You are on page 1of 73

Lecture notes on ergodic theory

Jesse Peterson

March 17, 2011


2
Contents

1 Representations of groups 5
1.1 Definitions and constructions . . . . . . . . . . . . . . . . . . . . 5
1.2 Functions of positive type . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Cocycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Induced representations . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Invariant vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6 Amenability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.6.1 Von Neumanns Mean Ergodic Theorem . . . . . . . . . . 21
1.7 Mixing properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.7.1 Compact representations . . . . . . . . . . . . . . . . . . . 24
1.7.2 Weak mixing for amenable groups . . . . . . . . . . . . . 26
1.8 Weak containment . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2 Group actions on measure spaces 31


2.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2 The Koopman representation . . . . . . . . . . . . . . . . . . . . 35
2.2.1 Ergodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.2.2 Weak mixing . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.2.3 Compact actions . . . . . . . . . . . . . . . . . . . . . . . 39
2.3 A remark about measure spaces . . . . . . . . . . . . . . . . . . . 41
2.4 Gaussian actions . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.5 Ergodic theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.6 Recurrence theorems . . . . . . . . . . . . . . . . . . . . . . . . . 50

3 Uniform multiple recurrence 53


3.1 Multiple recurrence for compact actions . . . . . . . . . . . . . . 54
3.2 Multiple recurrence for weak mixing actions . . . . . . . . . . . . 55
3.3 Joinings and the basic construction . . . . . . . . . . . . . . . . . 58
3.3.1 Joinings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3.2 The basic construction . . . . . . . . . . . . . . . . . . . . 59
3.4 Ergodic, weak mixing, and compact extensions . . . . . . . . . . 64

3
4 CONTENTS
Chapter 1

Representations of groups

1.1 Definitions and constructions


Definition 1.1.1. Let be a group, a unitary representation (resp. an
orthogonal representation) of is a homomorphism : U(H) (resp.
: O(H)) from into the unitary (resp. orthogonal) group of a complex
(resp. real) Hilbert space H.
Example 1.1.2. Let H be a Hilbert space and suppose is a group. The
trivial representation of on H is given by the homomorphism which takes
any group element to the identity operator on H.
Example 1.1.3. Let be a group, the left-regular representation (resp.
right-regular representation) is given by : U(2 ), (resp. :
U(2 )) which is defined by the formula ()(x x x ) = x x x (resp.
()(x x x ) = x x x 1 ).
Example 1.1.4. More generally, if a group acts on a set I then there is an as-
sociated unitary representation on 2 I defined by ()(xI x x ) = xI x x .
When < is a subgroup, then acts by left multiplication on the quotient
space / and the associated unitary representation is the quasi-regular rep-
resentation.
Suppose is a group and < is a subgroup. If : U(H) is a
unitary representation, then we obtain the reduced representation | of
by restricting to .
Given a complex Hilbert space H the adjoint Hilbert space H is the
Hilbert space which agrees with H as a set, has scalar multiplication given by
= and whose inner-product is given by h, iH = h, iH .
Given a unitary representation : U(H) we define the adjoint rep-
resentation : U(H) by setting () = (). Note that we have the
natural identification = .
If : O(H) is an orthogonal representation where H is a real Hilbert
space then we will use the convention H = H and = .

5
6 CHAPTER 1. REPRESENTATIONS OF GROUPS

If : U(H), and : U(K) are two unitary representations, then


is contained in if there is an isometry U : H K such that for all
we have that U () = ()U . If U can be chosen to be a unitary from H to K
then the representations are isomorphic and we will write
= .
Given a family of unitary representations : U(H ), with I, the
direct-sum representation is given by the map I : U(I H )
defined by
(I )() = I ( ()).
A unitary representation : U(H) is reducible if it contains a non-
trivial sub-representations, or equivalently, there exists a closed -invariant sub-
space K H such that K = 6 {0}, H. In this case it is easy to see that K is also
-invariant and the natural isomorphism from K K induces an equivalence
between the representations |K |K and . If a unitary representation has
no non-trivial invariant subspace we say that it is irreducible.
A vector H is a cyclic vector if the smallest -invariant subspace of H
which contains is H itself. Note that if is irreducible then every non-zero
vector is cyclic.
Lemma 1.1.5 (Schurs Lemma). Let : U(H), and : U(K)
be two irreducible unitary representations of a group , if T B(H, K) is -
invariant then either T = 0, or else T is a scalar multiple of a unitary. In
particular, B(H, K) has a non-zero -invariant operator if and only if and
are isomorphic.
Proof. Let and be as above and suppose T B(H, K) is -invariant. Thus,
T T B(H) is -invariant and hence any spectral projection of T T gives
a -invariant subspace. Since is irreducible it then follows that T T C.
If T T 6= 0 then my multiplying T by a scalar we may assume that T is an
isometry. Hence, T T B(K) is a non-zero -invariant projection, and since
is irreducible it follows that T T = T T = 1.
If I is finite, then the tensor product representation is given by the map
I : U(I H ) defined by
(I )() = I ( ()).
Lemma 1.1.6 (Fells Absorption Principle). Let : U(H) be a unitary
representation of a group , let 1H denote the trivial representation of on
H, and let : U(2 ) denote the left-regular representation. Then the
representations and 1H are isomorphic.
Proof. Consider the unitary U U(2 H) determined by U (x ) = x
(x), for all x , H. Then for all , x , and H we have that
U ( )()U (x ) = U ( )()(x (x))
= U (x ()(x))
= x ((x)1 )()(x) = ( 1H )()(x ).
1.1. DEFINITIONS AND CONSTRUCTIONS 7

Example 1.1.7 (Hilbert-Schmidt operators). Given two Hilbert spaces H and


K with respective orthonormal bases {i } H and {j } K we define the
space of Hilbert-Schmidt operators from H to K by

HS(H, K) = {S B(H, K) | kSk2HS = i kSi k2 < }.

Parsevals identity gives

i kSi k2 = i,j |hSi , j i|2 = j kS j k2 ,

hence we see that this sum is independent of the orthonormal bases chosen.
Moreover, this shows that if A B(H0 , H), B B(K, K0 ), and S HS(H, K)
then BSA HS(H0 , K0 ) and

kBSAk kBSAkHS kBkkSkHSkAk.

The space of Hilbert-Schmidt operators is an inner-product space with inner-


product given by
hS, T iHS = i,j hSi , j ihj , T i i,
which is well defined by the Cauchy-Schwarz inequality, and does not depend
on the bases by applying Persevals identity.
The inner-product defined above turns HS(H, K) into a Hilbert space. To
see this suppose Sn HS(H, K) is a sequence of Hilbert-Schmidt operators,
which is Cauchy in the Hilbert-Schmidt norm, then Sn is also Cauchy in the
operator norm and hence converges in the operator norm to an operator S :
H K.
For any finite dimensional subspace K0 K it is easy to see that

kPK0 S PK0 Sn kHS 0,

and thus we have that

kSkHS = sup kPK0 SkHS


K0 K, dim(K0 )<

lim sup kSn kHS < .


n

Hence, S HS(H, K) and so all that remains to show is that kS Sn kHS 0.


Suppose > 0 is given. Since S Sn is Cauchy in the Hilbert-Schmidt norm,
there exists some N > 0 such that for all n N we have that kSN Sn kHS <
/2. Take K0 K a finite dimensional subspace such that kPK 0
(S SN )kHS <
/2. Then for all n N we have

kS Sn kHS

kPK0 S PK0 Sn kHS + kPK


0
S PK
0
SN kHS + kPK
0
SN PK
0
Sn kHS .
The first term above tends to zero while the others are each at most /2, hence

lim sup kS Sn kHS .


n
8 CHAPTER 1. REPRESENTATIONS OF GROUPS

Since > 0 was arbitrary it follows that kS Sn kHS 0.


Note that because this does not depend on the choice of bases, we have that
unitaries U U(H) and V U(K) define a unitary on HS(H, K) by the map
S 7 U SV .
Suppose now that is a group and : U(H), : U(K) are
two unitary representations. We then have a unitary representation of on
HS(H, K) given by T 7 ()T ( 1 ), for each .
Consider the linear map between KH and HS(H, K) which takes a vector
KH to the operator () determined by the formula
h(), i = h, i,
for all H and K.
Then is a unitary since if , HS(H, K) we have that
h(), ( )iHS = i,j h()i , j ihj , ( )i i
= i,j h, j i ihj i , i = h, i.
It is easy to see that this isomorphism in turn gives an isomorphism between
and the representation of on HS(H, K) described above.

1.2 Functions of positive type


Definition 1.2.1. Given a group , a function of positive type on is a
map : C such that for all u C we have
, (1 ) 0.
One way in which function of positive types appear is when we have a unitary
representation : U(K), together with a vector K. Then function
defined by:
() = h(), i, for all ,
describes a function of positive type on . Indeed, if u C then
, (1 ) = , h(1 ), i
= k ()k2 0.
Every function of positive type arrises in this way. Specifically, if : C
is a function of positive type then we may place a pseudo-inner product on C
by the formula:
h u , u i = , (1 ).
The positivity of this inner product is given by the fact that is of positive
type. After quotienting by the kernel and completion we obtain a Hilbert space
K. Moreover, there is a natural unitary representation of on K given by
(0 ) u = u0 .
1.3. COCYCLES 9

The fact that this is well defined and describes a unitary follows from the
fact that for all 0 , and u C we have

k u0 k2 = , ((0 )1 (0 ))

= , (1 ) = k u k2 .
One has to check that this action is well defined and preserves the inner
product structure, and then the cyclic vector = ue gives the formula above.
If we start with a unitary representation : U(H), with a non-zero
vector H, consider the function of positive type , and then consider
the representation . Then the map u 7 () defines an
isometry from K to H which is equivariant. If is a cyclic vector for then
this gives a natural isomorphism between and .
One can easily check that the correspondence described above satisfies the
following relationships:

functions of positive type pointed unitary representations


=1 the trivial representation
= e the left regular representation on 2 , = e
= , < the quasi-regular representation on 2 (/), =
a character the one dimensional representation given by the character
addition direct sum
product tensor product

1.3 Cocycles
Definition 1.3.1. Suppose yX is an action of a group on a set X, and
is a group. A cocycle for the action into is a map : X such that

(1 2 , x) = (1 , 2 x)(2 , x),

for all 1 , 2 , and x X. Two cocycles , : X are cohomologous


if there is a map : X such that

(, x) = (x)(, x)(x)1 ,

for all , x X. A cocycle is trivial if it is cohomologous to the


cocycle which takes constant value e . A cocycle untwists if there is a
homomorphism : such that is cohomologous to the cocycle given by
(, x) 7 (), for all , and x X.
The set of all cocycles for the action yX with values in is denoted by
Z 1 (yX; ), the set of trivial cocycles is denoted by B 1 (yX; ), and the
set of equivalence classes of cohomologous cocycles is denoted by H 1 (yX; ).
Note that if is abelian then Z 1 (yX; ) is an abelian group under pointwise
multiplication, B 1 (yX; ) is a subgroup and H 1 (yX; ) is the quotient
group.
10 CHAPTER 1. REPRESENTATIONS OF GROUPS

Example 1.3.2. If : X is a cocycle for an action yX on a group


, and if S is a generating set, then the cocycle relation implies that the
cocycle is completely determined by its values on the set S X X. In
particular, for the group Z, a cocycle : Z X for an action ZyX is
completely determined by the function x 7 (1, x). Moreover, any function
: X determines a cocycle such that (1, x) = (x).
For instance if T : X X is a bijection and ZyX by n x = T n (x), and if
: X R is a function, then the corresponding cocycle
n1
(n, x) = k=0 T k (x),
for n > 0. Hence, if we denote by Sn : X R the function given by
Sn (x) = (n, T n (x)),
1
then n Sn is the average of the functions T k for 1 k n.
Example 1.3.3. Suppose and are two groups and consider the space
[, ] = {f : | f (e) = e} .
We have an action of on this set by
( f )(x) = f (x)f ()1 .
Note that the fixed points of this action are precisely the set of homomor-
phisms from to .
We then have a cocycle for this action : [, ] given by
(, f ) = f ().
Note that this action preserves the subsets of injective, surjective, and bijec-
tive maps.
Exercise 1.3.4. Think about the previous example and verify all the claims.
Example 1.3.5. One natural way in which cocycles arise is if we have a set X
together with a pair of actions yX, and yX of groups , and .
If the action of is free (if 6= e then x 6= x for all x X), and if the
orbits of are contained in the orbits of (x x for each x X), then we
can define a cocycle : X by setting (, x) to be the unique element
in such that
x = (, x)x.
We can verify that is a cocycle since for each 1 , 2 and x X we have
(1 2 , x)x = 1 2 x
= (1 , 2 x)(1 x) = (1 , 2 x)(2 , x)x.
Since the action yX is free this then implies that
(1 2 , x) = (1 , 2 x)(2 , x).
1.3. COCYCLES 11

Definition 1.3.6. If yY is an action of a group on a set Y then a fun-


damental domain for the action is a subset D Y such that each orbit of
contains exactly one element in D, i.e., for all x Y we have |x D| = 1.
Note that if D Y is a fundamental domain for the action yY then we
have that the map from D to Y given by (, x) 7 x is sujective. Moreover,
if the action yY is free then the map is also injective and hence the space Y
decomposes as a disjoint union of orbits

Y = xD x.

We thus obtain a bijection between the fundamental domain D and the orbit
space \X given by x 7 x.
Example 1.3.7. Another way in which cocycles arise is if Y is a set and we have
a left action of on Y and a right action of such that the actions commute,
i.e., for each , , and x Y we have

(x) = (x).

As in Example 1.3.5 we will require the action of to be free.


Since the actions of and commute, the action of on Y passes to the
space of orbits Y /. If D Y is a fundamental domain for the action Y x
then as mentioned above the map from D to Y / given by x 7 x is a bijection.
We will denote by : Y / D the inverse map.
We then obtain a cocycle : (Y /) by assigning to each and
x Y / the value of (, x) to be the unique element in such that

(x) = (x)(, x).

If 1 , 2 and x Y / then we have

(1 2 x)(1 2 , x) = 1 2 (x)

= 1 (2 x)(2 , x)
= (1 2 x)(1 , 2 x)(2 , x).
Since the action Y x is free this shows that is indeed a cocycle.
Suppose D Y is also a fundamental domain for the action Y x and
: (Y /) is the corresponding cocycle. If we let : Y / D be the
corresponding selection map for D , and we define : Y / so that for all
x Y / we have
(x) = (x)(x)1 ,
then its easy to see that for each and x Y / we have

(x)(, x)(x)1 = (, x).

Hence, the cohomology class of is independent of the fundamental domain D.


12 CHAPTER 1. REPRESENTATIONS OF GROUPS

Example 1.3.8. A special case of the above example to keep in mind is when
is a group and < is a subgroup. Then y x is a pair of commuting
actions given by group multiplication. Thus, we obtain a cocycle (unique up to
cohomology) : (/) .
Exercise 1.3.9. Show that in Example 1.3.8 if the action x has a fun-
damental domain 0 which is a subgroup of , then we have that
is a normal subgroup and splits as a semidirect product = 0 . Also,
compute the corresponding cocycle for this fundamental domain.

1.4 Induced representations


Given an action yX of a group on a set X, a cocycle : X into
a group , and a unitary representation : U(K), we obtain an induced
representation Ind 2
: U( XK) by linearly extending the formula

Ind
()(x ) = x (((, x))).

We can easily check that this is a representation since for all 1 , 2 , x X,


and K we have

Ind
(1 2 )(x ) = 1 2 x (((1 2 , x)))

= 1 2 x (((1 , 2 x))((2 , x)))


= Ind
(1 )(2 x (((2 , x)))

= Ind
(1 ) Ind (2 )(x ).

If : X is cocycle which is cohomologous to , and : X such


that (, x) = (x)(, x)(x)1 , for all , and x X, then we obtain a
unitary U U(2 XK), by linearly extending the formula

U (x ) = x ((x)1 ).

We then see easily that for all , x X, and K we have

U Ind ()U (x ) = U Ind ()(x ((x)1 ))

= U (x ((, x))((x)1 ))
= x ((x))((, x))((x)1 )
= Ind
()(x ).

Hence, the representations Ind and Ind


are isomorphic.

Lemma 1.4.1. Let yX be an action of a group on a set X, let be a group,


and suppose : X is a cocycle. If : U(H) and : U(K)
are unitary representations such that
= then Ind
= Ind .
1.4. INDUCED REPRESENTATIONS 13

Proof. If W : H K is a unitary such that W () = ()W for all .


Then id W : 2 X H 2 X K, and it follows easily that

(id W ) Ind
() = Ind ()(id W ).

Example 1.4.2. Consider a group acting on itself by left multiplication,


and let : U(H) be a unitary representation. Then we obtain cocycles
, : given by (, x) = , and (, x) = e.
By considering the function (x) = x we see that these cocycles are coho-
mologous, for every , x we have

(, x) = = (x)x1 = (x)(, x)(x)1 .

We therefore recover Fells absorption principle by obtaining an isomorphism



between the representations = Ind and 1H = Ind .

If is a group, and < is a subgroup, then we may consider the cocycle


: (/) associated to some fundamental domain for as is described
in Example 1.3.8. Hence if : U(H) is a unitary representation of then
we obtain the induced representation Ind : U(2 (/)H) as the
induced representation associated to the cocycle . Since the cohomology class
of does not depend on the fundamental domain, we have that the induced
representation is well defined up to unitary equivalence.
Observe that if 0 H, then we may consider the vector = 0
2
(/)H. We then have that the positive definite function : C is
given by 
() if ;
() =
0 otherwise.
Also observe that if = 1H : U(H) is the trivial representation then
Ind = / 1H is a multiple of the quasi-regular representation correspond-
ing to .

Remark 1.4.3. Suppose Y is a set and we have commuting left and right
actions yY x of groups and such that the action of is free. If :
U(H) is a unitary representation, and : (Y /) is the cocycle coming
from a fundamental domain D for Y x as explained in Example 1.3.7, then
we obtain the induced representation Ind . Here we will explain an alternate
way to obtain this representation which makes no reference to a fundamental
domain.
Consider a function : X H which is -equivariant, i.e., (x1 ) =
()(x), for all x X, and . Because, is -equivariant we have that
the function x 7 k(x)k is constant on the -orbits. We may therefore consider
the well defined space

L2 (Y ; H) = { : Y H | is equivariant, and xY / k(x)k2 < }.


14 CHAPTER 1. REPRESENTATIONS OF GROUPS

Since this space consists of -equivariant functions, this becomes a Hilbert space
when it is endowed with the well defined inner-product

h, i = xY / h(x), (x)i.

We then obtain the induced representation of on L2 (Y ; H) by requiring


that for all and x Y we have

(Ind ())(x) = ( 1 x).

If D Y is a fundamental domain for the action Y x , then any func-


tion 2 DH = 2 (D; H) extends uniquely to a -equivariant function
(Y ; H) by setting (x1 ) = ()(x) for each and x D.
2

By identifying Y / with D by the map we then obtain a unitary W :


2 (Y /)H 2 (Y ; H). If : (Y /) is the cocycle from Exam-
ple 1.3.7, then unwinding the definitions gives

W Ind
() = Ind ()W,

for all .
Lemma 1.4.4. Let be a group and < < subgroups. If : U(H)
is a unitary representation then

Ind
= Ind Ind
.

Proof. Using the equivalence in the previous remark, we may consider the map
W : L2 (; L2 (; H) ) L2 (; H) given by

(W f )() = (f ())(e).

Note that if then we have

(W f )( 1 ) = (f ( 1 ))(e) = (Ind
()f ())(e)

= (f ())( 1 )
= ()((f ())(e)) = ()((W f )()).
Also, if D is a set of coset representatives for , and E is a set of coset
representatives for then ED is a set of coset representatives for , hence
just as above we see that for all f L2 (; L2 (; H) ) we have

kW f k2 = (,)ED k(W f )()k2

= E D k((W f )())()k2 = kf k2 .
Thus, W is a well defined isometry, which is easy to see is a unitary. An easy
calculation then shows that for all we have

Ind ()W = W Ind Ind


().
1.4. INDUCED REPRESENTATIONS 15

Proposition 1.4.5. If : U(H) is a unitary representation of and


: U(K) is a unitary representation of , then

Ind
= Ind (| ).

Proof. Consider the map W : HL2 (; K) L2 (; HK) such that for each
, H, and f L2 (; K) we have

(W ( f ))() = (( 1 )) f ().

Then, if we have

(W ( f ))( 1 ) = (()) f ( 1 )

= ( )()(W ( f )).
Hence, it follows easily that W is a well defined unitary operator. A routine
check then shows that

Ind (| )()W = W ( Ind )(),

for all .
If , , and are groups, yX, and yY are actions, and : X ,
and : Y are cocycles then just as we induced representations above
we may induce the action yX to an action yX Y by the formula

(x, y) = (x, (, x)y).

We then may define the composition of cocycles : (X Y ) by


the formula
(, (x, y)) = ((, x), y).
We can verify that this is indeed a cocycle since for all 1 , 2 , and (x, y)
X Y we have

(1 2 , (x, y)) = ((1 , 2 x)(2 , x), y)

= ((1 , 2 x), (2 , x)y)((2 , x), y)


= (1 , (2 x, (2 , x)y))(2 , (x, y))
= (1 , 2 (x, y))(2 , (x, y)).
Exercise 1.4.6. If we have an inclusion of groups < < and we consider
the cocycles < , < , and < as described in Example 1.3.8, then show
that by identifying the sets (/) (/) and / we obtain an identification
of < < and < .
In light of the previous exercise we then have that the following lemma is an
extension of Lemma 1.4.4.
16 CHAPTER 1. REPRESENTATIONS OF GROUPS

Lemma 1.4.7. Let , , and be groups, yX, and yY be actions, and


: X , and : Y be cocycles. If : U(H) is a unitary
representation, then
Ind
= Ind Ind .

Proof. If we consider the natural identification V : 2 X 2 Y 2 (X Y ),


then it is easy to see that V idH implements an equivalence between the
representations Ind Ind
and Ind .

Lemma 1.4.8. Let yX be a action of a group on a set X, let be a group,


and let : X be a cocycle, if i : U(Hi ), i I is a family of
unitary representations, then

Ind
(i i ) = iI Ind (i ).

Proof. This follows easily by considering the natural unitary from iI (2 X


Hi ) to 2 X (iI Hi ).

1.5 Invariant vectors


Definition 1.5.1. Let be a group, a unitary representation : U(H)
contains invariant vectors if there exists a non-zero vector H such that
() = for all . The representation contains almost invariant vectors
if for each F , and > 0, there exists H, such that

k() k < kk, for all F.

We will also say that a representation : U(H) is ergodic if it does


not contain invariant vectors. In general, if we denote by H0 H the subspace
of -invariant vectors, then we say the representation has spectral gap if
H0 6= H and the sub-representation |H 0
does not contain almost invariant
vectors.

Proposition 1.5.2. Let be a group, and : U(H) a unitary represen-


tation. If there exists H and c > 0 such that (h(), i) > ckk2 for all
, then contains an invariant vector.

Proof. Let K be the closed convex hull of the orbit (). We therefore have
that K is -invariant and (h, i) ckk2 for every K. Let 0 K be the
unique element of minimal norm, then since acts isometrically we have that
for each , ()0 is the unique element of minimal norm for ()K = K,
and hence ()0 = 0 for each . Since 0 K we have that (h0 , i) 6= 0,
and hence 0 6= 0.

and : U(H) a unitary representation.


Corollary 1.5.3. Let be a group,
If there exists H and c < 2 such that k() k < ckk for all ,
then contains an invariant vector.
1.5. INVARIANT VECTORS 17

Proof. For each we have


2(h(), i) = 2kk2 k() k2 (2 c2 )kk2 .
Hence, we may apply Proposition 1.5.2.
Lemma 1.5.4. Let be a group, and : U(H) a unitary representation.
Then contains almost invariant vectors if and only if n contains almost
invariant vector, where n 1 is any cardinal number.
Proof. If does not contain almost invariant vectors then there exists c > 0,
and S finite, such that for all H we have
ckk2 S k() k2 .
If I is a set |I| = n, and i H for i I, such that iI ki k2 < , then
ck iI i k2 = iI cki k2
iI S k()i i k2 = S k n ()(iI i ) iI i k2 .
Hence, n does not contain almost invariant vectors. The converse is trivial
since is contained in .
Proposition 1.5.5. Let be a group, and : U(H) a unitary represen-
tation. Then contains almost invariant vectors if and only if for any finite
1
symmetric set S the operator TS = |S| S () satisfies kTS k = 1.
Proof. If contains almost invariant vectors then the triangle inequality easily
implies that 1 is in the spectrum of TS , for each finite symmetric set S .
Conversely, if S is a finite symmetric set with e S and kTS k = 1, then
since TS is self-adjoint either 1 or 1 is contained in its spectrum, however since
e S it is easy to see that 1 6 (TS ), hence for any > 0 there exists H,
kk = 1 such that
|1 hTS , i| < 2 /2|S|.
We then have that for each S
k ()k2 = 2|1 h(), i|
2|S||1 hTS , i| < 2 .
Since, and S were arbitrary this shows that contains almost invariant vectors.

Proposition 1.5.6. Let be a group, and : U(H) a unitary representa-


tion. The following are equivalent:
(1). The representation contains invariant vectors.
(2). The representation contains invariant vectors for some unitary
representation : U(K).
(3). The representation contains a finite dimensional sub-representation.
18 CHAPTER 1. REPRESENTATIONS OF GROUPS

Proof. (1) = (2) is obvious.


To show (2) = (3) suppose : U(K) is a unitary representation such
that contains invariant vectors. Identifying HK with the space of Hilbert-
Schmidt operators HS(K, H) we then have that there exists T HS(K, H),
non-zero, such that ()T ( 1 ) = T , for all . Then T T B(H, H)
is positive, non-zero, compact, and ()T T ( 1 ) = T T , for all . By
taking the range of a non-trivial spectral projection of T T we then obtain a
finite dimensional invariant subspace of .
(3) = (1) follows because if L H is a finite dimensional invariant
subspace then ProjL is a finite rank projection such that () ProjL ( 1 ) =
ProjL , for all . By identifying HS(H, H) with HH, we then obtain a
non-zero invariant vector for .

1.6 Amenability
Much of the material of this section has been taken from Section 2.6 in the book
of Brown and Ozawa [BO08].

Definition 1.6.1. Let be a group, a Flner net is a net of non-empty finite


subsets Fi such that |Fi Fi |/|Fi | 0, for all .

Note that we do not require that = i Fi , nor do we require that Fi are


increasing, however, if || = then it is easy to see that any Flner net {Fi }i
must satisfy |Fi | .

Exercise 1.6.2. Let be a group, show that if has a Flner net, then for
each finite set E , and > 0, there exists a finite set F such that
E F , F = F 1 , and |F F |/|F |, |F F |/|F | < .

Definition 1.6.3. A mean m on a non-empty set X is a finitely additive


probability measure on 2X , i.e., m : 2X [0, 1] such that m(X) = 1, and if
A1 , . . . , An X are disjoint then m(nj=1 An ) = nj=1 m(An ).

Given a mean m on X it is possible to define an integral over X just as in


the case if m wereRa measure. We therefore obtain a state m ( X) by the
formula m (f ) = X f dm. Conversely, if ( X) is a state, then restricting
to characteristic functions defines a corresponding mean.
If yX is an action, then an invariant mean m on X is a mean such that
m(A) = m(A) for all A X. An approximately invariant mean is a net of
probability measures i Prob(X), such that k i i k1 0, for all .

Definition 1.6.4. Let be a group, is amenable if it has a mean which


is invariant under the action of left multiplication, or equivalently, if there is a
state on which is invariant under the action of left multiplication.

Amenable groups were first introduced by von Neumann [vN29]. The term
amenable was coined by M. M. Day.
1.6. AMENABILITY 19

Theorem 1.6.5. Let be a group. The following conditions are equivalent.


(1). is amenable.
(2). has an approximate invariant mean under the action of left multipli-
cation.
(3). has a Flner net.
(4). The left regular representation : U(2 ) contains almost invariant
vectors.
1
(5). For any finite symmetric set S the operator TS = |S| S ()
satisfies kTS k = 1.
(6). There exists a state (B(2 )) such that (()T ) = (T ()) for
all , T B(2 ).
(7). The continuous action of on its Stone-Cech compactification which
is induced by left-multiplication admits an invariant Radon probability measure.
(8). Any continuous action yK on a compact Hausdorff space K admits
an invariant Radon probability measure.

Proof. We show (1) = (2) using the method of Day [Day57]. Since =
(1 ) , the unit ball in 1 is wk -dense in the unit ball of ( ) = (1 ) .
It follows that Prob() 1 is wk -dense in the state space of .
Let S , be finite and let K S 1 be the wk-closure of the set
{S ( ) | Prob()}. From the remarks above, since has a left
invariant state on , we have that 0 K. However, K is convex and so by
the Hahn-Banach Separation Theorem the wk-closure coincides with the norm
closure. Thus, for any > 0 there exists Prob() such that

S k k1 < .

We show (2) = (3) using the method of Namioka [Nam64]. Let S be


a finite set, and denote by Er the characteristic function on the set (r, ). If
Prob() then we have

S k k1 = S x | (x) (x)|
Z
= S x |Er ( (x)) Er ((x))| dr
R0
Z
= S x |Er ( (x)) Er ((x))| dr
R0
Z
= S kEr ( ) Er ()k1 dr.
R0
20 CHAPTER 1. REPRESENTATIONS OF GROUPS

By hypothesis, if > 0 then there exists Prob() such that S k


k1 < , and hence for this we have
Z Z
S kEr ( ) Er ()k1 dr < = kEr ()k1 dr.
R0 R0

Hence, if we denote by Fr the (finite) support of Er (), then for some


particular r > 0 we must have

S |Fr Fr | = S kEr ( ) Er ()k1 < kEr ()k1 = |Fr |.

For (3) = (4) just notice that if Fi is a Flner net, then |Fi1|1/2 xFi x
2 is a net of almost invariant vectors.
(4) (5) follows from Proposition 1.5.5.
For (4) = (6) let i 2 be a net of almost invariant vectors for .
We define states i on B(2 ) by i (T ) = hT i , i i. By wk-compactness of the
state space, we may take a subnet and assume that this converges in the weak
topology to B(2 ) . We then have that for all T B(2 ) and ,

|(()T T ())| = lim |h(()T T ())i , i i|


i

= lim |hT i , ( 1 )i i hT ()i , i i|


i

lim kT k(k( 1)i i k + k()i i k) = 0.


i

For (6) = (1), we have a natural embedding M : B(2 ) as


diagonal matrices, i.e., for a function f we have Mf (x x x ) =
x x f (x)x . Moreover, for f and we have ()Mf ( 1 ) =
Mf . Thus, if B(2 ) is a state which is invariant under the conjugation
by (), then restricting this state to gives a state on which is -
invariant.
For (1) = (7), the map : C() which takes a bounded function
on to its unique continuous extension on , is a C -algebra isomorphism,
which is -equivariant. Hence amenability of implies the existence of a -
invariant state on C(). The Riesz Representation Theorem then gives an
invariant probability measure on .
For (7) (8), suppose acts continuously on a compact Hausdorff space
K, and fix a point x0 K. Then the map f () = x0 on extends uniquely
to a continuous map f : K, moreover since f is -equivariant, so is f .
If is an invariant Radon probability measure for the action on then we
obtain the invariant Radon probability measure f on K. Since itself is
compact, the converse is trivial.
For (7) = (1), if there is a -invariant Radon probability measure on
, then we obtain an invariant mean m on my setting m(A) = (A).
Example 1.6.6. Any finite group is amenable, and any group which is locally
amenable (each finitely generated subgroup is amenable) is also amenable. The
1.6. AMENABILITY 21

group of integers Z is amenable (consider the Flner sequence Fn = {1, . . . , n}


for example). From this it follows easily that all finitely generated abelian groups
are amenable, and hence all abelian groups are.
It is also easy to see from the definition that subgroups of amenable groups
are amenable. A bit more difficult is to show that if 1 1 is
an exact sequence of groups then is amenable if and only if both and are
amenable. Thus all solvable groups, and even all nilpotent groups are amenable.
There are also finitely generated amenable groups which cannot be con-
structed from finite, and abelian groups using only the operations above [Gri84].
Example 1.6.7. Let F2 be the free group on two generators a, and b. Let A+
be the set of all elements in F2 whose leftmost entry in reduced form is a, let
A be the set of all elements in F2 whose leftmost entry in reduced form is a1 ,
let B + , and B be defined analogously, and consider C = {e, b, b2, . . .}. Then
we have that
F2 = A+ A (B + \ C) (B C)
= A+ aA
= b1 (B + \ C) (B C).
If m were a left-invariant mean on F2 then we would have

m(F2 ) = m(A+ ) + m(A ) + m(B + \ C) + m(B C)

= m(A+ ) + m(aA ) + m(b1 (B + \ C)) + m(B C)


= m(A+ aA ) + m(b1 (B + \ C) (B C)) = 2m(F2 ).
Hence, F2 and also any group containing F2 is not-amenable. There are also
finitely generated nonamenable groups which do not contain F2 [Ols80], and even
finitely presented nonamenable groups which do not contain F2 [OS02].

1.6.1 Von Neumanns Mean Ergodic Theorem


Amenable groups allow for nice averaging properties, we give such an example
here.

Theorem 1.6.8. Let be an amenable group with Flner net Fi , and let
: U(H) be a unitary representation. Let P0 B(H) be the projection
onto the subspace of invariant vectors. Then for each H we have that
1
k 1 () P0 ()k 0.
|Fi | Fi

Proof. By considering the vector P0 () instead, we may assume that P0 () =


0.
Note that if , and H then () is orthogonal to the space of
invariant vectors. Indeed, if H is an invariant vector, then h (), i =
h, ( 1 )i = 0.
22 CHAPTER 1. REPRESENTATIONS OF GROUPS

Denote by L the closure of the subspace of H spanned by vectors of the form


() for H, . Then if L we have that 0 = h, ()i =
h ( 1 ), i for all , and H, thus () = 0, for all .
Therefore we have shown that L is precisely the space of invariant vectors.
Fix 0 , and H, then if = (0 ) we have that

1 1
k F 1 ()k = k 1 (() (0 ))k
|Fi | i |Fi | Fi

(|Fi 0 Fi |/|Fi |)kk 0.

Since |F1i | F 1 () B(H) is always a contraction we may then pass to the


i
closure of the span to conclude that for all L we have

1
k 1 ()k 0.
|Fi | Fi

In the case when the representation is ergodic, = Z, and the we consider


the Flner sequence Fn = {0, 1, . . . , n + 1}, the above theorem then gives
the following corollary.

Corollary 1.6.9. Let : Z U(H) be a unitary representation. If is ergodic,


then for each H we have that

1 n1
lim k k=0 (k)k = 0.
n n

We remark that from the perspective of Example 1.3.2 another possible


generalization of Corollary 1.6.9 could be given in terms of cocycles. This per-
spective would then lead to Corollary 3.3 in [dCTV07] the proof of which is
quite similar to the proof of von Neumanns Ergodic Theorem given above.

1.7 Mixing properties


Definition 1.7.1. Let be a group, a unitary representation : U(H) is
weak mixing if for each finite set F H, and > 0 there exists such
that
|h(), i| < ,

for all F .
The representation is (strong) mixing if || = , and for each finite set
F H, we have
lim |h(), i| = 0.

1.7. MIXING PROPERTIES 23

These definitions should be compared with the characterization of ergodicity


found in Proposition 1.5.2. Hence we see that mixing implies weak mixing, which
in turn implies ergodic. It is also easy to see that if : U(H) is mixing
(resp. weak mixing) then so is , and if is mixing then so is for any
representation . Well see below in Corollary 1.7.6 that weak mixing is also
stable under tensoring.

Exercise 1.7.2. Let be a group and < a subgroup. Show that the
quasi-regular representation of on 2 (/) is weak mixing if and only if it is
ergodic, if and only if [ : ] = . Show that it is mixing if and only if || =
and || < .
Exercise 1.7.3. Let yX be an action of a group on a set X, let be
a group and let : X be a cocycle. Suppose : U(H) is a
unitary representation. Find necessary and sufficient conditions for the induced
representation of Section 1.4 to be mixing.

Lemma 1.7.4. Let be a group, a unitary representation : U(H) is


weak mixing if and only if for each finite set F H, and > 0 there exists
such that
|h(), i| < ,
for all , F .
The representation is mixing if || = and for each finite set F H, we
have
lim |h(), i| = 0,

for all , F .
Proof. This follows from the polarization identity. For each , H, and
we have
1
h(), i = 3k=0 ik h()( + ik ), ( + ik )i.
4

We now add another equivalent condition to Proposition 1.5.6.

Proposition 1.7.5. Let be a group, and let : U(H) be a unitary repre-


sentation. Then is weak mixing if and only if contains no finite dimensional
sup-representations.

Proof. If is weak mixing then if L H is a non-trivial, finite dimensional sub-


p with orthonormal basis F H, there exists such that |h(), i| <
space
1/ dim(L), for all , F . Hence, if F then k ProjL (())k < 1 = kk.
Thus, L is not an invariant subspace.
Conversely, If has no finite dimensional invariant subspaces, L H is a
finite dimensional subspace, and > 0, then there exists such that for all
L we have that
k ProjL (())k < .
24 CHAPTER 1. REPRESENTATIONS OF GROUPS

Indeed, if this is not the case then we would have that there exists c > 0 such
that

h( 1 ) ProjL (), ProjL iHS sup k ProjL (())k2 c.


L,kk=1

It would then follow Proposition 1.5.2 that there is a non-zero Hilbert-Schmidt


operator T such that ()T ( 1 ) = T , for all . Proposition 1.5.6 would
then give a contradiction.
Thus, if F H is finite such that kk 1 for each F , then by considering
the finite dimensional subspace L spanned by F we have shown that there exists
such that for all , F we have

|h(), i| k ProjL (())kkk < .

From the above proposition together with the equivalence between conditions
2) and 3) in Proposition 1.5.6 we obtain the following.

Corollary 1.7.6. Let be a group and let : U(H) be a unitary repre-


sentation. Then is weak mixing if and only if is weak mixing, if and
only if is weak mixing for all unitary representations .

Corollary 1.7.7. Let be a group and let : U(H) be a weak mixing


unitary representation. If < is a finite index subgroup then | is also weak
mixing.

Proof. Let D be a set of coset representatives for . If | is not mixing,


then by Proposition 1.7.5 there is a finite dimensional subspace L H which
is -invariant. We then have that D ()(L) H is finite dimensional and
-invariant. Hence, again by Proposition 1.7.5, is not weak mixing.

1.7.1 Compact representations


If H is a Hilbert space the strong operator topology on B(H) is such that
limi Ti = T if and only if limi k(Ti T )k = 0, for all H. The weak operator
topology on B(H) is such that limi Ti = T if and only if limi h(Ti T ), i = 0, for
all , H. The unitary group U(H) then becomes a topological group when
endowed with the strong operator topology. We also note that if Ui U(H)
is a net of unitaries such that Ui U U(H) in the weak operator topology,
then we also have that Ui U in the strong operator topology. Indeed, for any
H we have

k(Ui U )k2 = 2kk2 2(hUi , U i) 0.

Definition 1.7.8. Let be a group, a unitary representation : U(H) is


compact if () U(H) is pre-compact in the strong operator topology.
1.7. MIXING PROPERTIES 25

Lemma 1.7.9. Let (X, d) be a compact metric space, then Isom(X, d) with the
topology of point-wise convergence is compact.
Proof. By Tychonoffs Theorem X X is compact, thus we need only to show
that Isom(X, d) X X is closed. Suppose g Isom(X, d), for all x, y X we
have that {f X X | d(f (x), f (y)) = d(x, y)} is closed and contains Isom(X, d),
hence g is isometric.
For each x X denote by dx the distance from x to g(X). Then we have
that for all m N, d(x, g m (x)) dx , hence for all n, m N, n < m we have
d(g n (x), g m (x)) = d(x, g mn (x)) dx . Since X is compact, it must be totally
bounded and hence we must have that dx = 0. Hence g is surjective and thus
Isom(X, d) is compact.
Lemma 1.7.10. Let : U(H) be a unitary representation of a group .
Then is compact if and only if for each H, the orbit () is pre-compact
in H.
Proof. If () is pre-compact in the strong operator topology and G = (),
then for each H we have that () = G is compact, being the continuous
image of the compact set G.
Conversely, suppose that each orbit () H is pre-compact. By Zorns
Lemma we can find a collection of vectors J H such that H = J sp().
We therefore have a strong operator topology continuous embedding of ()
into the compact space J Isom((), dH ) U(J sp()), hence ()
is pre-compact.
The following is part of the Peter-Weyl Theorem.
Theorem 1.7.11. Let G be a compact group, and let : G U(H) be a strong
operator topology continuous unitary representation. Then decomposes as a
direct sum of finite dimensional representations.
Proof. Let denote the Haar measure on G. Since every representation decom-
poses into a direct sum of cyclic representations we may assume that the rep-
resentation has a cyclic vector H. We then define an operator K B(H)
such that for all , H we have
Z
hK, i = h(g), ih, (g)i d(g).
G

First, note that from left invariance of the Haar measure we have that
(g)K(g 1 ) = K, for all g G. Also, if H, then we have
Z
hK, i = |h(g), i|2 d(g) 0.
G

Thus, K is positive.
Moreover, if hK, i = 0 then we have that h(g), i = h, (g 1 )i = 0,
for all g G. This then implies that = 0 since is a cyclic vector. Thus K is
strictly positive.
26 CHAPTER 1. REPRESENTATIONS OF GROUPS

If n H is a bounded sequence which weakly converges to 0, then


Z
lim kKn k2 = lim h(g)n , ih, (g)Kn i d(g)
n n G
Z Z
= lim h(g)n , ih, (h)n ih(h)(g 1 ), i d(g) d(h) = 0.
n G G
We therefore have shown that K is a compact operator. Hence, H decom-
poses as a direct sum of the (finite dimensional) eigenspaces of K, each of which
is G-invariant.
Lemma 1.7.10 and Theorem 1.7.11 then give us a structural result for com-
pact representations.
Corollary 1.7.12. Let : U(H) be a representation of a group . Then
is compact if and only if decomposes as a direct sum of finite dimensional
representations.
Proposition 1.7.13. Let be a group and let : U(H) be a unitary
representation. Then there is a unique -invariant closed subspace K H such
that |K is compact and |K is weak mixing.
Proof. Let Z be the set of all orthonormal sets J H such that sp(()) is
finite dimensional for all J and sp(()) sp(()) for all , J ,
6= . If we order Z by inclusion then it is easy to see that the union of any
increasing chain in Z is again in Z.
If is not weakly mixing then by Proposition 1.7.5 we have that Z =
6 , hence
by Zorns Lemma there is a maximal element J Z. Let K = J sp(())
H, then |K = J |sp(()) is compact, and by maximality of J we have that
|K contains no finite dimensional sub-representation, and hence is weakly
mixing.
If K0 H is a finite dimensional -invariant subspace, then ProjK (K0 )
K is also a finite dimensional -invariant subspace. Since |K is weak mixing
it then follows that ProjK (K0 ) = {0} and hence K0 K. This then implies
uniqueness of the decomposition.

1.7.2 Weak mixing for amenable groups


Proposition 1.7.14. Let be an amenable group with a Flner net Fi ,
and let : U(H) be a unitary representation. Then is weak mixing if
and only if for each , H we have
1 2
1 |h(), i| 0.
|Fi | Fi
Proof. If is not weak mixing, then there exists a finite set F H, and > 0
such that ,F |h(), i|2 > , for all . It then follows that
1 2
lim inf 1 ,F |h(), i| .
i |Fi | Fi
1.7. MIXING PROPERTIES 27

Hence, by the pigeonhole principle, for some , F we have that


1 2 2
lim sup 1 |h(), i| /|F | > 0.
i |Fi | Fi
Conversely, if is weak mixing then by Propositions 1.5.6 and 1.7.5 we have
that is ergodic. Hence by von Neumanns Ergodic Theorem the operators
1
|Fi | Fi ( )() converge in the strong operator topology (and hence also
1

in the weak operator topology) to 0. If , H we then have


1 1
F 1 |h(), i|2 = 1 h( )()( ), i 0.
|Fi | i |Fi | Fi

Lemma 1.7.15 (Koopman, von Neumann 1932). Let X be a set, and let Fi
X, i I be a net of non-empty finite subsets of X. Suppose a X, then
1
xFi |a(x)| 0
|Fi |
if and only if
1
xFi |a(x)|2 0.
|Fi |
Proof. By the Cauchy-Schwarz inequality we have
1 1 1
xFi |a(x)| ( xFi |a(x)|2 )1/2 kak1/2
( xFi |a(x)|)1/2 .
|Fi | |Fi | |Fi |

Corollary 1.7.16. Let be an amenable group with a Flner net Fi , and


let : U(H) be a unitary representation. Then is weak mixing if and
only if for each , H we have
1
1 |h(), i| 0.
|Fi | Fi
Specializing to representations of the integers we also have the following
spectral characterization of weak mixing.
Proposition 1.7.17. Let : Z U(H) be a unitary representation. Then
is weak mixing if and only if (1) has no eigenvalues.
Proof. If is weak mixing then by Proposition 1.7.5 there is no finite dimen-
sional subspace of H which is Z invariant. In particular, this shows that (1)
has no eigenvalues.
Conversely, suppose (1) has no eigenvalues. Then has no 1-dimensional
invariant subspaces. However, since any unitary matrix in Mn (C) has an eigen-
value, it then follows that has no finite dimensional invariant subspaces.
Hence, is weak mixing by Proposition 1.7.5.
28 CHAPTER 1. REPRESENTATIONS OF GROUPS

1.8 Weak containment


Definition 1.8.1. Let be a group, and : U(H), : U(K) two
unitary representations. The representation is weakly contained in the
representation (written ) if for each H, F finite, and > 0,
there exists 1 , . . . , n K such that

|h(), i nj=1 h()j , j i| < ,

for all F . The representations and are weakly equivalent (written )


if and .
It follows easily from the definition that for any representation .
Also, it is easy to see that containment implies weak containment, and weak
containment is a partial order. We also have that if i , and i , i I are families
of representations such that i i for each i I, then iI i iI i .
Exercise 1.8.2. Show that if 1 , 2 , 1 , and 2 are unitary representations of
a group such that i i , for i {1, 2}, then 1 2 1 2 .
Example 1.8.3. If a representation : U(H) weakly contains the trivial
representation then for any finite symmetric set S and > 0 there exists
1 , . . . , n H such that

|1 nj=1 h()j , j i| < ,

for all S {e}. Hence, if we denote by = nj=1 j Hn H , then we


have |1 kk2 | < , and hence for each S we have

k ()k2 = 2(kk2 (h (), i)) < 4.

It follows that contains almost invariant vectors by Lemma 1.5.4.


Conversely, if contains almost invariant vectors, then it is easy to see that
weakly contains the trivial representation.
The following lemma from [Fel63] is a useful tool for checking if one repre-
sentation is weakly contained in another.
Lemma 1.8.4. Let : U(H) and : U(K) be two unitary representa-
tions of a group . Let L H be a set such that sp()L = H. Then if
and only if for each L, F finite, and > 0, there exists 1 , . . . , n K
such that
|h(), i nj=1 h()j , j i| < ,
for all F .
Proof. Suppose L H is as above, and consider X H the set of vectors H
such that the positive definite function 7 h(), i can be approximated
arbitrarily well on finite sets by sums of positive definite functions associated to
. By hypothesis L X , and we need to show that X = H.
1.8. WEAK CONTAINMENT 29

If X , K , and x x ux C, then from the formula

|h()(x x (x)), x x (x)i h()(x x (x)), x x (x)i|

x,y |y x ||h(y 1 x), i h(y 1 x), i|,


we see that x x (x) X . In particular, X is -invariant.
It is also easy to see that X is a closed set. Moreover, if , X are such
that sp(()) sp(() ), then it is easy to see that + X .
In general, we then have that if , X then

+ = ( + Projsp(()) ( )) + ( Projsp(()) ( )) X .

We therefore have shown that X is a closed -invariant subspace which contains


L and hence X = H.

If : C is a function of positive type and : U(H ) is the


corresponding representation described in Section 1.2, then is generated by
a single vector. We therefore obtain the following corollary.

Corollary 1.8.5. If : U(K) is a representation of a group , :


C is a function of positive type, and : U(H ) is the corresponding
representation. Then if and only if F finite, and > 0, there exists
1 , . . . , n K such that

|() nj=1 h()j , j i| < ,

for all F .

Exercise 1.8.6. Suppose yX is an action of a group on a set X, and


: X is a cocycle into a group . If : U(H), and : U(K)
are two representations such that , then show that Ind
Ind .
Conclude that if < is an amenable subgroup of a group then /
.

For further properties of weak containment a good place to look is Appendix


F in [BdlHV08].

Definition 1.8.7. [Bek90] Let : U(H) be a unitary representation of


a group , then is amenable if there exists a tate (B(H)) such that
(()T ) = (T ()) for all , T B(H).

Note that is amenable if and only if the left-regular representation is


amenable. We also have an analogue of Theorem 1.6.5 for amenable represen-
tations. The proof is similar, however we will not present it here.

Theorem 1.8.8. [Bek90] Let : U(H) be a unitary representation of a


group , then the following conditions are equivalent.
(1). is amenable.
30 CHAPTER 1. REPRESENTATIONS OF GROUPS

(2). There exists a net of trace class operators Ti B(H) such that kTi kTr =
1, and kTi () ()Ti kTr 0, for all .
(3). There exists a net of finite rank projections Pi B(H) such that
1
kPi kHS kPi ()
()Pi kHS 0, for all .

(4). contains almost invariant vectors.


1
(5). For any finite symmetric set S the operator TS = |S| S ()
satisfies kTS k = 1.
Chapter 2

Group actions on measure


spaces

2.1 Examples
Definition 2.1.1. Let (X, B, ) be a -finite measure space, and let be a
countable group, an action yX such that 1 (B) = B, for all is mea-
sure preserving if for each measurable set A B we have ( 1 A) = (A).
Alternately, we will say that is an invariant measure for the action y(X, B).
Well say that preserves the measure class of , or alternately, is a quasi-
invariant measure, if for each , and A B we have that (A) = 0 if and
only if ( 1 A) = 0, i.e., for each , the push-forward measure defined
by (A) = ( 1 A) for all A B is absolutely continuous with respect to .

Note, that since the action yX is measurable we obtain an action :


Aut(M(X, B)) of on the B-measurable functions by the formula (f ) =
f 1 .
If y(X, B, ) preserves the measure class of , then for each there
exists the Radon-Nikodym derivative dd : X [0, ), such that for each
measurable set A B we have
Z
d
( 1 A) = 1A d.
d

The Radon-Nikodym derivative is unique up to measure zero for .


If 1 , 2 , and A B then we have
Z
d2
((1 2 )1 A) = 1 1 A d
1 d
Z Z
d2 d2 d1
= 1A 1 ( ) d1 = 1A 1 ( ) d.
d d d

31
32 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

Hence, we have almost everywhere

d(1 2 ) d2 d1
= 1 ( ) .
d d d
d
Or, in other words, the map (, x) 7 d (x) R is a cocycle almost every-
where.

Definition 2.1.2. Let y(X, B, ), and y(Y, A, ) be measure class preserv-


ing actions of a countable group on -finite measure spaces y(X, B, ), and
y(Y, A, ). The actions are isomorphic (or conjugate) if there exists an
isomorphism of measure spaces1 such that for all and almost every x X
we have
(x) = (x).

Example 2.1.3. Any action yX of a countable group on a countable set X


can be viewed as a measure preserving action on X with the counting measure.

Example 2.1.4. Consider the torus T with the Borel -algebra, and Lebesgue
measure, if a [0, 1) then we obtain a measure preserving transformation T :
T T by T (ei ) = ei(+2a) . This then induces a measure preserving action of
Z.

Example 2.1.5 (The odometer action). Consider the space {0, 1}, with the
uniform probability measure, and consider X = {0, 1}N with the product mea-
sure. We obtain a measure preserving transformation T : X X given by
adding one. That is to say that T applied to a sequence a1 a2 a3 will be
the sequence 000 01an+1 an+2 where an is the first position in which a 0
occurs in a1 a2 a3 . Then T induces a probability preserving action of Z on
{0, 1}N.

Example 2.1.6 (Bernoulli shift). Let (X0 , B0 , 0 ) be a probability space, let


be a countable group, and consider X = X0 with the product measure. Then
we have a measure preserving action of on X by x = x 1 for each x X0 .

Example 2.1.7 (Generalized Bernoulli shift). Let (X0 , B0 , 0 ) be a probability


space, let yI be an action of a countable group on a non-empty countable
set I, and consider X = X0I with the product measure. Then just as in the case
of the Bernoulli shift we have a measure preserving action of on X given by
x = x 1 for each x X0 .

Exercise 2.1.8 (The bakers map). Let X = [0, 1] [0, 1] with Lebesgue
measure, consider the map T : X X defined by

(2x, y2 ), 0 x 12 ;
T (x, y) =
(2x 1, y+1
2 ),
1
2 < x 1.
1 A map : X Y is an isomorphism of measure spaces if is almost everywhere a

bijection such that and also 1 is measure preserving.


2.1. EXAMPLES 33

Then T and T 1 are both measure preserving and hence give a measure pre-
serving action of Z on X.
Show that the map : {0, 1}Z [0, 1] [0, 1] given by

(x) = (n0 x(n)2(n+1) , n>0 x(n)2n )

is an isomorphism of measure spaces which implements a conjugacy between


the Bernoulli shift Zy{0, 1}Z and the bakers action Zy[0, 1] [0, 1].

Example 2.1.9. Let G be a -compact, locally compact group, and let be a


countable group. Then any homomorphism from to G induces a Haar measure
preserving action of on G by left multiplication.
If G is compact, then we obtain a measure preserving action of on a
probability space.

Example 2.1.10. Let G be a -compact, locally compact group, let H < G be


a closed subgroup, and let < G be a countable subgroup. There always exists
a G-quasi-invariant measure on the homogeneous space G/H, (see for example
Section 2.6 in [Fol95]). Thus we always obtain a measure class preserving action
of on G/H.
A lattice in G is a discrete subgroup such that G/ has an invariant
probability measure. In this case we obtain a probability measure preserving
action of on G/.

Example 2.1.11. Suppose A is an abelian group, and is a group of automor-


phisms of A, then not only does preserve the Haar measure of A, but also
has a Haar measure preserving action of the dual group A given by x = x 1 .
As an example, we can consider the action of SLn Z on Zn given by matrix mul-
tiplication, this then induces a probability measure preserving action of SLn Z
on the dual group Tn .

Example 2.1.12 (Uniform ordering [MOP79, Kie75]). Let be a countable


group, and consider O() the set of all total orders of . By interpreting total
orders <t on with functions from to {0, 1} which take the value 1 at
(1 , 2 ) if and only if 1 <t 2 , we may consider O() as a subset of {0, 1},
and we consider the corresponding -algebra.
We have an action of on O() by requiring that x <t y if and only if
x <t y. We may place a probability measure on O() by requiring that for
each pairwise distinct x1 , x2 , . . . , xn we have

1
({<t O() | x1 <t x2 <t <t xn }) = .
n!
By Caratheodorys Extension Theorem, this extends to an invariant measure
for the action of . If yX is an action of on a countable set X, then we can
of course also consider the corresponding probability measure preserving action
of on O(X).
34 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

Example 2.1.13 (Furstenbergs Correspondence Principle [Fur77]). Let be


a countable amenable group and let Fn be a Flner sequence. Then each
set Fn gives rise to a probability measure n Prob() Prob(), given as
the uniform probability measure on Fn . Since the Stone-Cech compactification
is compact, the Arzela-Ascoli Theorem implies that the sequence n has a
cluster point Prob().
Since Fn is a Flner sequence we have that is an invariant probability
measure for the action of on . Moreover, if A then we have that

|A Fn | |A Fn |
lim inf (A) lim sup .
n |Fn | n |Fn |

We also remark that if A1 , A2 , . . . , Ak then it is easy to see that


kj=1 Aj= kj=1 Aj .
Since is not second countable L2 (, ) will not be separable in general.
However, since is countable, if A then we can consider the -invariant
sub--algebra generated by A , this then gives a separable Hilbert space.
Example 2.1.14. Similar to Furstenbergs Correspondence Principle, suppose
G is a locally compact group. The Baire -algebra Baire on G is the -algebra
generated by the G sets which are compact. Alexanderoff showed that there
is a Banach space isomorphism between Cb (G) and the space ba(G, Baire ) of
regular, bounded, finitely additive means m on Baire with norm given by total
valuation (see for example Theorem IV.6.2 in [DS88]).
Hence if m is a regular, finitely additive mean on the Baire -algebra of G
which is invariant under the action of G. Then m gives a state on Cb (G) =
C(G) which by the Riesz Representation Theorem gives a Radon probability
measure on G, hence, if is a countable group and we have a homomorphism
from to G then left multiplication induces a -probability measure preserving
action of on G.
Example 2.1.15 (Randomorphisms [Mon06]). Let and be two countable
groups and consider the action of on the space [, ] = {f | f (e) = e}
as described in Example 1.3.3. We consider with the Polish space structure
given by the product topology where is discrete, we then endow with the
Borel -algebra. A randomorphism from to is a -invariant probability
measure on on this -algebra. Note that homomorphism from to is
just a fixed point in , hence the Dirac measure at such a point gives a
random homomorphism.
If a random homomorphism is supported on the space of maps which are
injective then we say that is a randembedding. There is also of course the
corresponding notion of a random surjection, and a random bijection.
Notice that we can identify the space of bijections in [, ] with the space of
bijections in [, ] by the inverse map. Under this identification we then obtain
an action of on the space of bijections in [, ] given by

(f )(x) = f (xf 1 ())1 .


2.2. THE KOOPMAN REPRESENTATION 35

Example 2.1.16. Let y(X, B, ), and y(Y, A, ) be measure class preserv-


ing actions of a countable group on -finite measure spaces (X, B, ), and
(Y, A, ). Then we obtain a diagonal action y(X Y, B A, ) given by
(x, y) = (x, y).
If actions y(X, B, ), and y(Y, A, ) are measure preserving, then so is the
diagonal action.

2.2 The Koopman representation


Definition 2.2.1 ([Koo31]). Let y(X, B, ) be an action on a measure space
which preserves the measure class . The Koopman representation of
associated to this action is the unitary representation : U(L2 (X, ))
given by
d 1/2
(()f )(x) = f ( 1 x)( ) (x).
d
Note that for all 1 , 2 , and f L2 (X, ) we have
d(1 2 ) 1/2
(1 2 )f = 1 2 (f )( )
d
d2 1/2 d1 1/2
= 1 (2 (f )( ) )( ) = (1 )((2 )f ).
d d
Also, for all , and f L2 (X, ) we have
Z
d
k()f k22 = | (f )|2 d
d
Z
= | (f )|2 d = kf k22 .

Hence, is indeed a unitary representation.


If (X, B, ) is a finite measure space, and y(X, B, ) is measure preserving,
then 1 L2 (X, ) and ()(1) = 1 is an invariant vector. For this reason, in
this setting the Koopman representation usually denotes the restriction of the
above representation to the orthogonal complement
Z
L20 (X, ) = {f L2 (X, ) | f d = 0}.

Exercise 2.2.2. Let y(X, B, ) be an action on a measure space which pre-


serves the measure class . Show that the Koopman representation is isomor-
phic to its conjugate representation .
Exercise 2.2.3. Let y(X, B, ), and y(Y, A, ) be measure preserving ac-
tions of a countable group on probability spaces (X, B, ), and (Y, A, ). Show
that the Koopman representation XY for the product action decomposes as
XY = X Y (X Y ) where X and Y are the Koopman representa-
tions for y(X, B, ), and y(Y, A, ).
36 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

2.2.1 Ergodicity
Definition 2.2.4. Let y(X, B, ) be a measure class preserving action of a
countable group on a -finite measure space (X, B, ). The action y(X, B, )
is ergodic, if for any E B which is -invariant, we have that either (E) = 0
or (X \ E) = 0.

Lemma 2.2.5. Let y(X, B, ) be a measure preserving action of a countable


group on a probability space (X, B, ). The action y(X, B, ) is ergodic if
and only if the Koopman representation is ergodic.

Proof. If E B is -invariant then E (E) L20 (X, ) is also -invariant,


which is non-zero if (E) 6= 0, 1. On the other hand, if L20 (X, ) is a non-
zero -invariant function then Et = {x X | |(x)| < t} is -invariant for all
t > 0, and we must have (Et ) 6= 0, 1 for some t > 0.

Theorem 1.6.5 can be adapted to the setting of actions on measure spaces


as follows.

Theorem 2.2.6. Let y(X, B, ) be a measure preserving action of a count-


able group on an infinite, -finite measure space (X, B, ). The following
conditions are equivalent.
(1). There exists a state (L (X, B, )) such that for all , and
f L (X, B, ) we have (()(f )) = (f ).
(2). For every > 0, and F finite, there exists Prob(X, B), such
that is absolutely continuous with respect to and

d d
k k1 < .
d d

(3). For every > 0, and F finite, there exists a measurable set A X
such that (A) < , and

(AA) < (A).

(4). The Koopman representation : U(L2 (X, B, )) has almost in-


variant vectors.
(5). The Koopman representation : U(L2 (X, B, )) is amenable.

Exercise 2.2.7. Adapt the proof of Theorem 1.6.5 to prove Theorem 2.2.6.

If acts by measure preserving transformations on a probability space, then


the conditions above are trivially satisfied, however, we still have the following,
non-trivial, adaptation.
2.2. THE KOOPMAN REPRESENTATION 37

Theorem 2.2.8. Let y(X, B, ) be a measure preserving action of a countable


group on a probability space (X, B, ). The following conditions are equivalent.
R
(1). There exists a state (L (X, B, )) , different than f 7 f d,
such that for all , and f L (X, B, ) we have (()(f )) = (f ).
(2). There exits A X measurable such that (A) < 1, and for every
> 0, and F finite, there exists Prob(X, B), such that is absolutely
continuous with respect to and

d d
(A) > 1 , and k k1 < .
d d

(3). For every > 0, and F finite, there exists a measurable set A X
such that (A) 1/2, and

(AA) < (A).

(4). The Koopman representation : U(L20 (X, B, )) has almost in-


variant vectors.
(5). The Koopman representation : U(L20 (X, B, )) is amenable.

If y(X, B, ) is an ergodic measure preserving action of a countable group


on a probability space (X, B, ) such that none of the conditions of the previous
theorem hold, then we say that the action y(X, B, ) has spectral gap.

2.2.2 Weak mixing


Definition 2.2.9. Let y(X, B, ) be a measure preserving action of a count-
able group on a probability space (X, B, ). A sequence En B of measurable
subsets is an asymptotically invariant sequence if (En En ) 0, for
all . Such a sequence is said to be non-trivial if lim inf n (En )(1
(En )) > 0. The action y(X, B, ) is strongly ergodic if there does not
exist a non-trivial asymptotically invariant sequence.

Note that if y(X, B, ) has a non-trivial invariant set E B, then the con-
stant sequence En = E is a non-trivial asymptotically invariant sequence, hence
strongly ergodic implies ergodic. Also, a non-trivial asymptotically invariant
sequence will show that condition (3) in Theorem 2.2.8 is satisfied, hence an
ergodic action with spectral gap must be strongly ergodic.

Definition 2.2.10. Let y(X, B, ) be a measure preserving action of a count-


able group on a probability space (X, B, ). The action y(X, B, ) is weak
mixing, if || = , and for all E B finite, we have

lim inf A,BE |(A B) (A)(B)| = 0.



38 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

Proposition 2.2.11. Let y(X, B, ) be a measure preserving action of a


countable group on a probability space (X, B, ). The following conditions
are equivalent:
(1). The action y(X, B, ) is weak mixing.
(2). The Koopman representation : U(L20 (X, B, )) is weak mixing.
(3). The diagonal action y(X X, B B, ) is ergodic.
(4). For any ergodic, measure preserving action y(Y, A, ) on a probability
space, the diagonal action y(X Y, B A, ) is ergodic.
(5). The diagonal action y(X X, B B, ) is weak mixing.
(6). For any weak mixing, measure preserving action y(Y, A, ) on a prob-
ability space, the diagonal action y(X Y, B A, ) is weak mixing.

Proof. For (1) (2), if E B is a finite set then we can consider the finite
set F = {E (E) | E E} L20 (X, B, ). For each we therefore have

,F |h(), i| = A,BE |h (B ) (B), A (A)i|

= A,BE |hB (B), A (A)i|

= A,BE |(A B) (A)(B)|.

If the Koopman representation is weak mixing then from this we see imme-
diately that the action is weak mixing. A similar calculation shows that if the
action is weak mixing and F L20 (X, B, ) is a finite set of simple functions,
then
lim inf ,F |h(), i| = 0.

Since simple functions are dense in L20 (X, B, ) this shows that the Koopman
representation is weak mixing.
The remaining equivalences then easily follow from Exercise 2.2.3, together
with the corresponding properties of weak mixing for unitary representations in
Sections 1.5 and 1.7.

Corollary 1.7.7 then shows that weak mixing is preserved under taking finite
index.

Corollary 2.2.12. Let y(X, B, ) be a weak mixing, measure preserving ac-


tion of a countable group on a probability space (X, B, ). If < is a finite
index subgroup, then the restriction of the action of to is also weak mixing.

Corollary 1.7.16 gives the following result for weak mixing actions of amenable
groups.
2.2. THE KOOPMAN REPRESENTATION 39

Corollary 2.2.13. Let y(X, B, ) be a measure preserving action of a count-


able amenable group on a probability space (X, B, ). Let Fn , be a Flner
sequence for . The action y(X, B, ) is weak mixing if and only if for all
A, B B, we have

1
Fn |(A B) (A)(B)| 0.
|Fn |

Definition 2.2.14. Let y(X, B, ) be a measure preserving action of a count-


able group on a probability space (X, B, ). The action y(X, B, ) is
(strong) mixing, if || = , and for any A, B X measurable we have

lim |(A B) (A)(B)| = 0.


The same proof as in Proposition 2.2.11 yields the following proposition for
mixing actions.

Proposition 2.2.15. Let y(X, B, ) be a measure preserving action of a


countable group on a probability space (X, B, ). The action y(X, B, )
is mixing if and only if the Koopman representation is mixing.

2.2.3 Compact actions


Definition 2.2.16. Let (X, B, ) be a probability space. We denote by Aut(X, B, )
the group of automorphisms of (X, B, ), where we identify two automorphisms
if they agree almost everywhere. The weak topology on Aut(X, B, ) is the
smallest topology such that the maps T 7 (T (A)B) are continuous for all
A, B B.

Exercise 2.2.17. Show that the weak topology endows Aut(X, B, ) with a
topological group structure.

Exercise 2.2.18. The Koopman representation for Aut(X, B, ) is defined as


before, i.e., : Aut(X, B, ) U(L20 (X, B, )) is defined by (T )(f ) = f
T 1 . Show that the image of is closed and that is homeomorphism from
Aut(X, B, ) with the the weak topology onto (Aut(X, B, )) U(L20 (X, B, ))
with the strong operator topology.

Definition 2.2.19. Let y(X, B, ) be a measure preserving action of a count-


able group on a probability space (X, B, ). The action y(X, B, ) is com-
pact if the image of in Aut(X, B, ) is precompact in the weak topology.

An immediate consequence of Exercise 2.2.18 is the following.

Proposition 2.2.20. Let y(X, B, ) be a measure preserving action of a


countable group on a probability space (X, B, ). Then y(X, B, ) is compact
if and only if the Koopman representation : U(L20 (X, B, )) is compact.
40 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

Definition 2.2.21. Let y(X, B, ) be a measure preserving action of a count-


able group on a probability space (X, B, ). A function f L2 (X, B, ) is
almost periodic if the orbit { (f ) | } is pre-compact in the k k2 -
topology.

Proposition 2.2.22. Let y(X, B, ) be a measure preserving action of a dis-


crete group on a standard probability space (X, B, ). Let A B be the
-algebra generated by all almost periodic functions in L2 (X, B, ). Then A
is -invariant, and f L2 (X, B, ) is almost periodic if and only if f is A
measurable.

Proof. For each f L2 (X, B, ) denote by Sf the -orbit of f , and consider


the set AP (L2 (X, B, )) of all almost periodic functions. Note that it is obvious
that AP (L2 (X, B, )) is -invariant, contains the scalars, and is closed under
scalar multiplication.
Since addition is continuous with respect to k k2 , if f, g AP (L2 (X, B, ))
we have that Sf + Sg is compact being the continuous image of the compact
set Sf Sg . We therefore have that Sf +g Sf + Sg is precompact, hence
AP (L2 (X, B, )) is closed under addition. Similarly, if f, g AP (L2 (X, B, ))
then we have that

|f |, f , max{|f |, |g|}, min{|f |, |g|} AP (L2 (X, B, )).

Also, if we also have that g L (X, B, ) then f g AP (L (X, B, )).


If fn AP (L2 (X, B, )) and f L2 (X, B, ) such that kfn f k2 0, then
fix > 0, and take n N such that kfn f k2 < /2. Since fn is almost periodic
we have that Sfn is totally bounded, hence there exists a finite set L2 (X, B, )
such that inf hF k (fn ) hk2 < /2 for all . By the triangle inequality
we then have inf hF k (f ) hk2 < . This shows that Sf is totally bounded
and hence f AP (L2 (X, B, )).
The operations above then generate all A-measurable functions and so we
obtain the result.

Definition 2.2.23. Let y(X, B, ) be a measure preserving action of a dis-


crete group on a standard probability space (X, B, ). Suppose A B is a
-invariant -algebra, then we say that the action y(X, A, ) is a factor of
the action y(X, B, ).

Note that the Koopman representation of a factor y(X, A, ) is a sub-


representation of the Koopman representation of y(X, B, ). Hence, Proposi-
tion 2.2.11 together with Proposition 2.2.22 gives the following.

Proposition 2.2.24. Let y(X, B, ) be a measure preserving action of a dis-


crete group on a standard probability space (X, B, ). Then there exists a
unique maximal factor y(X, A, ) which is compact. Moreover, y(X, B, )
is weak mixing if and only if A is trivial, i.e., when A consists only of null or
co-null sets.
2.3. A REMARK ABOUT MEASURE SPACES 41

2.3 A remark about measure spaces


So far we have been considering actions of a countable group on a general
probability space (X, B, ). For most aspects of ergodic theory this is the proper
setting. However, occasionally this level of generality can be problematic. For
example, consider the action of Z by rotation on the circle T. We can consider
T equipped with either the Borel -algebra, or the Lebesgue -algebra. If we
are not concerned with sets of measure 0, then both of these systems contain
the same information, however the identity map from the Borel -algebra to the
Lebesgue -algebra is not measurable, and these two systems are not isomorphic
under our notion of isomorphism.
For another example, consider the case when y(X, B, ) and y(Y, A, )
are measure preserving actions. Then the diagonal action y(XY, BA, )
has the factor y(X Y, B {Y, }, ), which we would like to identify with
the action y(X, B, ). However, a problem arises because the projection map
from X Y to X is not almost everywhere 1-1 and hence is not an isomorphism
of actions.
One way to overcome this problem is to restrict ourselves to only consider
actions on nice measure spaces. Specifically, one considers the class of stan-
dard probability spaces. A standard probability space (X, B, ) is a proba-
bility space such that the underlying -algebra space (X, B) is isomorphic (as
-algebra spaces) to a Polish2 topological space with its Borel -algebra. In
this setting we do not allow actions on T with the Lebesgue measure, or actions
on the space (X Y, B {Y, }, ) since these are not standard probability
spaces, and so the problems above do not arise.
An alternate approach, which we will take here, is to continue to allow gen-
eral probability spaces, but instead generalize our notion of equivalence so that
the spaces above are equivalent under this more general notion. Thus, from now
on we will say that two probability spaces (X, B, ) and (Y, A, ) are isomor-
phic if there is an integral preserving unital -isomorphism from L (Y, A, ) to
L (X, B, ). We will also say that two measure preserving actions y(X, B, )
and y(Y, A, ) are isomorphic (or conjugate) if there exists a -isomorphism
from L (Y, A, ) to L (X, B, ) such that = for all . Note
that if : X Y is a bijection such that and 1 is measure preserving such
that = for all , then : L (Y, A, ) L (X, B, ) given by
(f ) = f 1 implements an isomorphism between the two actions.
We also have a process (essentially the GNS-construction) which takes us
from a general action on a separable measure space to an isomorphic action on
a standard Borel space.

Proposition 2.3.1. Let A be a unital -algebra with a state3 such that for
all a A there exists K > 0 such that (a abb ) K(bb ) for all b A. Let
2 A topological space is Polish if it is separable and has a complete metric which induces

the topology
3 A state on a unital -algebra is a linear map : A C such that (1) = 1, and (a a) 0

for all a A.
42 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

be a countable group and suppose : Aut(A) is an action of on A by


unital -homomorphisms such that ( (a)) = (a) for all a A, and .
Then there exists a compact Hausdorff space X, a continuous action yX,
a -invariant Radon probability measure on X, and a unital R-homomorphism
: A L (X, ) such that = for all , and (a) d = (a),
for all a A.
Moreover, if A is countable generated as an algebra, then X is separable (and
hence a Polish space).
Proof. We endow A with the inner-product ha, bi = (b a), by quotienting out
by the kernel of this inner-product and then taking the completion, we obtain
a Hilbert space L2 (A, ). Given a A we denote by a the equivalence class of
a in L2 (A, ). If a, b A, then we may consider a acting on L2 (A, ) by left
multiplication. The formula
b 2 = (a abb ) Kkbk2
kabk

shows that left multiplication is well defined and bounded, we therefore obtain
a -homomorphism 0 : A B(L2 (A, )).
We denote by C (A) the abelian C -algebra generated by 0 (A) in B(L2 (A, )),
and we denote by X the Gelfand spectrum of C (A). By Gelfands Theorem
we have that X is a compact Hausdorff space and we obtain a -isomorphism
from C (A) to C(X). We therefore obtain a -homomorphism : A C(X)
by applying the Gelfand transform to the image 0 (A).
On C(X) = C (A) we may consider the state given by (x) = hx1, 1i . By
the Riesz Representation Theorem
R the state corresponds to a Radon measure
on X such that (x) = x d for all x C(X). We therefore have that for
all a A Z
(a) d = ((a))

= h0 (a)1, 1i = ha, 1i = (a).


Since : Aut(A) preserves the state , we have that for all , and
a, b A

k\ 2 \ 2
(a)bk = ( (a a)bb ) = (a a 1 (bb )) = ka 1 (b)k .

Hence, if we define on 0 (A) by (0 (a)) = 0 ( (a)) then this is well


defined and preserves the operator norm, hence extends to a map (which we
still denote by ) from to Aut(C (A)). The Gelfand transform then gives a
continuous action of on X, such that (x) = x 1 for all x C(X).
An easy calculation then shows that = for all , and we
have that for all , and x C(X)
Z Z Z
x d = x 1 d = ( (x)) = (x) = x d.

Hence is -invariant.
Finally, if A is countably generated as an algebra then C (A) is a separable
C -algebra and hence the Gelfand spectrum X is separable.
2.4. GAUSSIAN ACTIONS 43

Corollary 2.3.2. Suppose y(X, B, ) is a measure preserving action of a


countable group on a probability space (X, B, ) such that L2 (X, B, ) is sep-
arable. Then the action y(X, B, ) is isomorphic to a measure preserving
action of on a standard probability space (Y, A, ).
Proof. Since L2 (X, B, ) is separable there exists a countably generated -subalgebra
A L (X, B, ) L2 (X, B, ) which is dense in L2 (X, B, ). Since is count-
able we may also assume that A is -invariant. By Proposition 2.3.1 there exists
a measure preserving action of on a standard probability space (Y, A, ) and
homomorphism : A L (Y, A, ) which is -equivariant and preserves the
integrals.
Since preserves the integrals we may extend it to a unitary U : L2 (X, B, )
2
L (Y, A, ) which also preserves the integrals, and is again -equivariant. The
map f 7 U f U then gives a -isomorphism between L (Y, A, ) and L (X, B, )
which preserves the integrals and is again -equivariant.

2.4 Gaussian actions


Let : O(H) be an orthogonal representation of a countable group . The
aim of this section, which is taken from [PS], is to describe the construction of
a measure-preserving action of on a non-atomic standard probability space
(X, ) such that H is realized as a subspace of L2R (X, ) and is contained in
the Koopman representation yL20 (X, ). The action y(X, ) is referred to
as the Gaussian action associated to .
Given a Hilbert space H, the n-symmetric tensor Hn is the subspace of
n
H fixed by the action of the symmetric group Sn by permuting the indices.
. , n H, we define their symmetric tensor product 1 n Hn
For 1 , . .P
1
to be n! Sn (1) (n) . Denote

M
S(H) = (R Hn ) R C,
n=1

with renormalized inner product such that kk2S(H) = n!kk2 , for Hn .


For H let x be the symmetric creation operator,

x () = , x (1 k ) = 1 k ,

and its adjoint,


k
X
x () = 0, x (1 k ) = hi , i1 bi k .
i=1

Let
1
(x + x ),
s() =
2
and note that it is an unbounded, self-adjoint operator on S(H).
44 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

The moment generating function M (t) for the Gaussian distribution is de-
fined to be
Z
1
M (t) = exp(tx) exp(x2 /2)dx = exp(t2 /2).
2

It is easy to check that if kk = 1 then

(2k)!
hs()n , i = M (n) (0) = ,
2k k!
if n = 2k and 0 if n is odd. Hence, s() may be regarded as a Gaussian random
variable. Note that if , H then s() and s() commute, moreover, if ,
then
hs()m s()n , i = hs()m , ihs()n , i,
for all m, n N; thus, s() and s() are independent random variables.
From now on we will use the convention 1 2 k to denote the symmetric
tensor 1 2 k . Let be a basis for H and

S() = {} {s(1 )s(2 ) s(k ) | 1 , 2 , . . . , k }.

Lemma 2.4.1. The set S() is a (non-orthonormal) basis of S(H).

Proof. We will show that 1 k sp(S()), for all 1 , . . . , k H. We have


sp(S()). Also, since s() = , H sp(S()). Now as s(1 ) s(k ) =
P (1 , . . . , k ) is a polynomial in 1 , . . . , k of degree k with top term 1 k ,
the result follows by induction on k.

Let u(1 , . . . , k ) = exp(is(1 ) s(k )) and u(1 , . . . , k )t = exp(its(1 ) s(k )).


Denote by A the von Neumann algebra generated by all such u(1 , . . . , k ), which
is the same as the von Neumann algebra generated by the spectral projections
of the unbounded operators s(1 ) s(k ).

Theorem 2.4.2. We have that L2 (A, ) = S(H), and A is a maximal abelian


-subalgebra of B(S(H)) with faithful state = h , i.

Proof. By Lemma 2.4.1, A 7 A is an embedding of A into S(H). By Stones


Theorem
u(1 , . . . , k )t 1
lim = s(1 ) s(k );
t0 it
hence, A is dense in S(H). This implies that A is maximal abelian in B(S(H)).

There is a natural strong operator topology continuous embedding O(H)


U(S(H)) given by

M
S
T 7 T = 1 T n .
n=1
2.5. ERGODIC THEOREMS 45

It follows that there is an embedding O(H) Aut(A, ), T 7 T , which can


be identified on the unitaries u(1 , . . . , k ) by

T (u(1 , . . . , k )) = Ad(T S )(u(1 , . . . , k )) = u(T (1 ), . . . , T (k )).

Thus for an orthogonal representation : O(H), there is a natural ac-


tion : Aut(A, ) given by (u(1 , . . . , k )) = u( (1 ), . . . , (k )) =
Ad(S )(u(1 , . . . , k )). Applying Proposition 2.3.1 we then obtain a measure
preserving probability space (X, B, ) and an action of on this space so that
we can identify A with L (X, ) in a state preserving -equivariant manor.
The action of on (X, B, ) is the Gaussian action associated to .
We have an explicit description of the Koopman representation of y(X, B, )
given by S () = (())S 1. Hence, we have that ergodic properties which
remain stable with respect to tensor products transfer from to .
Proposition 2.4.3. Let : O(H) be an orthogonal representation of a
countable group . Then the Gaussian action y(X, B, ) is ergodic, if and
only if it is weak mixing, if and only if is weak mixing.
Proof. If is not weak mixing, then there exists H2 such that for all
, 2 ()() = . Viewing as a Hilbert-Schmidt operator on H, let
|| = ( )1/2 . Since the map 7 is the same as taking the adjoint
of the corresponding Hilbert-Schmidt operator, we have that || H2 and
2 ()(||) = ||. Since 2 embeds into the Koopman representation of the
Gaussian action, it follows from Lemma 2.2.5 that y(X, B, ) is not ergodic.
Conversely, if y(X, B, ) is not ergodic then neitherLis thenKoopman rep-
S S
resentation
L . But is a sub-representation of n=1 = (1
n
n=1 ). Hence if S has non-trivial invariant vectors then is not weak
mixing by Proposition 1.5.6.

2.5 Ergodic theorems


Using the Koopman representation, von Neumanns Ergodic Theorem for Hilbert
spaces can be rephrased for actions.
Theorem 2.5.1 (Von Neumanns Ergodic Theorem [vN32]). Let y(X, ) be
a measure preserving action of a countable amenable group on a probability
space (X, B, ). Let EI B(L2 (X, B, )) be the projection onto the subspace
of -invariant functions. Let Fn be a Flner sequence, then for each f
L2 (X, B, ) we have that
1
k 1 (f ) EI (f )k2 0.
|Fn | Fn
To state Birkoffs Ergodic Theorem we first need to discuss conditional ex-
pectations. Note that if (X, B, ) is a probability space, then we may consider
a function f L (X, B, ) as a bounded operator on L2 (X, B, ) by multipli-
cation, i.e., g 7 f g. To avoid confusion in the next lemmas, when we consider
46 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

a function f L (X, B, ) as an element in L2 (X, B, ) we will continue to


use the notation f , however, when we consider f as a bounded operator we will
denote this operator by Mf .

Lemma 2.5.2. Let (X, B, ) be a probability space, then L (X, B, ) B(L2 (X, B, ))
is a maximal abelian self-adjoint subalgebra.

Proof. Suppose T B(L2 (X, B, )) commutes with L (X, B, ), we want to


show that there exists some f L2 (X, B, ) such that T (g) = f g for all g
L2 (X, B, ).
Let f = T (1) L2 (X, B, ), if g L (X, B, ) L2 (X, B, ) then we have

T (g) = T Mg (1) = Mg T (1) = Mg (f ) = f g.

Hence, if we consider the functions ft = 1[0,t) (|f |)f L (X, B, ), then we


have
kft k = sup kft gk2
gL2 (X,B,),kgk2 =1

= sup kT (1[0,t)(|f |)g)k2


gL2 (X,B,),kgk2 =1

sup kT kk1[0,t)(|f |)gk2 kT k.


gL2 (X,B,),kgk2 =1

Therefore, kf k = limt kft k kT k, and so f L (X, B, ). Moreover,


since L (X, B, ) L2 (X, B, ) is dense, we have that T (g) = f g for all g
L2 (X, B, ).

If A B is -subalgebra we will denote by EA B(L2 (X, B, )) the orthog-


onal projection onto L2 (X, A, ). Note that if g L (X, A, ) then we can
think of Mg as a bounded operator on L2 (X, A, ) or on L2 (X, B, ), and since
the product of two A-measurable functions is again A-measurable it follows that

Mg EA = EA Mg .

Then for each f L (X, B, ) and g L (X, A, ) we have that

(EA Mf EA )Mg = EA Mf Mg EA = Mg (EA Mf EA ).

By Lemma 2.5.2 there exists a function f0 L (X, B, ) such that EA Mf EA



=
Mf0 . It also follows from the proof of Lemma 2.5.2 that actually f0 = EA (f )
and so we see that EA (L (X, B, )) L (X, A, ), and kEA (f )k kf k
for all f L (X, B, ).
The map EA : L (X, B, ) L (X, A, ) is called the conditional ex-
pectation with respect to A.

Lemma 2.5.3. Let (X, B, ) be a probability space and let A B be a -


subalgebra, then the conditional expectation EA : L (X, B, ) L (X, A, )
satisfies the following properties:
2.5. ERGODIC THEOREMS 47
R R
(1). EA preserves the integral, i.e., EA (f ) d = f d, for all f
L (X, B, ).
(2). EA preserves positivity and for each f L (X, B, ) we have |EA (f )|
EA (|f |).
(3). EA is L (X, A, )-bimodular, i.e., if f L (X, B, ) and g L (X, A, )
then EA (f g) = EA (f )g.
(4). EA extends continuously to a contraction on Lp (X, B, ) for all 1
p .

Proof. (1) follows because if f L (X, B, ) then we have


Z Z
EA (f ) d = hEA (f ), 1iL2 (X,B,) = hf, EA (1)iL2 (X,B,) = f d.

We have (2) because for all f L (X, B, ) and g L2 (X, A, ) we have

hEA M|f | EA g, gi = hM|f | EA g, EA gi 0,

and

h((M|EA (f )| )2 (EA M|f | EA )2 )g, gi = h(EA M|f | (1 EA )M|f | EA )g, gi

= h(1 EA )|f |g, |f |gi 0.


Hence EA (|f |) 0, and |EA (f )|2 EA (|f |)2 .
(3) follows because as noted above the product of two A-measurable func-
tions is again A-measurable.
Finally, we have (4) because if 1 p < , then for every f Lp (X, B, )
and g L (X, A, ) we have
Z Z Z
EA (f )g d = EA (f g) d = f g d,

1 1
hence if 1 < q such that p + q = 1 then
Z
kEA (f )kp = sup | EA (f )g d|
gL (X,A,),kgkq 1

Z
sup | f g d| = kf kp .
gL (X,B,),kgkq 1

Note that because EA extends continuously to a contraction on Lp (X, B, )


for any 1 p we will use the same notation EA for any of these extensions.
48 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

Exercise 2.5.4. If y(X, B, ) is a measure preserving action, and A is -


invariant, then show that

(EA (f )) = EA ( (f )),

for all , and f Lp (X, B, ), 1 p .

Theorem 2.5.5 (Birkoffs Ergodic Theorem [Bir31]). Let ZyT (X, B, ) be a


measure preserving action of Z on a probability space (X, B, ). Let I B
denote the -subalgebra of -invariant sets. Then for each f L1 (X, B, ) we
have k k1 , and almost everywhere convergence

1 N 1
n (f ) EI (f ).
N n=0
We will model the proof of this theorem by the proof of von Neumanns
Ergodic Theorem (this is the approach taken in [PY98], that is to say we will
prove the theorem for functions of the form g n (g) which are easier to handle,
and we will also show that we can approximate an arbitrary function f with
EI (f ) = 0 by some linear combination of functions of the above type.

Lemma 2.5.6. Let ZyT (X, B, ) be a measure preserving action of Z on a


probability space (X, B, ). Let I B denote the -subalgebra of -invariant
sets. Then for each > 0, and f L1 (X, B, ) with EI (f ) = 0, there exists
g L (X, B, ) such that kf (1 (g) g)k1 < .

Proof. Let B be the k k1 closure


R of the subspace {1 (g) g | g L (X, B, )}.

If h L (X, B, ) such that (1 (g) g)h d = 0 for all g L (X, B, ) then
we have Z Z
1 (g)(h 1 (h)) d = (1 (g) g)h d = 0.

Thus 1 (h) = h and hence n (h) = h for all n Z. We therefore have that
h L (X, I, ) and hence
Z Z Z
f h d = EI (f h) d = EI (f )h d = 0.

Since L (X, B, ) = (L1 (X, B, )) by the Hahn-Banach Theorem this shows


that f B.

For any f L1 (X, B, ) and > 0, we will denote

1 N 1
E (f ) = {x X | lim sup | n (f )(x)| }.
N N n=0

Lemma 2.5.7.R Using the notation above, for each f L1 (X, B, ) we have
(E2 (f )) 1 |f | d.
2.5. ERGODIC THEOREMS 49

Proof. Suppose f 0, and for each M 1 consider


1 N 1
EM (f ) = {x X | sup n (f )(x) }.
1N M N n=0
We will first show that if P > M then we have the inequality
P 1 P M1
n=0 n (f )(x) n=0 n (1EM (f ) )(x).
Indeed, if x X, then we may consider the first 0 n0 P M such that
x T n0 (EM (f )), we then have that there exists some n0 +1 N0 n0 +M 1,
such that N n=n0 n (f )(x) (N0 n0 ). Hence, we have that
0

P 1 P 1
n=0 n (f )(x) n0 (f ) + n=N (f )(x)
0 n

= (nn=0
0 1 P 1
n (1EM (f ) )(x)) + n0 (f )(x) + n=N (f )(x)
0 n

(N 0 1 P 1
n=0 n (1EM (f ) )(x)) + n=N0 n (f )(x).

We may now consider the first N0 n1 P M such that x T n1 (EM (f ))


and the corresponding n1 + 1 N1 n1 + M 1, the same argument then gives
P 1
n=0 n (f )(x) (N 1 1 P 1
n=0 n (1EM (f ) )(x)) + n=N1 n (f ).

Continuing in this manor we obtain the inequality


P 1 P M1
n=0 n (f )(x) n=0 n (1EM (f ) )(x),
for all x X, and P > M .
Integrating this inequality we obtain
Z Z
P 1
P f d = n=0 n (f )(x) d
Z
P M1
n=0 n (1EM (f ) )(x) d = (P M )(EM (f )).

Dividing by P and taking the limit as P we have that


Z
f d (EM (f )).

for all M 1.
Since EM (f ) are increasing as M increases, and since E (f ) M
M=1 E (f ),
we have that Z
f d (E (f )).

For general f L1 (X, B, ) we may consider f = f+ f where f+ , f 0,


and f+ f = 0. We then have
Z
|f | d ((E (f+ )) + (E (f ))) (E2 (f )).
50 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

Proof of Birkoff s Ergodic Theorem. Let f L1 (X, B, ) be given, and note


that by subtracting EI (f ) we may assume that EI (f ) = 0.
Let , > 0, and consider E (f ) as defined above. By Lemma 2.5.6 there
exists g L (X, B, ) such that kf (1 (g) g)k1 < /4 < . For each x X
we have that
1 N 1 1
lim sup |n=0 n (1 (g) g)(x)| lim sup |(N (g) g)(x)|
N N N N

2kgk
lim sup = 0.
N N
Hence E/2 (1 (g) g) = .
By Lemma 2.5.7 we therefore have that

(E (f )) (E/2 (1 (g) g)) + (E/2 (f (1 (g) g)))

4
kf (1 (g) g)k1 < ,

and also
1 N 1
lim sup k n (f )k1
N N n=1
1 N 1
lim sup k (1 (g) g)k1 + kf (1 (g) g)k1 < .
N N n=1
1 N 1
Since and are arbitrary this shows that N n=1 n (f ) 0 almost every-
where and in k k1 .

2.6 Recurrence theorems


Theorem 2.6.1 (Poincares Recurrence Theorem [Poi90]). Let ZyT (X, B, )
be a measure preserving action on a probability space (X, B, ). If A X is
measurable, such that (A) > 0 then for almost every point x A, the orbit Zx
returns to A infinitely often.

Proof. Let F A be the set of points x such that T n (x) 6 A, for all n >
0. Then F = A \ (nN T n (A)) is measurable and if m > n then we have
T m (F ) T n (F ) = . Indeed, if T m (x) F A then by the definition of F
we have that T n (x) = T mn (T m (x)) 6 A. We therefore have that for any
N N
1 1 1
(F ) = N m
n=1 (T (F )) = (N m
n=1 T (F )) .
N N N
Hence (F ) = 0 and the theorem follows easily.

If ZyT (X, B, ) is a measure preserving action on a probability space (X, B, ),


and A X is measurable, such that (A) > 0, then for each x A we let
nA (x) N {} be the smallest natural number such that T nA (x) (x) A. It
2.6. RECURRENCE THEOREMS 51

is easy to see that nA is measurable and by Poincares Recurrence Theorem, nA


is almost everywhere finite.
For n N we let An A be the set of points x A such that T n (x) A
but T k (x) 6 A for all 0 < k < n, so that nA = nN n1An . If 0 k1 < n1 < ,
and 0 k2 < n2 < , then we have that T k1 (An1 ) T k2 (An2 ) = unless
n1 = n2 and k1 = k2 . Indeed, if k1 < k2 and x T k1 (An1 ) T k2 (An2 ),
then T k2 (x) An2 and T k2 k1 (T k2 (x)) = T k1 (x) An1 A, hence k2
k2 k1 n2 > k2 which cannot happen. On the other hand, if k1 = k2 , and
n1 6= n2 then T k1 (An1 ) T k2 (An2 ) = T k1 (An1 An2 ) = T k1 () = .
There is a nice picture to go along with this disjoint decomposition known as
the Kakutani tower, however I am not TEX savvy enough to draw it, so instead
I will refer to page 45 in [Pet83].
Theorem 2.6.2 ([Kac47]). Let ZyT (X, B, ) be a measure preserving action
on a probability space (X, B, ). If A X is measurable, such that (A) > 0,
then Z
nA d 1,
A
with equality when the action is ergodic.
Proof. Using the notation above, we have that
Z
nA d = nN n(An )
A

= nN k = 0n1 (T k (An )) = (nN k=0


n1 k
T (An )) 1.
n1 k
The set nN k=0 T (An ) is a Z-invariant set which contains A, hence if
the action is ergodic then we have
Z
n1 k
nA d = (nN k=0 T (An )) = 1.
A

Theorem 2.6.3 ([Roh48]). Let ZyT (X, B, ) be an ergodic measure preserving


action on a probability space (X, B, ). Then for any N N, and > 0, there
exists A X measurable, such that A, T (A), . . . , T N 1 (A) are pairwise disjoint,
and (N 1 j N
j=0 T (A)) > 1 , and (AT (A)) < .

Proof. This theorem is an easy exercise if (X, B, ) is not a diffuse subspace,


hence we will assume that it is diffuse. Given B X with (B) > 0 then we can
consider the Kakutani tower {T k (Bn )}nN,0k<n associated to B. If we then
consider A = nN 0k<n/N T kN (Bn ) then we have that A, T (A), . . . , T N 1(A)
are pairwise disjoint. Since the action is ergodic, we also have that

(N 1 j k
j=0 T (A)) 1 (nN max{nN,0}k<n T (Bn ))

1 N nN (Bn ) = 1 N (B),
52 CHAPTER 2. GROUP ACTIONS ON MEASURE SPACES

and
(AT N (A)) 2nN (Bn ) 2(B).
Hence, if we choose B X measurable, such that 0 < (B) < /N (which is
possible since (X, B, ) is diffuse), then we obtain a set A which satisfies the
desired properties.
Theorem 2.6.4 ([Khi35]). Let y(X, ) be a measure preserving action of
a countable amenable group on a probability space (X, B, ). Suppose f
L2 (X, ), f > 0. Then for each > 0, there exists F finite such that for
all 0 we have that
Z
0 F { | (f )f d kf k21 } 6= .

Proof. Let Fn be a Flner sequence for . By von Neumanns Ergodic


Theorem there exists n N, such that
1
k 1 (f ) EI (f )k2 < /(1 + kf k2 ).
|Fn | Fn

Therefore, we have that


1 1
|h 1 (f ) EI (f ), f i| = |h 1 (f ) EI (f ), 1 (f )i|
|Fn | 0 Fn |Fn | Fn 0

1
k 1 (f ) EI (f )k2 kf k2 < .
|Fn | Fn
Hence,
1
h 1 (f ), f i hEI (f ), f i
|Fn | 0 Fn
= kEI (f )k2 |hEI (f ), 1i|2 = kf k21 .

Corollary 2.6.5. Let ZyT (X, B, ) be a measure preserving action on a prob-


ability space (X, B, ), then for each A X measurable, (A) > 0, and > 0,
there exists K N such that for all j N we have

{j, j + 1, . . . , j + K} {k N | (T k (E) E) (E)2 } =


6 .

Proof. This follows by applying Theorem 2.6.4 to the Flner sequence Fn =


{1, 2, . . . , n} for Z, and the function f = E .
Chapter 3

Uniform multiple
recurrence

Furstenbergs Recurrence Theorem [Fur77] states that if ZyT (X, B, ) is a mea-


sure preserving action on a probability space (X, B, ) and A X is a measur-
able subset such that (A) > 0, then for each k 1 there exists n 1 such
that
(A T n (A) T kn (A)) > 0.
This was then generalized by Furstenberg and Katznelson [FK78] where it is
shown that if k N, and ZyT (X, B, ) is a measure preserving action on a
probability space (X, B, ), then for each measurable set A X, such that
(A) > 0, we actually have uniform multiple recurrence

1 N 1
lim inf (T n (A) T 2n (A) T kn (A)) > 0.
N N n=0
A further generalization of these theorems which is conjectured by Bergelson
is the following.

Conjecture 3.0.6 ([Ber96]). Let be a countable amenable group with a


j
Flner sequence Fn . Suppose k N and y (X, B, ) are measure
preserving actions of on a probability space (X, B, ), for 0 j < k. Assume
moreover that the actions pairwise commute, i.e., i j = j i , for all
i 6= j, and , . Then for each measurable subset A X with (A) > 0 we
have
1 0 0 1 0 k1
lim inf 1 ( (A) (A) (A)) > 0.

n |Fn | Fn

The purpose of the next two sections is to establish this conjecture in the
extreme cases when either all the actions involved are compact or they are all
weak mixing.

53
54 CHAPTER 3. UNIFORM MULTIPLE RECURRENCE

3.1 Multiple recurrence for compact actions


We will first introduce a bit of notation. If (X, B, ) is a probability space,
k1 k1 k1
we will denote by X k = j=0 X, B k = j=0 B, and k = j=0 . Given
f0 , f1 , . . . , fk1 L (X, B, ) we denote by f0 f1 fk1 the function
on X k given by

(f0 fk1 )(x0 , . . . , xk1 ) = f0 (x0 )f1 (x1 ) fk1 (xk1 ).

We also denote by L (X, B, )alg k L (X k , B k , k ) the algebra generated


by all functions of the form f0 fk1 . Note that this extends to a unitary
between L2 (X, B, )k and L2 (X k , B k , k ), hence we will identify these two
spaces.

Theorem 3.1.1. Let be a countable amenable group with a Flner sequence


j
Fn . Suppose k N and y (X, B, ) are compact, measure preserving
actions of on a probability space (X, B, ), for 0 j < k. Suppose f
L (X, B, ), f 6= 0. Then we have

1 0 1 k1
lim inf 1 k (f ) (f )
(f )k22 > 0.
n |Fn | Fn

Proof. Suppose f L (X, B, ) as above, note that we may assume that f 0,


and kf k2 = 1. Let f = f f f L (X k , B k , k ), and fix > 0
to be chosen later. Since each action j is compact, we have that j ()
Aut(X, B, ) is precompact in the weak topology. Thus the diagonal action
y (X k , B k , k ) is also compact since () Aut(X k , B k , k ) is contained
in a product of compact groups. Hence by Lemma 1.7.10 there exists a finite
set E such that inf 0 E k (f) 0 (f)k2 < , for all .
Therefore by the pigeon hole principle, for each finite set F there exists
0 E such that (F, 0 ) = { F | k (f) g0 (f)k2 < } satisfies

|(F, 0 )| |F |/|E |.

Note that if 01 (F, 0 ) then for each 0 i < k we have


k1 j
2hi (f ), f i 2j=0 h (f ), f i

= 2 k (f) fk22 2 2 ,
and hence
kf i (f )k22 = 2 2h (f ), f i 2 .
We therefore have
k0 (f )1 (f ) k1 (f )k2
k1
kf k k2 j=0 k1
kf k kj (f ) f k2

kf k k2 kkf k
k1
.
3.2. MULTIPLE RECURRENCE FOR WEAK MIXING ACTIONS 55

Thus, if we set = kf k k2 /2kkf k


k1
then we have

1 0 k1
lim inf 1 k (f )
(f )k22
n |Fn | Fn

|01 (F, 0 ) Fn | k 2
lim inf sup kf k2 /4
n 0 E |Fn |

|01 (F, 0 ) 01 Fn | k 2
= lim inf sup kf k2 /4
n 0 E |Fn |

kf k k22 /4|E | > 0.

Note that if i is compact for all 0 i < k and the actions pairwise commute,
then an easy exercise shows that the actions 0 1 i are also compact, hence
by considering f = 1A , the above theorem verifies Conjecture 3.0.6 in this case.

3.2 Multiple recurrence for weak mixing actions


Definition 3.2.1. Suppose y(X, B, ) is a measure preserving action of a
countable amenable group on a probability space (X, B, ). Let Fn be
a Flner sequence, and suppose A L (X, B, ) is a -invariant unital, self-
adjoint subalgebra. A state on A is said to be generic with respect to Fn ,
and if Z
1
lim 1 ( (f )) = f d,
n |Fn | Fn

for all f A.

The example to keep in mind is when Zy(T, Borel , ) is the rotation action
of Z on the circle by an irrational (modulo 2) angle. Where A is the algebra
of continuous functions on T, and A is given by evaluation at some point
z0 T. Then it follows from Birkoffs Ergodic Theorem that for almost all
choices of z0 , this state is generic. This justifies the terminology.

Lemma 3.2.2. Suppose y(X, B, ) is an ergodic, measure preserving action


of a countable amenable group on a probability space (X, B, ). Let Fn
be a Flner sequence, and suppose A L (X, B, ) is a -invariant unital,
self-adjoint subalgebra. If is a state on A which is generic with respect to Fn ,
and then Z
1
lim Fn1 (f ) = f d,
n |Fn |

for all f A, where the convergence is in L2 (A, ).


56 CHAPTER 3. UNIFORM MULTIPLE RECURRENCE
R
Proof. Note, that when is given by (f ) = f d then this is just a restate-
ment of von Neumanns Ergodic Theorem.
Let > 0 be given, and suppose
R that f A. Note that by subtracting the
integral we may assume that f d = 0. Since the action is ergodic it follows
from von Neumanns Ergodic Theorem that there exists n0 N such that
1 2
k 1 (f )k 2
L (X,B,) < .
|Fn0 | 0 Fn0 0
Using the triangle, followed by the Cauchy-Schwartz inequality ((| N1 N 1 2
n=0 an |)
( N1 N 1
n=0 |an |)
2 1 N 1 2
N n=0 |an | ), and then using the fact that is generic we
then have
1 1 2
lim sup (| Fn1 1 0 (f )| )
n |Fn | |Fn0 | 0 Fn0
1 1
lim sup Fn1 ( (| 0 Fn1 0 (f )|2 ))
n |F n | |Fn0 | 0

1 2
=k L (X,B,) < .
1 (f )k 2
|Fn0 | 0 Fn0 0
Also, for each 0 Fn0 we have that
1 |Fn 01 Fn |
lim sup k Fn1 ( (f ) 0 (g))k lim sup = 0.
n |Fn | n |Fn |
Hence, combining this with the above inequality we have
1 2
lim sup (| 1 (f )| ) < .
n |Fn | Fn
Since > 0 was arbitrary this shows that
1 2
lim sup k 1 (f )k 2
L (A,) = 0.
n |Fn | Fn

Exercise 3.2.3. Generalize the above lemma to arbitrary measure preserving


actions. That is to say, suppose y(X, B, ) is a measure preserving action of
a countable amenable group on a probability space (X, B, ). Let Fn
be a Flner sequence, and suppose A L (X, B, ) is a -invariant unital,
self-adjoint subalgebra such that if L (X, I, ) is the algebra of bounded -
invariant functions then L (X, I, ) A is dense in L2 (X, I, ). Suppose is
a state on A such that for all f A we have
Z
1
lim 1 ( (f )) = f d.
n |Fn | Fn

Then show that for all f A we have


1
lim 1 (f ) = EI (f ),
n |Fn | Fn
where the convergence is in L2 (A, ).
3.2. MULTIPLE RECURRENCE FOR WEAK MIXING ACTIONS 57

Theorem 3.2.4. Let be a countable amenable group with Flner sequence


j
Fn . Suppose k N and y (X, B, ) are weak mixing, measure preserving
actions of on a probability space (X, B, ), for 0 j < k. Assume moreover
that the actions pairwise commute, i.e., i j = j i , for all i 6= j,
and , , and that the actions 7 i i+1 j are weak mixing for all

0 i j < k. Suppose f0 , . . . , fk1 L (X, B, ). Then we have
Z
1 0 0 1 0 k1 k1
lim 1 (f0 )( (f1 )) ( (fk1 )) = j=0 ( fj d),

n |Fn | Fn

where the convergence is in L2 (X, B, ).


Proof. Consider the action y(X k , B k , k ) given by (x0 , x1 , . . . , xk1 ) =
(0 x0 , 0 1 x1 , . . . , 0 k1 xk1 ).
Denote by the diagonal measure on X k given by (A) = ({x X | (x, x, . . . , x)
A}). Then we obtain a well defined state on A = L (X, B, )alg k by
Z
(f ) = f d,

for all f A.
By Lemma 3.2.2, in order to prove the theorem it is enough to show that
is generic with respect to Fn , and k . We will do by induction on k.
Note that k = 1 is trivial. If this holds for k 1, and f0 , f1 , . . . , fk
L (X, B, ), then by Lemma 3.2.2 we have that
Z
1
lim Fn1 1 (f1 ) (1 k (fk )) = kj=1 ( fj d),
n |Fn |

in L2 (X, B, ).
Multiplying this by f0 and integrating we obtain
Z Z
1 1 1 k k
lim 1 f0 (f1 ) ( (fk )) d = j=0 ( fj d).
n |Fn | Fn

Applying 0 then gives the result.

Corollary 3.2.5. Let ZyT (X, B, ) be a weak mixing, measure preserving ac-
tion on a probability space (X, B, ). Then for each k N, and f0 , f1 , . . . , fk1
L (X, B, ) we have
Z
1 N 1 k1
lim n=0 n (f0 )2n (f1 ) kn (fk1 ) = j=0 ( fj d),
N N

where the convergence is in L2 (X, B, ).


Proof. If we consider the action i : Z Aut(X, B, ) which takes the generator
of Z to T i , then it follows from Corollary 1.7.7 that these actions satisfy the
hypotheses of the above theorem.
58 CHAPTER 3. UNIFORM MULTIPLE RECURRENCE

3.3 Relatively independent joinings and the ba-


sic construction
Having established Furstenbergs multiple recurrence for the cases of compact
and weak mixing actions of Z, it follows from Proposition ?? that every probabil-
ity measure preserving action Zy(X, B, ) has a non-trivial factor Zy(X, A, )
for which Furstenbergs multiple recurrence holds. To establish this principle
in the general case we must show how to extend this result to larger factors.
The strategy for doing this is to replace C with L (X, A, ) and to mimic the
previous constructions and proofs.

3.3.1 Joinings
Let y(X, B, ), and y(Y, A, ) be two measure preserving action of a count-
able group on probability spaces (X, B, ), and (Y, A, ). For the moment let
us denote by B = L (X, B, ) and A = L (Y, A, ). Suppose : B A is
a map which is positive R ((f ) 0, R whenever f 0), unital ((1) = 1), pre-
serves the integrals ( (f ) d = f d, for all f B), and is -equivariant
(( (f )) = ((f )), for all , and f B).
Then on the algebra B alg A we obtain a state such that for nj=1 bj aj
B alg A we have
Z
(nj=1 bj aj ) = nj=1 (bj )aj d.

Since is -equivariant, it follows from Proposition 2.3.1 that there exists a


probability space (Z, C, ) with a measure preserving action of , and there
exists a -homomorphism
R : B alg A L (Z, C, ) such that = ,
for all , and (x) d = (x) for all a B alg A.
Because is unital, and preserves the integral, it follows that |B1 and
|1A injective and integral preserving. Thus, L (Z, C, ) contains isomorphic
copies of L (X, B, ) and L (Y, A, ). Specifically, if we denote by CX C
(resp. CY C) the -subalgebra generated by the image of |B1 (resp. |1A )
then CX , and CY are -invariant, and we have that |B1 , and |1A give -
eqivariant isomrophism of L (X, B, ) onto L (Z, CX , ), and L (Y, A, ) onto
L (Z, CY , ).
Moreover, after this identification, we recover the map by the formula
(f ) = ECY (f ) L (Z, CY , ), for all f L (Z, CX , ). Indeed, this follows
since if b B, and a A, then we have
Z
(b)a d = (b a)
Z Z
= (b)(a) d = ECY ((b))(a) d.

Given b B, and a A, we will often abuse notation and denote by b a


the function (b a) L (Z, C, ). One should be careful however with this
abuse of notation because need not be faithful in general.
3.3. JOININGS AND THE BASIC CONSTRUCTION 59

Definition 3.3.1. Let y(X, B, ), and y(Y, A, ) be two measure preserv-


ing action of a countable group on probability spaces (X, B, ), and (Y, A, ).
A joining of these two actions is a measure preserving action y(Z, C, ) to-
gether with -equivariant, integral preserving embeddings of L (X, B, ) and
L (Y, A, ) in to L (Z, C, ), such that C is the -algebra generated by these
embeddings.
If y(Y, A, ) is the same action as y(X, B, ) then we say that a joining
is a self joining of the action y(X, B, ).

From the discussion above, joinings are in 1-1 correspondence with -equivariant,
unital, integral preserving, positive maps : L (X, B, ) L (Y, A, ).
If y(Z0 , C, ) is a probability measure preserving action, and we have two
probability measure preserving actions y(X, B, ) and y(Y, A, ), and -
equivariant embeddings : L (Z0 , C, ) L (X, B, ) and : L (Z0 , C, )
L (Y, A, ), then we obtain a -equivariant, positive, unital, integral preserv-
ing map from L (X, B, ) to L (Y, A, ) by first taking the conditional expec-
tion from L (X, B, ) to (L (Z0 , C, )) and then applying the isomorphism
1 . The joining corresponding to this map is called the relatively inde-
pendent joining over (Z0 , C, ). We denote the new space on which acts by
X (C)=(C) Y , or simply by X C Y if the embeddings and are clear from
the context. We also denote by L (X, B, ) alg
(C)=(C) L (X, B, ) the vector
space generated by functions of the type b a where b B, and a A.
A special case to consider is when y(X, B, ) contains a -invariant -
subalgebra A, and we have (Z0 , C, ) = (X, A, ), then we may consider the
relatively independent self joining over (X, A, ). Note, however that a rel-
atively independent joinings consists not only of invariant -subalgebras, but
also the ways in which we are including these subalgebras into the larger alge-
bras. For instance, if Aut(L (X, A, )) is a -equivariant automorphism,
then we obtain a new relatively independent joining by considering the alter-
nate embedding : L (X, A, ) L (X, A, ) L (X, B, ). In general,
the actions yX A X and yX A=(A) X need not be isomorphic.

3.3.2 The basic construction


Relatively independent joinings over the trivial -subalgebra corresponds to
taking a diagonal action on a product space. We have seen previously that a
useful tool in analyzing the structure of product actions yX X was the
identification between L2 (X X) and the Hilbert-Schmidt operators on L2 (X).
This allowed us to use tools such as functional calculus. There is an analog of
the Hilbert-Schmidt operators in the setting of relatively independent joinings
which we will now describe.
Suppose (X, B, ) is a probability space and A B is a -subalgebra. The
basic construction associated to the inclusion L (X, A, ) L (X, B, )
is the algebra L (X, A, ) B(L2 (X, B, )), of operators in B(L2 (X, B, ))
which commute with L (X, A, ). We will denote the basic construction by
60 CHAPTER 3. UNIFORM MULTIPLE RECURRENCE

hL (X, B, ), eA i, where eA B(L2 (X, B, )) is the orthogonal projection onto


L2 (X, A, ).
Note that we have

L (X, B, )eA L (X, B, ) = sp{feA g | f, g L (X, B, )} hL (X, B, ), eA i,

and that this is an algebra since if f L (X, B, ) we have

eA f eA = EA (f )eA .

Note that we distinguish here the the projection eA , which is an operator on


L2 (X, A, ), from the conditional expectation EA , which is an operator on
L (X, A, ).
Exercise 3.3.2. Suppose S hL (X, B, ), eA i, has polar decomposition S =
V |S|. Show that V hL (X, B, ), eA i.
Lemma 3.3.3. Suppose (X, B, ) is a probability space and A B is a -
subalgebra. For each S hL (X, B, ), eA i there exists a unique (S)
L (X, A, ) such that
eA SeA = (S)eA .
Moreover, the map S 7 (S) is a unital, positivity preserving extension of
EA , which is L (X, A, )-bimodular, and continuous with respect to the weak
operator topology.
Proof. By Lemma 2.5.2 L (X, A, ) is a maximal abelian subalgebra of B(L2 (X, A, )).
Thus, since eA SeA restricted to L2 (X, A, ) commutes L (X, A, ), there exists
a unique element (S) L (X, A, ) such that eA SeA f = (S)f = (S)eA f ,
for all f L2 (X, A, ). If f L2 (X, A, ) L2 (X, B, ) then we have
eA SeA f = 0 = (S)eA .
That x 7 (x) is unital, positivity preserving, and weak operator topology
continuous, follows from the fact that it is the composition of the map x 7
eA xeA and the -homomorphism aeA 7 a.
Exercise 3.3.4 (Generalized Cauchy-Schwartz inequality). Prove that for all
x, y hL (X, B, ), eA i we have

|A (y x)|2 A (y y)A (x x).

In the case where the -algebra A is trivial we have that eA is the rank 1
projection on to the subspace C1 L2 (X, B, ), and thus operators of the form
f eA g were rank 1 projections. Rather than working with a Hilbert space basis
{i } L2 (X, B, ) as before, we could have just as easily worked with the family
of partial isometries from C1 to Ci . This motivates the following lemma.
Lemma 3.3.5. Suppose (X, B, ) is a probability space and A B is a -
subalgebra. There exists a family of partial isometries {vi }iI hL (X, B, ), eA i
such that
(a). vi vi eA , for all i I;
3.3. JOININGS AND THE BASIC CONSTRUCTION 61

(b). vi vj = 0, for all i, j I, i 6= j;


(c). iI vi eA vi = 1.
Proof. A simple argument with Zorns Lemma shows that there is a maximal
(with respect to inclusion) family of partial isometries {vi }iI satisfying condi-
tions (a) and (b) above.
Let P = iI vi eA vi , and consider S (1 P ) hL (X, B, ), eA i eA . By
considering the polar decomposition S = V |S|, we have that V V = ProjRange(S)
eA , V V = ProjRange(S ) (1 P ), and V hL (X, B, ), eA i. Thus by max-
imality of the family {vi }iI we must have that V = 0, and hence S = 0.
Thus {0} = (1P )hL (X, B, ), eA ieA L2 (X, B, ) (1P )L (X, B, )
1. Since L (X, B, ) 1 is dense in L2 (X, B, ) this shows that P = 1.
A family of partial isometries which satisfy the conditions above will be
called an operator basis for hL (X, B, ), eA i. Note that since vi vi eA we
have that vi vi = A (vi vi ) L (X, A, ) for each i I.
Definition 3.3.6. Suppose (X, B, ) is a probability space and A B is a -
subalgebra. Let {vi }iI be an operator basis for hL (X, B, ), eA i. An operator
S hL (X, B, ), eA i is Hilbert-Schmidt class with Hilbert-Schmidt norm
if Z
kSk2HS = iI A (vi S Svi ) d < .

The quantity kSkHS is the Hilbert-Schmidt norm of S.


This definition does not depend on the operator basis {vi }iI , this can be
seen from the following analogue of Persevals identity. If {wj }jJ is another
operator basis, then we have
Z Z

iI A (vi S Svi ) d = iI jJ A (vi S (wj eA wj )Svi ) d
Z
= iI jJ A (vi S wj )A (wj Svi ) d
Z
= jJ iI A (vj Svi )A (vi S wj ) d
Z
= jJ A (wj S Swj ).

Exercise 3.3.7. Show that A is a contraction from the Hilbert-Schmidt norm


to L2 (X, A, ).
Note that it follows from above that we may approximate S in the Hilbert-
Schmidt norm with finite sums of the form i,j wj (wj Svi )vi . Also, if > 0, and
we consider gj , hi L (X, B, ) such that kgj wj (1)k2 , kvi vi hi vi (1)k2 < ,
then we have

= kA ((wj Svi ) (wj gj A (wj Svi )eA hi vi ))k2


62 CHAPTER 3. UNIFORM MULTIPLE RECURRENCE

= kA (wj Svi ) A (wj gj )A (wj Svi )A (hi vi )k2


kA (wj Svi )k (kwj wj A (wj gj )k2 + kvi vi A (hi vi )k2 )
2kSk.
It also follows that

(k,l)IJ,(k,l)6=(i,j) kA (wl gj )k22 kA (hi vk )k22 < 22 + 4 ,

and hence

kwj (wj Svi )vi gj A (wj Svi )eA hi k2HS 2(kSk + 1)2 + 4 .

Since > 0 was arbitrary we may then use the triangle inequality to conclude
that the -algebra L (X, B, )eA L (X, B, ) is dense in the Hilbert-Schmidt
norm.
If T hL (X, B, ), eA i then T T kT k2, and since A is positivity pre-
serving it follows that
Z
kT Sk2HS = iI A (vi S T T Svi ) d kT k2 kSk2HS.

Also, it follows from the argument above that the adjoint operator S 7 S is
an anti-linear isometry, and hence we also have

kST k2HS kT k2kSk2HS .

In particular, we see that the Hilbert-Schmidt class in hL (X, B, ), eA i is a


two sided ideal.

Exercise 3.3.8. Given a Hilbert-Schmidt class operator S hL (X, B, ), eA i


show that kSkHS = 0 if and only if S = 0.

The Hilbert-Schmidt norm has an associated inner product


Z
hS, T iHS = iI A (vi T Svi ) d,

which is well defined by the generalized Cauchy-Schwartz inequality, and does


not depend on the operator basis from the arguments above.
Thus, the class of Hilbert-Schmidt operators is an inner-product space. This
space is not complete in general1, we denote the Hilbert space completion by
L2 hL (X, B, ), eA i.
Even though the class of Hilbert-Schmidt operators is not a complete space
in general we do have that it is complete when we restrict to the Hilbert-Schmidt
operators whose uniform norm is bounded by some fixed constant.
1 Consider the case when A = B, then it is easy to see that the class of Hilbert-Schmidt

operators coincides with L (X, B, ), and the inner-product structure is the usual inner-
product on L2 (X, B, ).
3.3. JOININGS AND THE BASIC CONSTRUCTION 63

Proposition 3.3.9. Suppose (X, B, ) is a probability space and A B is a


-subalgebra. Suppose K > 0 and consider the convex set

BK = {S hL (X, B, ), eA i | kSk K}.

Then BK is complete in the Hilbert-Schmidt norm.


Proof. Fix an operator basis {vi }iI for hL (X, B, ), eA i. Since A is a con-
traction from the Hilbert-Schmidt norm to L2 (X, A, ), if Sn BK is Cauchy
in the Hilbert-Schmidt norm, then for all i, j I, (vj Sn vi ) L (X, A, ) is
Cauchy in L2 (X, A, ) and we also have that k(vj Sn vi )k K. Hence, there
exists gi,j L (X, A, ) such that kgi,j k K, and

lim kvj (vj Sn vi )vi vj gi,j vi kHS = lim kvj (vj Sn vi )eA vi vj gi,j vi kHS
n n

lim k(vj Sn vi )eA gi,j eA kHS = 0.


n

Since Range(vj ) are pairwise orthogonal subspaces we may then consider


the sum
S = i,jI vj gi,j vi B(L2 (X, B, )).
Then kSk K and it is easy to see that S hL (X, B, ), eA i. More over for
any finite set I0 I it follows from the triangle inequality that

lim ki,jI0 vj (vj Sn vi )vi vj (vj Svi )vi kHS = 0.


n

Since Sn is Cauchy in the Hilbert-Schmidt norm this implies that S is Hilbert-


Schmidt class and kSn SkHS 0.
If y(X, B, ) is a measure preserving action of a countable group , such
that A is -invariant, then we may consider the relatively independent self join-
ing yX A X defined above. We may then define a map : L (X, B, ) A
L (X, B, ) L (X, B, )eA L (X, B, ) by linearly extending the formula

(b a) = beA a,

for all a, b L (X, B, ).


If {vi }iI is an operator basis for hL (X, B, ), eA i, we then have that for
all nk=1 bk ak L (X, B, ) A L (X, B, )
Z
knk=1 bk ak k2L2 (XA X) = nk,l=1 EA (bk bl )al ak d
Z
= nk,l=1 A (bk bl )A (al ak ) d
Z
= i,jI nk,l=1 A (bk vi eA vi bl )A (al vj eA vj ak ) d
Z
= i,jI nk,l=1 A (bk vi )A (vi bl )A (al vj )A (vj ak ) d
64 CHAPTER 3. UNIFORM MULTIPLE RECURRENCE
Z
= i,jI nk,l=1 A (vi bl eA al vj )A (vj ak eA bk vi ) d

= k(nk=1 bk ak )k2HS .
Hence is well defined and extends to a unitary operator (which we will also
denote by ) from L2 (X A X) to L2 hL (X, B, ), eA i.
Moreover, this unitary implements an equivalence between the Koopman
representation of yXA X and the representation of on L2 hL (X, B, ), eA i,
given by S 7 S 1 .
Suppose y(X, B, ), and y(Y, A, ) are two probability measure pre-
serving actions and we have -equivariant embeddings : L (Z0 , C, )
L (X, B, ) and : L (Z0 , C, ) L (Y, A, ). We may consider the class of
operators S B(L2 (X, B, ), L2(Y, A, )) such that S(g) = (g)S and |S S|1/2
is in the Hilbert-Schmidt class of hL (X, B, ), e(C) i. In this case we can con-
sider the norm given by kSkHS = k|S S|1/2 kHS , and consider the completion
under this norm.
It then follows that this is a Hilbert space, and we may consider the map
which linearly extends the formula (b a) = be(C)=(C) a (here we view
e(C)=(C) as an operator from L2 (X, B, ) to L2 (Y, A, )). Then just as above
extends to a unitary operator which implements an isomorphism between the
Koopman representation yY (C)=(C) X and the representation given by
S 7 A SB1 .

Exercise 3.3.10. Fill in the details to the previous paragraphs.

3.4 Ergodic, weak mixing, and compact exten-


sions
Definition 3.4.1. Let be a countable group. An extension y(X, B, ) of a
probability measure preserving action y(X, A, ) is an ergodic extension if
the -algebra I of -invariant sets is contained in A.

Definition 3.4.2. Let be a countable group. An extension y(X, B, )


of a probability measure preserving action y(X, A, ) is a weak mixing
extension if for any > 0 and any finite set F L2 (X, B, ) such that EA (f ) =
0 for all f F , there exists such that for all f, g F we have

kEA ( (f )g)k1 < .

Definition 3.4.3. Let be a countable group and let y(X, B, ) be an ex-


tension of a probability measure preserving action y(X, A, ). A function
f L2 (X, B, ) is almost periodic relative to L (X, A, ) if for all > 0
there exist g1 , . . . , gn L (X, B, ) such that for all we have

(f ) L2 (X, A, )g1 + + L2 (X, A, )gn .


3.4. ERGODIC, WEAK MIXING, AND COMPACT EXTENSIONS 65

That is to say, for each there exist k1 , . . . , kn L2 (X, A, ) such that

k (f ) nj=1 kj gj k2 < .

The proof of the following proposition is a straight forward generalization of


the proof of Proposition 2.2.22 and we leave the details to the reader.

Proposition 3.4.4. Let be a countable group and let y(X, B, ) be an


extension of a probability measure preserving action y(X, A, ). Let C B
be the -algebra generated by all functions which are almost periodic relative to
L (X, A, ). Then A C, C is -invariant, and f L2 (X, B, ) is almost
periodic relative to L (X, A, ) if and only if f is C measurable.

Definition 3.4.5. Let be a countable group. An extension y(X, B, ) of a


probability measure preserving action y(X, A, ) is a compact extension if
every f L2 (X, B, ) is almost periodic relative to L (X, A, ).

Note that Proposition 3.4.4 shows that there is a unique maximal compact
extension of y(X, A, ) which is a factor of y(X, B, ).
We can now generalize the relationship between ergodic, weak mixing, and
compact actions that we have seen before.

Theorem 3.4.6. Let be a countable group and let y(X, B, ) be an ex-


tension of a probability measure preserving action y(X, A, ). The following
conditions are equivalent:
(1). The extension is weak mixing.
(2). There are no non-trivial functions f L2 (X, B, ) which are almost
periodic relative to L (X, A, ).
(3). The relatively independent self joining yX A X is an ergodic exten-
sion of y(X, A, ).
(4). For any ergodic extension y(Y, C, ) of y(X, A, ), the relatively
independent self joining yX A Y is an ergodic extension of y(X, A, ).
(5). The relatively independent self joining yX A X is a weak mixing
extension of y(X, A, ).
(6). For any weak mixing extension y(Y, C, ) of y(X, A, ), the rel-
atively independent self joining yX A Y is a weak mixing extension of
y(X, A, ).

Proof. For (1) = (2) suppose that y(X, B, ) is a weak mixing extension
and f L (X, B, ) is almost periodic relative to L (X, A, ), such that
EA (f ) = 0. Consider > 0, then there exists g1 , . . . , gn L (X, B, ) such
that for every , there are k1 , . . . , kn L2 (X, A, ) such that

k (f ) nj=1 kj gj k2 < . (3.1)


66 CHAPTER 3. UNIFORM MULTIPLE RECURRENCE

Note that since sup k (f )k = kf k we may assume that

K= sup kkj k < .


1jn,

Also, by replacing gj with gj EA (gj ) we may assume that EA (gj ) = 0 for


1 j n.
Since y(X, B, ) is a weak mixing extension there exists such that
for all 1 j n we have

kEA ( (f )gj )k1 < .

Hence, for all 1 j n we also have

|h (f ), kj gj i| kEA ( (f )gj kj )k1 = kEA ( (f )gj )kj k1 K. (3.2)

Combining (3.1) and (3.2) we then have

kf k22 k (f )k22 + knj=1 kj gj k22

= k (f ) nj=1 kj gj k22 2h (f ), nj=1 kj gj i 2 + 2Kn.


Since > 0 was arbitrary this shows that f = 0.
For (2) = (3) suppose that the relatively independent self joining yXA
X is not an ergodic extension. Hence, there exists a non-trivial -invariant
function f L2 (X A X) such that EA (f ) = 0. Using the identification from
Section ?? we then produce a non-trivial element S L2 hL (X, B, ), eA i such
that A (S) = 0, and Ad( )(S) = S for all . Note that we may assume
kSkHS = 1, and hence by considering an operator S0 hL (X, B, ), eA i such
that kS0 kHS = 1, and kS0 SkHS < 1/4, we have that kAd( )(S0 ) S0 kHS <
1/2, for all .
If we consider then the closed (in the Hilbert-Schmidt norm) convex hull
of {Ad( )(S0 ) | }, then the element T of minimal norm is Ad( )-
invariant, satisfies T 6 L (X, A, ), and is bounded in the uniform norm by
Proposition 3.3.9. Consider the unit ball B1 = {f L2 (X, B, ) | kf k2 1},
then T (B1 ) L2 (X, B, ) is -invariant.
If > 0 is given and we consider an operator basis {vi }iI for hL (X, B, ), eA i
then we may find a finite set I0 I such that

kT i,jI0 vj A (vj T vi )vi kHS < /2kT k.

It is then not hard to see that if we consider fj L (X, B, ) such that kfj
vj (1)k2 < /2|I0 |, then for all f T (B1 ) we have

f jI0 L2 (X, A, )fj .

This shows that every function in T (B1 ) is almost periodic relative to L (X, A, ).
The same argument also shows the same for every function in T (B1 ). Since
3.4. ERGODIC, WEAK MIXING, AND COMPACT EXTENSIONS 67

T 6= eA T eA this shows that there exists a function which is not in L2 (X, A, )


which is almost periodic relative to L (X, A, ).
For (3) = (1) if the action is not a weak mixing extension, then there
exists c > 0 and a finite set F L2 (X, B, ) such that EA (f ) = 0 for all f F ,
and for all we have

f,gF kEA ( (f )g)k22 f,gF kEA ( (f )g)k21 c.

If we consider the operator T = f F f eA f hL (X, B, ), eA i then we have


that T has finite Hilbert-Schmidt norm, and for each we have

hAd( )(T), TiHS

= f,gF h (f )eA (f ), geA giHS


= f,gF kEA ( (f )g)k22 c
Therefore, by Proposition 1.5.2 there exists a non-trivial Ad( )-invariant vector
in L2 hL (X, B, ), eA i. From Section ?? it then follows that yX A X is not
an ergodic extension.
The proof that (3) (4) then follows from the remarks at the end of Sec-
tion ??. If yX A Y is not an ergodic extension then just as above we
construct an operator T B(L2 (Y, C, ), L2 (X, B, )) which commutes with the
embeddings of A measurable functions, intertwines the Koopman representa-
tions of , and such that T 6= eA T eA . by considering the operators (T T )1/2
and (T T )1/2 , the result follows easily.
The equivalence of (1)-(4) shows that if yX A X is an ergodic extension
then y(X A X) A (X A X) is also an ergodic extension. This then shows
that (1)-(5) are equivalent, and a similar argument works for (6) as well.

Corollary 3.4.7. Let be a countably infinite amenable group with a Flner


sequence Fn , and let y(X, B, ) be an extension of a probability measure
preserving action y(X, A, ). Then y(X, B, ) is a weak mixing extension
if and only if for all f, g L2 (X, B, ) such that EA (f ) = EA (g) = 0 we have
that Z
1
lim Fn1 |EA ( (f )g)|2 d = 0.
n |Fn |

Proof. It is clear that this condition implies that y(X, B, ) is a weak mixing
extension and so we need only prove the converse.
Assume that y(X, B, ) is a weak mixing extension then by Theorem 3.4.6
the relatively independent self joining yX A X is an ergodic extension, and
hence if f, g L (X, B, ) are such that EA (f ) = EA (g) = 0, then we have
that kEI (f g)k2 kEA (f g)k2 = 0, where I is the -algebra of -invariant
sets for X A X.
By von Neumanns Ergodic Theorem we then have
Z
1
lim Fn1 |EA ( (f )g)|2 d
n |Fn |
68 CHAPTER 3. UNIFORM MULTIPLE RECURRENCE

1
= lim 1 h( A )(f f ), g gi
n |Fn | Fn
= hEI (f f ), g gi = 0.
We then obtain the general case when f, g L2 (X, B, ) by continuity.
Index

action invariant vectors, 16


compact, 39
ergodic, 16, 36 joining, 59
mixing, 39
strongly ergodic, 37 Kacs Theorem, 51
weak mixing, 37 Kakutani tower, 51
adjoint Hilbert space, 5 Khintchines Theorem, 52
almost invariant vectors, 16 Koopman representation, 35
almost periodic function, 40
amenable group, 18 lattice, 33
asymptotically invariant sequence, 37
mean, 18
basic construction, 59
odometer action, 32
Bernoulli shift, 32
Birkoffs Ergodic Theorem, 48
Poincares Recurrence Theorem, 50
cocycle, 9
quasi-invariant measure, 31
composition of cocycles, 15
conditional expectation, 46 randomorphism, 34
conjugacy, 32 relatively independent joining, 59
cyclic vector, 6 representation, 5
adjoint, 5
ergodic extension, 64
amenable, 29
Flner net, 18 compact, 24
factor, 40 induced, 12, 13
functions of positive type, 8 irreducible, 6
fundamental domain, 11 left-regular, 5
Furstenbergs Correspondence Princi- mixing, 22, 23
ple, 34 quasi-regular, 5
tensor product, 6
Gaussian action, 43 trivial, 5
generalized Bernoulli shift, 32 weak mixing, 22, 26, 27
generic measure, 55 Rohlins Theorem, 51

Hilbert-Schmidt class, 7, 61 spectral gap, 16, 37


standard probability space, 41
invariant measure, 31 symmetric creation operator, 43

69
70 INDEX

uniform ordering, 33

Von Neumanns Ergodic Theorem, 21,


45

weak containment, 28
weak mixing, 64
weak topology, 39
Bibliography

[BdlHV08] B. Bekka, P. de la Harpe, and A. Valette, Kazhdans property (T).


new math. monographs, no. 11, CUP, 2008.
[Bek90] B. Bekka, Amenable unitary representations of locally compact
groups, Invent. Math. 100 (1990), no. 2, 383401.
[Ber96] Vitaly Bergelson, Ergodic Ramsey theoryan update, Ergodic the-
ory of Zd actions (Warwick, 19931994), London Math. Soc. Lecture
Note Ser., vol. 228, Cambridge Univ. Press, Cambridge, 1996, pp. 1
61.
[Bir31] G. D. Birkhoff, Proof of the ergodic theorem, Proc. Natl. Acad. Sci.
U. S. A. 17 (1931), no. 12, 656660.
[BO08] N. Brown and N. Ozawa, C -algebras and finite-dimensional approx-
imations, Graduate Studies in Mathematics, vol. 88, AMS, 2008.
[Day57] Mahlon M. Day, Amenable semigroups, Illinois J. Math. 1 (1957),
509544.
[dCTV07] Yves de Cornulier, Romain Tessera, and Alain Valette, Isometric
group actions on Hilbert spaces: growth of cocycles, Geom. Funct.
Anal. 17 (2007), no. 3, 770792.
[DS88] Nelson Dunford and Jacob T. Schwartz, Linear operators. Part I,
Wiley Classics Library, John Wiley & Sons Inc., New York, 1988,
General theory, With the assistance of William G. Bade and Robert
G. Bartle, Reprint of the 1958 original, A Wiley-Interscience Publi-
cation.
[Fel63] J. M. G. Fell, Weak containment and Kronecker products of group
representations, Pacific J. Math. 13 (1963), 503510.
[FK78] H. Furstenberg and Y. Katznelson, An ergodic Szemeredi theorem
for commuting transformations, J. Analyse Math. 34 (1978), 275
291 (1979).
[Fol95] Gerald B. Folland, A course in abstract harmonic analysis, Studies
in Advanced Mathematics, CRC Press, Boca Raton, FL, 1995.

71
72 BIBLIOGRAPHY

[Fur77] Harry Furstenberg, Ergodic behavior of diagonal measures and a


theorem of Szemeredi on arithmetic progressions, J. Analyse Math.
31 (1977), 204256.

[Gri84] R. I. Grigorchuk, Degrees of growth of finitely generated groups and


the theory of invariant means, Izv. Akad. Nauk SSSR Ser. Mat. 48
(1984), no. 5, 939985.

[Kac47] M. Kac, On the notion of recurrence in discrete stochastic processes,


Bull. Amer. Math. Soc. 53 (1947), 10021010.

[Khi35] A. Khintchine, Eine Verscharfung des Poincareschen


Wiederkehrsatzes, Compositio Math. 1 (1935), 177179.

[Kie75] J. C. Kieffer, A generalized Shannon-McMillan theorem for the ac-


tion of an amenable group on a probability space, Ann. Probability
3 (1975), no. 6, 10311037.

[Koo31] B. O. Koopman, Hamiltonian systems and transformation in Hilbert


space, Proc. Natl. Acad. Sci. U. S. A. 17 (1931), no. 5, 233238.

[Mon06] Nicolas Monod, An invitation to bounded cohomology, International


Congress of Mathematicians. Vol. II, Eur. Math. Soc., Zurich, 2006,
pp. 11831211.

[MOP79] Jean Moulin Ollagnier and Didier Pinchon, Dynamical systems of


total orders, Ergodic theory (Proc. Conf., Math. Forschungsinst.,
Oberwolfach, 1978), Lecture Notes in Math., vol. 729, Springer,
Berlin, 1979, pp. 153162.

[Nam64] I. Namioka, Flners conditions for amenable semi-groups, Math.


Scand. 15 (1964), 1828.

[Ols80] Alexander Yu. Olshanskii, On the question of the existence of an in-


variant mean on a group, Uspekhi Mat. Nauk 35 (1980), no. 4(214),
199200.

[OS02] Alexander Yu. Olshanskii and Mark V. Sapir, Non-amenable finitely


presented torsion-by-cyclic groups, Publ. Math. Inst. Hautes Etudes
Sci. (2002), no. 96, 43169 (2003).

[Pet83] Karl Petersen, Ergodic theory, Cambridge Studies in Advanced


Mathematics, vol. 2, Cambridge University Press, Cambridge, 1983.

[Poi90] Henri Poincare, Sur le probleme des trois corps et les equations de
la dynamique, Acta Math. 13 (1890), 1270.

[PS] J. Peterson and T. Sinclair, On cocycle superrigidity for gaussian


actions, Preprint 2009. arXiv:0910.3958.
BIBLIOGRAPHY 73

[PY98] M. Pollicott and M. Yuri, Dynamical systems and ergodic theory,


London Mathematical Society student texts, vol. 40, Cambridge
University Press, Cambridge, 1998.
[Roh48] V. Rohlin, A general measure-preserving transformation is not
mixing, Doklady Akad. Nauk SSSR (N.S.) 60 (1948), 349351.
[vN29] John von Neumann, Zur allgemeinen theorie des maes, Fund.
Math. 13 (1929), 73111.
[vN32] , Proof of the quasi-ergodic hypothesis, Proc. Natl. Acad. Sci.
U. S. A. 18 (1932), no. 1, 7082.