Professional Documents
Culture Documents
see [Voi96, Section 6] and [Voi98]). We also note that the above correspondence
principle for 0 is probably less useful as the amenable part can always be handled
uniformly using the hyperfiniteness of von Neumann algebra LG3 (see in partic-
ular [Jun2]; for example, the proof in [BDJ] in the group case doesnt rely on
quasi-tilings). Concerning pmp equivalence relations, a free product formula has
been established by Gaboriau [Gab] for the cost, allowing for amalgamations over
amenable subrelations, and by Shlyakhtenko [Shl] for 0 , for free product with
trivial amalgamation.
Question 2. Can the assumption that G3 is amenable in Theorem 1 be weakened
(for example to 1 (G3 ) = 1)?
The paper is organized as follows: 1 and 2 establish basic facts about pmp
groupoids and their actions. In the case of s(G), the correspondence groupoids
equivalence relations is achieved by using the formula s(G) = s(G X0G ), where
G X0G is a Bernoulli shift, see Theorem 28 in 7 (other applications of the
correspondence principle are given in 9). The proof of this formula uses the idea
in a result of L. Bowen [Bow, Theorem 8.1] for the sofic entropy, as explained in
4. Other difficulties inherent to the groupoid setting are dealt with in 3, 5, 6
(these difficulties were avoided in [DKP2] by working with groups and their actions
rather than with general groupoids). The proof of Theorem 28 is given in Section
7. In 8 we prove a scaling formula for s(G). In 9 we prove Theorem 1 by putting
together these ingredients.
1. Review of s(G)
Recall that a discrete standard Borel groupoid G with base (=set of objects) G0 ,
source map s G G0 and range map r G G0 , is said to be probability measure
preserving (pmp) with respect to a Borel probability measure on G0 if the left
and right Haar measures h and h1 on G coincide:
h = h1
where
h(A) = Ae d(e) and h1 (A) = Ae d(e)
G0 G0
and A G is a Borel set with Ae = r1 (e) A and Ae = s1 (e) A for e G0 . Then
hG0 = h1 G0 = so we simply denote by h the measure on G .
0
If G = {1, . . . , d}2 then the uniform distance is the normalized Hamming distance
on JdK.
The (von Neumann) trace on JGK is given by
(s) = h(s Ge ) = h{e G0 se = e}.
It is the restriction to JGK LG of the finite trace on the von Neumann algebra LG
of G.
We have
s t = (s1 s) + (t1 t) (s1 st1 t) (st1 ).
We will write tr for the trace on JdK. So
tr() = d1 number of fixed points of JdK.
If F JGK is a finite subset and n Z1 then Fn denotes the set of all products
of at most n elements of F = F F 1 {Id}. For F JGK we let F JGK denote
the set of sums of elements of F with pairwise orthogonal domains and pairwise
orthogonal ranges.
By definition, G is sofic if its full inverse semigroup JGK is sofic:
Definition 3. A pmp groupoid G is called sofic if for every finite set F JGK, > 0
and n Z1 there exist d Z1 and a map
Fn JdK
such that
(i) (st) (s)(t) <
for every s, t Fn such that st Fn ( is -multiplicative) and
(ii) tr (s) (s) <
for every s Fn ( is -trace-preserving).
Remark 4. If G is a group, one can replace JdK by Sym(d) as is easily seen.
Remark 5. The notion of sofic pmp equivalence relations was introduced by Elek
and Lippner in [EL] in terms of graph approximation and in [Oz] by requiring
that JGK is sofic as in the definition above (compare [DKP1, DKP2]). The sofic
property was first considered for groups by Gromov and Weiss in terms of (Cayley)
graph approximation and was studied by Elek and Szabo. It is a simultaneous
generalization of amenability and the LEF property of Vershik and Gordon (see
[Pe] for more details).
Let
SA(F, n, , d) = { Fn JdK satisfying (i), (ii)}
and define
SA(F, n, , d)E = {E SA(F, n, , d)}
for E F .
Definition 6. For E F JGK finite, n Z1 and > 0 define successively
log SA(F, n, , d)E
sE (F, n, ) = lim sup
d d log d
sE (F, n) = inf sE (F, n, )
>0
sE (F ) = inf sE (F, n)
nZ1
A FREE PRODUCT FORMULA FOR THE SOFIC DIMENSION 5
2. Actions of groupoids
Let X be a standard Borel space endowed with a Borel fibration p X G0 where
G is the base of G. If (e )eG0 is a Borel field of probability measures on X we
0
Haar measure on G0 . Recall that a pmp action of G on the fibered space (X, ) is
a measurable map
G G0 X (g, x) gx X
(where G fibers via the source map s G XG ) satisfying the usual axioms of an
action, and such that
g s(g) = r(g)
for ae g G. Groupoid actions are denoted G X. The crossed product groupoid is
the fiber bundle GG0 X endowed with groupoid law defined by (s, x)(t, y) = (st, y)
whenever t(y) = x.
Example 11 (Bernoulli shifts). Given a pmp groupoid G with invariant Haar
measure h and a probability space (X0 , 0 ), consider the probability space
with the fibration X0G G0 and field of measures (0 G )eG0 . Every element
e
x X0G to
e
These maps are measure preserving and if we consider the Bernoulli action of H
with base X0IJ then we see it is H-equivariant: for h H e say h = h1
1 s(d, e)
1
where h1 H(e)
d (h(x)) = (h(xh0 gj (e)s(e,1 f ) )(i,j)IJ,h0 H(e),(e,f )RH )
= ((xs(d,e)h1 h0 gj (e)s(e,1
i
= ((xs(d,e)h1 h0 gj (e)s(e,1
i f)
)(i,j)IJ )
h0 H(e),(e,f )RH
s(s) (Fix(s)) = 0
where
Fix(s) = {x X s(s) sx = x}.
Lemma 14. If the pmp groupoid action G X is essentially free then G X is a
pmp equivalence relation.
Proof. Since G X is essentially free and Ge is countable the set
X0e = {x X e sx x, s Ge , s e}
has measure 1 in X e for every e A G0 a measurable subset with h(A) = 1. Let
(X0A , ) (A, h) be the measure fibration corresponding to (X0e )eA . Since G is
isomorphic to GA X0A , we may assume that sx x for all s G G0 , x X s(s).
Then for (s, x) G X
r(s, x) = s(s, x) x Fix(s)
so r(s, x) s(s, x) for every s G G0 and x X. This shows that G X is an
equivalence relation. It is an easy exercise to check that it is pmp (more generally
if G is pmp and G X is a pmp action, then G X is a pmp groupoid).
Lemma 15. If G is transversally finitely generated then so is GX for any ergodic
pmp action G X.
Proof. Let R be the orbit equivalence relation of G X. Since R is ergodic, there
exists an ergodic automorphism JRK, which is orbit equivalent to a Bernoulli shift
Z {0, 1}Z ([Dye]). For i {0, 1} let Bi be the cylinder set {x {0, 1}Z x0 = i}.
Let p1 , p2 be the projection in L (X) corresponding to B0 and B1 in Z {0, 1}Z.
If F JGK be a finite transversally generating set for G, then F {, p1 , p2 } is a
finite generating set for G X.
Lemma 16. The Bernoulli action G X0G is essentially free if G has infinite
fibers (i.e. Ge = for ae e G0 ), and the support of 0 contains at least two
points. If in addition 0 is diffuse then the action is essentially free.
Proof. Let s Ge , s e, such that Ge = . We show that e (Fix(s)) = 0. If
r(s) s(s) this is clear so we assume s G(e). Note that (xt )tGe is s-invariant
if and only if xsn t = xt for all n Z, t Ge . Thus we can find an infinite family of
pairwise disjoint pairs {si , ti }iI , si ti Ge , such that xsi = xti for every x Fix(s)
8 ROBERT GRAHAM AND MIKAEL PICHOT
and i I. Since I = , the set of x X0G such that xsi = xti is negligible, so
e
G3 X3
G1 X1 GX G2 X2
f1 k f2
3. Overlapping generators
Let F JGK be a finite subset, be a partition of F , and F JdK be a map.
Denote F the set of e G0 such that
e ss1
sF
and
s, t F, es = et (s) = (t).
Here and below we will view a partition of F as a map from F to {1, ..., } = ran
for some Z0 .
Similarly let F be the set of e {1, . . . d} (identified with the base space of the
transitive relation on {1, . . . , d}) such that
e (s)(s)1
sF
and
s, t F, e(s) = e(t) (s) = (t).
The following lemma will be useful in the proof of Theorem 28. We denote a
projection onto a set A by pA .
Lemma 19. Given F JGK, n 1, let
Fn = F {pF0 F0 Fn , partition of F0 }
If SA(Fn , 4nFn + 1, , d), F0 Fn , and is a partition of F0 then
F0
h(F0 ) < c1 (F, n).
d
where c1 (F, n) = 176(32n)F 2n .
Proof. Fix some F0 Fn and let 1 , 2 , . . . , m be the family of partitions of F0
(so m = Be(F0 ) is the Bell number of F0 ). For convenience let pj = pF0j and
pj = pF0
.
If j (s) = j (t), pj = st1 pj so
j
Moreover
(pj ) (pj )(pj ) = (p2j ) (pj )(pj ) < .
m
and since pj = ss1 using [DKP1, Lemma 3.5]
j=1 sF0
RRR m RRR RRR RRR
RR R R R
RRR (pj ) (s)(s)1 RRRR < RRRR ss1 (s)(s)1 RRRR + 150(2F + 1)2n
RRRj=1 sF0 R R
RR RR sF0 sF0 RRR
F0 4 + F0 2 + + 150(2F + 1)2n
157(32n)F 2n .
If V denotes the set of integers 1 k d such that for all 1 j m and s, t F0 :
(s)(t)1 (pj )k = (pj )k
(s)(t)1 (pj )k k
(pj )k = (pj )(pj )k
m
( (pj ))k = ( (s)(s)1 )k
j=1 sF0
then
V (1 7Fn 7Fn Fn Fn 157(32n)F 2n ) d
(1 175(32n)F 2n ) d
So, for all j we have that (pj )V is a projection and
j (s) = j (t) (s)(t)1 (pj )V = (pj )V and tr ((s)(t)1 (pj )V ) = 0
Hence ran((pj )pV ) ran(pj pV ), but since we also have that
m m
( (pj ))pV = ( (s)(s)1 )pV = ( pj )pV
j=1 sF0 j=1
and
A = (s)A(s)
sF
where F JdK.
The next lemma adapts the proof of Theorem 8.1 from L. Bowens paper [Bow].
Lemma 20. If d is large enough (depending on F , n and ) then there is a partition
{A1 , A2 , . . . , Aq } of {1 . . . d} such that if SA(Fn , 4nFn + 1, , d) then for every
subset F Fn and every function F {1, . . . , q}
A
(B ) < c2 (F, n)
d
where c2 (F, n) = 2c1 (F, n) Be(F n ) and Be is the Bell number.
Proof. Fix F Fn and F {1, . . . , q}. Create a random partition
{A1 , A2 , . . . , Aq }
of {1 . . . d} using the following scheme: for each k {1, . . . , d} place k in Ai with
probability 0 (i).
We will find the probability that (B ) A /d < c2 (F, n).
Let 1 , 2 , . . . , m be the family of partitions of F . Let sj ran j F be an
arbitrary section of j . If
0 if s, t F such that j (s) = j (t) and (s) (t)
j =
1 otherwise
then
(B ) = sB(s)
sF
m
= j h(Fj ) 0 ((sj (r)))
j=1 rran j
0 otherwise
We wish to compute E(Zk ).
12 ROBERT GRAHAM AND MIKAEL PICHOT
For k F
we have E(Zk ) = 0. Otherwise, k F
A if and only if
sj (r) k A(sj (r)) for all r ran j and since
j j
1
s1
j (r)
k s1 k, r r ran j
j (r )
we obtain
E(Zk ) = 0 ((sj (r)))
rran j
are at most Fn 2 F non independent pairs (k, l). Now clearly for these pairs
j
1k,ld 1k,ld
= E(Z 2 ) + Fn 2 F
j
E(Z 2 ) = Fn 2 F
j
Now we can apply Chebyshevs inequality to Z
d
for a > 0
var( Zd ) Fn 2 F
Fn 2
Pr ( a)
Z E(Z)
j
d d a2 a2 d2 a2 d
Since Z = A F
j
and E(Z) = F
j
rran j 0 ((sj (r))), and, by Lemma
19,
RR F RRR
RRR j R
RRRh(Fj ) RR < c1 (F, n)
RRR d RRRR
R
we have
R RRR
RRR RRR Fn 2
Pr RRRRh(Fj ) 0 ((sj (k))) F
/d R
RRR (F, 2
RRR
+
A a c n)
R
1
R R
kran j
j
a d
Let a = c1 (F, n) so then
R RR
RRR RRR Fn 2
Pr RRRRh(Fj ) 0 ((sj (k))) F
/d RRR 2c1 (F, n)
RRR RRRR
(c1 (F, n))2 d
A
R
j
kran j
Indeed for large enough d nearly all partitions will satisfy this and hence we will
have non zero probability that for all F Fn and F {1, . . . , q} the inequality
(B ) A /d < c2 (F, n) holds.
then
pA ai pAi < c3 (F, n)
i=1 2
where c3 (F, n) = ((1 + (3 + ) Be()N ))c2 (F, n) and N = 2 .
Proof. For I {1, . . . } define
DI = {x X x Bi i I}
and similarly
DI = {1 k d k Ai i I}.
We want to bound (DI ) DI /d.
Note that for any I {1, . . . }
{x X i I, x Bi } = ( Bi ) .
iI
Now either for some s Fi1 Fi2 we have i1 (s) i2 (s) and so iI Bi = or
otherwise iI Bi = B for some F {1, . . . , q}. In the first case iI Ai =
as well and in the second case by Lemma 20
(B ) A /d < c2 (F, n)
14 ROBERT GRAHAM AND MIKAEL PICHOT
and
A /d = Ai /d = {1 k d i I, k Ai }/d
iI
(DI ) = (1)kI ({x X i I , x Bi })
k=I I I
I =k
and similarly
DI
= (1)kI {1 k d i I , k Ai }
1
d k=I I I
d
I =k
Now
(B ) = (DI ).
I{1,...};iI ai =1
I = {k A x Ai i I}
D
then
(DI ) D
I /d c2 (F, n)N
so
DI D
I 2c2 (F, n)N .
d d
A FREE PRODUCT FORMULA FOR THE SOFIC DIMENSION 15
Let a(I) = iI ai .
pA ai pAi
X X
i=1 2
X
X
X X
X
X
=XX
X ( )p X
X
X
X
X
p a I X
X
X
X X
X
A i D
X I{1,...} iI
X2
X
X
X X
X
X X
X
X X
X
X
X
X X
X X
X X
X
XX
X X
X X
X ( )p X
X
X I XX X
X I XX
p p + a
X X X X
X X X X
A D i D
X X2 X X2
a(I)=1 a(I){0,1} iI
X
X
X X
X
X X
X
X X
X
X
X X X X
XX
X
X D I X X
X
X
X
X
X
X D I DI X X
X
X + DI /d
X X X X
p p + p p
X X
X2 X X
X X X X
A
X Xa(I)=1 X2
a(I)=1 a(I)=1 a(I){0,1}
RRR A D RRRR D
RRR I RR I DI + Be()c (F, n)N
RR
RR +
RRRR d a(I)=1 d RRR a(I)=1
R 2
d d
A RRR RRRR
D
R
(B ) + RRRR (DI ) I RR
RR
RRRa(I)=1 RRR
R R
d d
a(I)=1
where E and F range over finite subsets of K and P , Q range over finite subparti-
tions of R. Set s(G, X) = s(JGK, L (X, {0, 1})).
Since JGK L (X, {0, 1}) is a transversally generating set of the crossed product
groupoid we obtain by Theorem 9 (if the action G X is essentially free) and
[DKP2, Theorem 2.11] in general (compare [DKP2, Proposition 5.2]):
Proposition 24. s(G X) = s(G, X).
A FREE PRODUCT FORMULA FOR THE SOFIC DIMENSION 17
Remark 26. The set HA differs from the set HA introduced in [DKP2] in that we
do not assume to be a strict homomorphism. Furthermore the maps in [DKP2]
are defined on L (X) with values in Md using the 2-norm. It is often convenient to
adopt the latter point of view for computational purposes and we will do so below.
The definition of HA given above has the advantage of being purely finitary, in the
spirit of the sofic property. We have maintained the notation HA in view of (iii)
and (iv).
RR
R
2 = min RRRR ai bj RRRR , 1 , .... , a1 ..., a , T, T {1, . . . , } and
RR RRR
RRiT,jT R
pB = ai pBi , pB = bi pBi } /{0}
i=1 i=1
= min{1 , 2 , 3 }
so that depends only on F and n and 1. We want to find such that
(, ) HA(F, P, n, 9PFn 2 2 5 2 qc3 (F, n)2 , d)
1
(for sufficiently large d). Using Proposition 24 this will complete the proof.
Take a partition {A1 , . . . , Aq } such that the conclusion of Lemma 20 holds
(namely a random partition for d large). For each pBi , i = 1 . . . , let
0 (pBi ) = pAi
18 ROBERT GRAHAM AND MIKAEL PICHOT
Thus
0 (spBi ) (s)0 (pBi )2
0 (spBi ) (s)pAi 2 + (s)pAi (s)(1)pAi 2
+ (s)(1)pAi (s)0 (pBi )2
< 2c3 (F, n) + = 3c3 (F, n)
(3) Let p1 = i=1 ai pBi and p2 = i=1 bi pBi then
0 (p1 p2 ) 0 (p1 )0 (p2 )2 = ai bi (0 (pBi pBj ) pAi pAj )2
i,j=1
2 0 (pBi pBj ) pAi pAj 2
i,j=1
(4) pX = i=1
q
pBi and Id = i=1 pAi
q
so by the triangle inequality it suffices to
show
0 (pBi ) pAi 2 < 2c3 (F, n)
By Lemma 21
0 (pBi ) (1)pAi 2 < c3 (F, n)
so
0 (pBi ) pAi 2 0 (pBi ) (1)pAi 2 + (1)pAi pAi 2
< 2c3 (F, n) + < 2c3 (F, n) .
We now show how to define using 0 . For each pB
0 (pB ) pA 2 c3 (F, n)
by Lemma 21. Let M be the number of x {1 . . . , d} such that 0 (pB )x = pA x
then
0 (pB ) pA 2
dM 2
d
hence
M d(1 c3 (F, n)2 2 )
1
Similarly for pB , pB we have
0 (pB )0 (pB ) (pB )(pB )2 22 c3 (F, n)
20 ROBERT GRAHAM AND MIKAEL PICHOT
by (3) so
since
inf sE (F, n) = inf sE (F, 4nFn + 1).
n n
However
SA(Fn , 4nFn + 1, , d)E HA(F, P, n, 9PFn 2 (F,
1 5 2 2
qc 3 n) , d)E
2
so s(G) s(GX) by Proposition 24. Replacing lim sup by lim inf or limd above,
we get a similar inequality for s and s .
Theorem 28. Let G be a pmp groupoid, (X0 , 0 ) be a standard probability space,
and G X0G be the corresponding a Bernoulli action. Then s(G X0G ) = s(G).
The same holds true for s and s .
Proof. Again we only need to prove s(G) s(G X0G ). Let U = {U1 , U2 , . . . , Uq }
be any finite partition of X0 into measurable sets. Let
BU = {BU1 , . . . , BUq }
where
BUi = {x X0G e G0 , xe Ui }
Let M = {pUi } for all possible choices of U and all i, so that M is dynamically
generating *-subalgebra of L (X, ). Now if PU = {pU1 , . . . , pUq } then
s(G X) = sup sup inf inf sE,Q (F, PU ) = sup sup inf inf sE,Q (F, PU )
E Q F PU E Q PU F
Proof. Since
s(G) = s(G X0G ) s(G X0G ) = s(G)
s(G) = s(G) if and only if s(G X0G ) = s(G X0G ).
s(pGp)
1 1
s(G) + 1 .
h(p) h(p)
Furthermore the same inequality holds for s and s .
Proof. Write h(p) = N k
N
and choose s1 , s2 ...sk JGK with
i si p, h(si si ) =
1
s1 1
N
and
k
si si = 1 p.
1
i=1
Let S = {s1 , ...sk } so pGp S is generating G and let E, F JpGpK be finite subsets.
We may assume that p E F and s1 i si E F for all i. Let SApGp (F, 4n +
5, , d). In light of [DKP2, Lemma 2.13] we may also assume that N kd. Let
d = NNk d and partition {1, . . . , d } into sets A0 ...Ak with A0 = {1, ...d} and Ai = N
d
.
Since
i si )
h(s1
Fix ((s1 i si )) d d
h(p)
we may choose a subset Bi A0 with exactly d
N
elements such that
i si )) Bi
Fix ((s1 < d .
i sj = 0 i j 1
s1
ssi = s1
i s = 0 i 1, s pGp
i si F
s1
each element of n (JpGpK S)n can be written (not necessarily uniquely) as si f s1
for i, j 0 where s0 = p and f JpGpK. Let (s0 ) = pA0 We define a map
j
(F S)2n+5
Jd K
by
(si f s1
j ) = (si ) (si si f sj sj )(sj ) .
1 1 1
A FREE PRODUCT FORMULA FOR THE SOFIC DIMENSION 23
we have i1 = i2 j1 = j2 so si1 f1 s1
j1 = si1 f2 sj1 hence si1 si1 f1 sj1 sj1 = si1 si1 f2 sj1 sj1
1 1 1 1 1
so that
SAG (F S, n, , d).
Let us first see what this implies. Note if 1 (f ) 2 (f ) then 1 (f ) = pA0 1 (f )pA0 =
1 (f )
2 (f ). Next note if 1 (si ) 2 (si ) Bi Ai are two bijections then if we
have
1 (si ) = 1 (si ) (p) = 2 (si ) (p) = 2 (si )
we must have 1 (si )f ix((p)) = 2 (si )f ix((p)) . But since Fix((p)) {1, .., .d}
d we have at most JdK (d + 1)d choices for 2 such that 2 (si ) 1 (si )
and 2 (si ) = 1 (si ). Therefore there are at most (d+1)dk choices for 2 such
k
(N
k
d
!)
SApGp (F, n, , d)E SAG (F S, 4n + 5, , d )ES /
(d + 1)dk
So let us show that SAG (F S, n, , d ). We first show that is approxi-
mately linear. Suppose in the span we have
M
si f sj = si f sj ,
=1
M
i ( si f sj ) sj ) (sj )
= (si ) (s1 1 1
=1
M
= (si ) ( s1
i si f sj sj ) (sj )
1 1
=1
=1 =1
Thus if f = M
=1 f , then
RRR RRR
R R
(f ) (f ) = RRRR ( si f sj ) ( si f sj )RRRR
M N M N
=1 R
RR i,j=1 =1 i,j=1 RRR
N M
(si f sj ) (si f sj )
i,j=1 =1
RR i,j j RR
R i R
RRR RRR
R R
RRRR (ai bj ) (ai ) (bj )RRRR + 2 (150N 2 (2F S + 1)2(2n+5) )
2
s1
j1 si2 = (sj1 ) (si2 ) = 0 so we are done. Suppose otherwise, then
1
i1 si1 f1 sj1 sj1 )(sj1 ) (sj1 ) (sj1 sj1 f2 sj2 sj2 )(sj2 )
(si1 ) (s1 1 1 1 1 1
2
and for g1 , g2 F2n+5 such that g1 g2 F2n+5
then
j ) = tr((si ) (si si f sj sj )(sj ) ) = 0,
tr(si f s1 1 1 1
A FREE PRODUCT FORMULA FOR THE SOFIC DIMENSION 25
= tr ( (s1
i si f si si )(si ) (si )) (f si si )
1 1 1
tr ( (s1
i si f si si )(si ) (si ))) tr ( (si si f si si ))
1 1 1 1
(N k) (f s1
i si )
+ tr ( (s1
i si f si si ))
1
N h(p)
(2 + ) + <
where we used the computation above.
So we have proved
(N
k
d
!)
SApGp (F, 4n + 5, , d)E SAG (F S, n, , d )ES /
(d + 1)dk
So
k
Let SAG (F S, n+4, , d) and assume that d is large enough so d = h(p)d >
. Let B0 = Fix (p) so
1
B0
h(p) < .
d
Thus we may arbitrarily extend or shrink B0 to a subset B such that either B B0
or B B0 and B = d where
(p) pB (p) pFix (p) + pFix (p) pB
d PB PFix (s1
i
si ) (si ) (si ) < d6
1
( ) SApGp (F, n, , d )E
d
d , d1 , . . . , dk+1
i si ) + )
d(h(s1
k k
( ) di !
i=1 di i=1
28 ROBERT GRAHAM AND MIKAEL PICHOT
where the former term accounts for the choice of the Ai and B and the latter terms
for the choice of the Bi and (si ) respectively.
Now note that for f E
pB (f )pB (f ) (6 + 2) = 8
and similarly for i j
pAi (si )pBi (si ) pAi (si )1 (si ) + pBi (si )(si )1 + 3
(14 + 9k + 6) + (14 + 9k + 6) + + 3, for d sufficiently large
= (44 + 18k)
so we can find > 0 with 0 as 0 so that for a given choice of pB (f )pB
there are dd choices for (f ) and similarly for (si ) (see [DKP2, Lemma 2.5]) so
SAG (F S, n + 4, , d)E,S
( ) SApGp (F, n, , d )E
d
d , d1 , . . . , dk+1
i si ) + )
d(h(s1
k k
( ) (di !)ddES .
i=1 di i=1
Therefore
log SAG (F S, n + 4, , d)E,S log + log SApGp (F, n, , d )E
d!
d !dk+1 !
i si ) + )!
d(h(s1
k
+ log + dE S log d
i=1 i si ) + ) di )!
di !(d(h(s1
hence
log SAG (F S, n + 4, , d)E,S + log d ! log d! + log SApGp (F, n, , d )E + (, d).
with
i si ) + )!
d(h(s1
k
(, d) = log + dE S log d
i=1 di !(d(h(si si ) + ) di )!
1
by the previous lemma so lim supn s(qn Gqn ) s(pGp). However the two lemmas
combined imply
s(G) = h(pn )s(pn Gpn ) + 1 h(pn ).
so
lim sup s(pn Gpn ) = (s(g) 1) + 1
1
n h(p)
and
s(G) = h(qn )s(qn Gqn ) + 1 h(qn )
so
lim inf s(qn Gqn ) = (s(G) 1) + 1.
1
n h(p)
Therfore
s(pGp) = (s(G) 1) + 1
1
h(p)
and the same argument works for s and s .
So
s(G) = s(G) s(pGp) = s(pGp)
so G is s-regular if and only if pGp is s-regular.
using invariance of .
Remark 35. An analog of the so-called fixed price problem [Gab] for s is the
question of whether s(G X) depends on G only and not on the essentially free
pmp action G X. This fixed sofic dimension problem holds for example for
groupoids with fixed price 1 (=all their free pmp actions have cost 1) provided all
their free pmp actions are sofic, but it is open in general. Also open is whether a
groupoid is sofic if and only if all its free pmp actions are sofic.
Proof of Proposition 33. Since the Bernoulli action G [0, 1]G is essentially free
G [0, 1]G is a pmp equivalence relation so by Theorem 28 and [DKP1, Corollary
5.2]
s(G) = s(G [0, 1]G ) = 1 (D) = 1
1
G Ge
dh(e)
0
where D is the fundamental domain of the set of finite classes of the amenable pmp
equivalence relation G [0, 1]G .
If G X is a pmp action then the action GX [0, 1]GX is pmp and essentially
free so by Theorem 28
s(G X) = s ((G X) [0, 1]GX ) .
Now the measure isomorphism
(G X) [0, 1]GX G (X [0, 1]GX )
((s, x), y) (s, (x, y))
is an isomorphism of pmp equivalence relations, where the action G X [0, 1]GX
is diagonal and G [0, 1]GX on the first coordinate.
So
s(G X) = s (G (X [0, 1]GX ))
= 1 (D )
=1
1
G0 Ge
dh(e).
where D is the is the fundamental domain of the set of finite classes of the amenable
pmp equivalence relation G X [0, 1]GX . So
s(G) = s(G X).
Another example is the proof that the invariance of s under orbit equivalence
(Theorem 9) established in [DKP1, Theorem 4.1] is equivalent to the statement for
groupoids in [DKP2, Theorem 2.1].
Theorem 36. Let G be a pmp groupoid, E, F be transversally generating sets, then
s(E) = s(F ), s(E) = s(F ) and s (E) = s (F ).
A FREE PRODUCT FORMULA FOR THE SOFIC DIMENSION 31
References
[BCG] Biane, P., Capitaine, M., Guionnet, A. Large deviation bounds for matrix Brownian mo-
tion. Invent. Math., 152(2), 433459. (2003)
[Bow] L. Bowen. Measure conjugacy invariants for actions of countable sofic groups. J. Amer.
Math. Soc. 23 (2010), 217245.
[Bow2] , Entropy theory for sofic groupoids I: the foundations arXiv:1210.1992
32 ROBERT GRAHAM AND MIKAEL PICHOT
[BDJ] N.P. Brown, K. Dykema, and K. Jung, Free entropy dimension in amalgamated free prod-
ucts, Proc. London Math. Soc. 97 (2008), no. 2, 339367.
[Dye] H.A. Dye, On groups of measure preserving transformations. I, Amer. J. Math. 81 (1959),
no. 1, 119159.
[DKP1] K. Dykema, D. Kerr, M. Pichot, Orbit equivalence and sofic approximation,
arXiv:1102.2556
[DKP2] , Sofic dimension for discrete measure groupoids, arXiv:1111.2842
[EL] G. Elek and G. Lippner, Sofic equivalence relations, J. Funct. Anal. 258 (2010), no. 5,
16921708.
[FSZ] Feldman, J., Sutherland, C. E., Zimmer, R. J. Subrelations of ergodic equivalence relations.
Erg. Th. Dyn. Sys., 1988, 9(02), 239269.
[Gab] D. Gaboriau, Cot des relations dquivalence et des groupes, Invent. Math. 139 (2000),
no. 1, 4198.
[GL] Gaboriau, D., Lyons, R. A measurable-group-theoretic solution to von Neumanns problem.
Invent. Math., 177(3), (2009) 533540.
[Jun1] Jung, K. . A free entropy dimension lemma. Pacific J. Math. (2003), 177, 265271.
[Jun2] , The free entropy dimension of hyperfinite von Neumann algebras. Trans. AMS,
34. (2003).
[MS] Mineyev, I., Shlyakhtenko, D. Non-microstates free entropy dimension for groups, GAFA
(December 2005), 113.
[OW] D. S. Ornstein and B. Weiss. Entropy and isomorphism theorems for actions of amenable
groups. J. Analyse Math. 48 (1987), 1141.
[Oz] N. Ozawa, Hyperlinearity, sofic groups and applications to group theory, Notes from a 2009
talk available at http://people.math.jussieu.fr/pisier/taka.talk.pdf
[Pe] V. Pestov, Hyperlinear and sofic groups: a brief guide. Bull. Symbolic Logic 14 (2008), no.
4, 449480.
[Ram] Ramsay, A. Virtual groups and group actions. Advances in Mathematics, 322 (1971), 253
322.
[Shl] D. Shlyakhtenko, Microstates free entropy and cost of equivalence relations, Duke Math. J.
118 (2003), no. 3, 375426.
[Voi91] D. Voiculescu, Limit laws for random matrices and free products, Invent. Math. 104
(1991), no. 1, 201220.
[Voi96] , The analogues of entropy and of Fishers information measure in free probability
theory III: The absence of Cartan subalgebras, Geom. Funct. Anal. 6 (1996), no. 1, 172199.
[Voi98] , A strengthened asymptotic freeness result for random matrices with applications
to free entropy, Int. Math. Res. Not. 1998 (1998), no. 1, 4163.