Mathematical systems of simple con-

struction are found capable of

highly complex behavior
with many universal
features.
 by Stephen Wolfram

t appears that the basic laws of physics relevant to everyday phenomena are now known. Yet there are many

I everyday natural systems whose complex structure and behavior have so far defied even qualitative analysis. For
example, the laws that govern the freezing of water and the conduction of heat have long been known, but
analyzing their consequences for the intricate patterns of snowflake growth has not yet been possible. While many
complex systems may be broken down into identical components, each obeying simple laws, the huge number of
components that make up the whole system act together to yield very complex behavior.
In some cases this complex behavior may be simulated numerically with just a few components. But in most cases
the simulation requires too many components, and this direct approach fails. One must instead attempt to distill
the mathematical essence of the process by which complex behavior is generated. The hope in such an
approach is to identify fundamental mathematical mechanisms that are common to many different
natural systems. Such commonality would correspond to universal features in the behavior of
very different complex natural systems.
To discover and analyze the mathematical basis for the generation of complexity,
one must identify simple mathematical systems that capture the essence of
the process. Cellular automata are a candidate class of such systems. This
article surveys their, nature and properties, concentrating on funda-
mental mathematical features. Cellular automata prorrlise to
provide mathematical models for a wide variety of
complex phenomema, from turbulence in fluids to
patterns in biological growth. The general
features of their behavior discussed here
should form a basis for future
detailed studies of such
specific systems.
 The Nature of Cellular Automata
and a Simple Example

Cellular automata are simple mathemati-

cal idealizations of natural systems. They
consist of a lattice of discrete identical sites,
each site taking on a finite set of, say, integer
values. The values of the sites evolve in
discrete time steps according to deterministic
• ...
10110100011010110100

......
• rules that specify the value of each site in
~ terms of the values of neighboring sites.
Cellular automata may thus be considered as

I discrete idealizations of the partial differen-

tial equations often used to describe natural
systems. T.heir discrete nature also allows an
important analogy with digital computers :
Fig. 1. A typical c01ifiguration in the simple cellular automaton described by Eq. 1,
consisting of a sequence of sites with values 0 or 1. Sites with value 1 are represented
by squares: those with value 0 are blank.
cellular automata may be viewed as parallel-
(
,
processing computers of simple construction.
As a first example of a cellular automaton,
consider a line of sites, each with value 0 or I
(Fig. I). Take the value of a site at position i

l• on time step t to be a (t) . One very simple rule

I
for the time evolution of these site values is

•• a(l+l)
i
= a (l) +
i- I
a(l)
i+l
mod 2 , (I)

• where mod 2 indicates that the 0 or I

• remainder after division by 2 is taken. Ac-

I
•
cording to this rule, the value of a particular
site is given by the sum modulo 2 (or,
equivalently, the Boolean algebra "exclusive
or") of the values of its left- and right-hand

~ nearest neighbor sites on the previous time

step. The rule is implemented simultaneously
at each site.· Even with this very simple rule
• quite complicated behavior is nevertheless

~ found.

I
Fractal Patterns Grown from Cellular Au-
tomata. First of all, consider evolution ac-
Fig. 2. A few time steps in the evolution of the simple cellular automaton difrned by
Eq. 1, starting from a "seed" containing a single nonzero site. Successive lines are

I ·In the very simplest computer implementation a

obtained by successive applications of Eq. 1 at each site. According to this rule, the
separate array of updated site values must be value of each site is the sum modulo 2 of the values of its two nearest neighbors on the
maintained and copied back to the original site previous time step. The pattern obtained with this simple seed is Pascal's triangle of
value array when the updating process is com-
plete. binomial cotdjzeients, reduced modulo 2.

4 Fall 1983 LOS ALAMOS SCIENCE

Cellular Automata

cording to Eq. 1 from a "seed" consisting of

a single site with value 1, all other sites
having value O. The pattern generated by
evolution for a few time steps already
exhibits some structure (Fig. 2). Figure 3
shows the pattern generated after 500 time
steps. Generation of this pattern required
application of Eq. 1 to a quarter of a million
site values. The pattern of Figs. 2 and 3 is an
intricate one but exhibits some striking reg-
ularities. One of these is "self-similarity." As
illustrated in Fig. 3, portions of the pattern,
when magnified, are indistinguishable from
the whole. (Differences on small scales be-
tween the original pattern and the magnified
portion disappear when one considers the
limiting pattern obtained after an infinite
number of time steps.) The pattern is there-
fore invariant under rescaling of lengths.
Such a self-similar pattern is often called a
fractal and may be characterized by a fractal
dimension. The fractal dimension of the
pattern in Fig. 3, for example, is log23 =
log3/ log2 :,; 1.59. Many natural systems,
including snowflakes, appear to exhibit frac-
tal patterns. (See Benoit B. Mandelbrot, The
Fractal Geometry of Nature, W. H. Freeman
and Company, 1982.) It is very possible that
in many cases these fractal patterns are
generated through evolution of cellular
automata or analogous processes.

Self-Organization in Cellular Automata.

Figure 4 shows evolution according to Eq. 1
from a "disordered" initial state. The values
of sites in this initial state are randomly
chosen: each site takes on the value 0 or 1
with equal probability, independently of the
values of other sites. Even though the initial
state has no structure, evolution of the
cellular automaton does manifest some
structure in the form of many triangular
"clearings." The spontaneous appearance of
these clearings is a simple example of "self-
Fig. 3. Many time steps in the evolution of the cellular automaton of Fig. 2, generated organization. "
by applying the rule of Eq. 1 to about a quarter of a million site values. The pattern The pattern of Fig. 4 is strongly reminis-
obtained is "self similar": a part of the pattern, when magnified, is indistinguishable cent of the pattern of pigmentation found on
from the whole. The pattern has afractal dimension of log23 ~ 1.59. the shells of certain mollusks (Fig. 5). It is

LOS ALAMOS SCIENCE Fall 1983 5

 quite possible that the growth of these --.-,' - ;o:'J - ' .. , . .." -~~Z."
~,~-
.~

pigmentation patterns follows cellular au- ~ .6(

~~~ Ii{"~ W. ~ ~!"! ~
~ ::lI
tomaton rules.
In systems that follow conventional
"...c: '" -,.;
L'!i
[>-...
A
~
:i!:j!J ~
!:',
thermodynamics , the second law of ~I!I
~ )::Ii(>} ~ ~~:if:~ I
~ ~
JI: ~
thermodynamics implies a progressive deg- ~.tl ~",
....
~ g
.....
radation of any initial structure and a univer- ~~ ~
~:
~
l."_ ~ ..t< r.i
sal tendency to evolve with time to states of ~~
~~. I:J! ti folt cro;:
maximum entropy and maximum disorder. ~ <;1;
~
~ ~-~ l~
While many natural systems do tend toward ,A~
~.
B
ili ~
disorder, a large class of systems, biological
ones being prime examples, show a reverse
i l".i
M ~
~tIj
~~ If ',:~
io' ~
trend: they spontaneously generate structure ~ ~ ~ . ~~ ~ ~
~
os R!~
with time, even when starting from dis- V'O\:.
\.1. ;"':7:1 ~
~ i.~ 1;1

M~
~
ordered or structureless initial states. The
cellular automaton in Fig. 4 is a simple ., .I;.
~
(I-;: ;'d ~ u: ~ ~ ,,(

(
example of such a self-organizing system.
The mathematical basis of this behavior is , ~ ~
11t(,
;.- i;:!:
'.

, revealed by considering global properties of

the cellular automaton. Instead of following
evolution from a particular initial state, as in
~
~Aj,

~ ~
~l
",".lZI;J ...
l~ .I!lCA.'

Fig. 4. Evolution of the simple cellular automaton difrned by Eq. 1,from a disordered
'IW!t.
~~ZI
....... ~
~

~
Fig. 4, one follows the overall evolution of an initial state in which each site is taken to have value 0 or 1 with equal, independent
ensemble of many different initial states. probabilities. Evolution of the cellular automaton even from such a random initial
p It is convenient when investigating global state yields some simple structure.
• properties to consider finite cellular autom-
ata that contain a finite number N of sites
whose values are subject to periodic bound-

•p ary conditions. Such a finite cellular automa-

ton may be represented as sites arranged, for
example, around a circle. If each site has two

I
•
possible values, as it does for the rule of Eq.
1, there are a total of 2N possible states, or
configurations, for the complete finite cellu-
lar automaton. The global evolution of the

~ cellular automaton may then be represented

by a finite state transition graph plotted in
the "state space" of the cellular automaton.
Each of the 2N possible states of the com-
~
,
plete cellular automaton (such as the state
110 10 110 10 10 for a cellular automaton with

I twelve sites) is represented by a node, or

point, in the graph, and a directed line
connects each node to the node generated by
a single application of the cellular automaton
rule. The trajectory traced out in state space

II by the directed lines connecting one

particular node to its successors thus cor-
responds to the time evolution of the cellular
Fig. 5. A "cone shell" with a pigmentation pattern reminiscent of the pattern generated
by the cellular automaton of Fig. 4. (Shell courtesy of P. Hut.)

6 Fall 1983 LOS ALAMOS SCIENCE

Cellular Automata

Fig. 6. The global state transition graph increases those for others. For example, after
for afinite cellular automaton consisting just one time step the probabilities for states
of twelve sites arranged around a circle on the periphery of the state transition graph
and evolving according to the simple rule in Fig. 6 are reduced to zero; such states
of Eq. 1. Each node in the graph repre- may be given as initial conditions, but may
never be generated through evolution of the
sents one of the 4096 possible states, or
cellular automaton. After many time steps
sequences of the twelve site values, of the
only a small number of all the possible
cellular automaton. Each node is joined configurations actuaJly occur. Those that do
by a directed line to a successor node occur may be considered to lie on "attrac-
60 Copies that corresponds to the state obtained by tors" of the cellular automaton evolution.
one time step of cellular automaton Moreover, if the attractor states have special
evolution. The state transition graph "organized" features, these features will ap-
consists of many disconnected pieces, pear spontaneously in the evolution of the
many of identical structure. Only one cellular automaton. The possibility of self-
copy of each structurally identical piece organization is therefore a consequence of
is shown explicitly. Possible paths the irreversibility of the cellular automaton
evolution, and the structures obtained
through the state transition graph rep-
through self-organization are determined by
resent possible trajectories in the state
the characteristics of the attractors.
space of the cellular automaton. The
The irreversibility of cellular automaton
6 Copies merging of these trajectories reflects the
evolution revealed by Fig. 6 is to be con-
irreversibility of the cellular automaton trasted with the intrinsic reversibility of sys-
evolution. Any initial state of this tems described by conventional thermo-
cellular automaton ultimately evolves to dynamics. At a microscopic level the trajec-
an "attractor" represented in the graph tories representing the evolution of states in
by a cycle. For this particular cellular such systems never merge: each state has a
automaton all collfigurations evolve to unique predecessor, and no information is
attractors in at most three time steps. lost with time. Hence a completely dis-
(From O. Martin, A. Odlyzko, and S. ordered ensemble, in which all possible states
occur with equal probabilities, remains dis-
Wolfram, "A 1gebraic Properties of
ordered forever. Moreover, if nearby states
4 Copies Cellular Automata," Bell Laboratories
are grouped (or "coarse-grained") together,
report (January 1983) and to be pub-
as by imprecise measurements, then with
lished in Communications in Mathemat- time the probabilities for different groups of
ical Physics.) states will tend to equality, regardless of their
initial values. In this way such systems tend
automaton from the initial state represented of the state transition graph). The merging of with time to complete disorder and max-
by that particular node. The state transition trajectories implies that information is lost in imum entropy, as prescribed by the second
graph of Fig. 6 shows all possible trajectories the evolution of the cellular automaton: law of thermodynamics. Tendency to dis-
in state space for a cellular automaton with knowledge of the state attained by the sys- order and increasing entropy are universal
twelve sites evolving according to the simple tem at a particular time is not sufficient to features of intrinsically reversible systems in
rule of Eq. 1. determine its history uniquely, so that the statistical mechanics. Irreversible systems,
A notable feature of Fig. 6 is the presence evolution is irreversible. Starting with an such as the cellular automaton of Figs. 2, 3,
of trajectories that merge with time. While initial ensemble in which all configurations and 4, counter this trend, but universal laws
each state has a unique successor in time, it occur with any distribution of probabilities, have yet to be found for their behavior and
may have several predecessors or no pred- the irreversible evolution decreases the for the structures they may generate. One
ecessors at all (as for states on the periphery probabilities for some configurations and hopes that such general laws may ultimately

LOS ALAMOS SCIENCE Fall 1983 7

•
~

H be abstracted from an investigation of the ing for each configuration a characteristic II N~2(2k+ I ) 2II N~2k+ I
~
= .

, comparatively simple examples provided by

cellular automata.
While there is every evidence that the
polynomial

N- l For odd N, II may be shown to divide

fundamental microscopic laws of physics are A(x) = L aix i ,

I intrinsically reversible (information-preserv-

ing, though not precisely time-reversal in-
variant), many systems behave irreversibly
on a macroscopic scale and are ap-
i~O

where x is a dummy variable, and the

coefficient of Xi is the value of the site at
and in fact is almost always equal to this
value (the first exception occurs for N = 37).
propriately described by irreversible laws. position i. In terms of characteristic poly- Here sordJ2) is a number theoretical func-
For example, while the microscopic molecu- nomials, the cellular automaton rule of Eq. 1 tion defined to be the minimum positive
lar interactions in a fluid are entirely re- takes on the particularly simple form integer j for which 2J = ± I modulo N. The
versible, macroscopic descriptions of the maximum value of sord N (2), typically
average velocity field in the fluid, using, say, achieved ~hen N is prime, is (N-I)/ 2. The
the Navier-Stokes equations, are irreversible maximal cycle length is thus of order 2N12,
and contain dissipative terms. Cellular au- approximately the square root of the total
tomata provide mathematical models at this where number of possible states 2N.
macroscopic level. An unusual feature of this analysis is the
appearance of number theoretical concepts.
T(x) = (x + X- I)
Number theory is inundated with complex
Mathematical Analysis of a Simple results based on very simple premises. It
Cellular Automaton and all arithmetic on the polynomial coeffi- may be part of the mathematical mechanism
cients is performed modulo 2. The reduction by which natural systems of simple construc-
modulo ~-I implements periodic boundary tion yield complex behavior.
The cellular automaton rule of Eq. 1 is conditions. The structure of the state tran-
particularly simple and admits a rather com- sition diagram may then be deduced from
plete mathematical analysis. algebraic properties of the polynomial rex). More General Cellular Automata

,• The fractal patterns of Figs. 2 and 3 may

be characterized in a simple algebraic man-
ner. If no reduction modulo 2 were per-
For even N one finds, for example, that the
fraction of states on attractors is 2- D2 (N),
where DiN) is defined as the largest integral
The discussion so far has concentrated on

I
formed, then the values of sites generated power of 2 that divides N (for example,
the particular cellular automaton rule given
from a single nonzero initial site would Di12) = 4).
by Eq. 1. This rule may be generalized in
simply be the integers appearing in Pascal's Since a finite cellular automaton evolves
• triangle of binomial coefficients. The pattern deterministically with a finite total number of
several ways. One family of rules is obtained

!
by allowing the value of a site to be an
of nonzero sites in Figs. 2 and 3 is therefore possible states, it must ultimately enter a
arbitrary function of the values of the site
::,. the pattern of odd binomial coefficients in cycle in which it visits a sequence of states
itself and of its two nearest neighbors on the
Pascal's triangle. (See Stephen Wolfram, repeatedly. Such cycles are manifest as
previous time step:
"Geometry of Binomial Coefficients," to be closed loops in the state transition graph.
published in American Mathematical The algebraic analysis of Martin et al. shows
Monthly.) that for the cellular automaton of Eq.1 the a (l+ l ) = F(a(I)I ' a(I) , a(I» ) .
I 1- I i+ l
This algebraic approach may be extended maximal cycle length II (of which all other
to determine the structure of the state tran- cycle lengths are divisors) is given for even N
sition diagram of Fig. 6. (See O . Martin, A. by A convenient notation illustrated in Fig. 7,
Odlyzko, and S. Wolfram, "Algebraic assigns a "rule number" to each of the 256
Properties of Cellular Automata," Bell Labo- rules of this type. The rule number of Eq. I is
ratories report (January 1983) and to be 90 in this notation.
published in Communications in Mathemati- Further generalizations allow each site in
cal Physics.) The analysis proceeds by writ- or a cellular automaton to take on an arbitrary

8 Fall 1983 LOS ALAMOS SCIENCE

Cellular Automata

Universality Classes in Cellular

Automata

Rule 100 all 010 001

1 a To proceed in analyzing universality in
cellular automata, one must first give more
quantitative definitions of the classes identi-
fied above. One approach to such definitions
is to consider the degree of predictability of
the outcome of cellular automaton evolution,
given knowledge of the initial state. For class
I cellular automata complete prediction is
Rule trivial: regardless of the initial state, the
Number 0101101~ = 90'0 system always evolves to a unique homoge-
neous state. Class 2 cellular automata have
the feature that the effects of particular site
values propagate only a finite distance, that
Fig. 7. Assignment of rule numbers to cellular automata for which k = 2 and is, only to a finite number of neighboring
sites. Thus a change in the value of a single
r = 1. The values of sites obtained from each of the eight possible three-site
initial site affects only a finite region of sites
neighborhoods are combined to form a binary number that is quoted as a decimal
around it, even after an infinite number of
integer. The example shown is for the rule given by Eq. 1.
time steps. This behavior, illustrated in Fig.
9, implies that prediction of a particular final
site value requires knowledge of only a finite
number k of values and allow the value of a rated and simple; set of initial site values. In contrast, changes
site to depend on the values of sites at a o Class 3-evolution leads to a chaotic of initial site values in class 3 cellular autom-
distance up to r on both sides, so that pattern; ata, again as illustrated in Fig. 9, almost
o Class 4-evolution leads to complex always propagate at a finite speed forever
structures, sometimes long-lived. and therefore affect more and more distant
a(t+ l) = F(a(l) , ... ,a(I) .
I l- r Hr
sites as time goes on. The value of a
Examples of these classes are indicated in particular site after many time steps thus
The number of different rules with given k Fig. 8. depends on an ever-increasing number of
and r grows as k k2r+l and therefore becomes The existence of only four qualitative initial site values. If the initial state is dis-
immense even for rather small k and r. classes implies considerable universality in ordered, this dependence may lead to an
Figure 8 shows examples of evolution the behavior of cellular automata; many apparently chaotic succession of values for a
according to some typical rules with various features of cellular automata depend only on particular site. In class 3 cellular automata,
k and r values. Each rule leads to patterns the class in which they lie and not on the therefore, prediction of the value of a site at
that differ in detail. However, the examples precise details of their evolution. Such uni- infinite time would require knowledge of an
suggest a very remarkable result: all patterns versality is analogous, though probably not infinite numbt;r of initial site values. Class 4
appear to fall into only four qualitative mathematically related, to the universality cellular automata are distinguished by an
classes. These basic classes of behavior may found in the equilibrium statistical mechanics even greater degree of unpredictability, as
be characterized empirically as follows: of critical phenomena. In that case many discussed below.
systems with quite different detailed con- Class 2 cellular automata may be con-
o Class I-evolution leads to a homogene- struction are found to lie in classes with sidered as "filters" that select particular
ous state in which, for example, all sites have critical exponents that depend only on gen- features of the initial state. For example, a
value 0; eral, primarily geometrical features of the class 2 cellular automata may be constructed
o Class 2-evolution leads to a set of systems and not on their detailed construc- in which initial sequences III survive, but
stable or periodic structures that are sepa- tion. sites not in such sequences eventually attain

LOS ALAMOS SCIENCE Fall 1983 9

(
,
1

,•
I•
!

Fig. 8. Evolution of some typical cellular automata from on the values of sites up to r sites distant on both sides.
disordered initial states. Each group of six patterns shows the Different colors represent different site values: black cor-
evolution of various rules with particular values of k and r. responds to a value of 0, red to 1, green to 2, blue to 3, and
Sites take on k possible values, and the value of a site depends yellow to 4. Thefact that these and other examples exhibit only

10 Fall 1983 LOS ALAMOS SCIENCE

four qualitative classes of behavior (see text) suggests consider- examples on page 10 for which r = 2 evolve according to rules
able universality in the behavior of cellular automata. The in which the value of a site depends only on the sum of the
examples on page 10 for which r = 1 are labeled by rule values of the 2r + 1 sites in its neighborhood on the previous
number (in the notation of Fig. 7) and behavior class. The time step. Such rules may be specified by numerical codes C

LOS ALAMOS SCIENCE Fall 1983 II

,•
I
•
~
\

I such that the co~cient of 2j in the binary decomposition of C and J. Condon of Bell Laboratoriesfor their help in preparing
gives the value attained by a site if its neighborhood had total these and other color pictures of cellular automata.)
value j on the previous time step. These examples are labeled
by code number and behavior class. (/ am gratiful to R. Pike

12 Fall 1983 LOS ALAMOS SCIENCE

Fig. 9. Difference patterns showing the differences between the rule:for class 2 rules the tifJects havefinite range,'for class
corifigurations generated by evolution, according to various 3 rules the tifJects propagate to neighboring sites indifznitely at
cellular automaton rules, from initial states that differ in the a fIXed speed; and for class 4 rules the tifJects also propagate
value of a single site. Each difference pattern is labeled by the to neighboring sites indifznitely but at various speeds. The
behavior class of the cellular automaton rule. The effects of difference patterns shown here are analogues of Green 's
changes in a single site value depend on the behavior class of functions for cellular automata.

value O. Such cellular automata are of prac- random. In fact, as mentioned for the exam- but the corresponding configurations are
tical importance for digital image processing : ple of Eq. 1, they may exhibit important self- not). Instead, the configurations of an infinite
they may be used to select and enhance organizing behavior. In addition and again in cellular automaton form a Cantor set. Figure
particular patterns of pixels. After a suffi- contrast to class 2 cellular automata, the 10 illustrates two constructions for a Cantor
ciently long time any class 2 cellular automa- statistical properties of the states generated set. In construction (a) of Fig. 10, one starts
ton evolves to a state consisting of blocks by many time steps of class 3 cellular with the set of real numbers in the interval 0
containing nonzero sites separated by re- automaton evolution are the same for almost to I. First one excludes the middle third of
gions of zero sites. The blocks have a simple all possible initial states. The large-time the interval, then the middle third of each
form, typically consisting of repetitions of behavior of a class 3 cellular automaton is interval remaining, and so on. In the limit the
particular site values or sequences of site therefore determined by these common set consists of an infinite number of discon-
values (such as 101010 ... ). The blocks statistical properties. nected points. If positions in the interval are
either do not change with time (yielding The configurations of an infinite cellular represented by ternimals (base 3 fractions,
vertical stripes in the patterns of Fig. 8) or automaton consist of an infinite sequence of analogous to base 10 decimals), then the
cycle between a few states (yielding "railroad site values. These site values could be con- construction is seen to retain only points
track" patterns). sidered as digits in a real number, so that whose positions are represented by ternimals
While class 2 cellular automata evolve to each complete configuration would cor- containing no l 's (the point 0.2202022 is
give persistent structures with small periods, respond to a single real number. The topol- therefore included; 0.2201022 is excluded).
class 3 cellular automata exhibit chaotic ogy of the real numbers is, however, not An important feature of the limiting set is its
aperiodic behavior, as shown in Fig. 8. exactly the same as the natural one for the self-similarity, or fractal form : a piece of the
Although chaotic, the patterns generated by configurations (the binary numbers set, when magnified, is indistinguishable
class 3 cellular automata are not completely 0.111111 . .. and 1.00000... are identical, from the whole. This self-similarity is math-

LOS ALAMOS SCIENCE Fall 1983 13

 o 2
ematically analogous to that found for the o
o 2 2
limiting two-dimensional pattern of Fig. 3.
In construction (b) of Fig. 10, the Cantor o 2 o 2 o 2 o 2
set is formed from the "leaves" of an infinite
binary tree. Each point in the set is reached 02 02 02 02 02 02 02 02
by a unique path from the "root" (top as
drawn) of the tree. This path is specified by
an infinite sequence of binary digits, in which
successive digits determine whether the left-
(a)
or right-hand branch is taken at each suc-
cessive level in the tree. Each point in the
Cantor set corresponds uniquely to one
infinite sequence of digits and thus to one
configuration of an infinite cellular automa-
ton. Evolution of the cellular automaton then
corresponds to iterated mappings of the
Cantor set to itself. (The locality of cellular
automaton rules implies that the mappings
are continuous.) This interpretation of cellu-
lar automata leads to analogies with the
theory of iterated mappings of intervals of
the real line. (See Mitchell J. Feigenbaum,
"Universal Behavior in Nonlinear Systems,"
Los Alamos Science , Vol. I, No. 1(1980):
4-27.)
Cantor sets are parameterized by their
"dimensions." A convenient definition of
dimension, based on construction (a) of Fig.
10, is as follows. Divide the interval from 0
to 1 into k n bins, each of width k - n. Then let
N(n) be the number of these bins that
contain points in the set. For large n this
number behaves according to
(b)
N(n) ~ 0 n , (2)

and d is defined as the "set dimension" of the

Cantor set. If a set contained all points in the Fig. 10. Steps in two constructions of a Cantor set. At each step in construction (a),
interval 0 to 1, then with this definition its
the middle third of all intervals is excluded. The first step thus excludes all points
dimension would simply be 1. Similarly, any
whose positions, when expressed as base 3 fractions, have a 1 in the first "ternimal
finite number of segments of the real line
would form a set with dimension I. How-
place" (by analogy with decimal place), the second step excludes all points whose
ever, the Cantor set of construction (a), positions have a 1 in the second ternimal place, and so on. The limiting set obtained
which contains an infinite number of discon- qfter an infinite number of steps consists of an infinite number of disconnected points
nected pieces, has a dimension according to whose positions contain no 1 's. The set may be assigned a dimension, according to Eq.
Eq. 2 of log32 ~ 0.63. 2, that equals log; ~ 0.63. Construction (b) rtiflects the topological structure of the
An alternative definition of dimension, Cantor set. Infinite sequences of digits, representing cellular automaton configura-
agreeing with the previous one for present tions, are seen to correspond uniquely with points in the Cantor set.

14 Fall 1983 LOS ALAMOS SCIENCE

Cellular Automata

RULE BBB1BB1B (18)

equation Z3 - Z2 + 2z - 1 = O. (See D. A.
Lind, "Applications of Ergodic Theory and
2
Sofic Systems to Cellular Automata," Uni-
3
versity of Washington preprint (April 1983)
4
and to be published in Physica D; see also
5
Martin et ai., op. cit.) The greater the
6
? ;
irreversibility in the cellular automaton evo-
lution, the smaller is the dimension of the
8
9
Cantor set corresponding to the attractors
for the evolution. If the set of attractors for a
cellular automaton has dimension 1, then
essentially all the configurations of the
cellular automaton may occur at large times.
Fig. 11. Time evolution of the probabilities for each of the 1024 possible configura- If the attractor set has dimension less than 1,
tions of a typical class 3 cellular automaton with k = 2 and r = 1 and of size 10, then a vanishingly small fraction of all
possible configurations are generated after
starting from an initial ensemble in which each possible configuration occurs with
many time steps of evolution. The attractor
equal probability. The configurations are specified by integers whose binary digits
sets for most class 3 cellular automata have
form the sequence of site values. The probability for a particular configuration is given
dimensions less than 1. For those class 3
on successive lines in a vertical column: a dot appears at a particular time step if the cellular automata that generate regular pat-
configuration occurs with nonzero probability at that time step. In the initial ensemble terns, the more regular the pattern, the
all configurations occur with equal nonzero probabilities, and dots appear in all smaller is the dimension of the attractor set;
positions. The cellular automaton evolution modifies the probabilities for the these cellular automata are more irreversible
configurations, making some occur with zero probability and yielding gaps in which and are therefore capable of a higher degree
no dots appear. This "thinning" is a consequence of the irreversibility of the cellular of self-organization.
automaton evolution and is reflected in a decrease of entropy with time. In the limit of The dimension of a set of cellular automa-
cellular automata of infinite size, the configurations appearing at large times form a ton configurations is directly proportional to
Cantor set. For the rule shown (rule 18 in the notation of Fig. 7) the limiting the limiting entropy (or information) per site
of the sequence of site values that make up
dimension of this Cantor set isfound to be approximately 0.88.
the configurations. (See Patrick Billingsley,
Ergodic Theory and Information, John
purposes, is based on self-similarity. Take starting from such a disordered initial Wiley & Sons, 1965.) If the dimension of the
the Cantor set of construction (a) in Fig. 10. ensemble. As expected from the irre- set was 1, so that all possible sequences of
Contract the set by a magnification factor versibility of cellular automaton evolution, site values could occur, then the entropy of
k - m• By virtue of its self-similarity, the whole exemplified by the state transition graph of these sequences would be maximal. Di-
set is identical to a number, say M(m), of Fig. 6, different configurations attain dif- mensions lower than 1 correspond to sets in
copies of this contracted copy. For large m, ferent probabilities as evolution proceeds, which some sequences of site values are
M(m) ::::: kdm, where again d is defined as the and the probabilities for some configurations absent, so that the entropy is reduced. Thus
set dimension. decrease to zero. This phenomenon is mani- the dimension of the attractor for a cellular
With these definitions the dimension of the fest in the "thinning" of configurations on automaton is directly related to the limiting
Cantor set of all possible configurations for successive time steps apparent in Fig. 11. entropy attained in its evolution, starting
an infinite one-dimensional cellular automa- The set of configurations that survive with from a disordered ensemble of initial states.
ton is I. A disordered ensemble, in which nonzero probabilities after many time steps Dimension gives only a very coarse
each possible configuration occurs with of cellular automaton evolution constitutes measure of the structure of the set of con-
equal probability, thus has dimension 1. the "attractors" for the evolution. This set is figurations reached at large times in a
Figure II shows the behavior of the again a Cantor set; for the example of Fig. cellular automaton. Formal language theory
probabilities for the configurations of a typi- 11 its dimension is log2K ~ 0.88, where K ~ may provide a more complete characteriza-
cal cellular automaton as a function of time, 1. 755 is the real solution of the polynomial tion of the set. "Languages" consist of a set

LOS ALAMOS SCIENCE Fall 1983 15

 of words, typically infinite in number,
formed from a sequence of letters according
to certain grammatical rules. Cellular
automaton configurations are analogous to
words in a formal language whose letters are
the k possible values of each cellular automa-
ton site. A grammar then gives a succinct
specification for a set of cellular automaton
configurations.
Languages may be classified according to
the complexity of the machines or computers
necessary to generate them. A simple class
of languages specified by "regular gram-
mars" may be generated by finite state
machines. A finite state machine is repre-
sented by a state transition graph (analogous
to the state transition graph for a finite
cellular automaton illustrated in Fig. 6). The
possible words in a regular grammar are
generated by traversing all possible paths in
the state transition graph. These words may
be specified by "regular expressions" consist-
ing of finite length sequences and arbitrary
repetitions of these. For example, the regular
expression 1(00)*1 represents all sequences
containing an even number of O's (arbitrary
repetition of the sequence 00) flanked by a
pair of I's. The set of configurations ob-
tained at large times in class 2 cellular
automata is found to form a regular lan-
guage. It is likely that attractors for other
classes of cellular automata correspond to
more complicated languages.

Analogy with Dynamical

Systems Theory

The three classes of cellular automaton

behavior discussed so far are analogous to
three classes of behavior found in the solu-
tions to differential equations (continuous
dynamical systems). For some differential
equations the solutions obtained with any
initial conditions approach a fixed point at
large times. This behavior is analogous to
class I cellular automaton behavior. In a
second class of differential equations, the
limiting solution at large times is a cycle in
which the parameters vary periodically with
time. These equations are analogous to class
2 cellular automata. Finally, some differen-
tial equations have been found to exhibit
complicated, apparently chaotic behavior de-
Fig. 12. Evolution of a class 4 cellular automaton from several disordered initial
pending in detail on their initial conditions.
states. The bottom example has been reproduced on a larger scale to show detail. In
With the initial conditions specified by deci-
this cellular automaton,for which k = 2 and r = 2, the value of a site is 1 only if a total
mals, the solutions to these differential equa-
tions depend on progressively higher and of two or four sites out of the five in its neighborhood have the value 1 on the previous
higher order digits in the initial conditions. time step. For some initial states persistent structures are formed, some of which
This phenomenon is analogous to the de- propagate with time. This cellular automaton is believed to support universal
pendence of a particular site value on pro- computation, so that with suitable initial states it may implement any finite algorithm.

16 Fall 1983 LOS ALAMOS SCIENCE

L
Cellular Automata

'\
.~ ~ ~
~

... ..
. [;.. ' .......
:~
.~ I

Fig. 13. Persistent structures exhibited by the class 4 cellular structures are almost sufficient to demonstrate a universal
automaton of Fig. 12 as it evolves from initial states with computation capability for the cellular automaton.
nonzero sites in a region of twenty or fewer sites. These

gressively more distant initial site values in simple patterns. In addition, the propagating configurations must encode all possible pro-
the evolution of a class 3 cellular automaton. structures allow site values at one position to grams.
The solutions to this final class of differential affect arbitrarily distant sites after a suffi- The only known method of proving that a
equations tend to "strange" or "chaotic" ciently long time. No analogous behavior system may act as a universal computer is to
attractors (see Robert Shaw, "Strange At- has yet been found in a continuous show that its computational capabilities are
tractors, Chaotic Behavior, and Information dynamical system. equivalent to those of another system al-
Flow," Zeitschrift fur Naturforschung The complexity apparent in the behavior ready classified as a universal computer. The
36A(1981):80), which form Cantor sets in of class 4 cellular automata suggests the Church-Turing thesis states that no system
direct analogy with those found in class 3 conjecture that these systems may be may have computational capabilities greater
cellular automata. The correspondence be- capable of universal computation. A com- than those of universal computers. The thesis
tween classes of behavior found in cellular puter may be regarded as a system in which is supported by the proven equivalence of
automata and those found in continuous definite rules are used to transform an initial computational models such as Turing ma-
dynamical systems supports the generality of sequence of, say, l's and O's to a final chines, string-manipulation systems, ideal-
these classes. Moreover, the greater mathe- sequence of 1's and O's. The initial sequence ized neural networks, digital computers, and
matical simplicity of cellular automata sug- may be considered as a program and data cellular automata. While mathematical sys-
gests that investigation of their behavior may stored in computer memory, and part of the tems with computational power beyond that
elucidate the behavior of continuous final sequence may be considered as the of universal computers may be imagined, it
dynamical systems. result of the computation. Cellular automata seems likely that no such systems could be
may be considered as computers; their initial built with physical components. This conjec-
A Universal Computation Class configurations represent programs and initial ture could in principle be proved by showing
of Cellular Automata data, and their configurations after a long that all physical systems could be simulated
time contain the results of computations. by a universal computer. The main obstruc-
Figure 12 shows patterns obtained by A system is a universal computer if, given tion to such a proof involves quantum me-
evolution from disordered initial states ac- a suitable initial program, its time evolution chanics.
cording to a class 4 cellular automaton rule. can implement any finite algorithm. (See A cellular automaton may be proved
Complicated behavior is evident. In most Frank S. Beckman, Mathematical Founda- capable of universal computation by identify-
cases all sites eventually "die" (attain value tions of Programming, Addison-Wesley Pub- ing structures that act as the essential com-
0). In some cases, however, persistent struc- lishing Co., 1980.) A universal computer ponents of digital computers, such as wires,
tures that survive for an infinite time are need thus only be "reprogrammed," not NAND gates, memories, and clocks. The
generated, and a few of these persistent "rebuilt," to perform each possible calcula- persistent structures illustrated in Fig. 13
structures propagate with time. Figure 13 tion. (All modern general-purpose electronic provide many of the necessary components,
shows all the persistent structures generated digital computers are, for practical purposes, strongly suggesting that the cellular automa-
from initial states with nonzero sites in a universal computers; mechanical adding ma- ton of Figs. 12 and 13 is a universal
region of twenty or fewer sites. Unlike the chines were not.) If a cellular automaton is to computer. One important missing compo-
periodic structures of class 2 cellular au- be a universal computer, then, with a fixed nent is a "clock" that generates an infinite
tomata, these persistent structures have no rule for its time evolution, different initial sequence of "pulses"; starting from an initial

LOS ALAMOS SCIENCE Fall 1983 17

 configuration containing a finite number of universal computers, then their behavior structure should occur. After a sufficiently
nonzero sites, such a structure would give may be considered completely unpredictable. long time this configuration may reproduce
rise to an ever-increasing number of nonzero For class 3 cellular automata the values of many times, so that it ultimately dominates
sites. If such a structure exists, it can un- particular sites after a long time depend on the behavior of the cellular automaton. Even
doubtedly be found by careful investigation, an ever-increasing number of initial sites. For though the a priori probabililty for the
although it is probably too large to be found class 4 cellular automata this dependence is occurrence of a self-reproducing structure in
by any practical exhaustive search. If the by an algorithm of arbitrary complexity, and the initial state is very small, its a posteriori
cellular automaton of Figs. 12 and 13 is the values of the sites can essentially be probability after many time steps of cellular
indeed capable of universal computation, found only by explicit observation of the automaton evolution may be very large. The
then, despite its very simple construction, it cellular automaton evolution. The apparent possibility that arbitrarily complex behavior
is in some sense capable of arbitrarily com- unpredictability of class 4 cellular automata seeded by features of the initial state can
plicated behavior. introduces a new level of uncertainty into the occur in class 4 cellular automata with
Several complicated cellular automata behavior of natural systems. indefinitely low probability prevents the tak-
have been proved capable of universal com- The unpredictability of universal com- ing of meaningful statistical averages over
putation. A one-dimensional cellular autom- puter behavior implies that propositions con- infinite volume (length). It also suggests that
aton with eighteen possible values at each cerning the limiting behavior of universal in some sense any class 4 cellular automaton
site (and nearest neighbor interactions) has computers at indefinitely large times are with an infinite disordered initial state is a
been shown equivalent to the simplest known formally undecidable. For example, it is microcosm of the universe.
universal Turing machine. In two dimensions undecidable whether a particular universal In extensive samples of cellular automaton
several cellular automata with just two states computer, given particular input data, will rules, it is found that as k and r increase,
per site and interactions between nearest reach a special "halt" state after a finite time class 3 behavior becomes progressively more
neighbor sites (including diagonally adjacent or will continue its computation forever. dominant. Class 4 behavior occurs only for k
sites, giving a nine-site neighborhood) are Explicit simulations can be run only for finite > 2 or r > 1; it becomes more common for
known to be equivalent to universal digital times and thus cannot determine such infinite larger k and r but remains at the few percent
computers. The best known of these cellular time behavior. Results may be obtained for level. The fact that class 4 cellular automata
automata is the "Game of Life" invented by some special input data, but no general exist with only three values per site and
Conway in the early 1970s and simulated (finite) algorithm or procedure may even in nearest neighbor interactions implies that the
extensively ever since. (See Elwyn R. principle be given. If class 4 cellular autom- threshold in complexity of construction
Berlekamp, John H. Conway, and Richard ata are indeed universal computers, then it is necessary to allow arbitrarily complex
K. Guy, Winning Ways, Academic Press, undecidable (in general) whether a particular behavior is very low. However, even among
1982 and Martin Gardner, Wheels, Life, and initial state will ultimately evolve to the null systems of more complex construction, only
Other Mathematical Amusements, W. H. configuration (in which all sites have value 0) a small fraction appear capable of arbitrarily
Freeman and Company, October 1983. or will generate persistent structures. As is complex behavior. This suggests that some
The Life rule takes a site to have value 1 if typical for such generally undecidable physical systems may be characterized by a
three and only three of its eight neighbors are propositions, particular cases may be de- capability for class 4 behavior and universal
I or if four are I and the site itself was 1 on cided. In fact, the halting of the cellular computation; it is the evolution of such
the previous time step.) Structures analogous automaton of Figs. 12 and 13 for all initial systems that may be responsible for very
to those of Fig. 13 have been identified in the states with nonzero sites in a region of complex structures found in nature.
Game of Life. In addition, a clock structure, twenty sites has been determined by explicit The possibility for universal computation
dubbed the glider gun, was found after a long simulation. In general, the halting prob- in cellular automata implies that arbitrary
search. ability, or fraction of initial configurations computations may in principle be performed
By definition, any universal computer may ultimately evolving to the null configuration, by cellular automata. This suggests that
in principle be simulated by any other uni- is a noncomputable number. However, the cellular automata could be used as practical
versal computer. The simulation proceeds by explicit results for small initial patterns sug- parallel-processing computers. The mech-
emulating the elementary operations in the gest that for the cellular automaton of Figs. anisms for information processing found in
first universal computer by sets of operations 12 and 13, this halting probability is approx- most natural systems (with the exception of
in the second universal computer, as in an imately 0.93. those, for example, in molecular genetics)
"interpreter" program. The simulation is in In an infinite disordered configuration all appear closer to those of cellular automata
general only faster or slower by a fixed finite possible sequences of site values appear at than to those of Turing machines or conven-
factor, independent of the size or duration of some point, albeit perhaps with very small tional serial-processing digital computers.
a computation. Thus the behavior of a uni- probability. Each of these sequences may be Thus one may suppose that many natural
versal computer given particular input may considered to represent a I?ossible "pro- systems could be simulated more efficiently
be determined only in a time of the same gram"; thus with an infinite disordered initial by cellular automata than by conventional
order as the time required to run that state, a class 4 automaton may be con- computers. In practical terms the
universal computer explicitly. In general the sidered to execute (in parallel) all possible homogeneity of cellular automata leads to
a
behavior of universal computer cannot be programs. Programs that generate structures simple implementation by integrated circuits.
predicted and can be determined only by a of arbitrarily great complexity occur, at least A simple one-dimensional universal cellular
procedure equivalent to observing the univer- with indefinitely small probabilities. Thus, automaton with perhaps a million sites and a
sal computer itself. for example, somewhere on the infinite line a time step as short as a billionth of a second
If class 4 cellular automata are indeed sequence that evolves to a self-reproducing could perhaps be fabricated with current

18 Fall 1983 LOS ALAMOS SCIENCE

Cellular Automata

Fig. 14. Simulation networkfor symmetric cellular automaton Simulations are included in the network shown only when the
rules with k = 2 and r = 1. Each rule is specified by the number necessary blocks are three or fewer sites long. Rules 90 and
obtained as shown in Fig. 7, and its behavior class is indicated 150 are additive class 3 rules, rule 204 is the identity rule, and
by shades of gray: light gray corresponds to class 1, medium rules 170 and 240 are ltift- and right-shift rules, respectively.
gray to class 2, and dark gray to class 3. Rule A is considered Attractive simulation paths are indicated by bold lines.
to simulate rule B if there exist blocks of site values that evolve (Network courtesy of J. Milnor.)
under rule A as single sites would evolve under rule B.

technology on a single silicon wafer (the one- cellular automata, should be considered first. empirical result. Techniques from computa-
dimensional homogeneous structure makes tion theory may provide a basis, and ulti-
defects easy to map out). Conventional pro- mately a proof, of this result.
gramming methodology is, of course, of little A Basis for Universality? The first crucial observation is that with
utility for such a system. The development of special initial states one cellular automaton
a new methodology is a difficult but impor- may behave just like another. In this way
tant challenge. Perhaps tasks such as image The existence of four classes of cellular one cellular automaton may be considered to
processing, which are directly suitable for automata was presented above as a largely "simulate" another. A single site with a

LOS ALAMOS SCIENCE Fall 1983 19

particular value in one cellular automaton
may be simulated by a fixed block of sites in
another ; after a fixed number of time steps,
the evolution of these blocks imitates the
single time-step evolution of sites in the first
cellular automaton. For example, sites with
value 0 and 1 in the first cellular automaton
may be simulated by blocks of sites 00 and
11 , respectively, in the second cellular
automaton, and two time steps of evolution
in the second cellular automaton correspond
Fig. 15. Evolution of the class 3 cellular automaton rule 18 from a disordered initial
to one time step in the first. Then, with a
state with pairs of sites combined. The pair of site values ()() is shown as black, 01 as
special initial state containing 11 and 00 but red, 10 as green, and 11 as blue. At large times two phases are clearly evident,
not Oland 10 blocks, the second cellular separated by "difects" that execute approximately random walks and ultimately
automaton may simulate the first. annihilate in pairs. In each phase alternate sites have value 0, and the other sites
Figure 14 gives the network that repre- evolve according to the additive rule 90. Thus for almost all initial states rule 18
sents the simulation capabilities of sym- behaves like rule 90 at large times. Rule 18 therifore follows an attractive simulation
metric cellular automata with k = 2 and r = path to rule 90.
1. (Only simulations involving blocks of
length less than four sites were included in
the construction of the network.) If a cellular
automaton is computationally universal,
then with a sufficiently long encoding it
should be able to simulate any other cellular
a utomaton, so that a path should exist from
the node that represents its rule to nodes
representing all other possible rules.
An example of the simulation of one
cellular automaton by another is the simula-
tion of the additive rule 90 (Eq. 1) by the
class 3 rule 18. A rule 18 cellular automaton
behaves exactly like a rule 90 cellular
automaton if alternate sites in the initial
configuration have value 0 (so that 0 and 1
in rule 90 are represented by 00 and 01 in
rule 18) and alternate time steps are con-
sidered. Figure 15 shows evolution accord-
ing to rule 18 from a disordered initial state.
Two " phases" are clearly evident: one in
which sites at even-numbered positions have
value 0 and one in which sites at odd-
Fig. 16. Evolution of the class 2 cellular automaton rule 94 from an initial state in
numbered positions have value O. The
boundaries between these regions execute
which the members of most pairs of sites have the same values, so that the digrams 00
approximately random walks and eventually and 11 predominate and the sequences 010 and 101 are nearly absent. (Color
annihilate in pairs, leaving a system consist- designations are the same as in Fig. 15.) Class 3 behavior occurs, but is unstable;
ing of blocks of sites that evolve according to class 2 behavior is "seeded" by 10 and 01 digrams and ultimately dominates. Rule 94
the additive rule 90. (el P. Grassberger, exhibits a repulsive simulation path to the class 3 additive rule 90 and an attractive
"Chaos and Diffusion in Deterministic path to the identity rule 204.

20 Fall 1983 LOS ALAMOS SCIENCE

Cellular Automata

Cellular Automata," to be published in

Physica D.) Thus the simulation of rule 90
by rule 18 may be considered an "attractive"
one: starting from almost all initial states,
rule 18 evolves toward states in which it
simulates rule 90. In general, one expects
that some paths in the network of Fig. 14 are
attractive, while the rest are repulsive. The
consequences of a repulsive simulation path
are illustrated in Fig. 16: with special initial
states rule 94 behaves like rule 90, but any
impurities in the initial states grow and
eventually dominate the evolution of the
system.
Class I cellular automata have an attract-
ive simulation path to rule 0 (or its equiv-
alents). Class 2 cellular automata have at-
tractive simulation paths to the identity rule
204. A conjecture for which some evidence
exists is that all class 3 rules exhibit attrac-
tive simulations has to additive rules such
Stephen Wolfram was born in London, England in 1959. He was educated at Eton College, Oxford
as 90 or 150. Simulation by blocking of site University, and the California Institute of Technology where he received his Ph.D. in theoretical physics
values is analogous to a block spin or in 1979. He was a member of the faculty at Caltech from 1980 until he resigned in 1982. At that time he
renormalization group transformation; addi- joined the Institute for Advanced Study in Princeton, New Jersey. He has worked in various areas of
theoretical physics, including high-energy physics, cosmology, and statistical mechanics and has also
tive rules have the special property that they worked in computer science, particularly in the area of symbolic computation. He received a MacArthur
are invariant under such transformations. As Fellowship in 1981 and since 1982 has been a Visiting Staff Member of the Theoretical Division at Los
mentioned earlier, class 4 cellular automata Alamos.
are distinguished by the presence of simula-
tion paths leading to every other cellular
Acknowledgments
automaton rule. It is likely that no specific
path is distinguished as attractive. I am grateful to many people for discussions and suggestions. I thank in particular my collaborators in
various cellular automaton investigations: O. Martin, J. Milnor, and A. Odlyzko. The research described
Cellular automata of different classes may here was supported in part by the Office of Naval Research under contract number NOOOI4-80-C-0657.
thus be distinguished by their limiting
behavior under simulation transformations.
This approach suggests that classification of Further Reading
the qualitative behavior of cellular automata John von Neumann . Edited and completed by Arthur W. Burks. Theory of S elf-Reproducing Automata.
may be related to determinations of equiv- Urbana : University of Illinois Press, 1966.
alence of systems and problem classes in
computation theory. In general, one may Arthur W. Burks, editor. Essay s on Cellular A utomata. Urbana: University of Illinois Press, 1970.
hope for fundamental connections between
computation theory and the theory of com- Stephen Wolfram. " Statistical Mechanics of Cellular Automata." Reviews of Modern Phy sics
55(1983):601.
plex nonequilibrium statistical systems. In-
formation theory forms a mathematical basis
Stephen Wolfram. " Universality and Complexity in Cellular Automata." The Institute for Advanced
for equilibrium statistical mechanics. Com-
Study preprint (May 1983) and to be published in Phy sica D.
putation theory, which addresses time-de-
pendent processes, may be expected to play Stephen Wolfram, J. Doyne Farmer, and Tommaso Toffoli, editors. "Cellular Automata: Proceedings of
a fundamental role in nonequilibrium statis- an Interdisciplinary Workshop (Los Alamos; March 7- 11, 1983)." To be published in Physica D and to
tical mechanics. • be available separately from North-Holland Publishing Company.

LOS ALAMOS SCIENCE Fall 1983 21