle
ABC
Author
Robert Edward (Rufus) Bowen
(19471978)
University of California at Berkeley
USA
Preface by
David Ruelle
IHÉS
35, Route de Chartres
91440 BuressurYvette
France
ruelle@ihes.fr
Editor
JeanRené Chazottes
Centre de Physique Théorique
CNRSÉcole Polytechnique
91128 Palaiseau Cedex
France
jeanrene@cpht.polytechnique.fr
ISBN: 9783540776055 eISBN: 9783540776956
DOI: 10.1007/9783540776956
Lecture Notes in Mathematics ISSN print edition: 00758434
ISSN electronic edition: 16179692
Library of Congress Control Number: 2008922889
Mathematics Subject Classiﬁcation (2000): Primary: 37D20, Secondary: 37D35
c 2008, 1975 SpringerVerlag Berlin Heidelberg
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, speciﬁcally the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microﬁlm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations are
liable to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a speciﬁc statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general use.
Printed on acidfree paper
9 8 7 6 5 4 3 2 1
springer.com
Cover design: WMXDesign GmbH
Preface
The Greek and Roman gods, supposedly, resented those mortals endowed with
superlative gifts and happiness, and punished them. The life and achievements
of Rufus Bowen (1947–1978) remind us of this belief of the ancients. When
Rufus died unexpectedly, at age thirtyone, from brain hemorrhage, he was a
very happy and successful man. He had great charm, that he did not misuse,
and superlative mathematical talent. His mathematical legacy is important,
and will not be forgotten, but one wonders what he would have achieved if he
had lived longer. Bowen chose to be simple rather than brilliant. This was the
hard choice, especially in a messy subject like smooth dynamics in which he
worked. Simplicity had also been the style of Steve Smale, from whom Bowen
learned dynamical systems theory.
Rufus Bowen has left us a masterpiece of mathematical exposition: the slim
volume Equilibrium States and the Ergodic Theory of Anosov Diﬀeomorphisms
(Springer Lecture Notes in Mathematics 470 (1975)). Here a number of results
which were new at the time are presented in such a clear and lucid style that
Bowen’s monograph immediately became a classic. More than thirty years
later, many new results have been proved in this area, but the volume is as
useful as ever because it remains the best introduction to the basics of the
ergodic theory of hyperbolic systems.
The area discussed by Bowen came into existence through the merging of
two apparently unrelated theories. One theory was equilibrium statistical me
chanics, and speciﬁcally the theory of states of inﬁnite systems (Gibbs states,
equilibrium states, and their relations as discussed by R.L. Dobrushin, O.E.
Lanford, and D. Ruelle). The other theory was that of hyperbolic smooth dy
namical systems, with the major contributions of D.V. Anosov and S. Smale.
The two theories came into contact when Ya.G. Sinai introduced Markov par
titions and symbolic dynamics for Anosov diﬀeomorphisms. This allowed the
poweful techniques and results of statistical mechanics to be applied to smooth
dynamics, an extraordinary development in which Rufus Bowen played a ma
jor role. Some of Bowen’s ideas were as follows. First, only onedimensional
statistical mechanics is discussed: this is a richer theory, which yields what is
VI Preface
needed for applications to dynamical systems, and makes use of the powerful
analytic tool of transfer operators. Second, Smale’s Axiom A dynamical sys
tems are studied rather than the less general Anosov systems. Third, Sinai’s
Markov partitions are reworked to apply to Axiom A systems and their con
struction is simpliﬁed by the use of shadowing. The combination of simpli
ﬁcations and generalizations just outlined led to Bowen’s concise and lucid
monograph. This text has not aged since it was written and its beauty is as
striking as when it was ﬁrst published in 1975.
JeanRen´e Chazottes has had the idea to make Bowen’s monograph more
easily available by retyping it. He has scrupulously respected the original
text and notation, but corrected a number of typos and made a few other
minor corrections, in particular in the bibliography, to improve usefulness
and readability. In his enterprise he has been helped by Jerˆ ome Buzzi, Pierre
Collet, and Gerhard Keller. For this work of love all of them deserve our
warmest thanks.
Bures sur Yvette, mai 2007 David Ruelle
Contents
0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1 Gibbs Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
A. Gibbs Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
B. Ruelle’s PerronFrobenius Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 8
C. Construction of Gibbs Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
D. Variational Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
E. Further Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2 General Thermodynamic Formalism . . . . . . . . . . . . . . . . . . . . . . . 29
A. Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
B. Pressure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
C. Variational Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
D. Equilibrium States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3 Axiom a Diﬀeomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
A. Deﬁnition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
B. Spectral Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
C. Markov Partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
D. Symbolic Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4 Ergodic Theory of Axiom a Diﬀeomorphisms . . . . . . . . . . . . . . 61
A. Equilibrium States for Basic Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
B. The Case φ = φ
(u)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
C. Attractors and Anosov Diﬀeomorphisms . . . . . . . . . . . . . . . . . . . . . . 69
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
VIII Contents
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
These notes came out of a course given at the University of Minnesota and
were revised while the author was on a Sloan Fellowship.
0
Introduction
The main purpose of these notes is to present the ergodic theory of Anosov and
Axiom A diﬀeomorphisms. These diﬀeomorphisms have a complicated orbit
structure that is perhaps best understood by relating them topologically and
measure theoretically to shift spaces. This idea of studying the same example
from diﬀerent viewpoints is of course how the subjects of topological dynamics
and ergodic theory arose from mechanics. Here these subjects return to help
us understand diﬀerentiable systems.
These notes are divided into four chapters. First we study the statistical
properties of Gibbs measures. These measures on shift spaces arise in modern
statistical mechanics; they interest us because they solve the problem of de
termining an invariant measure when you know it approximately in a certain
sense. The Gibbs measures also satisfy a variational principle. This princi
ple is important because it makes no reference to the shift character of the
underlying space. Through this one is led to develop a “thermodynamic for
malism” on compact spaces; this is carried out in chapter two. In the third
chapter Axiom A diﬀeomorphisms are introduced and their symbolic dynam
ics constructed: this states how they are related to shift spaces. In the ﬁnal
chapter this symbolic dynamics is applied to the ergodic theory of Axiom A
diﬀeomorphisms.
The material of these notes is taken from the work of many people. I have
attempted to give the main references at the end of each chapter, but no doubt
some are missing. On the whole these notes owe most to D. Ruelle and Ya.
Sinai.
To start, recall that (X, B, µ) is a probability space if B is a σﬁeld of
subsets of X and µ is a nonnegative measure on B with µ(X) = 1. By an
automorphism we mean a bijection T : X → X for which
(i) E ∈ B iﬀ T
−1
E ∈ B,
(ii) µ(T
−1
E) = µ(E) for E ∈ B.
2 0 Introduction
If T : X → X is a homeomorphism of a compact metric space, a natural
σﬁeld B is the family of Borel sets. A probability measure on this σﬁeld is
called a Borel probability measure. Let M(X) be the set of Borel probability
measures on X and M
T
(X) the subset of invariant ones, i.e. µ ∈ M
T
(X) if
µ(T
−1
E) = µ(E) for all Borel sets E. For any µ ∈ M(X) one can deﬁne
T
∗
µ ∈ M(X) by T
∗
µ(E) = µ(T
−1
E).
Remember that the realvalued continuous functions C(X) on the compact
metric space X form a Banach space under f = max
x∈X
f(x). The weak
∗topology on the space C(X)
∗
of continuous linear functionals α : C(X) →R
is generated by sets of the form
U(f, ε, α
0
) = {α ∈ C(X)
∗
: α(f) −α
0
(f) < ε}
with f ∈ C(X), ε > 0, α
0
∈ C(X)
∗
.
Riesz Representation. For each µ ∈ M(X) deﬁne α
µ
∈ C(X)
∗
by α
µ
(f) =
fdµ. Then µ ↔ α
µ
is a bijection between M(X) and
{α ∈ C(X)
∗
: α(1) = 1 and α(f) ≥ 0 whenever f ≥ 0} .
We identify α
µ
with µ, often writing µ when we mean α(µ). The weak ∗
topology on C(X)
∗
carries over by this identiﬁcation to a topology on M(X)
(called the weak topology).
Proposition. M(X) is a compact convex metrizable space.
Proof. Let {f
n
}
∞
n=1
be a dense subset of C(X). The reader may check that
the weak topology on M(X) is equivalent to the one deﬁned by the metric
d(µ, µ
) =
∞
n=1
2
−n
f
n
−1
f
n
dµ −
f
n
dµ
.
Proposition. M
T
(X) is a nonempty closed subset of M(X).
Proof. Check that T
∗
: M(X) →M(X) is a homeomorphism and note that
M
T
(X) = {µ ∈ M(X) : T
∗
µ = µ}. Pick µ ∈ M(X) and let µ
n
=
1
n
(µ +
T
∗
µ +· · · +(T
∗
)
n−1
µ). Choose a subsequence µ
n
k
converging to µ
∈ M(X).
Then µ
∈ M
T
(X).
Proposition. µ ∈ M
T
(X) if and only if
(f ◦ T) dµ =
fdµ for all f ∈ C(X) .
Proof. This is just what the Riesz Representation Theorem says about the
statement T
∗
µ = µ.
1
Gibbs Measures
A. Gibbs Distribution
Suppose a physical system has possible states 1, . . . , n and the energies of these
states are E
1
, . . . , E
n
. Suppose that this system is put in contact with a much
larger “heat source” which is at temperature T. Energy is thereby allowed to
pass between the original system and the heat source, and the temperature T
of the heat source remains constant as it is so much larger than our system.
As the energy of our system is not ﬁxed any of the states could occur. It is
a physical fact derived in statistical mechanics that the probability p
j
that
state j occurs is given by the Gibbs distribution
p
j
=
e
−βE
j
¸
n
i=1
e
−βE
i
,
where β =
1
kT
and k is a physical constant.
We shall not attempt the physical justiﬁcation for the Gibbs distribution,
but we will state a mathematical fact closely connected to the physical rea
soning.
1.1. Lemma. Let real numbers a
1
, . . . , a
n
be given. Then the quantity
F(p
1
, . . . , p
n
) =
n
¸
i=1
−p
i
log p
i
+
n
¸
i=1
p
i
a
i
has maximum value log
¸
n
i=1
e
a
i
as (p
1
, . . . , p
n
) ranges over the simplex
¦(p
1
, . . . , p
n
) : p
i
≥ 0, p
1
+ + p
n
= 1¦ and that maximum is assumed
only by
p
j
= e
a
j
¸
i
e
a
i
−1
.
4 1 Gibbs Measures
This is proved by calculus. The quantity H(p
1
, . . . , p
n
) =
¸
n
i=1
−p
i
log p
i
is
called the entropy of the distribution (p
1
, . . . , p
n
) (note: ϕ(x) = −xlog x is
continuous on [0, 1] if we set ϕ(0) = 0.) The term
¸
n
i=1
p
i
a
i
is of course
the average value of the function a(i) = a
i
. In the statistical mechanics
case a
i
= −βE
i
, entropy is denoted S and average energy E. The Gibbs
distribution then maximizes
S −βE = S −
1
kT
E,
or equivalently minimizes E − kTS. This is called the free energy. The prin
ciple that “nature minimizes entropy” applies when energy is ﬁxed, but when
energy is not ﬁxed “nature minimizes free energy.” We will now look at a
simple inﬁnite system, the onedimensional lattice. Here one has for each in
teger a physical system with possible states 1, 2, . . . , n. A conﬁguration of the
system consists of assigning an x
i
∈ ¦1, . . . , n¦ for each i:
x
−2
x
−1
x
0
x
1
x
2
x
3
Thus a conﬁguration is a point
x = ¦x
i
¦
+∞
i=−∞
∈
¸
Z
¦1, . . . , n¦ = Σ
n
.
We now make assumptions about energy:
(1) associated with the occurrence of a state k is a contribution Φ
0
(k) to the
total energy of the system independent of which position it occurs at;
(2) if state k
1
occurs in place i
1
, and k
2
in i
2
, then the potential energy due to
their interaction Φ
∗
2
(i
1
, i
2
, k
1
, k
2
) depends only on their relative position,
i.e., there is a function Φ
2
: Z ¦1, . . . , n¦ ¦1, . . . , n¦ → R so that
Φ
∗
2
(i
1
, i
2
, k
1
, k
2
) = Φ
2
(i
1
−i
2
; k
1
, k
2
)
(also: Φ
2
(j; k
1
, k
2
) = Φ
2
(−j; k
2
, k
1
)).
(3) all energy is due to contributions of the form (1) and (2).
Under these hypotheses the energy contribution due to x
0
being in the 0th
place is
φ
∗
(x) = Φ
0
(x
0
) +
¸
j=0
1
2
Φ
2
(j; x
j
, x
0
).
(We “give” each of x
0
and x
k
half the energy due to their interaction). We
now assume that Φ
2

j
= sup
k
1
,k
2
[Φ(j; k
1
, k
2
)[ satisﬁes
∞
¸
j=1
Φ
2

j
< ∞.
A. Gibbs Distribution 5
Then φ
∗
(x) ∈ R and depends continuously on x when ¦1, . . . , n¦ is given the
discrete topology and Σ
n
=
¸
Z
¦1, . . . , n¦ the product topology.
If we just look at x
−m
. . . x
0
. . . x
m
we have a ﬁnite system (n
2m+1
possible
conﬁgurations) and an energy
E
m
(x
−m
, . . . , x
m
) =
m
¸
j=−m
Φ
0
(x
j
) +
¸
−m≤j<k≤m
Φ
2
(k −j; x
k
, x
j
)
and the Gibbs distribution µ
m
assigns probabilities proportional to
e
−βE
m
(x
−m
,...,x
m
)
. Now just suppose that for each x
−m
, . . . , x
m
the limit
µ(x
−m
. . . x
m
) = lim
k→∞
¸
¦µ
k
(x
−k
. . . x
k
) : x
i
= x
i
∀[i[ ≤ m¦
exists. Then µ ∈ M(Σ
n
) and it would be natural to call µ the Gibbs distrib
ution on Σ
n
(for the given energy and β). If we are given x = ¦x
i
¦
∞
i=−∞
, then
instead of E
m
(x
−m
, . . . , x
m
) one might add in the contributions by interac
tions of x
j
(−m ≤ j ≤ m) with all other x
k
’s, i.e.,
m
¸
j=−m
Φ
0
(x
j
) +
∞
¸
k=−∞
1
2
Φ
2
(k −j; x
k
, x
j
)
.
If we deﬁne the (left) shift homeomorphism σ : Σ
n
→ Σ
n
by σ¦x
i
¦
∞
i=−∞
=
¦x
i+1
¦
∞
i=−∞
, then this expression is just
¸
m
j=−m
φ
∗
(σ
j
x). This expression
diﬀers from E
m
(x
−m
, . . . , x
m
) by at most
m
¸
j=−m
⎛
⎝
∞
¸
k=j+m+1
1
2
Φ
2

k
+
∞
¸
k=m−j+1
1
2
Φ
2

k
⎞
⎠
≤
∞
¸
k=1
kΦ
2

k
.
Thus, if C =
¸
∞
k=1
kΦ
2

k
< ∞ then E
m
(x
−m
, . . . , x
m
) diﬀers from
¸
m
j=−m
φ
∗
(σ
j
x) by at most C. If we used
¸
m
j=−m
φ
∗
(σ
j
x) instead of
E
m
(x
−m
, . . . , x
m
) in the Gibbs distribution µ
m
, the probabilities would
change by factors in [e
−2C
, e
2C
]. The point is that taking x
i
into consider
ation for i / ∈ [−m, m] may change µ
m
, but not drastically if one assumes
¸
∞
k=1
kΦ
2

k
< ∞.
We want now to state a theorem we have been leading up to. For φ : Σ
n
→
R continuous deﬁne
var
k
φ = sup¦[φ(x) −φ(y)[ : x
i
= y
i
∀ [i[ ≤ k¦ .
As φ is uniformly continuous, lim
k→∞
var
k
φ = 0.
1.2. Theorem. Suppose φ : Σ
n
→ R and there are c > 0, α ∈ (0, 1) so that
var
k
φ ≤ cα
k
for all k. Then there is a unique µ ∈ M
σ
(Σ
n
) for which one can
ﬁnd constants c
1
> 0, c
2
> 0, and P such that
6 1 Gibbs Measures
c
1
≤
µ¦y : y
i
= x
i
∀ i = 0, . . . , m¦
exp
−Pm+
¸
m−1
k=0
φ(σ
k
x)
≤ c
2
for every x ∈ Σ
n
and m ≥ 0.
This measure µ is written µ
φ
and called Gibbs measure of φ. Up to
constants in [c
1
, c
2
] the relative probabilities of the x
0
. . . x
m
’s are given
by exp
¸
m−1
k=0
φ(σ
k
x). For the physical system discussed above one takes
φ = −βφ
∗
. In statistical mechanics Gibbs states are not deﬁned by the above
theorem. We have ignored many subtleties that come up in more complicated
systems (e.g., higher dimensional lattices), where the theorem will not hold.
Our discussion was a gross one intended to motivate the theorem; we refer to
Ruelle [9] or Lanford [6] for a reﬁned outlook.
For later use we want to make a small generalization of Σ
n
before we prove
the theorem. If A is an n n matrix of 0’s and 1’s, let
Σ
A
= ¦x ∈ Σ
n
: A
x
i
x
i+1
= 1 ∀i ∈ Z¦ .
That is, we consider all x in which A says that x
i
x
i+1
is allowable for every i.
One easily sees that Σ
A
is closed and σΣ
A
= Σ
A
. We will always assume
that A is such that each k between 1 and n occurs at x
0
for some x ∈ Σ
A
.
(Otherwise one could have Σ
A
= Σ
B
with B an mm matrix and m < n.)
1.3. Lemma. σ : Σ
A
→ Σ
A
is topologically mixing (i.e., when U, V are non
empty open subsets of Σ
A
, there is an N so that σ
m
U ∩ V = ∅ ∀m ≥ N) if
and only if A
M
> 0 (i.e., A
M
i,j
> 0 ∀i, j) for some M.
Proof. One sees inductively that A
m
i,j
is the number of (m + 1)strings
a
0
a
1
. . . a
m
of integers between 1 and n with
(a) A
a
k
a
k+1
= 1 ∀k,
(b) a
0
= i, a
m
= j.
Let U
i
= ¦x ∈ Σ
A
: x
0
= i¦ = ∅.
Suppose Σ
A
is mixing. Then ∃N
i,j
with U
i
∩ σ
n
U
j
= ∅ ∀n ≥ N
i,j
. If
a ∈ U
i
∩ σ
n
U
j
, then a
0
a
1
. . . a
n
satisﬁes (a) and (b); so A
m
i,j
> 0 ∀i, j when
m ≥ max
i,j
N
i,j
.
Suppose A
M
> 0 for some M. As each number between 1 and n occurs as
x
0
for some x ∈ Σ
A
, each row of A has at least one positive entry. From this
it follows by induction that A
m
> 0 for all m ≥ M.
Consider open subsets U, V of Σ
A
with a ∈ U, b ∈ V . There is an r so
that
U ⊃ ¦x ∈ Σ
A
: x
k
= a
k
∀[k[ ≤ r¦
V ⊃ ¦x ∈ Σ
A
: x
k
= b
k
∀[k[ ≤ r¦ .
For t ≥ 2r + M, m = t − 2r ≥ M and A
m
> 0. Hence ﬁnd c
0
, . . . , c
m
with
c
0
= b
r
, c
m
= a
−r
, A
c
k
c
k+1
= 1 for all k. Then
x = b
−2
b
−1
b
0
b
r
c
1
c
m−1
a
−r
a
0
a
1
is in Σ
A
and x ∈ σ
t
U ∩ V . So Σ
A
is topologically mixing. ¯.
A. Gibbs Distribution 7
Let F
A
be the family of all continuous φ : Σ
A
→ R for which var
k
φ ≤ bα
k
(for all k ≥ 0) for some positive constants b and α ∈ (0, 1). For any β ∈ (0, 1)
one can deﬁne the metric d
β
on Σ
A
by d
β
(x, y) = β
N
where N is the largest
nonnegative integer with x
i
= y
i
for every [i[ < N. Then F
A
is just the set
of functions which have a positive H¨ older exponent with respect to d
β
. The
theorem we are interested in then reads
1.4. Existence of Gibbs measures. Suppose Σ
A
is topologically mixing and
φ ∈ F
A
. There is unique σinvariant Borel probability measure µ on Σ
A
for
which one can ﬁnd constants c
1
> 0, c
2
> 0 and P such that
c
1
≤
µ¦y : y
i
= x
i
for all i ∈ [0, m)¦
exp
−Pm+
¸
m−1
k=0
φ(σ
k
x)
≤ c
2
for every x ∈ Σ
A
and m ≥ 0.
This theorem will not be proved for some time. The ﬁrst step is to reduce
the φ’s one must consider.
Deﬁnition. Two functions ψ, φ ∈ C(Σ
A
) are homologous with respect to σ
(written ψ ∼ φ) if there is a u ∈ C(Σ
A
) so that
ψ(x) = φ(x) −u(x) +u(σx) .
1.5. Lemma. Suppose φ
1
∼ φ
2
and Theorem 1.4 holds for φ
1
. Then it holds
for φ
2
and µ
φ
1
= µ
φ
2
.
Proof.
m−1
¸
k=0
φ
1
(σ
k
x) −
m−1
¸
k=0
φ
2
(σ
k
x)
=
m−1
¸
k=0
u(σ
k+1
x) −u(σ
k
x)
= [u(σ
m
x) −u(x)[ ≤ 2u .
The exponential in the required inequality changes by at most a factor of
e
2u
when φ
1
is replaced by φ
2
. Thus the inequality remains valid with c
1
,
c
2
changed and P, µ unchanged. ¯.
1.6. Lemma. If φ ∈ F
A
, then φ is homologous to some ψ ∈ F
A
with ψ(x) =
ψ(y) whenever x
i
= y
i
for all i ≥ 0.
Proof. For each 1 ≤ t ≤ n pick ¦a
k,t
¦
∞
k=−∞
∈ Σ
A
with a
0,t
= t. Deﬁne
r : Σ
A
→ Σ
A
by r(x) = x
∗
where
x
∗
k
=
x
k
for k ≥ 0
a
k,x
0
for k ≤ 0 .
Let
8 1 Gibbs Measures
u(x) =
∞
¸
j=0
(φ(σ
j
x) −φ(σ
j
r(x))) .
Since σ
j
x and σ
j
r(x) agree in places from −j to +∞,
[φ(σ
j
x) −φ(σ
j
r(x))[ ≤ var
j
φ ≤ bα
j
.
As
¸
∞
j=0
bα
j
< ∞, u is deﬁned and continuous. If x
i
= y
i
for all [i[ ≤ n, then,
for j ∈ [0, n],
[φ(σ
j
x) −φ(σ
j
y)[ ≤ var
n−j
φ ≤ bα
n−j
and
[φ(σ
j
r(x)) −φ(σ
j
r(y))[ ≤ bα
n−j
.
Hence
[u(x) −u(y)[ ≤
[
n
2
]
¸
j=0
[φ(σ
j
x) −φ(σ
j
y) +φ(σ
j
r(x)) −φ(σ
j
r(y))[ + 2
¸
j>
[
n
2
]
α
j
≤ 2b
⎛
⎜
⎝
[
n
2
]
¸
j=0
α
n−j
+
¸
j>
[
n
2
]
α
j
⎞
⎟
⎠ ≤
4b α
[
n
2
]
1 −α
This shows that u ∈ F
A
. Hence ψ = φ −u +u ◦ σ is in F
A
also. Furthermore
ψ(x) = φ(x) +
∞
¸
j=−1
φ(σ
j+1
r(x)) −φ(σ
j+1
x)
+
∞
¸
j=0
φ(σ
j+1
x) −φ(σ
j
r(σx))
= φ(r(x)) +
∞
¸
j=0
φ(σ
j+1
r(x)) −φ(σ
j
r(σx))
.
The ﬁnal expression depends only on ¦x
i
¦
∞
i=0
, as we wanted. D. Lind cleaned
up the above proof for us. ¯.
Lemmas 1.5 and 1.6 tell us that in looking for a Gibbs measure µ
φ
for
φ ∈ F
A
(i.e., proving Theorem 1.4) we can restrict our attention to functions
φ for which φ(x) depends only on ¦x
i
¦
∞
i=0
.
B. Ruelle’s PerronFrobenius Theorem
We introduce now onesided shift spaces. One writes x for ¦x
i
¦
∞
i=0
(we will
continue to write x for ¦x
i
¦
∞
i=−∞
but never for both things at the same time).
Let
Σ
+
A
=
x ∈
∞
¸
i=0
¦1, . . . , n¦ : A
x
i
,x
i+1
= 1 for all i ≥ 0
¸
B. Ruelle’s PerronFrobenius Theorem 9
and deﬁne σ : Σ
+
A
→ Σ
+
A
by σ(x)
i
= x
i+1
. σ is a ﬁnitetoone continuous
map of Σ
+
A
onto itself. If φ ∈ C(Σ
+
A
) we get φ ∈ C(Σ
A
) by φ(¦x
i
¦
∞
i=−∞
) =
φ(¦x
i
¦
∞
i=0
). Suppose φ ∈ C(Σ
A
) satisﬁes φ(x) = φ(y) whenever x
i
= y
i
for all i ≥ 0. Then one can think of φ as belonging to C(Σ
+
A
) as follows:
φ(¦x
i
¦
∞
i=0
) = φ(¦x
i
¦
∞
i=−∞
) where x
i
for i ≤ 0 are chosen in any way subject
to ¦x
i
¦
∞
i=−∞
∈ Σ
A
. The functions C(Σ
+
A
) are thus identiﬁed with a certain
subclass of C(Σ
A
). We saw in Lemmas 1.5 and 1.6 that one only needs to get
Gibbs measures for φ ∈ C(Σ
+
A
) ∩ F
A
in order to get them for all φ ∈ F
A
.
In this section we will prove a theorem of Ruelle that will later be used
to construct and study Gibbs measures. For φ ∈ C(Σ
+
A
) deﬁne the operator
L = L
φ
on C(Σ
+
A
) by
(L
φ
f)(x) =
¸
y ∈σ
−1
x
e
φ(y)
f(y) .
It is the fact that σ is not onetoone on Σ
+
A
that will make this operator
useful.
1.7. Ruelle’s PerronFrobenius Theorem [10, 11]. Let Σ
A
be topolog
ically mixing, φ ∈ F
A
∩ C(Σ
+
A
) and L = L
φ
as above. There are λ > 0,
h ∈ C(Σ
+
A
) with h > 0 and ν ∈ M(Σ
+
A
) for which Lh = λh, L
∗
ν = λν,
ν(h) = 1 and
lim
m→∞
λ
−m
L
m
g −ν(g)h = 0 for all g ∈ C(Σ
+
A
) .
Proof. Because L is a positive operator and L1 > 0, one has that G(µ) =
(L
∗
µ(1))
−1
L
∗
µ ∈ M(Σ
+
A
) for µ ∈ M(Σ
+
A
). There is a ν ∈ M(Σ
+
A
) with
G(ν) = ν by the SchauderTychonoﬀ Theorem (see Dunford and Schwartz,
Linear Operators I, p. 456): Let E be a nonempty compact convex subset of
a locally convex topological vector space. Then any continuous G : E → E
has a ﬁxed point. In our case G(ν) = ν gives L
∗
ν = λν with λ > 0.
We will prove 1.7 via a sequence of lemmas. Let b > 0 and α ∈ (0, 1) be any
constants so that var
k
φ ≤ bα
k
for all k ≥ 0. Set B
m
= exp
¸
∞
k=m+1
2bα
k
and deﬁne
Λ = ¦f ∈ C(Σ
+
A
) : f ≥ 0, ν(f) = 1 , f(x) ≤ B
m
f(x
),
whenever x
i
= x
i
for all i ∈ [0, m]¦ .
1.8. Lemma. There is an h ∈ Λ with Lh = λh and h > 0.
Proof. One checks that λ
−1
Lf ∈ Λ when f ∈ Λ. Clearly λ
−1
Lf ≥ 0 and
ν(λ
−1
Lf) = λ
−1
L
∗
ν(f) = ν(f) = 1 .
Assume x
i
= x
i
for i ∈ [0, m]. Then
Lf(x) =
¸
j
e
φ(jx)
f(jx)
10 1 Gibbs Measures
where the sum ranges over all j with A
jx
0
= 1. For x
the expression runs
over the same j; as jx and jx
agree in places 0 to m+ 1
e
φ(jx)
f(jx) ≤ e
φ(jx
)
e
bα
m+1
B
m+1
f(jx
) ≤ B
m
e
φ(jx
)
f(jx
)
and so
Lf(x) ≤ B
m
Lf(x
) .
Consider any x, z ∈ Σ
+
A
. Since A
M
> 0 there is a y
∈ σ
−M
x with y
0
= z
0
.
For f ∈ Λ
L
M
f(x) =
¸
y∈σ
−M
x
exp
M−1
¸
k=0
φ(σ
k
y)f(y)
≥ e
−Mφ
f(y
) .
Let K = λ
M
e
Mφ
B
0
. Then 1 = ν(λ
−M
L
M
f) ≥ K
−1
f(z) gives f ≤ K as z
is arbitrary. As ν(f) = 1, f(z) ≥ 1 for some z and we get inf λ
−M
L
M
f ≥ K
−1
.
If x
i
= x
i
for i ∈ [0, m] and f ∈ Λ, one has
[f(x) −f(x
)[ ≤ (B
m
−1)K → 0
as m → ∞, since B
m
→ 1. Thus Λ is equicontinuous and compact by
the ArzelaAscoli Theorem. Λ = ∅ as 1 ∈ Λ. Applying SchauderTychonoﬀ
Theorem to λ
−1
L : Λ → Λ gives us h ∈ Λ with Lh = λh. Furthermore
inf h = inf λ
−M
L
M
h ≥ K
−1
. ¯.
1.9. Lemma. There is an η ∈ (0, 1) so that for f ∈ Λ one has λ
−M
L
M
f =
ηh + (1 −η)f
with f
∈ Λ.
Proof. Let g = λ
−M
L
M
f −ηh where η is to be determined. Provided ηh ≤
K
−1
we will have g ≥ 0. Assume x
i
= x
i
for all i ∈ [0, m]. We want to pick η
so that g(x) ≤ B
m
g(x
), or equivalently
() η(B
m
h(x
) −h(x)) ≤ B
m
λ
−M
L
M
f(x
) −λ
−M
L
M
f(x) .
We saw above that Lf
1
(x) ≤ B
m+1
e
bα
m+1
Lf
1
(x
) ≤ B
m+1
e
bα
m
Lf
1
(x
) for
any f
1
∈ Λ. Applying this to f
1
= λ
−M+1
L
M−1
f one has
λ
−M
L
M
f(x) ≤ B
m+1
e
bα
m
λ
−M
L
M
f(x
) .
Now h(x) ≥ B
−1
m
h(x
) because h ∈ Λ. To get () it is therefore enough to
have
η(B
m
−B
−1
m
)h(x
) ≤ (B
m
−B
m+1
e
bα
m
) λ
−M
L
M
f(x
)
or
η(B
m
−B
−1
m
)h ≤ (B
m
−B
m+1
e
bα
m
)K
−1
.
B. Ruelle’s PerronFrobenius Theorem 11
There is an L so that the logarithms of B
m
, B
−1
m
and B
m+1
e
bα
m
are in [−L, L]
for all m. Let u
1
, u
2
be positive constants such that
u
1
(x −y) ≤ e
x
−e
y
≤ u
2
(x −y) for all x, y ∈ [−L, L], x > y .
For () to hold it is enough for η > 0 to satisfy
ηhu
1
(log B
m
+ log B
m
) ≤ K
−1
u
2
(log B
m
−log(B
m+1
e
bα
m
))
or
ηhu
1
4bα
m+1
1 −α
≤ K
−1
u
2
bα
m
or
η ≤ u
2
(1 −α)(4αu
1
hK)
−1
. ¯.
1.10. Lemma. There are constants A > 0 and β ∈ (0, 1) so that
λ
−n
L
n
f −h ≤ Aβ
n
for all f ∈ Λ, n ≥ 0.
Proof. Let n = Mq + r, 0 ≤ r < M. Inductively one sees from Lemma 1.9
and Lh = λh that, for f ∈ Λ,
λ
−Mq
L
Mq
f = (1 −(1 −η)
q
)h + (1 −η)
q
f
q
where f
q
∈ Λ. As f
q
 ≤ K one has
λ
−Mq
L
Mq
f −h ≤ (1 −η)
q
(h +K)
and
λ
−n
L
n
f −h = λ
−r
L
r
(λ
−Mq
L
Mq
f −h)
≤ A(1 −η)
q+1
≤ Aβ
n
,
where
A = (1 −η)
−1
(h +K) sup
0≤r<M
λ
−r
L
r

and
β
M
= 1 −η . ¯.
1.11. Lemma. Let C
r
= ¦f ∈ C(Σ
+
A
) : var
r
f = 0¦. If F ∈ Λ, f ∈ C
r
, f ≥ 0
and fF ≡ 0, then ν(fF)
−1
λ
−r
L
r
(fF) ∈ Λ.
12 1 Gibbs Measures
Proof. Assume x
i
= x
i
for i ∈ [0, m]. Then
L
r
(fF)(x) =
¸
j
1
···j
r
x
exp
r−1
¸
k=0
φ(σ
k
(j
1
j
r
x))
f(j
1
j
r
x)F(j
1
j
r
x)
where j
1
j
r
runs over all rstrings of symbols for which j
1
j
r
x ∈ Σ
+
A
.
In the expression for L
r
(fF)(x
) one has j
1
j
r
running over the same
rstrings. Now f(j
1
j
r
x) = f(j
1
j
r
x
) as f ∈ C
r
, F(j
1
j
r
x) ≤
B
m+r
F(j
1
j
r
x
), and φ(σ
k
(j
1
j
r
x)) ≤ φ(σ
k
(j
1
j
r
x
)) + var
m+r−k
φ.
Since
B
m+r
exp
r−1
¸
k=0
var
m+r−k
φ
≤ B
m+r
exp
⎛
⎝
m+r
¸
j=m+1
bα
j
⎞
⎠
≤ B
m
,
each term in the above expression for L
r
(fF)(x) is bounded by B
m
times the
corresponding term for L
r
(fF)(x
). Hence L
r
(fF)(x) ≤ B
m
L
r
(fF)(x
).
One must still show ν(fF) > 0. Reasoning as in the proof of 1.8 (with
L
r
(fF) in place of f) we get
λ
r
ν(fF) = ν(λ
−M
L
M+r
(fF)) ≥ K
−1
L
r
(fF)(z),
for any z. But (fF)(w) > 0 gives L
r
(fF)(σ
r
w) > 0 and so ν(fF) > 0. ¯.
1.12. Lemma. For f ∈ C
r
, F ∈ Λ and n ≥ 0,
λ
−n−r
L
n+r
(fF) −ν(fF)h ≤ Aν([fF[)β
n
.
For g ∈ C(Σ
+
A
) one has lim
m→∞
λ
−m
L
m
g −ν(g)h = 0.
Proof. Write f = f
+
−f
−
with f
+
, f
−
≥ 0 and f
+
, f
−
∈ C
r
. Then
λ
−n−r
L
n+r
(f
±
F) −ν(f
±
F)h ≤ Aν([f
±
F[)β
n
.
For f
±
F ≡ 0, this is obvious; for f
±
F ≡ 0 we apply Lemmas 1.11 and 1.10.
These inequalities add up to give us the ﬁrst statement of the lemma.
Given g and ε > 0 one can ﬁnd r and f
1
, f
2
∈ C
r
so that f
1
≤ g ≤ f
2
and
0 ≤ f
2
−f
1
≤ ε. As [ν(f
i
) −ν(g)[ < ε, the ﬁrst statement of the lemma with
F = 1 gives
λ
−m
L
m
(f
i
) −ν(g)h ≤ ε(1 +h)
for large m. Since λ
−m
L
m
f
1
≤ λ
−m
L
m
g ≤ λ
−m
L
m
f
2
,
λ
−m
L
m
g −ν(g)h ≤ ε(1 +h)
for large m. ¯.
The proof of 1.7 is ﬁnished.
C. Construction of Gibbs Measures 13
C. Construction of Gibbs Measures
We continue to assume that φ ∈ F
A
∩ C(Σ
+
A
) and ν, h, λ are as in Ruelle’s
PerronFrobenius Theorem. Then µ = hν is a probability measure on Σ
+
A
;
µ(f) = ν(hf) =
f(x)h(x) dν(x).
1.13. Lemma. µ is invariant under σ : Σ
+
A
→ Σ
+
A
.
Proof. We need to show that µ(f) = µ(f ◦ σ) for f ∈ C(Σ
+
A
). Notice that
((Lf) g)(x) =
¸
y∈σ
−1
x
e
φ(y)
f(y)g(x)
=
¸
y∈σ
−1
x
e
φ(y)
f(y)g(σy)
= L(f (g ◦ σ))(x) .
Using this
µ(f) = ν(hf)
= ν(λ
−1
Lh f)
= λ
−1
ν(L(h (f ◦ σ)))
= λ
−1
(L
∗
ν)(h (f ◦ σ))
= ν(h (f ◦ σ))
= µ(f ◦ σ) . ¯.
Because µ is σinvariant on Σ
+
A
there is a natural way to make µ into a
measure on Σ
A
. For f ∈ C(Σ
A
) deﬁne f
∗
∈ C(Σ
+
A
) by
f
∗
(¦x
i
¦
∞
i=0
) = min¦f(y) : y ∈ Σ
A
, y
i
= x
i
for all i ≥ 0¦ .
Notice that for m, n ≥ 0 one has
(f ◦ σ
n
)
∗
◦ σ
m
−(f ◦ σ
n+m
)
∗
 ≤ var
n
f .
Hence
[µ((f ◦σ
n
)
∗
) −µ((f ◦σ
n+m
)
∗
)[ = [µ((f ◦σ
n
)
∗
◦σ
m
) −µ((f ◦σ
n+m
)
∗
)[ ≤ var
n
f
which approaches 0, as n → ∞, since f is continuous.
Hence ˜ µ(f) = lim
n→∞
µ((f ◦ σ
n
)
∗
) exists by the Cauchy criterion. It is
straightforward to check that ˜ µ ∈ C(Σ
A
)
∗
. By the Riesz Representation
Theorem we see that ˜ µ deﬁnes a probability measures on Σ
A
, which we will
denote by µ despite the possible ambiguity. Note that
˜ µ(f ◦ σ) = lim
n→∞
µ((f ◦ σ
n+1
)
∗
) = ˜ µ
14 1 Gibbs Measures
proving that ˜ µ is σinvariant. Also ˜ µ(f) = µ(f) for f ∈ C(Σ
A
)
∗
.
Recall that µ is ergodic if µ(E) = 0 or 1 whenever E is a Borel set with
σ
−1
E = E. One calls µ mixing if
lim
n→∞
µ(E ∩ σ
−n
F) = µ(E)µ(F),
for all Borel sets E and F. It is clear that mixing implies ergodicity and a
standard argument shows that the mixing condition only need be checked for
E and F in a basis for the topology.
1.14. Proposition. µ is mixing for σ : Σ
A
→ Σ
A
.
Proof. Writing S
m
φ(x) =
¸
m−1
k=0
φ(σ
k
x) one checks inductively that for f,
g ∈ C(Σ
+
A
) one has
(L
m
f)(x) =
¸
y∈σ
−m
x
e
S
m
φ(y)
f(y) .
Then
((L
m
f) g)(x) =
¸
y∈σ
−m
x
e
S
m
φ(y)
f(y)g(σ
m
y)
= L
m
(f (g ◦ σ
m
)) .
Let
E = ¦y ∈ Σ
A
: y
i
= a
i
, r ≤ i ≤ s¦ ,
F = ¦y ∈ Σ
A
: y
i
= b
i
, u ≤ i ≤ v¦ .
In checking the mixing condition for E and F we may assume r = u = 0
because µ is σinvariant. We want to calculate
µ(E ∩ σ
−n
F) = µ(χ
E
χ
σ
−n
F
)
= µ(χ
E
(χ
F
◦ σ
n
))
= ν(hχ
E
(χ
F
◦ σ
n
))
= λ
−n
L
∗n
ν(hχ
E
(χ
F
◦ σ
n
))
= ν(λ
−n
L
n
(hχ
E
(χ
F
◦ σ
n
)))
= ν(λ
−n
L
n
(hχ
E
) χ
F
) .
Now
[µ(E ∩ σ
−n
F) −µ(E)µ(F)[ = [µ(E ∩ σ
−n
F) −ν(hχ
E
)ν(hχ
F
)[
= [ν
(λ
−n
L
n
(hχ
E
) −ν(hχ
E
)h)χ
F
[
≤ λ
−n
L
n
(hχ
E
) −ν(hχ
E
)h ν(F) .
Because χ
E
∈ C
s
Lemma 1.12 gives, for n ≥ s,
C. Construction of Gibbs Measures 15
λ
−n
L
n
(hχ
E
) −ν(hχ
E
)h ≤ Aµ(E)β
n−s
where β ∈ (0, 1). One then has
[µ(E ∩ σ
−n
F) −µ(E)µ(F)[ ≤ A
µ(E)µ(F)β
n−s
for n ≥ s where A
= A (inf h)
−1
. Thus µ(E ∩ σ
−n
F) → µ(E)µ(F). ¯.
1.15. Lemma. Let a =
¸
∞
k=0
var
k
φ < ∞. If x, y ∈ Σ
A
with x
i
= y
i
for
i ∈ [0, m), then
[S
m
φ(x) −S
m
φ(y)[ ≤ a .
Proof. Deﬁne y
by
y
i
=
y
i
for i ≥ 0
x
i
for i ≤ 0 .
Since φ ∈ C(Σ
+
A
), φ(σ
k
y
) = φ(σ
k
y) for k ≥ 0. Hence
[S
m
φ(x) −S
m
φ(y)[ ≤
m−1
¸
k=0
[φ(σ
k
x) −φ(σ
k
y
)[
≤
m−1
¸
k=0
var
m−1−k
φ
≤ a . ¯.
We now complete the proof of 1.4.
1.16. Theorem. µ is a Gibbs measure for φ ∈ F
A
∩ C(Σ
+
A
).
Proof. Let E = ¦y ∈ Σ
A
: y
i
= x
i
for i ∈ [0, m)¦. For any z ∈ Σ
+
A
there is
at most one y
∈ σ
−m
z with y
∈ E. Thus, using 1.15,
L
m
(hχ
E
)(z) =
¸
y∈σ
−m
z
e
S
m
φ(y)
h(y)χ
E
(y)
≤ e
S
m
φ(x)
e
a
h
and so
µ(E) = ν(hχ
E
)
= λ
−m
ν(L
m
(hχ
E
))
≤ λ
−m
e
S
m
φ(x)
e
a
h .
Thus take c
2
= e
a
h. On the other hand, for any z ∈ Σ
+
A
there is at least
one y
∈ σ
−m−M
z with y
∈ E. Then
L
m+M
(hχ
E
)(z) ≥ e
S
m+M
φ(y
)
h(y
)
≥ e
−Mφ−a
(inf h) e
S
m
φ(x)
16 1 Gibbs Measures
and
µ(E) = λ
−m−M
ν(L
m+M
(hχ
E
)) ≥ c
1
λ
−m
e
S
m
φ(x)
where c
1
= λ
−M
e
−Mφ−a
. We have veriﬁed the desired inequalities on mea
sures of cylinder sets given in 1.4 with P = log λ.
Suppose now that µ
is any other measure satisfying the inequalities in
Theorem 1.4 with constants c
1
, c
2
, P
. For x ∈ Σ
A
let E
m
(x) = ¦y ∈ Σ
A
:
y
i
= x
i
for all i ∈ [0, m)¦. Let T
m
be a ﬁnite subset of Σ
A
so that Σ
A
=
¸
x∈T
m
E
m
(x) disjointly. Then
c
1
e
−P
m
¸
x∈T
m
e
S
m
φ(x)
≤
¸
x∈T
m
µ
(E
m
(x))
= 1
≤ c
2
e
−P
m
¸
x∈T
m
e
S
m
φ(x)
.
From this one sees that P
= lim
m→∞
1
m
log
¸
x∈T
m
e
S
m
φ(x)
. One can
apply the same reasoning to µ; hence P
= P as they equal the same limit.
The estimates on µ
(E
m
(x)) and µ(E
m
(x)) give us µ
(E
m
(x)) ≤ dµ(E
m
(x))
where d = c
2
c
−1
1
. Taking limits this extends to µ
(E) ≤ dµ(E
m
) for all Borel
sets E. In particular µ
(E) = 0 when µ(E) = 0. By the RadonNikodym
Theorem µ
= fµ for some µintegrable f. Applying σ one has
µ
= σ
∗
µ
= (f ◦ σ) σ
∗
µ
= (f ◦ σ)µ.
As the RadonNikodym derivative is unique up to µequivalence, f ◦ σ
a.e.
= f.
Because µ is ergodic this gives f equivalent to some constant c.
1 = µ
(Σ
A
) =
c dµ = c and µ = µ
. ¯.
D. Variational Principle
We will describe Gibbs measures as those maximizing a certain quantity, in a
way analogous to Lemma 1.1. If C = ¦C
1
, . . . , C
k
¦ is a partition of a measure
space (X, B, µ) (i.e., the C
i
’s are pairwise disjoint and X =
¸
k
i=1
C
i
), one
deﬁnes the entropy
H
µ
(C) =
k
¸
i=1
(−µ(C
i
) log µ(C
i
)) .
If D is another (ﬁnite) partition,
C ∨ D = ¦C
i
∩ D
j
: C
i
∈ C , D
j
∈ D¦.
D. Variational Principle 17
1.17. Lemma. H
µ
(C ∨ D) ≤ H
µ
(C) +H
µ
(D).
Proof.
H
µ
(C ∨ D)−H
µ
(C) =
¸
i,j
(−µ(C
i
∩ D
j
) log µ(C
i
∩ D
j
))−
¸
i
(−µ(C
i
) log µ(C
i
))
=
¸
i,j
−µ(C
i
∩ D
j
) log
µ(C
i
∩ D
j
)
µ(C
i
)
=
¸
j
¸
i
µ(C
i
)
−µ(C
i
∩ D
j
)
µ(C
i
)
log
µ(C
i
∩ D
j
)
µ(C
i
)
The function ϕ(x) = −xlog x (ϕ(0) = 0) is con
cave on [0, 1] as ϕ
(x) < 0 for x ∈ (0, 1). From this
it follows that
ϕ(a
1
x
1
+a
2
x
2
) ≥ a
1
ϕ(x
1
) +a
2
ϕ(x
2
)
when x
1
, x
2
∈ [0, 1], a
1
+ a
2
= 1. Inductively
one sees that ϕ(
¸
n
i=1
a
i
x
i
) ≥
¸
n
i=1
a
i
ϕ(x
i
) when
¸
n
i=1
a
i
= 1 and a
i
≥ 0.
0 0.5 1.0
0
0.5
1.0
x
ϕ
Applying this to a
i
= µ(C
i
) and x
i
=
µ(C
i
∩D
j
)
µ(C
i
)
we get
¸
i
µ(C
i
) ϕ
µ(C
i
∩ D
j
)
µ(C
i
)
≤ ϕ
¸
i
µ(C
i
∩ D
j
)
= ϕ(µ(D
j
)) .
So
H
µ
(C ∨ D) −H
µ
(C) ≤
¸
j
ϕ(µ(D
j
)) = H
µ
(D) . ¯.
1.18. Lemma. Suppose ¦a
m
¦
∞
m=1
is a sequence satisfying inf
a
m
m
> −∞,
a
m+n
≤ a
m
+a
n
for all m, n. Then lim
m→∞
a
m
m
exists and equals inf
m
a
m
m
.
Proof. Fix m > 0. For j > 0, write j = km+n with 0 ≤ n < m. Then
a
j
j
=
a
km+n
km+n
≤
a
km
km
+
a
n
km
≤
ka
m
km
+
a
n
km
Letting j → ∞, k → ∞ one gets
limsup
j
a
j
j
≤
a
m
m
Thus limsup
j
a
j
j
≤ inf
m
a
m
m
. As liminf
j
a
j
j
≥ inf
m
a
m
m
, one gets that lim
j
a
j
j
exists and equals inf
m
a
m
m
¯ .
18 1 Gibbs Measures
1.19. Lemma. If D is a (ﬁnite) partition of (X, B, µ) and T an automor
phism of (X, B, µ), then
h
µ
(T, D) = lim
m→∞
1
m
H
µ
(D∨ T
−1
D∨ ∨ T
−m+1
D)
exists.
Proof. Let a
m
= H
µ
(D∨ T
−1
D∨ ∨ T
−m+1
D). Then
a
m+n
≤H
µ
(D∨T
−1
D∨ ∨T
−m+1
D)+H
µ
(T
−m
D∨ ∨T
−m−n+1
D)≤a
m
+a
n
since
T
−m
D∨ ∨ T
−m−n+1
D = T
−m
(D∨ ∨ T
−n+1
D)
and µ is Tinvariant. ¯.
Deﬁnition. Let µ ∈ M
σ
(Σ
A
) and U = ¦U
1
, . . . , U
n
¦ where U
i
= ¦x ∈ Σ
A
:
x
0
= i¦. Then s(µ) = h
µ
(σ, U) is called the entropy of µ.
Suppose now that φ ∈ C(Σ
A
) and that a
0
a
1
a
m−1
are integers between
1 and n satisfying A
a
k
a
k+1
= 1. Write
sup
a
0
a
1
···a
m−1
S
m
φ = sup
m−1
¸
k=0
φ(σ
k
x) : x ∈ Σ
A
, x
i
= a
i
for all 0 ≤ i < m
¸
and
Z
m
(φ) =
¸
a
0
a
1
···a
m−1
exp
sup
a
0
a
1
···a
m−1
S
m
φ
.
1.20. Lemma. For φ ∈ C(Σ
A
), P(φ) = lim
m→∞
1
m
log Z
m
(φ) exists (called
the pressure of φ).
Proof. Notice that
sup
a
0
a
1
···a
m+n−1
S
m+n
φ ≤ sup
a
0
a
1
···a
m−1
S
m
φ + sup
a
m
···a
m+n−1
S
n
φ.
From this one gets Z
m+n
(φ) ≤ Z
m
(φ)Z
n
(φ); the terms in Z
m+n
(φ) are
bounded by terms in Z
m
(φ)Z
n
(φ) and Z
m
(φ)Z
n
(φ) may have more terms,
all positive. Apply Lemma 1.18 to a
m
= log Z
m
(φ) (notice a
m
≥ −mφ). ¯.
1.21. Proposition. Suppose φ ∈ C(Σ
A
) and µ ∈ M
σ
(Σ
A
). Then
s(µ) +
φ dµ ≤ P(φ) .
D. Variational Principle 19
Proof. As
φ ◦ σ
k
dµ =
φdµ,
1
m
S
m
φdµ =
φdµ where S
m
φ(x) =
¸
m−1
k=0
φ(σ
k
x). Hence
s(µ) +
φ dµ ≤ lim
m→∞
1
m
H
µ
(U ∨ ∨ σ
−m+1
U) +
S
m
φ dµ
.
Now U ∨ ∨ σ
−m+1
U partitions points x ∈ Σ
A
according to x
0
x
1
x
m−1
.
Thus
H
µ
(U ∨ ∨ σ
−m+1
U) +
S
m
φ dµ
≤
¸
a
0
···a
m−1
µ(a
0
a
m−1
)(−log µ(a
0
a
m−1
)) + sup
a
0
a
1
···a
m−1
S
m
φ
≤ log Z
m
(φ) by Lemma 1.1 .
Now let m → ∞. ¯.
1.22. Theorem. Let φ ∈ F
A
, Σ
A
topologically mixing and µ
φ
the Gibbs mea
sure of φ. Then µ
φ
is the unique µ ∈ M
σ
(Σ
A
) for which
s(µ) +
φ dµ = P(φ) .
Proof. Given a
0
a
m−1
, pick x with x
i
= a
i
(i = 0, . . . , m−1) and
S
m
φ(x) = sup
a
0
a
1
···a
m−1
S
m
φ.
Now, as µ = µ
φ
is the Gibbs measure,
µ¦y ∈ Σ
A
: y
i
= a
i
∀ 0 ≤ i < m¦
exp(−Pm+S
m
φ(x))
∈ [c
1
, c
2
] .
Summing the measure of these sets over all possible a
0
a
m−1
’s gives 1; so
c
1
exp(−Pm)Z
m
(φ) ≤ 1 ≤ c
2
exp(−Pm)Z
m
(φ)
or
Z
m
(φ)
exp(Pm)
∈ [c
−1
2
, c
−1
1
] .
It follows that P(φ) = lim
m→∞
1
m
log Z
m
(φ) = P.
If y
i
= x
i
for all i = 0, . . . , m−1, then
[S
m
φ(y) −S
m
φ(x)[ ≤
m−1
¸
k=0
[φ(σ
k
y) −φ(σ
k
y)[
≤ var
0
φ + var
1
φ + + var
[
m
2
]
φ + var
m−
[
m
2
]
φ + + var
0
φ
≤ 2c
[
m
2
]
¸
k=0
α
k
≤
2c
1 −α
= d .
20 1 Gibbs Measures
Hence, if B = ¦y ∈ Σ
A
: y
i
= a
i
for all i = 0, . . . , m−1¦, then for x ∈ B,
−µ(B) log µ(B) +
B
S
m
φ dµ ≥ −µ(B) [ log µ(B) −S
m
φ(x) +d ]
≥ −µ(B) [ log
c
2
e
−Pm+S
m
φ(x)
−S
m
φ(x) +d ]
≥ µ(B)(Pm−log c
2
−d) .
Since
H
µ
(U ∨ ∨ σ
−m+1
U) +
S
m
φ dµ
=
¸
B
−µ(B) log µ(B) +
B
S
m
φ dµ
≥
¸
B
µ(B)(Pm−log c
2
−d) = Pm−log c
2
−d,
we get
s(µ) +
φ dµ = lim
m→∞
1
m
H
µ
(U ∨ ∨ σ
−m+1
U) +
S
m
φ dµ
≥ lim
m→∞
1
m
(Pm−log c
1
−d) = P = P(φ) .
The reverse inequality was in Proposition 1.21. So
s(µ
φ
) +
φ dµ
φ
= P(φ) .
To prove uniqueness we will need a couple of lemmas.
1.23. Lemma. Let X be a compact metric space, µ ∈ M(X), and D =
¦D
1
, . . . , D
n
¦ a Borel partition of X. Suppose ¦C
m
¦
∞
m=1
is a sequence of par
titions so that diam(C
m
) = max
C∈C
m
diam(C) → 0 as m → ∞. Then there
are partitions ¦E
m
1
, . . . , E
m
n
¦ so that
1. each E
m
i
is a union of members of C
m
,
2. lim
m→∞
µ(E
m
i
∆D
i
) = 0 for each i.
Proof. Pick compacts K
1
, . . . , K
n
with K
i
⊂ D
i
and µ(D
i
`K
i
) < ε. Let
δ = inf
i=j
d(K
i
, K
j
) and consider m with diam(C
m
) ≤
δ
2
. Divide the elements
C ∈ C
m
into groups whose unions are E
m
1
, . . . , E
m
n
so that
C ⊂ E
m
i
if C ∩ K
i
= ∅ .
As diam(C
m
) ≤
δ
2
any C ∈ C
m
can intersect at most one K
i
. Put a C hitting
no K
i
in any E
m
i
you like. Then E
m
i
⊃ K
i
and
µ(E
m
i
∆D
i
) = µ(D
i
`E
m
i
)+µ(E
m
i
`D
i
) ≤ ε +µ
X`
n
¸
i=1
K
i
≤ (n+1) ε . ¯.
D. Variational Principle 21
1.24. Lemma. Suppose 0 ≤ p
1
, . . . , p
m
≤ 1, s = p
1
+ + p
m
≤ 1 and
a
1
, . . . , a
m
∈ R. Then
m
¸
i=1
p
i
(a
i
−log a
i
) ≤ s
log
m
¸
i=1
e
a
i
−log s
.
Proof. This generalizes 1.1. One shows by calculus that the left side is maxi
mized at p
i
=
se
a
i
¸
j
e
a
j
¯ .
Proof of Theorem 1.22 (continued). Let ν ∈ M
σ
(Σ
A
) satisfy s(ν)+
φ dν = P.
First suppose ν is singular with respect to µ. Then there is a Borel set B with
σ(B) = B, µ(B) = 0 and ν(B) = 1. Let C
m
= σ
−
[
m
2
]
+1
U∨ ∨U∨ ∨σ
[
m
2
]
U.
Then diam(C
m
) → 0 (use d
β
metric). Applying 1.23 to ¦B, X`B¦ one ﬁnds
sets E
m
which are unions of elements of C
m
and satisfy (µ+ν)(B∆E
m
) → 0.
As µ + ν is σinvariant and σ
−m+
[
m
2
]
B = B one has (µ + ν)(B∆F
m
) → 0
where F
m
= σ
−m+
[
m
2
]
E
m
is a union of members of U ∨ ∨ σ
−m+1
U. Since
s(ν) = inf
1
m
H
ν
(U ∨ ∨ σ
−m+1
U) one has
P = P(φ) = s(ν) +
φ dν ≤
1
m
H
ν
(U ∨ ∨ σ
−m+1
U) +
S
m
φ dν
or
mP ≤
¸
B∈U∨···∨σ
−m+1
U
¸
−ν(B) log ν(B) +
B
S
m
φ dν
.
Picking x
B
∈ B one has S
m
φ ≤ S
m
φ(x
B
) +d on B and so
mP ≤ d +
¸
B
ν(B)(S
m
φ(x
B
) −log ν(B))
≤ d +
¸
B⊂F
m
ν(B)(S
m
φ(x
B
) −log ν(B))
+
¸
B⊂X\F
m
ν(B)(S
m
φ(x
B
) −log ν(B)).
Applying 1.24
mP −d ≤ ν(F
m
) log
¸
B⊂F
m
exp(S
m
φ(x
B
))
+ ν(X`F
m
) log
¸
B⊂X\F
m
exp(S
m
φ(x
B
)) + 2K
∗
.
where K
∗
= sup
0≤s≤1
(−s log s). Rearranging terms:
22 1 Gibbs Measures
−2K
∗
−d ≤ ν(F
m
) log
¸
B⊂F
m
exp
S
m
φ(x
B
) −mP
+ ν(X`F
m
) log
¸
B⊂X\F
m
exp
S
m
φ(x
B
) −mP
≤ ν(F
m
) log
¸
B⊂F
m
c
−1
2
µ(B) + ν(X`F
m
) log
¸
B⊂X\F
m
c
−1
2
µ(B)
≤ log c
−1
2
+ν(F
m
) log µ(F
m
) +ν(X`F
m
) log µ(X`F
m
) .
Letting m → ∞, ν(F
m
) → 1, µ(F
m
) → 0 and the above inequality is contra
dictory.
In general, for ν
∈ M
σ
(Σ
A
), write ν
= βν + (1 − β)µ
where β ∈ (0, 1),
ν ∈ M
σ
(Σ
A
) is singular w.r.t. µ and µ
∈ M
σ
(Σ
A
) is absolutely continuous
w.r.t. µ. As ν and µ
are supported on disjoints sets
P
ν
(φ) = βP
ν
(φ) + (1 −β)P
µ
(φ),
where P
ν
(φ) = s(ν) +
φ dν. Suppose P
ν
(φ) = P. Since P
ν
(φ) ≤ P and
P
µ
(φ) ≤ P (Prop. 1.21), we have P
ν
(φ) = P or β = 0. We just saw that
P
ν
(φ) = P. Thus ν
= µ
and write ν
=
dν
dµ
µ. Then the
dν
dµ
is σinvariant up
to equivalence as ν
, µ are both invariant and the RadonNikodym derivative
is unique up to equivalence. So
dν
dµ
is constant and ν
= µ. ¯.
E. Further Properties
In this section we look at more examples of the good behavior of Gibbs mea
sures. Throughout we assume µ = µ
φ
with φ ∈ F
A
and σ[
Σ
A
topologically
mixing.
Two partitions P and Q are called εindependent if
¸
P∈P,Q∈Q
[µ(P ∩ Q) −µ(P)µ(Q)[ < ε .
Let U = ¦U
1
, . . . , U
n
¦ be the partition of Σ
A
with
U
j
= ¦x ∈ Σ
A
: x
0
= j¦ .
The partition U is called weakBernoulli (for σ and µ) if for every ε > 0 there
is an N(ε) so that
P = U ∨ σ
−1
U ∨ ∨ σ
−s
U and Q = σ
−t
U ∨ ∨ σ
−t−r
U
are εindependent for all s ≥ 0, r ≥ 0, t ≥ s + N(ε). A wellknown theorem
of Friedman and Ornstein [4] states that if U is weakBernoulli, then (σ, µ) is
conjugate to a Bernoulli shift.
E. Further Properties 23
1.25. Theorem. U is weakBernoulli for the Gibbs measure µ = µ
φ
.
Proof. We may assume that φ ∈ C(Σ
+
A
) as before. For P ∈ P we have χ
P
∈
C
r
. As in the proof of 1.14, for Q ∈ Q
[µ(P ∩ Q) −µ(P)µ(Q)[ ≤ A
µ(P)µ(Q)β
t−s
where β ∈ (0, 1). Summing over P, Q
¸
P,Q
[µ(P ∩ Q) −µ(P)µ(Q)[ ≤ A
β
t−s
≤ ε
when t −s is large. ¯.
Because µ is mixing, µ(f (g ◦ σ
n
)) → µ(f)µ(g) as n → ∞ for f, g contin
uous (in fact L
2
) functions. The above proof used that this convergence was
exponentially fast for characteristic functions of cylinder sets. This exponen
tial convergence will now be carried over to functions in F
A
. For α ∈ (0, 1)
let H
α
be the family of f ∈ C(Σ
A
) with var
k
f ≤ cα
k
for some c. H
α
is a
Banach space under the norm
f
α
= f + sup
k≥0
(α
−k
var
k
f) .
1.26. Exponential Cluster Property. For ﬁxed α ∈ (0, 1) there are con
stants D and γ ∈ (0, 1) so that
[µ(f (g ◦ σ
n
)) −µ(f)µ(g)[ ≤ Df
α
g
α
γ
n
for all f, g ∈ H
α
, n ≥ 0.
Proof. For k ≥ 0 and x ∈ Σ
A
, let
E
k
(x) = ¦y ∈ Σ
a
: y
i
= x
i
for all [i[ ≤ k¦ .
Deﬁne f
k
(x) = µ(E
k
(x))
−1
E
k
(x)
f dµ. Then µ(f
k
) = µ(f) and f − f
k
 ≤
f
α
α
k
. Hence
[µ(f (g ◦ σ
n
)) −µ(f)µ(g)[ ≤
[µ(f
k
(g
k
◦σ
n
)) −µ(f
k
)µ(g
k
)[ +[µ((f −f
k
) (g ◦σ
n
))[ +[µ(f
k
((g −g
k
) ◦σ
n
))[
≤ [µ(f
k
(g
k
◦ σ
n
)) −µ(f
k
)µ(g
k
)[ + 2α
k
f
α
g
α
.
Now f
k
is measurable with respect to the partition P = ¦E
k
(x)¦
x
; i.e., f
k
=
¸
P∈P
a
P
χ
P
. Also g
k
=
¸
P∈P
b
P
χ
P
. Hence
[µ(f
k
(g
k
◦ σ
n
)) −µ(f
k
)µ(g
k
)[ ≤
¸
P,Q∈P
[a
P
b
P
[ [µ(P ∩ σ
−n
Q) −µ(P)µ(Q)[
≤ fgA
β
n−2k
≤ f
α
g
α
A
β
n−2k
.
Letting k = [n/3] we get the result with γ = max(α
1/3
, β
1/3
). ¯.
24 1 Gibbs Measures
1.27. Central Limit Theorem. For ψ ∈ F
A
there is a ξ = ξ(ψ) ∈ [0, ∞) so
that
µ
x ∈ Σ
A
:
1
√
n
(S
n
ψ(x) −nµ(ψ)) < r
n→∞
−→
1
ξ
√
2π
r
−∞
e
−x
2
/2ξ
2
dx.
For ξ = 0, convergence is asserted for r = 0 and the expression on the right
is taken to be 0 for r < 0 and 1 for r > 0.
Remark. We omit the proof, referring the reader to M. Ratner [13]. It is
interesting to know when ξ(ψ) = 0. This happens (see [13]) precisely when
ψ −µ(ψ) = u ◦ σ −u
has a solution u ∈ L
2
(µ). It is interesting that in case such u can be found
one can ﬁnd u ∈ F
A
and so ψ is homologous to a constant. The reasoning for
this is very roundabout and it would be good to ﬁnd a nice direct proof.
1.28. Theorem. Let Σ
A
be topologically mixing and φ, ψ ∈ F
A
. The following
are equivalent:
(i) µ
φ
= µ
ψ
.
(ii) There is a constant K so that S
m
φ(x)−S
m
ψ(x) = mK whenever σ
m
x = x.
(iii)There are a constant K and a u ∈ F
A
so that
φ(x) = ψ(x) +K +u(σx) −u(x) for all x ∈ Σ
A
.
(iv)There are constants K and L so that [S
m
φ(x) − S
m
ψ(x) − mK[ ≤ L for
all x and all m > 0.
If these conditions hold, then K = P(φ) −P(ψ).
Proof. (iii) ⇒ (iv) is obvious and (iv) ⇒ (i) is just like Lemma 1.5. Assume
µ
φ
= µ
ψ
and σ
m
x = x. From the deﬁnition of a Gibbs measure and µ
φ
= µ
ψ
one sees that
exp(−P(φ)j +S
j
φ(x))
exp(−P(ψ)j +S
j
ψ(x))
∈ [d
1
, d
2
],
where d
1
> 0, d
2
> 0 are independent of x and j. This is equivalent to
[S
j
φ(x) −S
j
ψ(x) −j(P(φ) −P(ψ))[ ≤ M
for some M independent of j, x. If σ
m
x = x, letting j = km, S
j
φ(x) =
kS
m
φ(x) and
M > k [S
m
φ(x) −S
m
ψ(x) −m(P(φ) −P(ψ))[ .
Letting k → ∞ we get (ii) with K = P(φ) −P(ψ).
Now assume (ii). In proving (iii) we will need the following standard
lemma.
E. Further Properties 25
1.29. Lemma. If T : X → X is a topologically transitive continuous map of
a compact metric space, then there is a point x ∈ X so that
if U = ∅ is open, N > 0, then T
n
x ∈ U for some n ≥ N.
Proof. As X is 2nd countable, let U
1
, U
2
, . . . be a basis for the topology. By
transitivity, the open set
V
i,N
=
¸
n≥N
T
−n
U
i
is dense in X. By Baire Category Theorem there is an x ∈
¸
i,N
V
i,N
. ¯.
Continuing the proof of (iii) from (ii), let x be as in the lemma for X = Σ
A
,
T = σ (topological mixing is stronger than transitivity). Let η = φ−ψ−K ∈
F
A
. Let Γ = ¦σ
k
x : k ≥ 0¦ and deﬁne u : Γ → R by
u(σ
k
x) =
k−1
¸
j=0
η(σ
j
x) .
As Γ is dense in Σ
A
, Γ must be inﬁnite (except in the trivial case of Σ
A
=one
point) and x is not periodic. Thus σ
k
x = σ
m
x for m = k and u is well deﬁned
on Γ. We will estimate var
r
(u[Γ). Suppose y = σ
k
x, and z = σ
m
x (m > k)
agree in places −r to r. Then x
k+s
= x
m+s
for all [s[ ≤ r. Deﬁne w ∈ Σ
A
by
w
i
= x
t
for i ≡ t (mod m−k) , k ≤ t ≤ m.
Then σ
m−k
w = w and w, x agree in places k − r to m + r; hence σ
j
x, σ
j
w
agree in places k −r −j through m+r −j. Now
u(w) −u(y) =
m−1
¸
j=k
η(σ
j
x) .
Since (ii) gives
m−1
¸
j=k
η(σ
j
w) = 0 ,
[u(z) −u(y)[ ≤
m−1
¸
j=k
[η(σ
j
x) −η(σ
j
w)[
≤ var
r
η + var
r+1
η + + var
r+1
η + var
r
η
≤ 2
∞
¸
s=r
var
s
η.
Since η ∈ F
A
, var
s
η ≤ cα
s
for some α ∈ (0, 1) and
26 1 Gibbs Measures
var
r
(u[
Γ
) ≤ 2c
∞
¸
s=r
α
s
=
2c
1 −α
α
r
.
So u is uniformly continuous on Γ and therefore extends uniquely to a con
tinuous u : Σ
A
= Γ → R. Because var
r
u = var
r
(u[
Γ
), u ∈ F
A
. For z ∈ Γ,
u(σz) −u(z) = η(z)
and this equation extends to Σ
A
by continuity. ¯.
References
For discussions of Gibbs’ states and statistical mechanics we refer to Ruelle’s
book [9] and Lanford [6]. The deﬁnition of Gibbs state in statistical mechanics
does not coincide with what we gave in Section 1. There Gibbs states are de
ﬁned for more general φ and then our theorem corresponds to one in statistical
mechanics about uniqueness of Gibbs states [3, 10, 6]. In those papers one deal
with Σ
n
instead of slightly more general Σ
A
. Proofs of Theorem 1.4 for Σ
A
appear in [2, 11, 12].
The variational characterization (Theorem 1.22) of the Gibbs measure µ
φ
is due
to LanfordRuelle [7] in the Σ
n
case. For Σ
A
one can ﬁnd it in [2] or [11]. We
have followed the proof in [2], which in turn was based on Adler and Weiss’
proof [1] of a theorem of Parry (Theorem 1.22 with φ = 0).
The weak Bernoulli condition (Theorem 1.25) was veriﬁed for the Σ
n
case ﬁrst
by Gallavotti [5]. It was extended to Σ
A
independently by each of Bunimovich,
Ratner, Ruelle and this author. Theorem 1.28 is taken from Livˇsic [8] and Sinai
[12]. Theorem 1.26 is from [11].
Ruelle gave two proofs of his PerronFrobenius Theorem, for diﬀerent circum
stances [10, 11]. We have followed parts of each.
1. R. L. Adler, B. Weiss: Similarity of automorphisms of the torus, Memoirs of
the American Mathematical Society 98, American Mathematical Society, Prov
idence, R.I. 1970.
2. R. Bowen: Some systems with unique equilibrium states, Math. Systems The
ory 8 (1974/75), no. 3, 193–202.
3. R. L. Dobrushin: The problem of uniqueness of a Gibbsian random ﬁeld and the
problem of phase transitions, Func. Anal. and its Appl. 2 (1968) no. 4, 44–57.
4. N. A. Friedman, D. S. Ornstein: On isomorphism of weak Bernoulli transforma
tions, Advances in Math. 5 (1970) 365–394.
5. G. Gallavotti: Ising model and Bernoulli schemes in one dimension, Comm.
Math. Phys. 32 (1973), 183–190.
6. O. E. Lanford: Entropy and equilibrium states in classical statistical mechanics.
In “statistical mechanics and mathematical problems” (ed. A. Lenard), Lecture
Notes in Physics 20, pp. 1–113. Springer, Berlin, 1973.
7. O. E. Lanford, D. Ruelle: Observables at inﬁnity and states with short range
correlations in statistical mechanics, Comm. Math. Phys. 13 (1969) 194–215.
References 27
8. A. N. Livˇsic: Certain properties of the homology of Ysystems, Mat. Zametki 10
(1971), 555–564. English translation: Math. Notes Acad. Sci. USSR 10 (1971),
758–763.
9. D. Ruelle: Statistical mechanics: Rigorous results, W. A. Benjamin, Inc., New
YorkAmsterdam 1969. (
1
)
10. D. Ruelle: Statistical mechanics of a onedimensional lattice gas, Comm. Math.
Phys. 9 (1968) 267–278.
11. D. Ruelle: A measure associated with axiomA attractors, Amer. J. Math. 98
(1976), no. 3, 619–654.
12. Ya. Sinai: Gibbs measures in ergodic theory, Uspehi Mat. Nauk 27 (1972), no.
4, 21–64. English translation: Russian Math. Surveys 27 (1972), no. 4, 21–69.
13. M. Ratner: The central limit theorem for geodesic ﬂows on ndimensional man
ifolds of negative curvature, Israel J. Math. 16 (1973), 181–197.
1
Reprint: World Scientiﬁc Publishing Co., Inc., River Edge, NJ; Imperial College
Press, London, 1999 (note of the editor).
2
General Thermodynamic Formalism
A. Entropy
In Section D of Chapter 1, we deﬁned the number h
µ
(T, D) when T is an
endomorphism of a probability space and D a ﬁnite measurable partition. We
now deﬁne the entropy of µ w.r.t. T by
h
µ
(T) = sup
D
h
µ
(T, D),
where D ranges over all ﬁnite partitions. We will now turn to some computa
tional lemmas.
We deﬁne
H
µ
(C[D) = H
µ
(C ∨ D) −H
µ
(D)
= −
¸
i
µ(D
i
)
¸
j
µ(C
j
∩ D
i
)
µ(D
i
)
log
µ(C
j
∩ D
i
)
µ(D
i
)
≥ 0 .
Lemma 1.17 says that H
µ
(C[D) ≤ H
µ
(C). We write C ⊂ D if each set in C
is a union of sets in D.
2.1. Lemma.
(a) H
µ
(C[D) ≤ H
µ
(C[E) if D ⊃ E.
(b) H
µ
(C[D) = 0 if D ⊃ C.
(c) H
µ
(C ∨ D[E) ≤ H
µ
(C[E) +H
µ
(D[E).
(d) H
µ
(C) ≤ H
µ
(D) +H
µ
(C[D).
Proof. Letting ϕ(x) = −xlog x, H
µ
(C[D) =
¸
j
¸
i
µ(D
i
) ϕ
µ(C
j
∩D
i
)
µ(D
i
)
.
Since E ⊂ D, one can rewrite this as
H
µ
(C[D) =
¸
j
¸
E∈E
µ(E)
¸
D
i
⊂E
µ(D
i
)
µ(E)
ϕ
µ(C
j
∩ D
i
)
µ(D
i
)
30 2 General Thermodynamic Formalism
By the concavity of ϕ (see the proof of Lemma 1.17) one has ϕ(
¸
a
i
x
i
) ≥
¸
a
i
ϕ(x
i
) where
a
i
=
µ(D
i
)
µ(E)
, x
i
=
µ(C
j
∩ D
i
)
µ(D
i
)
Hence
H
µ
(C[D) ≤
¸
j
¸
E∈E
µ(E) ϕ
µ(C
j
∩ E)
µ(E)
= H
µ
(C[E) .
To see (b) one notes that C ∨ D = D when D ⊃ C. For (c) one writes
H
µ
(C ∨ D[E) = H
µ
(C ∨ D∨ E) −H
µ
(D∨ E) +H
µ
(D∨ E) −H
µ
(E)
= H
µ
(C[D∨ E) +H
µ
(D[E)
≤ H
µ
(C[E) +H
µ
(D[E)
by (a). Finally
H
µ
(C) = H
µ
(C ∨ D) −H
µ
(D[C)
≤ H
µ
(C ∨ D) = H
µ
(D) +H
µ
(C[D) . ¯.
2.2. Lemma. Let T be an endomorphism of a probability space (X, B, µ), C
and D ﬁnite partitions. Then
(a) H
µ
(T
−k
C[T
−k
D) = H
µ
(C[D) for k ≥ 0,
(b) h
µ
(T, C) ≤ h
µ
(T, D) +H
µ
(C[D),
(c) h
µ
(T, C ∨ ∨ T
−n
C) = h
µ
(T, C).
Proof. As µ is Tinvariant,
H
µ
(T
−k
C[T
−k
D) = H
µ
(T
−k
C ∨ T
−k
D) −H
µ
(T
−k
D)
= H
µ
(C ∨ D) −H
µ
(D) = H
µ
(C[D) .
Using Lemma 2.1
H
µ
(C ∨ ∨ T
−m+1
C) ≤ H
µ
(D∨ ∨ T
−m+1
D)
+H
µ
(C ∨ ∨ T
−m+1
C[D∨ ∨ T
−m+1
D)
≤ H
µ
(D∨ ∨ T
−m+1
D)
+
m−1
¸
k=0
H
µ
(T
−k
C[D∨ ∨ T
−m+1
D)
≤ H
µ
(D∨ ∨ T
−m+1
D) +
m−1
¸
k=0
H
µ
(T
−k
C[T
−k
D)
= H
µ
(D∨ ∨ T
−m+1
D) +mH
µ
(C[D) .
Dividing by m and letting m → ∞,
A. Entropy 31
h
µ
(T, C) ≤ h
µ
(T, D) +H
µ
(C[D) .
Set D = C ∨ ∨ T
−n
C. Then
1
m
H
µ
(D∨ ∨ T
−m+1
D) =
1
m
H
µ
(C ∨ ∨ T
−m−n+1
C) .
Letting m → ∞, (as
m
m+n
→ 1) we get
h
µ
(T, D) = h
µ
(T, C) . ¯.
2.3. Lemma. Let X be a compact metric space, µ ∈ M(X), ε > 0 and C a
ﬁnite Borel partition. There is a δ > 0 so that H
µ
(C[D) < ε whenever D is a
partition with diam(D) < δ.
Proof. Let C = ¦C
1
, . . . , C
n
¦. In Lemma 1.23 we showed that, for any α > 0,
one could ﬁnd δ > 0 such that whenever D satisﬁes diam(D) < δ there is a
E = ¦E
1
, . . . , E
n
¦ ⊂ D with
µ(E
i
∆C
i
) < α.
The expression
H
µ
(C[E) =
¸
i,j
µ(E
j
) ϕ
µ(C
j
∩ E
i
)
µ(E
i
)
depends continuously upon the numbers
µ(C
j
∩ E
i
) and µ(E
i
) =
¸
j
µ(C
j
∩ E
i
)
and vanishes when µ(C
j
∩ E
i
) = δ
ij
µ(E
i
). Hence, for α small, H
µ
(C[E) < ε.
Then H
µ
(C[D) ≤ H
µ
(C[E) < ε by 2.1 (a). ¯.
2.4. Proposition. Suppose T : X → X is a continuous map of a com
pact metric space, µ ∈ M
T
(X) and that D
n
is a sequence of partitions with
diam(D
n
) → 0. Then
h
µ
(T) = lim
n→∞
h
µ
(T, D
n
) .
Proof. Of course h
µ
(T) ≥ limsup
n
h
µ
(T, D
n
). Consider any partition C. By
Lemmas 2.2 (b) and 2.3
h
µ
(T, C) ≤ liminf
n
h
µ
(T, D
n
) .
Varying C, h
µ
(T) ≤ liminf
n
h
µ
(T, D
n
). ¯.
A homeomorphism T : X → X is called expansive if there exists ε > 0 so
that
d(T
k
x, T
k
y) ≤ ε for all k ∈ Z ⇒ x = y .
32 2 General Thermodynamic Formalism
2.5. Proposition. Suppose T : X → X is a homeomorphism with expansive
constant ε. Then h
µ
(T) = h
µ
(T, D) whenever µ ∈ M
T
(X), and diam(D) ≤ ε.
Proof. Let D
n
= T
n
D ∨ ∨ D ∨ ∨ T
−n
D. Then diam(D
n
) → 0 using
expansiveness. Hence h
µ
(T) = lim
n
h
µ
(T, D
n
). But h
µ
(T, D
n
) = h
µ
(T, D) by
Lemma 2.2 (c). ¯.
Consider the case of σ : Σ
A
→ Σ
A
and standard partition U=¦U
1
, . . . , U
n
¦
where U
i
= ¦x ∈ Σ
A
: x
0
= i¦. Then σ is expansive and 2.5 gives that
h
µ
(σ) = h
µ
(σ, U) for µ ∈ M
σ
(Σ
A
). Now h
µ
(σ, U) is what we denoted by
s(µ) in Chapter 1. That s(µ) = h
µ
(σ) implies that the number s(µ) does
not depend on the homeomorphism σ and partition U, but only on σ as an
automorphism of the probability space (Σ
A
, B, µ) (because of the deﬁnition
of h
µ
(σ)).
2.6. Lemma. h
µ
(T
n
) = nh
µ
(T) for n > 0.
Proof. Let C be a partition and E = C ∨ ∨ T
−n+1
C. Then
nh
µ
(T, C) = lim
m→∞
n
nm
H
µ
(C ∨ ∨ T
−nm+1
C)
= lim
m→∞
1
m
H
µ
(E ∨ T
−n
E ∨ ∨ T
(−m+1)n
E)
= h
µ
(T
n
, E) ≤ h
µ
(T
n
) = nh
µ
(T) .
Varying C, nh
µ
(T) ≤ h
µ
(T
n
). On the other hand
h
µ
(T
n
, C) ≤ h
µ
(T
n
, E)
by 2.2 (b) and 2.1 (b). Hence
h
µ
(T
n
) = sup
C
h
µ
(T
n
, C) ≤ nsup
C
h
µ
(T, C) = nh
µ
(T) . ¯.
B. Pressure
Throughout this section T : X → X will be a ﬁxed continuous map on the
compact metric space X. We will deﬁne the pressure P(φ) of φ ∈ C(X) in a
way which generalizes Section D in Chapter 1.
Let U be a ﬁnite open cover of X, W
m
(U) the set of all mstrings
U = U
i
0
U
i
1
U
i
m−1
of members of U. One writes m = m(U),
X(U) = ¦x ∈ X : T
k
x ∈ U
i
k
for k = 0, . . . , m−1¦
B. Pressure 33
S
m
φ(U) = sup
m−1
¸
k=0
φ(T
k
x) : x ∈ X(U)
¸
.
In case X(U) = ∅, we let S
m
φ(U) = −∞. We say that Γ ⊂ W
m
(U) covers X
if X =
¸
U∈Γ
X(U). Finally one deﬁnes
Z
m
(φ, U) = inf
Γ
¸
U∈Γ
exp(S
m
φ(U)),
where Γ runs over all subsets of W
m
(U) covering X.
2.7. Lemma. The limit
P(φ, U) = lim
m→∞
1
m
log Z
m
(φ, U)
exists and is ﬁnite.
Proof. If Γ
m
⊂ W
m
(U) and Γ
n
⊂ W
n
(U) each cover X, then
Γ
m
Γ
n
= ¦UV : U ∈ Γ
m
, V ∈ Γ
n
¦ ⊂ W
m+n
(U)
covers X. One sees that
S
m+n
φ(UV) ≤ S
m
φ(U) +S
n
φ(V)
and so
¸
UV∈Γ
m
Γ
n
exp(S
m+n
φ(UV)) ≤
¸
U∈Γ
m
exp(S
m
φ(U))
¸
V∈Γ
n
exp(S
n
φ(V)).
Thus
Z
m+n
(φ, U) ≤ Z
m
(φ, U) Z
n
(φ, U)
and Z
m
(φ, U) ≥ e
−mφ
. Hence a
m
= log Z
m
(φ, U) satisﬁes the hypotheses of
Lemma 1.18. ¯.
2.8. Proposition. The limit
P(φ) = lim
diam(U)→0
P(φ, U)
exists (but may be +∞).
Proof. Suppose V is an open cover reﬁning U, i.e., every V ∈ V lies in some
U(V ) ∈ U. For V ∈ W
m
(V) let U(V) = U(V
i
0
) U(V
i
m−1
). If Γ
m
⊂ W
m
(V)
covers X, then U(Γ
m
) = ¦U(V) : V ∈ Γ
m
¦ ⊂ W
m
(U) covers X.
Let γ = γ(φ, U) = sup¦[φ(x) −φ(y)[ : x, y ∈ U for some U ∈ U¦.
Then S
m
φ(U(V)) ≤ S
m
φ(V) + mγ and so Z
m
(φ, U) ≤ e
mγ
Z
m
(φ, V), which
gives
34 2 General Thermodynamic Formalism
P(φ, U) ≤ P(φ, V) +γ .
Now for any U, all V with small diameter reﬁne U and so
P(φ, U) −γ(φ, U) ≤ liminf
diam(V)→0
P(φ, V) .
Letting diam(U) → 0, γ(φ, U) → 0 and
limsup
diam(U)→0
P(φ, U) ≤ liminf
diam(V)→0
P(φ, V) .
We are done. ¯.
In cases where confusion may arise we write the topological pressure P(φ)
as P
T
(φ).
2.9. Lemma. Let S
n
φ(x) =
¸
n−1
k=0
φ(T
k
x). Then
P
T
n(S
n
φ) = nP
T
(φ) for n > 0 .
Proof. Let V = U ∨ ∨ T
−n+1
U. Then W
m
(V) and W
mn
(U) are in one
toone correspondence; for U = U
i
0
U
i
1
U
i
mn−1
let V = V
i
0
V
i
m−1
where
V
i
k
= U
i
kn
∩ T
−1
U
i
kn+1
∩ ∩ T
−n+1
U
i
kn+n−1
. One sees that X(U) = X(V)
and S
T
mn
φ(U) = S
T
n
m
(S
n
φ)(V). Thus one gets
Z
T
mn
(φ, U) = Z
T
n
m
(S
n
φ, V) and nP
T
(φ, U) = P
T
n(S
n
φ, V) .
As diam(U) → 0, diam(V) → 0 and so nP
T
(φ) = P
T
n(S
n
φ). ¯.
We now come to our ﬁrst interesting result about the pressure P(φ).
2.10. Theorem. Let T : X → X be a continuous map on a compact metric
space and φ ∈ C(X). Then
h
µ
(T) +
φ dµ ≤ P
T
(φ),
for any µ ∈ M
T
(X).
We will ﬁrst need a couple of lemmas.
2.11. Lemma. Suppose D is a Borel partition of X such that each x ∈ X is
in the closures of at most M members of D. Then
h
µ
(T, D) +
φ dµ ≤ P
T
(φ) + log M .
B. Pressure 35
Proof. Let U be a ﬁnite open cover of X each member of which intersects
at most M members of D. Let Γ
m
⊂ W
m
(U) cover X. For each B ∈ D
m
=
D∨ ∨ T
−m+1
D pick x
B
∈ B with
B
S
m
φ dµ ≤ µ(B) S
m
φ(x
B
). Now
h
µ
(T, D) +
φ dµ ≤
1
m
H
µ
(D
m
) +
S
m
φ dµ
≤
1
m
¸
B
µ(B)(−log µ(B) +S
m
φ(x
B
))
≤
1
m
log
¸
B
exp(S
m
φ(x
B
))
by Lemma 1.1. For each x
B
pick U
B
∈ Γ
m
with x
B
∈ X(U
B
). This map
B → U
B
is at most M
m
to one. As S
m
φ(x
B
) ≤ S
m
φ(U
B
), one has
h
µ
(T, D) +
φ dµ ≤
1
m
log
¸
U∈Γ
m
M
m
exp(S
m
φ(U))
≤ log M +
1
m
log Z
m
(φ, U) .
Letting m → ∞ and then diam(U) → 0, we obtain the desired inequality. ¯.
2.12. Lemma. Let A be a ﬁnite open cover of X. For each n > 0 there is a
Borel partition D
n
of X so that
(a) D lies inside some member of T
−k
A for each D ∈ D
n
and k = 0, . . . , n−1,
(b) at most n[A[ sets in D
n
can have a point in all their closures.
Proof. Let A = ¦A
1
, . . . , A
m
¦ and g
1
, . . . , g
m
be a partition of unity subor
dinate to A. Then G = (g
1
, . . . , g
m
) : X → s
m−1
⊂ R
m
where s
m−1
is an
m− 1 dimensional simplex. Now U = ¦U
1
, . . . , U
m
¦ is an open cover where
U
i
= ¦x ∈ s
m−1
: x
i
> 0¦ and G
−1
U
i
⊂ A
i
. Since (s
m−1
)
n
is nm − n
dimensional, there is a Borel partition D
∗
n
of (s
m−1
)
n
so that
(a’) each member of D
∗
n
lies in some U
i
1
U
i
n
, and
(b’) at most nm members of D
∗
n
can have a common point in all their closures.
Then D
n
= L
−1
D
∗
n
works where
L = (G, G◦ T, . . . , G◦ T
n−1
) : X → (s
m−1
)
n
. ¯.
Proof of 2.10. Let C be a Borel partition and ε > 0. By Lemma 2.3 ﬁnd an
open cover A so that H
µ
(C[D) < ε whenever D is a partition every member of
which is contained in some member of A. Fix n > 0, let E = C∨ ∨T
−n+1
C
and D
n
as in Lemma 2.12. Then (see the proof of 2.6)
36 2 General Thermodynamic Formalism
h
µ
(T, C) +
φ dµ ≤
1
n
h
µ
(T
n
, E) +
S
n
φ dµ
≤
1
n
h
µ
(T
n
, D
n
) +
S
n
φ dµ
+
1
n
H
µ
(E[D
n
)
≤
1
n
(P
T
n(S
n
φ) + log(n[A[)) +
1
n
H
µ
(E[D
n
)
by Lemma 2.11. Now
H
µ
(E[D
n
) ≤
n−1
¸
k=0
H
µ
(T
−k
C[D
n
) .
Since D
n
reﬁnes T
−k
A for each k, one has H
µ
(T
−k
C[D
n
) < ε (since µ is
Tinvariant, T
−k
A bears the same relation to T
−k
C as A to C). Hence, using
2.9,
h
µ
(T, C) +
φ dµ ≤ P
T
(φ) +
1
n
log(n[A[) +ε .
Now let n → ∞ and then ε → 0. ¯.
2.13. Proposition. Let T
1
: X
1
→ X
1
, T
2
: X
2
→ X
2
be continuous maps on
compact metric spaces, π : X
1
→ X
2
continuous and onto satisfying π ◦ T
1
=
T
2
◦ π. Then
P
T
2
(φ) ≤ P
T
1
(φ ◦ π)
for φ ∈ C(X
2
).
Proof. For U an open cover of X
2
one sees that
P
T
2
(φ, U) = P
T
1
(φ ◦ π, π
−1
U) .
As in the proof of 2.8
P
T
1
(φ ◦ π, π
−1
U) ≤ P
T
1
(φ ◦ π) +γ(φ ◦ π, π
−1
U) .
But γ(φ◦π, π
−1
U) = γ(φ, U) → 0 as diam(U) → 0. Hence, letting diam(U) →
0 we get P
T
2
(φ) ≤ P
T
1
(φ ◦ π). ¯.
C. Variational Principle
Let U be a ﬁnite open cover of X. We say that Γ ⊂ W
∗
(U) =
¸
m>0
W
m
(U)
covers K ⊂ X if K ⊂
¸
U∈Γ
X(U). For λ > 0 and Γ ⊂ W
∗
(U) deﬁne
Z(Γ, λ) =
¸
U∈Γ
λ
m(U)
exp(S
m(U)
φ(U)).
C. Variational Principle 37
2.14. Lemma. Let P = P(φ, U) and λ > 0. Suppose that Z(Γ, λ) < 1 for
some Γ covering X. Then λ ≤ e
−P
.
Proof. As X is compact we may take Γ ﬁnite and Γ ⊂
¸
M
m=1
W
m
(U). Then
Z(Γ
n
, λ) ≤ Z(Γ, λ)
n
where Γ
n
= ¦U
1
U
2
U
n
: U
i
∈ Γ¦. Letting Γ
∗
=
¸
∞
n=1
Γ
n
, one has
Z(Γ
∗
, λ) =
∞
¸
n=1
Z(Γ
n
, λ) < ∞.
Fix N and consider x ∈ X. Since Γ covers X, one can ﬁnd U = U
1
U
2
U
n
∈
Γ
∗
with
(a) x ∈ X(U), and
(b) N ≤ m(U) < N +M.
Let U
∗
be the ﬁrst N symbols of U. Then
S
N
φ(U
∗
) ≤ S
m(U)
φ(U) +Mφ .
For Γ
N
the set of U
∗
so obtained,
λ
N
¸
Γ
N
exp S
N
φ(U
∗
) ≤ max
¸
1, λ
−M
¸
e
Mφ
Z(Γ
∗
, λ),
or λ
N
Z
N
(φ, U) ≤ constant. It follows that λ ≤ e
−P
. ¯.
Let δ
x
be the unitmeasure concentrated on the point x. Deﬁne
δ
x,n
= n
−1
(δ
x
+δ
Tx
+ +δ
T
n−1
x
)
and V (x) = ¦µ ∈ M(X) : δ
x,n
k
→ µ for some n
k
→ ∞¦ .
V (x) = ∅ as M(X) is a compact metric space. Now T
∗
δ
x,n
= δ
Tx,n
and for
f ∈ C(X), [T
∗
δ
x,n
(f)−δ
x,n
(f)[ = n
−1
[f(T
n
x)−f(x)[ ≤ 2n
−1
f. This shows
V (x) ⊂ M
T
(X).
Let E be a ﬁnite set, a = (a
0
, . . . , a
k−1
) ∈ E
k
. One deﬁnes the distribution
µ
a
on E by
µ
a
(e) = k
−1
(number of j with a
j
= e)
and H(a) = −
¸
e∈E
µ
a
(e) log µ
a
(e) .
2.15. Lemma. Let x ∈ X, µ ∈ V (x), U a ﬁnite open cover of X and ε > 0.
There are m and arbitrarily large N for which one can ﬁnd U ∈ W
N
(U)
satisfying the following
(a) x ∈ X(U),
(b) S
N
φ(U) ≤ N(
φdµ +γ(U) +ε),
38 2 General Thermodynamic Formalism
(c) U contains a subword of length km ≥ N − m which, when viewed as a =
a
0
, . . . , a
k−1
∈ (U
m
)
k
satisﬁes
1
m
H(a) ≤ h
µ
(T) +ε.
Proof. Let U = ¦U
1
, . . . , U
q
¦. Recall that
γ(U) = sup¦[φ(y) −φ(z)[ : y, z ∈ U
i
for some i¦ .
Pick a Borel partition C = ¦C
1
, . . . , C
q
¦ with C
i
⊂ U
i
. There is an m so that
1
m
H
µ
(C ∨ ∨ T
−m+1
C) ≤ h
µ
(T, C) +
ε
2
≤ h
µ
(T) +
ε
2
Let δ
x,n
j
→ µ. For n
> n one has
δ
x,n
=
n
n
δ
x,n
+
n
−n
n
δ
T
n
x,n
−n
.
If we replaced n
k
by the nearest multiple of m, this formula shows that µ
would still be the limit. Thus we assume n
j
= mk
j
.
Let D
1
, . . . , D
t
be the nonempty members of C∨ ∨T
−m+1
C and for each
D
i
ﬁnd a compact K
i
⊂ D
i
with µ(D
i
`K
i
) < β (β > 0 small). Each D
i
is
contained in some member of U∨ ∨T
−m+1
U and one can ﬁnd an open set
V
i
⊃ K
i
for which this is still true. Furthermore we may assume V
i
∩ V
j
= ∅
for i = j. Now enlarge each V
i
to a Borel set V
∗
i
still contained in a member
of U ∨ ∨ T
−m+1
U and so that ¦V
∗
1
, . . . , V
∗
t
¦ is a Borel partition of X.
Now ﬁx n
j
= mk
j
. Let M
i
be the number of s ∈ [0, n
j
) with T
s
x ∈ V
∗
i
and M
i,r
the number of such s ≡ r (mod m).
Deﬁne
p
i,r
= M
i,r
/k
j
and p
i
= M
i
/n
j
=
1
m
(p
i,0
+ +p
i,m−1
) . As δ
x,n
j
→ µ, one has
liminf
j→∞
p
i
≥ µ(K
i
) ≥ µ(D
i
) −β,
and limsup
j→∞
p
i
≤ µ(K
i
)+tβ ≤ µ(D
i
)+tβ. For β small enough and j large
enough one has
1
m
−
¸
i
p
i
log p
i
≤
1
m
−
¸
i
µ(D
i
) log µ(D
i
)
+
ε
2
≤ h
µ
(T) +ε .
By the concavity of ϕ(x) = −xlog x (see 1.17)
ϕ(p
i
) ≥
m−1
¸
r=0
1
m
ϕ(p
i,r
)
C. Variational Principle 39
and so
¸
i
ϕ(p
i
) ≥
1
m
m−1
¸
r=0
¸
i
ϕ(p
i,r
) .
For some r ∈ [0, m) one must have
¸
i
ϕ(p
i,r
) ≤
¸
i
ϕ(p
i
) and so
1
m
¸
i
ϕ(p
i,r
) ≤ h
µ
(T) +ε .
For N = n
j
+r with j large we form U = U
0
U
1
U
N−1
∈ U
N
as follows. For
s < r pick U
s
∈ U containing T
s
x. For each V
∗
i
we choose U
0,i
∩ T
−1
U
1,i
∩
∩ T
−m+1
U
m−1,i
⊃ V
∗
i
. For s > r we write s = r + mp + q with p ≥ 0,
m > q ≥ 0, pick i with T
r+mp
x ∈ V
∗
i
and let U
s
= U
q,i
. Letting
a
p
= U
0,i
U
1,i
U
m−1,i
we have
U = U
0
U
r−1
a
0
a
1
a
k
j
−1
.
Now a = (a
0
a
1
a
k
j
−1
) has its distribution µ
a
on U
m
given by the proba
bilities ¦p
i,r
¦
t
i=1
and some zeros.
So
1
m
H(a) =
1
m
¸
i
ϕ(p
i,r
) ≤ h
µ
(T) +ε.
We have yet to check (b). Since δ
x,n
j
→ µ, for j large we will have
1
N
δ
x,N
(φ)−
φ dµ
< ε or S
N
φ(x) ≤ N(
φ dµ +ε). As x ∈ X(U), S
N
φ(U) ≤ S
N
φ(x) +
Nγ(U). ¯.
2.16. Lemma. Fix a ﬁnite set E and h ≥ 0. Let R(k, h)=¦a ∈ E
k
: H(a)≤h¦.
Then
limsup
k→∞
1
k
log [R(k, h)[ ≤ h.
Proof. For any distribution ν on E and α ∈ (0, 1) consider
R
k
(ν) = ¦a ∈ E
k
: [µ
a
(e) −ν(e)[ < α ∀e ∈ E¦ .
Let µ be the Bernoulli measure on Σ =
¸
∞
i=0
E with the distribution
µ(e) = (1 −α)ν(e) +α/[E[ .
Each a ∈ R
k
(ν) corresponds to a cylinder set C
a
of Σ. Since each e ∈ E
occurs in a at most k(ν(e) +α) times,
µ(C
a
) ≥
¸
e
µ(e)
k(ν(e)+α)
.
As the C
a
are disjoint and have total measure 1,
40 2 General Thermodynamic Formalism
1 ≥ [R
k
(ν)[
¸
e
µ(e)
k(ν(e)+α)
,
or
1
k
log [R
k
(ν)[ ≤
¸
e
−(ν(e) +α) log µ(e)
≤ H(µ) +
¸
e
3α[ log µ(e)[ .
As µ(e) ≥ α/[E[, we get
1
k
log [R
k
(ν)[ ≤ H(µ) + 3α[E[(log [E[ −log α) .
When α → 0, the second term on the right approaches 0 and H(µ) → H(ν)
uniformly in ν. Hence, for any ε > 0 one can ﬁnd α small enough that
1
k
log [R
k
(ν)[ ≤ H(µ) +ε,
for all k and ν.
Once α is so chosen, let ^ be a ﬁnite set of distributions on E so that
(a) H(ν) ≤ h for ν ∈ ^, and
(b) if H(ν
) ≤ h then for some ν ∈ ^ one has
[ν
(e) −ν(e)[ < α for all e.
Then R(k, h) ⊂
¸
ν∈N
R
k
(ν),
1
k
log [R(k, h)[ ≤
1
k
log [^[ +h +ε
and limsup
k→∞
1
k
log [R(k, h)[ ≤ h +ε.
Now let ε → 0. ¯.
2.17. Variational Principle. Let T : X → X be a continuous map on a
compact metric space and φ ∈ C(X). Then
P
T
(φ) = sup
µ
h
µ
(T) +
φ dµ
where µ runs over M
T
(X).
Proof. Let U be a ﬁnite cover of X and > 0. For each m > 0 let X
m
be the
set of points x ∈ X for which 2.15 holds with this m and some µ ∈ V (x). By
2.15 X =
¸
m
X
m
since V (x) = ∅. For u ∈ R let Y
m
(u) be the set of x ∈ X
m
for which 2.15 holds for some µ ∈ V (x) with
φdµ ∈ [u −ε, u +ε]. Set
C. Variational Principle 41
c = sup
µ
h
µ
(T) +
φ dµ
.
For x ∈ Y
m
(u) the µ satisﬁes h
µ
(T) ≤ c −u +ε.
The a ∈ (U
m
)
k
appearing in 2.15 (c) for x ∈ Y
m
(u) lie in R(k, m(c −u +
2ε), U
m
). The number of possibilities for U for any ﬁxed N = km is hence at
most
b(N) = [E[
m
[R(k, m(c −u + 2ε), U
m
)[ .
By 2.16
limsup
N→∞
1
N
log b(N) ≤ c −u + 2ε .
Let Γ = Γ
m,u
be the collection of all U showing up in the present situation
for some N greater than a ﬁxed N
0
. Then Γ covers Y
m
(u) and
Z(Γ, λ) ≤
∞
¸
N=N
0
λ
N
b(N) exp(N(u + 2ε +γ(U))) .
For large enough N
0
, b(N) ≤ exp(N(c −u + 3ε)) and
Z(Γ, λ) ≤
∞
¸
N=N
0
λ
N
exp(N(c + 5ε +γ(U))) .
≤
∞
¸
N=N
0
β
N
=
β
N
0
1 −β
,
where β = λexp(c + 5ε +γ(U)) < 1.
We have seen that for λ < exp(−(c + 5ε + γ(U))) any Y
m
(u) can be
covered by Γ ⊂ W
∗
(U) with Z(Γ, λ) as small as desired. As X =
¸
∞
m=1
X
m
and X
m
= Y
m
(u
1
) ∪ ∪Y
m
(u
r
) where u
1
, . . . , u
r
are εdense in [−φ, φ],
taking unions of such Γ
s we obtain a Γ covering X with Z(Γ, λ) < 1. By
Lemma 2.14, λ ≤ e
−P(φ,U)
or
P(φ, U) ≤ c + 5ε +γ(U) .
As ε was arbitrary, P(φ, U) ≤ c +γ(U).
Finally
P(φ) ≤ lim
diam(U)→0
P(φ, U)
≤ lim
diam(U)→0
(c +γ(U)) = c .
The inequality c ≤ P(φ) follows from Theorem 2.10. ¯.
2.18. Corollary. Suppose ¦X
α
¦
α∈Λ
is a family of compact subsets of X and
TX
α
⊂ X
α
for each α. Then
P
T
(φ) = sup
α
P
T
X
α
(φ[
X
α
) .
42 2 General Thermodynamic Formalism
Proof. If µ ∈ M
T
(X
α
), then µ ∈ M
T
(X) and
P
T
(φ) ≥ h
µ
(T) +
φ dµ.
Hence
P
T
(φ) ≥ sup
µ∈M
T
(X
α
)
h
µ
(T) +
φ dµ
= P
T
X
α
(φ[
X
α
) .
If x ∈ X
α
, then V (x) ⊂ M
T
(X
α
) and so
c
= sup
h
µ
(T) +
φ dµ : µ ∈
¸
x∈X
V (x)
¸
≤ sup
α
P
T
X
α
(φ[
X
α
) .
In the proof of 2.17 what was actually used about the number c was c ≥
h
µ
(T) +
φ dµ for µ ∈ V (x). So c
would work just as well there to yield
P
T
(φ) ≤ c
. ¯.
D. Equilibrium States
If µ ∈ M
T
(X) satisﬁes h
µ
(T)+
φ dµ = P
T
(φ), then µ is called an equilibrium
state for φ (w.r.t. T). The Gibbs state µ
φ
of φ ∈ F
A
in Chapter 1 was shown
to be the unique equilibrium state for such a φ.
2.19. Proposition. Suppose that for some ε > 0 one has h
µ
(T, D) = h
µ
(T)
whenever µ ∈ M
T
(X) and diam(D) < ε. Then every φ ∈ C(X) has an
equilibrium state.
Proof. We show that µ → h
µ
(T) is upper semicontinuous on M
T
(X). Then
µ → h
µ
(T) +
φdµ will be also, and the proposition follows from 2.17 and
the fact that an u.s.c. function on a compact space assumes its supremum.
Fixing µ ∈ M
T
(X), α > 0, and D = ¦D
1
, . . . , D
n
¦ with diam(D) < ε, one
has
1
m
H
µ
(D∨ ∨T
−m+1
D) ≤ h
µ
(T) +α for some m. Let β > 0 and pick a
compact set K
i
0
,...,i
m−1
⊂
¸
m−1
k=0
T
−k
D
i
k
with
µ
¸
k
T
−k
D
i
k
K
i
0
,...,i
m−1
< β .
Then D
i
⊃ L
i
=
¸
m−1
j=0
¸
¦T
j
K
i
0
,...,i
m−1
: i
j
= i¦. As the L
i
are disjoint
compact sets, one can ﬁnd a partition D
= ¦D
1
, . . . , D
n
¦ with diam(D
) < ε
and L
i
⊂ int(D
i
). One then has
K
i
0
,...,i
m−1
⊂ int
¸
k
T
−k
D
i
k
.
References 43
If ν is close to µ in the weak topology, one will have
ν
¸
k
T
−k
D
i
k
≥ µ(K
i
0
,...,i
m−1
) −β
and
ν
¸
k
T
−k
D
i
k
− µ
¸
k
T
−k
D
i
k
≤ 2βn
m
. For β small enough, this
implies
h
ν
(T) = h
ν
(T, D
) ≤
1
m
H
ν
(D
∨ ∨ T
−m+1
D
)
≤
1
m
H
µ
(D∨ ∨ T
−m+1
D) +α ≤ h
µ
(T) + 2α. ¯.
2.20. Corollary. If T is expansive, every φ ∈ C(X) has an equilibrium state.
Proof. Recall 2.5. ¯.
One notices that the condition in 2.19 has nothing to do with φ. Taking
φ = 0, one deﬁnes the topological entropy of T by
h(T) = P
T
(0) .
The motivation for this chapter comes from two places: the theory of Gibbs
states from statistical mechanics and topological entropy from topological dy
namics (see references). Conditions on φ become important for the uniqueness
of equilibrium state and then only after stringent conditions have been placed
on the homeomorphism T. The Axiom A diﬀeomorphisms will be close enough
to the subshifts σ : Σ
A
→ Σ
A
so that one can prove uniqueness statements.
References
The deﬁnition of h
µ
(T) is due to Kolmogorov and Sinai (see [2]). For expansive T
Ruelle [15] deﬁned P
T
(φ) and proved Theorems 2.10, 2.17 and 2.20. For general
T the deﬁnition and results are due to Walters [16].
In the transition fromΣ
A
to a general compact metric space X, most of the work
has to do with the more complicated topology of X. Walters’ paper is therefore
closely related to earlier work on the topological entropy h(T), i.e., the case φ =
0. The deﬁnition of h(T) was made by Adler, Konheim and McAndrew [1]. The
theorems for this case are due to Goodwyn [10] (Theorem 2.10), Dinaburg [6]
(X ﬁnite dimensional, 2.17), Goodman [8] (general X, 2.17), and Goodman [9]
(2.20). For subshifts these results were proved earlier by Parry [14]. The proofs
we have given in these notes are adaptations of [4].
Gureviˇc [11] gives a T where φ = 0 has no equilibrium states and Misiurewicz
[13] gives such a T which is a diﬀeomorphism. The condition in 2.19 is satisﬁed by
a class of maps which includes all aﬃne maps on Lie groups [3] and Misiurewicz
[13] showed that equilibrium states exist under a somewhat weaker condition.
44 2 General Thermodynamic Formalism
Ruelle [15] showed that for expansive T a Baire set of φ have unique equilibrium
states. Goodman [9] gives a minimal subshift where φ = 0 has more than one
equilibrium state. I believe mathematical physicists know of speciﬁc φ on Σ
n
which do not have unique equilibrium states; in statistical mechanics one looks at
Z
m
actions instead of just homeomorphisms and gets nonuniqueness for m ≥ 2
even with simple φ’s. Uniqueness was proved in [5] for certain φ when T satisﬁes
expansiveness and a very restrictive condition called speciﬁcation; this result has
been carried over to ﬂows by FrancoSanchez [7].
Finally we mention a very interesting result in a diﬀerent direction. Let T :
M → M be a continuous map on a compact manifold and λ an eigenvalue of
the map T
∗
: H
1
(M) → H
1
(M) on onedimensional homology. Then Manning
[12] showed h(T) ≥ log λ. It is conceivable that this inequality is true for λ for
any H
k
(M) (not just k = 1) provided T is a diﬀeomorphism.
1. R. L. Adler, A. G. Konheim and M. H. McAndrew: Topological entropy, Trans.
Amer. Math. Soc. 114 (1965), 309319.
2. P. Billingsley: Ergodic Theory and Information, Wiley (1965). (
1
)
3. R. Bowen: Entropyexpansive maps, Trans. Amer. Math. Soc. 164 (1972), 323
331.
4. R. Bowen: Topological entropy for noncompact sets, Trans. Amer. Math. Soc.
184 (1973), 125136.
5. R. Bowen: Some systems with unique equilibrium states, Math. Systems Theory
8 (1974/75), no. 3, 193202.
6. E. I. Dinaburg: On the relations among various entropy characteristics of dy
namical systems, Math. USSR Izvestia 5 (1971), 337378.
7. E. FrancoSanchez: Flows with unique equilibrium states, Ph.D. thesis. Univ. of
California, Berkeley, 1974.
8. T. N. T. Goodman: Relating topological entropy and measure entropy, Bull.
London Math. Soc. 3 (1971), 176180.
9. T. N. T. Goodman: Maximal measures for expansive homeomorphisms, J. Lon
don Math. Soc. (2) 5 (1972), 439444.
10. L. W. Goodwyn: Topological entropy bounds measuretheoretic entropy, Proc.
Amer. Math. Soc. 23 (1969), 679688.
11. B. M. Gureviˇc: Topological entropy of denumerable Markov chains, Soviet Math.
Dokl. 10 (1969), 911915.
12. A. Manning: Topological entropy and the ﬁrst homology group, in “Dynamical
Systems”  Warwick 1974 (Proc. Symp. Appl. Topology and Dynamical Systems,
Univ. Warwick, Coventry, 1973/1974), pp. 185190. Lect. Notes in Math. 468,
Springer, Berlin, 1975.
13. M. Misiurewicz: Diﬀeomorphism without any measure with maximal entropy,
Bull. Acad. Polon. Sci. S´er. Sci. Math. Astronom. Phys. 21 (1973), 903–910.
14. W. Parry: Intrinsic Markov chains, Trans. Amer. Math. Soc. 112 (1964), 5566.
15. D. Ruelle: Statistical mechanics on a compact set with Z
ν
action satisfying
expansiveness and speciﬁcation, Trans. Amer. Math. Soc. 185 (1973), 237251.
16. P. Walters: A variational principle for the pressure of continuous transforma
tions, Amer. J. Math. 97 (1975), no. 4, 937971.
1
Reprint: Robert E. Krieger Publishing Co., Huntington, N.Y., 1978 (note of the
editor).
3
Axiom a Diﬀeomorphisms
A. Deﬁnition
We now suppose that f : M → M is a diﬀeomorphism of a compact C
∞
Riemannian manifold M. Then the derivative of f can be considered a map
df : TM → TM where TM =
¸
x∈M
T
x
M is the tangent bundle of M and
df
x
: T
x
M → T
f(x)
M.
Deﬁnition. A closed subset Λ ⊂ M is hyperbolic if f(Λ) = Λ and each
tangent space T
x
M with x ∈ Λ can be written as a direct sum
T
x
M = E
u
x
⊕E
s
x
of subspaces so that
(a) Df(E
s
x
) = E
s
f(x)
, Df(E
u
x
) = E
u
f(x)
;
(b) there exist constants c > 0 and λ ∈ (0, 1) so that
Df
n
(v) ≤ cλ
n
v when v ∈ E
s
x
, n ≥ 0
and
Df
−n
(v) ≤ cλ
n
v when v ∈ E
u
x
, n ≥ 0;
(c) E
s
x
, E
u
x
vary continuously with x.
Remark. Condition (c) actually follows from the others. E
u
=
¸
x∈Λ
E
u
x
and E
s
=
¸
x∈Λ
E
s
x
are continuous subbundles of T
Λ
M =
¸
x∈Λ
T
x
M and
T
Λ
M = E
u
⊕E
s
.
A point x ∈ M is nonwandering if
U ∩
¸
n>0
f
n
U = ∅,
for every neighborhood U of x. The set Ω = Ω(f) of all nonwandering points
is seen to be closed and finvariant. A point x is periodic if f
n
x = x for some
n > 0; clearly such an x is in Ω(f).
46 3 Axiom a Diﬀeomorphisms
Deﬁnition. f satisﬁes Axiom A if Ω(f) is hyperbolic and
Ω(f) = ¦x : x is periodic¦.
This deﬁnition is due to Smale [14]. A type of f studied extensively by
Russian mathematicians is the Anosov diﬀeomorphism: f is Anosov if M
is hyperbolic [2]. We shall see a little later that such diﬀeomorphisms always
satisfy Axiom A. Right now we mention that it is unknown whether Ω(f) = M
for every Anosov f. The reader should study the examples in [14].
The Riemannian metric on M is used to state condition (b) in the def
inition of hyperbolic set. The truth of this condition does not depend upon
which metric is used although the constants c and λ do. A metric is adapted
(to an Axiom A f) if Ω(f) is hyperbolic with respect to it with c = 1.
3.1. Lemma. Every Axiom A diﬀeomorphism has an adapted metric.
Proof. This lemma is due to Mather. See [8] for a proof. ¯.
We will always use an adapted metric. This will keep various estimates a
bit simpler. For x ∈ M deﬁne
W
s
(x) = ¦y ∈ M : d(f
n
x, f
n
y) → 0 as n → ∞¦
W
s
ε
(x) = ¦y ∈ M : d(f
n
x, f
n
y) ≤ ε for all n ≥ 0¦
W
u
(x) = ¦y ∈ M : d(f
−n
x, f
−n
y) → 0 as n → ∞¦
W
u
ε
(x) = ¦y ∈ M : d(f
−n
x, f
−n
y) ≤ ε for all n ≥ 0¦ .
The following stable manifold theorem is the main analytic fact behind the
behavior of Axiom A diﬀeomorphisms.
3.2. Theorem. Let Λ be a hyperbolic set for a C
r
diﬀeomorphism f. For
small ε > 0
(a) W
s
ε
(x), W
u
ε
(x) are C
r
disks for x ∈ Λ with T
x
W
s
ε
(x) = E
s
x
, T
x
W
u
ε
(x) =
E
u
x
;
(b) d(f
n
x, f
n
y) ≤ λ
n
d(x, y) for y ∈ W
s
ε
(x), n ≥ 0, and
d(f
−n
x, f
−n
y) ≤ λ
n
d(x, y) for y ∈ W
u
ε
(x), n ≥ 0;
(c) W
s
ε
(x), W
u
ε
(x) vary continuously with x (in C
r
topology).
Proof. See Hirsch and Pugh [8]. ¯.
One consequence of 3.2 is that W
s
ε
(x) ⊂ W
s
(x) for x ∈ Λ. One then sees
that
W
s
(x) =
¸
n≥0
f
−n
W
s
ε
(f
n
x)
for x ∈ Λ. Similarly
W
u
(x) =
¸
n≥0
f
n
W
u
ε
(f
−n
x) .
B. Spectral Decomposition 47
3.3. Canonical Coordinates. Suppose f satisﬁes Axiom A. For any small
ε > 0 there is a δ > 0 so that W
s
ε
(x) ∩ W
u
ε
(y) consists of a single point [x, y]
whenever x, y ∈ Ω(f) and d(x, y) ≤ δ. Furthermore [x, y] ∈ Ω(f) and
[, ] : ¦(x, y) ∈ Ω(f) Ω(f) : d(x, y) ≤ δ¦ −→ Ω(f)
is continuous.
Proof. See Smale [14]. The ﬁrst statement follows because the intersection
W
s
ε
(x) ∩ W
u
ε
(x) = ¦x¦ is transversal and such intersections are preserved
under small perturbation. To get [x, y] ∈ Ω(f) uses that the periodic points
are dense in Ω(f). In the Anosov case one of course has [x, y] ∈ M and thus
canonical coordinates on M (instead of Ω(f)) without any assumption on
periodic points. ¯.
3.4. Lemma. Let Λ be a hyperbolic set. Then there is an ε > 0 so that Λ is
expansive in M, i.e., if x ∈ Λ and y ∈ M with y = x, then
d(f
k
x, f
k
y) > ε for some k ∈ Z.
Proof. Otherwise y ∈ W
s
ε
(x) ∩ W
u
ε
(x). x is also in this intersection and so
y = x by 3.3. ¯.
B. Spectral Decomposition
From now on f will always be an Axiom A diﬀeomorphism.
3.5. Spectral Decomposition. One can write Ω(f) = Ω
1
∪ Ω
2
∪ ∪ Ω
s
where the Ω
i
are pairwise disjoint closed sets with
(a) f(Ω
i
) = Ω
i
and f[
Ω
i
is topologically transitive;
(b) Ω
i
= X
1,i
∪ ∪X
n
i
,i
with the X
j,i
’s pairwise disjoint closed sets, f(X
j,i
) =
X
j+1,i
(X
n
j
+1,i
= X
1,i
) and f
n
i
[
X
j,i
topologically mixing.
Proof ([14, 4]). For p ∈ Ω periodic let X
p
= W
u
(p) ∩ Ω. Let δ be as in 3.3.
We claim that
X
p
= B
δ
(X
p
) = ¦y ∈ Ω : d(y, X
p
) < δ¦ .
As periodic points are dense in Ω, it is enough to see that a periodic q ∈
B
δ
(X
p
) is in X
p
. Pick x ∈ W
u
(p) ∩ Ω with d(x, q) < δ and consider x
=
[x, q] ∈ W
u
(p) ∩ W
s
(q) ∩ Ω. Letting f
m
p = p and f
n
q = q, one has
f
kmn
x
∈ f
kmn
W
u
(p) = W
u
(f
kmn
p) = W
u
(p)
and d(f
kmn
x
, q) = d(f
kmn
x
, f
kmn
q) → 0 as k → ∞.
So q ∈ X
p
.
48 3 Axiom a Diﬀeomorphisms
Notice that fX
p
= X
f(p)
since fW
u
(p) = W
u
(f(p)). If q ∈ X
p
as above,
then W
u
δ
(q) ⊂ X
p
and
W
u
(q) =
¸
k≥0
f
kmn
W
u
δ
(q)
⊂
¸
k≥0
f
kmn
X
p
= X
p
.
(Note that y ∈ W
u
(q) iﬀ f
−kmn
y → q as k → ∞.) It follows that X
q
⊂ X
p
.
If x
is as above, then f
kmn
x
∈ X
q
for large k as X
q
= B
δ
(X
q
) is open in Ω.
As f
imn
X
q
= X
f
imn
q
= X
q
, one has f
jmn
x
∈ X
q
for all j and
p = lim
j→−∞
f
jmn
x
∈ X
q
= X
q
.
The above argument with the roles of p and q reversed gives X
p
⊂ X
q
. In
summary, if q ∈ X
p
with p, q periodic, X
p
= X
q
.
Now any two X
p
, X
q
are either disjoint or equal. For if X
q
∩ X
q
= ∅,
then this intersection is open in Ω and hence contains a periodic point r; then
X
p
= X
r
= X
q
. Now
Ω =
¸
p periodic
B
δ
(X
p
) =
¸
p
X
p
,
and so by compactness (the X
p
are open) let
Ω = X
p
1
∪ ∪ X
p
t
with the X
p
j
’s pairwise disjoint. Then f(X
p
j
) = X
fp
j
intersects and hence
equals some X
p
i
. So f permutes the X
p
j
’s and the Ω
i
are just the union of
the X
p
j
’s in the various cycles of the permutation.
The transitivity in (a) is implied by the mixing in (b). We ﬁnish by showing
f
N
: X
r
→ X
r
is mixing whenever r is periodic and N positive with f
N
X
r
=
X
r
. Suppose U, V are nonempty subsets of X
r
open in X
r
(i.e., in Ω). Pick
periodic points p ∈ U and q ∈ V , say f
m
p = p, f
n
q = q. For each 0 ≤ j < mn
with f
j
p ∈ X
r
one can ﬁnd a point x
j
as in the beginning of this proof so
that
x
j
∈ f
j
U and f
kmn
x
j
∈ V for large k .
Writing tN = kmn +j, 0 ≤ j ≤ mn, we have f
j
p = f
tN
p ∈ X
r
and
f
kmn
x
j
= f
tN
(f
−jN
x
j
) ∈ f
tN
U ∩ V
provided k is large. Then f
tN
U ∩V = ∅ for large t and f
N
[
X
r
is topologically
mixing. ¯.
The Ω
i
in the spectral decomposition of Ω(f) are called the basic sets of
f. Notice that if g = f
n
and n is a multiple of every n
i
, then the basic sets
B. Spectral Decomposition 49
of g are the X
j,i
’s and g[
X
j,i
is mixing. We will at times restrict our attention
to mixing basic sets and recover the general case by considering f
n
.
A sequence x = ¦x
i
¦
b
i=a
(a = −∞ or b = +∞ is permitted) of points in
M is an αpseudoorbit if
d(fx
i
, x
i+1
) < α for all i ∈ [a, b −1) .
A point x ∈ M βshadows x if
d(f
i
x, x
i
) ≤ β for all i ∈ [a, b] .
3.6. Proposition. For every β > 0 there is an α > 0 so that every αpseudo
orbit ¦x
i
¦
b
i=a
in Ω (i.e., every x
i
∈ Ω) is βshadowed by a point x ∈ Ω.
Proof. Let ε > 0 be a small number to be determined later and choose δ ∈
(0, ε) as in 3.3, i.e., W
s
ε
(x)∩W
u
ε
(y)∩Ω = ∅ whenever x, y ∈ Ω with d(x, y) ≤ δ.
Pick M so large that λ
M
ε < δ/2 and then α > 0 so that:
if ¦y
i
¦
M
i=0
is an α−pseudoorbit in Ω, then
d(f
j
y
0
, y
j
) < δ/2 for all j ∈ [0, M] .
Consider ﬁrst an αpseudoorbit ¦x
i
¦
rM
i=0
with r > 0. Deﬁne x
kM
recursively
for k ∈ [0, r] by x
0
= x
0
and
x
(k+1)M
= W
u
ε
(f
M
x
kM
) ∩ W
s
ε
(x
(k+1)M
) ∈ Ω.
This makes sense: d(f
M
x
kM
, f
M
x
kM
) ≤ λ
M
ε < δ/2 and d(f
M
x
kM
, x
(k+1)M
) <
δ/2 by the choice of α; so d(f
M
x
kM
, x
(k+1)M
) < δ and x
(k+1)M
exists. Now
set x = f
−rM
x
rM
. For i ∈ [0, rM] pick s with i ∈ [sM, (s + 1)M), then
d(f
i
x, f
i−sM
x
sM
) ≤
r
¸
t=s+1
d(f
i−tM
x
tM
, f
i−tM+M
x
(t−1)M
)
≤
r
¸
t=s+1
ε λ
tM−i
≤
ελ
1 −λ
where we used x
tM
∈ W
u
ε
(f
M
x
(t−1)M
). Since x
sM
∈ W
s
ε
(x
sM
)
d(f
i−sM
x
sM
, f
i−sM
x
sM
) ≤ ε .
By the choice of α one has
d(f
i−sM
x
sM
, x
i
) < δ/2 .
By the triangle inequality
d(f
i
x, x
i
) ≤
ελ
1 −λ
+ε +
δ
2
50 3 Axiom a Diﬀeomorphisms
For small ε this is less that the given β.
Now any αpseudoorbit ¦x
i
¦
n
i=0
in Ω extends to ¦x
i
¦
rM
i=0
when rM ≥ n
by setting x
i
= f
i−n
x
n
for i ∈ (n, rM]. An x ∈ Ω shadowing this extended
pseudoorbit will shadow the original one. If ¦x
i
¦
b
i=a
is a ﬁnite αpseudo orbit,
then so is ¦x
j+a
¦
b−a
j=0
and x shadowing this one yields f
−a
x shadowing the
original. Thus the proposition holds for ﬁnite pseudoorbits. If ¦x
i
¦
+∞
i=−∞
is
an αpseudoorbit in Ω, then ﬁnd x
(m)
∈ Ω βshadowing ¦x
i
¦
m
i=−m
and let x
be a limit point of the sequence x
(m)
. Then x ∈ Ω βshadows ¦x
i
¦
+∞
i=−∞
. ¯.
3.7. Corollary. Given any β > 0 there is an α > 0 so that the following
holds: if x ∈ Ω and d(f
n
x, x) < α, then there is an x
∈ Ω with f
n
x
= x
and
d(f
k
x, f
k
x
) ≤ β for all k ∈ [0, n] .
Proof. Let x
i
= f
k
x for i ≡ k (mod n), k ∈ [0, n). Then ¦x
i
¦
∞
i=−∞
is an
αpseudoorbit. Let x
∈ Ω βshadow it. Then d(f
i
x
, f
i
f
n
x
) ≤ d(f
i
x
, x
i
) +
d(x
i
, f
i+n
x
) ≤ 2β and by expansiveness (Lemma 3.4) f
n
x
= x
. ¯.
3.8. Anosov’s Closing Lemma. If f is an Anosov diﬀeomorphism, then f
satisﬁes Axiom A.
Proof. We must show that the periodic points are dense in Ω(f). We have
been assuming f satisﬁes Axiom A; however 3.3 is true also for Anosov dif
feomorphisms and so then is 3.6 and 3.7, using M in place of Ω(f). If y is
a nonwandering point for an Anosov f, then for any γ one can ﬁnd x with
d(x, y) < γ and d(f
n
x, x) < γ for some n. The periodic points x
constructed
in 3.7 for such x converge to y. ¯.
3.9. Fundamental Neighborhood. Let f satisfy Axiom A. There is a
neighborhood U of Ω(f) so that
¸
n∈Z
f
n
U = Ω(f) .
Proof. Let β be small and α as in 3.6. Pick γ < α/2 so that
∀x, y ∈ M , d(x, y) < γ implies d(fx, fy) < α/2 .
Let U = ¦y ∈ M : d(y, Ω) < γ¦. If y ∈
¸
n∈Z
f
n
U, pick x
i
∈ Ω with
d(f
i
y, x
i
) < γ. Then
d(f
i
y, f
i
x) < β +γ for all i .
For small β and γ this implies y = x ∈ Ω by 3.4. ¯.
For Ω
j
a basic set of an Axiom A diﬀeomorphism one let
W
s
(Ω
j
) = ¦x ∈ M : d(f
n
x, Ω
j
) → 0 as n → ∞¦
and W
u
(Ω
j
) = ¦x ∈ M : d(f
−n
x, Ω
j
) → 0 as n → ∞¦ .
C. Markov Partitions 51
Using the deﬁnition of nonwandering sets it is easy to check that f
n
x → Ω
and f
−n
x → Ω as n → ∞. As Ω = Ω
1
∪ ∪ Ω
s
is a disjoint union of closed
invariant sets one then sees that
M =
s
¸
j=1
W
s
(Ω
j
) =
s
¸
j=1
W
u
(Ω
j
)
and that there are disjoint unions.
3.10. Proposition. W
s
(Ω
j
) =
¸
x∈Ω
j
W
s
(x) and W
u
(Ω
j
) =
¸
x∈Ω
j
W
u
(x).
For ε > 0 there is a neighborhood U
j
of Ω
j
so that
¸
k≥0
f
−k
U
j
⊂ W
s
ε
(Ω
j
) =
¸
x∈Ω
j
W
s
ε
(x)
and
¸
k≥0
f
k
U
j
⊂ W
u
ε
(Ω
j
) =
¸
x∈Ω
j
W
u
ε
(x) .
Proof. Suppose f
n
y → Ω
j
as n → ∞; say d(f
n
y, Ω
j
) < γ for all n ≥ N. Pick
x
n
∈ Ω
j
for n ≥ N with d(x
n
, f
n
y) ≤ γ; for n < N let x
n
= f
n−N
x
N
. The
¦x
n
¦
∞
n=−∞
is a pseudoorbit in Ω
j
. Letting x ∈ Ω
j
shadow it, one gets
f
N
y ∈ W
s
ε
(f
N
x) ⊂ W
s
(f
N
x)
(provided γ was small enough). Then y ∈ f
−N
W
s
(f
N
x) = W
s
(x). The reverse
inclusion, W
s
(Ω
j
) ⊃
¸
x∈Ω
j
W
s
(x), is clear.
The proof for W
u
(Ω
j
) is similar and we have proved the second statement
with U
j
= ¦y ∈ M : d(y, Ω
j
) < γ¦. ¯.
C. Markov Partitions
A subset R ⊂ Ω
s
is called a rectangle if it has small diameter and
[x, y] ∈ R whenever x, y ∈ R.
R is called proper if R is closed and R = int(R) (int(R) is the interior of R as
a subset of Ω
s
). For x ∈ R, let
W
s
(x, R) = W
s
ε
(x) ∩ R and W
u
(x, R) = W
u
ε
(x) ∩ R
where ε is small and the diameter of R is small compared to ε.
3.11. Lemma. Let R be a closed rectangle. As a subset of Ω
s
, R has boundary
∂R = ∂
s
R ∪ ∂
u
R
where
52 3 Axiom a Diﬀeomorphisms
∂
s
R = ¦x ∈ R : x / ∈ int(W
u
(x, R))¦
∂
u
R = ¦x ∈ R : x / ∈ int(W
s
(x, R))¦
and the interiors of W
u
(x, R), W
s
(x, R) are as subsets of W
u
ε
(x)∩Ω, W
s
ε
(x)∩
Ω.
Proof. If x ∈ int(R), then W
u
(x, R) = R∩(W
u
ε
(x)∩Ω) is a neighborhood of x
in W
u
ε
(x)∩Ω since R is a neighborhood of x in Ω. Similarly x ∈ int(W
u
(x, R)).
Suppose x ∈ int(W
u
(x, R)) and x ∈ int(W
s
(x, R)). For y ∈ Ω
s
near x the
points
[x, y] ∈ W
s
ε
(x) ∩ Ω and [x, y] ∈ W
u
ε
(x) ∩ Ω
depend continuously on y. Hence for y ∈ Ω
s
close enough to x, [x, y] ∈ R and
[y, x] ∈ R. Then
y
= [[y, x], [x, y]] ∈ R ∩ W
s
ε
(y) ∩ W
u
ε
(y)
and y
= y as W
s
ε
(y) ∩ W
u
ε
(y) = ¦y¦. Thus x ∈ int(R). ¯.
Deﬁnition. A Markov partition of Ω
s
is a ﬁnite covering R = ¦R
1
, . . . , R
m
¦
of Ω
s
by proper rectangles with
(a) int(R
i
) ∩ int(R
j
) = ∅ for i = j,
(b) fW
u
(x, R
i
) ⊃ W
u
(fx, R
j
) and
fW
s
(x, R
i
) ⊂ W
s
(fx, R
j
) when x ∈ int(R
i
), fx ∈ int(R
j
).
3.12. Theorem. Let Ω
s
be a basic set for an Axiom A diﬀeomorphism f.
Then Ω
s
has Markov partitions R of arbitrarily small diameter.
Proof. Let β > 0 be very small and choose α > 0 small as in Proposition 3.6,
i.e., every αpseudoorbit in Ω
s
is βshadowed in Ω
s
. Choose γ < α/2 so that
d(fx, fy) < α/2 when d(x, y) < γ .
Let P = ¦p
1
, . . . , p
r
¦ be a γdense subset of Ω
s
and
Σ(P) =
q ∈
∞
¸
−∞
P : d(fq
j
, q
j+1
) < α for all j
¸
.
For each q ∈ Σ(P) there is a unique θ(q) ∈ Ω
s
which βshadows q; for each
x ∈ Ω
s
there are q with x = θ(q).
For q, q
∈ Σ(P) with q
0
= q
0
we deﬁne q
∗
= [q, q
] ∈ Σ(P) by
q
∗
j
=
q
j
for j ≥ 0
q
j
for j ≤ 0 .
Then d(f
j
θ(q
∗
), f
j
θ(q)) ≤ 2β for j ≥ 0 and d(f
j
θ(q
∗
), f
j
θ(q)) ≤ 2β for j ≤ 0.
So θ(q
∗
) ∈ W
s
2β
(θ(q)) ∩ W
u
2β
(θ(q
)), i.e.,
C. Markov Partitions 53
θ[q, q
] = [θ(q), θ(q
)] .
We now see that T
s
= ¦θ(q) : q ∈ Σ(P), q
0
= p
s
¦ is a rectangle. For
x, y ∈ T
s
we write x = θ(q), y = θ(q
) with q
0
= p
s
= q
0
. Then
[x, y] = θ[q, q
] ∈ T
s
.
Suppose x = θ(q) with q
0
= p
s
and q
1
= p
t
. Consider y ∈ W
s
(x, T
s
), y = θ(q
),
q
0
= p
s
. Then
y = [x, y] = θ[q, q
] and
fy = θ(σ[q, q
]) ∈ T
t
as σ[q, q
] has q
= p
t
in its zeroth position. Since fy ∈ W
s
ε
(fx) (diam(T
s
) ≤
2β is small compared to ε), fy ∈ W
s
ε
(fx, T
t
). We have proved
(i) fW
s
(x, T
s
) ⊂ W
s
(fx, T
t
).
A similar proof shows f
−1
W
u
(fx, T
t
) ⊂ W
u
(x, T
s
), i.e.,
(ii) fW
u
(x, T
s
) ⊃ W
u
(fx, T
t
).
Each T
s
is closed; this follows from the following lemma.
3.13. Lemma. θ : Σ(P) → Ω
s
is continuous.
Proof. Otherwise there is a γ > 0 so that for every N one can ﬁnd q
N
, q
N
∈
Σ(P) with q
j,N
= q
j,N
for all j ∈ [−N, N] but d(θ(q
N
), θ(q
N
)) ≥ γ. If x
N
=
θ(q
N
) , y
N
= θ(q
) one has
d(f
j
x
N
, f
j
y
N
) ≤ 2β ∀j ∈ [−N, N] .
Taking subsequences we may assume x
N
→ x and y
N
→ y as N → ∞. Then
d(f
j
x, f
j
y) ≤ 2β for all j and d(x, y) ≥ γ; this contradicts expansiveness of
f[
Ω
s
. ¯.
Now T = ¦T
1
, . . . , T
r
¦ is a covering by rectangles and (i) and (ii) above
are like the Markov condition (b). However the T
j
’s are likely to overlap and
not be proper. For each x ∈ Ω
s
let
T(x)=¦T
j
∈ T : x ∈ T
j
¦ and T
∗
(x)=¦T
k
∈ T : T
k
∩T
j
=∅for some T
j
∈ T(x)¦.
As T is a closed cover of Ω
s
, Z = Ω
s
`
¸
j
∂T
j
is an open dense subset of Ω
s
.
In fact, using arguments similar to 3.11, one can show that
Z
∗
= ¦x ∈ Ω
s
: W
s
ε
(x) ∩ ∂
s
T
k
= ∅ and W
u
ε
(x) ∩ ∂
u
T
k
= ∅ for all T
k
∈ T
∗
(x)¦
is open and dense in Ω
s
.
For T
j
∩ T
k
= ∅, let
T
1
j,k
= ¦x ∈ T
j
: W
u
(x, T
j
) ∩ T
k
= ∅, W
s
(x, T
j
) ∩ T
k
= ∅¦ = T
j
∩ T
k
T
2
j,k
= ¦x ∈ T
j
: W
u
(x, T
j
) ∩ T
k
= ∅, W
s
(x, T
j
) ∩ T
k
= ∅¦
T
3
j,k
= ¦x ∈ T
j
: W
u
(x, T
j
) ∩ T
k
= ∅, W
s
(x, T
j
) ∩ T
k
= ∅¦
T
4
j,k
= ¦x ∈ T
j
: W
u
(x, T
j
) ∩ T
k
= ∅, W
s
(x, T
j
) ∩ T
k
= ∅¦.
54 3 Axiom a Diﬀeomorphisms
T
4
j,k T
3
j,k
T
2
j,k T
1
j,k
T
k
T
j
W
s
W
u
If x, y ∈ T
j
, then W
s
([x, y], T
j
) = W
s
(x, T
j
) and W
u
([x, y], T
j
) =
W
u
(y, T
j
); this implies T
n
j,k
is a rectangle open in Ω
s
and each x ∈ T
j
∩ Z
∗
lies in int(T
n
j,k
) for some n. For x ∈ Z
∗
deﬁne
R(x) =
¸
¦int(T
n
j,k
) : x ∈ T
j
, T
k
∩ T
j
= ∅ and x ∈ T
n
j,k
¦ .
Now R(x) is an open rectangle (x ∈ Z
∗
). Suppose y ∈ R(x) ∩ Z
∗
. Since
R(x) ⊂ T(x) and R(x) ∩ T
j
= ∅ for T
j
/ ∈ T(x), one gets T(y) = T(x). For
T
j
∈ T(x) = T(y) and T
k
∩ T
j
= ∅, y lies in the same T
n
j,k
as x does since
T
n
j,k
⊃ R(x); hence R(y) = R(x). If R(x) ∩ R(x
) = ∅ (x, x
∈ Z
∗
), there is a
y ∈ R(x) ∩ R(x
) ∩ Z
∗
; then R(x) = R(y) = R(x
). As there are only ﬁnitely
many T
n
j,k
’s there are only ﬁnitely many distinct R(x)’s. Let
R = ¦R(x) : x ∈ Z
∗
¦ = ¦R
1
, . . . , R
m
¦ .
For x
∈ Z
∗
, R(x
) = R(x) or R(x
) ∩R(x) = ∅; hence (R(x) `R(x)) ∩Z
∗
= ∅.
As Z
∗
is dense in Ω
s
, R(x) `R(x) has no interior (in Ω
s
) and R(x) = int(R(x)).
For R(x) = R(x
)
int
R(x)
∩ int
R(x
)
= R(x) ∩ R(x
) = ∅ .
To show that R is Markov we are left to verify condition (b).
Suppose x, y ∈ Z
∗
∩ f
−1
Z
∗
, R(x) = R(y) and y ∈ W
s
ε
(x). We will show
R(fx) = R(fy). First T(fx) = T(fy). Otherwise assume fx ∈ T
j
, fy / ∈ T
j
.
Let fx = θ(σq) with q
1
= p
j
and q
0
= p
s
. Then x = θ(q) ∈ T
s
and by
inclusion (i) above
C. Markov Partitions 55
fy ∈ fW
s
(x, T
s
) ⊂ W
s
(fx, T
j
) ,
contradicting fy / ∈ T
j
. Now let fx, fy ∈ T
j
and T
k
∩T
j
= ∅. We want to show
that fx, fy belong to the same T
n
j,k
. As fy ∈ W
s
ε
(fx) we have W
s
(fy, T
j
) =
W
s
(fx, T
j
). We will derive a contradiction from
W
u
(fy, T
j
) ∩ T
k
= ∅ , fz ∈ W
u
(fx, T
j
) ∩ T
k
.
Recall that fx = θ(σq), q
1
= p
j
, q
0
= p
s
. Then by inclusion (ii)
fz ∈ W
u
(fx, T
j
) ⊂ fW
u
(x, T
s
) or z ∈ W
u
(x, T
s
) .
Let fz = θ(σq
); q
1
= p
k
and q
0
= p
t
. Then z ∈ T
t
and fW
s
(z, T
t
) ⊂
W
s
(fz, T
k
). Now T
s
∈ T(x) = T(y) and z ∈ T
t
∩ T
s
= ∅.
Now z ∈ W
u
(x, T
s
) ∩ T
t
and so there is some z
∈ W
u
(y, T
s
) ∩ T
t
as x, y are
in the same T
n
s,t
. Then
z
= [z, y] = [z, z
] ∈ W
s
(z, T
t
) ∩ W
u
(y, T
s
),
and fz
= [fz, fy] ∈ W
s
(fz, T
k
)∩W
u
(fy, T
j
) (using fz, fy ∈ T
j
a rectangle),
a contradiction. So R(fx) = R(fy).
For small δ > 0 the sets
Y
1
=
¸
⎧
⎨
⎩
W
s
δ
(z) : z ∈
¸
j
∂
s
T
j
⎫
⎬
⎭
and Y
2
=
¸
⎧
⎨
⎩
W
u
δ
(z) : z ∈
¸
j
∂
u
T
j
⎫
⎬
⎭
are closed and nowhere dense (like in the proof of 3.11). Now Z
∗
⊃ Ω
s
`(Y
1
∪Y
2
)
is open and dense. Furthermore if x / ∈ (Y
1
∪ Y
2
) ∩ f
−1
(Y
1
∪ Y
2
) then x ∈
Z
∗
∩f
−1
Z
∗
and the set of y ∈ W
s
(x, R(x)) with y ∈ Z
∗
∩f
−1
Z
∗
is open and
dense in W
s
(x, R(x) ) (as a subset of W
s
ε
(x) ∩Ω). By the previous paragraph
R(fy) = R(fx) for such y; by continuity
fW
s
(x, R(x) ) ⊂ R(fx) .
As fW
s
(x, R(x) ) ⊂ W
s
ε
(fx), fW
s
(x, R(x) ) ⊂ W
s
(fx, R(fx) ).
If int(R
i
) ∩ f
−1
int(R
j
) = ∅, then this open subset of Ω
s
contains some x
satisfying the above conditions, R
i
= R(x) and R
j
= R(fx). For any x
∈
R
i
∩ f
−1
R
j
one has W
s
(x
, R
i
) = ¦[x
, y] : y ∈ W
s
(x, R
i
)¦ and
fW
s
(x
, R
i
) = ¦[fx
, fy] : y ∈ W
s
(x, R
i
)¦
⊂ ¦[fx
, z] : z ∈ W
s
(fx, R
j
)¦
⊂ W
s
(fx
, R
j
) .
This completes the proof of half of the Markov conditions (b). The other
half is proved similarly and the proof is omitted. Alternatively one could apply
the above to f
−1
, noting that W
u
f
= W
s
f
−1
. ¯.
56 3 Axiom a Diﬀeomorphisms
D. Symbolic Dynamics
Throughout this section R = ¦R
1
, . . . , R
m
¦ will denote a Markov partition of
a basic set Ω
s
. One deﬁnes the transition matrix A = A(R) by
A
ij
=
1 if int(R
i
) ∩ f
−1
int(R
j
) = ∅
0 otherwise .
3.14. Lemma. Suppose x ∈ R
i
, fx ∈ R
j
, A
ij
= 1. Then fW
s
(x, R
i
) ⊂
W
s
(fx, R
j
) and fW
u
(x, R
i
) ⊃ W
u
(fx, R
j
).
Proof. This is just the same as the last part of the proof of 3.12. ¯.
Deﬁnition. ∂
s
R =
¸
j
∂
s
R
j
and ∂
u
R =
¸
j
∂
u
R
j
.
3.15. Proposition. f(∂
s
R) ⊂ ∂
s
R and f
−1
(∂
u
R) ⊂ ∂
u
R.
Proof. The set
¸
j
(int(R
i
) ∩ f
−1
int(R
j
)) is dense in R
i
. For any x ∈ R
i
one
can therefore ﬁnd some j and x
n
∈ int(R
i
) ∩f
−1
int(R
j
) with lim
n→∞
x
n
= x.
Then A
ij
= 1, x ∈ R
ij
and fx ∈ R
j
. Hence fW
u
(x, R
i
) ⊃ W
u
(fx, R
j
). If
fx / ∈ ∂
s
R, then W
u
(fx, R
j
) is a neighborhood of fx in W
s
ε
(fx) ∩ Ω and so
W
u
(x, R
i
) is a neighborhood of x in W
s
ε
(x) ∩Ω
s
– that is x / ∈ ∂
s
R
i
. We have
shown f(∂
s
R) ⊂ ∂
s
R. One gets f
−1
(∂
u
R) ⊂ ∂
u
R by a similar argument or
by applying the ﬁrst argument to f
−1
in place of f. ¯.
3.16. Lemma. Let D ⊂ W
s
δ
(x) ∩Ω and C ⊂ W
u
δ
(x) ∩Ω. Then the rectangle
[C, D] is proper iﬀ D = int(D) and C = int(C) as subsets of W
s
δ
(x) ∩ Ω and
W
u
δ
(x) ∩ Ω respectively.
Proof. This is like 3.11. ¯.
Deﬁnition. Let R, S be two rectangles. S will be called a usubrectangle of
R if
(a) S = ∅, S ⊂ R, S is proper, and
(b) W
u
(y, S) = W
u
(y, R) for y ∈ S.
3.17. Lemma. Suppose S is a usubrectangle of R
i
and A
ij
= 1. Then f(S)∩
R
j
is a usubrectangle of R
j
.
Proof. Pick x ∈ R
i
∩ f
−1
R
j
and set D = W
s
(x, R
i
) ∩ S. Because S is a
usubrectangle (condition (b)) one has
S =
¸
y∈D
W
u
(y, R
i
) = [W
u
(x, R
i
), D] .
As S is proper and nonempty, by 3.16 D = ∅ and D = int(D). Now
D. Symbolic Dynamics 57
f(S) ∩ R
j
=
¸
y∈D
(fW
u
(y, R
i
) ∩ R
j
) .
By 3.14, f(y) ∈ R
j
and fW
u
(y, R
i
) ∩ R
j
= W
u
(fy, R
j
).
So f(S) ∩ R
j
=
¸
y
∈f(D)
W
u
(y
, R
j
) = [W
u
(fx, R
j
), f(D)].
Since R
j
= [W
u
(fx, R
j
), W
s
(fx, R
j
)] is proper, one has W
u
(fx, R
j
) proper.
As f maps W
s
ε
(x)∩Ω homeomorphically onto a neighborhood in W
s
ε
(fx)∩Ω,
f(D) = int(f(D)) and so f(S) ∩ R
j
is proper by 3.16. f(S) ∩ R
j
= ∅ as
f(D) = ∅; if y
∈ f(S) ∩ R
j
, then y
∈ W
u
(y
, R
j
) for some y
∈ f(D) and
W
u
(y
, R
j
) = W
u
(y
, R
j
) ⊂ f(S) ∩ R
j
. So f(S) ∩ R
j
is a usubrectangle of
R
j
. ¯.
3.18. Theorem. For each a ∈ Σ
A
the set
¸
j∈Z
f
−j
R
a
j
consists of a single
point, denoted π(a). The map π : Σ
A
→ Ω
s
is a continuous surjection, π◦σ =
f ◦ π, and π is onetoone over the residual set Y = Ω
s
`
¸
j∈Z
f
j
(∂
s
R∪∂
u
R).
Proof. If a
1
a
2
a
n
is a word with A
a
j
a
j+1
= 1, then inductively using 3.17
one sees that
n
¸
j=1
f
n−j
R
a
j
= R
a
n
∩ f
⎛
⎝
n−1
¸
j=1
f
n−1−j
R
a
j
⎞
⎠
is a usubrectangle of R
a
n
. From this one gets that
K
n
(a) =
n
¸
j=−n
f
−j
R
a
j
is nonempty and the closure of its interior. As K
n
(a) ⊃ K
n+1
(a) ⊃ we
have
K(a) =
∞
¸
j=−∞
f
−j
R
a
j
=
∞
¸
n=1
K
n
(a) = ∅ .
If x, y ∈ K(a), then f
j
x, f
j
y ∈ R
a
j
are close for all j ∈ Z and so x = y by
expansiveness. As
K(σa) =
¸
j
f
−j
R
a
j+1
= f
⎛
⎝
¸
j
f
−j
R
a
j
⎞
⎠
= fK(a),
one has π ◦ σ = f ◦ π. That π is continuous is proved like 3.13. As ∂
s
R∪∂
u
R
is nowhere dense, Y is residual. For x ∈ Y pick a
j
with f
j
x ∈ R
a
j
. As x ∈ Y ,
f
j
x ∈ int(R
a
j
) and so A
a
j
a
j+1
= 1. Thus a = ¦a
j
¦ ∈ Σ
A
and x = π(a). If
x = π(b), then f
j
x ∈ R
b
j
and b
j
= a
j
because f
j
x / ∈ ∂
s
R ∪ ∂
u
R; so π is
injective over Y . As π(Σ
A
) is a compact subset of Ω
s
containing a dense set
Y , π(Σ
A
) = Ω
s
. ¯.
58 3 Axiom a Diﬀeomorphisms
3.19. Proposition. σ : Σ
A
→ Σ
A
is topologically transitive. If f[
Ω
s
is topo
logically mixing so is σ : Σ
A
→ Σ
A
.
Proof. Let U, V be nonempty open in Σ
A
. For some a, b ∈ Σ
A
and N one has
U ⊃ U
1
= ¦x ∈ Σ
A
: x
i
= a
i
∀ i ∈ [−N, N]¦
V ⊃ V
1
= ¦x ∈ Σ
A
: x
i
= b
i
∀ i ∈ [−N, N]¦ .
Now
∅ = int(K
N
(a)) =
N
¸
j=−N
f
−j
int(R
a
j
) = U
2
and ∅ = int(K
N
(b)) =
N
¸
j=−N
f
−j
int(R
b
j
) = V
2
.
Also, if x = π(x) ∈ U
2
, then f
j
x ∈ R
x
j
and f
j
x ∈ int(R
a
j
) imply x
j
= a
j
; so
π
−1
(U
2
) ⊂ U
1
. Similarly π
−1
(V
2
) ⊂ V
1
. Since f[
Ω
s
is transitive, f
n
U
2
∩V
2
= ∅
for various large n. Then
∅ = π
−1
(f
n
U
2
∩ V
2
) = π
−1
(f
n
U
2
) ∩ π
−1
(V
2
)
⊂ f
n
U ∩ V .
This same argument shows that σ[
Σ
A
is mixing if f[
Ω
s
is. ¯.
References
The basic references are Anosov [2] for Anosov diﬀeomorphisms and Smale [14]
for Axiom A diﬀeomorphisms. The stable manifold theorem for hyperbolic sets
is due to Hirsch and Pugh [8]. Canonical coordinates and spectral decomposition
are from [14], with the mixing part of 3.5 from [4].
The idea of pseudoorbit has probably occurred to many people. Proposition
3.6 is explicitly proved in [6], though earlier similar statements are in [4] and
for Anosov diﬀeomorphisms in [2]. Sinai [12] stated 3.6 explicitly for Anosov
diﬀeomorphisms. Corollary 3.7 is in [2] for Anosov diﬀeomorphisms and [3] for
Axiom A diﬀeomorphisms. As the name implies, 3.8 is due to Anosov. Results
3.9 and 3.10 are from [7], [15] and for their proofs we have followed [6].
Symbolic dynamics for certain geodesic ﬂows goes back to Hadamard and was
developed by Morse [9]. It was carried out for the horseshoe by Smale [13]
and for automorphisms of the torus by Adler and Weiss [1]. Sinai [10], [11] did
Sections 3 and 3 for Anosov diﬀeomorphisms and this was carried over to Axiom
A diﬀeomorphisms in [5].
1. R. Adler, B. Weiss: Similarity of automorphisms of the torus, Memoirs of the
American Mathematical Society 98, American Mathematical Society, Provi
dence, R.I. 1970.
References 59
2. D. Anosov: Geodesic ﬂows on closed Riemann manifolds with negative curva
ture, Proc. Steklov Inst. Math. 90 (1967).
3. R. Bowen: Topological entropy and Axiom A, Proc. Sympos. Pure Math., Vol.
XIV, Berkeley, Calif., 1968, pp. 23–41, Amer. Math. Soc., Providence, R.I.,
1970.
4. R. Bowen: Periodic points and measures for Axiom A diﬀeomorphisms, Trans.
Amer. Math. Soc. 154 (1971), 377397.
5. R. Bowen: Markov partitions for Axiom A diﬀeomorphisms, Amer. J. Math. 92
(1970), 725747.
6. R. Bowen: ωlimit sets for Axiom A diﬀeomorphisms, J. Diﬀerential Equa
tions 18 (1975), no. 2, 333–339.
7. M. Hirsch, J. Palis, C. Pugh and M. Shub: Neighborhoods of hyperbolic sets,
Invent. Math. 9 1969/1970 121–134.
8. M. Hirsch and C. Pugh: Stable manifolds and hyperbolic sets, Proc. Sympos.
Pure Math., Vol. XIV, Berkeley, Calif., 1968, pp. 133–163, Amer. Math. Soc.,
Providence, R.I., 1970.
9. M. Morse: A OnetoOne Representation of Geodesics on a Surface of Negative
Curvature, Amer. J. Math. 43 (1921), no. 1, 33–51.
10. Ya. Sinai: Markov partitions and Cdiﬀeomorphisms, Func. Anal. and its Appl.
2 (1968), no. 1, 6489.
11. Ya. Sinai: Construction of Markov partitions, Func. Anal. and its Appl. 2 (1968),
no. 2, 7080.
12. Ya. Sinai: Gibbs measures in ergodic theory, Uspehi Mat. Nauk 27 (1972), no.
4, 21–64. English translation: Russian Math. Surveys 27 (1972), no. 4, 21–69.
13. S. Smale: Diﬀeomorphisms with many periodic points, in “Diﬀerential and Com
binatorial Topology”. A Symp. in Honor of Marston Morse, pp. 6380, Princeton
Univ. Press, Princeton, N.J., 1965.
14. S. Smale: Diﬀerentiable dynamical systems, Bull. Amer. Math. Soc. 73 (1967),
747817.
15. S. Smale: The Ωstability theorem, Proc. Sympos. Pure Math., Vol. XIV, Berke
ley, Calif., 1968, pp. 289–298, Amer. Math. Soc., Providence, R.I.
4
Ergodic Theory of Axiom a Diﬀeomorphisms
A. Equilibrium States for Basic Sets
Recall that a function φ is H¨older continuous if there are constants a, θ > 0
so that
[φ(x) −φ(y)[ ≤ a d(x, y)
θ
.
4.1. Theorem. Let Ω
s
be a basic set for an Axiom A diﬀeomorphism f and
φ : Ω
s
→ R H¨older continuous. Then φ has a unique equilibrium state µ
φ
(w.r.t. f[
Ω
s
). Furthermore µ
φ
is ergodic; µ
φ
is Bernoulli if f[
Ω
s
is topologically
mixing.
4.2. Lemma. There are ε > 0 and α ∈ (0, 1) for which the following are true:
if x ∈ Ω
s
, y ∈ M, and d(f
k
x, f
k
y) ≤ ε for all k ∈ [−N, N], then d(x, y) < α
N
.
Proof. See p. 140 of [12]. ¯.
Proof of 4.1. Let R be a Markov partition for Ω
s
of diameter at most ε, A
the transition matrix for R and π : Σ
A
→ Ω
s
as in 3.D. Let φ
∗
= φ ◦ π. If
x, y ∈ Σ
A
have x
k
= y
k
for k ∈ [−N, N], then
f
k
π(x) , f
k
π(y) ∈ R
x
k
= R
y
k
for k ∈ [−N, N] .
This gives d(π(x), π(y)) < α
N
, [φ
∗
(x) −φ
∗
(y)[ ≤ a (α
θ
)
N
and φ
∗
∈ F
A
.
First we assume f[
Ω
s
is mixing. Then σ[
Σ
A
is mixing by 3.19 and we have a
Gibbs measure µ
φ
∗ as in Chapter 1. Let D
s
= π
−1
(∂
s
R) and D
u
= π
−1
(∂
u
R).
Then D
s
and D
u
are closed subsets of Σ
A
, each smaller than Σ
A
, and
σD
s
⊂ D
s
, σ
−1
D
u
⊂ D
u
. As µ
φ
∗ is σinvariant, µ
φ
∗(σ
n
D
s
) = µ
φ
∗(D
s
);
using σ
n+1
D
s
⊂ σ
n
D
s
one has
µ
φ
∗
⎛
⎝
¸
n≥0
σ
n
D
s
⎞
⎠
= µ
φ
∗(D
s
) .
62 4 Ergodic Theory of Axiom a Diﬀeomorphisms
Now
¸
n≥0
σ
n
D
s
has measure 0 or 1 as it is σinvariant and µ
φ
∗ is ergodic (see
1.14); since its complement (a nonempty open set) has positive measure by 1.4,
one gets µ
φ
∗(D
s
) = 0. Similarly one sees µ
φ
∗(D
u
) = 0. Now let µ
φ
= π
∗
µ
φ
∗,
i.e., µ
φ
(E) = µ
φ
∗(π
−1
E). Then µ
φ
is finvariant and the automorphisms of
the measure spaces (σ, µ
φ
∗), (f, µ
φ
) are conjugate since π is onetoone except
on the null set
¸
n∈Z
σ
n
(D
s
∪ D
u
). In particular h
µ
φ
(f) = h
µ
φ
∗
(σ) and so (
1
)
h
µ
φ
(f) +
φ dµ
φ
= h
µ
φ
∗
(σ) +
φ
∗
dµ
φ
∗
= P
σ
(φ
∗
) ≥ P
f
(φ) .
Hence (
2
) P
σ
(φ
∗
) = P
f
(φ) and µ
φ
is an equilibrium state for φ by Chapter 2.
Now µ
φ
is Bernoulli because of 1.25.
4.3. Lemma. For any µ ∈ M
f
(Ω
s
) there is a ν ∈ M
σ
(Σ
A
) with π
∗
ν = µ.
Proof. This wellknown fact is proved as follows. F(g ◦ π) = µ(g) =
gdµ
deﬁnes a positive linear functional on a subspace of C(Σ
A
). By a modiﬁcation
of the HahnBanach Theorem F extends to C(Σ
A
), still positive. As F(1) =
F(1 ◦ π) = 1, F is identiﬁed with some β ∈ M(Σ
A
). By compactness let
ν = lim
k→∞
1
n
k
(β + σ
∗
β + + (σ
n
k
−1
)
∗
β). Then σ
∗
ν = ν and π
∗
ν = µ
(using π
∗
(σ
k
)
∗
β = (f
k
)
∗
π
∗
β = (f
k
)
∗
µ = µ). ¯.
Proof of 4.1 (continued). Suppose µ is any equilibrium state of φ and pick
ν ∈ M
σ
(Σ
A
) with π
∗
ν = µ. Then h
ν
(σ) ≥ h
µ
(f) and so
h
ν
(σ) +
φ
∗
dν ≥ h
µ
(f) +
φ dµ = P(φ) = P(φ
∗
) .
Thus ν is an equilibrium state for φ
∗
and ν = µ
φ
∗ by 1.22. Then µ = π
∗
µ
φ
∗ =
µ
φ
.
We have left the case Ω
s
= X
1
∪ ∪X
m
with fX
k
= X
k+1
and f
m
[
X
1
mix
ing. For µ ∈ M
f
(Ω
s
), one has µ(X
1
) =
1
m
and so µ
= m µ[
X
1
∈ M
f
m(X
1
).
Conversely, if µ
∈ M
f
m(X
1
), then µ ∈ M
f
(Ω
s
) where
µ(E) =
1
m
m−1
¸
k=0
µ
(X
1
∩ f
k
E) .
One checks that µ ↔ µ
deﬁnes a bijection M
f
(Ω
s
) ↔M
f
m(X
1
), h
µ
(f
m
) =
mh
µ
(f), and
S
m
φ dµ
= m
φ dµ. Finding µ maximizing h
µ
(f) +
φ dµ
is equivalent therefore to ﬁnding µ
maximizing h
µ
(f
m
) +
S
m
φ dµ
. For φ
H¨older on Ω
s
, S
m
φ will be H¨ older on X
1
and therefore one is done since X
1
is a mixing basic set of f
m
. ¯.
1
The second equality follows from Theorem 1.22; The inequality comes from
Proposition 2.13 (note of the editor).
2
Using the variational principle 2.17 (note of the editor).
A. Equilibrium States for Basic Sets 63
4.4. Proposition. Let φ : Ω
s
→ R be H¨ older continuous and P = P
f
Ω
s
(φ).
For small ε > 0 there is a b
ε
> 0 so that, for any x ∈ Ω
s
and for all n,
µ
φ
¸
y ∈ Ω
s
: d(f
k
y, f
k
x) < ε ∀ k ∈ [0, n]
¸
≥ b
ε
exp(−Pn +S
n
φ(x)) .
Proof. Choose the Markov partition R above to have diam(R) < ε. Assume
ﬁrst f[
Ω
s
is mixing. Pick x ∈ Σ
A
with π(x) = x. Then
B =
¸
y ∈ Ω
s
: d(f
k
y, f
k
x) < ε ∀ k ∈ [0, n]
¸
⊃ π
¸
y ∈ Σ
A
: y
k
= x
k
∀ k ∈ [0, n]
¸
.
Applying 1.4 and P(φ
∗
) = P(φ) = P one gets
µ
φ
(B) ≥ c
1
exp(−Pn +S
n
φ(x)) .
We leave it to the reader to reduce the general case to the mixing one as in
the proof of 4.1. ¯.
4.5. Proposition. Let φ, ψ : Ω
s
→ R be two H¨older continuous functions.
Then the following are equivalent:
(i) µ
φ
= µ
ψ
.
(ii) There are constants K and L so that [S
m
φ(x) − S
m
ψ(x) − Km[ ≤ L for
all x ∈ Ω
s
and all m ≥ 0.
(iii) There is a constant K so that S
m
φ(x) − S
m
ψ(x) = Km when x ∈ Ω
s
with f
m
x = x.
(iv) There is a H¨ older function u : Ω
s
→ R and a constant K so that φ(x) −
ψ(x) = K +u(fx) −u(x).
If these conditions hold, K = P(φ) −P(ψ).
Proof. Let φ
∗
= φ ◦ π and ψ
∗
= ψ ◦ π. We assume f[
Ω
s
is mixing and leave
the reduction to this case to the reader. If µ
φ
= µ
ψ
, then µ
φ
∗ = µ
ψ
∗ and by
Theorem 1.28 there are K and L so that
[S
m
φ
∗
(x) −S
m
ψ
∗
(x) −Km[ ≤ L
for x ∈ Σ
A
. For x ∈ Ω
s
, picking x ∈ π
−1
(x), this gives us (ii).
Assume (ii) and f
m
x = x. Then
L ≥ [S
mj
φ(x) −S
mj
ψ(x) −mjK[ = j[S
m
φ(x) −S
m
ψ(x) −mK[ .
Letting j → ∞ we get (iii). If (iv) is true, then
φ
∗
(x) −ψ
∗
(x) = K +u(π(σx)) −u(π(x))
and µ
φ
∗ = µ
ψ
∗ by Theorem 1.28. One then has µ
φ
= π
∗
µ
φ
∗ = π
∗
µ
ψ
∗ = µ
ψ
.
Now we assume (iii) and prove (iv). Let η(x) = φ(x) −ψ(x) −K and pick
x ∈ Ω
s
with dense forward orbit (Lemma 1.29). Let A = ¦f
k
x : k ≥ 0¦ and
deﬁne u : A → R by
64 4 Ergodic Theory of Axiom a Diﬀeomorphisms
u(f
k
x) =
k−1
¸
j=0
η(f
j
x) .
For z ∈ A one has u(fz) − u(z) = η(z). Pick ε and α as in 4.2. By 3.7 there
is δ > 0 so that if y ∈ Ω
s
and d(f
n
y, y) < δ, then there is a y
∈ Ω
s
with
f
n
y
= y
and d(f
k
y, f
k
y
) <
ε
2
for all k ∈ [0, n].
Let R be the maximum ratio that f expands any distance. Suppose y =
f
k
x, z = f
m
x with k < m and d(y, z) < /2R
N
. Providing N is large one has
z = f
n
y, n = m−k, and d(f
n
y, y) < δ. Then ﬁnd y
as above with f
n
y
= y
.
Then, as S
n
η(y
) = 0,
[u(z) −u(y)[ =
S
n
η(y) −S
n
η(y
)
≤
n−1
¸
j=0
[η(f
j
y) −η(f
j
y
)[ .
By the choice of R and y
one sees that
d(f
j
y, f
j
y
) < ε for all j ∈ [−N, n +N] .
For j ∈ [0, n) Lemma 4.2 gives
d(f
j
y, f
j
y
) < α
min{j+N,N+n−j}
.
Because η is H¨older,
[η(f
j
y) −η(f
j
y
)[ ≤ a α
θmin{j+N,N+n−j}
[u(y) −u(z)[ ≤ 2a
∞
¸
r=N
α
θr
≤ a
α
θN
.
Pick N so that d(y, z) ∈ [ε/2R
N+1
, ε/2R
N
]. Taking γ > 0 so that (1/R)
γ
≥ α
θ
one has
[u(y) −u(z)[ ≤ a
d(y, z)
γ
.
Thus u is H¨older on A and extends uniquely to a H¨ older function on
¯
A = Ω
s
.
The formula η(z) = u(fz) −u(z) extends to Ω
s
by continuity. ¯.
B. The Case φ = φ
(u)
Recall that M has a Riemannian structure and this induces a volume measure
m on M. We will assume for the remainder of this chapter that f : M → M
is a C
2
Axiom A diﬀeomorphism and Ω
s
is a basic set for f. For x ∈ Ω
s
let
φ
(u)
(x) = −log λ(x) where λ(x) is the Jacobian of the linear map
Df : E
u
x
→ E
u
fx
using inner products derived from the Riemannian metric.
B. The Case φ = φ
(u)
65
4.6. Lemma. If Ω
s
is a C
2
basic set, then φ
(u)
: Ω
s
→ R is H¨ older continu
ous.
Proof. The map x → E
u
x
is H¨ older (see 6.4 of [12]) and E
u
x
→ φ
(u)
(x) is
diﬀerentiable, so the composition x → φ
(u)
(x) is H¨ older. ¯.
By Theorem 4.1 the function φ
(u)
has a unique equilibrium state which we
denote µ
+
= µ
φ
(u) . While φ
(u)
depends on the metric used, when f
m
x = x
S
m
φ
(u)
(x) = −log Jac(Df
m
: E
u
x
→ E
u
x
)
does not depend on the metric (this Jacobian is the absolute value of the
determinant). By 4.5 one sees that the measure µ
+
on Ω
s
and P(φ
(u)
) do not
depend on which metric is used.
4.7. Volume Lemma. Let
B
x
(ε, m) =
¸
y ∈ M : d(f
k
x, f
k
y) ≤ ε for all k ∈ [0, m)
¸
.
If x ∈ Ω
s
is a C
2
basic set and ε > 0 is small, then there is a constant C
ε
so
that
m(B
x
(ε, m)) ∈ [C
−1
ε
, C
ε
] exp(S
m
φ
(u)
(x))
for all x ∈ Ω
s
.
Proof. See 4.2 of [9]. ¯.
4.8. Proposition. Let Ω
s
be a C
2
basic set.
(a) Letting B(ε, n) =
¸
x∈Ω
s
B
x
(ε, n), one has (for small ε > 0)
P
f
Ω
s
(φ
(u)
) = lim
n→∞
1
n
log m(B(ε, n)) ≤ 0 .
(b) Let W
s
ε
(Ω
s
) =
¸
x∈Ω
s
W
s
ε
(x). If m(W
s
ε
(Ω
s
)) > 0, then
P
f
Ω
s
(φ
(u)
) = 0 and h
µ
+(f) = −
φ
(u)
dµ
+
.
Proof. Call E ⊂ M (n, ε)separated if whenever y, z are two distinct points
in E, one can ﬁnd k ∈ [0, n) with d(f
k
y, f
k
z) > δ. Choose E
n
(δ) maximal
among the (n, ε)separated subsets of Ω
s
. For x ∈ Ω
s
one has x ∈ B
y
(ε, n) for
some y ∈ E
n
(δ); otherwise E
n
(δ) ∪ ¦x¦ is (n, ε)separated. Then B
x
(ε, n) ⊂
B
y
(δ +ε, n), B(ε, n) ⊂
¸
y∈E
n
(δ)
B
y
(δ +ε, n) and by 4.7
() m(B(ε, n)) ≤ C
δ+ε
¸
y∈E
n
(δ)
exp(S
n
φ
(u)
(y)).
For δ ≤ ε,
¸
y∈E
n
(δ)
B
y
(δ/2, n) ⊂ B(ε, n) is a disjoint union and so
66 4 Ergodic Theory of Axiom a Diﬀeomorphisms
() m(B(ε, n)) ≥ C
−1
δ/2
¸
y∈E
n
(δ)
exp(S
n
φ
(u)
(y)).
Since φ
(u)
is H¨older, we have
[φ
(u)
(x) −φ
(u)
(y)[ ≤ a d(x, y)
θ
for some a, θ > 0 and all x, y ∈ Ω
s
. Suppose x ∈ B
y
(ε, n) ∩ Ω
s
. Then for
j ∈ [0, n)
d(f
j
x, f
j
y) < α
min{j,n−j−1}
by Lemma 4.2. Hence
[S
n
φ(x) −S
n
φ(y)[ ≤
n−1
¸
j=0
[φ
(u)
(f
j
x) −φ
(u)
(f
j
y)[
≤ 2a
∞
¸
k=0
α
kθ
= γ .
Fix δ ≤ ε and let U be an open cover of Ω
s
with diam(U) < δ. Let
Γ ⊂ U
n
cover Ω
s
. For each y ∈ E
n
(δ) pick U
y
∈ Γ with y ∈ X(U
y
). Then
S
n
φ
(u)
(U
y
) ≥ S
n
φ
(u)
(y). If U
y
= U
y
, then d(f
k
y, f
k
y
) ≤ diam(U) < δ and
y = y
as E
n
(δ) is (n, δ)separated. Thus
¸
U∈Γ
exp(S
n
φ
(u)
(U)) ≥
¸
y∈E
n
(δ)
exp(S
n
φ
(u)
(y)).
Using this together with () above one gets
P(φ
(u)
, U) = lim
n→∞
1
n
log inf
Γ
¸
U∈Γ
exp(S
n
φ
(u)
(U))
≥ limsup
n→∞
1
n
log m(B(ε, n)) .
Letting diam(U) → 0, one replaces P(φ
(u)
, U) with P(φ
(u)
).
Now let U be an open cover and let δ be a Lebesgue number for U. For
each y ∈ E
n
(δ) one can pick U
y
∈ U
n
with B
y
(δ, n) ∩ Ω
s
⊂ X(U
y
).
Let Γ = ¦U
y
: y ∈ E
n
(δ)¦. Then Γ covers Ω
s
since every x ∈ Ω
s
lies in some
B
y
(δ, n) with y ∈ E
n
(δ). Also
S
n
φ
(u)
(U
y
) ≤ S
n
φ
(u)
(y) +γ
and so
Z
n
(φ
(u)
, U) ≤
¸
U
y
∈Γ
n
exp(S
n
φ
(u)
(U
y
))
≤ e
γ
¸
y∈E
n
(δ)
exp(S
n
φ
(u)
(y)) .
B. The Case φ = φ
(u)
67
Using () we get
P(φ
(u)
, U) = lim
n→∞
1
n
log Z
n
(φ
(u)
, U)
≤ liminf
n→∞
1
n
log m(B(ε, n)) .
Noting that m(B(ε, n)) ≤ m(M), the proof of (a) is complete.
Now m(B(ε, n)) ≥ m(W
s
ε
(Ω
s
)) as W
s
ε
(Ω
s
) ⊂ B(ε, n). If m(W
s
ε
(Ω
s
)) > 0,
the formula from (a) yields P(φ
(u)
) = 0. Since µ
+
is an equilibrium state for
φ
(u)
,
h
µ
+(f) +
φ
(u)
dµ
+
= P(φ
(u)
) = 0 . ¯.
A basic set Ω
s
is an attractor if it has small neighborhood U with f(U) ⊂
U. By Proposition 3.10 this is equivalent to W
s
ε
(Ω
s
) being a neighborhood of
Ω
s
.
4.9. Lemma. Let Ω
s
be a C
1
basic set. If W
u
ε
(x) ⊂ Ω
s
for some x ∈ Ω
s
,
then Ω
s
is an attractor. If Ω
s
is not an attractor there exists γ > 0 such that
for every x ∈ Ω
s
, there is y ∈ W
u
ε
(x) with d(y, Ω
s
) > γ.
Proof. If W
u
ε
(Ω
s
) ⊂ Ω
s
, then
U
x
=
¸
¦W
s
ε
(y) : y ∈ W
u
ε
(x)¦
is a neighborhood of x in M (see Lemma 4.1 of [11]). Choose a periodic point
p ∈ U
x
, say f
m
x = x. For some small β one has W
u
β
(p) ⊂ U
x
; if z ∈ W
u
β
(p)
lies in W
s
ε
(y) (y ∈ W
u
ε
(Ω
s
)), then d(f
n
z, f
n
y) < ε and d(f
−n
z, f
−n
p) < β
for n ≥ 0. By Theorem 3.9 one has z ∈ Ω
s
and W
u
β
(p) ⊂ Ω
s
. Then also
W
u
(p) =
¸
k≥0
f
mk
W
u
β
(p) ⊂ Ω
s
.
Now X
p
= W
u
(p) and Ω
s
= X
p
∪fX
p
∪ ∪f
N
X
p
for some N. For each
x ∈
¸
N
k=0
f
k
W
u
(p) = Y one has W
u
ε
(x) ⊂ Ω
s
and so U
x
as deﬁned above
is a neighborhood of x in M. Since W
s
ε
(x), W
u
ε
(x) depend continuously on
x ∈ Ω
s
, one can ﬁnd a δ > 0 independent of x so that U
x
contains the 2δball
B
x
(2δ) about x in M for all x ∈ Y (see [11, Lemma 4.1]). Then
B
Ω
s
(δ) ⊂
¸
¦U
x
: x ∈ Y ¦ ⊂ W
s
ε
(Ω
s
)
and Ω
s
is an attractor.
To prove (b) notice that the set
V
γ
= ¦x ∈ Ω
s
: d(y, Ω
s
) > γ for some y ∈ W
u
ε
(x)¦
is open because W
u
ε
(x) varies continuously with x. Also V
γ
increases when γ
decreases and
¸
γ>0
V
γ
= Ω
s
by statement (a). By compactness V
γ
= Ω
s
for
some γ > 0. ¯.
68 4 Ergodic Theory of Axiom a Diﬀeomorphisms
4.10. Second Volume Lemma. Let Ω
s
be a C
2
basic set. For small ε, δ > 0
there is a d = d(ε, δ) > 0 so that
m(B
y
(δ, n)) ≥ d m(B
x
(ε, n))
whenever x ∈ Ω
s
and y ∈ B
x
(ε, n).
Proof. See 4.3 of [9]. ¯.
4.11. Theorem. Let Ω
s
be a C
2
basic set. The following are equivalent:
(a) Ω
s
is an attractor.
(b) m(W
s
(Ω
s
)) > 0.
(c) P
f
Ω
s
(φ
(u)
) = 0.
Proof. As W
s
(Ω
s
) =
¸
∞
n=0
f
−n
W
s
ε
(Ω
s
), (b) is equivalent to m(W
s
ε
(Ω
s
)) > 0.
If Ω
s
is an attractor, then (b) is true since W
s
ε
(Ω
s
) is a neighborhood of Ω
s
.
(c) follows from (b) by Proposition 4.8 (b). We ﬁnish by assuming Ω
s
is not
an attractor and showing P(φ
(u)
) < 0.
Given a small ε > 0 choose γ as in 4.9. Pick N so that
f
N
W
u
γ/4
(x) ⊃ W
u
ε
(f
N
x)
for all x ∈ Ω
s
. Let E ⊂ Ω
s
be (γ, n)separated. For x ∈ E there is a y(x, n) ∈
B
x
(γ/4, n) with
d(f
n+N
y(x, n), Ω
s
) > γ
(since f
n
B
x
(γ/4, n) ⊃ W
u
γ/4
(f
n
x) and f
N
W
u
γ/4
(f
n
x) ⊃ W
u
ε
(f
N+n
x)). Choose
δ ∈ (0, γ/4) so that d(f
N
z, f
N
y) < γ/2 whenever d(z, y) < δ. Then
B
y(x,n)
(δ, n) ⊂ B
x
(γ/2, n),
f
n+N
B
y(x,n)
(δ, n) ∩ B
Ω
s
(γ/2) = ∅ .
Hence B
y(x,n)
(δ, n) ∩ B(γ/2, n +N) = ∅. Using the Second Volume Lemma
m(B(γ/2, n)) −m(B(γ/2, n +N)) ≥
¸
x∈E
m(B
y(x,n)
(δ, n))
≥ d(3γ/2, δ)
¸
x∈E
m(B
x
(3γ/2, n))
≥ d(3γ/2, δ) m(B(γ/2, n)) .
Therefore, setting d = d(3γ/2, δ)
m(B(γ/2, n +N)) ≤ (1 −d) m(B(γ/2, n))
and by Proposition 4.8 (a)
P
f
Ω
s
(φ
(u)
) ≤
1
N
log(1 −d) < 0 . ¯.
Remark. It is possible to ﬁnd a C
1
basic set (a horseshoe) which is not an
attractor but nevertheless has m(W
s
(Ω
s
)) > 0 [8].
C. Attractors and Anosov Diﬀeomorphisms 69
C. Attractors and Anosov Diﬀeomorphisms
Because M =
¸
r
k=1
W
s
(Ω
k
), Theorem 4.11 implies that malmost all x ∈ M
approach an attractor under the action of a C
2
Axiom A diﬀeomorphism f.
This leads us next to the following result.
4.12. Theorem. Let Ω
s
be a C
2
attractor. For malmost all x ∈ W
s
(Ω
s
) one
has
lim
n→∞
1
n
n−1
¸
k=0
g(f
k
x) =
g dµ
+
for all continuous g : M → R (i.e., x is a generic point for µ
+
).
Proof. Let us write g(n, x) =
1
n
¸
n−1
k=0
g(f
k
x) and g =
g dµ
+
. Fix δ > 0 and
deﬁne the sets
C
n
(g, δ) = ¦x ∈ M : [g(n, x) −g[ > δ¦
E(g, δ) = ¦x ∈ M : [g(n, x) −g[ > δ for inﬁnitely many n¦
=
∞
¸
N=1
∞
¸
n=N
C
n
(g, δ) .
Choose ε > 0 so that [g(x) −g(y)[ < δ when d(x, y) < ε.
Now ﬁx N > 0 and choose sets R
N
, R
N+1
, . . . successively as follows.
Let R
n
(n ≥ N) be a maximal subset of Ω
s
∩ C
n
(g, 2δ) satisfying the condi
tions:
(a) B
x
(ε, n) ∩ B
y
(ε, k) = ∅ for x ∈ R
n
, y ∈ R
k
, N ≤ k < n,
(b) B
x
(ε, n) ∩ B
x
(ε, n) = ∅ for x, x
∈ R
n
, x = x
.
If y ∈ W
s
ε
(Ω
s
) ∩ C
n
(g, 3δ) (n ≥ N) and y ∈ W
s
ε
(z) with z ∈ Ω
s
, then
z ∈ C
n
(g, 2δ) by the choice of ε. By the maximality of R
n
one has
B
z
(ε, n) ∩ B
x
(ε, k) = ∅ for some x ∈ R
k
, N ≤ k ≤ n.
Then y ∈ B
z
(ε, n) ⊂ B
z
(ε, k) ⊂ B
x
(2ε, k) and so
W
s
ε
(Ω
s
) ∩
∞
¸
n=N
C
n
(g, 3δ) ⊂
∞
¸
k=N
¸
x∈R
k
B
x
(2ε, k) .
Using the Volume Lemma 4.7 one gets
m
W
s
ε
(Ω
s
) ∩
∞
¸
n=N
C
n
(g, 3δ)
≤ c
2ε
∞
¸
k=N
¸
x∈R
k
exp(S
k
φ
(u)
(x)).
The deﬁnition of R
n
shows that V
N
=
¸
∞
k=N
¸
x∈R
k
B
x
(ε, k) is a disjoint
union. The choice of ε gives B
x
(ε, k) ⊂ C
k
(g, δ) for x ∈ R
k
⊂ C
k
(g, 2δ) and so
70 4 Ergodic Theory of Axiom a Diﬀeomorphisms
V
N
⊂
¸
∞
k=N
C
k
(g, δ). Since the measure µ
+
is ergodic, the Ergodic Theorem
implies
0 = µ
+
(E(g, δ)) = lim
n→∞
µ
+
∞
¸
n=N
C
n
(g, δ)
and thus lim
N→∞
µ
+
(V
N
) = 0. By 4.8 (b) one has P
f
Ω
s
(φ
(u)
) = 0 and then
by 4.4
µ
+
(V
N
) ≥ b
ε
∞
¸
k=M
¸
x∈R
k
exp(S
k
φ
(u)
(x)).
As µ
+
(V
N
) → 0, the sum on the right approaches 0 as N → ∞. Using the
inequality of the preceding paragraph one sees
lim
N→∞
m
W
s
ε
(Ω
s
) ∩
∞
¸
n=N
C
n
(g, 3δ)
= 0 .
This in turn implies m(W
s
ε
(Ω
s
) ∩ E(g, 3δ)) = 0.
For δ
> 3δ the set E(g, δ
) ∩ f
−n
W
s
ε
(Ω
s
) ⊂ f
−n
(E(g, 3δ) ∩ W
s
ε
(Ω
s
)) has
measure 0 since f preserves measure (w.r.t. m). Thus
m(E(g, δ
) ∩ W
s
(Ω
s
)) ≤
∞
¸
n=0
m
E(g, δ
) ∩ f
−n
W
s
ε
(Ω
s
)
= 0 .
Fixing g still but letting δ
=
1
m
→ 0 one gets lim
n
g(n, x) = g for all x ∈
W
s
(Ω
s
) outside an mnull set A(g). Let ¦g
k
¦
∞
k=1
be a dense subset of C(M);
for x ∈ W
s
(Ω
s
)
¸
∞
k=1
A(g
k
) one gets that lim
n
g(n, x) = g for all g ∈ C(M).
¯.
4.13. Corollary. Suppose f : M → M is a transitive C
2
Anosov diﬀeo
morphism. If f leaves invariant a probability measure µ which is absolutely
continuous with respect to m, then µ = µ
+
.
Proof. In this case M = Ω = Ω
1
is the spectral decomposition. Let g ∈ C(M).
By the Ergodic Theorem there is a function g
∗
: M → R so that
lim
n→∞
1
n
n−1
¸
k=0
g(f
k
x) = g
∗
(x)
for µalmost all x. Let A be the set of x ∈ M with
lim
n→∞
1
n
n−1
¸
k=0
g(f
k
x) =
gdµ
+
.
Because m(A) = 1 and µ < m, µ(A) = 1. It follows that g
∗
(x) =
gdµ
+
for
µalmost all x. Then
C. Attractors and Anosov Diﬀeomorphisms 71
gdµ =
g
∗
dµ =
gdµ
+
.
As this holds for all g ∈ C(M), µ = µ
+
. ¯.
Remark. If M is connected, then M = Ω
1
= X
1
and f is mixing. So µ above
is Bernoulli.
4.14. Theorem. Let f be a transitive C
2
Anosov diﬀeomorphism. The fol
lowing are equivalent:
(a) f admits an invariant measure of the form dµ = hdm with h a positive
H¨older function.
(b) f admits an invariant measure µ absolutely continuous w.r.t. m.
(c) Df
n
: T
x
M → T
x
M has determinant 1 whenever f
n
x = x.
Proof. Clearly (a) implies (b). Assume (b) holds. Let λ
(s)
(x) be the Jacobian
of Df : E
s
f
−1
x
→ E
s
x
and φ
(s)
(x) = log λ
(s)
(x). Now f
−1
is Anosov with
E
u
x,f
−1
= E
s
x,f
an E
s
x,f
−1
= E
u
x,f
. Also
λ
(u)
f
−1
(x) = Jacobian Df
−1
: E
s
x
→ E
s
f
−1
x
= λ
(s)
(x)
−1
and so φ
(u)
f
−1
(x) = −log λ
(u)
f
−1
(x) = φ
(s)
(x). There is an invariant measure µ
−
so that
lim
n→∞
1
n
n−1
¸
k=0
g(f
−k
x) =
g dµ
−
for malmost all x; µ
−
is the unique equilibrium state for φ
(u)
f
−1
w.r.t. f
−1
.
Notice that equilibrium states w.r.t. f
−1
are the same as those w.r.t. f;
for M
f
(M) = M
f
−1(M) and h
ν
(f) = h
ν
(f
−1
). So µ
−
= µ
φ
(s) . Applying 4.13
to both f and f
−1
we see
µ
φ
(u) = µ
+
= µ = µ
−
= µ
φ
(s) .
By 4.8 (b), P(φ
(u)
) = 0 = P(φ
(s)
). By 4.5 one has, for f
n
x = x,
n−1
¸
k=0
φ
(u)
(f
k
x) −
n−1
¸
k=0
φ
(s)
(f
k
x) = 0.
Exponentiating,
1 = (det Df
n
[
E
u
x
) (det Df
n
[
E
s
x
) = (det Df
n
[
T
x
M
) .
Now assume (c) and let φ(x) = log Jac (Df : T
x
M → T
fx
M). Then
72 4 Ergodic Theory of Axiom a Diﬀeomorphisms
S
n
φ(x) = log
n−1
¸
k=0
Jac (Df : T
f
k
x
M → T
f
k+1
x
M)
= log Jac (Df
n
: T
x
M → T
f
k
x
M) = 0
when f
k
x = x. By Proposition 4.5 (with ψ = 0, K = 0) there is a H¨older
u : M → R with φ(x) = u(fx) −u(x). Let h(x) = e
u(x)
.
Thinking of µ = hdm as the absolute value of a form
f
∗
(hdm)(fx) = h(x) e
φ(x)
dm(fx)
= h(fx) dm(fx) = (h dm)(fx) .
So µ is finvariant. ¯.
Remark. Actually h above will be C
1
. See [13, 14].
4.15. Corollary. Among the C
2
Anosov diﬀeomorphisms the ones that admit
no invariant measures µ < m are open and dense.
Proof. [18], page 36. We use condition 4.14 (c). Suppose f
n
x = x and
det(Df
n
[
T
x
M
) = 1. For f
1
near f, f
1
will be Anosov and have a periodic point
x
1
near x with f
n
1
x
1
= x
1
. Then det(Df
n
1
[
T
x
1
M
) will be near det(Df
n
[
T
x
M
)
and not equal to one. We are using stability theory not covered in these notes.
On the other hand, if f
n
x = x with det(Df
n
[
T
x
M
) = 1, this condition is de
stroyed by the right small perturbation of f. ¯.
In the case where f admits no invariant µ < m, the measure m is actually
dissipative [10].
References
This chapter contains the main theorems of these notes. The notes as a whole
constitute a version of Sinai’s program [18] for applying statistical mechanics
to diﬀeomorphisms. It was Ruelle who carried Sinai’s work on Anosov diﬀeo
morphisms over to Axiom A attractors [11] and brought in the formalism of
equilibrium states [15].
Theorem 4.1 is from [6, 7]; see [18] for the Anosov case. Results 4.5, 4.13, 4.14
and 4.15 came from [13, 14, 18]. Section 4 is taken verbatim from [9]. Theorem
4.12 is due to Ruelle [16] (we followed the proof in [9]). Ruelle [16] along these
lines also proved that f
n
µ → µ
+
when µ m has support in a neighborhood
of an attractor; for the Anosov case this result is due to Sinai [17]. That µ
+
is Bernoulli in the transitive Anosov case is due to Azencott [4]. The ergodic
theory of Anosov diﬀeomorphisms with µ
+
m has been studied in many
papers; see for instance [2] or [3].
In case φ
(u)
is a constant function µ
+
is the unique invariant measure which
maximizes entropy (φ = 0 and φ
(u)
have the same equilibrium state). For hy
perbolic automorphisms of the 2torus µ
+
is Haar measure and construction
References 73
in 4.1 is due to Adler and Weiss [1]. This paper is very important in the devel
opment of the subject and is good reading. When φ
(u)
= C the measure µ which
maximizes entropy still has the following geometrical signiﬁcance: the periodic
points of Ω
s
are equidistributed with respect to µ [5]. K. Sigmund [19] studied
the generic properties of measures on Ω
s
.
1. R. Adler, B. Weiss: Similarity of automorphisms of the torus, Memoirs of the
American Mathematical Society 98, American Mathematical Society, Provi
dence, R.I. 1970.
2. D. Anosov: Geodesic ﬂows on closed Riemann manifolds with negative curva
ture, Proc. Steklov Inst. Math. 90 (1967).
3. D. Anosov, Ya. Sinai: Certain smooth ergodic systems, Uspehi Mat. Nauk 22
(1967) no. 5 (137), 107–172. English translation: Russ. Math. Surveys 22 (1967),
no. 5, 103167.
4. R. Azencott: Diﬀ´eomorphismes d’Anosov et sch´emas de Bernoulli, C. R. Acad.
Sci. Paris, S´er. AB 270 (1970), A1105A1107.
5. R. Bowen: Periodic points and measures for Axiom A diﬀeomorphisms, Trans.
Amer. Math. Soc. 154 (1971), 377397.
6. R. Bowen: Some systems with unique equilibrium states, Math. Systems The
ory 8 (1974/75), no.3, 193202.
7. R. Bowen: Bernoulli equilibrium states for Axiom A diﬀeomorphisms, Math.
Systems Theory 8 (1974/75), no. 4, 289–294.
8. R. Bowen: A horseshoe with positive measure, Invent. Math. 29 (1975), no. 3,
203204.
9. R. Bowen, D. Ruelle: The ergodic theory of Axiom A ﬂows, Invent. Math. 29
(1975), no. 3, 181202.
10. B. Gureviˇc, V. Oseledec: Gibbs distributions, and the dissipativity of
Cdiﬀeomorphisms, Soviet Math. Dokl. 14 (1973), 570–573.
11. M. Hirsch, J. Palis, C. Pugh and M. Shub: Neighborhoods of hyperbolic sets,
Invent. Math. 9 (1969/1970), 121–134.
12. M. Hirsch and C. Pugh: Stable manifolds and hyperbolic sets, Proc. Sympos.
Pure Math., Vol. XIV, Berkeley, Calif., 1968, pp. 133–163, Amer. Math. Soc.,
Providence, R.I., 1970.
13. A. N. Livˇsic: Cohomology of dynamical systems, Math. USSRIzv. 6 (1972),
12781301.
14. A. N. Livˇsic, Ya. Sinai: Invariant measures that are compatible with smoothness
for transitive Csystems, Soviet Math. Dokl. 13 (1972), 16561659.
15. D. Ruelle: Statistical mechanics on a compact set with Z
ν
action satisfying
expansiveness and speciﬁcation, Trans. Amer. Math. Soc. 185 (1973), 237251.
16. D. Ruelle: A measure associated with Axiom A attractors, Amer. J. Math. 98
(1976), no. 3, 619654.
17. Ya. Sinai: Markov partitions and Cdiﬀeomorphisms, Func. Anal. and its Appl.
2 (1968), no. 1, 6489.
18. Ya. Sinai: Gibbs measures in ergodic theory, Uspehi Mat. Nauk 27 (1972), no.
4, 21–64. English translation: Russian Math. Surveys 27 (1972), no. 4, 21–69.
19. K. Sigmund: Generic properties of invariant measures for Axiom A diﬀeomor
phisms, Invent. Math. 11 (1970), 99109.
Index
Anosov, 46, 71, 72
Attractor, 67–69
Axiom A, 46
Basic set, 48
Canonical coordinates, 47
Central limit theorem, 24
Entropy, 4, 16, 18, 29, 43
Equilibrium state, 42, 61
Ergodic, 14
Exponential clustering, 23
F
A
, 7
Free energy, 4
Gibbs measure, 6, 7, 15
Homologous, 7
Hyperbolic, 45
Markov partition, 52
Mixing, 6, 14
Nonwandering, 45
Pressure, 18, 33, 34
Pseudoorbit, 49
Rectangle, 51
Σ
A
, 6
Separated, 65
Shift, 5, 9
Transition matrix, 56
Variational principle, 19, 40
Weak Bernoulli, 22
Lecture Notes in Mathematics
For information about earlier volumes
please contact your bookseller or Springer
LNM Online archive: springerlink.com
Vol. 1755: J. Azéma, M. Émery, M. Ledoux, M. Yor
(Eds.), Séminaire de Probabilités XXXV (2001)
Vol. 1756: P. E. Zhidkov, Korteweg de Vries and Nonlin
ear Schrödinger Equations: Qualitative Theory (2001)
Vol. 1757: R. R. Phelps, Lectures on Choquet’s Theorem
(2001)
Vol. 1758: N. Monod, Continuous Bounded Cohomology
of Locally Compact Groups (2001)
Vol. 1759: Y. Abe, K. Kopfermann, Toroidal Groups
(2001)
Vol. 1760: D. Filipovi´ c, Consistency Problems for Heath
JarrowMorton Interest Rate Models (2001)
Vol. 1761: C. Adelmann, The Decomposition of Primes in
Torsion Point Fields (2001)
Vol. 1762: S. Cerrai, Second Order PDE’s in Finite and
Inﬁnite Dimension (2001)
Vol. 1763: J.L. Loday, A. Frabetti, F. Chapoton, F. Goi
chot, Dialgebras and Related Operads (2001)
Vol. 1764: A. Cannas da Silva, Lectures on Symplectic
Geometry (2001)
Vol. 1765: T. Kerler, V. V. Lyubashenko, NonSemisimple
Topological QuantumField Theories for 3Manifolds with
Corners (2001)
Vol. 1766: H. Hennion, L. Hervé, Limit Theorems for
Markov Chains and Stochastic Properties of Dynamical
Systems by QuasiCompactness (2001)
Vol. 1767: J. Xiao, Holomorphic Q Classes (2001)
Vol. 1768: M. J. Pﬂaum, Analytic and Geometric Study of
Stratiﬁed Spaces (2001)
Vol. 1769: M. AlberichCarramiñana, Geometry of the
Plane Cremona Maps (2002)
Vol. 1770: H. GluesingLuerssen, Linear Delay
Differential Systems with Commensurate Delays: An
Algebraic Approach (2002)
Vol. 1771: M. Émery, M. Yor (Eds.), Séminaire de Prob
abilités 19671980. A Selection in Martingale Theory
(2002)
Vol. 1772: F. Burstall, D. Ferus, K. Leschke, F. Pedit,
U. Pinkall, Conformal Geometry of Surfaces in S
4
(2002)
Vol. 1773: Z. Arad, M. Muzychuk, Standard Integral
Table Algebras Generated by a Nonreal Element of Small
Degree (2002)
Vol. 1774: V. Runde, Lectures on Amenability (2002)
Vol. 1775: W. H. Meeks, A. Ros, H. Rosenberg, The
Global Theory of Minimal Surfaces in Flat Spaces.
Martina Franca 1999. Editor: G. P. Pirola (2002)
Vol. 1776: K. Behrend, C. Gomez, V. Tarasov, G. Tian,
Quantum Comohology. Cetraro 1997. Editors: P. de Bar
tolomeis, B. Dubrovin, C. Reina (2002)
Vol. 1777: E. GarcíaRío, D. N. Kupeli, R. Vázquez
Lorenzo, Osserman Manifolds in SemiRiemannian
Geometry (2002)
Vol. 1778: H. Kiechle, Theory of KLoops (2002)
Vol. 1779: I. Chueshov, Monotone Random Systems
(2002)
Vol. 1780: J. H. Bruinier, Borcherds Products on O(2,1)
and Chern Classes of Heegner Divisors (2002)
Vol. 1781: E. Bolthausen, E. Perkins, A. van der Vaart,
Lectures on Probability Theory and Statistics. Ecole d’
Eté de Probabilités de SaintFlour XXIX1999. Editor:
P. Bernard (2002)
Vol. 1782: C.H. Chu, A. T.M. Lau, Harmonic Functions
on Groups and Fourier Algebras (2002)
Vol. 1783: L. Grüne, Asymptotic Behavior of Dynamical
and Control Systems under Perturbation and Discretiza
tion (2002)
Vol. 1784: L. H. Eliasson, S. B. Kuksin, S. Marmi, J.C.
Yoccoz, Dynamical Systems and Small Divisors. Cetraro,
Italy 1998. Editors: S. Marmi, J.C. Yoccoz (2002)
Vol. 1785: J. Arias de Reyna, Pointwise Convergence of
Fourier Series (2002)
Vol. 1786: S. D. Cutkosky, Monomialization of Mor
phisms from 3Folds to Surfaces (2002)
Vol. 1787: S. Caenepeel, G. Militaru, S. Zhu, Frobenius
and Separable Functors for Generalized Module Cate
gories and Nonlinear Equations (2002)
Vol. 1788: A. Vasil’ev, Moduli of Families of Curves for
Conformal and Quasiconformal Mappings (2002)
Vol. 1789: Y. Sommerhäuser, YetterDrinfel’d Hopf alge
bras over groups of prime order (2002)
Vol. 1790: X. Zhan, Matrix Inequalities (2002)
Vol. 1791: M. Knebusch, D. Zhang, Manis Valuations
and Prüfer Extensions I: A new Chapter in Commutative
Algebra (2002)
Vol. 1792: D. D. Ang, R. Gorenﬂo, V. K. Le, D. D. Trong,
Moment Theory and Some Inverse Problems in Potential
Theory and Heat Conduction (2002)
Vol. 1793: J. Cortés Monforte, Geometric, Control and
Numerical Aspects of Nonholonomic Systems (2002)
Vol. 1794: N. Pytheas Fogg, Substitution in Dynamics,
Arithmetics and Combinatorics. Editors: V. Berthé, S. Fer
enczi, C. Mauduit, A. Siegel (2002)
Vol. 1795: H. Li, FilteredGraded Transfer in Using Non
commutative Gröbner Bases (2002)
Vol. 1796: J.M. Melenk, hpFinite Element Methods for
Singular Perturbations (2002)
Vol. 1797: B. Schmidt, Characters and Cyclotomic Fields
in Finite Geometry (2002)
Vol. 1798: W.M. Oliva, Geometric Mechanics (2002)
Vol. 1799: H. Pajot, Analytic Capacity, Rectiﬁability,
Menger Curvature and the Cauchy Integral (2002)
Vol. 1800: O. Gabber, L. Ramero, Almost Ring Theory
(2003)
Vol. 1801: J. Azéma, M. Émery, M. Ledoux, M. Yor
(Eds.), Séminaire de Probabilités XXXVI (2003)
Vol. 1802: V. Capasso, E. Merzbach, B. G. Ivanoff,
M. Dozzi, R. Dalang, T. Mountford, Topics in Spatial
Stochastic Processes. Martina Franca, Italy 2001. Editor:
E. Merzbach (2003)
Vol. 1803: G. Dolzmann, Variational Methods for Crys
talline Microstructure – Analysis and Computation (2003)
Vol. 1804: I. Cherednik, Ya. Markov, R. Howe, G. Lusztig,
IwahoriHecke Algebras and their Representation Theory.
Martina Franca, Italy 1999. Editors: V. Baldoni, D. Bar
basch (2003)
Vol. 1805: F. Cao, Geometric Curve Evolution and Image
Processing (2003)
Vol. 1806: H. Broer, I. Hoveijn. G. Lunther, G. Vegter,
Bifurcations in Hamiltonian Systems. Computing Singu
larities by Gröbner Bases (2003)
Vol. 1807: V. D. Milman, G. Schechtman (Eds.), Geomet
ric Aspects of Functional Analysis. Israel Seminar 2000
2002 (2003)
Vol. 1808: W. Schindler, Measures with Symmetry Prop
erties (2003)
Vol. 1809: O. Steinbach, Stability Estimates for Hybrid
Coupled Domain Decomposition Methods (2003)
Vol. 1810: J. Wengenroth, Derived Functors in Functional
Analysis (2003)
Vol. 1811: J. Stevens, Deformations of Singularities
(2003)
Vol. 1812: L. Ambrosio, K. Deckelnick, G. Dziuk,
M. Mimura, V. A. Solonnikov, H. M. Soner, Mathematical
Aspects of Evolving Interfaces. Madeira, Funchal, Portu
gal 2000. Editors: P. Colli, J. F. Rodrigues (2003)
Vol. 1813: L. Ambrosio, L. A. Caffarelli, Y. Brenier,
G. Buttazzo, C. Villani, Optimal Transportation and its
Applications. Martina Franca, Italy 2001. Editors: L. A.
Caffarelli, S. Salsa (2003)
Vol. 1814: P. Bank, F. Baudoin, H. Föllmer, L.C.G.
Rogers, M. Soner, N. Touzi, ParisPrinceton Lectures on
Mathematical Finance 2002 (2003)
Vol. 1815: A. M. Vershik (Ed.), Asymptotic Com
binatorics with Applications to Mathematical Physics.
St. Petersburg, Russia 2001 (2003)
Vol. 1816: S. Albeverio, W. Schachermayer, M. Tala
grand, Lectures on Probability Theory and Statistics.
Ecole d’Eté de Probabilités de SaintFlour XXX2000.
Editor: P. Bernard (2003)
Vol. 1817: E. Koelink, W. Van Assche (Eds.), Orthogonal
Polynomials and Special Functions. Leuven 2002 (2003)
Vol. 1818: M. Bildhauer, Convex Variational Problems
with Linear, nearly Linear and/or Anisotropic Growth
Conditions (2003)
Vol. 1819: D. Masser, Yu. V. Nesterenko, H. P. Schlick
ewei, W. M. Schmidt, M. Waldschmidt, Diophantine
Approximation. Cetraro, Italy 2000. Editors: F. Amoroso,
U. Zannier (2003)
Vol. 1820: F. Hiai, H. Kosaki, Means of Hilbert Space
Operators (2003)
Vol. 1821: S. Teufel, Adiabatic Perturbation Theory in
Quantum Dynamics (2003)
Vol. 1822: S.N. Chow, R. Conti, R. Johnson, J. Mallet
Paret, R. Nussbaum, Dynamical Systems. Cetraro, Italy
2000. Editors: J. W. Macki, P. Zecca (2003)
Vol. 1823: A. M. Anile, W. Allegretto, C. Ring
hofer, Mathematical Problems in Semiconductor Physics.
Cetraro, Italy 1998. Editor: A. M. Anile (2003)
Vol. 1824: J. A. Navarro González, J. B. Sancho de Salas,
C
∞
– Differentiable Spaces (2003)
Vol. 1825: J. H. Bramble, A. Cohen, W. Dahmen, Mul
tiscale Problems and Methods in Numerical Simulations,
Martina Franca, Italy 2001. Editor: C. Canuto (2003)
Vol. 1826: K. Dohmen, Improved Bonferroni Inequal
ities via Abstract Tubes. Inequalities and Identities of
InclusionExclusion Type. VIII, 113 p, 2003.
Vol. 1827: K. M. Pilgrim, Combinations of Complex
Dynamical Systems. IX, 118 p, 2003.
Vol. 1828: D. J. Green, Gröbner Bases and the Computa
tion of Group Cohomology. XII, 138 p, 2003.
Vol. 1829: E. Altman, B. Gaujal, A. Hordijk, Discrete
Event Control of Stochastic Networks: Multimodularity
and Regularity. XIV, 313 p, 2003.
Vol. 1830: M. I. Gil’, Operator Functions and Localization
of Spectra. XIV, 256 p, 2003.
Vol. 1831: A. Connes, J. Cuntz, E. Guentner, N. Hig
son, J. E. Kaminker, Noncommutative Geometry, Martina
Franca, Italy 2002. Editors: S. Doplicher, L. Longo (2004)
Vol. 1832: J. Azéma, M. Émery, M. Ledoux, M. Yor
(Eds.), Séminaire de Probabilités XXXVII (2003)
Vol. 1833: D.Q. Jiang, M. Qian, M.P. Qian, Mathemati
cal Theory of Nonequilibrium Steady States. On the Fron
tier of Probability and Dynamical Systems. IX, 280 p,
2004.
Vol. 1834: Yo. Yomdin, G. Comte, Tame Geometry with
Application in Smooth Analysis. VIII, 186 p, 2004.
Vol. 1835: O.T. Izhboldin, B. Kahn, N.A. Karpenko,
A. Vishik, Geometric Methods in the Algebraic Theory
of Quadratic Forms. Summer School, Lens, 2000. Editor:
J.P. Tignol (2004)
Vol. 1836: C. Nˇ astˇ asescu, F. Van Oystaeyen, Methods of
Graded Rings. XIII, 304 p, 2004.
Vol. 1837: S. Tavaré, O. Zeitouni, Lectures on Probabil
ity Theory and Statistics. Ecole d’Eté de Probabilités de
SaintFlour XXXI2001. Editor: J. Picard (2004)
Vol. 1838: A.J. Ganesh, N.W. O’Connell, D.J. Wischik,
Big Queues. XII, 254 p, 2004.
Vol. 1839: R. Gohm, Noncommutative Stationary
Processes. VIII, 170 p, 2004.
Vol. 1840: B. Tsirelson, W. Werner, Lectures on Probabil
ity Theory and Statistics. Ecole d’Eté de Probabilités de
SaintFlour XXXII2002. Editor: J. Picard (2004)
Vol. 1841: W. Reichel, Uniqueness Theorems for Vari
ational Problems by the Method of Transformation
Groups (2004)
Vol. 1842: T. Johnsen, A. L. Knutsen, K
3
Projective Mod
els in Scrolls (2004)
Vol. 1843: B. Jefferies, Spectral Properties of Noncom
muting Operators (2004)
Vol. 1844: K.F. Siburg, The Principle of Least Action in
Geometry and Dynamics (2004)
Vol. 1845: Min Ho Lee, Mixed Automorphic Forms, Torus
Bundles, and Jacobi Forms (2004)
Vol. 1846: H. Ammari, H. Kang, Reconstruction of Small
Inhomogeneities from Boundary Measurements (2004)
Vol. 1847: T.R. Bielecki, T. Björk, M. Jeanblanc, M.
Rutkowski, J.A. Scheinkman, W. Xiong, ParisPrinceton
Lectures on Mathematical Finance 2003 (2004)
Vol. 1848: M. Abate, J. E. Fornaess, X. Huang, J. P. Rosay,
A. Tumanov, Real Methods in Complex and CR Geom
etry, Martina Franca, Italy 2002. Editors: D. Zaitsev, G.
Zampieri (2004)
Vol. 1849: Martin L. Brown, Heegner Modules and Ellip
tic Curves (2004)
Vol. 1850: V. D. Milman, G. Schechtman (Eds.), Geomet
ric Aspects of Functional Analysis. Israel Seminar 2002
2003 (2004)
Vol. 1851: O. Catoni, Statistical Learning Theory and
Stochastic Optimization (2004)
Vol. 1852: A.S. Kechris, B.D. Miller, Topics in Orbit
Equivalence (2004)
Vol. 1853: Ch. Favre, M. Jonsson, The Valuative Tree
(2004)
Vol. 1854: O. Saeki, Topology of Singular Fibers of Dif
ferential Maps (2004)
Vol. 1855: G. Da Prato, P.C. Kunstmann, I. Lasiecka,
A. Lunardi, R. Schnaubelt, L. Weis, Functional Analytic
Methods for Evolution Equations. Editors: M. Iannelli,
R. Nagel, S. Piazzera (2004)
Vol. 1856: K. Back, T.R. Bielecki, C. Hipp, S. Peng,
W. Schachermayer, Stochastic Methods in Finance, Bres
sanone/Brixen, Italy, 2003. Editors: M. Fritelli, W. Rung
galdier (2004)
Vol. 1857: M. Émery, M. Ledoux, M. Yor (Eds.), Sémi
naire de Probabilités XXXVIII (2005)
Vol. 1858: A.S. Cherny, H.J. Engelbert, Singular Stochas
tic Differential Equations (2005)
Vol. 1859: E. Letellier, Fourier Transforms of Invariant
Functions on Finite Reductive Lie Algebras (2005)
Vol. 1860: A. Borisyuk, G.B. Ermentrout, A. Friedman,
D. Terman, Tutorials in Mathematical Biosciences I.
Mathematical Neurosciences (2005)
Vol. 1861: G. Benettin, J. Henrard, S. Kuksin, Hamil
tonian Dynamics – Theory and Applications, Cetraro,
Italy, 1999. Editor: A. Giorgilli (2005)
Vol. 1862: B. Helffer, F. Nier, Hypoelliptic Estimates and
Spectral Theory for FokkerPlanck Operators and Witten
Laplacians (2005)
Vol. 1863: H. Führ, Abstract Harmonic Analysis of Con
tinuous Wavelet Transforms (2005)
Vol. 1864: K. Efstathiou, Metamorphoses of Hamiltonian
Systems with Symmetries (2005)
Vol. 1865: D. Applebaum, B.V. R. Bhat, J. Kustermans,
J. M. Lindsay, Quantum Independent Increment Processes
I. From Classical Probability to Quantum Stochastic Cal
culus. Editors: M. Schürmann, U. Franz (2005)
Vol. 1866: O.E. BarndorffNielsen, U. Franz, R. Gohm,
B. Kümmerer, S. Thorbjønsen, Quantum Independent
Increment Processes II. Structure of Quantum Lévy
Processes, Classical Probability, and Physics. Editors: M.
Schürmann, U. Franz, (2005)
Vol. 1867: J. Sneyd (Ed.), Tutorials in Mathematical Bio
sciences II. Mathematical Modeling of CalciumDynamics
and Signal Transduction. (2005)
Vol. 1868: J. Jorgenson, S. Lang, Pos
n
(R) and Eisenstein
Series. (2005)
Vol. 1869: A. Dembo, T. Funaki, Lectures on Probabil
ity Theory and Statistics. Ecole d’Eté de Probabilités de
SaintFlour XXXIII2003. Editor: J. Picard (2005)
Vol. 1870: V.I. Gurariy, W. Lusky, Geometry of Müntz
Spaces and Related Questions. (2005)
Vol. 1871: P. Constantin, G. Gallavotti, A.V. Kazhikhov,
Y. Meyer, S. Ukai, Mathematical Foundation of Turbu
lent Viscous Flows, Martina Franca, Italy, 2003. Editors:
M. Cannone, T. Miyakawa (2006)
Vol. 1872: A. Friedman (Ed.), Tutorials in Mathemati
cal Biosciences III. Cell Cycle, Proliferation, and Cancer
(2006)
Vol. 1873: R. Mansuy, M. Yor, Random Times and En
largements of Filtrations in a Brownian Setting (2006)
Vol. 1874: M. Yor, M. Émery (Eds.), In Memoriam Paul
André Meyer  Séminaire de Probabilités XXXIX (2006)
Vol. 1875: J. Pitman, Combinatorial Stochastic Processes.
Ecole d’Eté de Probabilités de SaintFlour XXXII2002.
Editor: J. Picard (2006)
Vol. 1876: H. Herrlich, Axiom of Choice (2006)
Vol. 1877: J. Steuding, Value Distributions of LFunctions
(2007)
Vol. 1878: R. Cerf, The Wulff Crystal in Ising and Percol
ation Models, Ecole d’Eté de Probabilités de SaintFlour
XXXIV2004. Editor: Jean Picard (2006)
Vol. 1879: G. Slade, The Lace Expansion and its Applica
tions, Ecole d’Eté de Probabilités de SaintFlour XXXIV
2004. Editor: Jean Picard (2006)
Vol. 1880: S. Attal, A. Joye, C.A. Pillet, Open Quantum
Systems I, The Hamiltonian Approach (2006)
Vol. 1881: S. Attal, A. Joye, C.A. Pillet, Open Quantum
Systems II, The Markovian Approach (2006)
Vol. 1882: S. Attal, A. Joye, C.A. Pillet, Open Quantum
Systems III, Recent Developments (2006)
Vol. 1883: W. Van Assche, F. Marcellàn (Eds.), Orthogo
nal Polynomials and Special Functions, Computation and
Application (2006)
Vol. 1884: N. Hayashi, E.I. Kaikina, P.I. Naumkin,
I.A. Shishmarev, Asymptotics for Dissipative Nonlinear
Equations (2006)
Vol. 1885: A. Telcs, The Art of Random Walks (2006)
Vol. 1886: S. Takamura, Splitting Deformations of Dege
nerations of Complex Curves (2006)
Vol. 1887: K. Habermann, L. Habermann, Introduction to
Symplectic Dirac Operators (2006)
Vol. 1888: J. van der Hoeven, Transseries and Real Differ
ential Algebra (2006)
Vol. 1889: G. Osipenko, Dynamical Systems, Graphs, and
Algorithms (2006)
Vol. 1890: M. Bunge, J. Funk, Singular Coverings of
Toposes (2006)
Vol. 1891: J.B. Friedlander, D.R. HeathBrown,
H. Iwaniec, J. Kaczorowski, Analytic Number Theory,
Cetraro, Italy, 2002. Editors: A. Perelli, C. Viola (2006)
Vol. 1892: A. Baddeley, I. Bárány, R. Schneider, W. Weil,
Stochastic Geometry, Martina Franca, Italy, 2004. Editor:
W. Weil (2007)
Vol. 1893: H. Hanßmann, Local and SemiLocal Bifur
cations in Hamiltonian Dynamical Systems, Results and
Examples (2007)
Vol. 1894: C.W. Groetsch, Stable Approximate Evaluation
of Unbounded Operators (2007)
Vol. 1895: L. Molnár, Selected Preserver Problems on
Algebraic Structures of Linear Operators and on Function
Spaces (2007)
Vol. 1896: P. Massart, Concentration Inequalities and
Model Selection, Ecole d’Été de Probabilités de Saint
Flour XXXIII2003. Editor: J. Picard (2007)
Vol. 1897: R. Doney, Fluctuation Theory for Lévy
Processes, Ecole d’Été de Probabilités de SaintFlour
XXXV2005. Editor: J. Picard (2007)
Vol. 1898: H.R. Beyer, Beyond Partial Differential Equa
tions, On linear and QuasiLinear Abstract Hyperbolic
Evolution Equations (2007)
Vol. 1899: Séminaire de Probabilités XL. Editors:
C. DonatiMartin, M. Émery, A. Rouault, C. Stricker
(2007)
Vol. 1900: E. Bolthausen, A. Bovier (Eds.), Spin Glasses
(2007)
Vol. 1901: O. Wittenberg, Intersections de deux
quadriques et pinceaux de courbes de genre 1, Inter
sections of Two Quadrics and Pencils of Curves of Genus
1 (2007)
Vol. 1902: A. Isaev, Lectures on the Automorphism
Groups of KobayashiHyperbolic Manifolds (2007)
Vol. 1903: G. Kresin, V. Maz’ya, Sharp RealPart Theo
rems (2007)
Vol. 1904: P. Giesl, Construction of Global Lyapunov
Functions Using Radial Basis Functions (2007)
Vol. 1905: C. Prévˆ ot, M. Röckner, A Concise Course on
Stochastic Partial Differential Equations (2007)
Vol. 1906: T. Schuster, The Method of Approximate
Inverse: Theory and Applications (2007)
Vol. 1907: M. Rasmussen, Attractivity and Bifurcation for
Nonautonomous Dynamical Systems (2007)
Vol. 1908: T.J. Lyons, M. Caruana, T. Lévy, Differential
Equations Driven by Rough Paths, Ecole d’Été de Proba
bilités de SaintFlour XXXIV2004 (2007)
Vol. 1909: H. Akiyoshi, M. Sakuma, M. Wada,
Y. Yamashita, Punctured Torus Groups and 2Bridge Knot
Groups (I) (2007)
Vol. 1910: V.D. Milman, G. Schechtman (Eds.), Geo
metric Aspects of Functional Analysis. Israel Seminar
20042005 (2007)
Vol. 1911: A. Bressan, D. Serre, M. Williams,
K. Zumbrun, Hyperbolic Systems of Balance Laws.
Cetraro, Italy 2003. Editor: P. Marcati (2007)
Vol. 1912: V. Berinde, Iterative Approximation of Fixed
Points (2007)
Vol. 1913: J.E. Marsden, G. Misiołek, J.P. Ortega,
M. Perlmutter, T.S. Ratiu, Hamiltonian Reduction by
Stages (2007)
Vol. 1914: G. Kutyniok, Afﬁne Density in Wavelet
Analysis (2007)
Vol. 1915: T. Bıyıkoˇ glu, J. Leydold, P.F. Stadler, Laplacian
Eigenvectors of Graphs. PerronFrobenius and Faber
Krahn Type Theorems (2007)
Vol. 1916: C. Villani, F. Rezakhanlou, Entropy Methods
for the Boltzmann Equation. Editors: F. Golse, S. Olla
(2008)
Vol. 1917: I. Veseli´ c, Existence and Regularity Properties
of the Integrated Density of States of RandomSchrödinger
(2008)
Vol. 1918: B. Roberts, R. Schmidt, Local Newforms for
GSp(4) (2007)
Vol. 1919: R.A. Carmona, I. Ekeland, A. Kohatsu
Higa, J.M. Lasry, P.L. Lions, H. Pham, E. Taﬂin,
ParisPrinceton Lectures on Mathematical Finance 2004.
Editors: R.A. Carmona, E. Çinlar, I. Ekeland, E. Jouini,
J.A. Scheinkman, N. Touzi (2007)
Vol. 1920: S.N. Evans, Probability and Real Trees. Ecole
d’Été de Probabilités de SaintFlour XXXV2005 (2008)
Vol. 1921: J.P. Tian, Evolution Algebras and their Appli
cations (2008)
Vol. 1922: A. Friedman (Ed.), Tutorials in Mathematical
BioSciences IV. Evolution and Ecology (2008)
Vol. 1923: J.P.N. Bishwal, Parameter Estimation in
Stochastic Differential Equations (2008)
Vol. 1924: M. Wilson, LittlewoodPaley Theory and
ExponentialSquare Integrability (2008)
Vol. 1925: M. du Sautoy, L. Woodward, Zeta Functions of
Groups and Rings (2008)
Vol. 1926: L. Barreira, V. Claudia, Stability of Nonauto
nomous Differential Equations (2008)
Vol. 1927: L. Ambrosio, L. Caffarelli, M.G. Crandall,
L.C. Evans, N. Fusco, Calculus of Variations and Non
Linear Partial Differential Equations. Cetraro, Italy 2005.
Editors: B. Dacorogna, P. Marcellini (2008)
Vol. 1928: J. Jonsson, Simplicial Complexes of Graphs
(2008)
Vol. 1929: Y. Mishura, Stochastic Calculus for Fractional
Brownian Motion and Related Processes (2008)
Vol. 1930: J.M. Urbano, The Method of Intrinsic Scaling.
A Systematic Approach to Regularity for Degenerate and
Singular PDEs (2008)
Vol. 1931: M. Cowling, E. Frenkel, M. Kashiwara,
A. Valette, D.A. Vogan, Jr., N.R. Wallach, Representation
Theory and Complex Analysis. Venice, Italy 2004.
Editors: E.C. Tarabusi, A. D’Agnolo, M. Picardello
(2008)
Vol. 1932: A.A. Agrachev, A.S. Morse, E.D. Sontag,
H.J. Sussmann, V.I. Utkin, Nonlinear and Optimal Control
Theory. Cetraro, Italy 2004. Editors: P. Nistri, G. Stefani
(2008)
Vol. 1933: M. Petkovic, Point Estimation of Root Finding
Methods (2008)
Vol. 1934: C. DonatiMartin, M. Émery, A. Rouault,
C. Stricker (Eds.), Séminaire de Probabilités XLI (2008)
Vol. 1935: A. Unterberger, Alternative Pseudodifferential
Analysis (2008)
Vol. 1936: P. Magal, S. Ruan (Eds.), Structured Population
Models in Biology and Epidemiology (2008)
Vol. 1937: G. Capriz, P. Giovine, P.M. Mariano (Eds.),
Mathematical Models of Granular Matter (2008)
Vol. 1938: D. Auroux, F. Catanese, M. Manetti, P. Seidel,
B. Siebert, I. Smith, G. Tian, Symplectic 4Manifolds
and Algebraic Surfaces. Cetraro, Italy 2003. Editors:
F. Catanese, G. Tian (2008)
Vol. 1939: D. Bofﬁ, F. Brezzi, L. Demkowicz, R.G. Durán,
R.S. Falk, M. Fortin, Mixed Finite Elements, Compati
bility Conditions, and Applications. Cetraro, Italy 2006.
Editors: D. Bofﬁ, L. Gastaldi (2008)
Vol. 1940: J. Banasiak, V. Capasso, M.A.J. Chaplain,
M. Lachowicz, J. Mie¸kisz, Multiscale Problems in the Life
Sciences. From Microscopic to Macroscopic. Be¸dlewo,
Poland 2006. Editors: V. Capasso, M. Lachowicz (2008)
Vol. 1941: S.M.J. Haran, Arithmetical Investigations.
Representation Theory, Orthogonal Polynomials, and
Quantum Interpolations (2008)
Vol. 1942: S. Albeverio, F. Flandoli, Y.G. Sinai, SPDE in
Hydrodynamic. Recent Progress and Prospects. Cetraro,
Italy 2005. Editors: G. Da Prato, M. Röckner (2008)
Vol. 1943: L.L. Bonilla (Ed.), Inverse Problems and Imag
ing. Martina Franca, Italy 2002 (2008)
Vol. 1944: A. Di Bartolo, G. Falcone, P. Plaumann,
K. Strambach, Algebraic Groups and Lie Groups with
Few Factors (2008)
Recent Reprints and New Editions
Vol. 1702: J. Ma, J. Yong, ForwardBackward Stochas
tic Differential Equations and their Applications. 1999 –
Corr. 3rd printing (2007)
Vol. 830: J.A. Green, Polynomial Representations of GL
n
,
with an Appendix on Schensted Correspondence and Lit
telmann Paths by K. Erdmann, J.A. Green and M. Schoker
1980 – 2nd corr. and augmented edition (2007)
Vol. 1693: S. Simons, From HahnBanach to Monotonic
ity (Minimax and Monotonicity 1998) – 2nd exp. edition
(2008)
Vol. 470: R. Bowen, Equilibrium States and the Ergodic
Theory of Anosov Diffeomorphisms. With a preface by
D. Ruelle. Edited by J.R. Chazottes. 1975 – 2nd rev.
edition (2008)
Vol. 523: S.A. Albeverio, R.J. HøeghKrohn,
S. Mazzucchi, Mathematical Theory of Feynman Path
Integral. 1976 – 2nd corr. and enlarged edition (2008)
Rufus Bowen
Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms
Second revised edition
JeanRené Chazottes
Editor
Preface by David Ruelle
ABC
Author Robert Edward (Rufus) Bowen (19471978)
University of California at Berkeley USA
Editor JeanRené Chazottes
Centre de Physique Théorique CNRSÉcole Polytechnique 91128 Palaiseau Cedex France jeanrene@cpht.polytechnique.fr
Preface by David Ruelle
IHÉS 35, Route de Chartres 91440 BuressurYvette France ruelle@ihes.fr
ISBN: 9783540776055 eISBN: 9783540776956 DOI: 10.1007/9783540776956
Lecture Notes in Mathematics ISSN print edition: 00758434 ISSN electronic edition: 16179692 Library of Congress Control Number: 2008922889 Mathematics Subject Classiﬁcation (2000): Primary: 37D20, Secondary: 37D35 c 2008, 1975 SpringerVerlag Berlin Heidelberg This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, speciﬁcally the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microﬁlm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a speciﬁc statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use.
Cover design: WMXDesign GmbH
Printed on acidfree paper 987654321 springer.com
Preface
The Greek and Roman gods, supposedly, resented those mortals endowed with superlative gifts and happiness, and punished them. The life and achievements of Rufus Bowen (1947–1978) remind us of this belief of the ancients. When Rufus died unexpectedly, at age thirtyone, from brain hemorrhage, he was a very happy and successful man. He had great charm, that he did not misuse, and superlative mathematical talent. His mathematical legacy is important, and will not be forgotten, but one wonders what he would have achieved if he had lived longer. Bowen chose to be simple rather than brilliant. This was the hard choice, especially in a messy subject like smooth dynamics in which he worked. Simplicity had also been the style of Steve Smale, from whom Bowen learned dynamical systems theory. Rufus Bowen has left us a masterpiece of mathematical exposition: the slim volume Equilibrium States and the Ergodic Theory of Anosov Diﬀeomorphisms (Springer Lecture Notes in Mathematics 470 (1975)). Here a number of results which were new at the time are presented in such a clear and lucid style that Bowen’s monograph immediately became a classic. More than thirty years later, many new results have been proved in this area, but the volume is as useful as ever because it remains the best introduction to the basics of the ergodic theory of hyperbolic systems. The area discussed by Bowen came into existence through the merging of two apparently unrelated theories. One theory was equilibrium statistical mechanics, and speciﬁcally the theory of states of inﬁnite systems (Gibbs states, equilibrium states, and their relations as discussed by R.L. Dobrushin, O.E. Lanford, and D. Ruelle). The other theory was that of hyperbolic smooth dynamical systems, with the major contributions of D.V. Anosov and S. Smale. The two theories came into contact when Ya.G. Sinai introduced Markov partitions and symbolic dynamics for Anosov diﬀeomorphisms. This allowed the poweful techniques and results of statistical mechanics to be applied to smooth dynamics, an extraordinary development in which Rufus Bowen played a major role. Some of Bowen’s ideas were as follows. First, only onedimensional statistical mechanics is discussed: this is a richer theory, which yields what is
Bures sur Yvette. JeanRen´ Chazottes has had the idea to make Bowen’s monograph more e easily available by retyping it. and Gerhard Keller. mai 2007 David Ruelle . For this work of love all of them deserve our warmest thanks. This text has not aged since it was written and its beauty is as striking as when it was ﬁrst published in 1975.VI Preface needed for applications to dynamical systems. Second. Pierre o Collet. Third. In his enterprise he has been helped by Jerˆme Buzzi. and makes use of the powerful analytic tool of transfer operators. in particular in the bibliography. The combination of simpliﬁcations and generalizations just outlined led to Bowen’s concise and lucid monograph. but corrected a number of typos and made a few other minor corrections. to improve usefulness and readability. Smale’s Axiom A dynamical systems are studied rather than the less general Anosov systems. He has scrupulously respected the original text and notation. Sinai’s Markov partitions are reworked to apply to Axiom A systems and their construction is simpliﬁed by the use of shadowing.
. . . . . . . . . . . . . . . . . . . . . . . . . Deﬁnition . . Axiom a Diﬀeomorphisms . Attractors and Anosov Diﬀeomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E. . . . . . . . . . . . . . . . . Gibbs Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B. . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . Gibbs Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Equilibrium States for Basic Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ergodic Theory of Axiom a Diﬀeomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Contents 0 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Variational Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . 1 3 3 8 13 16 22 26 29 29 32 36 42 43 45 45 47 51 56 58 61 61 64 69 72 2 3 4 . . . . Variational Principle . . . . . . . . . . . . . B. . . . . . . . . . . . . . . . . . . . . . . . . . D. . . . . . C. . . . . . . . . . . . . . . . . . . . A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C. . . General Thermodynamic Formalism . . . . . . . . . . . References . . C. . . . . B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Spectral Decomposition . . . . . Entropy . . . . . . . . . . . . . . . . . . . A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ruelle’s PerronFrobenius Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . Construction of Gibbs Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. . . . . . . . Pressure . . . . . . . D. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Case φ = φ(u) . . . C. . . Symbolic Dynamics . . Markov Partitions . . . . . . Equilibrium States . . . . . . . . . . . . . . . . . . . . . . . . . A. . . . . . . . . . . . . . . . . . . . . . . . . . Further Properties . . . . . . . . . . . . B. . . . . . . . . . . . D. . . . . . . . .
. . . . . . . .VIII Contents Index . . . . 75 These notes came out of a course given at the University of Minnesota and were revised while the author was on a Sloan Fellowship. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
To start. These measures on shift spaces arise in modern statistical mechanics. Ruelle and Ya. First we study the statistical properties of Gibbs measures. On the whole these notes owe most to D. In the ﬁnal chapter this symbolic dynamics is applied to the ergodic theory of Axiom A diﬀeomorphisms. Sinai. The Gibbs measures also satisfy a variational principle. The material of these notes is taken from the work of many people. µ) is a probability space if B is a σﬁeld of subsets of X and µ is a nonnegative measure on B with µ(X) = 1. These diﬀeomorphisms have a complicated orbit structure that is perhaps best understood by relating them topologically and measure theoretically to shift spaces. but no doubt some are missing. I have attempted to give the main references at the end of each chapter.0 Introduction The main purpose of these notes is to present the ergodic theory of Anosov and Axiom A diﬀeomorphisms. This idea of studying the same example from diﬀerent viewpoints is of course how the subjects of topological dynamics and ergodic theory arose from mechanics. µ(T −1 E) = µ(E) for E ∈ B . By an automorphism we mean a bijection T : X → X for which (i) (ii) E∈B iﬀ T −1 E ∈ B. B. Here these subjects return to help us understand diﬀerentiable systems. This principle is important because it makes no reference to the shift character of the underlying space. recall that (X. this is carried out in chapter two. These notes are divided into four chapters. . Through this one is led to develop a “thermodynamic formalism” on compact spaces. In the third chapter Axiom A diﬀeomorphisms are introduced and their symbolic dynamics constructed: this states how they are related to shift spaces. they interest us because they solve the problem of determining an invariant measure when you know it approximately in a certain sense.
M (X) is a compact convex metrizable space. Proposition. MT (X) is a nonempty closed subset of M (X).2 0 Introduction If T : X → X is a homeomorphism of a compact metric space. Let M (X) be the set of Borel probability measures on X and MT (X) the subset of invariant ones. The weak ∗topology on the space C (X)∗ of continuous linear functionals α : C (X) → R is generated by sets of the form U (f. Choose a subsequence µnk converging to µ ∈ M (X). i. ε. A probability measure on this σﬁeld is called a Borel probability measure. Proof. Riesz Representation. Let {fn }∞ be a dense subset of C (X). Proof. We identify αµ with µ. ε > 0. α0 ) = {α ∈ C (X)∗ : α(f ) − α0 (f ) < ε} with f ∈ C (X). T µ + · · · + (T ) Then µ ∈ MT (X). µ ) = n=1 2−n fn −1 fn dµ − fn dµ . . Remember that the realvalued continuous functions C (X) on the compact metric space X form a Banach space under f = maxx∈X f (x). often writing µ when we mean α(µ). The reader may check that n=1 the weak topology on M (X) is equivalent to the one deﬁned by the metric ∞ d(µ. Proof. µ ∈ MT (X) if and only if (f ◦ T ) dµ = f dµ for all f ∈ C (X) . The weak ∗topology on C (X)∗ carries over by this identiﬁcation to a topology on M (X) (called the weak topology). This is just what the Riesz Representation Theorem says about the statement T ∗ µ = µ. Check that T ∗ : M (X) → M (X) is a homeomorphism and note that 1 MT (X) = {µ ∈ M (X) : T ∗ µ = µ}. a natural σﬁeld B is the family of Borel sets. For any µ ∈ M (X) one can deﬁne T ∗ µ ∈ M (X) by T ∗ µ(E) = µ(T −1 E). Proposition.e. α0 ∈ C (X)∗ . Then µ ↔ αµ is a bijection between M (X) and {α ∈ C (X)∗ : α(1) = 1 and α(f ) ≥ 0 whenever f ≥ 0} . For each µ ∈ M (X) deﬁne αµ ∈ C (X)∗ by αµ (f ) = f dµ. Proposition. µ ∈ MT (X) if µ(T −1 E) = µ(E) for all Borel sets E. Pick µ ∈ M (X) and let µn = n (µ + ∗ ∗ n−1 µ).
.1 Gibbs Measures A. pn ) ranges over the simplex {(p1 . and the temperature T of the heat source remains constant as it is so much larger than our system. . Gibbs Distribution Suppose a physical system has possible states 1. . an be given. pn ) : pi ≥ 0. . We shall not attempt the physical justiﬁcation for the Gibbs distribution.1. Energy is thereby allowed to pass between the original system and the heat source. Lemma. Suppose that this system is put in contact with a much larger “heat source” which is at temperature T . but we will state a mathematical fact closely connected to the physical reasoning. . . . . n and the energies of these states are E1 . . Then the quantity n n F (p1 . . . . . . p1 + · · · + pn = 1} and that maximum is assumed only by −1 pj = e aj i e ai . 1 where β = kT and k is a physical constant. . . 1. En . . It is a physical fact derived in statistical mechanics that the probability pj that state j occurs is given by the Gibbs distribution pj = e−βEj n −βEi i=1 e . pn ) = i=1 n −pi log pi + i=1 pi ai has maximum value log i=1 eai as (p1 . . Let real numbers a1 . . . . . As the energy of our system is not ﬁxed any of the states could occur. . . .
2 j=0 (We “give” each of x0 and xk half the energy due to their interaction). k2 ) = Φ2 (−j. k2 ) depends only on their relative position. the onedimensional lattice. . . .. k1 . k2 ) satisﬁes ∞ Φ2 j=1 j < ∞. The Gibbs distribution then maximizes S − βE = S − 1 E.e.” We will now look at a simple inﬁnite system. but when energy is not ﬁxed “nature minimizes free energy.) The term i=1 pi ai is of course the average value of the function a(i) = ai . k2 . . n} × {1. . . . . . k1 . . there is a function Φ2 : Z × {1. . . kT or equivalently minimizes E − kT S. k2 ) = Φ2 (i1 − i2 . i2 . . n. then the potential energy due to their interaction Φ∗ (i1 . We now make assumptions about energy: (1) associated with the occurrence of a state k is a contribution Φ0 (k) to the total energy of the system independent of which position it occurs at. . . 2 i. entropy is denoted S and average energy E. A conﬁguration of the system consists of assigning an xi ∈ {1. . (2) if state k1 occurs in place i1 . . . 2. Under these hypotheses the energy contribution due to x0 being in the 0th place is 1 φ∗ (x) = Φ0 (x0 ) + Φ2 (j. . . In the statistical mechanics case ai = −βEi . . k2 ) 2 (also: Φ2 (j. . xj . k1 . .4 1 Gibbs Measures n This is proved by calculus. . . . . n} for each i: · · x−2 · x−1 · x0 · x1 · x2 · x3 · · · Thus a conﬁguration is a point x = {xi }+∞ ∈ i=−∞ Z {1. k1 . pn ) (note: ϕ(x) = −x log x is n continuous on [0. and k2 in i2 . Here one has for each integer a physical system with possible states 1. i2 . pn ) = i=1 −pi log pi is called the entropy of the distribution (p1 . 1] if we set ϕ(0) = 0. k1 )). . . n} → R so that Φ∗ (i1 . The quantity H(p1 . k1 . We now assume that Φ2 j = supk1 . This is called the free energy. (3) all energy is due to contributions of the form (1) and (2). x0 ).k2 Φ(j. The principle that “nature minimizes entropy” applies when energy is ﬁxed. n} = Σn .
. . i.e. . . xm ) diﬀers from m m ∗ j ∗ j j=−m φ (σ x) by at most C. n} is given the discrete topology and Σn = Z {1. 2 2 j=−m k=j+m+1 ∞ k=m−j+1 k=1 Thus. The point is that taking xi into consideration for i ∈ [−m. . Then there is a unique µ ∈ Mσ (Σn ) for which one can ﬁnd constants c1 > 0. Suppose φ : Σn → R and there are c > 0. . α ∈ (0. . . . . . .. Gibbs Distribution 5 Then φ∗ (x) ∈ R and depends continuously on x when {1. . . 1) so that vark φ ≤ cαk for all k. .2. . If we just look at x−m . . . x0 . the probabilities would change by factors in [e−2C . e2C ]. . If we used j=−m φ (σ x) instead of Em (x−m . . xm ) one might add in the contributions by interactions of xj (−m ≤ j ≤ m) with all other xk ’s. . 1. . .A. . xk . As φ is uniformly continuous. . xk . but not drastically if one assumes / ∞ k Φ2 k < ∞. . .. For φ : Σn → R continuous deﬁne vark φ = sup{φ(x) − φ(y) : xi = yi ∀ i ≤ k} . . . k=1 We want now to state a theorem we have been leading up to.xm ) . If we are given x = {xi }∞ i=−∞ . If we deﬁne the (left) shift homeomorphism σ : Σn → Σn by σ{xi }∞ i=−∞ = m . xm ) = j=−m Φ0 (xj ) + −m≤j<k≤m Φ2 (k − j. . and P such that . if C = k=1 k Φ2 k < ∞ then Em (x−m .. Then µ ∈ M (Σn ) and it would be natural to call µ the Gibbs distribution on Σn (for the given energy and β). xm ) in the Gibbs distribution µm . .. m ∞ Φ0 (xj ) + j=−m k=−∞ 1 Φ2 (k − j. . . xm the limit µ(x−m . then this expression is just φ∗ (σ j x). . Now just suppose that for each x−m . This expression {xi+1 }∞ i=−∞ j=−m diﬀers from Em (x−m . . . xm we have a ﬁnite system (n2m+1 possible conﬁgurations) and an energy m Em (x−m . xm ) = lim k→∞ {µk (x−k . Theorem. xj ) 2 . xm ) by at most ⎛ ⎞ m ∞ ∞ ∞ 1 1 ⎝ Φ2 k + Φ2 k ⎠ ≤ k Φ2 k . m] may change µm . xk ) : xi = xi ∀i ≤ m} exists. . c2 > 0. . then instead of Em (x−m . xj ) and the Gibbs distribution µm assigns probabilities proportional to e−βEm (x−m . limk→∞ vark φ = 0.. . . n} the product topology.
. then a0 a1 . cm with c0 = br . we consider all x in which A says that xi xi+1 is allowable for every i.j with Ui ∩ σ n Uj = ∅ ∀n ≥ Ni. c2 ] the relative probabilities of the x0 . i. . (Otherwise one could have ΣA = ΣB with B an m × m matrix and m < n.e. So ΣA is topologically mixing. Lemma. .e. Hence ﬁnd c0 . Let Ui = {x ∈ ΣA : x0 = i} = ∅. j) for some M . AM > 0 ∀i. .j Ni. We have ignored many subtleties that come up in more complicated systems (e. when U. higher dimensional lattices). For t ≥ 2r + M . From this it follows by induction that Am > 0 for all m ≥ M . σ : ΣA → ΣA is topologically mixing (i. Consider open subsets U . let ΣA = {x ∈ Σn : Axi xi+1 = 1 ∀i ∈ Z} . V are nonempty open subsets of ΣA . so Am > 0 ∀i. . . am of integers between 1 and n with (a) Aak ak+1 = 1 ∀k. . One sees inductively that Am is the number of (m + 1)strings i. there is an N so that σ m U ∩ V = ∅ ∀m ≥ N ) if and only if AM > 0 (i. This measure µ is written µφ and called Gibbs measure of φ. . In statistical mechanics Gibbs states are not deﬁned by the above theorem. V of ΣA with a ∈ U . . Suppose AM > 0 for some M . b ∈ V . (b) a0 = i. Then x = · · · b−2 b−1 b0 · · · br c1 · · · cm−1 a−r · · · a0 a1 · · · is in ΣA and x ∈ σ t U ∩ V . .. Then ∃Ni. Our discussion was a gross one intended to motivate the theorem. For the physical system discussed above one takes ∗ φ = −βφ . For later use we want to make a small generalization of Σn before we prove the theorem. As each number between 1 and n occurs as x0 for some x ∈ ΣA .3. . each row of A has at least one positive entry. Suppose ΣA is mixing. . m} exp −P m + m−1 k=0 φ(σ k x) ≤ c2 for every x ∈ Σn and m ≥ 0..j a0 a1 . . we refer to Ruelle [9] or Lanford [6] for a reﬁned outlook.j m ≥ maxi.j . One easily sees that ΣA is closed and σΣA = ΣA .6 1 Gibbs Measures c1 ≤ µ{y : yi = xi ∀ i = 0.j Proof.j .) 1. That is. . If a ∈ Ui ∩ σ n Uj . an satisﬁes (a) and (b). Up to constants in [c1 . j when i.. cm = a−r . xm ’s are given m−1 by exp k=0 φ(σ k x). where the theorem will not hold.g. m = t − 2r ≥ M and Am > 0. If A is an n × n matrix of 0’s and 1’s. There is an r so that U ⊃ {x ∈ ΣA : xk = ak ∀k ≤ r} V ⊃ {x ∈ ΣA : xk = bk ∀k ≤ r} . Ack ck+1 = 1 for all k. We will always assume that A is such that each k between 1 and n occurs at x0 for some x ∈ ΣA . . am = j.
The ﬁrst step is to reduce the φ’s one must consider. 1.x0 for k ≥ 0 for k ≤ 0 . Lemma. y) = β N where N is the largest nonnegative integer with xi = yi for every i < N . m)} exp −P m + m−1 k=0 φ(σ k x) ≤ c2 for every x ∈ ΣA and m ≥ 0. Then it holds for φ2 and µφ1 = µφ2 . Proof. Existence of Gibbs measures. Gibbs Distribution 7 Let F be the family of all continuous φ : ΣA → R for which vark φ ≤ bαk A (for all k ≥ 0) for some positive constants b and α ∈ (0.A. For each 1 ≤ t ≤ n pick {ak. m−1 m−1 m−1 φ1 (σ k x) − k=0 k=0 φ2 (σ k x) = k=0 u(σ k+1 x) − u(σ k x) = u(σ m x) − u(x) ≤ 2 u . This theorem will not be proved for some time. For any β ∈ (0. . µ unchanged. Deﬁnition. c2 > 0 and P such that c1 ≤ µ{y : yi = xi for all i ∈ [0.t = t. The exponential in the required inequality changes by at most a factor of e2 u when φ1 is replaced by φ2 . Lemma. Suppose ΣA is topologically mixing and φ ∈ F . If φ ∈ F .t }∞ k=−∞ ∈ ΣA with a0. Proof.6. Then F is just the set A of functions which have a positive H¨lder exponent with respect to dβ . φ ∈ C (ΣA ) are homologous with respect to σ (written ψ ∼ φ) if there is a u ∈ C (ΣA ) so that ψ(x) = φ(x) − u(x) + u(σx) . Thus the inequality remains valid with c1 . The o theorem we are interested in then reads 1. Suppose φ1 ∼ φ2 and Theorem 1.5. then φ is homologous to some ψ ∈ F with ψ(x) = A A ψ(y) whenever xi = yi for all i ≥ 0. Deﬁne r : ΣA → ΣA by r(x) = x∗ where x∗ = k Let xk ak.4 holds for φ1 . Two functions ψ. 1) one can deﬁne the metric dβ on ΣA by dβ (x. c2 changed and P .4. 1). 1. There is unique σinvariant Borel probability measure µ on ΣA for A which one can ﬁnd constants c1 > 0.
. for j ∈ [0.5 and 1. Let ∞ + ΣA = x∈ i=0 {1. .xi+1 = 1 for all i ≥ 0 . u is deﬁned and continuous. . then. n} : Axi . Ruelle’s PerronFrobenius Theorem We introduce now onesided shift spaces. Hence [n] 2 u(x) − u(y) ≤ j=0 ∞ φ(σ j x) − φ(σ j y) + φ(σ j r(x)) − φ(σ j r(y)) + 2 ⎛ [n] 2 α j=0 n−j αj j>[ n ] 2 ⎞ + j>[ n ] 2 n 4b α[ 2 ] · α ⎠≤ 1−α ⎜ ≤ 2b ⎝ j⎟ This shows that u ∈ F . Lemmas 1. as we wanted. n]. The ﬁnal expression depends only on {xi }∞ .6 tell us that in looking for a Gibbs measure µφ for φ ∈ F (i. D. . If xi = yi for all i ≤ n. proving Theorem 1.4) we can restrict our attention to functions A φ for which φ(x) depends only on {xi }∞ .e. φ(σ j x) − φ(σ j y) ≤ varn−j φ ≤ bαn−j and φ(σ j r(x)) − φ(σ j r(y)) ≤ bαn−j . φ(σ j x) − φ(σ j r(x)) ≤ varj φ ≤ bαj . Since σ x and σ r(x) agree in places from −j to +∞. One writes x for {xi }∞ (we will i=0 continue to write x for {xi }∞ i=−∞ but never for both things at the same time). Lind cleaned i=0 up the above proof for us.8 1 Gibbs Measures ∞ u(x) = j=0 j j (φ(σ j x) − φ(σ j r(x))) . i=0 B.. Furthermore A A ∞ ∞ ψ(x) = φ(x) + j=−1 ∞ φ(σ j+1 r(x)) − φ(σ j+1 x) + j=0 φ(σ j+1 x) − φ(σ j r(σx)) = φ(r(x)) + j=0 φ(σ j+1 r(x)) − φ(σ j r(σx)) . Hence ψ = φ − u + u ◦ σ is in F also. As j=0 bαj < ∞.
Set Bm = exp k=m+1 2bα and deﬁne + Λ = {f ∈ C (ΣA ) : f ≥ 0. ν(f ) = 1 . 1. If φ ∈ C (ΣA ) we get φ ∈ C (ΣA ) by φ({xi }∞ i=−∞ ) = φ({xi }∞ ). There are λ > 0. In our case G(ν) = ν gives L∗ ν = λν with λ > 0. We saw in Lemmas 1. p. 456): Let E be a nonempty compact convex subset of a locally convex topological vector space. m]} . σ is a ﬁnitetoone continuous + + map of ΣA onto itself. Clearly λ−1 Lf ≥ 0 and ν(λ−1 Lf ) = λ−1 L∗ ν(f ) = ν(f ) = 1 . 11].B. φ ∈ F ∩ C (ΣA ) and L = Lφ as above. We will prove 1. 1.6 that one only needs to get + Gibbs measures for φ ∈ C (ΣA ) ∩ F in order to get them for all φ ∈ F . A + + h ∈ C (ΣA ) with h > 0 and ν ∈ M (ΣA ) for which Lh = λh. one has that G(µ) = + + + (L∗ µ(1))−1 L∗ µ ∈ M (ΣA ) for µ ∈ M (ΣA ). Proof. Then one can think of φ as belonging to C (ΣA ) as follows: ∞ ∞ φ({xi }i=0 ) = φ({xi }i=−∞ ) where xi for i ≤ 0 are chosen in any way subject + to {xi }∞ i=−∞ ∈ ΣA . ν(h) = 1 and m→∞ lim + λ−m Lm g − ν(g)h = 0 for all g ∈ C (ΣA ) . Because L is a positive operator and L1 > 0. Proof. + It is the fact that σ is not onetoone on ΣA that will make this operator useful. For φ ∈ C (ΣA ) deﬁne the operator + L = Lφ on C (ΣA ) by (Lφ f )(x) = y ∈σ −1 x eφ(y) f (y) .7 via a sequence of lemmas. f (x) ≤ Bm f (x ). Ruelle’s PerronFrobenius Theorem 9 + + and deﬁne σ : ΣA → ΣA by σ(x)i = xi+1 .8. L∗ ν = λν. whenever xi = xi for all i ∈ [0. Then any continuous G : E → E has a ﬁxed point. There is a ν ∈ M (ΣA ) with G(ν) = ν by the SchauderTychonoﬀ Theorem (see Dunford and Schwartz. 1) be any ∞ k constants so that vark φ ≤ bαk for all k ≥ 0. Suppose φ ∈ C (ΣA ) satisﬁes φ(x) = φ(y) whenever xi = yi i=0 + for all i ≥ 0. One checks that λ−1 Lf ∈ Λ when f ∈ Λ.7. Ruelle’s PerronFrobenius Theorem [10.5 and 1. There is an h ∈ Λ with Lh = λh and h > 0. Linear Operators I. Let ΣA be topolog+ ically mixing. Lemma. m]. Assume xi = xi for i ∈ [0. A A In this section we will prove a theorem of Ruelle that will later be used + to construct and study Gibbs measures. The functions C (ΣA ) are thus identiﬁed with a certain subclass of C (ΣA ). Then Lf (x) = j eφ(jx) f (jx) . Let b > 0 and α ∈ (0.
since Bm → 1. Furthermore inf h = inf λ−M LM h ≥ K −1 . m+1 Bm+1 f (jx ) ≤ Bm eφ(jx ) f (jx ) Since AM > 0 there is a y ∈ σ −M x with y0 = z0 . There is an η ∈ (0. as jx and jx agree in places 0 to m + 1 eφ(jx) f (jx) ≤ eφ(jx ) ebα and so Lf (x) ≤ Bm Lf (x ) . As ν(f ) = 1. Thus Λ is equicontinuous and compact by the ArzelaAscoli Theorem. Assume xi = xi for all i ∈ [0. one has f (x) − f (x ) ≤ (Bm − 1)K → 0 as m → ∞.10 1 Gibbs Measures where the sum ranges over all j with Ajx0 = 1. m −1 Now h(x) ≥ Bm h(x ) because h ∈ Λ. m . z ∈ For f ∈ Λ + ΣA . or equivalently ( ) η(Bm h(x ) − h(x)) ≤ Bm λ−M LM f (x ) − λ−M LM f (x) .9. 1. We want to pick η so that g(x) ≤ Bm g(x ). Let K = λM eM φ B0 . Then 1 = ν(λ−M LM f ) ≥ K −1 f (z) gives f ≤ K as z is arbitrary. m] and f ∈ Λ. Consider any x. M −1 LM f (x) = y∈σ −M x exp k=0 −M φ φ(σ k y)f (y) ≥e f (y ) . f (z) ≥ 1 for some z and we get inf λ−M LM f ≥ K −1 . Let g = λ−M LM f − ηh where η is to be determined. m+1 m Lf1 (x ) ≤ Bm+1 ebα Lf1 (x ) for We saw above that Lf1 (x) ≤ Bm+1 ebα −M +1 M −1 L f one has any f1 ∈ Λ. Applying SchauderTychonoﬀ Theorem to λ−1 L : Λ → Λ gives us h ∈ Λ with Lh = λh. m]. Lemma. Applying this to f1 = λ λ−M LM f (x) ≤ Bm+1 ebα λ−M LM f (x ) . Provided η h ≤ K −1 we will have g ≥ 0. 1) so that for f ∈ Λ one has λ−M LM f = ηh + (1 − η)f with f ∈ Λ. If xi = xi for i ∈ [0. Proof. To get ( ) it is therefore enough to have m −1 η(Bm − Bm )h(x ) ≤ (Bm − Bm+1 ebα ) λ−M LM f (x ) or −1 η(Bm − Bm ) h ≤ (Bm − Bm+1 ebα )K −1 . For x the expression runs over the same j. Λ = ∅ as 1 ∈ Λ.
+ 1.11. Bm and Bm+1 ebα are in [−L. n ≥ 0. where A = (1 − η)−1 ( h + K) sup βM = 1 − η . f ≥ 0 −1 −r r and f F ≡ 0. Lemma. If F ∈ Λ. y ∈ [−L.B. 1. Let u1 . 1) so that λ−n Ln f − h ≤ Aβ n for all f ∈ Λ. There are constants A > 0 and β ∈ (0. As fq ≤ K one has λ−M q LM q f − h ≤ (1 − η)q ( h + K) and λ−n Ln f − h = λ−r Lr (λ−M q LM q f − h) ≤ A(1 − η)q+1 ≤ Aβ n . 0 ≤ r < M .9 and Lh = λh that. f ∈ Cr .10. Let n = M q + r. L] for all m. u2 be positive constants such that u1 (x − y) ≤ ex − ey ≤ u2 (x − y) for all x. Inductively one sees from Lemma 1. For ( ) to hold it is enough for η > 0 to satisfy η h u1 (log Bm + log Bm ) ≤ K −1 u2 (log Bm − log(Bm+1 ebα )) m or η h u1 or 4bαm+1 1−α ≤ K −1 u2 bαm η ≤ u2 (1 − α)(4αu1 h K)−1 . for f ∈ Λ. Let Cr = {f ∈ C (ΣA ) : varr f = 0}. x > y . λ−M q LM q f = (1 − (1 − η)q )h + (1 − η)q fq where fq ∈ Λ. Lemma. then ν(f F ) λ L (f F ) ∈ Λ. Proof. L]. Ruelle’s PerronFrobenius Theorem m 11 −1 There is an L so that the logarithms of Bm . 0≤r<M λ−r Lr and .
Then r−1 Lr (f F )(x) = j1 ···jr x exp k=0 φ(σ k (j1 · · · jr x)) f (j1 · · · jr x)F (j1 · · · jr x) + where j1 · · · jr runs over all rstrings of symbols for which j1 · · · jr x ∈ ΣA . Lemma. . r In the expression for L (f F )(x ) one has j1 · · · jr running over the same rstrings. One must still show ν(f F ) > 0. For f ∈ Cr . Now f (j1 · · · jr x) = f (j1 · · · jr x ) as f ∈ Cr . Proof. 1. the ﬁrst statement of the lemma with F = 1 gives λ−m Lm (fi ) − ν(g)h ≤ ε(1 + h ) for large m. Since λ−m Lm f1 ≤ λ−m Lm g ≤ λ−m Lm f2 .7 is ﬁnished. Assume xi = xi for i ∈ [0. for any z. f − ∈ Cr . Reasoning as in the proof of 1. As ν(fi ) − ν(g) < ε. f − ≥ 0 and f + . + For g ∈ C (ΣA ) one has limm→∞ λ−m Lm g − ν(g)h = 0. For f ± F ≡ 0. But (f F )(w) > 0 gives Lr (f F )(σ r w) > 0 and so ν(f F ) > 0.11 and 1.12. The proof of 1.8 (with Lr (f F ) in place of f ) we get λr ν(f F ) = ν(λ−M LM +r (f F )) ≥ K −1 Lr (f F )(z). Write f = f + − f − with f + . m]. and φ(σ k (j1 · · · jr x)) ≤ φ(σ k (j1 · · · jr x )) + varm+r−k φ. Hence Lr (f F )(x) ≤ Bm Lr (f F )(x ). F (j1 · · · jr x) ≤ Bm+r F (j1 · · · jr x ). Since ⎞ ⎛ r−1 Bm+r exp k=0 varm+r−k φ ≤ Bm+r exp ⎝ m+r bαj ⎠ ≤ Bm . this is obvious. Then λ−n−r Ln+r (f ± F ) − ν(f ± F )h ≤ Aν(f ± F )β n . λ−n−r Ln+r (f F ) − ν(f F )h ≤ Aν(f F )β n . j=m+1 each term in the above expression for Lr (f F )(x) is bounded by Bm times the corresponding term for Lr (f F )(x ).12 1 Gibbs Measures Proof. These inequalities add up to give us the ﬁrst statement of the lemma. Given g and ε > 0 one can ﬁnd r and f1 .10. for f ± F ≡ 0 we apply Lemmas 1. F ∈ Λ and n ≥ 0. f2 ∈ Cr so that f1 ≤ g ≤ f2 and 0 ≤ f2 − f1 ≤ ε. λ−m Lm g − ν(g)h ≤ ε(1 + h ) for large m.
Construction of Gibbs Measures + We continue to assume that φ ∈ F ∩ C (ΣA ) and ν. For f ∈ C (ΣA ) deﬁne f ∗ ∈ C (ΣA ) by f ∗ ({xi }∞ ) = min{f (y) : y ∈ ΣA . It is ˜ straightforward to check that µ ∈ C (ΣA )∗ . Note that µ(f ◦ σ) = lim µ((f ◦ σ n+1 )∗ ) = µ ˜ ˜ n→∞ . Hence µ(f ) = limn→∞ µ((f ◦ σ n )∗ ) exists by the Cauchy criterion. λ are as in Ruelle’s A + PerronFrobenius Theorem.13. Construction of Gibbs Measures 13 C. + Because µ is σinvariant on ΣA there is a natural way to make µ into a + measure on ΣA . Notice that ((Lf ) · g)(x) = y∈σ −1 x eφ(y) f (y)g(x) eφ(y) f (y)g(σy) y∈σ −1 x = = L(f · (g ◦ σ))(x) . Hence µ((f ◦ σ n )∗ ) − µ((f ◦ σ n+m )∗ ) = µ((f ◦ σ n )∗ ◦ σ m ) − µ((f ◦ σ n+m )∗ ) ≤ varn f which approaches 0. yi = xi for all i ≥ 0} . + Proof. since f is continuous. h. i=0 Notice that for m. We need to show that µ(f ) = µ(f ◦ σ) for f ∈ C (ΣA ). which we will ˜ denote by µ despite the possible ambiguity. Then µ = hν is a probability measure on ΣA . By the Riesz Representation ˜ Theorem we see that µ deﬁnes a probability measures on ΣA . Lemma. µ(f ) = ν(hf ) = f (x)h(x) dν(x). Using this µ(f ) = ν(hf ) = ν(λ−1 Lh · f ) = λ−1 ν(L(h · (f ◦ σ))) = λ−1 (L∗ ν)(h · (f ◦ σ)) = ν(h · (f ◦ σ)) = µ(f ◦ σ) . as n → ∞.C. n ≥ 0 one has (f ◦ σ n )∗ ◦ σ m − (f ◦ σ n+m )∗ ≤ varn f . µ is invariant under σ : ΣA → ΣA . + + 1.
14. for n ≥ s. In checking the mixing condition for E and F we may assume r = u = 0 because µ is σinvariant.12 gives. . F = {y ∈ ΣA : yi = bi . ˜ ˜ Recall that µ is ergodic if µ(E) = 0 or 1 whenever E is a Borel set with σ −1 E = E. µ is mixing for σ : ΣA → ΣA . for all Borel sets E and F .14 1 Gibbs Measures proving that µ is σinvariant. Because χE ∈ Cs Lemma 1. Proof. r ≤ i ≤ s} . Writing Sm φ(x) = + g ∈ C (ΣA ) one has m−1 k=0 φ(σ k x) one checks inductively that for f . (Lm f )(x) = y∈σ −m x eSm φ(y) f (y) . It is clear that mixing implies ergodicity and a standard argument shows that the mixing condition only need be checked for E and F in a basis for the topology. u ≤ i ≤ v} . Let E = {y ∈ ΣA : yi = ai . Then ((Lm f ) · g)(x) = y∈σ −m x eSm φ(y) f (y)g(σ m y) = Lm (f · (g ◦ σ m )) . Also µ(f ) = µ(f ) for f ∈ C (ΣA )∗ . We want to calculate µ(E ∩ σ −n F ) = µ(χE · χσ−n F ) = µ(χE · (χF ◦ σ n )) = ν(hχE · (χF ◦ σ n )) = λ−n L∗n ν(hχE · (χF ◦ σ n )) = ν(λ−n Ln (hχE · (χF ◦ σ n ))) = ν(λ−n Ln (hχE ) · χF ) . 1. Now µ(E ∩ σ −n F ) − µ(E)µ(F ) = µ(E ∩ σ −n F ) − ν(hχE )ν(hχF ) = ν (λ−n Ln (hχE ) − ν(hχE )h)χF  ≤ λ−n Ln (hχE ) − ν(hχE )h ν(F ) . One calls µ mixing if n→∞ lim µ(E ∩ σ −n F ) = µ(E)µ(F ). Proposition.
+ Thus take c2 = ea h . For any z ∈ ΣA there is −m at most one y ∈ σ z with y ∈ E. Hence m−1 Sm φ(x) − Sm φ(y) ≤ k=0 m−1 φ(σ k x) − φ(σ k y ) varm−1−k φ k=0 ≤ ≤ a. Construction of Gibbs Measures 15 λ−n Ln (hχE ) − ν(hχE )h ≤ Aµ(E)β n−s where β ∈ (0.16. m)}.C. φ(σ k y ) = φ(σ k y) for k ≥ 0. Thus µ(E ∩ σ −n F ) → µ(E)µ(F ).4. m).15. Theorem. Let a = i ∈ [0. y ∈ ΣA with xi = yi for Sm φ(x) − Sm φ(y) ≤ a . One then has µ(E ∩ σ −n F ) − µ(E)µ(F ) ≤ A µ(E)µ(F )β n−s for n ≥ s where A = A (inf h)−1 . ∞ k=0 vark φ < ∞. We now complete the proof of 1. Let E = {y ∈ ΣA : yi = xi for i ∈ [0. 1. then Proof. On the other hand. A + Proof. + Since φ ∈ C (ΣA ). 1). using 1. Thus. for any z ∈ ΣA there is at least −m−M z with y ∈ E. Lm (hχE )(z) = y∈σ −m z eSm φ(y) h(y)χE (y) ≤ eSm φ(x) ea h and so µ(E) = ν(hχE ) = λ−m ν(Lm (hχE )) ≤ λ−m eSm φ(x) ea h . µ is a Gibbs measure for φ ∈ F ∩ C (ΣA ). Then one y ∈ σ Lm+M (hχE )(z) ≥ eSm+M φ(y ) h(y ) ≥ e−M φ −a (inf h) eSm φ(x) . Deﬁne y by yi = yi for i ≥ 0 xi for i ≤ 0 . If x. + 1.15. Lemma.
1. C ∨ D = {Ci ∩ Dj : Ci ∈ C . Let Tm be a ﬁnite subset of ΣA so that ΣA = x∈Tm Em (x) disjointly. Taking limits this extends to µ (E) ≤ dµ(Em ) for all Borel 1 sets E. a. Suppose now that µ is any other measure satisfying the inequalities in Theorem 1.. For x ∈ ΣA let Em (x) = {y ∈ ΣA : yi = xi for all i ∈ [0. By the RadonNikodym Theorem µ = f µ for some µintegrable f . .16 1 Gibbs Measures and µ(E) = λ−m−M ν(Lm+M (hχE )) ≥ c1 λ−m eSm φ(x) where c1 = λ−M e−M φ −a . m)}. f ◦ σ = f . µ) (i. . B. . Then c1 e−P m x∈Tm eSm φ(x) ≤ x∈Tm µ (Em (x)) =1 ≤ c2 e−P From this one sees that P = limm→∞ 1 m m x∈Tm eSm φ(x) . P . If C = {C1 . one deﬁnes the entropy k Hµ (C) = i=1 (−µ(Ci ) log µ(Ci )) . hence P = P as they equal the same limit.e. As the RadonNikodym derivative is unique up to µequivalence.e. D.4 with constants c1 . log x∈Tm eSm φ(x) . . Because µ is ergodic this gives f equivalent to some constant c. in a way analogous to Lemma 1. Variational Principle We will describe Gibbs measures as those maximizing a certain quantity. 1 = µ (ΣA ) = c dµ = c and µ = µ . the Ci ’s are pairwise disjoint and X = i=1 Ci ). If D is another (ﬁnite) partition. One can apply the same reasoning to µ. In particular µ (E) = 0 when µ(E) = 0. We have veriﬁed the desired inequalities on measures of cylinder sets given in 1. The estimates on µ (Em (x)) and µ(Em (x)) give us µ (Em (x)) ≤ dµ(Em (x)) where d = c2 c−1 . Dj ∈ D}. .4 with P = log λ. Ck } is a partition of a measure k space (X. c2 . Applying σ one has µ = σ∗ µ = (f ◦ σ) σ ∗ µ = (f ◦ σ)µ .
m m Proof. Proof.0 x Applying this to ai = µ(Ci ) and xi = µ(Ci ) ϕ i µ(Ci ∩Dj ) µ(Ci ) we get = ϕ(µ(Dj )) .j (−µ(Ci ∩ Dj ) log µ(Ci ∩ Dj ))− i (−µ(Ci ) log µ(Ci )) = i. 1. Hµ (C ∨ D) ≤ Hµ (C) + Hµ (D). Inductively one sees that ϕ( n ai xi ) ≥ n ai ϕ(xi ) when i=1 i=1 n i=1 ai = 1 and ai ≥ 0. k → ∞ one gets lim sup j am aj ≤ · j m aj j Thus exists and a lim supj jj ≤ inf m am . write j = km + n with 0 ≤ n < m. m equals inf m am · m As lim inf j ≥ inf m am m . x2 ∈ [0. Then limm→∞ am exists and equals inf m am . Lemma. Suppose {am }∞ is a sequence satisfying inf am > −∞. 0 1.D. 1] as ϕ (x) < 0 for x ∈ (0.17.5 1.18. n. 1). Hµ (C ∨ D)−Hµ (C) = i. a1 + a2 = 1.j −µ(Ci ∩ Dj ) log µ(Ci ) j i µ(Ci ∩ Dj ) µ(Ci ) · = −µ(Ci ∩ Dj ) µ(Ci ∩ Dj ) log µ(Ci ) µ(Ci ) ϕ The function ϕ(x) = −x log x (ϕ(0) = 0) is concave on [0. For j > 0. 1]. one gets that limj aj j . Lemma.5 0 0. Then aj akm+n akm an kam an = ≤ + ≤ + · j km + n km km km km Letting j → ∞. m=1 m am+n ≤ am + an for all m.0 0. µ(Ci ∩ Dj ) µ(Ci ) ≤ϕ i µ(Ci ∩ Dj ) So Hµ (C ∨ D) − Hµ (C) ≤ j ϕ(µ(Dj )) = Hµ (D) . Fix m > 0. Variational Principle 17 1. From this it follows that ϕ(a1 x1 + a2 x2 ) ≥ a1 ϕ(x1 ) + a2 ϕ(x2 ) when x1 .
Let am = Hµ (D ∨ T −1 D ∨ · · · ∨ T −m+1 D). the terms in Zm+n (φ) are bounded by terms in Zm (φ)Zn (φ) and Zm (φ)Zn (φ) may have more terms. µ). D) = lim exists. Let µ ∈ Mσ (ΣA ) and U = {U1 . all positive. B. Notice that a0 a1 ···am+n−1 log Zm (φ) exists (called sup Sm+n φ ≤ a0 a1 ···am−1 sup Sm φ + am ···am+n−1 sup Sn φ .18 1 Gibbs Measures 1. Suppose φ ∈ C (ΣA ) and µ ∈ Mσ (ΣA ). . Proof. then hµ (T.18 to am = log Zm (φ) (notice am ≥ −m φ ). Apply Lemma 1. U) is called the entropy of µ. For φ ∈ C (ΣA ). Lemma. µ) and T an automorphism of (X. Then am+n ≤ Hµ (D∨T −1 D∨· · ·∨T −m+1 D)+Hµ (T −m D∨· · ·∨T −m−n+1 D) ≤ am +an since T −m D ∨ · · · ∨ T −m−n+1 D = T −m (D ∨ · · · ∨ T −n+1 D) 1 Hµ (D ∨ T −1 D ∨ · · · ∨ T −m+1 D) m→∞ m and µ is T invariant.21. Suppose now that φ ∈ C (ΣA ) and that a0 a1 · · · am−1 are integers between 1 and n satisfying Aak ak+1 = 1. . Then s(µ) = hµ (σ. P (φ) = limm→∞ the pressure of φ).20. Lemma. . Deﬁnition. . 1. Un } where Ui = {x ∈ ΣA : x0 = i}. xi = ai for all 0 ≤ i < m and Zm (φ) = a0 a1 ···am−1 exp a0 a1 ···am−1 1 m sup Sm φ . Proposition. From this one gets Zm+n (φ) ≤ Zm (φ)Zn (φ). . B. Proof. Then s(µ) + φ dµ ≤ P (φ) .19. 1. Write m−1 a0 a1 ···am−1 sup Sm φ = sup k=0 φ(σ k x) : x ∈ ΣA . If D is a (ﬁnite) partition of (X.
ΣA topologically mixing and µφ the Gibbs meaA sure of φ. pick x with xi = ai (i = 0. Variational Principle 19 Proof. as µ = µφ is the Gibbs measure. If yi = xi for all i = 0. µ{y ∈ ΣA : yi = ai ∀ 0 ≤ i < m} ∈ [c1 . m − 1) and Sm φ(x) = a0 a1 ···am−1 sup Sm φ . exp(−P m + Sm φ(x)) Summing the measure of these sets over all possible a0 · · · am−1 ’s gives 1. Proof. 2 1 exp(P m) 1 It follows that P (φ) = limm→∞ m log Zm (φ) = P .1 . . . Then µφ is the unique µ ∈ Mσ (ΣA ) for which s(µ) + φ dµ = P (φ) . then m−1 Sm φ(y) − Sm φ(x) ≤ k=0 φ(σ k y) − φ(σ k y) 2 2 ≤ var0 φ + var1 φ + · · · + var[ m ] φ + varm−[ m ] φ + · · · + var0 φ [m] 2 ≤ 2c k=0 αk ≤ 2c = d. so c1 exp(−P m)Zm (φ) ≤ 1 ≤ c2 exp(−P m)Zm (φ) or Zm (φ) ∈ [c−1 . .D. 1. Theorem. Thus Hµ (U ∨ · · · ∨ σ −m+1 U) + ≤ a0 ···am−1 Sm φ dµ a0 a1 ···am−1 µ(a0 · · · am−1 )(− log µ(a0 · · · am−1 )) + sup Sm φ ≤ log Zm (φ) by Lemma 1. Given a0 · · · am−1 . . .22. As φ ◦ σ k dµ = m−1 k k=0 φ(σ x). . . c2 ] . 1−α . Now. Now U ∨ · · · ∨ σ −m+1 U partitions points x ∈ ΣA according to x0 x1 · · · xm−1 . Now let m → ∞. Let φ ∈ F . Hence s(µ) + φ dµ ≤ lim m→∞ φdµ. 1 m 1 m Sm φdµ = φdµ where Sm φ(x) = Hµ (U ∨ · · · ∨ σ −m+1 U) + Sm φ dµ . m − 1. c−1 ] . .
Kn with Ki ⊂ Di and µ(Di \Ki ) < ε. . . . . 1. . and D = {D1 . Let X be a compact metric space. . . . Dn } a Borel partition of X. Then there m m are partitions {E1 . m 2. each Ei is a union of members of Cm . Let δ δ = inf i=j d(Ki . Suppose {Cm }∞ is a sequence of parm=1 titions so that diam(Cm ) = maxC∈Cm diam(C) → 0 as m → ∞. . . δ As diam(Cm ) ≤ 2 any C ∈ Cm can intersect at most one Ki . En so that m C ⊂ Ei if C ∩ Ki = ∅ .20 1 Gibbs Measures Hence. −µ(B) log µ(B) + B Sm φ dµ ≥ −µ(B) [ log µ(B) − Sm φ(x) + d ] ≥ −µ(B) [ log c2 e−P m+Sm φ(x) − Sm φ(x) + d ] ≥ µ(B)(P m − log c2 − d) . . m The reverse inequality was in Proposition 1.21. Divide the elements m m C ∈ Cm into groups whose unions are E1 . To prove uniqueness we will need a couple of lemmas. . m − 1}. . µ ∈ M (X). . Proof. limm→∞ µ(Ei ∆Di ) = 0 for each i. . Lemma. Then Ei ⊃ Ki and n m m m µ(Ei ∆Di ) = µ(Di \Ei )+µ(Ei \Di ) ≤ ε + µ X\ i=1 Ki ≤ (n+1) ε . . Kj ) and consider m with diam(Cm ) ≤ 2 . . So s(µφ ) + φ dµφ = P (φ) .23. then for x ∈ B. . . if B = {y ∈ ΣA : yi = ai for all i = 0. . En } so that m 1. Since Hµ (U ∨ · · · ∨ σ −m+1 U) + = B Sm φ dµ − µ(B) log µ(B) + B Sm φ dµ ≥ B µ(B)(P m − log c2 − d) = P m − log c2 − d. . Put a C hitting m m no Ki in any Ei you like. we get s(µ) + ≥ lim φ dµ = lim m→∞ 1 m Hµ (U ∨ · · · ∨ σ −m+1 U) + Sm φ dµ m→∞ 1 (P m − log c1 − d) = P = P (φ) . Pick compacts K1 .
D. Variational Principle
21
1.24. Lemma. Suppose 0 ≤ p1 , . . . , pm ≤ 1, s = p1 + · · · + pm ≤ 1 and a1 , . . . , am ∈ R. Then
m m
pi (ai − log ai ) ≤ s log
i=1 i=1
eai − log s
.
Proof. This generalizes 1.1. One shows by calculus that the left side is maxiai mized at pi = seeaj ·
j
Proof of Theorem 1.22 (continued). Let ν ∈ Mσ (ΣA ) satisfy s(ν)+ φ dν = P . First suppose ν is singular with respect to µ. Then there is a Borel set B with m m σ(B) = B, µ(B) = 0 and ν(B) = 1. Let Cm = σ −[ 2 ]+1 U∨· · ·∨U∨· · ·∨σ [ 2 ] U. Then diam(Cm ) → 0 (use dβ metric). Applying 1.23 to {B, X\B} one ﬁnds sets E m which are unions of elements of Cm and satisfy (µ + ν)(B∆E m ) → 0. m As µ + ν is σinvariant and σ −m+[ 2 ] B = B one has (µ + ν)(B∆F m ) → 0 m where F m = σ −m+[ 2 ] E m is a union of members of U ∨ · · · ∨ σ −m+1 U. Since 1 s(ν) = inf m Hν (U ∨ · · · ∨ σ −m+1 U) one has P = P (φ) = s(ν) + or mP ≤
B∈U∨···∨σ −m+1 U
φ dν ≤
1 m
Hν (U ∨ · · · ∨ σ −m+1 U) +
Sm φ dν
−ν(B) log ν(B) +
B
Sm φ dν .
Picking xB ∈ B one has Sm φ ≤ Sm φ(xB ) + d on B and so mP ≤ d +
B
ν(B)(Sm φ(xB ) − log ν(B)) ν(B)(Sm φ(xB ) − log ν(B))
B⊂F m
≤ d+ +
ν(B)(Sm φ(xB ) − log ν(B)).
B⊂X\F m
Applying 1.24 mP − d ≤ ν(F m ) log
B⊂F m
exp(Sm φ(xB )) exp(Sm φ(xB )) + 2K ∗ .
B⊂X\F m
+
ν(X\F m ) log
where K ∗ = sup0≤s≤1 (−s log s). Rearranging terms:
22
1 Gibbs Measures
−2K ∗ −d ≤ ν(F m ) log
B⊂F m
exp Sm φ(xB ) − mP exp Sm φ(xB ) − mP
+ ≤ ≤
ν(X\F m ) log
B⊂X\F m
ν(F m ) log
B⊂F m
c−1 µ(B) + ν(X\F m ) log 2
B⊂X\F m m m m
c−1 µ(B) 2
log c−1 2
+ ν(F ) log µ(F ) + ν(X\F ) log µ(X\F m ) .
Letting m → ∞, ν(F m ) → 1, µ(F m ) → 0 and the above inequality is contradictory. In general, for ν ∈ Mσ (ΣA ), write ν = βν + (1 − β)µ where β ∈ (0, 1), ν ∈ Mσ (ΣA ) is singular w.r.t. µ and µ ∈ Mσ (ΣA ) is absolutely continuous w.r.t. µ. As ν and µ are supported on disjoints sets Pν (φ) = βPν (φ) + (1 − β)Pµ (φ), where Pν (φ) = s(ν) + φ dν. Suppose Pν (φ) = P . Since Pν (φ) ≤ P and Pµ (φ) ≤ P (Prop. 1.21), we have Pν (φ) = P or β = 0. We just saw that Pν (φ) = P . Thus ν = µ and write ν = dν µ. Then the dν is σinvariant up dµ dµ to equivalence as ν , µ are both invariant and the RadonNikodym derivative is unique up to equivalence. So dν is constant and ν = µ. dµ
E. Further Properties
In this section we look at more examples of the good behavior of Gibbs measures. Throughout we assume µ = µφ with φ ∈ F and σΣA topologically A mixing. Two partitions P and Q are called εindependent if µ(P ∩ Q) − µ(P )µ(Q) < ε .
P ∈P,Q∈Q
Let U = {U1 , . . . , Un } be the partition of ΣA with Uj = {x ∈ ΣA : x0 = j} . The partition U is called weakBernoulli (for σ and µ) if for every ε > 0 there is an N (ε) so that P = U ∨ σ −1 U ∨ · · · ∨ σ −s U and Q = σ −t U ∨ · · · ∨ σ −t−r U
are εindependent for all s ≥ 0, r ≥ 0, t ≥ s + N (ε). A wellknown theorem of Friedman and Ornstein [4] states that if U is weakBernoulli, then (σ, µ) is conjugate to a Bernoulli shift.
E. Further Properties
23
1.25. Theorem. U is weakBernoulli for the Gibbs measure µ = µφ .
+ Proof. We may assume that φ ∈ C (ΣA ) as before. For P ∈ P we have χP ∈ Cr . As in the proof of 1.14, for Q ∈ Q
µ(P ∩ Q) − µ(P )µ(Q) ≤ A µ(P )µ(Q)β t−s where β ∈ (0, 1). Summing over P, Q µ(P ∩ Q) − µ(P )µ(Q) ≤ A β t−s ≤ ε
P,Q
when t − s is large. Because µ is mixing, µ(f · (g ◦ σ n )) → µ(f )µ(g) as n → ∞ for f, g continuous (in fact L2 ) functions. The above proof used that this convergence was exponentially fast for characteristic functions of cylinder sets. This exponential convergence will now be carried over to functions in F . For α ∈ (0, 1) A let Hα be the family of f ∈ C (ΣA ) with vark f ≤ cαk for some c. Hα is a Banach space under the norm f
α
= f + sup(α−k vark f ) .
k≥0
1.26. Exponential Cluster Property. For ﬁxed α ∈ (0, 1) there are constants D and γ ∈ (0, 1) so that µ(f · (g ◦ σ n )) − µ(f )µ(g) ≤ D f for all f, g ∈ Hα , n ≥ 0. Proof. For k ≥ 0 and x ∈ ΣA , let Ek (x) = {y ∈ Σa : yi = xi for all i ≤ k} . Deﬁne fk (x) = µ(Ek (x))−1 Ek (x) f dµ. Then µ(fk ) = µ(f ) and f − fk ≤ f α αk . Hence µ(f · (g ◦ σ n )) − µ(f )µ(g) ≤ µ(fk · (gk ◦ σ n )) − µ(fk )µ(gk ) + µ((f − fk ) · (g ◦ σ n )) + µ(fk · ((g − gk ) ◦ σ n )) ≤ µ(fk · (gk ◦ σ n )) − µ(fk )µ(gk ) + 2αk f
α α
g
αγ
n
g
α.
Now fk is measurable with respect to the partition P = {Ek (x)}x ; i.e., fk = P ∈P aP χP . Also gk = P ∈P bP χP . Hence µ(fk · (gk ◦ σ n )) − µ(fk )µ(gk ) ≤
P,Q∈P
aP bP  µ(P ∩ σ −n Q) − µ(P )µ(Q) g A β n−2k n−2k . α g αA β
≤ f ≤ f
Letting k = [n/3] we get the result with γ = max(α1/3 , β 1/3 ).
24
1 Gibbs Measures
1.27. Central Limit Theorem. For ψ ∈ F there is a ξ = ξ(ψ) ∈ [0, ∞) so A that 1 µ x ∈ ΣA : √ (Sn ψ(x) − nµ(ψ)) < r n
n→∞
1 −→ √ ξ 2π
r −∞
e−x
2
/2ξ 2
dx .
For ξ = 0, convergence is asserted for r = 0 and the expression on the right is taken to be 0 for r < 0 and 1 for r > 0. Remark. We omit the proof, referring the reader to M. Ratner [13]. It is interesting to know when ξ(ψ) = 0. This happens (see [13]) precisely when ψ − µ(ψ) = u ◦ σ − u has a solution u ∈ L2 (µ). It is interesting that in case such u can be found one can ﬁnd u ∈ F and so ψ is homologous to a constant. The reasoning for A this is very roundabout and it would be good to ﬁnd a nice direct proof. 1.28. Theorem. Let ΣA be topologically mixing and φ, ψ ∈ F . The following A are equivalent: (i) µφ = µψ . (ii) There is a constant K so that Sm φ(x)−Sm ψ(x) = mK whenever σ m x = x. (iii)There are a constant K and a u ∈ F so that A φ(x) = ψ(x) + K + u(σx) − u(x) for all x ∈ ΣA . (iv)There are constants K and L so that Sm φ(x) − Sm ψ(x) − mK ≤ L for all x and all m > 0. If these conditions hold, then K = P (φ) − P (ψ). Proof. (iii) ⇒ (iv) is obvious and (iv) ⇒ (i) is just like Lemma 1.5. Assume µφ = µψ and σ m x = x. From the deﬁnition of a Gibbs measure and µφ = µψ one sees that exp(−P (φ)j + Sj φ(x)) ∈ [d1 , d2 ], exp(−P (ψ)j + Sj ψ(x)) where d1 > 0, d2 > 0 are independent of x and j. This is equivalent to Sj φ(x) − Sj ψ(x) − j(P (φ) − P (ψ)) ≤ M for some M independent of j, x. If σ m x = x, letting j = km, Sj φ(x) = kSm φ(x) and M > k Sm φ(x) − Sm ψ(x) − m(P (φ) − P (ψ)) . Letting k → ∞ we get (ii) with K = P (φ) − P (ψ). Now assume (ii). In proving (iii) we will need the following standard lemma.
x agree in places k − r to m + r. Deﬁne w ∈ ΣA by wi = xt for i ≡ t (mod m − k) . let U1 . Let Γ = {σ k x : k ≥ 0} and deﬁne u : Γ → R by A k−1 u(σ k x) = j=0 η(σ j x) . We will estimate varr (uΓ ). T = σ (topological mixing is stronger than transitivity). U2 . . Thus σ k x = σ m x for m = k and u is well deﬁned on Γ . is dense in X. By transitivity. As Γ is dense in ΣA .N .E. .N Continuing the proof of (iii) from (ii). By Baire Category Theorem there is an x ∈ i. be a basis for the topology. Lemma.N = n≥N T −n Ui Vi. As X is 2nd countable. vars η ≤ cαs for some α ∈ (0. Now m−1 u(w) − u(y) = j=k η(σ j x) . j=k m−1 u(z) − u(y) ≤ j=k ∞ η(σ j x) − η(σ j w) ≤ varr η + varr+1 η + · · · + varr+1 η + varr η ≤2 s=r vars η. let x be as in the lemma for X = ΣA . Then σ m−k w = w and w.29. σ j w agree in places k − r − j through m + r − j. Suppose y = σ k x. Proof. . Γ must be inﬁnite (except in the trivial case of ΣA =one point) and x is not periodic. N > 0. Let η = φ − ψ − K ∈ F . then T n x ∈ U for some n ≥ N . and z = σ m x (m > k) agree in places −r to r. k ≤ t ≤ m . Further Properties 25 1. hence σ j x. 1) and A . then there is a point x ∈ X so that if U = ∅ is open. Since (ii) gives m−1 η(σ j w) = 0 . Then xk+s = xm+s for all s ≤ r. the open set Vi. If T : X → X is a topologically transitive continuous map of a compact metric space. Since η ∈ F .
Math.25) was veriﬁed for the Σn case ﬁrst by Gallavotti [5]. 6].22) of the Gibbs measure µφ is due to LanfordRuelle [7] in the Σn case. R. S. 1−α So u is uniformly continuous on Γ and therefore extends uniquely to a continuous u : ΣA = Γ → R. B. Lecture Notes in Physics 20. 6. 7.I. 32 (1973). D. 3. no. 10. R. 4. L. Phys. N. Lanford: Entropy and equilibrium states in classical statistical mechanics. 1970. Dobrushin: The problem of uniqueness of a Gibbsian random ﬁeld and the problem of phase transitions. 1. Func. In those papers one deal with Σn instead of slightly more general ΣA . 1–113.28 is taken from Livˇic [8] and Sinai s [12]. References For discussions of Gibbs’ states and statistical mechanics we refer to Ruelle’s book [9] and Lanford [6]. Ruelle gave two proofs of his PerronFrobenius Theorem. . We have followed parts of each. E. Proofs of Theorem 1. American Mathematical Society. Math. 2 (1968) no. 5 (1970) 365–394.4 for ΣA appear in [2. 11]. Gallavotti: Ising model and Bernoulli schemes in one dimension. for diﬀerent circumstances [10. In “statistical mechanics and mathematical problems” (ed. A. 44–57. 183–190. Comm. We have followed the proof in [2].26 1 Gibbs Measures ∞ varr (uΓ ) ≤ 2c s=r αs = 2c αr . R. 1973. Providence. G.26 is from [11]. The weak Bernoulli condition (Theorem 1. L. R. 3. The variational characterization (Theorem 1. 13 (1969) 194–215. Theorem 1. Friedman. Bowen: Some systems with unique equilibrium states. Systems Theory 8 (1974/75). Springer. For z ∈ Γ . Adler. 193–202. D. u ∈ F . Ratner. It was extended to ΣA independently by each of Bunimovich. pp. A. Ruelle: Observables at inﬁnity and states with short range correlations in statistical mechanics.22 with φ = 0). O. Memoirs of the American Mathematical Society 98. O. Theorem 1. Anal. and its Appl. Phys. 11. The deﬁnition of Gibbs state in statistical mechanics does not coincide with what we gave in Section 1. A u(σz) − u(z) = η(z) and this equation extends to ΣA by continuity. E. Berlin. There Gibbs states are deﬁned for more general φ and then our theorem corresponds to one in statistical mechanics about uniqueness of Gibbs states [3. Lenard). Ruelle and this author. 2. Weiss: Similarity of automorphisms of the torus. Because varr u = varr (uΓ ). 4. Ornstein: On isomorphism of weak Bernoulli transformations. which in turn was based on Adler and Weiss’ proof [1] of a theorem of Parry (Theorem 1. Advances in Math. 12]. Comm. For ΣA one can ﬁnd it in [2] or [11]. Math. Lanford. 5.
River Edge. Notes Acad. NJ. 619–654. Surveys 27 (1972). no. New YorkAmsterdam 1969. D. Phys. Sci. Sinai: Gibbs measures in ergodic theory. English translation: Russian Math. no.. Imperial College Press. W. no. Livˇic: Certain properties of the homology of Ysystems. 11. 555–564. Amer. 4. (1 ) 10.. 13. USSR 10 (1971).References 27 8.. Nauk 27 (1972). Uspehi Mat. 16 (1973). 758–763. A. . 12. 9 (1968) 267–278. 98 (1976). 1 Reprint: World Scientiﬁc Publishing Co. Inc. D. Comm. London. 3. 181–197. Zametki 10 s (1971). Mat. Ya. Math. Inc. 21–64. English translation: Math. Math. Ruelle: Statistical mechanics of a onedimensional lattice gas. Benjamin. M. 4. D. N. A. 21–69. 1999 (note of the editor). Ruelle: Statistical mechanics: Rigorous results. Israel J. J. Ruelle: A measure associated with axiomA attractors. 9. Math. Ratner: The central limit theorem for geodesic ﬂows on ndimensional manifolds of negative curvature.
Proof. T by hµ (T ) = sup hµ (T. (a) Hµ (CD) ≤ Hµ (CE) if D ⊃ E. (d) Hµ (C) ≤ Hµ (D) + Hµ (CD). (c) Hµ (C ∨ DE) ≤ Hµ (CE) + Hµ (DE). one can rewrite this as Hµ (CD) = j E∈E j i µ(Cj ∩ Di ) µ(Di ) µ(Di ) ϕ µ(Cj ∩Di ) µ(Di ) . We write C ⊂ D if each set in C is a union of sets in D. 2. Lemma.2 General Thermodynamic Formalism A. We deﬁne Hµ (CD) = Hµ (C ∨ D) − Hµ (D) µ(Cj ∩ Di ) log µ(Di ) =− µ(Di ) i j ≥ 0.r. D where D ranges over all ﬁnite partitions. We now deﬁne the entropy of µ w. D). D) when T is an endomorphism of a probability space and D a ﬁnite measurable partition. Hµ (CD) = Since E ⊂ D. Lemma 1. (b) Hµ (CD) = 0 if D ⊃ C.17 says that Hµ (CD) ≤ Hµ (C). Letting ϕ(x) = −x log x. Entropy In Section D of Chapter 1.t.1. we deﬁned the number hµ (T. We will now turn to some computational lemmas. µ(E) Di ⊂E µ(Di ) ϕ µ(E) µ(Cj ∩ Di ) µ(Di ) · .
2. As µ is T invariant. (b) hµ (T. Let T be an endomorphism of a probability space (X. Lemma. To see (b) one notes that C ∨ D = D when D ⊃ C. B.30 2 General Thermodynamic Formalism By the concavity of ϕ (see the proof of Lemma 1. . For (c) one writes Hµ (C ∨ DE) = Hµ (C ∨ D ∨ E) − Hµ (D ∨ E) + Hµ (D ∨ E) − Hµ (E) = Hµ (CD ∨ E) + Hµ (DE) ≤ Hµ (CE) + Hµ (DE) by (a).1 Hµ (C ∨ · · · ∨ T −m+1 C) ≤ Hµ (D ∨ · · · ∨ T −m+1 D) + Hµ (C ∨ · · · ∨ T −m+1 CD ∨ · · · ∨ T −m+1 D) ≤ Hµ (D ∨ · · · ∨ T −m+1 D) m−1 + k=0 Hµ (T −k CD ∨ · · · ∨ T −m+1 D) m−1 ≤ Hµ (D ∨ · · · ∨ T −m+1 D) + k=0 Hµ (T −k CT −k D) = Hµ (D ∨ · · · ∨ T −m+1 D) + mHµ (CD) .2. Finally Hµ (C) = Hµ (C ∨ D) − Hµ (DC) ≤ Hµ (C ∨ D) = Hµ (D) + Hµ (CD) . xi = · µ(E) µ(Di ) Hence Hµ (CD) ≤ j E∈E ai xi ) ≥ µ(E) ϕ µ(Cj ∩ E) µ(E) = Hµ (CE) . Hµ (T −k CT −k D) = Hµ (T −k C ∨ T −k D) − Hµ (T −k D) = Hµ (C ∨ D) − Hµ (D) = Hµ (CD) . C). C and D ﬁnite partitions.17) one has ϕ( ai ϕ(xi ) where µ(Di ) µ(Cj ∩ Di ) ai = . (c) hµ (T. D) + Hµ (CD). C) ≤ hµ (T. µ). Using Lemma 2. Proof. Then (a) Hµ (T −k CT −k D) = Hµ (CD) for k ≥ 0. Dividing by m and letting m → ∞. C ∨ · · · ∨ T −n C) = hµ (T.
4. T k y) ≤ ε for all k ∈ Z ⇒ x = y . Then Hµ (CD) ≤ Hµ (CE) < ε by 2. By Lemmas 2. Proposition. C) . The expression Hµ (CE) = i. 2. En } ⊂ D with µ(Ei ∆Ci ) < α . Lemma. ε > 0 and C a ﬁnite Borel partition. Let X be a compact metric space.3 hµ (T. for any α > 0. Dn ) . D) = hµ (T. Consider any partition C. . In Lemma 1. Proof. µ ∈ M (X). Dn ).j µ(Ej ) ϕ µ(Cj ∩ Ei ) µ(Ei ) depends continuously upon the numbers µ(Cj ∩ Ei ) and µ(Ei ) = j µ(Cj ∩ Ei ) and vanishes when µ(Cj ∩ Ei ) = δij µ(Ei ). C) ≤ lim inf hµ (T. Hµ (CE) < ε. . Of course hµ (T ) ≥ lim supn hµ (T.1 (a).3. D) + Hµ (CD) . for α small. Cn }.23 we showed that. . Entropy 31 hµ (T. Let C = {C1 . There is a δ > 0 so that Hµ (CD) < ε whenever D is a partition with diam(D) < δ.2 (b) and 2. Dn ) . Then 1 1 Hµ (D ∨ · · · ∨ T −m+1 D) = Hµ (C ∨ · · · ∨ T −m−n+1 C) . n Varying C. C) ≤ hµ (T. one could ﬁnd δ > 0 such that whenever D satisﬁes diam(D) < δ there is a E = {E1 . 2. . Then hµ (T ) = lim hµ (T. Dn ). . m m Letting m → ∞. Hence. . Set D = C ∨ · · · ∨ T −n C. n→∞ Proof. . (as m m+n → 1) we get hµ (T. .A. hµ (T ) ≤ lim inf n hµ (T. A homeomorphism T : X → X is called expansive if there exists ε > 0 so that d(T k x. Suppose T : X → X is a continuous map of a compact metric space. . µ ∈ MT (X) and that Dn is a sequence of partitions with diam(Dn ) → 0.
. .6. E) by 2. Let C be a partition and E = C ∨ · · · ∨ T −n+1 C. Suppose T : X → X is a homeomorphism with expansive constant ε. Now hµ (σ. We will deﬁne the pressure P (φ) of φ ∈ C (X) in a way which generalizes Section D in Chapter 1. Dn ) = hµ (T.5. Dn ). m→∞ Varying C. nhµ (T ) ≤ hµ (T n ). C) = lim n Hµ (C ∨ · · · ∨ T −nm+1 C) nm 1 Hµ (E ∨ T −n E ∨ · · · ∨ T (−m+1)n E) = lim m→∞ m = hµ (T n .1 (b). hµ (T n ) = nhµ (T ) for n > 0. X(U) = {x ∈ X : T k x ∈ Uik for k = 0.2 (b) and 2. Then hµ (T ) = hµ (T. Then nhµ (T. m − 1} . U) is what we denoted by s(µ) in Chapter 1. . . B. One writes m = m(U). U) for µ ∈ Mσ (ΣA ). . E) ≤ hµ (T n ) = nhµ (T ) . Proposition. Then σ is expansive and 2. D) by Lemma 2. On the other hand hµ (T n . That s(µ) = hµ (σ) implies that the number s(µ) does not depend on the homeomorphism σ and partition U. . 2.32 2 General Thermodynamic Formalism 2. C) ≤ hµ (T n . Then diam(Dn ) → 0 using expansiveness. Proof. Consider the case of σ : ΣA → ΣA and standard partition U = {U1 . Un } where Ui = {x ∈ ΣA : x0 = i}. . Pressure Throughout this section T : X → X will be a ﬁxed continuous map on the compact metric space X. C) ≤ n sup hµ (T. D) whenever µ ∈ MT (X). Hence hµ (T ) = limn hµ (T. µ) (because of the deﬁnition of hµ (σ)). C C B. Wm (U) the set of all mstrings U = Ui0 Ui1 · · · Uim−1 of members of U. But hµ (T. and diam(D) ≤ ε. Proof. C) = nhµ (T ) . Hence hµ (T n ) = sup hµ (T n . but only on σ as an automorphism of the probability space (ΣA . Let U be a ﬁnite open cover of X.2 (c). Let Dn = T n D ∨ · · · ∨ D ∨ · · · ∨ T −n D.5 gives that hµ (σ) = hµ (σ. Lemma. .
Let γ = γ(φ. In case X(U) = ∅. U) ≥ e−m Lemma 1. Suppose V is an open cover reﬁning U.18. If Γm ⊂ Wm (V) covers X. The limit P (φ) = exists (but may be +∞).8. Hence am = log Zm (φ. U) = sup{φ(x) − φ(y) : x. One sees that Sm+n φ(UV) ≤ Sm φ(U) + Sn φ(V) and so exp(Sm+n φ(UV)) ≤ UV∈Γm Γn U∈Γm m→∞ 1 log Zm (φ. Then Sm φ(U (V)) ≤ Sm φ(V) + mγ and so Zm (φ. U) ≤ Zm (φ. We say that Γ ⊂ Wm (U) covers X if X = U∈Γ X(U). i. Proposition.B. If Γm ⊂ Wm (U) and Γn ⊂ Wn (U) each cover X. V). 2. every V ∈ V lies in some U (V ) ∈ U. For V ∈ Wm (V) let U (V) = U (Vi0 ) · · · U (Vim−1 ). U) = inf Γ U∈Γ exp(Sm φ(U)). then Γm Γn = {UV : U ∈ Γm . Thus Zm+n (φ. U) satisﬁes the hypotheses of 2. we let Sm φ(U) = −∞. which gives diam(U)→0 lim P (φ. y ∈ U for some U ∈ U}. Lemma. U) ≤ emγ Zm (φ. V ∈ Γn } ⊂ Wm+n (U) covers X. then U (Γm ) = {U (V) : V ∈ Γm } ⊂ Wm (U) covers X.7. Proof. where Γ runs over all subsets of Wm (U) covering X.. U) . Finally one deﬁnes Zm (φ. Pressure m−1 33 Sm φ(U) = sup k=0 φ(T k x) : x ∈ X(U) . U) and Zm (φ. The limit P (φ. U) = lim exists and is ﬁnite.e. U) Zn (φ. Proof. U) m exp(Sm φ(U)) V∈Γn exp(Sn φ(V)). φ .
D) + φ dµ ≤ PT (φ) + log M . diam(V) → 0 and so nPT (φ) = PT n (Sn φ). Lemma. Let T : X → X be a continuous map on a compact metric space and φ ∈ C (X). Let Sn φ(x) = n−1 k=0 φ(T k x). diam(V)→0 Letting diam(U) → 0. In cases where confusion may arise we write the topological pressure P (φ) as PT (φ).9. γ(φ. U) → 0 and lim sup P (φ. Then hµ (T ) + for any µ ∈ MT (X). for U = Ui0 Ui1 · · · Uimn−1 let V = Vi0 · · · Vim−1 where Vik = Uikn ∩ T −1 Uikn+1 ∩ · · · ∩ T −n+1 Uikn+n−1 .10. Let V = U ∨ · · · ∨ T −n+1 U. We will ﬁrst need a couple of lemmas. V) . U) − γ(φ. 2. Then Wm (V) and Wmn (U) are in onetoone correspondence. 2. Then hµ (T. Suppose D is a Borel partition of X such that each x ∈ X is in the closures of at most M members of D. U) ≤ lim inf P (φ. V) n and nPT (φ. U) ≤ P (φ. U) = PT n (Sn φ. One sees that X(U) = X(V) T Tn and Smn φ(U) = Sm (Sn φ)(V). Then PT n (Sn φ) = nPT (φ) for n > 0 . Theorem.11. diam(U)→0 diam(V)→0 We are done. Now for any U. U) = Zm (Sn φ. As diam(U) → 0. φ dµ ≤ PT (φ). V) . Proof.34 2 General Thermodynamic Formalism P (φ. 2. all V with small diameter reﬁne U and so P (φ. . V) . U) ≤ lim inf P (φ. Thus one gets T T Zmn (φ. We now come to our ﬁrst interesting result about the pressure P (φ). V) + γ . Lemma.
Let A = {A1 . For each B ∈ Dm = D ∨ · · · ∨ T −m+1 D pick xB ∈ B with B Sm φ dµ ≤ µ(B) Sm φ(xB ).12. . For each xB pick UB ∈ Γm with xB ∈ X(UB ). . This map B → UB is at most M m to one. 2. there is a Borel partition D∗ of (sm−1 )n so that n (a’) each member of D∗ lies in some Ui1 × · · · × Uin . U) . . Lemma. . . n Then Dn = L−1 D∗ works where n L = (G. For each n > 0 there is a Borel partition Dn of X so that (a) D lies inside some member of T −k A for each D ∈ Dn and k = 0. .1. n−1. gm be a partition of unity subordinate to A. . D) + φ dµ ≤ 1 log m M m exp(Sm φ(U)) U∈Γm ≤ log M + 1 log Zm (φ. Now hµ (T. Then G = (g1 . Pressure 35 Proof. Fix n > 0. Since (sm−1 )n is nm − n dimensional. . . . . .12. we obtain the desired inequality. . Let C be a Borel partition and ε > 0. . Then (see the proof of 2. m Letting m → ∞ and then diam(U) → 0. (b) at most nA sets in Dn can have a point in all their closures. G ◦ T n−1 ) : X → (sm−1 )n . Let A be a ﬁnite open cover of X. G ◦ T. Proof.3 ﬁnd an open cover A so that Hµ (CD) < ε whenever D is a partition every member of which is contained in some member of A. let E = C ∨ · · · ∨ T −n+1 C and Dn as in Lemma 2. . . Now U = {U1 . Am } and g1 . . .10.B.6) . Let U be a ﬁnite open cover of X each member of which intersects at most M members of D. and n (b’) at most nm members of D∗ can have a common point in all their closures. . gm ) : X → sm−1 ⊂ Rm where sm−1 is an m − 1 dimensional simplex. . As Sm φ(xB ) ≤ Sm φ(UB ). . . Proof of 2. . By Lemma 2. . D) + φ dµ ≤ 1 m 1 ≤ m ≤ Hµ (Dm ) + Sm φ dµ µ(B)(− log µ(B) + Sm φ(xB )) B 1 log m exp(Sm φ(xB )) B by Lemma 1. Let Γm ⊂ Wm (U) cover X. Um } is an open cover where Ui = {x ∈ sm−1 : xi > 0} and G−1 Ui ⊂ Ai . one has hµ (T.
C. U) → 0 as diam(U) → 0. But γ(φ ◦ π. Then PT2 (φ) ≤ PT1 (φ ◦ π) for φ ∈ C (X2 ). π −1 U) ≤ PT1 (φ ◦ π) + γ(φ ◦ π. Proposition. We say that Γ ⊂ W ∗ (U) = m>0 Wm (U) covers K ⊂ X if K ⊂ U∈Γ X(U). T2 : X2 → X2 be continuous maps on compact metric spaces.8 PT1 (φ ◦ π. λ) = U∈Γ λm(U) exp(Sm(U) φ(U)). . C) + φ dµ ≤ 1 hµ (T n . For λ > 0 and Γ ⊂ W ∗ (U) deﬁne Z(Γ. Proof.13. Let T1 : X1 → X1 . π −1 U) . n Now let n → ∞ and then ε → 0. Since Dn reﬁnes T −k A for each k. 2. 1 hµ (T. π : X1 → X2 continuous and onto satisfying π ◦ T1 = T2 ◦ π.11. T −k A bears the same relation to T −k C as A to C). Hence. U) = PT1 (φ ◦ π.36 2 General Thermodynamic Formalism hµ (T. π −1 U) . π −1 U) = γ(φ. Variational Principle Let U be a ﬁnite open cover of X. using 2. For U an open cover of X2 one sees that PT2 (φ. As in the proof of 2.9. letting diam(U) → 0 we get PT2 (φ) ≤ PT1 (φ ◦ π). Hence. Now n−1 Hµ (EDn ) ≤ k=0 Hµ (T −k CDn ) . Dn ) + Sn φ dµ + Hµ (EDn ) n n 1 1 ≤ (PT n (Sn φ) + log(nA)) + Hµ (EDn ) n n by Lemma 2. E) + Sn φ dµ n 1 1 ≤ hµ (T n . one has Hµ (T −k CDn ) < ε (since µ is T invariant. C) + φ dµ ≤ PT (φ) + log(nA) + ε .
one has Z(Γ ∗ .n = δT x. (b) SN φ(U) ≤ N ( φdµ + γ(U) + ε). λ). Suppose that Z(Γ.n (f ) = n−1 f (T n x)−f (x) ≤ 2n−1 f . Let U∗ be the ﬁrst N symbols of U. . or λN ZN (φ. . λ) = ∞ M Z(Γ n . λN ΓN exp SN φ(U∗ ) ≤ max 1. Let δx be the unitmeasure concentrated on the point x. n=1 Fix N and consider x ∈ X. Letting Γ ∗ = ∞ n n=1 Γ . Since Γ covers X. Let E be a ﬁnite set. Then Z(Γ n .15. There are m and arbitrarily large N for which one can ﬁnd U ∈ WN (U) satisfying the following (a) x ∈ X(U).nk → µ for some nk → ∞} . Lemma. . µ ∈ V (x). Then SN φ(U∗ ) ≤ Sm(U) φ(U) + M φ .n (f )−δx. and (b) N ≤ m(U) < N + M . λ) < ∞ . As X is compact we may take Γ ﬁnite and Γ ⊂ m=1 Wm (U).n = n−1 (δx + δT x + · · · + δT n−1 x ) and V (x) = {µ ∈ M (X) : δx. ak−1 ) ∈ E k . λ)n where Γ n = {U1 U2 · · · Un : Ui ∈ Γ }. Let x ∈ X. λ) < 1 for some Γ covering X. λ−M eM φ Z(Γ ∗ . one can ﬁnd U = U1 U2 · · · Un ∈ Γ ∗ with (a) x ∈ X(U). 2. It follows that λ ≤ e−P . For Γ N the set of U∗ so obtained. a = (a0 . λ) ≤ Z(Γ.14. Then λ ≤ e−P . Variational Principle 37 2. This shows V (x) ⊂ MT (X). Deﬁne δx. Now T ∗ δx.C. U) and λ > 0. Let P = P (φ. U a ﬁnite open cover of X and ε > 0. Lemma. T ∗ δx. One deﬁnes the distribution µa on E by µa (e) = k −1 (number of j with aj = e) and H(a) = − e∈E µa (e) log µa (e) . Proof. .n and for f ∈ C (X). . V (x) = ∅ as M (X) is a compact metric space. U) ≤ constant.
. . Now enlarge each Vi to a Borel set Vi∗ still contained in a member of U ∨ · · · ∨ T −m+1 U and so that {V1∗ .nj → µ. .m−1 ) . . C) + ≤ hµ (T ) + · m 2 2 Let δx. . Furthermore we may assume Vi ∩ Vj = ∅ for i = j. . Let D1 . z ∈ Ui for some i} . this formula shows that µ would still be the limit. Now ﬁx nj = mkj . For β small enough and j large enough one has 1 m − i pi log pi ≤ 1 m − i µ(Di ) log µ(Di ) + ε 2 ≤ hµ (T ) + ε .n = n n −n δx. Each Di is contained in some member of U ∨ · · · ∨ T −m+1 U and one can ﬁnd an open set Vi ⊃ Ki for which this is still true. . one has lim inf pi ≥ µ(Ki ) ≥ µ(Di ) − β.n + δT n x. n n If we replaced nk by the nearest multiple of m. . nj ) with T s x ∈ Vi∗ and Mi.r /kj and pi = Mi /nj = 1 m (pi. . Deﬁne pi. when viewed as a = a0 . . Vt∗ } is a Borel partition of X. . .r = Mi.38 2 General Thermodynamic Formalism (c) U contains a subword of length km ≥ N − m which. . Recall that γ(U) = sup{φ(y) − φ(z) : y. .n −n .17) m−1 ϕ(pi ) ≥ r=0 1 ϕ(pi. Pick a Borel partition C = {C1 . . There is an m so that 1 ε ε Hµ (C ∨ · · · ∨ T −m+1 C) ≤ hµ (T.r ) m . j→∞ and lim supj→∞ pi ≤ µ(Ki ) + tβ ≤ µ(Di ) + tβ.nj → µ. By the concavity of ϕ(x) = −x log x (see 1. . ak−1 ∈ (Um )k satisﬁes 1 H(a) ≤ hµ (T ) + ε. Thus we assume nj = mkj . . Let Mi be the number of s ∈ [0. m Proof. .0 + · · · + pi. Dt be the nonempty members of C∨· · ·∨T −m+1 C and for each Di ﬁnd a compact Ki ⊂ Di with µ(Di \Ki ) < β (β > 0 small). Uq }. For n > n one has δx. As δx. Cq } with C i ⊂ Ui . . .r the number of such s ≡ r (mod m). Let U = {U1 .
As x ∈ X(U).r ) .i . Now a = (a0 a1 · · · akj −1 ) has its distribution µa on Um given by the probabilities {pi.N (φ)− φ dµ < ε or SN φ(x) ≤ N ( φ dµ + ε).r ) ≤ hµ (T ) + ε. m > q ≥ 0. h) = {a ∈ E k : H(a) ≤ h}. µ(Ca ) ≥ e µ(e)k(ν(e)+α) . k→∞ k Proof. For each Vi∗ we choose U0.r ) ≤ ϕ(pi ) and so ϕ(pi. Let R(k. i For N = nj + r with j large we form U = U0 U1 · · · UN −1 ∈ UN as follows. Each a ∈ Rk (ν) corresponds to a cylinder set Ca of Σ. Lemma. r=0 i i i For some r ∈ [0. m m i 1 We have yet to check (b). h) ≤ h .i ∩ T −1 U1. Fix a ﬁnite set E and h ≥ 0. i=1 So 1 1 H(a) = ϕ(pi. m) one must have 1 m ϕ(pi. 2.i we have U = U0 · · · Ur−1 a0 a1 · · · akj −1 .r ) ≤ hµ (T ) + ε . pick i with T r+mp x ∈ Vi∗ and let Us = Uq.C. SN φ(U) ≤ SN φ(x) + N γ(U).r }t and some zeros.nj → µ.i · · · Um−1.16. For s > r we write s = r + mp + q with p ≥ 0.i U1. Then 1 lim sup log R(k. 1) consider Rk (ν) = {a ∈ E k : µa (e) − ν(e) < α ∀e ∈ E} .i ⊃ Vi∗ . As the Ca are disjoint and have total measure 1. Let µ be the Bernoulli measure on Σ = ∞ i=0 E with the distribution µ(e) = (1 − α)ν(e) + α/E . For any distribution ν on E and α ∈ (0. Variational Principle 39 and so ϕ(pi ) ≥ i 1 m m−1 ϕ(pi. For s < r pick Us ∈ U containing T s x. Letting ap = U0.i ∩ · · · ∩ T −m+1 Um−1. . Since each e ∈ E occurs in a at most k(ν(e) + α) times. for j large we will have N δx. Since δx.
k→∞ k and Now let ε → 0. Proof. or 1 log Rk (ν) ≤ k −(ν(e) + α) log µ(e) e ≤ H(µ) + e 3α log µ(e) . and (b) if H(ν ) ≤ h then for some ν ∈ N one has ν (e) − ν(e) < α Then R(k. h) ≤ log N  + h + ε k k 1 lim sup log R(k. h) ≤ h + ε. the second term on the right approaches 0 and H(µ) → H(ν) uniformly in ν. For u ∈ R let Ym (u) be the set of x ∈ Xm for which 2.15 holds for some µ ∈ V (x) with φdµ ∈ [u − ε.15 X = m Xm since V (x) = ∅. Hence. 1 1 log R(k. Let T : X → X be a continuous map on a compact metric space and φ ∈ C (X). we get 1 log Rk (ν) ≤ H(µ) + 3αE(log E − log α) . Then PT (φ) = sup hµ (T ) + µ φ dµ where µ runs over MT (X). Rk (ν). Variational Principle. k When α → 0.40 2 General Thermodynamic Formalism 1 ≥ Rk (ν) e µ(e)k(ν(e)+α) . Set .15 holds with this m and some µ ∈ V (x). u + ε]. k for all k and ν. Once α is so chosen. For each m > 0 let Xm be the set of points x ∈ X for which 2. As µ(e) ≥ α/E. h) ⊂ ν∈N for all e.17. 2. for any ε > 0 one can ﬁnd α small enough that 1 log Rk (ν) ≤ H(µ) + ε. Let U be a ﬁnite cover of X and > 0. let N be a ﬁnite set of distributions on E so that (a) H(ν) ≤ h for ν ∈ N . By 2.
15 (c) for x ∈ Ym (u) lie in R(k. As ε was arbitrary. m(c − u + 2ε). λ) ≤ N =N0 ∞ λN exp(N (c + 5ε + γ(U))) . Um ). Variational Principle 41 c = sup hµ (T ) + µ φ dµ . U) (c + γ(U)) = c . U) ≤ c + γ(U).18.C. λ) as small as desired. 2. The number of possibilities for U for any ﬁxed N = km is hence at most b(N ) = Em R(k. λ ≤ e−P (φ. φ ]. . For x ∈ Ym (u) the µ satisﬁes hµ (T ) ≤ c − u + ε. The a ∈ (Um )k appearing in 2. P (φ. By 2. b(N ) ≤ exp(N (c − u + 3ε)) and ∞ Z(Γ. 1−β where β = λ exp(c + 5ε + γ(U)) < 1. Um ) .14.u be the collection of all U showing up in the present situation for some N greater than a ﬁxed N0 . Finally P (φ) ≤ ≤ diam(U)→0 diam(U)→0 lim P (φ. α . U) ≤ c + 5ε + γ(U) . taking unions of such Γ s we obtain a Γ covering X with Z(Γ. Then PT (φ) = sup PT Xα (φXα ) .10. As X = m=1 Xm and Xm = Ym (u1 ) ∪ · · · ∪ Ym (ur ) where u1 . . By Lemma 2. λ) ≤ N =N0 λN b(N ) exp(N (u + 2ε + γ(U))) . . ur are εdense in [− φ . N Let Γ = Γm. For large enough N0 . Then Γ covers Ym (u) and lim sup N →∞ ∞ Z(Γ. Suppose {Xα }α∈Λ is a family of compact subsets of X and T Xα ⊂ Xα for each α. m(c − u + 2ε). βN = N =N0 ≤ β N0 . lim The inequality c ≤ P (φ) follows from Theorem 2.U) or P (φ. Corollary.16 1 log b(N ) ≤ c − u + 2ε . . We have seen that for λ < exp(−(c + 5ε + γ(U))) any Ym (u) can be ∞ covered by Γ ⊂ W ∗ (U) with Z(Γ. λ) < 1.
one 1 has m Hµ (D ∨ · · · ∨ T −m+1 D) ≤ hµ (T ) + α for some m. then µ is called an equilibrium state for φ (w.im−1 <β. hµ (T ) + φ dµ = PT Xα (φXα ) . The Gibbs state µφ of φ ∈ FA in Chapter 1 was shown to be the unique equilibrium state for such a φ.r. α In the proof of 2....im−1 ⊂ k=0 T −k Dik with µ k m−1 T −k Dik Ki0 . .im−1 ⊂ int k T −k Dik . then V (x) ⊂ MT (Xα ) and so c = sup hµ (T ) + φ dµ : µ ∈ x∈X V (x) ≤ sup PT Xα (φXα ) ... and the proposition follows from 2..s.. . Proof. Then µ → hµ (T ) + φdµ will be also. If µ ∈ MT (Xα ). function on a compact space assumes its supremum.19. one can ﬁnd a partition D = {D1 . .im−1 : ij = i}. .. As the Li are disjoint compact sets. Dn } with diam(D) < ε..t. . Suppose that for some ε > 0 one has hµ (T. Proposition..17 what was actually used about the number c was c ≥ hµ (T ) + φ dµ for µ ∈ V (x)... If x ∈ Xα . ...17 and the fact that an u. and D = {D1 . . Equilibrium States If µ ∈ MT (X) satisﬁes hµ (T )+ φ dµ = PT (φ). One then has Ki0 . 2. then µ ∈ MT (X) and PT (φ) ≥ hµ (T ) + Hence PT (φ) ≥ sup µ∈MT (Xα ) φ dµ . Let β > 0 and pick a m−1 compact set Ki0 . Then Di ⊃ Li = j=0 {T j Ki0 . D) = hµ (T ) whenever µ ∈ MT (X) and diam(D) < ε. . α > 0. Then every φ ∈ C (X) has an equilibrium state. D. We show that µ → hµ (T ) is upper semicontinuous on MT (X). So c would work just as well there to yield PT (φ) ≤ c . Dn } with diam(D ) < ε and Li ⊂ int(Di ).42 2 General Thermodynamic Formalism Proof. T ).. Fixing µ ∈ MT (X).. .c.
The deﬁnition of h(T ) was made by Adler. References The deﬁnition of hµ (T ) is due to Kolmogorov and Sinai (see [2])...10.10). The condition in 2. one deﬁnes the topological entropy of T by h(T ) = PT (0) .20). Conditions on φ become important for the uniqueness of equilibrium state and then only after stringent conditions have been placed on the homeomorphism T . Proof.20.17 and 2. For expansive T Ruelle [15] deﬁned PT (φ) and proved Theorems 2. i. and Goodman [9] (2. The theorems for this case are due to Goodwyn [10] (Theorem 2. The Axiom A diﬀeomorphisms will be close enough to the subshifts σ : ΣA → ΣA so that one can prove uniqueness statements. The proofs we have given in these notes are adaptations of [4].References 43 If ν is close to µ in the weak topology. most of the work has to do with the more complicated topology of X.17).5. Corollary. the case φ = 0. 2. D ) ≤ 1 Hν (D ∨ · · · ∨ T −m+1 D ) m 1 ≤ Hµ (D ∨ · · · ∨ T −m+1 D) + α ≤ hµ (T ) + 2α . every φ ∈ C (X) has an equilibrium state.19 has nothing to do with φ.e. this and ν implies k T −k Dik − µ k T −k Dik hν (T ) = hν (T. 2. Konheim and McAndrew [1]. Goodman [8] (general X.. m 2. Recall 2. 2. Dinaburg [6] (X ﬁnite dimensional. Gureviˇ [11] gives a T where φ = 0 has no equilibrium states and Misiurewicz c [13] gives such a T which is a diﬀeomorphism. One notices that the condition in 2.. Walters’ paper is therefore closely related to earlier work on the topological entropy h(T ). In the transition from ΣA to a general compact metric space X. For subshifts these results were proved earlier by Parry [14]. one will have ν k T −k Dik ≥ µ(Ki0 . For general T the deﬁnition and results are due to Walters [16].im−1 ) − β ≤ 2βnm . The motivation for this chapter comes from two places: the theory of Gibbs states from statistical mechanics and topological entropy from topological dynamics (see references). If T is expansive.19 is satisﬁed by a class of maps which includes all aﬃne maps on Lie groups [3] and Misiurewicz [13] showed that equilibrium states exist under a somewhat weaker condition. For β small enough. ..20. Taking φ = 0.17).
. Trans. Ph.Y. P. 185190. Soc. Amer. Gureviˇ: Topological entropy of denumerable Markov chains. Let T : M → M be a continuous map on a compact manifold and λ an eigenvalue of the map T∗ : H1 (M ) → H1 (M ) on onedimensional homology. this result has been carried over to ﬂows by FrancoSanchez [7]. no. Sci. Konheim and M. Math.Warwick 1974 (Proc. Math. Bull. Amer. Goodman [9] gives a minimal subshift where φ = 0 has more than one equilibrium state. N. (1 ) 3. 23 (1969).44 2 General Thermodynamic Formalism Ruelle [15] showed that for expansive T a Baire set of φ have unique equilibrium states. G. (2) 5 (1972). 97 (1975). Math. 1978 (note of the editor). Symp. D. 9. 1973/1974). Soc. S´r. Walters: A variational principle for the pressure of continuous transformations. 11. Coventry. Soc. 8. Springer. R. P. 237251. Notes in Math. 3 (1971). Goodman: Maximal measures for expansive homeomorphisms. 1. Trans. 10. It is conceivable that this inequality is true for λ for any Hk (M ) (not just k = 1) provided T is a diﬀeomorphism. Bowen: Entropyexpansive maps. Topology and Dynamical Systems. Manning: Topological entropy and the ﬁrst homology group. Billingsley: Ergodic Theory and Information. of California. 2. Soc. A. Amer. 309319. Soc. 679688. Soviet Math. 184 (1973). 1 Reprint: Robert E. FrancoSanchez: Flows with unique equilibrium states. 176180. I believe mathematical physicists know of speciﬁc φ on Σn which do not have unique equilibrium states. 21 (1973). 16. 7. 114 (1965). N. E. T. 5. McAndrew: Topological entropy. e 14. M. in statistical mechanics one looks at Zm actions instead of just homeomorphisms and gets nonuniqueness for m ≥ 2 even with simple φ’s. pp. 10 (1969). Warwick. . W. I. Soc. Math. Misiurewicz: Diﬀeomorphism without any measure with maximal entropy. Math. R. E. 337378. Bull. Systems Theory 8 (1974/75). Math. 13. Krieger Publishing Co. 4. 903–910. Astronom. T. Polon. 1975. in “Dynamical Systems” . Soc. Trans. Berkeley. Amer. 323331. USSR Izvestia 5 (1971). B. Trans. Trans. Goodman: Relating topological entropy and measure entropy. M. Amer. W. Univ. Amer. Goodwyn: Topological entropy bounds measuretheoretic entropy. 12. L. Math. 185 (1973). 439444. R. A. Lect. 6. Appl. Berlin. Math. Proc. J. 3. Math. Finally we mention a very interesting result in a diﬀerent direction. T. Bowen: Topological entropy for noncompact sets. Bowen: Some systems with unique equilibrium states. Dinaburg: On the relations among various entropy characteristics of dynamical systems. Parry: Intrinsic Markov chains. 911915. 193202. Then Manning [12] showed h(T ) ≥ log λ. N. c Dokl. thesis. Amer. Phys. H. Ruelle: Statistical mechanics on a compact set with Zν action satisfying expansiveness and speciﬁcation. 164 (1972). Soc. T. L.. J. R. Uniqueness was proved in [5] for certain φ when T satisﬁes expansiveness and a very restrictive condition called speciﬁcation. Sci. Math. 15. Huntington. 112 (1964). Adler. 125136. no. London Math. Acad. 1974. 5566. London Math. 468.D. Univ. 4. 937971. Wiley (1965).
Df (Ex ) = Ef (x) . for every neighborhood U of x. clearly such an x is in Ω(f ). E u = x∈Λ Ex s s and E = x∈Λ Ex are continuous subbundles of TΛ M = x∈Λ Tx M and TΛ M = E u ⊕ E s . . Then the derivative of f can be considered a map df : T M → T M where T M = x∈M Tx M is the tangent bundle of M and dfx : Tx M → Tf (x) M . (b) there exist constants c > 0 and λ ∈ (0. Deﬁnition We now suppose that f : M → M is a diﬀeomorphism of a compact C ∞ Riemannian manifold M . The set Ω = Ω(f ) of all nonwandering points is seen to be closed and f invariant. n ≥ 0 u when v ∈ Ex . A point x is periodic if f n x = x for some n > 0. Ex vary continuously with x. Deﬁnition. A closed subset Λ ⊂ M is hyperbolic if f (Λ) = Λ and each tangent space Tx M with x ∈ Λ can be written as a direct sum u s Tx M = Ex ⊕ Ex of subspaces so that s s u u (a) Df (Ex ) = Ef (x) . n ≥ 0. 1) so that Df n (v) ≤ cλn v and Df −n (v) ≤ cλn v s when v ∈ Ex . Condition (c) actually follows from the others. s u (c) Ex . u Remark.3 Axiom a Diﬀeomorphisms A. A point x ∈ M is nonwandering if U∩ n>0 f n U = ∅.
A type of f studied extensively by Russian mathematicians is the Anosov diﬀeomorphism: f is Anosov if M is hyperbolic [2].2. and −n −n n u d(f x. A metric is adapted (to an Axiom A f ) if Ω(f ) is hyperbolic with respect to it with c = 1. f satisﬁes Axiom A if Ω(f ) is hyperbolic and Ω(f ) = {x : x is periodic}. Proof. We shall see a little later that such diﬀeomorphisms always satisfy Axiom A. s u (c) Wε (x). For x ∈ M deﬁne W s (x) = {y ∈ M s Wε (x) = {y ∈ M W u (x) = {y ∈ M u Wε (x) = {y ∈ M : : : : d(f n x. For small ε > 0 s u s s u (a) Wε (x). 3. Tx Wε (x) = u Ex . Wε (x) are C r disks for x ∈ Λ with Tx Wε (x) = Ex . f y) ≤ λ d(x. Theorem.46 3 Axiom a Diﬀeomorphisms Deﬁnition. Proof. The truth of this condition does not depend upon which metric is used although the constants c and λ do. 3. The reader should study the examples in [14]. f n y) → 0 as n → ∞} d(f n x. Every Axiom A diﬀeomorphism has an adapted metric. The Riemannian metric on M is used to state condition (b) in the definition of hyperbolic set. This lemma is due to Mather. The following stable manifold theorem is the main analytic fact behind the behavior of Axiom A diﬀeomorphisms.2 is that Wε (x) ⊂ W s (x) for x ∈ Λ. y) for y ∈ Wε (x). This will keep various estimates a bit simpler. . One then sees that s f −n Wε (f n x) W s (x) = n≥0 for x ∈ Λ. Right now we mention that it is unknown whether Ω(f ) = M for every Anosov f . f −n y) ≤ ε for all n ≥ 0} . See Hirsch and Pugh [8]. Lemma. f −n y) → 0 as n → ∞} d(f −n x. f n y) ≤ ε for all n ≥ 0} d(f −n x. f n y) ≤ λn d(x. y) for y ∈ Wε (x). n ≥ 0. n ≥ 0. Similarly W u (x) = n≥0 u f n Wε (f −n x) . This deﬁnition is due to Smale [14]. Wε (x) vary continuously with x (in C r topology). s One consequence of 3. s (b) d(f n x. We will always use an adapted metric. See [8] for a proof. Let Λ be a hyperbolic set for a C r diﬀeomorphism f .1.
Furthermore [x. Otherwise y ∈ Wε (x) ∩ Wε (x). y] ∈ Ω(f ) and [·. f kmn q) → 0 as k → ∞ . For any small s u ε > 0 there is a δ > 0 so that Wε (x) ∩ Wε (y) consists of a single point [x. .i ) = Xj+1. We claim that Xp = Bδ (Xp ) = {y ∈ Ω : d(y. The ﬁrst statement follows because the intersection s u Wε (x) ∩ Wε (x) = {x} is transversal and such intersections are preserved under small perturbation.i ∪· · ·∪Xni .5. y] ∈ Ω(f ) uses that the periodic points are dense in Ω(f ). y) ≤ δ. B. q] ∈ W u (p) ∩ W s (q) ∩ Ω. y ∈ Ω(f ) and d(x. Pick x ∈ W u (p) ∩ Ω with d(x. i. ·] : {(x. 3. Suppose f satisﬁes Axiom A. Spectral Decomposition From now on f will always be an Axiom A diﬀeomorphism. Let Λ be a hyperbolic set. For p ∈ Ω periodic let Xp = W u (p) ∩ Ω. Canonical Coordinates. As periodic points are dense in Ω. y] ∈ M and thus canonical coordinates on M (instead of Ω(f )) without any assumption on periodic points.i topologically mixing. 4]).e. f k y) > ε for some k ∈ Z . y) ∈ Ω(f ) × Ω(f ) : d(x. Proof ([14. then d(f k x. 3. Spectral Decomposition. d(f kmn x . Lemma.i with the Xj. y] whenever x. if x ∈ Λ and y ∈ M with y = x.i (Xnj +1. q) = d(f kmn x .3. One can write Ω(f ) = Ω1 ∪ Ω2 ∪ · · · ∪ Ωs where the Ωi are pairwise disjoint closed sets with (a) f (Ωi ) = Ωi and f Ωi is topologically transitive. s u Proof..4. See Smale [14]. Proof. y) ≤ δ} −→ Ω(f ) is continuous. f (Xj.3. x is also in this intersection and so y = x by 3.i ) and f ni Xj.3.i ’s pairwise disjoint closed sets. it is enough to see that a periodic q ∈ Bδ (Xp ) is in Xp . Then there is an ε > 0 so that Λ is expansive in M . Xp ) < δ} . one has f kmn x ∈ f kmn W u (p) = W u (f kmn p) = W u (p) and So q ∈ Xp . Let δ be as in 3. q) < δ and consider x = [x.i = X1. In the Anosov case one of course has [x. (b) Ωi = X1. To get [x. Spectral Decomposition 47 3. Letting f m p = p and f n q = q.B.
In summary. q periodic. V are nonempty subsets of Xr open in Xr (i. j→−∞ The above argument with the roles of p and q reversed gives Xp ⊂ Xq . Notice that if g = f n and n is a multiple of every ni . we have f j p = f tN p ∈ Xr and f kmn xj = f tN (f −jN xj ) ∈ f tN U ∩ V provided k is large. The transitivity in (a) is implied by the mixing in (b). one has f jmn x ∈ Xq for all j and p = lim f jmn x ∈ X q = Xq . For each 0 ≤ j < mn with f j p ∈ Xr one can ﬁnd a point xj as in the beginning of this proof so that xj ∈ f j U and f kmn xj ∈ V for large k . Pick periodic points p ∈ U and q ∈ V . So f permutes the Xpj ’s and the Ωi are just the union of the Xpj ’s in the various cycles of the permutation.e. Xq are either disjoint or equal. Writing tN = kmn + j. then Xp = Xr = Xq . say f m p = p. If x is as above. Xp = Xq . and so by compactness (the Xp are open) let Ω = Xp1 ∪ · · · ∪ Xpt with the Xpj ’s pairwise disjoint.. u then Wδ (q) ⊂ Xp and W u (q) = k≥0 u f kmn Wδ (q) ⊂ k≥0 f kmn Xp = Xp .) It follows that Xq ⊂ Xp .48 3 Axiom a Diﬀeomorphisms Notice that f Xp = Xf (p) since f W u (p) = W u (f (p)). (Note that y ∈ W u (q) iﬀ f −kmn y → q as k → ∞. Then f (Xpj ) = Xf pj intersects and hence equals some Xpi . then this intersection is open in Ω and hence contains a periodic point r. then the basic sets . in Ω). 0 ≤ j ≤ mn. We ﬁnish by showing f N : Xr → Xr is mixing whenever r is periodic and N positive with f N Xr = Xr . Now Ω= p periodic Bδ (Xp ) = p Xp . f n q = q. For if Xq ∩ Xq = ∅. then f kmn x ∈ Xq for large k as Xq = Bδ (Xq ) is open in Ω. Suppose U. Then f tN U ∩ V = ∅ for large t and f N Xr is topologically mixing. if q ∈ Xp with p. The Ωi in the spectral decomposition of Ω(f ) are called the basic sets of f . If q ∈ Xp as above. As f imn Xq = Xf imn q = Xq . Now any two Xp .
By the triangle inequality d(f i x. (s + 1)M ). f i−tM +M x(t−1)M ) ε λtM −i ≤ t=s+1 ≤ ελ 1−λ u s where we used xtM ∈ Wε (f M x(t−1)M ). A point x ∈ M βshadows x if d(f i x. We will at times restrict our attention to mixing basic sets and recover the general case by considering f n .i ’s and gXj. y ∈ Ω with d(x. Since xsM ∈ Wε (xsM ) d(f i−sM xsM . i. y) ≤ δ. ε) as in 3. xi ) ≤ δ ελ +ε+ · 1−λ 2 . yj ) < δ/2 for all j ∈ [0. rM ] pick s with i ∈ [sM. M Pick M so large that λ ε < δ/2 and then α > 0 so that: if {yi }M is an α−pseudoorbit in Ω. Deﬁne xkM recursively i=0 for k ∈ [0.3. Let ε > 0 be a small number to be determined later and choose δ ∈ s u (0. Now set x = f −rM xrM . For i ∈ [0. xi ) < δ/2 . r] by x0 = x0 and u s x(k+1)M = Wε (f M xkM ) ∩ Wε (x(k+1)M ) ∈ Ω .6. f i−sM xsM ) ≤ t=s+1 r d(f i−tM xtM .e. For every β > 0 there is an α > 0 so that every αpseudoorbit {xi }b in Ω (i. xi ) ≤ β for all i ∈ [a. x(k+1)M ) < δ and x(k+1)M exists. f M xkM ) ≤ λM ε < δ/2 and d(f M xkM . then i=0 d(f j y0 .e. f i−sM xsM ) ≤ ε . Consider ﬁrst an αpseudoorbit {xi }rM with r > 0.. A sequence x = {xi }b (a = −∞ or b = +∞ is permitted) of points in i=a M is an αpseudoorbit if d(f xi ..i is mixing. b] . xi+1 ) < α for all i ∈ [a. every xi ∈ Ω) is βshadowed by a point x ∈ Ω. Spectral Decomposition 49 of g are the Xj. By the choice of α one has d(f i−sM xsM . M ] . b − 1) . x(k+1)M ) < δ/2 by the choice of α. 3. then r d(f i x. This makes sense: d(f M xkM . Proposition. so d(f M xkM . i=a Proof.B. Wε (x)∩Wε (y)∩Ω = ∅ whenever x.
If y is a nonwandering point for an Anosov f .7 for such x converge to y. pick xi ∈ Ω with for all i . Now any αpseudoorbit {xi }n in Ω extends to {xi }rM when rM ≥ n i=0 i=0 by setting xi = f i−n xn for i ∈ (n.3 is true also for Anosov diffeomorphisms and so then is 3. The periodic points x constructed in 3. Then {xi }∞ i=−∞ is an αpseudoorbit. n). Let xi = f k x for i ≡ k (mod n). For small β and γ this implies y = x ∈ Ω by 3.4) f n x = x . rM ]. then there is an x ∈ Ω with f n x = x and d(f k x. y) < γ and d(f n x. Then x ∈ Ω βshadows {xi }+∞ .7. 3. Thus the proposition holds for ﬁnite pseudoorbits. xi ) + d(xi . There is a neighborhood U of Ω(f ) so that f n U = Ω(f ) . Let f satisfy Axiom A. i=−∞ 3. x) < γ for some n. Proof. Given any β > 0 there is an α > 0 so that the following holds: if x ∈ Ω and d(f n x. then for any γ one can ﬁnd x with d(x. An x ∈ Ω shadowing this extended pseudoorbit will shadow the original one. f i+n x ) ≤ 2β and by expansiveness (Lemma 3. Anosov’s Closing Lemma. Let x ∈ Ω βshadow it. y) < γ implies d(f x. Ω) < γ}.4. We have been assuming f satisﬁes Axiom A. Then d(f i y. If {xi }b is a ﬁnite αpseudo orbit. Proof.50 3 Axiom a Diﬀeomorphisms For small ε this is less that the given β. Ωj ) → 0 as n → ∞} and W u (Ωj ) = {x ∈ M : d(f −n x. If f is an Anosov diﬀeomorphism. f i f n x ) ≤ d(f i x .6 and 3. Fundamental Neighborhood. f y) < α/2 . k ∈ [0.6. then f satisﬁes Axiom A. n] . x) < α. For Ωj a basic set of an Axiom A diﬀeomorphism one let W s (Ωj ) = {x ∈ M : d(f n x. f i x) < β + γ f n U . If y ∈ d(f i y. Pick γ < α/2 so that ∀x. Ωj ) → 0 as n → ∞} . f k x ) ≤ β for all k ∈ [0. If {xi }+∞ is i=−∞ an αpseudoorbit in Ω. 3. xi ) < γ.7. . however 3. i=a then so is {xj+a }b−a and x shadowing this one yields f −a x shadowing the j=0 original. Corollary. y ∈ M .8. Let β be small and α as in 3. Then d(f i x . We must show that the periodic points are dense in Ω(f ). then ﬁnd x(m) ∈ Ω βshadowing {xi }m i=−m and let x be a limit point of the sequence x(m) . n∈Z Proof. using M in place of Ω(f ).9. d(x. n∈Z Let U = {y ∈ M : d(y.
Ωj ) < γ for all n ≥ N . C. Let R be a closed rectangle. R has boundary ∂R = ∂ s R ∪ ∂ u R where . Suppose f n y → Ωj as n → ∞. Lemma. Proposition. Then y ∈ f −N W s (f N x) = W s (x). let s W s (x. for n < N let xn = f n−N xN . The proof for W u (Ωj ) is similar and we have proved the second statement with Uj = {y ∈ M : d(y. f n y) ≤ γ. As a subset of Ωs . Markov Partitions A subset R ⊂ Ωs is called a rectangle if it has small diameter and [x. W s (Ωj ) ⊃ x∈Ωj W s (x). R) = Wε (x) ∩ R u and W u (x. Ωj ) < γ}. Letting x ∈ Ωj shadow it.C.11. As Ω = Ω1 ∪ · · · ∪ Ωs is a disjoint union of closed invariant sets one then sees that s s M= j=1 W s (Ωj ) = j=1 W u (Ωj ) and that there are disjoint unions. Markov Partitions 51 Using the deﬁnition of nonwandering sets it is easy to check that f n x → Ω and f −n x → Ω as n → ∞.10. 3. 3. R is called proper if R is closed and R = int(R) (int(R) is the interior of R as a subset of Ωs ). y ∈ R . one gets s f N y ∈ Wε (f N x) ⊂ W s (f N x) (provided γ was small enough). y] ∈ R whenever x. is clear. R) = Wε (x) ∩ R where ε is small and the diameter of R is small compared to ε. The {xn }∞ n=−∞ is a pseudoorbit in Ωj . and k≥0 f Uj ⊂ k u Wε (Ωj ) = Proof. The reverse inclusion. W s (Ωj ) = x∈Ωj W s (x) and W u (Ωj ) = For ε > 0 there is a neighborhood Uj of Ωj so that s f −k Uj ⊂ Wε (Ωj ) = k≥0 x∈Ωj s Wε (x) u Wε (x) . x∈Ωj x∈Ωj W u (x). For x ∈ R. say d(f n y. Pick xn ∈ Ωj for n ≥ N with d(xn .
q ] ∈ Σ(P ) by ∗ qj = qj qj for for j≥0 j ≤ 0. . If x ∈ int(R). i. y) < γ . for each x ∈ Ωs there are q with x = θ(q).6. Similarly x ∈ int(W u (x. R)). Proof. A Markov partition of Ωs is a ﬁnite covering R = {R1 . Then d(f j θ(q ∗ ). f j θ(q)) ≤ 2β for j ≤ 0. R))} / u ∂ R = {x ∈ R : x ∈ int(W s (x. pr } be a γdense subset of Ωs and ∞ Σ(P ) = q∈ −∞ P : d(f qj . Let Ωs be a basic set for an Axiom A diﬀeomorphism f .e. x] ∈ R. Let β > 0 be very small and choose α > 0 small as in Proposition 3. . For y ∈ Ωs near x the points s u [x. Theorem. Hence for y ∈ Ωs close enough to x. W s (x. x]. .e. . every αpseudoorbit in Ωs is βshadowed in Ωs .12. R) = R∩(Wε (x)∩Ω) is a neighborhood of x u in Wε (x)∩Ω since R is a neighborhood of x in Ω. Thus x ∈ int(R). s u So θ(q ∗ ) ∈ W2β (θ(q)) ∩ W2β (θ(q )). Rj ) and f W s (x. . i. Ri ) ⊂ W s (f x. For q. y] ∈ Wε (x) ∩ Ω depend continuously on y. q ∈ Σ(P ) with q0 = q0 we deﬁne q ∗ = [q. . R))} / u s and the interiors of W u (x. Then Ωs has Markov partitions R of arbitrarily small diameter. Choose γ < α/2 so that d(f x. R) are as subsets of Wε (x)∩Ω. Then s u y = [[y. qj+1 ) < α for all j . Rj ) when x ∈ int(Ri ). Ri ) ⊃ W u (f x. [x. y] ∈ Wε (x) ∩ Ω and [x. For each q ∈ Σ(P ) there is a unique θ(q) ∈ Ωs which βshadows q. R)) and x ∈ int(W s (x. Suppose x ∈ int(W u (x. Let P = {p1 . f x ∈ int(Rj ). . R). y]] ∈ R ∩ Wε (y) ∩ Wε (y) s u and y = y as Wε (y) ∩ Wε (y) = {y}. 3. Rm } of Ωs by proper rectangles with (a) int(Ri ) ∩ int(Rj ) = ∅ for i = j. . [x.. u Proof. (b) f W u (x. y] ∈ R and [y. Deﬁnition. R)). f j θ(q)) ≤ 2β for j ≥ 0 and d(f j θ(q ∗ ). . f y) < α/2 when d(x.52 3 Axiom a Diﬀeomorphisms ∂ s R = {x ∈ R : x ∈ int(W u (x.. then W u (x. Wε (x)∩ Ω.
Consider y ∈ W s (x. . y] = θ[q. Now T = {T1 . Otherwise there is a γ > 0 so that for every N one can ﬁnd q N . Tr } is a covering by rectangles and (i) and (ii) above are like the Markov condition (b).k = {x ∈ Tj : W u (x. y ∈ Ts we write x = θ(q). y = θ(q ) with q0 = ps = q0 . For each x ∈ Ωs let T(x) = {Tj ∈ T : x ∈ Tj } and T ∗ (x) = {Tk ∈ T : Tk ∩Tj = ∅ for some Tj ∈ T(x)}. Tj ) ∩ Tk = ∅.C. q ] ∈ Ts . θ : Σ(P ) → Ωs is continuous. N ] . Tj ) ∩ Tk = ∅} 4 Tj. y = θ(q ). For Tj ∩ Tk = ∅. Tj ) ∩ Tk = ∅. q ] and f y = θ(σ[q. Z = Ωs \ j ∂Tj is an open dense subset of Ωs . q ] = [θ(q). W s (x. one can show that s u Z ∗ = {x ∈ Ωs : Wε (x) ∩ ∂ s Tk = ∅ and Wε (x) ∩ ∂ u Tk = ∅ for all Tk ∈ T ∗ (x)} is open and dense in Ωs .k = {x ∈ Tj : W u (x.N = qj. θ(q )] . In fact. Then d(f j x. f j y) ≤ 2β for all j and d(x.13. We have proved (i) f W s (x. θ(q N )) ≥ γ. Ts ).k = {x ∈ Tj : W u (x. Lemma. However the Tj ’s are likely to overlap and not be proper. Each Ts is closed. this contradicts expansiveness of f Ωs . Tj ) ∩ Tk = ∅}. Then y = [x. Tj ) ∩ Tk = ∅. Tt ). (ii) f W u (x. using arguments similar to 3. N ] but d(θ(q N ). Ts ) ⊂ W s (f x. We now see that Ts = {θ(q) : q ∈ Σ(P ). i. . let 1 Tj. q N ∈ Σ(P ) with qj. q0 = ps } is a rectangle. f y ∈ Wε (f x.11. Tj ) ∩ Tk = ∅. this follows from the following lemma. W s (x. As T is a closed cover of Ωs . Tt ). Ts ).k = {x ∈ Tj : W u (x. Markov Partitions 53 θ[q. q ]) ∈ Tt s as σ[q. yN = θ(q ) one has d(f j xN . W s (x. . If xN = θ(q N ) . Then [x. A similar proof shows f −1 W u (f x.. f j yN ) ≤ 2β ∀j ∈ [−N. For x. Tj ) ∩ Tk = ∅} = Tj ∩ Tk 2 Tj. W s (x. Proof.e. y) ≥ γ. Ts ) ⊃ W u (f x. Since f y ∈ Wε (f x) (diam(Ts ) ≤ s 2β is small compared to ε).N for all j ∈ [−N. q0 = ps . Tj ) ∩ Tk = ∅} 3 Tj. . Tt ) ⊂ W u (x. y] = θ[q. Tt ). Suppose x = θ(q) with q0 = ps and q1 = pt . . Taking subsequences we may assume xN → x and yN → y as N → ∞. 3. q ] has q = pt in its zeroth position.
Tj ) and W u ([x. Rm } . Suppose y ∈ R(x) ∩ Z ∗ . To show that R is Markov we are left to verify condition (b). Now R(x) is an open rectangle (x ∈ Z ∗ ). . f y ∈ Tj . Tk ∩ Tj = ∅ and x ∈ Tj. As there are only ﬁnitely n many Tj. Since / R(x) ⊂ T(x) and R(x) ∩ Tj = ∅ for Tj ∈ T(x). R(x) = R(y) and y ∈ Wε (x). s Suppose x.k 3 Tj. then W s ([x. For R(x) = R(x ) int R(x) ∩ int R(x ) = R(x) ∩ R(x ) = ∅ .k as x does since n Tj. As Z ∗ is dense in Ωs . Tj ) = n W u (y. this implies Tj. R(x) \R(x) has no interior (in Ωs ) and R(x) = int(R(x)). hence (R(x) \R(x)) ∩ Z ∗ = ∅. For x ∈ Z ∗ . y ∈ Tj . there is a y ∈ R(x) ∩ R(x ) ∩ Z ∗ . Tj ) = W s (x.k ⊃ R(x). y lies in the same Tj. We will show / R(f x) = R(f y). then R(x) = R(y) = R(x ). R(x ) = R(x) or R(x ) ∩ R(x) = ∅. y ∈ Z ∗ ∩ f −1 Z ∗ . .54 3 Axiom a Diﬀeomorphisms Ws Wu 4 Tj. For x ∈ Z ∗ deﬁne R(x) = n n {int(Tj. If R(x) ∩ R(x ) = ∅ (x.k Tj Tk If x. Then x = θ(q) ∈ Ts and by inclusion (i) above .k is a rectangle open in Ωs and each x ∈ Tj ∩ Z ∗ n lies in int(Tj.k 2 Tj. First T(f x) = T(f y).k } .k ) : x ∈ Tj . Let R = {R(x) : x ∈ Z ∗ } = {R1 .k ’s there are only ﬁnitely many distinct R(x)’s. Let f x = θ(σq) with q1 = pj and q0 = ps . y]. one gets T(y) = T(x). Tj ). For n Tj ∈ T(x) = T(y) and Tk ∩ Tj = ∅. hence R(y) = R(x). . Otherwise assume f x ∈ Tj . y].k 1 Tj.k ) for some n. . x ∈ Z ∗ ).
Now Z ∗ ⊃ Ωs \(Y1 ∪Y2 ) is open and dense. Tj ) = s W (f x. Tt ) ⊂ W s (f z. Ri ) = {[f x . Tt ) ∩ W u (y. Tj ) ⊂ f W u (x.t . f y] ∈ W s (f z. Tj ) . and f z = [f z. Tk ).11). Ri ) = {[x . Let f z = θ(σq ). z] : z ∈ W s (f x. Then z = [z. y are n in the same Ts. z ] ∈ W s (z.C. R(f x) ). Tj ) ∩ Tk = ∅ . This completes the proof of half of the Markov conditions (b). Furthermore if x ∈ (Y1 ∪ Y2 ) ∩ f −1 (Y1 ∪ Y2 ) then x ∈ / Z ∗ ∩ f −1 Z ∗ and the set of y ∈ W s (x. Ts ). q0 = ps . Then by inclusion (ii) f z ∈ W u (f x.k . Ri )} and f W s (x . R(x)) with y ∈ Z ∗ ∩ f −1 Z ∗ is open and s dense in W s (x. y] : y ∈ W s (x. noting that Wf = Wf −1 . Tj ). So R(f x) = R(f y). f y ∈ Tj a rectangle). For small δ > 0 the sets ⎧ ⎧ ⎫ ⎫ ⎨ ⎨ ⎬ ⎬ s u Y1 = Wδ (z) : z ∈ ∂ s Tj Wδ (z) : z ∈ ∂ u Tj and Y2 = ⎩ ⎩ ⎭ ⎭ j j are closed and nowhere dense (like in the proof of 3. By the previous paragraph R(f y) = R(f x) for such y. R(x) ) (as a subset of Wε (x) ∩ Ω). R(x) ) ⊂ R(f x) . f z ∈ W u (f x. Ri )} ⊂ {[f x . As f y ∈ Wε (f x) we have W s (f y. Then z ∈ Tt and f W s (z. contradicting f y ∈ Tj . Tk )∩W u (f y. f y belong to the same Tj. Alternatively one could apply u s the above to f −1 . f y] : y ∈ W s (x. We want to show / n s that f x. q1 = pk and q0 = pt . Rj )} ⊂ W s (f x . by continuity f W s (x. . Now z ∈ W u (x. Now let f x. R(x) ) ⊂ W s (f x. Ts ) ∩ Tt as x. Tj ) ∩ Tk . For any x ∈ Ri ∩ f −1 Rj one has W s (x . f y ∈ Tj and Tk ∩ Tj = ∅. q1 = pj . s As f W s (x. f W s (x. Rj ) . then this open subset of Ωs contains some x satisfying the above conditions. Ri = R(x) and Rj = R(f x). R(x) ) ⊂ Wε (f x). Ts ) . Ts ) ⊂ W s (f x. y] = [z. The other half is proved similarly and the proof is omitted. We will derive a contradiction from W u (f y. Now Ts ∈ T(x) = T(y) and z ∈ Tt ∩ Ts = ∅. Ts ) or z ∈ W u (x. a contradiction. Ts ) ∩ Tt and so there is some z ∈ W u (y. Recall that f x = θ(σq). Tj ) (using f z. −1 If int(Ri ) ∩ f int(Rj ) = ∅. Markov Partitions 55 f y ∈ f W s (x.
Then Aij = 1. Then the rectangle s [C.14. Then f (S)∩ Rj is a usubrectangle of Rj .16. R) for y ∈ S. Proof. For any x ∈ Ri one can therefore ﬁnd some j and xn ∈ int(Ri ) ∩ f −1 int(Rj ) with limn→∞ xn = x.11. Let D ⊂ Wδ (x) ∩ Ω and C ⊂ Wδ (x) ∩ Ω. S ⊂ R. Lemma. then W u (f x. Because S is a usubrectangle (condition (b)) one has S= y∈D W u (y. Proposition. Rj ).15. Ri ) = [W u (x.16 D = ∅ and D = int(D). Let R. This is like 3. by 3. . S be two rectangles. Now . One gets f −1 (∂ u R) ⊂ ∂ u R by a similar argument or by applying the ﬁrst argument to f −1 in place of f .12. Deﬁnition. Proof. f x ∈ Rj . Lemma. Ri ) ⊂ W s (f x.56 3 Axiom a Diﬀeomorphisms D. We have shown f (∂ s R) ⊂ ∂ s R. Suppose S is a usubrectangle of Ri and Aij = 1. Symbolic Dynamics Throughout this section R = {R1 . f (∂ s R) ⊂ ∂ s R and f −1 (∂ u R) ⊂ ∂ u R. Proof. Suppose x ∈ Ri . Proof. 3. Hence f W u (x. D] . . The set j (int(Ri ) ∩ f −1 int(Rj )) is dense in Ri . If s f x ∈ ∂ s R. Ri ) is a neighborhood of x in Wε (x) ∩ Ωs – that is x ∈ ∂ s Ri . Rj ). One deﬁnes the transition matrix A = A(R) by Aij = 1 0 if int(Ri ) ∩ f −1 int(Rj ) = ∅ otherwise . Pick x ∈ Ri ∩ f −1 Rj and set D = W s (x. Lemma. and (b) W u (y. Ri ) ⊃ W u (f x. 3. Ri ) ∩ S. S will be called a usubrectangle of R if (a) S = ∅. As S is proper and nonempty. s u 3. Deﬁnition. D] is proper iﬀ D = int(D) and C = int(C) as subsets of Wδ (x) ∩ Ω and u Wδ (x) ∩ Ω respectively. S) = W u (y. Rm } will denote a Markov partition of a basic set Ωs . 3. Then f W s (x. This is just the same as the last part of the proof of 3.17. Ri ) ⊃ W u (f x. Ri ). ∂ s R = j ∂ s Rj and ∂ u R = j ∂ u Rj . . . x ∈ Rij and f x ∈ Rj . Rj ) and f W u (x. Aij = 1. Rj ) is a neighborhood of f x in Wε (f x) ∩ Ω and so / u s / W (x. S is proper.
As ∂ s R ∪ ∂ u R is nowhere dense. Rj ) for some y ∈ f (D) and W u (y . Proof. then f j x ∈ Rbj and bj = aj because f j x ∈ ∂ s R ∪ ∂ u R. one has W u (f x. Ri ) ∩ Rj = W u (f y. and π is onetoone over the residual set Y = Ωs \ j∈Z f j (∂ s R ∪ ∂ u R).18. Rj ). . As π(ΣA ) is a compact subset of Ωs containing a dense set Y . Thus a = {aj } ∈ ΣA and x = π(a). if y ∈ f (S) ∩ Rj .16. By 3.17 one sees that ⎞ ⎛ n f n−j Raj = Ran ∩ f ⎝ n−1 f n−1−j Raj ⎠ j=1 j=1 is a usubrectangle of Ran . Theorem. then y ∈ W u (y . As ⎞ ⎛ K(σa) = j f −j Raj+1 = f ⎝ j f −j Raj ⎠ = f K(a). s s As f maps Wε (x)∩Ω homeomorphically onto a neighborhood in Wε (f x)∩Ω. f (D) = int(f (D)) and so f (S) ∩ Rj is proper by 3. Rj ). Rj ) = W u (y . Ri ) ∩ Rj ) . then f j x. So f (S) ∩ Rj is a usubrectangle of Rj . If a1 a2 · · · an is a word with Aaj aj+1 = 1. f (y) ∈ Rj and f W u (y. For x ∈ Y pick aj with f j x ∈ Raj . n=1 If x. 3. Rj ) = [W u (f x. For each a ∈ ΣA the set j∈Z f −j Raj consists of a single point. The map π : ΣA → Ωs is a continuous surjection. W s (f x.D. Rj ). That π is continuous is proved like 3. Rj ) ⊂ f (S) ∩ Rj . From this one gets that n Kn (a) = j=−n f −j Raj is nonempty and the closure of its interior. denoted π(a). so π is injective over Y . π(ΣA ) = Ωs . y ∈ K(a). If / x = π(b). f j x ∈ int(Raj ) and so Aaj aj+1 = 1. f (S) ∩ Rj = ∅ as f (D) = ∅. one has π ◦ σ = f ◦ π. Symbolic Dynamics 57 f (S) ∩ Rj = y∈D (f W u (y. f (D)]. then inductively using 3. As x ∈ Y . f j y ∈ Raj are close for all j ∈ Z and so x = y by expansiveness. Y is residual. π ◦σ = f ◦ π.13. Since Rj = [W u (f x. Rj ) proper. So f (S) ∩ Rj = y ∈f (D) W u (y . As Kn (a) ⊃ Kn+1 (a) ⊃ · · · we have ∞ K(a) = j=−∞ f −j Raj = ∞ Kn (a) = ∅ . Rj )] is proper.14.
though earlier similar statements are in [4] and for Anosov diﬀeomorphisms in [2]. Adler.6 is explicitly proved in [6]. 3. then f j x ∈ Rxj and f j x ∈ int(Raj ) imply xj = aj . R. 1970. As the name implies.58 3 Axiom a Diﬀeomorphisms 3. References The basic references are Anosov [2] for Anosov diﬀeomorphisms and Smale [14] for Axiom A diﬀeomorphisms. This same argument shows that σΣA is mixing if f Ωs is. Proof. σ : ΣA → ΣA is topologically transitive. .6 explicitly for Anosov diﬀeomorphisms. N ]} V ⊃ V1 = {x ∈ ΣA : xi = bi ∀ i ∈ [−N. R. Corollary 3. 1. Then ∅ = π −1 (f n U2 ∩ V2 ) = π −1 (f n U2 ) ∩ π −1 (V2 ) ⊂ f nU ∩ V . Results 3. [11] did Sections 3 and 3 for Anosov diﬀeomorphisms and this was carried over to Axiom A diﬀeomorphisms in [5].8 is due to Anosov. Proposition 3.19. [15] and for their proofs we have followed [6]. Sinai [10].9 and 3. Weiss: Similarity of automorphisms of the torus. V be nonempty open in ΣA . If f Ωs is topologically mixing so is σ : ΣA → ΣA . Now N ∅ = int(KN (a)) = j=−N N f −j int(Raj ) = U2 f −j int(Rbj ) = V2 . Providence. For some a. so π −1 (U2 ) ⊂ U1 . f n U2 ∩ V2 = ∅ for various large n. if x = π(x) ∈ U2 .I. American Mathematical Society. with the mixing part of 3. The idea of pseudoorbit has probably occurred to many people. Proposition. N ]} . j=−N and ∅ = int(KN (b)) = Also. Memoirs of the American Mathematical Society 98. Symbolic dynamics for certain geodesic ﬂows goes back to Hadamard and was developed by Morse [9].7 is in [2] for Anosov diﬀeomorphisms and [3] for Axiom A diﬀeomorphisms. Similarly π −1 (V2 ) ⊂ V1 . The stable manifold theorem for hyperbolic sets is due to Hirsch and Pugh [8]. Sinai [12] stated 3. Let U. It was carried out for the horseshoe by Smale [13] and for automorphisms of the torus by Adler and Weiss [1]. b ∈ ΣA and N one has U ⊃ U1 = {x ∈ ΣA : xi = ai ∀ i ∈ [−N.10 are from [7]. Canonical coordinates and spectral decomposition are from [14]. B.5 from [4]. Since f Ωs is transitive.
Proc. Bowen: Markov partitions for Axiom A diﬀeomorphisms.. Sinai: Construction of Markov partitions. Calif. 6489. Anal. Soc. Trans. Func. Amer.. Proc. 9. pp. Pure Math.. Math. Nauk 27 (1972). 1968. 8. Sinai: Markov partitions and Cdiﬀeomorphisms. 2 (1968). Press. 3. 2. pp.J. J. 333–339. Amer. no.I.. Vol. Math. 11. Smale: The Ωstability theorem. 1970.References 59 2. Proc. and its Appl. R. 33–51. 133–163. 9 1969/1970 121–134. C. Smale: Diﬀeomorphisms with many periodic points. J. Anal. English translation: Russian Math. Berkeley. pp.I. in Honor of Marston Morse. Hirsch and C. Soc. XIV. 21–69. Providence. 21–64. 10. 12. Providence.. Math. R. 90 (1967). 1.. 725747. Vol. 747817. Sympos. XIV. Sympos. 2. Math. Math. Soc. no. 73 (1967). 92 (1970). pp. Amer. R. Amer. Ya.. S. Math. Palis. Providence. XIV. Princeton Univ. Calif. Ya.. Bowen: Topological entropy and Axiom A. 2 (1968). 5. R. J. 1965. Surveys 27 (1972). S. 1. Math. D. Pure Math. M. 7080. and its Appl. Princeton. no. Shub: Neighborhoods of hyperbolic sets. no. 4. 377397. M. N. 1968. 13. Math. Smale: Diﬀerentiable dynamical systems.. J. Berkeley.I. Proc. Pugh: Stable manifolds and hyperbolic sets. Calif. Amer. Pure Math. Func. 14. . 289–298. Sympos. 6. Math. Ya. 23–41.. M. Invent. Pugh and M. R. 7.. 154 (1971). no. A Symp. 6380.. 1970. Soc. Bowen: ωlimit sets for Axiom A diﬀeomorphisms. Amer. R. 15. no. Bull. Steklov Inst. Soc. Bowen: Periodic points and measures for Axiom A diﬀeomorphisms. Berkeley. Hirsch. Anosov: Geodesic ﬂows on closed Riemann manifolds with negative curvature. 4. Amer. R. Sinai: Gibbs measures in ergodic theory. 4. Uspehi Mat. Diﬀerential Equations 18 (1975). 43 (1921). Vol. S. Morse: A OnetoOne Representation of Geodesics on a Surface of Negative Curvature. in “Diﬀerential and Combinatorial Topology”. 1968.
1) for which the following are true: if x ∈ Ωs . y ∈ ΣA have xk = yk for k ∈ [−N. Furthermore µφ is ergodic. N ]. There are ε > 0 and α ∈ (0. then f k π(x) . y) < αN . µφ is Bernoulli if f Ωs is topologically mixing.19 and we have a Gibbs measure µφ∗ as in Chapter 1.D.t. µφ∗ (σ n Ds ) = µφ∗ (Ds ). N ]. Theorem. If x. and d(f k x. Let R be a Markov partition for Ωs of diameter at most ε. Proof of 4. Lemma. using σ n+1 Ds ⊂ σ n Ds one has ⎛ ⎞ µφ∗ ⎝ n≥0 σ n Ds ⎠ = µφ∗ (Ds ) . y ∈ M .2. and σDs ⊂ Ds . As µφ∗ is σinvariant. θ > 0 o so that φ(x) − φ(y) ≤ a d(x. y)θ . each smaller than ΣA . φ∗ (x) − φ∗ (y) ≤ a (αθ )N and φ∗ ∈ F . Proof. This gives d(π(x).4 Ergodic Theory of Axiom a Diﬀeomorphisms A. π(y)) < αN . . See p. f k y) ≤ ε for all k ∈ [−N. Then Ds and Du are closed subsets of ΣA .1. N ] . Equilibrium States for Basic Sets Recall that a function φ is H¨lder continuous if there are constants a. A the transition matrix for R and π : ΣA → Ωs as in 3.r. A First we assume f Ωs is mixing. then d(x. 4. Then σΣA is mixing by 3. Then φ has a unique equilibrium state µφ (w. 140 of [12]. Let Ds = π −1 (∂ s R) and Du = π −1 (∂ u R). Let Ωs be a basic set for an Axiom A diﬀeomorphism f and o φ : Ωs → R H¨lder continuous. σ −1 Du ⊂ Du . f k π(y) ∈ Rxk = Ryk for k ∈ [−N. f Ωs ).1. Let φ∗ = φ ◦ π. 4.
still positive. Then µφ is f invariant and the automorphisms of the measure spaces (σ. (f. Lemma.. one gets µφ∗ (Ds ) = 0. Now µφ is Bernoulli because of 1. Finding µ maximizing hµ (f ) + φ dµ is equivalent therefore to ﬁnding µ maximizing hµ (f m ) + Sm φ dµ . For any µ ∈ Mf (Ωs ) there is a ν ∈ Mσ (ΣA ) with π ∗ ν = µ.13 (note of the editor). Using the variational principle 2. µφ∗ ). In particular hµφ (f ) = hµφ∗ (σ) and so (1 ) hµφ (f ) + φ dµφ = hµφ∗ (σ) + φ∗ dµφ∗ = Pσ (φ∗ ) ≥ Pf (φ) . µφ (E) = µφ∗ (π −1 E). Proof. We have left the case Ωs = X1 ∪· · ·∪Xm with f Xk = Xk+1 and f m X1 mix1 ing.17 (note of the editor). Conversely.3. For φ o H¨lder on Ωs . Suppose µ is any equilibrium state of φ and pick ν ∈ Mσ (ΣA ) with π ∗ ν = µ. Similarly one sees µφ∗ (Du ) = 0. µφ ) are conjugate since π is onetoone except on the null set n∈Z σ n (Ds ∪ Du ). and Sm φ dµ = m φ dµ. if µ ∈ Mf m (X1 ). Then σ ∗ ν = ν and π ∗ ν = µ ∗ k ∗ (using π (σ ) β = (f k )∗ π ∗ β = (f k )∗ µ = µ). F is identiﬁed with some β ∈ M (ΣA ). Hence (2 ) Pσ (φ∗ ) = Pf (φ) and µφ is an equilibrium state for φ by Chapter 2. The inequality comes from Proposition 2. .62 4 Ergodic Theory of Axiom a Diﬀeomorphisms Now n≥0 σ n Ds has measure 0 or 1 as it is σinvariant and µφ∗ is ergodic (see 1. As F (1) = F (1 ◦ π) = 1. Proof of 4. k=0 One checks that µ ↔ µ deﬁnes a bijection Mf (Ωs ) ↔ Mf m (X1 ). hµ (f m ) = mhµ (f ). since its complement (a nonempty open set) has positive measure by 1. This wellknown fact is proved as follows. 1 2 The second equality follows from Theorem 1.25. For µ ∈ Mf (Ωs ).1 (continued). By a modiﬁcation of the HahnBanach Theorem F extends to C (ΣA ). Sm φ will be H¨lder on X1 and therefore one is done since X1 o is a mixing basic set of f m .22. 4. one has µ(X1 ) = m and so µ = m µX1 ∈ Mf m (X1 ). Now let µφ = π ∗ µφ∗ .22.14). F (g ◦ π) = µ(g) = gdµ deﬁnes a positive linear functional on a subspace of C (ΣA ). then µ ∈ Mf (Ωs ) where µ(E) = 1 m m−1 µ (X1 ∩ f k E) . Thus ν is an equilibrium state for φ∗ and ν = µφ∗ by 1. Then µ = π ∗ µφ∗ = µφ . By compactness let 1 ν = limk→∞ nk (β + σ ∗ β + · · · + (σ nk −1 )∗ β).4. Then hν (σ) ≥ hµ (f ) and so hν (σ) + φ∗ dν ≥ hµ (f ) + φ dµ = P (φ) = P (φ∗ ) .e. i.
29). f k x) < ε ∀ k ∈ [0. (iii) There is a constant K so that Sm φ(x) − Sm ψ(x) = Km when x ∈ Ωs with f m x = x. then µφ∗ = µψ∗ and by Theorem 1. If these conditions hold.4 and P (φ∗ ) = P (φ) = P one gets µφ (B) ≥ c1 exp(−P n + Sn φ(x)) . o For small ε > 0 there is a bε > 0 so that. µφ y ∈ Ωs : d(f k y. Letting j → ∞ we get (iii). If (iv) is true. If µφ = µψ . f k x) < ε ∀ k ∈ [0.5. n] ⊃ π y ∈ ΣA : yk = xk ∀ k ∈ [0. Let φ∗ = φ ◦ π and ψ ∗ = ψ ◦ π. We assume f Ωs is mixing and leave the reduction to this case to the reader. Proposition. this gives us (ii). Then B = y ∈ Ωs : d(f k y. Applying 1. (ii) There are constants K and L so that Sm φ(x) − Sm ψ(x) − Km ≤ L for all x ∈ Ωs and all m ≥ 0. n] . Pick x ∈ ΣA with π(x) = x. picking x ∈ π −1 (x). Let φ : Ωs → R be H¨lder continuous and P = Pf Ωs (φ). Proof.28. then φ∗ (x) − ψ ∗ (x) = K + u(π(σx)) − u(π(x)) and µφ∗ = µψ∗ by Theorem 1.1. Assume ﬁrst f Ωs is mixing. Equilibrium States for Basic Sets 63 4. Proposition. Then L ≥ Smj φ(x) − Smj ψ(x) − mjK = jSm φ(x) − Sm ψ(x) − mK . (iv) There is a H¨lder function u : Ωs → R and a constant K so that φ(x) − o ψ(x) = K + u(f x) − u(x). We leave it to the reader to reduce the general case to the mixing one as in the proof of 4. 4. for any x ∈ Ωs and for all n. Assume (ii) and f m x = x. Now we assume (iii) and prove (iv).4. o Then the following are equivalent: (i) µφ = µψ . One then has µφ = π ∗ µφ∗ = π ∗ µψ∗ = µψ . Choose the Markov partition R above to have diam(R) < ε. For x ∈ Ωs .28 there are K and L so that Sm φ∗ (x) − Sm ψ ∗ (x) − Km ≤ L for x ∈ ΣA . Let A = {f k x : k ≥ 0} and deﬁne u : A → R by . K = P (φ) − P (ψ). Proof. Let φ.A. n] ≥ bε exp(−P n + Sn φ(x)) . Let η(x) = φ(x) − ψ(x) − K and pick x ∈ Ωs with dense forward orbit (Lemma 1. ψ : Ωs → R be two H¨lder continuous functions.
n + N ] . The Case φ = φ(u) Recall that M has a Riemannian structure and this induces a volume measure m on M . u(z) − u(y) = Sn η(y) − Sn η(y ) n−1 ≤ j=0 η(f j y) − η(f j y ) . n = m − k. y) < δ. as Sn η(y ) = 0.2 gives d(f j y. f j y ) < αmin{j+N.7 there is δ > 0 so that if y ∈ Ωs and d(f n y. Because η is H¨lder.N +n−j} . . y) < δ. f k y ) < 2 for all k ∈ [0. Suppose y = f k x. Providing N is large one has z = f n y. n) Lemma 4. z) ∈ [ε/2R .64 4 Ergodic Theory of Axiom a Diﬀeomorphisms k−1 u(f x) = j=0 k η(f j x) . By the choice of R and y one sees that d(f j y. For x ∈ Ωs let φ(u) (x) = − log λ(x) where λ(x) is the Jacobian of the linear map u u Df : Ex → Ef x using inner products derived from the Riemannian metric. Then. and d(f n y. f j y ) < ε for all j ∈ [−N. Let R be the maximum ratio that f expands any distance. For z ∈ A one has u(f z) − u(z) = η(z). For j ∈ [0. z)γ . By 3. o o The formula η(z) = u(f z) − u(z) extends to Ωs by continuity. We will assume for the remainder of this chapter that f : M → M is a C 2 Axiom A diﬀeomorphism and Ωs is a basic set for f .N +n−j} ∞ u(y) − u(z) ≤ 2a r=N N +1 αθr ≤ a αθN . B. Pick ε and α as in 4. Then ﬁnd y as above with f n y = y . n]. z = f m x with k < m and d(y.2. ε/2RN ]. then there is a y ∈ Ωs with ε f n y = y and d(f k y. ¯ Thus u is H¨lder on A and extends uniquely to a H¨lder function on A = Ωs . Taking γ > 0 so that (1/R)γ ≥ αθ one has u(y) − u(z) ≤ a d(y. Pick N so that d(y. z) < /2RN . o η(f j y) − η(f j y ) ≤ a αθmin{j+N.
ε)separated. n Pf Ωs (φ(u) ) = lim s (b) Let Wε (Ωs ) = x∈Ωs n→∞ s s Wε (x).B.1 the function φ(u) has a unique equilibrium state which we denote µ+ = µφ(u) . Then Bx (ε. f k y) ≤ ε for all k ∈ [0. If Ωs is a C 2 basic set. n)) ≤ 0 . m) = y ∈ M : d(f k x. m(B(ε. so the composition x → φ (x) is H¨lder. The map x → Ex is H¨lder (see 6. n) for some y ∈ En (δ). Proof. If m(Wε (Ωs )) > 0. By (δ/2. m) .5 one sees that the measure µ+ on Ωs and P (φ(u) ) do not depend on which metric is used.4 of [12]) and Ex → φ(u) (x) is (u) o diﬀerentiable. n)) ≤ Cδ+ε y∈En (δ) y∈En (δ) exp(Sn φ(u) (y)). Volume Lemma. n) ⊂ By (δ + ε. ε)separated subsets of Ωs . (a) Letting B(ε.7 ( ) For δ ≤ ε. n) = x∈Ωs Bx (ε. n) with d(f k y. Let Bx (ε. Call E ⊂ M (n. m)) ∈ [Cε . Choose En (δ) maximal among the (n.8. n) is a disjoint union and so . Proof. z are two distinct points in E. Cε ] exp(Sm φ(u) (x)) for all x ∈ Ωs .6. n). one has (for small ε > 0) 1 log m(B(ε. one can ﬁnd k ∈ [0. By Theorem 4. 4. ε)separated if whenever y. While φ(u) depends on the metric used. 4. u u o Proof. Proposition. B(ε. Lemma. n) ⊂ y∈En (δ) By (δ + ε. For x ∈ Ωs one has x ∈ By (ε. f k z) > δ. The Case φ = φ(u) 65 4. Let Ωs be a C 2 basic set. then there is a constant Cε so that −1 m(Bx (ε. otherwise En (δ) ∪ {x} is (n.7. If x ∈ Ωs is a C 2 basic set and ε > 0 is small. n) and by 4. when f m x = x u u Sm φ(u) (x) = − log Jac(Df m : Ex → Ex ) does not depend on the metric (this Jacobian is the absolute value of the determinant). See 4. n). By 4. then φ(u) : Ωs → R is H¨lder continuo ous.2 of [9]. n) ⊂ B(ε. then Pf Ωs (φ(u) ) = 0 and hµ+ (f ) = − φ(u) dµ+ .
Let Γ ⊂ Un cover Ωs . Then Sn φ(u) (Uy ) ≥ Sn φ(u) (y). If Uy = Uy . Then for j ∈ [0. For each y ∈ En (δ) one can pick Uy ∈ Un with By (δ. then d(f k y. n)) . n) d(f j x. Fix δ ≤ ε and let U be an open cover of Ωs with diam(U) < δ. n)) ≥ Cδ/2 y∈En (δ) ( ) exp(Sn φ(u) (y)). n) ∩ Ωs ⊂ X(Uy ). we have φ(u) (x) − φ(u) (y) ≤ a d(x. y ∈ Ωs . δ)separated. y)θ for some a. n→∞ n Letting diam(U) → 0. U) ≤ Uy ∈Γ n exp(Sn φ(u) (Uy )) exp(Sn φ(u) (y)) .2. U) = lim n→∞ 1 log inf Γ n exp(Sn φ(u) (U)) U∈Γ 1 ≥ lim sup log m(B(ε. n) with y ∈ En (δ). Now let U be an open cover and let δ be a Lebesgue number for U. one replaces P (φ(u) . f k y ) ≤ diam(U) < δ and y = y as En (δ) is (n. Thus exp(Sn φ(u) (U)) ≥ U∈Γ y∈En (δ) exp(Sn φ(u) (y)). Let Γ = {Uy : y ∈ En (δ)}. f j y) < αmin{j. Hence n−1 Sn φ(x) − Sn φ(y) ≤ j=0 φ(u) (f j x) − φ(u) (f j y) ∞ ≤ 2a k=0 αkθ = γ . Using this together with ( ) above one gets P (φ(u) .66 4 Ergodic Theory of Axiom a Diﬀeomorphisms −1 m(B(ε. y∈En (δ) ≤ eγ . Suppose x ∈ By (ε. n) ∩ Ωs . θ > 0 and all x. U) with P (φ(u) ). For each y ∈ En (δ) pick Uy ∈ Γ with y ∈ X(Uy ). Also Sn φ(u) (Uy ) ≤ Sn φ(u) (y) + γ and so Zn (φ(u) . o Since φ(u) is H¨lder. Then Γ covers Ωs since every x ∈ Ωs lies in some By (δ.n−j−1} by Lemma 4.
then Ux = s u {Wε (y) : y ∈ Wε (x)} is a neighborhood of x in M (see Lemma 4. Lemma. By Theorem 3. f n y) < ε and d(f −n z. If Ωs is not an attractor there exists γ > 0 such that u for every x ∈ Ωs . Then also u mk u W (p) = k≥0 f Wβ (p) ⊂ Ωs . By compactness Vγ = Ωs for some γ > 0. Let Ωs be a C 1 basic set. n)) ≥ m(Wε (Ωs )) as Wε (Ωs ) ⊂ B(ε. s s s Now m(B(ε. To prove (b) notice that the set u Vγ = {x ∈ Ωs : d(y. hµ+ (f ) + φ(u) dµ+ = P (φ(u) ) = 0 .9. If Wε (Ωs ) ⊂ Ωs .B. Wε (x) depend continuously on x ∈ Ωs . Since µ is an equilibrium state for φ(u) . A basic set Ωs is an attractor if it has small neighborhood U with f (U ) ⊂ s U . (u) + the formula from (a) yields P (φ ) = 0. Lemma 4.1 of [11]). Then BΩs (δ) ⊂ s {Ux : x ∈ Y } ⊂ Wε (Ωs ) and Ωs is an attractor. u 4. Choose a periodic point u u p ∈ Ux . The Case φ = φ(u) 67 Using ( ) we get P (φ(u) . U) n 1 ≤ lim inf log m(B(ε. For some small β one has Wβ (p) ⊂ Ux . If m(Wε (Ωs )) > 0. Since Wε (x). n). Also Vγ increases when γ decreases and γ>0 Vγ = Ωs by statement (a).1]). f −n p) < β u for n ≥ 0. the proof of (a) is complete. then Ωs is an attractor. Ωs ) > γ for some y ∈ Wε (x)} u is open because Wε (x) varies continuously with x.9 one has z ∈ Ωs and Wβ (p) ⊂ Ωs . then d(f n z. n)) . For each N u x ∈ k=0 f k W u (p) = Y one has Wε (x) ⊂ Ωs and so Ux as deﬁned above s u is a neighborhood of x in M . n)) ≤ m(M ). say f m x = x.10 this is equivalent to Wε (Ωs ) being a neighborhood of Ωs . By Proposition 3. there is y ∈ Wε (x) with d(y. Ωs ) > γ. u Proof. If Wε (x) ⊂ Ωs for some x ∈ Ωs . n→∞ n n→∞ Noting that m(B(ε. one can ﬁnd a δ > 0 independent of x so that Ux contains the 2δball Bx (2δ) about x in M for all x ∈ Y (see [11. if z ∈ Wβ (p) s u lies in Wε (y) (y ∈ Wε (Ωs )). . U) = lim 1 log Zn (φ(u) . Now Xp = W u (p) and Ωs = Xp ∪ f Xp ∪ · · · ∪ f N Xp for some N .
. N Remark. n + N ) = ∅. n)) whenever x ∈ Ωs and y ∈ Bx (ε. Let Ωs be a C 2 basic set. Hence By(x.n) (δ. setting d = d(3γ/2. δ) m(B(γ/2. (c) follows from (b) by Proposition 4.n) (δ. Then By(x. γ/4) so that d(f z. (b) m(W s (Ωs )) > 0. Let E ⊂ Ωs be (γ. (c) Pf Ωs (φ(u) ) = 0. n)) x∈E ≥ d(3γ/2. Second Volume Lemma.3 of [9]. n + N )) ≤ (1 − d) m(B(γ/2. n). n)) ≥ d m(Bx (ε. δ) ≥ d(3γ/2. Proof. n)) and by Proposition 4.8 (a) Pf Ωs (φ(u) ) ≤ 1 log(1 − d) < 0 . δ) m(B(γ/2. Using the Second Volume Lemma m(B(γ/2.n) (δ. n) with d(f n+N y(x. Choose N N δ ∈ (0. n)) .9. n) ∈ Bx (γ/4. then (b) is true since Wε (Ωs ) is a neighborhood of Ωs .10. n). n)separated. δ) > 0 so that m(By (δ. Given a small ε > 0 choose γ as in 4. δ > 0 there is a d = d(ε. Let Ωs be a C 2 basic set. 4. It is possible to ﬁnd a C 1 basic set (a horseshoe) which is not an attractor but nevertheless has m(W s (Ωs )) > 0 [8]. For small ε.n) (δ. y) < δ. n) ∩ BΩs (γ/2) = ∅ .8 (b). s If Ωs is an attractor. For x ∈ E there is a y(x.68 4 Ergodic Theory of Axiom a Diﬀeomorphisms 4. n) ∩ B(γ/2. The following are equivalent: (a) Ωs is an attractor. n + N )) ≥ x∈E m(By(x. Pick N so that u u f N Wγ/4 (x) ⊃ Wε (f N x) ∞ for all x ∈ Ωs . n)) − m(B(γ/2. n)) m(Bx (3γ/2. n). f n+N By(x. See 4. f y) < γ/2 whenever d(z.11. Therefore. (b) is equivalent to m(Wε (Ωs )) > 0. As W s (Ωs ) = n=0 f −n Wε (Ωs ). Ωs ) > γ u u u (since f n Bx (γ/4. n) ⊂ Bx (γ/2. Theorem. We ﬁnish by assuming Ωs is not an attractor and showing P (φ(u) ) < 0. n) ⊃ Wγ/4 (f n x) and f N Wγ/4 (f n x) ⊃ Wε (f N +n x)). s s Proof.
δ) = {x ∈ M : g(n. n) = ∅ for x. δ) = {x ∈ M : g(n. This leads us next to the following result. 2δ) by the choice of ε. y) < ε. 3δ) ≤ c2ε k=N x∈Rk ∞ exp(Sk φ(u) (x)).C. By the maximality of Rn one has Bz (ε.12. x) − g > δ} E(g. k) = ∅ for some x ∈ Rk .11 implies that malmost all x ∈ M approach an attractor under the action of a C 2 Axiom A diﬀeomorphism f . 3δ) ⊂ k=N x∈Rk Bx (2ε. Attractors and Anosov Diﬀeomorphisms 69 C. k) = ∅ for x ∈ Rn . . Let Ωs be a C 2 attractor. 2δ) satisfying the conditions: (a) Bx (ε. k) ⊂ Ck (g. x = x . k) and so ∞ s Wε (Ωs ) ∩ n=N ∞ Cn (g.. n) ∩ Bx (ε. Let Rn (n ≥ N ) be a maximal subset of Ωs ∩ Cn (g. . Now ﬁx N > 0 and choose sets RN . then z ∈ Cn (g. RN +1 . The deﬁnition of Rn shows that VN = k=N x∈Rk Bx (ε. n) ∩ Bx (ε. Attractors and Anosov Diﬀeomorphisms Because M = k=1 W s (Ωk ). n) ⊂ Bz (ε. Theorem. x) = deﬁne the sets 1 n n−1 k=0 g(f k x) and g = g dµ+ . x) − g > δ for inﬁnitely many n} ∞ ∞ = N =1 n=N Cn (g. y ∈ Rk . s s If y ∈ Wε (Ωs ) ∩ Cn (g. . The choice of ε gives Bx (ε. Then y ∈ Bz (ε.7 one gets ∞ s m Wε (Ωs ) ∩ n=N ∞ Cn (g. δ) . n) ∩ By (ε. Using the Volume Lemma 4. N ≤ k ≤ n . 4. k) is a disjoint union. x ∈ Rn . δ) for x ∈ Rk ⊂ Ck (g. Choose ε > 0 so that g(x) − g(y) < δ when d(x. Let us write g(n.e. 3δ) (n ≥ N ) and y ∈ Wε (z) with z ∈ Ωs . k) ⊂ Bx (2ε. Proof. N ≤ k < n. successively as follows. Fix δ > 0 and Cn (g. 2δ) and so . x is a generic point for µ+ ). k) . Theorem 4. (b) Bx (ε. For malmost all x ∈ W s (Ωs ) one has n−1 1 lim g(f k x) = g dµ+ n→∞ n k=0 r for all continuous g : M → R (i.
the sum on the right approaches 0 as N → ∞. Proof. k=1 ∞ for x ∈ W s (Ωs ) k=1 A(gk ) one gets that limn g(n. Since the measure µ+ is ergodic. 3δ) ∩ Wε (Ωs )) has measure 0 since f preserves measure (w. 3δ) = 0. Let g ∈ C (M ). Corollary. gdµ+ for m. s This in turn implies m(Wε (Ωs ) ∩ E(g. the Ergodic Theorem ∞ 0 = µ (E(g. Let {gk }∞ be a dense subset of C (M ). δ).r.t. 3δ)) = 0. Let A be the set of x ∈ M with 1 lim n→∞ n Because m(A) = 1 and µ µalmost all x. m(E(g. Thus ∞ s m E(g. Then n−1 g(f k x) = k=0 gdµ+ . s s For δ > 3δ the set E(g. It follows that g ∗ (x) = . 4. then µ = µ+ . As µ (VN ) → 0.70 4 Ergodic Theory of Axiom a Diﬀeomorphisms ∞ k=N VN ⊂ implies Ck (g. δ ) ∩ f −n Wε (Ωs ) ⊂ f −n (E(g. x) = g for all x ∈ s W (Ωs ) outside an mnull set A(g). m).4 ∞ µ (VN ) ≥ bε + k=M x∈Rk + exp(Sk φ(u) (x)). µ(A) = 1. x) = g for all g ∈ C (M ). If f leaves invariant a probability measure µ which is absolutely continuous with respect to m. δ ) ∩ f −n Wε (Ωs ) = 0 . δ) and thus limN →∞ µ+ (VN ) = 0.13. Suppose f : M → M is a transitive C 2 Anosov diﬀeomorphism.8 (b) one has Pf Ωs (φ(u) ) = 0 and then by 4. δ)) = lim µ n→∞ + + n=N Cn (g. By the Ergodic Theorem there is a function g ∗ : M → R so that 1 n→∞ n lim n−1 g(f k x) = g ∗ (x) k=0 for µalmost all x. By 4. δ ) ∩ W s (Ωs )) ≤ n=0 1 Fixing g still but letting δ = m → 0 one gets limn g(n. Using the inequality of the preceding paragraph one sees ∞ N →∞ s lim m Wε (Ωs ) ∩ n=N Cn (g. In this case M = Ω = Ω1 is the spectral decomposition.
Also s s λf −1 (x) = Jacobian Df −1 : Ex → Ef −1 x (u) = λ(s) (x)−1 and so φf −1 (x) = − log λf −1 (x) = φ(s) (x). Notice that equilibrium states w.14.f .r. Assume (b) holds. Remark. Theorem. Let f be a transitive C 2 Anosov diﬀeomorphism. Clearly (a) implies (b). Now assume (c) and let φ(x) = log Jac (Df : Tx M → Tf x M ). µ = µ+ . There is an invariant measure µ− so that n−1 1 lim g(f −k x) = g dµ− n→∞ n (u) (u) k=0 for malmost all x. So µ above is Bernoulli.t. The following are equivalent: (a) f admits an invariant measure of the form dµ = hdm with h a positive H¨lder function. 4. (c) Df n : Tx M → Tx M has determinant 1 whenever f n x = x. Proof. P (φ(u) ) = 0 = P (φ(s) ). then M = Ω1 = X1 and f is mixing. f . f −1 .C.r. Now f −1 is Anosov with u s s u Ex.5 one has.f −1 = Ex. n−1 n−1 φ(u) (f k x) − k=0 k=0 φ(s) (f k x) = 0. m.t. As this holds for all g ∈ C (M ). µ− is the unique equilibrium state for φf −1 w.f an Ex.t. So µ− = µφ(s) .13 to both f and f −1 we see (u) µφ(u) = µ+ = µ = µ− = µφ(s) . f −1 are the same as those w.r. Then . Exponentiating. By 4. for f n x = x.f −1 = Ex. Applying 4. for Mf (M ) = Mf −1 (M ) and hν (f ) = hν (f −1 ). If M is connected.r. Attractors and Anosov Diﬀeomorphisms 71 gdµ = g ∗ dµ = gdµ+ . u s 1 = (det Df n Ex ) (det Df n Ex ) = (det Df n Tx M ) .t.8 (b). o (b) f admits an invariant measure µ absolutely continuous w. By 4. Let λ(s) (x) be the Jacobian s s of Df : Ef −1 x → Ex and φ(s) (x) = log λ(s) (x).
see [18] for the Anosov case. see for instance [2] or [3]. Actually h above will be C 1 . We use condition 4. It was Ruelle who carried Sinai’s work on Anosov diﬀeomorphisms over to Axiom A attractors [11] and brought in the formalism of equilibrium states [15]. Ruelle [16] along these m has support in a neighborhood lines also proved that f n µ → µ+ when µ of an attractor. For f1 near f . The ergodic m has been studied in many theory of Anosov diﬀeomorphisms with µ+ papers. m. By Proposition 4. 14].72 4 Ergodic Theory of Axiom a Diﬀeomorphisms n−1 Sn φ(x) = log k=0 Jac (Df : Tf k x M → Tf k+1 x M ) = log Jac (Df n : Tx M → Tf k x M ) = 0 when f k x = x. Then det(Df1 Tx1 M ) will be near det(Df n Tx M ) and not equal to one. Results 4. 14.14 (c). See [13. In the case where f admits no invariant µ dissipative [10]. For hyperbolic automorphisms of the 2torus µ+ is Haar measure and construction . Remark.1 is from [6. So µ is f invariant.12 is due to Ruelle [16] (we followed the proof in [9]). The notes as a whole constitute a version of Sinai’s program [18] for applying statistical mechanics to diﬀeomorphisms. Thinking of µ = hdm as the absolute value of a form f ∗ (hdm)(f x) = h(x) eφ(x) dm(f x) = h(f x) dm(f x) = (h dm)(f x) . Let h(x) = eu(x) . 4. Theorem 4.15.5.13. 7]. That µ+ is Bernoulli in the transitive Anosov case is due to Azencott [4]. 4. the measure m is actually References This chapter contains the main theorems of these notes. if f n x = x with det(Df n Tx M ) = 1.5 (with ψ = 0. page 36. f1 will be Anosov and have a periodic point n n x1 near x with f1 x1 = x1 . Corollary. this condition is destroyed by the right small perturbation of f .14 and 4. for the Anosov case this result is due to Sinai [17]. Among the C 2 Anosov diﬀeomorphisms the ones that admit no invariant measures µ m are open and dense. Section 4 is taken verbatim from [9]. [18]. K = 0) there is a H¨lder o u : M → R with φ(x) = u(f x) − u(x). We are using stability theory not covered in these notes. Proof. On the other hand. 18]. 4.15 came from [13. Theorem 4. In case φ(u) is a constant function µ+ is the unique invariant measure which maximizes entropy (φ = 0 and φ(u) have the same equilibrium state). Suppose f n x = x and det(Df n Tx M ) = 1.
R. 133–163. Anal. D.I. Ya. 11 (1970). Sympos. Ya. 11. A. 90 (1967). D. 4. S´r. R. no. 4. Soc. Hirsch. 13. 570–573. Palis. 4. Bowen: Bernoulli equilibrium states for Axiom A diﬀeomorphisms. Ya. Oseledec: Gibbs distributions. 17. M. Steklov Inst. no. C. 1. Ruelle: The ergodic theory of Axiom A ﬂows. 14 (1973). Paris. R. no. Math. 181202. 377397. Uspehi Mat. Bowen: Periodic points and measures for Axiom A diﬀeomorphisms.References 73 in 4. B. no. Azencott: Diﬀ´omorphismes d’Anosov et sch´mas de Bernoulli. 193202. no. Math.. Trans. e e Sci. Bowen: Some systems with unique equilibrium states. K. R. Providence. R. Soc. M. Math. 9. 4. 21–64. Providence. This paper is very important in the development of the subject and is good reading. Adler. 5 (137). 29 (1975). B. 2. and its Appl. e 5. American Mathematical Society. Dokl. J. R.I. Invent. USSRIzv. 16561659. D. When φ(u) = C the measure µ which maximizes entropy still has the following geometrical signiﬁcance: the periodic points of Ωs are equidistributed with respect to µ [5]. 7. 107–172. Sinai: Certain smooth ergodic systems. Nauk 22 (1967) no. R. Surveys 22 (1967).. R. J. Soc.3. no. English translation: Russian Math. Pugh and M. R. Gureviˇ. 3. Amer. no. Nauk 27 (1972). 21–69.. Math. Math. Sigmund [19] studied the generic properties of measures on Ωs .1 is due to Adler and Weiss [1]. Systems Theory 8 (1974/75). Berkeley. 99109. Bowen. Bowen: A horseshoe with positive measure. Soviet Math. 14. 8. 619654. Math. R. Math. Sinai: Invariant measures that are compatible with smoothness s for transitive Csystems. Memoirs of the American Mathematical Society 98. Amer. Ruelle: A measure associated with Axiom A attractors. 103167. 237251. 15. Anosov. 289–294. D. pp. no. Soviet Math. Vol. Livˇic. Invent. Sinai: Markov partitions and Cdiﬀeomorphisms. Amer. Invent. Livˇic: Cohomology of dynamical systems. 98 (1976). 1968. Sinai: Gibbs measures in ergodic theory. Ruelle: Statistical mechanics on a compact set with Zν action satisfying expansiveness and speciﬁcation. 6489. Calif. 121–134. Sigmund: Generic properties of invariant measures for Axiom A diﬀeomorphisms. N. . Func. Math. 154 (1971). 19. 1. 13 (1972). Hirsch and C. XIV. 1970. Acad. 2 (1968). and the dissipativity of c Cdiﬀeomorphisms. Dokl. V. 3. 10. Pugh: Stable manifolds and hyperbolic sets. C. 29 (1975). Amer. 1970. AB 270 (1970). 6 (1972). 9 (1969/1970). Proc. A1105A1107. Proc. A. English translation: Russ. Math. Weiss: Similarity of automorphisms of the torus. 203204. Ya. 3. Surveys 27 (1972). Math. Systems Theory 8 (1974/75). N. 5. Math. Shub: Neighborhoods of hyperbolic sets. Pure Math. 18. 16. no. K. Trans. Invent. 6. Math.. 12. 3. Anosov: Geodesic ﬂows on closed Riemann manifolds with negative curvature. Math. 185 (1973). s 12781301. D. Uspehi Mat.
33. 7. 67–69 Axiom A. 6. 49 Rectangle. 19. 65 Shift. 45 Markov partition. 48 Canonical coordinates. 40 Weak Bernoulli. 46 Basic set. 56 Variational principle. 18. 72 Attractor. 29. 43 Equilibrium state. 47 Central limit theorem. 6 Separated. 14 Exponential clustering. 4 Gibbs measure. 71. 45 Pressure. 61 Ergodic. 14 Nonwandering. 46. 15 Homologous. 22 . 5. 24 Entropy. 6. 34 Pseudoorbit. 16. 7 Free energy. 9 Transition matrix. 51 ΣA . 23 FA . 7 Hyperbolic. 42. 18.Index Anosov. 4. 52 Mixing.
Grüne. Italy 2001. R. Second Order PDE’s in Finite and Inﬁnite Dimension (2001) Vol. Berthé. Abe. 1799: H.M. Substitution in Dynamics. 1782: C. M. Tian. Kupeli. Ang. Burstall. The Decomposition of Primes in Torsion Point Fields (2001) Vol.). Loday. C. Martina Franca 1999. Editors: P. Lau. Runde. Bernard (2002) Vol. Xiao. T. M. N. Dozzi. 1767: J. S. Frobenius and Separable Functors for Generalized Module Categories and Nonlinear Equations (2002) Vol. Hervé. S. Marmi. T. Kopfermann. Analytic Capacity. R. 1772: F. Muzychuk. Goichot.C. 1789: Y. Mauduit. A. Menger Curvature and the Cauchy Integral (2002) Vol. M. 1765: T. 1770: H. Filipovi´ . 1795: H. Ros. Bolthausen. V. S. J. Analytic and Geometric Study of Stratiﬁed Spaces (2001) Vol. 1788: A. F. M. NonSemisimple Topological Quantum Field Theories for 3Manifolds with Corners (2001) Vol. Ledoux. 1761: C. L. 1779: I. Li. M. Dynamical Systems and Small Divisors. F. J. Ivanoff. A. 1793: J. Pajot.H. Monod. Toroidal Groups (2001) Vol. Osserman Manifolds in SemiRiemannian Geometry (2002) Vol. Chu. Leschke. Azéma.). Pinkall. Monotone Random Systems (2002) Vol. Ramero. Yoccoz. M.C. J. Monomialization of Morphisms from 3Folds to Surfaces (2002) Vol. Behrend. Yor (Eds. Control and Numerical Aspects of Nonholonomic Systems (2002) Vol. Zhan. Manis Valuations and Prüfer Extensions I: A new Chapter in Commutative Algebra (2002) Vol. 1787: S. Topics in Spatial Stochastic Processes. E. Chapoton. D. D. 1777: E.Lecture Notes in Mathematics For information about earlier volumes please contact your bookseller or Springer LNM Online archive: springerlink. Le. Editor: G. 1756: P. 1791: M. Arithmetics and Combinatorics. Trong. Ferus. de Bartolomeis. K.M. Bruinier. G. Oliva. Dalang. Capasso. Matrix Inequalities (2002) Vol.com Vol. 1764: A. 1760: D. FilteredGraded Transfer in Using Noncommutative Gröbner Bases (2002) Vol. Holomorphic Q Classes (2001) Vol. Sommerhäuser. 1800: O. Quantum Comohology. R. H. S. Zhang. Caenepeel. 1774: V. 1769: M. AlberichCarramiñana. Phelps. Rectiﬁability. D. Cetraro 1997. 1766: H. Knebusch. van der Vaart.). Séminaire de Probabilités XXXVI (2003) Vol. 1758: N. B. Pytheas Fogg. 1785: J. Perkins. Consistency Problems for Heathc JarrowMorton Interest Rate Models (2001) Vol. Editors: S. Linear DelayDifferential Systems with Commensurate Delays: An Algebraic Approach (2002) Vol. Émery. E. H. 1757: R. Moduli of Families of Curves for Conformal and Quasiconformal Mappings (2002) Vol. Pirola (2002) Vol. Borcherds Products on O(2. Séminaire de Probabilités 19671980. 1783: L. Arad. Ecole d’ Eté de Probabilités de SaintFlour XXIX1999. Almost Ring Theory (2003) Vol. Merzbach. D. Pﬂaum. Korteweg de Vries and Nonlinear Schrödinger Equations: Qualitative Theory (2001) Vol. 1776: K. 1801: J. GluesingLuerssen. Tarasov. Lyubashenko. G. F. Pedit. 1786: S. P. Ledoux. 1780: J. D. 1784: L. 1802: V. 1781: E. Zhu. Geometric. Limit Theorems for Markov Chains and Stochastic Properties of Dynamical Systems by QuasiCompactness (2001) Vol. Gomez. Geometry of the Plane Cremona Maps (2002) Vol. Merzbach (2003) . Gorenﬂo. 1775: W. 1778: H. 1790: X. K. V. Schmidt.1) and Chern Classes of Heegner Divisors (2002) Vol. 1798: W. Séminaire de Probabilités XXXV (2001) Vol. Arias de Reyna. Eliasson. E. Cannas da Silva. B. Yor (Eds. Cutkosky. 1762: S. M. Martina Franca. Azéma.L. Melenk. 1797: B. U. Chueshov. Yoccoz (2002) Vol. A. B. GarcíaRío. Conformal Geometry of Surfaces in S4 (2002) Vol. Hennion. Gabber. Adelmann. Ferenczi. Dubrovin. 1768: M. H. L. Kerler. Rosenberg. Characters and Cyclotomic Fields in Finite Geometry (2002) Vol. M. A Selection in Martingale Theory (2002) Vol. V. VázquezLorenzo. A. D. Editor: P. Militaru. Marmi. Lectures on Symplectic Geometry (2001) Vol. K. Standard Integral Table Algebras Generated by a Nonreal Element of Small Degree (2002) Vol. Lectures on Choquet’s Theorem (2001) Vol.M. Cetraro. Moment Theory and Some Inverse Problems in Potential Theory and Heat Conduction (2002) Vol. 1773: Z. Cerrai. Geometric Mechanics (2002) Vol. 1771: M. 1759: Y. Zhidkov. Émery. 1763: J. Dialgebras and Related Operads (2001) Vol. Siegel (2002) Vol. Cortés Monforte. D. Harmonic Functions on Groups and Fourier Algebras (2002) Vol. 1796: J. H. Vasil’ev. Émery. M. Mountford. 1794: N. Pointwise Convergence of Fourier Series (2002) Vol. Lectures on Amenability (2002) Vol. Yor (Eds. Kuksin. The Global Theory of Minimal Surfaces in Flat Spaces. hpFinite Element Methods for Singular Perturbations (2002) Vol. 1755: J. YetterDrinfel’d Hopf algebras over groups of prime order (2002) Vol. A. G. C. R. Kiechle. Editor: E. Continuous Bounded Cohomology of Locally Compact Groups (2001) Vol. Reina (2002) Vol. Meeks. Asymptotic Behavior of Dynamical and Control Systems under Perturbation and Discretization (2002) Vol. 1792: D. Italy 1998. Frabetti. C. Lectures on Probability Theory and Statistics. V. Theory of KLoops (2002) Vol. Editors: V.
Doplicher. Italy 2000. Milman.D. Noncommutative Geometry. Siburg. Yomdin. 1852: A. Cetraro. Editors: V. Tumanov. VIII. 2004.Vol. Qian. Cohen. Vol. 118 p. Martina Franca. Solonnikov. Italy 2002. Caffarelli. Reichel. 1835: O. Inequalities and Identities of InclusionExclusion Type. M.Q. Talagrand. Geometric Curve Evolution and Image Processing (2003) Vol. Tsirelson. 1828: D. Editors: P. 1823: A. Bielecki. 1853: Ch. Vol. 1816: S. Björk. 254 p. 1848: M. Funchal. G. Nussbaum. 256 p. P. Bildhauer. 113 p. Föllmer. 1833: D. Variational Methods for Crystalline Microstructure – Analysis and Computation (2003) Vol. M. Ganesh. Spectral Properties of Noncommuting Operators (2004) Vol. Lusztig. A. J. N.). Kaminker. X. Barbasch (2003) Vol. Editors: S. 1847: T. Séminaire de Probabilités XXXVII (2003) Vol. Ammari. M. Russia 2001 (2003) Vol. A. Ya. The Principle of Least Action in Geometry and Dynamics (2004) Vol. A. Mathematical Problems in Semiconductor Physics. Connes. Schachermayer. W. F. Fornaess. P. E. Vishik. R. 1819: D. Bank. 1810: J. Favre. 1840: B. Navarro González. Hoveijn. Lunther. A. Lectures on Probability Theory and Statistics. I. 1814: P. Guentner. VIII. Israel Seminar 20002002 (2003) Vol. N. IX. Cetraro. Operator Functions and Localization of Spectra. M. Derived Functors in Functional Analysis (2003) Vol. L. Topics in Orbit Equivalence (2004) Vol. Dolzmann. MalletParet. Karpenko. 170 p.J. M. Ledoux. D. Tavaré. Convex Variational Problems with Linear. 1841: W. Kahn. Xiong. Johnson. H.R. T. 1851: O. Masser. XII. 2003. Abate. St. Computing Singularities by Gröbner Bases (2003) Vol. Rosay. Dohmen. Vol. Uniqueness Theorems for Variational Problems by the Method of Transformation Groups (2004) Vol. Editors: D. Soner. Deformations of Singularities (2003) Vol. Orthogonal Polynomials and Special Functions. 1829: E. M. G. Dziuk. W. Portugal 2000. XII. 1817: E. Ecole d’Eté de Probabilités de SaintFlour XXXII2002. R. W. Martina Franca.A. Amoroso. 280 p. Van Assche (Eds. J. 1807: V. Reconstruction of Small Inhomogeneities from Boundary Measurements (2004) Vol. Summer School. Martina Franca. I. Steinbach. Deckelnick. Ecole d’Eté de Probabilités de SaintFlour XXX2000. Vol. V. Noncommutative Stationary Processes. Vol. Rodrigues (2003) Vol. 1812: L. G. M. Bifurcations in Hamiltonian Systems. Canuto (2003) Vol. B. Torus Bundles. Improved Bonferroni Inequalities via Abstract Tubes. Stevens. N. H. Anile (2003) Vol. Zampieri (2004) Vol.A. O’Connell.T. Israel Seminar 20022003 (2004) Vol. Émery. Ecole d’Eté de Probabilités de SaintFlour XXXI2001. 1806: H. 1845: Min Ho Lee. Italy 2002.G. Cherednik.). Tame Geometry with Application in Smooth Analysis. E. Measures with Symmetry Properties (2003) Vol. VIII. C. Jefferies. 1834: Yo. Rogers. Conti. Pilgrim. Wischik. R. Editors: J. F. Jonsson. Vol. Huang. M.F. 2003. J. XIV. Dynamical Systems. A. Milman. 1832: J. Leuven 2002 (2003) Vol. Kechris. Bernard (2003) Vol. B. Picard (2004) Vol. Zaitsev. Hiai. H. M. IX. nearly Linear and/or Anisotropic Growth Conditions (2003) Vol. Villani. J. A. G. Howe. Ambrosio. Koelink. H. Geometric Methods in the Algebraic Theory of Quadratic Forms. 2000. D. 1831: A. Brenier. Ringhofer.P. 1811: J. Wengenroth. M. A. 1846: H.J. G. 1830: M. M. Optimal Transportation and its Applications. Allegretto. Asymptotic Combinatorics with Applications to Mathematical Physics. Editor: J. Caffarelli. 1821: S. Schechtman (Eds. Jeanblanc. Italy 2001. Broer. Big Queues. Anile. A. Longo (2004) Vol. 2004. Cao. Izhboldin. Comte. Editor: C. Qian. XIV. 1843: B. ParisPrinceton Lectures on Mathematical Finance 2003 (2004) Vol. 1836: C. 1842: T. Nesterenko. M. 2003. 1827: K. Picard (2004) Vol. Combinations of Complex Dynamical Systems. Vershik (Ed. Editor: J. 304 p. W. Italy 2000. 1804: I. Touzi. M. Rutkowski. Diophantine Approximation. Multiscale Problems and Methods in Numerical Simulations. G. 1825: J. Vegter. Azéma. W. S. 1803: G. Buttazzo. F. V. Schlickewei. 186 p. Albeverio. Gohm. Dahmen. Vol. and Jacobi Forms (2004) Vol. 1809: O. B. L. Kosaki. Petersburg.C. Markov. IwahoriHecke Algebras and their Representation Theory. G. Zannier (2003) Vol. Cuntz. Werner. Means of Hilbert Space Operators (2003) Vol. 1849: Martin L. Baldoni. Colli. 1818: M. 1813: L. Geometric Aspects of Functional Analysis. Knutsen. M. 1844: K. Mixed Automorphic Forms. Yor (Eds. J. L. 2004. Martina Franca. Jiang. The Valuative Tree (2004) . Heegner Modules and Elliptic Curves (2004) Vol. M. Mathematical Aspects of Evolving Interfaces. L.). Editor: J. Schmidt. Green. Altman. 1822: S. Gaujal. On the Frontier of Probability and Dynamical Systems.). C ∞ – Differentiable Spaces (2003) Vol. Zeitouni. Schindler. D. J. Nˇ stˇ sescu. Bramble. Chow. Van Oystaeyen. E. 1808: W. B. 1815: A. O.P. Mathematical Theory of Nonequilibrium Steady States. 313 p. Lens. A. R.W. Adiabatic Perturbation Theory in Quantum Dynamics (2003) Vol. M. Zecca (2003) Vol. H. Real Methods in Complex and CR Geometry. Schechtman (Eds. 1820: F. Madeira. XIII.). Catoni. Waldschmidt. W. Johnsen. Italy 1999. Gröbner Bases and the Computation of Group Cohomology. 1850: V. Geometric Aspects of Functional Analysis. DiscreteEvent Control of Stochastic Networks: Multimodularity and Regularity. 1837: S. U. 138 p.S. N. M. 2004. 1805: F. Yu. Italy 2001. Teufel. J. 1838: A. Editors: L. ParisPrinceton Lectures on Mathematical Finance 2002 (2003) Vol. K3 Projective Models in Scrolls (2004) Vol. J. Y. Ambrosio. Macki. C. G. M. 1839: R.N. P. Cetraro. Gil’. Salsa (2003) Vol. 1826: K. 1824: J. G. J. Editor: A. Vol. Vol. Italy 1998. Hordijk. Kang. Brown. D. Baudoin. 2003. Martina Franca. Miller. Tignol (2004) Vol. Scheinkman. W. 2003. Sancho de Salas. Lectures on Probability Theory and Statistics. 2004. Stability Estimates for Hybrid Coupled Domain Decomposition Methods (2003) Vol. Statistical Learning Theory and Stochastic Optimization (2004) Vol. Soner. K. Higson. W. Vol. Editors: F. Mimura. H. Lectures on Probability Theory and Statistics. Methods of a a Graded Rings. Editor: P.
1999. Italy. Introduction to Symplectic Dirac Operators (2006) Vol. Proliferation. Bolthausen. Kunstmann. Graphs. Editor: A. 1854: O.A. H. Construction of Global Lyapunov Functions Using Radial Basis Functions (2007) . 1856: K.E. 2002. A. J.B. E. 1891: J. Classical Probability. Italy. A. Picard (2007) Vol. Fluctuation Theory for Lévy Processes. Schürmann. 1862: B. Bárány. Giorgilli (2005) Vol. Pillet. 1885: A. Pillet. W. I. B. Splitting Deformations of Degenerations of Complex Curves (2006) Vol. 1889: G. Open Quantum Systems II. 1865: D. Cell Cycle. Bhat. Results and Examples (2007) Vol. 1903: G.J. Cetraro. 1893: H. Saeki. Open Quantum Systems I. Attal. Editor: Jean Picard (2006) Vol. Editors: M. S. Piazzera (2004) Vol. Takamura. Recent Developments (2006) Vol.S. and Physics. Martina Franca. Habermann. Dembo. Hayashi. Combinatorial Stochastic Processes. M. Tutorials in Mathematical Biosciences I. Italy. Martina Franca. Picard (2007) Vol. Singular Stochastic Differential Equations (2005) Vol. Ecole d’Été de Probabilités de SaintFlour XXXV2005. M. A. 1900: E. The Hamiltonian Approach (2006) Vol. Axiom of Choice (2006) Vol. Iwaniec. Bovier (Eds. Franz (2005) Vol.Séminaire de Probabilités XXXIX (2006) Vol. 1884: N. Editors: C. Local and SemiLocal Bifurcations in Hamiltonian Dynamical Systems. Efstathiou. Ecole d’Été de Probabilités de SaintFlour XXXIII2003. Führ. Joye. A. U. 1898: H. Viola (2006) Vol. R. Mansuy. Weil. Bunge. R. Meyer. U. Lang. Abstract Harmonic Analysis of Continuous Wavelet Transforms (2005) Vol. DonatiMartin. 1872: A.I. 2003. Sneyd (Ed. L. R. T.A. 1880: S.). M. Doney. Henrard. 1873: R. Dynamical Systems. Picard (2006) Vol. Italy. R. Kaikina. 1858: A. In Memoriam PaulAndré Meyer . (2005) Vol. Peng. Osipenko. U. Franz. Pillet. 1877: J. Letellier. C. D. 2003. Nier. 1870: V. The Markovian Approach (2006) Vol.). Editor: J. W.Vol. 1860: A. Editors: A. Lectures on Probability Theory and Statistics. T. 1867: J. and Cancer (2006) Vol. Émery. 1859: E. 1888: J. A. 1901: O. Stricker (2007) Vol. Da Prato. Joye. S. (2005) Vol. 1886: S. Friedman (Ed. 1861: G. Spin Glasses (2007) Vol. L. Random Times and Enlargements of Filtrations in a Brownian Setting (2006) Vol. Computation and Application (2006) Vol. J. Cannone. Stochastic Methods in Finance. 1895: L. 1866: O. Posn (R) and Eisenstein Series. Habermann. Van Assche. Pitman. Hypoelliptic Estimates and Spectral Theory for FokkerPlanck Operators and Witten Laplacians (2005) Vol.W. 1857: M. Structure of Quantum Lévy Processes. S. Molnár. Kümmerer. BarndorffNielsen. Miyakawa (2006) Vol. Massart. Lusky. Telcs. 1879: G. (2005) Vol. Joye. The Wulff Crystal in Ising and Percolation Models. Ecole d’Eté de Probabilités de SaintFlour XXXIV2004. Attal. Editors: M. Analytic Number Theory. C. A.R. A. Naumkin. B. Back. Thorbjønsen. Editor: W. Intersections of Two Quadrics and Pencils of Curves of Genus 1 (2007) Vol. Wittenberg. Groetsch. Italy. Franz. Mathematical Modeling of Calcium Dynamics and Signal Transduction. 1874: M. M. Applebaum. J. Kresin. C. 1875: J. Beyond Partial Differential Equations. Schürmann. F. Gallavotti.I. Jorgenson. Ecole d’Eté de Probabilités de SaintFlour XXXIV2004. 1881: S. Editor: J. Rouault. Quantum Independent Increment Processes II. Geometry of Müntz Spaces and Related Questions. The Lace Expansion and its Applications. Singular Coverings of Toposes (2006) Vol. C. Benettin. Kazhikhov. 1876: H.V. Friedman. Yor. Fritelli. Stochastic Geometry. Ecole d’Eté de Probabilités de SaintFlour XXXII2002. Kustermans. S. Editor: J. Orthogonal Polynomials and Special Functions. 1897: R. Perelli. Borisyuk. Cherny. Mathematical Neurosciences (2005) Vol. Editors: M. 1868: J. 1894: C. 1869: A.A. 1883: W. Kaczorowski. P. 1871: P. Quantum Independent Increment Processes I. Transseries and Real Differential Algebra (2006) Vol.B. Iannelli. 1887: K. Schnaubelt. 1882: S. Editor: Jean Picard (2006) Vol. Yor. Engelbert. Maz’ya. and Algorithms (2006) Vol. Intersections de deux quadriques et pinceaux de courbes de genre 1. Value Distributions of LFunctions (2007) Vol. Tutorials in Mathematical Biosciences III.C.). 1902: A. 1864: K. 1896: P. Asymptotics for Dissipative Nonlinear Equations (2006) Vol. Tutorials in Mathematical Biosciences II. Picard (2005) Vol. 1904: P. Editor: J. R. Friedlander. Lunardi. Funaki. I. Marcellàn (Eds. G. Shishmarev. Baddeley. (2005) Vol. A. C. Hanßmann. J. 2004. M. Ukai. Weis. Kuksin. Nagel.). Herrlich. H. Funk. HeathBrown. W. S. Bielecki. Giesl. Séminaire de Probabilités XXXVIII (2005) Vol. Helffer. Selected Preserver Problems on Algebraic Structures of Linear Operators and on Function Spaces (2007) Vol. van der Hoeven. Gurariy. Metamorphoses of Hamiltonian Systems with Symmetries (2005) Vol. Attal. T. Terman. Ledoux. 1863: H. Schneider. Functional Analytic Methods for Evolution Equations. Émery. F.I. Editors: M. Beyer. Hipp. 1890: M.V. From Classical Probability to Quantum Stochastic Calculus. W. J. Stable Approximate Evaluation of Unbounded Operators (2007) Vol. Yor (Eds. Bressanone/Brixen. I. Slade. Cetraro. C. 1855: G. V. Runggaldier (2004) Vol.R. Steuding. Lectures on the Automorphism Groups of KobayashiHyperbolic Manifolds (2007) Vol. S. D. Ecole d’Eté de Probabilités de SaintFlour XXXIII2003. 1892: A. Fourier Transforms of Invariant Functions on Finite Reductive Lie Algebras (2005) Vol. Constantin. Hamiltonian Dynamics – Theory and Applications.A. M. Ermentrout. Lindsay. The Art of Random Walks (2006) Vol.). 1878: R. Sharp RealPart Theorems (2007) Vol.). On linear and QuasiLinear Abstract Hyperbolic Evolution Equations (2007) Vol. Concentration Inequalities and Model Selection. G. Weil (2007) Vol. Gohm. Lasiecka. Y. Isaev. Cerf. Topology of Singular Fibers of Differential Maps (2004) Vol. 1899: Séminaire de Probabilités XL. Mathematical Foundation of Turbulent Viscous Flows.R. Editors: M. P. Émery (Eds. Schachermayer. Open Quantum Systems III.
Representation Theory and Complex Analysis. 1702: J.. 1931: M. ParisPrinceton Lectures on Mathematical Finance 2004. Afﬁne Density in Wavelet Analysis (2007) Vol. Yong. Capasso. M. and augmented edition (2007) Vol. Jonsson. Schechtman (Eds.M. Stochastic Calculus for Fractional Brownian Motion and Related Processes (2008) Vol. S. Y. M. Italy 2002 (2008) Vol. Wilson.F. Punctured Torus Groups and 2Bridge Knot Groups (I) (2007) Vol. G. Mathematical Models of Granular Matter (2008) Vol.E. J. Morse.A. Algebraic Groups and Lie Groups with Few Factors (2008) Recent Reprints and New Editions Vol.S. Marsden. Evans. Capasso. 1926: L. Cowling. Differential Equations Driven by Rough Paths. M. A. Alternative Pseudodifferential Analysis (2008) Vol. Séminaire de Probabilités XLI (2008) Vol. Gastaldi (2008) Vol. 1999 – Corr. Hamiltonian Reduction by Stages (2007) Vol. Ambrosio. Edited by J. 1916: C. 1919: R. Carmona. Symplectic 4Manifolds and Algebraic Surfaces. T. Italy 2006. T. Williams. Ekeland. M.D. Stability of Nonautonomous Differential Equations (2008) Vol. du Sautoy. Simplicial Complexes of Graphs (2008) Vol. L. Bishwal. Fusco. Yamashita. Ma. Sussmann. SPDE in Hydrodynamic.).J. G. 1914: G. Cetraro. Calculus of Variations and NonLinear Partial Differential Equations. Tian. C. R.I. Tian. Woodward. edition (2008) Vol. J. Bofﬁ. 1936: P.A. Claudia. Stricker (Eds.). Schmidt. Magal. J. 1905: C. Prévˆ t.S. Falcone. Ortega. A. K. Tian (2008) Vol. 1922: A. Italy 2005.N. Italy 2005. 1924: M. Mishura. Albeverio. Italy 2004. ¸ Poland 2006. G.G. V. Roberts. Chazottes. Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms. M. 1918: B. Laplacian g Eigenvectors of Graphs. Cetraro. Taﬂin. Rasmussen. Iterative Approximation of Fixed Points (2007) Vol. P. Serre. D.A. DonatiMartin. Demkowicz. Caruana. F. Ruan (Eds. 1934: C. N. Sontag. Chaplain. Jr. 1913: J. N. L. Lasry.P. M.J. 1906: T. Catanese. Editors: V. and enlarged edition (2008) . Golse. 1915: T. P. Milman. A. Lions. Sakuma. Durán. Crandall.A. 1944: A.A.R. Miekisz. A Systematic Approach to Regularity for Degenerate and Singular PDEs (2008) Vol. Israel Seminar 20042005 (2007) Vol. Friedman (Ed.J. Lachowicz (2008) Vol. Nonlinear and Optimal Control Theory. Ecole d’Été de Probabilités de SaintFlour XXXV2005 (2008) Vol. 1933: M. A Concise Course on o Stochastic Partial Differential Equations (2007) Vol. E. 1917: I. KohatsuHiga. Probability and Real Trees. Structured Population Models in Biology and Epidemiology (2008) Vol. Bressan. V. I. Giovine. Existence and Regularity Properties c of the Integrated Density of States of Random Schrödinger (2008) Vol. Çinlar. Arithmetical Investigations. 1927: L. ForwardBackward Stochastic Differential Equations and their Applications. 1930: J. Venice.J. M. Erdmann. Fortin. Smith. With a preface by D. Seidel. E. Rezakhanlou. G. HøeghKrohn. 1920: S. J. Editors: G. Urbano. 1941: S. Evolution and Ecology (2008) Vol. Ruelle.C. Lyons. Valette. Carmona. From HahnBanach to Monotonicity (Minimax and Monotonicity 1998) – 2nd exp. D’Agnolo. Petkovic. P. D. Auroux. F. 3rd printing (2007) Vol. 1921: J.). K.L. Akiyoshi. Mariano (Eds. Villani. Hyperbolic Systems of Balance Laws. Cetraro. Editors: R. 1912: V. Misiołek. Bedlewo. Leydold. Bonilla (Ed. Röckner. Lévy. Local Newforms for GSp(4) (2007) Vol. P. 1693: S. Agrachev. 1937: G. Plaumann. Cetraro. Brezzi. Green. Ecole d’Été de Probabilités de SaintFlour XXXIV2004 (2007) Vol. Evans. Picardello (2008) Vol. Capriz. Parameter Estimation in Stochastic Differential Equations (2008) Vol. Albeverio. Italy 2003. Catanese. 1908: T.G. LittlewoodPaley Theory and ExponentialSquare Integrability (2008) Vol.L. R.N. Marcati (2007) Vol. 1939: D.G. Flandoli. Perlmutter. Röckner (2008) Vol. Olla (2008) Vol. 1940: J. Editors: F. Evolution Algebras and their Applications (2008) Vol. L.M.J. Italy 2003. 1925: M. Scheinkman.P.A. Zumbrun. Barreira. M. with an Appendix on Schensted Correspondence and Littelmann Paths by K. Cetraro. F.P. 1929: Y. Marcellini (2008) Vol. M. 1911: A. Polynomial Representations of GLn . Tarabusi. and Applications. P. Stefani (2008) Vol. P.M.A. The Method of Intrinsic Scaling. The Method of Approximate Inverse: Theory and Applications (2007) Vol. 1976 – 2nd corr. 1932: A. Editor: P. Banasiak. M. Editors: E. 1938: D. Inverse Problems and Imaging. Bofﬁ. Frenkel. M. Manetti. L. G. Orthogonal Polynomials. N. 830: J. Falk.C. M. H. Kashiwara.).S. E. 1928: J. Mixed Finite Elements. Editors: P. 1943: L. Siebert. PerronFrobenius and FaberKrahn Type Theorems (2007) Vol. 1942: S. From Microscopic to Macroscopic. Nistri. 1910: V. I. 470: R. Editors: B. Pham. Sinai. A. Stadler. Utkin. Simons. Tutorials in Mathematical BioSciences IV. M. M. Rouault. Geometric Aspects of Functional Analysis. I. J. E. B. Recent Progress and Prospects. edition (2008) Vol. R. Émery.Vol. Entropy Methods for the Boltzmann Equation. G. H. Zeta Functions of Groups and Rings (2008) Vol. Di Bartolo. Ratiu. J. P. Y. Representation Theory. L. Green and M. Da Prato. E.D. Ekeland. Bowen. Martina Franca. Berinde. 523: S. Bıyıkoˇ lu. Schuster. Editors: D. M. Strambach. and Quantum Interpolations (2008) Vol. 1935: A. Multiscale Problems in the Life ¸ Sciences.M. Haran. Touzi (2007) Vol. F. Vogan. Unterberger. Wada. Compatibility Conditions. Mathematical Theory of Feynman Path Integral. Caffarelli. 1923: J.A.R. R. A. Dacorogna. Kutyniok. Editors: F. Attractivity and Bifurcation for Nonautonomous Dynamical Systems (2007) Vol. 1975 – 2nd rev. Wallach. Italy 2004. 1907: M. 1909: H. Point Estimation of Root Finding Methods (2008) Vol. J. S.). Lachowicz. S. Mazzucchi. V.A. Veseli´ . Jouini.). Schoker 1980 – 2nd corr. Cetraro. M.
This action might not be possible to undo. Are you sure you want to continue?