Universit

`
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

(where is a nitely generated group and A is a nite alphabet). Indeed for

these sets is still possible to dene a structure of a cellular automaton.
More precisely, given a nitely generated group , one can consider as uni-
verse the Cayley graph of . A conguration is an element of the space A

(the
socalled full Ashift), 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,
3
where the topology in A is the discrete one. A subset X of 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

. For example, we prove that

the topological denition of shift is equivalent to the combinatorial one of set
of congurations avoiding some forbidden blocks.
Moreover we prove that a function between shift spaces is local if and only if
it is continuous and commutes with the action. This fact, together with the
compactness of the shift spaces (notice that in 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.
A fundamental notion we give is that of irreducibility for a shift; in the one
dimensional case, this means that given any pair of words u, v in the language
of the shift (i.e. the set of all nite words appearing in some biinnite con-
guration), there is a word w such that the concatenation uwv still belongs to
the language. We generalize this denition to a generic shift, using the patterns
and their supports rather than the words.
Then, in terms of forbidden blocks the notion of shift of nite type is given
and we prove that, as in the onedimensional case, such a shift has a useful
overlapping property that will be necessary in later chapters. Moreover 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.
As we have noticed, cellular automata have mainly been investigated in
the Euclidean case and for the full shifts. The dierence between the one
dimensional cellular automata and the higher dimensional ones, is very deep.
For example, Amoroso and Patt have shown in [AP] that there are algorithms
to decide surjectivity and injectivity of local maps for onedimensional cellular
automata. On the other hand Kari has shown in [K1] and [K2] that both the
injectivity and the surjectivity problems are undecidable for Euclidean cellular
automata of dimension at least two. Here we extend the AmorosoPatts results
to the onedimensional cellular automata over shifts of nite type.
Finally, the notion of entropy for a generic shift as dened by Gromov in
[G] is also 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], we recall its basic properties in the onedimensional
case, also stating the principal result of the PerronFrobenius theory to compute
4
it.
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. Using the GOE theorem and the compactness of the space
A
Z
n
, Richardson has proved in [R] that a transition function : A
Z
n
A
Z
n
is surjunctive. In Chapter 2, 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 some
generalities about periodic congurations, we recall 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

is surjunctive. In fact, we prove more generally that if the pe-

riodic congurations of a subshift X A

are dense, then a transition function

on X is surjunctive. We also prove that the density of the periodic congura-
tions is an invariant of the shifts, as is the number of the periodic conguration
with a xed period.
The remaining part of the chapter is devoted to establish for which class of
shifts the periodic congurations are dense. We prove the density of the periodic
congurations for an irreducible subshift of nite type of A
Z
and hence, a soc
shift being the image under a local map of a shift of nite type, the density of
the periodic congurations for an irreducible soc subshift of A
Z
. We see that
these results cannot be generalized to higher dimensions.
If the alphabet A is a nite group G and is abelian, the periodic cong-
urations of a subshift which is also a subgroup of G

(namely a group shift),

are dense (this result is a consequence of a more general theorem in [KitS2]).
Moreover, this further hypothesis is not still included in the result about irre-
ducible shifts of nite type; indeed a group shift is of nite type but there are
examples of group shifts which are not irreducible. Finally, we list some other
wellknown decision problems for Euclidean shifts proving that in the special
case of a onedimensional shift they can be solved; more generally they can be
solved for the class of group shifts using some results due to Wang [Wa] and
Kitchens and Schmidt [KitS1].
In Chapter 3, we consider generalizations of the Moores property and My-
hills property to a generic shift. In details, the GOEtheorem has been proved
by Mach` and Mignosi [MaMi] more generally for cellular automata in which
the space of congurations is the whole Ashift A

and the group has non

exponential growth; more recently it has been proved by CeccheriniSilberstein,
Mach` and Scarabotti [CeMaSca] for the wider class of the amenable groups.
Instead of the nonexistence of mutually erasable patterns, we deal with the
notion of preinjectivity (a function : X 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

. We generalize this result

showing that the MMproperty holds for strongly irreducible shifts of nite type
of 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

= + N is the neighborhood of obtained from

N by translation.
Our purpose is, given a nitely generated group , to consider as universe
the Cayley graph of . A conguration is an element of the space A

, 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

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 specifying 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
).
Sections 1.1 and 1.2 of this chapter are dedicated to give a formal denition
of 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

. For example, we prove that the topological denition of shift

7
is equivalent to the combinatorial one of set of congurations avoiding some
forbidden blocks. Moreover we prove that a function between shift spaces is
local if and only if it is continuous and commutes with the action. This fact,
together with the compactness of the shift spaces (notice that in 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

are two congurations, we dene the distance

dist(c
1
, c
2
) :=
1
n + 1
where n is the least natural number such that c
1
,= c
2
in D
n
. If such an n does
not exist, that is if c
1
= c
2
, we set their distance equal to zero.
Observe that the group acts on A

on the right as follows:

(c

)
|
:= c
|
for each c A

and each , (where c

|
is the value of c at ).
Now we give a topological denition of a shift space (briey shift); we will
see in the sequel that this denition is equivalent (in the Euclidean case) to the
classical combinatorial one.
10
Denition 1.2.1 A subset X of A

is called a shift if it is topologically closed

and invariant (i.e. X

= 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

, one can always consider X as a shift

over the alphabet A B.
Denition 1.2.3 Let be a nitely generated group with a xed symmetric
set of generators A, let A be a nite alphabet with at least two element and
let D
M
the disk in centered at 1 and with radius M. A cellular automaton is
a triple (X, D
M
, ) where X is a subshift of the compact space A

, D
M
is the
neighborhood of 1 and : X X is an Mlocal function.
Let : X A

be a Mlocal function; if c is a conguration of X and E is

a subset of , (c)
|E
only depends on c
|E
+M. Thus we have a family of functions
(
E
+M : X
E
+M (X)
E
)
E
.
The following characterization of local functions is, in the onedimensional
case, known as the CurtisLyndonHedlund theorem. Here we generalize it to
a general local function.
Proposition 1.2.4 A function : X A

is local if and only if it is continu-

ous and commutes with the action (i.e. for each c X and each , one
has (c

) = (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

that associates with

each c X its translated conguration c

. In general, this function does not

commute with the action (and therefore it is not local). Indeed, if is not
abelian and ,= , then (c

,= (c

. However, this function is continuous.

In order to see this, if n 0, x a number m 0 such that D
n
D
m
; if
dist(c, c)
1
m+1
, then c and c agree on D
m
and hence on D
n
. If D
n
, we
have c
|
= c
|
and then 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

are conjugate if there exists a local

bijective function between them (namely a conjugacy). The invariants are the
properties of a shift invariant under conjugacy.
Now we prove that the topological denition of a shift space is equivalent
to a combinatorial one involving the avoidance of certain forbidden blocks: this
fact is wellknown in the Euclidean case.
Theorem 1.2.6 A subset X A

is a shift if and only if there exists a subset

T

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

is irreducible if for each pair of patterns

p
1
X
E
and p
2
X
F
, there exists an element such that E F = and
a conguration c X such that c
|E
= p
1
and c
|F
= p
2
.
In other words, a shift is irreducible if whenever we have p
1
, p
2
L(X), there
exists a conguration c X in which these two blocks appear simultaneously
on disjoint supports. This denition could seem weaker than Denition 1.2.7,
in fact in this latter we establish that each word u L(X) must always appear
in a conguration on the left of each other word of the language. In order to
prove that the two denitions agree, suppose that X A
Z
is an irreducible
14
shift according with Denition 1.2.8. If u, v are words in L(X), there exists a
conguration c X such that c
|E
= u and c
|F
= v where E and F are nite and
disjoint intervals. If max E < min F then there exists a word w such that uwv
L(X) (where w = c
|I
and I is the interval [max E+1, minF 1]). If, otherwise,
max F < min E there exists a word w such that vwu L(X); consider the word
vwu two times, there exists another word x such that vwu x vwu L(X) and
hence uxv L(X).
1.3 Shifts of Finite Type
Now we give the fundamental notion of shift of nite type. The basic denition
is in terms of forbidden words; in a sense we may say that a shift is of nite
type if we can decide whether or not a conguration belongs to the shift only
checking its words of a xed (and only depending on the shift) length. This fact
implies an useful characterization of the onedimensional shifts of nite type,
a sort of overlapping property for the words of the language. We prove that
this overlapping property still holds for a generic shift of nite type.
Denition 1.3.1 A shift is of nite type if it admits a nite set of forbidden
blocks.
If X is a shift of nite type, since a nite set T of forbidden blocks of X has a
maximal support, we can always suppose that each block of T has the disk D
M
as support (indeed each block that contains a forbidden block is forbidden). In
this case the shift X is called Mstep and the number M is called the memory
of X. If X is a subshift of A
Z
, we dene the memory of X as the number M,
where M + 1 is the maximal length of a forbidden word.
For the shifts of nite type in A
Z
, we have the following useful property.
Proposition 1.3.2 [LinMar, Theorem 2.1.8] A shift X A
Z
is an Mstep
shift of nite type if and only if whenever uv, vw L(X) and [v[ M, then
uvw L(X).
Now we prove that this overlapping property holds more generally for
subshifts of nite type of 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

that agrees with c

1
on E and with c
2
on E is still in X.
Proof Let c
1
, c
2
and c be three congurations as in the statement; if T is a
nite set of forbidden blocks for X, each having D
M
=
1
, . . . ,
m
as support,
we have to prove that for each , c

|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

which is factor of a shift of nite type.

1.5.1 Minimal Deterministic Presentations of a Soc Shift
In automata theory, a nitestate automaton is deterministic if, given a state
Q and a letter a, there is at most one successive state

Q (determined by Q and
a). This corresponds, in the nite graph representing the automata, to the fact
that from a vertex i (the state) there is at most one edge carrying the label a.
Although this restrictive condition, one can prove that for each regular language
there is a deterministic automaton accepting it. This property holds true for
soc shifts: each soc shift admits a deterministic presentation. A minimal
deterministic presentation of a soc shift is a deterministic presentation with
the least possible number of vertices. We will see that the irreducibility of the
soc shift implies the existence of only one (up to labeled graphs isomorphism)
minimal deterministic presentation of it.
Denition 1.5.3 A labeled graph ( = (G, L) is deterministic if, for each vertex
i of G, the edges starting at i carry dierent labels.
Proposition 1.5.4 [LinMar, Theorem 3.3.2] Every soc shift has a determin-
istic presentation.
Denition 1.5.5 A minimal deterministic presentation of a soc shift X is a
deterministic presentation of X having the least number of vertices among all
deterministic presentations of X.
One can prove that any two minimal deterministic presentations of an ir-
reducible soc shift are isomorphic as labeled graphs (see [LinMar, Theorem
3.3.18]), so that one can speak of the minimal deterministic presentation of an
irreducible soc shift.
In the following propositions, we clarify the relation between the irreducibil-
ity of a soc (or edge) shift and the strong connectedness of its presentations.
19
On the other hand, it can be easily seen that the strong connectedness of G
is equivalent to the following property of the n n adjacency matrix A of G:
for each i, j 1, . . . , n, there exists an m N such that the (i, j)entry of
A
m
is not zero.
Proposition 1.5.6 [LinMar, Lemma 3.3.10] Let X be an irreducible soc shift
and ( = (G, L) the minimal deterministic presentation of X. Then G is a
strongly connected graph.
Proposition 1.5.7 [LinMar, Proposition 2.2.14] If G is a strongly connected
graph, then the edge shift X
G
is irreducible.
As a consequence of this two facts, we have the following corollary that will
be useful in the sequel.
Corollary 1.5.8 Let X be an irreducible soc shift and ( = (G, L) the minimal
deterministic presentation of X. Then the edge shift X
G
is irreducible.
As we have seen, a shift is soc if and only if it is the image under a local
function of a shift of nite type. From the previous result it follows a stronger
property.
Corollary 1.5.9 A shift is an irreducible soc shift if and only if it is the image
under a local function of an irreducible shift of nite type.
Notice that, in general, the image under a local function of an irreducible
shift is also irreducible.
1.6 Decision Problems
Two natural decision problems arising in the theory of cellular automata con-
cern the existence of eective procedures to establish the surjectivity and the
injectivity of the transition function. For Euclidean cellular automata over full
shifts, Amoroso and Patt have shown in [AP] that there are algorithms to decide
surjectivity and injectivity of local maps for onedimensional cellular automata
(see also [Mu]). On the other hand Kari has shown in [K1] and [K2] that both
the injectivity and the surjectivity problems are undecidable for Euclidean cel-
lular automata of dimension at least two.
In this section we extend the problem to cellular automata over subshifts
of nite type of A
Z
, giving in both cases a positive answer to the existence of
decision procedures.
1.6.1 A Decision Procedure for Surjectivity
If X is a shift of nite type, the problem of deciding whether or not a function
: X X given in terms of local map is surjective, is decidable. In fact we
can decide if a local function : X A
Z
is such that (X) X.
20
Consider a transition function : X A

dened by a local rule ; suppose

(in this assumption there is no loose of generality) that X has memory 2M and
that is Mlocal. The function can be represented in this way. Consider the
presentation G of the edge shift X
[2M+1]
constructed in Section 1.4, the label of
the edge u a v L(X) (where u, v are two words in the language of X of length
M), is the letter (u a v), that is the letter to write under a in the output tape
in correspondence with u a v:
u
1
. . . u
M
a v
1
. . . v
M1
u
2
. . . u
M
a v
1
. . . v
M
q
(u1 . . . uM a v1 . . . vM)
In this way we get a labeled graph ( which is the presentation of the (soc)
shift (X). On this presentation we can check if any forbidden block of X
appears in (X); if none of them appears, we have that : X X and hence
we deal with a genuine cellular automaton.
To see whether or not the function is surjective, rst suppose that the shift
X is irreducible. Hence it can be easily seen that also (X) is irreducible and
hence we can construct the minimal deterministic presentation of X and (X),
respectively. These two presentations are isomorphic (as labeled graphs) if and
only if X = (X).
In the general case, Lind and Marcus give in [LinMar, Theorem 3.4.13] an
eective procedure to decide whether two labeled graph generate the same shifts.
1.6.2 A Decision Procedure for Injectivity
If X is a shift of nite type, the problem of deciding whether or not a function
: X X given in terms of local map is injective, is decidable.
As we have seen in the previous subsection, we can construct a labeled graph
( which is the presentation of the shift (X). From (, we construct another
graph ((. The vertices of it are the couples (i, j) where i, j 1(() are vertices
of (. There is an edge (i, j)
a
(h, k) labeled a, if in ( there are two edges in
c(() labeled a of kind:
i
a
h and j
a
k.
Notice that, in general, X
G
= X
GG
and hence, in our case, (( is a presentation
of (X).
A vertex (i, j) of ( ( is diagonal if i = j. Now, notice that the function
is noninjective if and only if on the graph ( there are two dierent biinnite
21
paths with the same label. This fact is equivalent to the existence of a biinnite
path on ( ( that involves a nondiagonal vertex. Hence, starting from the
graph ( ( we construct an essential graph that accepts the same biinnite
paths . Now it suces to check, on this latter graph, if some nondiagonal
vertex is involved.
1.7 Entropy
The entropy of a shift is the rst invariant we deal with in the present work. It
is a concept introduced by Shannon [Sha] in information theory that involves
probabilistic concepts. Later Adler, Konheim and McAndrew [AdlKoM] intro-
duced the topological entropy for dynamical systems. The entropy we deal with
is a special case of topological entropy and is independent on probabilities.
In this section we give the general denition of entropy for a generic shift.
We will see that this denition involves the existence of a suitable sequence of
sets that, in the case of nonexponential growth of the group can be taken as
balls centered at 1 and with increasing radius.
Then, following Lind and Marcus [LinMar, Chapter 4], we recall its basic
properties in the onedimensional case, also stating the principal result of the
PerronFrobenius theory to compute it.
Let (E
n
)
n1
a sequence of subsets of such that

nN
E
n
= and
lim
n
[
M
E
n
[
[E
n
[
= 0; (1.5)
if X A

is a shift, the entropy of X respect to (E

n
)
n
is dened as
ent(X) := limsup
n
log [X
En
[
[E
n
[
.
Condition (1.5) will be necessary in the proof of Theorem 1.7.1 and other as-
pects of its importance will be claried in Chapter 3. Notice that (1.5) implies
lim
n
|
+
M
En|
|En|
= 0 and lim
n
|

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

a local function. Then

ent((X)) ent(X) (that is, the entropy is invariant under conjugacy).
Proof Let be Mlocal and let Y := (X); we have that the function

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

is surjunctive. In fact, in Section 2.4 we prove that if the pe-

riodic congurations of a subshift X A

are dense, then a transition function

on X is surjunctive. We also prove that the density of the periodic congura-
tions is an invariant of the shifts, as is the number of the periodic conguration
with a xed period.
The remaining part of the chapter is devoted to establish for which class of
shifts the periodic congurations are dense. We prove in Section 2.5 the density
of the periodic congurations for an irreducible subshift of nite type of A
Z
and
hence, a soc shift being the image under a local map of a shift of nite type,
the density of the periodic congurations for an irreducible soc subshift of A
Z
.
We see that these results cannot be generalized to higher dimensions.
In Section 2.6 we introduce the notion of a group shift and (as a conse-
quence of a more general theorem in [KitS2]), we prove that for this class of
shifts the periodic congurations are dense. Finally, we list some wellknown
decision problems for Euclidean shifts proving that in the special case of a one
dimensional shift they can be solved; more generally they can be solved for the
class of group shifts using some results due to Wang [Wa] and Kitchens and
Schmidt [KitS1].
25
2.1 Surjunctivity
In this section we investigate under which hypotheses an injective selfmapping
of a set is also surjective, that is we study the class of selfmappings that are
either surjective or noninjective. This property is the socalled surjunctivity
and is due to Gottschalk (see [Gott]). Here we prove a sucient condition for
which a selfmapping of a topological space is surjunctive.
Denition 2.1.1 Let X be a set and : X X a function; is surjunctive if
it is either surjective or noninjective.
The simplest example is that of a nite set X and a selfmapping : X X;
clearly every function of this kind is surjunctive; in other words, for a selfmap-
ping of X we have injectivity surjectivity. Another example of surjunctive
function is given by an endomorphism of a nitedimensional vector space and
by a regular selfmapping of a complex algebraic variety (see [Ax]); many others
examples are given in [G]. Moreover, Richardson proves in [R] that each transi-
tion function of an Euclidean cellular automaton on the full shift is surjunctive.
Lemma 2.1.2 Let X be a topological space, let : X X be a closed function
and let (X
i
)
iI
be a family of subsets of X such that
X =

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

. In the Euclidean case, we imagine that a periodic conguration is

obtained repeating in each direction the same nite block. Hence translating
such a conguration, we get only a nite number of new congurations; this
property leads us to dene periodic a conguration whose orbit is nite.
Denition 2.2.1 A conguration c A

is nperiodic if its orbit c

= 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

is constant on the right cosets of its own

stabilizer H
c
(i.e. the subgroup of all h such that c
h
= c). Indeed, if h H
c
,
we have
c
|h
= (c
h
)
|
= c
|
.
Hence, if c is periodic, c is constant on the right cosets of a subgroup of nite
index. Now we prove that this property characterizes periodic congurations.
Proposition 2.2.2 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

be a shift whose set C

p
(X) := C
p
X of periodic
congurations of X is dense in X. Then every transition function : X X
is surjunctive.
29
Proof By Corollary 2.2.4, we have that the set P
n
(X) := P
n
X is nite;
now we prove that if is local then (P
n
(X)) P
n
(X). Indeed if c P
n
(X)
then H
c
H
(c)
because if h H
c
((c))
h
= (c
h
) = (c).
Hence, the index of H
c
being a divisor of n, the index of H
(c)
also divides n
and (c) P
n
(X).
By Lemma 2.1.2, is surjunctive. 2
Corollary 2.4.2 If is a residually nite group and : 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

is a shift such that C

p
(X) is dense in X and
: X A

is a local function, then C

p
((X)) is dense in (X).
Proof Set Y := (X); we rst prove that (C
p
(X)) C
p
(Y ). Indeed if
c X the stabilizer H
c
is contained in H
(c)
and if H
c
has nite index, then
also H
(c)
has nite index. Then C
p
(Y ) (C
p
(X)) (C
p
(X)) = (X) = Y .
2
Corollary 2.4.4 The density of the periodic congurations is an invariant of
the shift.
In general, given a shift X, denote with Q
n
(X) the set of the periodic con-
gurations of X with period n and let q
n
(X) be the cardinality of Q
n
(X). We
have that q
n
(X) is an invariant of X. Indeed, suppose that : X Y is a
30
conjugacy and let c Q
n
(X). We prove that H
c
= H
(c)
; as we have already
seen, H
c
H
(c)
. If h H
(c)
, we have (c
h
) = (c)
h
= (c); being injective,
we have c
h
= c.
From this fact it is clear that also the number p
n
(X) (that is, the cardinality
of P
n
(X)), is an invariant of X.
2.5 Periodic Congurations of Euclidean Shifts
In this section we concentrate our attention on the density of the periodic con-
gurations for a Euclidean shift. In the onedimensional case we prove this
density for the irreducible soc shifts. The situation in the twodimensional
case is deeply dierent: there are counterexamples of irreducible shifts of -
nite type for which the set C
p
is not dense. At the end of the section also the
notion of mixing shift is given and this property is strictly stronger than irre-
ducibility. Nevertheless, Weiss have proved in [Wei2]) the existence of mixing
twodimensional shifts of nite type X and of local functions : X X which
are injective and not surjective.
Proposition 2.5.1 If X A
Z
is an irreducible shift of nite type, then the set
C
p
(X) of periodic congurations of X is dense in X.
Proof Suppose that X has memory M; let c X and let u
n
:= c
|[n,n]
. Fix
a L(X) with [a[ = M; X being irreducible, there exist two words v
n
, w
n

L(X) such that
a v
n
u
n
w
n
a L(X).
Let c
n
be the periodic conguration
. . . a v
n
u
n
w
n
a v
n
u
n
w
n
a = a v
n
u
n
w
n
;
by Proposition 1.3.2, c
n
X. Moreover c
n|[n,n]
= c
|[n,n]
and hence lim
n
c
n
= c. 2
Corollary 2.5.2 If X A
Z
is an irreducible soc shift, then the set C
p
(X) of
periodic congurations of X is dense in X.
Proof By Corollary 1.5.9, we have that every irreducible soc shift is the
image under a local map of an irreducible shift of nite type. Hence Propositions
2.5.1 and 2.4.3 apply. 2
Counterexample 2.5.3 We now dene a reducible shift X A
Z
which is of
nite type but whose set C
p
(X) is not dense.
Let A = 0, 1 and let X be the shift of nite type with set of forbidden
blocks 01. Then the elements of X are the congurations

0,

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,

1 which is closed (and so not dense) in X. Notice

that, for this shift, a transition function is injective if and only if it is surjective
and hence surjunctivity holds even if the set of periodic congurations is not
dense. 2
Observe that, if X is a subshift of A
Z
, it is always possible to dene an irre-
ducible subshift X
2
of A
Z
2
consisting of copies of X; more precisely, a congu-
ration c belongs to X
2
if and only if each horizontal line of c (i.e. the biinnite
word (c
|(z,t)
)
zZ
, for each xed t Z), belongs to X. The irreducibility of X
2
can be easily seen: given two blocks of the language, it suces to translate one
of them in the vertical direction in such a way that the supports are far enough.
Moreover, it is obvious that the shift X
2
is of nite type if X is.
Counterexample 2.5.4 We show, using the above example, that Proposition
2.5.1 no longer holds for the irreducible shifts of nite type of A
Z
2
.
Let X
2
be the shift over the alphabet A = 0, 1 generated by the shift X of
the previous counterexample; then X
2
is irreducible and of nite type. The set
C
p
(X
2
) is in this case contained in the set of all those congurations assuming
constant value at each horizontal line. It is then clear that a conguration
assuming the value 1 at (0, 0) and 0 at (1, 0), cannot be approximated with any
sequence of periodic congurations. 2
Denition 2.5.5 A shift X A

is mixing if for each pair of blocks p

1
X
E
and p
2
X
F
, there exists an M > 0 such that for each / D
M
there is a
conguration c X such that c
|E
= p
1
and c
|F
= p
2
(notice that if M is big
enough, then E F = ).
In other words, a shift X is mixing if and only if for each pair of open sets
U, V X there is an M > 0 such that U V

,= for all / D
M
. Indeed,
given a pattern p with support E, consider the set U := c X [ c
|E
= p; U
is open because if E =
1
, . . . ,
n
then U =

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 also a group with product

dened as in the direct product of innitely many copies of G. Endowed with
this operation the space G

is a compact metric topological group.

Denition 2.6.1 If G is a nite group, a subset X 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

from X to X is, as we have seen, continuous

and it is also a group homomorphism because
(c
1
c
2
)

|
= (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

and G is a compact Lie group. Notice that, in this con-

text, a conjugation is a continuous groups isomorphism that commutes with the
action.
A consequence of this fact is the following theorem.
Theorem 2.6.2 [KitS2, Corollary 7.4] Let X be a compact group and
Aut(X) a nitely generated, abelian group; if acts expansively on X then the
set of periodic points is dense in X.
The following result gives an answer to the problems arising from Coun-
terexample 2.5.4.
Corollary 2.6.3 Let G be a nite group and let be an abelian group; if X
G

is a group shift, then the set C

p
(X) of periodic congurations of X is dense
in X.
Proof We have to prove that the group acts expansively on 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

is a closed invariant subgroup there exists a nite set D such that

X = c 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

. By Theorem 2.3.4 we have that is residually

nite.
2.6.1 Decision Problems for Group Shifts
Now we list some other decision problems arising in the case of Euclidean sub-
shifts of nite type.
The tiling problem: given a nite list T of forbidden blocks is X
F
empty
or nonempty? In fact the tiling problem is an equivalent formulation of
the domino problem, proposed by Wang [Wa].
A problem strictly related with this latter is the following: given a nite
list T of forbidden blocks, is there a periodic conguration in X
F
?
Given a nite list T of forbidden blocks, are the periodic congurations
dense in X
F
?
The extension problem: given a nite list T of forbidden blocks and given
an allowable block (that is a block in which does not appear any forbidden
block), is there a conguration in X
F
in which it appears? Clearly a
positive answer to the extension problem would imply a positive answer
to the tiling problem.
Now we prove that the answers for subshifts of nite type of A
Z
are all
positive: there are algorithms to decide, the tiling and the extension problems
and there is an algorithm to decide whether or not the periodic congurations
are dense in X. In order to see the rst two algorithms, consider, more generally,
a soc shift. If ( is a labeled graph ( accepting X (and we may assume that (
is essential), it can be easily seen that X is nonempty if and only if it exists
a cycle on the graph. Hence the shift is nonempty if and only if it contains a
periodic conguration. On the other hand, the language of X is the language
accepted by ( (( being essential); hence an allowable word is a word of the
language if and only if it is accepted by (.
To establish the density of the periodic congurations, suppose that X is
of nite type with memory M; one has that C
p
(X) is dense in X if and only
if C
p
(X
[M+1]
) is dense in X
[M+1]
. The shift X
[M+1]
an edge shift accepted
by the graph G constructed in Section 1.5 and hence the set C
p
(X
[M+1]
) is
dense in X
[M]
if and only if each edge of G is contained in a strictly connected
component of G, that is if the graph G has no edges connecting two dierent
connected components.
34
For the subshifts of nite type of A
Z
2
the answers are quite dierent; in this
setting Berger proved in [B] the existence of a nonempty shift of nite type
containing no periodic congurations and the undecidability of the tiling prob-
lem. Sucient conditions to the decidability of tiling and extension problems
are the following.
Theorem 2.6.5 (Wang [Wa]) If every nonempty subshift of nite type of
A
Z
2
contains a periodic conguration then there is an algorithm to decide the
tiling problem.
Theorem 2.6.6 (Kitchens and Schmidt [KitS1]) If every subshift of nite
type of A
Z
2
has dense periodic congurations then there is an algorithm to decide
the extension problem.
As a consequence of these facts, we have from Corollary 2.6.3 that if X G
Z
2
is a group shift, then the tiling and extension problems are decidable for X.
35
3. The MooreMyhill
Property
For Euclidean cellular automata, Moore [Moo] has given a sucient condition
for the existence of the socalled Garden of Eden (GOE) patterns, that is those
patterns 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.
In this chapter, we consider generalizations of the Moores property and
Myhills property to a generic shift. In details, the GOEtheorem has been
proved by Mach` and Mignosi [MaMi] more generally for cellular automata in
which the space of congurations is the whole Ashift A

and the group

has nonexponential growth; more recently it has been proved by Ceccherini
Silberstein, Mach` and Scarabotti [CeMaSca] for the wider class of the amenable
groups.
Instead of the nonexistence of mutually erasable patterns, we deal with
the notion of preinjectivity (a function : X A

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

(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 in Section 3.6 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.
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

be a transition function; two patterns p

1
and p
2
with the same support F are called mutually erasable if p
1
,= p
2
and
if for every c
1
, c
2
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

be two shift and : X Y be a function; a

pattern p Y
E
is Garden of Eden (briey GOE) with respect to , if for every
c X one has (c)
|E
,= p.
The Garden of Eden (GOE) theorem is the union of the following two theo-
rems.
Theorem 3.1.3 (E. F. Moore - 1962) If : A
Z
2
A
Z
2
is a transition
function and there exist two mutually erasable patterns, then there exists a
GOE pattern.
Theorem 3.1.4 (J. Myhill - 1963) If : A
Z
2
A
Z
2
is a transition func-
tion and there exists a GOE pattern, then there exist two mutually erasable
patterns.
In order to consider GOElike theorems not in the whole of A

but in a sub-
shift X A

, notice rst that two patterns of X are not necessarily extendible

by the same conguration of X. Therefore it could happen that two patterns
with support F for which there does not exist a common extension c
|F
, are
mutually erasable although the function is bijective. The notion that seems
to be a good generalization of the nonexistence of mutually erasable patterns,
is that of preinjectivity; we will see that if X = A

then the nonexistence of

mutually erasable patterns is equivalent to the preinjectivity of .
Denition 3.1.5 A function : X A

is called preinjective if when-

ever c
1
, c
2
X and c
1
,= c
2
only on a nite nonempty subset of , then
(c
1
) ,= (c
2
).
38
Proposition 3.1.6 Let : A

be a transition function; then is pre

injective if and only if there are no mutually erasable patterns.
Proof Let p
1
and p
2
be two mutually erasable patterns with support F.
Fix a A and dene c
1
, c
2
A

that coincide, respectively, with p

1
and p
2
on
F and such that
(c
1
)
|
= (c
2
)
|
= a
if / F. Then c
1
and c
2
dier only on a nonempty nite set (since this set is
contained in F), and (c
1
) = (c
2
), so that is not preinjective.
Suppose, conversely, that is not preinjective; there exist two congura-
tions c
1
and c
2
such that, for some non empty nite set F we have:
c
1|F
,= c
2|F
c
1|F
= c
2|F
(c
1
) = (c
2
).
Set p
1
:= c
1|F
+2M and p
2
:= c
2|F
+2M, where M is such that is Mlocal;
then we prove that p
1
and p
2
are mutually erasable. First p
1
,= p
2
and if c
1
,
c
2
are two congurations such that
c
1|F
+2M = p
1
c
2|F
+2M = p
2
c
1
= c
2
out of F
+2M
we have that ( c
1
) = ( c
2
).
Indeed, set D
M
:=
1
, . . . ,
m
and
i
:=
i
; if and D
M

F
+2M
we have (( c
1
))
|
= ( c
1|
1
, . . . , ) = (p
1
|
1
, . . . , ) = (c
1|
1
, . . . , ) =
((c
1
))
|
= ((c
2
))
|
= (c
2|
1
, . . . , ) = (p
2|
1
, . . . , ) = ( c
2|
1
, . . . , ) =
(( c
2
))
|
.
If, otherwise, D
M
F and if we suppose, for example, that
1
, . . . ,
i

F
+2M
and
i+1
, . . . ,
n
(F
+2M
), then (( c
1
))
|
= ( c
1|
1
, . . . , c
1|
n
) =
(c
1|
1
, . . . , c
2|
n
) = (c
2|
1
, . . . , c
2|
n
) = ( c
2|
1
, . . . , c
2|
n
) = (( c
2
))
|
. 2
One can prove (see [MaMi, Theorem 5]) that a transition function on 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

has the MooreMyhill property (briey MM

property), if for every cellular automaton (X, D
M
, ) the transition function
is preinjective if and only if it is surjective. The Mooreproperty is surjective
preinjective and the Myhillproperty is preinjective surjective.
In the sequel we will distinguish between these properties and the GOE
theorems for a local function. Indeed the former are properties of a single shift
but, on the other hand, we will speak of GOEtheorem whenever we have a
GOElike theorem for a local function between two possibly dierent shifts.
Notice that if a shift X has the Myhillproperty, than each transition func-
tion : X X is surjunctive (because, obviously, injectivity preinjectivity).
Proposition 3.1.10 The MMproperty is invariant under conjugacy.
Proof A conjugacy being both surjective and preinjective, it suces to
prove that the composition of two local preinjective function is still a (local)
preinjective function. Hence suppose that : X Y and : Y Z are
preinjective functions; if c
1
, c
2
X with c
1
,= c
2
and there exists a nite
subset F such that c
1|F
= c
2|F
, we prove that if is Mlocal then
(c
1
)
|(F
+M
)
= (c
2
)
|(F
+M
)
. Actually, if / F
+M
then D
M
F and hence
(c
1
)
|
= (c
2
)
|
. The set F
+M
being nite, we have ((c
1
)) ,= ((c
2
)) so
that is preinjective. 2
3.2 Amenable Groups
In this section we give the denition of amenability for a group ; using a char-
acterization of it due to Flner (see [F], [Gr] and [N]), CeccheriniSilberstein,
Mach` and Scarabotti have proved that the MMproperty holds for the full shift
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

be a transition function. Then is

surjective if and only if there are no mutually erasable patterns.
As a consequence of this theorem we have the following statement.
Corollary 3.2.6 (MMproperty for the full shift) For an amenable group
, the full shift A

has the MMproperty.

42
3.3 The MooreMyhill Property for Subshifts of
A
Z
As far as irreducible shifts of nite type are concerned, we have the following
result that stems from the works of Hedlund [H] and Coven and Paul [CovP].
Theorem 3.3.1 [LinMar, Theorem 8.1.16] Let X be an irreducible shift of nite
type, Y a shift and : X Y a local function. Then is preinjective if and
only if ent(X) = ent((X)).
Corollary 3.3.2 (MMproperty for irreducible subshifts of nite type
of A
Z
) An irreducible subshift of nite type of A
Z
has the MMproperty.
Proof If is preinjective, then by Theorem 3.3.1 we have ent(X) =
ent((X)). By Theorem 1.7.4, there does not exist a proper subshift of X whose
entropy equals that of X. Thus (X) = X and is surjective. Conversely, if
is surjective we have ent(X) = ent((X)) and Theorem 3.3.1 applies. 2
Now we show that the irreducibility condition in the above theorem cannot
be dropped.
Counterexample 3.3.3 Myhillproperty no longer holds for a subshift of -
nite type of A
Z
but not irreducible.
Let X be the full shift over the alphabet A = 0, 1; clearly X is irreducible
and of nite type. Consider the set

X 0, 1, 2
Z
given by the union X

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(

) consisting of all restriction to

each nite subset F of of the left translations by an element of

; a generic
element of 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(

) dened on F to a (nite) subset

E F, is still in 1(

).
Two elements , in are

equivalent if there exists

such that
= , that is

=

. Then the equivalence classes are the right cosets

[ . The pseudogroup 1(

) is dense if each equivalence class

is
such that for some R = R() N one has
_

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(

) is a pseudogroup of holonomies, that is

1. id
XF
: X
F
X
F
is an element of 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(

), then their composition is still in 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(

) is dense (that is, if

has nite index), we have dened a
dense pseudogroup of holonomies.
If Y A

is another stable and

invariant space, a function : X
Y is of bounded propagation M if it is the limit of a family of functions

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(

) commute with if commutes with the

action:
X
F
HF,
X
F
F

_
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(

) is dense, we say that the function admits a

dense holonomy.
Under these hypotheses and supposing that is amenable, we have the
following theorem.
53
Theorem 3.4.2 Let X, Y A

be stable spaces of bounded propagation and

: X Y a map of bounded propagation admitting a dense holonomy, then
ent(X) = ent(Y ) implies that is surjective if and only if it is preinjective.
Suppose that is a bounded propagation M function between two

invariant stable spaces and commutes with the

action, if the pseudogroup
1(

) is dense, we can write each as = d (

, d D
R
) and
(c)
|
= (c)
|d
= ((c

))
|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

shift spaces of bounded propagation and : X Y a local function,

then ent(X) = ent(Y ) implies surjective preinjective.
Corollary 3.4.4 (MMproperty for shifts of bounded propagation) If
X A

is a shift space of bounded propagation then X has the MMproperty.

3.5 Strongly Irreducible Shifts
In this section we give the denition of strong irreducibility for a shift. In general,
as we have seen at the end of Section 1.2, it is possible to give a denition
of irreducibility that generalizes the onedimensional one. But although we
can prove the MMproperty for irreducible shifts of nite type of A
Z
, this
irreducibility is too weak in the general case of subshifts of nite type of A

(see
Counterexample 3.3.5). We prove the MMproperty for the strongly irreducible
shifts of nite type of A

. On the other hand, we will see that a shift of bounded

propagation (that has, by Gromovs theorem, the MMproperty ), is strongly
irreducible and of nite type, but the converse does not hold.
Denition 3.5.1 A shift X is called Mirreducible if for each pair of nite sets
E, F such that dist(E, F) > M and for each pair of patterns p
1
X
E
and
p
2
X
F
, there exists a conguration c X that satises c = p
1
in E and c = p
2
in F. The shift X is called strongly irreducible if it is Mirreducible for some
M N.
54
In the particular case = Z, it can be easily seen that a shift X A
Z
is
Mirreducible if for each n M and for each pair of words u, v L(X), there
exists a word w L(X) with [w[ = n, such that uwv L(X).
The following theorem will be proved in the next section in the case of groups
with nonexponential growth and semistrongly irreducible shift.
Proposition 3.5.2 Let be an amenable group. Let X be a strongly irreducible
shift of nite type and let : X A

be a local and preinjective function. Then

ent((X)) = ent(X).
Proof Suppose that the memory of X is M, that X is Mirreducible and
that is Mlocal. Set Y := (X) and x an amenable sequence (E
n
)
n
; we have
[Y
E
+2M
n
[ [Y
En
[[A[
|
+
2M
En|
and then
log [Y
E
+2M
n
[
[E
n
[

log [Y
En
[
[E
n
[
+
[
+
2M
E
n
[ log [A[
[E
n
[
.
Taking the maximum limit and being
lim
n
[
+
2M
E
n
[
[E
n
[
= 0,
we have
limsup
n
log [Y
E
+2M
n
[
[E
n
[
ent(Y ).
Suppose that ent(Y ) < ent(X); then
limsup
n
log [Y
E
+2M
n
[
[E
n
[
< limsup
n
log [X
En
[
[E
n
[
;
so that there exists n N such that
log [Y
E
+2M
n
[
[E
n
[
<
log [X
En
[
[E
n
[
that is
[Y
E
+2M
n
[ < [X
En
[.
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 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

has the MMproperty.

We conclude this section proving that the property of bounded propagation
for a shift is strictly stronger than the union of strong irreducibility and nite
type condition.
Lemma 3.5.9 A shift X is of nite type with memory M if and only if each
conguration c 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 a shift of bounded propagation, then X is

strongly irreducible and of nite type.
Proof Suppose that X has bounded propagation M; if E, F are such
that dist(E, F) > M and p
1
X
E
, p
2
X
F
are two patterns of X, consider
the pattern p with support E F given by the union of the functions p
1
and
p
2
. Clearly p X
EF
because if E F and, for instance E, we have
(E F) D
M
E and hence p
|(EF)DM
X
(EF)DM
. A conguration
in X extending p is such that c
|E
= p
1
and c
|F
= p
2
. Hence X is Mirreducible.
60
Now suppose that c 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

(provided that is amenable).

In this section we dene another form of irreducibility: the semistrong irre-
ducibility. For soc shifts, this notion is equivalent to the general irreducibility in
the onedimensional case and, in all other cases, allows us to prove the Myhill
property for the subshifts of nite type (if has nonexponential growth).
Denition 3.6.1 A shift X is called (M, k)irreducible (where M, k are natural
numbers such that M k) if for each pair of nite sets E, D (the second
one is a ball centered at ) such that dist(E, D) > M and for each pair of
patterns p
1
X
E
and p
2
X
D
, there exists a conguration c X that satises
c = p
1
in E and c = p
2
in D, where is such that || k. The shift X
is called semistrongly irreducible if it is (M, k)irreducible for some M, k N.
Hence the dierence between this new property and the strong irreducibility
one, lies in the fact that the support of the second pattern must be a ball and
the conguration c merging the two patterns moves this support slightly.
Notice that this move is a translation and hence it make sense to say that the
conguration c restricted to D coincides with p
2
X
D
. Moreover, under
the previous hypotheses, the translated disk D is still contained in (D)
+M
;
indeed if D = D
r
and D
r
, then dist(, ) dist(, ) + dist(, )
r +|
1
| r +k. In particular we have that E D = .
In Denition 3.6.1 is in fact essential that, given a nite set F , it exists
such that the translated set F is still contained in F
+M
. If the group is
not abelian, the set F could be quite far from F. On the other hand the set
F is near to F, but it is not, in general, obtained from F under translation.
62
Consider, for example, the free group F
2
generated by the elements a and b. If
F = a
n
, b
n
with n > M, we have that does not exist an ,= 1 such that
F = a
n
, b
n
F
+M
= D(a
n
, M) D(b
n
, M).
Indeed, if the reduced form of is b or b
1
then dist(a
n
, a
n
) = |a
n
a
n
| =
|a
n
b
1
a
n
| n +1 and dist(a
n
, b
n
) = |b
n
a
n
| = |b
n
b
1
a
n
| n, that
is a
n
/ F
+M
. If, otherwise, the reduced form of is a
1
then b
n
/ F
+M
.
To have a counterexample also in the amenable case, consider the innite
dihedral group C
2
C
2
with the presentation a, b [ a
2
= b
2
= 1); we have that
if M = 1 there is no ,= 1 such that a, b D(a, 1) D(b, 1). Indeed if
= (ab)
n
with n > 0 then a = (ab)
n
a = a(ba)
n
, hence
dist(a, a) = |aa| = |(ba)
n
| = 2n > 1
and
dist(a, b) = |ba| = |ba(ba)
n
| = |(ba)
n+1
| = 2(n + 1) > 1.
In both cases, we have that a / D(a, 1) D(b, 1). If = (ab)
n
a with n > 0
then a = (ab)
n
, hence
dist(a, a) = |a(ab)
n
| = |(ba)
n1
b| = 2(n 1) + 1 = 2n 1
and
dist(a, b) = |b(ab)
n
| = |(ba)
n
b| = 2n + 1 > 1.
To have a D(a, 1)D(b, 1), it must be n = 1. Hence = aba and b = abab.
But, in this case, we have
dist(b, a) = |bab| = 3 and dist(b, b) = |babab| = 5.
For the symmetry between a and b, we have an analogous result if = (ba)
n
b
or = (ba)
n
.
This is the reason why, to avoid this problem, we require that the second set
in Denition 3.6.1 is a ball centered at . Then we consider the new center
(which is near to ). The ball D having the same radius as D, is obtained
by translation of it. As we have seen, if has nonexponential growth the sets
in the amenable sequence (E
n
)
n
are balls centered at 1. Hence if M is large
enough we have that E
n
E
+M
n
.
In the particular case = Z, it can be easily seen that a shift X A
Z
is
(M, k)irreducible if for each n M and for each pair of words u, v L(X),
there exists a word w L(X) with nk [w[ n+k, such that u w v L(X).
In order to see that in the onedimensional case irreducibility and semi
strong irreducibility are equivalent, we state the wellknown Pumping Lemma
as follows.
63
Lemma 3.6.2 (Pumping Lemma) Let L be a regular language. There exists
M 1 such that if uwv L and [w[ M, there exists a decomposition
w = xyz
with 0 < [y[ M so that for each n N we have uxy
n
zv L.
Moreover, one can take as M the number of vertices of a graph accepting L.
Corollary 3.6.3 If X A
Z
is a soc shift, then
X irreducible X semistrongly irreducible.
Proof If X is irreducible, we claim that X is (M, M)irreducible, where M
is given by the Pumping Lemma.
If n M and u, v L(X), there exists w L(X) such that uwv L(X).
We distinguish two cases.
If [w[ > n +M, then w = x
1
y
1
z
1
with 0 < [y
1
[ M and if w
1
:= x
1
z
1
, then
uw
1
v L(X) and [w[ M [w
1
[ [w[ 1. If [w
1
[ n+M, moreover we have
[w
1
[ [w[ M > n > nM. If [w
1
[ > n+M, we repeat the above construction
to obtain, for some i 1, a string of elements w
1
, . . . , w
i
, w
i+1
such that for
each j = 1, . . . , i + 1
1. uw
j
v L(X)
2. w
j
= x
j
y
j
z
j
with 0 < [y
j
[ M
3. w
j+1
= x
j
z
j
4. [w
j
[ > n +M for each j = 1, . . . , i
5. [w
i+1
[ n +M.
Then [w
i+1
[ [w
i
[ M > n > n M so that we can set w := w
i+1
.
In the second case, suppose that [w[ < n M; there exists w
1
L(X) such
that uwv w
1
uwv L(X) and [wv w
1
uw[ > [w[. In this way we obtain a
word w
i
L(X) such that uw
i
v L(X) and [w
i
[ n M. If, moreover,
[w
i
[ > n +M, we can apply the former case. 2
Now we prove a fundamental result that in the amenable case has been
proved in Proposition 3.5.2 of previous section. We refer to that proof for the
details.
Proposition 3.6.4 Let be a group of nonexponential growth. Let X be a
semistrongly irreducible shift of nite type and let : X 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

. If Y is a proper subshift of X then

ent(Y ) < ent(X).
Proof Let X be (M, k)irreducible. If Y X, there exists a conguration
c X which does not belong to Y and then there exists a disk D
r
such that
c
|Dr
X
Dr
Y
Dr
. Let (F
j
)
jN
= (D(
j
, r))
jN
be a (D
r
, 2M, R)net; then
c
|Dr
/ Y
D(
j
,r)
whenever D
k
; by Lemma 3.6.5, ent(Y ) < ent(X). 2
Proposition 3.6.7 Let be a group with nonexponential growth. Let X be a
shift, let Y be a semistrongly 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 previous hypotheses and let : X Y be
a local function. Clearly (X) is a subshift of Y . By Corollary 3.6.6, we have
that if (X) Y , then ent((X)) < ent(Y ). 2
Theorem 3.6.8 Let be a group of nonexponential growth, let X be a semi
strongly irreducible shift of nite type and let Y be a semistrongly irreducible
shift. If : X Y is a local function and ent(X) = ent(Y ), then preinjective
implies surjective.
Proof If is preinjective we have, by Proposition 3.6.4, that ent((X)) =
ent(X). Then ent((X)) = ent(Y ) so that, by Proposition 3.6.7, is surjective.
2
Hence we may conclude stating this (partial) generalization of Corollary
3.5.8.
Corollary 3.6.9 (Myhillproperty for semistrongly irreducible shifts
of nite type) Let be a group of nonexponential growth. Let X be a semi
strongly irreducible subshift of nite type of A

and let : X X be a transition

function. Then preinjective implies surjective.
68
Bibliography
[AdlKoM] R. Adler, A. Konheim and M. McAndrew, Topological Entropy,
Trans. Amer. Math. Soc. 114 (1965), 309-319.
[AP] S. Amoroso and Y. N. Patt, Decision Procedures for Surjectivity
and Injectivity of Parallel Maps for Tessellation Structures, J.
Comput. System Sci. 6 (1972), 448-464.
[Ax] J. Ax, The Elementary Theory of Finite Fields, Ann. of Math.
(2) 88 (1968), 239-271.
[B] R. Berger, The Undecidability of the Domino Problem, Mem.
AMS 66, 1966.
[CeMaSca] T. G. CeccheriniSilberstein, A. Mach` and F. Scarabotti, Ame-
nable Groups and Cellular Automata, Ann. Inst. Fourier (Greno-
ble) 49 2 (1999), 673-685.
[CovP] E. Coven and M. Paul, Endomorphisms of Irreducible Shifts of
Finite Type, Math. Systems Theory 8 (1974), 167-175.
[E] V. A. Efremovic, The Proximity Geometry of Riemannian Man-
ifolds (Russian), Uspehi Matem. Nauk 8 (1953), 189
[F] E. Flner, On Groups with full Banach Mean Value, Math.
Scand. 3 (1955), 243-254.
[Gott] W. Gottschalk, Some General Dynamical Notions, Recent Ad-
vances in Topological Dynamics, 120-125. Lecture Notes in Math.
318, Springer, Berlin, 1973.
[Gr] F. P. Greenleaf, Invariant Means on Topological Groups and their
Applications, van Nostrand Mathematical Studies 16, van Nos-
trand, New YorkTorontoLondon, 1969.
[G] M. Gromov, Endomorphisms of Symbolic Algebraic Varieties, J.
Eur. Math. Soc.1 (1999), 109-197.
[H] G. A. Hedlund, Endomorphisms and Automorphisms of the Shift
Dynamical System, Math. Systems Theory 3 (1969), 320-375.
69
[K1] J. Kari, Reversibility of 2D Cellular Automata is Undecidable,
Phys. D 45 1-3 (1990), 379-385.
[K2] J. Kari, Reversibility and Surjectivity Problems of Cellular Au-
tomata, J. Comput. System Sci. 48 1 (1994), 149-182.
[KitS1] B. Kitchens and K. Schmidt, Periodic Points, Decidability and
Markov Subgroups, Dynamical Systems, Proceedings of the spe-
cial year (J. C. Alexander, ed.) Springer Lecture Notes in Math.
1342 (1988), 440-454.
[KitS2] B. Kitchens and K. Schmidt, Automorphisms of Compact Groups,
Ergod. Th. and Dynam. Sys. 9 (1989), 691-735.
[LinMar] D. Lind and B. Marcus, An Introduction to Symbolic Dynamics
and Coding, Cambridge University Press, Cambridge, 1995.
[Moo] E. F. Moore, Machine Models of Self-Reproduction, Proc. Symp.
Appl. Math., AMS Providence R. I. 14 (1963), 17-34.
[MaMi] A. Mach` and F. Mignosi, Garden of Eden Congurations for
Cellular Automata on Cayley Graphs on Groups, SIAM J. Disc.
Math. 6 (1993), 44-56.
[Mil] J. Milnor, A Note on Curvature and Fundamental Groups, J.
Dierential Geometry 2 (1968), 1-7.
[Mu] G. Muratore, Automi Cellulari, Tesi di Laurea, University of
Palermo, Italy 1990.
[My] J. Myhill, The converse of Moores Garden of Eden Theorem,
Proc. Amer. Math. Soc. 14 (1963), 685-686.
[N] I. Namioka, Flners Condition for Amenable SemiGroups,
Math. Scand. 15 (1964), 18-28.
[R] D. Richardson, Tessellation with Local Transformations, J. Com-
put. Syst. Sci. 6 (1972), 373-388.
[Rot] J. J. Rotman, An Introduction to the Theory of Groups, Fourth
Edition, GTM 148, SpringerVerlag, New York, 1995.
[Sha] C. Shannon, A Mathematical Theory of Communication, Bell
System Tech. J. 27 (1948), 379-423, 623-656.
[

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