Вы находитесь на странице: 1из 12

Complex Sy st ems 5 (1991) 19- 30

De Bruijn Graphs and Linear Cellular Automata


Klaus Sutner
St evens Instit ute of Technology, Hoboken, N J 07030 USA
Abstract. De Bruij n graphs provide a convenient way to describe
configura tions of linear cellular automata (CAs). Using t hese graphs,
we give a simple quadrat ic t ime algorithm to det ermine whet her a
linear CA is reversible. Similarly, one can decide in quadrat ic time
whet her t he global map of the automaton is surjective. We also show
that every recursive configuration that has a predecessor on a linear
CA already has a recursive predecessor. By contradist inction, it is in
general impossible to comput e such a predecessor effectively.
1. Introduction
Cellul ar auto mata (CAs) are dynamic systems : t he space of all configur a-
ti ons is compact under t he usual pr oduct topology and t he global CA map
is a cont inuous shift-invariant operator on t his space. On t he ot her hand,
configurations of a one-dimensional or linear CA may also be const rued as
a biinfinit e word over t he state set of t he aut omat on. Corres pondingly, t he
global rule of the CA may be int erpret ed as a fini t e-st at e t ransducer t hat
act s on the space bi infinit e words. With t his int erpret at ion one can then
apply ideas from t he t heory of formal languages t o t he st udy of linear CAs.
Our work was motivat ed by t he search for t hose pr operti es of CAs t hat are
effect ively decidab le. It was shown by Amoroso and Pat t t hat it is decidable
whether t he global map of a one-dimensional CA is inj ect ive [1]. It is well
known t hat CAs are reversible- in t he sense t hat t here exists anot her CA
over t he same alphabet that reverses t he evolut ion of configur at ions on t he
first-if and only if t he global fun ct ion is inj ecti ve, t hat is, reversible in
t he set -t heoret ic sense. This was shown in [16]. Hence reversibi lit y of one-
dimensional CAs is decidable. By cont radist incti on reversibilit y of higher-
dimensional CAs fails to be decidable as was recentl y shown by Kar i [9].
The proo f is based on the existence of a combinatorial st ruct ure in t he plane,
namely a special direct ed t iling , which has no counterpart in one dimension.
Simi larl y one can show t hat sur ject ivity of t he global map is decidab le for
linear CAs but fails to be decidab le for all dimensions higher t han one.
The argument by Amor oso and Patt is purely combinat orial. Alt ern a-
t ive proofs were given by Culik [8] and Head [6] using aut omata t heoret ic
1991 Complex Systems Publications, Inc.
20 Klaus Sut ner
methods. Culik uses t he fact that a CA in dimension one is a special type
of generalized sequent ial mapping and br ings t o bear result s about t hese
mappin gs. In Head's work, the de Bruij n graphs intr oduced in [20] are used
t o reduce inj ecti vit y and surject ivity t o an ambiguity probl em for finit e au-
t omata.
Decidabilit y by itself is sti ll a rat her weak condit ion: t he exist ing algo-
ri thm may demand unr easonabl e computat ional resources. Thus a property
of a CA my be decidable but intract able. For example, t o test for surject iv-
it y one can det ermine whether a certain nondet erministic finit e automat on
(NFA) accepts all st rings over it s alphabet . The NFA can be described con-
venientl y in t erms of de Bruijn graphs, as shown in sect ion 2 below. The
st andard algorit hm t o determine whet her an NFA accepts all inpu ts involves
t he const ruc t ion of a det erministi c simulat ion automat on and in general re-
quires exponent ial t ime as well as exponenti al space. Indeed, for arbitrary
NFAs it is PSPACE-hard to test whether the automaton accept s all inputs
(see [3]).
However , we will show t hat for t he NFAs arising from linear CAs the
pr obl em can be solved in quadratic time. Hence sur ject ivity of a linear CA
map can be test ed in quadrat ic time. Our algorit hm uses only st andard graph
theoreti c met hods and is st raight forward to implement. Complet e pr ograms
written in Mat hematica are available from t he aut hor. A slight modificat ion
of our algor it hm allows one to det ermine reversibility of a linear CA quadr atic
t ime.
One of t he hist orically first probl ems to be st udied in connect ion with CAs
is t he existence of configurations wit hout pr edecessors. These configur at ions
are ofte n refer red t o as Gar dens-of-Eden [2, 12, 14]. Clearly a Garden-of-Eden
exists if and only if the global map of the automat on fails to be surject ive.
The de Bruijn graphs in our decision algorit hms can also be applied to the
st udy of pr edecessor configurat ions in linear CAs.
The existence of a pr edecessor configurat ion of a given t arget configu-
rat ion Y is t ant amount to t he exist ence of a path in the de Bruijn graph
whose label sequence is Y. For spat ially periodic configurat ions on linear
CAs, one can easily see t hat such a path exist ence problem can be solved in
nondet erministi c logarithmic space. It is shown in [18] t hat t he pr oblem is
indeed NLOG-complet e.
Ret urning t o infinit e CAs, Golze [4] st udied restrict ions of the global map
to vario us classes of configur at ions , namely finit e configur at ions, spacially
periodi c configurat ions, and recur sive configurat ions . It is shown t here t hat
for linear CAs any finit e configurat ion t hat has a predecessor already has a
periodic predecessor .
In this pap er we will demonstrat e t hat in any linear CA a recursive con-
figur ati on wit h a pr edecessor must already have a recurs ive predecessor. Our
pro of is based on a slight ly more general resul t about t he existence of re-
cursive pat hs in time-dependent graphs (see sect ion 3). Note t hat there is
no analogous result for dimensions larger than one . Indeed, it was shown by
Hanf and Myers [5, 13] t hat one can const ruct sets of t iles t hat permit a t iling
De Bruijn Graphs and Lin ear Cellular Automat a 21
of the whole pl ane, but any such tiling must be nonrecur sive. It follows t hat
for some two-dimensional CA the homogeneous t ar get configur ation consist-
ing of O' s only has no recursive pr edecessor , but fails t o be a Garden-of-Eden
(see also [4]).
At any rate, even in one dimension t he recur sive pr edecessor in general
cannot be effect ively const ructe d from the target cor-figuration. Even if it
is known that t here are only finit e possibl e pr edecessors, all of which are
recur sive, it is impossible t o choose the proper one algorit hmically.
Not e t hat a rough upper bound for t he complexity of a pr edecessor of a
recursive configur at ion can be deri ved from t he Kreisel Basis Theorem: t he
class of all pr edecessors of a recur sive configur ation Y is Hence, by t he
t heorem, Y must have a .6.g pr edecessor (see [17]). Our resul ts impr ove t his
upper bound to
Apart from the introduct ion t his pap er cont ains t hree sect ions. In sec-
tion 2 we pr esent quadrat ic t ime algorit hms for inject ivi ty and surject ivity.
Secti on 3 cont ains t he result s for recursive pr edecessors. Lastly, in sect ion 4
we summarize our result s and pose some open problems. To keep t his pa-
per reasonably short we will not review t he basic definiti ons from language
theory and recursion t heory. We refer t he reader to references [7] and [17].
2. Definitions
We will most ly consider one-dimensional cellul ar automata. Thus a configu-
ration of t he aut omat on is a map X : Z ----> from the set of all cells t o the
alphabet . Here Z denot es t he int egers and is the collect ion of states . For
our purposes it is convenient t o t hink of configur at ions as biinfinit e words.
For any alphab et denot es by and t he collect ion of finite
words, infini t e words, coinfinite words, and biinfinit e words over respec-
ti vely (our w notation is taken from [15]). For a word X let Xi be t he
ith symbol in X and for -00 :::; n :::; m :::; 00 let X[n :m] be the subword
X
nXn
+1 .. . X
m
of X. Define t he s-fusion operat ion on words in E" as follows:
U<::) v = w =} 3x E Uo , Vo E ( U= uox , v = xVo, w = uoxvo )
Not e t hat <::) is a partial operation. The s-blocking operation on is defined
by (3 (x ) := (x [l: s], . .. , x [n - s + l:n]) , where n := [z]. Similarly one defines
fusion and blocking on infinite words. Since t he app ropriate s will always be
clear from context , we have chosen not to burden our notation by displaying
it , say as a subscript .
A linear CA may now be described as follows. A local rul e is a map
p : ----> Here r 1 is t he radi us of t he rule. For X E define t he
global rule (also denot ed by p) by
p(X)(i ) := p((3(X)i )
An elegant represent at ion of linear CAs uses lab eled de Bruijn graphs
(see [20]). To this end let be an alphabet and s 1 a number , and define
22 Kl aus Sutner
Fi gure 1: The de Bruijn graphs B(2, {O, I }) and B
150
for rule 150.
t he de Bruijn graph B (s, I;) as follows: B (s, I;) has vertex set I;s and edges
(ax, xb) for all a,b E I; , x E I;s-l. Figure I shows B (2,{O,I}) .
Now consider a rule p of radius r ::::: lover alphabet I; . Define B
p
to be the
de Bruijn gr aph B(2r, I;) , where edge (ax , xb) is lab eled by p(axb) E I;. Not e
t hat t his is equivalent to lab eling edge (u, v) by p(u 0 v). Here 0 denot es
(21' - l j-fusion. A biinfi nit e path U in B
p
is but a biinfinit e word over I;2r.
We associate two biinfinit e words over I; , t he trace of U and t he labeling of
U, with U as follows:
tr (U) = QUi
lab(U) = p(U
i
0 Ui+l)
It is clear from the definiti ons t hat p(t1' (U)) = lab(U). Thus B
p
may be
const rued as a Fischer automaton (a nondet erminist ic finit e automat on where
every state is init ial as well as final ) that recognizes biinfinit e words in p(WI;W).
For t he purpose of our inj ecti vity test ing algorit hm, recall t hat a subset
V
o
of a dir ect ed graph G = (V, E ) is strongly connected if and only if for
any two vertices x and y in V
o
there exists a path from x to y. A strongly
connected component (SCC) is a maxi mal st rongly connected subset . It is
well known t hat t he SCCs of a graph can be computed in t ime linear in t he
size of t he graph (see [11]). Aft er computing t he SCCs, one can form a new
direct ed graph 1), t he collapse of G. 1) contains one node for each SCC of
G. There is an edge in 1) from C to C' if and only if t here is a path in G
from some vertex in C t o some vertex in C'. Pl ainl y, 1) is acyclic. Hence 1)
De Bruijn Graphs and Linear Cell ular Automata 23
can be topologically sorte; t hat is, one can enumerate t he vertices of D as
{C
1
, . . . , Cd in such a way t hat whenever t here is an edge (C
i
, C
j
) in D t hen
i < j . Indeed, t he sorting can again be accomplished in linear time (see [11]).
The top ological sort ing can be used to eliminate all transient points of G,
that is, vert ices t hat fail t o lie on at least one biinfinit e path in G. Call an
SCC t rivial if and only if its induced subgraph contains no edges. Observe
t hat a vertex is nontransient if and only if it lies in a nont rivial SCC or
it lies on a pat h from a nont rivial SCC to another. This gives rise to t he
following algorit hm to delet e all t he nontransient point s. First mark all t he
nont rivial SCCs in the collapse D. Then sweep across the graph in order of
t he topological sort . Whenever a marked vert ex is encountered, mark all its
immedi at e successors. Upon complet ion of t he sweep, delet e all unmarked
nodes. Then reverse all arcs in D and repeat t he process. It is easy to see
t hat upon complet ion of t he second step exactly t hose SCCs surv ive t hat
cont ain nont ransient vert ices of G. Both sweeps are linear in t he size of G.
For future reference, denot e by G t he graph obtained from G by delet ing all
transient point s.
An interesti ng property of CAs is t ha t t hey are reversible- in t he sense
t hat t here exists anot her CA over t he same alphabet t hat reverses t he evo-
luti on of configurat ions on t he first - if and only if t he global funct ion p is
inj ective. This was shown in [16]. Since inj ectivi t y can be t ested in quadratic
t ime, we have t he following t heorem.
Theorem 2.1. For one-dimensional cellular automata, reversibili ty is decid-
able in quadratic time.
Proof. It follows from t he previous remarks that rule p fails to be injecti ve
if and only if t here are t wo disti nct biinfinit e paths V and U in t he de Bruijn
graph B
p
wit h the same sequence of labels: lab(V ) = lab(U).
The lat t er condit ion can be test ed as follows. Define a new graph B;
t hat has vert ices ~ 2 r X ~ 2 r . Fur thermore, [(ax, by), (xa' , yb')J is an edge in
B; if and only if p(axa') = p(byb'). Also, let 6. := { (w,w) Iw E ~ 2 r } denot e
t he diagonal in B;. It is well known t hat de Bruij n graphs are Hamil tonian.
Since 6. is an isomorphi c copy of B
p
, it follows t hat 6. is contained in one
SCC Ce:,. of B; . As described in t he remar ks pr eceding t he t heorem, let B ~
be t he graph obtained from B;by removing all t ransient point s.
Claim: Rul e p is inject ive if and only if 6. = B ~ .
ote t hat any two biinfinit e paths U and V in B
p
wit h the same labeling
t race a biinfinite pat h in B; and hence in B ~ . The convers e also holds.
To verify the claim, first suppose p fails t o be inj ective. Then we can
choose U '" V but wit h p( tr(U) ) = p( tr (V ) ) , t ha t is, wit h t he same
labeling in B
p
. Then, however , t he corr esponding path in B ~ cannot lie
ent irely wit hin 6. , and it follows t hat 6. '" B ~ .
24 Klaus Sutner
For t he oppos ite directi on suppose we have a biinfinit e path W = ( (Vi,Ui ) )
in B ~ , not ent irely in t:. . Letting V = (Vi) and U = (U
i
) it is clear t hat U
and V ar e distinct , but p( tr(V) ) = p( t r(U) ) . Hence p is not injecti ve and
we are t hrough.
As pointed out above we can construct B ~ in ti me linear in t he size of B;
and t herefore quadratic in the size of B
p
. However , t he size of the de Bruijn
graph is pr ecisely t he same as t he size of the local rule p, spec ified as a table.
Thus , a one-dimensional rul e p can be t est ed for injecti vity in quadratic t ime .
Reversibi lity in t he last lemma refers to injecti vity of the global functi on
for arbitrar y configurations . One frequently st udies t he subclass of finit e
configurat ions , namely, biinfinit e words of t he form wozow, wher e Z E I;*
and 0 is a special symbol. Let us call p locally injective if and only if, for all
finit e configurati ons X and Y , p(X) = p(Y ) implies X = Y . Let X = * Y iff
X and Y disagree onl y in a finit e number of places. It is st raight forward to
show that p is locally injective if and only if, for all configurat ions X = * Y,
p(X) = p(Y) implies X = Y . It can be shown t hat local injectivity of p is
equivalent to surjectivity of p (see [16] and [10]). The lat t er characterizat ion
is t he basis for the next t heorem.
Theorem 2.2. For one-dimensional cellular automata, surjectivity of the
global map is decidable in quadratic time.
Proof. Let B ~ be t he nontr ansient part of the product de Bruijn graph as
in the last proof. Recall t hat Ceo. denot es t he SCC containing t:. .
Claim: Rule p is locally inj ect ive if and only if t:. = Ceo. .
To see this, first assume t hat p fails to be locally inj ecti ve. Then by the
remark preceding t he theorem, there are two configur at ions X 01= Y that
disagree only in a finit e number of places and map to the same configuration
under p. Let W be t he corresponding biinfinite pat h in B ~ . Plainly, W has
an infini te ini t ial segment and an infinit e final segment in t:. . However , since
X 01= Y t he path W cannot be entirely in t:. ; consequently, t:. 01= Cs .
For t he opposit e dir ect ion suppose t hat t:. 01= Ceo.. Then t here exists a
biinfinit e pat h W in B ~ such that W has an infinit e init ial segment and
an infinite final segment in t:. , but W does not lie ent irely in t:.. W read-
ily translat es int o two configurat ions t hat wit ness p's failur e to be locally
inject ive.
As in t he last theorem it can be seen that t he necessar y computat ions
take only quadrat ic t ime.
Ex ample. Consider the elementary rule p wit h code number 150 (addit ion
modulo 2). The SCC of B ~ 5 0 containing t:. consists only of t:. itself. Indeed,
t here are exactly two SCCs: t:. , and all ot her vertices. Hence, rule 150 is
locally injecti ve but not inject ive. Simi larl y for rule 90 one can see th at BJo
De Br uijn Graphs and Linear Cell ular Automata 25
is par ti ti oned int o exact ly three SCCs, one of t hem being 1::>.. Hence rul e 90
is also locally injective but not inj ective.
Rule 30 produces a product graph wit h t wo nont rivial SCCs, one of t hem
I::>. and t he ot her containing 8 vert ices. Two vertices in Bi o are transient, and
the collapse of Bjo has diamond shape. Again, rul e 30 is locall y injecti ve bu t
not inj ective.
As an example of an inj ective elementary CA, consider rule 15. Bis has
13 sccs, but only one nontrivial SCC, namely 1::>..
Lastl y, Bi 7 is st rongly connect ed. Thus rul e 37 is not even locall y inj ec-
tive.
3. Recursive predecessors
In answer t o a question posed by Golze [4J we now show t hat , in any linear
cellular automat a, a recurs ive configur at ion with a pr edecessor must already
have a recur sive predecessor. To demonstr ate the exist ence of a recursive
pr edecessor , we first establish a slight ly more general resul t about t he ex-
ist ence of recur sive infinit e paths in cert ain recurs ive graphs . The graphs
under considerat ion are given by a finit e st atic graph and a recursive func-
ti on t hat descr ibes the evolut ion of t he static graph in t ime. A similar type
of t ime-depend ent graph was used in [19J t o establish PSPACE-hardness of
moti on planning problems in t he presence of moving obstacles.
A time-dependent graph G(W) consist s of a finite direct ed graph G =
(V, E) and an alph ab et ~ C P(E). For any word W E ~ w the time-depend ent
graph G(W) is defined as follows. The vertices of G(W) are V x N, and t he
edges are given by E(W) := { ( (p,t) , (q, t + 1) ) I (p, q) E W(t) }.
G is the underl ying static graph. For lat er use let us define the reachabl e
set of a vertex (p,t) at time s as
R((p, t ), s) := {q E V I t here exists a path from (p,t) to (q,s)
in G(W) }
For a set P of points in V x N let R(P, s) = U XEP R(x , s ). We will be
int erest ed mostl y in t he case where W const rue d as a map from N to ~ is
a recursive function. G(W) is t hen called a recursive t ime-dependent graph.
Note t hat , in a recursive time-dependent graph, R((p, t ), s) is computable
uniformly in p, s, and t .
Example. Let G = (V, E ) be t he complete direct ed graph on two points
with self-loops; t hus V = {O, I} and E = {(O, 0), (0, 1) , (1,0) , (1, I n . Define
alphabet ~ = {S, B , L, R} by S := E , B := {(O, 0) , (1, In, L = {(O, on, and
R = {(I , In. Here S st ands for switch, B for both, L for left , and R for
righ t (see figur e 2).
Now let W E ~ w be an infinit e word. A moment ' s t hought reveals t hat
G(W) has an infinit e path if and only if W does not contain a subword of
the form LB kR or RB
k
L. Thus the collect ion of all t ime-dependent graphs
26
I
L
CD
CD
I
@
@
B
R
Kl aus Sutner
I
I
Figure 2: The four components of t he time-dependent graph G.
over G with an infinit e path forms a regular set in the sense of [15J. This is
t rue for all t ime-dep endent gr aphs.
We claim t hat for the st atic graph G defined as above any recur sive t ime-
dependent graph G(W) with an infinit e path has a recursive infinit e path.
To see t his, note t hat we can disti nguish t he following two cases.
Case 1. W contains only finit ely many symbols S . Then there is a recursive
path of t he form P = Poi
w
where Po is finit e and i = 0 or i = 1.
Case 2. W contains symbol S infini tely oft en. In t his sit uation we can ob-
t ain a recursive path by sear ching for t he next occurrence of S and
choosing, say, t he lexicographically first finit e path t o t hat next occur-
rence of S . More pr ecisely, suppose we have already const ructed an
ini ti al segment of t he recursive path P up to some t ime t such t hat
W(t) = S . Then we can compute t' > t mini mal such t hat W (t' ) = S .
Since W[t:t'J E S( B + L) *S+S(B+R)*S, we can exte nd P up to t ime
t
l
It should be noted, however , t hat t he recurs ive path cannot be generated
effect ively from W; see also lemma 3.1 below.
Theorem 3.1. Every recursive time-dependent graph with an infini te path
bas a recursive infinite path .
Proof. Let G(W) be a recursive t ime-dependent graph with an infinite path.
For t .:::: 0 define the support of t by spt(t) := { P(t ) I P an infini t e path}.
A set Q c V is persist ent if and only if there are infinit ely many t such t hat
De Bruijn Graphs and Lin ear Cellular Automata 27
spt(t) = Q. Since G(W) cont ains an infinit e pat h, we may pick a nonempty
persist ent set Q. Not e t hat {t I spt(t) not persist ent} is necessari ly finit e.
Hence t here exists a to 2: 0 such t hat spt (t
o
) = Q, and for all t 2: t o the
support of t is persist ent. Let us define a pr edicate P (T, s) on N x N as
follows:
P(T,S ) <===? Vq E R(QX {T},S)- Q3 t> s ( R(( q,s) , t) = 0 )
Thus P (T, s) holds if and only if no pat h in G(W) st arting at some node in
Q at time T and passing through a node not in Q at t ime t has an infini te
ext ensi on. Clearly P is recursively enumerable. Also note that t he domain
of P contains all t imes t such that spt(t) = Q. We have t he following claim.
Claim: For all t 2: to, P (t
o
, t) if and only if spt( t) = Q.
First suppose t hat P (t
o
, t) holds. Since every infini t e pat h must pass
through Q at time to, no infini te path can fail t o pass through Q at t ime t .
All paths through Q x {to} and (V - Q) x {t} are finite by our assumption.
Hence spt (t ) C Q, and t her efore spt (t ) = Q by t he minimality of Q.
For t he opposit e directi on let spt (t ) = Q. Consider a point q not in Q, but
with (q, t) reachable from Q x {t o}. If we had Vt' > t (R ((q,t),t' ) # 0) t hen,
by compactness , t here would be an infinit e path start ing at (q, t ). However ,
q would t hen be in the support oft and hence in Q, which is a cont radict ion.
Thus P (t
o
, t) holds as required .
Using L; I-uniformization we may choose a parti al recursive functi on f
that uniformizes P. Since the domain of f is the same as t he domain of P,
it must include all ti mes t such t hat spt(t) = Q. Define t
n
:= r(to) for all
n 2: o. It follows from the claim that the sequence (ti)i>Ois well defined.
Also note that (ti)i>O is st rict ly increasing.
We can now const ruct a recur sive infini t e path through G(W) as follows.
First choose an init ial segment from time t = 0 to a point Po in Q at t ime
t o. Suppose a par t of t he path up t o (Pi, ti) has been const ructed, so Pi E
Q. Since spt (t
i
) = Spt (ti+l) = Q, it follows that there exists a pat h from
(Pi, t
i)
to (Pi+I, ti+!) for some Pi+! E Q. Choose one such pat h, say t he
lexicographically first . Since the sequence (ti)i>Ois computable, we obtain a
comput able infini t e path t hrough G(W). -
Theorem 3.2. In any linear cellular automata the exist ence of a predecessor
of a recursive configuration impli es the exi stence of a recursive predecessor.
Proof. In light of t he last theorem it suffices to show how t he existence of
recursive pr edecessors can be translated int o t he existence of infini te recur sive
paths in a suitable time-dependent graph. As we saw in the last sect ion, the
existence of a pr edecessor of some configurat ion X E wL;w is equivalent t o the
exist ence of an biinfinit e path in the de Bruijn graph of t he automaton. Now
let B; be t he graph obtained from B
p
by reversing all the edges. Clearly t here
is a biinfinite pat h labeled X in B
p
if and only if for every decomposition
28
Klaus Sutner
x = X
o
0 Z 0 Xl> X
o
E wL; , Z E L;2r - \ .x, E L;w, t here is an infinit e path
(one-way!) in B
p
starting at Z labeled X
o
, and an infinit e path in B; start ing
at Z labeled Xl'
Now fix some decomposition X = X
o
0 Z 0 Xl . For any symbol (J E L;
let E" := { e I e an edge in B
p
lab eled (J}. Thus, for any word Z 0 V E L;w,
we have a t ime-depend ent graph Bp(Z 0 V ) t hat contains an infini t e path
starting at Z if and only if t here exist s a biinfinit e word X such t hat p(X) =
U 0 Z 0 V. If X is recursive, so is Bp(Z 0 V ). Hence, by t he last t heorem
ther e exists a recursive infinit e pat h in Bp(Z 0 V) if t here exists any path at
all.
The argument for t he reverse graph B; is ent irely similar. Thus we have
two recursive paths in the two t ime-dependent graphs t hat can be combined
to produce one recursive biinfinit e path in t he original de Bruijn gr aph B
p
.
As we have seen, this path corresponds t o a recur sive predecessor of X .
The proof of t heorem 3.1 suggests t hat t he recursive path cannot in gen-
eral be obtained effect ively from t he target configuration. For example, it is
not hard to show t hat per sist ence is not a recurs ive proper ty. Thus one can-
not hope to comput e the paramet er Q used in t he definit ion of t he sequence
The same holds for paramet er to. The following lemma shows t hat
indeed recursive predecessors in one-dimensional CAs cannot in general be
const ructe d effect ively.
Lemma 3.1. Let p be a linear cell ular automaton. Then in general there
is no effective method to determine a recursive predecessor for a recursive
configuration with a predecessor.
Proof. It is routine to const ruct a linear CA with alphabet L; = {a, b, c, d}
such that
p(Waba
W
)
p(WcdcW)
p(WaW)
All other blocks are mapped t o d. For e 2: 0, define a wor d 1";, E wL;w as
follows .
Y, (n) {
if n :::; 0 is mi nimal such t hat {e}n(e) 0,
if n :::; 0 is minimal such that {e}n(e) 1,
ot herwise
Here {e}n(x) y means t he et h parti al recursive fun cti on on input x con-
verges afte r n steps wit h output y. Similarl y {e}(x ) y means t he et h partial
recursive fun ct ion on input x converges on out put y . To det er mine whether
{e}(e) y for some y is the Halting problem and well known to be undecid-
able (see [17]) . Not e t hat Y
e
= waw, 1";, = W abor', or 1";, = W aca" , depending on
whet her a certai n computat ion diverges, converges on 0, or converges on 1.
De Bruijn Graphs and Linear Cellular A utomata 29
Thus every configurat ion 1';, has a recursive predecessor. Now supp ose t hat
such a predecessor (or rather, an index t hereof) could be computed effec-
t ively. Then t he set A := { e I X( O) = a, X the pr edecessor of 1';, } would be
recursive. By t he S;;:' t heorem there is a pr imit ive recursive function p such
t hat {p(e)}(x) ~ 1 for all x if and only if {e}(e) converges and {p(e)}(x ) di-
verges for all x ot herwise. But t hen {e}(e) converges if and only if p(e) 'f- A.
It follows from the undecidability of t he Halt ing problem t hat A is not re-
curs ive, and we have the desired cont radict ion.
Not e t hat t he configurat ions Y
e
from above provide a simple proo f for t he
assert ion that in two-dimensional CAs a recursi ve t arget configurat ion may
fail to have a recur sive predecessor , but st ill not be a Garden-of-Eden. To
this end one arranges all t hese configurations int o a two-dimensional t ar-
get configurat ion, say parallel t o t he z-axis in t he upper half-plane and
separate d by blanks. Any pr edecessor of this configurat ion rest ricte d to
the y-axis is essent ially t he characteristic func tion of t he diago nal set K =
{ e I {e}(e) converges }. Hence no predecessor can be recursive.
4. Conclusion
We have shown t hat reversibility of a one-dimensional cellular automaton
is decidab le in quadrat ic time using standard graph theoretic algor it hms.
Simil arly one can decide in quadratic ti me whet her the limi t set of t he au-
tomaton is the whole space of configurations . Both propert ies ar e known t o
be undecidable even in two dimensions (see [9]) .
Our pro of shows t hat one can det ermine in polynomial t ime whet her
cert ain restrict ed nondet erminist ic automata accept every string over t heir
input alphab et . The automata under consideration here have t he special
propert y t hat all states are ini ti al and final. Such devices are usually referr ed
to as Fischer automata. We do not know whether acceptance of all input s can
be t ested in polynomial time for ar bit rary Fischer automata. For arbit rary
nondet erministi c aut omata the problem is known to be PSPACE-complete .
By unfolding de Bruijn graphs int o an associated infinite graph, we have
shown that all recursive configurations wit h an ar bit rary predecessor must
already have a recur sive predecessor. However , t he recur sive pr edecessor in
general cannot be effect ively comput ed from the target configurat ion. Again ,
there are no analogous resul t s for two-dimensional CAs; in a two-dimensional
CA even a t rivial t ar get configurat ion may have only nonrecur sive pr edeces-
sors (see [5, 13]).
References
[1] S. Amoroso and Y. N. Patt , "Decision P rocedures for Sur je ct ivity and Injec-
t ivity of Parallel Maps for Tesselation Struct ur es ," Journal of Computers and
System Science, 6 (1972) 448-464.
[2] A. W. Burks, Essays on Cellular A utomat a (Universi ty of Illinois P ress,
Urbana-Champaign, Illinois, 1970) .
30 Klaus Sutner
[3] M. R. Garey and D. S. Johnson , Computers and Intractabili ty (Freeman,
1979).
[4] U. Golze, "Differences between 1- and 2-Dimensional Cell Spaces," in
Automat a, Languages and Development , edited by A. Lindenmayer and
G. Rozenberg, pp. 369-384 (North-Holland, 1976).
[5] W. Hanf, "Nonrecursive Tilings of t he Plane I ," Journal of Symbolic Logic,
39(2) (1974) 283-285.
[6] T. Head , "Linear Cellular Automata: Inj ect ivity from Ambiguit y," fort hcom-
ing.
[7] J. E. Hopcroft and J. D. Ullman , Introduction to A utomata Th eory, Lan-
guages and Computation (Addison-Wesley, 1979).
[8] K. Culik II , "On Inverti ble Cellular Automata," Complex Systems, 1(6)
(1987) 1035-1044.
[9] J . Kari , "The Undecidability of Inj ectivity and Surj ecti vit y of Two-
Dimensional CAs," CA'89, Los Alamos, New Mexi co.
[10] A. Maruoka and M. Kimura, "Injec t ivit y and Surj ectivity of Par allel Maps
for Cellular Aut omat a," Journal of Computers and Syst em Science, 18 (1979)
47-57.
[11] K. Mehlhorn , Da ta St ructures and Algoritbms 2: Grapb Algorit bms and NP-
Complet eness (Springer-Verl ag, Berlin, 1984) .
[12] E. F . Moore, "Machine Models of Self-Reproduction," in Essays on Cell ular
A utomata [2] .
[13] D. Myers , "Nonr ecur sive Tilings of the Plane II," Journal of Symbolic Logic,
39(2) (1974) 286-294.
[14] J. Myhill, "The Converse of Moore's Garden-of-Eden Theorem," in Essays on
Cellular Automata [2].
[15] M. Nivat and D. Pe rrin, "Ensembles Reconna isable de Mot s Biinfinis," Tech-
nical Repor t 84-68, Universite Paris 7.
[16] D. Richar dso n, "Tesselat ion wit h Local Transformation," Journ al of Compu t-
ers and System Science, 6 (1972) 373-388.
[17] J . R. Shoenfield, Me ihemeii csl Logic (Addison-Wesley, 1967).
[18] K. Sutner , "The Complexity of Finit e Cellular Automata," forthcoming.
[19] K. Sutner and W. Maass, "Mot ion Plan ning among Time-Dependent Obst a-
cles," Acta Informati ca, 26 (1988) 93-122.
[20] S. Wolfram, "Comp utation Theory of Cellular Aut omat a," Communicat ions
in Mat bemat ical Physics, 96 (1) (1984) 15- 57.

Вам также может понравиться