Lecture 3

In this lecture we prove a general version of von Neumann’s famous mean ergodic
theorem, an assertion about the long-term behavior of measure-preserving transfor-
mations. As an application we show Khintchin’s refinement of Poincaré’s recurrence
theorem from the first lecture. In the second part of the lecture we then explore the
connection between ergodic theory and topological dynamics.

3.1 Averaging and the Mean Ergodic Theorem

Mathematical ergodic theory originated in statistical mechanics. A measure-pre-
serving map τ : X → X can model how a state x of a physical dynamical system
(describing, e.g., the position and impulse of all particles of a gas or fluid) is trans-
formed, by the principles of physics, into a new state τ (x). From this perspective, a
function f : X → C can be seen as an “observable” giving some measurable data f (x)
about any given state x of the system. By making measurements “after each time
step”, we obtain a sequence of observables f, f ◦ τ, f ◦ τ 2 , f ◦ τ 3 , . . . Von Neumann’s
famous mean ergodic theorem from 1931 provides information on the asymptotic
behavior of this sequence.

Theorem 3.1.1P(von Neumann). Let τ : X → X be a measure-preserving map.

N −1 n
Then limN →∞ N1 n=0 Uτ f exists in L2 (X) for every f ∈ L2 (X) and is an element
of fix(Uτ ).
We deduce Theorem 3.1.1 from a general operator theoretic result. The necessary
notation, concepts and tools from basic Hilbert space theory are collected in Ap-
pendix A.2. We start from the following definition.

Definition 3.1.2. Let H be a Hilbert space. A family S ⊆ L (H) is a semigroup

(of operators) if U V ∈ S for all U, V ∈ S . It is a contraction semigroup if,
in addition, ∥U ∥ ≤ 1 for all U ∈ S .
Generalizing the notion of the fixed space from Example 2.2.12, we introduce the

35
36 LECTURE 3.

following definition.

Definition 3.1.3. For a semigroup S ⊆ L (H) on a Hilbert space H we call

\
fix(S ) := fix(U ) = {f ∈ H | U f = f for every U ∈ S }
U ∈S

the fixed space of S .

Remark 3.1.4. For every bounded linear operator U ∈ L (H) on a Hilbert space
H we obtain a semigroup SU := {U n | n ∈ N0 }. Since the operator norm is
submultiplicative, SU is a contraction semigroup precisely when ∥U ∥ ≤ 1. Note
further that fix(SU ) = fix(U ).
As recalled in the Appendix (see Theorem A.2.2), every closed linear subspace E ⊆
H of a Hilbert space can be orthogonally complemented to obtain a decomposition
H = E ⊕ E ⊥ . For a contraction semigroup S ⊆ L (H) on a Hilbert space H there
is a nice description of the orthogonal complement of the fixed space fix(S ) in terms
of the ranges of the operators IdH − U for U ∈ S . Here and in the following, for a
subset A ⊆ H of a Hilbert space H,
(i) the closed convex hull co A is the closure of the set of all convex combinations
of elements of A, and
(ii) the closed linear hull lin A is the closure of the set of all linear combinations
of elements of A.

Theorem 3.1.5 (Abstract Mean Ergodic Theorem). For every contraction semi-
group S ⊆ L (H) on a Hilbert space H we have an orthogonal decomposition
[
H = fix(S ) ⊕ lin (IdH − U )(H).
U ∈S

Moreover, the orthogonal projection P onto fix(S ) has the following properties.
(i) P U = U P = P for every U ∈ S .
(ii) For every f ∈ H the vector P f ∈ H is the unique element of co {U f | U ∈ S }
of minimal norm.
The proof uses the following elementary observation on the adjoint operator of a
contraction.

Lemma 3.1.6. Let U ∈ L (H) with ∥U ∥ ≤ 1. Then fix(U ) = fix(U ∗ ).

Proof. Recall that the adjoint operator U ∗ ∈ L (H) satisfies ∥U ∗ ∥ = ∥U ∥ ≤ 1. We

first show the inclusion “⊆”. By the “Pythagorean Theorem” we obtain for f ∈ fix(U )
3.1. AVERAGING AND THE MEAN ERGODIC THEOREM 37

that

0 ≤ ∥U ∗ f − f ∥2 = ∥U ∗ f ∥2 − 2Re (U ∗ f |f ) + ∥f ∥2
= ∥U ∗ f ∥2 − 2Re (f |U f ) + ∥f ∥2 = ∥U ∗ f ∥2 − 2∥f ∥2 + ∥f ∥2 = ∥U ∗ f ∥2 − ∥f ∥2
≤ ∥f ∥2 − ∥f ∥2 = 0.

This yields U ∗ f = f . Thus, fix(U ) ⊆ fix(U ∗ ). Since (U ∗ )∗ = U , the same argument

applied to U ∗ shows fix(U ∗ ) ⊆ fix((U ∗ )∗ ) = fix(U ).

Proof of Theorem 3.1.5. We prove (i) and (ii), before we show the claimed decompo-
sition of H. For part (i), pick U ∈ S . Since P H = fix(S ), it is clear that U P = P
holds. By Lemma 3.1.6 we have fix(U ) = fix(U ∗ ), and therefore also U ∗ P = P .
Taking adjoints and using that P ∗ = P , we thus obtain

P U = ((P U )∗ )∗ = (U ∗ P ∗ )∗ = (U ∗ P )∗ = P ∗ = P.

For part (ii), take f ∈ H and set C := co {U f | U ∈ S }. As a special case of

Theorem A.2.1, every non-empty closed convex set of a Hilbert space contains a
unique element of minimal norm. Let g be the unique element of minimal norm in
C. We show that g = P f .
For U ∈ S we have U C ⊆ C since U is continuous and linear, and S is a semigroup.
In particular, U g ∈ C. Since ∥U g∥ ≤ ∥g∥, we must have U g = g by choice of g. This
means g ∈ fix(S ), and therefore g = P g ∈ P C. But, using that P is continuous
and linear, we have

P C ⊆ co {P U f | U ∈ S } = {P f }

by part (i). Hence, P f = g as claimed, and we have established (ii).

WeSfinally prove the claimed orthogonal decomposition by showing that fix(S )⊥ =
lin U ∈S (IdH − U )(H).
For the inclusion “⊆” observe that part (i) implies P (IdH − U ) = 0 for every U ∈
S .SSince P −1 ({0}) = fix(S )⊥ is a closed linear subspace of H, we thus obtain
lin U ∈S (IdH − U )(H) ⊆ fix(S )⊥ .
For the converse inclusion “⊇” pick f ∈ P −1 ({0}). For every Pm ε > 0 we find by
Pelements U1 , . . . , Um ∈ S and c1 , . . . , cm ∈ [0, 1] with j=1 cj = 1 such that
(ii)
∥ m i=1 ci Ui f ∥ ≤ ε. But this means

m
X
f− ci (IdH − Ui )f ≤ ε.
i=1
S
Therefore f is in the closure of the subspace lin U ∈S (IdH − U )(H).
38 LECTURE 3.

The abstract (and general) Theorem 3.1.5 implies the convergence of many different
“ergodic averages”. To make this precise, we introduce the notion of a net, gen-
eralizing the classical concept of a sequence, see also [Sin19, Section 4.2] for more
details.
A set I ̸= ∅ with a relation ≤ is a directed set if
(i) i ≤ i for every i ∈ I,
(ii) i1 ≤ i2 and i2 ≤ i3 for i1 , i2 , i3 ∈ I implies i1 ≤ i3 , and
(iii) for all i1 , i2 ∈ I there is i ∈ I with i1 ≤ i and i2 ≤ i.
Basic examples are N, Z, or R with their natural ordering, or the power set P(X)
of a set X equipped with set inclusion “⊆” as its relation.
A map I → Ω, i 7→ ωi from a directed set I into a set Ω is usually written as (ωi )i∈I
and is called a net in Ω. If Ω is a topological space, (ωi )i∈I is a net in Ω, and ω ∈ Ω
is some point, then we say, just as for sequences, that
(i) (ωi )i∈I converges to ω if for every open set O with ω ∈ O there is some i0 ∈ I
with xi ∈ O for all i ≥ i0 . In this case, ω is called a limit of (ωi )i∈I .
(ii) ω is an accumulation point of (ωi )i∈I if for every open set O with ω ∈ O
and each i0 ∈ I there is an i ≥ i0 with xi ∈ O.
In a Hausdorff space1 Ω, a net (ωi )i∈I can have at most one limit ω (see [Sin19,
Theorem 4.2.4], and we write limi∈I ωi := ω in this case.
Many well-known characterizations of topological notions in metric spaces transfer
over to topological spaces when one replaces the word “sequence” with “net”. For
example, a Hausdorff space Ω is compact2 precisely when every net in Ω has an
accumulation point (see, e.g., [Sin19, Theorems 4.2.10 and 5.1.17]).
The following definition uses the concept of nets in the context of the mean ergodic
theorem.

Definition 3.1.7. Let S ⊆ L (H) be a semigroup on a Hilbert space H. A net

(Vi )i∈I in L (H) is an ergodic net for S if for every f ∈ H we have
(i) Vi f ∈ co {U f | U ∈ S } for each i ∈ I, and
(ii) limi∈I Vi (f − U f ) = 0 for every U ∈ S .
We discuss some basic examples.

Examples 3.1.8. Let U ∈ L (H) with ∥U ∥ ≤ 1 and consider the induced contrac-
tion semigroup SU = {U n | n ∈ N0 } of Remark 3.1.4.
1
Recall that a topological space Ω is a Hausdorff space if for distinct x, y ∈ Ω we can always
find open, disjoint subsets Ox , Oy ⊆ Ω with x ∈ Ox and y ∈ Oy .
2
In this course, the term compact will always include the Hausdorff property.
3.1. AVERAGING AND THE MEAN ERGODIC THEOREM 39
P −1 n
(i) The classical Cesàro means VN := N1 N n=0 U for N ∈ N define an ergodic
net (VN )N ∈N for SU : For f ∈ H we obtain by telescopic summing,
N −1 N
1 X n 1 X n 1 2∥f ∥
∥VN (f − U f )∥ = U f− U f = ∥f − U N f ∥ ≤
N n=0 N n=1 N N

for every N ∈ N, hence limN →∞ ∥VN (f − U f )∥ = 0. Again using telescopic

summing this implies limN →∞ ∥VN (f − U n f )∥ = 0 for every n ∈ N.
(ii) Turn N2 into a directed set by setting (N1 , M1 ) ≤ (N2 , M2 ) if M1 ≤ M2 for
+M −1 n
(N1 , M1 ), (N2 , M2 ) ∈ N . Similarly to (i), the means VN,M := M N
2 1
P
n=N U
for (N, M ) ∈ N2 define an ergodic net (VN,M )(N,M )∈N2 for SU .
(iii) If we equip the set (1, ∞) P∞with−nits nnatural order, we obtain that the Abel
means Vr := (1 − /r) n=0 r U for r ∈ (1, ∞) define an ergodic net
1

(Vr )r∈(1,∞) for SU , see Exercise 3.3.

A general construction of ergodic nets involves so-called Følner nets of groups.
Here, for a subset A ⊆ G of a group G and x ∈ G we use the intuitive notation
Ax := {yx | y ∈ A}. Analogously, we introduce xA and A−1 (which will appear at
a later point).

Definition 3.1.9. For a group G call a net (Fi )i∈I of non-empty finite subsets of G
a (right) Følner net if limi∈I |Fi |F
∆ Fi x|
i|
= 0 for every x ∈ G.
Simple examples for the group G = Z are the sequences (EN )N ∈N and (FN )N ∈N with
EN := {0, . . . , N − 1} and FN := {−N + 1, . . . , N − 1} for N ∈ N. We obtain the
following general result.

Proposition 3.1.10. Every abelian group G has a Følner net.

We first prove a helpful lemma.

Lemma 3.1.11. Let E ⊆ G be a non-empty finite subset of an abelian group G and

consider the sets E n := {x1 · · · xn | x1 , . . . , xn ∈ E} for n ∈ N. For every r > 1
there is some n ∈ N with |E n+1 | ≤ r|E n |.
Proof. We write N := |E| for the number of elements of E. For n ∈ N there are
 
n+N −1 (n + N − 1) · · · (n + 1)
p(n) := =
n (N − 1)!
many possibilities to chose n elements out of N distinct elements with replacement
(see, e.g., [Bó06, Theorem 3.21]). This implies that the set E n has at most p(n)
many elements for every n ∈ N (here we use that G is abelian!). Assuming that
there is some r > 1 such that |E n+1 | > r|E n | for all n ∈ N, we obtain that
rn−1 |E| < |E n | ≤ p(n) for every n ∈ N.
40 LECTURE 3.

But this contradicts the fact that power functions with base greater than 1 grow
faster than any polynomial.

Proof of Proposition 3.1.10. Let I be the set of pairs (E, m) of non-empty finite
subsets E ⊆ G and elements m ∈ N. By setting

(E1 , m1 ) ≤ (E2 , m2 ) :⇔ E1 ⊆ E2 and m1 ≤ m2

for (E1 , m1 ), (E2 , m2 ) ∈ I, we turn I into a directed set. For every (E, m) ∈ I we
use Lemma 3.1.11 to find n(E, m) ∈ N with |E n(E,m)+1 | ≤ (1 + m1 )|E n(E,m) | and set
FE,m := E n(E,m) . We claim that (FE,m )(E,m)∈I is then a Følner net.
So let x ∈ G and n ∈ N. We take (E, m) ∈ I with (E, m) ≥ ({x, x−1 }, n) and show
that |FE,m ∆ FE,m x|/|FE,m | ≤ n2 . Since x ∈ E, we have FE,m x = E n(E,m) x ⊆ E n(E,m)+1 .
This implies

|FE,m x \ FE,m | |E n(E,m)+1 | − |E n(E,m) | 1 1

≤ n(E,m)
≤ ≤ .
|FE,m | |E | m n

But, since multiplication from the right with x−1 is a bijection, we also have

|FE,m \ FE,m x| |(FE,m \ FE,m x)x−1 | |(FE,m x−1 ) \ FE,m | 1

= = ≤
|FE,m | |FE,m | |FE,m | n

by the same reasoning and the fact that x−1 ∈ E. Combining both estimates we
obtain the claim.

We now construct ergodic nets for certain group representations.

Definition 3.1.12. For a Hilbert space H and a group G we call a group homomor-
phism U : G → U (H), x 7→ Ux to the group U (H) of unitary operators a unitary
representation of G on H.
We are mostly interested in the following class of examples.

Example 3.1.13. Let (X, T ) be a measure-preserving system. Then the Koopman

representation UT : Γ → U (L2 (X)) is a unitary representation of the group Γ on
the Hilbert space L2 (X).

Proposition 3.1.14. Let U : G → U (H) be a unitary representation of a group

1
P
G. If (Fi )i∈I is a Følner net for G, then by setting Vi := |Fi | x∈Fi Ux for i ∈ I we
obtain an ergodic net (Vi )i∈I for the image group {Ux | x ∈ G}.
3.1. AVERAGING AND THE MEAN ERGODIC THEOREM 41

Proof. For y ∈ G and f ∈ H we have

1 X 1 X 1 X
Vi (f − Uy f ) = Ux f − Ux f = Ux f
|Fi | x∈F |Fi | x∈F y |Fi | x∈Fi ∆ Fi y
i i

1 X |Fi ∆ Fi y|
≤ ∥Ux f ∥ ≤ · ∥f ∥
|Fi | x∈F ∆ Fi y
|Fi |
i

for all i ∈ I which implies the claim.

Thus, by Proposition 3.1.10, ergodic nets always exist for (the image of) a unitary
representation of an abelian group. In Exercise 3.5 we even construct ergodic nets
for any abelian contraction semigroup S ⊆ L (H) on a Hilbert space H.
We now obtain the following consequence of Theorem 3.1.5.

Theorem 3.1.15. Let S ⊆ L (H) be a contraction semigroup on a Hilbert space

H, and P ∈ L (H) the orthogonal projection onto fix(S ). If (Vi )i∈I is an ergodic
net for S , then limi∈I Vi f = P f for every f ∈ H.
In the proof we use the following standard and very useful lemma of operator theory
(cf. [Haa14, Exercise 9.10]). It is an easy exercise when using linearity and applying
the triangle inequality.

Lemma 3.1.16. Let E and F be normed spaces, and further D ⊆ E such that the
linear hull lin D is dense in E. For (Ui )i∈I a net in L (E, F ) with supi∈I ∥Ui ∥ < ∞
and U ∈ L (E, F ) the following assertions are equivalent.
(i) (Ui f )i∈I converges to U f for every f ∈ D.
(ii) (Ui f )i∈I converges to U f for every f ∈ E.

Proof of Theorem 3.1.15. Take f ∈ H. Using the orthogonal

S decomposition from
3.1.5 we can reduce to the cases f ∈ fix(S ) and f ∈ lin U ∈S (IdH − U )(H).
In the first situation we have U f = f for every U ∈ S . This implies co {U f | U ∈
S } = {f }, and therefore also Vi f = f for every i ∈ I by Definition 3.1.7 (i). This
implies limi∈I Vi f = f .
For the second case, observe that for g ∈ H, we obtain by Definition 3.1.7 (ii) that
i (IdH − U )g = 0. By Lemma 3.1.16 we obtain that limi∈I Vi f = 0 for every
limi∈I VS
f ∈ lin U ∈S (IdH − U )(H).

In particular, by taking the Cesàro means (see Example 3.1.8 (ii)) of a Koopman
operator, we recover von Neumann’s Theorem 3.1.1. The following refinement of
the Poincaré’s recurrence theorem, Theorem 1.1.6, is an application.
42 LECTURE 3.

Corollary 3.1.17 (Khintchin’s refinement). Let T : Σ(X) → Σ(X) be a measure

algebra homomorphism on a probability space X and A ∈ Σ(X). Then
N −1
2 1 X
µX (A) ≤ lim µX (A ∩ T n (A)).
N →∞ N
n=0

In particular, if µX (A) > 0, there are infinitely many n ∈ N with µX (A∩T n (A)) > 0.
Proof. Write RP for the Rorthogonal projection onto fix(UT ) = fix(SUT ) (cf. Remark
3.1.4). Since X UT f = X f for every f ∈ L2 (X) (see Lemma 1.3.3), we obtain
Z N −1 Z Z
1 X
µX (A) = 1A = lim UT 1A =
n
P 1A = (P 1A |1).
X N →∞ N
n=0 X X

By the Cauchy-Schwarz inequality we thus have

µX (A)2 ≤ ∥P 1A ∥22 · ∥1∥22 = (P 1A |P 1A ) = (P ∗ P 1A |1A )
N −1 Z N −1
1 X 1 X
= (P 1A |1A ) = lim UT 1A · 1A = lim
n
µX (A ∩ T n (A)).
N →∞ N N →∞ N
n=0 X n=0

3.2 Invariant Measures

In many examples the underlying maps of a concrete measure-preserving system
are not only measurable, but even continuous. Ergodic theory is therefore closely
related to the area of topological dynamics. We introduce its central objects. Here,
for a compact space K, write Homeo(K) for the group of all homeomorphisms
τ : K → K.

Definition 3.2.1. A topological dynamical system (K, τ ) consists of a compact

space K ̸= ∅ and a group homomorphism τ : Γ → Homeo(K), γ 7→ τγ .
As an important example we discuss a topological version of the Bernoulli shift. For
this, recall the product of topological spaces (see, e.g., [Sin19, Section 2.2]).

Definition 3.2.2. Let I be an index set and Ωi a topological space for every i ∈ I.
For a finite set {i1 , . . . , im } ⊆ I and open sets Oj ⊆ Ωij for each j ∈ {1, . . . , m} we
call
 Y 
Z(i1 , . . . , im ; O1 , . . . , Om ) := (ωi )i∈I ∈ Ωi | ωij ∈ Oj for all j ∈ {1, . . . , m}
i∈I
Q
an (open) cylinder set. The topology on i∈I Ωi generated by all such cylinder
sets is the product topology.
3.2. INVARIANT MEASURES 43

Remark
Q 3.2.3. The product topology Q is the smallest topology on the product
Ω
i∈I i making all projection maps prj : i∈I Ωi → Ωj , (ωi )i∈I 7→ ωj for j ∈ I con-
tinuous (see [Sin19, Proposition 2.2.7]) A different way to think about this topology
is via convergence of nets (and, in particular, sequences): A net converges in the
product topology to some limit precisely when it converges in each component to
that limit (see [Sin19, Theorem 4.2.7]).
From now on, we equip the product of topological spaces with the product topology.
If we start with compact spaces, then we still obtain a compact space it that way
(see, e.g., [Sin19, Theorem 5.1.14]):

Theorem 3.2.4 (Tychonoff). Let I Q be an index set and Ki be a compact space for
every i ∈ I. Then the product space i∈I Ki is also compact.
With the help of Tychonoff’s Theorem we now obtain a topological version of Ex-
ample 2.1.8.

Example 3.2.5. Let K ̸= ∅ be a compact space, e.g., K = {0, . . . , k − 1} equipped

with the discrete topology (i.e., every set is open). Then τ : Γ → Homeo(K Γ ), γ 7→
τγ with τγ (xδ )δ∈Γ = (xδ+γ )δ∈Γ for (xδ )δ∈Γ ∈ K Γ and γ ∈ Γ defines a topological
dynamical system (K Γ , τ ). In fact, with arguments similar as in Example 2.1.8,
we obtain that τγ : K Γ → K Γ is a homeomorphism for every γ ∈ Γ, and obviously
τγ1 +γ2 = τγ1 ◦ τγ2 for all γ1 , γ2 ∈ Γ.
We now transform topological dynamical systems into (concrete) measure-preserving
systems by adding a suitable probability measure. Recall that a probability measure
µ : B(K) → [0, 1] on the Borel σ-algebra B(K) of a compact space K is regular
if
µ(A) = sup{µ(L) | L ⊆ K compact with L ⊆ A}
= inf{µ(O) | O ⊆ K open with A ⊆ O}
for all A ∈ B(K), see [Rud87, Paragraph 2.4]. Note that, since we only consider
probability measures µ here, the second equality actually follows from the first by
considering complements. The Lebesgue measure on [0, 1] is an important example
of a regular Borel probability measure. We write P(K) for the set of all regular
Borel probability measures on a compact space K.
There is an equivalent way to look at regular Borel probability measures based on
the following observation: If µ : B(K) → [0, 1] is a regular Borel probability measure
on a compact space K, every continuous function f : K → C is Borel measurable and
bounded, hence integrable with respect to µ. We therefore obtain an “integration
map”
Z
φµ : C(K) → C, f 7→ f dµ.
K
44 LECTURE 3.

Moreover, φ = φµ is a linear map which is

(i) positive, i.e., φ(f ) ≥ 0 for all f ∈ C(K) with f ≥ 0, and
(ii) unital, i.e., φ(1) = 1 for the constant one-function 1 : K → C.
The following representation result (see, e.g., [Rud87, Paragraph 2.3]) gives a con-
verse.

Theorem 3.2.6 (Riesz–Markov–Kakutani). Let K be a compact space. For every

unital, positive, linear map φ : C(K) → C there is a unique µ ∈ P(K) with φ = φµ .
From now on, we identify for a compact space K each measure µ ∈ P(K) R with the
corresponding linear form φµ . In particular, we write µ(f ) := φµ (f ) = K f dµ for
f ∈ C(K) and µ ∈ P(K). R In this way, every µ ∈ P(K) becomes a linear map from
C(K) to C, and since | K f dµ| ≤ ∥f ∥∞ for every f ∈ C(K), it is bounded, hence
an element of the dual Banach space C(K)′ with norm ∥µ∥ ≤ 1. This allows us to
topologize the set P(K).
On any normed space E the weak* topology is the smallest topology making all
point evaluations
evf : E ′ → C, φ 7→ φ(f )
for f ∈ E continuous.3 A net (φi )i∈I in E ′ converges to φ ∈ E ′ with respect to the
weak* topology precisely when the net (φi (f ))i∈I converges to φ(f ) in C for every
f ∈ E. We refer, e.g., to [Ped89, Sections 2.4 and 2.5] for more information on the
weak* topology. The following is an important consequence of Tychonoff’s theorem,
Theorem 3.2.4 above (see, e.g., [Ped89, Theorem 2.5.2]).

Theorem 3.2.7 (Banach–Alaoglu). Let E be a normed space. Then the dual unit
ball {φ ∈ E ′ | ∥φ∥ ≤ 1} is compact with respect to the weak* topology.
In the context of regular Borel probability measures, this yields:

Proposition 3.2.8. If K is a compact space, then P(K) is compact with respect to

the weak* topology.
Proof. In view of Theorem 3.2.7, since closed subsets of compact spaces are again
compact, it suffices to show that P(K) is a closed subset of the dual unit ball
B := {φ ∈ C(K)′ | ∥φ∥ ≤ 1}. However, we can write
\
P(K) = B ∩ ev−1 −1
f ([0, ∞)) ∩ ev1 ({1}),
f ∈C(K)
f ≥0

and this implies that P(K) is closed as an intersection of closed subsets.

In other words, if E ′ is considered as a subset of CE , the space of all maps from E to C, the
3

weak* topology is simply the product topology.

3.2. INVARIANT MEASURES 45

To obtain measure-preserving systems from topological ones, we introduce the fol-

lowing definition.

Definition 3.2.9. Let (K, τ ) be a topological dynamical system. A regular Borel

probability measure µ : B(K) → [0, 1] is invariant if τγ is measure-preserving with
respect to µ for every γ ∈ Γ, i.e., µ(τγ−1 (A)) = µ(A) for all Borel sets A ⊆ K. We
write P(K, τ ) ⊆ P(K) for the set of all invariant probability measures.
For the following lemma notice that continuous maps between topological spaces are
Borel measurable (by Proposition 1.1.4 (i)).

Lemma 3.2.10. Let τ : K → L be a continuous map between compact spaces and

µ ∈ P(K). Then the push-forward measure
R τ∗ µ : B(L)R → [0, 1], A 7→ µ(τ −1 (A)) is a
regular Borel probability measure with K f ◦ τ dµ = L f dτ∗ µ for all f ∈ C(L).
The proof is left as Exercise 3.7. This leads to the following characterization of
invariant measures.

Proposition 3.2.11. For a topological dynamical system (K, τ ) and µ ∈ P(K) the
following assertions are equivalent.
(a) µ is invariant.
R R
(b) K f ◦ τγ dµ = K f dµ for all f ∈ C(K) and γ ∈ Γ.
Does every topological dynamical system admit an invariant measure? To answer
this question, we first introduce the following concept closely related to the notion
of ergodic nets from the previous section.

Definition 3.2.12. For a topological dynamical

R system (K, τ ) we call a net (µi )i∈I in
P(K) asympotically invariant if limi∈I K f − (f ◦ τγ ) dµi = 0 for every f ∈ C(K)
and each γ ∈ Γ.

Example 3.2.13. Let (Fi )i∈I be a Følner net for the group Γ. Then for a topological
dynamical system (K,Pτ ) and any µ ∈ P(K) (e.g., µ = δx a point measure for some
x ∈ K), set µi := |F1i | γ∈Fi (τγ )∗ µ for i ∈ I. One can check, with similar arguments
as in the proof of Proposition 3.1.14, that (µi )i∈I is asymptotically invariant, see
Exercise 3.8.

Proposition 3.2.14. Let (K, τ ) be a topological dynamical system. Every weak*

accumulation point µ of an asymptotically invariant net (µi )i∈I in P(K) is invariant.
R
Proof. Take γ ∈ Γ and f ∈ C(K). For ε > 0 we find some i0 ∈ I with | K f − (f ◦
τγ ) dµi | < ε/3 for each i ≥ i0 . By the definition of the weak* topology, the set
O := {ν ∈ P(K) | |µ(f ) − ν(f )| < ε/3} ∩ {ν ∈ P(K) | |µ(f ◦ τγ ) − ν(f ◦ τγ )| < ε/3}
is an open set containing µ. By the definition of accumulation points, we thus find
46 LECTURE 3.

some i ≥ i0 with µi ∈ O. But then |µ(f ) − µ(f ◦ τγ )| ≤ ε by the triangle inequality.

Thus, µ(f ) = µ(f ◦ τγ ), and µ is invariant by Proposition 3.2.11.

Corollary 3.2.15 (Krylov–Bogolyubov). Every topological dynamical system (K, τ )

has an invariant measure µ ∈ P(K, τ ).
Proof. By Proposition 3.1.10 and Example 3.2.13 the system (K, τ ) has an asymp-
totically invariant net (µi )i∈I in P(K). By Proposition 3.2.8 the net (µi )i∈I has
a weak* accumulation point µ ∈ P(K), and then µ is invariant by Proposition
3.2.14.
3.3. COMMENTS AND FURTHER READING 47

3.3 Comments and Further Reading

Von Neumann proved his mean ergodic theorem in [vN32b], but in a completely dif-
ferent way than what is presented here. His work was a mathematical formalization
of the ergodic hypothesis from statistical mechanics going back to Ludwig Boltz-
mann, see, e.g., [Bad06] for more information. Even though not part of the lecture,
we also mention the pointwise ergodic theorem of George Birkhoff
P −1 showing point-
wise almost everywhere convergence of the Cesàro means N1 N n=0 Uτ
n
f (instead of
2
L -convergence), see [Bir31]. Both results had a tremendous impact on the further
development of mathematical ergodic theory, but also caused a dispute between von
Neumann and Birkhoff - more on this can be read, e.g., in [Ber04].
The “abstract mean ergodic theorem” of this lecture is a special case of a general
result by Alaoglu and Birkhoff, see [AB40]. Our exposition is based on [EFHN15,
Supplement of Chapter 8]. There are many works on mean ergodic theorems for
operators and operator semigroups on Banach spaces, see, e.g., [Nag73], [Sat78],
[Kre85, Chapter 2], [EFHN15, Chapter 8], [Sch13], and [Kre18].
Groups having a Følner net are called amenable. There are many equivalent char-
acterizations of this property, and interesting examples, see, e.g., [Pat88], [Run02],
[CSD21], and [Jus22]. Our proof of the fact that abelian groups are amenable re-
lies on the property that these groups have “subexponential growth”, see [CSD21,
Corollary 14.7] or [Jus22, Subsection 2.6].
An introduction to topological dynamics is given in [Aus88], see also the classical
monograph [GH55], as well as [Bro79] and [dV93]. Here we are mostly interested
in constructing measure-preserving systems from topological ones. This will be a
crucial procedure in the next lecture.
48 LECTURE 3.

3.4 Exercises
Exercise 3.1. A sequence
PN (an )n∈N of complex numbers is Cesàro convergent to
1
a ∈ C if limN →∞ N n=1 an = a, and in this case a is the Cesàro limit of (an )n∈N .
(i) Show that for every sequence of real numbers (an )n∈N the inequalities
N N
1 X 1 X
lim inf an ≤ lim inf an ≤ lim sup an ≤ lim sup an
n→∞ N →∞ N N →∞ N n→∞
n=1 n=1

hold.
(ii) Show that if a sequence (an )n∈N of complex numbers converges to a ∈ C, then
it is also Cesàro convergent to a.
(iii) Let an = λn for n ∈ N, where λ ∈ C. Determine (depending on λ), if (an )n∈N
is Cesàro convergent, and in this case compute the Cesàro limit.
(iv) Give an example of a bounded sequence in C which is not Cesàro convergent.
(v) Show that a sequence (an )n∈N of complex numbers converges if and only if
each subsequence of (an )n∈N is Cesàro convergent.

Exercise 3.2. Give an example of a Banach space E, a bounded linear opera-

tor PU ∈ L (E) with ∥U ∥ ≤ 1, and an element f ∈ E such that the sequence
−1 n
( N1 Nn=0 U f )N ∈N does not converge in E.

Exercise 3.3. Let U ∈ L (H) be a bounded linear operator on a Hilbert space

H with ∥U ∥ ≤ 1 and SU the generated contraction semigroup from Remark 3.1.4.
Equip (1, ∞) with its usual order and define
∞
X N
X
−n
Vr := (1 − 1/r) r n
U = lim (1 − 1/r) r−n U n
N →∞
n=0 n=0

for r ∈ (1, ∞) where the limit exists with respect to the operator norm. Show that
(Vr )r∈(1,∞) is an ergodic net for SU .

Exercise 3.4. Let (Fi )i∈I be a Følner net for Γ. Show that for a measure-preserving
system (X, T ) the following assertions are equivalent.
(a) (X, T ) is ergodic.
1
UTγ f = ( X f ) · 1 in L2 (X).
R
(b) For all f ∈ L2 (X) we have limα
P
|Fα | γ∈Fα

Exercise 3.5. Let ∅ ̸= S ⊆ L (H) be an abelian contraction semigroup on a

Hilbert space H, i.e., U V = V U for all U, V ∈ S . Equip the convex hull I := co S
with the relation ≤ defined by
U ≤V :⇔ U = V or there is W ∈ I with V = W U
3.4. EXERCISES 49

for U, V ∈ I. Show the following assertions.

(i) I is a directed set.
(ii) The map I → L (H),P V−1 7→nV is an ergodic net.
n=0 U ∈ I for every N ∈ N and U ∈ S .
Hint: Note that N1 N

Exercise 3.6. A topological dynamical system (K, τ ) is topologically transitive

if for all non-empty open subsets O1 , O2 ⊆ K we there is some γ ∈ Γ such that
τγ−1 (O1 )∩O2 ̸= ∅. Show that if Γ is an infinite abelian group, then for any non-empty
compact space K the shift (K Γ , τ ) from Example 3.2.5 is topologically transitive.

Exercise 3.7. Prove Lemma 3.2.10.

Exercise 3.8. Consider a Følner net (Fi )i∈I for the group Γ, a topological
Pdynamical
1
system (K, τ ), and µ ∈ P(K). Show that the net (µi )i∈I with µi := |Fi | γ∈Fi (τγ )∗ µ
for i ∈ I is asymptotically invariant.
50 LECTURE 3.
A.2. BASIC HILBERT SPACE THEORY 55

A.2 Basic Hilbert Space Theory

The following are some basic concepts and tools from Hilbert space theory, see, e.g.,
[Con85, Chapter 1], [Rud87, Chapter 4], [Ped89, Sections 3.1 and 3.2], and [Haa14,
Chapters 1 and 8, and Section 12.2].
Given a Hilbert space H, two elements f, g ∈ H are called orthogonal (in symbols:
f ⊥ g) if (f |g) = 0. For orthogonal f, g ∈ H the Pythagorean theorem from above
becomes ∥f + g∥2 = ∥f ∥2 + ∥g∥2 . For a subset M ⊆ H we call
M ⊥ := {f ∈ M | f ⊥ g for every g ∈ M }
the orthogonal complement of M . This is always a closed subspace of H. Two
subsets M, N ⊆ H are orthogonal if f ⊥ g = 0 for all f ∈ M and g ∈ N , i.e.,
M ⊆ N ⊥.
The following geometric result is used in Chapter 3 (with x = 0) and is fundamental
in Hilbert space theory (see, e.g., [Con85, Theorems I.2.5 and I.2.6]. Recall here
that a subset C ⊆ E of a vector space E is convex if tf + (1 − t)g ∈ C for all
f, g ∈ C and t ∈ [0, 1].

Theorem A.2.1. Let C ⊆ H be a non-empty, closed, convex subset of a Hilbert

space H and x ∈ H.
(i) There is a unique y0 ∈ C with ∥y0 − x∥ = inf{∥y − x∥ | y ∈ C}.
(ii) If C ⊆ H is even a closed linear subspace, then y0 of (i) is the unique element
z ∈ C with x − z ∈ C ⊥ .
The following result makes use of Theorem A.2.1, see, e.g., [Ped89, Theorem 3.1.7].

Theorem A.2.2. Let M ⊆ H be a closed linear subspace of a Hilbert space H.

Then H = M ⊕ M ⊥ is a decomposition of H into closed and orthogonal subspaces
M and M ⊥ .
In the situation of Theorem A.2.2, the projection map PM : H → H induced by this
decomposition, i.e., the unique linear map sending elements x ∈ M to itself, and
elements x ∈ M ⊥ to 0, is called the orthogonal projection onto M .
By Theorem A.2.2 we obtain that for any subset M ⊆ H of a Hilbert space H, we
have the identity (M ⊥ )⊥ = lin M (see [Ped89, Corollary 3.1.8]). Thus, the linear
hull lin M is dense in H precisely when M ⊥ = {0}.
Another important consequence is the following famous representation theorem for
the dual of a Hilbert space, see, e.g., [Con85, Theorem I.3.4].

Theorem A.2.3 (Riesz–Fréchet). Let H be a Hilbert space. Then the map

H → H ′, g 7→ g
56 APPENDIX A. SOME FUNCTIONAL ANALYSIS

where g(f ) := (f |g) for all f, g ∈ H is a bijection. Moreover, ∥g∥ = ∥g∥ for every
g ∈ H.
The Riesz–Fréchet theorem has an interesting consequence for bounded linear op-
erators: If U ∈ L (H, K) for Hilbert spaces H and K, then for every y ∈ K the
map

y ◦ U : H → C, x 7→ (U x|y)

is a bounded linear map by the Cauchy–Schwarz inequality, and hence there is a

unique vector z ∈ H with y ◦ U = z. We set U ∗ y := z, i.e., U ∗ y is characterized
by the identity (x|U ∗ y) = (U x|y) for all x ∈ H and y ∈ K. The map U ∗ : K → H
is called the adjoint operator of U , and is itself a bounded linear operator with
∥U ∗ ∥ = ∥U ∥. Observe that (U ∗ )∗ = U . Moreover, if H, K, L are Hilbert spaces,
then (U V )∗ = V ∗ U ∗ holds for all U ∈ L (K, L) and V ∈ L (H, K).
One can check that a bounded linear operator U ∈ L (H, K) between Hilbert spaces
H and K is
(i) a linear isometry precisely when U ∗ U = IdH .
(ii) a unitary operator precisely when U ∗ U = IdH and U U ∗ = IdK .
A bounded linear operator U ∈ L (H) on a Hilbert space H is self-adjoint if
U ∗ = U . A self-adjoint operator P ∈ L (H) is an orthogonal projection if in
addition P 2 = P . For every closed subspace M ⊆ H the orthogonal projection PM
onto M from above is indeed an orthogonal projection in this sense. Conversely, if
P is an orthogonal projection, then the image M := P H is a closed linear subspace
and P = PM . See [Ped89, Section 3.2] for these and further assertions.
Finally, we will also need the concept of orthonormal bases in some of the lectures.
A subset E ⊆ H of a Hilbert space H is orthonormal if its elements are pairwise
orthogonal and ∥e∥ = 1 for every e ∈ E. An orthonormal subset E ⊆ H which is
maximal with respect to set inclusion, i.e., there is no strictly larger orthonormal
subset F ⊆ H, is called an orthonormal basis of H. An orthonormal subset
E ⊆ H is an orthonormal basis of H precisely when lin E = H. An application of
Zorn’s lemma yields the following (see [Ped89, Proposition 3.1.12]):

Theorem A.2.4. Every orthonormal subset of a Hilbert space H is contained in an

orthonormal basis of H.
In particular, each Hilbert space has an orthonormal basis. This allows us to repre-
sent vectors in a Hilbert space via the following result, see, e.g., [Con85, Theorem
I.4.13]. Note here that the finite subsets Pfin (A) of any set A are directed by set
inclusion, and so we can use the notion of net convergence from Chapter 3.
A.2. BASIC HILBERT SPACE THEORY 57

Theorem A.2.5 (Fourier Series/Parseval Identity). Let E ⊆ H be an orthonormal

basis of a Hilbert space H. Then
X X
f= (f |e)f := lim (f |e)e and
F ⊆E
e∈E finite e∈F
X X X
∥f ∥2 = |(f |e)|2 := lim |(f |e)|2 = sup |(f |e)|2
F ⊆E F ⊆E
e∈E finite e∈F finite e∈F

for every f ∈ H.
By extending an orthonormal subset to an orthonormal basis we obtain the following
corollary.

Corollary A.2.6 (Representation of Projections/Bessel Inequality). Let E ⊆ H be

an orthonormal subset of a Hilbert space H, M := lin E its closed linear hull and
PM the orthogonal projection onto M . Then
X X
PM f = (f |e)f := lim (f |e)e and
F ⊆E
e∈E finite e∈F
X X X
∥PM f ∥2 = |(f |e)|2 := lim |(f |e)|2 = sup |(f |e)|2 ≤ ∥f ∥2
F ⊆E F ⊆E
e∈E finite e∈F finite e∈F

for every f ∈ H.
58 APPENDIX A. SOME FUNCTIONAL ANALYSIS
Bibliography

[AB40] L. Alaoglu and G. Birkhoff. General ergodic theorems. Ann. of Math.,

41:293–309, 1940.
[AE05] H. Amann and J. Escher. Analysis I. Birkhäuser, 2005.
[Aus88] J. Auslander. Minimal Flows and their Extensions. Elsevier, 1988.
[Bad06] M. Badino. The foundational role of ergodic theory. Found. Sci., 11:323–
347, 2006.
[Bar95] R.G. Bartle. The Elements of Integration and Lebesgue Measure. Wiley,
1995.
[BBP16] H. Bercovici, A. Brown, and C. Pearcy. Measure and Integration.
Springer, 2016.
[Ber04] V. Bergelson. Some historical comments and modern questions around
the ergodic theorem. In Dynamics of Complex Systems, pages 1–11.
Research Institute for Math. Sciences, Kyoto, 2004.
[Bil95] P. Billingsley. Probability and Measure. Wiley, 3rd edition, 1995.
[Bir31] G. D. Birkhoff. Proof of the ergodic theorem. Proc. Nat. Acad. Sci.
U.S.A., 17:656–660, 1931.
[Bro79] I. U. Bronšteǐn. Extensions of Minimal Transformation Groups. Sijthoff
& Noordhoff, 1979.
[Bó06] M. Bóna. A Walk Through Combinatorics. An Introduction to Enumer-
ation and Graph Theory. Worlds Scientific, 2nd edition, 2006.
[Con85] J.B. Conway. A Course in Functional Analysis. Springer, 1985.
[CSD21] T. Ceccherini-Silberstein and M. D’Adderio. Topics in Groups and Ge-
ometry Growth, Amenability, and Random Walks. Springer, 2021.
[DNP87] R. Derndinger, R. Nagel, and G. Palm. Ergodic Theory in the Perspective
of Functional Analysis. Unpublished book manuscript, 1987.
[dV93] J. de Vries. Elements of Topological Dynamics. Springer, 1993.

59
60 BIBLIOGRAPHY

[EF19] T. Eisner and B. Farkas. 22nd Internet Seminar: Ergodic Theorems,

2018–2019.
[EFHN09] T. Eisner, B. Farkas, M. Haase, and R. Nagel. 12th Internet Seminar:
Ergodic Theory – An Operator Theoretic Approach, 2008–2009.
[EFHN15] T. Eisner, B. Farkas, M. Haase, and R. Nagel. Operator Theoretic Aspects
of Ergodic Theory. Springer, 2015.
[EW11] M. Einsiedler and T. Ward. Ergodic Theory with a view towards Number
Theory. Springer, 2011.
[Fur14] H. Furstenberg. Recurrence in Ergodic Theory and Combinatorial Num-
ber Theory. Princeton University Press, 2014.
[GH55] W. H. Gottschalk and G. A. Hedlund. Topological Dynamics. American
Mathematical Society, 1955.
[Gla03] E. Glasner. Ergodic Theory via Joinings. American Mathematical Soci-
ety, 2003.
[Haa14] M. Haase. Functional Analysis: An Elementary Introduction. American
Mathematical Society, 2014.
[Hal56] P. Halmos. Lectures on Ergodic Theory. American Mathematical Society,
1956.
[HS65] E. Hewitt and K. Stromberg. Real and Abstract Analysis. Springer, 1965.
[JT22] A. Jamneshan and T. Tao. An uncountable Mackey-Zimmer theorem.
Studia Math., 266(3):241–289, 2022.
[JT23a] A. Jamneshan and T. Tao. Foundational aspects of uncountable mea-
sure theory: Gelfand duality, Riesz representation, canonical models, and
canonical disintegration. Fund. Math., 261(1):1–98, 2023.
[JT23b] A. Jamneshan and T. Tao. An uncountable Moore-Schmidt theorem.
Ergodic Theory Dynam. Systems, 43(7):2376–2403, 2023.
[Jus22] K. Juschenko. Amenability of Discrete Groups by Examples. American
Mathematical Society, 2022.
[Kal97] O. Kallenberg. Foundations of Modern Probability. Springer, 1997.
[Koo31] B. O. Koopman. Hamiltonian systems and transformation in Hilbert
space. Proc. Natl. Acad. Sci. U.S.A., 18:315–318, 1931.
[Kre85] U. Krengel. Ergodic Theorems. de Gruyter, 1985.
[Kre18] H. Kreidler. Compact operator semigroups applied to dynamical systems.
Semigroup Forum, 97:523–547, 2018.
[Nag73] R. Nagel. Mittelergodische Halbgruppen linearer Operatoren. Ann. Inst.
BIBLIOGRAPHY 61

Fourier, 23:75–87, 1973.

[Pat88] A. L. T. Paterson. Amenability. AMS, 1988.
[Ped89] G. K. Pedersen. Analysis Now. Springer, 1989.
[Poi90] H. Poincaré. Sur le problème des trois corps et les équations de la dy-
namique. Acta Math., 13:1–270, 1890.
[Ran02] I. K. Rana. An Introduction to Measure and Integration. American
Mathematical Society, 2nd edition, 2002.
[Rud87] W. Rudin. Real and Complex Analysis. McGraw-Hill, 1987.
[Run02] V. Runde. Lectures on Amenability. Springer, 2002.
[Sat78] R. Sato. On abstract mean ergodic theorems. Tôhoku Math. Journ.,
38:575–581, 1978.
[Sch13] M. Schreiber. Uniform families of ergodic operator nets. Semigroup
Forum, 86:321–336, 2013.
[Sin19] T. B. Singh. Introduction to Topology. Springer, 2019.
[Tao11] T. Tao. An Introduction to Measure Theory. American Mathematical
Society, 2011.
[vN32a] J. von Neumann. Einige Sätze über messbare Abbildungen. Ann. Math.
(2), 33:574–586, 1932.
[vN32b] J. von Neumann. Proof of the quasi-ergodic hypothesis. Proc. Natl.
Acad. Sci. U.S.A., 18:70–82, 1932.
[VO16] M. Viana and K. Oliveira. Foundations of Ergodic Theory. Cambridge
University Press, 2016.
[Wal75] P. Walters. An Introduction to Ergodic Theory. Springer, 1975.