Bernoulli actions are weakly contained in any free

action
Mikl os Abert and Benjamin Weiss
March 5, 2011
Abstract
Let be a countable group and let f be a free probability measure
preserving action of . We show that all Bernoulli actions of are weakly
contained in f.
It follows that for a nitely generated group , the cost is maximal
on Bernoulli actions for and that all free factors of i.i.d. of have the
same cost.
We also show that if f is ergodic, but not strongly ergodic, then f
is weakly equivalent to f I where I denotes the trivial action of on
the unit interval. This leads to a relative version of the Glasner-Weiss
dichotomy.
1 Introduction
Let be a countable group. By a probability measure preserving (p.m.p.) action
f of we mean a representation of by f

in the group of measure preserving

transformations of (X, B, ). Here (X, B, ) is a standard Borel probability
space. The action f is free, if for -almost every x X and ,=

we have
f

(x) ,= f

(x).
Let f and g be p.m.p. actions of on (X, B, ) and (Y, (, ), respectively.
Following the denition of Kechris [7], we say that f weakly contains g (f _ g)
if for all Borel subsets Y
1
, . . . , Y
n
(, nite sets S and > 0 there exist
Borel subsets X
1
, . . . , X
n
B such that
[(f

X
i
X
j
) (g

Y
i
Y
j
)[ < (1 i, j n, S).
This means that the way g acts on nite partitions of the underlying space can be
simulated by f with arbitrarily small error. A special case of weak containment
is when g is a factor of f, that is, there exists a surjective measure preserving
-equivariant map from (X, B, ) to (Y, (, ). We call f and g weakly equivalent
if f _ g and g _ f.
Let be a probability space. The Bernoulli action

is dened as the set

of maps from to , endowed with the product measure and the shift action
by . These actions are also called i.i.d. processes on .
1
The main result of this paper is the following:
Theorem 1 Let be a countable innite group and let f be a free p.m.p. action
of . Then f weakly contains every Bernoulli action of . In particular, all free
factors of i.i.d.-s of are weakly equivalent.
Note that one can obtain the original Rokhlin lemma for Z as a quick corol-
lary of Theorem 1. In this language, the lemma says that every free action of
Z weakly contains the cyclic action of Z on n points for all n > 0. Since weak
containment is transitive, by Theorem 1, it is enough to prove this for 0, 1
Z
with p(0) = p(1) = 1/2. For a natural number k, 0 l n 1 and 0, 1
Z
let
||
k,l
=
k1

i=0
(in +l)
and let
A
k
=
_
0, 1
Z
[ ||
k,0
> ||
k,l
+n for all 1 l n 1
_
.
Let T denote the shift operator on 0, 1
Z
. Then the sets A
k
, TA
k
, . . . , T
n1
A
k
are disjoint, while the measure of A
k
converges to 1/n as k tends to innity,
since for a xed k, the ||
k,l
are independent binomial variables.
As a corollary of Theorem 1, one gets a new result on cost. The cost is
a numeric invariant of a p.m.p. action of a countable group. Its study was
initiated by G. Levitt [9] and has been investigated in depth by D. Gaboriau
[3]. One of the major open problems here is the Fixed Price problem, that is,
whether every free action of a countable group has the same cost. A. Kechris
and B. Miller showed that for nitely generated groups and free actions, the
cost is monotonic with respect to weak containment [8], so Theorem 1 leads to
the following.
Corollary 2 Let be a nitely generated group. Then, among free p.m.p.
actions of , the cost is maximal on free factors of i.i.d.-s. In particular, all
free factors of i.i.d.-s have the same cost.
Since the cost of any free action of an innite group is at least 1, we get that
an innite group has xed price 1 if and only if one of its Bernoulli actions
have cost 1. We will generalize the monotonicity result of A. Kechris and B.
Miller to actions that are not necessarily free, but for that we need to use an
extension of the notion of cost called the groupoid cost. For details see Section
3.
In the language of weak containment, one can express the well-known notion
of strong ergodicity as follows. An action of is strongly ergodic, if it is ergodic
and it does not weakly contain the identity action of on two points with equal
mass.
In the presence of ergodicity, the absence of strong ergodicity forces various
structural restrictions on the action. An example of this is the theorem of V.
2
Jones and K. Schmidt [6] that says that the equivalence relation dened by such
an action factors onto a nontrivial hypernite equivalence relation. We will give
another kind of restriction in the language of weak containment.
For a measure preserving action f on (X, B, ) let f I denote the diagonal
action of on (X, B, ) [0, 1] where acts on the second coordinate trivially.
Theorem 3 Let f be an ergodic p.m.p. action of a countable group. Then the
following are equivalent:
1) f is not strongly ergodic;
2) f is weakly equivalent to f I.
This leads to the following convexity result, that can be viewed as a relative
Glasner-Weiss theorem.
For a compact topological space K let K

denote the shift of with base

set K. This is a continuous action on a compact space. Let M(, K) denote
the set of -invariant Borel measures on K

, endowed with the weak* topology

and let E(, K) M(, K) denote the set of ergodic measures. Then M(, K)
is a simplex where the set of extreme points equals E(, K) (see ).
For an ergodic p.m.p. action f on (X, B, ) any Borel map : X K
denes a -equivariant map : X K

by setting, for x X and

(x)() = (f

(x)).
The measure is an invariant measure on K

.
Let
E(f, K) = [ a Borel map from X to K .
One can also describe E(f, K) as the set of invariant measures on K

that are
factors of f. We have the following dichotomy.
Theorem 4 Let f be an ergodic p.m.p. action of the countable group . If f
is strongly ergodic, then E(f, K) E(, K). If f is not strongly ergodic, then
E(f, K) is convex.
The Glasner-Weiss theorem [5] says that when has property (T), then
E(, K) is closed, and when it does not have property (T), then E(, K) =
M(, K). We can deduce this result from Theorem 4 as follows. By the Glasner-
Thouvenot-Weiss theorem [4], for every countable group there exists an er-
godic p.m.p. action f of that weakly contains all ergodic actions of . This
means that E(, K) E(f, K).
If has property (T), then by K. Schmidt [11], every ergodic action of
is strongly ergodic, in particular, f is strongly ergodic. So by Theorem 4
E(f, K) = E(, K) and hence E(, K) is closed.
If does not have property (T), then by the Connes-Weiss theorem, has an
ergodic action that is not strongly ergodic. But that action is weakly contained
in f, so f is also not strongly ergodic. So by Theorem 4, E(f, K) is convex
and since it contains all of E(, K), it must equal M(, K), the convex hull of
E(, K).
3
The paper is organized as follows. In Section 2 we prove Theorem 1. In
Section 3 we dene the groupoid cost and show that for nitely generated groups,
it is monotonic with respect to weak containment. Finally, in Section 4 we prove
Theorem 3 and Theorem 4.
Acknowledgement. The authors thank the Glasner family (Eli and Yair) for
establishing a link between them, that led to this paper.
2 Weak containment of i.i.d. actions
In this section we prove Theorem 1. Throughout the paper, when it is conve-
nient, we will assume that the p.m.p. action is continuous on a compact metric
space. This can be easily seen by taking a Borel bijection between X and the
Cantor set K and then considering the shift action on K

with the invariant

measure .
For positive numbers p
1
, . . . , p
d
with
d

i=1
p
i
= 1
let (p
1
, . . . , p
d
) denote the probability space on 1, . . . , d where P(i) = p
i
.
For any T there is a natural projection
T
from1, . . . , d

to 1, . . . , d
T
.
The sets of the form
1
T
(D) with D 1, . . . , d
T
are called the cylinder sets
with respect to T.
Lemma 5 Let f be a p.m.p. action of the countable group on (X, B, ). Then
f weakly contains (p
1
, . . . , p
d
)

if for all nite subsets F of and > 0 there

exists a Borel map : X 1, . . . , d such that for all 1, . . . , d
F

(x X [ (f

x) = () for all F)

F
p
()

< .
Proof. Assume that the condition holds. Let g denote the shift action of on
(p
1
, . . . , p
d
)

and let denote the measure on (p

1
, . . . , p
d
)

.
Let Y
1
, . . . , Y
n
1, . . . , d

be Borel subsets, let S be a nite symmetric

set containing the identity and let > 0.
Then for any r > 0, there exists a nite subset T of , and subsets D
i

1, . . . , d
T
such that for all 1 i n we have
(Y
i

1
T
(D
i
)) < r.
Here denotes symmetric dierence. We now apply the condition of the lemma
for F = TS and a > 0 to be chosen later.
For 1 i n let
X
i
=
1
(
1
T
(D
i
)).
4
Then for suciently small r and it follows that
[(f

X
i
X
j
) (g

Y
i
Y
j
)[ < (1 i, j n, S).
The lemma is proved.
Lemma 6 Let f be a free p.m.p. action of the countable group on the compact
metric space (X, d, ). Then for any nite symmetric subset F of and any
> 0 there exists s > 0 such that
(x X [ d(f

x, f

x) > s for all ,=

F) > 1 (*)
and
((x, x

) X X [ d(f

x, f

) > s for all ,

F) > 1 (**)
In other terms, for this s, most of the F-neighbourhoods of pairs of points
in X are s-apart. We leave the details to the reader.
We are ready to prove Theorem 1.
Proof of Theorem 1. We will verify that the conditions of Lemma 5 hold.
Let F be a nite subset of and let > 0. Let > 0 be determined later.
Use Lemma 6 with this to nd s that satises the conclusions of the lemma.
Let B
j
[ 1 j J be a Borel partition of X into sets of diameter less than s.
Let U
1
, . . . , U
J
be independent random variables with distribution (p
1
, . . . , p
d
).
Dene a random map : X 1, . . . , d by setting
(x) = U
j
when x B
j
.
We shall show that with high probability, this map will satisfy the conditions
of Lemma 5. In particular, this will imply the existence of such a map.
We shall use the so-called second moment method. For 1, . . . , d
F
let
G

= x X [ (f

x) = () for all F .
Now (G

) is a random variable. The expected value

E(G

) = E
_
X
1
G
(x)d(x) =
_
X
E1
G
(x)d(x).
For x X such that that the F-neighbourhood of x is s-apart (that is, the set
dened in (*) of Lemma 6) the expected value E1
G
(x) =

F
p
()
because the
-values at the neighbours are independent. It follows that

E(G

F
p
()

< 2.
5
We shall now estimate the variance of (G

). We have
E
2
(G

) = E
_
X
1
G
(x)d(x)
_
X
1
G
(x

)d(x

) =
= E
_
X
_
X
1
G
(x)1
G
(x

)d( )(x, x

) =
=
_
X
_
X
E(1
G
(x)1
G
(x

))d( )(x, x

)
For x, x

X such that that the F-neighbourhoods of x and x

are s-apart (that

is, when both x, x

satisfy (*) and (x, x

) satises (**) of Lemma 6) then

E1
G
(x)1
G
(x

) =
_
_

F
p
()
_
_
2
because the -values at the neighbours are independent. The -measure of
these pairs of points is at least 1 3. It follows that
Var((G

)) = E
2
(G

) (E(G

))
2
< 12.
Now Chebyshevs inequality says that the probability
P([(G

) E(G

)[ > c)
Var((G

))
c
2
Setting c =
1/3
we get
P([(G

) E(G

)[ >
1/3
) 12
1/3
Setting small enough, we get that [(G

) E(G

)[ is arbitrarily small for

all 1, . . . , d
F
simultaneously. Hence the conditions of Lemma 5 hold and
f weakly contains (p
1
, . . . , p
d
)

for any nite distribution (p

1
, . . . , p
d
).
Since any probability distribution can be approximated by nite distribu-
tions, we get that f weakly contains all Bernoulli actions and the theorem has
been established.
3 Groupoid cost and weak containment
In this section we will introduce the groupoid cost of a p.m.p. action and
show that for nitely generated groups, it is monotonic with respect to weak
containment. Groupoid cost has been introduced in [1] as an extension of cost
for non-free p.m.p. actions. The monotonicity result was proved by A. Kechris
and B. Miller [8] for the original cost and free actions.
Let f be a p.m.p. action of the countable group on (X, B, ). Without loss
of generality, we can assume that the space X is compact. Then the groupoid (
f
6
of the action is dened as follows. Endow with the discrete topology and the
counting measure. Consider X , endowed with the product Borel structure
and the product measure . Elements of (
f
are now dened as Borel subsets of
X .
We dene a partial product on X as follows. The product of (x
1
,
1
)
and (x
2
,
2
) is only dened, when f
1
x
1
= x
2
; in this case let (x
1
,
1
) (x
2
,
2
) =
(x
1
,
1

2
). The inverse of (x, ) is dened as (f

x,
1
) and is denoted by
(x, )
1
, so we have (x, ) (x, )
1
= (x, e). For A, B (
f
let
A B = a b [ a A, b B when a b is dened .
Let E = X e where e is the identity element of . Then for all A (
f
we have AE = EA = A. When f is a continuous action on a compact space,
the partial product is continuous as well, which in particular implies that the
product of open elements of (
f
is open and the product of compact elements is
compact.
We say that A (
f
generates the action f, if

_
n=1
_
A A
1
E
_
n
= X .
In particular, for any generating set S of , X
S
= XS generates f. Note that
this is a topological condition and is independent of the measure.
The groupoid cost of f is dened as
gcost(f) = inf
A generates f
(A).
When f is a free action, X can be identied with the equivalence relation
generated by f on X, elements of (
f
are graphings in this relation and A (
f
generates f if and only if the correspoding graphing generates the equivalence
relation. Hence, for free actions, gcost(f) = cost(f) and so the groupoid cost is
indeed an extension of cost to non-free actions. Note that the original notion of
cost required that the graphing generates the relation only up to a nullset, but
it is easy to see that without loss of generality, one can assume full generation.
Trivially, for every generator A (
f
and > 0 there exists an open generator
B (
f
such that A B and (B) (A) + . So, in the denition of gcost,
it is enough to consider open generators. The following lemma is essentially
contained in [1].
Lemma 7 Assume that X is compact, totally disconnected and is nitely
generated. Then
gcost(f) = inf (A)
where A varies over compact, clopen generators for f.
Proof. Let be generated by the nite set S. Let ( be a countable clopen
basis for the topology of X. List the elements of ( as C
1
, C
2
, . . . For an open
7
set O B and n > 0 let
O
n
=
_
1in
CiO
C
i
.
Let A (
f
be an open generating element for (
f
with (A) < . Then
A =

A() where A() = A(X ) and A() is open. For

n > 0 let A
n
=

A()
n
. Then A
n
is compact (since (A) < ) and
clopen. Let B
n
= A
n
A
1
n
E. Then

_
n=1
B
n
n
= X
since every possible nite product in A A
1
E is realized in B
n
for large
enough n. In particular,

_
n=1
(B
n
n
X
S
) = X
S
but since B
n
n
X
S
is open and X
S
is compact, there exists an n > 0 with
X
S
B
n
n
. Since X
S
generates f, we get that already A
n
generates f.
It follows that any open generator for f with nite measure contains a com-
pact, clopen generator. The lemma is proved.
For a compact topological space K let K

denote the shift of with base

set K. This is a continuous action on a compact space. Let M(, K) denote
the set of -invariant Borel measures on K

, endowed with the weak* topology

and let E(, K) M(, K) denote the set of ergodic measures.
For an ergodic p.m.p. action f on (X, B, ) any Borel map : X K
denes a -equivariant map : X K

by setting, for x X and

(x)() = (f

(x)).
The measure is an invariant measure on K

.
Let
E(f, K) = [ a Borel map from X to K .
One can also describe E(f, K) as the set of invariant measures on K

that are
factors of f.
Lemma 8 Let f and g be p.m.p. actions of the countable group on (X, B, )
and (Y, (, ), respectively. Then f weakly contains g if and only if
E(f, K) E(g, K) (*)
for any compact topological space K. The same result holds if we only consider
the class of nite spaces.
8
Proof. The (*) condition for all nite spaces K is easily seen to be equivalent
with the original denition of weak containment.
If L is a nite subset of K, then L

is contained in K

. The invariant
measures concentrated on L

as L varies over the nite subsets of K, are dense

in M(, K).
Moreover, it follows that
_
LK nite
E(g, L) E(g, K).
This is because
E(g, K) = [ a Borel map from Y to K .
and the Borel map can be well approximated by maps
n
with nite range
L
n
in such a way that
n
converges to . This yields
E(g, K)
_
LK nite
E(g, L)
_
LK nite
E(f, L) E(f, K).
The lemma is proved.
Theorem 9 Let f and g be p.m.p. actions of the nitely generated group
on (X, B, ) and (Y, (, ), respectively, such that f weakly contains g. Then
gcost(f) gcost(g).
Proof. Let K be the standard Cantor set. Let be a Borel isomorphism from
Y to K. Then the shift action on K

with the invariant measure

=
is isomorphic to g. Let h be the shift action of on K

with the invariant

measure

. Clearly, gcost(h) = gcost(g). Using Lemma 7, for any > 0 there

exists a clopen, compact A (
h
that generates h and

(A) < gcost(h) + .

Since weak convergence of probability measures
n
on K

implies convergence
of
n
on clopen compact subsets of K

, using Lemma 8 there exists a map

: X K such that

(A)

(A) +.
The condition of generating is independent of the invariant measure, so A gen-
erates the shift action on K

with the invariant measure as well. This

action is a factor of f, and hence its gcost is an upper bound for the gcost of f.
This implies
gcost(f)

(A)

(A) + < gcost(g) + 2.

The theorem holds.
Proof of Corollary 2. Let be a nitely generated group and let f be a free
p.m.p. action of . Then, using Theorem 1, f weakly contains all Bernoulli
actions, so by Theorem 9 the gcost of f is at most the gcost of any Bernoulli
9
action of , in particular, the cost of f is at most the cost of any free Bernoulli
action of .
Note that any nontrivial Bernoulli action of an innite group is free. Also,
it is easy to see that when is torsion free, then every nontrivial factor of i.i.d.
is free. However, in the presence of torsion, this last statement fails to hold.
4 Ergodic actions and convexity
Let f be a p.m.p. action of the countable group on (X, B, ). A sequence of
subsets A
n
B is almost invariant, if
lim
n
(f

A
n
A
n
) = 0 for all .
The sequence is trivial, if lim
n
(A
n
)(1(A
n
)) = 0. We say that the action
f is strongly ergodic, if every almost invariant sequence is trivial.
The following lemma is due to K. Schmidt [10].
Lemma 10 (Schmidt) Let f be a p.m.p. action of the countable group on
(X, B, ), such that f is ergodic, but not strongly ergodic. Then for all 0 < < 1
there exists an almost invariant sequence A
n
such that (A
n
) = (n 0).
In particular, the action is strongly ergodic if and only if it is ergodic and
does not weakly contain the identity action of on two points with equal mass.
The following lemma forms a part of the proof of Schmidts lemma. For
completeness, we provide a short proof.
Lemma 11 Let f be an ergodic p.m.p. action of the countable group on
the compact space (X, B, ), let 0 < < 1 and let A
n
be an almost invariant
sequence such that
lim
n
(A
n
) = .
Let
n
be the measure on X dened by

n
(Y ) = (A
n
Y ) (Y B).
Then
n
weak* converges to .
Proof. Let n
k
be a sequence such that
An
k
is weakly convergent and let be
the limit measure. Then is invariant under and is bounded by /, so by
ergodicity, is a scalar multiple of . But (X) = lim(A
n
) = which proves
the lemma.
Proof of Theorem 3. Let f be an ergodic p.m.p. action of the countable
group on (X, B, ).
Assume that f is weakly equivalent to f I. Since f I factors on the
identity action b of on two points with equal mass, this means that f weakly
10
contains b. Hence f can not be strongly ergodic.
Now assume f is not strongly ergodic. We claim that f weakly contains f b.
We can assume that the space (X, B, ) is totally disconnected and compact and
acts by continuous maps. Let ( denote the set of clopen sets in B. Then ( is
a nitely additive algebra and for any B B and > 0 there exists C ( with
(B C) < .
Hence, it is enough to check the weak containment condition for clopen
subsets. In addition, by Schmidts lemma, there exists an almost invariant
sequence A
n
( with
lim
n
(A
n
) = 1/2.
Lemma 11 implies that for every clopen set C (, we have
lim
n
(C A
n
) = (C)/2.
Indeed, weak* convergence of measures implies convergence on clopen sets.
Let Y

1
, . . . , Y

k
, Y

1
, . . . , Y

k
(, let S be a nite subset and let > 0.
Then by the almost invariance of A
n
, for all S we have
lim
n
(f

A
n
A
n
) = 0
Also for all 1 i, j k and S we have
lim
n
(A
n
f

i
Y

j
) (f

i
Y

j
)/2 = 0
and
lim
n
(A
n
f

i
Y

j
) (f

i
Y

j
)/2 = 0
Let n be large enough such that all the above quantities (listed in the above
three displays) are less than . Now for 1 i k let
X
i
= (A
n
Y

i
) (A
c
n
Y

i
)
where A
c
n
denotes the complement of A
n
.
Now let Y
1
, . . . , Y
k
be Borel subsets of X 0, 1. Let
Y

i
0 = Y
i
(X 0)
and let
Y

i
1 = Y
i
(X 1).
Applying the above argument for Y

i
and Y

i
we obtain the desired measurable
subsets X
1
, . . . , X
n
. The claim holds and f weakly contains f b.
Iterating the argument, we get that f weakly contains fb
2
n where b
2
n is the
trivial action of on 1, . . . , 2
n
with uniform measure. An easy approximation
argument yields that f weakly contains f I. Since f is a factor of f I, we
get that f is weakly equivalent to f I.
11
The theorem is proved.
Proof of Theorem 4. Let f be an ergodic p.m.p. action of the countable
group .
Assume that f is strongly ergodic. Assume by contradiction that there exists
a sequence of Borel maps
n
: X K such that
n
weak* converges to a non-
ergodic measure on K

. Let A be a -invariant Borel set with 0 < (A) < 1.

Let us approximate A with a clopen sequence A
n
with respect to . Then there
exists k(n) such that
1
k(n)
(A
n
) gives us a nontrivial almost invariant sequence
for f, a contradiction. Hence E(f, K) E(, K).
Assume that f is not strongly ergodic. Let

be the product measure

1/2, 1/2. Let
0
,
1
: X K be Borel maps and let : X 0, 1 K
be dened by
((x, i)) =
i
(x) for i = 0, 1.
Then by Theorem 3, f weakly contains f b, so by Lemma 4,

is in
E(f, K). But

=
1
2

0
+
1
2

1
.
This gives that for ,

E(f, K),
1
2
+
1
2

E(f, K). Hence E(f, K) is

convex.
The theorem holds.
References
[1] M. Ab ert and N. Nikolov, Rank gradient, cost of groups and the rank
versus Heegard genus problem, to appear in J. of the EMS
[2] A. Connes and B. Weiss, Property T and asymptotically invariant se-
quences, Israel Journal of Math. 37 1980 209210.
[3] D. Gaboriau, Co ut des relations dequivalence et des groupes. (French)
[Cost of equivalence relations and of groups] Invent. Math. 139 (2000), no.
1, 4198.
[4] E. Glasner, J.-P. Thouvenot and B. Weiss, Every countable group
has the weak Rohlin property, Bull. London Math. Soc. 38, 2006, 932936.
[5] E. Glasner and B. Weiss, Kazhdans property T and the geometry of
the collection of invariant measures, Geom. Funct. Anal. 7 (1997), no. 5,
917935.
[6] Jones and K. Schmidt, Asymptotically invariant sequences and approx-
imate niteness, Am. J. of Math. 109 (1987) 91-114.
[7] A. Kechris, Global aspects of ergodic group actions, to appear in the
series Mathematical Surveys and Monographs of the AMS
12
[8] A.S. Kechris and B.D. Miller, Topics in orbit equivalence, Springer
2004
[9] G. Levitt, On the cost of generating an equivalence relation, Ergodic
Theory Dynam. Systems 15 (1995), no. 6, 11731181.
[10] K. Schmidt, Amenability, Kazhdans property T, strong ergodicity and
invariant means for ergodic group-actions, Ergodic Theory Dynamical Sys-
tems 1 (1981), no. 2, 223236.
[11] K. Schmidt, Asymptotically invariant sequences and an action of SL(2, Z)
on the 2-sphere, Israel J. of Math. 37 (1980), 193-208.
13