Omri Sarig

Lecture Notes on Ergodic Theory

April 3, 2023

(Prepared using the Springer svmono author package)

 v

Special thanks to Snir Ben-Ovadia, Keith Burns, Yair Daon, Dimitris Gatzouras,
Yair Hartman, Qiujie Qiao, Abhishek Khetan, Ian Melbourne, Tom Meyerovitch,
Ofer Shwartz and Andreas Strömbergsson for indicating typos and mistakes in ear-
lier versions of this set of notes.
If you find additional errors please let me know! O.S.
Contents

1 Basic definitions and constructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 What is ergodic theory and how it came about . . . . . . . . . . . . . . . . . . . 1
1.2 The abstract setup of ergodic theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 The probabilistic point of view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Ergodicity and mixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5.1 Circle rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5.2 The angle doubling map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5.3 Bernoulli Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5.4 Finite Markov Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5.5 The geodesic flow on a hyperbolic surface . . . . . . . . . . . . . . . 17
1.6 Basic constructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.6.1 Skew-products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.6.2 Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.6.3 The natural extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.6.4 Induced transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.6.5 Suspensions and Kakutani skyscrapers . . . . . . . . . . . . . . . . . . 30
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2 Ergodic Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.1 The Mean Ergodic Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.2 The Pointwise Ergodic Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3 The non-ergodic case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.3.1 Conditional expectations and the limit in the ergodic theorem 40
2.3.2 Conditional probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.3 The ergodic decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.4 An Ergodic Theorem for Zd -actions . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.5 The Subadditive Ergodic Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.6 The Multiplicative Ergodic Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.6.1 Preparations from Multilinear Algebra . . . . . . . . . . . . . . . . . . 52

vii
viii Contents

2.6.2 Proof of the Multiplicative Ergodic Theorem . . . . . . . . . . . . . 57

2.6.3 The Multiplicative Ergodic Theorem for Invertible Cocycles 66
2.7 The Karlsson-Margulis ergodic theorem . . . . . . . . . . . . . . . . . . . . . . . . 69
2.7.1 The boundary of a non-compact proper metric space . . . . . . . 69
2.7.2 An ergodic theorem for isometric actions on CAT(0) spaces . 77
2.7.3 A geometric proof of the multiplicative ergodic theorem . . . 81
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

3 Spectral Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.1 The spectral approach to ergodic theory . . . . . . . . . . . . . . . . . . . . . . . . 89
3.2 Weak mixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.2.1 Definition and characterization . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.2.2 Spectral measures and weak mixing . . . . . . . . . . . . . . . . . . . . . 92
3.3 The Koopman operator of a Bernoulli scheme . . . . . . . . . . . . . . . . . . . 95
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

4 Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.1 Information content and entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.2 Properties of the entropy of a partition . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.2.1 The entropy of α ∨ β . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.2.2 Convexity properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.2.3 Information and independence . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.3 The Metric Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.3.1 Definition and meaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.3.2 The Shannon–McMillan–Breiman Theorem . . . . . . . . . . . . . . 109
4.3.3 Sinai’s Generator theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.4.1 Bernoulli schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.4.2 Irrational rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.4.3 Markov measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.4.4 Expanding Markov Maps of the Interval . . . . . . . . . . . . . . . . . 115
4.5 Abramov’s Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.6 Topological Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.6.1 The Adler–Konheim–McAndrew definition . . . . . . . . . . . . . . 118
4.6.2 Bowen’s definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.6.3 The variational principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.7 Ruelle’s inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.7.1 Preliminaries on singular values . . . . . . . . . . . . . . . . . . . . . . . . 124
4.7.2 Proof of Ruelle’s inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Contents ix

A The isomorphism theorem for standard measure spaces . . . . . . . . . . . . 133

A.1 Polish spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
A.2 Standard probability spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
A.3 Atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
A.4 The isomorphism theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

A The Monotone Class Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Chapter 1
Basic definitions and constructions

1.1 What is ergodic theory and how it came about

Dynamical systems and ergodic theory. Ergodic theory is a part of the theory of
dynamical systems. At its simplest form, a dynamical system is a function T defined
on a set X. The iterates of the map are defined by induction T 0 := id, T n := T ◦T n−1 ,
and the aim of the theory is to describe the behavior of T n (x) as n → ∞.
More generally one may consider the action of a semi-group of transformations,
namely a family of maps Tg : X → X (g ∈ G) satisfying Tg1 ◦ Tg2 = Tg1 g2 . In the
particular case G = R+ or G = R we have a family of maps Tt such that Tt ◦Ts = Tt+s ,
and we speak of a semi-flow or a flow.
The original motivation was classical mechanics. There X is the set of all pos-
sible states of given dynamical system (sometimes called configuration space) or
phase space), and T : X → X is the law of motion which prescribes that if the sys-
tem is at state x now, then it will evolve to state T (x) after one unit of time. The orbit
{T n (x)}n∈Z is simply a record of the time evolution of the system, and the under-
standing the behavior of T n (x) as n → ∞ is the same as understanding the behavior
of the system at the far future. Flows Tt arise when one insists on studying contin-
uous, rather than discrete time. More complicated group actions, e.g. Zd –actions,
arise in material science. There x ∈ X codes the configuration of a d–dimensional
lattice (e.g. a crystal), and {Tv : v ∈ Zd } are the symmetries of the lattice.
The theory of dynamical systems splits into subfields which differ by the struc-
ture which one imposes on X and T :
1. Differentiable dynamics deals with actions by differentiable maps on smooth
manifolds;
2. Topological dynamics deals with actions of continuous maps on topological
spaces, usually compact metric spaces;
3. Ergodic theory deals with measure preserving actions of measurable maps on a
measure space, usually assumed to be finite.

1
2 1 Basic definitions, examples, and constructions

It may seem strange to assume so little on X and T . The discovery that such
meagre assumptions yield non trivial information is due to Poincaré, who should be
considered the progenitor of the field.
Poincaré’s Recurrence Theorem and the birth of ergodic theory. Imagine a box
filled with gas, made of N identical molecules. Classical mechanics says that if
we know the positions qi = (q1i , q2i , q3i ) and momenta pi = (p1i , p2i , p3i ) of the i-th
molecule for all i = 1, . . . , N, then we can determine the positions and momenta of
each particle at time t by solving Hamilton’s equations

ṗij (t) = −∂ H/∂ qij

(1.1)
q̇ij (t) = ∂ H/∂ pij .

H = H(q1 , . . . , qN ; p1 , . . . , pN ), the Hamiltonian, is the total energy of the system.

It is natural to call (q, p) := (q1 , . . . , qN ; p1 , . . . , pN ) the state of the system. Let
X denote the collection of all possible states. If we assume (as we may) that the
total energy is bounded above, then for many reasonable choices of H this is a open
bounded subset of R6N . Let

Tt : (q, p) 7→ (q(t), p(t))

denote the map which gives solution of (1.1) with initial condition (q(0), p(0)). If
H is sufficiently regular, then (1.1) had a unique solution for all t and every ini-
tial condition. The uniqueness of the solution implies that Tt is a flow. The law of
conservation of energy implies that x ∈ X ⇒ Tt (x) ∈ X for all t.
Question: Suppose the system starts at a certain state (q(0), p(0)), will it eventually
return to a state close to (q(0), p(0))?
For general H, the question seems intractable because (1.1) is strongly coupled
system of an enormous number of equations (N ∼ 1024 ). Poincaré’s startling discov-
ery is that the question is trivial, if viewed from the right perspective. To understand
his solution, we need to recall a classical fact, known as Liouville’s theorem: The
Lebesgue measure m on X satisfies m(Tt E) = m(E) for all t and all measurable
E ⊂ X (problem 1.1).
Here is Poincaré’s solution. Define T := T1 , and observe that T n = Tn . Fix ε > 0
and consider the set W of all states x = (q, p) such that d(x, T n (x)) > ε for all n ≥ 1
(here d is the Euclidean distance). Divide W into finitely many disjoint pieces Wi of
diameter less than ε.
For each fixed i, the sets T −n (Wi ) (n ≥ 1) are pairwise disjoint: Otherwise
T (Wi ) ∩ T −(n+k) (Wi ) ̸= ∅, so Wi ∩ T −k (Wi ) ̸= ∅, and there exists x ∈ Wi ∩
−n

T −k (Wi ). But this leads to a contradiction:

1. x ∈ T −k (Wi ) implies that T k (x) ∈ Wi whence d(x, T k x) ≤ diam(Wi ) < ε, whereas
2. x ∈ Wi ⊂ W implies that d(x, T k x) > ε by the definition of W .
So T −n (Wi ) (n ≥ 1) are pairwise disjoint.
1.2 The abstract setup of ergodic theory 3

Since {T −nWi }n≥1 are pairwise disjoint, m(X) ≥ ∑k≥1 m(T −kWi ). But T −k (Wi )
all have the same measure (Liouville theorem), and m(X) < ∞, so we must have
m(Wi ) = 0. Summing over i we get that m(W ) = 0. In summary, a.e. x has the
property that d(T n (x), x) < ε for some n ≥ 1. Considering the countable collection
ε = 1/n, we obtain the following result:
Poincaré’s Recurrence Theorem: For almost every x = (q(0), p(0)), if the system
is at state x at time zero, then it will return arbitrarily close to this state infinitely
many times in the arbitrarily far future.
Poincaré’s Recurrence Theorem is a tour de force, because it turns a problem
which looks intractable to a triviality by simply looking at it from a different angle.
The only thing the solution requires is the existence of a finite measure on X such
that m(T −1 E) = m(E) for all measurable sets E. This startling realization raises
the following mathematical question: What other dynamical information can one
extract from the existence of a measure m such that m = m ◦ T −1 ? Of particular
interest was the justification of the following “assumption” made by Boltzmann in
his work on statistical mechanics:
The Ergodic Hypothesis: For certain invariant measures µ, many functions f : X →
R, and many states x = (q, p), the time average lim T1 0T f (Tt (x))dt exists, and
R
T →∞
1 R
equals the space average µ(X) f dµ.
(This is not Boltzmann’s original formulation.) The ergodic hypothesis is a quanti-
tative version of Poincaré’s recurrence theorem: If f is the indicator of the ε–ball
around a state x, then the time average of f is the frequency of times when Tt (x) is
ε–away from x, and the ergodic hypothesis is a statement on its value.

1.2 The abstract setup of ergodic theory

The proof of Poincaré’s Recurrence Theorem suggests the study of the following
setup.
Definition 1.1. A measure space is a triplet (X, B, µ) where
1. X is a set, sometime called the space.
2. B is a σ –algebra: a collection of subsets of X which contains the empty set,
and which is closed under complements, countable unions and countable inter-
sections. The elements of B are called measurable sets.
3. µ : B → [0, ∞], called the measure,
U
is a σ –additive function: if E1 , E2 , . . . ∈ B
are pairwise disjoint, then µ( i Ei ) = ∑i µ(Ei ).
If µ(X) = 1 then we say that µ is a probability measure and (X, B, µ) is a proba-
bility space.
In order to avoid measure theoretic pathologies, we will always assume that
(X, B, µ) is the completion (see problem 1.2) of a standard measure space: a mea-
4 1 Basic definitions, examples, and constructions

sure space (X, B ′ , µ ′ ), where X is a complete and separable metric space and B ′ is
its Borel σ –algebra.
It can be shown that such spaces are Lebesgue spaces: They are isomorphic
to the union of a compact interval equipped with the Lebesgue’s σ –algebra and
Lebesgue’s measure, and a finite or countable or empty collection of points with
positive measure. See the appendix for details.

Definition 1.2. A measure preserving transformation (mpt) is a quartet (X, B, µ, T )

where (X, B, µ) is a measure space, and
1. T is measurable: E ∈ B ⇒ T −1 E ∈ B;
2. m is T –invariant: µ(T −1 E) = µ(E) for all E ∈ B.
A probability preserving transformation (ppt) is a mpt on a probability space.

This is the minimal setup needed to prove (problem 1.3):

Theorem 1.1 (Poincaré’s Recurrence Theorem). Suppose (X, B, µ, T ) is a p.p.t.
If E is a measurable set, then for almost every x ∈ E there is a sequence nk → ∞
such that T nk (x) ∈ E.
Poincaré’s theorem is not true for general infinite measure preserving transforma-
tions, as the example T (x) = x + 1 on Z demonstrates.
Having defined the objects of the theory, we proceed to declare when do we
consider two objects to be isomorphic:
Definition 1.3. Two m.p.t. (Xi , Bi , µi , Ti ) are called measure theoretically isomor-
phic, if there exists a measurable map π : X1 → X2 such that
1. there are Xi′ ∈ Bi such that mi (Xi \ Xi′ ) = 0 and such that π : X1′ → X2′ is invertible
with measurable inverse;
2. for every E ∈ B2 , π −1 (E) ∈ B1 and m1 (π −1 E) = m2 (E);
3. T2 ◦ π = π ◦ T1 on X1 .
One of the main aims of ergodic theorists is to develop tools for deciding whether
two mpt’s are isomorphic.

1.3 The probabilistic point of view.

Much of the power and usefulness of ergodic theory is due to the following prob-
abilistic interpretation of the abstract set up discussed above. Suppose (X, B, µ, T )
is a ppt.
1. We imagine X to be a sample space: the collection of all possible outcomes ω of
a random experiment.
2. We interpret B as the collection of all measurable events: all sets E ⊂ X such
that we have enough information to answer the question “is ω ∈ E?.”
3. We use µ to define the probability law: Pr[ω ∈ E] := µ(E);
 1.4 Ergodicity and mixing 5

4. Measurable functions f : X → R are random variables f (ω);

5. The sequence Xn := f ◦ T n (n ≥ 1) is a stochastic process, whose distribution is
given by the formula
!
k
\
Pr[Xi1 ∈ Ei1 , . . . , Xik ∈ Eik ] := µ {ω ∈ X : f (T i j ω) ∈ Ei j } .
j=1

The invariance of µ guarantees that such stochastic processes are always station-
ary: Pr[Xi1 +m ∈ Ei1 +m , . . . , Xik ∈ Eik +m ] = Pr[Xi1 ∈ Ei1 , . . . , Xik ∈ Eik +m ] for all m.
The point is that we can ask what are the properties of the stochastic processes
{ f ◦ T n }n≥1 arising out of the ppt (X, B, µ, T ), and bring in tools and intuition from
probability theory to the study of dynamical systems.
Note that we have found a way of studying stochastic phenomena in a context
which is, a priori, completely deterministic (if we know the state of the system at
time zero is x, then we know with full certainty that its state at time n is T n (x)). The
modern treatment of the question “how come a deterministic system can behave
randomly” is based on this idea.

1.4 Ergodicity and mixing

Suppose (X, B, µ, T ) is a mpt. A measurable set E ∈ B is called invariant, if

T −1 (E) = E. Evidently, in this case T can be split into two measure preserving
transformations T |E : E → E and T |E c : E c → E c , which can be analyzed separately.
Definition 1.4. A mpt (X, B, µ, T ) is called ergodic, if every invariant set E satisfies
µ(E) = 0 or µ(X \ E) = 0. We say µ is an ergodic measure.

Proposition 1.1. Suppose (X, B, µ, T ) is a mpt on a complete measure space, then

the following are equivalent:
1. µ is ergodic;
2. if E ∈ B and µ(T −1 E△E) = 0, then µ(E) = 0 or µ(X \ E) = 0;
3. if f : X → R is measurable and f ◦ T = f a.e., then there is c ∈ R s.t. f = c a.e.

Proof. Suppose µ is ergodic, and E is measurable s.t. µ(E△T −1 E) = 0. We con-

struct a measurable set E0 such that T −1 E0 = E0 and µ(E0 △E) = 0. By ergodicity
µ(E0 ) = 0 or µ(X \ E0 ) = 0. Since µ(E△E0 ) = 0 implies that µ(E) = µ(E0 ) and
µ(X \ E) = µ(X \ E0 ) we get that either µ(E) = 0 or µ(X \ E) = 0.
The set E0 we use is E0 := {x ∈ X : T k (x) ∈ E infinitely often.}. It is obvious that
this set is measurable and invariant. To estimate µ(E0 △E) note that
(a) if x ∈ E0 \ E, then there exists some k s.t. x ∈ T −k (E) \ E;
(b) if x ∈ E \ E0 , then there exists some k s.t. x ̸∈ T −k (E), whence x ∈ E \ T −k (E).
6 1 Basic definitions, examples, and constructions

Thus E0 △E ⊂ k≥1 E△T −k (E).

S

We now use the following “triangle inequality” :

µ(A1 △A3 ) ≤ µ(A1 △A2 ) + µ(A2 △A3 ) (Ai ∈ B)

(This is because µ(Ai △Ai ) = ∥1Ai − 1A j ∥1 where 1Ai is the indicator function of Ai ,
equal to one on Ai and to zero outside Ai ). The triangle inequality implies that
∞ ∞ k−1
µ(E0 △E) ≤ ∑ µ(E△T −k E) ≤ ∑ ∑ µ(T −i E△T −(i+1) E)
k=1 k=1 i=0
∞
= ∑ kµ(E△T −1 E) (∵ µ ◦ T −1 = µ).
k=1

Since µ(E△T −1 E) = 0, µ(E0 △E) = 0 and we have shown that (1) implies (2).
Next assume (2). and let f be a measurable function s.t. f ◦ T = f almost every-
where. For every t, [ f > t]△T −1 [ f > t] ⊂ [ f ̸= f ◦ T ], so

µ([ f > t]△T −1 [ f > t]) = 0.

By assumption, this implies that either µ[ f > t] = 0 or µ[ f ≤ t] = 0. In other words,

either f > t a.e., or f ≤ t a.e. Define c := sup{t : f > t a.e.}, then f = c almost
everywhere, proving (3). The implication (3)⇒(1) is obvious: take f = 1E . ⊔
⊓

An immediate corollary is that ergodicity is an invariant of measure theoretic iso-

morphism: If two mpt are isomorphic, then the ergodicity of one implies the ergod-
icity of the other.
The next definition is motivated by the probabilistic notion of independence. Sup-
pose (X, B, µ) is a probability space. We think of elements of B as of “events”, we
interpret measurable functions f : X → R as “random variables”, and we view µ as
a “probability law” µ(E) = P[x ∈ E]. Two events E, F ∈ B are called independent,
if µ(E ∩ F) = µ(E)µ(F) (because in the case µ(E), µ(F) ̸= 0 this is equivalent to
saying that µ(E|F) = µ(E), µ(F|E) = µ(F)).

Definition 1.5. A probability preserving transformation (X, B, µ, T ) is called mix-

ing (or strongly mixing), if for all E, F ∈ B, µ(E ∩ T −k F) −−−→ µ(E)µ(F).
k→∞

In other words, T −k (F) is “asymptotically independent” of E. It is easy to see that

strong mixing is an invariant of measure theoretic isomorphism.
It can be shown that the sets E, F in the definition of mixing can be taken to be
equal (problem 1.12).

Proposition 1.2. Strong mixing implies ergodicity.

Proof. Suppose E is invariant, then µ(E) = µ(E ∩ T −n E) −−−→ µ(E)2 , whence

n→∞
µ(E)2 = µ(E). It follows that µ(E) = 0 or µ(E) = 1 = µ(X). ⊔
⊓
1.4 Ergodicity and mixing 7

Just like ergodicity, strong mixing can be defined in terms of functions. Before
we state the condition, we recall a relevant notion from statistics. The correlation
coefficient of non-constant f , g ∈ L2 (µ) is defined to be
R R R
f g dµ − f dµ · gdµ
ρ( f , g) := R R .
∥ f − f dµ∥2 ∥g − gdµ∥2

The numerator is equal to

Z  Z  Z
f− g − g dµ,

Cov( f , g) := f

called the covariance of f , g. This works as follows: If f , g are weakly correlated

then they will not always deviate from their means in the same direction, leading to
many cancelations in the integral, and a small net result. If f , g are strongly corre-
lated, there will be less cancelations, and a larger absolute value for the net result.
Positive covariance signifies that the deviations from the mean are often in the same
direction, and negative covariance indicates that that they are often in opposite di-
rections. The denominator in the definition of ρ is not important. It is there to force
ρ( f , g) to have values in [−1, 1] (Cauchy-Schwarz).

Proposition 1.3. ARppt (X, B, µ, T ) is strongly mixing iff for every f , g ∈ L2 ,

f g ◦ T dµ −−−→ f dµ gdµ, equivalently Cov( f , g ◦ T n ) −−−→ 0.
n
R R
n→∞ n→∞

Proof. We need the following trivial observations:

1. Since µ ◦ T −1 = µ, ∥ f ◦ T ∥ p = ∥ f ∥ p for all f ∈ L p and 1 ≤ p ≤ ∞;
2. Cov( f , g) is bilinear Rin f , g; R
3. |Cov( f , g)| ≤ 4∥ f − f ∥2 ∥g − g∥2 .
The first two statements are left as an exercise.
R
ForR the third we use the Cauchy-
Schwarz inequality: |Cov( f , g)| ≤ ∥ f − f ∥2 ∥g − g∥2 ≤ (∥ f ∥2 + ∥ f ∥1 )(∥g∥2 +
! !
∥g∥1 ) ≤ (2∥ f ∥2 )(2∥g∥2 ), where ≤ is because ∥ϕ∥1 = ∥ϕ · 1∥1 ≤ ∥ϕ∥2 ∥1∥2 = ∥ϕ∥2 .
Now for the proof of the proposition. The condition that Cov( f , g ◦ T n ) −−−→ 0
n→∞
for all f , g ∈ L2 implies mixing by substituting f = 1E , g = 1F . For the other di-
rection, assume that µ is mixing, and let f , g be two elements of L2 . If f , g are
indicators of measurable sets, then Cov( f , g ◦ T n ) → 0 by mixing. If f , g are finite
linear combinations of indicators, Cov( f , g ◦ T n ) → 0 because of the bilinearity of
the covariance. For general f , g ∈ L2 , we can find for every ε > 0 finite linear combi-
nations of indicators fε , gε s.t. ∥ f − fε ∥2 , ∥g − gε ∥2 < ε. By the observations above,

|Cov( f , g ◦ T n )| ≤ |Cov( f − fε , g ◦ T n )| + |Cov( fε , gε ◦ T n )| + |Cov( fε , (g − gε ) ◦ T n )|

≤ 4ε∥g∥2 + o(1) + 4(∥ f ∥2 + ε)ε, as n → ∞.

It follows that lim sup |Cov( f , g ◦ T n )| ≤ 4ε(∥ f ∥2 + ∥g∥2 + ε). Since ε is arbitrary,
the limsup, whence the limit itself, is equal to zero. ⊔
⊓
8 1 Basic definitions, examples, and constructions

1.5 Examples

1.5.1 Circle rotations

Let T := [0, 1) equipped with the Lebesgue measure m, and define for α ∈ [0, 1)
Rα : T → T by Rα (x) = x + α mod 1. Rα is called a circle rotation, because the
map π(x) = exp[2πix] is an isomorphism between Rα and the rotation by the angle
2πα on the unit circle S1 .
Proposition 1.4.
1. Rα is measure preserving for every α;
2. Rα is ergodic iff α ̸∈ Q;
3. Rα is never strongly mixing.

Proof. A direct calculation shows that the Lebesgue measure m satisfies m(R−1 α I) =
m(I) for all intervals I ⊂ [0, 1). Thus the collection M := {E ∈ B : m(R−1 α E) =
m(E)} contains the algebra of finite disjoint unions of intervals. It is easy to check
M is a monotone class, so by the monotone class theorem (see appendix) M con-
tains all Borel sets. It clearly contains all null sets. Therefore it contains all Lebesgue
measurable sets. Thus M = B and (1) is proved.
We prove (2). Suppose first that α = p/q for p, q ∈ N. Then Rqα = id. Fix some x ∈
[0, 1), and pick ε so small that the ε–neighborhoods of x +kα for k = 0, . . . , q −1 are
disjoint. The union of these neighborhoods is an invariant set of positive measure,
and if ε is sufficiently small then it is not equal to T. Thus Rα is not ergodic.
Next assume that α ̸∈ Q. Suppose E is an invariant set, and set f = 1E . Expand
f to a Fourier series:

f (t) = ∑ fb(n)e2πint ( convergence in L2 ).

n∈Z

The invariance of E dictates f = f ◦ Rα . The Fourier expansion of f ◦ Rα is

( f ◦ Rα )(t) = f (t + α mod 1) = ∑ e2πinα fb(n)e2πint .

n∈Z

Equating coefficients, we see that fb(n) = fb(n) exp[2πinα]. Thus either fb(n) = 0
or exp[2πinα] = 1. Since α ̸∈ Q, fb(n) = 0 for all n ̸= 0. We obtain that f = fb(0)
a.e., whence 1E = m(E) almost everywhere. This can only happen if m(E) = 0 or
m(E) = 1, proving the ergodicity of m.
To show that m is not mixing, we consider the function f (x) = exp[2πix]. This
function satisfies f ◦ Rα = λ f with λ = exp[2πiα] (such a function is called an
eigenfunction). For every α there is a sequence nk → ∞ s.t. nk α mod 1 → 0 (Dirich-
n
let theorem), thus ∥ f ◦ Rαk − f ∥2 = |λ nk − 1| −−−→ 0. It follows that F := Re( f ) =
k→∞
n n
F ◦ Rαk Fdm −−−→ F 2 dm ̸=
R R
cos(2πx) satisfies ∥F ◦ Rαk − F∥2 −−−→ 0, whence
k→∞ k→∞
( F)2 , and m is not mixing.
R
⊔
⊓
1.5 Examples 9

1.5.2 The angle doubling map

Again, we work with T := [0, 1] equipped with the Lebesgue measure m, and define
T : T → T by T (x) = 2x mod 1. T is called the angle doubling map, because the
map π(x) := exp[2πix] is an isomorphism between T and the map eiθ 7→ e2iθ on S1 .

Proposition 1.5. The angle doubling map is probability preserving, and strong mix-
ing, whence ergodic.

Proof. It is convenient to work with binary expansions x = (0.d1 d2 d3 . . .), (di =

0, 1), because with this representation T (0.d1 d2 . . .) = (0.d2 d3 . . .). For every finite
n–word of zeroes and ones (d1 , . . . , dn ), define the sets (called “cylinders”)

[d1 , . . . , dn ] := {x ∈ [0, 1) : x = (0.d1 · · · dn ε1 ε2 . . .), for some εi ∈ {0, 1}}.

This is a (dyadic) interval, of length 1/2n .

It is clear that T −1 [d1 , . . . , dn ] = [∗, d1 , . . . , dn ] where ∗ stands for “0 or 1”. Thus,
m(T −1 [d]) = m[0, d] + m[1, d] = 2 · 2−(n+1) = 2−n = m[d]. We see that M := {E ∈
B : m(T −1 E) = m(E)} contains the algebra of finite disjoint unions of cylinders.
Since M is obviously a monotone class, and since the cylinders generate the Borel
σ –algebra (prove!), we get that M = B, whence T is measure preserving.
We prove that T is mixing. Suppose f , g are indicators of cylinders: f = 1[a1 ,...,an ] ,
g = 1[b1 ,...,bm ] . Then for all k > n,
Z
f · g ◦ T k dm = m[a, ∗| ·{z
· · ∗}, b] = m[a]m[b].
k−n

Thus Cov( f , g ◦ T k ) −−−→ 0 for all indicators of cylinders. Every L2 –function can be
k→∞
approximated in L2 by a finite linear combination of indicators of cylinders (prove!).
One can therefore proceed as in the proof of proposition 1.3 to show that Cov( f , g ◦
T k ) −−−→ 0 for all L2 functions. ⊔
⊓
k→∞

1.5.3 Bernoulli Schemes

Let S be a finite set, called the alphabet, and let X := SN be the set of all one–sided
infinite sequences of elements of S. Impose the following metric on X:

d((xn )n≥0 , (yn )n≥0 ) := 2− min{k:xk ̸=yk } . (1.2)

The resulting topology is generated by the collection of cylinders:

[a0 , . . . , an−1 ] := {x ∈ X : xi = ai (0 ≤ i ≤ n − 1)}.

10 1 Basic definitions, examples, and constructions

It can also be characterized as being the product topology on SN , when S is given

the discrete topology. In particular this topology is compact.
The left shift is the transformation T : (x0 , x1 , x2 , . . .) 7→ (x1 , x2 , . . .). The left shift
is continuous.
Next fix a vector p = (pa )a∈S of positive numbers whose sum is equal to one.
Definition 1.6. The Bernoulli measure corresponding to p is the unique measure on
the Borel σ –algebra of X such that µ[a0 , . . . , an−1 ] = pa0 · · · pan−1 for all cylinders.
It is useful to recall why such a measure exists. Here is a review of the necessary
tools from measure theory.
Definition 1.7. Let X be a non-empty set.
1. A semi-algebra on X is a collection S os subsets of X such that
a. ∅, X ∈ S
b. S is closed under intersections, and
c. for every A ∈ S , X \ A is a finite disjoint union of elements from S .
2. The σ -algebra generated by S is the smallest σ -algebra of subsets of X which
contains S . Equivalently, this is the σ -algebra equal to the intersection of all
σ -algebras which contain S (e.g. the power set of X).
3. A function µ : S → [0, ∞] is called σ -finite if X is a countable disjoint union of
elements Ai ∈ S such that µ(Ai ) < ∞ for all i. This always happens if µ(X) < ∞.
4. A function µ : S → [0, ∞] is called σ -additive on SU if for all pairwise disjoint
countable collection of Ai ∈ S , if Ai ∈ S then µ( Ai ) = ∑ µ(Ai ).
U

Theorem 1.2 (Carathéodory’s Extension Theorem). Let X be a set and S a semi-

algebra of subsets of X. Every σ –additive σ -finite µ : S → [0, ∞] has a unique
extension to a σ –additive σ -finite function on the σ –algebra generated by S .
In our case X = {(x0 , x1 , . . .) : xi ∈ S for all i}, S = {∅, X} ∪ {cylinders}, and
µ : S → [0, ∞] is defined by µ(∅) := 0, µ(X) := 1, µ[a0 , . . . , an−1 ] := pa0 · · · pan−1 .
S is a semi-algebra, because the intersection of cylinders is empty or a cylinder,
and because the complement of a cylinder [a0 , . . . , an−1 ] is the finite disjoint union
of cylinders [b0 , . . . , bn−1 ] such that for some i, bi ̸= ai .
It is also clear that µ : S → [0, ∞] is σ -finite. Indeed it is finite (µ(X) < ∞)! It
remain to check that µ is σ -additive on S .
Suppose [a] is a countable disjoint union of cylinders [b j ]. Each cylinder is open
and compact (prove!), so such unions are necessarily finite. Let N be the max-
imal length of the cylinder [b j ]. Since [b j ] ⊆ [a], we can write [b j ] = [a, β j ] =
U j
|c|=N−|b j | [a, β , c], and a direct calculation shows that

 N−|b j |
j j
∑ µ[a, β , c] = µ[a, β ] ∑ pc = µ[a, β j ] ≡ µ[b j ].
|c|=N−|b j | c

Summing over j, we get that ∑ j µ[b j ] = ∑ j ∑|c|=N−|b j | µ[a, β j , c].

1.5 Examples 11

Now [a] = j [b j ] = j |c|=N−|b j | [a, β j , c], so the collection of (β j , c) is equal

U U U

to the collection of all possible words w of length N − |a| (otherwise the right hand
side misses some sequences). Thus
 N−|a|
j
∑ µ[b ] = ∑ µ[a, w] = µ[a] ∑ pc = µ[a],
j |w|=N−|a| c

proving the σ –additivity of µ on S .

It now follows from the Carathéodory’s Extension Theorem that µ : S → [0, ∞]
has a unique extension to a probability measure on the smallest σ -algebra containing
the cylinders.
Call this σ -algebra B. Since the cylinders generate the topology of X, every open
set is a union of cylinders. This union is countable, because there are only countably
many cylinders. So every open set belongs to B. It follows that B contains the Borel
σ -algebra of X (which equals by definition to the smallest σ -algebra containing all
open sets). So µ extends uniquely to a non-negative σ -additive function on the Borel
σ -algebra of X. Definition probability measure. Definition 1.6 is justified.

Proposition 1.6. Suppose X = {0, 1}N , µ is the ( 12 , 21 )–Bernoulli measure, and σ is

the left shift, then (X, B(X), µ, T ) is measure theoretically isomorphic to the angle
doubling map.

Proof. The isomorphism is π(x0 , x1 , . . . , ) := ∑ 2−n xn . This is a bijection between

X ′ := {x ∈ {0, 1}N :̸ ∃n s.t. xm = 1 for all m ≥ n}

and [0, 1) (prove that µ(X ′ ) = 1), and it is clear that π ◦ σ = T ◦ π. Since the image
of a cylinder of length n is a dyadic interval of length 2−n , π preserves the measures
of cylinders. The collection of measurable sets which are mapped by π to sets of the
same measure is a σ –algebra. Since this σ –algebra contains all the cylinders and all
the null sets, it contains all measurable sets. ⊔
⊓

Proposition 1.7. Every Bernoulli scheme is mixing, whence ergodic.

The proof is the same as in the case of the angle doubling map. Alternatively, it
follows from the mixing of the angle doubling map and the fact that the two are
isomorphic.

1.5.4 Finite Markov Chains

We saw that the angle doubling map is isomorphic to a dynamical system acting
as the left shift on a space of sequences (a Bernoulli scheme). Such representations
appear frequently in the theory of dynamical systems, but more often than not, the
space of sequences is slightly more complicated than the set of all sequences.
12 1 Basic definitions, examples, and constructions

1.5.4.1 Subshifts of finite type

Let S be a finite set, and A = (ti j )S×S a matrix of zeroes and ones without columns
or rows made entirely of zeroes.
Definition 1.8. The subshift of finite type (SFT) with alphabet S and transition ma-
trix A is the set ΣA+ = {x = (x0 , x1 , . . .) ∈ SN : txi xi+1 = 1 for all i}, together with the
metric d(x, y) := 2− min{k:xk ̸=yk } and the action σ (x0 , x1 , x2 , . . .) = (x1 , x2 , . . .).
This is a compact metric space, and the left shift map σ : ΣA+ → ΣA+ is continuous.
We think of ΣA+ as of the space of all infinite paths on a directed graph with vertices
S and edge a → b connecting a, b ∈ S such that tab = 1.
Let ΣA+ be a SFT with set of states S, |S| < ∞, and transition matrix A = (Aab )S×S .
• A stochastic matrix is a matrix P = (pab )a,b∈S with non-negative entries, such
that ∑b pab = 1 for all a, i.e. P1 = 1. The matrix is called compatible with A, if
Aab = 0 ⇒ pab = 0.
• A probability vector is a vector p = ⟨pa : a ∈ S⟩ of non-negative entries, s.t.
∑ pa = 1
• A stationary probability vector is a probability vector p = ⟨pa : a ∈ S⟩ s.t. pP = p:
∑a pa pab = pb .
Given a probability vector p and a stochastic matrix P compatible with A, one can
define a Markov measure µ on ΣA+ (or ΣA ) by

µ[a0 , . . . , an−1 ] := pa0 pa0 a1 · · · pan−2 an−1 ,

where [a0 , . . . , an−1 ] = {x ∈ ΣA+ : xi = ai (i = 0, . . . , n − 1)}. The stochasticity of P

guarantees that this measure is finitely additive on the algebra of cylinders, and σ –
subadditivity can be checked as for Bernoulli measures using compactness. Thus
this gives a Borel probability measure on ΣA+ .
Proposition 1.8. µ is shift invariant iff p is stationary w.r.t. P. Any stochastic matrix
has a stationary probability vector.
Proof. To see the first half of the statement, we note that µ is shift invariant iff
µ[∗, b] = µ[b] for all [b], which is equivalent to

∑ pa pab0 pb0 b1 · · · pbn−2 bn−1 = pb0 pb0 b1 · · · pbn−2 bn−1 .

a

Canceling the identical terms on both sides gives ∑a pa pab0 = pb0 . Thus µ is shift
invariant iff p is P–stationary.
We now show that every stochastic matrix has a stationary probability vector.
Consider the right action of P on the simplex ∆ of probability vectors in RN :

∆ := {(x1 , . . . , xN ) : xi ≥ 0, ∑ xi = 1} , T (x) = xP.

We have T (∆ ) ⊆ ∆ , since ∑a (T x)a = ∑a ∑b xb pba = ∑b xb ∑a pba = ∑b xb = 1. Re-

call Brouwer’s fixed point theorem: A continuous mapping of a compact convex
1.5 Examples 13

subset of Rd into itself has a fixed point. Applying this to T : ∆ → ∆ we find x ∈ ∆

such that xP = x. This is the stationary probability vector. ⊔
⊓

Thus every stochastic matrix determines (at least one) shift invariant measure.
Such measures are called Markov measures We ask when is this measure ergodic,
and when is it mixing.

1.5.4.2 Ergodicity and mixing of Markov measures

There are obvious obstructions to ergodicity and mixing. To state them concisely,
we introduce some terminology. Suppose P = (pab )S×S is a stochastic matrix. We
n
say that a connects to b in n steps, and write a → − b, if there is a path of length n + 1
(a, ξ1 , . . . , ξn−1 , b) ∈ S n+1 s.t. paξ1 pξ1 ξ2 · . . . · pξn−1 b > 0 (see problem 1.5).
Definition 1.9. A stochastic matrix P = (pab )a,b∈S is called irreducible, if for every
n
a, b ∈ S there exists an n s.t. a →
− b.
n
Lemma 1.1. If A is irreducible, then p := gcd{n : a →
− a} is independent of a.
(gcd =greatest common divisor).
n n n
Proof. Let pa := gcd{n : a →
− a}, pb := gcd{n : b →
− b}, and Λb := {n : b →
− b}. Then
Λb +Λb ⊂ Λb , and therefore Λb −Λb is a subgroup of Z. Necessarily Λb −Λb = pb Z.
α β α n β
By irreducibility, there are α, β s.t. a −
→b−→ a. So for all n ∈ Λb , a −
→b→− b− → a,
whence pa | gcd(α + β + Λb ). This clearly implies that pa divides every number in
Λb − Λb = pb Z. So pa |pb . Similarly one shows that pb |pa , whence pa = pb . ⊔
⊓

Definition 1.10. The period of an irreducible stochastic matrix P is the number p :=

n
gcd{n : a →
− a} (this is independent of a by the lemma). An irreducible stochastic
matrix is called aperiodic if its period is equal to one.


For example, the SFT with transition matrix 01 10 is irreducible with period two.
If P is not irreducible, then any Markov measure of P and a strictly positive
stationary probability vector of P is non-ergodic. To see why, pick a, b ∈ S s.t. a
does not connect to b in any number of steps. The set

E := {x ∈ ΣA+ : xi ̸= b for all i sufficiently large}

is a shift invariant set which contains [a], but which is disjoint from [b]. So E is an
invariant set with such that 0 < pa ≤ µ(E) ≤ 1 − pb < 1. So µ is not ergodic.
If P is irreducible, but not aperiodic, then any Markov measure on ΣA+ is non-
mixing, because the function fR:= 1[a] satisfies f f ◦ T n ≡ 0 for all n not divisible
by the period. This means that f f ◦ T n dµ is equal to zero on a subsequence, and
therefore cannot converge to µ[a]2 .
We claim that these are the only possible obstructions to ergodicity and mixing.
The proof is based on the following fundamental fact, whose proof will be given at
the next section.
14 1 Basic definitions, examples, and constructions

Theorem 1.3 (Ergodic Theorem for Markov Chains). Suppose P is a stochastic

(n)
matrix, and write Pn = (pab ), then P has a stationary probability vector p, and
n (k)
1
1. if P is irreducible, then n ∑ pab −−−→ pb (a, b ∈ S);
k=1 n→∞
(k)
2. if P is irreducible and aperiodic, then pab −−−→ pb (a, b ∈ S).
n→∞

Corollary 1.1. A shift invariant Markov measure on a SFT ΣA+ is ergodic iff its
stochastic matrix is irreducible, and mixing iff its stochastic matrix is irreducible
and aperiodic.

Proof. Let µ be a Markov measure with stochastic matrix P and stationary proba-
bility vector p. For all cylinders [a] = [a0 , . . . , an−1 ] and [b] = [b0 , . . . , bm−1 ],
 

µ([a] ∩ σ −k [b]) = µ  [a, ξ , b] , Wℓ := {ξ = (ξ1 , . . . , ξℓ ) : [a, ξ , b] ̸= ∅}

]

ξ ∈Wk−n
(k−n)
µ[b] pa b
= µ[a] · ∑ pan−1 ξ1 · · · pξk−n b0 · = µ[a]µ[b] · n−1 0 .
ξ ∈Wk−n
pb0 pb0

If P is irreducible, then by theorem 1.3, n1 ∑n−1 −k

k=0 µ([a] ∩ σ [b]) −
−−→ µ[a]µ[b].
n→∞
We claim that this implies ergodicity. Suppose E is an invariant
fix ε > 0,
set, and
arbitrarily small. There are cylinders A1 , . . . , AN ∈ S s.t. µ E△ Ni=1 Ai < ε.1 Thus
U

N N
µ(E) = µ(E ∩ σ −k E) = ∑ µ(Ai ∩ σ −k E) ± ε = ∑ µ(Ai ∩ σ −k A j ) ± 2ε.
i=1 i, j=1

Averaging over k, and passing to the limit, we get

N
1 n −k
N
µ(E) = lim
∑ n→∞ n ∑ µ(Ai ∩ σ A j ) ± 2ε = ∑ µ(Ai )µ(A j ) ± 2ε
i, j=1 k=1 i, j=1
!2
N
= ∑ µ(Ai ) ± 2ε = [µ(E) ± ε]2 ± 2ε.
i=1

Passing to the limit ε → 0+ , we obtain µ(E) = µ(E)2 , whence µ(E) = 0 or 1.

Now assume that that P is irreducible and aperiodic. The ergodic theorem for
Markov chains says that µ([a] ∩ σ −k [b]) −−−→ µ[a]µ[b]. Since any measurable sets
k→∞
E, F can approximated by finite disjoint unions of cylinders, an argument similar to
the previous one shows that µ(E ∩ σ −k F) −−−→ µ(E)µ(F) for all E, F ∈ B. This
k→∞
is is mixing. ⊔
⊓
1 Proof: The collection of sets E satisfying this approximation property is a σ –algebra which
contains all cylinders, therefore it is equal to B.
1.5 Examples 15

Remark 1. The ergodic theorem for Markov chains can be visualized as follows.
Imagine that we distribute mass on the states of S according to a probability dis-
tribution q = (qa )a∈S . Now shift mass from one state to another using the rule that
a pab –fraction of the mass at a is moved to b. The new mass distribution is qP
(check). After n steps, the mass distribution is qPn . The previous theorem says that,
in the aperiodic case, the mass distribution converges to the stationary distribution
—- the equilibrium state. It can be shown that the rate of convergence is exponential
(problem 1.7).
Remark 2: The ergodic theorem for Markov chains has an important generaliza-
tion to all matrices with non-negative entries, see problem 1.6.

1.5.4.3 Proof of the Ergodic Theorem for Markov chains

Suppose first that P is an irreducible and aperiodic stochastic matrix. This implies
that there is some power m such that all the entries of Pm are strictly positive.2
Let N := |S| denote the number of states, and consider the set of all probability
vectors ∆ := {(x1 , . . . , xN ) : xi ≥ 0, ∑ xi = 1}. Since P is stochastic, the map T (x) =
xP maps ∆ continuously into itself. By the Brouwer Fixed Point Theorem, there is
a probability vector p s.t. pP = p (this is the stationary probability vector).
The irreducibility of P implies that all the coordinates of p are strictly positive.
Indeed, ∑ pk = 1 so at least one coordinate pi is positive. For any other coordinate
p j there is by irreducibility a path (i, ξ1 , . . . , ξn−1 , j) such that

piξ1 pξ1 ξ2 · · · pξn−2 ξn−1 pξn−1 j > 0.

(n)
So p j = (pPn ) j = ∑k pk pk j ≥ pi piξ1 pξ1 ξ2 · · · pξn−2 ξn−1 pξn−1 j > 0.
So p lies in the interior of ∆ , and the set C := ∆ − p is a compact convex neigh-
borhood of the origin such that T (C) ⊂ C , T m (C) ⊂ int[C]. (We mean the relative
interior in the (N − 1)-dimensional body C, not the ambient space RN .)
Now consider L := span(C) (an N − 1–dimensional space). This is an invariant
space for T , whence for Pt (the transpose of P). We claim that all the eigenvalues of
Pt |L have absolute value less than 1:
1. Eigenvalues of modulus larger than one are impossible, because P is stochastic,
so ∥vP∥1 ≤ ∥v∥1 , so the spectral radius of Pt cannot be more than 1.
2. Roots of unity are impossible, because in this case for some k, Pkm has a real
eigenvector v with eigenvalue one. There is no loss of generality in assuming
that v ∈ ∂C, otherwise multiply v by a suitable scalar. But Pkm cannot have fixed
points on ∂C, because Pkm (C) ⊂ int[C]
3. Eigenvalues eiθ with θ ̸∈ 2πQ are impossible, because if eiθ is an eigenvalue then
there are two real eigenvectors u, v ∈ ∂C such that the action of P on span{u, v}
2 n
Begin by proving that if A is irreducible and aperiodic, then for every a there is an Na s.t. a →
− a
n
for all n > Na . Use this to show that for all a, b there is an Nab s.t. a →
− b for all n > Nab . Take
m = max{Nab }.
16 1 Basic definitions, examples, and constructions
 
cos θ sin θ
is conjugate to , an irrational rotation. This means ∃kn → ∞ s.t.
− sin θ cos θ
uPmkn → u ∈ ∂C. But this cannot be the case because Pm [C] ⊂ int[C], and by
compactness, this cannot intersect ∂C.
In summary the spectral radius of Pt |L is less than one.
But RN = L ⊕ span{p}. If we decompose a general vector v = q + t p with q ∈
L, then the above implies that vPn = t p + O(∥Pn |L ∥)∥q∥ −−−→ t p. It follows that
n→∞
(n)
pab −−−→ pb for all a, b.
n→∞
This is almost the ergodic theorem for irreducible aperiodic Markov chains, the
only thing which remains is to show that P has a unique stationary vector. Suppose
(n)
q is another probability vector s.t. qP = q. We can write pab → pb in matrix form
as follows:
Pn −−−→ Q, where Q = (qab )S×S and qab = pb .
n→∞

This means that q = qPn → qQ, whence qQ = q, so qa = ∑b qb qba = ∑b qb pa = pa .

Consider now the periodic irreducible case. Let A be the transition matrix associ-
ated to P (with entries tab = 1 when pab > 0 and tab = 0 otherwise), and let p denote
the period of P. Fix once and for all a state v. Working with the SFT ΣA+ , we let
n
Sk := {b ∈ S : v →
− b for some n = k mod p} (k = 0, . . . , p − 1).

Sk are pairwise disjoint, because if b ∈ Sk1 ∩ Sk2 , then ∃αi = ki mod p and ∃β
αi β
s.t. v −→ b−→ v for i = 1, 2. By the definition of the period, p|αi + β for i = 1, 2,
whence k1 = α1 = −β = α2 = k2 mod p.
It is also clear that every path of length ℓ which starts at Sk , ends at Sk′ where
k′ = k + ℓ mod p. In particular, every path of length p which starts at Sk ends at Sk .
(p)
This means that if pab > 0, then a, b belong to the same Sk .
p
It follows that P is conjugate, via a coordinate permutation, to a block matrix
(p)
with blocks (pab )Sk ×Sk . Each of the blocks is stochastic, irreducible, and aperiodic.
Let π (k) denote the stationary probability vectors of the blocks.
(ℓp) (k)
By the first part of the proof, pab −−−→ πb for all a, b in the same Sk , and
ℓ→∞
(n)
pab = 0 for n ̸= 0 mod p. More generally, if a ∈ Sk1 and b ∈ Sk2 , then
(ℓp+k2 −k1 ) (k −k1 +p) ((ℓ−1)p)
lim p = lim ∑ paξ2 pξ b
ℓ→∞ ab ℓ→∞
ξ ∈Sk2
(k −k1 +p) (k2 )
= ∑ paξ2 πb (by the above)
ξ ∈Sk2
(k2 ) (k −k1 +p) (k2 ) (k −k1 +p)
= πb ∑ paξ2 = πb . (∵ paξ2 = 0 when ξ ̸∈ Sk2 )
ξ ∈S

(ℓp+α)
Similarly, pab = 0 when α ̸= k2 − k1 mod p. Thus
1.5 Examples 17

1 n (k) 1 (k)
lim
n→∞ n
∑ pab = p πb for the unique k s.t. Sk ∋ b
k=0

p (k)
The limiting vector π is a probability vector, because ∑k=1 ∑b∈Sk 1p πb = 1.
We claim that π is the unique stationary probability vector of P. The limit the-
(n) n−1 k
orem for pab can be written in the form n1 ∑k=0 P → Q where Q = (qab )S×S and
qab = πb . As before this implies that πP = π and that any probability vector q such
that qP = q, we also have qQ = q, whence q = p. ⊔
⊓

1.5.5 The geodesic flow on a hyperbolic surface

The hyperbolic plane is the surface H := {z ∈ C : Im(z) > 0} equipped with the
Riemannian metric ds = |dz|/Im(z), which gives it constant curvature (−1).
It is known that the orientation preserving isometries (i.e. distance preserving
maps) are the Möbius transformations which preserve H. They form the group
 
az + b
Möb(H) = z 7→ : a, b, c, d ∈ R, ad − bc = 1
cz + d
  
ab
≃ : a, b, c, d ∈ R, ad − bc = 1 {±id} =: PSL(2, R).
cd



(We quotient by {±id} to identify ac db , −a −b
−c −d which represent the same Möbius
transformation.)
The geodesics (i.e. length minimizing curves) on the hyperbolic plane are vertical
half–lines, or circle arcs which meet ∂ H at right angles. Here is why: Suppose ω ∈
T M is a unit tangent vector which points directly up, then it is not difficult to see
that the geodesic at direction ω is a vertical line. For general unit tangent vectors
ω, find an element ϕ ∈ Möb(H) which rotates them so that dϕ(ω) points up. The
geodesic in direction ω is the ϕ–preimage of the geodesic in direction dϕ(ω) (a
vertical half–line). Since Möbius transformations map lines to lines or circles in a
conformal way, the geodesic of ω is a circle meeting ∂ H at right angles.
The geodesic flow of H is the flow gt on the unit tangent bundle of H,

T 1 H := {tangent vectors with length one}

which moves ω along the geodesic it determines at unit speed.

To describe this flow it useful to find a convenient parametrization for T 1 H. Fix
ω0 ∈ T 1 H (e.g. the unit vector based at i and pointing up). For every ω, there is a
unique ϕω ∈ Möb(H) such that ω = dϕω (ω0 ), thus we can identify

T 1 H ≃ Möb(H) ≃ PSL(2, R).

It can be shown that in this coordinate system the geodesic flow takes the form
18 1 Basic definitions, examples, and constructions

et/2 0
    
ab ab
gt = .
cd cd 0 e−t/2
 
10
To verify this, it is enough to calculate the geodesic flow on ω0 ≃ .
01
Next we describe the Riemannian volume measure on T 1 H (up to normaliza-
tion). Such a measure must be invariant under the action of all isometries. In our
coordinate system, the isometry ϕ(z) = (az + b)/(cz + d) acts by
    
x y ab x y
ϕ = .
zw cd zw

Since PSL(2, R) is a locally compact topological group, there is only one Borel
measure on PSL(2, R) (up to normalization), which is left invariant by all left trans-
lations on the group: the Haar measure of PSL(2, R). Thus the Riemannian volume
measure is a left Haar measure of PSL(2, R), and this determines it up to normal-
ization.
It is a convenient feature of PSL(2, R) that its left Haar measure is also invari-
ant under right translations. It follows that the geodesic flow preserves the volume
measure on T 1 H. But this measure is infinite, and it is not ergodic (prove!).
To obtain ergodic flows, we need to pass to compact quotients of H. These are
called hyperbolic surfaces.
A hyperbolic surface is a Riemannian surface M such that every point in M has a
neighborhood V which is isometric to an open subset of H. Recall that a Riemannian
surface is called complete, if every geodesic can be extended indefinitely in both
directions.

Theorem 1.4 (Killing–Hopf Theorem). Every complete connected hyperbolic sur-

face is isometric to Γ \H := {Γ z : z ∈ H}, where
1. Γ is a subgroup of Möb(H) and Γ z := {g(z) : g ∈ Γ }.
2. Every point z ∈ H is in the interior of some open set U ⊂ PSL(2, R) s.t. {g(U) :
g ∈ Γ } are pairwise disjoint. So Γ z′ 7→ the unique point in Γ z′ ∩U is a bijection.
3. the Riemannian structure on {Γ z′ : z′ ∈ U} is the one induced by the Riemannian
structure on U.

If we identify Γ with a subgroup of PSL(2, R), then we get the identification

T 1 (Γ \H ) ≃ Γ \PSL(2,R) . It is clear that the Haar measure on PSL(2, R) induces a
unique locally finite measure on Γ \PSL(2,R) , and that the geodesic flow on T 1 (Γ \H )
takes the form  t/2 
t e 0
g (Γ ω) = Γ ω ,
0 e−t/2
and preserves this measure.

Definition 1.11. A measure preserving flow gt : X → X is called ergodic, if every

measurable set E such that g−t (E) = E for all t satisfies m(E) = 0 or m(E c ) = 0.
1.6 Basic constructions 19

Theorem 1.5. If Γ \H is compact, then the geodesic flow on T 1 (Γ \H ) is ergodic.

Proof. If Γ \H is compact, then Γ \H has finite volume. In this case, ergodicity is

equivalent to the following property: Every square integrable function f such that
for a.e. x, f (gt x) = f (x) for all t ∈ R, is equal a.e. to a constant function (prove!).
Consider the following flows:
 
1t
htst (Γ ω) = Γ ω
01
 
t 10
hun (Γ ω) = Γ ω
t 1
If we can show that any geodesic invariant function f is also invariant under these
flows then we are done, because it is known that
     
1t 10 λ 0
, , = PSL(2, R)
01 t 1 0 λ −1

(prove!), and any PSL(2, R)–invariant function on Γ \PSL(2,R) is constant.

Since our measure is induced by the the Haar measure, the flows htun , htst are
measure preserving. A matrix calculation shows:
−s
gs htst g−s = hte
st −−−→ id
s→∞
s
g−s htun gs = hte
un −
−−−→ id
s→−∞

−s L2
Step 1. If f ∈ L2 , then f ◦ hte −−−→ f .
s→∞

Proof. Approximate by continuous functions of compact support, and observe that

ht is an isometry of L2 .
Step 2. If f ∈ L2 and f ◦ gs = f , then f ◦ htun = f and f ◦ htst = f .
Proof. ∥ f ◦ htst − f ∥2 = ∥ f ◦ gs ◦ htst − f ∥2 = ∥ f ◦ gs ◦ htst ◦ g−s − f ∥2 → 0.
Thus f is htst –invariant. A similar calculation shows that it is htun –invariant, and we
are done. ⊔
⊓

1.6 Basic constructions

In this section we discuss several standard methods for creating new measure pre-
serving transformations from old ones. These constructions appear quite frequently
in applications.
20 1 Basic definitions, examples, and constructions

Products

Recall that the product of two measure spaces (Xi , Bi , µi ) (i = 1, 2) is the measure
space (X1 × X2 , B1 ⊗ B2 , µ1 × µ2 ) where B1 × B2 is the smallest σ –algebra which
contains all sets of the form B1 × B2 where Bi ∈ Bi , and µ1 × µ2 is the unique
measure such that (µ1 × µ2 )(B1 × B2 ) = µ1 (B1 )µ2 (B2 ).
This construction captures the idea of independence from probability theory: if
(Xi , Bi , µi ) are the probability models of two random experiments, and these experi-
ments are “independent”, then (X1 × X2 , B1 ⊗ B2 , µ1 × µ2 ) is the probability model
of the pair of the experiments, because for every E1 ∈ B1 , E2 ∈ B2 ,

F1 := E1 × X2 is the event “in experiment 1, E1 happened”

F2 := X1 × E2 is the event “in experiment 2, E2 happened”
F1 ∩ F2 = E1 × E2

and F1 ∩F2 = E1 ×E2 ; so (µ1 × µ2 )(F1 ∩F2 ) = (µ1 × µ2 )(F1 )(µ1 × µ2 )(F2 ), showing
that the events F1 , F2 are independent.

Definition 1.12. The product of two measure preserving systems (Xi , Bi , µi , Ti ) (i =

1, 2) is the measure preserving system (X1 × X2 , B1 ⊗ B2 , µ1 × µ2 , T1 × T2 ), where
(T1 × T2 )(x1 , x2 ) = (T1 x1 , T2 x2 ).

(Check that S is measure preserving.)

Proposition 1.9. The product of two ergodic mpt is not always ergodic. The product
of two mixing mpt is always mixing.

Proof. The product of two (ergodic) irrational rotations S := Rα × Rα : T2 → T2 ,

S(x, y) = (x + α, y + α) mod 1 is not ergodic: F(x, y) = y − x mod 1 is a non-
constant invariant function. (See problem 1.8.)
The product of two mixing mpt is however mixing. To see this set µ = µ1 × µ2 ,
S = T1 × T2 , and S := {A × B : A ∈ B1 , B ∈ B2 }. For any E1 := A1 × B1 , E2 :=
A2 × B2 ∈ S ,

µ(E1 ∩ S−n E2 ) = µ (A1 × B1 ) ∩ (T1 × T2 )−n (A2 × B2 )

= µ (A1 ∩ T −n A2 ) ∩ (B1 ∩ T −n B2 )

= µ1 (A1 ∩ T −n A2 )µ2 (B1 ∩ T −n B2 )

−−−→ µ1 (A1 )µ2 (B1 )µ1 (A2 )µ2 (B2 ) = µ(A1 × B1 )µ(A2 × B2 ).
n→∞

Since S is a semi-algebra which generates B1 ⊗ B2 , any element of B1 ⊗ B2 can

be approximated by a finite disjoint elements of S , and a routine approximation
argument shows that µ(E ∩ S−n F) → µ(E)µ(F) for all E, F ∈ B. ⊔
⊓
1.6 Basic constructions 21

1.6.1 Skew-products

We start with an example. Let µ be the ( 12 , 12 )–Bernoulli measure on the two shift
Σ2+ := {0, 1}N . Let f : Σ2+ → Z be the function f (x0 , x1 , . . .) = (−1)x0 . Consider the
transformation

T f : Σ2+ × Z → Σ2+ × Z , T f (x, k) = (σ (x), k + f (x)),

where σ : Σ2+ → Σ2+ is the left shift. This system preserves the (infinite) measure
µ × mZ where mZ is the counting measure on Z. The n-th iterate is

T fn (x, k) = (σ n x, k + X0 + · · · + Xn−1 ), where Xi := (−1)xi .

What we see in the second coordinate is the symmetric random walk on Z, started at
k, because (1) the steps Xi take the values ±1, and (2) {Xi } are independent because
of the choice of µ. We say that the second coordinate is a “random walk on Z driven
by the noise process (Σ2+ , B, µ, σ ).”
Here is a variation on this example. Suppose T0 , T1 are two measure preserving
transformations of the same measure space (Y, C , ν). Consider the transformation
(X ×Y, B ⊗ C , µ × ν, T f ), where

T f (x, y) = (T x, T f (x) y).

The n–th iterate takes the form T fn (x, y) = (σ n x, Txn−1 · · · Tx0 y). The second coordi-
nate looks like the random concatenation of elements of {T0 , T1 }. We say that T f is
a “random dynamical system driven by the noise process (X, B, µ, T ).”
These examples suggest the following abstract constructions.
Suppose (X, B, µ, T ) is a mpt, and G is a locally compact Polish3 topological
group, equipped with a left invariant Haar measure mG . Suppose f : X → G is mea-
surable.
Definition 1.13. The skew–product with cocycle f over the basis (X, B, µ, T ) is the
mpt (X × G, B ⊗ B(G), µ × mG , T f ), where T f : X × G → X × G is the transforma-
tion T f (x, g) = (T x, g · f (x)).
(Check, using Fubini’s theorem, that this is a mpt.) The n–th iterate T fn (x, g) =
(T n−1 x, g · f (x) · f (T x) · · · f (T n−1 x)), is a “random walk on G driven by the noise
process (X, B, µ, T ).”
Now imagine that the group G is acting in a measure preserving way on some
space (Y, C , ν). This means that there are measurable maps Tg : Y → Y such that
ν ◦ Tg−1 = ν, Tg1 g2 = Tg1 Tg2 , and (g, y) 7→ Tg (y) is a measurable from X × G to Y .

Definition 1.14. The random dynamical system on (Y, C , ν) with action {Tg : g ∈
G}, cocycle f : X → G, and noise process (X, B, µ, T ), is the system (X × Y, B ⊗
C , µ × ν, T f ) given by T f (x, y) = (T x, T f (x) y).
3 “Polish”=has a topology which makes it a complete separable metric space.
22 1 Basic definitions, examples, and constructions

(Check using Fubini’s theorem that this is measure preserving.) Here the n-th iterate
is T fn (x, y) = (T n x, T f (T n x) · · · T f (T x) T f (x) y).
It is obvious that if a skew–product (or a random dynamical system) is ergodic or
mixing, then its base is ergodic or mixing. The converse is not always true. The
ergodic properties of a skew–product depend in a subtle way on the interaction
between the base and the cocycle.
Here are two important obstructions to ergodicity and mixing for skew–products.
In what follows G is a polish group and G b is its group of characters,

b := {γ : G → S1 : γ is a continuous homomorphism}.
G

Definition 1.15. Suppose (X, B, µ, T ) is a ppt and f : X → G is Borel.

1. f is called arithmetic w.r.t. µ, if ∃h : X → S1 measurable, and γ ∈ G
b non-constant,
such that γ ◦ f = h/h ◦ T a.e.
2. f is called periodic w.r.t. µ, if ∃h : X → S1 measurable, |λ | = 1, and γ ∈ G
b non-
constant, such that γ ◦ f = λ h/h ◦ T a.e.
Proposition 1.10. Let (X, B, µ, T ) be a ppt, f : X → G Borel, and (X × G, B ⊗
B(G), µ × mG , T f ) the corresponding skew–product. If f is arithmetic, then T f is
not ergodic, and if f is periodic, then T f is not mixing.
Proof. Suppose f is arithmetic. The function F(x, y) := h(x)γ(y) satisfies

F(T x, y f (x)) = h(T x)γ(y)γ( f (x)) = h(T x)γ(y)h(x)/h(T x) = F(x, y),

and we have a non-constant invariant function. So arithmeticity⇒non-ergodicity.

Similarly, if f is periodic, then F(x, y) = h(x)γ(y) satisfies F(T x, y f (x)) = λ F(x, y),
n
whence F ◦ T f = λ F. Pick nk → ∞ s.t. λ nk → 1, then Cov(F, F ◦ T f k ) −−−→ F 2 −
R
k→∞
( F)2 . Since F ̸= F a.e., the limit is non-zero and we get a contradiction to mix-
R R

ing. So periodicity⇒non-mixing. ⊔
⊓

1.6.2 Factors

When we construct skew-products over a base, we enrich the space. A factor is a

constructions which depletes the space.
Definition 1.16. A mpt transformation (X, B, µ, T ) is called a (measure theoretic)
factor of a mpt transformation (Y, C , ν, S), if there are sets of full measure X ′ ⊂
X,Y ′ ⊂ Y such that T (X ′ ) ⊂ X ′ , S(Y ′ ) ⊂ Y ′ , and a measurable onto map π : Y ′ → X ′
such that ν ◦ π −1 = µ and π ◦ S = T ◦ π on Y ′ . We call π the factor map.
S
Y ′ −−−−→ Y′
 
π π
y y
T
X ′ −−−−→ X ′
1.6 Basic constructions 23

If (X, B, µ, T ) is a factor of (Y, C , ν, S), then it is customary to call (Y, C , ν, S)

an extension of (X, B, µ, T ) and π the factor map.
There are three principle examples:
1. Any measure theoretic isomorphism between two mpt is a factor map between
them. But some factor maps are not isomorphisms because they are not injective.
2. A skew product T f : X × Y → X × Y is an extension of its base T : X → X. The
factor map is π : X × G → X, π(x, y) = x.
3. Suppose (X, B, µ, T ) is an mpt and T is measurable w.r.t. a smaller σ –algebra
C ⊂ B (i.e. T −1 C ⊂ C ), then (X, C , µ, T ) is a factor of (X, B, µ, T ). The factor
map is the identity.
We dwell a bit more on the third example. In probability theory, σ –algebras
model information: a set E is “measurable”, if we can answer the question “is ω in
E?” using the information available to use. For example, if a real number x ∈ R is
unknown, but we can “measure” |x|, then the information we have on x is modeled
by the σ –algebra {E ⊂ R : E = −E}, because we can determined whether x ∈ E
only for symmetric sets E. By decreasing the σ –algebra, we are forgetting some
information. For example if instead of knowing |x|, we only know whether 0 ≤ |x| ≤
1 or not, then our σ –algebra is the finite σ –algebra {∅, R, [−1, 1], R \ [−1, 1]}.
Here is a typical example. Suppose we have a dynamical system (X, B, µ, T ),
and we cannot “measure” x, but we can “measure” f (x) for some measurable
f : X → R. Then the information we have by observing the dynamical system is
encoded in the smallest σ –algebra C ⊂ B with respect to which f ◦ T n are all
measurable.4 The dynamical properties we feel in this case are those of the fac-
tor (X, C , µ, T ) and not of the (X, B, µ, T ). For example, it could be the case that
µ(E ∩ T −n F) → µ(E)µ(F) for all E, F ∈ C but not for all E, F ∈ B — and then we
will observe “mixing” simply because our information is not sufficient to observe
the non-mixing in the system.

1.6.3 The natural extension

An invertible mpt is a mpt (X, B, µ, T ) such that for some invariant set X ′ ⊂ X of full
measure, T : X ′ → X ′ is invertible, and T, T −1 : X ′ → X ′ are measurable. Invertible
mpt are more convenient to handle than general mpt because they have the following
properties: If (X, B, µ) is a complete measure space, then the forward image of a
measurable set is measurable, and for any countable collection of measurable sets
T T
Ai , T ( Ai ) = T (Ai ) up to a set of measure zero. This is not true in general for
non-invertible maps. Luckily, every “reasonable” non-invertible mpt is the factor of
an invertible mpt. Here is the construction.

4 Such a σ –algebra exists: take the intersection of all sub-σ –algebras which make f ◦ T n all mea-
surable, and note that this intersection is not empty because it contains B.
24 1 Basic definitions, examples, and constructions

Definition 1.17. Suppose (X, B, µ, T ) is a ppt defined on a Lebesgue measure

space, and assume T (X) = X. The natural extension of (X, B, µ, T ) is the system
e B,
(X, e µe , Te), where
1. Xe := {x = (..., x−1 , x0 , x1 , x2 , . . .) : xi ∈ X, T (xi ) = xi+1 for all i};
2. B e is the smallest σ -algebra which contains all sets of the form {x ∈ Xe : xi ∈ E}
with i ≤ 0 and E ∈ B (see below);
3. µe is the unique probability measure on B e such that µ{x ∈ Xe : xi ∈ Ei } = µ(Ei )
for all i ≤ 0 and Ei ∈ T B; −i

4. Te is the left shift.

Lemma 1.2. The measure µ

e exists and is unique.

Proof (the proof can be omitted on first reading). Let S denote the collection of all
sets of the form [E−n , . . . , E0 ] := {x ∈ Xe : x−i ∈ E−i (i = 0, . . . , n)} where n ≥ 0 and
E−n , . . . , E−1 , E0 ∈ B. We call the elements of S cylinders.
It is easy to see that S is a semi-algebra. Our plan is to define µ e on S and then
apply Carathéodory’s extension theorem. To do this we first observe the following
important identity:
n
T −(n−i) E−i }.
\
[E−n , . . . , E0 ] = {x ∈ Xe : x−n ∈ (1.3)
i=0

The inclusion ⊆ is because for every x ∈ [E−n , . . . , E0 ], x−n ≡ T −(n−i) x−i ∈ T −(n−i) E−i
by the definition of X; e and ⊇ is because if x−n ∈ T −(n−i) E−i then x−i ≡ T (n−i) (x−n ) ∈
E−i for i = 0, . . . , n. Motivated by this identity we define µ e : S → [0, ∞] by
!
n
T −(n−i) E−i .
\
e [E−n , . . . , E0 ] := µ
µ
i=0

e (X)
Since µ e = µ(X) = 1, µ e is σ -finite on X.
e We will check that µ
e is σ -additive on
S and deduce the lemma from Carathéodory’s extension theorem. We begin with
simpler finite statements.
S TEP 1. Suppose C1 , . . . ,Cα are pairwise disjoint cylinders and D1 , . . . , Dβ are pair-
Uβ
wise disjoint cylinders. If αi=1 Ci = i=1 Di , then ∑ µ
U
e (Ci ) = ∑ µ
e (Di ).
Proof. Notice that [E−n , . . . , E0 ] ≡ [X, . . . , X, E−n , . . . , E0 ] and (by the T -invariance
e [E−n , . . . , E0 ] = µ
of µ) µ e [X, . . . , X, E−n , . . . , E0 ], no matter how many X’s we add to
the left. So there is no loss of generality in assuming that Ci , Di all have the same
(i) (i) (i) (i)
length: Ci = [C−n , . . . ,C0 ], D j = [D−n , . . . , D0 ].
By (1.3), since Ci are pairwise disjoint, −(n−k)C(i) are pairwise disjoint.
Tn
k=0 T
−k
(i)
Similarly, nk=0 T −(n−k) D−k are pairwise disjoint. Since αi=1 Ci = i=1 Di , the iden-
T U Uβ
(i) (i)
tity (1.3) also implies that αi=1 nk=0 T −(n−k)C−k = j=1 nk=0 T −(n−k) D−k . So
U T Uβ T
1.6 Basic constructions 25
! !
α α n β n α
(i) ( j)
T −(n−k)C−k −(n−k)
\ \
∑ µe (Ci ) = ∑ µ = ∑µ T D−k = ∑ µe (D j ),
i=1 i=1 k=0 j=1 k=0 j=1

which proves our claim.

Un
S TEP 2. Suppose C,C1 , . . . ,Cn are cylinders. If C ⊂ e (C) ≤ ∑ni=1 µ
i=1 Ci , then µ e (Ci ).
Proof. Notice that ni=1 Ci = ni=1 (Ci ∩ i−1 j=1 C j ). Using the fact that S is a semi-
S U cT
Ti−1 c
algebra, it is not difficult to see that Ci ∩ j=1 C j is a finite pairwise disjoint union
Uni
of cylinders: Ci ∩ i−1 c
T
j=1 C j = k=1 Cik . So

n
[ ni
n ]
] ni
]
Ci = Cik where Cik ⊂ Ci .
i=1 i=1 k=1 k=1

Look at the inclusion C ⊂ ni=1 Ci ≡ ni=1 nk=1 i

S U U
Cik . The set difference of the two
sides of the equation is a finite pairwise disjoint union
of cylinders because S is a
semi-algebra.USo by step 1 µ e (C) ≤ ∑ni=1 ∑nk=1
i e (Cik ) .
µ
ni
Similarly, k=1 Cik ⊂ Ci and the set difference of the two sides of the inclusion is
a finite pairwise disjoint union of cylinders. So by step 1, ∑nj=1 i e (Cik ) ≤ µ
µ e (Ci ). In
summary µ e (C) ≤ ∑ µe (Ci ).
(k) (k)
We are finally ready to prove σ -additivity. Suppose Ck = [E−n
Uk
, . . . , E0 ] are
a countable collection of pairwise disjoint cylinders such that ∞ k=1 Ck = C =
[E−n , . . . , E0 ]. Our aim is to show that µ(C) = ∑ µ(Ck ).
S TEP 3. µ(C) ≥ ∑ µ(Ck )
Proof. Without loss of generality nk ≥ n for all k, otherwise we can replace Ck by
(k) (k)
the equal cylinder [X, . . . , X, E−nk , . . . , E0 ] with X repeated n − nk − 1 times.
For every K, let NK := max{nk : 1 ≤ k ≤ K}, then NK ≥ n. Define E−m := X for
(k)
m > n and E−m := X for m > nk . If m > NK and 1 ≤ k ≤ K then C = [E−m , . . . , E0 ]
(k) (k) (k) (k)
and Ck = [E−m , . . . , E0 ]. So [E−m , . . . , E0 ] ⊇ Kk=1 [E−m , . . . , E0 ].
U
m T−(m−i) E ⊇ K U Tm −(m−i) E (k) , whence
ItTnow follows from (1.3)that i=0 T −i  k=1 i=0 T −i
m −(m−i) K m −(m−i) (k)
E−i ≥ µ
U T
µ i=0 T k=1 i=0 T E−i .
Since µ is σ -additive on B and µ ◦ T −1 = µ, the left-hand-side equals µ
e (C) and
K e
the right-hand-side equals ∑k=1 µ (Ck ).
S TEP 2. µ(C) ≤ ∑ µ(Ck ).
Proof. It is here that we use the assumption that (X, B, µ) is a Lebesgue measure
space. The assumption allows us to assume without loss of generality that X is a
compact metric space and B is the completion of the Borel σ -algebra (because
any union of an interval and a countable collection of atoms can be isomorphically
embedded in such a space). In this case µ is regular: for every E ∈ B and for
every ε > 0 there is an open set U and a compact set F such that F ⊂ E ⊂ U and
µ(E \ F), µ(U \ E) < ε.
26 1 Basic definitions, examples, and constructions

In particular, given ε > 0 there is no problem in finding a compact set F ⊂

Tn −(n−i) E and open sets U ⊃ Tnk T −(nk −i) E (k) open so that
i=0 T −i k i=0 −i

ε
µ(F) ≥ µ
e (C) − ε and µ(Uk ) ≤ µ
e (Ck ) + .
2k
U∞ S∞
By (1.3), [F, X, . . . , X ] ⊂ C = k=1 Ck ⊂ k=1 [Uk , X, . . . , X ], and with respect to
| {z } | {z }
n nk
the product topology on X, e [F, X, . . . , X] is compact and [Uk , X, . . . , X] are open. So
there is a finite N s.t. [F, X, . . . , X ] ⊂ Nk=1 [Uk , X, . . . , X ]. By step 2
S
| {z } | {z }
n nk

N
e [F, X, . . . , X ] ≤ ∑ µ
µ e [Uk , X, . . . , X ]
| {z } k=1 | {z }
n nk

So µ(F) ≤ ∑Nk=1 µ(Uk ), whence µe (C) ≤ ∑Nk=1 µ

e (Ck ) + 2ε ≤ ∑∞ e (Ck ) + 2ε.
k=1 µ
The step follows by taking ε → 0. ⊔
⊓

Theorem 1.6. The natural extension of (X, B, µ, T ) is an invertible extension of

(X, B, µ, T ), and is the factor of any other invertible extension of (X, B, µ, T ). Te is
ergodic iff T is ergodic, and Te is mixing iff T is mixing.

Proof. It is clear that the natural extension is an invertible ppt. Let π : Xe → X denote
the map π(x) = x0 , then π is measurable and π ◦ Te = T ◦ π. Since T X = X, every
point has a pre-image, and so π is onto. Finally, for every E ∈ B,

e (π −1 (E)) = µ
µ x ∈ Xe : x0 ∈ E}) = µ(E)
e ({e

by construction. So (X, e B,
e µe , Te) is an invertible extension of (X, B, µ, T ).
Suppose (Y, C , ν, S) is another invertible extension, and let πY : Y → X be the
factor map (defined a.e. on Y ). We show that (Y, C , ν, S) extends (X, e B,
e µe , Te).
Let (Ye , Ce, νe, S)
e be the natural extension of (Y, C , ν, S). It is isomorphic to
(Y, C , ν, S), with the isomorphism given by ϑ (y) = (yk )k∈Z , yk := Sk (y). Thus it
is enough to show that (X, e B,e µ e , Te) is a factor of (Ye , Ce, νe, Te). Here is the factor
map: θ : (yk )k∈Z 7→ (πY (yk ))k∈Z .
If Te is ergodic, then T is ergodic, because every T -invariant set E lifts to a Te-
invariant set Ee := π −1 (E). The ergodicity of Te implies that µ e (E)
e = 0 or 1, whence
µ(E) = µ −1
e (π (E)) = µ e (E) = 0 or 1.
e
To see the converse (T is ergodic ⇒ Te is ergodic) we make use of the following
observation:
 
C LAIM : Let B en := {e x ∈ Xe : x−n ∈ E} : E ∈ B , then

1. B
en are σ -algebras
2. B
e1 ⊂ Be2 ⊂ · · · and S B
n≥0 n generate B
e e
1.6 Basic constructions 27

3. Te−1 (Ben ) ⊂ B e B
en and (X, en , µ
e , Te) is a ppt
e B
4. if T is ergodic then (X, en , µ
e , Te) is ergodic.
Proof. We leave the first three items as exercises to the reader. To see the last item
suppose T is ergodic and Ee ∈ B en is Te-invariant. By the definition of B en , Ee = {e
x:
xe−n ∈ E} with E ∈ B, and it is not difficult to see that E must be T -invariant. Since
T is ergodic, µ(E) = 0 or 1. So µ e (E)
e = µ(E) = 0 or 1. So (X, e B en , µ
e , Te) is ergodic.
We can now prove the ergodicity of (X, e B,
e µe , Te) as follows. Suppose fe is abso-
lutely integrable and T -invariant, and let
e

fen := E( fe|B
en )

(readers who are not familiar with conditional expectations can find their definition
in section 2.3.1).
We claim that fen ◦ Te = fen . This is because for every bounded B
en –measurable test
function ϕ,
e
Z Z Z Z
(1) (2)
[ϕe · fen ◦ Te−1 ]d µ
e= (ϕe · fen ◦ Te−1 ) ◦ Ted µ
e= ϕe ◦ Te · fen d µ
e= ϕe ◦ Te · fed µ
e
Z Z Z
(3) (4) (5)
= (ϕe ◦ Te · fe◦ Te)d µ
e= e=
ϕe fed µ e.
ϕe fen d µ

Justifications: (1) is because µe ◦ Te−1 = µ e ; (2) is because fen = E( fe|B

en ) and ϕe ◦ Te
is Bn –measurable by part (3) of the claim; (3) is because fe◦ Te = fe; (4) is because
e
e ◦ Te−1 = µ
µ e ; and (5) is by the definition of the conditional expectation.
In summary [ϕe · fen ◦ Te−1 ]d µ e for all bounded B
R R
e = ϕe fen d µ en –measurable func-
tions, whence fn ◦ T = fn a.e. as claimed.
e e e
We saw above that (X, e B e , Te) is ergodic. Therefore, the invariance B
en , µ en –
R
measurability of fen imply that fen = fed µ e almost everywhere. By the martingale
convergence theorem,

fen −−−→ fe almost everywhere.

n→∞

So fe is constant almost everywhere. This shows that every Te-invariant integrable

function is constant a.e., so Te is ergodic.
Next we consider the mixing of T , Te. Te mixing ⇒ T mixing because for ev-
ery A, B ∈ B, µ(A ∩ T −n B) = µ
e (π −1 A ∩ T −n π −1 B) −−−→ µ
e (π −1 A)µ
e (π −1 B) ≡
n→∞
µ(A)µ(B). We show the other direction. Suppose T is mixing and consider A, e Be ∈
Bn with Bn defined as before. Then A = {e
e x : x−n ∈ A} and B = {e
e x : x−n ∈ B} with
A, B ∈ B and using the identity (Texe)i = T (e
xi ) it is easy to see that

e ∩ Te−k Be = {x ∈ Xe : xe−n ∈ A ∩ T −k B}.

A

So µ e ∩ Te−k B)
e (A e = µ(A ∩ T −k B) −−−→ µ(A)µ(B) ≡ µ
e (A)
eµ e (B).
e
k→∞
28 1 Basic definitions, examples, and constructions

e Be ∈ Bn . To get mixing for general A,

This proves mixing for A, e Be ∈ B we note
Bn generates B
e so for any Ee and for every ε > 0 there are n and Ee′ ∈ B
S e en such
e Ee′ ) < ε (because the collection of sets Ee with such n and Ee′ forms a
e (E△
that µ
σ -algebra which contains B
S e
n ). A standard approximation argument now shows
that µ e ∩ Te−k B)
e (A e −−−→ µ
e (A)
eµ e (B) e Be ∈ B.
e for all A, e ⊔
⊓
k→∞

1.6.4 Induced transformations

Suppose (X, B, µ, T ) is a probability preserving transformation, and let A ∈ B be a

set of positive measure. By Poincaré’s Recurrence Theorem, for a.e. x ∈ A there is
some n ≥ 1 such that T n (x) ∈ A. Define

ϕA (x) := min{n ≥ 1 : T n x ∈ A},

with the minimum of the empty set being understood as infinity. Note that ϕA < ∞
a.e. on A, hence A0 := {x ∈ A : ϕA (x) < ∞} is equal to A up to a set of measure zero.
Definition 1.18. The induced transformation on A is (A0 , B(A), µA , TA ), where
A0 := {x ∈ A : ϕA (x) < ∞}, B(A) := {E ∩ A0 : E ∈ B}, µA is the measure µA (E) :=
µ(E|A) = µ(E ∩ A)/µ(A), and TA : A0 → A0 is TA (x) = T ϕA (x) (x).
Theorem 1.7. Suppose (X, B, µ, T ) is a ppt, and A ∈ B has positive finite measure.
1. µA ◦ TA−1 = µA ;
2. if T is ergodic, then TA is ergodic (but the mixing of T ̸⇒ the mixing of TA );
R ϕA −1
f ◦ T k dµ for every f ∈
R
3. Kac Formula: If µ is ergodic, then f dµ = A ∑k=0
1
R
L (X). In particular A ϕA dµA = 1/µ(A).
Proof. Given E ⊂ A measurable, µ(E) = µ(T −1 E ∩ A) + µ(T −1 E ∩ Ac ) =
| {z }
µ(TA−1 E∩[ϕA =1])

= µ(T −1 E ∩ A) + µ(T −2 E ∩ T −1 Ac ∩ A) + µ(T −2 E ∩ T −1 Ac ∩ Ac )

| {z } | {z }
µ(TA−1 E∩[ϕA =1]) µ(TA−1 E∩[ϕA =2])
N N−1
∑ µ(TA−1 E ∩ [ϕA = j]) + µ(T −N E ∩ T − j Ac ).
\
= ··· =
j=1 j=0

Passing to the limit as N → ∞, we see that µ(E) ≥ µA (TA−1 E). Working with A \
E, and using the assumption that µ(X) < ∞, we get that µ(A) − µ(E) ≥ µ(A) −
µ(TA−1 E) whence µ(E) = µ(TA−1 E). Since µA is proportional to µ on B(A), we get
µA = µA ◦ TA−1 .
We assume that T is ergodic, and prove that TA is ergodic. The set

Ω := {x : T n (x) ∈ A for infinitely many n ≥ 0}

1.6 Basic constructions 29

is a T –invariant set of non-zero measure (bounded below by µ(A)), so it must has

full measure. Thus a.e. x ∈ X has some n ≥ 0 s.t. T n (x) ∈ A, and

rA (x) := min{n ≥ 0 : T n x ∈ A} < ∞ a.e. in X.

Suppose f : A0 → R is a TA –invariant L2 –function. Define

F(x) := f (T rA (x) x).

This makes sense a.e. in X, because rA < ∞ almost everywhere. This function is T –
invariant, because either x, T x ∈ A and then F(T x) = f (T x) = f (TA x) = f (x) = F(x)
or one of x, T x is outside A and then F(T x) = f (T rA (T x) T x) = f (T rA (x) x) = F(x).
Since T is ergodic, F is constant a.e. on X, and therefore f = F|A is constant a.e. on
A. Thus the ergodicity of T implies the ergodicity of TA .
Here is an example showing that the mixing of T does not imply the mixing of
TA . Let Σ + be a SFT with states {a, 1, 2, b} and allowed transitions

a → 1; 1 → 1, b; b → 2; 2 → a.

Let A = {x : x0 = a, b}. Any shift invariant Markov measure µ on Σ + is mixing,

because Σ + is irreducible and aperiodic (1 → 1). But TA is not mixing, because
TA [a] = [b] and TA [b] = [a], so [a] ∩ TA−n [a] = ∅ for all n odd.
Next we prove the Kac formula. Suppose first that f ∈ L∞ (X, B, µ) and f ≥ 0.
Z Z Z Z Z
f dµ = f dµ + f · 1Ac dµ = f dµ + f ◦ T · 1T −1 Ac dµ
ZA Z A
Z
= f dµ + f ◦ T · 1T −1 Ac ∩A dµ + f ◦ T · 1T −1 Ac ∩Ac dµ
ZA Z Z
= f dµ + f ◦ T · 1[ϕA >1] dµ + f ◦ T 2 · 1T −2 Ac ∩T −1 Ac dµ
A A
Z N−1 Z
= ··· = ∑ f ◦ T j · 1[ϕA > j] dµ + f ◦ T N · 1TN − j Ac dµ.
A j=0 j=1 T

The first term tends, as N → ∞, to

Z ∞ ∞ Z ϕA −1
j
∑ f ◦T ∑ 1[ϕA =i] dµ ≡ ∑ f ◦ T j dµ.
A j=0 i= j+1 A j=0

The second term is bounded by ∥ f ∥∞ µ{x : T j (x) ̸∈ A for all k ≤ N}. This bound
tends to zero, because µ{x : T j (x) ̸∈ A for all k} = 0 because T is ergodic and re-
current (fill in the details). This proves the Kac formula for all L∞ functions.
Every non-negative L1 –function is the increasing limit of L∞ functions. By the
monotone convergence theorem, the Kac formula must hold for all non-negative
L1 –function. Every L1 –function is the difference of two non-negative L1 –functions
( f = f · 1[ f >0] − | f | · 1[ f <0] ). It follows that the Kac formula holds for all f ∈ L1 . ⊔
⊓
30 1 Basic definitions, examples, and constructions

1.6.5 Suspensions and Kakutani skyscrapers

The operation of inducing can be “inverted”, as follows. Let (X, B, µ, T ) be a ppt,

and r : X → N an integrable measurable function.
Definition 1.19. The Kakutani skyscraper with base (X, B, µ, T ) and height func-
tion r is the system (Xr ), B(Xr ), ν, S), where
1. Xr := {(x, n) : x ∈ X, 0 ≤ n ≤ r(x) − 1};
2. B(Xr ) = {E ∈ B(X) ⊗ B(N) : E ⊆ Xr }, where B(N) R= 2N ;
3. ν is the unique measure such that ν(B × {k}) = µ(B)/ rdµ;
4. S is defined by S(x, n) = (x, n+1), when n < r(x)−1, and S(x, n) = (T x, 0), when
n = r(x) − 1.
(Check that this is a ppt.)
We think of Xr as a skyscraper made of stories {(x, k) : r(x) > k}; the orbits of S
climb up the skyscraper until the reach the top floor possible, and then move to the
ground floor according to T .
If we induce a Kakutani skyscraper on {(x, 0) : x ∈ X}, we get a system which is
isomorphic to (X, B, µ, T ).

Proposition 1.11. A Kakutani skyscraper over an ergodic base is ergodic, but there
are non-mixing skyscrapers over mixing bases.

The proof is left as an exercise.

There is a straightforward important continuous–time version of this construc-
tion: Suppose (X, B, µ, T ) is a ppt, and r : X → R+ is a measurable function such
that inf r > 0.

Definition 1.20. The suspension semi-flow with base (X, B, µ, T ) and height func-
tion r is the semi-flow (Xr , B(Xr ), ν, Ts ), where
1. Xr := {(x,t) ∈ X × R : 0 ≤ t < r(x)};
2. B(Xr ) = {E ∈ B(X) ⊗ B(R) : E ⊆ Xr };
R R R r(x) R
3. ν is the measure such that Xr f dν = X 0 f (x,t)dtdµ(x)/ X rdµ;
n−1 n−1
4. Ts (x,t) = (T n x,t + s − ∑ r(T k x)), where n is s.t. 0 ≤ t + s − ∑ r(T k x) < r(T n x).
k=0 k=0

(Check that this is a measure preserving semi-flow.)

Suspension flows appear in applications in the following way. Imagine a flow
Tt on a manifold X. It is often possible to find a submanifold S ⊂ X such that (al-
most) every orbit of the flow intersects S transversally infinitely many times. Such
a submanifold is called a Poincaré section. If it exists, then one can define a map
TS : S → S which maps x ∈ S into Tt x with t := min{s > 0 : Ts (x) ∈ S}. This map is
called the Section map. The flow itself is isomorphic to a suspension flow over its
Poincaré section.
 1.6 Basic constructions 31

Problems

1.1. Proof of Liouville’s theorem in section 1.1

(a) Write x := (q, p) and y := Tt (q, p). Use Hamilton’s equations to show that the
∂y
Jacobian matrix of y = y(x) satisfies ∂ x = I + tA + O(t 2 ) as t → 0, where tr(A) = 0.
(b) Show that for every matrix A, det(I + tA + O(t 2 )) = 1 + ttr(A) + O(t 2 ) as t → 0.
(c) Prove that the Jacobian of Tt is equal to one for all t. Deduce Liouville’s theorem.

1.2. The completion of a measure space. Suppose (X, B, µ) is a measure space. A

set N ⊂ X is called a null set, if there is a measurable set E ⊇ N such that µ(E) = 0.
A measure space is called complete, if every null set is measurable. Every measure
space can be completed, and this exercise shows how to do this.
(a) Let B0 denote the the collection of all sets of the form E ∪ N where E ∈ B and
N is a null set. Show that B0 is a σ –algebra.
(b) Show that µ has a unique extension to a σ –additive measure on B0 .

1.3. Prove Poincaré’s Recurrence Theorem for a general probability preserving

transformation (theorem 1.1).

1.4. Fill in the details in the proof above that the Markov measure corresponding to
a stationary probability vector and a stochastic matrix exists, and is shift invariant
measure.

1.5. Suppose ΣA+ is a SFT with stochastic matrix P. Let A = (tab )S×S denote the
matrix of zeroes and ones where tab = 1 if pab > 0 and tab = 0 otherwise. Write
(n) (n)
An = (tab ). Prove that tab is the number of paths of length n starting at a and ending
n (n)
at b. In particular: a →
− b ⇔ tab > 0.

1.6. The Perron–Frobenius Theorem5 : Suppose A = (ai j ) is a matrix all of whose

entries are non-negative, and let B := (bi j ) be the matrix bi j = 1 if ai j > 0 and
bi j = 0 if ai j = 0. Assume that B is irreducible, then A has a positive eigenvalue λ
with the following properties:
(i) There are positive vectors r and ℓ s.t. ℓA = λ ℓ, Ar = λ r.
(ii) The eigenvalue λ is simple.
(iii) The spectrum of λ −1 A consists of 1, several (possibly zero) roots of unity, and a
finite subset of the open unit disc. In this case the limit lim 1n ∑n−1 −k k
k=0 λ A exists.
n→∞
(iv) If B is irreducible and aperiodic, then the spectrum of λ −1 A consists of 1 and a
finite subset of the open unit disc. In this case the limit lim λ −n An exists.
n→∞

1. Prove the Perron–Frobenius theorem in case A is stochastic, first in the aperiodic

case, then in the general case.
2. Now consider the case of a non-negative matrix:
5 Perron first proved this in the aperiodic case. Frobenius later treated the periodic irreducible case.
32 1 Basic definitions, examples, and constructions

a. Use a fixed point theorem to show that λ , ℓ, r exist;

b. Set 1 := (1, . . . , 1) and let V be the diagonal matrix such that V 1 = r. Prove
that λ −1V −1 AV is stochastic.
c. Prove the Perron-Frobenius theorem.
(n)
1.7. Suppose P = (pab )S×S is an irreducible aperiodic stochastic matrix. Use the
(n)
spectral description of P obtained in problem 1.6 to show that pab → pb exponen-
tially fast.
1.8. Show that the product of n irrational rotations Rα1 , . . . , Rαn is ergodic iff
(α1 , . . . , αn ) are independent over the irrationals.
1.9. Suppose gt : X → X is a measure preserving flow. The time one map of the
flow is the measure preserving map g1 : X → X. Give an example of an ergodic flow
whose time one map is not ergodic.
1.10. The adding machine
Let X = {0, 1}N equipped with the σ –algebra B generated by the cylinders, and the
Bernoulli ( 21 , 21 )–measure µ. The adding machine is the ppt (X, B, µ, T ) defined by
the rule T (1n 0∗) = (0n 1∗), T (1∞ ) = 0∞ . Prove that the adding machine is invertible
and probability preserving. Show that T (x) = x ⊕ (10∞ ) where ⊕ is “addition with
carry to the right”.
1.11. Prove proposition 1.11.
1.12. Show that a ppt (X, B, µ, T ) is mixing whenever µ(A ∩ T −n A) −−−→ µ(A)2
n→∞
for all A ∈ B. Guidence:
1A f ◦ T n dµ −−−→ µ(A) f dµ for all f ∈ spanL2 {1, 1A ◦ T, 1A ◦ T 2 , . . .}.
R R
1.
n→∞
1A f ◦ T n dµ −−−→ µ(A) f dµ for all f ∈ L2 .
R R
2.
n→∞ R
g f ◦ T n dµ −−−→ gdµ f dµ for all f , g ∈ L2 .
R R
3.
n→∞
1.13. Show that a Kakutani skyscraper over an invertible transformation is invert-
ible, and find a formula for its inverse.
1.14. Conservativity
Let (X, B, µ, T ) be a measure preserving transformation on an infinite σ –finite,
measure space.6 A set W ∈ B is called wandering, if {T −nW : n ≥ 0} are pairwise
disjoint. A mpt is called conservative, if every wandering set has measure zero.
1. Show that any ppt is conservative. Give an example of a non-conservative mpt
on a σ –finite infinite measure space.
2. Show that Poincaré’s recurrence theorem extends to conservative mpt.
3. Suppose (X, B, µ, T ) is a conservative ergodic mpt, and let A be a set of fi-
nite positive measure. Show that the induced transformation TA : A → A is well–
defined a.e. on A, and is an ergodic ppt.
4. Prove Kac formula for conservative ergodic transformations under the previous
set of assumptions.
6 A measure space is called σ –finite, if its sample space is the countable union of finite measure
sets.
References 33

Notes for chapter 1

The standard references for measure theory are [7] and [4], and the standard refer-
ences for ergodic theory of probability preserving transformations are [6] and [8].
For ergodic theory on infinite measure spaces, see [1]. Our proof of the Perron-
Frobenius theorem is taken from [3]. Kac’s formula has very simple proof when T
is invertible. The proof we use (taken from [5]) works for non-invertible transfor-
mations, and extends to the conservative infinite measure setting. The ergodicity of
the geodesic flow was first proved by E. Hopf by other means. The short proof we
gave is due to Gelfand & Fomin and is reproduced in [2].

References

1. Aaronson, J.: Introduction to infinite ergodic theory, Mathematical Surveys and Monographs
50, AMS, 1997. xii+284pp.
2. Bekka, M.B. and Mayer, M.: Ergodic theory and topological dynamics of group actions on
homogeneous spaces. London Mathematical Society LNM 269, 2000. x+200pp.
3. Brin, M. and Stuck, G.: Introduction to dynamical systems. Cambridge University Press,
Cambridge, 2002. xii+240 pp.
4. Halmos, P. R.: Measure Theory. D. Van Nostrand Company, Inc., New York, N. Y., 1950.
xi+304 pp.
5. Krengel, U.: Ergodic theorems. de Gruyter Studies in Mathematics 6 1985. viii+357pp.
6. Petersen, K.: Ergodic theory. Corrected reprint of the 1983 original. Cambridge Studies in
Advanced Mathematics, 2. Cambridge University Press, Cambridge, 1989. xii+329 pp.
7. Royden, H. L.: Real analysis. Third edition. Macmillan Publishing Company, New York,
1988. xx+444 pp.
8. Walters, P.: An introduction to ergodic theory. Graduate Texts in Mathematics, 79. Springer-
Verlag, New York-Berlin, 1982. ix+250 pp.
Chapter 2
Ergodic Theorems

2.1 The Mean Ergodic Theorem

Theorem 2.1 (von Neumann’s Mean Ergodic Theorem). Suppose (X, B, µ, T ) is

L2
a ppt. If f ∈ L2 , then 1
N ∑N−1
k=0 f ◦T − → f where f ∈ L2 is invariant. If T is ergodic,
k −−
R n→∞
then f = f dµ.
R R
Proof. Observe that since T is measure preserving, then f ◦ T dµ = f dµ for
every f ∈ L1 , and ∥ f ◦ T ∥2 = ∥ f ∥2 for all f ∈ L2 (prove this, first for indicator
functions, then for all L2 –functions).
Suppose f = g − g ◦ T where g ∈ L2 (in this case we say that f is a coboundary
with transfer function g ∈ L2 ), then it is obvious that

1 N−1 1
∑ f ◦Tk
N k=0
=
N
∥g ◦ T n − g∥2 ≤ 2∥g∥2 /N −−−→ 0.
N→∞
2

Thus the theorem holds for all elements of C := {g − g ◦ T : g ∈ L2 }.

We claim that the theorem holds for all elements of C (L2 –closure). Suppose
f ∈ C , then for every ε > 0, there is an F ∈ C s.t. ∥ f − F∥2 < ε. Choose N0 such
that for every N > N0 , ∥ N1 ∑N−1 k
k=0 F ◦ T ∥2 < ε, then for all N > N0

1 N−1 1 N−1 1 N−1

∑ f ◦Tk
N k=0
≤ ∑ ( f − F) ◦ T k
N k=0
+ ∑ F ◦Tk
N k=0
2 2 2
1 N−1
≤ ∑ ∥( f − F) ◦ T k ∥2 + ε < 2ε.
N k=0

This shows that 1

N ∑N−1
k=0 f ◦ T
k −−−→ 0.
2 N→∞
⊥
Next we claim that C = {invariant functions}. Suppose f ⊥ C , then

35
36 2 Ergodic theorems

∥ f − f ◦ T ∥22 = ⟨ f − f ◦ T, f − f ◦ T ⟩ = ∥ f ∥2 − 2⟨ f , f ◦ T ⟩ + ∥ f ◦ T ∥22
= 2∥ f ∥22 − 2⟨ f , f − ( f − f ◦ T )⟩ = 2∥ f ∥22 − 2∥ f ∥22 = 0 =⇒ f = f ◦ T a.e.
⊥
So C ⊆ {invariant functions}. Conversely, if f is invariant then for every g ∈ L2

⟨ f , g − g ◦ T ⟩ = ⟨ f , g⟩ − ⟨ f , g ◦ T ⟩ = ⟨ f , g⟩ − ⟨ f ◦ T, g ◦ T ⟩ = ⟨ f , g⟩ − ⟨ f , g⟩ = 0,

so f ⊥ C , whence f ⊥ C . In summary, C = {invariant functions}, and

L2 = C ⊕ {invariant functions}.

We saw above that the MET holds for all elements of C with zero limit, and
holds for all invariant functions f with limit f . Therefore the MET holds for all L2 –
functions, and the limit f is the orthogonal projection of f on the space of invariant
functions. R
In particular f is invariant. If T is ergodic, then f is constant and f = f dµ
1 N
almost everywhere. Also, since N ∑n=1 f ◦ T n → f in L2 , then

1 N 1 N
Z Z
f dµ = ∑ ⟨1, f ◦ T n ⟩ = ⟨1, ∑ f ◦ T n ⟩ → ⟨1, f ⟩ = f dµ,
N n=1 N n=1
R R R
so f dµ = f dµ whence f = f dµ almost everywhere. ⊔
⊓

Remark 1. The proof shows that the limit f is the projection of f on the space of
invariant functions.
Remark 2. The proof only uses the fact that U f = f ◦T is an isometry of L2 . In fact
it works for all linear operators U : H → H on separable Hilbert spaces s.t. ∥U∥ ≤ 1,
see problem 2.1.
L2
Remark 3. If fn −−−→ f , then ⟨ fn , g⟩ −−−→ ⟨ f , g⟩ for all g ∈ L2 . Specializing to the
n→∞ n→∞
case fn = 1n ∑n−1 k
k=0 1B ◦ T , g = 1A we obtain the following corollary of the MET:

Corollary 2.1. A ppt (X, B, µ, T ) is ergodic iff for all A, B ∈ B,

1 n−1
∑ µ(A ∩ T −k B) −n→∞
n k=0
−−→ µ(A)µ(B).

So ergodicity is mixing “on the average.” We will return to this point when we
discuss the definition of weak mixing in the next chapter.
2.2 The Pointwise Ergodic Theorem 37

2.2 The Pointwise Ergodic Theorem

Theorem 2.2 (Birkhoff’s Pointwise Ergodic Theorem). Let (X, B, µ, T ) be a ppt.

N−1 N−1
1 1
If f ∈ L1 , then the limit f (x) := lim N ∑ f (T k x) exists for a.e. x, and N ∑ f◦
N→∞ k=0
k → f in L1 . The function f is T –invariant, absolutely integrable, and
R k=0
T
R R
f dµ =
f dµ. If T is ergodic, then f = f dµ almost everywhere.

Proof. Since every f ∈ L1 is the difference of two non-negative L1 -functions, there

is no loss of generality in assuming that f ≥ 0. Define

1 n−1
An (x) := ∑ f (T k x) , A(x) := lim
n k=0
sup An (x) , A(x) := lim inf An (x).
n→∞ n→∞

A(x), A(x) take values in [0, ∞], are measurable, and are T -invariant, as can be easily
1
checked by taking the limits on both sides of An (x) = n−1n An−1 (T x) + O( n ).
R R
S TEP 1. f dµ ≥ Adµ.
Proof. Fix ε, L, M > 0, and set AL (x) := A(x) ∧ L = min{A(x), L} (an invariant func-
tion). The following function is well-defined and finite everywhere:

τL (x) := min{n > 0 : An (x) > AL (x) − ε}.

For a given N, we “color” the time interval 0, 1, 2, . . . , N − 1 as follows:

• If τL (T 0 x) > M, color 0 red; If τL (T 0 x) ≤ M color the next τL (x) times blue, and
move to the first uncolored k
• If τL (T k x) > M, color k red; Otherwise color the next τL (T k x) times blue, and
move to the first uncolored k
Continue in this way until all times colored, or until τL (T k x) > N − k.
This partitions {0, 1, . . . , N − 1} into red segments, and (possibly consecutive)
blue segments of length ≤ M, plus (perhaps) one last segment of length ≤ M. Note:

• If k is red, then T k x ∈ [τL > M], so ∑ 1[τL >M] (T k x) ≥number of red k’s
red k’s
• The average of f on each blue segment of length τL (T k x) is larger than AL (T k x)−
ε = AL (x) − ε. So for each blue segment

∑ f (T k x) ≥ length of segment × (AL (x) − ε).

k∈blue segment

Summing over all segments: ∑ f (T k x) ≥ (AL (x) − ε)×number of blue k’s.

blue k’s

We can combine these two estimates as follows.

38 2 Ergodic theorems

N−1
∑ ( f + AL 1[τL >M] )(T k x) ≥ #(blues) · (AL (x) − ε) + #(reds) · AL (x)
k=0
≥ #(blues and reds) · (AL (x) − ε) ≥ (N − M)(AL (x) − ε).

Next we divide by N, integrate, and obtain from the T -invariance of µ that

Z  Z
! ML
Z Z
f dµ + AL dµ ≥ (1 − M
N ) A L dµ − ε ≥ AL dµ − − ε,
[τL >M] N
!
where ≥ is because
R
0 ≤ AL ≤ L. Subtracting from both sides of the inequality the
(finite) quantity [τL >M] AL dµ, we find that

ML
Z Z
f dµ ≥ AL dµ − − ε.
[τL ≤M] N

We now take the following limits in the following order: N → ∞; M → ∞; L → ∞;

and
R
ε → 0, using
R
the monotone convergence theorem where appropriate. The result
is f dµ ≥ Adµ.
R R
S TEP 2. f dµ ≤ Adµ
R
Proof. First observe, using Fatou’s Lemma, that Adµ ≤ ∥ f ∥1 < ∞. So A(x) < ∞
almost everywhere. For every ε > 0,

θ (x) := min{n > 0 : An (x) < A(x) + ε}

is well-defined and finite for a.e. x. We now repeat the coloring argument of step 1,
with θ replacing τ and A replacing AM . Let fM (x) := f (x) ∧ M, then as before
N−1
∑ fM (T k x)1[θ ≤M] (T k x) ≤ ∑ f (T k x) + ∑ 0+ ∑ M
k=0 k blue k red no color
≤ #(blues)(A(x) + ε) + M ≤ N(A(x) + ε) + M 2 .
2

R R
Integrating and dividing by N, we obtain 1[θ ≤M] f ∧M ≤ A+ε +O(1/N). Passing
to the
R
limits
R
N → ∞ and then M → ∞ (byR the monotone
R
convergence theorem), we
get f ≤ A + ε. Since ε was arbitrary, f ≤ A.
S TEP 3. limn→∞ n1 ∑n−1 k
k=0 f (T x) exists almost everywhere
R
Together, steps 1 and 2 imply that (A − A)dµ ≤ 0. But A ≥ A, so necessarily A = A
µ–almost everywhere (if µ[A > A] > 0, the integral would be strictly positive). So

1 n−1 1 n−1
lim sup ∑ f (T k x) = lim inf ∑ f (T k x) a.e.,
n→∞ n k=0 n→∞ n
k=0

which proves that the limit exists almost everywhere.

2.3 The non-ergodic case 39

S TEP 4. f (x) := limn→∞ 1n ∑n−1 k

k=0 f (T x) is T -invariant, absolutely integrable, and
1 n−1 k 1
R R
n ∑k=0 f ◦ T → f in L . Consequently, f dµ = f dµ and if T is ergodic, then
R
f = f dµ almost everywhere.
Proof. The T -invariance of f follows by taking the limit as n → ∞ in the identity
1
An (x) = n−1
n An−1 (T x) + n f (x). Absolute integrability is because of step 1.
L1
To see that n1 ∑n−1 k −−
k=0 f ◦ T − → f , fix ε > 0 and construct ϕ ∈ L∞ such that ∥ f −
n→∞
ϕ∥1 < ε. Notice that ∥ n1 ∑n−1 k
k=0 ϕ ◦ T ∥∞ ≤ ∥ϕ∥∞ , therefore by step 3 and the bounded
convergence theorem,
1 n−1 L1
∑ ϕ ◦ T k −n→∞
n k=0
−−→ ϕ (2.1)

for some bounded invariant function ϕ. So

1 n−1 1 n−1 1 n−1

∑ f ◦Tk − f
n k=0
≤ ∑ ( f − ϕ) ◦ T k
n k=0
+ ∑ ϕ ◦Tk −ϕ
n k=0
+ ∥ϕ − f ∥1
1 1 1
1 N−1
Z
≤ ∥ f − ϕ∥1 + o(1) + lim
n→∞ N
∑ |(ϕ − f ) ◦ T k |dµ, by (2.1)
k=0

1 n−1
Z
≤ ∥ f − ϕ∥1 + o(1) + lim
N→∞
∑ |(ϕ − f ) ◦ T k |dµ, by Fatou’s Lemma
N k=0
≤ 2ε + o(1).

Since ε was arbitrary, 1

n ∑n−1 k
k=0 f ◦ T − f 1
→ 0.
L1 R R R
It is easy to see that if gn −−−→ g, then gn → g. So f dµ = lim An ( f )dµ =
R n→∞ R R
f dµ. Necessarily, in the ergodic case, f = const = f dµ = f dµ. ⊔
⊓

2.3 The non-ergodic case

The almost sure limit in the pointwise ergodic theorem is clear when the map is
ergodic: N1 ∑N−1 k
R
k=0 f ◦ T −−−→ f dµ. In this section we ask what is the limit in the
N→∞
non-ergodic case.
If f belongs to L2 , the limit is the projection of f on the space of invariant func-
tions, because of the Mean Ergodic Theorem and the fact that every sequence of
functions which converges in L2 has a subsequence which converges almost every-
where to the same limit.1 But if f ∈ L1 we cannot speak of projections. The right
notion in this case is that of the conditional expectation.
L2
1 Proof: Suppose fn −−−→ f . Pick a subsequence nk s.t. ∥ fnk − f ∥2 < 2−k . Then ∑k≥1 ∥ fnk − f ∥2 <
n→∞
∞. This means that ∥ ∑ | fnk − f |∥2 < ∞, whence ∑( fnk − f ) converges absolutely almost surely. It
follows that fnk − f → 0 a.e.
40 2 Ergodic theorems

2.3.1 Conditional expectations and the limit in the ergodic theorem

Let (X, B, µ) be a probability space. Let F ⊂ B be a σ –algebra. We think of

F ∈ F as of the collection of all sets F for which we have sufficient information
to answer the question “ is x ∈ F?”. The functions we have sufficient information to
calculate are exactly the F –measurable functions, as can be seen from the formula

f (x) := inf{t : x ∈ [ f < t]}.

Suppose g is not F –measurable. What is the ‘best guess’ for g(x) given the infor-
mation F ?
Had g been in L2 , then the “closest” F –measurable function (in the L2 –sense) is
the projection of g on L2 (X, F , µ). The defining property of the projection Pg of g
is ⟨Pg, h⟩ = ⟨g, h⟩ for all h ∈ L2 (X, F , µ). The following definition mimics this case
when g is not necessarily in L2 :
Definition 2.1. The conditional expectation of f ∈ L1 (X, B, µ) given F is the
unique L1 (X, F , µ)–element E( f |F ) which is
1. E( f |F ) is F –measurable;
2. ∀ϕ ∈ L∞ F –measurable, ϕE( f |F )dµ = ϕ f dµ.
R R

Note: E( f |F ) is only determined almost everywhere.

Proposition 2.1. The conditional expectation exists for every L1 element, and is
unique up sets of measure zero.
Proof. Consider the measures ν f := f dµ|F and µ|F on (X, F ). Then ν f ≪ µ.
dν
The function E( f |F ) := dµf (Radon-Nikodym derivative) is F –measurable, and
it is easy to check that it satisfies the conditions of the definition of the conditional
expectation. The uniqueness of the conditional expectation is left as an exercise. ⊔ ⊓
Proposition 2.2. Suppose f ∈ L1 .
1. f 7→ E( f |F ) is linear, and a contraction in the L1 –metric;
2. f ≥ 0 ⇒ E( f |F ) ≥ 0 a.e.;
3. if ϕ is convex and ϕ ◦ f ∈ L1 , then E(ϕ ◦ f |F ) ≥ ϕ(E( f |F ));
4. if h is F –measurable and bounded, then E(h f |F ) = hE( f |F );
5. If F1 ⊃ F2 , then E[E( f |F) |F2 ] = E( f |F2 ).
We leave the proof as an exercise.
Theorem 2.3. Let (X, B, µ, T ) be a p.p.t, and f ∈ L1 (X, B, µ). Then

1 N−1
lim
N→∞ N
∑ f ◦ T k = E( f |Inv(T )) a.e. and in L1 ,
k=0

where Inv(T ) := {E ∈ B : E = T −1 E}. Alternatively, Inv(T ) is the σ –algebra

generated by all T –invariant functions.
2.3 The non-ergodic case 41

Proof. Set f := lim 1

∑N−1 k
k=0 f ◦ T on the set where the limit exists, and zero oth-
N→∞ N
erwise. Then f is Inv–measurable and T -invariant. For every T -invariant ϕ ∈ L∞ ,
!
1 N−1 1 N−1
Z Z
ϕ f dµ = ϕ ∑ f ◦ T k dµ + O ∥ϕ∥∞ ∑ f ◦Tk − f =
N k=0 N k=0
1
1 N−1
Z Z
= ∑ ϕ ◦ T k f ◦ T k dµ + o(1) −−−→ ϕ f dµ,
N k=0 N→∞

because the convergence in the ergodic theorem is also in L1 . ⊔

⊓

2.3.2 Conditional probabilities

Recall that a standard probability space is a probability space (X, B, µ) where X is

a complete, metric, separable space, and B is its Borel σ –algebra.
Theorem 2.4 (Existence of Conditional Probabilities). Let µ by a Borel proba-
bility measure on a standard probability space (X, B, µ), and let F ⊂ B be a
σ –algebra. There exist Borel probability measures {µx }x∈X s.t.:
1. x 7→ µx (E) is F –measurable for
R
every E ∈ B; R R R

2. if f is µ–integrable, then x 7→ f dµx is integrable, and f dµ = f dµx dµ;
R X X
3. if f is µ–intergable, then f dµx = E( f |F )(x) for µ–a.e. x.

Definition 2.2. The measures µx are called the conditional probabilities of F . Note
that they are only determined almost everywhere.

Proof. By the isomorphism theorem for standard spaces, there is no loss of general-
ity in assuming that X is compact. Indeed, we may take X to be a compact interval.
Recall that for compact metric spaces X, the space of continuous functions C(X)
with the maximum norm is separable.2
n=0 in C(X) s.t. f 0 ≡ 1. Let AQ be the algebra
Fix a countable dense set { fn }∞
generated by these functions over Q. It is still countable.
Choose for every g ∈ AQ an F –measurable version E(g|F ) of E(g|F ) (recall
that E(g|F ) is an L1 –function, namely not a function at all but an equivalence class
of functions). Consider the following collection of conditions:
1. ∀α, β ∈ Q, g1,2 ∈ AQ , E(αg1 + β g2 |F )(x) = αE(g1 |F )(x) + β E(g2 |F )(x)
2. ∀g ∈ AQ , min g ≤ E(g|F )(x) ≤ max g
2 Proof: Compact metric space are separable because they have finite covers by balls of radius
1/n. Let {xn } be a countable dense set of points, then ϕn (·) := dist(xn , ·) is a countable family
of continuous functions, which separates points in X. The algebra which is generated over Q by
{ϕn } ∪ {1X } is countable. By the Stone-Weierstrass Theorem, it is dense in C(X).
42 2 Ergodic theorems

This is countable collection of F –measurable conditions, each of which holds with

full µ–probability. Let X0 be the set of x’s which satisfies all of them. This is an
F –measurable set of full measure.
We see that for each x ∈ X0 , ϕx [g] := E(g|F )(x) is linear functional on AQ , and
∥ϕx ∥ ≤ 1. It follows that ϕx extends uniquely to a positive bounded linear functional
on C(X). This is a measure µx .
R R R
Step 1. ( f dµx )dµ(x) = f dµ for all f ∈ C(X).
Proof. This is true for all f ∈ AQ by definition, and extends to all C(X) because AQ
is dense in C(X). (But for f ∈ L1 it is not even clear that the statement makes sense,
because µx could live on a set with zero µ–measure!)
Step 2. x 7→ µx (E) is F –measurable for all E ∈ B.
Exercise: Prove this using the following steps
1. The indicator function of any open set is the pointwise limit of a sequence of
continuous functions 0 ≤ hn ≤ 1, thus the step holds for open sets.
2. The collection of sets whose indicators are pointwise limits of a bounded se-
quence of continuous functions forms an algebra. The step holds for every set in
this algebra.
3. The collection of sets for which step 1 holds is a monotone class which contains
a generating algebra.

Step 3. If f = g µ–a.e., then f = g µx –a.e. for µ–a.e. x.

Proof. Suppose µ(E) = 0. Choose open sets Un ⊇ E such that µ(Un ) → 0. Choose
continuous functions 0 ≤ hεn ≤ 1 s.t. hεn vanish outside Un , hεn are non-zero inside
Un , and hεn −−−→ 1Un (e.g. hεn (·) := [dist (x,Unc )/diam(X)]ε ).
ε→0+
By construction 1E ≤ 1Un ≡ lim hεn , whence
ε→0+

Z Z Z
µx (E)dµ(x) ≤ lim hεn dµx dµ ≤
ε→0+
Z Z Z Z
≤ lim hεn dµx dµ = lim hεn dµ ≤ lim hεn dµ ≤ µ(Un ) −−−→ 0.
ε→0+ ε→0+ ε→0+ n→∞

It follows that µx (E) = 0 a.e.

R
Step 4. For all f µ–absolutely integrable, E( f |F )(x) = f dµx µ–a.e.
Proof. Find gn ∈ C(X) such that
∞
f= ∑ gn µ-a.e., and ∑ ∥gn ∥L1 (µ) < ∞.
n=1

Then
2.3 The non-ergodic case 43
∞
E( f |F ) = ∑ E(gn |F ), because E(·|F ) is a bounded operator on L1
n=1
∞ Z
= ∑ gn dµx a.e., because gn ∈ C(X)
n=1 X
Z ∞
= ∑ gn dµx a.e., (justification below)
X n=1
Z
= f dµx a.e.
X
R
Here is the justification: ∑ |gn |dµx < ∞, because the integral of this expression, by
the monotone convergence theorem is less than ∑ ∥gn ∥1 < ∞. ⊔
⊓

2.3.3 The ergodic decomposition

Theorem 2.5 (The Ergodic Decomposition). Let µ be an invariant Borel proba-

bility measure of a Borel map T on a standard probability space X. Let {µx }x∈X be
the conditional probabilities w.r.t. Inv(T ). Then
1. µ = X µx dµ(x) (i.e. this holds when applies to L1 –functions or Borel sets);
R

2. µx is invariant for µ–a.e. x ∈ X;

3. µx is ergodic for µ–a.e. x ∈ X.

Proof. By the isomorphism theorem for standard probability spaces, there is no

loss of generality in assuming that X is a compact metric space, and that B is its
σ –algebra of Borel sets.
For every f ∈ L1 , f dµ = E( f |Inv(T ))dµ(x) = X X f dµx dµ(x). This shows
R R R R

(1). We have to show that µx is invariant and ergodic for µ–a.e. x.

Fix a countable set { fn } which is dense in C(X), and choose Borel versions
Eµ ( fn |Inv(T ))(x). By the ergodic theorem, there is a set of full measure Ω such
that for all x ∈ Ω ,

1 N−1
Z
fn dµx = Eµ ( fn |Inv(T ))(x) = lim
N→∞ N
∑ fn (T k x) for all n.
k=0

Step 1. µx is T –invariant for a.e. x ∈ Ω .

Proof. For every n,

1 N−1
Z
fn ◦ T dµx = lim
N→∞ N
∑ fn (T k+1 x) a.e. (by the PET)
k=0

1 N−1
Z
= lim
N→∞ N
∑ fn (T k x) = Eµ ( fn |Inv(T ))(x) = fn dµx .
k=0
44 2 Ergodic theorems

Let Ω ′ be the set of full measure for which theR

above holdsRfor all n, and fix x ∈ Ω ′ .
Since { fn } is ∥ · ∥∞ –dense in C(X), we have f ◦ T dµx = f dµx for all f ∈ C(X).
−1
Using the density of C(X) in L1 ( µx +µ2x ◦T ), it is routine to see that f ◦ T dµx =
R

f dµx for all µx –integrable functions. This means that µx ◦ T −1 = µx for all x ∈ Ω ′ .
R

Step 2. µx is ergodic for all x ∈ Ω .

Proof. With { fn }∞ ′′ ′ lim 1n ∑n−1 k
n=1 as above, let Ω := {x ∈ Ω : ∀k, n→∞ k=0 f k (T x) =
R
fk dµx }. This is a set of full measure because of the ergodic theorem. Now

1 N−1
Z
0 = lim ∑ fn ◦ T k − fn dµx (∵ L1 -convergence in the PET)
N→∞ N k=0 X
L1 (µ)

1 N−1
Z Z Z
= lim
N→∞ X
∑ fn ◦ T k −
N k=0 X
fn dµx dµ(x) (∵ µ =
X
µx dµ)
L1 (µx )

1 N−1
Z Z
≥ lim inf
X N→∞
∑ fn ◦ T k −
N k=0 X
fn dµx dµ(x) (∵ Fatou’s Lemma)
L1 (µx )

1
∑n−1 k
R
It follows that lim inf
n→∞ n k=0 f n ◦ T − X f n dµx L1 (µx ) = 0 µ–a.e. Let

1 n−1
 Z 
Ω := x ∈ Ω ′′ : lim inf ∑ f n ◦ T k
− fn dµx = 0 for all n, ,
n→∞ n k=0 X
L1 (µx )

then Ω has full measure.

Fix x ∈ Ω . Since { fn }n≥1 is dense in C(X), and C(X) is dense in L1 (µx ), { fn } is
dense in L1 (µx ). A standard approximation argument shows that

1 n−1
Z
lim inf ∑ f ◦Tk − f dµx = 0 for all f ∈ L1 (µx ).
n→∞ n k=0 X
L1 (µx )

In particular, every f ∈ L1 (µx ) such that f ◦ T = f µx -a.e. must be constant µx -

almost everywhere. So µx is ergodic for x ∈ Ω ′′′ . ⊔
⊓

2.4 An Ergodic Theorem for Zd -actions

Let T1 , . . . , Td denote d measure preserving transformations on a probability space

(Ω , F , µ). We say that T1 , . . . , Td commute if Ti ◦ T j = T j ◦ Ti for all i, j. Let Zd+ :=
(N ∪ {0})d and define for n = (n1 , . . . , nd ) ∈ Zd+
n
T n := T1n1 ◦ · · · ◦ Td d .
2.4 An Ergodic Theorem for Zd -actions 45

If T1 , . . . , Td commute, then T n ◦ T m = T n+m . An algebraist would say that the semi-

group (Zd+ , +) acts on (Ω , F , µ) by n·x = T n (x). This is called the Zd+ –semi-action
generated by T1 , . . . , Td . If Ti are invertible, this extends naturally to a Zd –action.
A pointwise ergodic theorem for d–commuting maps is an almost sure conver-
gence statement for averages of the type
1 1
SI f :=
|Ir | r ∑ f ◦Tn
|Ir | n∈Ir

where Ir is a sequence of subsets of Zd+ which “tends to Zd+ ”, and |Ir | =cardinality
of Ir . Such statements are not true for any choice of {Ir } (even when d = 1). Here
we prove the following pointwise ergodic theorem for increasing boxes : Boxes are
sets of the form

[n, m) := {k = (k1 , . . . , kd ) ∈ Zd+ : ni ≤ ki ≨ mi (1 ≤ i ≤ d)} (n, m ∈ Zd+ ).

A sequence of boxes {Ir }r≥1 is said to be increasing if SIr ⊂ Ir+1 for all r. An in-
creasing sequence of boxes is said to tend to Zd+ if Zd+ = r≥1 Ir .

Theorem 2.6 (Tempelman). Let T1 , . . . , Td be d–commuting probability preserving

maps on a probability space (Ω , F , µ), and suppose {Ir }r≥1 is an increasing se-
quence of boxes which tends to Zd+ . If f ∈ L1 , then

1
∑ f ◦ T n −r→∞
|Ir | n∈I
−−→ E( f |Inv(T1 ) ∩ · · · ∩ Inv(Td )) almost surely.
r

For convergence along more general sequences of boxes, see problem 2.8.
Proof. Fix f ∈ L1 . Almost sure convergence is obvious in the following two cases:
1. Invariant functions: If f ◦ Ti = f for i = 1, . . . , d, then |I1r | SIr f = f for all r, so the
limit exists and is equal to f .
2. Coboundries: Suppose f = g − g ◦ Ti for some g ∈ L∞ and some i,

1 1
SI f = SI gi − SIr +ei gi , where e1 , . . . ed is the standard basis of Rd
|Ir | r |Ir | r
1 |Ir △(Ir + ei )|
= S gi − S(Ir +ei )\Ir gi ≤ ∥gi ∥∞ .
|Ir | Ir \(Ir +ei ) |Ir |

Now |Ir △(Ir + ei )|/|Ir | −−−→ 0, because the lengths ℓ1 (r), . . . , ℓd (r) of the sides
r→∞
of the box Ir tend to infinity, and so

|Ir △(Ir + ei )| 2
= −−−→ 0.
|Ir | ℓi (r) r→∞

So the limit exists and is equal to zero.

46 2 Ergodic theorems

Step 1. Any f ∈ L1 can put in the form f = ∑di=1 (gi − gi ◦ Ti ) + h + ϕ, where gi ∈ L∞ ,

h is Ti –invariant for all i, and ∥ϕ∥1 < ε with ε arbitrarily small.
Proof. One checks, as in the proof of the Mean Value Theorem, that

span{g − g ◦ Ti : g ∈ L2 , (1 ≤ i ≤ d)}⊥ = { f ∈ L2 : f ◦ Ti = f (1 ≤ i ≤ d)},

whence L2 = { f ∈ L2 : f ◦Ti = f (1 ≤ i ≤ d)}⊕span{g−g◦Ti : g ∈ L2 , (1 ≤ i ≤ d)}

(orthogonal sum).
This means that any f ′ ∈ L2 can be put in the form f ′ = ∑di=1 (g′i −g′i ◦Ti )+h+ϕ ′ ,
where g′i ∈ L2 , h ∈ L2 is Ti –invariant for all i, and ∥ϕ ′ ∥2 < ε/3.
Since L∞ is dense in L2 , it is no problem to replace g′i by L∞ –functions gi so that
′
f = ∑di=1 (gi − gi ◦ Ti ) + h + ϕ, where ∥ϕ∥2 < ε/2. By Cauchy–Schwarz, ∥ϕ∥1 <
ε/2. This proves the step when f is in L2 . If f is in L1 , find f ′ ∈ L2 s.t. ∥ f − f ′ ∥1 <
ε/2 and apply the above to f ′ .
Step 2 (Maximal Inequality). For every non-negative ϕ ∈ L1 and t > 0,

2d ∥ϕ∥1
 
1
µ sup SIr ϕ > t ≤ . (2.2)
r |Ir | t

We give the proof later.

Step 3. How to use the maximal inequality to complete the proof.
For every f ∈ L1 , let ∆ ( f )(ω) := lim sup |I1r | (SIr f )(ω) − lim inf |I1r | (SIr f )(ω). To
r→∞ r→∞
show that lim |I1r | SIr f exists almost surely, one needs to show that ∆ ( f ) = 0 a.e.
Notice that ∆ is subadditive: ∆ ( f1 + f2 ) ≤ ∆ ( f1 ) + ∆ ( f2 ). If we express f as in
the first step, then we get ∆ ( f ) ≤ ∆ (ϕ). It follows that for every δ > 0,
 
1
µ [∆ ( f ) > δ ] ≤ µ [∆ (ϕ) > δ ] ≤ µ 2 sup SIr |ϕ| > δ
r |Ir |
2d ∥ϕ∥1
≤ by the maximal inequality.
(δ /2)
√
Taking δ := ε and recalling that ϕ was constructed so that ∥ϕ∥1 < ε, we see that
√  √
µ ∆ ( f ) > ε < 2d+1 ε.


But ε was arbitrary, so we must have ∆ ( f ) = 0 almost everywhere. In other words,

lim |I1r | SIr f exists almost everywhere. The proof also shows that the value of the limit
f equals h, so it is an invariant function.
To identify h, we argue as in the proof of Birkhoff’s theorem. First we claim that
1 L1
|Ir | SIr f −−−→ h. If f is bounded, this is a consequence of pointwise convergence
r→∞
and the bounded convergence theorem. For general L1 –functions, write f = f ′ + ϕ
with f ′ ∈ L∞ and ∥ϕ∥1 < ε. The averages of f ′ converge in L1 , and the averages of
 2.4 An Ergodic Theorem for Zd -actions 47

1
ϕ remain small in L1 norm. It follows that |Ir | SIr f converge in L1 . The limit must
agree with the pointwise limit f (see the footnote on page 39).
Integrate |I1r | SI1 f against a bounded invariant function g and pass to the limit. By
L1 –covergence, f g = hg. It follows that h = E( f | di=1 Inv(T
R R T
R i
)). In particular,
if the Zd+ –action generated by T1 , . . . , Td is ergodic, then h = f . This finishes the
proof, assuming the maximal inequality.
Proof of the maximal inequality. Let ϕ ∈ L1 be a non-negative function, fix N and
α > 0. We will estimate the measure of EN (α) := {ω ∈ Ω : max1≤k≤N |I1 | SIk ϕ > α}.
k
Pick a large box I, and let A(ω) := {n ∈ I : T n (ω) ∈ EN (α)}. For every n ∈ A(ω)
there is a 1 ≤ k(n) ≤ N such that (SIk(n) +n ϕ)(ω) > α|Ik(n) |.
Imagine that we were able to find a disjoint subcollection {n + Ik(n) : n ∈ A′ (ω)}
which is “large” in the sense that there is some global constant K s.t.
1
∑′ |n + Ik(n) | ≥ |A(ω)|. (2.3)
n∈A (ω)
K

The sets n+Ik(n) (n ∈ A′ (ω)) are included in the box J ⊃ I obtained by increasing
the sides of I by 2 max{diam(I1 ), . . . , diam(IN )}. This, the non-negativity of ϕ, and
the invariance of µ implies that
1 1
Z Z
∥ϕ∥1 ≥ (SJ ϕ)(ω) dµ ≥ ∑ (SIk(n) +n ϕ)(ω) dµ
|J| |J| n∈A′ (ω)
1
Z
≥ ∑ α|Ik(n) + n| dµ
|J| n∈A′ (ω)
Z Z
α α α|I|
≥
K|J|
|A(ω)|dµ =
K|J| ∑ 1EN (α) (T n ω)dµ = K|J| µ[EN (α)].
n∈I

It follows that µ[EN (α)] ≤ K |J|

|I| ∥ϕ∥1 /α. Now I was arbitrary, and by construction
|J| ∼ |I| as |I| → ∞, so µ[EN (α)] ≤ K∥ϕ∥1 /α. In the limit N → ∞, we get the
maximal inequality (except for the identification K = 2d ).
We now explain how to find the disjoint subcollection {n + Ik(n) : n ∈ A′ (ω)}. We
use the “greedy” approach by first adding as many translates of IN (the largest of
I1 , . . . , IN ) as possible, then as many translates of IN−1 as possible, and so on:
(N) Let MN be a maximal disjoint collection of sets of the form IN +n with k(n) = N.
(N − 1) Let MN−1 be a maximal disjoint collection of sets of the form IN−1 + n with
k(n) = N − 1 and such that all elements of MN−1 are disjoint from MN .
S

·········

(1) Let M1 be a maximal disjoint collection of sets of the form I1 + n where k(n) = 1
and such that all elements of M1 are disjoint from (MN ∪ · · · ∪M2 ).
S
48 2 Ergodic theorems

Now let A′ (ω) := {n : Ik(n) + n ∈ M1 ∪ · · · ∪ MN }. This is a disjoint collection.

We show that A(ω) ⊆ n∈A′ (ω) (n+Ik(n) −Ik(n) ) (where I −I = {n−m : n, m ∈ I}).
S

Suppose n ∈ A(ω). By the maximality of Mk(n) , either n + Ik(n) ∈ Mk(n) , or n + Ik(n)

intersects some m + Ik(m) ∈ Mk(m) s.t. k(m) ≥ k(n).
• In the first case n ∈ n + Ik(n) − Ik(n) ∈ n∈A′ (ω) (n + Ik(n) − Ik(n) )
S

• In the second case there are u ∈ Ik(n) , v ∈ Ik(m) s.t. n + u = m + v, and again we get
n ∈ m + Ik(m) − Ik(n) ⊆ m + Ik(m) − Ik(m) (since k(m) ≥ k(n) and Ir is increasing).
For a d–dimensional box I, |I − I| = 2d |I|. Since A(ω) ⊆ (n + Ik(n) − Ik(n) ),
S
n∈A′ (ω)
|A(ω)| ≤ ∑n∈A′ (ω) 2d |Ik(n) |, and we get (2.3) with K = 2d . ⊔
⊓

2.5 The Subadditive Ergodic Theorem

We begin with two examples.

Example 1 (Random walks on groups) Let (X, B, µ, T ) be the Bernoulli scheme
with probability vector p = (p1 , . . . , pd ). Suppose G is a group, and f : X → G is the
function f (x0 , x2 , . . .) = gx0 , where g1 , . . . , gn ∈ G. The expression

fn (x) := f (x) f (T x) · · · f (T n−1 x)

describes the position of a random walk on G, which starts at the identity, and whose
steps have the distribution Pr[step = gi ] = pi . What can be said on the behavior of
this random walk?
In the special case G = Zd or G = Rd , fn (x) = f (x) + f (T x) + · · · + f (T n−1 x),
3 1 R
and the ergodic theorem says that n fn (x) has an almost sure limit, equal to f dµ =
∑ pi gi . So: the random walk has speed ∥ ∑ pi gi ∥, and direction ∑ pi gi /∥ ∑ pi gi ∥.
(Note that if G = Zd , the direction need not lie in G.)
Example 2 (The derivative cocycle) Suppose T : V → V is a diffeomorphism act-
ing on an open set V ⊂ Rd . The derivative of T at x ∈ V is a linear transformation
(dT )(x) on Rd , v 7→ [(dT )(x)]v. By the chain rule,

(dT n )(x) = (dT )(T n−1 x) ◦ (dT )(T n−2 x) ◦ · · · ◦ (dT )(x).

If we write f (x) := (dT )(x) ∈ GL(d, R), then we see that

(dT n )(x) = f (T n−1 x) f (T n−2 x) · · · f (T x) f (x)

is a “random walk” on GL(d, R) := {invertible d × d matrices with real entries}.

Notice the order of multiplication!
3 applied to the each coordinate of the vector valued function f = ( f 1 , . . . , f d ).
2.5 The Subadditive Ergodic Theorem 49

What is the “speed” of this random walk? Does it have an asymptotic “direction”?
The problem of describing the “direction” of random walk on a group is deep, and
remain somewhat mysterious to this day, even in the case of groups of matrices. We
postpone it for the moment, and focus on the conceptually easier task of defining
the “speed.” Suppose G is a group of d × d matrices with real-entries. Then G can
be viewed to be group of linear operators on Rd , and we can endow A ∈ G with the
operator norm ∥A∥ := max{∥Av∥2 /∥v∥2 : 0 ̸= v ∈ Rd }. Notice that ∥AB∥ ≤ ∥A∥∥B∥.
We will measure the speed of fn (x) := f (x) f (T x) · · · f (T n−1 x) by analyzing

g(n) (x) := log ∥ fn (x)∥ as n → ∞.

Key observation: g(n+m) ≤ g(n) + g(m) ◦ T n , because

g(n+m) (x) = log ∥ fn+m (x)∥ = log ∥ fn (x) fm (T n x)∥ ≤ log(∥ fn (x)∥ · ∥ fm (T n x)∥)
≤ log ∥ fn (x)∥ + log ∥ fm (T n x)∥ = g(n) (x) + g(m) (T n x).

We say that {g(n) }n is a subadditive cocycle.

Theorem 2.7 (Kingman’s Subadditive Ergodic Theorem). Let (X, B, m, T ) be a

probability preserving transformation, and suppose g(n) : X → R is a sequence of
measurable functions such that g(n+m) ≤ g(n) + g(m) ◦ T n for all n, m, and g(1) ∈ L1 .
Then the limit g := lim g(n) /n exists almost surely, and is an invariant function.
n→∞

Proof. We begin by observing that it is enough to treat the case when g(n) are all
non-positive. This is because h(n) := g(n) − (g(1) + g(1) ◦ T + · · · + g(1) ◦ T n−1 ) are
non–positive, sub-additive, and differ from g(n) by the ergodic sums of g(1) whose
asymptotic behavior we know by Birkhoff’s ergodic theorem.
Assume then that g(n) are all non-negative. Define G(x) := lim inf g(n) (x)/n (the
n→∞
limit may be equal to −∞). We claim that G ◦ T = G almost surely.
Starting from the subadditivity inequality g(n+1) ≤ g(n) ◦ T + g(1) , we see that
G ≤ G ◦ T . Suppose there were a set of positive measure E where G ◦ T > G + ε.
Then for every x ∈ E, G(T n x) ≥ G(T n−1 x) ≥ · · · ≥ G(T x) > G(x) + ε. But this is
impossible, because by Poincaré’s Recurrence Theorem, for a.e. x there is some
n > 0 such that G(T n x) = ∞ or |G(T n x) − G(x)| < ε (prove!). This contradiction
shows that G = G ◦ T almost surely. Henceforth we work the set of full measure
\
X0 := [G ◦ T n = G].
n≥1

Fix M > 0, and define GM := G ∨ (−M). This is an invariant function on X0 .

(n)
We aim at showing lim sup g n ≤ GM a.s.. Since M is arbitrary, this implies that
n→∞
lim sup g(n) /n ≤ G = lim inf g(n) /n, whence almost sure convergence.
n→∞ n→∞
Fix x ∈ X0 , N ∈ N, and ε > 0. Call k ∈ N ∪ {0}
• “good”, if ∃ℓ ∈ {1, . . . , N} s.t. g(ℓ) (T k x)/ℓ ≤ GM (T k x) + ε = GM (x) + ε;
 50 2 Ergodic theorems

• “bad”, if it’s not good: g(ℓ) (T k x)/ℓ > GM (x) + ε for all ℓ = 1, . . . N.
Color the integers 0, . . . , n − 1 inductively as follows, starting from k = 1. Let k be
the smallest non-colored integer,
(a) If k ≤ n − N and k is “bad”, color it red;
(b) If k ≤ n − N and k is “good”, find the smallest 1 ≤ ℓ ≤ N s.t. g(ℓ) (T k x)/ℓ ≤
GM (T k x) + ε and color the segment [k, k + ℓ) blue;
(c) If k > n − N, color k white.
Repeat this procedure until all integers 0, . . . , n − 1 are colored.
The “blue” part can be decomposed into segments [τi , τi + ℓi ), with ℓi s.t.
g(ℓi ) (T τi x)/ℓi ≤ GM (x) + ε. Let b denote the number of these segments.
The “red” part has size ≤ ∑nk=1 1B(N,M,ε) (T k x), where

B(N, M, ε) := {x ∈ X0 : g(ℓ) (x)/ℓ > GM (x) + ε for all 1 ≤ ℓ ≤ N}.

Let r denote the size of the red part. The “white” part has size w ≤ N.
By the sub-additivity condition

g(n) (x) 1 b (ℓi ) τi 1 1

≤ ∑ g (T x) + ∑ g(1) (T k x) + ∑ g(1) (T k x)
n n i=1 n k red n k white
| {z }
non-positive
b b
1 1 #{blues}
≤ ∑ g(ℓi ) (T τi x) ≤ n ∑ (GM (x) + ε)ℓi = n (GM (x) + ε).
n i=1 i=1

Now #{blues} ≤ n − (r + w) = n − ∑nk=1 1B(N,M,ε) (T k x) + O(1), so by the Birkhoff

ergodic theorem, for almost every x, #{blues}/n −−−→ 1 − E(1B(N,M,ε) |Inv). Thus
n→∞

g(n) (x)
lim sup ≤ (GM (x) + ε)(1 − E(1B(N,M,ε) |Inv)) almost surely.
n→∞ n

Now N was arbitrary, and for fixed M and ε, B(N, M, ε) ↓ ∅ as N ↑ ∞, because GM ≥

G = lim inf g(ℓ) /ℓ. It is not difficult to deduce from this that E(1B(N,M,ε) |Inv) −−−→ 0
ℓ→∞ N→∞
almost surly.4 Thus

g(n) (x)
lim sup ≤ GM (x) + ε almost surely.
n→∞ n

Since ε was arbitrary, lim sup g(n) /n ≤ GM almost surely, which proves almost sure
n→∞
convergence by the discussion above. ⊔
⊓
4 Suppose 0 ≤ fn ≤ 1 and fn ↓ 0. The conditional expectation is monotone, so E( fn |F ) is decreas-
ing at almost every point. Let ϕ be its almost sure limit, then 0 ≤ ϕ ≤ 1 a.s., and by the BCT,
E(ϕ) = E(lim E( fn |F )) = lim E(E( fn |F )) = lim E( fn ) = E(lim fn ) = 0, whence ϕ = 0 almost
everywhere.
2.5 The Subadditive Ergodic Theorem 51

Proposition 2.3. Suppose m is ergodic, and g(n) ∈ L 1 for all n, then the limit in
R (n)
Kingman’s ergodic theorem is the constant inf[(1/n) g dm] ∈ [−∞, ∞).

Proof. Let G := lim g(n) /n. Subadditivity implies that G ◦ T ≤ G. Recurrence im-
plies that G◦T = T . Ergodicity implies that G = c a.e., for some constant c = c(g) ≥
−∞. We claim that c ≤ inf[(1/n) g(n) dm]. This is because
R

!
1 (kn) 1 g(n) g(n) n g(n) n(k−1)
c = lim g ≤ lim + ◦T +··· ◦T
k→∞ kn k→∞ k n n n
1
Z
= g(n) dm (Birkhoff’s ergodic theorem),
n

proving that c ≤ (1/n) g(n) dm for all n.

R

To prove the other inequality we first note (as in the proof of Kingman’s sub-
additive theorem) that it is enough to treat the case when g(n) are all non-positive.
Otherwise work with h(n) := g(n) − (g(1) + · · · + g(1) ◦ T n−1 ). Since g(1) ∈ L1 ,

1 (1)
Z
(g + · · · + g(1) ◦ T n−1 ) −−−→ g(1) dm pointwise and in L1 .
n n→∞

(n)
Thus c(g) = lim g n = c(h)+ g(1) = inf(1/n)[ h(n) + Sn g(1) ] = inf[(1/n) g(n) ]dm.
R R R R
(n)
Suppose then that g(n) are all non-positive. Fix N, and set gN := max{g(n) , −nN}.
This is, again, subadditive because
(n+m)
gN = max{g(n+m) , −(n + m)N} ≤ max{g(n) + g(m) ◦ T n , −(n + m)N}
(n) (m) (n) (m)
≤ max{gN + gN ◦ T n , −(n + m)N} ≡ gN + gN ◦ T n .

(n)
By Kingman’s theorem, gN /n converges pointwise to a constant c(gN ). By defini-
(n)
tion, −N ≤ gN /n ≤ 0, so by the bounded convergence theorem,

1 1 1
Z Z Z
(n) (n)
c(gN ) = lim gN dm ≥ inf gN dm ≥ inf g(n) dm. (2.4)
n→∞ n n n

Case 1: c(g) = −∞. In this case g(n) /n → −∞, and for every N there exists N(x) s.t.
(n)
n > N(x) ⇒ gN (x) = −N. Thus c(gN ) = −N, and (2.4) implies inf[(1/n) g(n) dm] =
R

−∞ = c(g).
Case 2: c(g) is finite. Take N > |c(g)| + 1, then for a.e. x, if n is large enough, then
(n)
g(n) /n > c(g) − ε > −N, whence gN = g(n) . Thus c(g) = c(gN ) ≥ inf n1 g(n) dm
R

and we get the other inequality. ⊔

⊓

Here is a direct consequence of the subadditive ergodic theorem (historically, it

predates the subadditive ergodic theorem):
52 2 Ergodic theorems

Theorem 2.8 (Furstenberg–Kesten). Let (X, B, µ, T ) be a ppt, and A : X → GL(d, R)

be a measurable function s.t. log ∥A±1 ∥ ∈ L1 . If An (x) := A(T n−1 x) · · · A(x), then the
following limit exists a.e. and is invariant: lim 1n log ∥An (x)∥.
n→∞

The following immediate consequence will be used in the proof of the Oseledets
theorem for invertible cocycles:
Remark: Suppose (X, B, m, T ) is invertible, and let g(n) be a subadditive cocycle
s.t. g(1) ∈ L1 . Then for a.e. x, lim g(n) ◦ T −n /n exists and equals lim g(n) /n.
n→∞ n→∞

Proof. Since g(n) is subadditive, g(n) ◦ T −n is subadditive:

g(n+m) ◦ T −(n+m) ≤ [g(n) ◦ T m + g(m) ] ◦ T −(n+m) = g(n) ◦ T −n + [g(m) ◦ T −m ] ◦ T −n .

R
Let m = my dπ(y) be the ergodic decomposition of m. Kingman’s ergodic theo-
rem and the previous remark say that for π–a.e. y,

g(n) ◦ T −n 1 1
Z Z
lim = inf g(n) ◦ T −n dmy = inf g(n) dmy my a.e.
n→∞ n n n
g(n)
= lim my a.e.
n→∞ n

Thus the set where the statement of the remark fails has zero measure with respect
to all the ergodic components of m, and this means that the statement is satisfied on
a set of full m–measure. □

2.6 The Multiplicative Ergodic Theorem

2.6.1 Preparations from Multilinear Algebra

Multilinear forms. Let V = Rn equipped with the Euclidean inner product ⟨v, w⟩ =
∑ vk wk . A linear functional on V is a linear map ω : V → R. The set of linear func-
tionals is denoted by V ∗ . Any v ∈ V determines v∗ ∈ V ∗ via v∗ = ⟨v, ·⟩. Any linear
function is of this form.
A k–multilinear function is a function T : V k → R such that for all i and
v1 , . . . , vi−1 , vi+1 , . . . , vk ∈ V , T (v1 , . . . , vi−1 , · , vi+1 , . . . , vk ) is a linear functional.
The set of all k–multilinear functions on V is denoted by T k (V ). The tensor
product of ω ∈ T k (V ) and η ∈ T ℓ (V ) is ω ⊗ η ∈ T k+ℓ (V ) given by

(ω ⊗ η)(v1 , . . . , vk+l ) := ω(v1 , . . . , vk )η(vk+1 , . . . , vk+l ).

The tensor product is bilinear and associative, but it is not commutative.

2.6 The Multiplicative Ergodic Theorem 53

The dimension of T k(V ) is nk . Here is a basis: {e∗i1 ⊗ · · · ⊗ e∗ik : 1 ≤ i1 , . . . , ik ≤ n}.

To see this note that every element in T k (Ω ) is completely determined by its action
on {(ei1 , · · · , eik ) : 1 ≤ i1 , . . . , ik ≤ n}.
Define an inner product on T k (V ) by declaring the above basis to be an orthogo-
nal collection of vectors of length √1k! (the reason for the normalization will become
clear later).
Alternating multilinear forms. A multilinear form ω is called alternating, if it
satisfies ∃i ̸= j(vi = v j ) ⇒ ω(v1 , . . . , vn ) = 0. Equivalently,

ω(v1 , . . . , vi , . . . , v j , . . . vn ) = −ω(v1 , . . . , v j , . . . , vi , . . . vn ).

(to see the equivalence, expand ω(v1 , . . . , vi + v j , . . . , v j + vi , . . . vn )). The set of all
k–alternating forms is denoted by Ω k (V ).
Any multilinear form ω gives rise to an alternating form Alt(ω) via

1
Alt(ω) := sgn(σ )σ · ω,
k! σ∑
∈S k

where Sk is the group of k–permutations, and the action of a permutation σ on

ω ∈ T k (V ) is given by (σ · ω)(v1 , . . . , vk ) = ω(vσ (1) , . . . , vσ (k) ). The normalization
k! is to guarantee Alt|Ω k (V ) = id, and Alt2 = Alt. Note that Alt is linear.

Lemma 2.1. Alt[Alt(ω1 ⊗ ω2 ) ⊗ ω3 ] = Alt(ω1 ⊗ ω2 ⊗ ω3 ).

Proof. We show that if Alt(ω) = 0, then Alt(ω ⊗ η) = 0 for all η. Specializing to

the case ω = Alt(ω1 ⊗ ω2 ) − ω1 ⊗ ω2 and η = ω3 , we get (since Alt2 = Alt)

Alt[(Alt(ω1 ⊗ ω2 ) − ω1 ⊗ ω2 ) ⊗ ω3 ] = 0,

which is equivalent to the statement of the lemma.

Suppose ω ∈ T k (V ), η ∈ T ℓ (V ), and Alt(ω) = 0. Let G := {σ ∈ Sk+ℓ : σ (i) =
i for all i = k + 1, . . . , k + ℓ}. This is a subgroup of Sk+l , and there is natural isomor-
phism σ 7→ σ ′ := σ |{1,...,k} from G to Sk . Let Sk+l = j Gσ j be the corresponding
U

coset decomposition, then

(k + ℓ)!Alt(ω ⊗ η)(v1 , . . . , vk+ℓ ) =

=∑ ∑ sgn(σ σ j )(σ σ j ) · (ω ⊗ η)(v1 , . . . , vk+ℓ )
j σ ∈G

= ∑ sgn(σ j )η(vσ j (k+1) , . . . , vσ j (k+ℓ) ) ∑ sgn(σ )(σ · ω)(vσ j (1) , . . . , vσ j (k) )

j σ ∈G

= ∑ sgn(σ j )η(vσ j (k+1) , . . . , vσ j (k+ℓ) ) ∑ sgn(σ ′ )(σ ′ · ω)(vσ j (1) , . . . , vσ j (k) )

j ′
σ ∈Sk

= ∑ sgn(σ j )η(vσ j (k+1) , . . . , vσ j (k+ℓ) )k!Alt(ω)(vσ j (1) , . . . , vσ j (k) ) = 0. ⊔

⊓
j
54 2 Ergodic theorems

Using this “antisymmetrization operator”, we define the following product, called

the exterior product or the wedge product: If ω ∈ Ω k (V ), η ∈ Ω l (V ), then

(k + l)!
ω ∧ η := Alt(ω ⊗ η).
k!ℓ!
The wedge product is bilinear, and the previous lemma shows that it is associative.
It is almost anti commutative: If ω ∈ Ω k (V ), η ∈ Ω ℓ (V ), then

ω ∧ η = (−1)kℓ η ∧ ω.

We’ll see the reason for the peculiar normalization later.

Proposition 2.4. {e∗i1 ∧ · · · ∧ e∗ik : 1 ≤ i1 < · · · < ik ≤ n} is an orthonormal basis for
Ω k (V ), whence dim Ω k (V ) = nk .

Proof. Suppose ω ∈ Ω k (V ), then ω ∈ T k (V ) and so ω = ∑ ai1 ,...,ik e∗i1 ⊗ · · · ⊗ e∗ik ,

where the sum ranges over all k–tuples of numbers between 1 and n. If ω ∈ Ω k (V ),
then Alt(ω) = ω and so

ω = ∑ ai1 ,...,ik Alt(e∗i1 ⊗ · · · ⊗ e∗ik ).

Fix ξ := e∗i1 ⊗ · · · ⊗ e∗ik . If iα = iβ for some α ̸= β , then the permutation σ0 which

switches α ↔ β preserves ξ . Thus for all σ ∈ Sk ,

sgn(σ σ0 )(σ σ0 ) · ξ = −sgn(σ )σ · ξ

and we conclude that Alt(ξ ) = 0. If, on the other hand, i1 , . . . , ik are all different,
then it is easy to see using lemma 2.1 that Alt(e∗i1 ⊗ e∗i2 ⊗ · · · ⊗ e∗ik ) = k!1 e∗i1 ∧ · · · ∧ e∗ik .
Thus ω = k!1 ∑ ai1 ,...,ik e∗i1 ∧ · · · ∧ e∗ik , and we have proved that the set of forms in
the statement spans Ω k (V ).
To see that this set is independent, we show that it is orthonormal. Suppose
{i1 , . . . , ik } ̸= { j1 , . . . , jk }, then the sets {σ · e∗i1 ⊗ · · · ⊗ e∗ik }, {σ · e∗j1 ⊗ · · · ⊗ e∗jk }
are disjoint, so Alt(e∗i1 ⊗ · · · ⊗ e∗ik ) ⊥ Alt(e∗j1 ⊗ · · · ⊗ e∗jk ). This proves orthogo-
nality. Orthonormality is because ∥e∗i1 ∧ · · · ∧ e∗ik ∥2 = k!∥Alt(e∗i1 ∧ · · · ∧ e∗ik )∥2 =
k! 1
k! ∑σ ∈Sk sgn(σ )σ · (e∗i1 ⊗ · · · ⊗ e∗ik ) = ∑σ ∈Sk sgn(σ )2 ( √1k! )2 = 1. (This explains
2
why we chose to define ∥e∗i1 ⊗ · · · ⊗ e∗ik ∥2 := √1 )
k!
⊔
⊓

Corollary 2.2. e∗1 ∧ · · · ∧ e∗n is the determinant. This is the reason for the peculiar
normalization in the definition of ∧.

Proof. The determinant is an alternating n–form, and dim Ω n (V ) = 1, so the deter-

minant is proportional to e∗1 ∧ · · · ∧ e∗n . Since the values of both forms on the standard
basis is one (because e∗1 ∧ · · · ∧ e∗n = n!Alt(e∗1 ⊗ · · · ⊗ e∗n )), they are equal. ⊔
⊓

We define an inner product on Ω k (V ) by declaring the basis in the proposition to

be orthonormal. Let ∥ · ∥ be the resulting norm.
2.6 The Multiplicative Ergodic Theorem 55

Lemma 2.2. For v ∈ V , let v∗ := ⟨v, · ⟩, then

(a) ∥ω ∧ η∥ ≤ ∥ω∥∥η∥.
(b) ⟨v∗1 ∧ · · · ∧ v∗k , w∗1 ∧ · · · ∧ w∗k ⟩ = det(⟨vi , w j ⟩).
(c) If {u1 , . . . , un } is an orthonormal basis for V , then {u∗i1 ∧ · · · ∧ u∗ik : 1 ≤ i1 < · · · <
ik ≤ n} is an orthonormal basis for Ω k (V ).
(d) If span{v1 , . . . , vk } = span{u1 , . . . , uk }, then v∗1 ∧ · · · ∧ v∗k and u∗1 ∧ · · · ∧ u∗k are
proportional.

Proof. Write for I = (i1 , . . . , ik ) such that 1 ≤ i1 < · · · < ik ≤ n, e∗I := e∗i1 ∧ · · · ∧ e∗ik .
Represent ω := ∑ αI e∗I , η := ∑ βJ e∗J , then
2 2
∥ω ∧ η∥2 = ∑ αI βJ e∗I ∧ e∗J = ∑ ±αI βJ e∗I∪J = ∑ αI2 βJ2 ≤ ∥ω∥2 ∥η∥2 .
I,J I∩J=∅ I∩J=∅

Take two multi indices I, J. If I = J, then the inner product matrix is the identity
matrix. If I ̸= J, then ∃α ∈ I \ J and then the α–row and column of the inner product
matrix will be zero. Thus the formula holds for any pair e∗I , e∗J . Since part (b) of
the lemma holds for all basis vectors, it holds for all vectors. Part (c) immediately
follows.
Next we prove part (d). Represent vi = ∑ αi j u j , then
!
v∗1 ∧ · · · ∧ v∗k = const. Alt(v∗1 ⊗ · · · ⊗ v∗k ) = const. Alt ∑ α1 j u∗j ⊗ · · · ⊗ ∑ αk j u∗j
j j

= const. ∑ α1 j1 · · · αk jk Alt(u∗j1 ⊗ · · · ⊗ u∗jk ).

The terms where j1 , . . . , jk are not all different are annihilated by Alt. The terms
where j1 , . . . , jk are all different are mapped by Alt to a form which proportional to
u∗1 ∧ · · · ∧ u∗k . Thus the result of the sum is proportional to u∗1 ∧ · · · ∧ u∗k . ⊔
⊓

Exterior product of linear operators Let A : V → V be a linear operator. The

k − th exterior product of A is A∧k : Ω k (V ) → Ω k (V ) given by

(A∧k ω)(v1 , . . . , vk ) := ω(At v1 , . . . , At vk ).

The transpose is used to get A∧k (v∗1 ∧ · · · ∧ v∗k ) = (Av1 )∗ ∧ · · · ∧ (Avk )∗ .

Theorem 2.9. ∥A∧k ∥ = λ1 · · · λk , where λ1 ≥ λ2 ≥ · · · ≥ λn are the eigenvalues of
(At A)1/2 , listed in decreasing order with multiplicities.

Proof. The matrix AAt is symmetric, so it can be orthogonally diagonalized. Let

{v1 , . . . , vn } be an orthonormal basis of eigenvectors, listed so that (AAt)vi = λi2 vi .
Then {v∗I : I ⊆ {1, . . . , d}, |I| = k} is an orthonormal basis for Ω k (Rd ), where we are
using the multi index notation

v∗I = v∗i1 ∧ · · · ∧ v∗ik ,

56 2 Ergodic theorems

where i1 < · · · < ik is an ordering of I.

Given ω ∈ Ω k (Rd ), write ω = ∑ ωI v∗I , then
* +
∥A∧k ω∥2 = ⟨A∧k ω, A∧k ω⟩ = ∑ ωI A∧k v∗I , ∑ ωJ A∧k v∗J
I J
D E
= ∑ ωI ωJ A∧k v∗I , A∧k v∗J .
I,J
D E
Now, A∧k v∗I , A∧k v∗J = ⟨(Avi1 )∗ ∧ · · · ∧ (Avik )∗ , (Av j1 )∗ ∧ · · · ∧ (Av jk )∗ ⟩
 
= det ⟨Aviα , Av jβ ⟩ (Lemma 2.2(b))
   
= det ⟨viα , At Av jβ ⟩ = det ⟨viα , λi2β v jβ ⟩
 
= ∏ λ j2 det ⟨viα , viβ ⟩
j∈J
(
∏i∈I λi2 I=J
= ∏ λ j2 ⟨v∗I , v∗J ⟩ = ,
j∈J 0 I ̸= J

Thus ∥A∧k ω∥2 = ∑I ωI2 ∏i∈I λi2 ≤ ∥ω∥2 ∏ki=1 λi2 . It follows that ∥A∧k ∥ ≤ λ1 · · · λk .
To see that the inequality is in fact an equality, consider the case ω = v∗I where
I = {1, . . . , k}: ∥A∧k ω∥ = ⟨v∗I , v∗I ⟩ = (λ1 · · · λk )2 = (λ1 · · · λk )2 ∥ω∥2 . ⊔
⊓

Exterior products and angles between vector spaces The angle between vector
spaces V,W ⊂ Rd is

∡(V,W ) := min{arccos ⟨v, w⟩ : v ∈ V, w ∈ W, ∥v∥ = ∥w∥ = 1}.

So if V ∩W ̸= {0} iff ∡(V,W ) = 0, and V ⊥ W iff ∡(V,W ) = π/2.

Proposition 2.5. If (w1 , . . . , wk ) is a basis of W , and (v1 , . . . , vℓ ) is a basis of W , then
∥(v∗1 ∧ · · · ∧ v∗ℓ ) ∧ (w∗1 ∧ · · · ∧ w∗k )∥ ≤ ∥v∗1 ∧ · · · ∧ v∗ℓ ∥ · ∥w∗1 ∧ · · · ∧ w∗k ∥ · | sin ∡(V,W )|.

Proof. If V ∩W ̸= {0} then both sides are zero, so suppose V ∩W = {0}, and pick
an orthonormal basis e1 , . . . , en+k for V ⊕ W . Let w ∈ W, v ∈ V be unit vectors s.t.
∡(V,W ) = ∡(v, w), and write v = ∑ vi ei , w = ∑ w j e j , then
2.6 The Multiplicative Ergodic Theorem 57
2 2
∥v∗ ∧ w∗ ∥2 = ∑ vi w j e∗i ∧ e∗j = ∑ (vi w j − v j wi )e∗i ∧ e∗j = ∑ (vi w j − v j wi )2
i, j i< j i< j
1
= ∑(vi w j − v j wi )2 (the terms where i = j vanish)
2 i, j
"  2 #
1 2 2 2 2

1 2 2
= ∑ vi w j + v j wi − 2vi wi · v j w j = 2 vi ∑ w j − 2 ∑ vi wi
2 i, j 2 ∑i j i

= ∥v∥2 ∥w∥2 − ⟨v, w⟩2 = 1 − cos2 ∡(v, w) = sin2 ∡(V,W ).

Complete v to an orthonormal basis (v, v′2 , . . . , v′ℓ ) of V , and complete w to an

orthonormal basis (w, w′2 , . . . , w′k ) of W . Then
∗ ∗ ∗
∥(v∗ ∧ v′2 ∧ · · · ∧ v′ℓ ) ∧ (w∗ ∧ w′2 ∧ · · · ∧ v′k )∥
∗ ∗ ∗ ∗
≤ ∥v∗ ∧ w∗ ∥ · ∥v′2 ∧ · · · ∧ v′ℓ ∥ · ∥w′2 ∧ · · · ∧ v′k ∥ = | sin ∡(V,W )| · 1 · 1,

because of orthonormality. By lemma 2.2

∗ ∗
v∗1 ∧ · · · ∧ v∗ℓ = ±∥v∗1 ∧ · · · ∧ v∗ℓ ∥ · v∗ ∧ v′2 ∧ · · · ∧ v′ℓ
∗ ∗
w∗1 ∧ · · · ∧ w∗k = ±∥w∗1 ∧ · · · ∧ w∗k ∥ · w∗ ∧ w′2 ∧ · · · ∧ w′k

and the proposition follows. ⊔

⊓

2.6.2 Proof of the Multiplicative Ergodic Theorem

Let (X, B, m, f ) be a ppt, and A : X → GL(d, R) some Borel map. We define An :=

A ◦ f n−1 · · · A, then the cocycle identity holds: An+m (x) = An ( f m x)Am (x).
Theorem 2.10 (Multiplicative Ergodic Theorem). Let (X, B, T, m) be a ppt, and
A : X → GL(d, R) a Borel function s.t. ln ∥A(x)±1 ∥ ∈ L1 (m), then

Λ (x) := lim [An (x)t An (x)]1/2n

n→∞

exists a.e., and lim n1 ln ∥An (x)Λ (x)−n ∥ = lim n1 ln ∥(An (x)Λ (x)−n )−1 ∥ = 0 a.s.
n→∞ n→∞
p
Proof. The matrix Bn (x) := An (x)t An (x) is symmetric, therefore it can be orthog-
s (x)
onally diagonalized. Let ∃λn1 (x) < · · · < λn n (x) be its different eigenvalues, and
s (x)
λ 1 (x) λ n (x)
Rd = Wn n (x) ⊕ · · · ⊕ Wn n (x) the orthogonal decomposition ot Rd into the
corresponding eigenspaces. The proof has the following structure:
p
Part 1: Let tn1 (x) ≤ · · · ≤ tnd (x) be a list of the eigenvalues of Bn (x) := An (x)t An (x)
with multiplicities, then for a.e. x, there is a limit ti (x) = lim [tni (x)]1/n , i =
n→∞
1, . . . , d.
58 2 Ergodic theorems

Part 2: Let λ1 (x) < · · · < λs(x) (x) be a list of the different values of {ti (x)}di=1 .
Divide {tni (x)}di=1 into s(x) subsets of values {tni (x) : i ∈ Inj }, (1 ≤ j ≤ s(x)) in
such a way that tni (x)1/n → λ j (x) for all i ∈ Inj . Let

Unj (x) := sum of the eigenspaces of tni (x), i ∈ Inj .

= the part of the space where Bn (x) dilates by approximately λ j (x)n .

We show that the spaces Unj (x) converge as n → ∞ to some limiting spaces U j (x)
(in the sense that the orthogonal projections on Unj (x) converge to the orthogonal
projection on U j (x)).
Part 3: The theorem holds with Λ (x) : Rd → Rd given by v 7→ λi (x)v on U i (x).
Part 1 is proved by applying the sub additive ergodic theorem for a cleverly chosen
sub-additive cocyle (“Raghunathan’s trick”). Parts 2 and 3 are (non-trivial) linear
algebra.
(n)
Part 1: Set gi (x) := ∑dj=d−i+1 lntnj (x). This quantity is finite, because Atn An is
invertible, so none of its eigenvalues vanish.
(n)
The sequence gi is subadditive! This is because the theory of exterior products
(n)
says that exp gi = product of the i largest e.v.’s of An (x)t An (x) = ∥An (x)∧i ∥, so
p

(n+m)
exp gi (x) = ∥A∧i n ∧i ∧i n ∧i ∧i
n+m (x)∥ = ∥Am (T x) An (x) ∥ ≤ ∥Am (T x) ∥∥An (x) ∥
(m) (n)
= exp[gi (T n x) + gi (x)],

(n+m) (n) (m)

whence gi ≤ gi + gi ◦ T n .
We want to apply Kingman’s subadditive ergodic theorem. First we need to check
(1)
that gi ∈ L1 . We use the following fact from linear algebra: if λ is an eigenvalue
of a matrix B, then ∥B−1 ∥−1 ≤ |λ | ≤ ∥B∥.5 Therefore

1
| lntni (x)| ≤ max{| ln ∥Atn An ∥|, | ln ∥(Atn An )−1 ∥|}
2
≤ max{| ln ∥An (x)∥|, | ln ∥An (x)−1 ∥|}
n−1  
≤ ∑ | ln ∥A(T k x)∥ + | ln ∥A(T k x)−1 ∥|
k=0
n−1  
(n)
∴ |gi (x)| ≤ i ∑ | ln ∥A(T k x)∥ + | ln ∥A(T k x)−1 ∥| . (2.5)
k=0

So ∥g(n) ∥1 ≤ n( ln ∥A∥ 1
+ ln ∥A−1 ∥ 1 ) < ∞.
(n)
Thus Kingman’s ergodic theorem says that lim n1 gi (x) exists almost surely, and
belongs to [−∞, ∞). In fact the limit is finite almost everywhere, because (2.5) and
5Proof: Let v be an eigenvector of λ with norm one, then |λ | = ∥Bv∥ ≤ ∥B∥ and 1 = ∥B−1 Bv∥ ≤
∥B−1 ∥∥Bv∥ = ∥B−1 ∥|λ |.
2.6 The Multiplicative Ergodic Theorem 59

the Pointwise Ergodic Theorem imply that

1 (n)
|g | ≤ iE | ln ∥A∥| + | ln ∥A−1 ∥| Inv < ∞ a.e.


lim
n→∞ n i
Taking differences, we see that the following limit exists a.e.:
1 (n) (n) 1
lnti (x) := lim [g (x) − gd−i (x)] = lim lntni (x).
n→∞ n d−i+1 n→∞ n

Thus [tni (x)]1/n −−−→ ti (x) almost surely, for some ti (x) ∈ R.
n→∞

Part 2: Fix x s.t. [tni (x)]1/n −−−→ ti (x) for all 1 ≤ i ≤ d. Henceforth we work with
n→∞
this x only, and write for simplicity An = An (x), ti = ti (x) etc.
Let s = s(x) be the number of the different ti . List the different values of these
quantities an increasing order: λ1 < λ2 < · · · < λs . Set χ j := log λ j . Fix 0 < δ <
1 i 1/n → λ , the following sets
2 min{χ j+1 − χ j }. Since for all i there is a j s.t. (tn ) j
eventually stabilize and are independent of n:

I j := {i : |(tni )1/n − λ j | < δ } ( j = 1, . . . , s).

p
Define, relative to Atn An ,

• Unj := ∑i∈I j [eigenspace of tni (x)] (this sum is not necessarily direct);
• Vnr := ⊕ j≤rUnj
• Venr := ⊕ j≥rUnj
The linear spaces Un1 , . . . ,Uns are orthogonal,
p since they are eigenspaces of different
eigenvalues for a symmetric matrix ( Atn An ). We show that they converge as n → ∞,
in the sense that their orthogonal projections converge.
The proof is based on the following technical lemma. Denote the projection of a
vector v on a subspace W by v|W , and write χi := log λi .
Technical lemma: For every δ > 0 there exists constants K1 , . . . , Ks > 1 and N s.t.
for all n > N, t = 1, . . . , s, k ∈ N, and u ∈ Vnr ,
r+t
u|Ven+k ≤ Kt ∥u∥ exp(−n(χr+t − χr − δt))

We give the proof later. First we show how it can be used to finish parts 2 and 3.
We show that Vnr converge as n → ∞. Since the projection on Uni is the projection
on Vni minus the projection on Vni−1 , it will then follow that the projections of Uni
converge.
Fix N large. We need it to be so large that
1. I j are independent of n for all n > N;
2. The technical lemma works for n > N with δ as above.
There will be other requirements below.
60 2 Ergodic theorems

Fix an orthonormal basis (v1n , . . . , vdnr ) for Vnr (dr = dim(Vnr ) = ∑ j≤r |I j |). Write

vin = αni win+1 + uin+1 , where win+1 ∈ Vn+1

r
, ∥win+1 ∥ = 1, uin+1 ∈ Ven+1
r+1
.
r+1
Note that ∥uin+1 ∥ = ∥vin |Ven+1 ∥ ≤ K1 exp(−n(χr+1 − χr − δ )). Using the identity
q
i i 2
αn = 1 − ∥un+1 ∥ , it is easy to see that for some constants C1 and 0 < θ < 1
independent of n and (vin ),

∥vin − win+1 ∥ ≤ C1 θ n .

(θ := maxr exp[−(χr+1 − χr − δ )] and C1 := 2K1 should work.)

The system {win+1 } is very close to being orthonormal:
j j j
⟨win+1 , wn+1 ⟩ = ⟨win+1 − vin , wn+1 ⟩ + ⟨vin , vnj ⟩ + ⟨vin , wn+1 − vnj ⟩ = δi j + O(θ n ),

because {vin } is an orthonormal system. It follows that for all n large enough, win+1
are linearly independent. A quick way to see this is to note that
j
∥(w1n+1 )∗ ∧ · · · ∧ (wdn+1
r
)∗ ∥2 = det(⟨win+1 , wn+1 ⟩) (lemma 2.2)
= det(I + O(θ n )) ̸= 0, provided n is large enough,

and to observe that wedge produce of a linearly dependent system vanishes.

It follows that {w1n+1 , . . . , wdn+1
r r . Since
} is a linearly independent subset of Vn+1
r r
dim(Vn+1 ) = ∑ j≤r |I j | = dr , this is a basis for Vn+1 .
Let (vin+1 ) be the orthonormal basis obtained by applying the Gram–Schmidt
procedure to (win+1 ). We claim that there is a global constant C2 such that

∥vin − vin+1 ∥ ≤ C2 θ n . (2.6)

Write vi = vin+1 , wi = win+1 , then the Gram–Schmidt process is to set vi = ui /∥ui ∥,

where ui are defined by induction by u1 := w1 , ui := wi − ∑ j<i ⟨wi , v j ⟩v j . We con-
struct by induction global constants C2i s.t. ∥vi − wi ∥ ≤ C2i θ n , and then take C2 :=
max{C2i }. When i = 1, we can take C21 := C1 , because v1 = w1 , and ∥w1 − v1n ∥ ≤
C1 θ n . Suppose we have constructed C21 , . . . ,C2i−1 . Then
!
∥ui − wi ∥ ≤ ∑ |⟨wi , v j ⟩| ≤ ∑ |⟨wi , w j ⟩| + ∥w j − v j ∥ ≤ 2C1 (i − 1) + ∑ C2j θ n ,
j<i j<i j<i

because |⟨wi , w j ⟩| = |⟨wi − vin , w j ⟩ + ⟨vin , w j − vnj ⟩ + ⟨vin , vnj ⟩| ≤ 2C1 θ n . Call the term
in the brackets K, and assume n is so large that Kθ n < 1/2, then |∥ui ∥ − 1| ≤ ∥ui −
wi ∥ ≤ Kθ n , whence

ui − ∥ui ∥wi ∥ui − wi ∥ + |1 − ∥ui ∥|

∥vi − wi ∥ = ≤ ≤ 4Kθ n
∥ui ∥ ∥ui ∥
2.6 The Multiplicative Ergodic Theorem 61

and we can take C2i := 4K. This proves (2.6).

Starting from the orthonormal basis (vin ) for Vnr , we have constructed an orthonor-
mal basis (vin+1 ) for Vn+1
r such that ∥vin − vin+1 ∥ ≤ C2 θ n . Continue this procedure by
induction, and construct the orthonormal bases (vin+k ) for Vn+k r . By (2.6), these bases

form Cauchy sequences: vin+k −−−→ vi .

k→∞
The limit vectors must also be orthonormal. Denote their span by V r . The pro-
jection on V r takes the form
dr dr
lim ∑ ⟨vin+k , · ⟩vin+k = lim projV r .
∑ ⟨vi , · ⟩vi = k→∞ k→∞ n+k
i=1 i=1

r → V r.
Thus Vn+k
Part 3: We saw that proj Uni (x) −−−→ proj U i (x) for some linear spaces Ui (x). Set
n→∞
Λ (x) ∈ GL(Rd ) to be the matrix representing

s(x)
Λ (x) = ∑ eχ j (x) proj Ui (x) .
j=1

Since Ui (x) are limits of Uni , they are orthogonal, and they sum up to Rd . It follows
that Λ is invertible, symmetric, and positive. p
Choose an orthogonal basis {v1n (x), . . . , vdn (x)} of Bn (x) := An (x)t An (x) so that
Bn vin = tni vin for all i, and let Wni := span{vin }. Then for all v ∈ Rd ,
d
(Atn An )1/2n v = ( Atn An )1/n v = ∑ tni (x)1/n proj Wni (v)
p
i=1
s
= ∑ ∑ tni (x)1/n proj Wni (v)
j=1 i∈I j
s
= ∑ eχ j (x) ∑ proj Wni (v) + o(∥v∥),
j=1 i∈I j

where o(∥v∥) denotes a vector with norm o(∥v∥)

s
= ∑ eχ j (x) proj Unj (v) + o(∥v∥) −n→∞
−−→ Λ (x)v.
j=1

Thus (Atn An )1/2n → Λ .

We show that n1 log ∥(AnΛ −n )±1 ∥ −−−→ 0. It’s enough to show that
n→∞

1
lim log ∥An v∥ = χr := log λr uniformly on the unit ball in Ur . (2.7)
n→∞ n
To see that this is enough, note that Λ v = ∑sr=1 eχr (v|Ur ); for all δ > 0, if n is large
enough, then for every v,
62 2 Ergodic theorems
s s
∥AnΛ −n v∥ ≤ ∑ e−nχr ∥An (v|Ur )∥ = ∑ e−nχr en(χr +δ ) ∥v∥ ≤ senδ ∥v∥ (v ∈ Rd )
r=1 r=1
∥AnΛ −n v∥ = e−nχr ∥An v∥ = e±nδ ∥v∥ (v ∈ Ur )

Thus ∥AnΛ −n ∥ ≍ e−nδ for all δ , whence n1 log ∥AnΛ −n ∥ → 0 a.e.

To see that 1n log ∥(AnΛ −n )−1 ∥ → 0, we use a duality trick.
Define for # := (C−1 )t , then (C C )# = C#C# . Thus (A# ) = (A )# ,
p a matrix C, C p p 1 2 1 2 n n
and Bn := (A# )tn (A# )n = ( Atn An )# = ( Atn An )−1 . Thus we have the following
#

relation between the objects associated to A# and A:

1. the eigenvalues of B#n are 1/tnd ≤ · · · ≤ 1/tn1 (the order is flipped)
2. the eigenspace of 1/tni for B#n is the eigenspace of tni for Bn
3. χ #j = −χs− j+1
4. (Unj )# = Uns− j+1 , (Vnr )# = Vens−r+1 , (Venr )# = Vns−r+1
5. Λ # = Λ −1 .
Thus ∥(Λ n A−1 n −1 t # # −n 1 n −1 −−
n )∥ = ∥(Λ An ) ∥ = ∥An (Λ ) ∥, so the claim n log ∥Λ An ∥ − →0
n→∞
a.e. follows from what we did above, applied to A# .
Here is another consequence of this duality: There exist K1# , . . . , Kt# s.t. for all δ ,
there is an N s.t. for all n > N, if u ∈ Unr , then for all k
r−t
∥u|Vn+k ∥ ≤ Kt# exp[−n(χr − χr−t − δ )]. (2.8)

r−t s−r+t+1 #
To see this note that Vn+k = (Ven+k ) and Unr ⊂ (Vns−r+1 )# , and apply the techni-
cal lemma to the cocycle generated by A# .
We prove (2.7). Fix δ > 0 and N large (we see how large later), and assume
n > N. Suppose v ∈ Ur and ∥v∥ = 1. Write v = lim vn+k with vn+k := v|Un+k k ∈ Ur .
n+k
Note that ∥vn+k ∥ ≤ 1. We decompose vn+k as follows
s−r
vn+k = vn+k |Vnr−1 + (vn+k |Unr ) + ∑ vn+k |Unr+t ,



t=1

and estimate the size of the image of each of the summands under An .
First summand:

∥An (vn+k |Vnr−1 )∥2 ≡ ⟨B2n (vn+k |Vnr−1 ), (vn+k |Vnr−1 )⟩

= e2n(χr−1 +o(1)) ∥vn+k |Vnr−1 ∥ ≤ e2n(χr−1 +o(1)) .

Thus the first summand is less than exp[n(χr−1 + o(1))].

Second Summand:
2.6 The Multiplicative Ergodic Theorem 63

∥An (vn+k |Unr )∥2 = ⟨B2n (vn+k |Unr ), (vn+k |Unr )⟩

= e2n(χr +o(1)) ∥vn+k |Unr ∥2 = e2n(χr +o(1)) (∥v|Unr ∥ ± (∥(vn+k − v)|Unr ∥)2
= e2n(χr ±δ ) (∥v|Unr ∥ ± ∥vn+k − v∥)2 = e2n(χr ±δ ) [1 + o(1)] uniformly in v.

Thus the second summand is [1 + o(1)] exp[n(χr ± δ )] uniformly in v ∈ Ur , ∥v∥ = 1.

Third Summand: For every t,

∥An (vn+k |Unt )∥2 = ⟨B2n (vn+k |Unt ), (vn+k |Unt )⟩

≤ e2n(χr+t +o(1)) ∥vn+k |Unr+t ∥2
!2
2n(χr+t +o(1))
≡e sup ⟨vn+k , u⟩ , because ∥x|W ∥ = sup ⟨x, w⟩
u∈Unr+t ,∥u∥=1 w∈W,∥w∥=1
!2
≤ e2n(χr+t +o(1)) sup sup ⟨v, u⟩
r ,∥v∥≤1
u∈Unr+t ,∥u∥=1 v∈Vn+k

= e2n(χr+t +o(1)) sup r

∥u|Vn+k ∥2
u∈Unr+t ,∥u∥=1

≤ (Kt# )2 e2n(χr+t +o(1)) exp[−2n(χr+t − χr − o(1))], by (2.8)

= (Kt# )2 e2n(χr +o(1)) .

Note that the cancellation of χr+t — this is the essence of the technical lemma. We
get: ∥An (vn+k |Unt )∥ = O(exp[n(χr + o(1))]). Summing over t = 1, . . . , s − r, we get
that third summand is O(exp[n(χr + o(1))]).
Putting these estimates together, we get that

∥An vn+k ∥ ≤ const. exp[n(χr + o(1))] uniformly in k, and on the unit ball in Ur .

“Uniformity” means that the o(1) can be made independent of v and k. It allows us
to pass to the limit as k → ∞ and obtain

∥An v∥ ≤ const. exp[n(χr + o(1))] uniformly on the unit ball in Ur .

On the other hand, an orthogonality argument shows that

∥An vn+k ∥2 = ⟨B2n vn+k , vn+k ⟩

= ∥1st summand∥2 + ∥2nd summand∥2 + ∥3rd summand∥2
≥ ∥2nd summand∥2 = [1 + o(1)] exp[2n(χr + o(1))].

Thus ∥An vn+k ∥ ≥ [1 + o(1)] exp[n(χr + o(1))] uniformly in v, k. Passing to the limit
as k → ∞, we get ∥An v∥ ≥ const. exp[n(χr + o(1))] uniformly on the unit ball in Ur .
These estimates imply (2.7).
64 2 Ergodic theorems

Proof of the technical lemma: We are asked to estimate the norm of the projection
r+t
of a vector in Vnr on Vn+k . We so this in three steps:
r+t
1. Vnr → Vn+1 , all t > 0;
r r+1
2. Vn → Vn+k , all k > 0;
r+t
3. Vnr → Vn+k , all t, k > 0.

Step 1. The technical lemma for k = 1: Fix δ > 0, then for all n large enough and
r′ ∥ ≤ ∥u∥ exp(−n(χ ′ − χ − δ )).
for all r′ > r, if u ∈ Vnr , then ∥u|Vn+1 r r

Proof. Fix ε, and choose N = N(ε) so large that tni = e±nε ti for all n > N, i = 1, . . . , d.
For every t = 1, . . . , s, if u ∈ Vnr , then
q
∥An+1 u∥ = ⟨Atn+1 An+1 u, u⟩
q
r+t r+t r+t−1 r+t−1
= ⟨Atn+1 An+1 (u|Ven+1 ), (u|Ven+1 )⟩ + ⟨Atn+1 An+1 (u|Vn+1 ), (u|Vn+1 )⟩
r+t−1 e r+t
(because Vn+1 , Vn+1 are orthogonal, Atn+1 An+1 –invariant,
and Rd = Vn+1
r+t−1 r+t
⊕ Ven+1 )
q
r+t 2 r+t−1 2
= ∥An+1 (u|Ven+1 )∥ + ∥An+1 (u|Vn+1 )∥
r+t
≥ ∥An+1 (u|Ven+1 )∥ = e(χr+t ±ε)(n+1) ∥u|Ven+1
r+t
∥.

On the other hand

p
∥An+1 u∥ = ∥A(T n x)An (x)u∥ ≤ ∥A(T n x)∥ ⟨Atn An u, u⟩
≤ ∥A(T n x)∥en(χr ±ε) ∥u∥
= en(χr ±ε)+o(n) ∥u∥,

because by the ergodic theorem

1 1 n 1 n−1
log ∥A(T n x)∥ = ∑ log ∥A(T k x)∥ − ∑ log ∥A(T k x)∥ −−−→ 0 a.e.
n n k=0 n k=0 n→∞

By further increasing N, we can arrange |o(n)| < nε, which gives

e(χr+t −ε)(n+1) ∥u|Ven+1

r+t
∥ ≤ en(χr +2ε) ,
r+t
whence ∥u|Ven+1 ∥ ≤ e−n(χr+t −χr −3ε) . Now take ε := δ /3.
Step 2. Fix δ > 0. Then for all n large enough and for all k, if u ∈ Vnr , then
∥u|Ve r+1 ∥ ≤ ∥u∥ ∑k−1 exp(−(n + j)(χr+1 − χr − δ )). Thus ∃K1 s.t.
n+k j=0

r+1
∥u|Ven+k ∥ ≤ K1 ∥u∥ exp[−n(χr+1 − χr − δ )].
2.6 The Multiplicative Ergodic Theorem 65

Proof. We use induction on k. The case k = 1 is dealt with in step 1. We assume by

induction that the statement holds for k − 1, and prove it for k. Decompose
r+1 r r+1 r+1 r+1
u|Ven+k = [(u|Vn+k−1 ) Ven+k ] + [(u|Ven+k−1 ) Ven+k ].

• First summand: u|Vn+k−1r r

∈ Vn+k−1 , so by step 1 the norm of the first sum-
r
mand is less than ∥u|Vn+k−1 ∥ exp[−(n + k − 1)(χr+1 − χr − δ )], whence less than
∥u∥ exp[−(n + k − 1)(χr+1 − χr − δ )].
r+1
• Second summand: The norm is at most ∥u|Ven+k−1 ∥. By the induction hypothesis,
k−2
this is less than ∥u∥ ∑ j=0 exp(−(n + j)(χr+1 − χr − δ )).
We get the statement for k, and step 2 follows by induction.
As a result, we obtain the existence of a constant K1 > 1 for which u ∈ Vnr implies
∥u|Ve r+1 ∥ ≤ K1 ∥u∥ exp(−n(χr+1 − χr − δ )).
n+k
Step 3. ∃K1 , . . . , Ks−1 > 1 s.t. for all n large enough and for all k, u ∈ Vnr implies
r+ℓ
∥u|Ven+k ∥ ≤ Kℓ ∥u∥ exp(−n(χr+ℓ − χr − ℓδ )) (ℓ = 1, . . . , s − r).
Proof. We saw that K1 exists. We assume by induction that K1 , . . . , Kt−1 exist, and
construct Kt . Fix 0 < δ0 < min j {χ j+1 − χ j } − δ ; the idea is to first prove that if
u ∈ Vnr , then
! ! !
t−1 k−1 k−1
r+t −δ0 j
∥u|V ∥ ≤ ∥u∥ ∑ Kτ
e
n+k ∑e ∑ exp[−(n + j)(χr+t − χr − tδ )] (2.9)
τ=1 j=0 j=0

Once this is done, step 3 follows with Kt := ∑t−1 −δ0 j 2 .

τ=1 Kτ ∑ j≥0 e
We prove (2.9) using induction on k. When k = 1 this is because of step 1. Sup-
pose, by induction, that (2.9) holds for k − 1. Decompose:
r+t r r′
u|Ven+k = u|Vn+k−1 |Ve r+t + ∑ u|Un+k−1 r+t
|Ven+k r+t
+ u|Ven+k−1 |Ve r+t
| {z n+k} r<r′ <r+t | {z n+k}
A | {z } C
B

• Estimate of ∥A∥: By step 1, ∥A∥ ≤ ∥u∥ exp(−(n + k − 1)(χr+t − χr − δ )).

• Estimate of ∥B∥: By step 1, and the induction hypothesis (on t):
r′
∥B∥ ≤ ∑ ∥u|Un+k−1 ∥ exp(−(n + k − 1)(χr+t − χr′ − δ ))
′
r<r <r+t
′
r
≤ ∑ ∥u|Ven+k−1 ∥ exp(−(n + k − 1)(χr+t − χr′ − δ ))
′
r<r <r+t
≤ ∑ Kr′ −r ∥u∥ exp(−n(χr′ − χr − (r′ − r)δ ))×
r<r′ <r+t
× exp(−(n + k − 1)(χr+t − χr′ − δ ))
66 2 Ergodic theorems
!
t−1
≤ ∥u∥ ∑ Kt ′ e−(k−1)(χr+t −χr′ −δ ) exp(−n(χr+t − χr − tδ ))
′
t =1
!
t−1
≤ ∥u∥ ∑ Kt ′ e−δ0 (k−1) exp(−n(χr+t − χr − tδ )).
′
t =1

r+t
• Estimate of ∥C∥: ∥C∥ ≤ ∥u|Ven+k−1 ∥. By the induction hypothesis on k,
! ! !
t−1 k−2 k−2
∥C∥ ≤ ∥u∥ ∑ Kt ′ ∑ e−δ0 j ∑ exp[−(n + j)(χr+t − χr − tδ )] .
t ′ =1 j=0 j=0

It is not difficult to see that when we add these bounds for ∥C∥, ∥B∥ and ∥A∥, the
result is smaller than the RHS of (2.9) for k. This completes the proof by induction
of (2.9). As explained above, step 3 follows by induction. ⊔
⊓

Corollary 2.3. Let χ1 (x) < · · · < χs(x) (x) denote the logarithms of the (different)
eigenvalues
L
of Λ (x). Let Uχi be the eigenspace of Λ (x) corresponding to exp χi . Set
Vχ := χ ′ ≤χ Uχ ′ .
1. χ(x, v) := lim n1 log ∥An (x)v∥ exists a.s, and is invariant.
n→∞
2. χ(x, v) = χi on Vχi \Vχi−1
3. If ∥A−1 ∥, ∥A∥ ∈ L∞ , then 1n log | det An (x)| = ∑ ki χi , where ki = dimUχi .
{χi (x)} are called the Lyapunov exponents of x. {Vχi } is called the Lyapunov fil-
tration of x. Property (2) implies that {Vχ } is A–invariant: A(x)Vχ (x) = Vχ (T x).
Property (3) is sometimes called regularity.

Remark: Vχi \ Vχi−1 is A–invariant, but if A(x) is not orthogonal, then UχiL doesn’t
need to be A–invariant. When T is invertible, there is a way of writing Vχi = j≤i H j
so that A(x)H j (x) = H j (T x) and χ(x, ·) = χ j on H j (x), see the next section.

2.6.3 The Multiplicative Ergodic Theorem for Invertible Cocycles

Suppose A : X → GL(n, R). There is a unique extension of the definition of An (x)

to non-positive n’s, which preserves the cocycle identity: A0 := id , A−n := (An ◦
T −n )−1 . (Start from An−n = A0 = id and use the cocycle identity.)
The following theorem establishes a compatibility between the Lyapunov spectra
and filtrations of An and A−n .
Theorem 2.11. Let (X, B, m, T ) be an invertible probability preserving transforma-
tion, and A : X → GL(d, R) a Borel function s.t. ln ∥A(x)±1 ∥. There are invariant
L p(x)
Borel functions p(x), χ1 (x) < · · · < χ p(x) (x), and a splitting Rd = i=1 H i (x) s.t.
1. An (x)H i (x) = H i (T n x) for all n ∈ Z
2.6 The Multiplicative Ergodic Theorem 67

1
2. lim log ∥An (x)v∥ = ±χi (x) on the unit sphere in H i (x).
n→±∞ |n|
1 i n j n
3. n log sin ∡(H (T x), H (T x)) − −−→ 0.
n→∞

Proof. Fix x, and let Let tn1 ≤ · · · ≤ tnd and t 1n ≤ · · · ≤ t dn be the eigenvalues of
(Atn An )1/2 and (At−n A−n )1/2 . Let ti := lim(tni )1/n , t i = lim(t in )1/n . These limits ex-
ists almost surely, and {logti }, {logt i } are lists of the Lyapunov exponents of An
and A−n , repeated with multiplicity. The proof of the Oseledets theorem shows that
d
1
lim log ∥A∧i
∑logt i ≡ n→∞ −n ∥
k=d−i+1 n
1  
= lim log | det A−n |·∥((A−n )−1 )∧(d−i) ∥ (write using e.v.’s)
n→∞ n
1 
∧(d−i)

≡ lim log |(det An ◦ T −n )|−1 ·∥An ◦ T −n ∥
n→∞ n
1 
∧(d−i)

≡ lim log | det An |−1 ·∥An ∥ (remark after Kingman’s Theorem)
n→∞ n
d d d−i
= ∑ logtk − ∑ logtk = − ∑ logtk .
k=d−i+1 k=1 k=1

Since this is true for all i, logti = − logt d−i+1 .

It follows that if the Lyapunov exponents of An are χ1 < . . . < χs , then the Lya-
punov exponents of A−n are −χs < · · · < −χ1 .
Let V 1 (x) ⊂ V 2 (x) ⊂ · · · ⊂ V s (x) be the Lyapunov filtration of An :

1
V i (x) := {v : lim log ∥An (x)v∥ ≤ χi (x)}.
n→∞ n
1 2 s
Let V (x) ⊃V (x) ⊃ · · · ⊃V (x) be the following decreasing filtration, given by

i 1
V (x) := {v : lim log ∥A−n (x)v∥ ≤ −χi (x)}.
n→∞ n

i i
These filtrations are invariant: A(x)V i (x) = V i (T x), A(x)V (x) = V (T x).
i
Set H i (x) := V i (x) ∩V (x). We must have A(x)H i (x) = H i (T x).
We claim that R = H i (x) almost surely. It is enough to show that for a.e. x,
d L
i+1
R = V i (x) ⊕V (x), because
d

1 1 2 2
Rd ≡ V = V ∩ [V 1 ⊕V ] (V 1 ⊕V = Rd )
1 2 2 1 2
= H 1 ⊕ [V ∩V ] = H 1 ⊕V (V ⊇ V )
2 3 3
= H 1 ⊕ [V ∩ (V 2 ⊕V )] (V 2 ⊕V = Rd )
3
= H 1 ⊕ H 2 ⊕V = · · · = H 1 ⊕ · · · ⊕ H s .
68 2 Ergodic theorems
i+1
Since the spectra of Λ , Λ agree with matching multiplicities, dimV i +dimV =
i+1
d. Thus it is enough to show that E := {x : V i (x) ∩V (x) ̸= {0}} has zero measure
for all i.
Assume otherwise, then by the Poincaré recurrence theorem, for almost every
x ∈ E there is a sequence nk → ∞ for which T nk (x) ∈ E. By the Oseledets theorem,
for every δ > 0, there is Nδ (x) such that for all n > Nδ (x),
i+1
∥An (x)u∥ ≤ ∥u∥ exp[n(χi + δ )] for all u ∈ V i ∩V , (2.10)
i+1
∥A−n (x)u∥ ≤ ∥u∥ exp[−n(χi+1 − δ )] for all u ∈ V i ∩V . (2.11)
i+1
If nk > Nδ (x), then Ank (x)u ∈ V i (T nk x) ∩V (T nk x) and T nk (x) ∈ E, so

∥u∥ = ∥A−nk (T nk x)Ank (x)u∥ ≤ ∥Ank (x)u∥ exp[−nk (χi+1 − δ )],

whence ∥Ank (x)u∥ ≥ ∥u∥ exp[nk (χi+1 − δ )]. By (2.10),

exp[nk (χi+1 + δ )] ≤ exp[nk (χi + δ )],

whence |χi+1 − χi | < 2δ . But δ was arbitrary, and could be chosen to be much
smaller than the gaps between the Lyapunov exponents. With this choice, we get a
contradiction which shows that m(E) = 0.
Thus Rd = H i (x). Evidently, V i = V i ⊇ j≤i H j and V i ∩ j>i H j ⊆ V i ∩
L L L
i+1 i
= {0}, so V i = j. j. It follows that H i ⊂
L L
V j≤i H In the same way V = j≥i H
i i+1
(V i \V i−1 ) ∩ (V \V ). Thus lim 1 log ∥An v∥ = ±χi on the unit sphere in H i .
n→±∞ |n|
Next we study the angle between H i (x) and H e i (x) := L j̸=i H j (x). Pick a basis
(vi1 , . . . , vimi ) for H i (x). Pick a basis (wi1 , . . . , wimi ) for He i (x). Since An (x) is invertible,
i
Ak (x) maps (v1 , . . . , vmi ) onto a basis of H (T x), and (wi1 , . . . , wimi ) onto a basis of
i i k

e i (T k x). Thus if v := V vi , w := V wi , then

H j j

∥An (x)∧d (v ∧ w)∥

| sin ∡(H i (T k x), H(T
e k x))| ≥ .
∥An (x)∧mi v∥ · ∥An (x)∧(d−mi ) w∥

We view A∧p ∗ ∗
n as an invertible matrix acting on span{ei1 ∧ · · · ∧ ei p : i1 < · · · < i p } via
∗ ∗
(An (x)ei1 ) ∧ · · · ∧ (An (x)ei p ) . It is clear
 ∧p
Λ p (x) := lim ((A∧p ∗ ∧p 1/2n
n ) (An )) = lim (A∗n An )1/2n = Λ (x)∧p ,
n→∞ n→∞

thus the eigenspaces of Λ p (x) are the tensor products of the eigenspaces of Λ (x).
This determines the Lyapunov filtration of An (x)∧p , and implies – by Oseledets
theorem – that if v j ∈ Vχk( j) \Vχk( j)−1 , and v1 , . . . , vk are linearly independent, then
2.7 The Karlsson-Margulis ergodic theorem 69
p
1
lim log ∥An (x)∧p ω∥ = ∑ χk( j) , for ω := v1 ∧ · · · ∧ v p .
n→∞ n
j=1

It follows that lim 1n log | sin ∡(H i (T n x), H

e i (T n x))| ≥ 0. ⊔
⊓
n→∞

2.7 The Karlsson-Margulis ergodic theorem

Suppose (X, d) is a “nice” metric space, and Isom(X) is the group of isometries of
X. Let (Ω , F , µ, T ) be Ra ppt and f : Ω → Isom(X) be a measurable map such that
for some (any) x0 ∈ X, X d( f (ω) · x, x0 )dµ(ω) < ∞.
Consider the “random walk” xn (ω) := f (T n−1 ω) ◦ · · · ◦ f (ω) · x0 . It is not
difficult to see that g(n) (ω) := d(x0 , xn (ω)) is a sub-additive cocycle. The sub-
additive ergodic theorem, implies the existence of almost sure asymptotic speed
s(ω) = lim 1n d(x0 , xn (ω)). The Karlsson-Margulis theorem provides (under addi-
n→∞
tional assumptions on (X, d)), the existence of an asymptotic velocity: A geodesic
ray γω (t) ⊂ X which starts at x0 s.t. d(xn (ω), γω (s(ω)n)) = o(n) as n → ∞.
Example 1 (Birkhoff Ergodic Theorem): Take an ergodic ppt (Ω , F , µ, T ), X =
Rd and f (ω) · v := v + f (ω) where f (ω) := ( f1 (ω), . . . , fd (ω)) and fi : Ω → R are
absolutely integrable with non-zero integral. Then

1 n−1 1 n−1
f1 (T i ω), . . . , ∑ fd (T i (ω)


xn (ω) = x0 + n ∑
n i=0 n i=0
R R
so s(ω) = ∥v∥2 a.e. and γω (t) = tv/∥v∥ a.e., where v := ( f1 dµ, . . . , fd dµ).
Example 2 (Multiplicative Ergodic Theorem): See section 2.7.3 below.

2.7.1 The boundary of a non-compact proper metric space

Let (X, d) be a metric space. We need some terminology:

1. X is called proper is every closed bounded subset of X is compact.
2. A curve is a continuous function γ : [a, b] → X. A curve is called rectifiable if
( )
n−1
ℓ(g) := sup ∑ d(γ(ti ), γ(ti+1 )) : a = t0 < t1 < · · · < tn = b, n ≥ 1 < ∞.
i=0

The number ℓ(γ) is called the length of γ.

3. A geodesic segment from A to B (A, B ∈ X) is a curve γ : [0, L] → X s.t. γ(0) = A,
γ(L) = B, and d(γ(t), γ(t ′ )) = |t −t ′ | for all 0 ≤ t,t ′ ≤ L. In particular, L = d(x, y).
We denote such segments by AB.
70 2 Ergodic theorems

4. A geodesic ray is a curve γ : [0, ∞) → X s.t. d(γ(t), γ(t ′ )) = |t − t ′ | for all t,t ′ .
5. A metric space is called a geodesic space, if every A, B ∈ X are connected by a
geodesic segment.
6. A geodesic metric space is called complete if every geodesic segment can be
extended to a geodesic ray in the positive direction.
Suppose X is a non–compact proper geodesic metric space. We are interested in
describing the different ways of “escaping to infinity” in X.
The idea is to compactify X by adding to it a “boundary” so that tending to
infinity in X in a certain “direction” corresponds to tending to a point in the boundary
of X in X.
b
Fix once and for all a reference point x0 ∈ X (the “origin”), and define for each
x ∈ X the function
Dx (z) := d(z, x) − d(x0 , x).
Dx (·) is has Lipschitz constant 1, and Dx (x0 ) = 0. It follows that {Dx (·) : x ∈ X} is
equicontinuous and uniformly bounded on compact subsets of X.
By the Arzela–Ascoli theorem, every sequence {Dxn }n≥1 has a subsequence
{Dxnk }k≥1 which converges pointwise (in fact uniformly on compacts). Let

Xb := { lim Dxn (·) : {xn }n≥1 ⊂ X, {Dxn }n≥1 converges uniformly on compacts}.
n→∞

We put a metric on Xb as follows:

∞
1
d( f , g) := ∑ 2n ϕ(d(xmax | f (x) − g(x)|),
n=1 0 ,x)≤n

where ϕ(t) := π1 arctan(t) + 12 (a homeomorphism ϕ : R → (0, 1)). This metric is

well-defined, since X is proper. The resulting topology is the topology of uniform
convergence on compacts.

Theorem 2.12. Suppose (X, d) is a proper geodesic space, then Xb is a compact

metric space, and if ı : X ,→ Xb is the map ı(x) = Dx (·), then
1. ı : X → ı(X) is a homeomorphism;
2. ı(X) is dense in X.
b

Proof. Xb is compact, because it is the closure of {Dx (·) : x ∈ X}, which is precom-
pact by Arzela–Ascoli.
The map ı : x 7→ Dx (·) is one-to-one because x can be read of Dx (·) as the unique
point where that function attains its minimum. The map ı is continuous, because if
d(xn , x) → 0, then

|Dxn (z) − Dx (z)| ≤ |d(z, xn ) − d(z, x)| + |d(xn , x0 ) − d(x, x0 )|

≤ 2d(xn , x) −−−→ 0.
n→∞
2.7 The Karlsson-Margulis ergodic theorem 71

To see that ı−1 is continuous, we first note that since Xb is metrizable, it is enough
to show that if Dxn → Dx uniformly on compacts, then xn → x in X. Suppose Dxn →
Dx uniformly on compacts, and fix ε > 0. Suppose by way of contradiction that
∃nk ↑ ∞ s.t. d(xnk , x) ≥ ε. Construct ynk on the geodesic segment connecting x to xnk
s.t. d(x, ynk ) = ε/2. We have

Dxnk (ynk ) = d(ynk , xnk ) − d(xnk , x0 ) = d(x, xnk ) − d(x, ynk ) − d(xnk , x0 )
ε
= Dxnk (x) − .
2
Since d(ynk , x) = ε/2 and X is proper, ynk lie in a compact subset of X. W.l.o.g.
ynk −−−→ y ∈ X. Passing to the limit we see that Dx (y) = Dx (x) − ε2 < Dx (x). But
k→∞
this is absurd, since Dx attains its minimum at x. It follows that xn → x. ⊔
⊓
Terminology: Xb is called the horofunction compactification of X, and ∂ X :=
Xb \ ı(X) is called the horofunction boundary of X. Elements of ∂ X are called horo-
functions.
The horofunction compactification has a very nice geometric interpretation in
case (X, d) has “non–positive curvature”, a notion we now proceed to make precise.
Suppose (X, d) is a geodesic space, then any three points A, B,C ∈ X determine
a geodesic triangle △ABC obtained by connecting A, B,C by geodesic segments.
A euclidean triangle △ABC ⊂ R2 is called a (euclidean) comparison triangle for
△ABC if it has the same lengths:

d(A, B) = dR2 (A, B), d(B,C) = dR2 (B,C), d(C, A) = dR2 (C, A).

A point x ∈ AB is called a comparison point for x ∈ AB, if d(x, A) = dR2 (x, A) and
d(x, B) = dR2 (x, B).
Definition 2.3. A geodesic metric space (X, d) is called a CAT(0) space if for any
geodesic triangle △ABC in X and points x ∈ AC, y ∈ BC, if △A BC is a euclidean
comparison triangle for △ABC, and x ∈ AC and y ∈ BC are comparison points to
x ∈ AB and y ∈ BC, then d(x, y) ≤ dR2 (x, y). (See figure 2.1.)

Theorem 2.13. Suppose (X, d) is a CAT(0) complete proper geodesic space.

1. If γ is a geodesic ray s.t. γ(0) = x0 , then the following limit exists, and is a
horofunction:
Bγ (z; x0 ) = lim [d(γ(t), z) − d(γ(t), x0 )].
t→∞

2. Every horofunction arises this way.

3. If two geodesic rays γ, γ ′ s.t. γ(0) = γ ′ (0) = x0 determine the same horofunction,
then they are equal.
Thus horofunctions are represented by geodesic rays emanating from x0 .

Proof.
72 2 Ergodic theorems

Fig. 2.1 The CAT(0) inequality

Part 1. Existence of Bγ (z; x0 ).

Suppose γ is a geodesic ray s.t. γ(0) = x0 .
1. t 7→ d(γ(t), z) − d(γ(t), x0 ) is decreasing: If t < s, then

d(γ(t), z) − d(γ(t), x0 ) ≥ [d(γ(s), z) − d(γ(s), γ(t))] − d(γ(t), x0 ) (triangle ineq.)

= d(γ(s), z) − (s − t) − t = d(γ(s), z) − s
= d(γ(s), z) − d(γ(s), x0 ).

2. t 7→ d(γ(t), z) − d(γ(t), x0 ) is bounded below, by d(x0 , z).

It follows that the limit which defines Bγ (z; x0 ) exists pointwise. By the Arzela–
Ascoli theorem, this limit holds uniformly on compacts, so Bγ ∈ X. b
To see that Bγ ∈ ∂ X, we note that Bγ (γ(s); x0 ) = −s −−−→ −∞ whereas every
s→∞
function of the form Dx (·) is bounded from below. Thus Bγ ∈ Xb \ ı(X) = ∂ X.
Part 2. Every horofunction is equal to Bγ for some geodesic ray γ.
Suppose D is a horofunction, and write D = lim Dxk . We must have xk → ∞ (i.e.
k→∞
xk leaves every compact set), otherwise, since X is proper, there is a convergent
subsequence xki → x. But in this case (cf. the proof of the previous theorem) D =
lim Dxki = Dx ∈ ı(X), whereas we are assuming that D is a horofunction.
We show that the geodesic segments x0 xn converge to a geodesic ray γ s.t. D(·) =
Bγ ( · ; x0 ), and then prove that D = Bγ ( · ; x0 ).
Step 1. Let γn (t) denote the geodesic ray which starts at x0 and passes through xn
(it exists since X is complete) , then γn (t) → γ(t) uniformly on compacts in [0, ∞),
where γ(t) is geodesic ray.
2.7 The Karlsson-Margulis ergodic theorem 73

Proof. Fix ε > 0 and N so large that if k > N, then d(xk , x0 ) > t and |Dxk (z)−D(z)| <
ε for all z s.t. d(z, x0 ) ≤ t. Let yk := γk (t), the point on the geodesic segment x0 xk at
distance t from x0 . We show that {yk }k≥1 is a Cauchy sequence.
Fix m, n > N, and construct the geodesic triangle △xm yn x0 . Let △xm yn x0 be its
euclidean comparison triangle. Let ym ∈ [xm , x0 ] be the comparison point to ym on
the geodesic segment from xm to x0 . By the CAT(0) property,

d(ym , yn ) ≤ |ym yn |.

Fig. 2.2

Working in the euclidean plane, we drop a height yn z to the line connecting xm to

x0 , and mark the point w at distance 2t from ym on the line passing through xm and
x0 (figure 2.2). Let θ := ∡yn ym x0 , then

|yn ym | 1 2t
= = .
ym z cos θ |yn ym |
p
It follows that |yn ym | ≤ 2t|ym z|.
74 2 Ergodic theorems

|ym z| = |xm z| − |xm ym | ≤ |xm yn | − |xm ym | (∵ |xm z| is the hypotenuse in △xm z yn )

= d(xm , yn ) − d(xm , ym ) = d(xm , yn ) − d(xm , x0 ) + d(xm , x0 ) − d(xm , ym )
= d(xm , yn ) − d(xm , x0 ) + t
= d(xm , yn ) − d(xm , x0 ) − [d(xn , yn ) − d(xn , x0 )]
= Dxm (yn ) − Dxn (yn )
≤ sup |Dxm (y) − Dxn (y)| −−−−→ 0 by assumption.
d(y,x0 )≤t m,n→∞

This shows that {γn (t)}n≥1 = {yn }n≥1 is a Cauchy sequence. Moreover, the Cauchy
criterion holds uniformly on compact subsets of t ∈ [0, ∞).
The limit γ(t) = lim γn (t) must be a geodesic ray emanating from x0 (exercise).
Step 2. D(z) = Bγ (z; x0 ).
(
γ(tn ) n is odd
Proof. Let tn := d(x0 , xn ), and define ξn := Then Dξ2n (z) → D(z)
xn n is even.
and Dξ2n−1 (z) → Bγ (z; x0 ).
We use the fact that the geodesic segments x0 ξn converge to γ(t) to show that
|Dξn − Dξn+1 | → 0 uniformly on compacts. It will follow that D(z) = Bγ (z; x0 ).
Fix ε > 0 small and r > ρ > 0 large. Let ηk denote the point on the segment ξk x0
at distance r from x0 , then

|Dξk (z) − Dξk+1 (z)| = |d(ξk , z) − d(ξk , x0 ) − d(ξk+1 , z) + d(ξk+1 , x0 )|

= d(ξk , z) − [d(ξk , ηk ) + r] − d(ξk+1 , z) + [d(ξk+1 , ηk+1 ) + r]
= d(ξk , z) − d(ξk , ηk ) − d(ξk+1 , z) + d(ξk+1 , ηk+1 )
≤ d(ξk , z) − [d(ξk , ηk ) + d(ηk , z)]
+ [d(ηk+1 , z) + d(ξk+1 , ηk+1 )] − d(ξk+1 , z)
+ |d(ηk , z) − d(ηk+1 , z)|

The last summand tends to zero because ηk → γ(r). We show that the other two
summands are small for all z s.t. d(z, x0 ) ≤ ρ.
Let △ξ k x0 z be a euclidean comparison triangle for △ξk x0 z, and let η k ∈ ξk x0
be a comparison point to ηk ∈ ξk x0 . Let z′ be the projection of z in the line passing
through ξ k x0 (figure 2.3).
By the CAT(0) property, d(ηk , z) ≤ d(η k , z), and so

[d(ξk , ηk ) + d(ηk , z)] − d(ξk , z) ≤ [d(ξ k , η k ) + d(η k , z)] − d(ξ k , z)

= [d(ξ k , z′ ) − d(ξ k , z)] + [d(η k , z) − d(η k , z′ )]

We now appeal to the following simple consequence of the Pythagorean Theorem:

In a triangle △ABC s.t. ∡ABC = 90◦ , if |AB| > r and |BC| ≤ ρ, then 0 ≤ |AC| −
|AB| < ρ 2 /r. Applying this to △ξ k z′ z and △η k z′ z, we see that
2.7 The Karlsson-Margulis ergodic theorem 75

Fig. 2.3

d(ξk , ηk ) + d(ηk , z) − d(ξk , z) ≤ 2ρ 2 /r.

Similarly, one shows that d(ξk+1 , ηk+1 )+d(ηk+1 , z)−d(ξk+1 , z) ≤ 2ρ 2 /r. Choosing
r > 2ρ 2 /ε sufficiently large, and k large enough so that d(xk , x0 ) > r we see that the
two remaining terms are less than ε, with the result that |Dξk+1 (z) − Dξk (z)| < 3ε for
all z s.t. d(z, x0 ) < ρ.
It follows that Bγ (z; x0 ) = lim Dξ2n−1 (z) = lim Dξ2n (z) = D(z).
Part 3. If Bγ1 ( · ; x0 ) = Bγ2 ( · ; x0 ), then γ1 = γ2 .
(
γ1 (tn ) n is odd
Proof. Fix tn ↑ ∞, and set xn := . The sequences Dx2n (·), Dx2n−1 (·)
γ2 (tn ) n is even
have the same limit, Bγ1 ( · ; x0 ) = Bγ2 ( · ; x0 ), therefore lim Dxn exists. Step 1 in Part
2 shows that the geodesic segments x0 xn must converge uniformly to a geodesic ray
γ(t). But these geodesic segments lie on γ1 for n odd and on γ2 for n even; it follows
that γ1 = γ2 . ⊔
⊓
Proposition 2.6. Let (X, d) be a proper geodesic space with the CAT(0) property,
and suppose xn ∈ X tend to infinity at speed s, i.e. n1 d(x0 , xn ) −−−→ s. If
n→∞

1
D(xn ) −−−→ −s for D ∈ ∂ X,
n n→∞

then d(xn , γ(sn)) = o(n), where γ is the geodesic ray s.t. D = Bγ .

Proof. Fix an origin x0 ∈ X, and suppose γ is the geodesic ray so that
76 2 Ergodic theorems

D(z) = Bγ (z; x0 ) := lim [d(γ(t), z) − d(γ(t), x0 )].

t→∞

Fix n and t large, and consider the geodesic triangle △xn x0 γ(st) and the point
γ(sn) on the segment from x0 to γ(st). Let △xn x0 γ(st) be the euclidean comparison
triangle, and let γ(sn) be the comparison point to γ(st). By the CAT(0) property,

d(xn , γ(sn)) ≤ d(xn , γ(sn)).

Let wt be the point on the segment connecting γ(st) to x0 at the same distance from
γ(st) as xn . Drop a height xn z to that segment (figure 2.4).

Fig. 2.4

1. • |wt γ(sn)| = |γ(st) wt | + |γ(sn) x0 | − |γ(st) x0 | = |d(γ(st), xn ) + sn − st| −−→

t→∞
D(xn ) + sn. By assumption, D(xn )/n → −s, so ∃N0 so that for all n ≥ N0 , for
every t > n, |wt γ(sn)| = o(n).
• For fixed n, it is easy to see that α(t) := ∡xn γ(st)z −−→ 0. It follows that
t→∞
|z wt | = |xn z| tan α(t)
2 −−→ 0. So ∃T (n) s.t. for all t > T (n), |z wt | = o(n).
t→∞
We see that for all n > N0 and t > T (n), |z x0 | = sn + o(n).
2. By assumption |xn x0 | = d(xn , x0 ) = sn + o(n), so if t > T (n) then

|xn z|2 = |xn x0 |2 − |x0 z|2 = [sn + o(n)]2 − [sn + o(n)]2 = o(n2 ),

whence |xn z| = o(n).

2.7 The Karlsson-Margulis ergodic theorem 77

3. |zγ(sn)| = |z x0 | − |γ(sn)x0 | = [sn + o(n)] − sn = o(n).

It follows from the above that if t > T (n), then
q q
d(xn , γ(sn)) ≤ |xn γ(sn)| = |xn z|2 + |zγ(sn)|2 = o(n)2 + o(n)2 = o(n). ⊔
⊓

Corollary 2.4. ∂ X = {D ∈ Xb : D is not bounded below}.

Proof. Dx (z) = d(z, x0 ) − d(x0 , x) ≥ −d(x, x0 ), but Bγ (γ(s); x0 ) = −s → −∞. ⊔
⊓

2.7.2 An ergodic theorem for isometric actions on CAT(0) spaces

Throughout this section, (X, d) is a metric space which is proper, geodesic, geodesi-
cally complete, and which has the CAT(0) property. We fix once and for all some
point x0 ∈ X (“the origin”).
A map ϕ : X → X is called an isometry, if it is invertible, and d(ϕ(x), ϕ(y)) =
d(x, y) for all x, y ∈ X. The collection of isometries is a group, which we will denote
by Isom(X).
Suppose (Ω , B, µ, T ) is a ppt and f : Ω → Isom(X) is measurable, in the sense
that (ω, x) 7→ f (ω)(x) is a measurable map Ω × X → X. We study the behavior of
fn (ω)x0 as n → ∞, where

fn (ω) := f (ω) ◦ f (T ω) ◦ · · · ◦ f (T n−1 ω).

The subadditive theorem implies that { fn (ω)x0 }n≥1 has “asymptotic speed”:

1
s(ω) := lim d(x0 , fn (ω)x0 ).
n→∞ n
We are interested in the existence of an “asymptotic direction.”
Theorem 2.14 (Karlsson–Margulis). Suppose (Ω , B, µ, T ) is Ra ppt on a standard
probability space, and f : Ω → Isom(X) is a measurable map. If Ω d(x0 , f (ω)x0 )dµ
is finite, then for a.e. ω ∈ Ω there exists a geodesic ray γω (t) emanating from x0 s.t.

1
d( f (ω) f (T ω) · · · f (T n−1 ω)x0 , γω (ns(ω)) −−−→ 0.
n n→∞

Proof (Karlsson and Ledrappier). Some reductions: w.l.o.g. Ω is a compact metric

space and B is the Borel σ -algebra of Ω (cf. appendix A). W.l.o.g. µ is ergodic
(otherwise work with its ergodic components). In the ergodic case, s(ω) = s a.e.
where s is a constant. W.l.o.g. s > 0, otherwise the theorem holds trivially.
By Proposition 2.6, to prove the theorem it is enough to find a horofunction Dω (·)
with the property that
1
lim Dω ( fn (ω)x0 ) = −s(ω). (2.12)
n→∞ n
78 2 Ergodic theorems

The trick is to write the expression D( fn (ω)x0 ) as an ergodic sum for some other
dynamical system, and then apply the pointwise ergodic theorem.
We first extend the action of an isometry ϕ on X to an action on X. b Recall that
every D ∈ Xb equals lim Dxn (·) where Dxn (·) = d(xn , ·) − d(xn , x0 ). Define
n→∞

ϕ(D)(z) := lim Dϕ(xn ) (z)

n→∞
= lim d(ϕ(xn ), z) − d(ϕ(xn ), x0 ) = lim d(xn , ϕ −1 (z)) − d(xn , ϕ −1 (x0 ))
n→∞ n→∞
= lim d(xn , ϕ −1 (z)) − d(xn , x0 ) + d(xn , x0 ) − d(xn , ϕ −1 (x0 ))
n→∞
= D(ϕ −1 (z)) − D(ϕ −1 (x0 )).

In particular, the definition is independent of the choice of Dxn , so the definition is

proper. The identity ϕ(D) := lim Dϕ(xn ) shows that ϕ(D) ∈ X. b We write

(ϕ · D)(z) := ϕ(D)(z) := D(ϕ −1 (z)) − D(ϕ −1 (x0 )).

This action preserves ∂ X, because by Corollary 2.4, D ∈ ∂ X iff infX D = −∞ iff

infX ϕ · D = inf D − D(ϕ −1 (x0 )) = −∞.
Define a map S : Ω × Xb → Ω × Xb by

S : (ω, D) 7→ (T (ω), f (ω)−1 · D).

Notice that the second coordinate of the iterate Sk (ω, D) = (T k (ω), fk (ω)−1 · D) is
the horofunction D( fk (ω)z) − D( fk (ω)x0 ). Define F : Ω × Xb → R by

F(ω, D) := D( f (ω)x0 ),

then (F ◦ Sk )(ω, D) = F(T k (ω), fk (ω)−1 · D) = ( fk (ω)−1 · D)( f (T k ω)x0 )

= D( fk+1 (ω)x0 ) − D( fk (ω)x0 ). Summing over k, we see that
n−1
∑ (F ◦ Sk )(ω, D) = D( fn (ω)x0 ).
k=0
2.7 The Karlsson-Margulis ergodic theorem 79

b0 on Ω × Xb such that
Technical step: To construct a probability measure µ
1. µ
b0 is S–invariant, S–ergodic, and µb0 (Ω × ∂ X) = 1;
2. µ b0 (E × X)
b0 projects to µ in the sense that µ b = µ(E) for all E ⊂ Ω measurable;
3. F ∈ L1 (µ
R
b0 ) and Fd µ b0 = −s.
We give the details later. First we explain how to use such µ
b0 to prove the theo-
rem. By the pointwise ergodic theorem,
 
1
Z
U := (ω, D) ∈ Ω × ∂ X : D( fn (ω)x0 ) −−−→ Fd µ0 = −s b
n n→∞

b0 –measure. Let E := {ω ∈ Ω : ∃D ∈ ∂ X s.t. (ω, D) ∈ U}, then E × ∂ X ⊇

has full µ
U, whence µ b0 (E × ∂ X) = 1. Since µ
b0 projects to µ, µ(E) = 1. So for a.e. ω ∈
Ω ,∃Dω ∈ ∂ X with property (2.12). The theorem now follows from Proposition 2.14.
We now explain how to carry out the technical step. This requires some functional-
theoretic machinery which we now develop.
Let C(X)
b denote the space of continuous functions on Xb equipped with the supre-
mum norm. Since Xb is compact and metrizable, C(X) b is a separable Banach space.
By the Riesz representation theorem, the dual space C(X) b ∗ can be identified with
the space of signed measures on X with finite total variation.
b
A function ϕ : Ω → C(X) b will be called measurable, if (ω, D) 7→ ϕ(ω)(D) is
measurable with respect to the Borel structures of C(X) b and Ω × X. b It is easy to
check using the separability of X that in this case ω 7→ ∥ϕ(ω)∥ is Borel.
b
Let
R
L 1 (Ω ,C(X))
b denote the space of measurable functions ϕ : Ω → C(X)
R
b such
that ∥ϕ(ω)∥dµ < ∞. Identifying ϕ, ψ s.t. ∥ϕ − ψ∥dµ = 0, we obtain a linear
vector space
L1 (Ω ,C(X))
b

of equivalence classes of measurable functions ϕ : Ω → C(X)

b with norm ∥ϕ∥1 :=
∥ϕ(ω)∥dµ < ∞. Here are some basic facts on L1 (Ω ,C(X))
R
b [1, chapter 1]:
1. L1 (Ω ,C(X))
b is a Banach space. We leave this as an exercise.
1 b is separable: Every ϕ ∈ L1 (Ω ,C(X))
2. L (Ω ,C(X)) b can be approximated in norm
n
by a function of the form ∑i=1 ϕi (D)1Ei (ω) where ϕi belong to a countable dense
b and Ei belongs to a countable subset of F .
subset of C(X)
b ∗ is sequentially compact with respect to the weak-
3. The unit ball in L1 (Ω ,C(X))
1 b ∗ . This follows from the Banach-Alaoglu Theorem.
star topology on L (Ω ,C(X))
Notice that F(ω, D) := D( f (ω)x0 ) belongs to L 1 (Ω ,C(X)): b (a) For fixed ω,
F(ω, ·) is continuous on X,b because if Dk → D in Xb then Dk → D pointwise (uni-
formly on compacts); (b) F(ω, D) is measurable, because it is not difficult to see that
F(ω, D) is a limit ofRsimple functions; (c) F ∈ L1 , because |D(z)| = |D(z)−D(x0 )| ≤
d(z, x0 ), so ∥F∥1 ≤ Ω d( f (ω)x0 , x0 )dµ(ω) < ∞.
We are now ready to construct µ
b0 .
R
We begin by noting that (−F)d νb0 ≤ s for allRS-ergodic invariant measures νb
b0 we seek maximizes (−F)d νb0 . To see this note that
which project to µ, so the µ
80 2 Ergodic theorems

for every x, z ∈ X, Dx (z) = d(z, x) − d(x0 , x) ≥ [d(x0 , x) − d(x0 , z)] − d(x, x0 ), so

Dx (z) ≥ −d(x0 , z) with equality iff z = x. (2.13)

Passing to limits we find that D(z) ≥ −d(x0 , z) for all D ∈ X.

b In particular,

1 n−1 1
Z Z Z
(−F)d νb = − ∑ (F ◦ Sk )d νb = − n
n k=0
D( fn (ω) · x0 )d νb(ω, D)

1
Z
≤ d(x0 , fn (ω) · x0 )dµ, by (2.13) and the assumption that νb projects to µ.
n
! !
As this holds for all n, (−F)d νb ≤ inf 1n d(x0 , fn (ω) · x0 )dµ = s. Equality = is due
R R

to the subadditive ergodic theorem for the cocycle d(x0 , fn (ω) · x0 ).

We see that the measure we seek maximizes the expectation value of (−F), or
equivalently, of − n1 ∑n−1 k
k=0 F ◦ S .
n−1
Recall that − n1 ∑k=0 F ◦ Sk ≡ − n1 D( fn (ω)x0 ). By (2.13), this expression is max-
imized at D = D fn (ω)x0 . Let
Z
bn :=
η δ(ω,D fn (ω)x ) dµ(ω), where δ(ω,D) :=point mass at (ω, D).
Ω 0

We have − n ∑n−1 1R k b 1 1
R R
k=0 F ◦S d ηn = − n D fn (ω)x0 ( f n (ω)x0 )dµ = n d(x0 , f n (ω)x0 )dµ.
So − 1n ∑n−1 k b
R
k=0 F ◦ S d ηn ≥ s. It is easy to see that ηn projects to µ.
b
But ηbn is not S–invariant. Let µ bn := 1n ∑n−1
k=0 η
b n ◦ S−k . This measure still projects
n−1
bn = − n1 ∑k=0 F ◦ Sk d η bn ◦ S−1 − µ
R R
to µ (check!); − Fd µ bn ≥ s; and |µ b | ≤ 2/n (total
variation norm).
We may think of µ bn as of positive bounded linear functional on L1 (Ω , X) b with
norm 1. By the Banach-Alaoglu Theorem, there is a subsequence µnk which con-
verges weak-star in L1 (Ω , X)b ∗ to some positive µ b ∈ L1 (Ω , X)b ∗ . This functional re-
∗
stricts to a functional in C(Ω × X) , and therefore can be viewed as a positive finite
b
measure on Ω × X. b Abusing notation, we call this measure µ b.
• µ
b is a probability measure, because µ b (1) = lim µ bnk (1) = 1
b projects to µ, because for every E ∈ F , gE (ω)(D) := 1E (ω) belongs to
• µ
L 1 (Ω ,C(X)),
b whence µ b (E) = µb (gE ) = lim µbnk (ϕE ) = lim µ
bnk (E) = µ(E).
b is S-invariant, because if g ∈ C(Ω × X),
• µ b then g, g ◦ S ∈ L 1 (Ω ,C(X)),b whence
|µ
b (g) − µ
b (g ◦ S)| = lim |µ
bnk (g) − µ
bnk (g ◦ S)| = 0.
b (−F) = s: The inequality ≤ is because of the S-invariance of µ
• µ b . The inequality
≥ is because F ∈ L 1 (Ω ,C(X)), b whence µ b (−F) = lim µ bnk (−F) ≥ s.
(These arguments require weak∗ convergence in L1 (Ω ,C(X)) b ∗ , not just weak∗ con-
vergence of measures, because in general gE , F, g ◦ S ̸∈ C(Ω × X).)
b
We found an S-invariant probability measure µ b which projects to µ and so that
b (−F) = s. But µ
µ b is not necessarily ergodic.
R
b = Ω ×∂ X µ
Let µ b be its ergodic decomposition. Then: (a) almost every
b(ω,D) d µ
ergodic component is ergodic and invariant; (b) Almost every ergodic component
2.7 The Karlsson-Margulis ergodic theorem 81

projects to µ (prove using the extremality µ); (c) Almost every ergodic component
gives (−F) integral s, because µb(ω,D) (−F) ≤ s for a.e. (ω, D) by S-invariance, and
the inequality cannot be strict on a set with positive measure as this would imply
b (−F) < s.
that µ
SoR
a.e. ergodic component µb0 of µ
b is ergodic, invariant, projects to µ, and satis-
fied Fd µ b0 = −s. It remains to check that µ
b0 is carried by Ω × ∂ X. This is because
b -a.e. (ω, D)
by the ergodic theorem, for µ

1 1 n−1
D( fn (ω)x0 ) ≡ − ∑ (F ◦ Sk )(ω, D) −−−→ −s < 0,
n n k=0 n→∞

whence infX D = −∞. By corollary 2.4, D ∈ ∂ X. The construction of µ

b0 is complete.
As explained above, the theorem follows. ⊔
⊓

2.7.3 A geometric proof of the multiplicative ergodic theorem

We explain how to obtain the multiplicative ergodic theorem as a special case of the
Karlsson-Margulis ergodic theorem. We begin with some notation and terminology.
• log := ln, log+ t = max{logt, 0} p
• Vectors in Rd are denoted by v, w etc. ⟨v, w⟩ = ∑ vi wi and ∥v∥ = ⟨v, v⟩.
• The space of d × d matrices with real entries will be denote by Md (R).
• diag(λ1 , . . . , λd ) := (ai j ) ∈ Md (R) where aii := λi and ai j = 0 for i ̸= j
• I :=identity matrix= diag(1, . . . , 1)
• For a matrix A = (ai j ) ∈ Md (R), Tr(A) := ∑ aii (the trace), and Tr1/2 (A) :=
p
Tr(A). At := (a ji ) (the transpose). ∥A∥ := sup∥v∥≤1 ∥Av∥.
∞
exp(A) := I + ∑ Ak /k!
k=1

• GL(d, R) := {A ∈ Md (R) : A is invertible}

• Sym(d, R) := {A ∈ Md (R) : A is symmetric, i.e. At = A}
• Od (R) := {A ∈ Md (R) : A is orthogonal, i.e. At A = i}
• Posd (R) := {A ∈ Symd (R) : A positive definite, i.e. ∀v ∈ Rd \ {0}, ⟨Av, v⟩ ≩ 0}.
• Orthogonal diagonalization: P ∈ Posd (R) ⇔ ∃O ∈ Od (R) such that OPOt =
diag(λ1 , . . . , λd ) and λi > 0 for all i.
• Positive definite matrices have roots: ∀P ∈ Posd (R), ∀m ∈ N ∃Q ∈ Posd (R) s.t.
1/m 1/m
P = Qm . Take Q := Ot diag(λ1 , . . . , λd )O. We write Q = P1/m
• Positive definite matrices have logarithms: ∀P ∈ Posd (R), ∃! S ∈ Symd (R) s.t.
P = exp(S). Take S := Ot diag(log λ1 , . . . , log λd )O. We write S := log P

Proposition 2.7. GL(d, R) acts transitively on Posd (R) by A · P = APAt , i.e.:

1. ∀A ∈ GL(d, R), ∀P ∈ Posd (R), A · P ∈ Posd (R)
82 2 Ergodic theorems

2. ∀A, B ∈ GL(d, R), ∀P ∈ Posd (R), (AB) · P = A · (B · P)

3. ∀P, Q ∈ Posd (R), ∃A ∈ GL(d, R) s.t. A · P = Q

Proof. Suppose A ∈ GL(d, R), P ∈ Posd (R). Clearly APAt is symmetric. It is pos-
itive definite, because for every v ̸= 0, ⟨APAt v, v⟩ = ⟨PAt v, At v⟩ > 0 by the positive
definiteness of P and the invertibility of At . This proves (1). (2) is trivial. (3) is
because A · P = Q for A := Q1/2 (P−1 )1/2 . Here we use the elementary fact that
P ∈ Posd (R) ⇒ P is invertible and P−1 ∈ Posd (R). ⊔
⊓

Theorem 2.15. There exists a unique metric d on Posd (R) as follows:

1. For all A ∈ GL(d, R), P 7→ A · P is an isometry of Posd (R) i.e.

d(APAt , AQAt ) = d(P, Q) (P, Q ∈ Posd (R))

2. If S is symmetric and Tr(S2 ) = 1, then γ(t) := exp(tS) is a geodesic ray. Any

geodesic which starts from the identity matrix has this form.
This metric turns Posd (R) into a proper CAT (0) geodesic space.

For details and proof, see Bridson & Haefliger: Metric spaces of non-positive cur-
vature, Springer 1999, chapter II.10.
q
Lemma 2.3. For every A ∈ GL(d, R), d(I, A · I) = 2 ∑di=1 (log λi )2 , where λi are
√
singular values of A (the eigenvalues of AAt ).

Proof. The idea is to find the geodesic from I to A · I. Such geodesics take the form
γ(τ) = exp(τS) for S ∈ Symd (R) s.t. Tr(S2 ) = 1. To find S we write A · I = AAt =
exp(log AAt ). This leads naturally to the solution

S := t −1 (log AAt ), and t := Tr1/2 [log AAt ]2 .

It follows that A · I = exp(tS) whence d(I, A · I) = t = Tr1/2 [log AAt ]2 .

To calculate the trace we use the well-known singular value decomposition of A:
A = O1 diag(λ1 , . . . , λd )Ot2 for some O1 , O2 ∈ Od (R). So

(log AAt )2 = O1 diag(4(log λ1 )2 , . . . , 4(log λd )2 )Ot1 ,

q
whence d(I, A · I) = t = Tr1/2 [log AAt ]2 = 2 ∑di=1 (log λi )2 . ⊔
⊓

Corollary 2.5. For every A ∈ GL(d, R) we have the estimate

√
max{| log ∥A∥|, | log ∥A−1 ∥|} ≤ d(I, A · I) ≤ 2 d max{| log ∥A∥|, | log ∥A−1 ∥|}
√
Proof. Let λ1 , . . . , λd denote the singular values of A (the eigenvalues of At A), and
let λmax := max{λ1 , . . . , λd } , λmin := min{λ1 , . . . , λd }.
2.7 The Karlsson-Margulis ergodic theorem 83

It is well-known that ∥A∥ = λmax , ∥A−1 ∥ = 1/λmin . 6 The corollary follows from
the lemma 2.3 and the identity max{(log λi )2 } = 2 max{| log λmax |, | log λmin |}. ⊔
⊓

Lemma 2.4. Suppose Pn , P ∈ Posd (R). If d(Pn , P) −−−→ 0 where d is the metric in
n→∞
Theorem 2.15, then ∥Pn − P∥ −−−→ 0.
n→∞

Proof. For every Q ∈ Posd (R), Q = Q1/2 · I. So d(Pn , P) → 0 implies that

1/2 1/2
δn := d(I, P−1/2 Pn P−1/2 ) = d(I, P−1/2 Pn ·I) = d(P1/2 ·I, Pn ·I) = d(P, Pn ) −−−→ 0.
n→∞

Let γn (t) denote the geodesic ray from I to P−1/2 Pn P−1/2 , then γn (t) = exp(tSn ) for
some symmetric matrix Sn such that Tr[Sn2 ] = 1. This gives us the identity

P−1/2 Pn P−1/2 = exp[δn Sn ], with δn → 0, Sn ∈ Symd (R), Tr[S2 ] = 1.

We claim that exp[δn Sn ] → I in norm. To see this we first diagonalize Sn orthogo-

nally: Sn = On diag(λ1 (n), . . . , λd (n))Otn where On ∈ Od (R), ∑di=1 λi (n)2 = Tr[Sn2 ] =
1. Using the identities exp diag(λ1 , . . . , λd ) = diag(eλ1 , . . . , eλd ) and exp(OAOt )
= O(exp A)Ot for O ∈ Od (R), we find that

P−1/2 Pn P−1/2 = exp[δn Sn ] = On diag(eδn λ1 (n) , . . . , eδn λd (n) )Otn .

Since ∑ λi (n)2 = 1, the middle term tends to I in norm, whence by the orthogonality
of On , ∥P−1/2 Pn P−1/2 − I∥ → 0. It follows that ∥Pn − P∥ → 0. ⊔
⊓

Multiplicative Ergodic Theorem: Let (Ω , B, µ, T ) be a ppt on a standard proba-

bility space, and A : Ω → GL(d, R) a measurable function s.t. log ∥A∥, log ∥A−1 ∥
are absolutely integrable. If An (ω) := A(ω)A(T ω) · · · A(T n−1 ω) then the limit
Λ := lim (An Atn )1/2n exists a.e., and n1 log ∥Λ −n An ∥, n1 log ∥A−1 n −−
n Λ ∥− → 0 a.e.
n→∞ n→∞

Proof. The idea is to apply the Karlsson-Margulis ergodic theorem to the space
X := Posd (R), the origin x0 := I, and the isometric action A(ω) · P := A(ω)PA(ω)t .
First we need to check the integrability condition:
R
Step 1. d(x0 , A(ω) · x0 )dµ < ∞.
This follows from by corollary 2.5.
Step 2. For a.e. ω the following limit exists: s(ω) := lim 1n d(I, An (ω) · I).
n→∞

Proof. It is easy to see using the triangle inequality and the fact that GL(d, R) acts
isometrically on Posd (R) that g(n) (ω) := d(I, An (ω) · I) is a sub-additive cocycle.
The step follows from the sub-additive ergodic theorem.
Step 3. Construction of Λ (ω) s.t. 1n log ∥An (ω)−1Λ (ω)n ∥ −−−→ 0 a.e.
n→∞

6This can be easily deduced from the singular decomposition: A = O1 diag(λ1 , . . . , λd )Ot2 where
O1 , O2 ∈ Od (R).
84 2 Ergodic theorems

The Karlsson-Margulis ergodic theorem provides, for a.e. ω, a geodesic ray γω (t)
emanating from I such that

d(An (ω) · I, γω (ns(ω)) = o(n).

By theorem 2.15, γω (t) = exp(tS(ω)), for S(ω) ∈ Symd (R) s.t. Tr[S2 ] = 1. Let
 
1
Λ (ω) := exp s(ω)S(ω)
2

Then Λ ∈ Posd (R), d(An · I,Λ 2n ) = o(n), and by Corollary 2.5

max{| log ∥A−1 n −n −1 n n

n Λ ∥|, | log ∥Λ An ∥|} ≍ d(I, An Λ · I) = d(An · I,Λ · I)
= d(An · I,Λ 2n ) ≡ d(An · I, γω (s(ω)n)) = o(n)

Step 4. For a.e. ω, (An (ω)An (ω)t )1/2n −−−→ Λ (ω).

n→∞

Proof. Fix ω such that d(An · I,Λ 2n )

= o(n). Given n, consider the geodesic tri-
angle with vertices I, An · I, Λ 2n = γω (ns).
• The geodesic connecting I, An · I equals γ1 (t) := exp[cnt log An Atn ], 0 ≤ t ≤
d(I, An Atn ), where cn := 1/d(I, An Atn ). It contains the point

1 1
A′n := (An Atn )1/2n = exp[ log An Atn ] = γ1 ( )
2n 2ncn

So d(I, A′n ) = d(γ1 (0), γ1 ( 2nc

1
n
1
)) = 2ncn
1
= 2n 1
d(I, An Atn ) = 2n d(I, An · I).
2n
• The geodesic connecting I,Λ is γω (t) = exp(tS), 0 ≤ t ≤ 2n. It contains the
1
point Λ = γω (s/2). Again, d(I,Λ 2 ) = s/2 = 2n d(I,Λ 2n ).
2n
• The geodesic connecting An · I,Λ has length d(An · I, γω (ns)) = o(n).
We now consider the Euclidean comparison triangle I, An · I,Λ 2n . The Euclidean
triangle generated by I, A′n , Λ is similar to the euclidean triangle I, A′n , Λ , and has
1
sides 2n × the lengths of the sides of that triangle. So

1 1
dR2 (A′n ,Λ ) = dR2 (An · I,Λ 2n ) ≡ d(An · I, γω (ns)) = o(1).
2n 2n

By the CAT(0) property d(A′n ,Λ ) ≤ dR2 (A′n ,Λ ) = o(1).

Equivalently, d((An Atn )1/2n ,Λ ) → 0. By Lemma 2.4, ∥(An Atn )1/2n −Λ ∥ → 0. ⊔
⊓

Problems

2.1. The Mean Ergodic Theorem for Contractions

Suppose H is a Hilbert space, and U : H → H is a bounded linear operator such that
2.7 The Karlsson-Margulis ergodic theorem 85

∥U∥ ≤ 1. Prove that N1 ∑N−1 k

k=0 U f converges in norm for all f ∈ H, and the limit is
the projection of f on the space { f : U f = f }.

2.2. Ergodicity as a mixing property

Prove that a ppt (X, B, µ, T ) is ergodic, iff for every A, B ∈ B, 1
N ∑N−1
k=0 µ(A ∩
T −k B) −−−→ µ(A)µ(B).
N→∞

2.3. Use the pointwise ergodic theorem to show that any two different ergodic in-
variant probability measures for the same transformation are mutually singular.

2.4. Ergodicity and extremality

An invariant probability measure µ is called extremal, if it cannot be written in the
form µ = tµ1 + (1 − t)µ2 where µ1 , µ2 are different invariant probability measures,
and 0 < t < 1. Prove that an invariant measure is extremal iff it is ergodic, using the
following steps.
1. Show that if E is a T –invariant set of non-zero measure, then µ( · |E) is a T –
invariant measure. Deduce that if µ is not ergodic, then it is not extremal.
2. Show that if µ is ergodic, and µ = tµ1 + (1 − t)µ2 where µi are invariant, and
0 < t < 1, then
a. For every E ∈ B, 1
N ∑N−1 k
k=0 1E ◦ T −−−→ µ(E) µi –a.e. (i = 1, 2).
N→∞
b. Conclude that µi (E) = µ(E) for all E ∈ B (i = 1, 2). This shows that ergod-
icity implies extremality.

2.5. Prove that the Bernoulli ( 21 , 12 )–measure is the invariant probability measure for
the adding machine (Problem 1.10), by showing that all cylinders of length n must
have the same measure as [0n ]. Deduce from the previous problem that the adic
machine is ergodic.

2.6. Explain why when f ∈ L2 , E( f |F ) is the projection of f on L2 (F ). Prove:

1. If F = {∅, A, X \ A}, then E(1B |F ) = µ(B|A)1A + µ(B|Ac )1Ac
2. If F = {∅, X}, then E( f |F ) = f dµ
R

3. If X = [−1, 1] with Lebesgue measure, and F = {A : A is Borel and − A = A},

then E( f |F ) = 12 [ f (x) + f (−x)]

2.7. Prove:
1. f :7→ E( f |F ) is linear, and a contraction in the L1 –metric
2. f ≥ 0 ⇒ E( f |F ) ≥ 0 a.e.
3. if ϕ is convex, then E(ϕ ◦ f |F ) ≤ ϕ(E( f |F ))
4. if h is F –measurable, then E(h f |F ) = hE( f |F )
5. If F1 ⊂ F2 , then E[E( f |F2 )|F1 ] = E( f |F1 )

2.8. A Zd –PET along more general sequences of boxes (Tempelman). A family

of boxes {Ir }r≥1 is called regular (with constant C) if there exists an increasing
sequence of boxes {Ir′ }r′ ≥1 which tends to Zd+ s.t. |Ir | ≤ C|Ir′ | for all r
86 2 Ergodic theorems

1. Show that the following sequences are regular:

a. Any increasing sequence of boxes which tends to Zd+
b. Ir := [r, r + r2 ] (in dimension one)
c. Ir := [0, n(r)) where n(k) ∈ Zd+ is a sequence of vectors which tends to infinity
“in a sector” in the sense that (a) min{n1 (k), . . . , nd (k)} −−−→ ∞, and (b) for
  k→∞
some constant K max1≤i, j≤d nnij(k)(k) ≤ K for all k.

2. Suppose T1 , . . . , Td are commuting measure preserving maps on a probability

space (Ω , F , µ), and {Ir }r≥1 is a regular sequence of boxes with constant C.
a. Prove the following maximal inequality: If ϕ ∈ L1 is non-negative, then for
all α > 0, µ[supr |I1r | SIr ϕ > α] < C2d ∥ϕ∥1 /α.
1
b. Deduce that if f ∈ L1 , then |Ir | SIr f −−−→ E( f |Inv(T1 ) ∩ · · · ∩ Inv(Td )) a.e.
r→∞

2.9. The Martingale Convergence Theorem (Doob)

Suppose (X, B, µ) is a probability space, and F1 ⊂ F2 ⊂ · · · are σ –algebras all of
which are contained in B. Let F := σ ( n≥1 Fn ) (the smallest σ –algebra contain-
S

ing the union). If f ∈ L1 , then E( f |Fn ) −−−→ E( f |F ) a.e. and in L1

n→∞

Prove this theorem, using the following steps (W. Parry). It is enough to consider
non-negative f ∈ L1 .
L1
1. Prove that E( f |Fn ) −−−→ E( f |F ) using the following observations:
n→∞
S 1 (X, F
a. The convergence holds for all elements of n≥1 L n , µ);
b. n≥1 L1 (X, Fn , µ) is dense in L1 (X, F , µ).
S

2. Set Ea := {x : max1≤n≤N E( f |Fn )(x) > a}. Show that µ(Ea ) ≤ 1a f dµ. (Hint:
R

E = n≥1 {x : E( f |Fn )(x) > λ , and E( f |Fk )(x) ≤ λ for k = 1, . . . , n − 1}.)

U

3. Prove that E( f |Fn ) −−−→ E( f |F ) a.e. for every non-negative f ∈ L1 , using the
n→∞
following steps. Fix f ∈ L1 . For every ε > 0, choose n0 and g ∈ L1 (X, Fn0 , µ)
such that ∥E( f |F ) − g∥1 < ε.
a. Show that |E( f |Fn ) − E( f |F )| ≤ E(| f − g| |Fn )| + |E( f |F ) − g| for all n ≥
n0 . Deduce that
√ 1√
   
µ lim sup |E( f |Fn ) − E( f |F )| > ε ≤ µ sup |E(| f − g||Fn )| > ε
n→∞ n 2
1√
 
+ µ |E( f |F ) − g| > ε
2
 √ 
b. Show that µ lim supn→∞ |E( f |Fn ) − E( f |F )| > ε −−−→ 0. (Hint: Prove
ε→0+
first that for every L1 function F, µ[|F| > a] ≤ 1a ∥F∥1 .)
c. Finish the proof.
2.7 The Karlsson-Margulis ergodic theorem 87

2.10. Hopf’s ratio ergodic theorem

Let (X, B, µ, T ) be a conservative ergodic mpt on a σ –finite measure space. If f , g ∈
∑n−1
k=0 f ◦T
k R
f dµ
L1 and g ̸= 0, then
R
−−−→ R gdµ almost everywhere.
∑n−1
k=0 g◦T k n→∞

Prove this theorem using the following steps (R. Zweimüller). Fix a set A ∈ B s.t.
0 < µ(A) < ∞, and let (A, BA , TA , µA ) denote the induced system on A (problem
1.14). For every function F, set

Sn F := F + F ◦ T + · · · + F ◦ T n−1
SnA F := F + F ◦ TA + · · · + F ◦ TAn−1

1. Read problem 1.14, and show that a.e. x has an orbit which enters A infinitely
many times. Let 0 < τ1 (x) < τ2 (x) < · · · be the times when T τ (x) ∈ A.
2. Suppose f ≥ 0. Prove that for every n ∈ (τk−1 (x), τk (x)] and a.e. x ∈ A,
A f )(x)
(Sk−1 (Sn f )(x) (SkA f )(x)
≤ ≤ .
(SkA 1A )(x) A 1 )(x)
(Sn 1A )(x) (Sk−1 A

1
3. Verify that SAj 1A = j a.e. on A, and show that (Sn f )(x)/(Sn 1A )(x) −−−→
R
µ(A) f dµ
n→∞
a.e. on A.
4. Finish the proof.

Notes for chapter 2

For a comprehensive reference to ergodic theorems, see [7]. The mean ergodic the-
orem was proved by von Neumann, and the pointwise ergodic theorem was proved
by Birkhoff. By now there are many proofs of Birkhoff’s theorem; the one we use
is taken from [6], where it is attributed to Kamae — who found it using ideas from
nonstandard analysis. The subadditive ergodic theorem was first proved by King-
man. The proof we give is due to Steele [11]. The proof of Tempelman’s pointwise
ergodic theorem for Zd is taken from [7]. For ergodic theorems for the actions of
other amenable groups see [8]. The multiplicative ergodic theorem is due to Os-
eledets. The “linear algebra” proof is due to Raghunathan and Ruelle, and is taken
from [10]. The geometric approach to the multiplicative ergodic theorem is due to
Kaimanovich [3] who used it to generalize that theorem to homogeneous spaces
other than Posd (R). The ergodic theorem for isometric actions on CAT(0) spaces
is due to Karlsson & Margulis [5]. We proof we give is due to Karlsson & Ledrap-
pier [4]. The Martingale convergence theorem (problem 2.9) is due to Doob. The
proof sketched in problem 2.9 is taken from [2]. The proof of Hopf’s ratio ergodic
theorem sketched in problem 2.10 is due to R. Zweimüller [12].
88 2 Ergodic theorems

References

1. Arnold, L.: Random dynamical systems, Springer, 2003.

2. Bridson, M.R. & Haefliger, A.: Metric spaces of non-positive curvature, Springer, 1999
3. Kaimanovich, V.A.: Lyapunov exponents, symmetric spaces, and multiplicative ergodic theo-
rems for semisimple Lie groups. J. Soviet Math. 47 (1989), 2387–2398.
4. Karlsson, A. and Ledrappier, F.: On laws of large numbers for random walks. Ann. Probab.
34 (2006) no. 5, 1693–1706.
5. Karlsson, A. and Margulis, G. A.: A multiplicative ergodic theorem and non-positively curved
spaces, Commun. Math. Phys. 208 (1999), no. 1, 107–123.
6. Keane, M.: Ergodic theory and subshifts of finite type. In Ergodic theory, symbolic dynamics,
and hyperbolic spaces (Trieste, 1989), 35–70, Oxford Sci. Publ., Oxford Univ. Press, New
York, 1991.
7. Krengel, U.: Ergodic theorems. de Gruyter Studies in Mathematics 6 1985. viii+357pp.
8. Lindenstrauss, E.: Pointwise ergodic theorems for amenable groups, Invent. Math. 146
(2001), 259–295.
9. Parry, W.: Topics in ergodic theory. Cambridge Tracts in Mathematics, 75 Cambridge Uni-
versity Press, Cambridge-New York, 1981. x+110 pp.
10. Ruelle, D.: Ergodic theory of differentiable dynamical systems. Inst. Hautes Études Sci. Publ.
Math. No. 50 (1979), 27–58.
11. Steele, J. M.: Kingman’s subadditive ergodic theorem. Ann. Inst. H. Poincaré Probab. Statist.
25 (1989), no. 1, 93–98.
12. Zweimüller, R.: Hopf’s ratio ergodic theorem by inducing. Colloq. Math. 101 (2004), no. 2,
289–292.
Chapter 3
Spectral Theory

3.1 The spectral approach to ergodic theory

A basic problem in ergodic theory is to determine whether two ppt are measure
theoretically isomorphic. This is done by studying invariants: properties, quantities,
or objects which are equal for any two isomorphic systems. The idea is that if two
ppt have different invariants, then they cannot be isomorphic. Ergodicity and mixing
are examples of invariants for measure theoretic isomorphism.
An effective method for inventing invariants is to look for a weaker equivalence
relation, which is better understood. Any invariant for the weaker equivalence rela-
tion is automatically an invariant for measure theoretic isomorphism. The spectral
approach to ergodic theory is an example of this method.
The idea is to associate to the ppt (X, B, µ, T ) the operator UT : L2 (X, B, µ) →
L2 (X, B, µ), UT f = f ◦ T . This is an isometry of L2 (i.e. ∥UT f ∥2 = ∥ f ∥2 and
⟨UT f ,UT g⟩ = ⟨ f , g⟩).
It is useful
R
here to think of L2 as a Hilbert space over C, with inner product
⟨ f , g⟩ := f gdµ. The need to consider complex numbers arises from the need to
consider complex eigenvalues, see below.

Definition 3.1. Two ppt (X, B, µ, T ), (Y, C , ν, S) are called spectrally isomorphic,
if their associated L2 –isometries UT and US are unitarily equivalent: There exists a
linear operator W : L2 (X, B, µ) → L2 (Y, C , ν) s.t.
1. W is invertible;
2. ⟨W f ,W g⟩ = ⟨ f , g⟩ for all f , g ∈ L2 (X, B, µ);
3. WUT = USW .

It is easy to see that any two measure theoretically isomorphic ppt are spectrally iso-
morphic, but we will see later that there are Bernoulli schemes which are spectrally
isomorphic but not measure theoretically isomorphic.

Definition 3.2. A property of ppt is called a spectral invariant, if whenever it holds

for (X, B, µ, T ), it holds for all ppt which are spectrally isomorphic to (X, B, µ, T ).

89
90 3 Spectral theory

Proposition 3.1. Ergodicity and mixing are spectral invariants.

Proof. Suppose (X, B, µ, T ) is a ppt, and let UT be as above. The trick is to phrase
ergodicity and mixing in terms of UT .
Ergodicity is equivalent to the statement “all invariant L2 –functions are con-
stant”, which is the same as saying that dim{ f : UT f = f } = 1. Obviously, this
is a spectral invariant.
Mixing is equivalent to the following statement:

⟨ f ,UTn g⟩ −−−→ ⟨ f , 1⟩⟨g, 1⟩ for all f , g ∈ L2 . (3.1)

n→∞

To see that (3.1) is a spectral invariant, we first note that (3.1) implies:
1. dim{g : UT g = g} = 1. One way to see this is to note that mixing implies ergod-
icity. Here is a more direct proof: If UT g = g, then N1 ∑N−1 n
n=0 UT g = g. By (3.1),
1 N−1 n w
N ∑n=0 UT g −−−→ ⟨g, 1⟩1. Necessarily g = ⟨g, 1⟩1 = const.
N→∞
2. For unitary equivalences W , W 1 = c with |c| = 1. Proof: W 1 is an eigenfunction
with eigenvalue one, so W 1 = const by 1. Since ∥W 1∥1 = ∥1∥2 = 1, |c| = 1.
Suppose now that W : L2 (X, F , µ) → L2 (Y, F , ν) is a unitary equivalence between
(X, F , µ, T ) and (Y, G , ν, S). Suppose T satisfies (3.1). For every F, G ∈ L2 (Y ) we
can write F = W f , G = W g with f , g ∈ L2 (Y ), whence

⟨F,USn G⟩ = ⟨W f ,USnW g⟩ = ⟨W f ,WUTn g⟩ = ⟨ f ,UTn g⟩ → ⟨ f , 1⟩⟨g, 1⟩.

Since W 1 = c with |c|2 = 1, ⟨ f , 1⟩⟨g, 1⟩ = ⟨W f ,W 1⟩⟨W g,W 1⟩ = ⟨F, c1⟩⟨G, c1⟩ =

c⟨F, 1⟩ · c⟨G, 1⟩ = ⟨F, 1⟩⟨G, 1⟩, and we see that US satisfies (3.1). ⊔
⊓

The spectral point of view immediately suggests the following invariant.

Definition 3.3. Suppose (X, B, µ, T ) is a ppt. If f : X → C, 0 ̸= f ∈ L2 satisfies

f ◦ T = λ f , then we say that f is an eigenfunction and that λ is an eigenvalue. The
point spectrum T is the set H(T ) := {λ ∈ C : λ is an eigenvalue}.

H(T ) is a countable subgroup of the unit circle (problem 3.1). Evidently H(T ) is a
spectral invariant of T .
It is easy to see using Fourier expansions that for the irrational rotation Rα ,
H(Rα ) = {eikα : k ∈ Z} (problem 3.2), thus irrational rotations by different angles
are non-isomorphic.
Here are other related invariants:
Definition 3.4. Given a ppt (X, B, µ, T ), let Vd := span{eigenfunctions}. We say
that (X, B, µ, T ) has
1. discrete spectrum (sometime called pure point spectrum), if Vd = L2 ,
2. continuous spectrum, if Vd = {constants} (i.e. is smallest possible),
3. mixed spectrum, if Vd ̸= L2 , {constants}.
3.2 Weak mixing 91

Any irrational rotation has discrete spectrum (problem 3.2). Any mixing transfor-
mation has continuous spectrum, because a non-constant eigenfunction f ◦ T = λ f
satisfies Z
⟨ f , f ◦ T nk ⟩ −−−→ ∥ f ∥22 ̸= | f |2
n→∞

λ nk ∥ f ∥2 ̸= | f dµ|2 for all non-constant

R
along any nk → ∞ s.t. → 1. (To see that
functions, apply Cauchy-Schwarz to f − f , or note that non-constant L2 functions
R

have positive variance.)

The invariant H(T ) is tremendously successful for transformations with discrete
spectrum:
Theorem 3.1 (Discrete Spectrum Theorem). Two ppt with discrete spectrum are
measure theoretically isomorphic iff they have the same group of eigenvalues.
But this invariant cannot distinguish transformations with continuous spectrum. In
particular - it is unsuitable for the study of mixing transformations.

3.2 Weak mixing

3.2.1 Definition and characterization

We saw that if a transformation is mixing, then it does not have non-constant eigen-
functions. But the absence of non-constant eigenfunctions is not equivalent to mix-
ing (see problems 3.8–3.10 for an example). Here we study the dynamical signifi-
cance of this property. First we give it a name.

Definition 3.5. A ppt is called weak mixing, if every f ∈ L2 s.t. f ◦ T = λ f a.e. is

constant almost everywhere.

Theorem 3.2. The following are equivalent for a ppt (X, B, µ, T ) on a Lebesgue
space:
1. weak mixing;
2. for all E, F ∈ B, 1
N ∑N−1
n=0 |µ(E ∩ T
−n F) − µ(E)µ(F)| −−−→ 0;
N→∞
3. for every E, F ∈ B, ∃N ⊂ N of density zero (i.e. |N ∩ [1, N]|/N −−−→ 0) s.t.
N→∞
µ(E ∩ T −n F) −−−−−→ µ(E)µ(F);
N ̸∋n→∞
4. T × T is ergodic.

Proof. We prove (2) ⇒ (3) ⇒ (4) ⇒ (1). The remaining implication (1) ⇒ (2)
requires additional preparation, and will be shown later.
(2) ⇒ (3) is a general fact from calculus (Koopman–von Neumann Lemma): If
an is a bounded sequence of non-negative numbers, then N1 ∑Nn=1 an → 0 iff there is
a set of zero density N ⊂ N s.t. an −−−−−→ 0 (Problem 3.3).
N ̸∋n→∞
92 3 Spectral theory

We show that (3) ⇒ (4). Let S be the semi-algebra {E × F : E, F ∈ B} which

generates B ⊗ B, and fix Ei × Fi ∈ S . By (3), ∃Ni ⊂ N of density zero s.t.

µ(Ei ∩ T −n Fi ) −−−−−→ µ(Ei )µ(Fi ) (i = 1, 2).

Ni ̸∋n→∞

The set N = N1 ∪ N2 also has zero density, and

µ(Ei ∩ T −n Fi ) −−−−−→ µ(Ei )µ(Fi ) (i = 1, 2).

N ̸∋n→∞

Writing m = µ × µ and S = T × T , we see that this implies that

m[(E1 × E2 ) ∩ S−n (F1 × F2 )] −−−−−→ m(E1 × F1 )m(E2 × F2 ),

N ̸∋n→∞

−n
whence 1
N ∑N−1
k=0 m[(E1 × F1 ) ∩ S (E2 × F2 )] −−−→ m(E1 × F1 )m(E2 × F2 ). In sum-
N→∞
mary, 1
N∑N−1 −n
k=0 m[A ∩ S B] − −−→ m(A)m(B) for all A, B ∈ S .
N→∞
Since S generates B ⊗ B, every A, B ∈ B ⊗ B can be approximated up to
measure ε by finite disjoint unions of elements of S . A standard approximation
argument shows that N1 ∑N−1 −n
k=0 m[A ∩ S B] − −−→ m(A)m(B) for all A, B ∈ B ⊗ B.
N→∞
and this implies that T × T is ergodic.
Proof that (4)⇒ (1): Suppose T were not weak mixing, then T has an non-
constant eigenfunction f with eigenvalue λ . The eigenvalue λ has absolute value
equal to one, because |λ |∥ f ∥2 = ∥| f | ◦ T ∥2 = ∥ f ∥2 . Thus

F(x, y) = f (x) f (y)

is T × T –invariant. Since f is non-constant, F is non-constant, and we get a contra-

diction to the ergodicity of T × T .
The proof that (1)⇒ (2) is presented in the next section. ⊔
⊓

3.2.2 Spectral measures and weak mixing

Suppose (X, B, µ, T ) is an invertible ppt. We are interested in the behavior of UTn f

as n → ±∞. Clearly, all the action takes place in the closed invariant subspace

H f := span{UTn f : n ∈ Z}.

It turns out that UT : H f → H f is unitarily equivalent to the operator M : g(z) 7→ zg(z)

on L2 (S1 , B(S1 ), ν f ) where ν f a special finite measure on S1 . This measure is called
the spectral measure of f , and it contains all the information on UT : H f → H f .
To construct it, we need the following important tool from harmonic analysis.
b (n) = S1 zn dµ.
R
Recall that The n-th Fourier coefficient of µ is the number µ
3.2 Weak mixing 93

Theorem 3.3 (Herglotz). A sequence {rn }n∈Z is the sequence of Fourier coeffi-
cients of a positive Borel measure on S1 iff r−n = rn and {rn } is positive definite:
N
∑ rn−m am an ≥ 0 for all sequences {an } and N. This measure is unique.
n,m=−N

It is easy to check that rn = ⟨UTn f , f ⟩ is positive definite (to see this expand
N
⟨Σn=−N anUTn f , Σm=−N
N amUTm f ⟩ noting that ⟨UTn f ,UTm f ⟩ = ⟨UTn−m f , f ⟩).
Definition 3.6. Suppose (X, B, µ, T ) is an invertible ppt, and f ∈ L2 \ {0}. The
spectral measure of f is the unique measure ν f on S1 s.t. ⟨ f ◦ T n , f ⟩ = S1 zn dν f
R

for n ∈ Z.
Proposition 3.2. Suppose (X, B, µ, T ) is an invertible ppt, f ∈ L2 \ {0}, and
H f := span{UTn f : n ∈ Z}. Then UT : H f → H f is unitarily equivalent to the op-
erator g(z) 7→ zg(z) on L2 (S1 , B(S1 ), ν f ).
Proof. By the definition of the spectral measure,
2 * +
N N N N Z
n n m
∑ an z = ∑ an z , ∑ am z = ∑ an am zn−m dν f (z)
n=−N n=−N m=−N n,m=−N S1
L2 (ν f)
2
N N N
= ∑ an am ⟨UTn−m f , f⟩ = ∑ an am ⟨UTn f ,UTm f ⟩ = ∑ anUTn f
n,m=−N n,m=−N n=−N L2 (µ)

N
In particular, if Σn=−N anUTn f = 0 in L2 (µ), then Σn=−N
N an zn = 0 in L2 (ν f ). It follows
that W : UT f 7→ z extends to a linear map from span{UTn f : n ∈ Z} to L2 (ν f ).
n n

This map is an isometry, and it is bounded. It follows that W extends to an linear

isometry W : H f → L2 (ν f ). The image of W contains all the trigonometric poly-
nomials, therefore W (H f ) is dense in L2 (ν f ). Since W is an isometry, its image is
closed (exercise). It follows that W is an isometric bijection from H f onto L2 (ν f ).
Since (WUt )[g(z)] = z[W g(z)] on span{UTn f : n ∈ Z}, WUT g(z) = zg(z) on H f ,
and so W is the required unitary equivalence. ⊔
⊓
Proposition 3.3. Suppose T is weak mixingRinvertible ppt on a Lebesgue space,
then all the spectral measures of f ∈ L2 s.t. f = 0 are non-atomic (this explains
the terminology “continuous spectrum”).
Proof. Suppose f ∈ L2 has integral zero and that ν f has an atom λ ∈ S1 . We con-
N−1 −n n
struct an eigenfunction (with eigenvalue λ ). Consider the sequence N1 ∑n=0 λ UT f .
This sequence is bounded in norm, therefore has a weakly convergent subsequence
(here we use the fact that L2 is separable — a consequence of the fact that (X, B, µ)
is a Lebesgue space):
1 Nk −1 −n n w
∑ λ UT f −N→∞
Nk n=0
−−→ g.

The limit g must satisfy ⟨UT g, h⟩ = ⟨λ g, h⟩ for all h ∈ L2 (check!), therefore it must
be an eigenfunction with eigenvalue λ .
94 3 Spectral theory

Next we claim that g is not constant, and obtain a contradiction to weak mixing:

1 Nk −1 −n n 1 Nk −1
Z
⟨g, f ⟩ = lim ∑ λ ⟨UT f , f ⟩ = lim ∑ λ −n zn dν f (z)
k→∞ Nk n=0 k→∞ Nk n=0

1 Nk −1
Z
= ν f {λ } + lim ∑ λ −n zn dν f (z)
k→∞ Nk n=0 S1 \{λ }

1 1 − λ −Nk zNk
Z
= ν f {λ } + lim −1 z
dν f (z).
k→∞ S1 \{λ } Nk 1 − λ

The limit is equal to zero, because the integrand tends to zero and is uniformly
bounded (by one). Thus ⟨g, f ⟩ = ν f {λ } ̸= 0, whence g ̸= const. ⊔
⊓

Lemma 3.1. Suppose T is an invertible ppt on a Lebesgue space. If T is weak mix-

N−1
1
ing, then for every real-valued f ∈ L2 , ∑ | f · f ◦ T n dµ − ( f dµ)2 | −−−→ 0.
R R
N N→∞
n=0
R
Proof. It is enough to treat the case when f dµ = 0. Let ν f denote the spectral
measure of f , then
2 2
1 N−1 1 N−1 1 N−1
Z Z
2
∑ f · f ◦ T n dµ = ∑ T |⟨U n
f , f ⟩| = ∑ S1 zn dν f (z)
N k=0 N k=0 N k=0
1 N−1
Z  Z 
n n dν (z)
= ∑ S1 z dν f (z) z f
N k=0 S1

1 N−1
Z Z
= ∑ zn wn dν f (z)dν f (w)
N k=0 S1 S1
!
N−1
1
Z Z
n n
= ∑zw dν f (z)dν f (w)
S1 S1 N k=0

The integrand tends to zero and is bounded outside ∆ := {(z, w) : z = w}. If we can
N−1
1 2
∑ | f · f ◦ T n dµ| −−−→ 0.
R
show that (ν f × ν f )(∆ ) = 0, then it will follow that N
k=0 N→∞
This is indeed the case: T is weak mixing,
R
so by the previous proposition ν f is
non-atomic, whence (ν f × ν f )(∆ ) = S1 ν f {w}dν f (w) = 0 by Fubini-Tonelli.
It remains to note that by the Koopman - von Neumann theorem, for every
bounded non-negative sequence an , N1 ∑Nk=1 a2n → 0 iff N1 ∑Nk=1 an → 0, because both
conditions are equivalent to saying that an converges to zero outside a set of indices
of density zero. ⊔
⊓

We can now complete the proof of the theorem in the previous section:
Proposition 3.4. If T is weak mixing (possibly non-invertible) ppt, then for all real-
valued f , g ∈ L2 ,
3.3 The Koopman operator of a Bernoulli scheme 95

1 N−1
Z Z  Z 
n
∑ g · f ◦ T dµ − f dµ gdµ −−−→ 0. (3.2)
N n=0 N→∞

Proof. Suppose
R
first that T is invertible. It is enough to consider f ∈ L2 real valued
such that f dµ = 0.
Set S( f ) := span{UTk f : k ≥ 0} ∪ {1}. Then L2 = S( f ) ⊕ S( f )⊥ , and
1. Every g ∈ S( f ) satisfies (3.2). To see this note that if g1 , . . . , gm satisfy (3.2)
then so does g = ∑ αk gk for any αk ∈ R. Therefore it is enough to check (3.2)
for g := UTRk f and g = const. Constant functions satisfy (3.2) because for such
functions, g f ◦ T n dµ = 0 for g for all n ≥ 0. Functions of the form g = UTk f
satisfy (3.2) because for such functions for all n > k,
Z Z Z
g f ◦ T n dµ = f ◦ T k f ◦ T n dµ = f · f ◦ T n−k dµ −−−−−→ 0
N ̸∋n→∞

for some N ⊂ N of density zero, by Lemma 3.1.

2. EveryR
g ⊥ S( f ) ⊕ {constants} satisfies (3.2) because ⟨g, f ◦ T n ⟩ is eventually zero
and g = ⟨g, 1⟩ = 0.
It follows that every g ∈ L2 satisfies (3.2).
This proves the proposition for invertible ppt. Now consider the case of a non-
invertible ppt. Let (X, e B, e , Te) be the natural extension. A close look at the def-
e µ
inition of B shows that if fe : Xe → R is a B–measurable
e e eigenfunction, then
fe(. . . , x−1 , x0 , x1 , . . .) can be approximated in L2 by functions which depend only
on x0 . Therefore fe(. . . , x−1 , x0 , x1 , . . .) is a.e. completely determined by x0 . Let
π : Xe → X be the natural projection π(. . . , x−1 , x0 , x1 , . . .) = x0 , then fe : Xe → C
is of the form f ◦ πe where f is B–measurable. Thus every eigenfunction for Te is a
lift of an eigenfunction for T . It follows that if T is weak mixing, then Te is weak
mixing. By the first part of the proof, Te satisfies (3.2). Since T is a factor of T , it
also satisfies (3.2).
For a direct proof of the proposition for non-invertible maps, which does not rely
on the natural extension, see Exercise 3.4. ⊔
⊓

3.3 The Koopman operator of a Bernoulli scheme

In this section we analyze the Koopman operator of an invertible Bernoulli scheme.

The idea is to produce an orthonormal basis for L2 which makes the action of UT
transparent.
We cannot expect to diagonalize UT : Bernoulli schemes are mixing, so they have
no non-constant eigenfunctions. But we shall we see that we can get the following
nice structure:
96 3 Spectral theory

Definition 3.7. An invertible ppt is said to have countable Lebesgue spectrum if

L2 has an orthonormal basis of the form {1} ∪ { fλ , j : λ ∈ Λ , j ∈ Z} where Λ is
countable, and UT fλ , j = fλ , j+1 for all i, j.
The reason for the terminology is that the spectral measure of each fλ , j is propor-
tional to the Lebesgue measure on S1 (problem 3.6).
Example. The invertible Bernoulli scheme with probability vector ( 21 , 12 ) has count-
able Lebesgue spectrum.
Proof. The phase space is X = {0, 1}Z . Define for every finite non-empty A ⊂ Z the
function ϕA (x) := ∏ j∈A (−1)x j . Define ϕ∅ := 1. Then,
R
1. if A ̸= ∅ then ϕA = 0 (exercise) R
2. if A ̸= B, then ϕA ⊥ ϕB because ⟨ϕA , ϕB ⟩ = ϕA△B = 0;
3. span{ϕA : A ⊂ Z finite} is algebra of functions which separates points, and con-
tains the constants.
By the Stone-Weierstrass theorem, span{ϕA : A ⊂ Z finite} = L2 , so {ϕA } is an
orthonormal basis of L2 . This is called the Fourier–Walsh system.
Note that UT ϕA = ϕA+1 , where A + 1 := {a + 1 : a ∈ A}. Take Λ the set of equiva-
lence classes of the relation A ∼ B ⇔ ∃c s.t. A = c + B. Let Aλ be a representative of
λ ∈ Λ . The basis is {1} ∪ {ϕAλ +n : λ ∈ Λ , n ∈ Z} = {Fourier Walsh functions}. ⊔ ⊓
It is not easy to produce such bases for other Bernoulli schemes. But they exist.
To prove this we introduce the following sufficient condition for countable Lebesgue
spectrum, which turns out to hold for all Bernoulli schemes as well as many smooth
dynamical systems:
Definition 3.8. An invertible ppt (X, B, µ, T ) is called a K automorphism if there is
a σ –algebra A ⊂ B s.t.
1. T −1 A ⊂ A ;
2. A generates B: σ ( n∈Z T −n A ) = B mod µ;1
S

3. the tail of A is trivial: ∞ −n A = {∅, X} mod µ.

T
n=0 T

Proposition 3.5. Every invertible Bernoulli scheme has the K property.

Proof. Let (SZ , B(SZ ), µ, T ) be a Bernoulli scheme, i.e. B(SZ ) is the sigma algebra
generated by cylinders −k [a−k , . . . , aℓ ] := {x ∈ SZ : xi = ai (−k ≤ i ≤ ℓ)}, T is the
left shift map, and µ(k [a−k , . . . , aℓ ]) = pa−k · · · paℓ .
Call a cylinder non-negative, if it is of the form 0 [a0 , . . . , an ]. Let A be the sigma
algebra generated by all non-negative cylinders. It is clear that T −1 A ⊂ A and
T −n A generates B(SZ ). We show that the measure of every element of
S
that
T∞ n∈Z
n=0 T
−n A is either zero or one.2
1 F ⊂ F mod µ is for all F ∈ F there is a set F ∈ F s.t. µ(F △F ) = 0, and F = F
1 2 1 2 2 2 1 2 1 2
mod µ iff F1 ⊂ F2 mod µ and F2 ⊂ F1 mod µ.
2 Probabilists call the elements of this intersection tail events. The fact that every tail event for

a sequence of independent identically distributed random variables has probability zero or one is
called “Kolmogorov’s zero–one law.”
3.3 The Koopman operator of a Bernoulli scheme 97

Two measurable sets A, B are called independent, if µ(A ∩ B) = µ(A)µ(B). For

Bernoulli schemes, any two cylinders with non-overlapping set of indices is inde-
pendent (check). Thus for every cylinder B of length |B|,

B is independent of T −|B| A for all non-negative cylinders A.

It follows that B is independent of every element of T −|B| A (the set of B’s like
thatTis a sigma-algebra). Thus every cylinder B is independent of every element
of
T n≥1
T −n A . Thus every element of B is independent of every element of
n≥1 T
−n A (another sigma-algebra argument).
This means that every E ∈ n≥1 T −n A is independent of itself. Thus µ(E) =
T

µ(E ∩ E) = µ(E)2 , whence µ(E) = 0 or 1. ⊔

⊓

Proposition 3.6. Every K automorphism on a non-atomic standard probability

space has countable Lebesgue spectrum.

Proof. Let (X, B, µ, T ) be a K automorphism of a non-atomic standard probability

space. Since (X, B, µ) is a non-atomic standard space, L2 (X, B, µ) is (i) infinite
dimensional, and (ii) separable.
Let A be a sigma algebra in the definition of the K property. Set V := L2 (X, A , µ).
This is a closed subspace of L2 (X, B, µ), and
1. U
ST
(V ) ⊆ V , because T −1 A ⊂ A ;
2. n∈Z UTn (V ) is dense in L2 (X, B, µ), because n∈Z T −n A generatesSB, so every
S

B ∈ B can be approximated by a finite disjoint union of elements of n∈Z T −n A ;

3. n=1 UTn (V ) = {constant functions}, because n≥1 T −n A = {∅, X} mod µ.
T∞ T

Now let W := V ⊖ UT (V ) (the orthogonal complement of UT (V ) in V ). For all

n > 0, UTn (W ) ⊂ UTn (V ) ⊂ UT (V ) ⊥ W . Thus W ⊥ UTn (W ) for all n > 0. Since UT−1
is an isometry, W ⊥ UTn (W ) for all n < 0. It follows that

L2 (X, B, µ) = {constants} ⊕
M
UTn (W ) (orthogonal sum).
n∈Z

If { fλ : λ ∈ Λ } is an orthonormal basis for W , then the above implies that

{1} ∪ {UTn fλ : λ ∈ Λ }

is an orthonormal basis of L2 (X, B, µ) (check!).

This is almost the full countable Lebesgue spectrum property. It remains to show
that |Λ | = ℵ0 . |Λ | ≤ ℵ0 because L2 (X, B, µ) is separable. We show that Λ is infi-
nite by proving dim(W ) = ∞. We use the following fact (to be proved later):

∀N ∃A1 , . . . , AN ∈ A pairwise disjoint sets, with positive measure. (3.3)

Suppose we know this. Pick f ∈ W \ {0} (W ̸= {0}, otherwise L2 = {constants}

and (X, B, µ) is atomic). Set wi := f 1Ai ◦ T with A1 , . . . , AN as above, then (i) wi
are linearly independent (because they have disjoint supports); (ii) wi ∈ V (because
98 3 Spectral theory

T −1 Ai ∈ T −1 A ⊂ A , so wi is A –measurable); and (iii) wi ⊥ UT (V ) (check, using

f ∈ W ). It follows that dim(W ) ≥ N. Since N was arbitrary, dim(W ) = ∞.
Here is the proof of (3.3). Since (X, B, µ) is non-atomic, ∃B1 , . . . , BN ∈ B pair-
wise disjoint with positive measure. By assumption, n∈Z T n A generates B, thus
S

we can approximate Bi arbitrarily well by elements of n∈Z T n A . By assumption,

S

A ⊆ T A . This means that we can approximate Bi arbitrarily well by sets from

T n A by choosing n sufficiently large. It follows that L2 (X, T n A , µ) has dimension
at least N. This forces T n A to contain at least N pairwise disjoint sets of positive
measure. It follows that A contains at least N pairwise disjoint sets of positive mea-
sure. ⊔
⊓

Corollary 3.1. All systems with countable Lebesgue spectrum, whence all invertible
Bernoulli schemes, are spectrally isomorphic.

Proof. Problem 3.7. ⊔

⊓

We see that all Bernoulli schemes are spectrally isomorphic. But it is not true that
all Bernoulli schemes are measure theoretically isomorphic. To prove this one needs
new (non-spectral) invariants. Enter the measure theoretic entropy, a new invariant
which we discuss in the next chapter.

Problems

3.1. Suppose (X, B, µ, T ) is an ergodic ppt on a Lebesgue space, and let H(T ) be
its group of eigenvalues.
1. show that if f is an eigenfunction, then | f | = const. a.e., and that if λ , µ ∈ H(T ),
then so do 1, λ µ, λ /µ.
2. Show that eigenfunctions of different eigenvalue are orthogonal. Deduce that
H(T ) is a countable subgroup of the unit circle.

3.2. Prove that the irrational rotation Rα has discrete spectrum, and calculate H(Rα ).

3.3. Koopman - von Neumann Lemma

Suppose an is a bounded sequence of non-negative numbers. Prove that N1 ∑Nn=1 an →
0 iff there is a set of zero density N ⊂ N s.t. an −−−−−→ 0. Guidance: Fill in the
N ̸∋n→∞
details in the following argument.
1. Suppose N ⊂ N has density zero and an −−−−−→ 0, then 1
N ∑Nn=1 an → 0.
N ̸∋n→∞
1
2. Now assume that N ∑Nn=1 an → 0.
a. Show that Nm := {k : ak > 1/m} form an increasing sequence of sets of den-
sity zero.
b. Fix εi ↓ 0, and choose ki ↑ ∞ such that if n > ki , then (1/n)|Ni ∩ [1, n]| < εi .
Show that N := i Ni ∩ (ki , ki+1 ] has density zero.
S
3.3 The Koopman operator of a Bernoulli scheme 99

c. Show that an −−−−−→ 0.

N ̸∋n→∞

3.4. Here is a sketch of an alternative proof of proposition 3.4, which avoids natural
extensions (B. Parry). Fill in the details.
1. Set H := L2 , V := n and W := H ⊖UT H := {g ∈ H, g ⊥ UT H}.
T
n≥0 UT (H),

a. H = V ⊕ [(UT H)⊥ + (UT2 )⊥ + · · · ]

decreasing, {(UTk H)⊥ } us increasing.
b. {UTk H} isL
c. H = V ⊕ ∞ k
k=1 UT W (orthogonal space decomposition).

2. UT : V → V has a bounded inverse (hint: use the fact from Banach space the-
ory that any bounded linear operator between mapping one Banach space onto
another Banach space which is one-to-one, has a bounded inverse).
3. (3.2) holds for any f , g ∈ V .
4. if g ∈ UTk W for some k, then (3.2) holds for all f ∈ L2 .
5. if g ∈ V , but f ∈ UTk W for some k, then (3.2) holds for f , g.
6. (3.2) holds for all f , g ∈ L2 .
3.5. Show that every invertible ppt with countable Lebesgue spectrum is mixing,
whence ergodic.
3.6. Suppose (X, B, µ, T ) has countable Lebesgue spectrum. Show that { f ∈ L2 :
R
f = 0} is spanned by functions f whose spectral measures ν f are equal to the
Lebesgue measure on S1 .
3.7. Show that any two ppt with countable Lebesgue spectrum are spectrally iso-
morphic.
3.8. Cutting and Stacking and Chacon’s Example
This is an example of a ppt which is weak mixing but not mixing. The example is
a certain map of the unit interval, which preserves Lebesgue’s measure. It is con-
structed using the method of “cutting and stacking” which we now explain.
Let A0 = [1, 32 ) and R0 := [ 23 , 1] (thought of as reservoir).
Step 1: Divide A0 into three equal subintervals of length 92 . Cut a subinterval B0
of length 29 from the left end of the reservoir.
• Stack the three thirds of A0 one on top of the other, starting from the left and
moving to the right.
• Stick B0 between the second and third interval.
• Define a partial map f1 by moving points vertically in the stack. The map is
defined everywhere except on R \ B0 and the top floor of the stack. It can be
viewed as a partially defined map of the unit interval.
Update the reservoir: R1 := R \ B0 . Let A1 be the base of the new stack (equal to
the rightmost third of A0 ).
Step 2: Cut the stack vertically into three equal stacks. The base of each of these
thirds has length 13 × 29 . Cut an interval B1 of length 13 × 29 from the left side of
the reservoir R1 .
100 3 Spectral theory

• Stack the three stacks one on top of the other, starting from the left and moving
to the right.
• Stick B1 between the second stack and the third stack.
• Define a partial map f2 by moving points vertically in the stack. This map is
defined everywhere except the union of the top floor floor and R1 \ B1 .
Update the reservoir: R2 := R1 \ B1 . Let A2 be the base of the new stack (equal to
the rightmost third of A1 ).
Step 3: Cut the stack vertically into three equal stacks. The base of each of these
thirds has length 312 × 29 . Cut an interval B2 of length 312 × 92 from the left side of
the reservoir R2 .
• Stack the three stacks one on top of the other, starting from the left and moving
to the right.
• Stick B2 between the second stack and the third stack.
• Define a partial map f3 by moving points vertically in the stack. This map is
defined everywhere except the union of the top floor floor and R2 \ B2 .
Update the reservoir: R3 := R2 \ B2 . Let A3 be the base of the new stack (equal to
the rightmost third of A2 )

A0 R0 B0
l1

B0 A1 R1

step 1 (cutting) step 1 (stacking)

B1
l2
B0
B1

A1 R1
A2 R2
step 2 (cutting) step 2 (stacking)

Fig. 3.1 The construction of Chacon’s example

Continue in this manner, to obtain a sequence of partially defined maps fn . There is

a canonical way of viewing the intervals composing the stacks as of subintervals of
3.3 The Koopman operator of a Bernoulli scheme 101

the unit interval. Using this identification, we may view fn as partially defined maps
of the unit interval.
1. Show that fn is measure preserving where it is defined (the measure is Lebesgue’s
measure). Calculate the Lebesgue measure of the domain of fn .
2. Show that fn+1 extends fn (i.e. the maps agree on the intersection of their do-
mains). Deduce that the common extension of fn defines an invertible probability
preserving map of the open unit interval. This is Chacon’s example. Denote it by
(I, B, m, T ).
3. Let ℓn denote the height of the stack at step n. Show that the sets {T i (An ) : i =
0, . . . , ℓn , n ≥ 1} generate the Borel σ –algebra of the unit interval.

3.9. (Continuation) Prove that Chacon’s example is weak mixing using the follow-
ing steps. Suppose f is an eigenfunction with eigenvalue λ .
1. We first show that if f is constant on An for some n, then f is constant every-
where. (An is the base of the stack at step n.)
a. Let ℓk denote the height of the stack at step k. Show that An+1 ⊂ An , and
T ℓn (An+1 ) ⊂ An . Deduce that λ ℓn = 1.
b. Prove that λ ℓn+1 = 1. Find a recursive formula for ℓn . Deduce that λ = 1.
c. The previous steps show that f is an invariant function. Show that any invari-
ant function which constant on An is constant almost everywhere.
2. We now consider the case of a general L2 – eigenfunction.
a. Show, using Lusin’s theorem, that there exists an n such that f is nearly con-
stant on most of An . (Hint: part 3 of the previous question).
b. Modify the argument done above to show that any L2 –eigenfunction is con-
stant almost everywhere.

3.10. (Continuation) Prove that Chacon’s example is not mixing, using the following
steps.
1. Inspect the image of the top floor of the stack at step n, and show that for every n
and 0 ≤ k ≤ ℓn−1 , m(T k An ∩ T k+ℓn An ) ≥ 31 m(T k An ).
2. Use problem 3.8 part 3 and an approximation argument to show that for every
Borel set E and ε > 0, m(E ∩ T ℓn E) ≥ 31 m(E) − ε for all n. Deduce that T cannot
be mixing.

Notes to chapter 3

The spectral approach to ergodic theory is due to von Neumann. For a thorough
modern introduction to the theory, see Nadkarni’s book [1]. Our exposition follows
in parts the books by Parry [2] and Petersen [1]. A proof of the discrete spectrum
theorem mentioned in the text can be found in Walters’ book [5]. A proof of Her-
glotz’s theorem is given in [2].
102 3 Spectral theory

References

1. Nadkarni, M. G.: Spectral theory of dynamical systems. Birkhäuser Advanced Texts:

Birkhäuser Verlag, Basel, 1998. x+182 pp.
2. Parry, W.: Topics in ergodic theory. Cambridge Tracts in Mathematics, 75. Cambridge Uni-
versity Press, Cambridge-New York, 1981. x+110 pp.
3. Petersen, K.: Ergodic theory. Corrected reprint of the 1983 original. Cambridge Studies in
Advanced Mathematics 2 Cambridge University Press, Cambridge, 1989. xii+329 pp.
4. Walters, P.: An introduction to ergodic theory. Graduate Texts in Mathematics, 79 Springer-
Verlag, New York-Berlin, 1982. ix+250 pp.
Chapter 4
Entropy

We saw at the end of the last chapter that every two Bernoulli schemes are spec-
trally isomorphic (because they have countable Lebesgue spectrum). The question
whether any two Bernoulli schemes are measure theoretically isomorphic was a ma-
jor open problem until it was answered negatively by Kolmogorov and Sinai, who
introduced for this purpose a new invariant: metric entropy. Later, Ornstein proved
that this invariant is complete within the class of Bernoulli schemes: Two Bernoulli
schemes are measure theoretically isomorphic iff their metric entropies are equal.

4.1 Information content and entropy

Let (X, B, µ) be a probability space. Suppose x ∈ X is unknown. How to quantify

the “information content” I(A) of the statement “x belongs to A?”
1. I(A) should be non-negative, and if µ(A) = 1 then I(A) = 0;
2. If A, B are independent, i.e. µ(A ∩ B) = µ(A)µ(B), then I(A ∩ B) = I(A) + I(B);
3. I(A) should be a Borel decreasing function of the probability of A.

Proposition 4.1. The only functions ϕ : [0, 1] → R+ such that I(A) = ϕ[µ(A)] sat-
isfies the above axioms for all probability spaces (X, B, µ) are c lnt with c < 0.

We leave the proof as an exercise. This leads to the following definition.

Definition 4.1 (Shannon). Let (X, B, µ) be a probability space.
1. The Information Content of a set A ∈ B is Iµ (A) := − log2 µ[A]
2. The Information Function of a countable measurable partition α of X is

Iµ (α)(x) := ∑ Iµ (A)1A (x) = − ∑ log2 µ(A)1A (x)

A∈α A∈α

3. The Entropy of a countable measurable

R
partition is the average of the information
content of its elements: Hµ (α) := X Iµ (α)dµ = − ∑A∈α µ(A) log2 µ(A).

103
104 4 Entropy

Conventions: Henceforth, log = log2 , ln = loge and 0 log 0 = 0.

The are important conditional versions of these notions:
Definition 4.2. Let (X, B, µ) be a probability space, and suppose F is a sub-σ -
algebra of B. We use the notation µ(A|F )(x) := E(1A |F )(x) (as L1 –elements).
1. The information content of A given F is Iµ (A|F )(x) := − log2 µ(A|F )(x)
2. The information function of a finite measurable partition α given F is Iµ (α|F ) :=
∑A∈α Iµ (A|F )1A
3. The conditional entropy of α given F is Hµ (α|F ) := Iµ (α|F )dµ.
R

Convention: Let α, β be partitions; We write Hµ (α|β ) for Hµ (α|σ (β )), where

σ (β ) :=smallest σ –algebra which contains β .
The following formulæ are immediate:
Z
Hµ (α|F ) = − ∑ µ(A|F )(x) log µ(A|F )(x)dµ(x)
X A∈α

µ(A ∩ B)
Hµ (α|β ) = − ∑ µ(B) ∑ µ(A|B) log µ(A|B), where µ(A|B) = .
B∈β A∈α µ(B)

4.2 Properties of the entropy of a partition

We need some notation and terminology. Let α, β be two countable partitions.

1. σ (α) is the smallest σ –algebra which contains α;
2. α ≤ β means that α ⊆ σ (β ) mod µ, i.e. every element of α is equal up to a set
of measure zero to an element of σ (β ). Equivalently, α ≤ β if every element of
α is equal up to a set of measure zero to a union of elements of β . We say that β
is finer than α, and that α is coarser than β .
3. α = β mod µ iff α ≤ β and β ≤ α.
4. α ∨ β is the smallest partition which is finer than both α and β . Equivalently,
α ∨ β := {A ∩ B : A ∈ α, B ∈ β }.
If F1 , F2 are two σ –algebras, then F1 ∨ F2 is the smallest σ –algebra which con-
tains F1 , F2 .

4.2.1 The entropy of α ∨ β

Suppose α = {A1 , . . . , An } is a finite measurable partition of a probability space

(X, B, µ). Suppose x ∈ X is unknown. It is useful to think of the information content
of α as of the partial information on x “which element of α contains an x.”
4.2 Properties of the entropy of a partition 105

We state and prove a formula which says that the information content of α and
β is the information content of α plus the information content of β given the knowl-
edge α.

Theorem 4.1 (The Basic Identity). Suppose α, β are measurable countable parti-
tions, and assume Hµ (α), Hµ (β ) < ∞, then
1. Iµ (α ∨ β |F ) = Iµ (α|F ) + Iµ (β |F ∨ σ (α));
2. Hµ (α ∨ β ) = Hµ (α) + Hµ (β |α).

Proof. We calculate Iµ (β |F ∨ σ (α)):

Iµ (β |F ∨ σ (α)) = − ∑ 1B log µ(B|F ∨ σ (α))

B∈β

F)
Claim: µ(B|F ∨ σ (α)) = ∑A∈α 1A µ(B∩A|
µ(A|F ) :

1. This expression is F ∨ σ (α)–measurable

2. Observe that F ∨ σ (α) = { A∈α A ∩ FA : FA ∈ F } (this is a σ –algebra which
U

contains α and F ). Thus every F ∨ σ (α)–measurable function is of the form

∑A∈α 1A ϕA with ϕA F –measurable. It is therefore enough to check test functions
of the form 1A ϕ with ϕ ∈ L∞ (F ). For such functions

µ(B ∩ A′ |F ) µ(B ∩ A|F )

Z Z
1A ϕ ∑ 1A′ dµ = 1A ϕ dµ =
′
A ∈α
µ(A′ |F ) µ(A|F )
µ(B ∩ A|F )
Z Z Z
= E(1A |F )ϕ dµ = ϕ µ(B ∩ A|F )dµ = 1A ϕ · 1B dµ.
µ(A|F )
F)
So µ(B|F ∨ σ (α)) = ∑A∈α 1A µ(B∩A|
µ(A|F ) as claimed.

Using the claim, we see that

µ(B ∩ A|F )
Iµ (β |F ∨ σ (α)) = − ∑ 1B log ∑ 1A
B∈β A∈α µ(A|F )
µ(B ∩ A|F )
=− ∑ ∑ 1A∩B log
B∈β A∈α
µ(A|F )
=− ∑ ∑ 1A∩B log µ(B ∩ A|F ) + ∑ ∑ 1A∩B log µ(A|F )
B∈β A∈α A∈α B∈β

= Iµ (α ∨ β |F ) − Iµ (α|F ).

This proves the first part of the theorem.

Integrating, we get Hµ (α ∨ β |F ) = Hµ (α|F ) + Hµ (β |F ∨ α). If F = {∅, X},
then Hµ (α ∨ β ) = Hµ (α) + Hµ (β |α). ⊔
⊓
106 4 Entropy

4.2.2 Convexity properties

Lemma 4.1. Let ϕ(t) := −t logt, then for every probability vector (p1 , . . . , pn ) and
x1 , . . . , xn ∈ [0, 1] ϕ(p1 x1 + · · · + pn xn ) ≥ p1 ϕ(x1 ) + · · · + pn ϕ(xn ), with equality iff
all the xi with i s.t. pi ̸= 0 are equal.
Proof. This is because ϕ(·) is strictly concave. Let m := ∑ pi xi . If m = 0 then the
lemma is obvious, so suppose m > 0. It is an exercise in calculus to see that ϕ(t) ≤
ϕ(m) + ϕ ′ (m)(t − m) for t ∈ [0, 1], with equality iff t = m. In the particular case
m = ∑ pi xi and t = xi we get

pi ϕ(xi ) ≤ pi ϕ(m) + ϕ ′ (m)(pi xi − pi m) with equality iff pi = 0 or xi = m.

Summing over i, we get ∑ pi ϕ(xi ) ≤ ϕ(m) + ϕ ′ (m)(∑ pi xi − m) = ϕ(m). There is

an equality iff for every i pi = 0 or xi = m. ⊔
⊓
Proposition 4.2 (Convexity properties). Let α, β , γ be countable measurable par-
titions with finite entropies, then
1. α ≤ β ⇒ Hµ (α|γ) ≤ Hµ (β |γ)
2. α ≤ β ⇒ Hµ (γ|α) ≥ Hµ (γ|β )
Proof. The basic identity shows that β ∨ γ has finite entropy, and so Hµ (β |γ) =
Hµ (α ∨ β |γ) = Hµ (α|γ) + Hµ (β |γ ∨ α) ≥ Hµ (α|γ).
For the second inequality, note that ϕ(t) = −t logt is strictly concave (i.e. its
negative is convex), therefore by Jensen’s inequality
Z Z
Hµ (γ|α) = ∑ ϕ[E(C|σ (α))]dµ = ∑ ϕ[E(E(C|σ (β ))|σ (α))]dµ ≥
C∈γ C∈γ
Z Z
≥ ∑ E(ϕ[E(1C |σ (β ))]|σ (α))]dµ = ∑ ϕ[E(1C |σ (β ))]dµ ≡ Hµ (γ|β ),
C∈γ C∈γ

proving the inequality. ⊔

⊓

4.2.3 Information and independence

We say that two partitions are independent, if ∀A ∈ α, B ∈ β , µ(A ∩ B) = µ(A)µ(B).

This the same as saying that the random variables α(x), β (x) are independent.
Proposition 4.3 (Information and Independence). Hµ (α ∨ β ) ≤ Hµ (α) + Hµ (β )
with equality iff α, β are independent.
Proof. Hµ (α ∨ β ) = Hµ (α) + Hµ (β ) iff Hµ (α|β ) = Hµ (α). But

Hµ (α|β ) = − ∑ µ(B) ∑ µ(A|B) log µ(A|B).

B∈β A∈α
4.3 The Metric Entropy 107

Let ϕ(t) = −t logt. We have:

∑ ∑ µ(B)ϕ[µ(A|B)] = ∑ ϕ[µ(A)].
A∈α B∈β A∈α

But ϕ is strictly concave, so ∑B∈β µ(B)ϕ[µ(A|B)] ≤ ϕ[µ(A)], with equality iff

µ(A|B) are equal for all B ∈ β s.t. µ(B) ̸= 0.
We conclude that µ(A|B) = c(A) for all B ∈ β s.t. µ(B) ̸= 0. For such B,
µ(A ∩ B) = c(A)µ(B). Summing over B, gives c(A) = µ(A) and we obtain the inde-
pendence condition. ⊔
⊓

4.3 The Metric Entropy

4.3.1 Definition and meaning

Definition 4.3 (Kolmogorov, Sinai). The metric entropy of a ppt (X, B, µ, T ) is

defined to be

hµ (T ) := sup{hµ (T, α) : α is a countable measurable partition s.t. Hµ (α) < ∞},

 
n−1
hµ (T, α) := lim 1n Hµ T −i α
W
where .
n→∞ i=0

It can be shown that the supremum is attained by finite measurable partitions (prob-
lem 4.9).
Proposition 4.4. The limit which defines hµ (T, α) exists.
Proof. Write αn := n−1
W −i
i=0 T α. Then an := Hµ (αn ) is subadditive, because an+m :=
Hµ (αn+m ) ≤ Hµ (αn ) + Hµ (T −n αm ) = an + am .
We claim that any sequence of numbers {an }n≥1 which satisfies an+m ≤ an + am
converges to a limit (possibly equal to minus infinity), and that this limit is inf[an /n].
Fix n. Then for every m, m = kn + r, 0 ≤ r ≤ n − 1, so

am ≤ kan + ar .

Dividing by m, we get that for all m > n

am kan + ar an ar
≤ ≤ + ,
m kn + r n m
whence lim sup(am /m) ≤ an /n. Since this is true for all n, lim sup am /m ≤ inf an /n.
But it is obvious that lim inf am /m ≥ inf an /n, so the limsup and liminf are equal,
and their common value is inf an /n.
We remark that in our case the limit is not minus infinity, because Hµ ( n−1
W −i
i=0 T α)
are all non-negative. ⊔
⊓
108 4 Entropy

Hµ (αn ) is the average information content in the first n–digits of the α–itinerary.
Dividing by n gives the average “information per unit time.” Thus the entropy mea-
sure the maximal rate of information production the system is capable of generating.
It is also possible to think of entropy as a measure of unpredictability. Let’s think
of T as moving backward in time. Then α1∞ := σ ( ∞
S −n α) contains the informa-
n=1 T
tion on the past of the itinerary. Given the future, how unpredictable is the present,
on average? This is measured by Hµ (α|α1∞ ).
−i α
S∞

Theorem 4.2. If Hµ (α) < ∞ , then hµ (T, α) = Hµ (α|α1∞ ), where α1∞ = σ i=1 T .

Proof. We show that hµ (T, α) = Hµ (α|α1∞ ). Let αmn :=

Wn −i α, then
i=m T

Hµ (α|α0n ) = Hµ (α0n ) − Hµ (T −1 α0n−1 ) = Hµ (α0n ) − Hµ (α0n−1 ).

Summing over n, we obtain

n
Hµ (αn ) − Hµ (α) = ∑ Hµ (α|α1k )
k=1

Dividing by n and passing to the limit we get

1 n
hµ (T, α) = lim
n→∞ n
∑ Hµ (α|α1k )
k=1

It is therefore enough to show that Hµ (α|α1k ) −−−→ Hµ (α|α1∞ ).

k→∞
This is dangerous! It is true that Hµ (α|α1k ) = Iµ (α|α1k )dµ and that by the mar-
R

tingale convergence theorem

Iµ (α|α1k ) −−−→ Iµ (α|α1∞ ) a.e. .

k→∞

But the claim that the integral of the limit is equal to the limit of the integrals requires
justification.
If |α| < ∞, then we can bypass the problem by writing
Z
Hµ (α|α1k ) = ∑ ϕ[µ(A|α1k )]dµ, with ϕ(t) = −t logt,
A∈α

and noting that this function is bounded (by |α| max ϕ). This allows us to apply the
bounded convergence theorem, and deduce that Hµ (α|α1k ) −−−→ Hµ (α|α1∞ ).
k→∞
If |α| = ∞ (but Hµ (α) < ∞) then we need to be more clever, and appeal to the
following lemma (proved below):
Lemma 4.2 (Chung–Neveu). Suppose α is a countable measurable partition with
finite entropy, then the function f ∗ := supn≥1 Iµ (α|α1n ) is absolutely integrable.
The result now follows from the dominated convergence theorem. ⊔
⊓
4.3 The Metric Entropy 109

Here is the proofUof the Chung Neveu Lemma. Fix A ∈ α, then we may decom-
pose A ∩ [ f ∗ > t] = m≥1 A ∩ Bm (t; A), where

Bm (t; A) := {x ∈ X : m is the minimal natural number s.t. − log2 µ(A|α1m ) > t}.

We have

µ[A ∩ Bm (t; A)] = Eµ 1A 1Bm (t;A) = Eµ Eµ 1A 1Bm (t;A) |σ (α1m )

= Eµ 1Bm (t;A) Eµ (1A |σ (α1m )) , because Bm (t; A) ∈ σ (α1m )

 m

≡ Eµ 1Bm (t;A) 2log2 µ (A|σ (α1 ))
≤ Eµ 1Bm (t;A) 2−t = 2−t µ[Bm (t; A)].

Summing over m we see that µ(A ∩ [ f ∗ > t]) ≤ 2−t . Of course we also have µ(A ∩
[ f ∗ > t]) ≤ µ(A). Thus µ(A ∩ [ f ∗ > t]) ≤ min{µ(A), 2−t }. R
We now use the following fact from measure theory: If g ≥ 0, then gdµ =
1
R∞
0 µ[g > t]dt:
Z Z ∞ Z ∞
f ∗ dµ = µ(A ∩ [ f ∗ > t])dt ≤ min{µ(A), 2−t }dt
A 0 0
2−t
Z − log µ(A) Z ∞ ∞
2
−t
≤ µ(A)dt + 2 dt = −µ(A) log2 µ(A) −
0 − log2 µ(A) ln 2 − log2 µ(A)
= −µ(A) log2 µ(A) + µ(A)/ ln 2.
R ∗
Summing over A ∈ α we get that f dµ ≤ Hµ (α) + (ln 2)−1 < ∞. ⊔
⊓

4.3.2 The Shannon–McMillan–Breiman Theorem

Theorem 4.3 (Shannon–McMillan–Breiman). Let (X, B, µ, T ) be an ergodic ppt,

and α a countable measurable partition of finite entropy, then
1
Iµ (α0n−1 ) −−−→ hµ (T, α) a.e.
n n→∞

In particular, if αn (x) :=element of αn which contains x, then

1
− log µ(αn (x)) −−−→ hµ (T, α) a.e.
n n→∞

Proof. Let α0n := ni=0 T −i α. Recall the basic identity Iµ (α0n−1 ) ≡ Iµ (α1n−1 ∨ α) =
W

Iµ (α1n−1 ) + Iµ (α|α1n−1 ). This gives

1
R R R∞ R∞R R∞
Proof: X gdµ = X 0 1[0≤t<g(x)] (x,t)dtdµ(x) = 0 X 1[g>t] (x,t)dµ(x)dt = 0 µ[g > t]dt.
110 4 Entropy

Iµ (α0n ) = Iµ (α|α1n ) + Iµ (α0n−1 ) ◦ T

= Iµ (α|α1n ) + [Iµ (α|α1n−1 ) + Iµ (α0n−2 ) ◦ T ] ◦ T
n−1
= ··· = ∑ Iµ (α|α1n−k ) ◦ T k + Iµ (α) ◦ T n
k=0
!
n n n−1
= ∑ Iµ (α|α1k ) ◦ T n−k + ∑ Iµ (α) ◦ T k
− ∑ Iµ (α) ◦ T k
k=1 k=0 k=0
n
= ∑ Iµ (α|α1k ) ◦ T n−k + o(n) a.e.,
k=1

because 1n ×brackets tends to zero a.e. by the pointwise ergodic theorem.

By the Martingale Convergence Theorem, Iµ (α|α1k ) −−−→ Iµ (α|α1∞ ). The idea of
k→∞
the proof is to use this to say

1 1 n
lim Iµ (α0n ) ≡ lim ∑ Iµ (α|α1k ) ◦ T n−k
n→∞ n n→∞ n
k=1

? 1 n n−k 1 n−1
= lim
n→∞ n
∑ Iµ (α|α ∞
1 ) ◦ T ≡ lim
n→∞ n
∑ Iµ (α|α1∞ ) ◦ T k
k=1 k=0
Z
= Iµ (α|α1∞ )dµ (Ergodic Theorem)
= Hµ (α|α1∞ ) = hµ (T, α).

The point is to justify the question mark. Write fn := Iµ (α|α1n ) and f∞ =

Iµ (α|α1∞ ). It is enough to show

1 n
Z
lim sup
n→∞
∑ | fk − f∞ | ◦ T n−k dµ = 0.
n k=1
(4.1)

(This implies that the limsup is zero almost everywhere.) Set Fn := supk>n | fk − f∞ |.
Then Fn → 0 almost everywhere. We claim that Fn → 0 in L1 . This is because of
the dominated convergence theorem and the fact that Fn ≤ 2 f ∗ := 2 supm fm ∈ L1
(Chung–Neveu Lemma). Fix some large N, then

1 n−1
Z
lim sup
n→∞
∑ | fn−k − f∞ | ◦ T k dµ =
n k=0
1 n−N−1 1 n−1
Z Z
= lim sup
n→∞
∑ | fn−k − f∞ | ◦ T k dµ +
n k=0
lim sup
n→∞
∑ | fn−k − f∞ | ◦ T k dµ
n k=n−N
!
1 n−N−1 1 N−1 ∗
Z Z
≤ lim
n→∞ n
∑ FN ◦ T k dµ + lim sup
n→∞
∑ 2 f ◦ T k ◦ T n−N dµ
n k=0
k=0
4.3 The Metric Entropy 111
!
1 N−1 ∗
Z Z Z
k
= FN dµ + lim sup ∑ 2 f ◦ T dµ = FN dµ.
n→∞ n k=0

Since FN → 0 in L1 , FN dµ → 0. So (4.1) holds.

R
⊔
⊓

4.3.3 Sinai’s Generator theorem

Let F1 , F2 be two sub-σ –algebras of a probability space (X, B, µ). We write F1 ⊆

F2 mod µ, if ∀F1 ∈ F1 , ∃F2 ∈ F2 s.t. µ(F1 △F2 ) = 0. We write F1 = F2 mod µ,
if both inclusions hold mod µ. For example, B(R) = B0 (R) Wmod Lebesgue’s
∞ = W∞ −i ∞ ∞ −i
measure. For every partition α, let α−∞ i=−∞ T α , α0 := i=0 T α denote
∞ ∞
T −i α and T −i α.
S S
the smallest σ –algebras generated by, respectively,
i=−∞ i=0

Definition 4.4. A countable measurable partition α is called a generator for an

B, T −i α = B mod µ, and a strong generator, if
W∞
invertible (X, µ, T ) if i=−∞
i=0 T α = B mod µ.
W∞ −i

(This latter definition makes sense in the non-invertible case as well)

Example: α = {[0, 21 ), [ 12 , 1]} is a strong generator for T x = 2x mod 1, because
W∞ −i S n−1
i=0 T α = σ ( n α0 ) is the Borel σ –algebra (it contains all dyadic intervals,
whence all open sets).

Theorem 4.4 (Sinai’s Generator Theorem). Let (X, B, µ, T ) be an invertible ppt.

If α is a generator of finite entropy, then hµ (T ) = hµ (T, α). A similar statement
holds for non-invertible ppt assuming that α is a strong generator.

Proof. We treat the inverible case only, and leave the non-invertible case as an ex-
ercise. Fix a finite measurable partition β ; Must show that hµ (T, β ) ≤ hµ (T, α).
Step 1. hµ (T, β ) ≤ hµ (T, α) + Hµ (β |α)

1 1 1
Hµ (β0n−1 ) ≤ Hµ (β0n−1 ∨ α0n−1 ) = Hµ (α0n−1 ) + Hµ (β0n−1 |α0n−1 )

n " n n#
n−1
1
≤ Hµ (α0n−1 ) + ∑ Hµ (T −k β |α0n−1 ) ∵ Hµ (ξ ∨ η|G ) ≤ Hµ (ξ |G ) + Hµ (η|G )
n k=0
" #
n−1
1
≤ Hµ (α0n−1 ) + ∑ Hµ (T −k β |T −k α)
n k=0
1
= Hµ (α0n−1 ) + Hµ (β |α).
n
Now pass to the limit.
112 4 Entropy

N )
Step 2. For every N, hµ (T, β ) ≤ hµ (T, α) + Hµ (β |α−N
N instead of α, and check that h (T, α N ) =
Repeat the previous step with α−N µ −N
hµ (T, α).
N ) −−−→ H (β |B) = 0.
Step 3. Hµ (β |α−N µ
N→∞

Z Z
N N N
Hµ (β |α−N )= Iµ (β |α−N )dµ = − ∑ 1B log µ(B|α−N )dµ
B∈β
Z
N N
=− ∑ µ(B|α−N ) log µ(B|α−N )dµ
B∈β
Z Z
N
= ∑ ϕ[log µ(B|α−N )]dµ −−−→ ∑ ϕ[log µ(B|B)]dµ = 0,
N→∞
B∈β B∈β

because µ(B|B) = 1B , ϕ[1B ] = 0, and |β | < ∞

This proves that hµ (T, α) ≥ sup{hµ (T, β ) : |β | < ∞}. Problem 4.9 says that this
supremum is equal to hµ (T ), so we are done. ⊔
⊓

The following variation on Sinai’s Generator Theorem is very useful. Suppose

(X, d) is a compact metric space. A finite Borel partition of X is finite collection
α = {A1 , . . . , AN } of pairwise disjoint Borel sets whose union covers X. Let

diam(α) := max{diam(A) : A ∈ α},

where diam(A) := sup{d(x, y) : x, y ∈ A}.

Proposition 4.5. Let T be a Borel map on a compact metric space (X, d), and sup-
pose αn is a sequence of finite Borel partitions such that diam(αn ) −−−→ 0. Then for
n→∞
every T -invariant Borel probability measure µ, hµ (T ) = limn→∞ hµ (T, αn ).

Proof. We begin, as in the proof of Sinai’s generator theorem, with the inequality

hµ (T, β ) ≤ hµ (T, α) + Hµ (β |α).

Step 1. For every ε and m there is a δ such that if β = {B1 , . . . , Bm } and γ =

{C1 , . . . ,Cm } are partitions such that ∑ µ(Bi △Ci ) < δ , then Hµ (β |γ) < ε:
Proof. We prepare a partition ξ with small entropy such that β ∨ γ = ξ ∨ γ and then
write Hµ (β |γ) = Hµ (β ∨ γ) − Hµ (γ) = Hµ (ξ ∨ γ) −SHµ (γ) ≤ Hµ (ξ ).
Here is the partition ξ : ξ := {Bi ∩C j : i ̸= j} ∪ { Bi ∩Ci }. Then:
1. β ∨ξ = β ∨γ: β ∨ξ ≤ β ∨γ by construction, and β ∨ξ ≥ β ∨γ because Bi ∩C j ∈
β ∨ ξ for i ̸= j and Bi ∩Ci = Bi \ k̸=ℓ Bk ∩Cℓ
S

2. γ ∨ ξ = β ∨ γ for the same reason

3. ξ contains m(m − 1) elements with total measure < δ and one element with total
measure > 1 − δ . So Hµ (ξ ) ≤ m(m − 1)δ | log δ | + (1 − δ )| log(1 − δ )|.
4.4 Examples 113

Taking δ sufficiently small gives Hµ (ξ ) < ε. So Hµ (β |γ) ≤ Hµ (ξ ) < ε.

Step 2. Suppose diam(αn ) −−−→ 0, then Hµ (β |αn ) −−−→ 0.
n→∞ n→∞

Proof. Fix δ , to be determined later. Write β = {B1 , . . . , Bm } and find compact sets
Ki and open sets Ui such that Ki ⊂ Bi ⊂ Ui and ∑ µ(Ui \ Ki ) < δ . Find δ ′ > 0 so
small that dist (Ki , K j ) > δ ′ for all i ̸= j, and Nδ ′ (Ki ) ⊂ Ui .
Fix N so that n > N ⇒ diam(αn ) < δ ′ /2. By choice of δ ′ , every A ∈ αn can
intersect at most one Ki . This gives rise to a partition γ = {C1 , . . . ,Cm+1 } where

[ m
[
Ci := {A ∈ αn : A ∩ Ki ̸= ∅} (i = 1, . . . , m) , Cm+1 := X \ Ci .
i=1

Note that Ki ⊂ Ci ⊂ Ui , whence µ(Bi △Ci ) ≤ µ(Ui \ Ki ). Also

m m m
µ(Cm+1 ) ≤ 1 − ∑ µ(Ki ) ≤ 1 − ∑ (µ(Bi ) − µ(Ui \ Ki )) = ∑ µ(Ui \ Ki ) < δ
i=1 i=1 i=1

If we extend β to the partition {B1 , . . . , Bm+1 } where Bm+1 = ∅, then we get

m+1 m
∑ µ(Bi △Ci ) ≤ ∑ µ(Ui \ Ki ) + µ(Cm+1 ) < 2δ .
i=1 i=1

We now use Claim 2 to choose δ so that ∑m+1 i=1 µ(Bi △Ci ) < 2δ ⇒ Hµ (β |γ) < ε.
Since γ ≥ αn , Hµ (β |αn ) < ε for all n > N.
Step 3. Proof of the proposition: Pick nk → ∞ such that hµ (T, αnk ) → lim inf hµ (T, αn ).
For every finite partition β , hµ (T, β ) ≤ hµ (T, αnk ) + Hµ (β |αnk ).
By step 1, Hµ (β |αnk ) −−−→ 0, so hµ (T, β ) ≤ lim infn→∞ hµ (T, αn ). Passing to the
k→∞
supremum over all finite partitions β , we find that

hµ (T ) ≤ lim inf hµ (T, αn ) ≤ lim sup hµ (T, αn ) ≤ hµ (T ),

n→∞ n→∞

which proves the proposition. ⊔

⊓

4.4 Examples

4.4.1 Bernoulli schemes

Proposition 4.6. The entropy of the Bernoulli shift with probability vector p is
− ∑ pi log pi . Thus the ( 31 , 31 , 13 )–Bernoulli scheme and the ( 12 , 21 )–Bernoulli scheme
are not isomorphic.
114 4 Entropy

Proof. The partition into 1-cylinders α = {[i]}
is a strong generator, and Hµ (α0n−1 ) =
− ∑x0 ,...,xn−1 px0 · · · pxn−1 log px0 + · · · log pxn−1 = −n ∑ pi log pi . ⊔
⊓

4.4.2 Irrational rotations

Proposition 4.7. The irrational rotation has entropy zero w.r.t. the Haar measure.

Proof. The reason is that it is an invertible transformation with a strong generator.

We first explain why any invertible map with a strong generator must have zero
entropy. Suppose α is such a strong generator. Then

hµ (T, α) = Hµ (α|α1∞ ) = Hµ (T α|T (α1∞ )) =

= Hµ (T α|α0∞ ) = Hµ (T α|B) = 0, because T α ⊂ B.

We now claim that α := {A0 , A1 } (the two halves of the circle) is a strong generator.
It is enough to show that for every ε, n≥1 α0n−1 contains open covers of the circle
S

by open arcs of diameter < ε (because this forces α0∞ to contain all open sets).
It is enough to manufacture one arc of diameter less than ε, because the transla-
tions of this arc by kα will eventually cover the circle.
But such an arc necessarily exits: Choose some n s.t. nα mod 1 ∈ (0, ε). Then
A1 \ T −n A1 = (A1 \ [A1 − nα] is an arc of diameter less than ε.

4.4.3 Markov measures

Proposition 4.8. Suppose µ is a shift invariant Markov measure with transition ma-
trix P = (pi j )i, j∈S and probability vector (pi )i∈S . Then hµ (σ ) = − ∑ pi pi j log pi j .

Proof. The natural partition α = {[a] : a ∈ S} is a strong generator.

Hµ (α0n ) = − ∑ µ[ξ ] log µ[ξ ]

ξ0 ,...,ξn ∈S


=− ∑ pξ0 pξ0 ξ1 · · · pξn−1 ξn log pξ0 + log pξ0 ξ1 + · · · + log pξn−1 ξn
ξ0 ,...,ξn ∈S
n−1
=−∑ ∑ pξ0 pξ0 ξ1 · · · pξn−1 ξn log pξ j ξ j+1
j=0 ξ0 ,...,ξn ∈S

− ∑ pξ0 pξ0 ξ1 · · · pξn−1 ξn log pξ0

ξ0 ,...,ξn ∈S
4.4 Examples 115

n−1
=−∑ ∑ pξ0 pξ0 ξ1 · · · pξ j−1 ξ j · pξ j+1 ξ j+2 · · · pξn−1 ξn pξ j ξ j+1 log pξ j ξ j+1
j=0 ξ0 ,...,ξn ∈S

− ∑ pξ0 log pξ0

ξ0 ∈S
n−1
=−∑ ∑ µ(σ − j [ξ j ]) · pξ j+1 ξ j+2 · · · pξn−1 ξn pξ j ξ j+1 log pξ j ξ j+1
j=0 ξ j ,...,ξn ∈S

− ∑ pξ0 log pξ0

ξ0 ∈S
n−1
=−∑ ∑ pξ j pξ j ξ j+1 log pξ j ξ j+1 ∑ pξ j+1 ξ j+2 · · · pξn−1 ξn
j=0 ξ j ,ξ j+1 ∈S ξ j+2 ,...,ξn−1 ∈S

− ∑ pξ0 log pξ0

ξ0 ∈S
n−1
=−∑ ∑ pξ j pξ j ξ j+1 log pξ j ξ j+1 − ∑ pξ0 log pξ0
j=0 ξ j ,ξ j+1 ∈S ξ0 ∈S
!
= n − ∑ pi pi j log pi j − ∑ pi log pi
i, j i

Now divide by n + 1 and pass to the limit. ⊔

⊓

4.4.4 Expanding Markov Maps of the Interval

Theorem 4.5 (Rokhlin formula). Suppose T : [0, 1] → [0, 1] and α = {I1 , . . . , IN }

is a partition into intervals s.t.
1. α is a Markov partition
2. The restriction of T to α is C1 , monotonic, and |T ′ | > λ > 1
3. T has an invariant measure µ.
dµ
R
Then hµ (T ) = − log dµ◦T dµ, where (µ ◦ T )(E) = ∑A∈α µ[T (A ∩ E)].

Proof. One checks that the elements of α0n−1 are all intervals of length O(λ −n ).
Therefore α is a strong generator, whence
Z
hµ (T ) = hµ (T, α) = Hµ (α|α1∞ ) = Iµ (α|α1∞ )dµ.

We calculate Iµ (α|α1∞ ).
First note that α1∞ = T −1 (α0∞ ) = T −1 B, thus Iµ (α|α1∞ ) = − ∑ 1A log µ(A|T −1 B).
A∈α
We need to calculate E(·|T −1 B). For this purpose, introduce the operator Tb :
dµ
L → L1 given by (Tb f )(x) = ∑Ty=x dµ◦T
1 (y) f (y).
116 4 Entropy

Exercise: Verify: ∀ϕ ∈ L∞ and f ∈ L1 , ϕ Tb f dµ = ϕ ◦ T · f dµ.

R R

We claim that E( f |T −1 B) = (Tb f ) ◦ T . Indeed, the T −1 B–measurable functions are

exactly the functions of the form ϕ ◦ T with ϕ B–measurable; Therefore (Tb f ) ◦ T
is T −1 B–measurable, and
Z Z Z
ϕ ◦ T · Tb f ◦ T dµ = ϕ · Tb f dµ = ϕ ◦ T · f dµ,

proving the identity.

We can now calculate and see that

Iµ (α|α1∞ ) = − ∑ 1A (x) log E(1A |T −1 B)(x)

A∈α
dµ dµ
=− ∑ 1A (x) log ∑ (y)1A (y) ≡ − ∑ 1A (x) log (x)
A∈α Ty=T x dµ ◦ T A∈α dµ ◦T
dµ
= − log (x).
dµ ◦ T
dµ
R
We conclude that hµ (T ) = − log dµ◦T (x)dµ(x). ⊔
⊓

4.5 Abramov’s Formula

Suppose (X, B, µ, T ) is a ppt. A set A is called spanning, if X = −n A

S∞
n=0 T mod µ.
If T is ergodic, then every set of positive measure is spanning.
Theorem 4.6 (Abramov). Suppose (X, B, µ, T ) is a ppt on a Lebesgue space, let A
be a spanning measurable set, and let (A, BA , µA , TA ) be the induced system, then
1
hµA (TA ) = µ(A) hµ (T ).

Proof. (Scheller) We prove the theorem in the case when T is invertible. The non-
invertible case is handled by passing to the natural extension.
The idea is to show, for as many partitions α as possible, that hµ (T, α) =
µ(A)hµA (TA , α ∩ A), where α ∩ A := {E ∩ A : E ∈ α}. As it turns out, this is the case
for all partitions s.t. (a) Hµ (α) < ∞; (b) Ac ∈ α; and (c) ∀n, TA [ϕA = n] ∈ σ (α).
Here, as always, ϕA (x) := 1A (x) inf{n ≥ 1 : T n x ∈ A} (the first return time).
To see that there are such partitions, we let

ξA := {Ac } ∪ TA ηA , where ηA := {[ϕA = n] : n ∈ N}

(the coarsest possible) and show that Hµ (ξA ) < ∞. A routine calculation shows that
Hµ (ξA ) = Hµ ({A, Ac }) + µ(A)HµA (TA ηA ) ≤ 1 + HµA (ηA ). It is thus enough to show
that − ∑ pn log2 pn < ∞, where pn := µA [ϕA = n]. This is because ∑ npn = 1/µ(A)
4.5 Abramov’s Formula 117

(Kac formula) and the following fact from calculus: probability vectors with finite
expectations have finite entropy.2
Assume now that α is a partition which satisfies (a)–(c) above. We will use
throughout the following fact:

A, Ac , [ϕA = n] ∈ α1∞ . (4.2)

Here is why: [ϕA = n] = T −n TA [ϕA = n] ∈ T −n α ⊂ α1∞ . Since AA = n≥1 [ϕA = n],

S

we automatically have A, Ac ∈ α1∞ .

Let α be a finite entropy countable measurable partition of X such that Ac is an
atom of α and such that α ≥ ξA . In what follows we use the notation A ∩ α :=
{B ∩ A : B ∈ α}, α1∞ ∩ A := {B ∩ A : B ∈ α1∞ }. Since Hµ (α) < ∞,

hµ (T, α) = Hµ (α|α1∞ )
Z Z
= ∑ 1B log µ(B|α1∞ )dµ + 1Ac log µ(Ac |α1∞ )dµ
B∈A∩α
Z
= ∑ 1B log µ(B|α1∞ )dµ, because Ac ∈ (ξA )∞ ∞
1 ⊆ α1
B∈A∩α
Z
= µ(A) ∑ 1B log µA (B|α1∞ )dµA ,
A B∈A∩α

because A ∈ α1∞ and B ⊂ A imply Eµ (1B |F ) = 1A EµA (1B |A ∩ F ).

It follows that hµ (T, α) = µ(A)HµA (A ∩ α|A ∩ α1∞ ). We will show later that
∞
TA−i (A ∩ α)
_
A ∩ α1∞ = (4.3)
i=1

This implies that hµ (T, α) = µ(A)hµA (TA , A ∩ α). Passing to the supremum over all
α which contain Ac as an atom, we obtain

µ(A)hµA (TA ) = sup{hµ (T, α) : α ≥ ξA , Ac ∈ α, Hµ (α) < ∞}

= Entropy of (X, B ′ , µ, T ), B ′ := σ ( {α−∞
[
∞
: Ac ∈ α, Hµ (α) < ∞}).

(See problem 4.11).

Now B ′ = B mod µ, because A is spanning, so ∀E ∈ B, E = ∞ −n (T n E ∩
S
n=0 T
A) mod µ, whence E ∈ B mod µ. This shows Abramov’s formula, given (4.3).
′

The proof of (4.3):

Proof of ⊆: Suppose B is an atom of A ∩ nj=1 T − j α, then B has the form A ∩
W
Tn −j
j=1 T A j where A j ∈ α. Let j1 < j2 < · · · < jN be an enumeration of the j’s s.t.
A j ⊂ A (possibly an empty list). Since Ac is an atom of α, A j = Ac for j not in
2Proof: Enumerate (pn ) in a decreasing order: pn1 ≥ pn2 ≥ · · · . If C = ∑ npn , then C ≥ ∑ki=1 ni pni ≥
pnk (1 + · · · + k), whence pnk = O(k−2 ). Since −x log x = O(x1−ε ) as x → 0+ , this means that
−pnk log pnk = O(k−(2−ε) ), and so − ∑ pn log pn = − ∑ pnk log pnk < ∞.
118 4 Entropy

this list, and so B = N−1 TA−k (A jk ∩ [ϕA = jk+1 − jk ]) ∩ TA−N [ϕA > n − jN ]. Since
T
W∞ k=1
−1
ηA ≤ α ∩ A, B ∈ i=1 TA (α ∩ A).
Proof of ⊇: TA−1 (α ∩ A) ≤ A ∩ −i α,
W∞
i=1 T because if B ∈ α ∩ A, then
∞ ∞
TA−1 B = T −n (B ∩ TA [ϕA = n]) ∈ T −n α (∵ TA ηA ≤ ξA ≤ A ∩ α).
[ _

n=1 n=1

The same proof shows that TA−1 (T −n α ∩ A) ≤ A ∩ −i α.

W∞
i=1 T It follows that
!
∞ ∞ ∞
TA−2 (α ∩ A) ≤ TA−1 A ∩ T −i α TA−1 (A ∩ T −i α) ⊆ A ∩ T −i α.
_ _ _
⊆ A∩
i=1 i=1 i=1

Iterating this procedure we see that TA−n (α ∩ A) ≤ A ∩ −i α

W∞
i=1 T for all n, and ⊇
follows. ⊔
⊓

4.6 Topological Entropy

Suppose T : X → X is a continuous mapping of a compact topological space (X, d).

Such a map can have many different invariant Borel probability measures. For ex-
ample, the left shift on {0, 1}N has an abundance of Bernoulli measures, Markov
measures, and there are many others.
Different measures may have different entropies. What is the largest value possi-
ble? We study this question in the context of continuous maps on topological spaces
which are compact and metric.

4.6.1 The Adler–Konheim–McAndrew definition

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

terminology and notation:
an open cover of X is a collection of open sets U = {Uα : α ∈ Λ } s.t. X =
1. S
α∈Λ Uα ;
2. if U = {Uα : α ∈ Λ } is an open cover, then T −k U := {T −kUα : α ∈ Λ }. Since
T is continuous, this is another open cover.
3. if U , V are open covers, then U ∨ V := {U ∩V : U ∈ U ,V ∈ V }.
Since X is compact, every open cover of X has a finite subcover. Define

N(U ) := min{#V : V ⊆ U is finite, and X = V }.

[

It easy to check that N(·) is subadditive in the following sense:

4.6 Topological Entropy 119

N(U ∨ V ) ≤ N(U ) + N(V ).

Definition 4.5. Suppose T : X → X is a continuous mapping of a compact metric

space (X, d), and let U be an open cover of X. The topological entropy of U is
n−1
1
htop (T, U ) := lim log2 N(U0n−1 ), where U0n−1 := T −k U .
_
n→∞ n
i=0

The limit exists because of the subadditivity of N(·): an := log N(U0n−1 ) satisfies
am+n ≤ am + an , so lim an /n exists.

Definition 4.6. Suppose T : X → X is a continuous mapping of a compact metric

space (X, d), then the topological entropy of T is the (possibly infinite)

htop (T ) := sup{htop (T, U ) : U is an open cover of X}.

The following theorem was first proved by Goodwyn.

Theorem 4.7. Suppose T is a continuous mapping of a compact metric space, then
every invariant Borel probability measure µ satisfies hµ (T ) ≤ htop (T ).

Proof. Eventually everything boils down to the following inequality, which can be
checked using Lagrange multipliers: For every probability vector (p1 , . . . , pk ),
k
− ∑ pi log2 pi ≤ log k, (4.4)
i=1

with equality iff p1 = · · · = pk = 1/k.

Suppose µ is an invariant probability measure, and let α := {A1 , . . . , Ak } be a
measurable partition.
We approximate α by a partition into sets with better topological properties. Fix
ε > 0 (to be determined later), and construct compact sets

B j ⊂ A j s.t. µ(A j \ B j ) < ε ( j = 1, . . . , k).

Let B0 := X \ kj=1 B j be the remainder (of measure less than kε), and define β =
S

{B0 ; B1 , . . . , Bk }.
Step 1 in the proof of Sinai’s theorem says that hµ (T, α) ≤ hµ (T, β ) + Hµ (α|β ).
We claim that Hµ (α|β ) can be made uniformly bounded by a suitable choice of
ε = ε(α):
120 4 Entropy

Hµ (α|β ) = − ∑ ∑ µ(A ∩ B) log2 µ(A|B)

A∈α B∈β

=− ∑ ∑ µ(A ∩ B) log2 µ(A|B) − ∑ µ(A ∩ B0 ) log2 µ(A|B0 )

B∈β \{B0 } A∈α A∈α
k
= ∑ µ(Bi ) log2 1 − ∑ µ(A ∩ B0 ) log2 µ(A|B0 )
i=1 A∈α
= −µ(B0 ) ∑ µ(A|B0 ) log2 µ(A|B0 ) ≤ −µ(B0 ) log(#α) ≤ kε · log2 k.
A∈α

If we choose ε < 1/(k log2 k), then we get Hµ (α|β ) ≤ 1, and

hµ (T, α) ≤ hµ (T, β ) + 1. (4.5)

We now create an open cover from β by setting U := {B0 ∪ B1 , . . . , B0 ∪ Bk }.

This is a cover. To see that it is open note that

B0 ∪ B j = B0 ∪ (A j \ B0 ) (∵ A j ∩ B0 = A j \ B j )
! !
[ [
= B0 ∪ A j = B0 ∪ X \ Ai = B0 ∪ X \ Bi .
i̸= j i̸= j

We compare the number of elements in U0n−1 to the number of elements in β0n−1 .

Every element of U0n−1 is of the form

(B0 ⊎ Bi0 ) ∩ T −1 (B0 ⊎ Bi1 ) ∩ · · · ∩ T −(n−1) (B0 ⊎ Bin−1 ).

This can be written as a pairwise disjoint union of 2n elements of β0n−1 (some of

which may be empty sets). Thus every element of U0n−1 contains at most 2n el-
ements of β0n−1 . Forming the union over a sub cover of U0n−1 with cardinality
N(U0n−1 ), we get that #β0n−1 ≤ 2n N(U0n−1 ).
We no appeal to (4.4): Hµ (β0n−1 ) ≤ log2 (#β0n−1 ) ≤ H(U0n−1 ) + n. Dividing by
n and passing to the limit as n → ∞, we see that hµ (T, β ) ≤ htop (U ) + 1. By (4.5),
hµ (T, α) ≤ htop (U ) + 2 ≤ htop (T ) + 2.
Passing to the supremum over all α, we get that hµ (T ) ≤ htop (T ) + 2, and this
holds for all continuous mappings T and invariant Borel measures µ. In particular,
this holds for T n (note that µ ◦ (T n )−1 = µ): hµ (T n ) ≤ htop (T n ) + 2. But hµ (T n ) =
nhµ (T ) and htop (T n ) = nhtop (T ) (problems 4.4 and 4.13). Thus we get upon division
by n that hµ (T ) ≤ htop (T ) + (2/n) −−−→ 0, which proves the theorem. ⊔
⊓
n→∞

In fact, htop (T ) = sup{hµ (T ) : µ is an invariant Borel probability measure}. But to

prove this we need some more preparations. These are done in the next section.
4.6 Topological Entropy 121

4.6.2 Bowen’s definition

We assume as usual that (X, d) is a compact metric space, and that T : X → X is

continuous. For every n we define a new metric dn on X as follows:

dn (x, y) := max d(T i x, T i y).

0≤i≤n−1

This is called Bowen’s metric. It depends on T . A set F ⊂ X is called (n, ε)–

separated, if for every x, y ∈ F s.t. x ̸= y, dn (x, y) > ε.

Definition 4.7.
1. sn (ε, T ) := max{#(F) : F is (n, ε)–separated}.
2. s(ε, T ) := lim sup n1 log sn (ε, T )
n→∞
3. htop (T ) := lim s(ε, T ).
ε→0+

Theorem 4.8 (Bowen). Suppose T is a continuous mapping of a compact metric

space X, then htop (T ) = htop (T ).

Proof. Suppose U is an open cover all of whose elements have diameters less than
ε. We claim that N(U0n−1 ) ≥ sn (ε, T ) for all n. To see this suppose F is an (n, ε)–
separated set of maximal cardinality. Each x ∈ F is contained in some Ux ∈ U0n−1 .
Since the d–diameter of every element of U is less than δ , the dn –diameter of every
element of U0n−1 is less than δ . Thus the assignment x 7→ Ux is one-to-one, whence

N(U0n−1 ) ≥ sn (ε, T ).

It follows that s(ε, T ) ≤ htop (T, U ) ≤ htop (T ), whence htop (T ) ≤ htop (T ).

To see the other inequality we use Lebesgue numbers: a number δ is called a
Lebesgue number for an open cover U , if for every x ∈ X, the ball with radius δ
and center x is contained in some element of U . (Lebesgue numbers exist because
of compactness.)
Fix ε and let U be an open cover with Lebesgue number bigger than or equal to
ε. It is easy to check that for every n, ε is a Lebesgue number for U0n−1 w.r.t. dn .
Let F be an (n, ε/2)–separated set of maximal cardinality, i.e. #F = sn (ε). Then
any point y we add to F will break its (n, ε)–separation property, and so

∀y ∈ X ∃x ∈ F s.t. dn (x, y) ≤ ε/2.

It follows that the sets Bn (x; ε/2) := {y : dn (x, y) ≤ ε/2} (x ∈ F) cover X.

Every Bn (x, ε/2) (x ∈ F) is contained in some element of U0n−1 , because U0n−1
has Lebesgue number ε w.r.t dn . The union of these elements covers X. We found a
sub cover of U0n−1 of cardinality at most #F = sn (ε). This shows that

N(U0n−1 ) ≤ sn (ε).
122 4 Entropy

We just proved that for every open cover U with Lebesgue number at least ε,
htop (T, U ) ≤ s(ε). It follows that

sup{htop (T, U ) : U has Lebesgue number at least ε} ≤ s(ε).

We now pass to the limit ε → 0+ . The left hand side tends to the supremum over all
open covers, which is equal to htop (T ). We obtain htop (T ) ≤ lim s(ε). ⊔
⊓
ε→0+

Corollary 4.1. Suppose T is an isometry, then all its invariant probability measures
have entropy zero.

Proof. If T is an isometry, then dn = d for all n, therefore s(ε, T ) = 0 for all ε >
0, so htop (T ) = 0. The theorem says that htop (T ) = 0. The corollary follows from
Goodwyn’s theorem. ⊔
⊓

4.6.3 The variational principle

The following theorem was first proved under additional assumptions by Dinaburg,
and then in the general case by Goodman. The proof below is due to Misiurewicz.

Theorem 4.9 (Variational principle). Suppose T : X → X is a continuous map of a

compact metric space, then htop (T ) = sup{hµ (T ) : µ is an invariant Borel measure}.

Proof. We have already seen that the topological entropy is an upper bound for the
metric entropies. We just need to show that this is the least upper bound.
Fix ε, and let Fn be a sequence of (n, ε)–separated sets of maximal cardinality
(so #Fn = sn (ε, T )). Let
1
νn := ∑ δx ,
#Fn x∈F n

where δx denotes the Dirac measure at x (i.e. δx (E) = 1E (x)). These measure are not
invariant, so we set
1 n−1
µn := ∑ νn ◦ T −k .
n k=0
Any weak star limit of µn will be T –invariant (check).
w∗ 1
Fix some sequence nk → ∞ s.t. µnk −−−→ µ and s.t. nk log snk (ε, T ) −−−→ s(ε, T ).
k→∞ n→∞
We show that the entropy of µ is at least s(ε, T ). Since s(ε, T ) −−−→ htop (T ), this
ε→0+
will prove the theorem.
Let α = {A1 , . . . , AN } be a measurable partition of X s.t. (1) diam(Ai ) < ε; and
(2) µ(∂ Ai ) = 0. (Such a partition can be generated from a cover of X by balls of
radius less than ε/2 and boundary of zero measure.) It is easy to see that the dn –
diameter of α0n−1 is also less than ε. It is an exercise to see that every element of
α0n−1 has boundary with measure µ equal to zero.
4.7 Ruelle’s inequality 123

We calculate Hνn (α0n−1 ). Since Fn is (n, ε)–separated and every atom of α has
dn –diameter less than ε, α0n−1 has #Fn elements whose νn measure is 1/#Fn , and the
other elements of α0n−1 have measure zero. Thus

Hνn (α0n−1 ) = log2 #Fn = log2 sn (ε, T ).

We now “play” with Hνn (α0n−1 ) with the aim of bounding it by something in-
q−1
volving a sum of the form ∑n−1
i=0 Hνn ◦T −i (α0 ). Fix q, and j ∈ {0, . . . , q − 1}, then

[n/q]−1
log2 sn (ε, T ) = Hνn (α0n−1 ) ≤ Hνn (α0j−1 ∨ T −(qi+ j) α0q−1 ∨ αq([n/q]−1)+
_
n−1
j+1 )
i=0
[n/q]−1
≤ ∑ Hνn ◦T −(qi+ j) (α0q−1 ) + 2q log2 #α.
i=0

Summing over j = 0, . . . , q − 1, we get

1 n−1
q log2 sn (ε, T ) ≤ n · ∑ Hνn ◦T −k (α0q−1 ) + 2q log2 #α
n k=0
≤ nHµn (α0q−1 ) + 2q log2 #α,

because µn = n1 ∑n−1 −i and φ (t) = −t log t is concave. Thus

i=0 νn ◦ T 2

1 1 2
log2 snk (ε, T ) ≤ Hµnk (α0q−1 ) + log2 #α, (4.6)
nk q nk
where nk → ∞ is the subsequence chosen above.
w∗
Since every A ∈ α0n−1 satisfies µ(∂ A) = 0, µnk (A) −−→ µ(A) for all A ∈ α0n−1 .3
k→ ∞
It follows that Hµnk (α0q−1 ) −−−→ Hµ (α0q−1 ). Passing to the limit k → ∞ in (4.6),
k→∞
we have s(ε, T ) ≤ q1 Hµ (α) −−−→ hµ (T, α) ≤ hµ (T ). Thus hµ (T ) ≥ s(ε, T ). Since
q→∞
s(ε, T ) −−−→ htop (T ) the theorem is proved. ⊔
⊓
ε→0+

4.7 Ruelle’s inequality

Theorem 4.10 (Ruelle). Let f be a C1 diffeomorphism of a compact C2 Rieman-

nian manifold M without boundary. Suppose µ is an f -invariant Borel probability
3
R
FixRδ and sandwich u ≤ 1A\∂ A ≤ 1A ≤ 1A∪∂ A ≤ v with u, v continuous s.t. | udµ − µ(E)| < δ
and | vdµ − µ(E)| R < δ . This Ris possible because A \ ∂ A is open, A ∪ ∂ A is open,
R and µ(∂RA) = 0.
Then µ(A) − δ ≤ udµ = lim udµnk ≤ lim inf µnk (E) ≤ lim sup µnk (E) ≤ lim vdµnk = vdµ ≤
µ(A) + δ . Since δ is arbitrary, µnk (E) −−−→ µ(E).
k→∞
124 4 Entropy

measure with Lyapunov exponents χ1 (x) ≥ · · · ≥ χd (x). Then4

Z
∑ χi (x)dµ ≥ ln 2 · hµ ( f ).
χi (x)>0

Corollary 4.2. Let µ be an ergodic invariant probability measure for a C1 diffeo-

morphism f on a compact smooth manifold without boundary. If hµ ( f ) > 0, then µ
has at least one positive Lyapunov exponent.

In other words, positive entropy implies exponential sensitivity to initial conditions

almost everywhere!

4.7.1 Preliminaries on singular values

√ of an invertible d × d matrix A are the eigenvalues e ≥ · · · ≥

The singular values χ1
χ
e of the matrix A A.
d t

Proposition 4.9. Let eχ1 ≥ · · · ≥ eχd denote the singular values of an invertible ma-
trix A listed with multiplicity. The ellipsoid {Av : v ∈ Rd , ∥v∥2 ≤ 1} can be inscribed
in an orthogonal box with sides 2eχ1 , . . . , 2eχd .
√
Proof. The matrix At A is symmetric and positive definite, therefore it can be di-
agonalized orthogonally. Let u1 , . . . , ud be an orthonormal basis of eigenvectors so
that Aui = eχi ui . Using this basis we can represent
( ) ( )
n o d d
d 2
Av : v ∈ R , ∥v∥2 ≤ 1 ≡ ∑ αi Aui : ∑ αi ≤ 1 ⊂ ∑ αi Aui : |αi | ≤ 1 .
i=1 i=1

The last set is an orthogonal box as requested, because Aui are orthogonal vectors
with lengths eχi : ⟨Aui , Au j ⟩ = ⟨At Aui , u j ⟩ = e2χi ⟨ui , u j ⟩ = e2χi δi j .

Proposition 4.10 (Minimax Principle). eχi (A) = max min ∥Av∥.

V ⊂Rd ,dimV =i v∈V,∥v∥=1
√
Proof. The matrix At A is symmetric and positive definite. Let u1 , . . . , ud be an
orthonomal basis of eigenvectors with eigenvalues eχ1 ≥ · · · ≥ eχd .
Fix i. For every subspace V of dimension i, dim(V ∩ span{ui , . . . , ud }) ≥ 1 by
d d
dimension count. Therefore ∃v ∈ V s.t. v = ∑ α j u j ∈ V , and ∥v∥ = ∑ α 2j = 1. We
j=i j=i
have ∥Av∥2 = ⟨Av, Av⟩ = ⟨At Av, v⟩ = ∑ j≥i e2χ j α 2j ≤ e2χi . This shows that
4 The ln 2 is because we followed the tradition in ergodic theory to use base two logarithms to define
the entropy, and we followed the tradition is dynamical systems to use natural base logarithms to
define Lyapunov exponents. In literature on dynamical systems, it is customary to use natural
logarithms also for the definition of the entropy, and then no such factor is necessary.
4.7 Ruelle’s inequality 125

max min ∥Av∥ ≤ exp χi .

V ⊂Rd ,dimV =i v∈V,∥v∥=1

For the other inequality, consider V0 := span{u1 , . . . , ui }. Every unit vector v ∈ V0

can be put in the form v = ∑ij=1 β j u j with ∑ij=1 β j2 = 1, whence

i i
∥Av∥2 = ⟨At Av, v⟩ = ∑ e2χ j β j2 ≥ e2χi ∑ β j2 = e2χi .
j=1 j=1

The lower bound is achieved for v = ui , so min ∥Av∥ = eχi , proving that
v∈V0 ,∥v∥=1
max min ∥Av∥ ≥ min ∥Av∥ = exp χi . ⊔
⊓
V ⊂Rd ,dimV =i v∈V,∥v∥=1 v∈V0 ,∥v∥=1

4.7.2 Proof of Ruelle’s inequality

Fix k, n ∈ N and ε > 0, to be determined later.

Step 1. A generating sequence of “nice” partitions.
Proof. Let Ek := {y1 , . . . , yηk } denote a maximal ε/k-separated set: “ε/k-separation”
means that i ̸= j ⇒ d(yi , y j ) > ε/k; “maximality” means that ∀x ∈ M∃i s.t. d(x, yi ) ≤
ε/k. Such a set exists by the compactness of M.
Each yi ∈ Ek defines a Dirichlet domain

A′i := {x ∈ M : d(x, yi ) ≤ d(x, y j ) for all j}.

The following properties are easy to verify:

1. A′i cover M
2. A′i have disjoint interiors and contain the pairwise disjoint balls B(yi , ε/2k)
3. A′i ⊂ B(yi , ε/k), otherwise we can find x ∈ A′i such that for all j, d(x, y j ) ≥
d(x, yi ) > ε/k, in contradiction to the maximality of Ek
4. diam(A′i ) ≤ 2ε/k
Let A1 := A′1 , A′i := Ai \ i−1 ′
j=1 A j . Then αk := {A1 , . . . , Aηk } is a partition of M
S

such that diam(αk ) −−−→ 0. So hµ (T, αk ) −−−→ hµ (T, α). Note that it is still the case
k→∞ k→∞
that B(yi , ε/k) ⊂ Ai ⊂ B(yi , 2ε/2k).
Step 2. hµ ( f , αk ) ≤ ∑A∈αk µ(A)( 1n log2 Kn (A, k)), where

Kn (A, k) := #{A′ ∈ αk : A′ ∩ f n (A) ̸= ∅}

Proof. hµ ( f , αk ) = 1n hµ ( f n , αk ) = 1n hµ ( f −n , αk ), due to exercises 4.4 and 4.5. By

Rokhlin’s formula,
126 4 Entropy
∞
hµ ( f −n , αk ) = Hµ (αk | f + jn αk ) ≤ Hµ (αk | f n αk ) = Hµ ( f −n αk |αk ).
_

j=1

Hµ ( f −n αk |αk ) = − ∑A∈Ak µ(A)(∑B∈ f −n αk µ(B|A) log2 µ(B|A)). The inner sum is

bounded by log2 #{B ∈ αk : µ( f −n (B) ∩ A) > 0} ≡ log Kn (A, k).
Step 3. There exists a constant C which depends on M but not on ε, n, k such that for
all n large enough, for all k large enough
d
Kn (A, k) ≤ C ∏(1 + si (DFyn ))
i=1

for all y ∈ A, where si (DFyn ) denote the singular values of the derivative matrix of
f n at y, expressed in coordinates.
We first explain the method on the heuristic level, and then develop the rigorous
implementation. Suppose n is fixed.
Observe that f n (A) has diameter at most Lip( f n )diamαk ≤ Lip( f n )2ε/k, and
therefore f n (A) and all the A′ which intersect it are included in a ball B of ra-
dius (Lip( f n ) + 1)ε/k). If ε is very small, this ball lies inside a single coordinate
chart ψ1 (Rd ). Similarly, A lies inside a ball of radius diamαk ≤ 2ε/k, which for ε
sufficiently small lies inside a coordinate chart ψ2 (Rd ). This allows us to express
f n | f −n (B) in coordinates: F n := ψ2−1 ◦ f n ◦ ψ1 : Rd → Rd .
If ε is sufficiently small, F n is very close to its linearization. Suppose for the
purpose of the explanation that F n is precisely equal to its linearization on f −n (B).
From now on we abuse notation and identify f n with its derivative matrix DF n in
coordinates.
Since A is contained in a ball of radius ε/k, DF n (A) is contained in a box with
orthogonal sides whose lengths are 2(ε/k)×the singular values
n n
eχ1 (DF ) ≥ · · · ≥ eχd (DF ) .

The A′ which intersect f n (A) are contained in an ε/k–neighborhood of this box, a

set whose (Lebesgue) volume is bounded above by
 d d
2ε n

k ∏(eχi (DF ) + 1)
i=1

By construction, A′i contain pairwise disjoint balls of radius ε/2k, and volume
n
const. (ε/k)d . Thus their number is at most const. ∏χi (DF n )≥0 (eχi (DF ) + 1), where
the constant comes from the volume of the unit ball in Rd .
We now explain how to implement this idea when f n is not linear. The details are
tedious, but are all routine.
We begin by recalling some facts from differential topology. Recall that M is a
compact smooth Riemannian manifold of dimension d. Let Tx M denote the tangent
space at x, and let T M denote the tangent bundle. Since M is a Riemannian manifold,
4.7 Ruelle’s inequality 127

every Tx M comes equipped with an inner product ⟨·, ·⟩x and a norm ∥ · ∥x . We denote
tangent vectors in Tx M by ⃗v, and vectors in Rd by v.
Let expx : Tx M → M denote the exponential map, defined by expx (⃗0) := x and

γ⃗v (t) is a geodesic with

expx (⃗v) = γ⃗v (∥⃗v∥), where
beginning x, direction ⃗v and speed 1.

It is a basic fact in differential topology, that since M is compact and C2 -smooth,

there are constants κ1 , κ2 > 0 s.t. for all x ∈ M,
1. (x,⃗v) 7→ expx (⃗v) is a smooth map T M → M
2. expx |B(⃗0,κ1 ) is a C2 diffeomorphism onto its image, and expx |B(⃗0,κ1 ) ⊃ B(x, κ2 )
3. the derivative of expx at 0 is id : Tx M → Tx M
4. for all ⃗v1 ,⃗v2 ∈ B(⃗0, κ1 ), 12 ∥⃗v1 −⃗v2 ∥x ≤ d(expx (⃗v1 ), expx (⃗v2 )) ≤ 2∥⃗v1 −⃗v2 ∥x
It is also a standard procedure in differential topology to cover M by a fi-
nite collection balls B(xi , κ2 ) ⊂ M of radius κ2 together with a continuous map
Θ : T B(xi , κ2 ) → Rd such that Θ |Tx M : Tx M → Rd is a linear isometry. If we glue
these identifications we can obtain a measurable (even piecewise continuous) map

ϑ : T M → Rd s.t. ϑx := ϑ |Tx M : Tx M → Rd is a linear isometry.

Now fix n ∈ N, x ∈ M, and choose κ3 < κ1 /(4Lip( f n )) where Lip( f n ) is the

Lipschitz constant of f n . The following map is well defined on {t ∈ Rd : ∥t∥ < κ3 }:

Fxn := ϑ f n (x) ◦ exp−1 n −1

f n (x) ◦ f ◦ expx ◦ϑx

Notice that Fxn maps a neighborhood of 0 ∈ Rd onto a neighborhood of 0 ∈ Rd ,

Fxn (0) = 0, and that the derivative at 0 is

(DFxn )0 = ϑ f n (x) ◦ (D f n )x ◦ ϑx−1 ∈ GL(d, R).

Here we identified the linear maps ϑx : Tx M → Rd with their derivatives at ⃗0, and
(D f n )x : Tx M → T f n (x) M is the derivative of f n at x.
Claim. For every ε > 0 there is a κ < κ3 such that for all x ∈ M and ∥t∥ < κ
d(Fxn (t), (DFxn )0t) < ε∥t∥.
Proof. Write G = Fxn , fix t, and let gt (s) = G(st) ∈ Rd . Since gt (0) = Fxn (0) = 0,
Z 1
∥G(t) − DG0t∥ = ∥gt (1) − gt (0) − gt′ (0)∥ ≤ gt′ (s)ds − gt′ (0)
0
Z 1
≤ ∥gt′ (s) − gt′ (0)∥ds ≤ ∥t∥ sup ∥(DFxn )s1 − (DFxn )s2 ∥
0 ∥s1 −s2 ∥≤∥t∥

Since f is C1 and ϑ f n (x) , ϑx are isometries, the supremum can be made smaller than
ε by choosing κ sufficiently small.
128 4 Entropy

Claim. Let si (D fxn ) := i-th singular value of ϑ f n (x) D fxn ϑx−1 , then there exists κ4 such
that |si (D fyn ) − si (D fzn )| < 1 for all d(y, z) < κ4 .
Proof. By the minimax theorem the fact that ϑx , ϑ f n (x) are isometries,

si (D fxn ) = max min ∥(D fxn )⃗v∥

V ⊂Tx M,dimV =i ⃗v∈V,∥⃗v∥x =1

The C1 –smoothness of f implies that (x,⃗v) 7→ ∥(D f n )x⃗v∥ f n (x) is uniformly continu-
ous. From this it is routine to deduce that si (D fxn ) is uniformly continuous.
We can now begin the estimation of Kn (A, k). Recall that A is contained in some
ball B(yi , ε/k) of radius ε/k. Therefore

f n (A) ⊂ B( f n (yi ), Lip( f n )ε/k).

So every A′ ∈ αk which intersects f n (A) satisfies

A′ ⊂ B( f n (yi ), (Lip( f n ) + 1)ε/k)

Assume k is so large that (Lip( f n ) + 1)ε/k < κ3 . Using the bound Lip(exp−1
x )≤
2, we can write A′ and A in coordinates as follows:

A = (expyi ◦ϑy−1
i
)(B), where B ⊂ B(0, 2ε/k) ⊂ Rd
A′ = (exp f n (yi ) ◦ϑ f−1 ′ ′ ′
n (y ) )(B ), where diam(B ) ≤ 2diam(A ) ≤ 4ε/k.
i

Since A′ ∩ f n (A) ̸= ∅ and f n (A) = (exp f n (yi ) ◦ϑ f−1 n

n (y ) )(Fyi (B)), it must be the case
i
that B′ ∩ Fyni (B) ̸= ∅, whence since diam(B′ ) ≤ 4ε/k

B′ ⊂ Nρ (Fyni (B)) := {x : d(x, Fyni (B)) < ρ}

where ρ := 4ε/k.
Claim. Choose k so large that 10ε/k < 1 and ε < 1/2. Then the Lebesgue volume
of Nρ (Fyni (B)) is less than 100d (ε/k)d ∏di=1 (1 + si (ϑ f−1 n
n (y ) D f yi ϑyi )).
i

Proof. B ⊂ B(0, 2ε/k), therefore

 
orthogonal box whose sides
(DFyni )0 (B) ⊂ =: Box
have lengths 2 · (2ε/k) · si [(DFyni )0 ]

Since d(Fyni (t), (DFyni )0t) ≤ ε∥t∥ < ε

k on B(0, 2ε/k), Fyni (B) ⊂ Nε/k (Box). Thus
 
orthogonal box whose sides
Nρ (Fyni (B)) ⊂ .
have lengths (100ε/k)(1 + si [(DFyni )0 ])

(100 is an over estimate).

We can now bound Kn (A, k): Each A′ contains a ball of radius ε/2k. Since
exp−1
f n (y )
has Lipschitz constant ≤ 2, each B′ contains a ball of radius ε/4k. These
i
4.7 Ruelle’s inequality 129

balls are disjoint, because the A′ are disjoint. We find that Kn (A, k) is bounded by
the maximal number of pairwise disjoint balls of radius ε/4k which can be fit inside
a set of volume 100d (ε/k)d ∏di=1 (1 + si (ϑ f−1 n
n (y ) D f yi ϑyi )). Thus
i

d
Kn (A, k) ≤ C ∏[1 + si (D fyni )],
i=1

where C is independent of ε, n, k: The (ε/k)d term cancels out!

Proof of Ruelle’s inequality. Steps 2 and 3 give us the inequality
1
hµ ( f , αk ) ≤ ∑ µ(A) log2 Kn (A, k)
n A∈α
k
d
1
≤ ∑ µ(A) ∑ log2 [1 + si (D fyni )] + logC for all k large enough
A∈αk i=1 n
d
1 1
Z
≤∑ log2 [1 + si (D fyn )]dµ(y) + logC,
i=1 n n

where the last estimate is because A ⊂ B(yi , ε/k) and

1
|si (D fyn ) − si (D fyni )| < for all y ∈ A ⊂ B(yi , ε/k).
2
Passing to the limit k → ∞ gives us that
d
1 1
Z
hµ ( f ) ≤ ∑ log2 [1 + si (D fyn )]dµ(y) + logC,
i=1 n n

d
1
Z
whence hµ ( f ) ≤ lim log2 [1 + si (D fyn )]dµ(y).
∑
n→∞
i=1 n
We analyze the last limit. Let A(x) = ϑ f (x) D fx ϑx−1 , then

A( f n−1 (x)) · · · A( f (x))A(x) = ϑ f n (x) D fxn ϑx−1

and therefore si (D fyn ) = i-th singular value of A( f n−1 (x)) · · · A( f (x))A(x). By the
multiplicative ergodic theorem, si (x)1/n −−−→ eχi (x) , where χ1 (x) ≥ · · · ≥ χd (x) are
n→∞
the Lyapunov exponents of A(x). Necessarily,
(
1 χi (x)/ ln 2 χi (x) > 0
log2 [1 + si (D fyn )] −−−→
n n→∞ 0 χi (x) ≤ 0,

where the ln 2 factor is due to the identity log2 (eχi ) = χi / ln 2.

130 4 Entropy

Since 1n log2 [1+si (D fyn )] is uniformly bounded (by (ln 2)−1 supy ∥D fy ∥), we have
by the bounded convergence theorem that
d
1 1
Z Z
hµ ( f ) ≤ ∑ lim log2 [1 + si (D fyn )]dµ(y) = ∑ χi (x)dµ,
i=1
n→∞ n ln 2 i:χi (x)>0

and we have Ruelle’s inequality. □

Problems
µ(A∩B)
4.1. Prove: Hµ (α|β ) = − ∑ µ(B) ∑ µ(A|B) log µ(A|B), where µ(A|B) = µ(B) .
B∈β A∈α

4.2. Prove: if Hµ (α|β ) = 0, then α ⊆ β mod µ.

4.3. Prove that hµ (T ) is an invariant of measure theoretic isomorphism.
4.4. Prove that hµ (T n ) = nhµ (T ).
4.5. Prove that if T is invertible, then hµ (T −1 ) = hµ (T ).
4.6. Entropy is affine
Let T be a measurable map on X, and µ1 , µ2 be two T –invariant probability mea-
sures. Set µ = tµ1 + (1 −t)µ2 (0 ≤ t ≤ 1). Show: hµ (T ) = thµ1 (T ) + (1 −t)hµ2 (T ).
Guidance: Start by showing that for all 0 ≤ x, y,t ≤ 1,

0 ≤ ϕ(tx + (1 − t)y) − [tϕ(x) + (1 − t)ϕ(y)] ≤ −tx logt − (1 − t)y log(1 − t)

4.7. Let (X, B, µ) be a probability space. If α, β are two measurable partitions of

X, then we write α = β mod µ if α = {A1 , . . . , An } and B = {B1 , . . . , Bn } where
µ(Ai △Bi ) = 0 for all i. Let P denote the set of all countable measurable partitions
of X, modulo the equivalence relation α = β mod µ. Show that

ρ(α, β ) := Hµ (α|β ) + Hµ (β |α)

induces a metric on P.
4.8. Let (X, B, µ, T ) be a ppt. Show that |hµ (T, α) − hµ (T, β )| ≤ Hµ (α|β ) +
Hµ (β |α).
4.9. Use the previous problem to show that

hµ (T ) = sup{hµ (T, α) : α finite measurable partition}

4.10. Suppose α = {A1 , . . . , An } is a finite measurable partition. Show that for every
ε, there exists δ = δ (ε, n) such that if β = {B1 , . . . , Bn } is measurable partition s.t.
µ(Ai △Bi ) < δ , then ρ(α, β ) < ε.
References 131

4.11. Entropy via generating sequences of partitions

Suppose (X, B, µ) is a probability space, and A is an algebra of F –measurable
subsets (namely a collection of sets which contains ∅ and which is closed under
finite unions, finite intersection, and forming complements). Suppose A generates
B (i.e. B is the smallest σ –algebra which contains A ).
1. For every F ∈ F and ε > 0, there exists A ∈ A s.t. µ(A△F) < ε.
2. For every F –measurable finite partition β and ε > 0, there exists an A –
measurable finite partition α s.t. ρ(α, β ) < ε.
3. If T : X → X is probability preserving, then

hµ (T ) = sup{hµ (T, α) : α is an A –measurable finite partition}.

4. Suppose α1 ≤ α ≤ · · · is an increasing sequence of finite measurable partitions

S 2
such that σ ( n≥1 αn ) = B mod µ, then hµ (T ) = lim hµ (T, αn ).
n→∞

4.12. Show that the entropy of the product of two ppt is the sum of their two en-
tropies.
4.13. Show that htop (T n ) = nhtop (T ).

Notes to chapter 4

The notion of entropy as a measure of information is due to Shannon, the father

information theory. Kolmogorov had the idea to adapt this notion to the ergodic the-
oretic context for the purposes of inventing an invariant which is able to distinguish
Bernoulli schemes. This became possible once Sinai has proved his generator theo-
rem — which enables the calculation of this invariant for Bernoulli schemes. Later,
in the 1970’s, Ornstein has proved that entropy is a complete invariant for Bernoulli
schemes: they are isomorphic iff they have the same entropy. The maximum of the
possible entropies for a topological Markov shift was first calculated by Parry, who
also found the maximizing measure. The material in this chapter is mostly classical,
[5] and [1] are both excellent references. For an introduction to Ornstein’s isomor-
phism theorem, see [4]. Our proof of Ruelle’s inequality is based on [3] and [2].

References

1. Petersen, K.: Ergodic theory. Corrected reprint of the 1983 original. Cambridge Studies in
Advanced Mathematics, 2. Cambridge University Press, Cambridge, 1989. xii+329 pp.
2. Katok, A. and Hasselblat, B.: Introduction to the modern theory of dynamical systems, with a
supplement by A. Katok and L. Mendoza, Encyclopedia of Mathematics and its Applications
54, Cambridge University Press, 1995, xviii+802.
3. Min Qian, Jian-Sheng Xie, Shu Zhu: Smooth ergodic theory for endomorphisms, LNM 1978
(2009)
132 4 Entropy

4. Rudolph, D.: Fundamentals of measurable dynamics: Ergodic theory on Lebesgue spaces,

Oxford Science Publications, 1990. x+168 pp.
5. Walters, P.: An introduction to ergodic theory. Graduate Texts in Mathematics, 79. Springer-
Verlag, New York-Berlin, 1982. ix+250 pp.
Appendix A
The isomorphism theorem for standard measure
spaces

A.1 Polish spaces

Definition A.1. A polish space is a metric space (X, d) which is

1. complete (every Cauchy sequence has a limit);
2. and separable (there is a countable dense subset).

Every compact metric space is polish. But a polish space need not be compact, or
even locally compact. For example,

NN := {x = (x1 , x2 , x3 , . . .) : xk ∈ N}

equipped with the metric d(x, y) := ∑k≥1 2−k |xk−1 − y−1

k | is a non locally compact
polish metric space.
Notation: B(x, r):={y ∈ X : d(x, y) < r} (the open ball with center x and radius r).

Proposition A.1. Suppose (X, d) is a polish space, then

1. Second axiom of countability: There exists a countable family of open sets U
such that every open set in X is a union of a subfamily of U .
2. Lindelöf property: Every cover of X by open sets has a countable sub-cover.
3. The intersection of any decreasing sequence of closed balls whose radii tend to
zero is a single point.

Proof. Since X is separable, it contains a countable dense set {xn }n≥1 . Define

U := {B(xn , r) : n ∈ N, r ∈ Q}.

This is a countable collection, and we claim that it satisfies (1). Take some open set
U. For every x ∈ U there are
1. R > 0 such that B(x, R) ⊂ U (because U is open);
2. xn ∈ B(x, R/2) (because {xn } is dense); and

133
134 A The isomorphism theorem for standard measure spaces

3. r ∈ Q such that d(x, xn ) < r < R/2.

It is easy to check that x ∈ B(xn , r) ⊂ B(x, 2r) ⊂ B(x, R) ⊂ U. Thus for every x ∈ U
there is Ux ∈ U s.t. x ∈ Ux ⊂ U. It follows that U is a union of elements from U
and (1) is proved. (2) is an immediate consequence.
To see (3), suppose Bn := B(zn , rn ) is a sequence of closed balls such that Bn ⊃
Bn+1 and rn → 0. It is easy to verify that {zn } is a Cauchy sequence. Since X is
complete, it converges to a limit z. This limit belongs to Bn because zk ∈ Bk ⊂ Bn for
all k > n, and Bn is closed. Thus the intersection of Bn contains at least one point.
It cannot contain more than one point, because its diameter is zero (because it is
bounded by diam[Bn ] ≤ rn → 0). ⊔
⊓

A.2 Standard probability spaces

Definition A.2. A standard probability space is a probability space (X, B, µ) where

X is polish, B is the σ –algebra of Borel sets of X, and µ is a Borel probability
measure.
Theorem A.1. Suppose (X, B, µ) is a standard probability space, then
1. Regularity: Suppose E ∈ B. For every ε > 0 there exists an open set U and a
closed set F such that F ⊂ E ⊂ U and µ(U \ F) < ε.
2. Separability: There exists a countable collection of measurable sets {En }n≥1
such that for every E ∈ B and ε > 0 there exists some n s.t. µ(E△En ) < ε.
Equivalently, L p (X, B, µ) is separable for some (and then for all) 1 ≤ p < ∞.
Proof. Say that a set E satisfies the approximation property if for every ε there are
a closed set F and an open set U s.t. F ⊂ E ⊂ U and µ(U \ E) < ε.
Open balls B(x, r) have the approximation property: Take U = B(x, r) and F =
B(x, r − 1n ) for n sufficiently large (these sets increase to B(x, r) so their measure
tends to that of B(x, r)).
Open sets U have the approximation property. The approximating open set is
the set itself. To find the approximating closed set use the second axiom of count-
ability to write the open set as the countable union of balls Bn , and approximate
each BSn from within by a closed setSFn such that µ(Bn \ Fn ) < ε/2n+1 . Then
µ(U \ n≥1 Fn ) < ε/2. Now take F := Ni=1 Fi for N large enough.
Thus the collection C := {E ∈ B : E has the approximation property} contains
the open sets. Since it is a σ –algebra (check!), it must be equal to B, proving (1).
We prove separability. Polish spaces satisfy the second axiom of countability, so
there is a countable family of open balls U = {Bn : n ∈ N} such that every open
set is the union of a countable subfamily of U . This means that every open set
can be approximated by a finite union of elements of U to arbitrary precision. By
the regularity property shown above, every measurable set can be approximated by
a finite union of elements of U to arbitrary precision. It remains to observe that
E = {finite unions of elements of U } is countable.
A.3 Atoms 135

The separability of L p for 1 ≤ p < ∞ follows from the above and the obvious fact
that the collection {∑Ni=1 αi 1Ei : N ∈ N, αi ∈ Q, Ei ∈ E } is dense in L p (prove!). The
other direction is left to the reader. ⊔
⊓

The following statement will be used in the proof of the isomorphism theorem.
Lemma A.1. Suppose (X, B, µ) is a standard probability space and E is a measur-
able set of positive measure, then there is a point x ∈ X such that µ[E ∩ B(x, r)] ̸= 0
for all r > 0.

Proof. Fix εn ↓ 0. Write X as a countable union of open balls of radius ε1 (second

axiom of countability). At least one of these, B1 , satisfies µ(E ∩ B1 ) ̸= 0. Write B1
as a countable union of open balls of radius ε2 . At least one of these, B2 , satisfies
µ[E ∩ B1 ∩ B2 ] ̸= 0. Continue in this manner. The result is a decreasing sequence
of open balls with shrinking diameters B1 ⊃ B2 ⊃ · · · which intersect E at a set of
positive measure.
The sequence of centers of these balls is a Cauchy sequence. Since X is polish, it
converges to a limit x ∈ X. This x belongs to the closure of each Bn .
For every r find n so large that εn < r/2. Since x ∈ Bn , d(x, xn ) ≤ εn , and this
implies that B(x, r) ⊇ B(xn , εn ) = Bn . Since Bn intersects E with positive measure,
B(x, r) intersects E with positive measure. ⊔
⊓

A.3 Atoms

Definition A.3. An atom of a measure space (X, B, µ) is a measurable set A of non-

zero measure with the property that for all other measurable sets B contained in A,
either µ(B) = µ(A) or µ(B) = 0. A measure space is called non-atomic, if it has no
atoms.

Proposition A.2. For standard spaces (X, B, µ), every atom is of the form {x}∪null
set for some x s.t. µ{x} ̸= 0.

Proof. Suppose A is an atom. SinceS

X can be covered by a countable collection of
open balls of radius r1 := 1, A = i≥1 Ai where Ai are measurable subsets of A of
diameter at most r1 . One of those sets, Ai1 , has non-zero measure. Since A is an
atom, µ(Ai1 ) = µ(A). Setting A(1) := Ai1 , we see that

A(1) ⊂ A , diam(A(1) ) ≤ r1 , µ(A(1) ) = µ(A).

Of course A(1) is an atom.

Now repeat this argument with A(1) replacing A and r2 := 1/2 replacing r1 . We
obtain an atom A(2) s.t.

A(2) ⊂ A , diam(A(2) ) ≤ r2 , µ(A(2) ) = µ(A).

136 A The isomorphism theorem for standard measure spaces

We continue in this manner, to obtain a sequence of atoms A ⊃ A(1)T⊃ A(2) ⊃ · · ·

of the same measure, with diameters rk = 1/k → 0. The intersection A(k) is non-
empty, because its measure is lim µ(A(k) ) = µ(A) ̸= 0. But its diameter is zero.
Therefore it is a single point x, and by construction x ∈ A and µ{x} = µ(A). ⊔
⊓

Lemma A.2. Suppose (X, B, µ) is a non-atomic standard probability space, and

r > 0. Every B–measurable set E can be written in the form E = ∞
i=1 i ⊎ N where
U
F
µ(N) = 0, Fi are closed, diam(Fi ) < r, and µ(Fi ) ̸= 0.

Proof. Since every measurable set is a finite or countable disjoint union of sets of
diameter less than r (prove!), it is enough to treat sets E such that diam(E) < r.
Standard spaces are regular, so we can find a closed set F1 ⊂ E such that µ(E \
F1 ) < 12 . If µ(E \ F1 ) = 0, then stop. Otherwise apply the argument to E \ F1 to
find a closed set F2 ⊂ E \ F1 of positive measure such that µ[E \ (F1 ∪ F2 )] < 212 .
Continuing in this manner we obtain pairwise disjoint closed sets Fi ⊂ E such that
µ(E \ ni=1 Fi ) < 2−n for all n, or until we get to an n such that µ(E \ ni=1 Fn ) = 0.
S U

If the procedure did not stop at any stage, then the lemma follows with N :=
E \ i≥1 Fi .
S

We show what to do in case the procedure stops after n steps. Set F = Fn , the last
closed set. The idea is to split F into countably many disjoint closed sets, plus a set
of measure zero.
Find an x such that µ[F ∩ B(x, r)] ̸= 0 for all r > 0 (previous lemma). Since X is
non-atomic, µ{x} = 0. Since B(x, r) ↓ {x}, µ[F ∩ B(x, r)] −−−→ 0. Choose rn ↓ 0 for
n→∞
which µ[F ∩ B(x, rn )] is strictly decreasing. Define

C1 := F ∩ B(x, r1 ) \ B(x, r2 ), C2 := F ∩ B(x, r2 ) \ B(x, r3 ) and so on.

This is an infinite sequence of closed pairwise disjoint sets of positive measure in-
side F. By the construction of F they are disjoint from F1 , . . . , Fn−1 and they are
contained in E.
Now consider E ′ := E \( n−1 i=0 Fi ∪ i Ci ). Applying the argument in the first para-
S S
′
graph to E , we write it as a finite or countable disjoint union of closed sets plus a
null set. Adding these sets to the collection {F1 , . . . , Fn−1 } ∪ {Ci : i ≥ 1} gives us the
required decomposition of E. ⊔
⊓

A.4 The isomorphism theorem

Definition A.4. Two measure spaces (Xi , Bi , µi ) (i = 1, 2) are called isomorphic if

there are measurable subsets of full measure Xi′ ⊂ Xi and a measurable bijection
π : X1′ → X2′ with measurable inverse such that µ2 = µ1 ◦ π −1 .

Theorem A.2 (Isomorphism theorem). Every non-atomic standard probability

space is isomorphic to the unit interval equipped with the Lebesgue measure.
A.4 The isomorphism theorem 137

Proof. Fix a decreasing sequence of positive numbers εn which tend to zero. Using
lemma A.2, decompose X = ∞j=1 F( j) ⊎ N where F( j) are pairwise disjoint closed
U

sets of positive measure and diameter less than ε1 , and N1 is a null set.
Applying lemma A.2 to each F( j), decompose F( j) = ∞
U
k=1 F( j, k)⊎N( j) where
F( j, k) are pairwise disjoint closed sets of positive measure and diameter less than
ε2 , and N( j) is a null set.
Continuing in this way we obtain a family of sets F(x1 , . . . , xn ), N(x1 , . . . , xn ),
(n, x1 , . . . , xn ∈ N) such that
1. F(x1 , . . . , xn ) are closed, have positive measure, and diam[F(x1 , . . . , xn )] < εn ;
2. F(x1 , . . . , xn−1 ) = y∈N F(x1 , . . . , xn−1 , y) ⊎ N(x1 , . . . , xn−1 );
U

3. µ[N(x1 , . . . , xn )] = 0.
Set X ′ := n≥1 x1 ,...,xn ∈N F(x1 , . . . , xn ). It is a calculation to see that µ(X \ X ′ ) =
T U

0. The set X ′ has tree-like structure: every x ∈ X ′ determines a unique sequence

(x1 , x2 , . . .) ∈ NN such that x ∈ F(x1 , . . . , xn ) for all n. Define π : X ′ → [0, 1] by

1
π(x) =
1
x1 +
x2 + · · ·

This map is one-to-one on X ′ , because if π(x) = π(y), then 1/(x1 + 1/(x2 +

· · · )) = 1/(y2 + 1/(y2 + · · · )) whence xk = yk for all k;1 this means that x, y ∈
F(x1 , . . . xn ) for all n, whence d(x, y) ≤ εn −−−→ 0.
n→∞
This map is onto [0, 1] \ Q, because every irrational t ∈ [0, 1] has an infinite con-
tinued
T
fraction expansion 1/(a1 + 1/(a2 + · · · )), so t = π(x) for the unique x in
n≥1 F(a 1 , . . . , an ). (This intersection is non-empty because it is the decreasing in-
tersection of closed sets of shrinking diameters in a complete metric space.)
We claim that π : X ′ → [0, 1] \ Q is Borel measurable. Let [a1 , . . . , an ] de-
note the collection of all irrationals in [0, 1] whose continued fraction expansion
starts with (a1 , . . . , an ). We call such sets “cylinders.” We have π −1 [a1 , . . . , an ] =
F(a1 , . . . , an ) ∩ X ′ , a Borel set. Thus

C := {E ∈ B([0, 1] \ Q) : π −1 (E) ∈ B}

contains the cylinders. It is easy to check that C is a σ –algebra. The cylinders gener-
ate B([0, 1]\ Q) (these are intervals whose length tends to zero as n → ∞). It follows
that C = B([0, 1] \ Q) and the measurability of π is proved.
Next we observe that π[F(a1 , . . . , an ) ∩ X ′ ] = [a1 , . . . , an ], so π −1 : [0, 1] \ Q → X ′
is Borel measurable by an argument similar to the one in the previous paragraph.
It follows that π : (X, B, µ) → ([0, 1] \ Q, B([0, 1] \ Q), µ ◦ π −1 ) is an isomor-
phism of measure spaces. There is an obvious extension of µ ◦ π −1 to B([0, 1])
obtained by declaring µ(Q) := 0. Let m denote this extension. Then we get an
isomorphism between (X, B, µ) to ([0, 1], B([0, 1]), m) where is m is some Borel
probability measure on [0, 1]. Since µ is non-atomic, m is non-atomic.
1 Hint: apply the transformation x 7→ [1/x] to both sides.
138 A The isomorphism theorem for standard measure spaces

We now claim that ([0, 1], B([0, 1]), m) is isomorphic to ([0, 1], B([0, 1]), λ ),
where λ is the Lebesgue measure.
Consider first the distribution function of m, s 7→ m[0, s). This is a monotone
increasing function (in the weak sense). We claim that it is continuous. Otherwise it
has a jump J at some point x0 :

m[0, x0 + ε) − m[0, x0 ) > J for all ε > 0.

This means that m{x0 } ≥ J, which cannot be the case since m is non-atomic.
Since Fm (s) = m([0, s)) is continuous, and Fm (0) = 0, Fm (1) = m([0, 1])−m{1} =
1, Fm (s) attains any real value 0 ≤ t ≤ 1. So the following definition makes sense:

ϑ (t) := min{s ≥ 0 : m([0, s)) = t} (0 ≤ t ≤ 1)

Notice that Fm (ϑ (t)) = t. We will show that ϑ : ([0, 1], B, m) → ([0, 1], B, λ ) is an
isomorphism of measure spaces.
Step 1. m([0, 1] \ ϑ [0, 1]) = 0.
Proof. s ∈ ϑ [0, 1] iff s = ϑ (Fm (s)) iff ∀s′ < s, m[s′ , s] > 0.
Conversely, s ̸∈ ϑ [0, 1], iff s belongs to an interval with positive length and zero
m measure. Let I(s) denote the union of all such intervals. This is a closed interval
(because m is non-atomic), m(I(s)) = 0, λ (I(s)) > 0, and for any s1 , s2 ∈ [0, 1] \
ϑ [0, 1], either I(s1 ) = I(s2 ) or I(s1 ) ∩ I(s2 ) = ∅.
There can be at most countably many different intervals I(s). It follows that
[0, 1] \ ϑ [0, 1] is a finite or countable union of closed intervals with zero m-measure.
In particular, m([0, 1] \ ϑ [0, 1]) = 0.
Step 2. ϑ is one-to-one on [0, 1].
Proof: Fm ◦ ϑ = id
Step 3. ϑ is measurable, with measurable inverse.
Proof. ϑ is measurable, because for every interval (a, b), ϑ −1 (a, b) = {t : t =
Fm (ϑ (t)) ∈ Fm (a, b)} = (Fm (a), Fm (b)), a Borel set.
To see that ϑ −1 is measurable, note that ϑ is strictly increasing, therefore
ϑ (a, b) = (ϑ (a), ϑ (b)) ∩ ϑ ([0, 1]). This set is Borel, because as we saw in the proof
of step 1, ϑ ([0, 1]) is the complement of a countable collection of intervals.
Step 4. m ◦ ϑ = λ .
Proof. By construction, m[0, ϑ (t)) = t. So m[ϑ (s), ϑ (t)) = t −s for all 0 < s < t < 1.
This the semi-algera of half-open intervals generates the Borel σ -algebra of the
interval, m ◦ ϑ = λ .
Steps 1–4 show that ϑ : ([0, 1], B([0, 1]), m) → ([0, 1], B([0, 1]), λ ) is an isomor-
phism. Composing this with π, we get an isomorphism between (X, B, µ) and the
unit interval equipped with Lebesgue’s measure. ⊔
⊓
A.4 The isomorphism theorem 139

We comment on the atomic case. A standard probability space (X, B, µ) can have
at most countably many atoms (otherwise it will contain an uncountable collection
of pairwise disjoint sets of positive measure, which cannot be the case). Let {xi : i ∈
Λ } be a list of the atoms, where Λ ⊂ N. Then

µ = µ ′ + ∑ µ{xi }δxi (δx = Dirac measure)

i∈Λ

where µ ′ is non-atomic.
Suppose w.l.o.g that X ∩ N = ∅. The map
(
x x ̸∈ {xi : i ∈ Λ }
π : X → X ∪ Λ , π(x) =
i x = xi

is an isomorphism between X and the measure space obtained by adding to (X, B, µ ′ )

atoms with right mass at points of Λ . The space (X, B, µ ′ ) is non-atomic, so it is iso-
morphic to [0, µ ′ (X)] equipped with the Lebesgue measure. We obtain the follow-
ing generalization of the isomorphism theorem: Every standard probability space
is isomorphic to the measure space consisting of a finite interval equipped with
Lebesgue’s measure, and a finite or countable collection of atoms.

Definition A.5. A measure space is called a Lebesgue space, if it is isomorphic to

the measure space consisting of a finite interval equipped with the Lebesgue mea-
surable sets and Lebesgue’s measure, and a finite or countable collection of atoms.

Note that the σ –algebra in the definition is the Lebesgue σ –algebra, not the Borel
σ –algebra. (The Lebesgue σ –algebra is the completion of the Borel σ –algebra with
respect to the Lebesgue measure, see problem 1.2.) The isomorphism theorem and
the discussion above say that the completion of a standard space is a Lebesgue space.
So the class of Lebesgue probability spaces is enormous!
Appendix A
The Monotone Class Theorem

Definition A.1. A sequence of sets {An } is called increasing (resp. decreasing) if

An ⊆ An+1 for all n (resp. An ⊇ An+1 for all n).

Notation: An ↑ A means that {An } is an increasing sequence of sets, and A =

S
An .
An ↓ A means that {An } is a decreasing sequence of sets, and A = An .
T

Proposition A.1. Suppose (X, B, µ) is a measure space, and An ∈ B.

1. if An ↑ A, then µ(An ) −−−→ µ(A);
n→∞
2. if An ↓ A and µ(An ) < ∞ for some n, then µ(An ) −−−→ µ(A).
n→∞

Proof. For (1), observe that A = n≥1 An+1 \ An and use σ –additivity. For (2), fix
U

n0 s.t. µ(An0 ) < ∞, and observe that An ↓ A implies that (An0 \ An ) ↑ (An0 \ A). ⊔
⊓

The example An = (n, ∞), µ =Lebesgue measure on R, shows that the condition in
(2) cannot be removed.

Definition A.2. Let X be a set. A monotone class of subsets of X is a collection M

of subsets of X which contains the empty set, and such that if An ∈ M and An ↑ A
or An ↓ A, then A ∈ M .

Recall that an algebra of subsets of a set X is a collection of subsets of X which

contains the empty set, and which is closed under finite unions, finite intersections,
and forming the complement.

Theorem A.1 (Monotone Class Theorem). A monotone class which contains an

algebra, also contains the sigma–algebra generated by this algebra.

Proof. Let M be a monotone class which contains an algebra A . Let M (A ) de-

note the intersection of all the collections M ′ ⊂ M such that (a) M ′ is a monotone
class, and (b) M ′ ⊇ A . This is a monotone class (check!). In fact it is the min-
imal monotone class which contains A . We prove that it is a σ –algebra. Since
M (A ) ⊂ M , this completes the proof.

141
142 A The Monotone Class Theorem

We begin by claiming that M (A ) is closed under forming complements. Sup-

pose E ∈ M (A ). The set

M ′ := {E ′ ∈ M (A ) : (E ′ )c ∈ M (A )}

contains A (because A is an algebra), and it is a monotone class (check!). But

M (A ) is the minimal monotone class which contains A , so M ′ ⊃ M (A ). It
follows that E ∈ M ′ , whence E c ∈ M (A ).
Next we claim that M (A ) has the following property:

E ∈ M (A ), A ∈ A =⇒ E ∪ A ∈ M (A ).

Again, the reason is that the collection M ′ of sets with this property contains A ,
and is a monotone class.
Now fix E ∈ M (A ), and consider the collection

M ′ := {F ∈ M (A ) : E ∪ F ∈ M (A )}.

By the previous paragraph, M ′ contains A . It is clear that M ′ is a monotone class.

Thus M (A ) ⊆ M ′ , and as a result E ∪ F ∈ M (A ) for all F ∈ M (A ). But E ∈
M (A ) was arbitrary, so this means that M (A ) is closed under finite unions.
Since M (A ) is closed under finite unions, and countable increasing unions, it is
closed under general countable unions.
Since M (A ) is closed under forming complements and taking countable unions,
it is a sigma algebra. By definition this sigma algebra contains A and is contained
in M . ⊔
⊓
Index

Abramov formula, 116 eigenvalue, 90

Action, 1 Entropy
action, 45 Ruelle inequality, 123
adding machine, 32, 85 entropy
algebra, 130 conditional, 104
Alt, 53 of a measure, 107
arithmeticity, 22 of a partition, 103
atom, 135 topological, 118, 119
ergodic, 5
Bernoulli scheme, 9, 95 decomposition, 43
entropy of, 113 hypothesis, 3
Bowen’s metric, 121 ergodicity and countable Lebesgue
box, 45 spectrum, 99
ergodicity and extremality, 85
Carathéodory Extension Theorem, 10 ergodicity and mixing, 85
CAT(0) space, 71 flows, 18, 32
Chacon’s example, 99 theory, 1
Chung-Neveu Lemma, 108 Ergodic Theorem
coboundary, 35 Mean, 35, 84
Commuting transformations, 44 Multiplicative, 57
comparison triangle, 71 Pointwise, 37
complete Ratio, 87
measure space, 31 Subadditive, 49
Riemannian surface, 18 ergodicity and mixing, 36
conditional extension, 23
entropy, 104 exterior product, 54
expectation, 40 and angles, 56
probabilities, 41 of linear operators, 55
configuration, 1 extremal measure, 85
conservative mpt, 32
covariance, 7 factor, 22
cutting and stacking, 99 factor map, 23
cylinders, 9 flow, 1
Fourier Walsh system, 95
Dynamical system, 1 Furstenberg-Kesten theorem, 52

eigenfunction, 90 generalized intervals

143
144 Index

regular family, 85 Lebesgue, 4, 139

generator, 111 non-atomic, 135
geodesic flow, 17 sigma finite, 32
geodesic metric space, 70 standard, 3
geodesic path, 69 mixing
geodesic ray, 70 and countable Lebesgue spectrum, 99
geodesic triangle, 71 weak, 91
GL(d, R), 81 mixing and ergodicity, 36
Goodwyn’s theorem, 119 monotone class, 141
mpt, 4
Herglotz theorem, 93 multilinear function, 52
horofunction compactification, 71 alternating, 53
horofunctions, 71
hyperbolic natural extension, 23
plane, 17
surface, 18 Od (R), 81
orbit, 1
independence, 6
for partitions, 106 partition
induced transformation, 28 finer or coarser, 104
Abramov formula, 116 wedge product of, 104
entropy of, 116 periodicity, 22
for infinite mpt, 33 Perron-Frobenius Theorem, 31
Kac formula, 28 Phase space, 1
Kakutani skyscraper, 30 Poincaré Recurrence Theorem, 3
information Poincaré section, 30
conditional, 104 Polish, 21
content, 103 polish space, 21, 133
function, 103 Posd (R), 81
invariant set, 5 positive definite, 81, 93
invertibility, 23 probability
isometry, 77 preserving transformation, 4
isomorphism, 4 stationary probability vector, 12
measure theoretic, 4 measure, 3
of measure spaces, 136 space, 3
spectral, 89 vector, 12
isomorphism theorem for measure spaces, 136 product
iterates, 1 of measure spaces, 20
of mpt, 20
K automorphism, 96 proper metric space, 69
Kac formula, 28, 33 PSL(2,R), 17
Kakutani skyscraper, 30
rectifiable curve, 69
Lebesgue number, 121 regular family
Liouville’s theorem, 2 generalized intervals, 85
regularity (of a measure space), 134
Markov measure, 13 Rokhlin formula, 115
ergodicity and mixing, 13 rotation, 8
Markov measures Rotations, 32
Entropy, 114 Entropy of, 114
Martingale convergence theorem, 86 Ruelle’s inequality, 123
measure, 3
measure preserving transformation, 4 section map, 30
measure space, 3 semi algebra, 10
Index 145

semi-flow, 1 ergodic theorem, 49

Shannon–McMillan–Breiman Theorem, 109 subshift of finite type, 12
sigma algebra, 3 suspension, 30
sigma finiteness, 32 Sym(d, R), 81
Sinai’s Generator Theorem, 111
skew product, 21 tail events, 96
spanning set, 116 tensor product, 52
spectral time one map, 32
invariant, 89 Topological entropy, 118
isomorphism, 89 definition using separated sets, 121
measure, 93 of isometries, 122
spectrum variational principle, 122
and the K property, 96 topological entropy, 119
continuous, 90, 93 transition matrix
countable Lebesgue, 95 aperiodic, 13
discrete, 90 irreducible, 13
Lebesgue, 99 period, 13
mixed, 90
point, 90 unitary equivalence, 89
pure point, 90
standard variational principle, 122
probability space, 134
stationary wandering set, 32
probability vector, 12 wedge product
stationary stochastic process, 5 of σ –algebras, 104
stochastic matrix, 12 of multilinear forms, 54
stochastic process, 5 of partitions, 111
subadditive
cocycle, 49 Zero one law, 96