Documente Academic
Documente Profesional
Documente Cultură
`
a degli Studi di Roma La Sapienza
Dottorato di Ricerca in Matematica
XII ciclo
tesi
Cellular Automata
and
Finitely Generated Groups
di
Francesca Fiorenzi
Dipartimento di Matematica
Universit` a di Roma La Sapienza
Piazzale Aldo Moro 2 - 00185 Roma
email orenzi@mat.uniroma1.it
Contents
Introduction 3
1. Cellular Automata 7
1.1 Cayley Graphs of Finitely Generated Groups . . . . . . . . . . . 8
1.2 Shift Spaces and Cellular Automata . . . . . . . . . . . . . . . . 10
1.3 Shifts of Finite Type . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4 Edge Shifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5 Soc Shifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5.1 Minimal Deterministic Presentations of a Soc Shift . . . 19
1.6 Decision Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.6.1 A Decision Procedure for Surjectivity . . . . . . . . . . . 20
1.6.2 A Decision Procedure for Injectivity . . . . . . . . . . . . 21
1.7 Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2. Surjunctivity and Density of Periodic Configurations 25
2.1 Surjunctivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 Periodic Congurations of a Shift . . . . . . . . . . . . . . . . . . 26
2.3 Residually Finite Groups . . . . . . . . . . . . . . . . . . . . . . 28
2.4 Density of Periodic Congurations . . . . . . . . . . . . . . . . . 29
2.5 Periodic Congurations of Euclidean Shifts . . . . . . . . . . . . 31
2.6 Group Shifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.6.1 Decision Problems for Group Shifts . . . . . . . . . . . . . 34
3. The MooreMyhill Property 36
3.1 The Garden of Eden Theorem and the MooreMyhill Property . 38
3.2 Amenable Groups . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 The MooreMyhill Property for Subshifts of A
Z
. . . . . . . . . 43
3.3.1 A counterexample to Mooreproperty for a soc subshift
of A
Z
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.4 Gromovs Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5 Strongly Irreducible Shifts . . . . . . . . . . . . . . . . . . . . . . 54
3.6 SemiStrongly Irreducible Shifts . . . . . . . . . . . . . . . . . . 62
Bibliography 69
2
Introduction
The notion of a cellular automaton has been introduced by Ulam [U] and von
Neumann [vN]. In this classical setting, the universe is the lattice of integers
Z
n
of Euclidean space R
n
. The set of states is a nite set A (also called the
alphabet) and a conguration is a function c : Z
n
A. Time t goes on in
discrete steps and represents a transition function : A
Z
n
A
Z
n
(if c is the
conguration at time t, then (c) is the conguration at time t + 1), which
is deterministic and local. Locality means that the new state at a point
Z
n
at time t + 1 only depends on the states of certain xed points in the
neighborhood of at time t. More precisely, if c is the conguration reached
from the automaton at time t then (c)
|
= (c
|N
), where : A
N
A is
a function dened on the congurations with support the nite set N (the
neighborhood of the point 0 Z
n
), and N
= + N is the neighborhood
of obtained from N by translation. For these structures, Moore [Moo] has
given a sucient condition for the existence of the socalled Garden of Eden
(GOE) patterns, that is those congurations with nite support that cannot be
reached at time t from another conguration starting at time t 1 and hence
can only appear at time t = 0. Moores condition (i.e. the existence of mutually
erasable patterns a sort of noninjectivity of the transition function on the
nite congurations) was also proved to be necessary by Myhill [My]. This
equivalence between local injectivity and local surjectivity of the transition
function is the classical wellknown Garden of Eden theorem.
The purpose of this work is to consider this kind of problems in the more
general framework of symbolic dynamic theory, with particular regard to sur-
junctivity theorems (and, in this case, to the density of the periodic congu-
rations) and to GOElike theorems restricted to the subshifts of the space A
(the
socalled full Ashift), that is a function dened on with values in a nite
alphabet A. The space A
which is
invariant and topologically closed is called subshift, shift space or simply shift.
In this setting a cellular automaton (CA) on a shift space X is given by speci-
fying a transition function : X X which is local (that is the value of (c),
where c X is a conguration, at a point only depends on the values of
c at the points of a xed neighborhood of ).
In Chapter 1 we formally dene all these notions, also proving that many
basic results for the subshifts of A
Z
given in the book of Lind and Marcus
[LinMar], can be generalized to the subshifts of A
.
Hence if is a residually nite group, a transition function of a cellular
automaton on A
is preinjective if when-
ever two congurations c, c X dier only on a nite nonempty subset of ,
then (c) ,= ( c)); this notion has been introduced by Gromov in [G]. In fact,
we prove that these two properties are equivalent for local functions dened on
5
the full shift, but in the case of proper subshifts the former may be meaningless.
On the other hand, we give a proof of the fact that the nonexistence of GOE
patterns is equivalent to the nonexistence of GOE congurations, that is to the
surjectivity of the transition function. Hence, in this language, the GOE theo-
rem states that is surjective if and only if it is preinjective. We call Moores
property the implication surjective preinjective and Myhills property the
inverse one.
Concerning onedimensional shifts, from the works of Hedlund [H] and Coven
and Paul [CovP] we prove that the MooreMyhill property (MMproperty)
holds for irreducible shifts of nite type of A
Z
. Moreover, using this result we
prove that Myhills property holds for irreducible soc shifts of A
Z
. On the
other hand, we give a counterexample of an irreducible soc shift X A
Z
but
not of nite type for which Moores property does not hold.
Concerning general cellular automata over amenable groups, from a result
of Gromov [G] in a more general framework, it follows that the MMproperty
holds for shifts of bounded propagation contained in A
(and we also show that strong irreducibility together with the nite type
condition is strictly weaker that the bounded propagation property).
The main dierence between irreducibility and strong irreducibility is easily
seen in the onedimensional case. Here the former property states that given
any two words u, v in the language of a shift, there exists a third word w such
that the concatenation uwv is still in the language. Strong irreducibility says
that we can arbitrarily x the length of this word (but it must be greater than
a xed constant only depending on the shift). The same properties for a generic
shift refers to the way in which two dierent patterns in the language of the shift
may appear simultaneously in a conguration. For irreducibility we have that
two patterns always appear simultaneously in some conguration if we translate
their supports. Strong irreducibility states that if the supports of the pattern
are far enough, than it is not necessary to translate them in order to nd a
conguration in which both the patterns appear.
These two irreducibility conditions are not equivalent, not even in the one
dimensional case. Hence our general results about strongly irreducible shifts of
nite type are strictly weaker than the onedimensional ones. In the attempt of
using weaker hypotheses in the latter case, we give a new notion of irreducibility,
the semistrong irreducibility. In the onedimensional case, this property means
that the above word w may almost be of the length we want (provided that it
is long enough): we must allow it to be a little longer or a little shorter; the
length of this dierence is bounded and only depends on the shift. In general,
semistrong irreducibility states that if the supports of the patterns are far
enough from each other, than translating them a little we nd a conguration
in which both the patterns appear. The reason for this choice lies in the fact that
using the Pumping Lemma we can prove that a soc subshift of A
Z
is irreducible
if and only if is semistrongly irreducible. Moreover Myhills property holds for
semistrongly irreducible shifts of nite type of A
if has nonexponential
growth.
6
1. Cellular
Automata
In this chapter we give the notion of a cellular automaton over a nitely gener-
ated group. This notion generalizes the denition given by Ulam [U] and von
Neumann [vN]. In that classical setting, the universe is the lattice of integers
Z
n
of Euclidean space R
n
. The set of states is a nite set A (also called the
alphabet) and a conguration is a function c : Z
n
A. Time t goes on in
discrete steps and represents a transition function : A
Z
n
A
Z
n
(if c is the
conguration at time t, then (c) is the conguration at time t + 1), which is
deterministic and local. Locality means that the new state at a point Z
n
at time t + 1 only depends on the states of certain xed points in the neigh-
borhood of at time t. More precisely, if c is the conguration reached from
the automaton at time t then (c)
|
= (c
|N
), where : A
N
A is a function
dened on the congurations with support the nite set N (the neighborhood
of the point 0 Z
n
), and N
, that
is a function dened on with values in a nite alphabet A. The space A
is
naturally endowed with a metric and hence with an induced topology, this latter
being equivalent to the usual product topology, where the topology in A is the
discrete one. A subset X of A
closeness
and compactness are equivalent), implies that the inverse of an invertible cellular
automaton is also a cellular automaton and allows us to call conjugacy each
invertible local function between two shifts. The invariants are those properties
which are invariant under conjugacy. Finally we extend to a general shift the
notion of irreducibility and of being of nite type.
In Sections 1.4 and 1.5, we recall from [LinMar] the notions of edge shift and
of soc shift, and restate many of the basic properties of these onedimensional
shifts strictly connected with the onedimensional shifts of nite type.
In Section 1.6, we generalize the AmorosoPatts results about the decision
problem of the surjectivity and the injectivity of local maps for onedimensional
cellular automata on the full shift (see [AP] and [Mu]). Our results concern one
dimensional cellular automata over shifts of nite type.
Finally, in Section 1.7 the notion of entropy for a generic shift as dened
by Gromov in [G] is given. We see that this denition involves the existence of
a suitable sequence of sets and we prove that, in the case of nonexponential
growth of the group, these sets can be taken as balls centered at 1 and with
increasing radius. Moreover, we see that the entropy is an invariant of the shifts.
Then, following Lind and Marcus [LinMar, Chapter 4], we recall basic prop-
erties of the entropy in the onedimensional case.
1.1 Cayley Graphs of Finitely Generated Groups
In this section, we give the basic notion of Cayley graph of a nitely generated
group; as we have said, this graphs will constitute the universe of our class
of cellular automata. Moreover, we recall the denition of growth of a nitely
generated group .
Let be a nitely generated group and A a xed nite set of generators for
; then each can be written as
= x
1
i1
x
2
i2
. . . x
n
in
(1.1)
where the x
ij
s are generators and
j
Z.
We dene the length of (with respect to A) as the natural number
||
X
:= min[
1
[ +[
2
[ + +[
n
[ [ is written as in (1.1)
so that is naturally endowed with a metric space structure, with the distance
given by
dist
X
(, ) := |
1
|
X
(1.2)
and
D
X
n
:= [ ||
X
n
8
is the disk of radius n centered at 1. Notice that D
X
1
is the set A A
1
1.
The asintotic properties of the group being independent on the choice of the
set of generators A, from now on we x a set A which is also symmetric (i.e.
A
1
= A) and we omit the index A in all the above denitions.
For each , the set D
n
provides, by left translation, a neighborhood
of , that is the set D
n
= D(, n). Indeed, if D
n
then = with
|| n. Hence dist(, ) = |
1
| = |
1
| n. Conversely, if D(, n)
then |
1
| n (that is
1
D
n
), and =
1
.
Now we dene, for each E and for each n N, the following subsets of
: the nclosure of E
E
+n
:=
_
E
D(, n);
the ninterior of E
E
n
:= E [ D(, n) E;
the nboundary of E
n
E := E
+n
E
n
;
the nexternal boundary of E
+
n
E := E
+n
E
and, nally, the ninternal boundary of E
n
E := EE
n
.
For all these sets, we will omit the index n if n = 1.
The Cayley graph of , is the graph in which is the set of vertices and
there is an edge from to if there exists a generator x A such that = x.
Obviously this graph depends on the presentation of . For example, we may
look at the classical cellular decomposition of Euclidean space R
n
as the Cayley
graph of the group Z
n
with the presentation a
1
, . . . , a
n
[ a
i
a
j
= a
j
a
i
).
If G = (1, c) is a graph with set of vertices 1 and set of edges c, the graph
distance (or geodetic distance) between two vertices v
1
, v
2
1 is the minimal
length of a path from v
1
to v
2
. Hence the distance dened in (1.2) coincides
with the graph distance on the Cayley graph of .
We recall that the function g : N N dened by
g(n) := [D
n
[
which counts the elements of the disk D
n
, is called growth function of (with
respect to A). One can prove that the limit
:= lim
n
g(n)
1
n
9
always exists; if > 1 then, for all suciently large n,
g(n)
n
,
and the group has exponential growth. If = 1, we distinguish two cases.
Either there exists a polynomial p(n) such that for all suciently large n
g(n) p(n),
in which case has polynomial growth, or has intermediate growth (i.e. g(n)
grows faster than any polynomial in n and slower then any exponential function
x
n
with x > 1). Moreover, it is possible to prove that the type of growth is
a property of the group (i.e. it does not depend on the choice of a set of
generators); for this reason we deal with the growth of a group. This notion has
been indipendently introduced by Milnor [Mil], Efremovic [E] and
Svarc [
S]; it
is very useful in the theory of cellular automata.
1.2 Shift Spaces and Cellular Automata
Let A be a nite alphabet; in the classical theory of cellular automata, the
universe is the Cayley graph of the free abelian group Z
n
and a conguration
is an element of A
Z
n
, that is a function c : Z
n
A assigning to each point of
the graph a letter of A. We generalize this notion taking as universe a Cayley
graph of a generic nitely generated group and taking suitable subsets of
congurations in A
.
Let A be a nite set (with at least two elements) called alphabet; on the set
A
(i.e. the set of all functions c : A), we have a natural metric and hence
a topology. This latter is equivalent to the usual product topology, where the
topology in A is the discrete one. By Tychonos theorem, A
is also compact.
An element of A
is called a conguration.
If c
1
, c
2
A
)
|
:= c
|
for each c A
= X).
For every X A
and E , we set
X
E
:= c
|E
[ c X;
a pattern of X is an element of X
E
where E is a nonempty nite subset of .
The set E is called the support of the pattern; a block of X is a pattern of X
with support a disk. The language of X is the set L(X) of all the blocks of X.
If X is a subshift of A
Z
, a conguration is a biinnite word and a block of X
is a nite word appearing in some conguration of X.
Hence a pattern with support E is a function p : E A. If , we have
that the function p : E A dened as p
|
= p
|
(for each E), is the
pattern obtained copying p on the translated support E. Moreover, if X is
a shift, we have that p X
E
if and only if p X
E
. For this reason, in the
sequel we do not make distinction between p and p (when the context makes it
possible). For example, a word a
1
. . . a
n
is simply a nite sequence of symbols
for which we do not specify (if it is not necessary), if the support is the interval
[1, n] or the interval [n, 1].
Denition 1.2.2 Let X be a subshift of A
; a function : X A
is Mlocal
if there exists : X
DM
A such that for every c X and
((c))
|
= ((c
)
|DM
) = (c
|1
, c
|2
, . . . , c
|m
),
where D
M
=
1
, . . . ,
m
.
In this denition, we have assumed that the alphabet of the shift X is the
same as the alphabet of its image (X). In this assumption there is no loss of
generality because if : X A
, D
M
is the
neighborhood of 1 and : X X is an Mlocal function.
Let : X A
) = (c)
).
11
Proof Suppose that is Mlocal and that D
M
=
1
, . . . ,
m
; then for
and c X,
((c
))
|
= ((c
)
|1
, (c
)
|2
, . . . , (c
)
|m
) = (c
|1
, . . . , c
|m
).
On the other hand
(((c))
)
|
= ((c))
|
= (c
|1
, . . . , c
|m
).
Hence commutes with the action. We prove that is continuous. A generic
element of a subbasis of A
is
:= c A
[ c
|
= a
with and a A. It suces to prove that the set
:=
1
() = c X [ (c)
|
= a
is open in X. Actually, if c
and
r := max(|
1
|, . . . , |
m
|), (1.3)
we claim that the ball B
X
(c,
1
r+1
) is contained in
. Indeed, if c B
X
(c,
1
r+1
)
then
dist(c, c) <
1
r + 1
and on D
r
we have c
|Dr
= c
|Dr
; since, from (1.3),
i
D
r
, we have ( c)
|
=
( c
|1
, . . . , c
|m
) = (c
|1
, . . . , c
|m
) = (c)
|
= a.
Conversely, suppose that is continuous and commutes with the action of
. Since X is compact, is uniformly continuous; x M in N such that for
every c, c X,
dist(c, c) <
1
M + 1
dist((c), ( c)) < 1.
Hence, if c and c agree on D
M
, then (c) and ( c) agree at 1, that is, for every
c X,
((c))
|1
= (c
|1
, . . . , c
|m
)
where D
M
=
1
, . . . ,
m
.
In general, since commutes with the action of , we have
((c))
|
= ((c))
|1
= (((c))
)
|1
= ((c
))
|1
= (c
|1
, . . . , c
|m
)
showing that is Mlocal. 2
From the previous theorem, it is clear that the composition of two local
functions is still local. In any case, this can be easily seen by a direct proof that
follows Denition 1.2.2.
12
Now, x and consider the function X A
,= (c
|
= c
|
that is c
and c
agree on D
n
so that
dist(c
, c
) <
1
n+1
.
This result will be necessary in the proof of Theorem 1.2.6, below.
Observe that if X is a subshift of A
and : X A
is a local function,
then, by Proposition 1.2.4, the image Y := (X) is still a subshift of A
. Indeed
Y is closed (or, equivalently, compact) and invariant:
Y
= ((X))
= (X
) = (X) = Y.
Moreover, if is injective then : X Y is a homeomorphism; if c Y then
c = ( c) for a unique c X and we have
1
(c
) =
1
(( c)
) =
1
(( c
)) = c
= (
1
(c))
that is,
1
commutes with the action. By Proposition 1.2.4,
1
is local.
Hence the wellknown theorem (see [R]), stating that the inverse of an invert-
ible Euclidean cellular automaton is a cellular automaton, holds also in this
more general setting. In the onedimensional case, Lind and Marcus [LinMar,
Theorem 1.5.14] give a direct proof of this fact.
This result leads us to give the following denition.
Denition 1.2.5 Two subshifts X, Y A
nN
A
Dn
such that X = X
F
, where
X
F
:= c A
[ c
|Dn
/ T for every , n N.
In this case, T is a set of forbidden blocks of X.
Proof Suppose that X is a shift. X being closed, for each c / X there exists
an integer n
c
> 0 such that the ball
B
A
(c,
1
n
c
) X.
13
Let T be the set
T := c
|Dnc
[ c / X;
we prove that X = X
F
. If c / X
F
, there exists c / X such that c
|Dn
c
= c
|Dn
c
for some . Then dist(c
, c) <
1
n c
and hence c
B
A
( c,
1
n c
) X which
implies c / X, by the invariance of X. This proves that X X
F
. For the
other inclusion, if c / X then, by denition of T, c
|Dnc
T and hence c / X
F
.
Now, for the converse, we have to prove that a set of type X
F
is a shift.
Observe that
X
F
=
pF
X
{p}
and, if supp(p) = D
n
=
1
, . . . ,
N
,
X
{p}
=
c A
[ c
|Dn
,= p =
_
_
iDn
c A
[ c
|i
,= p
|i
_
.
Thus, in order to prove that X
F
is closed, it suces to prove that for any
i = 1, . . . , N the set
c A
[ c
|i
,= p
|i
(1.4)
is closed. We have
(c A
[ c
|i
,= p
|i
)
= c A
[ c
|i
,= p
|i
and then the set in (1.4) is closed being the preimage of a closed set under a
continuous function. Finally we have to prove that X
F
is invariant. If
and c X
F
, for every and every n N we have c
|Dn
/ T that is
(c
|Dn
/ T and hence c
X
F
2
Now we give the rst, fundamental notion of irreducibility for a onedimen-
sional shift and we see how to generalize this notion to a generic shift.
Denition 1.2.7 A shift X A
Z
is irreducible if for each pair of words u, v
L(X), there exists a word w L(X) such that the concatenated word uwv
L(X).
The natural generalization of this property to any group is the following.
Denition 1.2.8 A shift X A
.
Proposition 1.3.3 Let X be an Mstep shift of nite type and let E be a sub-
set of . If c
1
, c
2
X are two congurations that agree on
+
2M
E, then the
conguration c A
|DM
/ T. Observe that either
D
M
E
+2M
or D
M
E;
15
indeed, if E
+M
then by denition D
M
(E
+M
)
+M
= E
+2M
. If / E
+M
,
suppose that there exists D
M
E. Then, for some i, =
i
and hence
=
1
i
D
M
E
+M
which is excluded; then D
M
E.
Now, if is such that D
M
E
+2M
, we have c
|DM
= (c
1
)
|DM
/ T. If
is such that D
M
E, we have c
|DM
= (c
2
)
|DM
/ T. 2
Corollary 1.3.4 Let X be an Mstep shift of nite type and let E be a nite
subset of ; if p
1
, p
2
X
E
+2M are two patterns that agree on
+
2M
E, than there
exist two extensions c
1
, c
2
X of p
1
and p
2
, respectively, that agree on E.
Proof Indeed, if c
1
and c
2
are two congurations extending p
1
and p
2
re-
spectively, then they agree on
+
2M
E. The conguration c of Proposition 1.3.3
and the conguration c
2
, coincide on E and they are extensions of p
1
and p
2
,
respectively. 2
1.4 Edge Shifts
A relevant class of onedimensional subshifts of nite type, is that of edge shifts.
This class is strictly tied up the one of nite graphs. This relation allows us
to study the properties of an edge shift (possible quite complex) studying the
properties of its graph. On the other hand, each onedimensional shift of nite
type is conjugate to an edge shift and hence they have the same invariants;
thus also in this case the properties of the shift depend on the structure of the
accepting graph.
Denition 1.4.1 Let G be a nite directed graph with edge set c. The edge
shift X
G
is the subshift of c
Z
dened by
X
G
:= (e
z
)
zZ
[ t(e
z
) = i(e
z+1
) for all z Z
where, for e c, i(e) denotes the initial vertex of e and t(e) the terminal one.
The structure of a nite graph (and hence of an edge shift) can be easily
described by a matrix, the socalled adjacency matrix of G.
More precisely, let G be a graph with vertex set 1 = 1, . . . , n; the adja-
cency matrix of G is the matrix A such that A
ij
is the number of edges in G
with initial state i and terminal state j.
It is easily seen that the (i, j)entry of the matrix A
m
(the product of A
with itself m times), is the number of paths in G with length m from i to j.
The edge shift X
G
is sometimes denoted as X
A
, where A is the adjacency
matrix of G. The fundamental role of the adjacency matrix will be claried in
the sequel.
Concerning the edge shifts, it is easy to see that every edge shift is a 1step
shift of nite type with set of forbidden blocks ef [ t(e) ,= i(f). On the other
hand, each shift of nite type is conjugate to an edge shift accepted by a suitable
graph G. Now we give an eective procedure to construct it.
16
Let X be a shift of nite type X with memory M, the vertices of G are the
words of L(X) of length M and the edges are the words of L(X) of length M+1.
There is an edge named a
1
. . . a
M
a
M+1
L(X) from a
1
. . . a
M
to a
2
. . . a
M+1
.
a
1
. . . a
M
a
2
. . . a
M+1
q
a
1
. . . a
M
a
M+1
The edge shift accepted by this graph is the (M + 1)th higher block shift of
X and is denoted by X
[M+1]
. Consider the function : X
[M+1]
X dened
setting, for each c X
[M+1]
and each z Z, (c)
|z
equal to the rst letter of
the word c
|z
; is a bijective local function and hence a conjugacy.
. . . a
1
a
0
. . . a
M
a
0
a
1
. . . a
M+1
a
1
a
2
. . . a
M+2
. . .
. . . a
1
a
0
a
1
. . .
The above table points out the behavior of the function .
1.5 Soc Shifts
The class of soc shifts has been introduced by Weiss in [Wei1] as the smallest
class of shifts containing the shifts of nite type and closed under factorization
(i.e. the image under a local map of a soc shift is soc). Equivalently, one
can see that a shift is soc if it is the set of all labels of the biinnite paths
in a nite labeled graph (or nitestate automaton). In automata theory, this
corresponds to the notion of regular language. Indeed a language (i.e. a set of
nite words over a nite alphabet) is regular if it is the set of all labels of nite
paths in a labeled graph.
Denition 1.5.1 Let A be a nite alphabet; a labeled graph ( is a pair (G, L),
where G is a nite graph with edge set c, and the labeling L : c A assigns
to each edge e of G a label L(e) from A.
We recall that a nitestate automaton is a triple (A, Q, ) where Ais a nite
alphabet, Qis a nite set whose elements are called states and : QA T(Q)
is a function dened on the Cartesian product QA with values in the family
of all subsets of Q. Suppose that
Q (Q, a); this means that when the auto-
maton is in the state Q and reads the letter a, then it can pass to the
state
Q. The automaton is deterministic when : Q A Q , that is
17
the state
Q above (if it exists) is determined from Q and a; the automaton is
complete when : QA T(Q), that is for each state Q and each letter
a there is at least a state
Q such that
Q (Q, a).
Clearly each general nitestate automaton determines a labeled graph set-
ting 1 := Q and drawing an edge labeled a from Q to
Q if and only if
Q (Q, a). Also the converse holds: given a labeled graph ( with set of
vertices 1, we can set Q := 1 and (i, a) is the set of all vertices j of the graph
for which there is an edge e such that i(e) = i, t(e) = j and L(e) = a.
In the sequel we will not distinguish between a labeled graph and the related
nitestate automaton.
If = . . . e
1
e
0
e
1
. . . is a conguration of the edge shift X
G
, dene the label
of the path to be
L() = . . . L(e
1
)L(e
0
)L(e
1
) A
Z
.
The set of labels of all congurations in X
G
is denoted by
X
G
:= c A
Z
[ c = L() for some X
G
= L(X
G
).
Denition 1.5.2 A shift X A
Z
is soc if X = X
G
for some labeled graph (.
A presentation of a soc shift X is a labeled graph ( for which X
G
= X.
Obviously each edge shift is soc. Indeed if G is a graph with edge set c,
we can consider the identity function on c as a labeling L : c c so that
X
G
= X
(G,idE)
. Moreover, each shift of nite type is soc. Consider the graph
Gintroduced in the previous section that accepts the edge shift X
[M+1]
; labeling
the edges of G setting L(a
1
. . . a
M
a
M+1
) = a
1
(that is L is the previous function
: X
[M+1]
X), it can be easily seen that the labeled graph ( so obtained is
such that X = X
G
.
a
1
. . . a
M
a
2
. . . a
M+1
a
3
. . . a
M+2
j
a
1
j
a
2
Notice that in a graph G (or in the labeled graph ( having G as support),
there can be a vertex from which no edges start or at which no edges end.
Such a vertex is called stranded. Clearly no biinnite paths in X
G
(or in X
G
),
involve a stranded vertex, hence the stranded vertices and the edges starting or
ending at them are inessential for the edge shift X
G
or for the soc shift X
G
.
18
Following Lind and Marcus [LinMar, Denition 2.2.9], a graph is essential if no
vertex is stranded. Removing step by step the stranded vertices of G, we get
an essential graph
G that gives rise to the same edge shift. This procedure is
eective, because G has a nite number of vertices. Moreover, this essential
form of G is unique. This property of the edge shift X
G
holds true for the
soc shift X
G
: the labeled graph
( = (
G, L
|E(
G)
) is such that X
G
= X
G
(indeed
X
G
= L(X
G
) = L(X
G
) = X
G
).
If we deal only with essential graphs, it is easy to see that the language of a
soc shift is regular.
Among the subshifts of A
Z
, the soc shifts are the images under a local
function (briey factors) of a shift of nite type. Indeed each labeling is a local
function dened on an edge shift and, on the other hand, the factor of a shift
of nite type is also the image under a local function of a suitable edge shift
(each shift of nite type being conjugate to an edge shift). This characterization
allows us to call soc a shift X A
nN
E
n
= and
lim
n
[
M
E
n
[
[E
n
[
= 0; (1.5)
if X A
M
En|
|En|
= 0; moreover it is equivalent to the
condition lim
n
|En|
|En|
= 0.
If X is a subshift of A
Z
, we choose as E
n
the interval [1, n] (or equivalently,
in order to have
nN
E
n
= , the interval [n, n]), so that X
En
is the set of
the words of X of length n. One can prove that in this onedimensional case,
the above maximum limit is a limit. Indeed [X
En+m
[ [X
En
[[X
Em
[ and hence
setting a
n
:= log [X
En
[, we have that the sequence (a
n
)
nN
is subadditive,
namely a
n+m
a
n
+a
m
. We want to prove that
lim
n
a
n
n
= inf
n1
a
n
n
.
22
Fix m 1; then there exist q, r N such that n = mq + r and we have
a
n
qa
m
+a
r
so that
an
n
qam
n
+
ar
n
. Now
lim
n
_
qa
m
n
+
a
r
n
_
= lim
n
_
q
n
a
m
+
a
r
n
_
=
a
m
m
and then limsup
n
an
n
am
m
. Hence
limsup
n
a
n
n
inf
m1
a
m
m
liminf
n
a
n
n
.
In general, if is a group of nonexponential growth, we choose as E
n
a suitable disk centered at 1 ; indeed, setting a
h
= [D
h
[, we have that
lim
h
h
a
h
= 1 hence liminf
h
a
h+1
a
h
= 1. From this fact it follows that for
a suitable sequence (a
h
k
)
k
we have that lim
k
a
h
k
+1
a
h
k
= 1. Hence
liminf
k
a
h
k
+1
a
h
k
1
= lim
k
a
h
k
+1
a
h
k
liminf
k
a
h
k
a
h
k
1
= 1.
Then, for a suitable sequence (a
h
kn
)
n
we have lim
n
a
h
kn
+1
a
h
kn
1
= 1, that is we
nd a sequence of disks E
n
:= D
h
kn
such that
lim
n
[E
+
n
[
[E
n
[
= 1.
Hence
|En|
|En|
=
|E
+
n
\E
n
|
|En|
|E
+
n
\E
n
|
|E
n
|
=
|E
+
n
|
|E
n
|
1
n
0.
In the following theorem we prove that the entropy is an invariant of the
shift.
Theorem 1.7.1 Let X be a shift and : X A
E
+M
n
: X
E
+M
n
Y
En
is surjective and hence
[Y
En
[ [X
E
+M
n
[ [X
En
[[X
+
M
En
[ [X
En
[[A
+
M
En
[.
From the previous inequalities we have
log [Y
En
[
[E
n
[
log [X
En
[
[E
n
[
+
[
+
M
E
n
[ log [A[
[E
n
[
and hence, taking the maximum limit, ent(Y ) ent(X). 2
Now we see how to compute the entropy for an irreducible soc subshift of
A
Z
.
23
Let X be a soc shift; given a presentation ( = (G, L) of X, there is an
eective procedure to construct, from ( a deterministic presentation
( of X.
This procedure is called subset construction. The vertices of
( are the non
empty subsets of the vertices of ( and there is an edge in
( labeled a from I to
J if J is the set of terminal vertices of edges in ( starting at some vertex in I
and labeled a.
Using combinatorial methods, it is easy to see that the entropy of a soc shift
X coincides with the entropy of the edge shift X
G
of a deterministic presentation
of X, as stated in the following theorem.
Proposition 1.7.2 [LinMar, Proposition 4.1.13] Let X be a soc shift and let
( = (G, L) be a deterministic presentation of X. Then ent(X) = ent(X
G
).
Given a nite graph G, it is obvious that the adjacency matrix Aof Gis non
negative and, by the PerronFrobenius Theorem, has a maximum eigenvalue
A
> 0, the socalled Perron eigenvalue of A. This eigenvalue gives us a way
to compute the entropy of an irreducible soc shift, as stated in the following
theorem.
Proposition 1.7.3 [LinMar, Theorem 4.3.3] Let X be an irreducible soc shift
and let ( = (G, L) be a strongly connected deterministic presentation of X.
Then ent(X) = log(
A
).
If X and ( are as in the above theorem, it is clear that the graph G is a
strongly connected deterministic presentation of X
G
. Hence we have another
proof of Proposition 1.7.2 (as far as irreducible soc shifts are concerned).
In fact, in [LinMar, Section 4.4] the above results are used to give a general
method of computing the entropy of a generic soc shift.
A fundamental result given in that section whose proof is based on the
proposition above, is the following theorem.
Theorem 1.7.4 [LinMar, Corollary 4.4.9] If X is an irreducible soc shift and
Y is a proper subshift of X, then ent(Y ) < ent(X).
In Chapter 3, we will prove a theorem of this kind in a much more general
setting.
24
2.
Surjunctivity and
Density of Periodic
Configurations
A selfmapping : X X on a set X is surjunctive if it is either noninjective or
surjective. In other words a function is surjunctive if the implication injective
surjective holds. In this chapter, we consider the surjunctivity of the transition
function in a general cellular automaton over a group .
A conguration of a shift is periodic if its orbit is nite; after proving in
Section 2.2 some generalities about periodic congurations, we recall in Section
2.3 the class of residually nite groups, proving that a group is residually
nite if and only if the periodic congurations are dense in A
.
Hence if is a residually nite group, a transition function of a cellular
automaton on A
iI
X
i
(X
i
) X
i
|Xi
: X
i
X
i
is surjunctive
then is surjunctive.
Proof If is injective then, for every i I, the restriction
|Xi
is also injec-
tive; by the hypotheses we have (X
i
) = X
i
and hence
iI
X
i
=
iI
(X
i
) =
(
iI
X
i
) (
iI
X
i
) = (X). Then X =
iI
X
i
(X), and being
closed we have X (X). 2
2.2 Periodic Congurations of a Shift
In this section we point out the fundamental subset of the periodic congura-
tions of A
= c
[
consists of n elements; in this case n is the period of c. A conguration
is periodic if it is nperiodic for some n N.
26
From now on, P
n
denotes the set of the periodic congurations whose period
divides n and C
p
is the set
n1
P
n
of all the periodic congurations.
In general, a conguration c A
belongs to P
n
if and only if there
exists a subgroup H of with nite index dividing n, such that c is constant on
the right cosets of H.
Proof Let c be mperiodic with m[n; by denition, the stabilizer H
c
has
nite index m and c is constant on the right cosets of H
c
. Conversely, if H has
nite index dividing n and c is constant on the right cosets of H, we prove that
H H
c
. Indeed, if h H and we have
(c
h
)
|
= c
|h
= c
|
and hence c
h
= c so that h H
c
. H being of nite index, H
c
also has nite
index and the index of H
c
divides that of H so that it divides n. Hence c P
n
.
2
The following result is well known (see, for example, [Rot]).
Lemma 2.2.3 Let be a nitely generated group; for every n N there are
nitely many subgroups of of nite index n.
Proof If H has nite index n, x a set
1
, . . . ,
n
of right coset
representatives of H and consider the function : S
n
from to the
symmetric group on n elements dened by:
(
i
) =
j
where H
i
= H
j
, that is
H
i
= H
(
i
).
This function is a group homomorphism, indeed from the above equality
H
i
= H
(
i
) = H
(
(
i
));
on the other hand we have, by denition,
H
i
= H
(
i
)
27
and then
=
.
Notice that ker() H. Indeed, if
= id
Sn
, then
(1) = 1 that is
H = H.
Now, the subgroups of containing ker() are as many as the subgroups of
/ ker() which is a nite group.
On the other hand, there are nitely many homomorphisms from to S
n
because such an homomorphism is completely determined by its value on the
(nitely many) generators of . 2
Corollary 2.2.4 The set P
n
is nite.
Proof By Proposition 2.2.2, a conguration c X belongs to P
n
(X) if
and only if it is constant on the right cosets of a subgroup H with nite index
dividing n. By Lemma 2.2.3, these subgroups H are in nite number. For a
xed H, there are nitely many functions from the right cosets of H to A, that
is [A[
[:H]
. 2
2.3 Residually Finite Groups
In this section we deal with the basic properties of residually nite groups. We
will see that the (nitely generated) residually nite groups are precisely those
groups such that for each nite set A, the set C
p
of periodic congurations is
dense in A
.
Denition 2.3.1 A group is residually nite if for every with ,= 1,
there exists H of nite index such that / H.
In other words, a group if residually nite if
H
[:H]<
H = 1.
Examples of residually nite groups are the groups Z
n
(n = 1, 2, . . . ) and, in
general, all nitely generated abelian groups. The free group F
n
of rank n is an
example of residually nite, nonabelian group. The additive group of rational
numbers Q is an example of abelian, nonnitely generated and nonresidually
nite group.
The proofs of the following Lemma 2.3.2 and Theorem 2.3.3 are due to T.
CeccheriniSilberstein and A. Mach`.
Lemma 2.3.2 If is a residually nite group and F =
1
, . . . ,
n
is a nite
subset of with 1 / F, then there exists a subgroup H
F
of nite index such
that
i
/ H
F
and H
F
i
,= H
F
j
(if i ,= j).
28
Proof For every i = 1, . . . , n let H
i
be a subgroup of nite index such
that
i
/ H
i
and let H
ij
be a subgroup of nite index such that
i
1
j
/ H
ij
(where i ,= j). The intersection H
F
of all these subgroups also has nite index;
moreover
i
/ H
F
(for every i) and
i
1
j
/ H
F
(i ,= j). 2
In particular, the set F of this Lemma can be extended to a set of right coset
representatives of the subgroup H
F
.
Theorem 2.3.3 Let be a nitely generated group and A a nite alphabet. If
is residually nite, then the set C
p
of periodic congurations is dense in A
.
Proof Suppose that is residually nite; we have to prove that
A
= C
p
.
Fix c A
and let H
Dn
be the subgroup of nite index whose existence is
guaranteed by Lemma 2.3.2 with F := D
n
1, and let D be a set of right coset
representatives of H
Dn
containing D
n
. If and = hd with h H
Dn
and
d D, dene a conguration c
n
such that (c
n
)
|
= c
|d
. This conguration being
constant on the right cosets of H
Dn
, is periodic. Moreover c and c
n
agree on
D
n
and hence dist(c, c
n
) <
1
n+1
. Then the sequence of periodic congurations
(c
n
)
n
converges to c. 2
The same result is also given by Yukita [Y]. The converse of this theorem
also holds.
Theorem 2.3.4 Let be a nitely generated group and A a nite alphabet.
Then is residually nite if and only if the set C
p
of periodic congurations is
dense in A
.
Proof If is not residually nite, let 1 ,= be an element belonging
to all the subgroups of of nite index; in particular
cCp
H
c
so that,
for c C
p
, c
= c and hence c
|
= c
|1
. Let c A
such that c
|
,= c
|1
, then
for each n such that D
n
and each c C
p
we have c
|Dn
,= c
|Dn
. Hence
dist( c, c)
1
n+1
and c / C
p
. 2
2.4 Density of Periodic Congurations
In this section we prove that the density of the periodic conguration is a su-
cient (but not necessary) condition for the surjunctivity of the transition func-
tion in a cellular automaton. By Theorem 2.3.4, we have that from the residual
niteness of it follows that a transition function : A
on a full shift
surjunctive. The groups Z
n
being residually nite, we have that this result
generalizes Richardsons theorem.
Theorem 2.4.1 Let X A
is a transi-
tion function, then is surjunctive.
In general, the implication injective surjective of Theorem 2.4.1 is not
invertible; the standard example is the following. Let A = 0, 1 and = Z; let
be the local function given by the local rule : A
3
A such that
(a
1
, a
2
, a
3
) = a
1
+a
3
(mod2).
Then is surjective and not injective. Indeed if (a
z
)
zZ
is a conguration in
A
Z
, dene a preimage of it in this way:
_
_
b
0
= 0
b
1
= 0
b
n+1
= a
n
b
n2
(mod2) if n 2
b
n
= a
n+1
b
n+2
(mod2) if n 0
that is
. . . a
2
a
0
a
1
a
0
0 0 a
1
a
2
a
3
a
1
a
4
a
2
. . .
. . . a
3
a
2
a
1
a
0
a
1
a
2
a
3
a
4
a
5
. . .
.
With b
0
= 1 = b
1
we can construct a preimage in an analogous way.
Proposition 2.4.3 If X A
1 constant in 0
and 1 and the congurations of the type . . . 1111110000000 . . . ; clearly X is not
irreducible because there are no words u L(X) such that 0u1 L(X). In this
31
shift we have C
p
(X) =
0,
n
i=1
c X [ c
|
i
= p
|
i
which
is a nite intersection of open sets.
Even if we strengthen the irreducibility hypothesis by assuming that the
shift is mixing, there are examples of a mixing subshifts of nite type X and
of local functions : X X which are injective and not surjective (see [Wei2,
Section 4]).
2.6 Group Shifts
If the alphabet is a nite group G, the full shift G
is a group shift if is
a subshift and a subgroup of G
.
32
Clearly a group shift is also a compact (metric) group. Hence it can be seen
as an example of dynamical system (X, ), where X is a compact group and
is a subgroup of the group Aut(X) of the automorphisms of X which are also
continuous. Indeed the action of denes a subgroup of Aut(X): for a xed
, the bijective function c c
|
= (c
1
c
2
)
|
= c
1|
c
2|
= c
1
|
c
2
|
= (c
1
c
2
)
|
.
If (X, ) is a dynamical system, the group acts expansively on X if there
exists a neighborhood U of the identity in X such that
(U) = 1
X
; the
set of periodic points is the set of points x X such that (x) [ is
nite. Clearly it coincides with the set C
p
(X) if X is a group shift.
In the hypotheses that X is metrizable and is an innite and nitely gen-
erated abelian group, Kitchens and Schmidt prove in [KitS2, Theorem 3.2] that
if acts expansively on X then (X, ) satises the descending chain condition
(i.e. each nested decreasing sequence of closed invariant subgroups is nite),
if and only if (X, ) is conjugate to a dynamical system (Y, ), where Y is a
group subshift of G
; indeed
the identity in X is the conguration c
1
assuming the constant value 1
G
, where
1
G
is the identity of G. Consider the neighborhood X
1
of c
1
consisting of all
those congurations of X assuming the value 1
G
at 1
. Obviously
[
c X
1
=
c X [ c
|
= 1
G
= c
1
. 2
In [KitS2] is also proved that if X is a group shift, then X is of nite type.
Indeed the following theorem is proved.
Theorem 2.6.4 [KitS2, Corollary 3.9] Let G be a compact Lie group. If X
G
[ c
|D
H for every ,
where H is a closed subgroup of G
D
.
33
Hence if G is nite and X is a group shift, the set G
D
H is nite and is a set
of forbidden blocks for X. Although this fact, X is not necessarily irreducible.
For example, consider in (Z/2Z)
Z
the group shift
0,
1, 01, 10.
Notice that an abelian, nitely generated group is also residually nite; we
have another proof of this fact xing a nite group G and applying Corollary
2.6.3 to the group shift G
is preinjective if
whenever two congurations c, c X dier only on a nite nonempty subset
of , then (c) ,= ( c)); this notion has been introduced by Gromov in [G].
In fact, we prove in Section 3.1 that these two properties are equivalent for
local functions dened on the full shift, but in the case of proper subshifts the
former may be meaningless. On the other hand, we give a proof of the fact that
the nonexistence of GOE patterns is equivalent to the nonexistence of GOE
congurations, that is to the surjectivity of the transition function. Hence, in
this language, the GOE theorem states that is surjective if and only if it is pre
injective. We call Moores property the implication surjective preinjective
and Myhills property the inverse one.
36
In Section 3.2, we recall the notion of an amenable group. We give the more
useful characterization of amenability in terms of Flners sequences. It will
follow from Section 1.7 that, if the group has nonexponential growth, the
Flners sequence can be taken as a suitable sequence of disks centered at 1 .
Concerning onedimensional shifts, in Section 3.3 we see that from the works
of Hedlund [H] and Coven and Paul [CovP] it follows that the MooreMyhill
(MM) property holds for irreducible shifts of nite type of A
Z
. Moreover, using
this result we prove that Myhills property holds for irreducible soc shifts of
A
Z
. On the other hand, we give a counterexample of an irreducible soc shift
X A
Z
but not of nite type for which Moores property does not hold.
Concerning general cellular automata over amenable groups, we see in Sec-
tion 3.4 that from a result of Gromov [G] in a more general framework, it follows
that the MMproperty holds for shifts of bounded propagation contained in A
.
In Section 3.5, we generalize this result showing that the MMproperty holds
for strongly irreducible shifts of nite type of A
if
has nonexponential growth.
37
3.1 The Garden of Eden Theorem and the Moo-
reMyhill Property
In this section we give the denition of a preinjective function, proving that
this property is equivalent to the notion of nonexistence of mutually erasable
patterns used in the original works of Moore [Moo] and Myhill [My]. Indeed they
prove that a transition function of a Euclidean cellular automaton on a full shift
admits two mutually erasable patterns if and only if it admits a Garden of Eden
pattern, that is a pattern without preimage. A result of this kind, concerning a
local function between two shifts, is what we call a Garden of Eden theorem;
we deal with the MooreMyhill property when we have a shift such that for each
transition function of a cellular automaton over it preinjectivity is equivalent
to surjectivity.
Denition 3.1.1 Let : A
such that c
1|F
= p
1
, c
2|F
= p
2
and c
1|F
= c
2|F
, one
has (c
1
) = (c
2
).
Denition 3.1.2 Let X, Y A
but in a sub-
shift X A
is surjective if and only if there are no GOE patterns. It is easy to prove that
this property holds also for the local functions between shifts, as proved in the
following Proposition.
Proposition 3.1.7 Let : X Y a local function; then is surjective if and
only if there are no GOE patterns.
Proof It is clear that the surjectivity of implies the nonexistence of GOE
patterns.
For the converse, suppose that for each nite set E and each p Y
E
there is a conguration c X such that (c)
|E
= p; we prove that is surjective.
39
If c Y , let c
n
X be such that (c
n
)
|Dn
= c
|Dn
; hence lim
n
(c
n
) = c. X
being compact, there is a subsequence (c
n
k
)
k
that converges to a conguration
c X. being continuous, we have that c = lim
k
(c
n
k
) = (c). 2
Hence we can restate the Garden of Eden Theorem [Moo] and [My] as follows.
Theorem 3.1.8 If : A
Z
2
A
Z
2
is a transition function, then is pre
injective if and only if it is surjective.
Denition 3.1.9 A shift X A
(see [CeMaSca]).
Denition 3.2.1 A group is called amenable if it admits a invariant prob-
ability measure, that is a function : P() [0, 1] such that for A, B and
for every
A B = (A B) = (A) +(B) (nite additivity)
(A) = (A) (invariance)
40
() = 1 (normalization).
Theorem 3.2.2 (Flner) A group is amenable if and only if for each nite
subset F and each > 0 there exists a nite subset K such that
[KFK[
[K[
< .
This characterization is equivalent to the following one.
For each pair of nite subsets F, H with 1 H and each > 0 there exists
a nite subset K H such that
[KFK[
[K[
< .
Indeed, suppose that there exists
K such that
[
KHF
K[
[
K[
< .
We have that
K
KH and hence
[
KHF
KH[
[
KH[
[
KHF
K[
[
K[
< ;
it suces to set K :=
KH.
An analogous equivalent form of Flners characterization is given by Namioka
in [N].
Theorem 3.2.3 Let be an amenable group; then there exists a sequence of
nite sets (E
n
)
n1
such that:
E
1
E
2
. . . E
n
. . .
n1
E
n
= ,
lim
n
|MEn|
|En|
= 0.
Proof First, notice that in Flner condition there is no loss of generality
if we suppose 1 K. Now we construct, by induction, a nested sequence
1 K
1
. . . K
n
. . . such that, for each n 1
[K
n
(D
+M
n
)K
n
[
[K
n
[
<
1
n
.
Let K
1
be a nite subset 1 K
1
such that
[K
1
(D
+M
1
)K
1
[
[K
1
[
< 1
41
whose existence is guaranteed by Theorem 3.2.2. Suppose to have found K
n
,
there exists K
n+1
K
n
such that
[K
n+1
(D
+M
n+1
)K
n+1
[
[K
n+1
[
<
1
n + 1
.
Observe that
K
n
(D
+M
n
) = (K
n
D
n
)
+M
K
n
K
n
(D
M
n
) (K
n
D
n
)
M
K
n
K
n
D
n
hence
[(K
n
D
n
)
+M
(K
n
D
n
)
M
[
[K
n
D
n
[
[K
n
(D
+M
n
)K
n
[
[K
n
[
<
1
n
.
Setting E
n
:= K
n
D
n
we have the stated properties because D
n
K
n
. 2
A sequence as in Theorem 3.2.3 is called amenable (or Flner sequence); from
now on we x the amenable sequence (E
n
)
n1
found above and the entropy of a
shift will be dened with respect to (E
n
)
n1
. As we have seen in Section 1.7, if
the group has nonexponential growth, the sequence (E
n
)
n1
can be replaced
by a suitable sequence of disks centered at 1.
Corollary 3.2.4 If (E
n
)
n1
is an amenable sequence, then
lim
n
[
+
M
E
n
[
[E
n
[
= 0
and
lim
n
[
M
E
n
[
[E
n
[
= 0.
Using the existence of an amenable sequence in the amenable group ,
CeccheriniSilberstein, Mach` and Scarabotti have proved the following theo-
rem for the full shift A
.
Theorem 3.2.5 [CeMaSca, Corollary of Theorem 3] Let be a nitely gen-
erated amenable group and : A
2,
where
2 is the biinnite word constant in 2. The set
X is a shift of nite type
over the alphabet
A = 0, 1, 2 with set of forbidden blocks:
02, 20, 12, 21.
Moreover
X is not irreducible; indeed we have 1, 2 L(
X) but for no word
w L(
X) the word 1w2 belongs to L(
X).
Consider the transition function :
X
X dened by:
(c) =
_
c if c X
0 if c =
2.
Clearly is 1local where the local rule is given by (a) = a if a ,= 2 and
(2) = 0. This function is not surjective because the word
2 has no preimages,
but it is preinjective. Actually, if c
1
and c
2
are dierent congurations which
only dier on a nite subset of Z, then they must belong to X and so their
images under are dierent. 2
Counterexample 3.3.4 Mooreproperty no longer holds for a shift of nite
type but not irreducible.
43
Let X be the shift over the alphabet A = 0, 1, 2 with set of forbidden
blocks 01, 02. The shift X is not irreducible; indeed for no word u L(X)
the word 0u1 belongs to L(X).
Consider the transition function : X X dened by the local rule:
(a
1
a
2
a
3
) =
_
a
2
if a
3
,= 0
0 if a
3
= 0.
The function is surjective because a generic word of X, for example,
. . . 1211122121212212121 0 0000000000000 . . .
has two preimages:
. . . 1211122121212212121 1 0000000000000 . . .
and
. . . 1211122121212212121 2 0000000000000 . . .
This also shows that is not preinjective. 2
Using the generalizations to dimension 2 of Counterexamples 3.3.3 and 3.3.4,
we obtain two irreducible shifts of nite type of A
Z
2
which give the following
counterexamples.
Counterexample 3.3.5 MMproperty no longer holds for an irreducible shift
of nite type contained in A
Z
2
.
Consider the cellular automaton of Counterexample 3.3.3; it is clear that the
1local function can be extended to a 1local function
2
:
X
2
X
2
with
the same local rule ((a) = a if a ,= 2 and (2) = 0). The function
2
is, as ,
preinjective and nonsurjective.
For the cellular automaton of Counterexample 3.3.4, we have similarly that
the extended function
2
is surjective and not preinjective 2.
We now prove that a result similar to Theorem 3.3.1 holds for irreducible
soc shifts.
Theorem 3.3.6 Let X be an irreducible soc shift, Y a shift and : X Y a
local function. Let ( = (G, L) be the minimal deterministic presentation of X.
Then L is preinjective if and only if ent(X) = ent((X)).
Proof The labeled graph ( = (G, L) being a presentation of X, we have
X = X
G
= L(X
G
). By Corollary 1.5.8, X
G
is an irreducible shift of nite
type. Moreover L : X
G
Y is a local function; thus, by Theorem 3.3.1,
L is preinjective if and only if ent(X
G
) = ent((L(X
G
))) = ent((X)). By
Proposition 1.7.2, ent(X
G
) = ent(X) and the claim is proved. 2
Corollary 3.3.7 (Myhillproperty for irreducible soc shifts) Let X be
an irreducible soc shift and : X X a transition function. Then pre
injective implies surjective.
44
Proof Let ( = (G, L) be the minimal deterministic presentation of X; we
prove that if L is not preinjective, then is not preinjective either. Suppose
that there exist two biinnite paths c
1
, c
2
X
G
which are dierent only on a
nite path and such that (L(c
1
))) = (L(c
2
)). Then one can write c
1
and c
2
,
respectively, as:
c
1
:
e2
i
1
e1
i
0
e0
i
1
e1
en1
i
n
en
i
n+1
en+1
. . .
and
c
2
:
e2
i
1
e1
i
0
f0
j
1
f1
fn1
j
n
fn
i
n+1
en+1
. . . ,
with e
0
,= f
0
. Setting a
i
:= L(e
i
) and b
i
:= L(f
i
), the graph ( being determin-
istic we have a
0
,= b
0
and hence
L(c
1
) = a
2
a
1
a
0
a
1
. . . a
n1
a
n
a
n+1
. . .
and
L(c
2
) = a
2
a
1
b
0
b
1
. . . b
n1
b
n
a
n+1
. . .
are two congurations in X which dier only on a nite (non empty) set and
whose images under are equal. Therefore is not preinjective.
Thus, if is preinjective, then L is also preinjective; by Theorem 3.3.6,
we have ent(X) = ent((X)). X being irreducible and by Theorem 1.7.4, (X)
cannot be a proper subshift of X. Hence (X) = X, i.e. is surjective. 2
3.3.1 A counterexample to Mooreproperty for a soc
subshift of A
Z
We now give an example of an irreducible soc shift not of nite type for
which the transition function is surjective but not preinjective (that is, Moore
property no longer holds in general if the nite type condition is dropped). Our
example will be the even shift X
e
, that is the subshift of 0, 1
Z
with forbidden
blocks
10
2n+1
1 [ n 0.
The shift X
e
is soc, indeed it is accepted by the following labeled graph.
I
q
0
i
0
~
1
45
We dene a transition function as follows:
(c)
|z
= (c
|z2
, c
|z1
, c
|z
, c
|z+1
, c
|z+2
)
where is the local rule:
(a
1
a
2
a
3
a
4
a
5
) =
_
1 if a
1
a
2
a
3
= 000 or a
1
a
2
a
3
= 111 or a
1
a
2
a
3
a
4
a
5
= 00100,
0 otherwise.
First we prove a Lemma.
Lemma 3.3.8 If a block 0
n
1 with n 3, has a preimage under of length
n + 5 in the language of X
e
a
1
a
2
a
3
a
4
. . . a
n+1
a
n+2
a
n+3
a
n+4
a
n+5
0 0 . . . 0 0 1
,
then this preimage can be only of type
1. (i) a
1
a
2
xx (1 x)(1 x) . . . 11 00 11 000a
n+4
a
n+5
,
(ii) a
1
a
2
xx (1 x)(1 x) . . . 11 00 11 00100,
(iii) a
1
a
2
(1 x)(1 x) xx. . . 00 11 00 111a
n+4
a
n+5
,
when n is even and for a suitable x 0, 1;
2. (i) a
1
a
2
(1 x) xx. . . 11 00 11 000a
n+4
a
n+5
,
(ii) a
1
a
2
(1 x) xx. . . 11 00 11 00100
(iii) a
1
a
2
x (1 x)(1 x) . . . 00 11 00 111a
n+4
a
n+5
when n is odd and for a suitable x 0, 1.
Proof We prove the statement by induction on n 3. Assume that
(a
1
a
2
a
3
a
4
a
5
a
6
a
7
a
8
) = 0001; we distinguish three cases.
a
4
a
5
a
6
= 000
a
1
a
2
a
3
0 0 0 a
7
a
8
0 0 0 1
.
Then a
3
= 1 otherwise (a
3
a
4
a
5
a
6
a
7
) = (0000a
7
) = 1 ,= 0.
a
4
a
5
a
6
a
7
a
8
= 00100
a
1
a
2
a
3
0 0 1 0 0
0 0 0 1
.
Then, for the same reasons as above, a
3
= 1.
a
4
a
5
a
6
= 111
46
a
1
a
2
a
3
1 1 1 a
7
a
8
0 0 0 1
.
Then a
3
= 0 otherwise (a
3
a
4
a
5
a
6
a
7
) = (1111a
7
) = 1 ,= 0.
Now let us suppose that the statement is true for n and that we have
(a
1
. . . a
n+6
) = 0
n+1
1:
a
1
a
2
a
3
a
4
. . . a
n+2
a
n+3
a
n+4
a
n+5
a
n+6
0 0 . . . 0 0 1
.
If n is even, by the inductive hypothesis one has either
a
4
. . . a
n+4
= xx (1 x)(1 x) . . . 11 000
or
a
4
. . . a
n+6
= xx (1 x)(1 x) . . . 11 00100
or
a
4
. . . a
n+4
= (1 x)(1 x) xx. . . 00 111
for a suitable x 0, 1.
In any case we have a
4
= a
5
. If a
3
= a
4
, then (a
3
a
4
a
5
a
6
a
7
) = (a
4
a
4
a
4
a
6
a
7
)
= 1 ,= 0. Thus a
3
,= a
4
.
It follows, in the three cases, that either
a
1
. . . a
n+6
= a
1
a
2
(1 x) xx (1 x)(1 x) . . . 11 000a
n+5
a
n+6
or
a
1
. . . a
n+6
= a
1
a
2
(1 x) xx (1 x)(1 x) . . . 11 00100
or
a
1
. . . a
n+6
= a
1
a
2
x (1 x)(1 x) xx. . . 00 111a
n+5
a
n+6
.
If n is odd, by the inductive hypothesis we have either
a
4
. . . a
n+4
= (1 x) xx. . . 11 000
or
a
4
. . . a
n+6
= (1 x) xx. . . 11 00100
or
a
4
. . . a
n+4
= x (1 x)(1 x) . . . 00 111
for a suitable x 0, 1.
In any case a
4
,= a
5
= a
6
. If a
3
,= a
4
, then a
3
a
4
a
5
= a
5
a
4
a
5
so that a
4
= 1
(otherwise we had a forbidden block). For the same reason, a
2
= a
3
= 0. This
implies (a
2
a
3
a
4
a
5
a
6
) = (00100) = 1 ,= 0. Thus a
3
= a
4
.
It follows, in the three cases, that either
a
1
. . . a
n+6
= a
1
a
2
(1 x)(1 x) xx. . . 11 000a
n+5
a
n+6
47
or
a
1
. . . a
n+6
= a
1
a
2
(1 x)(1 x) xx. . . 11 00100
or
a
1
. . . a
n+6
= a
1
a
2
xx (1 x)(1 x) . . . 00 111a
n+5
a
n+6
.
Then the statement is still true for 0
n+1
1. 2
Proposition 3.3.9 The local function is a transition function, that is (X
e
)
X
e
.
Proof (X
e
) being a subshift of 0, 1
Z
, it suces to prove that no forbidden
block 10
n
1 with n odd, has a preimage of length n + 6 in the language of X
e
.
First we prove that there is no block a
1
a
2
a
3
a
4
a
5
a
6
a
7
of length 7 such that
(a
1
a
2
a
3
a
4
a
5
a
6
a
7
) = 101:
a
1
a
2
a
3
a
4
a
5
a
6
a
7
1 0 1
.
We distinguish two cases.
a
3
a
4
= 00
a
1
a
2
0 0 a
5
a
6
a
7
1 0 1
Then a
2
= 1 otherwise (a
2
a
3
a
4
a
5
a
6
) =(000a
5
a
6
) = 1 ,= 0. Then (a
1
a
2
a
3
a
4
a
5
)
= (a
1
100a
5
a
6
) = 0 ,= 1.
a
3
a
4
a
5
= 111
a
1
a
2
1 1 1 a
6
a
7
1 0 1
Then a
2
= 0 otherwise (a
2
a
3
a
4
a
5
a
6
) = (1111a
6
) = 1 ,= 0. Thus (a
1
a
2
a
3
a
4
a
5
)
= (a
1
0111a
6
) = 0 ,= 1. We have proved that no block of length 7 goes to 101
under .
Let us now prove that no block a
1
. . . a
n+6
of length n+6 has 10
n
1 as image
under , where n N is odd and strictly greater than 1. If
a
1
a
2
a
3
a
4
a
5
. . . a
n+3
a
n+4
a
n+5
a
n+6
1 0 0 . . . 0 1
,
by Lemma 3.3.8 we have a
4
a
5
a
6
. . . =x(1x)(1x) . . . , and being (a
1
a
2
a
3
a
4
a
5
)
= 1, we distinguish two cases:
x = 0
48
a
1
a
2
a
3
0 1 1 . . . a
n+3
a
n+4
a
n+5
a
n+6
1 0 0 0 . . . 0 1
Then a
3
= 0 (otherwise we had a forbidden block) and a
2
= 1 because (a
2
a
3
a
4
a
5
a
6
) = (a
2
0011) = 0 and (00011) = 1. Then (a
1
a
2
a
3
a
4
a
5
) = (a
1
1001) =
0 ,= 1.
x = 1
a
1
a
2
a
3
1 0 0 . . . a
n+3
a
n+4
a
n+5
a
n+6
1 0 0 0 . . . 0 1
If a
3
= 0 then a
2
= 0 and (a
2
a
3
a
4
a
5
a
6
) = (00100) = 1 ,= 0. Thus a
3
=
1. Then (a
2
a
3
100) = (a
2
1100) and (a
2
1100) = 0 implies a
2
= 0. Thus
(a
1
a
2
a
3
10) = (a
1
0110) = 0 ,= 1. Hence 10
n
1 has no preimage under . 2
Proposition 3.3.10 The transition function : X
e
X
e
is surjective.
Proof To prove the surjectivity of it suces, as we have seen, to prove
the nonexistence of GOE patterns or, equivalently, the nonexistence of GOE
words. To this aim, as it can be easily seen, it is enough to prove that each block
of kind 10
n1
10
n2
. . . 10
n
k
1 (where n
1
, . . . n
k
are even integers), has a preimage
block. Indeed each word in L(X
e
) is contained in such a special word.
First we prove that every block of the type 10
n
1 where n is even, has a
preimage a
1
. . . a
n+6
in the language of X
e
of length n + 6
a
1
a
2
a
3
a
4
a
5
. . . a
n+2
a
n+3
a
n+4
a
n+5
a
n+6
1 0 0 . . . 0 0 1
,
in each of the three cases in which a
n+4
1.
If n = 0
0 0 0 0 a
5
a
6
1 1
,
0 0 0 1 0 0
1 1
,
and
1 1 1 1 a
5
a
6
1 1
.
If n = 2
a
1
a
1
1 0 0 0 a
5
a
6
1 0 0 1
,
49
a
1
a
1
1 0 0 1 0 0
1 0 0 1
,
and
0 0 0 1 1 1 a
5
a
6
1 0 0 1
.
If n 4, for a suitable x 0, 1,
1 x 1 x 1 x x x . . . 0 0 0 a
n+5
a
n+6
1 0 0 . . . 0 0 1
.
Similarly
1 x 1 x 1 x x x . . . 0 0 1 0 0
1 0 0 . . . 0 0 1
and, nally,
x x x 1 x 1 x . . . 1 1 1 a
n+5
a
n+6
1 0 0 . . . 0 0 1
.
Now, x a word of kind 10
n1
10
n2
. . . 10
n
k
1; we can construct a preimage of
this word starting from the rst on the right block 10
n
k
1 (over the rst on the
right 1 we can write, arbitrarily, 000**, 111** or 00100). In this way we get
a word a
1
. . . a
5
over the second on the left 1 and we can start from this word
over 1 to construct a preimage for the second on the right block 10
n
k1
1, and
so on:
. . . b
1
b
2
b
3
. . . 0 0 1
b
4
b
5
. . . a
1
a
2
0 0 . . . 0 0
. .
n
k1
a
3
1
a
4
a
5
. . . * *
0 0 . . . 0 0
. .
n
k
* * *
1
In each of the possible choices we can nd a block whose image under is
our xed word.
For what we have stated before, is surjective. 2
Proposition 3.3.11 The transition function : X
e
X
e
is not preinjective.
Proof Let us consider the conguration c
1
:
. . . 0 0 0 0 0 1 0 0 1 0 0 0 0 0 . . .
and the conguration c
2
:
50
. . . 0 0 0 0 0 0 1 1 1 0 0 0 0 0 . . . .
These congurations are dierent only on a nite subset of Z, but they have
the same image under , that is the conguration
. . . 1 1 1 1 1 1 0 0 1 0 0 1 1 1 . . . .
Thus is not preinjective. 2
3.4 Gromovs Theorem
Recently, Gromov has proved a GOElike theorem in a setting of graphs much
more general than Cayley graphs, for alphabets not necessarily nite and for
subset of the universe not necessarily invariant under translation. Because
of the weakness of these hypotheses, in his theorem are needed properties that
are stronger than ours (as we will see in next section), for example the bounded
propagation of the spaces. In this section we apply Gromovs theorem to our
cellular automata proving that all the properties required in the hypotheses of
this theorem are satised.
Denition 3.4.1 A closed subset X A
is of bounded propagation M if
for each pattern p A
F
with support F one has
p
|FD(,M)
X
FD(,M)
for each F p X
F
.
If , the left translation i
: dened by i
() = is an isometry.
Indeed
dist(, ) = |
1
1
| = |
1
| = dist(, ).
It is clear that not all the isometries are of this kind; for example consider
Z
2
= a, b [ ab = ba) and the unique homomorphism extending the function
i : Z
2
Z
2
dened by
_
i(a) = b
i(b) = a
;
clearly |i()| = ||. Then
dist(i(), i()) = |i()
1
i()| = |i(
1
)| = dist(, ).
But i is not a translation.
Consider a subgroup
and the set 1(
) is i
|F
: F F. The set 1(
) is a pseudogroup of partial
isometries that is, following Gromovs denition:
1. id
F
: F F is an element of 1(
), indeed id
F
= i
1|F
;
51
2. if
and i
|F
: F F is an element of 1(
), then (i
|F
)
1
: F F
is still in 1(
) because (i
|F
)
1
= i
1
|F
;
3. if i
1 |F
: F
1
F and i
2 |
1
F
:
1
F
2
1
F are two elements of 1(
),
then their composition is still in 1(
) because i
2|
1
F
i
1|F
= i
1|F
;
4. the restriction of each element of 1(
).
Two elements , in are
equivalent if there exists
such that
= , that is
=
. Then the equivalence classes are the right cosets
[ . The pseudogroup 1(
D(, R) = .
We prove that this property is equivalent to the existence of R N such that
_
D(, R) = (3.1)
Indeed if (3.1) holds and there exists
such that dist(, ) R
hence
dist(, ) R + dist(, ) = R +||
and then, xed a right coset representative of
, we have
D(, R +
||) = . Moreover, we can prove that (3.1) holds if and only if
has nite
index. Indeed, suppose that (3.1) holds; consider the right cosets
with
D
R
. It is clear that these are nitely many; furthermore
DR
=
because if by (3.1), we have D
R
with
. Hence
has nite
index.
Conversely, if
has nite index, x a set
1
, . . . ,
n
of cosets representa-
tives. Let R := max
i
|
i
|. If , we have that =
i
for some
and
some i. Hence dist(, ) = |
i
| R, that is D(, R).
Now consider a stable (i.e. closed) and
invariant space X A
; if we
consider the nite subsets of and the elements of
, a family of functions
H
F,
: X
F
X
F
= X
i(F)
which commute with the restriction (i.e. (H
F,
(c
|F
))
|E
= H
E,
(c
|E
) or, in
other words, the following diagram
X
F
HF,
X
F
_
X
E
HE,
X
E
52
commutes), gives rise to a set of holonomy maps. In particular, we have a set
of holonomy maps H(
) dening
H
F,
(c
|F
) := c
1
|F
;
indeed (H
F,
(c
|F
))
|E
= (c
1
|F
)
|E
= c
1
|E
= H
E,
(c
|E
).
The set H(
), indeed id
XF
= H
F,1
;
2. if
and H
F,
: X
F
X
F
is an element of H(
), then (H
F,
)
1
:
X
F
X
F
is still in H(
) because (H
F,
)
1
= H
F,
1 ;
3. if H
F,
1
: X
F
X
1
F
and H
1
F,
2
: X
1
F
X
1
F
are two elements of
H(
) because H
1
F,
2
H
F,
1
=
H
F,
2
1
;
4. the restriction of each element of H(
) dened on X
F
to X
E
(where E is
a nite subset E F), is still in H(
).
Finally, if 1(
F
: X
F
Y
F
M that commute with the restrictions; then a function of
bounded propagation is such that (c)
|
=
D(,M)
(c
|D(,M)
)
|
and, in gen-
eral,
F
(c
|F
) = (c)
|F
M.
If is a function of bounded propagation, we have that the holonomies in
H(
_
F
X
F
M
H
F
M
,
X
F
M
Indeed (notice that (F
M
) = (F)
M
),
F
(H
F,
(c
|F
)) =
F
((c
1
)
|F
)) = (c
1
)
|(F)
M
and
H
F
M
,
(
F
(c
|F
)) = H
F
M
,
((c)
|F
M) = ((c)
1
)
|F
M.
In this case, provided that 1(
))
|d
=
D(d,M)
(c
|D(d,M)
)
|d
=
DM+R
(c
|DM+R
)
|d
.
This means that in order to know the function it is sucient to know how
the image under of a conguration in X acts on D
R
. In other words, it is
sucient to know the function
DM+R
: X
DM+R
(X)
DR
.
On the other hand, if is Mlocal between two shift spaces, we have
(c)
|
= (c
)
|1
=
DM
(c
|DM
)
|1
that is it suces to know how the image under of a conguration in X acts
on the identity of , i.e. the local rule .
For this reasons, the notion of bounded propagation is a generalization of
the notion of local function as far as stable spaces not necessarily invariant
are concerned.
Hence, if is amenable, the next two theorems follow from Theorem 3.4.2.
Corollary 3.4.3 (GOEtheorem for shifts of bounded propagation) Let
X, Y A
(see
Counterexample 3.3.5). We prove the MMproperty for the strongly irreducible
shifts of nite type of A
+
2M
E
+M
n
; since dist(
+
2M
E
+M
n
, E
n
) = M + 1 > M for each u X
En
there exists a pattern p X
E
+3M
n
that coincides with u on E
n
and with v on
+
2M
E
+M
n
. Then
[p X
E
+3M
n
[ p
|
+
2M
E
+M
n
= v[ [X
En
[ > [Y
E
+2M
n
[.
Since
E
+3M
n
: X
E
+3M
n
Y
E
+2M
n
is surjective, there exist two patterns p
1
, p
2
X
E
+3M
n
such that p
1
,= p
2
but p
1
= v = p
2
on
+
2M
E
+M
n
and
E
+3M
n
(p
1
) =
55
E
+3M
n
(p
2
). By Corollary 1.3.4, there exist two congurations c
1
, c
2
X which
extend p
1
and p
2
and which coincide outside E
+M
n
. We prove that (c
1
) =
(c
2
), and hence that is not preinjective. If E
+2M
n
we have D
M
E
+3M
n
and hence, if D
M
=
1
, . . . ,
m
, (c
1
)
|
= (c
1|1
, . . . , c
1|m
) =
(p
1|1
, . . . , p
1|m
) =
E
+3M
n
(p
1
)
|
=
E
+3M
n
(p
2
)
|
= (p
2|1
, . . . , p
2|m
) =
(c
2|1
, . . . , c
2|m
) = (c
2
)
|
. If / E
+2M
n
, we have D
M
(E
+M
n
) and
hence (c
1
)
|
= (c
2
)
|
, since c
1
coincide with c
2
on (E
+M
n
). 2
Lemma 3.5.3 If is a nitely generated group, there exists a sequence of disks
(F
j
)
jN
obtained by translation of a disk D and at distance > M such that
jN
F
+R
j
= for a suitable R > 0.
Proof Let D be the disk centered at 1 and of radius ; dene the following
sequence of nite subsets of :
0
:= 1,
1
:= [ || = 2 +M + 1
and, in general,
n
:= [ || = n(2 +M + 1).
It is clear that for each n, dist(
n
,
n+1
) = 2 + M + 1. Inside the set
n
, x
n,1
and eliminate all the points in
n
whose distance from
n,1
is less than
2 +M + 1.
Next, x
n,2
among the remaining points and eliminate all the points whose
distance from
n,2
is less than 2 + M + 1. In this way, we will get a set
n
whose elements have mutual distance 2 + M + 1 and such that for each
element
n
of
n
there exists an element of
n
whose distance from
n
is less
than 2 +M + 1.
We now prove that, denoting by (
j
)
jN
the sequence of the elements of
nN
n
, the sequence (
j
D)
jN
is a (D, M, R)net with R := 2 + 2M; so
that we can set F
j
:=
j
D.
Let then ; there exists
n
n
such that dist(,
n
) + M. Since
n
belongs to
n
, there is
n
n
such that dist(
n
,
n
) 2 + M and hence
dist(,
n
) 3 + 2M; then (
n
D)
+(2+2M)
. 2
We call the above sequence a (D, M, R)net.
Lemma 3.5.4 Let be an amenable group and let (E
n
)
n
be a xed amenable
sequence of . Let (F
j
)
jN
be a (D
r
, 2M, R)net, let X be an Mirreducible
shift and let Y be a subset of X such that Y
Fj
X
Fj
for every j N. Then
ent(Y ) < ent(X).
Proof Let (p
j
)
jN
be a sequence of patterns such that p
j
X
Fj
Y
Fj
; let
N(n) be the number of F
j
s such that F
+M
j
E
n
and denote by F
j1
, . . . , F
jN
56
these disks. Set := [X
D
+M[ and denote by
ji
: X
En
X
Fj
i
the restriction
to F
ji
of the patterns of X
En
. We prove that
[X
En
N
_
i=1
1
ji
(p
ji
)[ (1
1
)
N
[X
En
[ (3.2)
by induction on m 1, . . . , N. We have
[X
En
[ [X
F
+M
j
1
[[X
En\F
+M
j
1
[
then
[X
En
[ [X
En\F
+M
j
1
[.
Since X is an Mirreducible shift and since dist(F
j1
, E
n
F
+M
j1
) > M, given a
pattern p X
En\F
+M
j
1
, there exists a pattern p dened on all E
n
that coincides
with p on E
n
F
+M
j1
and with p
j1
on F
j1
; then
[X
En\F
+M
j
1
[ [
1
j1
(p
j1
)[.
Hence
1
[X
En
[ [
1
j1
(p
j1
)[
and
[X
En
1
j1
(p
j1
)[ [X
En
[
1
[X
En
[ = (1
1
)[X
En
[.
Suppose
[X
En
m1
_
i=1
1
ji
(p
ji
)[ (1
1
)
m1
[X
En
[.
Since
[X
En
m1
_
i=1
1
ji
(p
ji
)[ [X
F
+M
jm
[[p
|En\F
+M
jm
[ p X
En
m1
_
i=1
1
ji
(p
ji
)[,
and, being [X
F
+M
jm
[ = ,
[X
En
m1
_
i=1
1
ji
(p
ji
)[ [p
|En\F
+M
jm
[ p X
En
m1
_
i=1
1
ji
(p
ji
)[.
Moreover, since X is Mirreducible,
[p
|En\F
+M
jm
[ p X
En
m1
_
i=1
1
ji
(p
ji
)[
[p X
En
m1
_
i=1
1
ji
(p
ji
) [ p
|Fjm
= p
jm
[.
57
Hence
1
[X
En
m1
_
i=1
1
ji
(p
ji
)[ [p X
En
m1
_
i=1
1
ji
(p
ji
) [ p
|Fjm
= p
jm
[
and then
[X
En
m
_
i=1
1
ji
(p
ji
)[ = [
_
X
En
m1
_
i=1
1
ji
(p
ji
)
_
1
jm
(p
jm
)[
[
_
X
En
m1
_
i=1
1
ji
(p
ji
)
_
p X
En
m1
_
i=1
1
ji
(p
ji
) [ p
|Fjm
= p
jm
[
[X
En
m1
_
i=1
1
ji
(p
ji
)[
1
[X
En
m1
_
i=1
1
ji
(p
ji
)[
(1
1
)(1
1
)
m1
[X
En
[ = (1
1
)
m
[X
En
[.
Hence (3.2) holds, and since [Y
En
[ [X
En
N
i=1
1
ji
(p
ji
)[, we have
log [Y
En
[
[E
n
[
N(n) log(1
1
)
[E
n
[
+
log [X
En
[
[E
n
[
. (3.3)
Observe that
E
n
N
_
i=1
F
+R
ji
(E
n
E
(R+2r+M)
n
). (3.4)
Indeed suppose that E
n
and /
N
i=1
F
+R
ji
; (F
j
)
j
being a (D
r
, 2M, R)
net, we have that F
+R
k
for some k, that is D
+R
with such that
D
+M
, E
n
. Hence dist(, ) r +R so that D
+R
. If E
(R+2r+M)
n
,
then D
+R
E
(r+M)
n
so that F
+M
k
= D
+M
E
n
which is excluded.
From (3.4), we have
[E
n
[ N(n)[D
+R
[ +[E
n
E
(R+2r+M)
n
[
so that
1
N(n)
[E
n
[
[D
+R
[ +
[
R+2r+M
E
n
[
[E
n
[
;
taking the minimum limit and being lim
n
|
R+2r+M
En|
|En|
= 0,
:= liminf
n
N(n)
[E
n
[
> 0
and then from (3.3) it follows
ent(Y ) log(1
1
) + ent(X) < ent(X). 2
58
Proposition 3.5.5 Let X be a strongly irreducible shift of nite type, let :
X A
Z
be a local function such that ent((X)) = ent(X). Then is pre
injective.
Proof Suppose that X has memory M, that is X is Mirreducible and that
is Mlocal. Moreover suppose that is not preinjective; then there exist
c
1
, c
2
X and a disk D contained in , such that c
1
,= c
2
on D, c
1
= c
2
out of
D and (c
1
) = (c
2
). Set (F
j
)
jN
= (
j
D
+2M
)
jN
a (D
+2M
, 2M, R)net and
denote by Y the subset of X dened by
Y := c X [ (c
j
)
|D
+2M ,= c
2|D
+2M for every j N,
that is the subset of X avoiding the pattern c
2|D
+2M on the disk D
+2M
and
on the translated disks F
j
=
j
D
+2M
. The set Y is a subset of X such that
Y
Fj
X
Fj
; we prove that (Y ) = (X). Indeed if c XY , there exists a
subset J N such that for every j J, we have (c
j
)
|D
+2M = c
2|D
+2M. Dene
c X in the following way:
c = c
1
1
j
on F
j
for every j J,
c = c out of the union
jJ
F
j
.
That is, c is obtained from c substituting all the occurences of c
2|D
+2M with
c
1|D
+2M.
By Proposition 1.3.3, we have c X and moreover c Y ; we prove that
( c) = (c).
If
j
D
+M
for some j J, we have D
M
F
j
and then ( c)
|
=
(c
1
1
j
)
|
= (c
1
)
|
1
j
= (c
2
)
|
1
j
= (c
2
1
j
)
|
= (c)
|
.
Suppose that /
j
D
+M
for every j J; then D
M
(
j
D) and hence
( c)
|
= (c)
|
. Indeed c and c coincide on
jJ
(
j
D): if j J and
+
2M
j
D = F
j
j
D, we have c
|
= (c
1
1
j
)
|
= c
1
|
1
j
. Since
1
j
+
2M
D
one has c
1
|
1
j
= c
2
|
1
j
= (c
2
1
j
)
|
= c
|
.
Then, by Lemma 3.5.4,
ent((X)) = ent((Y )) ent(Y ) < ent(X). 2
Proposition 3.5.6 Let be an amenable group. Let X be a shift, let Y be
a strongly irreducible shift and let : X Y be a local function such that
ent((X)) = ent(Y ). Then is surjective.
Proof Let X and Y be as in the hypotheses and let : X Y be a local
function. We prove that if (X) Y , then ent((X)) < ent(Y ). Indeed if
(X) Y , there exists a conguration c Y which does not belong to (X)
and then there exists a disk D such that c
|D
Y
D
((X))
D
. Let (F
j
)
jN
be
a (D, 2M, R)net; then ((X))
Fj
Y
Fj
; by Lemma 3.5.4, ent((X)) < ent(Y ).
2
59
Theorem 3.5.7 Let be an amenable group. Let X be a strongly irreducible
shift of nite type, let Y be a strongly irreducible shift and let : X Y be a
local function with ent(X) = ent(Y ). Then is preinjective if and only if is
surjective.
Proof If is preinjective we have, by Proposition 3.5.2, that ent((X)) =
ent(X). Then ent((X)) = ent(Y ) so that, by Proposition 3.5.6, is surjective.
If, conversely, is surjective then ent((X)) = ent(Y ) that is ent((X)) =
ent(X). By Proposition 3.5.5, is preinjective. 2
Corollary 3.5.8 (MMproperty for strongly irreducible shifts of nite
type) If is an amenable group, a strongly irreducible subshift of nite type of
A
such that c
|D(,M)
X
D(,M)
for every , belongs to
X.
Proof Let X be a shift of nite type with memory M, let T a nite set of
forbidden blocks each one with support D
M
and let c A
be a conguration
such that c
|D(,M)
X
D(,M)
for every .
We prove that c X; it is clear that for each , c
|DM
X
DM
and hence
we have c
|DM
/ T, that is c X
F
= X.
For the converse, suppose that each conguration c A
such that c
|D(,M)
X
D(,M)
for every , belongs to X. Dene
T := c
|DM
[ c
|DM
/ X
DM
;
if c X we have that for each , c
|DM
X
DM
c
|DM
/ T and c X
F
. If
c X
F
we have for each that c
|DM
X
DM
c
|DM
X
DM
and c X.
2
Now we can prove the following statement.
Proposition 3.5.10 If X A
is such that c
|D(,M)
X
D(,M)
for every .
Then if n M and D
n
we have
c
|DnD(,M)
= (c
|D(,M)
)
|DnD(,M)
X
DnD(,M)
;
X being of bounded propagation we have c
|Dn
X
Dn
. X being closed we have
c X. 2
If = Z and X is an edge shift, then also the converse of this theorem holds.
Proposition 3.5.11 If X A
Z
is a is strongly irreducible edge shift, then it
is of bounded propagation.
Proof Let G be a graph such that X = X
G
; notice that if uv, vw
L(X), then uvw L(X) for every word v L(X) such that [v[ 1. In-
deed if e
1
, . . . , e
n
, f
1
, . . . , f
m
, g
1
, . . . , g
l
c(G) are edges of G such that u =
e
1
. . . e
n
, v = f
1
. . . f
m
and w = g
1
. . . g
l
, we have:
uv : i
1
e1
i
2
e2
en
i
n+1
f1
i
n+2
f2
fm
i
n+m+1
vw : i
n+1
f1
j
n+2
f2
fm
i
n+m+1
g1
i
n+m+2
g2
g
l
i
n+m+l+1
.
Then it is clear that the word
uvw : i
1
e1
en
i
n+1
f1
fm
i
n+m+1
g1
g
l
i
n+m+l+1
belongs to L(X).
Suppose that X is Mirreducible; we prove that X has bounded propagation
M.
Let F be the interval [1, L] and let p A
F
a pattern such that for each
F we have p
|FD(,M)
X
FD(,M)
; then there exist q 0 and 0 r < M
for which L = qM +r + 1. Set
1
:= M + 1, then F D(
1
, M) = [1, 2M + 1]
and hence p
|[1,2M+1]
X
[1,2M+1]
. Set
2
:= 2M + 1, then F D(
2
, M) =
[M+1, 3M+1] and hence p
|[M+1,3M+1]
X
[M+1,3M+1]
. By the above property,
we have that p
|[1,3M+1]
X
[1,3M+1]
. In this way we can prove that p
|[1,qM+1]
X
[1,qM+1]
. Set
q
:= qM +1, then F D(
q
, M) = [(q 1)M +1, L] and hence
p = p
|[1,L]
X
[1,L]
= X
F
.
If F is the union of two disjoint intervals F
1
= [1, L
1
] and F
2
= [L
1
+n, L
2
]
at distance n M, we already have from the above case that p
|F1
X
F1
and
p
|F2
X
F2
. Set := L
1
, then F D(, M) = F
1
[L
1
+n, L
1
+M] and hence
p
|F1[L1+n,L1+M]
X
F1[L1+n,L1+M]
. Then there exists a word u of length n2
such that p
|F1
u p
|[L1+n,L1+M]
L(X) and the word p
|F2
L(X). For the above
property we have p
|F1
u p
|F2
L(X) and hence p X
F
.
Finally, if F is the union of two disjoint intervals F
1
and F
2
at distance > M,
we have that p X
F
for the Mirreducibility of X.
Because each nite subset F of Z is a nite union of intervals, we have that
the property holds for every F. 2
61
Now we prove that in general strong irreducibility and nite type condition
do not imply the bounded propagation property. Consider the subshift X
0, 1
Z
with set of forbidden blocks:
010, 111.
Clearly X is a strongly irreducible (in fact 2irreducible) shift of nite type; if
M 1 consider the following pattern p with F := supp(p)
0 1 1 1 . . . 1 1
. .
M copies of 1
1 0
In this case we have p
|FD(,M)
X
FD(,M)
but p / X
F
; hence X is not of
bounded propagation M for each M 1.
3.6 SemiStrongly Irreducible Shifts
As we have seen, the MMproperty holds for irreducible subshifts of nite type
of A
Z
. On the other hand we have that the MMproperty holds, in general, for
strongly irreducible subshifts of nite type of A
be a local and
preinjective function. Then ent((X)) = ent(X).
64
Proof Suppose that the memory of X is M, that X is (M, k)irreducible
and that is Mlocal. Set Y := (X) and x an amenable sequence of disks
(E
n
)
n
; as we have seen in the proof of Proposition 3.5.2 we have
limsup
n
log [Y
E
+2M
n
[
[E
n
[
ent(Y ).
Let l = l(k) be the number of s such that || k and suppose that ent(Y ) <
ent(X); then
limsup
n
log [Y
E
+2M
n
[
[E
n
[
< limsup
n
log [X
En
[
[E
n
[
= limsup
n
log(
|XEn
|
l
)
[E
n
[
;
so that there exists n N such that
log [Y
E
+2M
n
[
[E
n
[
<
log(
|XEn
|
l
)
[E
n
[
that is
[Y
E
+2M
n
[ <
[X
En
[
l
.
Fix v X
+
2M
E
+M
n
; since dist(
+
2M
E
+M
n
, E
n
) = M + 1 > M for each u X
En
there exists an D
k
and a pattern p X
E
+3M
n
that coincides with u on E
n
and with v on
+
2M
E
+M
n
. Then
[p X
E
+3M
n
[ p
|
+
2M
E
+M
n
= v[
[X
En
[
l
> [Y
E
+2M
n
[.
Since
E
+3M
n
: X
E
+3M
n
Y
E
+2M
n
is surjective, there exist two patterns p
1
, p
2
X
E
+3M
n
such that p
1
,= p
2
but p
1
= v = p
2
on
+
2M
E
+M
n
and
E
+3M
n
(p
1
) =
E
+3M
n
(p
2
). By Corollary 1.3.4, there exist two congurations c
1
, c
2
X which
extend p
1
and p
2
and which coincide outside E
+M
n
. As we have seen in the proof
of Proposition 3.5.2, one has (c
1
) = (c
2
); hence is not preinjective. 2
Observe that for semistrongly irreducible shifts, Lemma 3.5.4 does not nec-
essarily hold. Indeed consider the subshift X = 01, 10 A
Z
; this shift is of
nite type. Being accepted by the graph
I
q
0
i
1
65
the shift X is (2, 2)irreducible. Now (5j)
jZ
is a (0, 4, 10)net. If c is the
conguration = 0 on the even numbers and = 1 on the odd ones, set Y := c;
we have Y
{5j}
Y
{5j}
but ent(Y ) = ent(X) = 0.
The following lemma is very similar to Lemma 3.5.4 but as one can see the
hypotheses are quite stronger.
Lemma 3.6.5 Let be a group with nonexponential growth and let (E
n
)
n
be a xed amenable sequence of disks. Let (F
j
)
jN
= (D(
j
, r))
jN
be a
(D
r
, 2M, R)net, let X be an (M, k)irreducible shift and let Y be a subset of X
such that for each j N, there exists a pattern p
j
X
Fj
for which p
j
/ Y
D(
j
,r)
whenever D
k
. Then ent(Y ) < ent(X).
Proof Let N(n) be the number of F
j
s such that F
+M
j
E
n
and denote by
F
j1
, . . . , F
jN
these disks. Set := [X
D
+M[ and denote by P
jm
X
En
the set of
the blocks p of X
En
such that p
|D(
jm
,r)
= p
jm
for some D
k
; we prove that
[X
En
N
_
i=1
P
ji
[ (1
1
)
N
[X
En
[ (3.5)
by induction on m 1, . . . , N. We have
[X
En
[ [X
F
+M
j
1
[[X
En\F
+M
j
1
[
then
[X
En
[ [X
En\F
+M
j
1
[.
Since X is (M, k)irreducible and since dist(F
j1
, E
n
F
+M
j1
) > M, given a pat-
tern p X
En\F
+M
j
1
there exists a pattern p dened on all E
n
that coincides with
p on E
n
F
+M
j1
and with p
j1
on some D(
j1
, r); then
[X
En\F
+M
j
1
[ [P
j1
[.
Hence we have
1
[X
En
[ [P
j1
[
so that
[X
En
P
j1
[ [X
En
[
1
[X
En
[ = (1
1
)[X
En
[.
Suppose
[X
En
m1
_
i=1
P
ji
[ (1
1
)
m1
[X
En
[;
since
[X
En
m1
_
i=1
P
ji
[
66
[X
F
+M
jm
[[p X
En\F
+M
jm
[ p
|D(
j
i
,r)
,= p
ji
for each i = 1, . . . , m1 and each [
and being [X
F
+M
jm
[ = , we have
[X
En
m1
_
i=1
P
ji
[
[p X
En\F
+M
jm
[ p
|D(
j
i
,r)
,= p
ji
for each i = 1, . . . , m1 and each [.
Moreover, since X is (M, k)irreducible,
[p X
En\F
+M
jm
[ p
|D(
j
i
,r)
,= p
ji
for each i = 1, . . . , m1 and each [
[p X
En
[ p
|D(
j
i
,r)
,= p
ji
for each i = 1, . . . , m 1 and each
and p
|D(
jm
,r)
= p
jm
for some [.
Hence
1
[X
En
m1
_
i=1
P
ji
[ [p X
En
m1
_
i=1
P
ji
[ p
|D(
jm
,r)
= p
jm
for some [
and then
[X
En
m
_
i=1
P
ji
[ = [
_
X
En
m1
_
i=1
P
ji
_
P
jm
[
[
_
X
En
m1
_
i=1
P
ji
_
p X
En
m1
_
i=1
P
ji
[ p
|D(
jm
,r)
= p
jm
for some [
[X
En
m1
_
i=1
P
ji
[
1
[X
En
m1
_
i=1
P
ji
[
(1
1
)(1
1
)
m1
[X
En
[ = (1
1
)
m
[X
En
[.
Hence (3.5) holds, and since [Y
En
[ [X
En
N
m=1
P
jm
[, we have
log [Y
En
[
[E
n
[
N(n) log(1
1
)
[E
n
[
+
log [X
En
[
[E
n
[
. (3.6)
As we have proved in Lemma 3.5.4,
:= liminf
n
N(n)
[E
n
[
> 0;
then taking the maximum limit in (3.6), it follows
ent(Y ) log(1
1
) + ent(X) < ent(X). 2
The following statement is an easy consequence of Lemma 3.6.5 and gener-
alizes the result given in Theorem 1.7.4.
67
Corollary 3.6.6 Let be a group with nonexponential growth and let X be
a semistrongly irreducible subshift of A
S] A. S.
Svarc, A Volume Invariant of Coverings (Russian), Dokl.
Akad. Nauk SSSR 105 (1955), 32-34.
[U] S. Ulam, Random Processes and Transformations, Proc. Int.
Congr. Mathem. 2 (1952), 264-275.
70
[vN] J. von Neumann, The Theory of SelfReproducing Automata (A.
Burks, ed.), University of Illinois Press, Urbana and London,
1966.
[Wa] H. Wang, Proving Theorems by Pattern Recognition II, Bell Sys-
tem Tech. J. 40 (1961), 1-41.
[Wei1] B. Weiss, Subshifts of Finite Type and Soc Systems, Monats.
Math. 77 (1973), 462-474.
[Wei2] B. Weiss, Soc Groups and Dynamical Systems, preprint.
[Y] S. Yukita, Dynamics of Cellular Automata on Groups, IEICE
Trans. Inf. & Syst., Vol. E82-D 10 (1999), 1316-1323.
71