LECTURE NOTES ON ERGODIC THEORY

ALEJANDRO KOCSARD

1. The setup of Ergodic Theory

Ergodic theory can be roughly described as the study of measure preserving
dynamical systems.
More precisely, by a (discrete time) dynamical system we mean a map T : X Ñ X,
where X is its phase space and the dynamics is given by the iteration of T: T 0 “ id X
and T n`1 :“ T ˝ T n , for every n ě 0. When T is bijective, one can consider the
n
“past trajectories” taking negative iterates that are defined by T ´n :“ pT ´1 q , for
every n ě 0.
We say that T : X ý is a measurable system when X is a endowed with a σ-
algebra B Ă 2X and T is a B-measurable map.
When µ is a measure on the measurable space pX, Bq and T : X ý is a measurable
map, we say that T preserves µ, or that µ is a T-invariant measure, when it holds
µpAq “ µ T ´1 pAq , @A P B.
` ˘

So, a measure preserving system is the quadruple T : pX, B, µq ý where T : X ý

is a B-measurable map and µ is a T-invariant probability measure, i.e. where
µpXq “ 1.
Analogous concepts can be defined for continuous time dynamical systems, also
called semi-flows and flows, and more generally, for (measurable) actions of semi-
groups and groups.
Our main references for the course are the books of Viana and Oliveira [VO16]
and Mañé [Mañ87].

2. Motivations
2.1. Dynamical systems with a natural invariant measure. There are many classical
dynamical systems, which are used to model natural phenomena, that already arise
with a “natural” invariant (finite) measure. From the historical point of view, one of
the most important examples are Hamiltonian systems that are used to described
the evolution of conservative systems in Classical Mechanics.
So, the first main question in ergodic theory could be briefly and roughly sum-
marized as follows:
What dynamical information can be extracted from the fact that a
dynamical system leaves invariant a certain measure?
2.2. Statistical study of dynamical systems. A second motivation to study measure
preserving systems comes from the analysis of statistical aspects of dynamical sys-
tems. In its simplest form, let us suppose a dynamical system f : X ý is given and
for certain subset A Ă X we want to analyze the asymptotic to behavior (i.e. as
pnq
n Ñ 8) of the mean visiting time function to A τA : X Ñ r0, 1s given by
7 0 ď i ď n ´ 1 : f i pxq P A
␣ (
pnq
τA pxq :“ , @x P X, @n P N.
n
Date: October 9, 2023.
1
2 ALEJANDRO KOCSARD

` pnq ˘
As we shall see along the course, the asymptotic behavior of functions τA ně1
for most subsets A is determined the whole set of f -invariant measures.

2.3. Applications to combinatorics and number theory. As third motivation, we

will see that ideas and techniques of ergodic theory can be applied to get some nice
results on combinatorics and number theory. The main reference for this part of
the course in the book of Einsiedler and Ward [EW11].

3. Measure theory
3.1. Measurable spaces. All along these notes, X will denote a non-empty set. We
shall write 2X for the set of subsets of X. A σ-algebra Ť on X is a family B Ă 2
X

such hat H, X P B and given A1 , A2 , . . . P B , it holds ně1 An P B . A measurable

space is a pair pX, B q where X is a non-empty set and B is a σ-algebra on X.
There is a natural partial order among σ-algebras of X given by set inclusion.
Invoking this order, given any family C Ă 2X one can define the σ-algebra generated
by C , which will be denoted by σpC q, as the smallest σ-algebra that contains C .
When X is a topological space and τ Ă 2X is the topology of X, i.e. τ is the
family of open subsets in X, the σ-algebra σpτq generated by τ is called the Borel
σ-algebra of X is usually denoted by B pXq.
Given a measurable space pX, B q, and an arbitrary map f : X Ñ Y, one can
define the push-forward σ-algebra f ‹ B on Y by
! )
(1) f ‹ B :“ E Ă Y : f ´1 pEq P B .

On the other hand, given another map h : Z Ñ X, we defined the pull-back

σ-algebra f ‹ B on Z by
! )
(2) f ‹ B :“ E Ă Z : DA P B, f ´1 pAq “ E .

Given two measurable spaces pX, B q and pY, C q, a map f : X Ñ Y is said to be

measurable when C Ă f ‹ B .
Exercise 1. If pX, B q and pY, C q are measurable spaces and f : X Ñ Y, show that
f ‹ C Ă B if and only if C Ă f ‹ B .
A measure on a measurable space pX, B q is map µ : B Ñ r0, `8s such that
µpHq “ 0, µpXq ą 0 and
ˆğ8 ˙ 8
ÿ
µ An “ µpAn q,
n“1 n “1

for every sequence of pairwise disjoint sets A1 , A2 , . . . P B . The measure µ on pX, B q

is said to be finite when µpXq ă 8; a probability measure when µpXq “ 1; and
Ť exists a sequence X1 , X2 , . . . P B such that µpXn q ă 8 for every
σ-finite when there
n ě 1 and X “ ně1 Xn .
Given a measure space pX, B , µq, we say that two sets A, B P B are equal modulo
µ when µpA△ Bq “ 0, where A△ B :“ pAzBq Y pBzAq.
Exercise 2. Let pX, B , µq be a finite measure space. Show that the relation „ on
B given by A „ B if and only if A and B are equal modulo µ is an equivalence
relation. Then, let us consider the function dµ : B {„ ˆB {„Ñ R given by
dµ prAs, rBsq :“ µpA△ Bq, @A, B P B,
and where rAs, rBs denotes the equivalence classes of A and B, respectively. Show
that the function dµ is well defined and is a distance on B {„.
 ERGODIC THEORY 3

A measure space is triple pX, B , µq, where pX, B q is a measurable space and µ is
µ
a measure on it. The σ-algebra B0 Ă B of µ-null sets is given by
µ
B0 :“ tN P B : µpNq “ 0u.
µ
We say that the measure space pX, B , µq is complete when for every N P B0 and
any N 1 Ă N, it holds N P B .
Every measure space pX, B , µq admits a unique completion: it is a complete
metric space pX, B , µ̄q such that
! )
µ
B :“ B Y N1 zN2 : B P B , DN P B0 , N1 , N2 Ă N ,
and
µ
µ̄pB Y N1 zN2 q :“ µpBq, @B P B , @N P B0 , @N1 , N2 Ă N.
So, in order to avoid unnecessary technical difficulties, from now on we will
assume all our measure spaces are complete.
Given a measure space pX, B , µq and a map f : X Ñ Y, we can define the push-
forward measure f ‹ µ on the measurable space pY, f ‹ B q by
f ‹ µpAq :“ µ f ´1 pAq , @A P f ‹ B .
` ˘
(3)
Exercise 3. Let pX, B , µq be a measure space, pY, C q a measurable space and f : X Ñ
Y a measurable map. The it holds
ż ż
ϕ ˝ f dµ “ ϕ d f ‹ µ,
X Y
for any measurbale function ϕ : Y Ñ r0, 8s.
It is important to remark that, in general, it is not possible to pulled measures
back. In fact, it is only possible to do it when the corresponding map is invertible.
Given two measure spaces pX, B , µq and pY, C , νq, a measure preserving map or a
metric homomorphism is a measurable map h : X Ñ Y such that h‹ µ “ ν. Observe
that if we were pedantically formal, taking into account ˇ that the domain of h‹ µ
is h‹ B Ą C , last equality should be written as h‹ µˇC “ ν. In particular, a map
f : X Ñ X is said to be a metric endomorphism or a measure preserving map when
f is measurable and f ‹ µ “ µ. In such a case, we say that µ is an f -invariant
measure.
A measure preserving map h : pX, B , µq Ñ pY, C , νq is said to be a measure
isomorphism when there exist ˇ measurable sets X 1 P B and ˇY 1 P C such that
µpXzX q “ νpYzY q “ 0, h X1 : X 1 Ñ Y 1 is a bijection and h´1 ˇY1 is a measurable
1 1 ˇ
map.
Finally, we will say that a measure preserving map f : pX, B , µq Ñ pX, B , µq is
an automorphism or an invertible measure preserving map when f is a metric
isomorphism.

3.2. Borel and Lebesgue spaces. Let M denote a separable metric space and let
B M be its Borel σ-algebra, i.e. B M is the smallest σ-algebra containing every open
subset.
A probability space pX, B, µq is said to be a Borel space when there is a Borel
probability measure µ M on pM, B M q such that pX, B, µq is isomorphic to pM, B M , µ M q.
Observe that if pX, B, µq is Borel space, then there exists a countable family
A Ă B such that σpAq “ B, i.e. the σ-algebra B is countably generated.
Finally, we say that the probability space pX, B, µq is a Lebesgue space when
it is complete and is isomorphic to the completion of a Borel probability space.
See [EW11, Appendix A. 6] for more details about Borel and Lebesgue spaces.
4 ALEJANDRO KOCSARD

3.3. Smooth measures. Let M be a finite-dimensional compact oriented smooth d-

manifold M and ω be a smooth volume form, i.e. ω is a smooth d-form such that
for every x P M and every positively oriented base tX1 , X2 , . . . , Xd u of the tangent
space Tx M, it holds ωpX1 , . . . , Xd q ą 0.
Then, we can define the continuous linear functional Λω : C8 pM, Rq Ñ R by
ż
Λω pϕq :“ ϕω, @ϕ P C8 pM, Rq.

One can easily check that Λω can be uniquely extended to a continuous linear
functional Λω : C0 pM, Rq Ñ R, and by Riesz’ representation theorem we conclude
that there exists a finite Borel measure µω such that
ż
Λω pϕq “ ϕ dµω , @ϕ P C0 pM, Rq.
M
The measure µω is said to be a smooth measure.
Exercise 4. Can we define smooth measures on non-compact oriented manifolds?
Can smooth measures be finite on non-compact manifolds? Smooth measures can
be defined on non-orientable manifolds?

4. Recurrence on finite measure spaces

In this section we shall analyze the fact that finiteness of invariant measures
imposes recurrence of the dynamical system.
Our first result is the celebrated
Theorem 4.1 (Poincaré recurrence theorem). Let pX, B , µq be a probability space
and f : pX, B , µq ý be a measure preserving map. Let A P B be a measurable set
such that µpAq ą 0. Then, there exists n ě 1 such that
µ A X f ´n pAq ą 0.
` ˘

Proof. Let us suppose that A, f ´1 pAq, f ´2 pAq, . . . are essentially disjoint, i.e. sup-
pose that
µ f ´i pAq X f ´ j pAq “ 0, @i, j P N0 , i ‰ j.
` ˘
(4)
´Ť ¯
n ´1 ´ i
Let us define the sets A0 :“ A, and An :“ f ´n pAqz i“0 f pAq , for every
n ě 1. Then, the sets A0 , A1 , A2 , . . . are pairwise disjoint, and by (4), we have
nÿ
´1 `
µ f ´n pAq ě µpAn q ě µp f ´n pAqq ´ µ f ´n pAq X f ´i pAq “ µ f ´n pAq ,
` ˘ ˘ ` ˘

i “0

for every n ě 1, and hence it holds µ f ´n pAq “ µpAn q, for any n ě 0. Since µ is
` ˘

f -invariant, this implies µpAn q “ µpAq ą 0, for every n ě 0, getting a contradiction

because ˆğ 8 ˙ ÿ 8 8
ÿ
1 “ µpXq ě µ An “ µpAn q “ µpAq “ 8.
i “0 i “0 i “0
So, assumption (4) does not hold and hence, there exist i, j P N, with i ă j such
that
´ ˘¯
0 ă µ f ´i pAq X f ´ j pAq “ µ f ´i A X f i´ j pAq “ µpA X f ´p j´iq pAqq.
` ˘ `

□
There are a lot of important results that can be proven as a consequence of
Theorem 4.1. The firs one is indeed a very simple improvement of Theorem 4.1:
 ERGODIC THEORY 5

Corollary 4.2. Let f : pX, B , µq ý and A P B be as in Theorem 4.1. Then, there is

a natural number n ď µp1Aq such that

µ A X f ´n pAq ą 0.
` ˘

Exercise 5. Prove Corollary 4.2.

Corollary 4.3. Let f : pX, B , µq ý and A P B as in Theorem 4.1. Then the set
A f :“ tx P A : Dn ě 1, f n pxq P Au
belongs to B and µpAzA f q “ 0.
Proof. First notice that ď`
A X f ´n pAq .
˘
Af “
n ě1
So, A f P B .
Then, let us define A1 :“ AzA f . Observe that A1 X f ´n pA1 q “ H, for every
n ě 1. Hence, we have that f ´m pA1 q X f ´n pA1 q “ H, whenever 0 ď m ă n. By
Theorem 4.1, we conclude that µpA1 q “ 0. □
Then notice that Corollary 4.3 can be considerably improved:
Corollary 4.4. Let f : pX, B , µq ý and A P B as in Theorem 4.1. Then the set
A8 :“ x P A : Dn j Ò 8, f n j pxq P A, @j ě 1
␣ (

belongs to B and µpAzA8 q “ 0.

Exercise 6. Prove Corollary 4.4.
4.1. Topological recurrence. Let X be a topological space and f : X ý be an ar-
bitrary map. We say that a point x0 P X is recurrent for f when given any
neighborhood V of x0 , there is a natural number n ě 1 such that f n px0 q P V.
Recall that a base for the topology of X is a family of open subsets V such that
every open subset of X can be writte as the union of elements of V .
Then we have the following topological version of Poincaré recurrence theorem:
Theorem 4.5. Let X be a topological space admitting a countable base for its topology
and let B pXq denote its Borel σ-algebra. Let µ be a Borel probability measure
on X (i.e. a probability measure on pX, B pXqq) and f : pX, B pXq, µq ý a measure
preserving transformation. Then, µ-almost every point is recurrent for f .
Proof. Let U :“ tUn : n P Nu be a base for the topology of X. For each n P N
let us consider the set Wn :“ tx P Un : f k pxq R Un , @k ě 1u. ByŤCorollary 4.3 we
know that µpWn q “ 0, for every n ě 1. Hence, if we write W :“ ně1 Wn , we have
µpWq “ 0. Then we claim that every point in XzW is recurrent for f . In fact, let x
an arbitrary point of XzW and V any neighborhood of x. Since U is a base for the
topology of X, there is a natural number n such that x P Un Ă V. Since x R W, we
have x R Wn and then, there is k ě 1 such that f k pxq P Un Ă V. Since V was an
arbitrary neighborhood, we conclude x is recurrent for f , as desired. □

5. Some examples of measure preserving transformations

5.1. Periodic orbits. Let X be an arbitrary set and z P X an arbitrary point. Then
we can define the δz on the measurable space pX, 2X q by
#
1, if z P A;
δz pAq :“
0, otherwise.
6 ALEJANDRO KOCSARD

Now, if f : X ý is any map and there exists p P X which is periodic for f ,

i.e. there exists k P N such that f k pzq “ z, then the periodic measure
k´1 k´1
1 ÿ 1 ÿ j
νp :“ δ f j p pq “ f ‹ δp
k k
j “0 j “0

is f -invariant.

5.1.1. Homeomorphisms of the interval. Using periodic measures and Poincare’s

recurrence theorem (Theorem 4.5) one can easily determine all invariant measures
for homeomorphisms of a compact interval. . . at least for homeomorphisms with
finitely many periodic orbits.
In fact, if I Ă R denotes a compact interval and f : I ý is an orientation-
preserving homeomorphism (i.e. f is an strictly increasing map), then one can easily
show that x P I is a recurrent point for f if and only if f pxq “ x, and moreover, the
set of fixed points Fixp f q :“ tw P I : f pwq “ wu is non-empty.
On the other hand, if f is orientation-reversing, i.e. f is strictly decreasing, the
Fixp f q has exactly one point and any other recurrent point is periodic of period 2.
In any case, if f has finitely many periodic orbits, by Theorem 4.5 one have
that any f -invariant Borel probability measure is indeed a convex combination of
periodic ones. What can we say about the characterization of f -invariant measures
when f has infinitely many periodic orbits?

5.2. Expanding maps of the interval. Let k P N be an arbitrary number and con-
sider the map f : r0, 1s ý given by
f pxq :“ kx ´ tkxu , @x P r0, 1s
where txu :“ maxtℓ P Z : ℓ ď xu denotes the integer part of x.
We claim that the Lebesgue measure m on r0, 1s is f -invariant. If J Ă r0, 1s is a
subinterval, then one can easily check that f k´1 pJq is the disjoint union of k intervals
J1 , . . . , Jk Ă r0, 1s with mpJi q “ 1k mpJq, for each i P t1, . . . , ku. So, we conclude
k
ˆğ ˙ ÿk k
` ´1 ˘ ÿ 1
m f pJq “ m Ji “ mpJi q “ mpJq “ mpJq.
k
i“1 i “1 i “1

Thus, we have proved that the two probability measures m and f ‹ m coincide on
the subintervals of r0, 1s. In order to conclude that they are indeed the same, we
invoke the following general result:
Theorem 5.1 (Uniqueness of measures, Theorem 5.7 in [Sch17]). Let µ and ν two
measures on the measurable space pX, B q. Let C Ă 2X be a family that generates
B , i.e. B “ σpC q, and suppose that
‚ if A, B P C ùñ A X B P C ;
‚ there exists pCn qně1 Ă C such that Cn Ò X (i.e. ně1 Cn “ X) with µpCn q “
Ť
νpCn q ă 8, for every n ě 1.
Then, µ “ ν.
5.3. Circle rotations. Given any α P R, one can define the map Tα : r0, 1s ý by
Tα pxq :“ x ` α ´ tx ` αu , @x P r0, 1s.
We claim that the Lebesgue measure m on r0, 1s is again Tα -invariant, for every
α P R. In fact, it is enough to consider the case where α P p0, 1q and then, one can
 ERGODIC THEORY 7

easily show that given ra, bs Ă r0, 1s, it holds

Tα´1 ra, bs “ ra ´ α, b ´ αs, when α ď a ď b ď 1;
Tα´1 ra, bs “ ra ´ α ` 1, b ´ α ` 1s, when 0 ď a ď b ă α;
Tα´1 ra, bs “ ra ´ α ` 1, 1q X r0, b ´ αs, when 0 ď a ă α ď b ď 1.
Circle rotations are particular examples of translations on compact groups.
5.4. Translations on locally compact groups. A topological group is a ternary pG, T , ¨q
where G is a non-empty set, T is a Hausdorff topology on G and pG, ¨q is group
such that the map G ˆ G Q pg, hq ÞÑ g ¨ h´1 P G is a continuous map.
Analogously, a Lie group is a ternary pG, C , ¨q where G is a non-empty set, C is
a finite dimensional C8 Hausdorff structure on G, pG, ¨q is a group and the map
G ˆ G Q pg, hq ÞÑ g ¨ h´1 P G is C8 . Observe that any Lie group is a locally compact
topological group.
Given topological group G and any element g P G, one can define the left and
right translations by L g : G Q h ÞÑ gh P G and R g : G Q h ÞÑ hg´1 P G, respectively.
Notice that L g and R g are homeomorphism on G; and when G is a Lie group, they
are C8 diffeomorphisms.
Definition 1 (Haar measure). Given a locally compact group G, a (left) Haar mea-
sure is a Radon measure µ on G (i.e. µ is Borel and µpKq ă 8, for every compact
K Ă G) such that µ is L g -invariant, for every g P G, and there is a Borel set E Ă G
such that µpEq ą 0.
Theorem 5.2 (Existence and characterization of Haar measures). If G is locally
compact group, then there exists a left Haar measure µG . The support of µG is G
(i.e. µG pUq ą 0, for every non-empty open subset U Ă G) and µG pGq ă 8 if and
only if G is compact.
On the other hand, if ν is another left Haar measure, then there exists c ą 0
such that
µpEq “ cνpEq, @E P B pGq.
Proof. See for instance Halmos [Hal50, §58] □
Exercise 7. If G is Lie group, then show that any left Haar measure is a smooth
measures (as defined in §3.3).
Example 5.3 (Tori). The space Rd is an abelian Lie group and Zd ă Rd is a
(normal) subgroup. Hence, Td :“ Rd {Zd is a compact abelian Lie group, and there
exists a unique left Haar measure λd which is a probability measure. The measure
λd is left and right invariant is usually called the Haar or the Lebesgue measure on
Td .
Given a topological group G, we say that a map f : G ý is a topological group
endomorphism (or just an endomorphism for short) when f is continuous and
f pxyq “ f pxq f pyq, for every x, y P G. An endomorphism f : G ý is said to be
a topological group automorphism (or just an automorphism) when f is bijective
and the map f ´1 is an endomorphism as well.
Exercise 8. Let G be a compact topological group, µ be a the left Haar probability
measure and f : G ý be an automorphism.
(a) Show that µ is a right Haar measure too, i.e. it is invariant by every right
translation.
(b) Show that µ is an f -invariant measure.
(c) What happend when f is just an endomorphism?
(d) And what happend when G is non-compact and just locally compact?
8 ALEJANDRO KOCSARD

5.5. The Gauss map. Let us consider the map G : p0, 1q Ñ r0, 1q given by
Z ^
1 1
Gpxq :“ ´ , @x P p0, 1q.
x x
Exercise 9. Show that given x P p0, 1q there exists n P N such that Gi pxq ‰ 0, for
every i P t0, 1, . . . , n ´ 1u and G n pxq “ 0 if and only if x P Q.
ˇ
So, if we define X “ p0, 1qzQ, then it holds GˇX : X ý.
A remarkable fact is that there exists a smooth probability measure µ on p0, 1q
(or on X) which is G-invariant. In fact we have
Theorem 5.4. The measure µ given by
ż
1 dx
µpEq :“ , @E P B p0, 1q,
log 2 E 1 ` x
is a G-invariant Borel probability measure.
Proof. It clearly holds µp0, 1q “ 1. Let us prove that µ is G-invariant. For each
n P N, consider the interval
ˆ ˙
1 1
In :“ , , @n P N.
n`1 n
Observe that Gn :“ Gˇ I : In Ñ p0, 1q is a C8 diffeomorphism, for each n P N,
ˇ
´Ů n

and m n In q “ mp0, 1q “ 1, where m denotes the Lebesgue measure on p0, 1q.

Given any ρ P L1 pp0, 1q, mq, with ρ ě 0 m-almost everywhere, let us consider the
measure µρ given by
ż
µρ pEq :“ ρ dm, @E P Bp0, 1q.
E
If we define the so called Perron-Frobenious operator L by
ÿ8
ρpGk´1 pyqq ğ
Lρpyq :“ 1
` ´1 ˘ , @y P Ik ,
k“1 Gk Gk pyq k ě1

we claim that
ż
G‹ µρ pEq “ µLρ pEq “ Lρ dm, @E P Bp0, 1q.
E
In order to prove this, let E be an arbitrary Borel subset of p0, 1q and let us write
1E to denote its indicator function. By the classical change of variable formula we
have
ż1 ż1
µρ G´1 pEq “ 1G´1 pEq pxqρpxq dmpxq “ 1E Gpxq ρpxq dmpxq
` ˘ ` ˘
0 0
8 ż
ÿ
1E
` ˘
“ Gpxq ρpxq dmpxq
k“1 In
ÿ8 ż1 ˘1
1E pyqρ Gk´1 pyq Gk´1 pyq dmpyq
` ˘`
“
k “1 0
8 ż1
ρ Gk´1 pyq
` ˘
ÿ
“ 1E pyq 1 ` ´1 ˘ dmpyq
k “1 0 Gk Gk pyq
ż1
“ 1E pyqLpρqpyq dmpyq “ µLρ pEq,
0
 ERGODIC THEORY 9

where the last line of the chain of equalities follows by classical Levi’s monotone
convergence theorem.
1
So, if we consider ρpxq :“ 1` x , the fact that µρ is a G-invariant (finite) Borel
measure follows from the fact that Lρ “ ρ and this easily follows from a straight-
forward computation. □
5.6. Invariant smooth measures. Let M and N be a smooth manifolds, α be a k-
form on N and f : M Ñ N be a C1 map. The pullback of α by f is the k-form f ‹ α
on M given by
` ˘
f ‹ αpX1 , X2 , . . . , Xk q :“ α D f x pX1 q, D f x pX2 q, . . . , D f x pXk q ,
for all x P M, and every X1 , . . . , Xk P Tx M.
Exercise 10. Let M be a compact (oriented) smooth manifold, ω a smooth volume
form and f : M ý be a C1 diffeomorphism. Then, show that the induced smooth
measure µω is f -invariant if and only if f ‹ ω “ ω.
On the other hand, let X be smooth vector field on the compact manifold M.
Let us write Φ X : R ˆ M Ñ M for the flow induced by X. We say that a Borel
measure µ on M is X-invariant (or Φ X -invariant) when it is ΦtX -invariant, for every
t P R, where ΦtX pzq “ Φ X pz, tq, for every z P M and any t P R.
Then, given a smooth k-form α on M, one defines the Lie derivative of α with
respect to X by ` t ˘‹
ΦX α ´ α
L X α :“ lim .
t Ñ0 t
Exercise 11. Show that given a smooth vector field X on the compact oriented man-
ifold M, the induced smooth measure µω is X-invariant if and only if L X ω “ 0.
5.7. Symplectomorphisms and Hamiltonian flows.

6. Constructing new measure preserving systems from old ones

6.1. Product maps. Let pX1 , B1 , µ1 q and pX2 , B2 , µ2 q be two σ-finite measure spaces.
The product σ-algebra B1 b B2 on X1 ˆ X2 is defined as the σ-algebra generated
by the family of rectangles tE1 ˆ E2 : Ei P Bi , i P t1, 2uu. Then, one shows that
there exists a unique measure µ1 b µ2 on the measurable space pX1 ˆ X2 , B1 b B2 q,
called the product measure, such that
µ1 b µ2 pE1 ˆ E2 q “ µ1 pE1 qµ2 pE2 q, @E1 P B1 , @E2 P B2 .
When f i : pX1 , Bi , µi q ý, for i P t1, 2u, are measure preserving maps, then invok-
ing Theorem 5.1 one can easily show`that the measure ˘ µ1 b µ2 is f 1 ˆ f 2 -invariant,
where f 1 ˆ f 2 : X1 ˆ X2 Q px1 , x2 q ÞÑ f 1 px1 q, f 2 px2 q P X1 ˆ X2 .
By induction, this construction can be easily extended from two to any (finite)
number of factor maps. Moreover, this construction can be extended to countably
many factors, assuming that each factor is a probability space (see [VO16, Appendix
A.2] for details).
6.2. Induction or renormalization. Let pX, B, µq be a probability space, f : X ý be
a measure preserving map and A P B a measurable set with µpAq ą 0. By Poincaré
recurrence theorem (Corollary 4.4) we know that µ-almost every point x P A returns
to A infinitely many times. So we can define the function ρ A : A Ñ N by
ρ A pxq :“ min tn P N : f n pxq P Au ,
and this function is well-defined µ-a.e. on A. Then we define the first return map
to A by
f A pxq :“ f ρ A pxq pxq,
10 ALEJANDRO KOCSARD

which is well define µ-a.e. on A. In order to see that f A is a well defined dynamical
system, is enough to observe that if we define
˜ ¸
č ď
A f :“ A X f ´n pAq ,
m ě0 n ě m

then by Corollary 4.4 one have that µpA f q “ µpAq ą 0, with A f Ă A, and one can
easily check that ρ A pxq ă 8 and f A pxq P A f , for every x P A f . So, this means that
f A : A f ý is a well defined dynamical system. Moreover, if define the probability
measure µ A on the measure space pA, B A q by
µpEq
µ A pEq :“ , @E P B A ,
µpAq
and where B A :“ tB P B : B Ă A f u, we have the following
Theorem 6.1. The map f A : A f ý is a B A -measurable and the measure µ A is
f A -invariant.
Proof. For the sake of simplicity of the notation, from now on we shall assume that
A “ A f , and for each n P N, let us consider the set An :“ ρ´1 pnq Ă A. Observe
that ğ
A“ An ,
nPN
and for each E Ă A, it holds
ğ ğ
(5) f A´1 pEq “ f A´1 pEq X An “ f ´n pEq X An .
n PN n PN
On the other hand, one can easily check that
ˆ nč
´1 ˙
(6) An “ ρ´1 pnq “ A X f ´i pXzAq X f ´n pAq, @n P N.
i “1
From (6) we conclude ρ is a measurable function and then, by (5) we get that
f A is a B A -measurable map.
Firnally, in order to show that µ A is f A -invariant, it is enough to show that
µpEq “ µp f A´1 pEqq, for any E P B A . In order to prova that, let us fix a measurable
set E P B A and observe that, by (5) we get
8 `
ÿ
µ f A´1 pEq “ µ f ´n pEq X An .
` ˘ ˘

n“1

Let us prove that the right hand side of this equality is equal to µpEq. To do that,
let us start noticing that
µpEq “ µ f ´1 pEq “ µ f ´1 pEq X A ` µ f ´1 pEqzA
` ˘ ` ˘ ` ˘
(7)
“ µ f ´1 pEq X A1 ` µ f ´1 pEqzA ,
` ˘ ` ˘
` ˘ ` ˘
where the first equality follows from the fact f ´1 pEq “ f ´1 pEq X A \ f ´1 pEqzA ,
and the second one from f ´1 pEq X A “ f ´1 pEq X A1 .
Then, focusing on the last term of (7) we get
µ f ´1 pEqzA “ µ f ´2 pEqz f ´1 pAq
` ˘ ` ˘
´` ¯ ´` ¯
´2 ´1 ´2 ´1
˘
(8) “ µ f pEqz f pAq X A ` µ f pEqz f pAqqzA
´ ˘¯
“ µ f ´2 pEq X A2 q ` µ f ´2 pEqz A X f ´1 pAq ,
` `
 ERGODIC THEORY 11

where the first equality follows from the invariance of µ by f , the second one by the
partition argument
` we used above
˘ to prove (7), and the last equality follows from
the fact that f ´2 pEqz f ´1 pAq X A “ f ´1 pEq X A2 .
Combining equations (7) and (8) and repeating inductively the last argument,
one can show that
ÿn ´ ˘¯
µ f ´i pEq X Ai ` µ f ´n pEqz A X f ´1 pAq X ¨ ¨ ¨ X f ´pn´1q pAq ,
` ˘ `
µpEq “
i “1

for every n ě 1. ` ˘
So, in order to prove that µpEq “ µ f ´1 pEq it is enough to show that
ˆ ´ nč
´1 ¯˙ ˆ ´ nď
´1 ¯˙
µ f ´n pEqz f ´i pAq ď µ f ´n pAqz f ´i pAq Ñ0
i “1 i “1

as n Ñ `8. But this easily follows from the fact that f ´1 pAqzA, f ´2 pAqzpA Y
f ´1 pAqq . . . are pairwise disjoint and hence,
8 ˆ
ÿ ´ nď
´1 ¯˙
1 “ µpXq ě µ f ´n pAqz f ´i pAq .
n“1 i “1

6.3. Kakutani-Rokhlin towers. In this section we shall try to revert the previous
construction we performed in §6.2. To do that, for the time being let pX, B, µq be
an arbitrary measure space, ρ : X Ñ N0 be an integrable function and f : X ý be
a measure preserving map.
We first onsider the new measure space pXρ , Bρ , µρ q where each object is defined
as follows: we first define the space Xρ by
ğ ğ
Xρ :“ ρ´1 pnq ˆ tiu Ă X ˆ N0 ;
nPN0 0ďiďn

the σ-algebra Bρ is defined as the one generated by the family

! )
Bρ :“ σ A ˆ tiu Ă X ˆ N0 : A P B, Dn P N0 , 0 ď i ď n, A Ă ρ´1 pnq ;

and the measure µρ is characterized as the only measure on the measurable space
pXρ , Bρ q such that
ˆż ˙´1
µρ pA ˆ tiuq :“ ρ ` 1 dµ µpAq,
X

for every generating set A ˆ tiu of Bρ .

Exercise 12. Show that pXρ , Bρ , µρ q is a well-defined probability space.
Then we define the measure preserving map f ρ : Xρ ý by
#
px, i ` 1q, for x P X, 0 ď i ă ρpxq;
f ρ px, iq :“ ` ˘
f pxq, 0 , for x P X, i “ ρpxq.

Exercise 13. Show that the map f ρ : pXρ , Bρ , µρ q ý is indeed a measure preserv-
ing map. Now consider the first return map (also called induction as in §6.2) of
the map f ρ to the set A :“ X ˆ t0u. Show that µρ pAq ą 0 and the first return
map is “isomorphic” to the original f : pX, B, µq ý. Can you give a definition of
isomorphism between measure preserving maps?
12 ALEJANDRO KOCSARD

7. Existence of invariant measures for continuous maps

The main porpuse of this section consists in proving the following
Theorem 7.1. Let X be compact metric space and f : X ý be a continuous map.
Then, there exists an f -invariant Borel probability measure.
In order to prove Theorem 7.1 we first need to introduce some concepts in Func-
tional Analysis.
7.1. Linear operators and dual spaces. Let pE, }¨}E q and pF, }¨}F q be two Banach
spaces over the field K “ R, C. A linear map A : E Ñ F is continuous if and only
its norm
}Ax}F
(9) }A} :“ sup ă 8.
xPX zt0u }x}E

The space of linear continuous maps from E to F is a denoted by LpE, Fq and it

turns to be a Banach space when its endowed with the norm given by (9).
When F is the field itself K, a continuous linear operator η : E Ñ K is called
a continuous linear functional. The dual space E‹ is defined as the linear space
LpE, Kq. Then we can define the norm }¨}E‹ on E‹ by
}η}E‹ :“ sup t|ηpxq| : x P E, }x}E “ 1u , @η P E‹ .
The linear space E‹ turns into a Banach space when it is endowed with the norm
}¨}E‹ . The topology on E‹ induced by the norm }¨}E‹ is usually called the strong
topology.
On the other hand, any continuous linear operator A : E Ñ F naturally induces
the so callled adjoint operator A : F‹ Ñ E‹ given by
` ˘
A‹ ηpxq :“ η Apxq , @x P E, @η P F‹ .
Exercise 14. Show that the adjoint operator is indeed a linear continuous operator
and }A} ě }A‹ }.
7.2. Weak topologies. Given a topological space pY, τY q, a non-empty set X and an
arbitrary family of maps F “ t f i : X Ñ YuiP I , we define the weak topology on X
induced by τY and F as the weakest (i.e. the coarsest) topology τ on X such that
f i : pX, τq Ñ pY, τY q is continuous, for every i P I.
Now, if pE, }¨}E q denotes a Banach space over K “ R, C, and E‹ is its dual space,
we can define the so called weak topology on E as the weak topology induced by the
family E‹ , i.e. it is the weakest topology on E such that every η P E‹ is continuous
on E. In general, the weak topolgy is weaker than the Banach topology induced by
the norm }¨}E .
Exercise 15. Show that if E is a Banach space, then norm and the weak topologies
coincide if and only if E has finite dimension.
Now, we can consider the space E‹‹ :“ pE‹ q‹ which is the dual space of the
Banach space pE‹ , }¨}E‹ q. Analogously, we can endow the space E‹ with the weak
topology which is the weakest topolgy such that every element of E‹‹ is continuous.
However, there shall need to endow the space E‹ with a weaker topology which
is defined as follows. One can show that there is a natural injection of E into E‹‹ ,
i.e. a linear continuous injective map j : E Ñ E‹‹ which is given by
jpxqpηq :“ ηpxq, @x P E, @η P E‹ .
Then, the so called weak-‹ topology is the weakest topolgy on E‹ such that every
linear functional in jpEq is continuous. Given a sequence pϕn qně1 Ă E‹ , we shall
write
ηn á η P E‹ , as n Ñ 8,
 ERGODIC THEORY 13

when the sequence pηn qně1 converges to η in the weak-‹ topology.

Exercise 16. Show that the weak and weak-‹ topologies coincide on E‹ if and only
if E has finite dimension. On the other hand, show that that the closed unit ball
on a Banach space E, which is given by tx P E : }x}E ď 1u is compact if and only
if E has finite dimension.
We shall need the following
Theorem 7.2 (Banach-Alaoglu theorem). The closed unit ball tη P E‹ : }η}E‹ ď 1u
is a compact set when E‹ is with its weak-‹ topology.
Proof. See for instance Einsiedler and Ward [EW17, Theorem 8.10]. □
7.3. The space of continuous functions. Let pX, dq be a compact metric space and
write C0 pXq to denote the space of continuous real functions on X. Since X is
compact, any continuous function is bounded and hence, we can define the norm
}ϕ}0 :“ sup |ϕpxq| , @ϕ P C0 pXq.
xPX

Exercise 17. Show that C0 pXq is a Banach space when it is endowed with the norm
}¨}.
Given an arbitrary continuous map f : X ý, we can consider the induced pull-
back linear map, also called the Koopman operator U f : C0 pXq ý given by

U f pϕq :“ ϕ ˝ f , @ϕ P C0 pXq.
Exercise 18. Show that U f is a continuous linear operator and its norm
› › !› › )
›U f › :“ sup ›U f pϕq› : ϕ P C0 pXq, }ϕ}0 “ 1 “ 1.
› › › ›
0
› ›
Moreover, ›U f ϕ› “ }ϕ}, for every ϕ P C0 pXq if and only if f is surjective.
› ›

On the other hand, we have the following

Theorem 7.3 (Riesz representation theorem). If η : C0 pXq Ñ R is a continuous
linear functional on the Banach space pC0 pXq, }¨}q, then there exist two finite Borel
measures µ“ and µ“ on X such that
ż ż
ηpϕq “ ϕ dµ` ´ ϕ dµ´ , @ϕ P C0 pXq.
X X

Proof. See for instance Einsiedler and Ward [EW17, Theorem 7.44]. □
Then, let us write MpXq for the space of Borel probability measures on X. By
‹
Theorem 7.3 we can consider MpXq as a subset of unit ball of the dual space C0 pXq .
On the other hand, by Exercise 3 we know that the restriction of the adjoint operator
of U f to the space MpXq coincide with the push-forward operator f star : MpXq ý.
In particular, we conclude f ‹ : MpXq ý is continuous when MpXq is endowed with
‹
the weak-‹ topology that inherits from the inclusion MpXq Ă C0 pXq .
Exercise 19. Show that the space of ! Borel probability measures
) MpXq is a closed
0 ‹
subset of the closed unit ball B̄ :“ ξ P C pXq : }ξ} ď 1 when it is endowed with
the weak-‹ topology. Then, invoking Theorem 7.2, conclude that MpXq is compact
with the weak-‹ topology.
Now we can complete the
14 ALEJANDRO KOCSARD

Proof of Theorem 7.1. Let ν be any Borel probability measure on X. For each
n P N, let us consider the measure µn P MpXq given by
n ´1
1 ÿ j
µn :“ f ‹ ν.
n
j “0

By Exercise 19 we know that MpXq is compact for the weak-‹ topology. So,
there exists a subsequent pµn j q and µ P MpXq such that µn j á µ, as n j Ñ 8.
jě1
The, since f ‹ : MpXq ý is continuous, we get f ‹ µn j á f ‹ µ, as n j Ñ 8. So, in order
to show that µ is f -invariant, it is enough to show that f ‹ µn ´ µn á 0, as n Ñ 8.
In order to do that, let ϕ P C0 pXq be arbitrary and notice that
ż n ´1 ˆ ż ż ˙ ż
1 ÿ j `1 j 1
ϕ dp f ‹ µn ´ µn q “ ϕ d f‹ ν ´ ϕ d f‹ ν “ ϕ ˝ f n ´ ϕ dν Ñ 0,
x n X X n X
j “0

as n Ñ `8. □
Exercise 20. Let X and Y be two compact metric spaces, and f : X ý and g : X ˆ
Y Ñ Y two continuous functions. Let us consider the so called skew-product map
H : X ˆ Y ý given by
` ˘
Hpx, yq :“ f pxq, gpx, yq , @px, yq P X ˆ Y.
Given any invariant measure µ P Mp f q, let us show that there exists an invariant
measure µ̂ P MpHq such that π‹ µ̂ “ µ, where π : X ˆ Y Ñ X denotes the projection
on the first coordinate.

8. Ergodic theorems
Let pX, B, µq be a probability space and f : X ý a measure preserving endomor-
phism. Let E P B be a subset such that µpEq ą 0. Then by Poincare’s recurrence
theorem (Corollary 4.4) we know that the set ti P N0 : f i pxq P Eu is infinite for
µ-almost every point x P E. This can be considered as a qualitative information.
In this section we shall focus on the problem of finding quantitative estimates for
the frequency of sojourn
n ´1
1 1 ÿ
7ti P N0 : i ă n, f i pxq P Eu “ 1E ˝ f i pxq,
n n
i“0

as n Ñ 8, and where 1E : X Ñ t0, 1u denotes the indicator function of the set E.

To do that, we shall analyze the asymptotic behavior of so called Birkhoff averages
as the following one:
n ´1
1 ÿ
S nf ϕ :“ ϕ ˝ f j,
n
j “0
for an arbitrary integrable function ϕ : X Ñ R.
The first result concerning the asymptotic behavior of Birkhoff averages is the
celebrated
Theorem 8.1 (von Neumann Ergodic Theorem). Let px, B, µq be a probability space
and f : X ý be a measure preserving endomorphism. Then, for every ϕ P L2 pX, B, µq,
there is a ϕ̃ P L2 pX, B, µq such that
n ´1
1 ÿ
ϕ ˝ f j Ñ ϕ̃, in L2 pX, B, µq,
n
j “0
 ERGODIC THEORY 15

as n Ñ `8. Moreover, it holds

` ˘
ϕ̃pxq “ ϕ̃ f pxq , for µ-a.e. x P X.
In order to prove Theorem 8.1, we first need to revise some concepts on Lebesgue
and Hilbert spaces.

8.1. Revision on Lebesgue spaces. Let pX, B , µq be an arbitrary measure space. For
each real number p with 1 ď p ă 8, let us consider the space L p pX, B , µq of (equiv-
alence classes of) measurable functions ϕ : X Ñ C such that |ϕ| p is integrable, and
where two functions are equivalent if and only if they coincide µ-almost everywhere
on X. Then one can show that the L p pX, B, µq is a Banach space when it is endowed
with the norm
ˆż ˙1{ p
p
}ϕ} p :“ |ϕ| dµ , @ϕ P L p pX, B, µq.
X
On the other hand, we define the space L8 pX, B, µq as the space of (equivalence
classes of) measurable functions ϕ : X Ñ C such that there exists C ą 0 satisfying
|ϕpxq| ď C, for µ-almost every x P X, and where two functions are equivalent when
they coincide µ-almost everywhere. The space L8 pX, B, µq is a Banach space when
it is endowed with the norm
}ϕ}8 :“ inf tC ą 0 : |ϕpxq| ď C, for µ-a.e. x P Xu , @ϕ P L8 pX, B, µq.
Exercise 21. If pX, B, µq is a probability space, 1 ď p ă q ď 8 and ϕ P Lq pX, B, µq,
then show that ϕ P L p pX, B, µq and }ϕ} p ď }ϕ}q . On the other hand, show that

}ψ}8 “ lim }ψ} p , @ψ P L8 pX, B, µq.

pÑ8

Exercise 22. Consider the map x¨, ¨y : L2 pX, B, µq ˆ L2 pX, B, µq Ñ C given by

ż
xϕ, ψy :“ ϕψ dµ, @ϕ, ψ P L2 pX, B, µq.
X

Show that x¨, ¨y is well defined, is a scalar product and it turns L2 pX, B, µq into a
Hilbert space.
A Banach space is said to be Hibertlizable when one can define a scalar product
on it that turns it into a Hilbert space and whose topology coincides with the original
one given by the Banach structure.
Show that the Lebesgues space L p pX, B, µq is Hilbertlizable if and only if p “ 2.
Given a probability space pX, B, µq, a measure preserving map f : X ý, and a
measurbale function ϕ : X Ñ C, consider the new measurable function U f ϕ :“ ϕ ˝ f .
The map U f is called the Koopman operator.
Exercise 23. For every p P r1, 8s, show that U f ϕ P L p pX, B, µq, whenever ϕ P
› ›
L p pX, B, µq. Moreover, show that U f : L p pX, B, µq ý is a linear isometry, i.e. ›U f ϕ› “
› ›
p
}ϕ} p , for every ϕ P L p pX, B, µq and every p P r1, 8s.

8.2. Revision of Hilbert space theory. Let pH , x¨, ¨yq be a Hilbert over the field
K “ R, C. We will write E ă H when E is a closed linear subspace. Given any
non-empty subset F Ă H , we define its orthogonal complement FK by
FK :“ tx P H : xx, yy “ 0, @y P Fu.
Observe that FK ă H and F X FK is either empty or the singleton t0u.
We will need the following
16 ALEJANDRO KOCSARD

Theorem 8.2 (Projection on convex subsets). Let K Ă H be a non-empty convex

closed set. Then there is a continuous map PK : H Ñ K such that for every x P H
the vector PK pxq is the only one in K such that
}x ´ Pk pxq} “ inft}x ´ y} : y P Ku.
Moreover, if K is closed linear subspace, then PK is a linear continuous map such
that
xx ´ PK pxq, xy “ 0, @x P H .
In particular, it holds ker PK “ KK and K ‘ KK “ H .
Proof. See for instance [EW17, Theorem 3.13 and Corollary 3.17]. □
Let H ‹ denote its dual space, i.e. the space of continuous linear functionals on
H . We have the following
Theorem 8.3 (Riesz representation theorem for Hilbert spaces). For every η P H ‹ ,
there exists a unique vη P H such that
ηpxq “ xx, vη y, @x P H .
Proof. See for instance [EW17, Corollary 3.19]. □
Theorem 8.3 lets us redefine the concept of adjoint operator on Hilbert spaces.
Given two Hilbert spaces pH , x¨, ¨yH q and pJ , x¨, ¨yJ q over K “ R, C, and a
continuous linear operator A P LpH , J q, we define its adjoint operator A‹ : J Ñ
H by the following property:
xAx, yy “ xx, A‹ yy, @x P H , @y P J .
Exercise 24. Show that A‹ P LpJ , H q, for every A P LpH , J q.
A linear operator U P LpH , J q is said to be an isometry when
}Ux}J “ }x}H , @x P H .
Exercise 25. Let U : H Ñ J be a continuous linear operator. Show that the
following statements are equivalent:
(1) U is an isometry;
(2) xUx, UyyJ “ xx, yxH , for every x, y P H ;
(3) U ‹ U “ I, where I P LpH q denotes the identity map.
A surjective isometry U P LpH q is called a unitary operator. One can easily
check that U is a unitary operator if and only if U ‹ U “ UU ‹ “ I.
Now we can prove the so called von Neumann’s ergodic theorem for isometries
on Hilbert spacees:
Theorem 8.4. Let H be a Hilbert space over K “ R, C and U : H Ñ H be a linear
isometry. Let I :“ kerpU ´ Iq ă H and PI : H Ñ I the projection map given
by Theorem 8.2. Then, it holds
n ´1
1 ÿ i
U x Ñ PI pxq, @x P H ,
n
i “0
as n Ñ `8.
Proof. Let us start considering the closed linear space
E “ pU ´ IqH .
Let us show that I “ EK . In fact, first observe that for every x P H and any
y P I we have
xUx ´ x, yy “ xUx, yy ´ xx, yy “ xUx, Uyy ´ xx, yy “ 0,
 ERGODIC THEORY 17

where the last equality follows from the fact that U is an isometry and Exercise 25.
So, I Ă EK .
On the other hand, if y P EK , then we have
0 “ xUx ´ x, yy “ xUx, yy ´ xx, yy “ xx, U ‹ yy ´ xx, yy “ xx, U ‹ y ´ yy,
for every x P H . So, by Theorem 8.3 we have that U ‹ y ´ y “ 0. Then we have

}Uy ´ y}2 “ xUy ´ y, Uy ´ yy “ }Uy}2 ´ xUy, yy ´ xy, Uyy ` }y}2

“ }y}2 ´ xy, U ‹ yy ´ xU ‹ y, yy ´ }y}2
“ }y}2 ´ }y}2 ´ }y}2 ` }y}2 “ 0.

So, y P I and we have proved that I “ EK . Hence, it holds H “ I ‘ E and in

order to conclude the proof it is enough to show that
n ´1
1 ÿ i
(10) U x Ñ 0, as n Ñ `8,
n
i “0

for every x P E. To do that, let x denote an arbitrary point of E and ε ą 0 be any

positive real number. So, there exists a y P H such that
ε
}x ´ pUy ´ yq} ă ,
2
and we get
› › › ›
› 1 nÿ ´1 › 1 ›nÿ ´1 ´` ¯›
U i x› “ › U i Uy ´ y ` x ´ pUy ´ yq ›
› › › ˘ ›
›
›n › n› ›
i “0 i“0
› ›
n ´1 › 1 nÿ´1 › `
1 ›› ÿ i`1 ˘››
U y ´ U i y› `
› i
›U x ´ pUy ´ yq ›
›
ď ›
n› › n
i“0 i “0
nÿ
´1
}U n y ´ y} 1 2 ε
“ ` }x ´ pUy ´ yq} ď }y} ` ď ε,
n n n 2
i “0

4 }y}
for every n ě . Since ε ą 0 is arbitrary, we conclude that (10) holds. □
ε
Proof of Theorem 8.1. It is a straightforward consequence of Theorem 8.4 and the
fact that the Koopman operator U f : L2 pX, B, µq ý is an isometry (Exercise 23).
□

Here we review some interesting consequences of von Neumann’s ergodic theorem

(Theorem 8.1):
Corollary 8.5. Let pX, B, µq be a probability space and f : X ý be a measure preserv-
ing automorphism. Then, for every ϕ P L2 pX, B, µq, there exists ϕ̃ P L2 pX, B, µq
such that
n ´1 n ´1
1 ÿ 1 ÿ
lim ϕ ˝ f i “ lim ϕ ˝ f ´i “ ϕ̃,
nÑ`8 n nÑ´8 n
j “0 i “0

Where both limits are taken in L2 .

Proof. This is a straightforward consequence of Theorem 8.4 and the simple fact
that U f and U f ´1 “ U f ´1 are isometries, and kerpU f ´ Iq “ kerpU f ´1 ´ Iq □
18 ALEJANDRO KOCSARD

Corollary 8.6. Let pX, B, µq be a probability space, f : X ý be measure preserving

map and ϕ P L1 pX, B, µq. Then, there is ϕ̃ P L1 pX, B, µq such that
n ´1
1 ÿ
ϕ ˝ f i Ñ ϕ̃, in L1 ,
n
i “0
as n Ñ `8.
Proof. For the sake of simplicity of notation, given any ψ : X Ñ R and any n P N,
let us write S nf ψ for the Birkhoff average of order n of ψ over f , i.e. we have
n ´1
1 ÿ
S nf ψ “ ψ ˝ f j.
n
j “0

Let us fix a real number ε ą 0. Since L2 pX, B, µq is dense in L1 pX, B, µq, there
is ξ P L2 pX, B, µq such that
ε
}ϕ ´ ξ}1 ă ,
4
and then,
› › ε
› n
›S f ϕ ´ S nf ξ › ă , @n ě 1.
›
1 4
On the other hand, by Theorem 8.1 we know that there is ξ̃ P L2 pX, B, µq such
that › ›
› n
›S f ξ ´ ξ̃ › Ñ 0, as n Ñ `8,
›
2
and thus, there is N P N such that
› › › ›
› n
›S f ξ ´ ξ̃ › ď ›S nf ξ ´ ξ̃ › ă ε{4, @n ě N.
› › ›
1 2
Putting together all these estimates we get
› › › › › › › › › ›
› m
›S f ϕ ´ S nf ϕ› ď ›S m m › › m n › › n n ›
S S S S S
› › › ›
f ϕ ´ f ξ › ` › f ξ ´ ξ̃ › ` › ξ̃ ´ f ξ › ` › f ξ ´ f › ă ε,
ϕ
1 1 1 1 1

S nf ϕ L1 pX, B, µq
` ˘
for every m, n ě N. So, is a Cauchy sequence in and hence,
n ě1
there exists its limit ϕ̃ P L pX, B, µq.
1 □

Now we are going to prove that indeed pointwise convergence holds on Corol-
lary 8.6:
Theorem 8.7 (Birkhoff ergodic theorem). Let pX, B, µq be a probability space, f : X ý
a measure preserving map and ϕ : X Ñ R any integrable function. Then there is
ϕ̃ P L1 pX, B, µq such that
n ´1
1 ÿ ` j ˘
S nf ϕpxq “ ϕ f pxq Ñ ϕ̃pxq, as n Ñ 8,
n
j “0

for µ-almost every x P X.

Moreover, it holds
ż ż
ϕ dµ “ ϕ̃ dµ,
X X
and
` ˘
ϕ̃pxq “ ϕ̃ f pxq , for µ-a.e. x P X.
In order to prove Theorem 8.7 we first need some preelimiary results:
 ERGODIC THEORY 19

Lemma 8.8 (Borel-Cantelli). Let pX, B, µq be a probability space and A1 , A2 , . . . P B

be a sequence of measurable sets. If
ÿ
µpAn q ă 8,
ně1
then it holds ´ ¯
µ lim sup An “ 0,
n
where the limit superior of the sequence pAn q is defined by
č ď
lim sup An :“ An .
n m ě1 n ě m

Proof. For each m P N, let us consider the set

ď
Bm :“ An .
něm
ř
Observe that B1 Ą B2 Ą ¨ ¨ ¨ and since ně1 µpAn q ă 8, it holds µpBm q ď
ř
něm µpAn q Ñ 0, as m Ñ 8. Finally, notice that
´ ¯ ˆ č ˙
µ lim sup An “ µ Bm “ lim µpBm q “ 0.
n mÑ8
m ě1
□
Lemma 8.9. Let pX, B, µq be a probability space and ϕ P L1 pX, B, µq. Then it holds
ϕ f n pxq
` ˘
lim “ 0, for µ-a.e. x P X.
nÑ`8 n
Proof. For the sake of simplicity of notation, given any ψ : X Ñ R and any A Ă R
and any r P R let us write
rψ P As :“ ψ´1 pAq, and rψ ą rs :“ rψ P pr, 8qs.
Then, first observe that given an arbitrary point x P X, the sequence ϕp f n pxqq{n
does not converge to 0 if and only if the limit superior of |ϕp f n pxqq| {n is positive.
So, we have
# +
ϕ f n pxq
` ˘
ď č ď „ |ϕ ˝ f n | 1
ȷ
xPX: ­Ñ 0, as n Ñ `8 “ ą .
n něm
n p
pě1 mě1

Thus in order to prove this lemma, it is enough to show that given any ε ą 0,
|ϕ ˝ f n |
ˆ „ ȷ˙
µ lim sup ąε “ 0,
n n
and by Lemma 8.8, to prove this last equality it is enough to show that
ÿ „ |ϕ ˝ f n | ȷ
µ ą ε ă 8.
n
n ě1
Finally, to prove the convergence of this series, observe that
ÿ „ |ϕ ˝ f n | ȷ ÿ ˆ „ ȷ˙ ÿ „ |ϕ| ȷ
´n | ϕ |
µ ąε “ µ f ąε “ µ ąε
n n n
n ě1 n ě1 n ě1
ÿ „ |ϕ| ȷ ÿ ÿ „ |ϕ|
ȷ
µ ąn “ µ jă ď j`1
ε ε
n ě1 n ě1 j ě n
ÿ „ |ϕ|
ȷ ż
|ϕ|
“ jµ j ă ď j`1 ď dµ ă 8.
ε X ε
j ě1
20 ALEJANDRO KOCSARD

□
Theorem 8.10 (Maximal ergodic theorem). Let pX, B, µq, f and ϕ as in Theo-
rem 8.7. For any α ą 0, let us consider the set
« ff $ ,
nÿ
´1 `
& 1 .
Eα pϕq :“ sup S nf ϕ ą α “ x P X : sup ϕ f j pxq ą α .
˘
ně1 % n ě1 n -
j“0

Then it holds ż
` ˘
ϕ dµ ě αµ Eα pϕq .
E α pϕq

Proof of Theorem 8.10. Let us start noticing that given any α P R and any ϕ P
L1 pX, B, µq, it holds
Eα pϕq “ E0 pϕ ´ αq,
and therefore,
ż ż
ϕ ´ α dµ ě 0 ðñ ϕ dµ ě αµpEα pϕqq.
E0 pϕ´αq E α pϕq
So, it is enough to prove this theorem for α “ 0.
For each n ě 1, let us consider the function
$ ,
&kÿ´1 ` .
ϕ f j pxq : 1 ď k ď n ,
˘
ϕn pxq :“ sup @x P X.
% -
j “0

If we write rϕn ą 0s :“ tx P X : ϕn pxq ą 0u, so it holds rϕn ą 0s Ă rϕn`1 ą 0s,

for every n ě 1 and ď
E0 pϕq :“ rϕn ą 0s.
n ě1
ş
Thus, to prove that E0 pϕq ϕ dµ ě 0, by the dominated convergence theorem it is
enough to show that ż
ϕ dµ ě 0, @n ě 1.
rϕn ą0s
Then, given any n ě 1 and any m P t1, 2, . . . , nu, we know that
mÿ
´1
ϕ f j pxq ,
` ˘
ϕn pxq ě @x P X,
j“0

and hence,
m
ÿ
ϕ f j pxq ,
` ˘ ` ˘
(11) ϕn f pxq ` ϕpxq ě @x P X, @m P t1, . . . , nu.
j “0
` ˘
Fixing
` a˘point x P X, let us consider the two possible cases: either ϕn f pxq ě 0
or ϕn f pxq ă 0. ` ˘ ` ˘
In the first case, when ϕn f pxq ě 0, we have ϕn f pxq ` ϕpxq ě ϕpxq. Then, by
estimate (11) it follows that
` ˘
ϕn f pxq ` ϕpxq ě ϕn pxq, @x P rϕn ˝ f ě 0s.
` ˘
In the second case, when ϕn f pxq ă 0, by (11) we get
m
ÿ m
˘ ÿ
ϕ f j pxq ´ ϕn f pxq ą ϕ f j pxq ,
` ˘ ` ` ˘
ϕpxq ě @m P t1, . . . , nu,
j “0 j “0

and then, ϕn pxq “ ϕpxq, whenever x P rϕn ˝ f ă 0s.

 ERGODIC THEORY 21

Putting together these estimates we get

ż ż ż
ϕ dµ “ ϕ dµ ` ϕ dµ
rϕn ą0s rϕn ą0sXrϕn ˝ f ă0s rϕn ą0sXrϕn ˝ f ě0s
ż ż
ě ϕ dµ ` ϕn ´ ϕn ˝ f dµ
rϕn ą0sXrϕn ˝ f ă0s rϕn ą0sXrϕn ˝ f ě0s
ż ż
“ ϕn dµ ` ϕn ´ ϕn ˝ f dµ
rϕn ą0sXrϕn ˝ f ă0s rϕn ą0sXrϕn ˝ f ě0s
ż ż
“ ϕn dµ ´ ϕn ˝ f dµ
rϕn ą0s rϕn ą0sXrϕn ˝ f ě0s
ż ż
ě ϕn dµ ´ ϕn ˝ f dµ
rϕn ą0s rϕn ˝ f ě0s
ż ż
“ ϕn dµ ´ ϕn dp f ‹ µq “ 0.
rϕn ą0s rϕn ě0s

Corollary 8.11. If pX, B, µq, f , α and ϕ are as in Theorem 8.10 and A P B satisfies
A Ă Eα pϕq and f ´1 pAq “ A, then it holds
ż
ϕ dµ ě αµpAq.
A

Exercise 26. Prove Corollary 8.11.

Now we are ready for

Proof of Theorem 8.7. Let us start considering the functions ϕ, ϕ : X Ñ r´8, `8s
given by
ϕpxq :“ lim sup S nf ϕpxq and ϕpxq :“ lim inf S nf ϕpxq,
nÑ`8 nÑ`8

for every x P X.
Exercise 27. Using Lemma 8.9, show that the functions ϕ and ϕ are (essentially)
f -invariant.

Our purpose now consists in showing that the functions ϕ and ϕ coincide µ-a.e.
In order to prove that, given real numbers α ă β, consider the set
! )
β
Eα :“ x P X : ϕpxq ă α ă β ă ϕpxq ,

and observe that

ď β
rϕ ‰ ϕs “ rϕ ă ϕs “ Eα
α,βPQ, αă β
` β˘
So, to prove that ϕ and ϕ coincide µ-a.e. it is enough to show that µ Eα “ 0, for
every pair of real numbers α ă β. ´ ¯
β β
Since ϕ and ϕ are f -invariant, it is easy to show that f ´1 Eα “ Eα up to sets
β
of µ-measure zero. Then, observing that Eα Ă E β pϕq, where the last set is defined
as in Theorem 8.10, applying Corollary 8.11 we can conclude
ż ´ ¯
β
β
ϕ dµ ě βµ Eα .
Eα
22 ALEJANDRO KOCSARD

β
On the other hand, notice that Eα Ă E´α p´ϕq. So, we can again apply Corol-
lary 8.11 to the function ´ϕ to get
ż ´ ¯
β
β
´ϕ dµ ě ´αµ Eα .
Eα

Putting together these estimates we conclude

´ ¯ ´ ¯
β β
βµ Eα ď αµ Eα ,
´ ¯
β
and since α ă β, this can only happend when µ Eα “ 0.
The fact that ϕ̃ :“ ϕ “ ϕ belongs to L1 pX, B, µq and
ż ż
ϕ dµ “ ϕ̃ dµ,
X X

will be proved in Corollary 8.13 □

There are some important consequences of Birkhoff ergodic theorem:

Corollary 8.12. If pX, B, µq and f are as in Theorem 8.7 and ϕ P L p pX, B, µq with
1 ď p ă 8, then the limit function ϕ̃ given by Theorem 8.7 belongs to L p pX, B, µq
and it holds
› ›
›ϕ̃› ď }ϕ} .
p p

Proof. Notice that

n´1
1 ÿ
|ϕ̃pxq| ď lim ϕ ˝ f j pxq , for µ-a.e. x P X,
nÑ`8 n
j“0

and hence,
¨ ˛p
nÿ
´1
p 1
|ϕ̃pxq| ď lim ˝ ϕ ˝ f j pxq ‚ .
nÑ`8 n
j“0

Then, as a consequence of Fatou lemma and Minkowski inequality we get

» ˛ p fi 1p
¨
„ż ȷ1 nÿ
´1 ` ż
p p 1
ϕ f j pxq ‚ dµfl
› › ˘
›ϕ̃› “
p |ϕ̃pxq| dµ ď lim inf – ˝
X n X n
j “0
n ´1 ˆ ż ˙1
1 ÿ p p
ϕ f j pxq
` ˘
ď lim inf dµ “ }ϕ} p .
n n X j “0

Corollary 8.13 (L p -convergence of ergodic averages). If pX, B, µq, f , ϕ and ϕ̃ are

as in Corollary 8.12, then
› ›
› n
›S f ϕ ´ ϕ̃› Ñ 0,
›
p

1 ř n ´1
as n Ñ `8,and where S nf ϕ “ ϕ ˝ f j . Therefore, it holds
n j “0
ż ż
ϕ dµ “ ϕ̃ dµ.
X X
 ERGODIC THEORY 23

Proof. By Corollary 8.13 we know that ϕ̃ P L p pX, B, µq.› Then› to prove the con-
verges in L , let us start assuming ϕ P L px, B, µq. So, ›S f ϕ› ď }ϕ}8 , for every
p 8 › n ›
8
n ě 1, and hence, ϕ̃ P L8 pX, B, µq, as well. So,
› › › ›
› n
›S f ϕ ´ ϕ̃› ď ›S nf ϕ ´ ϕ̃› ď }ϕ}8 ` ›ϕ̃›8 ,
› › › › ›
p 8

and S nf ϕ Ñ ϕ̃, µ-a.e. as n Ñ 8. So, by dominated convergence theorem, it holds

› ›
› n
›S f ϕ ´ ϕ̃› Ñ 0, as n Ñ 8.
›
p
Now, to prove the convergence in L p without the additional assumption that ϕ
is bounded, let us consider a sequence of truncations of ϕ, i.e. for each k P N let us
consider the function ϕpkq pxq :“ ϕpxq1ϕ´1 r´k,ks pxq. Observe that ϕpkq P L8 pX, B, µq
for every k and ϕpkq Ñ ϕ as k Ñ 8, µ-almost everywhere
› and ›in L p .
So, given any ε ą 0 there exists k P N such that ›ϕ ´ ϕ› ă ε{3, and hence,
› pk q ›
› › p
› n pk q
›S f pϕ ´ ϕq› ă ε{3, for every n ě 1, and by Corollary 8.13, it holds
›
p
› › › › ε
› pk q
›ϕ̃ ´ ϕ̃› ď ›ϕpkq ´ ϕ› ă .
› › ›
p p 3
› ›
Then, by the previous argument, we have ›S nf ϕpkq ´ ϕ̃pkq › ă ε{3, whenever n is
› ›
p
sufficiently large, and so we get
› › › › › › › ›
› n
›S f ϕ ´ ϕ̃› ď ›S nf ϕ ´ S nf ϕpkq › ` ›S nf ϕpkq ´ ϕ̃pkq › ` ›ϕ̃pkq ´ ϕ̃›
› › › › › › ›
p p p p
ε ε ε
ď ` ` “ ε,
3 3 3
for n large enough. ş ş
Finally, in order to prove that x ϕ dµ “ X ϕ̃ dµ, it is enough to notice that if
ϕ P L p pX, B, µq, then ϕ P L1 pX, B, µq and it holds
ż ż ż ż
n
ϕ dµ ´ ϕ̃ dµ “ S f ϕ dµ ´ ϕ dµ
X X
żX X
› ›
S f ϕ ´ ϕ̃ dµ “ ›S nf ϕ ´ ϕ̃› Ñ 0,
n › ›
ď
X 1

as n Ñ `8. □

Exercise 28. Show that Corollary 8.13 is false for p “ 8.

Corollary 8.14 (Birkhoff theorem for continuous functions). Let X be compact metric
space and f : X ý be a continuous map. Then there exists a Borel subset X 1 Ă X
such that µpX 1 q “ 1, for every µ P Mp f q and the following limit exists for every
x P X 1 and every ϕ P C0 pX, Rq:
n ´1
1 ÿ ` j ˘
ϕ̃pxq :“ lim ϕ f pxq .
nÑ`8 n
j “0

Proof. Since X is compact, the Banach space C0 pX, Rq is separable, i.e. there exists
a sequence pϕk qkě1 which is dense in C0 pX, Rq. For each k P N we define
$ ,
nÿ
´1 `
& 1 .
ϕk f j pxq exists .
˘
Xk :“ x P X : lim
% nÑ`8 n -
j “0
24 ALEJANDRO KOCSARD

The set Xk is clearly a Borel set and, by Birkhoff theorem (Theorem 8.7) we know
that µpXk q “ 1, for every k P N and all µ P Mp f q. So, if we define
č
X 1 :“ Xk ,
k PN

it clearly holds µpX 1 q “ 1, for every µ P MpXq.

Then, let us consider an arbitrary point x P X 1 and a positive real number ε ą 0.
Given any ϕ P C0 pX, Rq, by denseness of the sequence pϕk q we know there exists
k P N such that }ϕ ´ ϕk }0 ă ε{2. So we have
n´1 n ´1
1 ÿ ` j ˘ 1 ÿ ` j ˘
0 ď lim sup ϕ f pxq ´ lim inf ϕ f pxq
n n n n
j“0 j “0

n ´1 n ´1
1 ÿ ` j ˘ 1 ÿ ` j ˘
ď lim sup ϕ f pxq ´ lim sup ϕk f pxq
n n n n
j “0 j “0

n ´1 n´1
1 ÿ ` j ˘ 1 ÿ ` j ˘
` lim inf ϕk f pxq ´ lim inf ϕ f pxq
n n n n
j “0 j“0
ε ε
ă ` “ ε.
2 2
1 ř n ´1
Therefore, the limn ϕp f j pxqq exists and the result is proved. □
n j “0

9. Ergodicity
Let pX, B, µq be a probability space and f : X ý be a measure preserving map.
We say that p f , µq is ergodic when given any A P B such that f ´1 pAq “ A, either
µpAq “ 0 or µpAq “ 1.
Ergodic systems satisfy the following properties:
Theorem 9.1. Let pX, B, µq be a probability space and f : X ý be a measure pre-
serving map. Then the following statements are equivalent:
(1) p f , µq is ergodic;
(2) given any B P B such that µ B△ f ´1 pBq “ 0, then µpBq P t0, 1u;
` ˘

(3) if ϕ P X Ñ R is a measurable function such that

` ˘
ϕ f pxq “ ϕpxq, for µ-a.e. x P X,
then there exists c P R such that ϕpxq “ c, for µ-a.e. x P X;
(4) for every A, B P B it holds
n ´1
1 ÿ ` ´j
µ f pAq X Bq Ñ µpAqµpBq,
n
j “0

as n Ñ `8;
1 1
(5) given p, q P p1, 8q such that ` “ 1, and ϕ P L p pX, B, µq and ψ P
p q
Lq pX, B, µq, it holds
n ´1 ż ż ż
1 ÿ
pϕ ˝ f j qψ dµ Ñ ϕ dµ ψ dµ,
n X X X
j “0

as n Ñ `8.
 ERGODIC THEORY 25

Exercise 29. Let pX, B, µq be a probability space and f : X ý˘be measure preserving
map. Let B P B be a measurable set such that µ B△ f ´1 pBq “ 0. Show that there
`

exists A P B satisfying f ´1 pAq “ A and µpA△ Bq “ 0.

Proof of Theorem 9.1. Clearly (2) implies (1). On the other hand, by Exercise 29
we know that (1) implies (2).
If p f , µq is not ergodic, then there exists A P B with f ´1 pAq “ A and 0 ă µpAq ă
1. Then, 1 A ˝ f “ 1 f ´1 p Aq “ 1 A and 1 A is not constant µ-a.e. So, (3) implies (1).
` ˘
On the other hand, if ϕ f pxq “ ϕpxq for µ-a.e. point x P X and ϕ is not constant
µ-almost everywhere,` then there ˘exists r P R such that B :“ ϕ´1 r´8, rs satisfies
0 ă µpBq ă 1, with µ B△ f ´1 pBq “ 0. Thus, (2) implies (3).
Then, if (4) holds and A P B is f -invariant, then we have
n ´1
1 ÿ `
µ A X f ´1 pAq Ñ µpAq2 ,
˘
µpAq “ as n Ñ `8.
n
j “0

So µpAq P t0, 1u and (1) holds.

On the other hand, assuming (5) we can easily prove (4) taking ϕ :“ 1 A and
ψ :“ 1B .
Finally, let us prove that (1) implies (5). To do that, let us star assuming
ϕ P L8 pX, B, µq and ψ P Lq pX, B, µq. Then, by Birkhoff ergodic theorem (Theo-
rem 8.7), and since (1) implies (3), we know that
n ´1 ż
1 ÿ ` i ˘
ϕ f pxq Ñ ϕ dµ, as n Ñ `8,
n X
i “0
and for µ-a.e. x P X, and we have
n ´1
1 ÿ ` i ˘
ϕ f pxq ψpxq ď }ϕ}8 |ψpxq| , @n P N,
n
i “0
where |ψ| is an integrable function. So, by dominated converge theorem we conclude
that
n ´1 ż ˆż ˙ ˆż ˙
1 ÿ i
lim pϕ ˝ f qψ dµ “ ϕ dµ ψ dµ ,
nÑ`8 n x x X
i “0
we wanted to prove. The general case, where ϕ P L p pX, B, µq can be proved taking
the sequence of truncations of ϕ given by ϕk :“ 1r|ϕ|ďks ϕ P L8 pX, B, µq, for every
k P N. Then we have ϕk Ñ ϕ in L p and µ-almost everywhere. Then, given any
ε ą 0 there is k P N such that }ϕ ´ ϕk } p ă ε and, by the previous argument, there
is N P N such that
n ´1 ż ż ż
1 ÿ
pϕ ˝ f i qψ dµ ´ ψ dµ ψ dµ ă ε, @n ě N.
n X X X
i “0
Hence we have
n ´1 ż ż ż
1 ÿ i
pϕ ˝ f qψ dµ ´ ϕ dµ ψ dµ
n X X X
i “0
n´1 ż n ´1 ż ż ż
1 ÿ 1 ÿ
pϕ ´ ϕk q ˝ f i ψ dµ ` pϕk ˝ f i qψ dµ ´
` ˘
ď ϕk dµ ψ dµ
n X n X X X
i “0 i “0
ż ż
` pϕk ´ ϕq dµ ψ dµ ď }ϕ ´ ϕk } p }ψ}q ` ε ` }ϕk ´ ϕ}1 }ψ}1
X X
` ˘
ď ε }ψ}q ` ε ` }ϕk ´ ϕ} p }ψ}q ď ε 2 }ψ}q ` 1 ,
26 ALEJANDRO KOCSARD

for every n ě N. □
We shall need a little improvement of condition (3) of Theorem 9.1:
Exercise 30. Let pX, B, µq be a probability space and f : X ý be a measure preserving
map, and let A Ă 2X be an algebra such that B is the σ-algebra generated by A
and suppose that
n ´1
1 ÿ ` ´i ˘
µ f pAq X B Ñ µpAqµpBq, as n Ñ `8,
n
i “0
for every A, B P A. Then, show that p f , µq is ergodic.
9.1. Examples of ergodic systems.
9.1.1. Circle rotations. Given any α P R, we define the rotation Tα : R{Z Ñ R{Z
by
Tα px ` Zq :“ x ` α ` Z, @x P R.
Since R{Z is a compact group, its Haar measure λ is Tα -invariant for every
α P R.
Exercise 31. Let π : R Ñ R{Z be a´ the natural
¯ quotient map, i.e. πpxq :“ x ` Z,
for every x P R. Show that λ “ π‹ Leb r0,1s .
ˇ
ˇ

Exercise 32. Given α P Q, show that pTα , λq is not ergodic.

Theorem 9.2. The system pTα , λq is ergodic for every α P RzQ.
9.1.2. Expanding maps on the interval. Given any k P N with k ě 2, consider the
map f : r0, 1s ý given by
f pxq :“ kx ´ tkxu ,
@x P r0, 1s.
ˇ
We have shown in §5.2 that the measure λ :“ Lebˇr0,1s is f -invariant. Moreover,
we have
Theorem 9.3. The system p f , λq is ergodic.
9.1.3. Bernoulli shifts.

10. Unique ergodicity

11. Ergodic decomposition
11.1. Ergodic measures as extreme points. Let pX, Bq be a measurable space and
f : X ý be a measurable map. We first observe that ergodic measures are “minimal
elements” with respect to the preorder given by absolute continuity among invariant
measures:
Proposition 11.1. If µ and ν are two f -invariant probability measures such that µ
is ergodic and ν ! µ (i.e. νpAq “ 0, for every A P B such that µpAq “ 0), then
ν “ µ.
Proof. Given an arbitrary set A P B, by Birkhoff ergodic theorem and since µ is
ergodic, we get that there exists a measure R A P B with µpR A q “ 1 such that
n ´1
1 ÿ
1 A ˝ f j pxq Ñ µpAq, as n Ñ `8, @x P R A .
n
i“0
Since ν ! µ and µpXzR A q “ 0, we get νpXzR A q “ 0. So the above Birkhoff averages
converge to µpAq for ν-a.e. x P R and consequently, µpAq “ νpAq. Since A was
arbitrary, we get µ “ ν. □
 ERGODIC THEORY 27

We get the following

Corollary 11.2. Let pX, Bq be a measurable space and f : X ý be a measurable map.
If µ denotes an f -invariant probability measure, then µ is not ergodic if and only
there exist two f -invariant measures ν1 and ν2 with ν1 ‰ ν2 and t P p0, 1q such that
µ “ tν1 ` p1 ´ tqν2 .
Proof. Let us start assumung µ is not ergodic. Then there is A P B with 0 ă
1 1
µpAq ă 1. So, let us define ν! pEq :“ µpA X Eq and ν2 :“ µpEzAq, for
µpAq µpXzAq
every E P B, and it clearly holds
` ˘
µ “ µpAqν1 ` µpXzAqν2 “ µpAqν1 ` 1 ´ µpAq ν2 .
Reciprocally, let us suppose that µ is ergodic and there are f -invariant measures
ν1 and ν2 and t P p0, 1q such that µ “ tν1 ` p1 ´ tqν2 . Then it clearly holds ν1 ! µ
and ν2 ! µ. By Proposition 11.1, we know that µ “ ν1 “ ν2 . □
11.2. Partitions of measure spaces. Let pX, B, µq be an arbitrary probability space.
A partition of X is a family ξ Ă B such that C X D “ H, for every C, D P ξ with
C ‰ D and ˆğ ˙
µ C “ 1.
C Pξ
Ů
The elements of ξ are called atoms of ξ. Given a point x P CPξ C, we define
ξpxq P ξ as the only atom of ξ that contains the point x. So, we can define the
natural projection map π ξ : X Ñ ξ by π ξ pxq :“ ξpxq and notice that π ξ is well
defined µ-almost everywhere.
We can use the natural projection map π ξ to endow the set ξ with a measure
space structure. Indeed, we shall always assume the partition ξ as the supporting
set of the measure space pξ, Bξ , µ̂ξ q where Bξ :“ π‹ B and µ̂ξ :“ π‹ µ denote the
ξ ξ

push-forward σ-algebra of B and the measure µ by π ξ , respectively.

We can define a partial order relation among partitions of a measure space. Given
two partitions ξ, η of the measure space pX, B, µq, we say that ξ is finer than η, or
that η is coarser that ξ, and we write η ď ξ, when there is a measurable subset
X 1 Ă X with µpX 1 q “ 1 and such that
ˆğ ˙ ˆğ ˙
1 1
ξpxq X X Ă ηpxq X X , @x P C X D .
C Pξ D Pη

We say that ξ and η coincide or are equal when η ď ξ and ξ ď η.

If we write P X to denote the set of all partitions of the measure space pX, B, µq,
we know that pP X , ďq is a partially order set.
Exercise 33. Show that the notions of infimum and supremum are well defined for
countable subsets on the partially order set pP X , ďq. More precisely, show that given
any countable family of partitions Ξ Ă P X , there exists two partitions inf Ξ, sup Ξ P
P X such that
η ď inf Ξ ď ξ ď sup Ξ ď η 1 , @ξ P Ξ,
and every η, η 1 P P X such that η ď ξ 1 ď η 1 , for every ξ 1 P P X .
We can use the notions of infimum and supremum on P X to define the operators
^ and _: given any countable (or finite) family Ξ Ă P X , we define
ľ
ξ :“ sup tη P P X : η ď ξ, @ξ P Ξu ,
ξ PΞ
28 ALEJANDRO KOCSARD

and
ł
(12) ξ :“ inf tη P P X : ξ ď η, @ξ P Ξu .
ξ PΞ

11.3. Disintegration of measures. Given a probability space pX, B, µq and a partion

ξ P P X , a disintegration of µ with respect to ξ is family tµC : C P ξu satisfying the
following properties:
(i) µC is a probability measure on the measurable space pX, Bq, for every C P ξ;
(ii) µC pCq “ 1, for µ̂ξ -almost every atom C P ξ;
(iii) for every E P B, the map ξ Q C ÞÑ µC pEq P r0, 1s is pB ξ , BR q-measurable;
(iv) it holds ż
µpEq “ µC pEq dµ̂ξ pCq, @E P B.
ξ
Let us observe that one can pull back the partion and push forward the disinte-
grations:
Exercise 34. Let pX, B, µq be a probability space, ξ P P X be a partion and f : X ý be
a measure preserving map. Then, show that f ‹ ξ :“ t f ´1 pCq : C P ξu is a partition
of pX, B, µq. On the other hand, if tµ D : D P f ‹ ξu is a disintegration of µ with
respect to the partition f ‹ ξ, show that tµC :“ f ‹ µ f ´1 pCq : C P ξu is a disintegration
of µ with respect to ξ.
First we shall prove that every partition admits at most one disintegration:
Theorem 11.3 (Uniqueness of disintegrations). If pX, B, µq is a Borel space (see
§3), ξ P P X is a partition, and tµC : C P ξu and tνC : C P ξu are two disintegrations
of µ with respect to ξ, then it holds
µC “ νC , for µ̂ξ -a.e. C P ξ.
Proof. Since pX, B, µq is a Borel space, there exists a countable algebra A Ă B
that generates B. For each A P A , let us consider the sets
M A :“ tC P ξ : µC pAq ă νC pAqu, and N A :“ tC P ξ : µC pAq ą νC pAqu.
Then, since tµC : C P ξu is a disintegration of µ with respect to ξ, we have
ż
ξ ´1 ´1
` ˘ ` ˘
µ AXπ pM A q “ µC A X π ξ pM A q dµ̂ξ pCq
ξ
ż
“ µC pAq dµ̂ξ pCq.
MA
Analogously, one can show that
ż
ξ ´1
` ˘
µ AXπ pM A q “ νC pAq dµ̂ξ pCq.
MA
Since µC pAq ă νC pAq, for all C P M A , we get µ̂ξ pM A q “ 0. Analogously, we can
show that µ̂ξ pN A q “ 0 and hence, µC “ νC , for µ̂ξ -almost every C P ξ. □
Theorem 11.3 can be used to show that certain partitions do not admit any
disintegration of the measure:
Example 11.4. Let α be an arbitrary real irrational number and f : R{Z ý be the
corresponding rotation f px ` Zq :“ x ` α ` Z, for every x P R. Let λ denote the
Haar measure on the circle R{Z and consider the partition ξ of the probability
space pR{Z, BR{Z , λq given by the orbits of f , i.e. we have
ξ :“ tt f n pxq P R{Z : n P Zu : x P R{Zu .
Then there is no disintegration of λ with respect to the partion ξ.
 ERGODIC THEORY 29

Now we can state the main result of this section:

Theorem 11.5 (Ergodic decomposition theorem). Let pX, B, µq be a Borel space and
f : X ý be a measure preserving map. Then there exists a unique partition ξ f P P X ,
called the ergodic partition of f , such ξ f admits a disintegration tµC : C P ξ f u
and the following conditions hold:
(1) f ‹ ξ f “ ξ f , i.e. f ´1 pCq “ C, for µ̂ξ f -almost every atom C P ξ f ;
(2) µC is an ergodic f -invariant measure, for µ̂ξ f -almost every atom C P ξ f .
The proof of Theorem 11.5 will be postpone and get as a consequence of Rokhlin
decomposition theorem.
11.4. Measurable partitions. By Example 11.4 we know that some partitions do not
admit disintegrations. In this section we shall characterize the family of partitions
admitting disintegrations.
Given a probability space pX, B, µq we say that a partition ξ P P X is a measur-
able partition when there exists a sequence pξ n qně1 Ă P X of finite partitions such
that ł
ξ“ ξn,
n ě1
Ž
where the equality among partitions is given in §11.2 and the operator is given
by (12)
Exercise 35. Let pX, B, µq be a probability space, M be separable metric space and
f : X Ñ M be a measurable map. Then the partition
! )
ξ f :“ f ´1 pyq : y P M P P X
is a measurable partion.
Then we have the main result about measurable partitions:
Theorem 11.6 (Rokhlin theorem). Let pX, B, µq be a Borel space and ξ P P X be
measurable partition. Then there exits a disintegration of µ with respect to ξ.
We postpone the proof of Theorem 11.6. Now we shall prove the Ergodic de-
composition theorem assuming Theorem 11.6.
Proof of Theorem 11.5. Since pX, B, µq is a Borel space, there exists a countable
algebra A Ă B that generates the σ-algebra B. For each A P A , let us consider
the function τA : X Ñ r0, 1s given by
n ´1
1 ÿ
1 A f i pxq ,
` ˘
τA pxq :“ lim
nÑ`8 n
i “0
whenever the limit exists. By Birkhoff ergodic theorem we know that τA is well
defined µ-almost everywhere.
Since A is countable, we get the set X0 :“ tx P X : τA pxq is well defined, @A P
A u is a measurable set with µpX0 q “ 1.
Then we define the partition ξ f P P X as the set of equivalence classes on X0 such
that x „ y if and only if τA pxq “ τA pyq, for every A P A .
Exercise 36. Show that ξ f is a measurable partition.
Then, by Rokhlin theorem (Theorem 11.6) we know that there is a disintegration
of µ with respect to ξ f tµC : C P ξ f u.
By the very same definition of functions τA we know that f ´1 pX0 q “ X0 and
τA pxq “ τA ˝ f pxq, for every x P X0 . So, it clearly holds f ´1 pCq “ C, for every
30 ALEJANDRO KOCSARD

C P ξ f . Thus, by uniqueness of disintegrations (Theorem 11.3), we conclude that

f ‹ µC “ µC , for µ̂ξ f -almost every atom C P ξ f .
So, it just remains to prove that the measure µC is ergodic for f , for every atom
C P ξ f . To do that, let us fix an atom C P ξ f and observe that to prove the
ergodicity of µC is enough to show that τE pxq “ µC pEq, for every E P B and µC -a.e.
x P C. Then, let us define the family
BC :“ tE P B : τE pxq “ µC pEq, for µC -a.e. x P Cu .
By our definition it clearly holds A Ă BC . On the other hand, if E1 , E2 P BC , and
E1 Ă E2 , then E2 zE1 P BC because
τE2 zE1 pxq “ τE2 pxq ´ τE1 pxq “ µC pE2 zE1 q,
for µC -almost every x P C; and if E1 , E2 , . . . P BC is a family of pairwise disjoint
sets, the it clearly holds
ÿ ˆğ ˙
τ j Ej pxq “
Ů τEj pxq “ µC Ej .
j j

Then, the ergodicity of µC is consequence of the following

Exercise 37. A family C Ă 2X is called a monotone class when given any sequence
A1 Ă AŞ2 Ă . . . and B1 Ą B2 Ą . . ., with An , Bn P C , for every n P N, it holds
Ť
n An , n Bn P C .
Then, if C is a monotone class and contains an algebra A , then C contains the
σ-algebra σpA q generated by A .
□
11.5. Conditional expectations. In order to prove Theore 11.6 we first need to con-
sider the notion of conditional expectations. Let pX, B, µq denote an arbitrary prob-
ability space and let A Ă B be a sub-σ-algebra of B. A real function ϕ : X Ñ R
is said to be A -measurable when ϕ´1 pEq P A for every measurable subset E Ă R.
Theorem 11.7 (Conditional expectation). There exists a continuous linear operator
Ep¨|A q : L1 pX, B, µq Ñ L1 pX, A , µq such that
ż ż
ϕ dµ “ Epϕ|A q dµ, @A P A .
A A
Proof of Theorem 11.7. Given an arbitrary function ϕ : X Ñ R, we can write ϕ “
ϕ` ´ ϕ´ , where ϕ` :“ maxpϕ, 0q and ϕ´ :“ maxp´ϕ, 0q. So, in order to define the
linear operator Ep¨|A q it is enough to define it on non-negative functions.
So, let us suppose ϕ P L1 pX, B, µq and ϕ ě 0. Consider the measure νϕ on the
measurable space pX, A q given by
ż
νϕ pAq :“ ϕ dµ, @A P A .
A

Since νϕ ! µˇA , by Radon-Nikodym theorem we know that there is Epϕ|A q P

ˇ

L1 pX, A , µq such that

ż ż
νϕ pAq “ ϕ dµ “ Epϕ|A q dµ, @A P A .
A A

Exercise 38. Show that Ep¨|A q can be uniquely extended as a linear operator to the
whole space L1 pX, B, µq and this extension is continuous.
□
 ERGODIC THEORY 31

References
[EW11] M. Einsiedler and T. Ward, Ergodic theory with a view towards number theory, Graduate
Texts in Mathematics, vol. 259, Springer-Verlag London, Ltd., London, 2011.
[EW17] , Functional analysis, spectral theory, and applications, Graduate Texts in Math-
ematics, vol. 276, Springer, Cham, 2017.
[Hal50] P. Halmos, Measure Theory, D. Van Nostrand Company, Inc., New York, N. Y., 1950.
[Mañ87] R. Mañé, Ergodic theory and differentiable dynamics, Ergebnisse der Mathematik und
ihrer Grenzgebiete (3), vol. 8, Springer-Verlag, Berlin, 1987, Translated from the Por-
tuguese by Silvio Levy.
[Sch17] R. Schilling, Measures, integrals and martingales, second ed., Cambridge University
Press, Cambridge, 2017.
[VO16] M. Viana and K. Oliveira, Foundations of ergodic theory, Cambridge Studies in Advanced
Mathematics, vol. 151, Cambridge University Press, Cambridge, 2016.

IMPA. Estrada Dona Castorina, 110. Rio de Janeiro, 22460–320, Brazil.

Email address: alejandro@impa.br