Etvs Lornd University

Faculty of Sciences

Sofic groups
MSc Thesis

Nra Gabriella Szke

Mathematics MSc

Supervisor

Mikls Abrt
Alfrd Rnyi Institute of Mathematics, Hungarian Academy of Sciences

Budapest, 2014
Acknowledgments
I would like to thank my advisor, Mikls Abrt, for introducing me to this subject.
I am grateful for the many discussions we had, for his helpful suggestions, and for
his patience with me.
I would also like to thank Bandi Szab for his immense help with the proof of
Theorem 2.2.
Table of contents 1

Contents

Introduction 2

1 Definitions and basic examples 3

1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Two characterizations of soficity . . . . . . . . . . . . . . . . . 4
1.1.2 Other approaches . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Some examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 The class of sofic groups 10

2.1 Free products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Amalgamated products over an amenable subgroup . . . . . . . . . . 18
2.3.1 Quasi-tilings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.2 The proof of the theorem . . . . . . . . . . . . . . . . . . . . . 26

3 Surjunctivity conjecture 34

Bibliography 37
2 Introduction

Introduction

The idea of sofic groups (originally: initially subamenable groups) was intro-
duced by Gromov ([8]) in 1999, as a common generalization of residually finite and
amenable groups. The name sofic, from the Hebrew word meaning finite, was
coined by Weiss ([11]).
Sofic groups are of considerable interest because several important general con-
jectures of group theory were shown to be true for them. Among these, we discuss
Gottschalks surjunctivity conjecture, that was proved by Gromov in [8].
Possibly the biggest open question in the subject is whether all countable groups
are sofic. So far we know that the class of sofic groups is closed under taking sub-
groups, direct products, inverse limits, direct limits, free products, extensions by
amenable groups ([5]), and amalgamation over amenable subgroups ([6], [10]).
The goal of this thesis is to summarize some results about sofic groups and to
clarify a proof of Elek and Szab about the amalgamated products of sofic groups
over amenable subgroups. The original proof in [6] contains an error and is incom-
plete. An alternate proof, using von Neumann algebras, appears in [10].
The structure is as follows: In the first chapter, we define sofic groups, give some
useful and interesting characterizations for them, and review some basic examples.
The second chapter contains the theorems about the class of sofic groups. Finally, in
the third chapter we present a proof of the surjunctivity conjecture for sofic groups.
Chapter 1: Definitions and basic examples 3

Chapter 1

Definitions and basic examples

Soficity has many different definitions and characterizations in several papers. In

this chapter we introduce some of them.

1.1 Definition
Our definition of sofic groups follows [5].

For a finite set A let Map(A) denote the monoid of self-maps of A acting on the
right. We write a f for f (a) and multiplication in Map(A) works as usual: a f g =
(a f ) g = g(f (a)). Let (0, 1) be a real number, then we say that two elements
e, f Map(A) are -similar, or e f , if the number of points a A with ae 6= af
is at most |A|. We say that e, f are (1 )-different if they are not (1 )-similar.
Let G be a group, (0, 1) a real number and F G a finite subset. An (F, )-
quasi-action of G on a finite set A is a function : G Map(A) with the following
properties:
(a) For any two elements e, f F the map (ef ) is -similar to (e)(f ).

(b) (1) is -similar to the identity map of A.

(c) For each e F \ {1} the map (e) is (1 )-different from the identity map of
A.
Definition 1.1. The group G is sofic if for all (0, 1) and all finite subsets F G
there exists an (F, )-quasi-action of G.
We can ask for more than this:
Lemma 1.2. If a group G is sofic then for each F G and for each (0, 1)
there is an (F, )-quasi-action of G on a finite set A satisfying the following extra
conditions:
4 Chapter 1: Definitions and basic examples

(b) (1) is the identity map of A, for each 1 6= g G the map (g) is a fixpoint
free bijection and (g 1 ) = (g)1 .

(c) For different elements e, f F {1} the map (e) is (1 )-different from
(f ).

It is clear from the definition that a non-finitely generated group is sofic if and only
if all of its finitely generated subgroups are sofic. Because of this, from now on we
will mostly consider finitely generated groups.

1.1.1 Two characterizations of soficity

We will use the following two characterizations of sofic groups:

1. For the definition used by Weiss in [11], we need the notion of Cayley graphs.
Let G be a finitely generated group, and S G a fixed finite, symmetric (i.e.
S = S 1 ) generating set. The Cayley graph of G is a directed graph Cay(G, S),
whose edges are labeled by the elements of S: the set of vertices equals G, and
the edges with label s S are the pairs (g, sg) for all g G. Let Br (1) denote
the r-ball around 1 Cay(G, S) (it is an edge-colored graph, and also a finite
subset in G).

Proposition 1.3. A finitely generated group G is sofic if there exists a finite,

symmetric generating set S, such that for each > 0 and each r N there is
a finite directed graph (V, E) edge-labeled by S, and a subset V0 V with the
following properties:

1. For each point v V0 there is a function v : Br (1) V which is an

isomorphism (of labeled graphs) between Br (1) Cay(G, S) and the r-ball
around v in V .
2. |V0 | (1 )|V |.

We call this finite graph V an (r, )-approximation of the Cayley graph.

Moreover, if G is sofic, then all finite, symmetric generating sets have this
property.

The above proposition, i.e. the equivalence of the two definitions was proven
in [4].

2. Let G be a finitely generated group: G = hSi, where S = S 1 and |S| < .

Let FS denote the free group generated by S, and f : FS G the factor map.
Chapter 1: Definitions and basic examples 5

Proposition 1.4. A finitely generated group G is sofic if there is a finite,

symmetric generating set S such that for every > 0 and k N there exists
n N, a homomorphism : FS Sn and a subset A [n] ([n] denotes the
set of integers from 1 to n, Sn acts on [n]), |A| > (1 )n, such that if w FS
and the length of w is at most k then for every x A, x(w) = x if and only
if f (w) = 1 in G. In this case we call this a (k, )-quasi-action.
If G is sofic then this is true for every generating set S.

Proof. First suppose we have a set S with this property. Let (0, 1) and
F G, |F | < be arbitrary. We can assume that 1 F . Since S is a
generating set, we can write every element of F as a product of elements of S.
Consider a shortest such word for each element in F F , and take the maximum
of the length of these words, where F F = {ef | e, f F }. Since F is a finite
set, this is a finite number, let us denote it by t.
We can find a (k, )-quasi-action of G, where k = 2t + 1. Call this quasi-action
: FS Sn . Let A [n] be the set described in the proposition. We define
: F F Map([n]) the following way. For e F F and x A, then we can
find w FS such that f (w) = e and the length of w is at most t. Define

x(e) = x(w).

This is well-defined because if w0 FS is another word that has length at most

t and f (w0 ) = e, then w0 w1 has length at most k and f (w0 w1 ) = 1 hence
x(w0 w1 ) = x, i.e. x(w0 ) = x(w1 )1 = x(w).
This way we defined (e) on the elements of A, on the rest of the elements we
can define it to be the identity. For g
/ F F let (g) be the identity. It is easy
to see that this : G Map([n]) is an (F, )-quasi-action of G.
For the other direction let G be a finitely generated sofic group and S a finite,
symmetric generating set. We would like to find a (k, )-quasi-action for kN
and > 0.
Let W FS be the set of words of length at most k, and let f (W ) = F G.
This F is clearly finite, denote the size of F by m. Let

= ,
m2 +m
and let : G Map(A) be an (F, )-quasi-action of G.
Then for each e1 , e2 F there is a subset Ae1 ,e2 A, |Ae1 ,e2 | > (1 )|A|
on which (e1 e2 ) = (e1 )(e2 ). Similarly, for every e F \ {1} there exists
6 Chapter 1: Definitions and basic examples

Ae such that |Ae | > (1 )|A| and (e) has no fixpoints in Ae . Let A0 be the
intersection of all these Ae1 ,e2 s and Ae s. Clearly
X X
|A0 | > |A| |A\Ae1 ,e2 | |A\Ae | (1m2 m)|A| = (1)|A|.
e1 ,e2 F eF \{1}

On this intersection the above statements hold simultaneously for all elements
of F .
For s S FS let us define (s)|A0 = (f (s))|A0 and extend it to A as a
bijection. It is possible since by the definition of A0 (e) is injective on A0 for
e F . Let |A| = n and identify A with [n]. This way we defined : S Sn ,
which gives a homomorphism : FS Sn . One can easily check that this
is a (k, )-quasi-action.

1.1.2 Other approaches

3. For two permutatuions , Sn their normalized Hamming distance dist(, )
is defined to be the number of points not fixed by 1 , divided by n. Clearly
we can reformulate the definition of sofic groups using this concept:
A group G is sofic if for all (0, 1) and every finite subset F G there is
some n N and a map : G Sn such that

for every g F \ {e}, dist((g), e) > 1 , where e is the identity element

of G,
for all g, h F , dist((g 1 h), (g)1 (h)) < .

The following is another useful characterization: Given positive integers n(k)

we denote by N the normal subgroup of
Q
k=1 Sn(k) consisting of all sequences
(k )
k=1 such that limk dist(k , idn(k) ) = 0, where idn(k) is the identity ele-
ment of the permutation group Sn(k) .

Proposition 1.5 (From [2]). Let G be a finitely generated group. Then G

is sofic if and only if for some sequence of integers n(k) there is a group
homomorphism

!
Y
:G Sn(k) /N,
k=1

given by (g) = [(k (g))

k=1 ] for some maps k : G Sn(k) such that

lim dist(k (g), idn(k) ) = 1

k

for all nontrivial elements g G.

Chapter 1: Definitions and basic examples 7

4. Let G be a finitely generated group, and let Sub(G) denote the set of subgroups
endowed with the Chabauty topology, which is just the product topology on
the set of subsets of G restricted to the set of subgroups. This is a compact
topological space, G acts on it by conjugation, and this action is by homeo-
morphisms.
An invariant random subgroup (IRS) of G is a random subgroup of G whose
distribution is a G-invariant probability measure on Sub(G). For example if
N / G then the Dirac measure N is an IRS. If H G, |G : H| < , then H
has finitely many conjugates, let H denote the uniform probability measure
on these conjugates. This is also an IRS.
Recall that a sequence {n }nN of probabilitiy measures weak? converges to
if for every real valued continuous function f we have
Z Z
f dn f d.

Proposition 1.6 (From [1]). Let G be a finitely generated group, G

= F/N ,
where F is a free group, N / F . Then G is sofic if and only if there exists
?
Hn F for every n N such that |F : Hn | < and Hn N .

1.2 Some examples

Let be a finitely generated group, S a finite, symmetric generating set.
Consider the Cayley graph Cay(, S). For a subgraph X of this graph let X denote
the boundary of X, i.e., the set of vertices that have at least one neighbor in the
complement of X.
A Flner sequence in the Cayley graph is a sequence of spanned subgraphs F =
{F1 , F2 , . . . } such that

[
Fn = Cay(, S),
n=1
|Fn |
lim = 0.
n |Fn |

Definition 1.7. A finitely generated group is amenable if one of its Cayley graphs
admits a Flner sequence. Note that this property does not depend on the generating
set.

Proposition 1.8. Let be a finitely generated amenable group. Then is sofic.

Proof. We choose a finite, symmetric generating set S and a Flner sequence in

Cay(, S), F = {F1 , F2 , . . . }. Let us fix and r. We would like to find a finite graph
described in Proposition 1.3.
8 Chapter 1: Definitions and basic examples

If |S| = 2d, then the Cayley graph is d-regular. We know that the Fn s have small
boundary for large enough n. Let

Fn0 = Cay(, S) \ Nr (Cay(, S) \ Fn ),

where Nr (X) denotes the r-neighborhood of a subgraph X. This means that we have
omitted those points from Fn that are close to its complement. All the remaining
points have the property that their r-neighborhood in Fn is isomorphic to an r-ball
in the Cayley graph. Clearly

|Fn0 | = |Fn | |Fn Nr (Cay(, S) \ Fn )| > |Fn | |Nr1 (Fn )| |Fn | |Fn |dr1 .

So  
r1 |Fn |
|Fn0 | > 1d |Fn |.
|Fn |
|Fn | 1
This means if we choose n such that |Fn |
< dr1 , then

|Fn0 | > (1 )|Fn |,

which is exactly what we wanted.

Remark 1.9. Generally we say that a group is amenable is for every > 0 and for
each finite set F there is a finite subset A such that F does not move A
too much, i.e.
|Ag \ A| < |A| for every g F.

Notice that for finitely generated groups, this is equivalent to the definition we used
above.
If G is not necessarily finitely generated, then G is amenable if and only is all of its
finitely generated subgroups are amenable. It follows that every amenable group is
sofic.

There is another large class of groups for which it is easy to see that they are
sofic: residually finite groups. Recall that a group G is residually finite if for every
1 6= g G there is a normal subgroup N / G, |G : N | < such that g
/ N . This is
equivalent to
\
N = {1}.
N /G,|G:N |<

Proposition 1.10. Residually finite groups are sofic.

Proof. Let G be residually finite and F G an arbitrary finite set. This means
we can find a normal subgroup N / G, |G : N | < such that the factor map
Chapter 1: Definitions and basic examples 9

: G G/N is injective on F F F {1}. Indeed, for each pair g, h F F F {1}

there is Ng,h such that g 1 h
/ Ng,h . Let N be the intersection of these.
Now let A = G/N , which is a finite set and let (g) be the right multiplication on
G/N by (g). This way clearly : G Map(A) and conditions (a)-(c) in Definition
1.1 hold for any (0, 1).
10 Chapter 2: The class of sofic groups

Chapter 2

The class of sofic groups

In this chapter we prove the following two theorems about the class of sofic groups.

Theorem 2.1 (Elek-Szab, [5]). The class of sofic groups is closed under the fol-
lowing constructions:

1. subgroups, direct products, inverse limits, direct limits;

2. free products

3. certain extensions: if N / G, N is sofic and G/N is amenable then G is also

sofic.

Theorem 2.2 (Elek-Szab, Punescu, [6] and [10]). Free products of sofic groups
amalgamated over arbitrary amenable subgroups are sofic.

Proof of part 1. of Theorem 2.1 by [5]. Suppose that G is sofic and H G. Let
F H finite and (0, 1). Then of course F G as well, and an (F, )-quasi-
action of G is an (F, )-quasi-action of H too. So H is sofic.
Next let G1 and G2 be sofic groups. We would like to prove that G1 G2 is sofic as
well. Let (0, 1), F G1 G2 be a finite subset, then we can choose finite sets
Fi Gi (i = 1, 2) such that F F1 F2 . Since G1 and G2 are sofic, there exist
(Fi , 2 )-quasi-actions i : Gi Map(Ai ). Let A = A1 A2 , and define : G1 G2
Map(A) the following way:

(a1 , a2 ) (g1 , g2 ) = (a1 1 (g1 ), a2 2 (g2 )).

It is easy to see that is an (F, )-quasi-action of G1 G2 . Iterating this argument

gives us that finite direct products of sofic groups are also sofic.
Q
Now let G = iI Gi where {Gi }iI are sofic groups. For a finite subset F G there
Q
exists a finite J I such that the natural projection : G iJ Gi is injective on
Chapter 2: The class of sofic groups 11

Q
F {1}. Hence all ((F ), )-quasi-action of iJ Gi gives us an (F, )-quasi-action of
G. This is just a finite direct product, and for that we already proved the statement.
An inverse limit of groups is by definition a subgroup of their direct product, so that
part of the theorem follows from the previous ones.
Assume we have a directed system of sofic groups {Gi }iI , let G = limiI Gi . Fix
(0, 1) and a finite set F G. Then there is an index i and a finite subset Fi Gi
such that the natural homomorphism i : Gi G is a bijection Fi F .
Denote by G b Gi that has the property
b the image of Gi in G. Choose a map s : G
that i (s(g)) = g for all g G.
b Let i : Gi Map(A) be an (Fi , )-quasi-action of
Gi . Then define : G Map(A) as follows. For a A let

a i (s(g)), if g G
b
a (g) =
a, if g G \ G.
b

This is clearly an (F, )-quasi-action of G, hence G is sofic.

2.1 Free products

In this section we present a proof for part 2. of Theorem 2.1 that uses a different
method as the one in [5]. We will use Proposition 1.4, the second characterization
of sofic groups mentioned in the first chapter.
Let G = hSi and H = hS 0 i be finitely generated sofic groups, f : FS G and
f 0 : FS 0 H the factor maps. If : FS Sn and : FS 0 Sn are (k, )-quasi-
actions of G and H respectively, then for Sn let us define b : FSS 0 Sn in
the following way: for s S b (s) = (s) and for s S 0 b (s) = 1 (s), and
extend this to a homomorphism from FSS 0 (which agrees with on FS and with
on FS 0 ).

Proposition 2.3. For arbitrary 0 < and k N, we can choose and n0 such that
if n > n0 and : FS Sn and : FS 0 Sn are as above, then the following holds.
There exists Sn such that b is a (k, )-quasi-action of G H.

Proof. Let f denote the factor map from FSS 0 to G H and let us define the set of
good points in [n] for a word w FSS 0 :

A,w = {x [n] : xb (w) = x if and only if f(w) = 1}.

The set good points for all words of length at most k is denoted by
\
A = A,w .
l(w)k
12 Chapter 2: The class of sofic groups

Our goal is to choose the parameters such that |A | > (1 )n.

We choose from Sn uniformly randomly. If the above holds in expected value,
i.e. E(|A |) > (1 )n, then we can find a suitable . An estimate on the expected
value is
X X
E(|A |) = n E(|[n] \ A |) n E(|[n] \ A,w |) = n (n E(|A,w |)).
l(w)k l(w)k

This means that we have to find a lower bound for E(|A,w |). For the identity
element A,1 = [n] for every , so any lower bound works here. Let us fix w FSS 0 ,
where w 6= 1 and has length at most k. The word w has the following form: w =
g1 h1 g2 h2 . . . gl hl , where gj FS and hj FS 0 for each j, and for j 6= 1 gj 6= 1 and
j 6= l hj 6= 1.
There are two cases: The first one is that no f (gi ) or f 0 (hi ) equals the identity
element of G or H (except when g1 or hl is the empty word in FSS 0 ). The second
is that this doesnt hold, which means that w is not one of the shortest words in
f1 (f(w)).

1. In this case A,w is the set of non-fixpoints of b (w). Assume that g1 6= 1 and
hl 6= 1.

b (w) = b (g1 h1 g2 h2 . . . gl hl ) = (g1 ) 1 (h1 )(g2 ) 1 (h2 ) . . . 1 (hl )

We call the sequence of points x, x(g1 ), x(g1 ) 1 , x(g1 ) 1 (h1 ),

x(g1 ) 1 (h1 ), x(g1 ) 1 (h1 )(g2 ), . . . , x(g1 ) . . . 1 (hl ) the trajec-
tory of x. Let

B,w = {x [n] : the trajectory of x consists of different points in [n]}.

Clearly B,w A,w , so E(|A,w |) E(|B,w |).

X
E(|B,w |) = P(x B,w )
x[n]

If no f (gi ) or f 0 (hi ) is the identity element of G or H, then clearly no (gi ) or

(hi ) is the identity in Sn , because and are (k, )-quasi-actions and the
length of gi and hi is at most k. So B,w 6= . We will find a lower bound for
P(x B,w ).
If x is a fixpoint of (g1 ) = b (g1 ) then this probability is zero, but this happens
at most in n cases, because is a (k, )-quasi-action. If x(g1 ) = x1 6= x, then
let (gi ) = 2i1 and (hi ) = 2i , so we can write

b (w) = 1 1 2 3 1 4 . . . 2l .
Chapter 2: The class of sofic groups 13

Let x0 = x, and for 1 t l let x4t3 = x1 1 . . . 2t1 , x4t1 =

x1 1 . . . 1 2t , x4t2 = x1 . . . 2t1 1 , x4t = x1 . . . 2t . This sequence
is the trajectory of x. Let Em denote the event that x0 , x1 , . . . , x2m1 are all
different.
We estimate P(Em+1 |Em ) as follows. Suppose Em happens, let 0 denote if
m is even and 1 if m is odd (since the distribution of 1 is the same as
of , the calculations are the same in both cases). Then x2m = x2m1 0 and
x2m+1 = x2m1 0 m+1 . We choose 0 uniformly randomly, so the images of any
tuple of different points are independent. In our case so far we know the image
of m 1 points, which implies that x2m1 0 is uniformly distributed among
the other n (m 1) points. However, there are some bad choices when Em+1
doesnt hold.

1. If x2m = xi for 0 i 2m 1.
2. When x2m is a fixpoint of m+1 , hence x2m+1 = x2m . There are at most n
such points. (Because m+1 6= id.)
1
3. If x2m+1 = xi for 0 i 2m 1, so xi m+1 are also bad points.

Summing these up we get that there are at most 4m + n bad points. In the
case when m + 1 = 2l there are no bad points of the second and third kind,
but if we subtract those as well, we still get a lower bound. So the probability
we are interested in is
n (m 1) (4m + n) n 5m n (1 )n 10l
P(Em+1 |Em ) .
n (m 1) n n
The last inequality holds because m 2l.
Since Em+1 Em ,
P(Em+1 )
P(Em+1 |Em ) = .
P(Em )
Using this we get a lower bound on E2l :
2l1 2l1
Y Y (1 )n 10l
P(E2l ) = P(E1 ) P(Ei+1 |Ei ) (1 ) =
i=1 i=1
n
 2l1
(1 )n 10l
= (1 ) ,
n
where we used the fact that E1 is the event that x 6= x1 . But E2l is the event
that the trajectory of x consists of different points, so
 2l1  k+1
(1 )n 10l (1 )n 5k 10
P(x B,w ) (1) (1) .
n n
14 Chapter 2: The class of sofic groups

The last inequality holds because the length of w is at most k so k 2l 2.

So in this case we have
 k+1
X (1 )n 5k 10
E(|A,w |) E(|B,w |) P(x B,w ) n(1 ) .
n
x[n]

In the case when g1 = 1 or hl = 1, the calculations work similarly and we get

slightly higher lower bounds for the probability.

2. Suppose w is of the second type. If for some i, f (gi ) = 1 in G, then w0 =

g1 h1 . . . hi1 hi gi+1 . . . gl hl , a shorter word in FSS 0 , goes to the same element
of G H. We can say even more: is a quasi-action and the length of gi is at
most k, f (gi ) = 1, so (gi ) = id Sn . Hence b (w0 ) = b (w). This is also true
when f 0 (hi ) = 1 in H for some i, we can delete hi as well.

Iterate this step until we get a word w which has the form (g1 )h1 g2 . . . gt (ht ),
where none of the (gi )s and (hi )s are equal to the identity. We can get
here in finitely many steps since the length of the word decreases in each step.

In other words, this w belongs to the first case. So we already have a lower
bound for E(|A,w |). As we saw, these steps do not change the b -image of
our word, so b (w) = b (w). This means that the same lower bound works for
E(|A,w |) too.

We showed that the above holds for each w of length at most k. From this we get
X
E(|A |) n (n E(|A,w |))
l(w)k

 k+1
X (1 )n 5k 10
n (n n(1 ) ),
n
l(w)k

 k+1 !
X (1 )n 5k 10
E(|A |) n 1 1 (1 ) .
n
l(w)k

Let r denote the number of words in FSS 0 that have length at most k. We could
calculate precisely but all that matters is that r is some finite number, that depends
only on |S S 0 | and k. Let 0 < /r.
Let  k+1
(1 )n 5k 10
(1 ) = an, .
n
As n goes to infinity, an, (1 )k+2 . Choose such that (1 )k+2 > 1 0 /2.
This way there exists n0 such that if n > n0 , then an, > 1 0 . So for n > n0 we
Chapter 2: The class of sofic groups 15

have

X X
E(|A |) n 1 (1 an, ) > n 1 0 = n(1 r0 ) > n(1 ).
l(w)k l(w)k

The proof of the proposition is complete.

Theorem 2.4. If G and H are sofic groups then their free product G H is also a
sofic group.

Proof. First suppose that G and H are finitely generated.

Fix some > 0 and k N. We want to find a (k, )-quasi-action of G H using the
method described in the previous proposition. According to Proposition 2.3 we can
choose > 0 such that for each large enough n there exists a good Sn . The only
things missing are a (k, )-quasi-action of G and a (k, )-quasi-action of H on
n points for some n > n0 .
Since G and H are sofic, there are (k, )-quasi-actions for them on n1 and n2 points
respectively.
Notice that if we have a quasi-action of a group on n points, then we have one on
any multiple of n: if we have mn points, divide it to m classes of n elements, and
we use the original action on every class.
So there are (k, )-quasi-actions of G and H on n1 n2 m points for every m N.
Choose m such that n = n1 n2 m > n0 . For this n there exist the needed quasi-
actions so we can find a good and a (k, )-quasi-action of G H on n points.
If G and H are not necessarily finitely generated, then the free product G H is the
direct limit of {G H }, where G and H are finitely generated groups. Since the
direct limit of sofic groups is sofic, G H is a sofic group.

2.2 Extensions
Proposition 2.5. Let N be a sofic group, and let N /G such that G/N is amenable.
Then G is sofic.

Remark 2.6. Before we begin the proof, let us get a picture of what we would like
to do. Consider the finitely generated case, when N = hXi, G = hX Y i such that
G/N = h(Y )i where : G G/N is the factor map. Fix the numbers r N and
> 0.
Because of the soficity of N we can choose an approximation of Cay(N, X), call it
B. SinceG/N is amenable, we can find a subgraph A Cay(G/N, (Y )) with small
boundary.
We construct a graph in the following way:
16 Chapter 2: The class of sofic groups

The vertex set equals V (A) V (B).

Into each copy of V (B) we can draw the graph B, these will be the X-colored
edges.

Now let {a1 } B and {a2 } B be two copies of this B such that a1 a2 is
a y-labeled edge for some y Y . First draw an y-edge between (a1 , b) and
(a2 , b) for some b B. This one edge determines the other y-colored edges
between {a1 } B and {a2 } B. Denote by b g the endpoint of the path
corresponding to g N starting from b. Then draw a y-edge between (a1 , b g)
and (a2 , b y 1 gy).)

The previously described strategy works for the edges of a spanning tree of A.
After that if we take an edge a1 a2 then this gives us a cycle. This means that we
have to be careful when drawing the first edge between {a1 } B and {a2 } B,
because that cycle in Cay(G/N, (Y )) might not be a cycle in Cay(G, X Y ).
However, it does us a well-defined element in N . This determines where we
should draw the first edge, and after this, the procedure is the same.

This method gives us a finite graph edge-labeled by X Y . It turns out that we can
choose these approximations to be so close to the corresponding Cayley graphs, so
that this final graph that we constructed is an (r, )-approximation.
What we described was the main idea of the proof, but in this graph-picture the
calculations would be much more complicated. So we present a combinatorial proof.

Proof of Proposition 2.5. Let F G be a finite set and (0, 1). We aim to find
an (F, )-quasi-action of G.
Denote by : G G/N the factor map, and choose a section of this homomorphism:
: G/N G, such that ((h)) = h for each h G/N . This is equivalent to saying
that g((g))1 N for all g G.
By the amenability of G/N we can find A G/N such that for each h (F ) we
have

|Ah \ A| < |A|.
3
Let A = (A) (so (A) = A) and H = (A F A1 ) N N . Clearly |H| < so
we can find an (H, /3)-quasi-action of N , : N B.
Now we define a map : G Map(A B) as follows. Note that this is the point
where we use the idea described before the proof,


((ag)), b (ag((ag))1 )
, if (ag) (A)
a, b (g) =
a, b
otherwise.
Chapter 2: The class of sofic groups 17

This definition makes sense because ag((ag))1 N . We will prove that this
is an (F, )-quasi-action of G on the set A B. Let us check conditions (a)-(c) of
Definition 1.1.

(b) If g = 1 then (a1) A, so

a, b (1) = ((a)), b (a((a))1 ) = a, b (1) ,

here (1) is /3-similar to the identity map of B, hence (1) is /3-similar to

the identity map of A B.

(c) Let e F \ {1}. For this e there are at most |A|/3 elements a A such that
(ae) = (a)(e)
/ A.
If e
/ N then (e) 6= 1 in N . Suppose that (ae) = (a)(e) A for some
a A. Then obviously (ae) 6= (a), hence ((ae)) 6= ((a)) = a. So
in this case, for any b B, (a, b) cannot be a fixpoint of (e), i.e., (e) is
(1 /3)-different from the identity map of A B.
Now consider the case when e N . Then (e) = 1. So (ae) = (a) A for
every a A, and

a, b (e) = ((a)), b (ae((ae))1 ) = a, b (aea1 ) .

Here aea1 H and aea1 6= 1, this means that (aea1 ) is (1 /3)-different

from the identity map of B, hence (e) is (1 /3)-different from the identity
map of A B.

(a) Let e, f F . As before, there are at most |A|/3 elements of A where (ae)
/
A, and the same amount for f . This means that altogether we have at most
2|A|/3 wrong elements in A.
Now consider those a A when (ae) A and (af ) A. In this case

a, b (e) (f ) = ((ae)), b (ae((ae))1 ) (f ) =

= ((((ae))f )), b (ae((ae))1 ) (((ae))f ((((ae)))f )1 ) .

Here
((((ae))f )) = ((((ae)))(f )) =

= ((ae)(f )) = ((aef )).

So using the notations g1 = ae((ae))1 and g2 = ((ae))f ((aef ))1 , we

get


a, b (e) (f ) =
18 Chapter 2: The class of sofic groups

= ((aef )), b (ae((ae))1 ) (((ae))f ((aef ))1 ) =

= ((aef )), b (g1 ) (g2 ) .

On the other hand, we have

a, b (ef ) = ((aef )), b (aef ((aef ))1 ) = ((aef )), b (g1 g2 ) .

As we can see, the first coordinate is the same, and the second coordinates
are equal too if b (g1 ) (g2 ) = b (g1 g2 ). Since g1 , g2 H and is an
(H, /3)-quasi-action, this is true for at least (1 /3)|B| values of b.
We got that there are at most 2|A|/3 wrong elements in A and at most |B|/3
wrong elements in B for each good a A. For all the other pairs (a, b) we have



a, b (e) (f ) = a, b (ef ).

So (e)(f ) is -similar to (ef ).

We proved that is an (F, )-quasi-action, hence G is a sofic group.

2.3 Amalgamated products over an amenable sub-

group
Definition 2.7. Let G be a finitely generated group with a symmetric generating
set S. Let B be a finite graph, such that each directed edge of B is labeled by
an element of S. We say that B is an r-approximation of the Cayley graph of G,
Cay(G, S) if there exists a subset W V (B) such that

|W | > 1 1r |V (B)| and

if p W then the r-neighborhood of p is rooted isomorphic to the r-

neighborhood of a vertex of the Cayley-graph (as an edge labeled graph).

In other words, B is an r-approximation if it is an r, 1r -approximation of the Cayley

graph.

Proposition 2.8. The group G is sofic if and only if for any r 1 there exists an
r-approximation of Cay(G, S) by a finite graph.
1
Proof. If G is sofic, then using Proposition 1.3 for = r
and r provides us such a
finite graph.
For the other direction choose R max{r, 1 }, this way an R-approximation is good
for r and in Proposition 1.3, proving that the group G is sofic.
Chapter 2: The class of sofic groups 19

Definition 2.9. Let E(A) denote the set of edges in a colored graph A. We say that
two colored graphs A and B are r-isomorphic for some r > 0 if there are subgraphs
A0 A and B 0 B such that
   
0 1 0 1
|E(A )| 1 |E(A)|, |E(B )| 1 |E(B)|
r r
and A0 is isomorphic to B 0 (as colored graphs). The isomorphism between A0 and
B 0 is called an r-isomorphism.

Lemma 2.10. Let G be a sofic group, and S a finite, symmetric generating set of G.
Then for each integer r > 0 there exists an integer Rapp (r) which has the following
property. Whenever a finite graph B is an Rapp (r)-approximation of Cay(G, S) and
A is Rapp (r)-isomorphic to B, then A is an r-approximation of Cay(G, S).

Proof. Suppose that for some R > 0 A and B are R-isomorphic and B is an R-
approximation of the Cayley graph. Then there are B 0 B and A0 A subgraphs
such that B 0
= A0 and |B 0 | > 1 R1 |B|, |A0 | > 1 R1 |A|. There is also B



1
B which is small: |B| < R
|B|, and each point which is in B \ B has a good R-
neighborhood in B.
As before, for a subgraph X let Nr (X) denote the r-neighborhood of X. Let

A00 = A \ Nr (A \ A0 ) A0 .

Let d = |S|, so our graphs are d-regular. Since |A\A0 | < |A|/R, clearly |Nr (A\A0 )| <
dr |A|/R. So
dr
 
00
|A | > 1 |A|.
R
Let : B 0 A0 be an isomorphism. Let A = (B B 0 ), the -image of the bad
points in B. Consider the subgraph A00 \ A. Its r-neighborhood is still isomorphic
to a subgraph of B and its 1 -image in B is disjoint from B. This means that the
r-ball around any point of A00 \ A is rooted isomorphic to an r-ball in the Cayley
graph.
We also need a lower bound for the size of this subgraph. Clearly
1 1 |B 0 | 1 1 1
|A| |B| < |B| < 1

= |B 0 | = |A0 | |A|.
R R 1 R R1 R1 R1
Hence
dr
 
00 1
|A \ A| > 1 |A|.
R R1
We can choose R = Rapp (r) such that
dr 1 1
+ < ,
R R r
so the proof of the lemma is complete.
20 Chapter 2: The class of sofic groups

Definition 2.11. Let T = (T1 , T2 , . . . , Tm ) be a finite sequence of colored graphs.

Their linear combination with coefficient vector = (1 , . . . , m ) Nm , denoted by
T , is the disjoint union of 1 copies of T1 , 2 copies of T2 , . . . , m copies of Tm .
As a special case, if m = 1 then we talk about the integer multiples of T1 .

Remark 2.12. Let be a finitely generated amenable group, S a finite symmetric

generating set. Then there exists a Flner sequence in the Cayley graph Cay(, S),
let us denote it by F = {F1 , F2 , . . . }. It is easy to see that for all r > 0, all but
finitely many of the Fn are r-approximations of G.

2.3.1 Quasi-tilings
Let us recall the notion of quasi-tilings from [9].
Let X be a finite set and {Ai }ni=1 be finite subsets of X. We say that {A1 , A2 , . . . , An }
are -disjoint if there exist subsets Ai Ai such that

For any 1 i n, |Ai |/|Ai | 1 .

If i 6= j then Ai Aj = .

On the other hand, if {Hj }m

j=1 are finite subsets of X, we say that they -cover X if

| m
j=1 Hj |
.
|X|
Definition 2.13. Let be a finitely generated amenable group with a symmetric
generating set S and let 1 F1 F2 . . . ,
n=1 Fn = be a Flner-exhaustion.
Let Br (x) denote the r-neighborhood of the vertex x in a graph. Let B be a finite
graph edge-labeled by the elements of S, and let L be a natural number. Let

QB
L = {x B : BL (x) ' BL (1) Cay(, S) as edge-labeled graphs}.

Let {Fn1 , Fn2 , . . . , Fns } be a finite collection of the Flner sets above such that for
any 1 i s, Fni BL/2 (1). Then for any x QB
L and 1 i s, Tx (Fni ) is the
image of Fni under the unique colored isomorphism BL (1) BL (x) mapping 1 to
x. We call such a subset a tile of type Fni and say that x is the center of Tx (Fni ). A
system of tiles -quasi-tile V (B) if they are -disjoint and form an (1 )-cover.

Proposition 2.14 (Theorem 2 from [3]). For any > 0, n > 0, there exist L > 0,
> 0 and a finite collection {Fn1 , Fn2 , . . . , Fns } BL (1) of Flner sets such that
ni > n and if
|QBL|
>1
|V (B)|
then V (B) can be -quasi-tiled by tiles of the form Tx (Fni ), x QB
L , 1 i s.
Chapter 2: The class of sofic groups 21

Definition 2.15. If T = {F1 , . . . , Fm } is a sequence of Flner sets and Nm a

nonzero lattice vector, then we say that a finite graph can be -quasi-tiled by T
if it can be -quasi-tiled by the union of 1 tiles of type T1 , 2 tiles of type T2 , . . . ,
m tiles of type Tm .
Suppose there exists some Nm such that i i for every 1 i m, |V (T )| <
|V ( T )| and a finite graph B can be -quasi-tiled by ( ) T . In this case we
say that T almost -quasi-tiles B.

Proposition 2.16. If a finite graph B is 1/2r-quasi-tiled

almost by
{A1 , A2 , . . . , An }, then B is r-isomorphic to the disjoint union ni=1 Ai .
S

1
Proof. We can assume that for m n {A1 , . . . , Am } 2r
-quasi-tiles B, and
m   n
X 1 X
|Ai | > 1 |Ai |.
i=1
2r i=1

Let Ai Ai for 1 i m such that |Ai | > (1 1/2r)|Ai | and Ai Aj = if i 6= j.

So we have a map : m
S
i=1 Ai B that is injective. Here

m   m  2 Xn   n
X 1 X 1 1 X
|Ai | > 1 |Ai | > 1 |Ai | > 1 |Ai |,
i=1
2r i=1 2r i=1
r i=1

and m   m  2  
X 1 X 1 1
|Ai | > 1 |Ai | > 1 |B| > 1 |B|.
i=1
2r i=1 2r r
So they are indeed r-isomorphic.

The following is a variation of Lemma 2.5 from [6].

Definition 2.17. For a vector 0 6= Rm we define the unit vector () = /kk

where kk denotes the (usual) length of . Also, for an arbitrary = (1 , . . . , m )
let bc = (b1 c, . . . , bm c) Zm .

Lemma 2.18. Let T = {T1 , T2 , . . . , Tm } be a finite sequence of colored (non-empty)

graphs. For each (0, 1) there is an integer M () > 0 (also depending on T )
such that whenever , Nm are nonzero lattice vectors with kk M ()kk and
k() ()k 1/M () then the graph T can be almost -quasi-tiled by (t) T
for some integer t.

Proof. Let v be the maximum number of vertices in Ti , and let t be the largest
integer such that ktk kk. Clearly M () t and kk 1, hence
1 2
k tk kk + tk() ()k kk + t kkt,
M () M ()
22 Chapter 2: The class of sofic groups

hence each coordinate of differs by at most 2kkt/M () from the corresponding

coordinate of t. Therefore T can be obtained from (t) T by adding or deleting
at most 2ktk/M () copies of each Ti . This means, that we can get T from
(t (2tkk/M ())1) T by adding at most 4ktk/M () copies of each Ti (where
1 = (1, 1, . . . , 1) Nm ). The maximum number of vertices involved in this operation
is 4mvtkk/M ().
The sum of the coordinates of is at least kk, hence the graph T has at
least kk tkk vertices and (t) T has at least tkk vertices. If we choose
M () > 4mv/ than we obtain that there are at most tkk vertices of T which
are not covered by (t (2tkk/M ())1) T . So this is an -quasi-tiling. Note that
the tiles are clearly -disjoint since they are disjoint.
On the other hand, the graph (2tkk/M ())1 T contains at most 2mvtkk/M ()
edges, which is clearly smaller than tkk. So (t) T almost -quasi-tiles T .

Lemma 2.19. Let B be a finite d-regular graph, and A = {A1 , . . . , An } an 2 -quasi-

tiling of B, where Ai B. Furthermore, suppose that for 1 i n, Ai is almost
1 -quasi-tiled by {B1i , B2i , . . . , Bki i } where Bji Ai , that is, we can omit some of them
such that we get an 1 -quasi-tiling.
Let B = {Bji : 1 i n, 1 j ki } and v = max{|Bji | : 1 i n, 1 j ki }.
Suppose that dv 2 + 2 + 41 < 1.
In this case B almost (dv 2 + 2 + 41 )-quasi-tiles B.

Proof. We can assume that Ai is 1 -quasi-tiled by {B1i , B2i , . . . , Blii } for some li ki .
Let us keep in mind that this implies lj=1 |Bji | > (1 1 ) kj=1 |Bji |.
Pi Pi

According to the definition of -disjointness, there are Ai Ai and Bji Bji sub-
graphs such that

|Ai | > (1 2 )|Ai | for 1 i n;

Ai Ai0 = if i 6= i0 ;

|Bji | > (1 1 )|Bji | for 1 i n, 1 j li ;

Bji Bji 0 = if j 6= j 0 .

Our problem is that these Bji s are not disjoint for different is. The plan is to omit
those which intersect each other and prove that we still have enough Bji s.
Since our graph is d-regular, if D is a subgraph then |Nr (D)| dr |D|. Now let

A0i = Ai \ Nv (Ai \ Ai ).
Chapter 2: The class of sofic groups 23

Clearly A0i Ai . The following calculations come from the previous observations:

|Ai \ Ai | < 2 |Ai |;

|Nv (Ai \ Ai )| < dv 2 |Ai |;

|A0i | > |Ai | dv 2 |Ai | = (1 dv 2 )|Ai |.

It is also clear that if for some j, A0i Bji 6= then Bji (Ai \ Ai ) = (because
v |V (Bji )|). Let us omit those Bji s that do not intersect their A0i . We can assume
that we are left with {Bji : 1 i n, 1 j mi } for some mi li . This way for
i 6= i0 we have
0
Bji Ai , Bji 0 Ai0 ,
0
so Bji Bji 0 = .

This means that our new system of tiles {Bji : 1 i n, 1 j mi } is 1 -disjoint.

We need to count two more things: How many vertices of B do they cover and how
many vertices did we omit from B = {Bji : 1 i n, 1 j ki }?

1. It is easier to count the vertices covered by the Bji s because they are disjoint.
First consider just one of the Ai s. Here we have {B1i , B2i , . . . , Bm
i
i
}. We know
that
li
X
|Bji | (1 1 )|Ai |,
j=1

and since |Bji | > (1 1 )|Bji |, we have

li
X
|Bji | > (1 1 )2 |Ai | > (1 21 )|Ai |.
j=1

Sli
So there are at most 21 |Ai | points not covered by j=1 Bji in Ai , consequently
there are at most the same number of points not covered by the same set in
A0i . If Bji intersects A0i then j mi . So every vertex in A0i which is in lj=1 Bji
Si

is contained in m
S i i
j=1 Bj as well. Hence
m 
 [i 
 Bji A0i  > |A0i | 21 |Ai |,
 
 
j=1
m 
 [i 
 Bji  > (1 dv 2 21 )|Ai |.
 
 
j=1

From this, clearly

 m   
[n [ i  [n 
Bji  > (1 dv 2 21 )  Ai  > (1 dv 2 21 )(1 2 )|B|,
   

   
i=1 j=1 i=1
24 Chapter 2: The class of sofic groups

 m 
[n i 
 [ i
so  Bj  > (1 dv 2 21 )(1 2 )|B| > (1 dv 2 2 21 )|B|.
 
i=1 j=1

2. We saw that for each i we have

m 
 [i 
 Bji  > (1 dv 2 21 )|Ai |.
 
 
j=1

Hence mi
X
|Bji | > (1 dv 2 21 )|Ai |.
j=1
Pli Sli
Since we know that j=1 |Bji | < | j=1 Bji |/(1 1 ) |Ai |/(1 1 ), we can
estimate
Pmi i
j=1 |Bj | (1 dv 2 21 )|Ai |
Pli i
> |Ai |
= (1 1 )(1 dv 2 21 ) > 1 dv 2 31 .
j=1 |Bj | 11

From the choice of li we know that

li
X ki
X
|Bji | > (1 1 ) |Bji |,
j=1 j=1

so
Pmi Pmi Pli
j=1 |Bji | j=1 |Bji |j=1 |Bji |
Pki = Pli Pki > (1 dv 2 31 )(1 1 ) > 1 dv 2 41 .
j=1 |Bji | j=1
i
|Bj | j=1 |Bji |

We can see that these Bji s form an (1 dv 2 2 21 )-cover of B and they contain
(1dv 2 41 ) times all the vertices of B. This means that B almost (dv 2 +2 +41 )-
quasi-tile B.

This is a stronger version of Proposition 2.8 from [6]:

Proposition 2.20. Let be a finitely generated amenable group, S a finite

symmetric generating set and F a Flner sequence in Cay(, S). Then for each
integer r > 0 there is an integer Rqt (r) (which depends also on F) and a finite
subsequence T F of say m members and a lattice vector 0 6= Nm with the
following property. Each member of T is an r-approximation of Cay(, S), and each
Rqt (r)-approximation of Cay(, S) can be almost 1r -quasi-tiled by an integer multiple
of T .

Proof. By discarding a few elements from F we may assume that the remaining
elements are all r-approximations of Cay(, S). We apply Proposition 2.14 with
parameters 1 and n = 0 (1 shall be given later).
Chapter 2: The class of sofic groups 25

We obtain T = {F1 , F2 , . . . , Fm } F, L and . Choose R such that

R > max(L, 1/). This way whenever a finite edge-colored graph B satisfies
|QB
L |/|V (B)| > 1 , it is an R-approximation of the Cayley-graph. So it follows
from the proposition that each R-approximation of Cay(, S) can be 1 -quasi-tiled
by some linear combination of T .
Let F 0 F be those Flner sets that are R-approximations of Cay(, S). For each
member F F 0 we choose a lattice vector 0 6= F Nm such that F T 1 -quasi-
tiles F .
Now {(F ) : F F 0 } is a sequence of unit vectors, so it has an accumulation point
u Rm . We apply Lemma 2.18 to T and 1 , and obtain the bound M (1 ) = M . Let
us fix a member H F 0 such that k(H ) uk < 1
2M
and then let

1
F 00 = {F F 0 : k(F ) uk < and kF k M kH k}.
2M

This F 00 is still a Flner sequence and for each F F 00 we have k(F ) (H )k <
1/M , hence F T can be almost 1 -quasi-tiled by (tF H ) T for some tF N.
Applying again Proposition 2.14 for F 00 and 2 we get a new finite subsequence
T 0 = {Q1 , Q2 , . . . , Qk } F 00 and L0 , 0 . Choose R0 such that R0 > max(L0 , 10 ).
Now let X be an R0 -approximation of Cay(, S). Then X can be 2 -quasi-tiled by
a linear combination of T 0 , say T 0 . Let us examine a tile of type Qi , it can be 1 -
quasi-tiled by Qi T . Since Qi F 00 , we can choose (tQi H (2tQi kH k/M (1 ))1)T
tiles among Qi T such that we omitted at most 1 |Qi T | vertices (according to
the proof of Lemma 2.18). This means that our new tiles cover at least (1 21 )|Qi |
vertices. In other words, (tQi H ) T almost 21 -quasi-tiles this tile of type Qi . We
can do the same for every tile.
So we have exactly the situation of the previous lemma with T 0 2 -quasi-tiling X
and (tQi H ) T almost 21 -quasi-tiling the tiles of type Qi , v = max{|F | : F T }.
Lemma 2.19 tells us that (1 tQ1 + 2 tQ2 + + k tQk ) H T almost (dv 2 +2 +81 )-
quasi-tile X. Here d = |S|/2, since S is a symmetric generating set.
Choose 1 to be 1/16r. This determines T and consequently v. Then we can choose
2 = 1/(2r(dv + 1)). This way

1
dv 2 + 2 + 81 = ,
r

so this integer multiple of H T almost 1r -quasi-tiles X.

Remark 2.21. Note that this proposition clearly implies Proposition 2.8 from [6].
26 Chapter 2: The class of sofic groups

2.3.2 The proof of the theorem

Definition 2.22. Let and partitions of a finite set S. The incidence graph of
and is a bipartite graph, whose two sets of vertices consist of the classes of
and the classes of , and the edges are the elements of S, each element connects its
-class with its -class.

Proposition 2.23 (From [5].). For each triple (a, b, r) of integers there is a finite
set S with two partitions and on it such that each -class has a elements, each
-class has b elements, an -class can meet a -class in at most one element, and
in the incidence graph of and each simple cycle is longer than 2r.

Remark 2.24. Suppose we have S as described in the proposition, and there is a

graph A whose vertex set is S and for every edge its endpoints are in the same
-class or in the same -class. The condition that in the incidence graph there are
no short cycles means the following here. If we have a simple cycle in A such that
its length is at most 2r then it is contained in an -class or in a -class.

Definition 2.25. Let A be a colored graph and Z a subset of the colors used in
A. Then A|Z denotes the subgraph obtained from A by omitting all edges whose
color does not belong to Z. Suppose that C is another colored graph and is an
r-isomorphism between A|Z and C|Z .
We build another graph called the enhancement of C with A along , denoted by

A C. We start from C and add new edges to it. Namely, for each edge a b of
A whose color does not belong to Z we add a new edge (a) (b) of the same
color, provided that is defined at the endpoints a and b.

Theorem 2.26. Let G and H be finitely generated sofic groups, G a finitely

generated amenable subgroup and : , H an injective homomorphism. Then the
amalgamated product G H is also a sofic group.

Proof. For simplicity we identify G and H with their canonical image in G H.

This identifies and () with their image too. Let X G, Y H and Z be
finite symmetric generating sets such that Z = X Y .
Let r > 0 be an integer. We aim to find an r-approximation of Cay(G H, X Y ).
Now we define recursively a sequence of numbers ki for 0 i 2r + 1. Consider
all the words containing letters from (X Y ) \ Z that have length at most 2r + 1.
Some of these words may belong to the amenable subgroup in G H. If a word
w , let us call the Z-length of w the length of the shortest word w such that w
consists of letters from Z and w = w in G H. Since we have finitely many such
words, the maximum of their Z-lengths is a finite number, let us call it k0 .
Chapter 2: The class of sofic groups 27

If we already have k0 , k1 , . . . , ki1 then we can find ki as follows. Consider all the
words that contain at most 2r + 1 i letters from (X Y ) \ Z and at most iki1
letters from Z, and consider the Z-lengths of those that are in . Let the maximum
of these Z-lengths be ki .
Let d = |X Y | = |X| + |Y | |Z|, K = k2r+1 + 2r + 1, and set

r0 > K,

r1 > 8r0 d2r0 +1 ,

X Y
r2 > max{Rapp (r1 ), Rapp (r1 )},
X
where Rapp (r1 ) denotes the number Lemma 2.10 provides for Cay(G, X) and r1 , and
Y
Rapp (r1 ) is what we get in the case of Cay(H, Y ).
We choose a Flner sequence F in Cay(, Z), we can assume that each F F
is connected. Then applying Proposition 2.20 to the number r2 provides a finite
T = {F1 , . . . , Fm } F and a nonzero vector Nm . Let us denote by D the graph
T . Let
r3 > max{Rqt (2r2 ), K + |D|}.

Let us start with r3 -approximations A resp. B of the Cayley graphs Cay(G, X)

resp. Cay(H, Y ). Then A|Z and B|Z are both r3 -approximations of Cay(, Z). Since
r3 > Rqt (2r2 ), Proposition 2.20 implies that both A|Z and B|Z can be almost 1/(2r2 )-
quasi-tiled by some integer multiples of T , say a( T ) and b( T ). In particular
this means that A|Z is r2 -isomorphic to a D and B|Z is r2 -isomorphic to b D.
Pm
Claim 2.27. Let N = i=1 i and A0 = N A, B 0 = N B. Then there exists
an r2 -isomorphism from N a D to A0 such that for each copy of D takes all
the connected components of this copy to different copies of A, and there exists an
r2 -isomorphism from N b D to B 0 with the same property.

Proof. We have an r2 -isomorphism : a D A. Let D1 , . . . , Da denote the copies

of D, so : ai=1 Di A. We construct an r2 -isomorphism from N a D to A0 in the
S

following way. Let us arrange N a D in a matrix form, let {Di,j : 1 i N, 1

j a} denote the copies of D.
Now fix some 1 j a and look at Di,j , which is isomorphic to T . Hence this
graph contains N of the Flner sets, so it has exactly N connected components, let
us denote them by T i,j , T i,j , . . . , T i,j . We can assume that T 1,j
1 2 N k =T
k
2,j
= ...
=Tk
N,j

for each k. For 1 l N let

Djl = {Tki,j : k i l mod N }.

28 Chapter 2: The class of sofic groups

Each Djl contains one Tki,j for a fixed k, so because of our assumption on the Tki,j s
Djl is isomorphic to D. Clearly Dj1 , . . . , DjN are pairwise disjoint and N l
S
l=1 Dj =
SN
i=1 Di,j . This way we ensured that the connected components of a Di,j lie in dif-
ferent Djl s. Now for a fixed l we know that aj=1 Djl is isomorphic to ai=1 Di , let 0
S S

denote the isomorphism which takes Djl to Dj . Then

a
[
0 : Djl A
j=1

is an r2 -isomorphism. Let us do the same for each l, except that the images should
be different copies of A, say we take aj=1 Djl to Al . This way we defined
S

a [
[ N N
[
: Djl Al ,
j=1 l=1 l=1

it is the same as
a [
[ N
: Di,j A0 .
j=1 i=1

This has the property that is takes all the connected components of a Di,j to
different copies of A and it is indeed an r2 -isomorphism.
The other r2 -isomorphism from N b D to B 0 can be constructed similarly.

This means that by replacing A by A0 we can assume that A|Z is r2 -isomorphic to

a D and this r2 isomorphism has the property that there is no path in A between
different connected components of a D-copy. Replace B by B 0 , hence B has the same
property.
Apply Proposition 2.23 for the triple (a, b, r + 1), we obtain a finite set S and
partitions and of S. Define the graph C0 = S D, the union of |S| disjoint
copies of D. Now for an -class S the subgraph D is exactly the graph
a( T ). So it is r2 -isomorphic to A|Z and we can enhance it with A along the
r2 -isomorphism we constructed above. Repeating this enhancement for all -classes
we obtain a new graph C1 on the same vertex set (S D). This graph C1 is clearly
r2 -isomorphic to (|S|/a) A and C1 |Z is still isomorphic to S D.
Next for each -class S we consider the subgraph C1 C1 spanned by the
subset D. Then C1 |Z is r2 -isomorphic to B|Z , so we can enhance it with B
along the other r2 -isomorphism we constructed. Repeating this enhancement for all
-classes we obtain a new graph C on the same vertex set S D. This C|X is
r2 -isomorphic to (|S|/a) A and C|Y is r2 -isomorphic to (|S|/b) B.
We would like to prove that this graph C is an r-approximation of the group G H.
We need to find a set of good points: those which have r-neighborhood isomorphic
to an r-ball in the Cayley graph.
Chapter 2: The class of sofic groups 29

In A there exists a subgraph A3 such that |A3 | > (1 1/r3 )|A| and for each
p A3 Br3 (p) in A is isomorphic to the r3 -ball in Cay(G, X). There is a set
B3 B with similar properties. For each -class in C there is an image of this
A3 , let us call the union of these images C3X . Similarly we get C3Y from the
images of B3 . For these we have
   
X 2 2
|C3 | > 1 |C|, |C3Y | > 1 |C|,
r3 r3

and for every p C3X the r3 -neighborhood of p in C|X\Z still looks good (like
the ball Br3 (1) in Cay(G, X)|X\Z ), and the same holds for every q C3Y in
C|Y \Z . There is 2 in the numerator because the size of the -classes might
differ from the size of A, but at most by |A|/r3 .

Each -class in C is r2 -isomorphic to A, this means we have A2 A that is

isomorphic to a subgraph of our -class. Let us call the union of these images
C2X , and similarly we get C2Y . Clearly
   
X 1 1
|C2 | > 1 |C|, |C2Y | > 1 |C|.
r2 r2

Since every -class is r2 -isomorphic to A and A is an r3 -approximation of

Cay(G, X), the -classes are r1 -approximations of Cay(G, X). So for an -
class there is a subset A1 D for which |A1 | > (1 1/r1 )| D| and for
each point in A1 the r1 -neighborhood of this point looks good. Let us denote
the union of these by C1X . We get C1Y similarly.

Consider the set

C 0 = C3X C3Y C2X C2Y C1X C1Y .

Clearly we have  
0 4 2 2
|C | > 1 |C|,
r3 r2 r1
and the points in this set still have the previously described properties.
Let C 00 = C \ N2r0 +1 (C \ C 0 ). This means that the 2r0 + 1-neighborhood of C 00 is in
C 0 . Clearly
8
|C \ C 0 | < |C|,
r1
hence
8 8 1
|N2r0 +1 (C \ C 0 )| < d2r0 +1 < d2r0 +1 2r +1
= |C|,
r1 8r0 d 0 r0
 
1
|C 00 | = |C \ N2r0 +1 (C \ C 0 )| > 1 |C|.
r0
30 Chapter 2: The class of sofic groups

Claim 2.28. For each point p C 00 the r-neighborhood of p in C is rooted isomorphic

to an r-ball in Cay(G H, X Y ).

Proof. Let us fix a vertex p C 00 . We need to show two things. First, if there is
a word with letters from X Y that is equal to the identity in the group and has
length at most 2r + 1, then the corresponding path starting from p in C must be a
cycle. Secondly, if we have a simple cycle containing p of length at most 2r + 1 then
the corresponding word must be the identity. These clearly imply our claim.
We will use the following lemma, which can be easily verified by the reader.

Lemma 2.29. Let w be a word with letters from X Y , and suppose that w is the
identity in G H. Then we can get the empty word from w using the following
operation finitely many times. We choose a subword u of w such that u contains
letters only from X or only from Y , and the group element u is in the amenable
subgroup . Then we change it to a word that is equal to u in the group and contains
letters only from Z. (If it is the identity then we can delete it.)

This is the analogue of the statement that in a free product G H a word

g1 h1 g2 h2 . . . gk hk can only be simplified if one of the gi s is the identity element
in G or one of the hi s is the identity of H.
Now we can prove the claim.

1. Let w be a word which has length at most 2r + 1 and w = 1 in G H, and let

us denote by P the corresponding path from p. Since the 2r0 + 1-neighborhood
of p is in C 0 , all points of the path P are also in C 0 .
According to the lemma, we have finitely many, say M operations of the above
kind. Let us call wi the word we have after the ith step, so wM is the empty
word. Suppose w has k letters from (X Y ) \ Z and l letters from Z (k + l
2r + 1). In this case clearly M k since in each step we decrease the number
of letters which are not in Z.
Let us denote by Pi the path in the graph C starting from p which corresponds
to the word wi . We will prove the followings by induction on i.

(a) The word wi has at most k i letters from (X Y ) \ Z and at most l + iki
letters from Z. So in particular the length of wi is less than K.
(b) The path Pi is contained in C 0 .
(c) Pi has the same endpoint as Pi1 .

If w0 = w, the statements are trivial for i = 0. Note that (b) follows from (a),
because the length of the path is less than K.
Chapter 2: The class of sofic groups 31

Now consider the step when we get wi from wi1 . Let wi1 = v1 uv2 and
wi = v1 u0 v2 . Let the endpoint of the path corresponding to v1 be q. Now we
know that u contains at most k (i 1) 2r + 1 (i 1) letters from
(X Y ) \ Z and at most l + (i 1)ki1 iki1 letters from Z, hence its
Z-length is at most ki according to the definition of ki . So the length of u0 is
at most ki , and wi contains at most ki + l + (i 1)ki1 < l + iki letters from
Z.

For the third statement, we know that q C 0 . Assume that u has letters only
from X (the proof in the other case is the same), then the path corresponding
to the word u and u0 both lie in C|X . They are equal in the group, both have
length at most K. We also know that the r0 -neighborhood of q in C|X looks
like an r0 -ball in Cay(G, X). So the paths corresponding to u and u0 must end
at the same point. So Pi1 and Pi have the same endpoint as well.

We know that PM is just the empty word, hence its endpoint is p. This means
that all Pi s end at p, in particular the original path P ends there too. So P
is a cycle, this is what we wanted to prove.

2. Now let P be a simple cycle starting and ending at p which has length at most
2r + 1. Let w be the corresponding word in the group. We would like to prove
that this word is equal to the identity in G H. Assume that w contains at
most 2r + 1 i letters from (X Y ) \ Z, and at most i iki1 letters from Z.

Take the map C S, where {s} D goes to the point s S. Consider the
image of P in S. For every edge of P , whose color does not belong to Z, we
draw an edge between the images of the endpoints. The images of the Z-edges
would be loops, that is why we do not consider them. This is a cycle, though
usually it is no longer simple.

But clearly we can find a simple part of this cycle, i.e., there exists P 0 P
such that the endpoints of P 0 are in the same copy of D in C. We can assume
/ P 0 , since there are always at least two such simple parts of a cycle,
that p
so we can choose one that does not contain p.

According to Remark 2.24, this P 0 is contained in an -class or in a -class.

Suppose that it lies in an -class, the proof in the other case goes similarly.

Let the corresponding word to P 0 be u, this is a subword in w. There are two

kinds of Y \ Z-colored edges in C: those that have endpoints in the same D-
copy, and those that have endpoints in different -classes. The path P 0 clearly
cannot contain the one of the latter, and since the image of P 0 is simple, it
32 Chapter 2: The class of sofic groups

cannot contain one of the former kind. So u does not contain elements from
Y \ Z.
Let w = v1 uv2 , and let us denote by q the endpoint of the path corresponding
to v1 . So P 0 is a path starting from q and ending in the same D-copy, let us call
this endpoint q 0 . Because of the assumption on the r2 -isomorphism between A
and the -classes, there are no X-colored paths between different components
of this copy of D. Since P 0 is an X-colored path with endpoints in the same
copy, its endpoints must be in the same connected component of D = T .
Since p C 00 , P C 0 . This means that P 0 has an image in A, call this image
P 0 , the endpoints q and q 0 . These endpoints are in the same tile of the quasi-
tiling, because q and q 0 are in the same componenet of D. The size of this tile
is at most the size of D, hence we have a path Q in the tile containing only
Z-edges between q 0 and q, that has length at most |D|. This means that P 0 Q
is a cycle in A.
We also know that q has a good r3 -neighborhood in A. We chose r3 to be
greater than |D| + K, and the length of P 0 Q is at most |D| + K, so the word
corresponding to this path is the identity in the group G.
Let us denote by u0 the corresponding word to Q1 . The previous observation
means that u0 = u in G. Since Q has only edges labeled by Z, u0 contains
only letters from Z. Since u had at most iki1 letters from Z and at most
2r + 1 i letters from X \ Z, the length of u0 is at most ki . This ki is smaller
than K < r0 , and in C the r0 -ball around q looks good, so the corresponding
path to u0 u1 is a cycle in C starting from q.
So we can change our word w = v1 uv2 to v1 u0 v2 . Let w1 be this new word, and
let P1 be the cycle corresponding to w1 . Then w1 is the identity if and only if
w is the identity, so we can continue with w1 . We also know that w1 contains
at most 2r +1i1 letters from (X Y )\Z and at most iki1 +ki (i+1)ki
letters from Z.
From now on, we can do the same as before, everything works similarly (in
other words, we can use induction). We get the words w2 , w3 , . . . , each of them
having length at most K and in the group G H they are all equal. After
2r + 1 i steps we have w2r+1i , that contains only letters from Z, and has
length at most K. The corresponding path P2r+1i is a cycle in C, hence it is
a cycle in A (or in B) as well, and this p has a good r0 -neighborhood in A. So
w2r+1i = 1 in , hence it is the identity element in G H.

We finished the proof of the claim.

Chapter 2: The class of sofic groups 33

So we found C 00 C, for which

   
00 1 1
|C | > 1 |C| > 1 |C|,
r0 r

and for each point p C 00 , the r-ball around p in C is rooted isomorphic to an

r-ball in the Cayley graph of the amalgamated product. This proves that C is an
r-approximation of the Cayley graph, so the group G H is sofic.

We still need to prove the statement in the case when our groups are not necessarily
finitely generated.

Theorem 2.30. Let G and H be sofic groups, G an amenable subgroup and

: , H an injective homomorphism. Then the amalgamated product G H is
also sofic.

Proof. The amalgamated product G H is the direct limit of amalgamated products

{G H } over amenable subgroups , where these G , H and are all finitely
generated groups. We know that direct limits of sofic groups are sofic, so this proves
the theorem.
34 Chapter 3: Surjunctivity conjecture

Chapter 3

Surjunctivity conjecture

The following conjecture is due to Gottschalk [7].

Conjecture 3.1 (Gottschalk). Let G be a countable group and X a finite set. Con-
sider the compact metrizable space X G of X-valued functions on G equipped with
the product topology. Let A : X G X G be a continuous map that commutes with
the natural right G-action. Then if A is injective, it is surjective as well.

We present the main ideas of the proof of this conjecture for sofic groups. The
following proof was communicated by Mikls Abrt.

Theorem 3.2 (Gromov). Let be a finitely generated sofic group, X a finite set
and let A : X X be a continuous -equivariant map, and suppose that A is
injective. Then A is surjective as well.

We will use the following two easy lemmas. We leave the proofs for the reader.

Lemma 3.3. Let A : X X be a continuous -equivariant map. This means that

A commutes with the natural right -action on X . Then the following statements
hold for A.

(a) We define A : X X such that for f : X, A(f ) = f (1). Here 1 is

the identity element of . Then A determines A.

(b) There exists a finite subset H (which depends on A), such that if f, f 0 :
X, f |H = f 0 |H then A(f )(1) = A(f 0 )(1). In other words, the restriction of
f to H determines the value of its A-image on the identity element.

(c) A(X ) X is a -invariant closed subset.

Lemma 3.4. Suppose that A : X X is not surjective, i.e., A(X ) 6= X . Then

there exists a forbidden pattern: T , |T | < and f : T X, such that for
any f A(X ), f |T 6= f.
Chapter 3: Surjunctivity conjecture 35

Proof of Theorem 3.2. Let = hSi be a sofic group, where S is a finite symmetric
generating set. For f X we regard f as a coloring of by the elements of X.
Part (b) of Lemma 3.3 provides a finite subset H . Then there exists an integer
L > 0 such that H BL (1) in Cay(, S). This means that for each element g
Cay(, S) and every f X the f -coloring of the L-neighborhood of g determines
A(f )(g), the A(f )-color of g.
Assume that A is not surjective. Then Lemma 3.4 gives us T and f : T X. Let
NL (T ) denote the L-neighborhood of T in Cay(, S) and choose K N such that
NL (T ) BK (1).
Choose {n }
n=1 such that limn n = 0, and {rn }n=1 such that limn rn =
and rn > K for each n N. Since is sofic, for every n we can find a finite graph
Gn that is an (rn , n )-approximation of Cay(, S).
There are subsets Vn V (Gn ) for every n such that for each p Vn , Brn (p) is
isomorphic to Brn (1) in Cay(, S), and we have

|Vn | > (1 n ) |V (Gn )|.

We define An : X Gn X Gn as follows. Let fn X Gn . If p Vn , then Brn (p)

is isomorphic to Brn (1) in Cay(, S). Since rn > L, we have H Brn (1). Then,
according to Lemma 3.3, this coloring determines the value of its A-image on the
centre of the rn -ball. Let this color be x X. Let An (fn )(p) = x. For p
/ Vn let
An (fn )(p) = x0 , where x0 X is a fixed element.
For every p Vn the rn -ball around p is isomorphic to Brn (1). Since rn > K, this
ball contains the L-neighborhood of the forbidden pattern T . So we have a copy
of T and its L-neighborhood in Brn (p). In the A-image of a coloring the color of a
point depends only on its L-neighborhood. This is true for An and p Vn too. Since
f : T X is a coloring that is not in the image of A, this pattern does not occur
in the image of An .
Let us denote by Bn X Gn the image of An .

Claim 3.5. Since Bn does not contain the forbidden pattern, there exists n0 N
and a constant c > 1, such that for n > n0 we have

X n  > c|V (Gn )| |Bn |.

 G 

This means that for n > n0 we can choose colorings fn Bn X Gn , such that
|A1
n (fn )| > c
|V (Gn )|
.
For n > n0 let xn A1 1
n (fn ). We would like to find yn An (fn ) and pn Vn such
that xn (pn ) 6= yn (pn ). Suppose that for some n there is no such coloring. This means
that for each x0 A1 0
n (fn ), x differs from xn only in the points of V (Gn ) \ Vn . Since
36 Chapter 3: Surjunctivity conjecture

we have |X||V (Gn )\Vn | ways to color these points, we can have at most this amount
of colorings in A1
n (fn ). So

c|V (Gn )| < |A1

n (fn )| < |X|
|V (Gn )\Vn |
< |X|n |V (Gn )| .

As n 0, there exists n1 such that for n > n1 , |X|n < c. So for n > n1 we have

c|V (Gn )| > |X|n |V (Gn )| ,

which is exactly the opposite of the above inequality. This implies that for n > n1
we can find the desired yn and pn .
Recall that a sequence of rooted graphs {(Gi , qi )}
i=1 is rooted convergent and con-
verges to the rooted graph (G, q) if for every R there exist i(R) such that for i > i(R)
the R-ball around qi in Gi is rooted isomorphic to the R-ball around q in G.
Now for n > n1 consider (Gn , pn ) as a rooted graph. Since pn Vn , i.e., pn has a
good rn -neighborhood in Gn , the rooted limit of {(Gn , pn )} is clearly (Cay(, S), 1).
We define three sequences of colorings of as follows. For n > n1 , Brn (pn ) is rooted
isomorphic to Brn (1) in Cay(, S). Using this isomorphism, xn gives us a coloring of
Brn (1). Extend this to the other vertices by x0 X, this way we obtain the coloring
xn X . We get yn from yn , and fn from fn similarly.
Since X is a compact space, every sequence has a convergent subsequence. So we
can find {nk } , such that {xn }, {yn } and {fn } are convergent. Let xn x and
k=1 k k k k

ynk y and fnk f . Since xn (pn ) 6= yn (pn ), we get that x(1) 6= y(1).

Claim 3.6. For these x, y, f X , we have A(x) = A(y) = f .

This finishes the proof, since x 6= y, A(x) = A(y) contradicts the injectivity of A.
BIBLIOGRAPHY 37

Bibliography

[1] M. Abrt, A. Thom, B. Virg, Benjamini-Schramm convergence and pointwise

convergence of the spectral measure, preprint at http://www.renyi.hu/~abert

[2] B. Collins and K. J. Dykema, Free products of sofic groups with amalgamation
over monotileably amenable groups, Mnster Journal of Mathematics, 2011,
vol. 4, pp. 101-118.

[3] G. Elek, The strong approximation conjecture holds for amenable groups, Jour-
nal of Functional Analysis, 2006, vol. 239, pp. 345-355.

[4] G. Elek and E. Szab, Sofic groups and direct finiteness, Journal of Algebra,
2004, vol. 280, no. 2, pp. 426-434.

[5] G. Elek and E. Szab, On sofic groups, Journal of group theory, 2006, vol. 9,
no. 2, pp. 161-171.

[6] G. Elek and E. Szab, Sofic representations of amenable groups, Proceedings of

the American Mathematical Society, 2011, vol. 139, pp. 4285-4291.

[7] W. Gottschalk, Some general dynamical notions, Lecture Notes in Mathematics,

1973, pp. 120-125.

[8] M. Gromov, Endomorphisms of symbolic algebraic varieties, Journal of the

European Mathematical Society, April 1999, vol. 1, no. 2, pp. 109-197.

[9] D. S. Ornstein and B. Weiss, Entropy and isomorphism theorems for actions of
amenable groups, Journal dAnalyse Mathmatique, 1987, vol. 48, no. 1, pp. 1-
141.

[10] L. Punescu, On sofic actions and equivalence relations, Journal of Functional

Analysis, 2011, vol. 261, no. 9, pp. 2461-2485.

[11] B. Weiss, Sofic groups and dynamical systems, Sankhya: The Indian Journal in
Statistics, special issue on Ergodic Theory and Harmonic Analysis, 2000, vol. 62,
series A, pt. 3, pp. 350-359.