Professional Documents
Culture Documents
Cellular - Automata 1 100
Cellular - Automata 1 100
Cellular Automata
A Volume in the Encyclopedia of
Complexity and Systems Science,
Second Edition
Encyclopedia of Complexity and
Systems Science Series
Editor-in-Chief
Robert A. Meyers
The Encyclopedia of Complexity and Systems Science Series of topical
volumes provides an authoritative source for understanding and applying the
concepts of complexity theory together with the tools and measures for
analyzing complex systems in all fields of science and engineering. Many
phenomena at all scales in science and engineering have the characteristics of
complex systems, and can be fully understood only through the transdisciplin-
ary perspectives, theories, and tools of self-organization, synergetics, dynam-
ical systems, turbulence, catastrophes, instabilities, nonlinearity, stochastic
processes, chaos, neural networks, cellular automata, adaptive systems,
genetic algorithms, and so on. Examples of near-term problems and major
unknowns that can be approached through complexity and systems science
include: the structure, history, and future of the universe; the biological basis of
consciousness; the integration of genomics, proteomics, and bioinformatics as
systems biology; human longevity limits; the limits of computing; sustainabil-
ity of human societies and life on earth; predictability, dynamics, and extent of
earthquakes, hurricanes, tsunamis, and other natural disasters; the dynamics of
turbulent flows; lasers or fluids in physics; microprocessor design; macromo-
lecular assembly in chemistry and biophysics; brain functions in cognitive
neuroscience; climate change; ecosystem management; traffic management;
and business cycles. All these seemingly diverse kinds of phenomena and
structure formation have a number of important features and underlying
structures in common. These deep structural similarities can be exploited to
transfer analytical methods and understanding from one field to another. This
unique work will extend the influence of complexity and system science to a
much wider audience than has been possible to date.
Cellular Automata
A Volume in the Encyclopedia of
Complexity and Systems Science,
Second Edition
This Springer imprint is published by the registered company Springer Science+Business Media,
LLC part of Springer Nature.
The registered company address is: 233 Spring Street, New York, NY 10013, U.S.A.
Series Preface
v
vi Series Preface
Each entry in each of the Series books was selected and peer reviews
organized by one of our university-based book Editors with advice and
Series Preface vii
1
Ulam S. M. A. Collection of Mathematical Problems (New York: Interscience, 1960), p. 30
ix
Cellular Automata Editorial
xi
xii Cellular Automata Editorial
xv
xvi Contents
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 749
About the Editor-in-Chief
Biography
Dr. Meyers has worked with more than 20 Nobel laureates during his career
and is the originator and serves as Editor-in-Chief of both the Springer Nature
Encyclopedia of Sustainability Science and Technology and the related and
supportive Springer Nature Encyclopedia of Complexity and Systems Science.
Education
Dr. Meyers holds more than 20 patents and is the author or Editor-in-Chief
of 12 technical books including the Handbook of Chemical Production
xix
xx About the Editor-in-Chief
xxi
xxii About the Volume Editor
xxiii
xxiv Contributors
# Springer-Verlag 2009 1
A. Adamatzky (ed.), Cellular Automata,
https://doi.org/10.1007/978-1-4939-8700-9_58
Originally published in
R. A. Meyers (ed.), Encyclopedia of Complexity and Systems Science, # Springer-Verlag 2009
https://doi.org/10.1007/978-0-387-30440-3_58
2 Cellular Automata in Triangular, Pentagonal, and Hexagonal Tessellations
Cellular Automata in Triangular, Pentagonal, and with exactly 2 or 3 live neighbors remain alive (otherwise
Hexagonal Tessellations, Fig. 1 Top: Each cell in a they die); dead cells with exactly 3 live neighbors come to
grid has 8 “neighbors”. The cells containing “n” are neigh- life (otherwise they remain dead)”. Let us now evaluate the
bors of the cell containing the “X”. Any cell in the grid can transition from generation 1 to generation 2. In our dia-
be either “dead” or “alive”. Bottom: Here we have outlined gram, cell “a” is dead. Since it does not have exactly 3 live
a specific area of what is presumably a much larger grid. At neighbors, it remains dead. Cell “b” is alive, but it needs
the left we have installed an initial shape. Shaded cells are exactly 2 or 3 live neighbors to remain alive; since it only
alive; all others are dead. The number within each cell has 1, it dies. Cell “c” is dead; since it has exactly 3 live
gives the quantity of live neighbors for that cell. (Cells neighbors, it comes to life. And cell “d” has 2 live neigh-
containing no numbers have zero live neighbors.) Depicted bors; hence it will remain alive. And so on. Notice that the
are three generations, starting with the configuration at form repeats every two generations. Such forms are called
generation 1. Generations 2 then 3 show the result when oscillators
we apply the following cellular automata rule: “Live cells
Other configurations can have much larger 1. All neighbors must be touching the candidate
periods, or can behave in a more chaotic fash- cell and all are treated the same.
ion. Motionless patterns can be thought of as 2. There must exist at least one translating oscil-
oscillators whose period is one. lator (called a “glider”).
Needless to say, there are a huge number of 3. Random configurations must eventually stabi-
rules that can be applied, and each rule will lize into zero or more oscillators.
cause a distinct action. The rule given above –
the most famous cellular automaton of all – For a more formal description of GL rule
specifies the “Game of Life”, discovered by requirements see (Bays 2005). It is important to
John Horton Conway in 1968. Game of Life note that CA can be represented in one, two,
(GL) rules must satisfy the following informal three or higher dimensions, but most work has
criteria. been done in one or two. Furthermore, neighbors
Cellular Automata in Triangular, Pentagonal, and Hexagonal Tessellations 3
E 1 , E 2 , . . . =F1 , F 2 , . . .
Cellular Automata in Triangular, Pentagonal, and Hexagonal Tessellations, Fig. 3 Examples of expanding rules.
The starting configurations are at the top
4 Cellular Automata in Triangular, Pentagonal, and Hexagonal Tessellations
Introduction
Almost all CA research in two dimensions has
been done using rectangular (Cartesian) coordi-
nates, and hence typically utilizes the square grid.
But there is no reason to limit ourselves to this
tessellation; the number of different possible grids
is almost endless. Here we shall briefly investigate
CA behavior in only three – triangular, hexagonal
and pentagonal.
Cellular Automata in Triangular, Pentagonal, and quite large. The plot at the lower left gives the number of
Hexagonal Tessellations, Fig. 5 Examples of bounded live cells at each generation. These values exhibit a normal
rules that churn endlessly. The total number of live cells can distribution (plot “A”). Note however that there are some
be employed as pseudorandom numbers that approximate a gaps. This is because the rule 1-8/6-8 tends to have “clumps”
normal distribution. Many candidate rules can be used. of living (and fairly large “holes” of non-living) cells.
Naturally the random like patterns eventually repeat, but Hence, before using this technique for generating random
with a sufficiently large initial shape, the period will be numbers, the candidate rule should be carefully investigated
Cellular Automata in Triangular, Pentagonal, and Hexagonal Tessellations 5
With 12 touching neighbors instead of 8 (as in For some rules we can start with bounded
the square grid), we can write more than 16 million forms whose innards churn endlessly forever;
distinct rules, most of which are probably of only these rules can, for example, be used to generate
marginal interest. Many however exhibit behavior random numbers (Fig. 5). Such rules differ some-
worthy of investigation. what from expanding rules in that all finite pat-
Some rules will generate a continually terns are bounded and will not expand
expanding collection of live cells – we shall call indefinitely, but an infinite grid of random live
such rules “expanding” or “unstable” rules. Thus 2, cells will never stabilize.
3/2; 2, 3/3, 4; 2, 3, 4/3 each produce an ever
increasing area of live cells – even with extremely Game of Life Rules in the Triangular Grid
small starting configurations (Fig. 3). A few As mentioned above, the most famous GL rule
expanding rules “barely” expand; i.e. several gen- is Conway’s game, which utilizes a square
erations are required and the initial live configura- grid. But GL rules are not limited to squares;
tion must be fairly large in order to observe quite a few exist in the triangular grid. Among
instability. For example 2, 3, 6/4, 5 can produce these are 4, 5, 6/4; 3, 4/4, 5; 4, 5/4, 5, 6; 2,
unbounded growth, while 2, 3, 8/4, 5 always even- 3/4, 5; 3, 4/4, 5, 6; 2/3; 2, 4/4, 6; 3, 5/4; 2, 4,
tually stabilizes. The fate of configurations under 2, 6/4, 6; 2, 7/3; 2, 7, 8/3. Further information
3, 7/4, 5 is uncertain, but the rule appears to produce about these and other rules can be found in
unbounded growth. Many rules will ultimately lead (Bays 2005) and (Bays 1994) (Figs. 6, 7, 8, 9,
to a stable pattern (Fig. 4), or no live cells at all. 10 and 11).
6 Cellular Automata in Triangular, Pentagonal, and Hexagonal Tessellations
Cellular Automata in Triangular, Pentagonal, and arranged to depict the number of touching neighbors for
Hexagonal Tessellations, Fig. 14 The 14 distinct con- each cell. Where more than one number is given, there are
vex pentagonal tilings (Bays 2005; Wolfram 2002). They some cells with each of those neighbor counts. For exam-
are based upon certain relationships between the angles ple the “67b” tiling is the second tiling where some cells
and lengths of the sides of the particular pentagon that have 6 neighbors and others 7. The Cairo tiling is at the
constitutes the tiling. A sample of the pentagon for that upper left and is topologically equivalent to 7a and 7b.
tiling is displayed at the right of each. The tilings have been Note that 7c and 7d are also topologically equivalent
Cellular Automata in Triangular, Pentagonal, and Hexagonal Tessellations 9
Cellular Automata in Triangular, Pentagonal, and (i 1; j); (i 1; j + 1); (i; j + 1); etc. We can even simulate
Hexagonal Tessellations, Fig. 16 Templates can be grids made up of different types of polygons. Here, we
used to simulate any grid with rectangular coordinates. determine the polygon type by examining the subscripts of
For example, if we are evaluating the neighbors for a the cell in question. Of course, appropriate graphics pro-
hexagonal cell at (i, j) (see “X”) they would be found at cedures must be employed in order to view our grid
many different tessellations – pentagonal and Bays C (2005) A note on the game of life in hexagonal and
otherwise. The ultimate conclusion is that pentagonal tessellations. Complex Syst 15:245–252
Preston K Jr, Duff MJB (1984) Modern cellular automata.
there is room for much work in the area of Plenum Press, New York
non-cartesian CA. Sugimoto T, Ogawa T (2000) Tiling problem of convex
pentagon. Forma 15:75–79
Wolfram S (2002) A new kind of science. Wolfram Media,
Champaign
Bibliography
Cellular Automata
S2 5
in Hyperbolic Spaces,
Fig. 1 The pentagrid:
regular pentagons with
P1
vertex angle p2 R3 4
S1
1
P0
R2 2
R1
14 Cellular Automata in Hyperbolic Spaces
2 3 4
1 1 1
0 0 0
0 1
5 6 7 8 9 10 11 12
1 1 1 1 1 1 1 1
0 0 0 0 0 0 0 0
0 0 1 0 0 0 1 1
0 1 0 0 0 1 0 0
0 1 0 0 1
13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 1
0 0 0 1 1 0 0 0 0 0 0 0 0 1 1 1 0 0 0 0 0
0 0 1 0 0 0 0 1 0 0 0 1 1 0 0 0 0 0 0 1 1
0 1 0 0 1 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0
0 1 0 0 1 0 1 0 0 1 0 0 1
of the pentagrid is called the standard Fibonacci longest representation with respect to the lexico-
tree, illustrated by Fig. 2. Note that on level k, the graphic order and call it the coordinate of the node
first node is numbered f2k and the last one f2kþ21 : to which the corresponding number is attached.
The above splitting induces a particular struc- First, we have that the set of coordinates is a
ture on the standard Fibonacci tree. Define white regular language, which is a corollary of a well-
nodes as nodes which have three sons and black known theorem (see Fraenkel 1985). Now we
nodes as nodes which have two sons. Black and have a more interesting property, which was
white are the two possible values of the status of first noticed in Margenstern (2000) and which
a node. Then, there is a rule to define the status of we call the preferred son property. Let ak. . .a0
the sons of a node. We can write them as follows, be the coordinate of a node n of the standard
in self-explained notations: Fibonacci tree, with a0 as the lightest digit of
the representation. Denote it by [n]. The property
W ! BWW
says that for each node n of the standard
B ! BW:
Fibonacci tree, there is exactly one son of
As initially performed in Margenstern (2000), n whose coordinate is [n]00. This son is called
let us represent the numbers attached to the nodes of preferred. Moreover, there is a rule to find out
the Fibonacci tree in the numeration basis defined the preferred son from the status of a node: in a
by the Fibonacci sequence itself, starting from f1. black node, the preferred son is the black son; in a
The representation is not unique. Choose the white node, the preferred son is the middle one.
Cellular Automata in Hyperbolic Spaces 15
for CAs in the Euclidean plane and in the 3D 2007c), it is also possible in the case of tessella-
Euclidean space. tions of the hyperbolic plane to get rid of the
The third condition already requires that we origin. We just mention this point, here, and
have a system of coordinates for the tiles at our refer the interested reader to the quoted papers
disposal. More than three centuries after Des- for a closer study.
cartes’ discovery of the system of coordinates Once again, we illustrate how we proceed by
which everybody uses for the Euclidean plane, the case of the pentagrid. It is repeated in the case
this condition is trivially fulfilled. This is not only of the heptagrid (see Margenstern 2006a, 2007a)
the three-century usage. This is also the case and in the case of the dodecagrid (see
because the mathematical structure of the group Margenstern 2006b).
of displacements which leaves the considered For the implementation, we first fix a basis of
tessellations of the Euclidean plane globally splitting and the representation of the tiling. As
invariant is a very simple structure. indicated in Margenstern (2000, 2007c), there are
The situation is very different in the case of a lot of choices with the same basis of splitting.
hyperbolic spaces. Before Margenstern (2000), Moreover, in the case of the pentagrid and of the
there was no convenient, at least fast, procedure heptagrid in which the standard Fibonacci tree is
to define the coordinates of the tiles in a way also a spanning tree, we have the choice between
which is in connection with the geometrical prop- using the Fibonacci sequence, as we did in
erties of the tiling. section “The Locating Problem in Hyperbolic
Now, the splitting method gives such a solu- Tilings,” and using the basis derived from
tion. First, it effectively exhibits a tree which the polynomial of the splitting. The difference is
generates the tiling. As Gromov pointed out that the Fibonacci pffiffi sequence is defined by the
(Gromov 1981), hyperbolic spaces are character- golden mean 1þ2 5 , while the sequence of the
ized by a tree structure. Second, it provides fast splitting pisffiffi defined by the square of the golden
algorithms to handle these coordinates. By fast, mean, 3þ2 5
we mean that the basic algorithms we need are Let us go on with the Fibonacci sequence, as is
linear in time with respect to the coordinate of it used in the majority of papers.
the initial point. Note that nobody really matters The preferred son property allows us to com-
with the fact that addition of vectors in Euclidean pute very easily the coordinates of the neighbors
coordinates is linear, while multiplication of of a cell n from [n]: the computation is linear in
coordinates by a scalar is not. Here, we have no time with respect to the length of [n] (see
addition, no multiplication, and no nice formula. Margenstern 2003, 2007c). Similarly, the path
We have algorithms only, but they turn out to from a cell n to the root of its tree can be com-
work in the best time. puted in a linear time with respect to the length of
The result of these considerations is that the [n] (again see Margenstern 2003, 2007c). As a
directions, north, south, east, and west, which simple example, if m is defined by ½n ¼ ½ma1 a0 ,
play a so nice role in the Euclidean case no more where a 0 is the lowest digit of [n], the father of
exist. In fact, we have infinitely many directions, n has m þ a1 as its number.
each of which defines an essential direction in the What we indicated up to now fixes the coordi-
space: if you follow other directions, you will nates for a cell whose supporting tile is in a given
never go to the area covered by this one. Of quarter. We can consider the central cell as the
course, an infinite amount of information is root of a tree whose sub-trees are exactly the five
ruled out in computer science. And so, we replace initial quarters which lie around the central
this basic indetermination of the direction by the pentagon. Now, it is enough to number the five
direction of the father. Of course, we are led to a sub-trees attached to these quarters in order to
root and a central cell, but nobody complains completely define the coordinates of a cell. The
about using an origin in the Euclidean case. central pentagon has 0 as a unique coordinate. All
Moreover, as shown in Margenstern (2006a, other cells are defined by two numbers: (a, n). The
Cellular Automata in Hyperbolic Spaces 17
first number, a is in {1. . .5} and defines the variables, is triggered at initial time. Once it is
quarter. The second number, v, defines the tile reached, the information comes back to the root
in the indicated quarter. Together with its coor- from the leaves of the tree, i.e., the nodes which
dinate, a cell is associated with other data: the are on the level n of the tree: each node computes
status of its supporting tile and the indication of the OR on the values of its left-hand side and
which side is shared with its father. On one hand, right-hand side sons. Accordingly, the root gives
note that the coordinate is a hardware feature: it is true if and only if there is a branch from it to a leaf
never known by the cell and it cannot be – it has along which the value is always true.
not a bounded size. Note that this is the same for From this, applying classical tools of the the-
CAs in Euclidean spaces. On the other hand, the ory of complexity, we obtain that any
status of the supporting node can be known by the NP-complete problem can be solved in polyno-
cell. As shown in Margenstern (2003), one can mial time by an appropriate cellular automaton of
define rules for a cellular automaton to dispatch the hyperbolic plane.
this information. As it is a finite information
which can be provided by the hardware, we may
P = NP in the Hyperbolic Plane
assume that the cell knows it.
From what we have seen previously, we have that
the classical class NP is contained in the class of
HCAs which work in polynomial time, denote it
Complexity of Cellular Automata by Ph. Now, it is also possible to define NPh for
in Hyperbolic Spaces
HCAs, taking the classical definition of nondeter-
ministic computations in polynomial time.
Now, we are ready to give the results about the
As shown in Iwamoto et al. (2002), it turns out
complexity classes of HCAs.
that Ph = NPh. The key point is that the compu-
tation of a nondeterministic Turing machine in
SAT and NP-Complete Problems
time O(t(n)), with t(n) n, can be computed by a
In Margenstern and Morita (1999), HCAs are
deterministic HCA in time O(t2(n)).
proved to be able to solve SAT in polynomial
From this theorem, the following surprising
time. Historically, the possibility to solve
result can easily be derived (see Iwamoto
NP-complete problems in the hyperbolic plane
et al. 2002):
was first announced in Morgenstein and
Kreinovich (1995). Although the authors of
Ph ¼ NPh ¼ PSPACE,
Margenstern and Morita (1999) were not aware
of paper (Morgenstein and Kreinovich 1995), the
where PSPACE is the classical class of functions
latter paper does not involve cellular automata computed in polynomial space by a Turing
and does not provide a precise description of
machine.
how SAT can be solved in the new frame. On
Of course, in these results, a basic ingredient is
the contrary, Margenstern and Morita (1999) the possibility, given by the hyperbolic plane, to
describe a HCA which is able to solve the prob-
occupy a working space of exponential area
lem. In Margenstern and Morita (1999), the com-
within a polynomial time. The above process for
putation is estimated as quadratic. In fact it can be solving SAT is a basic example of such a
proved to be linear in the size of the input.
possibility.
The solution for SAT is easy: it makes use of a
Fibonacci tree, in which only two nodes are
selected among the sons of a node. Each level Other Parts of the Complexity Hierarchy
represents the possible assignment of true and of HCAs
false values to the variable indexed by this In fact, if we look at the hierarchy of complexity
level. The computation of all possible assign- classes for HCAs, we get a landscape which is
ments until the level n, where n is the number of very different from the classical situation.
18 Cellular Automata in Hyperbolic Spaces
We have the following situation, described in cellular automaton A with states in Q on ℤ2, F A is
Iwamoto and Margenstern (2003): a continuous operator on C, fitted with the product
topology, which also commutes with the shifts on
DLOGh ¼ NLOGh ¼ Ph ¼ NPh ¼ PSPACE ℤ2. The remarkable property is that the converse is
⊈PSPACEh ¼ EXPTIMEh true. However, if we take a continuous operator F
¼ NEXPTIMEh ¼ EXPSPACE: on C which commutes with shifts, the proof of the
theorem does not allow us to obtain an effective
We can notice that compared to the Euclidean process that would provide a cellular automaton A
analogs, the hyperbolic hierarchy seems to be such that F A = F. The problem is that there is no
very flat. As, by construction, Ph ⊈EXPTIMEh ; algorithm which would compute the radius of the
there are indeed two classes on which the hierar- neighborhoods of A from F.
chy concentrates. In Margenstern (2008d) we proved that a sim-
We also have NPh ⊈APh ; unless ilar characterization exists for hyperbolic cellular
PSPACE ¼ NEXPTIME, where APh denotes automata on the pentagrid or the heptagrid, pro-
the class of alternate HCAs. As with classical vided we consider rotation invariant cellular
machines, an alternate HCA is defined on the automata. The characterization holds for rotation
set of configurations of a nondeterministic invariant cellular automata on all tilings of the
HCA. In the tree of these configurations, certain hyperbolic plane or of the hyperbolic 3D space
nodes are called existential; others are called whose tiling is algorithmically spanned by a tree.
universal. At an existential node, the node is
accepting if and only if it has at least one Synchronization of a HCA
accepting child. At a universal node, the node is Although no paper is especially devoted to this
accepting if and only if all its children are problem, we mention it because it has an analog
accepting. The result about APh indicates a sim- to the standard problem of the firing squad in
ilar situation with the Euclidean classes where one-dimensional CAs, and we shall use it in the
P⊈AP, unless P ¼ PSPACE: Accordingly, we next section.
may expect that alternating HCAs should be more In fact, as mentioned more or less explicitly in
powerful than HCAs, either deterministic or papers devoted to HCA (see Iwamoto and
nondeterministic. Margenstern (2003) and Margenstern (2008a)),
for instance, it is very easy to synchronize a disk
or a sector inside a disk, defined by a tree rooted
On Specific Problems of Cellular at the center of the disk. The idea is simply to
Automata simulate any classical algorithm of synchroniza-
tion of a one-dimensional CA on each branch of
Characterization of a HCA the tree for each radius of the disk.
For classical cellular automata, i.e., for cellular The synchronization is linear in the radius of
automata in a Euclidean space, there is a well- the disk or the height of the tree.
known characterization of cellular automata in
terms of operations on the space of configurations. Communications Between HCAs
Consider the most studied case of the square grid in Another problem, more specific to HCAs, is the
the Euclidean plane. The grid is most often identi- communication between HCAs, possibly distant
fied with ℤ2 so that if Q is the set of states of a ones. Two papers study the problem in different
cellular automaton, C = Qℤ is the set of all possible
2
settings (see Margenstern 2006a, 2007a).
configurations for a cellular automaton on ℤ2 with In Margenstern (2006a) the question is: how to
states in Q A cellular automaton A on ℤ2 with establish a contact between two cells of a HCA,
states in Q defines an operator on C called the possibly distant ones? The paper provides a solu-
global function of A denoted by F A . A famous tion based on a new system of coordinates in
theorem (see Hedlund 1969), says that if A is a which there is no more an origin. The new system
Cellular Automata in Hyperbolic Spaces 19
is based on the possibility to represent the hyper- HCAs with a Small Number of States”). There
bolic plane as a union of growing quarters. We fix was also a paper about an intrinsically universal
such a sequence in an appropriate way. Each term HCA (see subsection “An Intrinsically Universal
of the sequence is a Fibonacci tree, indexed by an HCA”). There was also very recently two papers
integer n, and it contains all the trees indexed by about strong universality of HCAs with a rather
m when m n. Inside a given Fibonacci tree, we small number of states (see subsection “Strong
use the standard system of coordinates, indicated Universal HCAs with a Small Number of States”)
in section “The Locating Problem in Hyperbolic
Tilings.” In the construction, the roots of the
mentioned trees belong to a line d. It is not diffi- Weakly Universal HCAs with a Small Number
cult to see that sending signals on d makes it of States
possible for the cell to establish a contact in a First, we have to notice that the just-mentioned
linear time with respect to their mutual distance. universal HCAs with a small number of states are
In Margenstern (2007a), another problem is in fact weakly universal HCAs. The term weak
considered. This time all cells may dispatch mes- refers to two conditions:
sages, and each cell forwards the messages it
receives and to which it does not want to reply. – The HCA needs an infinite initial
Accordingly, the same cell may be an emitter of configuration.
messages, a receiver of messages, and a relay in – The initial configuration is ultimately
the message system. The idea is to use the tree periodic.
property to be in bijection with the tiling as fol-
lows: each emitting cell considers that it is the Note that these conditions are standardly used
center of the hyperbolic plane, and the message is with ordinary CAs where universality with a
accompanied by an address which is updated by small number of states is investigated.
the relays and which is the address in the tree The second condition requires some explana-
whose root is the sender of the message. This tion. In the context of a hyperbolic space, the
allows any receiver willing to answer the mes- notion of periodicity is not as clear as it is in the
sage to send it to the right emitter. Again, the Euclidean case. Accordingly, we mean, by ulti-
complexity of the computation is linear in the mate periodicity that at large, i.e., outside a big
mutual distance of a sender and a receiver. Also enough domain, the configuration can be split
see subsection “Communications in a Network.” into finitely infinite domains in each of which it
is globally invariant under a shift depending on
the domain.
Universality in Cellular Automata The results accumulated in the recent years. In
in Hyperbolic Spaces the hyperbolic plane, there were a weakly univer-
sal HCA with nine states in the pentagrid (see
Of course, from the existence of universal cellu- Margenstern and Song 2009) and then two uni-
lar automata on the line, we conclude that there versal HCAs in the heptagrid: first with six states,
are universal HCAs. This means that there are (Margenstern and Song 2008), and then with four
HCAs which are able to simulate any universal states (see Margenstern 2011b). Very recently,
device, as a Turing machine, for instance. there was a weakly universal HCA in the tiling
There was recently a definite progress in the {13, 3} of the hyperbolic plane with two states
study of universal HCAs. From the first result only. In the dodecagrid, after the weakly univer-
about a universal HCA in the pentagrid with sal HCA with five states (see Margenstern
22 states (see Herrmann and Margenstern 2003), 2006b), there was a weakly universal HCA with
we arrive to universal HCAs with two states in three states (see Margenstern et al. 2010a) and
the hyperbolic plane and also in the hyperbolic then with two states (see Margenstern 2013),
3D space (see subsection “Weakly Universal which is the best result in this tiling.
20 Cellular Automata in Hyperbolic Spaces
Cellular Automata in Hyperbolic Spaces, memory switch and the flip-flop, we represented the left-
Fig. 4 Heptagrid and four states: the idle configurations hand side version only: the right-hand side ones can easily
at crossings and switches. From left to right: crossing, be devised from these ones
fixed switch, memory switch, and flip-flop. For the
F
E
C B
A
D A
Cellular Automata in Hyperbolic Spaces, roundabout, first branching: through E. Arrival at the
Fig. 5 Configuration at a roundabout in {13, 3}. Left- second and third branching: through A. Exit through F at
hand side: general view; right-hand side: zoom at a the third branching
branching. Right-hand side picture. Arrival at a
initial configuration of the HCA to be simulated, simulating cell, the new state is determined, and
say, A, and an encoding of the transition table of it is installed in the appropriate region, controlled
A. Figure 7 illustrates the mechanism of propaga- by the simulating cell. When this is performed,
tion of the scaled tree. the cell waits until it is informed by its simulating
Each step of the simulated HCA A is simulated sons that their step of computation is completed.
by a cycle of steps of the simulating HCA U. The When this is the case, the cells inform its father in
number of steps of U in a cycle is not constant. It the scaled tree that it finished its computation.
may be increasing, especially if the simulated Accordingly, when the central cell receives the
configuration is growing during its own compu- message of completion from all its neighbors of
tation. The synchronization algorithm of section the scaled tree, it knows that the computation of
“On Specific Problems of Cellular Automata” is this step of A is finished. Then the comparison
used to delimit the stages into which a cycle is with the previous configuration is performed,
split. These stages are the reception of the current thanks to a synchronization. Depending on the
states of the neighbors of the simulated cell of A, result of the comparison, the computation is
for each simulating cell of U. When this is stopped, if there was no difference, or it goes
achieved, possibly at different times for each on, when a difference was noticed.
22 Cellular Automata in Hyperbolic Spaces
Cellular Automata
in Hyperbolic Spaces,
Fig. 7 Propagation of a
scaled tree
Strong Universal HCAs with a Small Number been performed in Margenstern (2010) for the
of States weakly universal case. As underlined in that
The intrinsic cellular automaton of subsection paper, the implementation of one dimensional
“An Intrinsically Universal HCA” makes a into the pentagrid, the heptagrid, and the
natural transition to this subsection: that cellular dodecagrid is easy if not almost trivial. Then, it
automaton is strongly universal. It starts from is enough to implement the two-state weakly
a finite configuration, and it simulates a cellular universal cellular automaton of Cook (2004)
automaton starting from a finite configuration. As using the elementary cellular automaton defined
mentioned in Margenstern (2008a), the number by rule 110 (also see Wolfram et al. 2002).
of states of such an automaton is enormous. Paper To reach strong universality, it is not that
(Margenstern 2008a) does not even try to give an trivial: it is needed to go on an initial segment in
estimate to the number of states. such a way that the continuation of the segment
In this section, we consider the possibility to remains a segment supported by the same line. It
devise a small strongly universal cellular autom- is also desirable that the continuation is
aton. A simple idea would be to implement a performed at the same time as the computation
one-dimensional cellular automaton. This has itself. These constraints are satisfied in
Cellular Automata in Hyperbolic Spaces 23
that if the tiles are decorated, a solution must In order to go back, an additional facility is
satisfy the matching of any adjacent tiles along given: a compass, in the form of a point which
their common side. indicates the direction where the central cell lies
Robert Berger proved in 1966 (see Berger 1966) when it is no more visible in the disk. If the user
that the tiling problem is undecidable in the Euclid- wants to go back to the central cell, it is enough to
ean plane. In Robinson (1971), Raphael Robinson go to the direction of the compass as long as the
significantly simplified Berger’s solution and raised central cell is no more visible in the disk. What
the question about the status of the same problem can be seen is illustrated by the right-hand side
for the hyperbolic plane. Robinson himself gave a picture of Fig. 8.
partial answer, when the first tile is fixed, to this
problem in 1978 (see Robinson 1978). A Japanese Keyboard for Cell Phones
A few weeks after the time when the result Another application deals with cell phones with a
published in Margenstern (2008c) was announced possible way to write messages in Japanese. The
(see Margenstern et al. 2007b), another solution of idea is based on the fact that the Japanese lan-
the same problem was claimed (see Kari 2007). guage has two syllabic alphabets, hiraganas and
The solution given in Margenstern (2008c) turned katakanas, for phonetic purpose.
out to be very fruitful: the construction given in These syllabic alphabets are based on five
Margenstern (2008c) allows us to prove the vowels and the same series of consonants is
undecidability of the periodic tiling problem (see used for each vowel. The corresponding syllabic
Margenstern 2009a) and then the undecidability of signs are dispatched as indicated in Fig. 9. The
the finite tiling problem (see Margenstern 2008e). use of the keyboard is similar to that of the color
Also, another important result was obtained by a chooser. It can be noticed that any syllabic sign
refinement of the construction produced in can be reached in at most three clicks on the keys.
Margenstern (2008c): the injectivity of the global
function of a cellular automaton of the hyperbolic Communications in a Network
plane is also undecidable (see Margenstern The situation of Margenstern (2007a) was thor-
2009b). It also turned out that the classical theo- oughly studied in Margenstern (2012). The com-
rems connecting the surjectivity of the global munication protocol described in Margenstern
function of a cellular automaton with the (2007a) was implemented in the heptagrid with
injectivity and the injectivity on finite configura- an additional feature. In Margenstern (2007a), it
tions are no more true for hyperbolic cellular was decided that the decision by a cell to send
automata (see Margenstern 2009c). messages or to reply to messages it receives fol-
lows a Poisson law. Of course, the Poisson coef-
ficients are different. In Margenstern (2012), in
Possible Applications order to be more realistic, it was decided that a
message sent by a cell cannot run forever at infin-
A few applications of the theory described in the ity. When it is sent, it is dispatched within a disk of
previous sections were indicated, in particular in radius r, where r is a randomly fixed integer again
Margenstern (2008b). We mention three of them. following a Poisson law. Two experiments where
performed with a help of a simulation through a
Color Chooser computer program. The difference between the
The first one is a color chooser. It consists in a experiments was the size of the expansion radius
representation of the heptagrid in Poincaré’s disk, of a message and the coefficients of the Poisson
as illustrated by the left-hand side picture of laws followed by the different parameters. In
Fig. 8. At each step, the user selects a neighbor Margenstern (2012), an account of both experi-
v of the central cell. At the next time, v appears at ments is given. The experiments indicate that the
the central place. The motion is repeated until the model seems to be reasonable. See Margenstern
user finds out the color he/she wished. (2012) for more details.
Cellular Automata in Hyperbolic Spaces 25
Cellular Automata in Hyperbolic Spaces, Fig. 8 Left-hand side: idle position of the color chooser. Middle: the user
chooses to look at the blue colors. Right-hand side: the compass
Future Directions
Cook M (2004) Universality in elementary cellular autom- Margenstern M (2008b) Cellular automata in hyperbolic
ata. Complex Syst 15(1):1–40 spaces, volume 2, implementation and computation.
Fraenkel AS (1985) Systems of numerations. Am Math Old City Publishing, Philadelphia, 360p
Mon 92:105–114 Margenstern M (2008c) The domino problem of the
Gromov M (1981) Groups of polynomial growth and hyperbolic plane is undecidable. Theor Comp Sci
expanding maps. Publ Math l’IHES 53:53–73 407:29–84
Hedlund G (1969) Endomorphisms and automorphisms of Margenstern M (2008d) On a characterization of cellular
shift dynamical systems. Math Syst Theory 3:320–375 automata in tilings of the hyperbolic plane. Int J Found
Herrmann F, Margenstern M (2003) A universal cellular Comp Sci 19(5):1235–1257
automaton in the hyperbolic plane. Theor Comp Sci Margenstern M (2008e) The finite tiling problem is
296:327–364 undecidable in the hyperbolic plane. Int J Found
Iwamoto Ch, Margenstern M (2003) A survey on the Comp Sci 19(4):971–982
complexity classes in hyperbolic cellular automata. Margenstern M (2009a) The periodic domino problem is
Proceedings of SCI’2003, V, pp 31–35 undecidable in the hyperbolic plane. Lect Notes
Iwamoto Ch, Margenstern M, Morita K, Worsch Th Comput Sci 5797:154–165
(2002) Polynomial-time cellular automata in the Margenstern M (2009b) The injectivity of the global func-
hyperbolic plane accept exactly the PSPACE lan- tion of a cellular automaton in the hyperbolic plane is
guages. SCI’2002. Orlando, pp 411–416 undecidable. Fundam Inform 94(1):63–99
Kari J (2007) The tiling problem revisited. Lect Notes Margenstern M (2009c) About the garden of Eden theo-
Comput Sci 4664:72–79 rems for cellular automata in the hyperbolic plane.
Lindgren K, Nordahl MG (1990) Universal computations Electron Notes Theor Comp Sci 252:93–102
in simple one-dimensional cellular automata. Complex Margenstern M (2010a) A weakly universal cellular
Syst 4:299–318 automaton in the hyperbolic 3D space with three
Margenstern M (2000) New tools for cellular automata in states. Discrete Mathematics and Theoretical Com-
the hyperbolic plane. J Univ Comp Sci puter Science. Proceedings of AUTOMATA’2010,
6(12):1226–1252 pp 91–110
Margenstern M (2002a) A contribution of computer sci- Margenstern M (2010b) Towards the frontier between
ence to the combinatorial approach to hyperbolic decidability and undecidability for hyperbolic cellular
geometry, SCI’2002, 14–19 July 2002. Orlando automata. Lect Notes Comput Sci 6227:120–132
Margenstern M (2002b) Revisiting Poincaré’s theorem Margenstern M (2010c) An upper bound on the number of
with the splitting method, talk at Bolyai’200, Interna- states for a strongly universal hyperbolic cellular
tional Conference on Geometry and Topology, Cluj- automaton on the pentagrid, JAC’2010, 15–17 Dec
Napoca, Romania, 1–3 October 2002 2010. Turku, Finland, Proceedings, Turku Center for
Margenstern M (2003) Implementing cellular automata Computer Science 2010. ISBN 978-952-12-2503-1,
on the triangular grids of the hyperbolic plane 168-179
for new simulation tools, ASTC’2003. Orlando, Margenstern M (2011a) A new weakly universal cellular
29 Mar- 4 Apr automaton in the 3D hyperbolic space with two states.
Margenstern M (2004) The tiling of the hyperbolic 4D Lect Notes Comput Sci 6945:205–2017
space by the 120-cell is combinatoric. J Univ Comp Margenstern M (2011b) A universal cellular automaton on
Sci 10(9):1212–1238 the heptagrid of the hyperbolic plane with four states.
Margenstern M (2006a) A new way to implement cellular Theor Comp Sci 412:33–56
automata on the penta- and heptagrids. J Cell Autom Margenstern M (2012) A protocol for a message system
1(1):1–24 for the tiles of the heptagrid, in the hyperbolic plane.
Margenstern M (2006b) A universal cellular automaton Int J Satell Commun Policy Manag 1(2–3):206–219
with five states in the 3D hyperbolic space. J Cell Margenstern M (2013a) Small universal cellular in hyper-
Autom 1(4):317–351 bolic spaces, a collection of jewels. Emergence, Com-
Margenstern M (2007a) On the communication between plexity and Computation, Springer, p 320
cells of a cellular automaton on the penta- and Margenstern M (2013b) About strongly universal cellular
heptagrids of the hyperbolic plane. J Cell Autom automata. Proceedings of MCU’2013, (2013) EPTCS
(to appear) 128, 93–125
Margenstern M (2007b) About the domino problem in the Margenstern M, Morita K (1999) A polynomial solution
hyperbolic plane, a new solution. arXiv:cs/0701096, for 3-SAT in the space of cellular automata in the
(Jan 2007), p 60 hyperbolic plane. J Univ Comput Syst 5:563–573
Margenstern M (2007c) Cellular automata in hyperbolic Margenstern M, Morita K (2001) NP problems are tracta-
spaces, theory, vol 1. Old City Publishing, Philadel- ble in the space of cellular automata in the hyperbolic
phia, 422p plane. Theor Comp Sci 259:99–128
Margenstern M (2008a) A uniform and intrinsic proof that Margenstern M, Skordev G (2003a) The tilings {p, q} of
there are universal cellular automata in hyperbolic the hyperbolic plane are combinatoric, SCI’2003, V,
spaces. J Cell Autom 3(2):157–180 pp 42–46
Cellular Automata in Hyperbolic Spaces 27
Margenstern M, Skordev M (2003b) Tools for devising Martin B (2005) VirHKey: a virtual hyperbolic keyboard
cellular automata in the hyperbolic 3D space. Fundam with gesture interaction and visual feedback for mobile
Inform 58(2):369–398 devices, MobileHCI’05, Sept. Salzburg
Margenstern M, Song Y (2008) A universal cellular Robinson RM (1971) Undecidability and nonperiodicity
automaton on the ternary heptagrid. Electron Notes for tilings of the plane. Invent Math 12:177–209
Theor Comp Sci 223:167–185 Robinson RM (1978) Undecidable tiling problems in the
Margenstern M, Song Y (2009) A new universal cellular hyperbolic plane. Invent Math 44:259–264
automaton on the pentagrid. Parallel Process Lett Róka Z (1994) One-way cellular automata on Cayley
19(2):227–246 graphs. Theor Comp Sci 132:259–290
Morgenstein D, Kreinovich V (1995) Which algorithms Stewart I (1994) A subway named turing, mathematical
are feasible and which are not depends on the geome- recreations in scientific American. pp 90–92
try of space-time. Geombinatorics 4(3):80–97 Wolfram S (2002) A new kind of science. Wolfram Media
restricted-totalistic (RT) – but can be as varied
Structurally Dynamic Cellular as those for conventional cellular automata.
Automata Degree The degree of a node (or site, i) of a
graph is equal to the number of distinct nodes
Andrew Ilachinski to which i is linked, and where the links are
Center for Naval Analyses, Alexandria, VA, USA assumed to possess no directional information.
In general graphs, the in-degree (= number of
incoming links towards i) is distinguished from
Article Outline the out-degree (= number of outgoing links
originating at i).
Glossary Effective dimension A quantity used to approx-
Definition of the Subject imate the dimensionality of a graph. It is
Introduction defined as the ratio between the average num-
The Basic Model ber of next-nearest neighbors to the average
Emerging Patterns and Behaviors degree, both averaged over all nodes of the
SDCA as Models of Computation graph. The effective dimension equals the
Generalized SDCA Models Euclidean dimension d, in cases where the
Related Graph Dynamical Systems graph is the familiar d-dimensional hypercubic
SDCA as Models of Fundamental Physics lattice.
Future Directions and Speculations Graph A graph is a finite, nonempty set of nodes
Bibliography (referred to as “sites” throughout this article),
together with (a possibly empty) set of edges
Glossary (or links). The links may be either directed
(in which case the edge from a site i, say, is
Adjacency matrix The adjacency matrix of a directed away from i toward another site j, and
graph with N sites is an N N matrix [aij] is considered distinct from another directed
with entries aij = 1 if i and j are linked, and edge originating at j and pointed toward i) or
aij = 0 otherwise. The adjacency matrix is undirected (in which case if a link exists
symmetric (aij = aji) if the links in the graph between sites i and j it carries no directional
are undirected. information).
Coupler link rules Coupler rules are local rules Graph grammar Graph grammars (sometimes
that act on pairs of next-nearest sites of a graph also referred to as graph rewriting systems)
at time t to decide whether they should be apply formal language theory to networks.
linked at t + 1. The decision rules fall into Each language specifies the space of “valid
one of three basic classes – totalistic (T), structures”, and the production (or “rewrite”)
outer-totalistic (OT) or restricted-totalistic rules by which given graphs may be trans-
(RT) – but can be as varied as those for con- formed into other valid graphs.
ventional cellular automata. Graph metric function The graph metric func-
Decoupler link rules Decoupler rules are local tion defines the distance between any two
rules that act on pairs of linked sites of a graph nodes, i and j. It is equal to the length of the
at time t to decide whether they should be shortest path between i and j. If no path exists
unlinked at t + 1. As for coupler rules, the (such as when i and j are on two disconnected
decision rules fall into one of three basic components of the same graph), the distance is
classes – totalistic (T), outer-totalistic (OT) or assumed to be equal to 1.
# Springer-Verlag 2009 29
A. Adamatzky (ed.), Cellular Automata,
https://doi.org/10.1007/978-1-4939-8700-9_528
Originally published in
R. A. Meyers (ed.), Encyclopedia of Complexity and Systems Science, # Springer-Verlag 2009
https://doi.org/10.1007/978-0-387-30440-3_528
30 Structurally Dynamic Cellular Automata
Graph-rewriting automata Graph-rewriting auto- SDCA model hierarchy The SDCA model hier-
mata are generalized CA-like systems in which archy is a set of eight related structurally
both (the number of) nodes and links are allo dynamic cellular automata models, defined
wed to change. explicitly for studying their formal computa-
Next-nearest neighbor Two sites i and j are tional capabilities. The hierarchy is ordered
next-nearest neighbors in a graph if (1) they (from lowest to highest level) according to
are not directly linked (so that aij = 0; see their relative computational strength. For
adjacency matrix), and (2) there exists at least example, the SDCA model at the top of the
one other site k such that k 2 = {i, j}, and i and hierarchy is capable of simulating a conven-
j are both lined to k. tional CA with a speedup factor of two.
Random dynamics approximation The long-
term behavior of structurally dynamic cellular
automata may be approximated in certain cases Definition of the Subject
(in which the structure and value configura-
tions are both sufficiently random and Structurally dynamic cellular automata
uncorrelated) by a random dynamics approxi- (abbreviated, SDCA) are a generalized class of
mation: values of sites are replaced by the CA in which the topological structure of the
probability ps of a site having value s (and is (usually quiescent) underlying lattice is dynami-
assumed to be equal for all sites), and links cally coupled to the local site value configuration.
between sites are replaced by the probability The coupling is defined to treat geometry and
p‘ of being linked (and also assumed to be the value configurations on an approximately equal
same for all pairs of sites). The approximation footing: the lattice structure is altered locally as a
often yields qualitatively correct predictions function of individual site neighborhood value-
about how the real system evolves under a states and geometries, while the underlying local
specific set of rules; for example, to predict topology supports site-value evolution precisely
whether one expects unbounded growth or as in conventional nearest-neighbor CA models
that the lattice will eventually settle onto a defined on random lattices.
low periodic state or simply decay. SDCA provide a dynamical framework for a
Restricted totalistic rules Restricted totalistic CA-like analysis of the generation, transmission
rules are a generalized class of link rules and interaction of topological disturbances in a
(operating on pairs of sites, i and j), analogous lattice. Moreover, they provide a natural testbed
to “outer totalistic” rules (that operate on site for studying self organized geometry; by which
values) used in conventional CA. The local we mean true structural evolution, and not merely
neighborhood around i and j is first partitioned space-time patterns of value configurations that
into three sets: (1) the two sites, i and j; (2) sites may be interpreted geometrically (but are really
connected to either i or j, but not both; and just “bits” of information overlayed on top of an
(3) sites connected to both i and j. The otherwise static background lattice).
restricted totalistic rule is then completely
defined by associating a specific action with
each possible 3-tuple of site-value sums (where Introduction
the individual components represent a unique
sum in each of the three neighborhoods). SDCA were formally introduced in 1986 as part of
Structurally dynamic cellular auto- a physics doctoral dissertation by Ilachinski
mata Structurally dynamic cellular automata (1988), and developed further by Ilachinski and
are generalizations of conventional cellular Halpern (1987a, b), Halpern (1989, 1996),
automata models in which the underlying lat- Halpern and Caltagirone (1990), Majercik
tice structure is dynamically coupled to the (1994), and Alonso-Sanz and Martín (2006),
local site-value configurations. Alonso-Sanz (2006, 2007); in their original
Structurally Dynamic Cellular Automata 31
incarnation (Ilachinski 1986), and at least two discrete pre-geometric theories of space-time
subsequent papers (Halpern and Caltagirone (Meschini et al. 2005). Just as “value structure”
1990; Rose 1993), SDCA were called topological solitons are ubiquitous in conventional CA
automata. Pedagogical discussions appear in models (Ilachinski 2001; Wolfram 1984), “link
Adamatzky (1995) and Ilachinski (2001). Exten- structure” solitons might emerge in SDCA; phys-
sions of the basic SDCA model (all discussed in ical particles would, in such a scheme, be viewed
this article) include the addition of probabilistic asgeometrodynamic disturbances propagating
rules, memory and reversibility. within a dynamic lattice. Three SDCA-like theo-
Applications include the simulation of crystal ries of pregeometry have recently been proposed
growth (Krivovichev 2004), the study of pattern in which space-time is a self-organized emergent
formation of random cellular structures construct: Hillman (1995), Nowotny and
(Schliecker 1998), modeling synaptic plasticity Requardt (2006) and Wolfram (2002).
in neural network models (Gerstner and Kistler Finally, we briefly comment on ostensible over-
2002), phase transitions in chemical systems laps between SDCA and four other related fields of
(Rose et al. 1994), chemical self-assembly study: (1) Lindenmeyer (or L-) systems, (2) graph
(Hasslacher and Meyer 1998), and gene- grammars, (3) random graphs (abbreviated, RG),
regulatory networks (Halpern and Caltagirone and (4) dynamic network analysis (abbreviated,
1990). Majercik (1994) hasstudied SDCA as gen- DNA). L-systems (Prusinkiewicz and
eralized models of computation, and describes a Lindenmayer 1990) are generalized CA systems
CA-universal SDCA that can simulate any con- in which the number of sites can grow with time,
ventional CA of the same dimension. and consist of recursive rules for rewriting strings
More recently, O’Sullivan (2001) and Saidani of symbols. If interpreted graphically, abstract
(2003, 2004) have used graph-based CA models symbol strings can be used to model growth pro-
similar to SDCA to study urban dynamics and cesses of plants and evolving morphology of phys-
emergent behaviors of self-reconfigurable robots, ical organisms. Graph grammars (Grzegorz 1997;
respectively. Tomita et al. (2002, 2005, 2006a, b, c) Kniemeyer et al. 2004) apply formal language
have introduced graph-rewriting automata in theory to networks, and consist of production
which both links and (the number of) nodes are rules that define the set of “valid structures” in a
allowed to change; and show that these systems given graph language. The study of RG (Durrett
are capable of both self-replication and Turing 2006) was introduced by Erdos and Renyi in the
universality (among with many other emergent late 1950s (Erdos and Renyi 1960), and is a math-
behaviors). Since SDCA provide the basic formal- ematical framework for exploring the general topo-
ism for describing locally induced topological logical structures of computational systems and the
changes within arbitrary graphs, they are a poten- behavior of certain random dynamical systems.
tially powerful general tool for studying complex Like SDCA, RG describes evolving graphs, but
adaptive networks, such as communication and the dynamics are global and random. DNA
social networks (Alonso-Sanz and Martin 2006). (Mendes 2004; Newman et al. 2006) is an emerg-
The concept behind SDCA has also been used as a ing field that fuses traditional social network theory
foundation for philosophical musings about com- with statistical analysis and modeling; part of its
putationally emergent artificiality (Mustafa 1999). charter is to explore general properties of network
More ambitious applications of SDCA generation and evolution.
encroach on fundamental physics. Because While, conceptually speaking, there is a prima
SDCA are inherently self-modifying systems – facie relationship between SDCA and all four
in which physical events are not just dynamically fields of study, the elucidation of a more precise
coupled to, but are an integral part of the spatio- nature of the relationship between SDCA and
temporal arena on which their transformations are these other systems awaits a future study. (The
defined – they are a potentially powerful method- relationship appears particularly strong between
ological and ontological tool for exploring SDCA and a generalized L-system called the
32 Structurally Dynamic Cellular Automata
graph development system (abbreviated, GDS), Dij ¼ Minimum #links, l rs j fr,sg Pij , (2)
Paths,Pij
introduced by Doi (1984), but not developed fur-
ther since its original conception. Using incidence
matrices to represent arbitrary topologies, GDS is we can write a general r-neighborhood CA
essentially a grammar by which sub matrices of value-transition rule ‘f’ (which will from now on
the whole matrix are rewritten to describe topo- refer generically to as a s-rule) in the form
logical changes. SDCA also formally falls under hn o i
the broader rubrics of DNA and RG; however, stþ1
i ¼f r ðiÞ ,
stj j j S G (3)
there is no explicit reference to SDCA in the
current literature of either field.)
where S G r ðiÞ ¼ jj Dij r is the radius-r graph
sphere about the site i. In words, the value of stþ1 i
is some function, f, of the values stj in radius
The Basic Model r graph sphere around the site i. With this distance
measure, G becomes a discrete metric space.
Conventional CA are defined on fixed, and typically If G is a one-dimensional line, and r = 1, then
regular, lattices (one-dimensional lines, two- SGr ðiÞ ¼ fi 1,i, i þ 1g ; i.e., it is equal to the
dimensional Euclidean or hexagonal grids, etc.), conventional three site local neighborhood of
the sites of which are populated with discrete-valued elementary CA.
dynamic elements ( si {0, 1, . . ., k}, where We now formally extend a conventional CA’s
i labels a particular site on the lattice) that evolve dynamic arena – limited to the values
0
according to local transition functions, f : si ! si . sti f0,1, . . . , k 1g , i = 1,. . .,N – to one that
We emphasize that the dynamics of conventional includes the components of theunderlying lat-
CA are confined to the temporal evolution of the sis. tice’s adjacency matrix:
SDCA generalize conventional CA in two ways:
(1) they relax the assumption that the underlying
stþ1 ¼ F s ½fst g, f‘t g
lattice is uniform, allowing the local site $ site , (4)
‘tþ1 ¼ F ‘ ½fst g, f‘t g
connectivity pattern to vary throughout the lattice;
and (2) they allow both the set {si} and the lattice to
where Fs and F‘ are some functions (to be defined
evolve according to local transition rules. The most
explicitly below) that explicitly couple the chang-
obvious – also the most dramatic – conceptual
ing value states and geometries. The complete
change this entails over the dynamics of conven-
system at time t is specified by the state-vector
tional CA, is that the meaning of “local” itself
changes as a function of how the SDCA system n o
evolves: previously far separated sites may become jGit ¼ st1 , . . . , stN ; ‘tij : (5)
neighbors; and sites that are local at time t may
become far separated at some later time, t0. The time-evolution of |Gi proceeds according
To properly define SDCA, we first generalize to the following transition rules: (i) s-rules of the
regular lattices to mathematical graphs G () general form given above and familiar from CA
possessing arbitrary topology. Assuming G has simulations and (ii) ‘-rules, which are divided into
N lattice sites, and that G is (for now) an undi- site couplers, linking previously unconnected ver-
rected graph (meaning that none of G’s links carry tices and site decouplers, which disconnect linked
directional information), G is completely defined points. Because the topology can be altered only
by the N-by-N adjacency matrix, ‘ij: by either a deletion of existing links or an addition
of links between pairs of vertices ‘i’ and ‘j’ with
1 if i and j are linked; Dij = 2, the dynamics is strictly local.
‘ij ¼ (1)
0 otherwise: To be more precise, we first restrict the general
s-rule F1 to (maximally symmetric) totalistic (T)
Using the graph metric function, and outer-totalistic (OT) type. Since the underlying
Structurally Dynamic Cellular Automata 33
lattice is a fully dynamic object, |Gi will, in general, where we distinguish the operator fb i acting on the
tend towards having a complex local geometry with global value state from the actual local transition
an unspecified local directionality. The most general function f which transforms each site value.
rules which can therefore be applied are those which
are completely invariant under all rotation and
Link Rules
reflection symmetry transformations on local neigh-
Local geometry altering rules are constructed by
borhoods. T(OT) s-rules are then specified by list-
direct analogy: for any two selected sites i and j we
ing particular sums{a}(outer-sums{a0},{a1}
restrict attention to site values of vertices
corresponding to center site values ‘0’ and ‘1’
contained within a 1-sphere of either site; that is,
respectively) for which the value of the center
to all k S1(i, j) = S1(i) [ S1(j). Link operators,
sitebecomes ‘1’. Formally,
whose action on the state is represented by:
!
X E E
si ¼ ffag
tþ1
‘ij sj , si ,
t t t
(6) b ij ‘t ¼ ‘t ,...,‘tþ1 ¼ cij ,...,‘t
decouplers : c fbg ij 11 ij NN
E E
j
where b feijg ‘tij ¼ ‘t11 ,...,‘tþ1
couplers : o ij ¼ oij
,...,‘t
NN ,
(9)
ffag ðx,aÞ
P
P a dðx þ a,aÞ P $T either link or unlink two sites ‘i’ and ‘j’ depending
¼
a a1 dðx,a1 Þ þ ð1 aÞ a0 dðx, a0 Þ $ OT, on whether the actual sum of values in S1(i, j)
(7) matches any of those given in the {b} or {e}
lists, which completely define decouplers and
P
and d(x, y) is the Kronecker delta. Note that j ‘tij couplers, respectively.
stj sums the values of all sites ‘j’ linked to ‘i’ at time In order to construct classes of rules analogous
‘t’. The action on the state |Gi is represented by to the two types of s-rules defined above, we
partition the local neighborhood into 3 disjoint
b i jsi
f sets (see Fig. 1): S1(i, j) = Vij [ Aij [ Bij, where
fag t
X 8
t < _ij ¼ fi,jg,
¼ s1 ,...,stþ1
i ¼ ff ag ‘tij stj ,sti ,...,stN ,
Aij ¼ fkjk C 1 ðiÞ\C 1 ðjÞg,where C 1 ðiÞ ¼ S 1 ðiÞ fig,
(8) :
Bij ¼ S 1 ðiÞ[S 1 ðjÞ_ij Aij :
(10)
Structurally Dynamic Cellular Automata, belonging to one of the two sites but not both ( Bij). In
Fig. 1 Neighborhood partitioning. In the same way as this way we obtain the analogous totalistic (T), outer-
outer sites can be considered separately for s-transitions, totalistic (OT), and an additional type called restricted
we may, for topology transitions, distinguish between totalistic (RT)
those sites belonging to both i and j( Aij) and those
34 Structurally Dynamic Cellular Automata
The action of link operators is then conve- • RT rules are completely specified by giving the
niently expressed as a function of the sums within ‘k’ 3-tuples of values (xsi + sj, y = sum in A,
the individual partitions. Defining nij = si + sj, aij z = sum in B), for which the link operation
P P
¼ k Aij sk , and bij ¼ k Bij sk , we get between ‘i’ and ‘j’ is to be performed. For exam-
decouplers, cijfbg ¼ cijfbg nij , aij , bij , where ple, define ‘c’ by unlinking ‘i’ and ‘j’ for the
following values of partitioned sums: (0, 0, 1),
(0, 0, 2), (0, 1, 1), (1, 1, 1); we then have that
cijfbg ðx,y,zÞ (b1,1 = 0, b2,1 = 0, b3,1 = 1), (b1,2 = 0, b2,2 = 0,
8 P
>
> 1 k dðx þ y þ z,bk Þ ‘ij $T b3,2 = 2), (b1,3 = 0, b2,3 = 1, b3,3 = 1), and
<n P o
¼ n1 k dx, b1,k dy þ z, b2,k ‘ij o $ OT (b1,4 = 0, b2,4 = 1, b3,4 = 1).
>
: 1 P d x, b
>
‘ij $ RT,
k 1,k d y, b2,k d z, b3,k
oijfeg ðx,y,zÞ 8
< Fbfag jsi ¼ i fbfag jsi,
8 P > Q i
< d Dij ,2 Pk dðx þ y þ z, ek Þ $T
Q
¼ d Dij ,2 Pk d x, e1,k d y þ z, e2,k $ OT Ccfbg j‘i ¼ bi
nij cfbg j‘i,
(13)
:
d Dij ,2 $ RT: >
:c Q
k d x, e1,k d y, e2,k d z, e3,k O feg j‘i ¼ nnij obfei g j‘i,
(12)
Structurally Dynamic
Cellular Automata,
Fig. 2 Sample dynamic
update of a (5 5) lattice
from t = 1 to t = 2, obeying
a T-type s-rule with s ! s0
for local sums =1,3,5
(i.e. a {1, 3, 5}), and
OT-type ‘-rules: (i) link for
{e1,1 = 1, e2,1 = 3} and
(ii) unlink for
{b1,1 = 1, b2,1 = 3} and
{b1,2 = 1, b2,2 = 4}. Solid
sites indicate that s = 1
8
ðvalueÞ > Ffag : fagT ¼ fa81¼
1,a2 ¼ 3,a3 ¼ 5g, 9 ‘t¼2 ¼ þ st¼1
st¼1 h , sb þ sd þ sg
t¼1 t¼1 t¼1
>
> ch c c
>
< < b1,1 ¼ 1,b2,1 ¼ 3 =
ðtopologyÞ Cfbg : fbgOT ¼ , þst¼1
i ; þsm ‘ch
t¼1 t¼1
>
> : b ¼ 1,b ¼ 4 ;
>
>
1,2 2,2
¼ cð1,3Þð1Þ ¼ 0:
:
Ofeg : fegOT ¼ e1,1 ¼ 1,e2,1 ¼ 3 :
(17)
(15)
We evolve the system by systematically 3. All next-nearest neighbors ‘i’ and ‘j’: . . .
sweeping through all sites, linked pairs, and linking them only if the 2-tuple
next-nearest neighbors: (a, b) = {(1, 3)}. By “next-nearest neighbor”
we mean those pairs which are themselves
1. All Sites: . . . setting si = 1 only at those ‘i’ for unlinked but which share at least one other
which the sum of the values at ‘i’ and its linked neighbor: (a, g), (h, r) and (w, y), for
neighbors is equal to ‘2’ at t = 1. By “neigh- example, are all next-nearest neighbors at
bors” of any point ‘i’ we will always mean the t = 1. For ‘c’ and ‘g’ we find
set of vertices linked to ‘i’: (a, b), (h, m) and
‘t¼2
cg ¼ o sc
t¼1
þ st¼1
g , sb þ sd þ sf
t¼1 t¼1 t¼1
(x, y), for example, are all neighbors at t = 1.
Writing out a few value-changing terms explic- h þ sl
þ st¼1 t¼1
d Dcg ,2
itly, we find that ¼ oð1,3Þð1Þ ¼ 1:
(18)
st¼2 ¼ f st¼1
b þ sc
t¼1
þ st¼1
d þ sh
t¼1
Notice that although ‘t¼1
dn ¼ 0 ! ‘dn ¼ 1, it is
c t¼2
of the values of the points that are neighbors of c[f] = 42, c[c] = 23(3)+1 = 1024 and
one of the considered points, but not of the other c[o] = 23(4)+1 + 23(3)+1 = 9216. Note that ‘c’
(neighborhood B in Fig. 1). The final state |Git = 2 and ‘O’ are chosen always to be of the
emerges after the above process has been applied same type.
concurrently to all pairs, neighbors and next- Comment 3. Computer simulations of these sys-
nearest neighbors in |Git = 1. tems require that some measures be taken to
prevent possible memory overflows, such as
Comments would happen in cases either of pure coupling,
We conclude this section by making a few impor- where links are continually added and none
tant general comments: deleted, or in isolated regions of a graph
where for a few sites more neighbors are
Comment 1. As defined above, G consists of added than are allowed by memory. We thus
three operators acting simultaneously on the introduce working link transition rules
state |Gi. More generally, one may prescribe
any of 10 possible time-orderings to the opera- 0 0
cij $ d i or d j > d d min
c~
ij
tors O,C and F. That is, specify certain inter- (20)
1 $ else,
mediate state dependencies, so that, for example
G1|Gi (OC)(F|Gi) would in general be 0 0
oij $ d i or d j < D d max
expected to yield results different from, say, o~ij (21)
0 $ else,
G2|Gi O(F(C|Gi)). While we will be solely
concerned with the synchronous time ordering where di = degree(i) (i.e. number of neighbors of
defined above, we do not expect the qualitative i). In words: make a sweep of the lattice, tem-
results to depend critically on this choice. porarily storing the candidates to add and
Comment 2. A given rule G is completely defined delete for each point. If, for any point i, the
by the set of sums {a},{b} and {e}. Alterna- updated degree is greater than d then proceed
tively, we can conveniently summarize a cho- with deleting the stored deletion-candidates,
sen transition rule by its vector-code otherwise do not delete; similarly, provided
!
C ¼ ðc½f, c½c, c½oÞa,b , where that the updated degree is less than D proceed
with addition. Thus, it is sufficient that one of
(P two points allow a dynamic link change
a2
a
$T between them for that change to be enacted.
c½f ¼ P P In the following, the complete constrained
a0 2
2a0
þ a1 2ð2a1 þ1Þ $ OT !½d,D
8P dynamics will be quoted as C ða,bÞ . If con-
>
> k2
bk
$T
>
<P straints play no role in the actual evolution of
3b2,k þb1,k
c½ c ¼ k2 $ OT specific examples, they will be left out of the
>
>
> P 3ðb þab Þþb
: definition.
2,k 3,k 1,k
k2 $ RT Comment 4. Because each dynamic update
8 P
> k2
ek
$T involves three separate types of processing,
>
<P
3e2,k þe1,k
the number of possible rules is extraordinarily
c½o ¼ k2 $ OT large (see Table 1). Unlike pure s-transitions,
>
>
: P 3ðe2,k þbe3,k Þþe1,k however, the fraction of the total number
k2 $ RT
which yield interesting behavior (i.e. neither
(19)
immediately explosive, where the number of
where a = max {b2,k} + 1, b = max {e2,k} + 1, links increases without bound, nor immedi-
and must be specified only for RT-type topol- ately degenerative, where an initial graph rap-
ogy rules. The G appearing in the above exam- idly dwindles to a few isolated links) appears to
ple, therefore, can be summarized by be manageably smaller.
Structurally Dynamic Cellular Automata 37
Structurally Dynamic Cellular Automata, represent sites with s = 1. Notice how the link
Table 1 Numbers of possible rules for each of the three additions follow the emerging corrugated bound-
types of transition rules. d= maximum allowable degree
and a= maximum sum to be used from partition Aij. Exam- ary surface of the value configuration. Remember
ple: for d = 5, we have Nf = 4096, Nc = 224 2 107 and that link additions are more than passive markers
No = 221 2 106. We thus have NT = NfNcNo 1017 indicating particular correlations between local
possible type OT Gs value configurations and structure; their presence
Rule type f c o directly influences all subsequent value develop-
T 2d+1 22d 22d1 ment in their immediate vicinity.
OT 22d+2 26(d1) 23(2d3) Figure 4 (in which site values are suppressed for
RT – 23(a+1)(2d1) 23(a+1)(2d+1) clarity) shows the continued development of this
system. Though boundary effects begin to appear
by t = 25, thecharacteristic manner in which this
Comment 5. Although it is the intrinsic geometri- particular G restructures the initial graph is clear:
cal patterning whose generic behavioral prop-
erties we are trying to deduce, one may • There is a high degree of geometrical organi-
approach SDCA from an alternative point of zation (the symmetry of the initial state is triv-
view: maintain the emphasis on unraveling the ially preserved by the totally symmetric G).
value configurational behavior, and interpret • The lattice remains connected.
the presence of [C, O] as background opera- • The distribution of link changes made through-
tors inducing nonlocal spatial connectivities. out the lattice remains fairly uniform i.e. there
Whereas the systems defined above are is an approximate uniformity in the probability
completely abstract entities, in that locality is of appearance of particular local value states
strictly defined by the link structure, the alter- which induce a structural change.
native scheme would be to embed the discrete • Link-lengths do not get arbitrarily large.
networks in some specified manifold, and to
study the effects of dynamically allocated non-
local communication channels. The last point implies that for a system embed-
ded in the plane, communication channels remain
approximately local. The global pattern emerges as
Emerging Patterns and Behaviors a consequence of local ordering. On the other hand,
Gs for which link-lengths get arbitrarily large are
Consider patterns that emerge from simple value also easy to find.
seeds starting from ordered two dimensional Euclid- Some other varieties of behavior are shown in
ean lattices. A single non-zero site may represent a Figs. 5 and 6. Figure 5a, b are representative of the
small local disturbance that then propagates out- class of ‘-rules that only mildly perturb the under-
ward, restructuring the lattice. With appropriately lying lattice (and for which s states do not differ
chosen Gs one can induce a rich spectrum of differ- much from their conventional CA cousins). Other
ent time evolutions only slightly perturbed by very rules, of course, may have a stronger effect on the
few concurrent link changes to ones in which the lattice, giving rise to associated s states bearing little
initial geometry becomes radically altered. (The or no resemblance to their conventional CA
graphical representation of evolving one dimen- counterparts.
sional systems, in which link additions must be Figure 5c shows an example of a link rule that
shown as arcs to avoid overlap with existing links, accelerates the outward propagation of the value
is needlessly confusing and is not considered.) configuration. Compare the diameter of this pat-
Figure 3 shows the first five iterations of a tern to that in the earlier figures, both shown at
system starting from a four neighbor lattice with equal times. The outwardly oriented links that
a single non-zero site at its center, the link struc- emerge from sites along the boundary surface
ture is given explicitly and the solid circles become conduits by which non-zero values
38 Structurally Dynamic Cellular Automata
!
Structurally Dynamic Cellular Automata, Fig. 3 First RT ‘-rules: C ¼ ð26,69648,32904Þ½3,3 (see text for rule
five iterations of an SDCA system starting from a 4-neighbor
definitions and code). Solid sites have s = 1
Euclidean lattice seeded with a single non-zero site at the
center. The global transition rule G consists of T s-rule and
rapidly propagate. Had the underlying of the initial disturbance. Figure 6c, on the other
lattice topology been suppressed in this figure, hand, shows a typical state of a system whose
and attention focused exclusively on the devel- global connectivity becomes progressively more
oping s state, we could have interpreted the complicated.
result as showing an effective increase in infor- A typical evolution starting from an initial state
mation propagation speed due to non-local in which all sites are randomly assigneds = 1 with
connectivities (see comment 5 of the previous probability p = 1/2 is shown in Fig. 7. Notice the
section). rapid development of complex local connectivity
Figure 5d, on the other hand, gives an example patterns, the appearance of which points to a geo-
in which the link dynamics lags behind the s metrical self-organization.
development. The boundary proceeds outward In general, structural behaviors emerging from
essentially unaffected by changes in geometry, random s-states under typical Gs can be grouped
which are themselves confined to the interior into four basic classes (not to be confused with
parts of the lattice (at least at this early stage of Wolfram’s classification of elementary CA
this system’s development). (Wolfram 1984)):
Figure 6 shows snapshot views of a few system
undergoing a slightly more complex evolution. • Class-1, in which initial graphs decay into
Figure 6b, for example, shows a rule in which structurally much simpler final states: most
the outward s propagation rapidly deletes most links are destroyed, and graphs ‘tij , for suffi-
links from the original lattice but leaves a complex ciently large t, consist essentially of a large
(though structurally stable) geometry at the origin number of small local subgraphs.
Structurally Dynamic Cellular Automata 39
Structurally Dynamic Cellular Automata, have been suppressed for clarity. The boundaries of the
Fig. 4 Several further time frames in the structural evolution original lattice do not extend beyond the region shown so
of the same system shown in the preceding figure. The values that the development is strictly confined to a 31 31 graph
• Class-2, whose final states are characterized by values of both hdegi and Deffec increase without
periodic but globally connected geometries. bound for class-3 SDCA (unless an arbitrary
SDCA typically arise in this class either because upper constraint D is imposed on G).
of a specific class-2 Fs remaining unchanged by Because the s-density responds to the
the coupling to the lattice or class-3 Fs coupling changing local neighborhood structure, it is
with {C, O} in such a way as to induce a lattice possible that what at first appears to be an
structure that supports a periodic state. explosive growth in fact eventually leads to a
• Class-3, consisting of SDCA that tend to grow more sedate, if not static, behavior at some
in size and complexity, at least as measured by larger hdegi D. Fs that yield hsit constant
two basic metrics: the average degree, hdegi over a range of hdegi (such as the sum
(1/N) . i[|S1(i)| 1], and effective dimensional- modulo-2 rule; see below), when coupled
ity, Deffec hNnni/hdegi, where hNnni is the with link rules that themselves become pro-
average number of next-nearest neighbors. The gressively less active with increasinghdegi,
40 Structurally Dynamic Cellular Automata
Structurally Dynamic Cellular Automata, coupler c[o] = 32, b = 2; c OT c[f] = 1022 and RT
Fig. 5 Snapshot views of four typical developing states coupler c[o] = 8, b = 1; dT s- and OT ‘-rules
starting from a single non-zero site at the center of a !
C ¼ ð682,19634061312,133120Þ½2,8
4-neighbor graph. Gs are as follows: a OTc[f] = 1022
and RT coupler c[o] = 16, b = 1; bTc[f] = 22 and RT
may induce evolutions leading to only mild continually changes, the average ratio of the
changes within specific ranges of the local number of next-nearest to nearest neighbors
structural parameters. stays approximately constant over long periods
• Class-4, which is a provisional class (pending of time. While there is evidence to suggest this
stronger evidence) that denotes a set of rules class is real, simulations have unfortunately
that yield open-ended s- and ‘ changes, but been run for too short a time and on graphs
during which the value of Deffec remains containing too few sites to permit making any
roughly constant. Cs and Os belonging to this conclusive statements regarding the veracity of
class effectively induce a structural equilib- this class. Nonetheless, it is tempting to spec-
rium: despite the fact that large numbers of ulate that, for arbitrary values of D , there
link changes continue to be made, so that the exists at least one set of SDCA rules for
detailed structure of the evolving graph which Deffec D (within a desired ϵ > 0) as
Structurally Dynamic Cellular Automata 41
!
Structurally Dynamic Cellular Automata, Fig. 6 Four bT s- and
!
OT ‘-rules C ¼ ð42,589952,8192Þ½2,8; cT s- and
more examples of states emerging from simple seeds. ‘-rules C ¼ ð42,128,4Þ½0,10 ; dTc[f] = 682 and RT ‘-rules
Figure a, b, c start from 4-neighbor graphs and d from an defined explicitly by C(104),(114),(124),(103),(113),(123) and
8-neighbor graph ( 4-neighbor with diagonals). Gs are as V(111),(215)
!
follows: aT s- and RT ‘-rules C ¼ ð42,69648,32904Þ½3,3 ;
the size of the graph N ! 1. (Pseudo class-4 random state is even more difficult (although
behavior, of course, can always be artificially graph visualization algorithms may help). How-
induced either by imposing severe [d, D = d] ever, even in cases for which a direct visual
constraints, or, as must typically be done for inspection of the dynamics reveals little, one can
category-3 Gs, by deliberately impeding always indirectly keep abreast of a given system’s
growth with some threshold D.) properties by monitoring its core structural and
behavioral measures (a more detailed account is
given in Ilachinski (1988)).
Statistical Measures
Site value measures include the average den-
As evidenced by Fig. 7, it is already nontrivial to P
meaningfully visualize the short-time evolution of 1=N Þ ENi¼1
sity of sites with value s = 1, hsit ðD
(initially) regular lattices that start with random sti ; the local value correlation, C t sti stj
D E
initial value state. Visualizing the long-term ðrt Þ2 , where sti stj is averaged over all pairs
dynamics of systems that start from a completely
42 Structurally Dynamic Cellular Automata
Structurally Dynamic Cellular Automata, constraints are [d = 0, D = 10]. The appearance of localized
Fig. 7 Evolution of a 35 35 lattice, with randomly substructures is evidence of a geometrical self-organization
seeded sites. The development proceeds according to T s-
!
and OT ‘-rules defined by code C ¼ ð84,36864,2048Þ. The
i and j with ‘ij = 1; the fraction of sites whose that describe all complex networks – such as
value changes during
P of the evolution, Dt
one step connectivity, density, clustering, and path lengths
ð1=N Þ Ni¼1 st1 i 2 i ; where
s t
2 is a sum (2,6), are applicable to SDCA as well.
modulo-2. Link changes may be monitored by keeping
Geometry measures include the average track of (1) the total number of link changes
degree, hdegi; the average number of next-nearest (allowed under prescribednconstraintoconditions),
ðl Þ P PN
neighbors, hNnnit (1/N)i[|S2(i)| |S1(i)| 1]; Dt ð1=2Þ Ni¼1 t t1
j¼1 l ij 2 l ij ; (2) the
and Deffec. A measure of how the actual size of ðl Þ ðl Þ
constraint influence, f l Dt =N t , where N t
ðl Þ
local neighborhoods changes with time may be is the total number of link changes that would
obtained by embedding graphs into the two-
have occurred in the absence of constraints
dimensional plane and calculating the average (fl = 1 indicates that the evolution is pure, mean-
path length at time t. Of course, global features
ing it is unaffected by constraints; flsmall
Structurally Dynamic Cellular Automata 43
P P n t¼0 L n o
suggests that the imposed constraint window for alteration are selected; and (5) the link evolu-
[d, D] has resulted in observed structures that are tion index, gnL 1=N t¼0 l i l
j ij 2 l ij ,
impure); (3) the link creation- and link deletion- which gives the fraction of the initial lattice
ðl Þ ðl Þ
ratios, f C N C =Dt and f D N D =Dt , where remaining after n iterations.
NC and ND are the numbers of link created Figure 8 shows time series plots of Deffec for
and destroyed, respectively; (4) the activity levels, rules in each of the four behavioral classes defined
gtC N C =N t1
nn and gD N D =N l
t t1
(where N t1
l is above. The initial structure in each case is 35 35
the number of links at time t 1), which give the 4-neighbor Euclidean lattice, so that Dt¼0 effec 2 .
number of dynamic alterations relative to the Figure 8a gives an example of class-1 behavior, in
corresponding spaces from which the candidates which a short period of initial growth is followed
44 Structurally Dynamic Cellular Automata
by a decay into mostly disconnected clusters. The 1. Static state: this trivially occurs when the link
final state is characterized by hdegi < 1, and is rules are unable to take effect; namely, when
stable. Figure 8b shows a system that starts from m 7 and n 8.
the same initial state as in Fig. 8a but whose G 2. Rapid growth: for an entire range of m and n, the
leads to a periodic geometry. Just the right number average number of neighbors for each site of the
of links have been deleted to permit regions with lattice increases rapidly for 20–30 iterations.
isolated activity to emerge.
Figure 8c shows class-3 behavior in which This number would likely continue to increase,
Deffec steadily increases. The apparent leveling off were it not for the constraint conditions
seen toward the end of the run is due both to a ( [0, 10]). The “final state” is neither stable nor
decreased overall activity level and the increas- periodic. One sometimes also sees delayed growth
ingly effect of the D = 10 constraint. The system in this class of behavior, in which case the link
in Fig. 8d exhibits class-4 behavior, characterized structure is initially relatively quiescent (and the
by a ongoing structural development within a rel- behavior of the system as a whole mimics that of a
atively narrow interval of values of Deffec. Note that conventional CA). As coupler rules are triggered
the structural changes here are essentially pure, and by specific s states, the average degree of the
are not merely artifacts of any imposed constraints. lattice rapidly increases (at least until the con-
Ilachinski (3) explores a wide range of emergent straint conditions take effect).
behaviors across all four classes, and examines the
qualitative relationship between emergent behavior 3. Spontaneous decay: when decouplers are
and initial s- and ‘-seeding. stronger than couplers, the average degree
typically decreases. If this occurs too rapidly,
Phase Plots the structure surrounding the single nonzero
While it is of obvious interest to systematically valued site may become isolated from other
explore every possible combination of b’s and ϵ’s parts of the lattice. If a few non-zero values do
that define ‘-rules, Table 1 unfortunately suggests not leak out into the outlying regions, link
that the resulting rule space is simply too large. changes remain confined to the central sub-
Nonetheless, we can learn much even by focusing graph, leading to either rapid stability or
our attention on a small subset of the complete periodicity.
rule space, keeping F, the initial s-seeding, and 4. Initial growth, followed by periodicity: this is the
all other factors constant. Specifically, consider least common behavior, and requires a delicate
the subset of all possible ‘-rules that consists of balance between coupler and decoupler rules.
OT link rules consisting of a single coupler, o,
L It is interesting to compare these results with
and a single decoupler, c. Moreover, let f 2
(i.e., sum modulo-2 rule), demand that only pairs those obtained from a random s seed. In this case,
of s = 0 sites be considered for a link change, and the sharp divisions between characteristic behav-
consider ‘-rules belonging to the following set: iors disappear, and there is a pronounced increase
in the number of links for all m and n. However,
n o the inclusion of an additional decoupler, may
decouplers : b1,1 ¼ 0, b2,1 ¼ m , 1 m 0, induce decay and periodicity. For example, con-
couplers : ϵ 1,1 ¼ 0, ϵ 2,1 ¼ n , 1 n 0: sider the same initial lattice and F as used in
(22) Fig. 9, fix two OT ‘-rules C(0, 5) and O(0, 1), and
add the decoupler C(0, m) : {b1,1 = 0, b2,1 = m},
Figure 9 summarizes the behavior of a four 1 m 9. Surveying the emergent behaviors for
neighbor, 25 25 lattice with periodic boundary this range of m’s, one now finds decaying lattices
conditions, starting from an initial s-seed for m 2. In each case, the initial graph succumbs
consisting of a single nonzero site. Four basic to periodicity following a transient of between
kinds of structural behaviors emerge: 50 and 100 iterations. The evolving lattice is
Structurally Dynamic Cellular Automata 45
Structurally Dynamic Cellular Automata, couplers–{ϵ 1,1 = 0, ϵ 2,1 = n}. Grey areas in both plots
Figure 9 Phase plot that summarizes behavior of a four denote periodic states. White areas denote growth in the
neighbor, 25 25 lattice with periodic boundary condi- plot for link behavior, and a nonperiodic state for
tions, starting from an initial s-seed consisting of a single s-behavior. The black area that appears in the link-
nonzero site. G is defined by the sum modulo-2 s-rule and bevavior plot denotes decay. Numbers that appear in indi-
‘-rules of the form: decouplers–{b1,1 = 0, b2,1 = m}, vidual boxes denote period lengths
also more prone to break up into small discon- for describing physical processes (which is the
nected subgraphs. primary reason for which SDCA0 were first con-
Although, just as in conventional CA, small ceived). Motivated primarily by finding models of
changes to ‘-rules can lead to large differences in human brain function (for which one intuitively
emergent behavior, they generally appear to do so expects nonlocal neural connections to play a
in a more predictable and patterned manner. Of fundamental role in the rewiring of neural tissue),
course, particular classes of G may induce more Majercik shows that suitably generalized SDCA
complex phase plots; for example, isolated are not only capable of universal computation, but
pockets of anomalous (and rapidly shifting) actually represent a more efficient class of com-
behavior may appear within larger surrounding putational models than conventional
regions undergoing otherwise mutually consistent CA. Majercik also reports an SDCA that can
and slowly changing dynamics. A better sense of solve the firing squad problem in O(logt) time
the space of possible emergent behaviors, along (i.e., exponentially faster than the O(t) in conven-
with a deeper understanding of the relationship tional CA),and a class of CA-universal SDCA
between F and ‘-rules, awaits a future study. models that can simulate any conventional CA
with a speedup factor of two. (The firing squad
problem (Moore 1962) consists of finding a rule
SDCA as Models of Computation for which allsites in a CA evolve into a special
state after the exactly the same number of steps.)
The basic SDCA model, as outlined above (which Majercik proceeds by first identifying five
we will denote as SDCA0 to avoid possible con- properties of SDCA0 that, while reasonable from
fusion with the hierarchy of related SDCA models a physical modeling standpoint, make it difficult
introduced in this section), was modified and gen- to rigorously formulate and prove theorems:
eralized by Majercik (1994) into a form more
suitable for addressing its formal computational 1. Finiteness: The requirement that SDCA0 be
capabilities rather than as an exploratory toolkit strictly finite, both in time and space, is obviously
46 Structurally Dynamic Cellular Automata
necessary for computer experiments, but is locations remain unspecified. Finally, in MS, it is
unnecessarily restrictive for general theorem assumed that no information about the source of
proving. Likewise, the assumption that the sets the neighborhood states exists, and transition
a, b and ϵ must be finite is questioned. functions only know the number of neighbors in
2. Bidirectionality: While SDCA0 are defined a particular state.
with symmetric links, an obvious generaliza- Each of the three core models may be defined
tion that makes the basic model more readily in two versions: an unbounded links (abbreviated,
applicable to neural dynamics (among other UL) version, in which the number of neighbors a
kinds of physical and biological systems) is to given site can have is unbounded, and a bounded
allow for unidirectional links. links (abbreviated, BL) version, in which an
3. Link-rule Asymmetry: While SDCA0’s link explicit upper limit is imposed. In addition, there
decoupler function (Eq. (11)) contains the fac- is also one finite labels version of ML. Majercik
tor ‘ij to explicitly prevent the system from imposes certain mild conditions on the local tran-
inadvertently linking two unlinked sites, sition functions; for example, that local neighbor-
SDCA0 does not include an analogous term hoods always remain strictly finite, s-rules leave
for the coupler function (that is, a term to quiescent neighborhoods alone, and that links
prevent an evolving system from inadvertently between sites with quiescent neighborhood
unlinking two linked sites). remain unaltered.
4. Inconsistency: While s-rules effectively ignore
site positions, all three types of link rules Relative Location SDCA Model
assume that the various neighborhoods In the Relative Location model, the transition
surroundings individual sites (Aij, Bij, and functions all have access to the exact relative
Cij {k| Dik = 1 Djk = 1}, where ‘ ’ location and state of each neighbor site. Define a
denotes exclusive or) are all recognized as neighbor of site i, ni S Zd, as a pair that
such by the dynamics. That is, the link rules specifies the state (by a single label) and relative
effectively “ know” the positions of a site’s location of the neighboring site (as a d-tuple of
neighbors, while s-rules possess no such coordinates). Let W = S Zd be the set of all
information. possible neighbors, and F W (called the neighbor-
5. Small Rule Set: The class of s- and ‘-rules used hood function) be the set of all possible finite,
by SDCA0 may be generalized to include a far nonempty, partial functions that map Zd to W.
broader class of transition functions. The local state transition function s : F ! W
maps neighborhood functions to the state set of
On the basis of these observations, Majercik SDCA. The local link transition function
(1994) introduces a set of three core models todefine l : F F {0, 1, 2} ! {0, 1} maps pairs of
a hierarchy of eight alternative SDCA computational neighborhood functions (that define the neighbor-
systems, {SDCA(1), SDCA(2), . . ., SDCA(8)}. The hoods of two sites, i and j and a number that
three core models are (1) the relative location specifies the status of the link between i and j:
model (=MR), (2) the labeled links model value zero meaning that i and j are neither direct
(=ML), and (3) the symmetric links model neighbors nor next-nearest neighbors; value one
(=MS). They differ only in the degree to which meaning that i and j are immediate neighbors; and
their s- and ‘-transition functions depend on spe- value two meaning i and j are next-nearest neigh-
cific sites. For example, MR’s transition functions bors) to one of two link states: zero, meaning no
depend on the state and exact relative position of link between i and j, and one, meaning a link exists.
each neighbor (and therefore “knows” the exact
source of any state in a local neighborhood). In Labeled Links SDCA model
ML, links are labeled and the transition functions The Labeled Links model removes from MR’s
know both neighbor states and the label of the transition functions any dependency on the exact
links to given neighbors, but the exact neighbor relative location of a site’s neighbors, but allows
Structurally Dynamic Cellular Automata 47
to preserve the local topology of a simulated CA, it classes, C1 and C2, let C1sC2 denote the fact
is necessary to define separate encoding (=e) and that if, given any SDCA S1 C1 there exists an
decoding (=d) functions – e : SCA ! SSDCA trans- SDCA S2 C2 that simulates S1. Then, for exam-
forms the initial configuration of the CA systems to ple, since any BL SDCA can be simulated by an
configurations in the SDCA system being used to unbounded links version of the same system, and a
simulate it (where SCA and SSDCA are the configu- finite links version of ML can be simulated by
rations spaces of CA and SDCA, respectively); and a bounded links version, we know immediately
d : SSDCA ! SCA effectively performs the inverse that SDCA(7)s SDCA(8), SDCA(4) s SDCA(5)
transformation. Encoding (and decoding) functions s SDCA(6), and SDCA(2) s SDCA(3). Similar
are called structurally defined if they are recursive reasoning (Majercik 1994) leads to the general
and use a finite amount of information to encode relationship:
(or decode) a given configuration; and are otherwise
expected to transform quiescent states to quiescent SDCAð3Þ s SDCAð8Þ s SDCAð6Þ ,and
(25)
states. Majercik further assumes that (1) e has access SDCAð2Þ s SDCAð7Þ s SDCAð5Þ :
to the rule table of the conventional CA system
being simulated; (2) d does not have access to the Finally, since the unbounded links version of MR
rule tables of either system; and (3) e and d must has all the information necessary to construct the
together satisfy the relation: e d = Identity(SCA). neighborhood partitions used by SDCA0, and since
Denoting the global transition functions of the SDCA(8) s SDCA(6), we see that SDCA0 s
CA and SDCA systems by FCA and FSDCA, respec- SDCA(6) and SDCA0 s SDCA(8).
tively, FSDCA is said to simulateFCA if there exist Majercik’s two main results, which we state
m 1,n 1 and structurally-defined functions without proof, are:
e : SCA ! SSDCA and d : SSDCA ! SCA, such
that for any configuration ! s SCA and any k 1, Majercik Theorem 1: Given an arbitrary
1-dimensional conventional CA with radius
! km !
CA s ¼ d FSDCA e s
Fkn : (24) r = 1, there exists an unbounded links version
of MR (=SDCA(8) of the SDCA hierarchy) that
If m > n then FSDCA simulates FCA with a can simulate it with a speedup factor of two.
slowdown factor of m/n. If m < n then FSDCA Majercik Theorem 2: There exists a 1-dimensional
simulates FCA with a speedup factor of n/m. finite links version of ML (=SDCA(4)) that can
simulate an arbitrary k-state 1-dimensional
conventional CA with radius r = 1 with a
SDCA Hierarchy of Models pffiffiffiffiffiffiffiffiffiffiffiffiffi
slowdown factor O k 2r 2rlogk .
Majercik (1994) uses the three generalized models
introduced above (MR,ML, and MR) to define a Detailed proofs of these two theorems appear in
hierarchy of eight SDCA models of computation. Majercik (1994) (where they are called Theorems
At the top of his hierarchy (arranged from top-to- 4.4 and 4.5, respectively). In Chap. 5 of his thesis
bottom in roughly, but not completely, decreasing (Majercik 1994), Majercik presents an explicit con-
order of computational strength; see discussion that struction of a CA-universal SDCA(4) computa-
follows) are the UL and BL versions of MR: tional model, and compares it to Albert and
SDCA(8) and SDCA(7), respectively; followed by Culik’s (1987) construction of a 1-dimensional
SDCA(6) = UL version of ML; SDCA(5) = BL CA-universal conventional CA that simulates any
version of ML; SDCA(4)= a finite labels version 1-dimensional, k-state, radius r CA with an O(k8r)
of ML; SDCA(3) = UL version of MS; SDCA(2) = slowdown. Although Majercik’s CA-universal
BL version of MS; and, sitting on the lowest level SDCA uses more states than Albert and Culik’s
(computationally speaking), is SDCA(1)= SDCA0. universal CA, it is also markedly faster.
A little thought suffices to establish certain rela- The reason why the SDCA is faster is at least
tionships among the various classes. Give two intuitively clear. An SDCA’s dynamic links
Structurally Dynamic Cellular Automata 49
may – depending on the fitness function – converge eventually levels off. In the first stage, the fittest
prematurely to a sub-optimal solution. To reduce possible communities are first established; in the
the likelihood of this happening, Muhlenbein intro- second stage, connections with poorer candidate
duces a spatial population structure; restricting fit- solution are deleted; finally, in the third stage, the
ness and mating to neighborhoods called demes. system essentially “fine-tunes” its optimal solu-
Demes are geographically separate subpopulations tions. Halpern (2003) finds two different evolution-
in which candidate solutions evolve along dispa- ary paths toward high connectivity: (1) monotonic
rate trajectories; though occasional mixing still growth over time (for low mutation rates, pm), and
occurs through the process of migration. (2) a phase transition between low and high
In Halpern’s variant (1996), an otherwise con- degrees of connectivity (for some pm ). Using
ventional GA is placed within the structure of SDCA$GA hybrid model parameters N = 100
SDCA0 (i.e., the basic model defined by and fD = fC = 0.1, pm 0:05, for which Halpern
Eqs. (11), (12), (13), and (14)). Heuristically, this (2003) finds a sharp increase in the number of links
allows each candidate solution to “choose a com- per site between generations 350 and 450.
munity” with which to mate, during each genera- Despite the novelty of the approach, and the
tion. The choice of neighborhoods thus becomes promising link between optimization rates and
an integral component of the GA, and is deter- dynamic structure established in Halpern (1996),
mined dynamically by the evolving solutions. concrete applications of the algorithm – except for
Halpern’s algorithm proceeds as follows (2003): Weinert et al. (2002) work on a related hybrid GA
(Step 1) an initially random lattice (defined by adja- algorithm – have yet to be developed. One sugges-
ðt¼0Þ
cency matrix l ij ) is seeded with single- tion, from Halpern (2003), is to use the SDCA$GA
chromosome candidate solutions of fixed length, hybrid model for finding “optimal” connectivity
P
one per site; (Step 2) a fitness function, f i ¼ Ni¼1 patterns in parallel computers. The search algorithm
dij , is defined to assign a numerical measure of may be used to directly model how component
“optimality” to each site (N is the number of sites, processors are connected, and decide to keep or
and dij is the value – equal to 0 or 1 – of the jth gene sever existing links, or establish new ones, adap-
of the ith chromosome; (Step 3) each site i ranks tively as a function of local fitness criteria.
each of its nearest and next-nearest neighbors
according to fi; (Step 4) each site disconnects with
a fraction, fD, of its least-fit neighbors, and connects Generalized SDCA Models
with a fraction, fC, of its fittest next-nearest neigh-
bors; (Step 5) each site randomly mates with one of Despite SDCA being obviously more “complex”
its nearest neighbors (i.e., the usual processes of than conventional CA (and certainly more com-
mutation and crossover operations are applied plex to formally define, if only because one must
(Mitchell 1998)); (Step 6) the least fit members of specify both s and ‘ rules), the SDCA model
the population are replaced by the offsprings from nonetheless has more in common with elementary
Step 5; and (Step 7) loop through steps 5–7, until CA than with any of its brethren’s more “compli-
some suitable “optimality” threshold (or some other cated” variants. By “elementary” CA we mean the
convergence criterion) is satisfied. simplest one-dimensional CA with s {0, 1}
Halpern (1996, 2003) reports a wide range of and local neighborhoods consisting only of left
resulting behaviors, collectively suggesting a clear and right (i.e.,nearest) neighbors. Just as there are
relationship between the parameters defining the many generalizations of elementary CA – for
GA optimization and lattice connectivity. Of par- example, increasing the state space to include s’s
ticular interest are the dynamic conditions for that take on one of N values, larger-sized neigh-
which the fitness-based creation and deletion of borhoods, and memory, among many other
links increases the rate of growth of overall fitness. possibilities – so too there are natural extensions
The fastest convergence occurs when lattice con- of basic SDCA. In this section we discuss three
nectivity first increases, then decreases, then generalizations: (1) rules that are reversible in
Structurally Dynamic Cellular Automata 51
time, (2) rules that retain a memory of past states, a two dimensional hexagonal lattice, and values
and (3) probabilistic rules. evolve according to the three-state (i.e.,
s {0, 1, 2}), next-nearest neighborhood T bee-
hive rule. The beehive rule is defined explicitly by
Reversible SDCA
assigning one of three values (0, 1, or 2), to
The first generalization of the basic SDCA model,
each possible 3-tuple, (N0, N1, N2), that gives
explored extensively by Alonso-Sanz (2006), is
the number of local sites with N0 0s, N1 1s, and
to apply the Fredkin reversible-rule construction
N2 2s (Alonso-Sanz 2006): (0, 0, 6) ! 0,
to ‘ rules to render them reversible in time. Con-
(0, 1, 5) ! 1, (0, 2, 4) ! 2, (0, 3, 3) ! 1,
i that is first-order
sider a conventionalh CA system
(0, 4, 2) ! 2, (0, 5, 1) ! 0, (0, 6, 0) ! 0,
in time, si ¼ f sj N i , where N i is the
tþ1 t
(1, 0, 5) ! 0, (1, 1, 4) ! 2, (1, 2, 3) ! 2,
neighborhood around site i and, generally, si (1, 3, 2) ! 2, (1, 4, 1) ! 1, (1, 5, 0) ! 1,
Z k. The Fredkin construction converts this system (2, 0, 4) ! 0, (2, 1, 3) ! 0, (2, 2, 2) ! 2,
into an explicitly invertible one that is second- (2, 3, 1) ! 2, (2, 4, 0) ! 0, (3, 0, 3) ! 0,
order in time by subtracting the value of the center (3, 1, 2) ! 2, (3, 2, 1) ! 2, (3, 3, 0) ! 0,
site at time t 1: (4, 0, 2) ! 0, (4, 1, 1) ! 0, (4, 2, 0) ! 2,
h i (5, 0, 1) ! 2, (5, 1, 0) ! 0, (6, 0, 0) ! 0.
stþ1 ¼ f stj N i k st1 The top row of Fig. 10 shows the first four
i i , (26)
steps (t = 1,2,3, and 4) in the memoryless evolu-
tion of the initial “ring” of sites that appears at
where ‘k’ is subtraction modulo-k. Since
t = 1. The link rules used for this run are those
Eq. (26)h can i be trivially solved for
defined in Eq. (28). Since the decoupler removes
st1
i ¼ f stj N i k stþ1
i , we see that any links between pairs of sites whose values are equal
pair of consecutive configurations uniquely spec- to zero, most of the lattice disappears after a single
ifies the backwards trajectory of the system. More- time step, and both value and link activity is
over, this is true for arbitrary (and, in particular, confined to a small region. After two more steps
irreversible) functions f. of changes,
n the o system
n o attains a fixed
quickly
Now, exactly the same procedure may be point: st , ‘tij ¼ st¼4 , ‘t¼4 ij for all t 5.
applied to link functions:
While the frequency of states is not constrained
8 n o to total six for a dynamic lattice, the beehive rule is
< l tþ1
ij ¼ c st , l t 2 l t1
ij , unchanged; if the sum of frequencies at a given
k n ij o (27)
: l tþ1 ¼ o st , l t 2 l t1 , site exceeds six, the site value remains the same.
ij k ij ij
The bottom row shows the evolution of the
Fredkin reversible versions of the rules defined
where 2 is subtraction modulo-2 (since links are
in Eq. (28) (to simplify the visualization, links
obviously binary valued).
along the border sites are not shown). In contrast
Following Alonso-Sanz (2006, 2007), we con-
to the basic SDCA version, the initial lattice in this
sider these two specific SDCA link rules (which
case does not decay. Since, according to Eq. (27)
will also be used in a later example):
(which assumes that ‘t¼0ij ¼ ‘t¼1
ij ), the initial hex-
8 agonal lattice is subtracted from the evolved struc-
< c sti ,stj ,l tij ¼ 0 iff l tij ¼ 1 and sti þ stj ¼ 0,
ture at t = 1 (modulo-2), the original graph is
: o st ,st , lt ¼ 1 iff l t ¼ 0,st > 0,st > 0,and Dij ¼ 2: effectively restored, and the outlying regions
i j ij ij i j
appear undisturbed.
(28)
Structurally Dynamic Cellular Automata, of these rules (obtained by applying Eq. (27) to Eq. (28)).
Fig. 10 Comparison between first few time steps of a a In both cases, s’s evolve according to the beehive rule
memoryless SDCA, evolving according to link rules defined in the text. (Reproduced with permission from
defined in Eq. (28), and b the Fredkin reversible versions Alonso-Sanz (2006))
and Martín (2006) and Alonso-Sanz (2006, 2007), function of a given site’s current value, si,
is to endow both s-rules and ‘-rules with memory. alone, but is instead a function of the transformed
The rules for conventional memoryless CA and value, s ¼ Mf ðs;m,aÞ, obtained from si’s past
SDCA, depend only on neighborhood configura- m values: fm : s ! s0, where 0 a 1 is a
tions that appear on the immediately preceding numerical memory factor. The value transforming
time step. Therefore, rules may be said to possess memory function, M , assumes the following
a “memory” of depth m if they depend explicitly specific form (to avoid confusion, note that in
on values (in the case of CA), or on both values Eqs. (29) and (30), sxi means the value of si at
and link states (in the case of SDCA), that existed timet = x, and ax means the numerical quantity a
on m previous time steps. We note, in passing, that raised to the power x):
since the Fredkin construction couples states at 8 t
times t + 1, t and t 1, reversibility may be < 1 if sbi Þm > 1=2,
>
considered a specific form of memory that extends sti ¼ M f sti ;m,a ¼ sti if sbit Þm ¼ 1=2,
>
:
backwards a single step. 0 if sbit Þm < 1=2,
Of course, there is no unique prescription for (29)
introducing a dependency on past values; and a where
P
variety of alternative memory mechanisms have sti þ m aDt stDt
sbi t
¼ Pm
Dt¼1 i
: (30)
been proposed in the literature (for example, see m 1 þ Dt¼1 aDt
page 43 in Ilachinski (2001) and page 118 in
Wolfram (1984)). We focus our discussion on At any given time, t, the depth m can never
the approach proposed by Alonso-Sanz (14), and exceed t 1. Our discussion follows Alonso-Sanz
for the moment confine our attention to value (2007), and sets m(t) t 1 for all t; i.e., we assume
rules, f : s ! s0. Alonso-Sanz’s approach is to that M f ðs;m,aÞ yields a weighted mean value of all
preserve the form of the transition rule, but have it the previous values of a given site. In practice,
act on an effective site value that is a weighted memory becomes active only after a certain number
function of its m prior values. of initialization steps, here taken to be three; with
This is done by introducing a memory- seeded values s1i ¼ s1i and s2i ¼ s2i .
endowed value rule, fm, that – in contrast to its Memory can be added to link rules in a similar
memoryless version, f – is not, in general, a manner. The form of the link rules (c and o)
Structurally Dynamic Cellular Automata 53
remains the same, but rather than acting on a evolves according to the parity T s-rule (that
graph that is defined by its adjacency matrix, ‘tij , assigns a value zero to a site if the sum of the
c and o instead act on the memory-transformed values in its neighborhood is even, and assigns the
values, L ¼ Mðc,oÞ ð‘;m,aÞ: value one if the sum is odd) and the ‘ rules defined
above in Eq. (28). The first row of evolving pat-
8 t
terns (for each a) applies memory only to values;
>
> 1 if ^l ij > 1=2,
>
< t m the second applies memory only to links, and the
Ltij ¼ M ðc,oÞ l tij ;m,a ¼ l tij if ^l ij ¼ 1=2, third appliers memory to both. Figure 13 shows
>
> t m
>
:0 the reversible beehive SDCA shown in Fig. 10,
if ^l ij < 1=2,
m but with full memory (a = 1.0).
(31)
Probabilistic SDCA
where
Another natural extension of the basic SDCA model
t P is to replace the set of explicit s- and/or ‘-rules with
l tij þ m Dt¼1 a
Dt
l tDt
^l ¼ P ij
: (32) probabilities. In this way one can study the evolution
ij
m 1þ m Dt¼1 a Dt
of a system that undergoes random but s-dependent
lattice changes. For example, this may be useful for
As for memory-endowed s-rules, the memory studying genetic networks in which new links are
for link rules is activated only on the third iteration forged (with a given probability) only if both genes
step, and the system is initialized by setting L1i are active, and existing connections are broken if
¼ s1i and L2i ¼ s2i . both sites are inactive.
Figures 11 and 12 show the effects of applying Following Halpern and Caltagirone (1990),
partial memory weighting (a = 0.6) and full mem- consider the parity T s-rule and the following
ory (a = 0.6), respectively, to a SDCA that starts probabilistic versions of decoupler (cp) and cou-
with a Euclidean four-neighbor lattice, and pler rules (op):
Structurally Dynamic Cellular Automata, The first row of evolving patterns applies memory only to
Fig. 11 Sample runs of a SDCA with memory for mem- values; the second row applies memory only to links, and
ory weighting a = 0.6. The SDCA is initialized as a the third row shows the evolution when memory is applied
Euclidean four-neighbor lattice, and evolves according to to both. (Reproduced by permission from Alonso-Sanz
the parity T s-rule and the two ‘ rules defined in Eq. (28). (2007))
54 Structurally Dynamic Cellular Automata
Structurally Dynamic Cellular Automata, Fig. 12 Sample runs of the same SDCA shown in Fig. 11, but with
memory weighting a = 1.0. (Reproduced by permission from Alonso-Sanz (2007))
Structurally Dynamic Cellular Automata, Fig. 13 Sample runs of a reversible beehive SDCA with full memory
(a = 1.0); compare to Fig. 10. (Reproduced by permission from Alonso-Sanz (2007))
8
< l tþ1 ¼ cp l tij ; sti , stj , pD , and (3) PD = 1 decoupler rule applied 100% of the
ij
ðdecoupler Þ : time (consistent with non-probabilistic SDCA
: c 1 d st þ st ,0 dðp > rÞ,
p i j D
8 rules). We see that changing PD induces qualita-
< l tþ1
ij ¼ op l ij ; si , sj , pD ,
t t t
tively different s behavior, that ranges from small
ðcouplerÞ :
: op d Dij ,2 d st þ st ,2 dðp > rÞ,
i j C
fluctuations around hsi 0.5 (for PD = 0), to
decay to small static values (hsi = 0.05 for
(33)
PD = 1/2, and hsi = 0.12 for PD = 1).
where PD and PC are the decoupler and coupler Halpern and Caltagirone (1990) have studied a
probabilities, respectively, and r is a random num- wide range of probabilistic SDCA, using random
ber between 0 and 1. initial s configurations, step-function, parity, and
Thus, cp unlinks two previously linked sites Conway’s life s-rules, Cartesian and random initial
with probability PD if and only if the sum of their lattice structures, and various probabilities
site values is zero; and op links two previously 0 PD 1 and 0 PC 1. Some of their results
unlinked sites with probability PC if and only if are reproduced (with permission) in the behavioral
phase plots shown in Fig. 15. (The step-function rule
they are next-nearest neighbors and the sum of P t t
their site values is two. is defined by stþ1 ¼ 0 if and only if j ‘ij si > 2
i
P
Figure 14 shows time series plots of hsi as a and stþ1
i ¼ 1 if and only if ‘ t t
s
j ij i 2; Conway’s
function of time for three different cases: life rule assigns st + 1 = 1 to a site if and only if
(1) PD = 0 (no decoupling at all), (2) PD = 1/2, s(t) = 0 and the sum of values in its neighborhood at
Structurally Dynamic Cellular Automata 55
Structurally Dynamic
Cellular Automata,
Fig. 14 Time series of
average s value, st, for the
Halpern-Caltagirone rules
(defined in Eq. (33)) and
for three values of
decoupler probability:
PD = 0, PD = 1/2, and
PD = 1. (Reproduced with
permission from Halpern
and Caltagirone (1990))
time t is equal to 3 or st = 1 and the sum of values is the influence of dynamic topology on the zero-
equal to 2 or 3; otherwise st + 1 = 0.) temperature limit of ferromagnetic transitions.
Figure 15 shows a wide range of possible
behaviors. Consider, for example, the number of Random Dynamics Approximation
links per site for the case where the lattice is For cases in which the structure and value config-
updated with probabilistic ‘-rules and the s’s are urations are both sufficiently random and
all random (shown at the top left of the figure). uncorrelated, a random dynamics approximation
Four distinct classes of behavior appear, with (abbreviated, RDA) may suffice to qualitatively
growth dominant for most values of PD and PC. predict how the system will tend to evolve under
Pure decoupling (or pure coupling) leads to com- a specific rule set; for example, to predict whether a
plete decay (or growth to a stable state); a mixed state given rule is more (or less) likely to yield
of coupling/decoupling generally yields slow unbounded growth, to eventually settle into a low
growth. Periodic behavior occurs only for periodic state, or to simply decay. The idea is to
PD PC 1. Compare this behavior with the approximate the real SDCA as a mean-field; that is,
cases where the s-rule is either the parity value rule assume all local value and structural correlations
(shown in the middle of the top row of Fig. 15) or the are close to zero (and can thus be ignored), and
step-function rule (shown at left bottom of the fig- replace all specific site values and local link geom-
ure). While the parity rule also displays four similar etries with average, or effective, values.
phases (growth to stability, decay to stability, incom- More precisely, assuming that (1) the probability
plete growth, and incomplete decay), decaying struc- pðnsi Þ of a site ‘i’ having value s = 1 at time t = n is the
tures eventually reach a stable (not null) final state. same for all sites – so that pðnsi Þ ¼ pðnsÞ for all i – and
The step-function rule shows an even greater variety ðlij Þ
of possible behaviors, and appears more sensitive to (2) that the probability pn of two sites ‘i’ and ‘j’
small changes in link probabilities. being linked at t = n is the same for all pairs of sites –
ðlij Þ
The probabilistic SDCA system discussed in so that pn ¼ pðnlÞ for all i and j – the RDA evolution
this section adds a stochastic element specifically equations may be written formally as follows:
to SDCA. Of course, there are other ways of (
ð sÞ
injecting stochasticity into a CA with dynamic pnþ1 ¼ F RDA pðnsÞ , pðnlÞ ; GSDCA ,
ðl Þ (34)
topology. For example, Makowiec (2004) com- pnþ1 ¼ GRDA pðnsÞ , pðnlÞ ; GSDCA ,
bines the deterministic evolution of a conven-
tional CA with an asynchronous stochastic where SDCA’s rule GSDCA (defined in Eq. (14)) is
evolution of its underlying lattice (patterned after included, formally, to remind us that the functional
the Barabasi and Albert (2002) model of degree forms assumed by FRDA and GRDA will be different
distributions in small-world networks), to explore for different GSDCAs.
56 Structurally Dynamic Cellular Automata
Structurally Dynamic Cellular Automata, and small fluctuations (around a stable lattice), and peri-
Fig. 15 Behavioral phase plots summarizing the long odicity. The initial graph is a Cartesian four-neighbor lat-
term evolution for several different s and ‘-rules defined tice in each case except for the top-right plot (labeled
in Eq. (33). The x and y axes for each plot depict values Random connections/links) for which the initial graph is
( {0, .25, .5, .75, 1}) of PC and PD, respectively. There random. (Reproduced with permission from Halpern and
are six classes of behavior: growth, decay, stability, large Caltagirone (1990))
The first function, FRDA, is the easier of the two any site has exactly d neighbors. Since this means
to calculate. For any given site with degree d we that, out of a total of N 1 possible neighbors, a
simply count the total number of ways to distribute given site must have exactly d links, and not be
the local s-values among the d possible neighbor- connected to any of the remaining (N 1 d)
ing sites to obtain the desired sums that define a sites, we have by inspection:
given rule. In this way we find the average 0 1h id
N 1 N 1d
expected s density at t = n + 1, assuming all sites P d; pðnlÞ ¼ @ d A pðnlÞ 1 pðnlÞ :
in the lattice have the same degree d at time t = n:
ðsÞ
(36)
ðsÞ
pnþ1
8 d,p
X d þ 1 h ia
n
>
>
dþ1a
To calculate the second function in Eq. (34)
>
> pðnsÞ 1 pðnsÞ $T
>
> fag a (= GRDA), we first define the local transition
>
>
>
< X d h ðsÞ ia0 þ1 da0 functions
¼ pn 1 pðnsÞ 8 a
> a0
>
> fa0 g > p ðd 1 , d 2 ,lÞ
>
>
> X
d þ1
h i
a1 dþ1a1 < ¼n Prob l ¼ 1 ! l 0 ¼ 0j d ¼ d , d ¼ d , A ¼ l ,
>
>
> þ pðnsÞ 1 pðnsÞ $ OT i 1 j 2 ij
>
: a1 > pbn ðd 1 , d 2 ,lÞ
fa1 g >
:
¼ Prob D ¼ 2 ! l0 ¼ 1j d i ¼ d 1 , d j ¼ d 2 , Aij ¼ l ,
(35)
(37)
X
We then get
ð sÞ
F RDA ! pnþ1 ¼ P d; pðnlÞ which give the probabilities that any two sites –
d i and j – will be disconnected (pan) or connected (pbn)
pðnsÞ d,pðnsÞ as an average over all possible if they have prescribed degrees di = d1 and
degrees, where P d; pðnlÞ is the probability that dj = d2, and are each linked to the same l sites
Structurally Dynamic Cellular Automata 57
in the shared neighbor set, Aij (see Fig. 1). In the choosing d1 l sites from i, d2 l sites from j,
case of type-T s- and ‘-rules, pan and pbn are given and l sites from both:
explicitly by (OT versions of s- and ‘-rules, and
RT versions of ‘-rules are defined by similar, but ðN 3Þd1 þd 2 l2
P2 ðd 1 ,d 2 ,lÞ ¼
slightly more complicated, expressions): ðd 1 lÞ!ðd 2 lÞ!l!
ðlÞ d 1 þd 2 2ðN þl3Þd 1 d 2
8 p 1 pð l Þ :
> P d 1 þ d 2 l ð sÞ bk
>
> pn ðd 1 ,d 2 ,lÞ ¼ k
a
pn (42)
>
> bk
>
>
< ðsÞ 1 2 lbk
d þd
1 pn , The second (link-update) function of the pair
> P d þ d þ 2 l ek of functions in Eq. (34) is thus given by
>
> pbn ðd 1 ,d 2 ,lÞ ¼ k 1 2
pðnsÞ
>
> e
>
> k
: d 1 þd 2 þ2lek ðl Þ
1 pðnsÞ , GRDA ! pnþ1
(38) ¼ pðnlÞ 1 Pan þ 1 pðnlÞ
where bk and ek refer to the sums that appear in PD¼2 Pan , (43)
Eqs. (11) and (12). The total probability that any
two sites will be disconnected (l = 1 ! l0 = 0) or
where, assuming that two sites, i and j, are not
connected (D = 2 ! l0 = 1) – Pan and Pbn ,
themselves connected, PD=2 = probability that
respectively – may then be obtained by summing
there exists at least one site k, such that Dik = Djk = 1,
over all possible local topologies:
which implies that PD¼2 ¼ 1 Probðthere is no such k Þ
8 a 2 N 2
>
> Pn XhProb
X ¼ 1 ! l 0 ¼ 0Þ i
ðl X ¼ 1 1 pðnlÞ ; and P1 and P2 are defined
>
>
>
> ¼ P1 ðd 1 , d 2 ,lÞ pan ðd 1 , d 2 ,lÞ,
< in Eqs. (41) and (42).
d1 d2 l
Pbn X
hProb ¼ 2 ! l 0 ¼ 1Þ i
ðDX A structural equilibrium is established when
>
> X
>
>
ðl Þ
pnþ1 pðnlÞ, which happens when the average num-
>
> ¼ P2 ðd 1 , d 2 ,lÞ pbn ðd 1 , d 2 ,lÞ,
: ber of new connections is equal to the average
d1 d2 l
(39) number of link deletions: Pbn hN nn i ¼ Pan hdegi,
where
where hdegi ¼ pðnlÞ ðN 1Þ is the average degree,
8
>
> P1 ðd 1 ,d 2 ,lÞ ¼ Probðsites i,j j l ij ¼ 1 have
d i ¼ d 1 , and Nnn is the average number of next-nearest neigh-
< d ¼ d ,A ¼ l ,
j 2 ij bors. For SDCA rules that naturally tend to produce
>
> P2 ðd 1 ,d 2 ,lÞ ¼ Probðsites i,j j l ij ¼ 0 have
d i ¼ d 1 , graphs with minimal site value and structural corre-
:
d j ¼ d 2 ,Aij ¼ l : lations, the predicted ratio of RDA link creations to
(40) deletions, gc:d Pbn N nn =Pan hdegi, may be used
to predict qualitatively how the graphs will evolve.
To find P1 we need to count, from among the Since the averagenumber
remaining N 2 sites, the number of ways of of pairs of sites a distance
N
selecting disjoint sets S1, containing d1 1 l D = 2 apart ¼ PD¼2 ¼ hN nn i N =2, we
2
sites linked only to i; S2, consisting of d2 1 l find that:
sites connected only to j; and S3, with l sites
( )
linked to both i and j. But this is simply a multi-
Pb 1 pðnlÞ h i2 N 2
nomial coefficient, so we can write: gc::d ¼ na ðl Þ
1 1 pðnlÞ :
Pn pn
ðN 2Þd 1 þd 2 l2 (44)
P1 ðd 1 ,d 2 ,lÞ ¼
ðd 1 1 lÞ!ðd 2 1 lÞ!l!
ðlÞ d1 þd 2 2 2ðN 1Þd 1 d 2
p 1 pðlÞ , gc::d is also implicitly a function of site-value
(41) density, since pðnsÞ appears in both Pan and Pbn ,
defined in Eq. (39).
where (n)k n(n 1) (n k + 1). Similarly, for Figure 16 shows a grayscale density-plot of
P2, we need to count the number of ways of gc::d for an OT decoupler rule:
58 Structurally Dynamic Cellular Automata
Structurally Dynamic
Cellular Automata,
Fig. 16 Density-plot of
gc::d for an OT decoupler
rule: {(0, 0),
(1, 1), (1, 2), (2, 2)}; an
OT coupler rule: {(1, 1)};
0:1 pðnsÞ 0:9; and 0:1
pðnlÞ 0:9; the rectangular
area highlighted in black
denotes the “equilibrium
boundary” that separates
regions of growth and decay
{(0, 0), (1, 1), (1, 2), (2, 2)}; an OT coupler rules assume the familiar T and OT (and related
rule: {(1, 1)}; 0:1 pðnsÞ 0:9 ; and 0:1 pðnlÞ RT) forms, as defined in section “The Basic
0:9 . Areas that are close to white represent Model”. Indeed, while the preceding sections of
combinations of pðnsÞ , pðnlÞ for which gc::d 1, this article have introduced several
and which therefore predict “decay”; areas that are generalizations – such as the addition of probabi-
close to black represent combinations of listic rules, reversibility and memory – in each
pðnsÞ , pðnlÞ for which gc::d 1, and predict case, the basic form of the rules (as defined in
“growth”; the rectangular area highlighted in Eqs. (11), (12), and (13)) has remained essentially
black denotes the “equilibrium boundary” that the same. However, just as for conventional CA, an
separates regions of growth and decay. almost endless variety of different kinds of rules
can in principle be defined; including rules that
alter the geometry but are not functions of the s
states. In this section, we look at two illustrative
Related Graph Dynamical Systems examples of SDCA-like dynamical systems: one
that uses coupled s-ℓ rules, and another whose
The original SDCA model (Ilachinski 1986) repre-
rules depend only on topology.
sents one (albeit not entirely arbitrary) approach to
dynamically coupling site values({si}) and topol-
ogy ({lij}), of the normally quiescent lattice. Since Graph Rewriting Automata
this model was primarily introduced as a general Tomita et al. (2002, 2005, 2006a, b, c) have
tool to explore self-organized emergent geome- recently introduced graph rewriting automata
tries, s values are an integral dynamic component (abbreviated, GRA), in which both links and (the
only because SDCA’s original rules were con- number of) sites are allowed to change. Motivated
ceived to generalize conventional CA rules, not by CA models of self-reproduction, Tomita et al
replace them. Moreover, SDCA’s link rules are, suggest that fixed, two-dimensional lattices – used
by design, close analogs of their conventional-CA as static backdrops to most conventional models –
brethren; this is the reason why SDCA’s c and o are unnecessarily restrictive for describing
Structurally Dynamic Cellular Automata 59
Structurally Dynamic Cellular Automata, Fig. 17 Graphical representations of the actions of the GRA rules defined
in Eq. (45). (Reproduced from Tomita et al. (2006a) with permission)
in which the rules are applied is determined by an “tape” (of indefinite length, to record data), a
a priori priority ranking. Also, since applying “head” (that reads/writes symbols on the tape,
either commutation or annihilation rules to adja- and that can move left or right), and “state transi-
cent links yields inconsistency, whenever a local tion rules” (that tell the head which new symbols
context arises in which this might happen, the to write given the current state of the tape). The
application of these rules is temporarily tape is analogous to “memory” in a modern com-
suppressed. (This is done by sweeping through puter; the head is analogous to the microproces-
the link set twice: on the first pass, a temporary sor. A Turing machine is called “universal” if it
flag is set for each link that satisfies a rule condi- can simulate any other Turing machine.
tion; on the second pass, the link rule is applied if Tomita et al.’s (2002) Turing machine is modeled
and only if the four neighboring links did not raise as a ladder structure: the upper sites constitute the
flags during the first pass.) “tape” mechanism; the lower sites form the “tape
Figure 18 shows the first few steps in applying head” that reads the tape; both ends of the ladder are
one division and two commutation rules to a sim- single sites that define “end of tape”; and the two
ple initial graph. (Kohji Tomita provides several ends are joined to form a loop. Although the tape is
movies of GRA evolutions on his website: http:// initially finite, the ladder can grow to arbitrary
staff.aist.go.jp/k.tomita/ga/) length, as required, by using appropriate GRA
Tomita et al. (2002, 2005, 2006a, b, c) report a rules. Tomita et al. (2002) self-replicating Turing
variety of emergent behaviors, including (1) arbi- GRA consists of 20 states and 257 (2-symbol)
trary resolution (because GRA rules effectively rules. They also introduce a design for a universal
allow an arbitrary number of sites to “grow” out of Turing machine (Tomita et al. 2006a) that consists
any initial structure, these systems define their of 30 states and 955 rules for reproduction, and
own “boundary conditions” and graphs with arbi- 23 states and 745 rules for computation. While
trary resolution are possible); (2) repetitive struc- self-reproducing universal Turing machines can be
tures, in which some geometrical subset of an described using conventional CA, their expression
initial graph is reproduced, indefinitely, and con- using GRA rules are considerably more compact.
tinuously grafted onto the original structure; and
(3) self-replication, in which both site-value and Dynamic Graphs as Models of Self-
structure is replicated after N steps. In Tomita et al. Reconfigurable Robots
(2006b), Tomita et al. describe how genetic algo- In the context of looking for self-reconfiguration
rithms (Mitchell 1998) may be used for automat- algorithms that may be used to manufacture mod-
ing the search for self-replicating patterns. ular robots for industry, Saidani (2003, 2004) has
In Tomita et al. (2002), Tomita et al. also pre- recently introduced a dynamic graph calculus that
sent the design of a self-reproducing Turing includes rules similar to those that define SDCA;
Machine. Turing machines are abstract symbol but which depend only on the topology of (but not
manipulating devices that mimic the basic opera- the s-values living on) the lattice. Saidani and Piel
tions of a computer. Formally, they consist of a (2004) have also introduced an interactive
Structurally Dynamic Cellular Automata 61
Structurally Dynamic
Cellular Automata,
Fig. 19 Schematic
illustration of a tree
topology reconfiguring
itself into a linear chain
using a set of case-based
“if–then” topology rules
defined in Saidani (2004);
see text for details
programming environment for studying dynamic computed from its local topology, and the in-
graph simulations called Dynagraph, and and out-degrees of its nearest neighbors); (3) a
implemented in Smalltalk. site controls its outgoing links (and can connect
There are two basic approaches to designing or disconnect any outgoing links), but cannot
modular robots: (1) to develop a set of elementary sever incoming connections; (4) sites must main-
generic modules that can be rapidly assembled by tain at least one link throughout an evolution
humans to form robots that solve a specific problem, (so that the graph remains connected); and (5) all
and (2) to design a set of (otherwise identical) prim- sites are equipped with the same set of rules.
itive components that can adaptively reconfigure As in conventional CA and the basic SDCA
themselves. Focusing on the latter approach, Saidani model, the “reconfiguration” proceeds synchro-
(2004) formally reinterprets modular “robots” to nously throughout the graph. The decision pro-
mean modular networks; and proceeds to model cess includes an innate stochastic element: in the
adaptive robotic self-reconfigurations as a class of event that there is a rule that specifies that a site is
recursive graph dynamical systems. In contrast to to establish a link to a neighbor of one of its
other related dynamic graph models (Ferreira 2002; neighbors, but all neighboring sites have the
Harary and Gupta 1997), the “modules” same degree (which is the only dynamical dis-
(or subgraphs) of Saidani’s model use local knowl- criminant), the neighbor with which a new link
edge of their neighborhood topology to collectively will be forged is selected at random.
evolve to some goal configuration. Although the As a concrete example, Saidani (2004) presents a
dynamics transforms the global state, the evolution tree-to-chain algorithm that evolves an initial “tree”
remains strictly decentralized, and individual mod- graph to a linear chain of linked sites (see Fig. 19).
ules do not know the (desired) final state. While we do not reproduce the full algorithm here, it
Apart from restricting the dynamics to topol- is essentially a case-driven list of rules of the form if
ogy alone (indeed, none of the sites harbor infor- condition C1 (and condition C2, . . . and condition
mation states of any kind), Saidani (2003, 2004), Cn) then connect (or disconnect) site i to (from) the
Saidani and Piel (2004) further assumes that nth neighbor of i’s neighbor, n. For example, an
(1) connections between sites are directional explicit “rule” might be: if 1 deg(i) 2 and
(both to- and from-links may coexist between the deg+(i) = 1 and |t(i)| = 2 then link i to a neighboring
same two modular components); (2) “active” sites site j that has deg(j) = 0, where deg(i) and
reconfigure their local neighborhood by deg+(i) are the in- and out-degrees of site i, and
accepting, keeping, or removing their adjacent t(i) is the total number of sites to which i is currently
links according to rules that are functions of their linked (with either incoming or outgoing links).
current topology (defined as a given sites’ current Conceptually, the details of Saidani’s rules are
local neighborhood and the current neighborhood less important than what the unfolding process
of its neighbors: a site only knows about its own represents as a whole. An initial graph – which
in- and out-degree, which can obviously be we recall is to be viewed as a distillation of a
62 Structurally Dynamic Cellular Automata
“modular robot” – is transformed, by the individ- proposed the idea that particles be viewed as geo-
ual sites (or parts of the robot), into another metric disturbances of space time, called
desired structure; i.e., the graph is entirely self- geometrodynamic excitons.
reconfigured. Though the broader reverse- A priori, SDCA appear tailor-made for describ-
engineering problem (which includes asking ing pregeometric theories of space-time. Since in
such fundamental questions as “How can a SDCA, lattice and local s-values are explicitly
desired final state be mapped onto a specific coupled, and geometry and value configurations
cased-based list of graphical rules?”) remains, are treated on an approximately equal footing,
as yet, unanswered, and the Dynagraph work SDCA is certainly at least formally consistent
environment (Saidani and Piel 2004) is currently with Einstein’s geometrodynamic credo. The
limited to experimenting only with graphs that structure is altered locally as a function of indi-
have less than 30 sites, the basic model already vidual site neighborhood value-states and geom-
represents a viable new approach to using etries, while local site-connectivity supports the
dynamic graphs to describe self-reconfigurable site-value evolution in exactly the same way as in
robots; and is potentially more far-reaching as a conventional CA models defined on random lat-
general model of topologically-reconfigurable tices. The microphysical view of physics that
dynamical systems. emerges from this construction is one in which a
fundamentally discrete pregeometry continually
evolves in time as an amorphous structure but
SDCA as Models of Fundamental Physics with a globally well-defined dimensionality. Par-
ticles are constructs of that amorphous structure
Pregeometric Theories of Emergent Space- and can be viewed as locally persistent
Time substructures – i.e. geometrical or topological
Although SDCA are a natural formal extension of solitons – with dimensions that differ from the
conventional CA – and serve as general-purpose surrounding value. Just as “value structure” soli-
modeling tools – their conception was originally tons are ubiquitous in conventional CA models
motivated by fundamental physics; specifically, (Ilachinski 2001; Wolfram 1984), “link structure”
by a search for models of self-organized emergent solitons might emerge in SDCA; physical parti-
discrete space-time (Meschini et al. 2005). “Space cles would, in such a scheme, be viewed as
acts on matter, telling it how to move; . . . matter geometrodynamic disturbances propagating
reacts back on space, telling it how to curve”, within a dynamic lattice.
which is the central lesson of Einstein’s geometro- Of course, speculation regarding the ultimate
dynamics, as explained by Misner, Thorne and constituents of matter and space-time date back at
Wheeler in their classic text on Gravitation least as far as 500 BC when the philosopher
(Misner et al. 1973). Wheeler (1982) has been a Democritus mused on the idea that matter is
particularly eloquent spokesman for the need to made of indivisible units separated by void.
search for what he calls a pregeometry, or a set of Since then there have been countless attempts,
basic elements out of which what we normally with varying degrees of success, to fashion an
think of geometry is built, but which are them- entirely discrete theory of nature. We limit our
selves devoid of a specific dimensionality: discussion to a short survey of some recent work
“Space-time . . . often considered to be the ulti- that centers on ideas that are either direct out-
mate continuum of physics, evidences nowhere growths of, or are otherwise conceptually related
more clearly than at big bang and at collapse to, SDCA models. (A short history of pre-
that it cannot be a continuum. Obliterated in geometric theories appears in Chap. 12 of
those events is not only matter, but the space and Ilachinski (2001)).
time that envelope that matter . . . we are led to ask One of the earliest proponents of pregeometry
out of what ‘pregeometry’ the geometry of space is Zuse (1982), who speculated on what it would
and spacetime are built”. Wheeler has also take for a CA-like universe to sustain “digital
Structurally Dynamic Cellular Automata 63
particles” on a cellular lattice. He focused on two sum of tensor products of single-particle; or, in
main problems: (1) How does the universe’s this case, single-graph, Hilbert spaces) it shares
observed isotropy arise from a CA’s (Euclidean, two important properties with SDCA: (1) interac-
hexagonal, etc.) anisotropy?, and (2) What is the tions depend only on the local properties of the
information content of a physical particle? As an graph, and (2) interactions induce only minimal
answer to the first question, Zuse suggests . . . changes to the local metric function. An important
consequence of their theory is that the dimension
“. . . variable and growing automata. Irregularities
of the grid structure are a function of moving pat- of a graph is a scale dependent quantity that is
terns, which is represented by digital particles. generated by the dynamics.
Now, not only certain values are assigned to the
single crosspoints of the grid in the concept of the Combinatorial Space-Time
cellular automaton which are interrelated and
sequencing each other, but also the irregularities Hillman (1995) introduces a combinatorial
of the grid are itself functions of these values of the space-time, which he defines as a class of dynam-
just existing interlinking network. One can imagine ical systems in which finite pieces of space time
rather easily that in such a way the interdependence contain finite amounts of information. Space
of mass, energy, and curvature of space may logi-
cally result from the behavior of the grid structure.” time is modeled as a combinatorial object,
constructed by dynamically coupling copies of
Jourjine (1985) generalizes Euclidean lattice finitely many types of certain allowed neighbor-
field theory on a d-dimensional lattice to a cell hoods. There is no a priori metric, and no con-
complex. Using homology theory to replace cept of continuity, which is expected to emerge
points by cells of various dimensions and fields on the macroscale.
by functions on cells, he develops a formalism The construction (and evolution) of spaces
that treats space-time as a dynamical variable proceeds in three steps: (1) define a set X of com-
and describes the change in the dimension of binatorial n-dimensional spaces (examples are
space-time as a phase transition. conventional CA graphs, graphs with directional
Kaplunovsky and Weinstein (1985) develop a links, or some other kind of embedded symme-
field-theoretic formalism that treats the topology try); (2) define a set of local, invertible primitive
and dimension of the spacetime continuum as maps T : X $ Y between pairs of space sets, such
dynamically generated variables. Dimensionality that the maps do not all commute with one another
is introduced out of the characteristic behavior of (for example, a simple renaming of the sites or
the energy spectrum of a system of a large number links gives an invertible, local map); (3) generate
of coupled oscillators. an arbitrary set of local invertible graph transfor-
Dadic and Pisk (1979) introduce a self- mations by composing primitive maps with one
generating discrete-space model that is based on another. Since the primitive maps are deliberately
the local quantum-mechanics of graphs. Just as in chosen so that they do not all commute, the act of
SDCA, Dadic and Pisk’s spatial structure is dis- composition yields infinitely many nontrivial
crete but not static; it is fundamentally amorphous transformations. The orbits{T z(x)| z Z} (for
and evolves in time. Though the metric is essen- each space x in X) are (n + 1)-dimensional com-
tially the same one used to define SDCA (i.e., binatorial spacetimes; which include reversible
Deffec), it is generalized to unlabeled graphs by CA and SDCA-like networks in which geometry
referring to the topological description of the node evolves locally over time. Formally, Hillman
positions rather than their arbitrary labels. Though uses matrices of nonnegative integers, directed
their “graph dynamics” differs from what is used graphs, and symmetric tensors to describe these
by SDCA (and uses a symmetrized Fock space systems, so that local equivalences between
that is local in terms of their graph metric, where space sets are generated by simple matrix trans-
“Fock space” is a Hilbert space used to describe formations. Concrete examples of dynamic com-
quantum states with a variable, or unspecified, binatorial space-time graphs are given in Hillman
number of particles, and is made from the direct (1995).
64 Structurally Dynamic Cellular Automata
Structurally Dynamic Disordered Cellular Nowotny and Requardt (1998) introduce two
Networks network models: one in which connected sites that
As an explicit example of how dynamic graphs can have very different internal states typically lead to
be used to model pregeometry, consider structurally large local fluctuations (=SDDCN1), and another
dynamic disordered cellular networks (abbreviated, in which sites with similar internal states are
SDDCN), recently introduced by Nowotny and connected (=SDDCN2):
Requardt (1998, 1999, 2006) and Requardt (1998,
2003a, b). SDDCN are a class of models closely SDDCN
8 tþ11 X
related to SDCA but developed explicitly to >
> si ¼ sti þ J tji ,
>
>
describe a discrete, dynamic space-time fundamen- < j
tal physics. The main difference between the two $ J tþ1 ¼ sign Ds Dsij l2 ,or
>
> ij for
Dsij l1 ^J tij 6¼ 0,
models is that whereas link connections in SDCA > ij
>
: tþ1
are strictly local, SDDCN are capable of generating J ij ¼ 0, otherwise,
both local and translocal links. SDDCN
8 2 X
>
> stþ1 ¼ sti þ J tji ,
In contrast to more mainstream high-energy >
>
i
>
> j
>
theories of fundamental physics (which are dom- < 0 < Dsij < l1 ,or
$ J ij ¼ sign Dsij
tþ1
for
inated by string theory and/or loop quantum grav- > 0 < Dsij < l2 ^J tij 6¼ 0,
>
>
ity, both of which assume a certain level > J tþ1 ¼ J t ,
> for Dsij ¼ 0,
>
> ij ij
discretization at the Planck scale, but assume : J tþ1 ¼ 0, otherwise,
ij
that a discrete space-time emerges from an under-
(46)
lying continuum physics), SDDCN takes a
bottom-up approach. SDDCN assumes that there
is underlying dynamic, discrete and highly erratic where Dsij ¼ sti stj , and l2 l1 0. Since
network substratum that consists of (on a given SDDCN is intended to model pregeometric
scale) irreducible mutually interacting agents dynamics, Nowotny and Requardt (1998) caution
exchanging information via primordial channels that the t parameter that appears in these equations
(links). The known continuum structures are must not to be confused with the “true time” that
expected to emerge on a macroscopic (or, meso- (they expect) emerges on coarser scales. In keep-
scopic) scale, via a sequence of coarse graining ing with its physics-based motivation, SDDCN’s
and/or renormalization steps. dynamical laws depend only on the relative dif-
Like SDCA, SDDCN are defined on arbitrary ferences in site values, not on their absolute
graphs, G, initially defined by a specified set of values. Indeed, charge is nowhere either created
sites and links. Both sites and links are allowed to or destroyed,Xso that SDDCN conserves global
take on values. Site values, si (which represent a “charge”: sti ¼ constant , where the arbitrary
primitive “charge”), are taken from some discrete constant cani be set to zero.
set, q Z, where q is a discrete quantum of infor- Both models start out initially on a simplex
mation; link states assume the values Jij {1, 0, graph with N 200 nodes, so that the maximum
+1}, and represent an elementary coupling. The number of possible links is N(N 1)/2. The initial
Jij are equivalent to SDCA’s lij, but take on three s-seed consists of a uniform random distribution
values rather than two. Heuristically, Jij represent of values scattered over the interval {k, k + 1,
directed edges pointing either from site i to . . ., k 1, k}, where k 100. The initial values
j (if Jij = 1), or from j to i (if Jij = 1); or, in for link states, J t¼0
ij , are selected from {1, 1}
the case of Jij = 0, the absence of a link. At each with equal probability; i.e., the initial state is a
time step (representing an elementary quantum of maximally entangled nucleus of nodes and links.
time), an elementary quantum q is transported Nowotny and Requardt (2006) state that “. . . in a
along each existing directed link in the indicated sense, this is a scenario which tries to imitate the
direction. As for SDCA, SDDCN dynamically big bang scenario. The hope is, that from this
couples site values to links. nucleus some large-scale patterns may ultimately
Structurally Dynamic Cellular Automata 65
The bold claim made in Chap. 9 of NKS is that, loss of generality. With two connections, only
on an even more fundamental level, what under- very trivial graphs are possible; and it is easy to
lies all the laws of physics, as we currently under- show that any site with more than three links can
stand them, is a simple CA-like program, from always be redefined, locally, as a collection of
which, ultimately, all the phenomenologically sites with exactly three links each (see Fig. 20).
observed complexity in the universe naturally Wolfram (2002) gives several concrete exam-
emerges. As for the specific forms such a “pro- ples of evolving graphs (as models of pre-
gram” may take, Wolfram’s intellectual point of geometry), the dynamics of which are prescribed
departure echoes that of other proponents of a by a set of substitution rules; i.e., explicit lists of
discrete dynamic pregeometric theory: the topological configurations (of sites and links)
that are used to replace (at time t + 1) specific local
“. . . cellular automata . . . cells are always
configurations (as they appear at time t). However,
arranged in a rigid array in space. I strongly sus-
pect that in the underlying rule for our universe in contrast to SDCA rules, Wolfram’s substitution
there will be no such built-in structure. Rather . . . rules are strictly topological; no site-value infor-
my guess is that at the lowest level there will just be mation is used. Also, the number of sites in the
certain patterns of connectivity that tend to exist,
graph can change as the graph evolves; where, in
and that space as we know it will then emerge from
these patterns as a kind of large-scale limit.” SDCA, the number remains constant.
Figure 21 shows examples of rules in which
Wolfram introduces his network automata specific clusters of sites are replaced with other
(abbreviated, NA) with these basic assumptions clusters of sites. While the rules shown in the
(see additional notes in NKS (Wolfram 2002) on figure share the property that they all preserve
the evolution of networks: pp. 1037–1040): planarity, there is no particular reason for impos-
(1) features of our universe emerge solely from ing such a restriction; in fact, rules that generate
properties of space, (2) the underlying model non-planarity are just as easy to define. Wolfram
(and/or “rules”) must contain only a minimal speculates (2002, pp. 526–530) that “particle
underlying geometric structure, (3) the individual states” may be defined as mobile non-planar sub-
sites of emergent graphs must not be assigned any graphs that persist on an otherwise planar, but
intrinsic position, (4) sites are limited to randomly fluctuating topology. Reversible ver-
possessing purely topological information (that sions of these rules may also be constructed, by
defines the set of sites to which a given site is associating a “backward” version with each “for-
connected), (5) incoming and outgoing connec- ward” transformation.
tions need not be distinguished, and (6) all sites Some care must be taken while both defining
have exactly the same total number of links to and applying these rules consistently. For exam-
other sites (which is assumed equal to three). ple, if a cluster of sites contains a certain number
This last assumption – which is essentially the of links at t, one is not permitted to define a rule
same one made by Nowotny and Requardt that replaces that cluster with another one that has
(1998) as the basis of their SDDCN model; see a different number of connections. Another
subsection “Structurally Dynamic Disordered restriction is that rules must be independent of
Cellular Networks” above – does not lead to any orientation; that is, if a candidate rule requires
Structurally Dynamic Cellular Automata, Fig. 20 Illustration of how sites that have more than three links can always
be redefined as a set of sites with exactly three links each
Structurally Dynamic Cellular Automata 67
Structurally Dynamic Cellular Automata, Fig. 21 Examples of planarity-preserving network substitution rules.
(Reproduced from Wolfram (2002) with permission)
identifying the specific links (of, say, an otherwise subgraph (or subgraphs, in Fig. 22a) involved in
topologically symmetric n-link local subgraph) the replacement is highlighted at both top and
before activating a desired substitution, that rule bottom.
is likewise forbidden. However, even with these Wolfram also suggests that analogs of mobile
restrictions, a large number of rules are still pos- automata (Miramontes et al. 1993) can be defined
sible. For example, 419 distinct rules may be for evolving networks. By tagging a site i, say,
defined for clusters with no more than five sites. with a “charge”, si 1, substitution rules may be
In applying network rules, one cannot simply defined to replace clusters of sites around the
simultaneously replace all pertinent subgraphs charged site. The effect is that the charge itself
with their replacements, since, in general, two or appears to move, as its effective (relative) position
more subgraphs with the same topology may within the network changes as the geometric
overlap somewhere within the network. Since dynamics unfolds. (However, Wolfram also
there is no priori, or universally consistent, way notes – on page 1040 in Wolfram (2002) – that
of ordering the subgraphs, meta-rules must be “despite looking at several hundred cases I have
imposed to eliminate any possible ambiguities. not been able to find network mobile automata
For example, one method (m1) is to restrict with especially complicated behavior”).
replacements to a single subgraph per time step,
selecting the subgraph whose replacement entails
the minimal change to all recently updated sites. Future Directions and Speculations
Another method (m2) is to allow all possible
nonoverlapping replacements, while ignoring Although SDCA were first introduced over two
those that overlap. Wolfram reports that, although decades ago (Ilachinski 1986), much of their
the second method obviously produces larger behavior remains unexplored. Of course, this is
graphs in fewer steps, the two methods generally due largely to the difficulty of studying dynamical
produce qualitatively similar structures. systems that harbor an a priori vastly larger
Figure 22 traces the first few steps in the evo- coupled value-geometry space than the “merely”
lution of a simple graph under the action of a spatially-confined behavioral space of conven-
single substitution rule (defined at the center of tional CA. Only relatively recently have desktop
the figure). Figure 22a, b show the results of computers become sufficiently powerful, and
applying this rule using methods m1 and m2, visualization programs adept enough at rendering
respectively. In each case, the top row shows the multidimensional graphs (Chen 2004), to make a
form of the network before the substitution takes serious study of SDCA behaviors possible. For
place at that step, and the bottom row shows the example, the general-purpose math programs
network that results from the substitution. The Mathematica (http://www.wri.com) and Maple
68 Structurally Dynamic Cellular Automata
Structurally Dynamic Cellular Automata, Fig. 22 Examples of network evolutions using the substitution rule shown
at center. See text for explanation. (Reproduced from Wolfram (2002) with permission)
range of values, ‘ij {0, 1, 2, . . ., m} Chen C (2004) Graph drawing algorithms. In: Information
(where, say, determines “directionality”, and visualization. Springer, New York
Dadic I, Pisk K (1979) Dynamics of discrete-space struc-
absolute value, |‘ij|, represents either channel ture. Int J Theor Phys 18:345–358
capacity for information flow or some other innate Doi H (1984) Graph theoretical analysis of cleavage pat-
property); and (3) both sites and links may take on tern: graph developmental system and its application to
richer, and more explicitly “active”, roles of cleavage pattern in ascidian egg. Develop Growth Dif-
fer 26(1):49–60
agent-actors (Ferber 1999). Durrett R (2006) Random graph dynamics. Cambridge
Apart from these formal extensions, some obvi- University Press, New York
ous future applications include modeling communi- Erdos P, Renyi A (1960) On the evolution of random
cation and social network dynamics, studying the graphs. Publ Math Inst Hung Acad Sci 5:11–61
Ferber J (1999) Multi-agent systems: an introduction to dis-
dynamics of plasticity in artificial neural networks, tributed artificial intelligence. Addison-Wesley, New York
designing adaptive self-reconfiguring parallel- Ferreira A (2002) On models and algorithms for dynamic
computer networks (as well as “amorphous” com- communication networks: the case for evolving graphs.
puter chips), studying behaviors of gene-regulatory In: 4th Recontres Francophones sur les Aspects
Algorithmiques des Télécommunications (ALGOTEL
networks, and providing the conceptual core for 2002), Meze
fundamental pregeometric physical theories of dis- Gerstner W, Kistler WM (2002) Spiking neuron models.
crete, emergent space-times. Single neurons, populations, plasticity. Cambridge
University Press, New York
Grzegorz R (1997) Handbook of graph grammars and
computing by graph transformation. World Scientific,
Singapore
Bibliography
Halpern P (1989) Sticks and stones: a guide to structurally
dynamic cellular automata. Am J Phys 57(5):405–408
Primary Literature Halpern P (1996) Genetic algorithms on structurally
Adamatzky A (1995) Identification of cellular automata. dynamic lattices. In: Toffo T, Biafore M, Leao J (eds)
Taylor & Francis, London PhysComp96. New England Complex Systems Insti-
Albert J, Culik IIK (1987) A simple universal cellular tute, Cambridge, pp 135–136
automaton and its one-way and totalistic version. Com- Halpern P (2003) Evolutionary algorithms on a self-
plex Syst 1:1–16 organized, dynamic lattice. In: Bar-Yam Y, Minai
Ali SM, Zimmer RM (1995) Games of proto-life in masked A (eds) Unifying themes in complex systems,
cellular automata. Complex Int 2. http://www.complex vol 2. Proceedings of the second international confer-
ity.org.au ence on complex systems. Westview Press Cambridge
Ali SM, Zimmer RM (2000) A formal framework for Halpern P, Caltagirone G (1990) Behavior of topological
emergent panpsychism. In: Tucson 2000: conscious- cellular automata. Complex Syst 4:623–651
ness research abstracts. http://www.consciousness.ari Harary F, Gupta G (1997) Dynamic graph models. Math
zona.edu/tucson2000/. Accessed 14 Oct 2008 Comp Model 25(7):79–87
Alonso-Sanz R (2006) The beehive cellular automaton Hasslacher B, Meyer D (1998) Modeling dynamical geom-
with memory. J Cell Autom 1(3):195–211 etry with lattice gas automata. Int J Mod Phys C 9:1597
Alonso-Sanz R (2007) A structurally dynamic cellular Hillman D (1995) Combinatorial spacetimes. PhD disser-
automaton with memory. Chaos, Solitons Fractals tation, University of Pittsburgh
32(4):1285–1304 Ilachinski A (1986) Topological life-games I. Preprint.
Alonso-Sanz R, Martin M (2006) A structurally dynamic State University of New York at Stony Brook
cellular automaton with memory in the hexagonal tes- Ilachinski A (1988) Computer explorations of self organi-
sellation. In: El Yacoubi S, Chopard B, Bandini S (eds) zation in discrete complex systems. Diss Abstr Int
Lecture notes in computer science, vol 4173. Springer, B 49(12):5349
New York, pp 30–40 Ilachinski A (2001) Cellular automata: a discrete universe.
Applied Graph & Network Analysis software. http://benta. World Scientific, Singapore
addr.com/agna/. Accessed 14 Oct 2008 Ilachinski A, Halpern P (1987a) Structurally dynamic cel-
Barabasi AL, Albert R (2002) Statistical mechanics of lular automata. Preprint. State University of New York
complex networks. Rev Mod Phys 74:47–97 at Stony Brook
Bollobas B (2002) Modern graph theory. Springer, Ilachinski A, Halpern P (1987b) Structurally dynamic cel-
New York lular automata. Complex Syst 1(3):503–527
Borgatti SP (2002) NetDraw 1.0: network visualization Jourjine AN (1985) Dimensional phase transitions: coupling
software. Analytic Technologies, Harvard of matter to the cell complex. Phys Rev D 31:1443
70 Structurally Dynamic Cellular Automata
Kaplunovsky V, Weinstein M (1985) Space-time: arena or Nowotny T, Requardt M (1999) Pregeometric concepts on
illusion? Phys Rev D 31:1879–1898 graphs and cellular networks as possible models of
Kniemeyer O, Buck-Sorlin GH, Kurth W (2004) A graph space-time at the Planck-scale. Chaos, Solitons Fractals
grammar approach to artificial life. Artif Life 10:469–486
10(4):413–431 Nowotny T, Requardt M (2006) Emergent properties in
Krivovichev SV (2004) Crystal structures and cellular structurally dynamic disordered cellular networks.
automata. Acta Crystallogr A 60(3):257–262 arXiv:cond-mat/0611427. Accessed 14 Oct 2008
Lehmann KA, Kaufmann M (2005) Evolutionary algo- O’Sullivan D (2001) Graph-cellular automata: a general-
rithms for the self-organized evolution of networks. ized discrete urban and regional model. Environ Plan
In: Proceedings of the 2005 conference on genetic and B: Plan Des 28(5):687–705
evolutionary computation. ACM Press, Washington, Prusinkiewicz P, Lindenmayer A (1990) The algorithmic
DC/New York beauty of plants. Springer, New York
Love P, Bruce M, Meyer D (2004) Lattice gas simulations of Requardt M (1998) Cellular networks as models for
dynamical geometry in one dimension. Phil Trans Royal Planck-scale physics. J Phys A 31:7997–8021
Soc A: Math Phys Eng Sci 362(1821):1667–1675 Requardt M (2003a) A geometric renormalisation group in
Majercik S (1994) Structurally dynamic cellular automata. discrete quantum space-time. J Math Phys 44:5588–5615
Master’s thesis, Department of Computer Science, Uni- Requardt M (2003b) Scale free small world networks and
versity of Southern Maine the structure of quantum space-time. arXiv.org:gr-qc/
Makowiec D (2004) Cellular automata with majority rule 0308089
on evolving network. In: Lecture notes in computer Rose H (1993) Topologische Zellulaere Automaten. Mas-
science, vol 3305. Springer, Berlin, pp 141–150 ter’s thesis, Humboldt University of Berlin
Mendes RV (2004) Tools for network dynamics. Int Rose H, Hempel H, Schimansky-Geier L (1994) Stochastic
J Bifurc Chaos 15(4):1185–1213 dynamics of catalytic CO oxidation on Pt(100). Physica
Meschini D, Lehto M, Piilonen J (2005) Geometry, pre- A 206:421–440
geometry and beyond. Stud Hist Philos Mod Phys Saidani S (2003) Topodynamique de Graphe. Les Journées
36:435–464 Graphes, Réseaux et Modélisation. ESPCI, Paris
Miramontes O, Solé R, Goodwin B (1993) Collective Saidani S (2004) Self-reconfigurable robots
behavior of random-activated mobile cellular autom- topodynamic. In: IEEE international conference on
ata. Physica D 63:145–160 robotics and automation, vol 3. IEEE Press,
Misner CW, Thorne KS, Wheeler JA (1973) Gravitation. New York, pp 2883–2887
W.H. Freeman, New York Saidani S, Piel M (2004) DynaGraph: a Smalltalk environ-
Mitchell M (1998) An introduction to genetic algorithms. ment for self-reconfigurable robots simulation.
MIT Press, Boston European Smalltalk User Group conference. http://
Moore EF (1962) Sequential machines: selected papers. www.esug.org/
Addison-Wesley, New York Schliecker G (1998) Binary random cellular structures.
Muhlenbein H (1991) Parallel genetic algorithm, popula- Phys Rev E 57:R1219–R1222
tion dynamics and combinatorial optimization. In: Tomita K, Kurokawa H, Murata S (2002) Graph automata:
Schaffer H (ed) Third international conference on natural expression of self-reproduction. Physica
genetic algorithms. Morgan Kauffman, San Francisco D 171(4):197–210
Murata S, Tomita K, Kurokawa H (2002) System genera- Tomita K, Kurokawa H, Murata S (2005) Self-description
tion by graph automata. In: Ueda K (ed) Proceedings of for construction and execution in graph rewriting
the 4th international workshop on emergent synthesis automata. In: Lecture notes in computer science,
(IWES ‘02), Kobe University, pp 47–52 vol 3630. Springer, Berlin, pp 705–715
Mustafa S (1999) The concept of poiesis and its application Tomita K, Kurokawa H, Murata S (2006a) Two-state
in a Heideggerian critique of computationally emergent graph-rewriting automata. NKS 2006 conference,
artificiality. PhD thesis, Brunel University, London Washington, DC
Newman M, Barabasi A, Watts DJ (2006) The structure Tomita K, Kurokawa H, Murata S (2006b) Automatic
and dynamics of networks. Princeton University Press, generation of self-replicating patterns in graph autom-
New Jersey ata. Int J Bifurc Chaos 16(4):1011–1018
Nochella J (2006) Cellular automata on networks. Talk Tomita K, Kurokawa H, Murata S (2006c) Self-description
given at the wolfram science conference (NKS2006), for construction and computation on graph-rewriting
Washington, DC, 16–18 June automata. Artif Life 13(4):383–396
Nooy W, Mrvar A, Batagelj V (2005) Exploratory social Weinert K, Mehnen J, Rudolph G (2002) Dynamic neigh-
network analysis with Pajek. Cambridge University borhood structures in parallel evolution strategies.
Press, New York Complex Syst 13(3):227–244
Nowotny T, Requardt M (1998) Dimension theory of Wheeler JA (1982) The computer and the universe. Int
graphs and networks. J Phys A 31:2447–2463 J Theor Phys 21:557
Structurally Dynamic Cellular Automata 71
Wolfram S (1984) Universality and complexity in cellular Bornholdt S, Schuster HG (eds) (2003) Handbook of
automata. Physica D 10:1–35 graphs and networks. Wiley-VCH, Cambridge
Wolfram S (2002) A new kind of science. Wolfram Media, Breiger R, Carley K, Pattison P (2003) Dynamical social
Champaign, pp 508–545 network modeling and analysis. The National Acad-
Zuse K (1982) The computing universe. Int J Theor Phys emy Press, Washington, DC
21:589–600 Dogogovtsev SN, Mendes JF (2003) Evolution of net-
works. Oxford University Press, New York
Durrett R (2006) Random graph dynamics. Cambridge
Books and Reviews
University Press, New York
Battista G, Eades P, Tamassia R, Tollis IG (1999) Graph
Gross JL, Yellen J (eds) (2004) Handbook of graph theory.
drawing: algorithms for the visualization of graphs.
CRC Press, Boca Raton
Prentice Hall, New Jersey
system converges to a fixed point. Fixed points
Asynchronous Cellular play a special role in the theory of asynchro-
Automata nous cellular automata because synchronous
and (classical) asynchronous models have the
Nazim Fatès same set of fixed points. In some cases,
LORIA UMR 7503, Inria Nancy – Grand Est, reaching a fixed point can be interpreted as
Nancy, France the end of a randomized computation.
De Bruijn graph (or diagram) This is an ori-
ented graph which allows one to represent all
Article Outline the overlaps of length n 1 in words of length
n. This graph is used to find some elementary
Glossary properties of the convergence of asynchronous
Article Outline CA, in particular to determine the set of fixed
Definition of the Subject points of a rule.
Introduction Elementary cellular automata There are
Defining Asynchrony in the Cellular Models 256 one-dimensional binary rules defined
Convergence Properties of Simple Binary Rules with nearest-neighbor interactions; an update
Phase Transitions Induced by a-Asynchronous sets a cell state to a value that depends only on
Updating its three inputs – its own state and the states of
Other Questions Related to the Dynamics its left and right neighbors. Using the symme-
Openings tries that exchange 0 s and 1 s and left and
Cross-References right, these rules reduce to 88 equivalence
Bibliography classes.
Game of Life This cellular automaton was
invented by Conway in 1970. It is probably
Glossary the most famous rule, and it has been shown
that it can simulate a universal Turing machine.
Configurations These objects represent the The behavior of this rule shows interesting
global state of the cellular automaton under phenomena when it is updated asynchronously.
study. The set of configurations is denoted by Markov chain A stochastic process that does not
Qℒ, where Q is the set of states of the cells and keep memory of the past; the next state of the
ℒ is the space of cells. In this text, we mainly system depends only on the current state of the
consider finite configurations with periodic system.
boundary conditions. In one dimension, we Reversibility When the system always return to
use ℒ = ℤ/nℤ, the class of equivalence of its initial condition, we say that it is reversible
integers modulo n. or, more properly speaking, that it is recurrent.
Convergence When started from a given initial Various interpretations of the notion of revers-
condition, the system evolves until it attains a ibility can be given in the context of probabi-
set of configurations from which it will not listic cellular automata.
escape. It is a difficult problem to know in Updating scheme The function that decides
general what are the properties of these attrac- which cells are updated at each time step. In
tive sets and how long it takes for the system to this text, we focus on probabilistic updating
attain them. In this text, we are particularly schemes. Our cellular automata are thus partic-
interested in the case where these sets are lim- ular cases of probabilistic cellular automata or
ited to a single configuration, that is, when the interacting particle systems.
# Springer Science+Business Media LLC, part of Springer Nature 2018 73
A. Adamatzky (ed.), Cellular Automata,
https://doi.org/10.1007/978-1-4939-8700-9_671
Originally published in
R. A. Meyers (ed.), Encyclopedia of Complexity and Systems Science, # Springer Science+Business Media LLC 2018
https://doi.org/10.1007/978-3-642-27737-5_671-2
74 Asynchronous Cellular Automata
at some point, the hypothesis of perfectly syn- asynchrony (a = 0.98), and an evolution with a
chronous transitions may seem unrealistic, but stronger asynchrony (a = 0.5).
we cannot know a priori if its use introduces The first observation is that the introduction of
spurious effects. There are some cases where a a small degree of asynchrony does not modify the
given behavior of a cellular automaton will only qualitative behavior of the rule on the short term.
be seen for the synchronous case, and there are However, one can predict that the long-term
also cases where this behavior remains constant behavior of the rule will be perturbed because it
when the updating scheme is changed. In essence, is no longer possible to observe cycles. For exam-
without any information on the system, we have ple, the configuration with only three living cells
no means to tell what are the consequences of in a row oscillates in the classical Game of life, but
choosing one updating scheme or the other. these oscillations only exist with a synchronous
If we have a robust model, changes in the updating, and the configuration evolves to a
updating may only perturb slightly the global totally different pattern when this perfect simulta-
behavior of a system. On the contrary, if this neity is broken. Another important property to
modification induces a qualitative change on the remark is that the new (asynchronous) system
dynamics, the model will be called structurally has the same fixed points as the original
unstable or simply sensitive to the perturbations of (synchronous) system. In fact, this is a quite gen-
its updating Scheme. A central question about eral property that does not depend on the local
cellular automata is thus to know how to assess rule. The reason is simple: if a configuration is a
their degree of robustness to the perturbations of fixed point of the synchronous system, it means
their updating. Naturally, the same questions can that all its cells are stable under the application of
be raised about the other hypotheses of the model: the local rule. Hence, if we select a subset of cells
the homogeneity of the local rule, the regular for an update, this subset will also be stable.
topology, the discreteness of states, etc. (see, Reciprocally, if any choice of cells gives a stable
e.g., Problem 11 in Wolfram (1985)). situation, then the whole system is also stable.
The second important observation regards the
A First Experiment evolution with a = 0.5: the global behavior of the
In order to make things more concrete, we pro- system is completely overwhelmed! A new sta-
pose to start our examination with a simple asyn- tionary behavior appears, and a pattern which
chronous CA. We will employ the resembles a labyrinth forms. This pattern is stable
a-asynchronous updating scheme (Fatès and in some parts and unstable in some other parts of
Morvan 2005) and apply the following rule: at the grid. We will not enter here into the details on
each time step, each cell is updated with a proba- how this stability can be quantified, but it is suffi-
bility a and is left in the same state with probabil- cient to observe that, in most cases, this pattern
ity 1 – a. The parameter a is called the synchrony remains for a very long time.
rate (see the formal definitions below) (Note that
from the point of view of a given cell, all happens Questions
as if between two updates each cell was waiting a It may be argued that these observations are not
random time that follows a geometric law of that surprising, because if one modifies the basic
parameter a.). The advantage of this definition is definitions of a dynamical system, one naturally
to control the robustness of the model by varying expects to see effects on its behavior. However,
the synchrony rate continuously from the classical this statement is only partially true, as this type of
synchronous case a = 1 to a small value of a, radical modifications is not observed for all the
where most updates will occur sequentially. We rules. In fact, as Nakamura has shown, we can
thus propose to examine the behavior of the always modify a rule in order to make it insensi-
a-asynchronous Game of Life. Figure 1 shows tive to the variations of its updating scheme
three different evolutions of the rule: the synchro- (Nakamura 1974, 1981). Formally, this amounts
nous case (a = 1), an evolution with a little to show that any classical deterministic cellular
76 Asynchronous Cellular Automata
a= 1
a = 0.98
a = 0.50
t=0 t = 25 t = 50 t = 75 t = 100
Asynchronous Cellular Automata, Fig. 1 Configura- stable at t = 50; (middle) small asynchrony introduced,
tions obtained with the a-asynchronous Game of Life for the system is still evolving at t = 100; (bottom) a = 1/2, the
three values of the synchrony rate a and the same initial qualitative behavior of the system has changed
conditions. (Top) Synchronous updating, the system is
automaton may be simulated by an asynchronous this rule on a two-dimensional grid with periodic
one. By “simulated” we mean that the knowledge boundary conditions shows that it is robust to the
of the evolution of the stochastic asynchronous variations of a: roughly speaking, if we start from a
system allows one to know the evolution of the uniform random initial condition, for 0.5 < a < 1,
deterministic original rule with a simple transfor- the system seems to always stabilize quickly on a
mation (see Th. Worsch’s article). The idea of fixed point. For smaller values of a, the only notice-
Nakamura is that each cell should keep three reg- able effect is a slowdown of the converge time.
isters: one with its current state, one with its previ- However, a modification also exists at the vicinity
ous state, and one with a counter that tells if its local of a = 1: like for the Game of Life, as soon as a little
time is late, in advance or synchronized with the asynchrony is present, cycles disappear.
local time of its neighbors. There is of course an These experiments indicate that there is some-
overhead in terms of simulation time and number of thing about asynchronous systems that deserves to
states which are used, and one may want to reduce be investigated. Since the first numerical simula-
this overhead as much as possible (Lee et al. 2004), tions (Buvel and Ingerson 1984), a great number
but the point is that there are asynchronous rules of approaches have been adopted to gain insights
which will evolve as their synchronous determin- on asynchronous cellular automata. However, if
istic counterparts. As an extreme example, we can we want to be convinced that these systems can be
also think of the rule where each cell turns to a studied and understood theoretically, and despite
given state independently of its neighbor: the their randomness, we need some analytical tools.
global evolution is easily predicted. The purpose of the lines that follow is to give a
Partial robustness can also be observed with few indications on how the question of asyn-
some simple rules. For example, let us consider chrony in cellular automata can be dealt with
the majority rule: cells take the state that is the theoretical tools from computer science and prob-
most present in their neighborhood. Observing ability theory.
Asynchronous Cellular Automata 77
Defining Asynchrony in the Cellular 8c ℤ, xtþ1 ¼ . . . cþnk :
t t
c f x cþn1 , , x
Models
Literally, asynchronous is a word derived from the Now, to define an asynchronous cellular
Ancient Greek a᾿sunwronoB, which simply means automaton, we need to introduce an updating
“not same time”. From this etymology, it follows scheme. Such a function takes the form U : ℒ !
that we cannot speak of a single model of asyn- P ðℒÞ, where P ðSÞ denotes the parts of S, that is,
chrony in cellular automata, but there is an infinity the set of all subsets of S (also denoted by 2S). For
of models. In fact, one is allowed to speak of an a given time step t ℕ, the set of cells that are
asynchronous model as soon as there is some updated at time t is represented by U ðtÞ.
perturbation in the updating process of the cells We obtain a new global rule, denoted by
(Note that asynchrony and asynchronism have FU : ℕ Qℒ ! Qℒ where FU ðx, tÞ represents
been both used in the literature in an equivalent the image of x at time t given the updating
way. We will in general use the former for the scheme U: The evolution of (xt)t ℕ starting
modification of the updating and use the latter to from x Qℒ is now defined with x0 = x and
designate a topic of research.).We voluntarily stay xtþ1 ¼ FU ðxt Þ such that:
vague at this point in order to stress that one may (
imagine a great variety of situations where some f xtcþn1 , . . . , xtcþnk if c U ðtÞ,
8c ℤ, xc ¼
tþ1
t
irregularity occurs on the way the information is xc otherwise:
processed by the cells. For instance, we may
examine what happens if all the transitions do The type of function U defines the type of
occur at each time step but where the cells receive asynchronism in use. The first issue of distinction
the state of their neighbors imperfectly. is between deterministic and stochastic
In this text, we will restrict our scope to the (or probabilistic) functions. In this text, we will
most simple cases of asynchronous updating. focus on stochastic functions. Indeed, since asyn-
chronism is often thought of as an unpredictable
aspect of the system, stochastic systems have been
Mathematical Framework more intensively studied. One finds only a small
Let ℒ ℤd be the set of cells that compose a number of studies which use deterministic systems.
d-dimensional cellular automaton. The set of Examples of such studies can be found in Cornforth
states that each cell may hold is Q. The collection et al. (2005), Schönfisch and de Roos (1999) where
of all states at a given time is called a configura- the authors have considered, for example, the
tion, and the configuration space is thus Qℒ. effects caused by updating cells sequentially from
k
Let N ℤd be the neighborhood of the left to right. As one may expect, such approaches
cellular automaton, that is, for N ¼ ðn1 , . . . , nk Þ, often lead to curious phenomena: the information
ni represents the vector between the central cell spreads in a nonnatural way because a single
and its ith neighbor. sequence of updates from left to right suffices to
The local function of a cellular automaton is a change the state of the whole system. More inter-
function f: Qk ! Q which assigns to a cell c ℒ esting are even-odd updating schemes where one
its new state q0 = f (q1,. . ., qk), where the tuple updates the even cells and, in the next step, the odd
(q1,. . ., qk) represents the state of the neighbors of cells. A famous example of such model is the Q2R
a cell c. model (Vichniac 1984): although the local rule of
Starting from an initial configuration x Qℒ, this system is deterministic, using a random initial
the classical evolution of the system gives a condition makes it evolve with the same density as
sequence of configurations that we denote by the Ising model (see, e.g., Kari and Taati (2015) for
(xt)t ℕ. This sequence is obtained by the recursive a recent development).
application of the global rule F : Qℒ ! Qℒ In fact, we can remark that in general it is not
defined with x0 = x and xt+1 = F (xt) such that: difficult to transform an asynchronous system into
78 Asynchronous Cellular Automata
a synchronous one: in many cases, adding more of independent and identically distributed ran-
states is sufficient. For example, for the even-odd dom variables that select an element uniformly
updating, we may mark the even and odd cells with in ℒ. The evolution of the system is given by:
a flag up and down, respectively, and make this flag
“flip” at each time step. Similarly, an ordered x0 ¼ x and 8i ℤ, xtþ1 i
updating may be simulated in a synchronous (
model by moving a token in a given order. How- f xtiþn1 , . . . , xtiþnk if i ¼ St ,
¼
ever, such direct transformations are not always xtc otherwise:
possible: for example, Vielhaber has proposed an
original way of achieving computation universality Note that in most
cases,
authors do not use the
by selecting the cells to update (Vielhaber 2013), indices i and t for ℬti or (St) and simply consider
and this construction cannot be transformed into a that there is one function that is used at each time
deterministic cellular automaton by the mere addi- step and for each cell.
tion of a few internal states. We do not enter here into the details of how we
can generalize these definitions (see, e.g.,
Dennunzio et al. (2013)). We point the work of
Randomness in the Updating
Bouré et al. on asynchronous lattice-gas cellular
In the case where the updating scheme U is a
automata to underline that adding asynchrony to
random variable, then the evolution of the system
the cellular models which have more structure
is a stochastic process, and if U does not depend
than the classical ones can be a nontrivial opera-
on time, it is a Markov chain (a memoryless sys-
tion if one wants to maintain the properties of
tem). In order to be perfectly rigorous in the for-
these models (e.g., conservation of the number
mal description of the system, advanced tools
of particles) (Bouré et al. 2013a). Similar difficul-
from probability theory are necessary. A good
ties arise when agents can move on the cellular
example on how to properly use these mathemat-
grid, and one needs to define some procedures to
ical objects and their properties can be found in a
solve the conflicts that may occur when several
survey by Mairesse and Marcovici (2014). How-
agents want to modify simultaneously the same
ever, for the sake of simplicity, one may still use
cell (Belgacem and Fatès 2012; Chevrier and
the usual notations and consider that the
Fatès 2010).
sequences (xt)t ℕ are formed by configurations
rather than probability distributions.
We can now define the two major asynchro-
nous updating schemes: Convergence Properties of Simple
Binary Rules
• a-asynchronous updating scheme: let a (0, 1]
be constant called the synchrony rate. Let We have seen that a central question in the study
t
ℬi i ℒ, t ℕ be a sequence of independent of asynchronous cellular automata was to deter-
and identically distributed Bernoulli random mine their convergence properties. In particular
variables of parameter a. The evolution of the one may wonder, given a simple binary rule, what
system with an a-asynchronous updating we can predict about its possible behavior. Is it
scheme is then given by: converging to a given fixed point? In which time
in average? And if so, what kind of “trajectory”
the system will follow to attain a stable state
x0 ¼ x and 8i ℤ, xtþ1 i
( (if any)? The lines that follow aim at presenting
f xtiþn1 , . . . , xtiþnk if ℬti ¼ 1, the mathematical tools to answer these questions.
¼
xtc otherwise:
Expected Convergence Time to a Fixed Point
• Fully asynchronous updating scheme: in the Recall that one major modification caused by the
case where ℒ is finite, let (St)t ℕ be a sequence transformation of a cellular automaton from
Asynchronous Cellular Automata 79
b 1 f
001 011
0 1
0011001111100 1 1
abfgcbfhhhgca
↑ 0 a 1 e 1 d 0 h
1
0011001011100 000 010 0 101 111
abfgcbedfhgca 0 0
0 1
c 0 g
100 110
Asynchronous Cellular Automata, Fig. 2 (left): Exam- correspondence between binary sequences of length
ple of two binary configurations and their images by the 3 and transitions A, . . ., H. The label on the edges shows
transition code. The upper configuration is obtained by the next letter that is given in input when reading a binary
updating the lower configuration on the cell indicated sequence from left to right
with an arrow. (Right) De Bruijn graph with the
Asynchronous Cellular Automata, Fig. 3 Space-time bottom to top. Each row shows the state of the system
diagrams showing the evolution of the shift rule for a after n random updates. This convention is kept in the
ring of n cells, with n = 20. Cells in blue and white, following
respectively, represent states 0 and 1. Time goes from
Ti n3/8; in other words, for the configurations We thus have an upper bound on the WECT
with only two zones, the average number of which is WECT (n) n2/8, and, considering the
updates needed to attain a fixed point is at most initial condition x = 0n/21n/2, we obtain the lower
cubic in n. bound WECT (n) n2/8. We can thus write WECT
(n) = O(n2) where O expresses the equivalence up
to a constant. We thus say that the shift has a
Martingales
quadratic convergence time or, for short, that it
How can we deal with the other configurations? If
is quadratic.
we start from a configuration x with k 1-regions
and k > 1, the probability to increase or decrease
A Relationship with Computational Problems
by 1 the number of 1 s is kϵ. The evolution of the
In fact, since the convergence of the asynchro-
system can no longer be described by the Markov
nous shift depends on the initial density, one
chain in Fig. 4. Indeed, the value ϵ needs to be
may consider this process as a particular kind
replaced by ϵ 0 = kϵ, but, as k is not constant, this
of decentralized computation. For the sake of
process is no longer a Markov chain. As seen in
brevity, we will not develop this point here, but
Fig. 4, the frontiers of the regions will perform
we simply indicate to the readers interested by
random walks until a region disappears, which
this issue that similar stochastic rules have been
will make ϵ 0 decrease again and so on until we
used to solve the density classification problem
reach one of the two fixed points. In order to
(see, e.g., (de Oliveira 2014; Fatès 2013b; Fukś
determine the convergence time t (x), one could
2002) and de Oliveira’s article in this
estimate the average “living time” of a configura-
encyclopedia).
tion with k-regions. However, this is a difficult
problem because this living time strongly depends
From the Shift to Other Quadratic Rules
on the size of each region.
We now examine step by step how to generalize
It is easier to note that the process (Xt) defined
the example of the asynchronous shift given
with Xt = |xt|1 is a martingale, that is, a stochastic
above to a wider class of rules.
process whose average value is constant over
With the decomposition described above, we
time. The theory of martingales allows us to find
can readily deduce that the Markov chain
the probability p1(x) to reach the fixed point
described for counting the number of 1 s for
1 from x and the average time of convergence
the shift rule (BDEG) also applies for rule CG,
ftðxÞg. For the sake of brevity, we skip the details
for which the 10-frontier performs a non-biased
of the mathematical treatments and write down
random walk and for rule BCDEFG, for which
directly the results that are exposed in Fatès et al.
the two frontiers perform a random walk.
(2006a): (a) the probability of reaching the
In a second time, we can ask what happens if
fixed point 1 is still equal to the initial density,
we change the code of these rules by removing
p1(x) = |x|1/n, and (b) the rescaled average time
the transition D of their code, that is, we set
scales
also quadratically with n : ftðxÞg jxj1 101 ! 0 and make the transition D inactive.
n jxj1 =2.
This transformation implies that the 0-regions
can no longer disappear, while the 1-regions
may disappear if an isolated 1 is updated
ε ε ε (010 ! 0). As a consequence, the fixed point
0 1 2 n−1 n 1 is no longer reachable, and the system will
ε ε ε almost surely converge to the fixed point 0 for
1 1−2ε 1−2ε 1−2ε 1 an initial condition different from 1. The system
will thus most of the time behave as a regular
Asynchronous Cellular Automata, Fig. 4 Representa-
tion of the Markov chain that counts the number of 1 s. The
martingale, but sometimes it will “bounce” on
constant ϵ = 1/n represents the probability to update a cell an isolated 0. Is the average convergence time
at a given time step still quadratic? The answer is positive: even
Asynchronous Cellular Automata 83
though the behavior cannot be described “spectacular” than the quadratic rules. The size
(simply) by a martingale, it is possible to of the 1-regions regularly decreases until all the
“save” the previous results and still obtain a regions disappear and the system reaches the fixed
quadratic scaling of the WECT. Interested point 0. It is easy to see that in the case where the
readers may refer to our study on fully asyn- initial condition does not contain an isolated 0, the
chronous doubly quiescent (A quiescent state is evolution of the number of 1 s is a non-increasing
a state q such that the local rule f obeys f(q,. . ., function. Now, let us consider the function
q) = q.) rules for the mathematical details (Fatès f : Qℒ ! ℕ defined by f(x) = |x|1 + |x|01. Writing
et al. 2006a). (Xt) = f(xt), one can verify that the evolution of
(Xt) is non-increasing. Indeed, if a transition D is
Functions with a Potential applied, the number of 1 s increases by 1, but the
In the previous paragraph, we started from the number of regions also decreases by 1. Moreover,
shift rule (BDEG), showed that it had a quadratic we have that Xt = 0 implies that xt = 0. The
WECT, an then indicated that five other rules had function f can thus be named a potential: it is a
a similar qualitative behavior and a quadratic positive, non-increasing function of the current
WECT. The other rules were obtained by making state of the system, which equals to zero when
the transition D inactive or by changing the the system has attained its attractive fixed point.
behavior of the frontiers, as long as this move- This argument can be applied for showing a linear
ment remained a non-biased random walk WECT for the following four rules (G is active)
(Fig. 5). 136:EG, 140:G, 168:DEG, and 172:DG and the
We now propose to examine what happens if following four rules (F and G are active) 128:
we dare to “touch” a transition that breaks the EFG, 132:FG, 160:DEFG, and 164:DFG.
random movement of the frontiers. Concretely, Interestingly, a similar type of convergence can
let us make the transition B inactive: we obtain also be obtained by adding an active transition to
the minimal representative rule DEG (168). The the shift rule. For example, let us consider ECA
evolution of this rule is displayed in Fig. 6; it can BDEFG (162). Its evolution is shown in Fig. 6.
be seen that the evolution of the rule is less One should observe that the 01-frontiers perform
Asynchronous Cellular Automata, Fig. 5 Space-time diagrams showing the evolution of four rules with a quadratic
worse expected convergence time (WECT).
84 Asynchronous Cellular Automata
Asynchronous Cellular Automata, Fig. 6 Space-time diagrams showing two evolutions of two rules with a linear
worse expected convergence time (WECT)
a non-biased random walk, while the 10-frontier • The rules 134:BFG, 142:BG, 156:CG, and
tends to move to the left. This means that the 150:BCFG are non-converging. This is
1-regions have a tendency to decrease, but their because in all these rules, transitions D and
evolution is no longer monotonous as in the case E are inactive and, at the same time, the fron-
of rule DEG. It can be shown that if we take back tiers are not static.
the function f(x) = |x|1 + |x|01 and Xt = f(xt),
then (Xt) is a super-martingale, that is, its average Other Elementary Rules
value decreases in average. This property and The question of classifying the other ECA rules,
other conditions ensuring that it cannot stay too where no state or only one state is quiescent, is still
“static” imply that its convergence time scales open. Some conjectures have been stated from
linearly with the ring size n (Fatès et al. 2006a). experimental observations, but they still deserve
Indeed, for any configuration
that is not a fixed an in-depth analysis (Fatès 2013a). In particular,
point, the quantity Xtþ1 Xt jxt is negative. there are currently only partial results for all the
The same method can be applied for showing the rules which are conjectured to converge “very rap-
convergence in linear time for the rule 130: idly,” that is, in logarithmic time (Fatès 2014b).
BEFG.
From Fully Asynchronous to a-Asynchronous
Non-polynomial Types of Convergence. Updating
For the sake of brevity, we will not go here into the What happens if one uses a partially synchronous
details but only indicate the other classes of con- updating scheme instead of a totally asynchronous
vergence that were exhibited. Readers may con- one? Regnault et al. have extended the conver-
sult Fatès et al. (2006a) for detailed arguments. gence results of the doubly quiescent ECA to the
case of a-asynchronous updating (Fatès et al.
• The rules 200:E and 232:DE have a logarith- 2006b). The possibility of having simultaneous
mic WECT. This can be shown with the same updates of neighboring cells creates additional
techniques as for the convergence of the “local movements,” and the behavior of these
coupon-collector process (Fatès et al. 2006a). rules is more difficult to analyze. In particular,
• The rule 154:BCEG has an exponential the authors have identified four phenomena that
WECT. This comes from a kind of paradox: are specifically related to the a-asynchronous
the rule has a tendency to increase the number updating: the shift, the fork, the spawn, and the
of 1 s, but its only fixed point 1 is not reachable. annihilation. These phenomena are shown in
The only way it can converge is by reaching the Fig. 7.
fixed point 0, a phenomenon that is very The authors developed an interesting analytical
unlikely. framework (potential functions, masks, etc.) and
Asynchronous Cellular Automata 85
t +1
t
shift fork spawn annihilation
Asynchronous Cellular Automata, Fig. 7 New phenomena observed with the a-asynchronous updating of linear CA
(From the work of Fatès et al. (2006b))
succeeded in giving bounds on the convergence of remarkable exception was given by the epidemic
19 (minimal) doubly quiescent rules, leaving the rule, where a 0 turns into a 1 if it has a 1 in its
question open for five other rules. The various neighborhood and then will always remain a 1. This
pffiffiffi
rules show different kinds of scaling relations of rule has a WECT which scales as Yð nÞ: Even
the WECT, depending on a and n. If we consider though this scaling property can be intuitively
the dependence on n only, the families of functions understood from the dynamics of the rule, which
are the same as those obtained for fully asynchro- merely amounts to “contaminating” neighboring
nous dynamics, that is, logarithmic, linear, qua- cells, proving the class of convergence was a diffi-
dratic, exponential, and infinite. However, there are cult task. It is only recently that a proof has been
rules whose type of converge varies from the fully proposed by Gerin, who succeeded in applying
asynchronous updating to a-asynchronous subtle combinatorial arguments to obtain upper
updating. For example, rule 152:CEG, which is and lower bounds on the time of convergence
quadratic with a fully asynchronous updating (see (Gerin 2017).
above), becomes linear for a-asynchronous The minority rule received a special attention.
updating. Two rules, namely, ECA 146:BCEFG Indeed, when updated asynchronously, it has the
and 178:BCDEFG, were conjectured to display a ability to create patterns which can take the form
phase transition: their type of converge may change of checkerboard or stripes. The behavior of this
from polynomial to exponential depending on rule with an asynchronous updating was analyzed
whether the a is greater or lower than a particular in the case of von Neumann and Moore neighbor-
critical value. This property was partially proved by hood (the cell and its eight nearest neighbors)
Regnault in a thorough study of ECA 178, where the (Regnault et al. 2009, 2010). Regnault et al.
polynomial and exponential convergence times noticed that the convergence to the fixed point
were formally obtained for extrema values of the was not uniform: the process can be separated in
synchrony rate (Regnault 2013). Ramos and Leite two phases – first the “energy” decreases rapidly,
recently studied a generalization of this model and then the system stays in a low-energy state
where the asynchronous case appears as a special where it will progressively approach the fixed
case of the family of probabilistic cellular automata point by moving the unstable patterns, thanks to
that are studied (Ramos and Leite 2017). the random fluctuations. It is an open question to
know to which extent this type of behavior can be
Two-Dimensional Rules found in other contexts, e.g., lattice-gas cellular
The study of the convergence properties of simple automata (Bouré et al. 2013b).
two-dimensional rules has been carried out for the The convergence properties can thus be deter-
so-called totalistic cellular automata, where the local mined quite precisely but only for a family of simple
rule only depends on the number of 1 s in the binary cellular automata rules. It is an open problem
neighborhood (Fatès and Gerin 2009). For the von to find such analytical tools. As far as the
Neumann neighborhood (the cell and its four nearest a-asynchronous updating is concerned, the results
neighbors), there are 26 such rules. Their WECT are even more restricted. As we will see in the
were also analyzed for the fully asynchronous following, this is not so surprising because the
updating, and all rules but one was found to fall behavior of some rules sometimes requires the intro-
into the previous classes of convergence. One duction of tools from advanced statistical physics.
86 Asynchronous Cellular Automata
Phase Transitions Induced by at ac 0.911. Moreover, for the Game of Life, the
a-Asynchronous Updating critical phenomenon was shown to be robust to
the introduction of a small degree of irregularity in
The Game of Life the grid. This phase transition was also observed
We propose to come back to the phenomenon for other lifelike rules (Fatès 2010).
observed in Fig. 1 (see p. 4). Blok and Bergersen
were the first authors to give a precise explanation Elementary Cellular Automata
of the change of behavior in the Game of Life, the In the first experiment where the whole set of
phenomenon that was described in the introduc- ECAs was examined with an a-asynchronous
tory part of this article. They identified the exis- updating (Fatès and Morvan 2005), some rules
tence of a second-order phase transition were observed to display an abrupt variation of
(Informally, in statistical physics, phase transi- the density for a given value of the synchrony rate
tions are defined by the existence of a discontinu- a. This phenomenon was later studied in detail,
ity in the values taken by a macroscopic and this critical phenomenon was identified for
parameter, called the order parameter, when sys- ten (non-equivalent) rules. As for the Game of
tem is submitted to a continuous variation of a Life, we are here in the presence of second-order
control parameter. First-order transitions are those phase transitions which belong to the directed
for which the discontinuity appears directly on the percolation universality class (Fatès 2009). The
order parameter, while second-order phase transi- values of the measured critical synchrony rates are
tions (or continuous phase transitions) are those reported in Table 3.
where the derivative of the order parameter is It is a puzzling question to know why these ten
infinite.) which separates two qualitatively differ- rules are specifically producing such critical phe-
ent behaviors: a high-density steady state with nomena. Some insights to this question were given
vertical and horizontal stripes and low-density in a study of the local-structure approximations of
steady state with avalanches (Blok and Bergersen the rules, that is, a generalization of the mean-field
1999). They measured the critical value of the syn- approximation to correlations of higher order (Fukś
chrony rate at ac 0.906 and showed that near the and Fatès 2015). This study revealed that it was
critical point, the stationary density d1 obeyed a possible to predict the occurrence of a phase transi-
power law of the form d1 (a ac)b. It is well tion, but it was not possible to use it to correctly
known in the field of statistical physics that the approximate the value of the critical synchrony rate
values taken by the power laws are not arbitrary (Fig. 8). Another possible approach would be to
and that various systems of unrelated fields may analyze the branching-annihilating phenomenon in
display the same critical exponents (see F. Bagnoli’s a specific way, with small-size Markov chains, for
article in this encyclopedia). The class of systems instance, but this remains an open path of research.
which share the same values of exponents is called a
universality class, and in the case of the Game of
Life, Blok and Bergersen found that its phase tran- Other Questions Related to the
sition was likely to belong to the universality class Dynamics
of directed percolation (also called oriented perco-
lation or Reggeon field theory). In order to broaden our view of asynchronous
These measures were later confirmed by a set cellular automata, we now briefly mention some
of more precise experiments (Fatès 2010), and the other problems which have been studied with
critical value of the synchrony rate was measured analytical tools.
Asynchronous Cellular Automata, Table 3 Critical synchrony rates for the ECA with a phase transition
ECA 6 18 26 38 50 58 106 134 146 178
ac 0.283 0.714 0.475 0.041 0.628 0.340 0.815 0.082 0.675 0.410
Asynchronous Cellular Automata 87
ECA 6 ECA 50
0.5 0.5
0.45 0.45
0.4 0.4
0.35 0.35
0.3 0.3
P(1)
P(1)
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
α α
ECA 146
0.35
plots legend
0.3
k=2
0.25 k=3
0.2 k=4
P(1)
0.15 k=5
k=6
0.1
k=9
0.05 exp
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
α
Asynchronous Cellular Automata, Fig. 8 Local struc- k = 8 are omitted. The plot in red (labelled “exp”) shows
ture approximations obtained for various approximation the experimental steady-state density obtained for a ring
levels of order k (see Fukś and Fatès (2015) for details). size of n = 40,000 cells after 10,000 time steps
For the sake of readability of the results, the cases k = 7 and
Asynchronous Cellular Automata, Table 4 Wolfram Fig. 9: if we start from two different initial condi-
codes and transition codes of the 16 recurrent rules (From tions of the same size and apply the same updates
Fatès et al. (2017)). The two separate rules are recurrent for
n 6¼ 3 on the two systems, they quickly synchronize and
adopt the same evolution. This is a particular kind
35: 38:BDFGH 43:ABDEGH 46:BDGH
ABDEFGH of synchronization where no desynchronization is
51: 54:BCDFGH 57:ACDEGH 60:CDGH possible: after the coalescence has occurred, the
ABCDEFGH two trajectories remain identical as the local rules
62:BCDGH 105:ADEH 108:DH 134:BFG are deterministic. The question is to know under
142:BG 150:BCFG 156:CG 204:I which conditions coalescence happens and how
33:ADEFGH 41:ADEGH long does it take in average for two different initial
conditions to “merge” their trajectories.
Rouquier and Morvan have studied experi-
2017). These rules are listed in Table 4. For the
mentally this phenomenon for the 88 ECA with
recurrent rules, the structure of the transition
a-asynchronous updating (Rouquier and Morvan
graph was analyzed as well as the number of
2009). They discovered an unexpected richness of
connected components of this graph, that is, the
behavior: some rules coalesce rapidly and others
number of communication classes of the rules. It
slowly, some never coalesce, some even display
was found that the number of classes of commu-
phase transitions, etc. Insights have been given by
nication varies greatly from one rule to another:
Francès de Mas on this question, and a classifica-
some rules have an exponential number, while
tion of the convergence time has been given from
others have a constant number; the most interest-
both the observation of space-time diagrams and
ing examples were obtained for the rules with an
an analysis of the behavior (de Mas 2017). It is
“intermediary” behavior. For example, for rule
still an open question to provide a complete math-
105:ADEH, the number of communication clas-
ematical analysis of these systems and to issue a
ses is 2 for an odd ring size n and is equal to n/
proof that coalescence can indeed happen in a
2 + 3 when n is divisible by 4 and to n/2 when n is
linear time with respect to the ring size.
even and not a multiple of 4. It is an open question
to generalize these results to other types of rules or
Other Problems
to other types of updating schemes.
There are many other problems which have led to
These results are encouraging, and it is rather
various interesting experimental or theoretical
pleasant to note that contrarily to the problem of
works. For instance, Gacs (2001) and then
convergence seen above, deciding the recurrence
MacAuley and Mortveit (2010, 2013; Macauley
properties of an ECA can be achieved. It is thus
et al. 2008) have provided a deep analysis on the
interesting to see to which extent these results
independence of the trajectories of an asynchro-
apply to a broader class of systems, including
nous with regard to the updating sequence.
infinite-size systems.
Chassaing and Gerin analyzed the scaling relation-
ships that would lead to an infinite-size continuous
Coalescence framework (Chassaing and Gerin 2007). This
In the experimental study of the a-asynchronous framework is also analyzed in detail by Dennunzio
ECA (Fatès and Morvan 2005), a strange phe- et al., who examined how the theory of measure
nomenon was noticed for ECA 46, almost by can be applied to one-dimensional systems defined
chance: though this rule does not converge to a on an infinite line (Dennunzio et al. 2013, 2017).
fixed point and remains in a chaotic-like steady As an example of a possible application of the
state, its evolution does not seem to depend on the use of these dynamical systems, we mention the
initial condition. All seems to happen as if the work of Das et al., who proposed to use such
evolution of the rule was only dictated by the models for pattern classification (Sethi et al.
sequence of updates that is applied. This phenom- 2016), and the work of Takada et al., who designed
enon, named coalescence, can be observed in asynchronous self-reproducing loops (Takada
Asynchronous Cellular Automata 89
et al. 2007a). These are only some entry point to has led to propose the use of some measure-
the literature on this topic, and we refer again to theoretic tools to define m-asynchronous cellular
our survey paper for a wider scope (Fatès 2014a). automata to include the cases of nonhomogeneous
probabilities of updating, infinitesimal ones,
etc. (Dennunzio et al. 2012, 2013).
Openings To complete this point, let us underline that
Bouré et al. have proposed to examine the case
We have seen that the randomness involved in the where the randomness occurs not on the moments
asynchronous updating create an amazing source of updating but on the possibility to miss the
of new questions on cellular automata. After more information from one or several neighbors
than two decades of continued efforts, this topic (Bouré et al. 2012). Interestingly, the study of
shows signs of maturity, and although it remains these new updating schemes, named b- and
in large part a terra incognita, there are some g-asynchronous updating schemes, shows that
insights on how asynchronous cellular automata their behavior partially overlaps with
can be studied with a theoretical point of view. a-asynchronous systems but also reveals some
A set of analytical tools are now available, and novel and unexpected behaviors (e.g., other rules
when the analysis fails to answer all the questions, show a phase transitions).
one can carry out numerical simulations. Readers
should now be convinced that asynchronous cel- Asynchronous Models
lular automata are by no means some “exotic” The theoretical results obtained so far do not tell
mathematical objects but constitute a thriving us what is a good model of asynchrony in general.
field of research. The elements we presented Since cellular automata are defined with a discrete
here are only a small part of this field and should of time and space, it is not straightforward to
be completed by a more extensive bibliographical decide a priori to use a synchronous updating, or
work. Before closing this text, we want to present a fully asynchronous one, or a partially synchro-
a few questions that are currently investigated. nous one. In fact, the most reasonable position
would be to test various updating schemes on a
Defining Asynchrony rule and to examine if it is robust or sensitive to
As mentioned in the introduction, asynchrony is a these modifications. Although this critical atti-
concept that can be defined with a great variety of tude has been quite rare so far, a good example
forms. For example, the notion of a-asynchronous of such a study has been provided by Grilo and
updating scheme needs to be generalized to go Correia, who made a systematic study of the
beyond the simple homogeneous finite case. This effects of the updating in the spatially extended
Asynchronous Cellular Automata, Fig. 9 Rapid coa- middle). The right diagram shows the agreement and dis-
lescence phenomenon for ECA 46 with fully asynchronous agreement of the two systems. Cells in white and light gray,
updating. The same updates are applied on two systems respectively, show agreement on state 0 or 1, while red and
with two different random initial conditions (left and green show disagreement (the order is not important)
90 Asynchronous Cellular Automata
evolutionary games. This question rose after the asynchronous computations (Lee et al. 2016b;
criticisms made by Huberman and Glance (1993) Peper et al. 2010). They represent a potential
to the model proposed by Nowak and May (1992). source of major technical innovations, in particu-
We think that exploring more systematically these lar with the possibility of implementing such cir-
issues on real-world models could help us under- cuits with DNA reaction-diffusion systems
stand to which extent the simplifications operated (Yamashita et al. 2017) or single electron tunnel-
in a model are justified or are a potential source of ing techniques (Lee et al. 2016a).
artifacts (see Fatès (2014a) for other examples).
Cornforth D, Green DG, Newth D (2005) Ordered asyn- of view. To appear in a monography edited by Andrew
chronous processes in multi-agent systems. Physica Adamatzky – Preprint available on the HAL server, id:
D 204(1–2):70–82 hal-01571847
de Mas JF (2017) Coalescence in fully asynchronous ele- Fukś H (2002) Nondeterministic density classification with
mentary cellular automata. Technical report, HAL pre- diffusive probabilistic cellular automata. Phys Rev
print hal-01627454 E 66(6):066106
de Oliveira PPB (2014) On density determination with Fukś H, Fatès N (2015) Local structure approximation as a
cellular automata: results, constructions and directions. predictor of second-order phase transitions in asyn-
J Cell Autom 9(5–6):357–385 chronous cellular automata. Nat Comput
Dennunzio A, Formenti E, Manzoni L (2012) Computing 14(4):507–522
issues of asynchronous CA. Fundamenta Informaticae Gács P (2001) Deterministic computations whose history
120(2):165–180 is independent of the order of asynchronous updating.
Dennunzio A, Formenti E, Manzoni L, Mauri G (2013) Technical report – arXiv:cs/0101026
m-asynchronous cellular automata: from fairness to Gerin L (2017) Epidemic automaton and the eden model:
quasi-fairness. Nat Comput 12(4):561–572 various aspects of robustness. Text to appear in a mono-
Dennunzio A, Formenti E, Manzoni L, Mauri G, Porreca graphy on probabilistic cellular automata. Springer
AE (2017) Computational complexity of finite asyn- Huberman BA, Glance N (1993) Evolutionary games and
chronous cellular automata. Theor Comput Sci computer simulations. Proc Natl Acad Sci U S A
664:131–143 90:7716–7718
Fatès N (2009) Asynchronism induces second order phase Kari J, Taati S (2015) Statistical mechanics of surjective
transitions in elementary cellular automata. J Cell cellular automata. J Stat Phys 160(5):1198–1243
Autom 4(1):21–38 Lee J, Peper F (2008) On brownian cellular automata. In:
Fatès N (2010) Does life resist asynchrony? In: Adamatzky Adamatzky A, Alonso-Sanz R, Lawniczak AT, Martí-
A (ed) Game of life cellular automata. Springer, nez GJ, Morita K, Worsch T (eds) Proceedings of
London, pp 257–274 automata 2008. Luniver Press, Frome, pp 278–291
Fatès N (2013a) A note on the classification of the most Lee J, Adachi S, Peper F, Morita K (2004) Asynchronous
simple asynchronous cellular automata. In: Kari J, game of life. Phys D 194(3–4):369–384
Kutrib M, Malcher A (eds) Proceedings of autom- Lee J, Peper F, Cotofana SD, Naruse M, Ohtsu M,
ata’13, volume 8155 of lecture notes in computer sci- Kawazoe T, Takahashi Y, Shimokawa T, Kish LB,
ence. Springer, Netherlands, pp 31–45. https://doi.org/ Kubota T (2016a) Brownian circuits: designs. Int
10.1007/s11047-013-9389-2 J Unconv Comput 12(5–6):341–362
Fatès N (2013b) Stochastic cellular automata solutions to Lee J, Peper F, Leibnitz K, Ping G (2016b) Characteriza-
the density classification problem – when randomness tion of random fluctuation-based computation in cellu-
helps computing. Theory Comput Syst 53(2):223–242 lar automata. Inf Sci 352–353:150–166
Fatès N (2014a) A guided tour of asynchronous cellular Louis P-Y (2015) Supercritical probabilistic cellular
automata. J Cell Autom 9(5–6):387–416 automata: how effective is the synchronous updating?
Fatès N (2014b) Quick convergence to a fixed point: a note Nat Comput 14(4):523–534
on asynchronous elementary cellular automata. In: Macauley M, Mortveit HS (2010) Coxeter groups and
Was J, Sirakoulis GC, Bandini S (eds) Proceedings of asynchronous cellular automata. In: Bandini S,
ACRI’14, volume 8751 of lecture notes in computer Manzoni S, Umeo H, Vizzari G (eds) Proceedings of
science. Krakow, Poland, Springer, pp 586–595 ACRI’10, volume 6350 of lecture notes in computer
Fatès N, Gerin L (2009) Examples of fast and slow con- science. Springer, Ascoli Piceno, Italy, pp 409–418
vergence of 2D asynchronous cellular systems. Old Macauley M, Mortveit HS (2013) An atlas of limit set
City Publishing. J Cell Autom 4(4):323–337. http:// dynamics for asynchronous elementary cellular autom-
www.oldcitypublishing.com/journals/jca-home/ ata. Theor Comput Sci 504:26–37. Discrete mathemat-
Fatès N, Morvan M (2005) An experimental study of ical structures: from dynamics to complexity
robustness to asynchronism for elementary cellular Macauley M, McCammond J, Mortveit HS (2008) Order
automata. Complex Syst 16:1–27 independence in asynchronous cellular automata. J Cell
Fatès N, Morvan M, Schabanel N, Thierry E (2006a) Fully Autom 3(1):37–56
asynchronous behavior of double-quiescent elementary Mairesse J, Marcovici I (2014) Around probabilistic cellu-
cellular automata. Theor Comput Sci 362:1–16 lar automata. Theor Comput Sci 559:42–72. Non-
Fatès N, Regnault D, Schabanel N, Thierry E (2006b) uniform cellular automata
Asynchronous behavior of double-quiescent elemen- Moore EF (1962) Machine models of self-reproduction.
tary cellular automata. In: Correa JR, Hevia A, Kiwi Proc Symp Appl Math 14:17–33. (Reprinted in Essays
MA (eds) Proceedings of LATIN 2006, volume 3887 of on cellular automata, Burks AW (ed), University of
lecture notes in computer science. Valdivia, Chile, Illinois Press, 1970)
Springer, pp 455–466 Morita K (2008) Reversible computing and cellular
Fatès N, Sethi B, Das S (2017) On the reversibility of ecas automata – a survey. Theor Comput Sci
with fully asynchronous updating: the recurrence point 395(1):101–131
92 Asynchronous Cellular Automata
Nakamura K (1974) Asynchronous cellular automata and Sethi B, Roy S, Das S (2016) Asynchronous cellular
their computational ability. Syst Comput Controls automata and pattern classification. Complexity
5(5):58–66 21:370–386
Nakamura K (1981) Synchronous to asynchronous transfor- Silva F, Correia L (2013) An experimental study of noise
mation of polyautomata. J Comput Syst Sci 23(1):22–37 and asynchrony in elementary cellular automata with
Nowak MA, May RM (1992) Evolutionary games and sampling compensation. Nat Comput 12(4):573–588
spatial chaos. Nature 359:826–829 Silva F, Correia L, Christensen AL (2015) Modelling syn-
Peper F, Lee J, Adachi S, Mashiko S (2003) Laying out chronisation in multirobot systems with cellular autom-
circuits on asynchronous cellular arrays: a step towards ata: analysis of update methods and topology
feasible nanocomputers? Nanotechnology 14(4):469 perturbations. In: Sirakoulis GC, Adamatzky A (eds)
Peper F, Lee J, Isokawa T (2010) Brownian cellular autom- Robots and lattice automata, volume 13 of emergence,
ata. J Cell Autom 5(3):185–206 complexity and computation. Springer International
Ramos AD, Leite A (2017) Convergence time and phase Publishing, Springer. pp 267–293
transition in a non-monotonic family of probabilistic Takada Y, Isokawa T, Peper F, Matsui N (2007a) Asyn-
cellular automata. J Stat Phys 168(3):573–594 chronous self-reproducing loops with arbitration capa-
Regnault D (2013) Proof of a phase transition in probabi- bility. Phys D Nonlinear Phenom 227(1):26–35
listic cellular automata. In: Béal MP and Carton O (eds) Takada Y, Isokawa T, Peper F, Matsui N (2007b) Asyn-
Proceedings of developments in language theory, vol- chronous self-reproducing loops with arbitration capa-
ume 7907 of lecture notes in computer science. bility. Phys D 227(1):26–35
Springer, Marne-la-Vallée, France, pp 433–444 Vichniac GY (1984) Simulating physics with cellular
Regnault D, Schabanel N, Thierry E (2009) Progresses in automata. Phys D Nonlinear Phenom 10(1):96–116
the analysis of stochastic 2D cellular automata: a study Vielhaber M (2013) Computation of functions on n bits by
of asynchronous 2D minority. Theor Comput Sci asynchronous clocking of cellular automata. Nat
410(47–49):4844–4855 Comput 12(3):307–322
Regnault D, Schabanel N, Thierry E (2010) On the analysis Wacker S, Worsch T (2013) On completeness and decid-
of “simple” 2d stochastic cellular automata. Discrete ability of phase space invertible asynchronous cellular
Math Theor Comput Sci 12(2):263–294 automata. Fundam Informaticae 126(2–3):157–181
Rouquier J-B, Morvan M (2009) Coalescing cellular Wolfram S (1985) Twenty problems in the theory of cellu-
automata: synchronization by common random source lar automata. Phys Scr T9:170
for asynchronous updating. J Cell Autom 4(1):55–78 Worsch T (2013) Towards intrinsically universal asynchro-
Schönfisch B, de Roos A (1999) Synchronous and asyn- nous CA. Nat Comput 12(4):539–550
chronous updating in cellular automata. Biosystems Yamashita T, Isokawa T, Peper F, Kawamata I, Hagiya
51:123–143 M (2017) Turing-completeness of asynchronous non-
Sethi B, Fatès N, Das S (2014) Reversibility of elementary camouflage cellular automata. In: Dennunzio A,
cellular automata under fully asynchronous update. In: Formenti E, Manzoni L, Porreca AE (eds) Proceed-
Gopal TV, Agrawal M, Li A, Cooper B (eds) Proceed- ings of AUTOMATA 2017, volume 10248 of lecture
ings of TAMC’14, volume 8402 of lecture notes in notes in computer science. Springer, Milan, Italy,
computer science. Springer, Chennai, India, pp 39–49 pp 187–199
Pauli operator The three Pauli operators are sx ¼
Quantum Cellular Automata 0 1 0 i 1 0
,sy ¼ ,sz ¼
1 0 i 0 0 1
Karoline Wiesner Phase gate The one-qubit unitary gate
School of Mathematics, University of Bristol, 1 0
Bristol, UK U¼
0 eif
QMA complexity class Quantum Merlin-
Arthur, the class of decision problems such
Article Outline that a “yes” answer can be verified by a
1-message quantum interactive proof
Glossary (verifiable in BQP).
Definition of the Subject Quantum Turing machine A quantum version
Introduction of a Turing machine – an abstract computa-
Cellular Automata tional model able to compute any computable
Early Proposals sequence.
Models of QCA Qubit Two-state quantum system, representable
Computationally Universal QCA as vector a|0i + b|1i in complex space with
Modeling Physical Systems a2 + b2 = 1.
Implementations Schrödinger picture Time evolution is
Future Directions represented by a quantum state evolving in
Bibliography time according to a time-independent unitary
operator acting on it.
Glossary Space homogeneous The transition function/
update table is the same for each cell.
BQP complexity class Bounded error, quantum Swap operation The one-qubit unitary gate
probabilistic, the class of decision problems 0 1
solvable by a quantum computer in polyno- U¼
1 0
mial time with an error probability of at most Time homogeneous The transition function/
one third. update table is time independent.
Configuration The state of all cells at a given Update table Takes the current state of a cell and
point in time. its neighborhood as an argument and returns
Hadamard gate The one-qubit unitary gate
the cell’s state at the next time step.
1 1
U ¼ p1ffiffi2
1 1
Heisenberg picture Time evolution is Definition of the Subject
represented by observables (elements of an
operator algebra) evolving in time according Quantum cellular automata (QCA) are a general-
to a unitary operator acting on them. ization of (classical) cellular automata (CA) and in
Neighborhood All cells with respect to a given particular of reversible CA. The latter are
cell that can affect this cell’s state at the next reviewed shortly. An overview is given over
time step. A neighborhood always contains a early attempts by various authors to define one-
finite number of cells. dimensional QCA. These turned out to have seri-
ous shortcomings which are discussed as well.
Various proposals subsequently put forward by a Neumann presented a detailed analysis of the
number of authors for a general definition of one- above question in his book Theory of Self-
and higher-dimensional QCA are reviewed, and Reproducing Automata (von Neumann 1966).
their properties such as universality and revers- Thus, von Neumann initiated the field of cellular
ibility are discussed. automata. He also made central contributions to the
Quantum cellular automata (QCA) are a quanti- mathematical foundations of quantum mechanics,
zation of classical cellular automata (CA); and in a sense von Neumann’s quantum logic
d-dimensional arrays of cells with a finite- ideas were an early attempt at defining a computa-
dimensional state space; and a local, spatially homo- tional model of physics. But he did not pursue this
geneous, discrete-time update rule. For QCA, each and did not go in the directions that have led to
cell is a finite-dimensional quantum system, and the modern ideas of quantum computing in general or
update rule is unitary. CA as well as some versions quantum cellular automata in particular.
of QCA have been shown to be computationally The idea of quantum computation is generally
universal. Apart from a theoretical interest in a attributed to Feynman who, in his now famous
quantized version of CA, QCA are a natural frame- lecture in 1981, proposed a computational scheme
work for what is most likely going to be the first based on quantum mechanical laws (Feynman
application of quantum computers – the simulation 1982). A contemporary paper by Benioff contains
of quantum physical systems. In particular, QCA the first proposal of a quantum Turing machine
are capable of simulating quantum dynamical sys- (Benioff 1980). The general idea was to devise a
tems whose dynamics are uncomputable by classi- computational device based on and exploiting quan-
cal means. QCA are now considered one of the tum phenomena that would outperform any classi-
standard models of quantum computation next to cal computational device. These first proposals
quantum circuits and various types of were sequentially operating quantum mechanical
measurement-based quantum computational machines imitating the logical operations of classi-
models. (For details on these and other aspects of cal digital computation. The idea of parallelizing the
quantum computation, see the article by Kendon in operations was found in classical cellular automata.
this encyclopedia.) Unlike their classical counter- However, how to translate cellular automata into a
part, an axiomatic, all-encompassing definition of quantum mechanical framework turned out not to
(higher-dimensional) QCA is still missing. be trivial. And to a certain extent how to do this in
general remains an open question until today.
The study of quantum cellular automata
Introduction (QCA) started with the work of Grössing and
Zeilinger who coined the term QCA and provided
Automata theory is the study of abstract comput- a first definition (Grössing and Zeilinger 1988).
ing devices and the class of functions they can Watrous developed a different model of QCA
perform on their inputs. The original concept of (Watrous 1995). His work led to further studies
cellular automata is most strongly associated with by several groups (van Dam 1996; Dürr and
John von Neumann (1903, †1957), a Hungarian Santha 2002; Dürr et al. 1997). Independently of
mathematician who made major contributions to a this, Margolus developed a parallelizable quan-
vast range of fields including quantum mechanics, tum computational architecture building on
computer science, functional analysis, and many Feynman’s original ideas (Margolus 1991). For
others. According to Burks, an assistant of von various reasons to be discussed below, none of
Neumann (1966), von Neumann had posed the these early proposals turned out to be physical.
fundamental questions: “What kind of logical The study of QCA gained new momentum with
organization is sufficient for an automaton to the work by Richter, Schumacher, and Werner
reproduce itself?” It was Stanislaw Ulam who (Richter 1996; Schumacher and Werner 2004)
suggested to use the framework of cellular autom- and others (Arrighi and Fargetton 2007; Arrighi
ata to answer this question. In 1966, von et al. 2007a; Perez-Delgado and Cheung 2007)
Quantum Cellular Automata 95
who avoided unphysical behavior allowed by the Quantum Cellular Automata, Table 1 Update table for
early proposals (Arrighi et al. 2007a; Schumacher CA rule “110” (the second row is the decimal number
“110” in binary notation)
and Werner 2004). It is important to notice that in
spite of the over two-decade-long history of QCA, M 110 ¼ 111
0
110 101 100 011 010 001 000
1 1 0 1 1 1 0
Early Proposals
translation matrix Tk for some k ℤ, times a crucial to avoid an infinite product of unitaries
phase” (Meyer 1996a). and, thus, to obtain a well-defined QCA.
Grössing and Zeilinger also introduced QCA The Watrous QCA, however, allows for non-
where the unitarity constraint is relaxed to only physical dynamics. It is possible to define tran-
approximate unitarity. After each update, the con- sition functions that do not represent unitary
figuration can be normalized which effectively evolution of the configuration, either by produc-
causes nonlocal interactions. ing superpositions of configurations which do
The properties of Grössing-Zeilinger QCA not preserve the norm or by inducing a global
were studied by Grössing and coworkers in transition function which is not unitary. This
some more detail in following years (see Fussy leads to nonphysical properties such as super-
et al. 1993, p. and references therein). This luminal signaling (Schumacher and Werner
pioneering definition of QCA, however, was not 2004). The set of Watrous QCA is not closed
studied much further, mostly because the “non- under composition and inverse (Schumacher
local” behavior renders the Grössing-Zeilinger and Werner 2004).
definition nonphysical. In addition, it has little in Watrous defined a restricted class of QCA by
common with the concepts developed in quantum introducing a partitioning scheme.
computation later on. The Grössing-Zeilinger def-
inition really concerns what one would call today Definition (Partitioned Watrous QCA) A
a quantum random walk (for further details, see partitioned Watrous QCA is a Watrous QCA with
the review by Kempe 2003). S = Sl Sc Sr for finite sets Sl, Sc, and Sr
The first model of QCA researched in depth was and matrix L of size S S. For any state, s = (sl,
that introduced by Watrous (1995), whose ideas sc, sr) S define transition function f as
were further explored by van Dam (1996), Dürr
et al. (1997), Dürr and Santha (2002), and Arrighi f ðs1 , s2 , s3 , sÞ ¼ Lðsl3 , sm2 , sr1 , sÞ, (2)
(2006). A Watrous QCA is defined over an infinite
1-dimensional lattice, a finite set of states including a with matrix element Lsi,sj.
quiescent state. The transition function maps a
neighborhood of cells to a single quantum state In a partitioned Watrous QCA, each cell is
instantaneously and simultaneously. divided into three sub-cells – left, center, and
right. The neighborhood scheme is then a nearest-
Definition (Watrous QCA) A Watrous QCA is a neighbor interaction confined to each cell. The
four-tuple ðL,S,N ,f Þ which consists of (1) a transition function consists of a unitary acting on
1-dimensional lattice L ℤ, (2) a finite set of each partitioned cell and swap operations among
cell states S including a quiescent state e, (3) a sub-cells of different cells. Figure 2 illustrates the
finite neighborhood scheme N , and (4) a local swap operation between neighboring cells.
transition function f :SN ! ℋS. For the class of partitioned Watrous QCA,
Watrous provides the first proof of computational
Here, ℋS denotes the Hilbert space spanned universality of a QCA by showing that any quan-
by the cell states S. This model can be viewed as a tum Turing machine can be efficiently simulated
direct quantization of a CA where the set of pos- by a partitioned Watrous QCA with constant slow-
sible configurations of the CA is extended to down and that any partitioned Watrous QCA can
include all linear superpositions of the classical be simulated by a quantum Turing machine with
cell configurations and the local transition func- linear slowdown.
tion now maps the cell configurations of a given
neighborhood to a quantum state. One cell is Theorem (Watrous 1995) Given any quantum
labeled “accept” cell. The quiescent state is used Turing machine MTM, there exists a partitioned
to allow only a finite number of states to be active Watrous QCA MCA which simulates MTM with
and renders the lattice effectively finite. This is constant slowdown.
98 Quantum Cellular Automata
Models of QCA
Reversible QCA
Schumacher and Werner used the Heisenberg picture
rather than the Schrödinger picture in their model
(Schumacher and Werner 2004). Thus, instead of
associating a d-level quantum system with each
Quantum Cellular Automata, Fig. 2 Each cell is cell, they associated an observable algebra with
divided into three sub-cells labeled l, c, and r for left, center, each cell. Taking a quasi-local algebra as the tensor
and a right, respecti vely. The update rule consists of
product of observable algebras over a finite subset of
swapping left and right sub-cells of neighboring cells and
then updating each cell internally using a unitary operation cells, a QCA is then a homomorphism of the quasi-
acting on the left, center, and right part of each cell local algebra, which commutes with lattice transla-
tions and satisfies locality on the neighborhood.
The observable-based approach was first used
Theorem (Watrous 1995) Given any partitioned in Richter (1996) with focus on the irreversible
Watrous QCA MCA, there exists a quantum Turing case. However, this definition left questions open
machine MTM which simulates MCA with linear such as whether the composition of two QCA will
slowdown. again form a QCA. The following definition does
avoid this uncertainty.
Watrous’ model was further developed by Consider an infinite d-dimensional lattice L ℤd
van Dam (1996), who defined a QCA as an of cells x ℤ2, where each cell is associated with
assignment of a product vector to every basis the observable algebra Ax and each of these algebras
state in the computational basis. Here the qui- is an isomorphic copy of the algebra of complex
escent state is eliminated, and thus, the QCA is d d-matrices. When L ℤd is a finite subset of
made explicitly finite. Van Dam showed that the cells, denote by A(L) the algebra of observables
finite version is also computationally universal. belonging to all cells in L, i.e., the tensor product
N
Efficient algorithms to decide whether a given x LAx. The completion of this algebra is called
1-dimensional QCA is unitary was presented by a quasi-local algebra and will be denoted by A(ℤd).
Dürr et al. (1997), Dürr and Santha (2002). Due
to substantial shortcomings such as non- Definition (Reversible QCA) A quantum cellu-
physical behavior, these early proposals were lar automaton with neighborhood scheme N
replaced by a second wave of proposals to be ℤd is a homomorphism T:A(ℤd) ! A(ℤd) of the
discussed below. quasi-local algebra, which commutes with
Today, there is not a generally accepted QCA lattice translations, and satisfies the locality con-
model that has all the attributes of the CA model: dition T ðAðLÞÞ T ðAðL þ N ÞÞ for every finite
unique definition, simple to describe, and compu- set L ℤd. The local transition rule of a cellular
tationally powerful. In particular, there is no axi- automaton is the homomorphism T0:A0 ! A(N).
omatic definition, contrary to its classical
counterpart, that yields an immediate way of Schumacher and Werner presented and proved
constructing/enumerating all of the instances of the following theorem on one-dimensional QCA.
this model. Rather, each set of authors defines
QCA in their own particular fashion. Theorem (Structure Theorem (Schumacher
The states s S are basis states spanning a and Werner 2004)) Let T be the global transition
finite-dimensional Hilbert space. At each point homomorphism of a one-dimensional nearest-
in time, a cell represents a finite-dimensional neighbor QCA on the lattice ℤd with single-cell
quantum system in a superposition of basis states. algebra A0 = Md. Then T can be represented in
The unitary operators represent the discrete-time the generalized Margolus partitioning scheme,
evolution of strictly finite propagation speed. i.e., T restricts to an isomorphism
Quantum Cellular Automata 99
T : Að□Þ ! ℬs , (3)
sS