Diagrammatic approach to cellular automata

and the emergence of form with inner structure

Vladimir Garcı́a-Morales
Departament de Termodinàmica, Universitat de València,
E-46100 Burjassot, Spain
garmovla@uv.es

We present a diagrammatic method to build up sophisticated cellular automata (CAs) as models of

complex physical systems. The diagrams complement the mathematical approach to CA modeling,
whose details are also presented here, and allow CAs in rule space to be classified according to their
hierarchy of layers. Since the method is valid for any discrete operator and only depends on the
alphabet size, the resulting conclusions, of general validity, apply to CAs in any dimension or order
in time, arbitrary neighborhood ranges and topology. We provide several examples of the method,
illustrating how it can be applied to the mathematical modeling of the emergence of order out of
disorder. Specifically, we show how the the majority CA rule can be used as a building block to
construct more complex cellular automata in which separate domains (with substructures having
different dynamical properties) are able to emerge out of disorder and coexist in a stable manner.

1
I. INTRODUCTION

Cellular automata (CAs) [1–11] are fully discrete dynamical systems with finite local and global
phase spaces, evolving in discrete time and space. CAs provide an efficient and computationally
inexpensive means of modeling complex physical systems. Although they usually constitute crude
models of the physical reality, the general question arises as whether there exist increasingly
sophisticated CA models able to capture both essential and subtle properties of experimental physical
systems. A good deal of effort has been devoted to the identification of appropriate CA models in
the huge computational space for specific needs [5]. Building on very recent work [12, 13] this article
advances a general and systematic method as a pathway to tackle this problem.

Let a CA rule work on an alphabet of p symbols, where p ∈ N, p ≥ 2 is the alphabet size. If p

is a composite number (i.e. not a prime number), such CA rule is p-decomposable [12]. We have
presented in [12] a systematic way in which elementary layers can be superimposed to construct
graded CA rules where the layers are all independent. This article completes the theory in [12] by
presenting the general construction of non-graded p-decomposable CA rules. In these CAs the layers
interact in complex ways and are generally coupled yielding a wide variety of classes of models and
possibilities. These couplings, however, can themselves be designed so that the prominent features
and correlations of emergent structures (that arise when iterating the models) can be predicted
beforehand. Although the output of these CA models is highly nontrivial, the systematic method
presented makes them amenable to mathematical treatment and the resulting models are often easy
to formulate and study. Our construction is aided by diagrams which make explicit how a given layer
decomposes in sublayers and how the layers are coupled with each other. These diagrams specify
universality classes of CA models that can then be automatically translated into mathematical
expressions so that specific CA models belonging to these classes are constructed. This work,
thus, completes the development in [12]. It provides, together with that paper, a toolbox for the
analysis and construction of sophisticated CA rules that can be used in the analysis of complex systems.

The outline of this article is as follows. In Section II we present the general theory, the concept of
p-decomposability and the diagrammatic approach that aids in the construction of models for complex
systems. The relationship between the diagrams and mathematical expressions is explained. We then
classify all CA rules working on alphabets where p = p0 p1 has two divisors, p0 and p1 , only. In
Section III we give several examples where we show how CA rules can be endowed with algebraic
structure, explaining how this structure is linked to the observed spatiotemporal evolution of the CA.
The article is completed by giving a number of examples in which we use the majority CA rule as
ground layer to build more complex CA, according to the method advanced in this article. We show
how structures containing substructures of smaller size (whose dynamical behavior can be specified)
arise out of disorder, thus providing a general framework to model natural processes in which form
with inner structure emerges.

2
II. CELLULAR AUTOMATA: DIAGRAMMATIC APPROACH TO
p-DECOMPOSABLE RULES

A. General theory

We first give a brief self-contained account of the main result in [12] that we shall need, and then
proceed to describe the diagrammatic approach. In this work S shall denote the set of integers in the
interval [0, p − 1], with p ≥ 2 being a natural number called the alphabet size.
Let us first consider, for simplicity, a ring with Ns sites (the following theory easily generalizes to
higher dimensions and other topologies as it shall be shown in the examples). The initial condition
of a CA is the vector x0 = (x00 , ..., xN 0
s −1
), with xj0 ∈ S, ∀j ∈ [0, Ns − 1] (here j ∈ Z specifies,
modulo Ns , the position of a site in the ring). At each discrete time t, the state of the CA is given
by xt = (x0t , ..., xtNs −1 ) (with xjt ∈ S). The local spatiotemporal dynamics of the CA is a map
f : S l+r+1 → S which provides for all j the value xjt+1 as a function of some xj+k t , where k ∈ [−r, l]
specifies the position of the sites within the neighborhood of j. The quantity ρ ≡ l + r + 1 = 2ξ is the
neighborhood range and ξ is the (average) neighborhood radius. A neighborhood is called symmetric
if l = r = ξ. The CA map f : S l+r+1 → S is explicitly given by the universal expression [12]
l
! $ % $ %
R R
xjt+1 = dp pk+r xtj+k , R =
X
Pl k+r j+k
−p Pl k+r j+k
(1)
k=−r p k=−r p xt p1+ k=−r p xt
for each j ∈ [0, Ns − 1]. Here, we have introduced the digit function [12, 14, 15] which is defined, for
p ∈ N, k ∈ Z and x ∈ R as    
x x
dp (k, x) = k − p k+1 (2)
p p
with b. . .c denoting the floor (lower closest integer) function. In Eq. (1) we have also introduced the
l+r+1
Wolfram code, R ∈ Z, R ∈ [0, pp − 1]
pr+l+1
X−1
R≡ an pn . (3)
n=0

where the an ’s (all ∈ S) are the coefficients which specify the CA rule. These quantities are the
components of the rule vector a. We see that R is obtained from a through Eq. (3). Conversely, we
can obtain a from the knowledge of R, since

an = dp (n, R) (4)

i.e., the coefficients an are the digits of the radix-p representation of R. For any x ∈ R we have

dp (k, pk x) = dp (0, x) = dp (0, x + np) (5)

and a simple calculation shows that Eq. (1) is equivalent to

pl+r+1
X−1 l
!
xjt+1 = pk+r xj+k
X
an dp n− t ,1 (6)
n=0 k=−r

3
since, at time t, only the term an for which n = lk=−r pk+r xj+k
P
t contributes to the sum, all others
being zero. Note that for n, m ∈ Z, dp (n − m, 1) is a representation of the Kronecker delta δnm : We
have δnm = 1 for n = m and δnm = 0 for n 6= m.
We use the notation [9, 16]
l
!
l r j k+r j+k
X
Rp (xt ) ≡ dp p xt , R (7)
k=−r
to refer concisely to a CA rule on p symbols, neighborhood radii l and r and Wolfram code R. Thus,
for example 1 3012 is Wolfram’s elementary CA rule with code 30, and neighborhood consisting of one
site to the left and one to the right of the site which is updated at the next time step. There are 256
such rules with p = 2 and l = r = 1.
Totalistic CA rules are a subset of the above ones and depend only on the sum over neighborhood
values. In this specific case, the universal map above becomes,
l
!
j j j+k
X
xt+1 ≡ l RTpr (xt ) = dp xt , R (8)
k=−r
or, alternatively,
(p−1)(l+r+1) l
!
xjt+1 xj+k
X X
= a n dp n− t ,1 (9)
n=0 k=−r

The digit function has a decomposition property [12] which constitutes the main mathematical
tool of this article. Let p be a composite number, i.e., the product of N factors p = p0 p1 . . . pN −1 =
QN −1 j j j
h=0 pm . Let xt ∈ [0, p − 1]. We have xt = dp (0, xt ). Then,
$ % $ %
j j xjt j xjt
dp (0, xt ) = xt − p = xt − p0 p1 . . . pN −1
p p0 p1 . . . pN −1
$ % $ % $ % $ %
j j j j
x x x x
= xjt − p0 t
+ p0 t
− p0 p1 t
+ p0 p1 t
− ...
p0 p0 p0 p1 p0 p1
$ %
xjt
−p0 p1 . . . pN −1
p0 p1 . . . pN −1
! !

j
 xjt xjt
= dp0 0, xt + p0 dp1 0, + . . . + p0 p1 . . . pN −2 dpN −1 0,
p0 p0 p1 . . . pN −2
N −1 h−1
!
X xjt Y
= dph 0, Qh−1 pm (10)
h=0 m=0 pm m=0
Hence, by defining local integer variables that take values in the interval [0, ph − 1]
!
(h),j xjt
yt ≡ dph 0, Qh−1 h = 0, 1, . . . , N − 1 (11)
m=0 pm
we can write [12]
N −1 h−1
(0),j (1),j (2),j (N −1),j (h),j
xjt = yt
X Y
+ p0 yt + p0 p1 yt + . . . + p0 p1 . . . pN −2 yt = yt pm (12)
h=0 m=0

4
This expression coincides with Cantor’s mixed-radix representation of the integer number xjt in terms
of the factors ph [17]. Of course, at time t + 1 we have, similarly,
N −1 h−1
(h),j
xjt+1 =
X Y
yt+1 pm (13)
h=0 m=0

and, hence, the transformation xjt → xjt+1 on p symbols is equivalent to the transformation
(0),j (h),j (N −1),j (0),j (h),j (N −1),j (h),j
(yt , . . . , yt , . . . , yt ) → (yt+1 , . . . , yt+1 , . . . , yt+1 ) on N -tuples of variables yt ∈
[0, ph − 1]. From Eqs. (1), (11) and (13), we have
!
(h),j 1
yt+1 = dph 0, Qh−1 xjt+1 (14)
p
m=0 m
l
"N −1
X (s),j+k s−1
# !!
1 X
k+r
Y
= dph 0, Qh−1 dp p yt pm , R (15)
m=0 p m k=−r s=0 m=0

which we call the h-layer value of the CA l Rpr (xjt ).

We note that the above decomposition is independent of the topology and dimensionality of the
(h),j
model CA and is only based on the alphabet structure. If yt+1 on layer h only depends at time t + 1
(h),j
on the yt variables of that same layer h, Eq. (15) reduces to
l
!!
(h),j 1 X
k+r (h),j+k
yt+1 = dph 0, Qh−1 dp p yt , R (16)
m=0 pm k=−r
l l
! ! !
X
k+r (h),j+k
X xjt
= dph ph yt (h)
, Aph = dph pk+r
h dph 0, Qh−1 , A(h)
ph
k=−r k=−r m=0 pm

(h)
i.e. each layer value behaves as a CA with Wolfram code Aph that is independent of the other layers.
We call them layer CAs. If all layers are layer CAs the CA rule l Rpr is called a graded rule [12]. In
general, however, the layer values are nonlinearly coupled as seen in Eq. (15). In this general case we
say that l Rpr is a p-decomposable rule. All graded CA rules are p-decomposable rules but the contrary
is not true.
(h)
We observe that any factor ph of p can also be composite. In that case, any layer Aph can be
further decomposed in sublayers following the above scheme. We thus introduce the pair of integers
(g, h) called the rank g and the layer index h. We order the layers in ranks. If the rank of a layer is
g, the rank of a sublayer is g + 1. The rank g = 0 is the CA model of the system. On each rank g,
the layer with h = 0 is called the ground layer and any layer with h > 0 a lifted layer. If the layers
are CA then we speak about a ground CA layer and a lifted CA layer respectively.
The following are the rules to construct diagrams:

• At the top of the diagram (rank g = 0) we put the CA rule l Rpr that constitutes the model.

• Layers with increasing rank are arranged vertically from top to bottom.

• Layers with increasing index h are arranged from left to right. Thus, the ground layer is the
leftmost node of a certain rank.

5
 • We label any layer CA with a solid node and an arbitrary layer that is not a CA with a hollow
node.
• Undirected edges are vertically drawn to relate layers of rank g to their decompositions into
layers of rank g + 1. Such edges are to be read as “layer with rank g decomposes as layers
A, B, C, . . . with rank g + 1”.
• We omit the radii (l and r) of any layers different to the model (at the top) of the diagram. If
the radii of the other nodes are layer-dependent, i.e., if one has l(g,h) , r(g,h) , for the radii, then
the model has l = max l(g,h) and r = max r(g,h) . For each layer, only the ’coordinates’ (g, h)
and the alphabet size ph are indicated.
• Directed edges indicate the dependence of a layer on other layers. An outgoing arrow from a
node A to a node B means that A influences B but is independent of B. The incoming arrow
at B means that the value of B depends on the value (or configuration) at A.
In Fig. 1 a) we see the full decomposition of a graded CA rule on p = 630 symbols. It contains
only undirected edges and layer CAs. The rule has two nonzero ranks, g = 1 with a ground layer CA
on p0 = 30 symbols and a lifted layer CA on p1 = 21 symbols. The former layer is further decomposed
(2,0) (2,1) (2,2) (1,1)
into three rank-2 layer CAs, A3 , A5 , A2 . Layer A21 is decomposed in two rank-2 layer CAs.
We note that we could have stopped the decomposition at any of the nodes with rank g = 1. If the
decomposition is shown up to the non-decomposable sublayers (over a prime number of symbols) we
say that the CA diagram is fully decomposed. There is a unique fully decomposed diagram for each
given CA model. If some layers over composite ph are not further decomposed, we say that the CA
diagram is partially decomposed. Any fully decomposed diagram can be thus embedded in a partially
decomposed diagram and the former are a subset of the latter. Often, only partial decompositions are
relevant. In Fig. 1 c) a partially decomposed diagram that contains the one of Fig. 1 a) is plotted.
In Fig. 1 b) a p-decomposable non-graded CA rule is shown. We see that, although the number of
nodes and the alphabets of the layers are the same as in Fig. 1 a), not all layers are layer CAs, and
(2,2)
some layers depend on others: We see that layer CA A2 influences two of the layers, while layers
(2,0) (2,1)
A3 and A5 are interdependent. In Fig. 1 d) a partially decomposed diagram that contains the
one of Fig. 1 b) as a specific case, is plotted.
We summarize some observations:
• The top rule in a diagram is the CA model. The inputs of this model are xj−r t , . . . , xjt , . . . , xj+l
t
and the output xjt+1 at every j and t. For the top layer, we always write xjt and xjt+1 instead of
(0,0),j (0,0),j
yt , yt+1 , respectively.
• A graded CA rule contains only layer CAs and undirected edges.
• The product of the ph ’s of rank-(g + 1) layers coming from a single node at rank g equals p of
that node.
• From a layer CA only outgoing arrows or undirected edges are possible.
• From a layer that is not a CA, undirected edges and incoming and outgoing arrows are possible.
If it has only undirected edges then it must exhibit an incoming arrow in at least one partial
decomposition of the tree.

6
FIG. 1: a) A fully decomposed diagram of a graded CA rule with p = 630. b) A possible p-decomposable non-graded CA rule
derived from a). In c) and d) two partially decomposed diagrams compatible with a) and b), respectively, are shown.

• If a node corresponds to a layer on a prime number of symbols it cannot be further decomposed.

Since graded rules constitute the backbone of any tree, it is useful first to see how a diagram
representing an arbitrary graded rule at a node (g, h) can be translated to a mathematical model.
Such arbitrary building block from which to construct trees as the one in Figs. 1 is shown in Fig. 2.
Let us assume that a node with coordinates (g, h) in a diagram splits as shown in Fig. 2. From Eqs.
(11) to (16) we have
N −1 s−1
(g,h),j (g+1,n+s),j
X Y
yt+1 = yt+1 pm (17)
s=0 m=0

where
l
!
(g+1,n+s),j (g+1,n+s),j+k
X
yt+1 = dp s psk+r yt , A(g+1,n+s)
ps s = 0, 1, . . . , N − 1 (18)
k=−r

7
 (g,h)
FIG. 2: A diagram for an arbitrary graded rule Ap at node (g, h) of a tree.

and !
(g,h),j
(g+1,n+s),j yt
yt ≡ dp s 0, Qs−1 s = 0, 1, . . . , N − 1 (19)
m=0 pm

These expressions are thus the mathematical transposition of the tree in Fig. 2 and are to be
(g,h),j (g,h)
interpreted as follows. The value yt+1 on the node governed by Ap is known from the previous
(g,h),j (g+1,n+s),j
value yt , by replacing the latter in Eq. (19) to obtain all the yt ’s on the rank g + 1 at
(g+1,n+s),j
time t. Then, all these quantities are replaced in Eq. (18) to obtain all N quantities yt+1 at
(g,h),j
the next time step and, from them, the value of yt+1 is finally obtained from Eq. (17). Because of
the independence of the layer CAs within the CA model, we say that the latter is the direct sum of
its layer CAs and write
−1)
A(g,h)
p = A(g+1,n)
p0 ⊕ Ap(g+1,n+1)
1
⊕ . . . ⊕ A(g+1,n+N
p0 (20)

The couplings between layers are represented through arrows as described above. Let an arrow
start from a node with coordinates (g 0 , h0 ) and end in a node with coordinates (g, h). Thus, layer h
is influenced by layer h0 . The mathematical general way to express this relationship is obtained, by
using Eq. (6), as
pl+r+1
h0
−1 l
! l
!
(g,h),j (g,h),j+k (g,h) (g 0 ,h0 ),j+k
X X X
yt+1 = dph phk+r yt , An0 ,ph dph0 n0 − pk+r
h0 yt ,1 (21)
n0 =0 k=−r k=−r

(g 0 ,h0 ) pl+r+1 (g 0 ,h0 )

where the An0 ,ph are integers in the interval [0, phh − 1]. If all An0 ,ph are independent of n0 , then,
(g,h) (g,h)
the layer (g 0 , h0 ) becomes a layer CA with Wolfram code An0 ,ph = Aph . In this case, from this latter
Ppl+r+1 −1

(g 0 ,h0 ),j+k

dph0 n0 − lk=−r pk+r
0 P
expression, since nh0 =0 h0 yt , 1 = 1, we have

l
!
(g 0 ,h0 ),j (g 0 ,h0 ),j+k
X
yt+1 = d p h0 pk+r
h0 yt , A(g,h)
ph (22)
k=−r

Eq. (21) thus describes any hollow node in the tree and Eq. (22) a solid node (as we have seen above).
If there are more incoming arrows at node (g, h) coming from nodes e.g. at (g 0 , h0 ) and (g 00 , h00 ), Eq.

8
(21) generalizes as
pl+r+1
h00
−1 pl+r+1
h0
−1 l
!
(g,h),j (g,h),j+k (g,h)
X X X
yt+1 = dph pk+r
h yt , An0 n00 ,ph × (23)
n00 =0 n0 =0 k=−r
l l
! !
(g 0 ,h0 ),j+k (g 00 ,h00 ),j+k
X X
0 00
×dph0 n − pk+r
h0 yt ,1 dph00 n − pk+r
h00 yt ,1
k=−r k=−r

(g 0 ,h0 ) pl+r+1
where, again, the An0 n00 ,ph are integers in the interval [0, ph − 1]. The above expression generalizes
h

to an arbitrary number of incoming arrows.

(g,h)
A remark is in order on how to calculate the integer numbers An0 ,ph from an arbitrary function
fn0 : S l+r+1 → S, 0 ≤ fn0 ≤ p − 1 (fn0 ∈ Z)
 
(g,h),j−r (g,h),j−r+1 (g,h),j−r (g,h),j+l−1 (g,h),j+l
fn0 yt , yt , . . . , yt , . . . , yt , yt (24)
Pl (g,h),j+k (g,h),j+k
Let n = k=−r pk+r
h yt . We have yt = dph (k + r, n). Therefore,

pl+r+1
h −1
(g,h)
X
An0 ,ph = pnh fn0 (dph (0, n), . . . , dph (r, n), . . . , dph (l + r, n)) (25)
n=0

since
l
!
 
(g,h),j+k (g,h) (g,h)
X
dp h phk+r yt , An0 ,ph = dph n, An0 ,ph = fn0 (26)
k=−r

Having explained how to pass from diagrams to mathematical expressions and vice versa, this
abstract theory is now completed. The CA modeling through the diagrammatic approach proceeds
through the following steps

• 1. Determine how many layers are needed and construct the diagram of a graded CA rule with
all those layers. Transpose the diagram to a mathematical model employing Eqs. (17) to (19)
as building blocks for each branch in the diagram.

• 2. Draw the arrows needed to couple the corresponding layers. For each layer (g 0 , h0 ) with an
incoming arrow, replace the corresponding mathematical expression obtained from the graded
(g 0 ,h0 ),j
rule for yt+1 by Eq. (22).

Let us illustrate this approach with reference to the trees in Fig. 1. The tree in Fig. 1 a) is a

9
graded CA rule and we can write out the mathematical model as follows
(1,0),j (1,1),j
xjt+1 = yt+1 + 30yt+1 (27)
(1,0),j (2,0),j (2,1),j (2,2),j
yt+1 = yt+1 + 3yt+1 + 15yt+1 (28)
(1,1),j (2,3),j (2,4),j
yt+1 = yt+1 + 7yt+1 (29)
l
!
(2,0),j (2,0),j+k (2,0)
X
yt+1 = d3 3k+r yt , A3 (30)
k=−r
l
!
(2,1),j (2,1),j+k (2,1)
X
yt+1 = d5 5k+r yt , A5 (31)
k=−r
l
!
(2,2),j (2,2),j+k (2,2)
X
yt+1 = d2 2k+r yt , A2 (32)
k=−r
l
!
(2,3),j (2,3),j+k (2,3)
X
yt+1 = d7 7k+r yt , A7 (33)
k=−r
l
!
(2,4),j (2,4),j+k (2,4)
X
yt+1 = d3 3k+r yt , A3 (34)
k=−r

! !
(1,0),j (1,0),j
(2,0),j

(1,0),j

(2,1),j yt (2,2),j yt
yt = d3 0, yt yt = d5 0, yt = d2 0,
3 15
(35)
!
(1,1),j
(2,3),j

(1,1),j

(2,4),j yt
yt = d7 0, yt yt = d3 0, (36)
7
!
(1,0),j
 
(1,1),j xjt
yt = d30 0, xjt yt = d21 0, (37)
30

Thus, computations with such a model are carried from bottom to top, starting from Eqs. (37)
(2,0) l+r+1
for an input value of xjt ∈ [0, 629], ∀j ∈ [0, Ns − 1] and the five integers A3 ∈ [0, 33 − 1],
(2,1) 5l+r+1 (2,2) 2l+r+1 (2,3) 7l+r+1 (2,4) 3l+r+1
A5 ∈ [0, 5 − 1], A2 ∈ [0, 2 − 1], A7 ∈ [0, 7 − 1] and A3 ∈ [0, 3 − 1].
Thus all all quantities are calculated from bottom to top so that, at the end xjt+1 ∈ [0, 629] is obtained
from Eq. (27). This system of equations is the mathematical equivalent to the diagram in Fig. 1 a).
It is now straightforward to construct the family of CA rules described by the tree in Fig. 1 b).
(g,h),j
We have just only to provide expressions for yt+1 for all those layers that have an incoming arrow.
(2,0),j
In the tree in Fig. 1 b) there are three such layers, whose values at time t + 1 are given by yt+1 ,

10
 (2,1),j (2,3),j
yt+1 and yt+1 . Thus, by using Eqs. (21) and (23)
5l+r+1
X−1 l l
! !
(2,0),j (2,0),j+k (2,0) (2,1),j+k
X X
0
yt+1 = d3 3k+r yt , An0 ,ph d5 n − 5k+r yt ,1
n0 =0 k=−r k=−r
2l+r+1 l+r+1
X−1 3 X−1 l
!
(2,1),j (2,1),j+k (2,1)
X
yt+1 = d5 5k+r yt , An0 n00 ,ph × (38)
n00 =0 n0 =0 k=−r
l l
! !
(2,0),j+k (2,2),j+k
X X
0 00
×d3 n − 3k+r yt ,1 d2 n − 2k+r yt ,1
k=−r k=−r
2l+r+1
X−1 l l
! !
(2,3),j (2,3),j+k (2,3) (2,2),j+k
X X
yt+1 = d7 7k+r yt , An0 ,ph d2 n0 − 2k+r yt ,1
n0 =0 k=−r k=−r

and these expressions replace the corresponding ones within the graded rule above. In this way we
account for all possible couplings among layers described by the tree in Fig. 1 b). Note that all other
equations remain unchanged, as in the graded rule.
Models of partially decomposed diagrams are considerably simplified compared to the ones corre-
sponding to fully developed diagrams. Therefore, the specific choice to which extent a diagram is to
be developed depends on the specific problem at hand. The diagram in Fig. 1 c) is mathematically
described by the following system of equations
(1,0),j (1,1),j
xjt+1 = yt+1 + 30yt+1 (39)
(1,1),j (2,0),j (2,1),j
yt+1 = yt+1 + 7yt+1 (40)
l
!
(2,0),j (2,0),j+k (2,0)
X
yt+1 = d7 7k+r yt , A7 (41)
k=−r
l
!
(2,1),j (2,1),j+k (2,1)
X
yt+1 = d3 3k+r yt , A3 (42)
k=−r
l
!
(1,0),j (1,0),j+k (1,0)
X
yt+1 = d30 30k+r yt , A30 (43)
k=−r
!
(1,1),j
(2,0),j

(1,1),j

(2,1),j yt
yt = d7 0, yt yt = d3 0, (44)
7
!
(1,0),j
 
(1,1),j xjt
yt = d30 0, xjt yt = d21 0, (45)
30

11
Finally, the diagram in Fig. 1 d) corresponds to the following mathematical model
(1,0),j (1,1),j
xjt+1 = yt+1 + 30yt+1 (46)
(1,1),j (2,0),j (2,1),j
yt+1 = yt+1 + 7yt+1 (47)
30l+r+1 −1 l l
! !
(2,0),j (2,0),j+k (2,0) (1,0),j+k
X X X
yt+1 = d7 7k+r yt , An0 ,ph d30 n0 − 30k+r yt ,1
n0 =0 k=−r k=−r
(48)
l
!
(2,1),j (2,1),j+k (2,1)
X
yt+1 = d3 3k+r yt , A3 (49)
k=−r
l
!
(1,0),j (1,0),j+k (1,0)
X
yt+1 = d30 30k+r yt , A30 (50)
k=−r

!
(1,1),j
(2,0),j

(1,1),j

(2,1),j yt
yt = d7 0, yt yt = d3 0, (51)
7
!
(1,0),j
 
(1,1),j xjt
yt = d30 0, xjt yt = d21 0, (52)
30

In this way we obtain simple constructions of CA models on 630 symbols. We devote the rest of
the article to examples to illustrate the above approach. These are presented in the next section.

B. Classification of CA rules for p having a small number of proper divisors

All CA rules can be classified according to the above approach and we carry out now this in the
simplest cases for p having a small number of divisors. It is clear, from all above, that the ‘tree’ of
a non-decomposable rule, i.e. a rule where the alphabet size p is a prime number, has the trivial
representation

The mathematical expression for these CA is just Eq. (1) because p has not proper divisors and,
hence, no layers other than the trivial one (i.e. the CA). All Boolean CA (p = 2) fall in this category.
We now consider decomposable CA rules l Rpr on p = p0 p1 symbols, with p0 and p1 both prime (not
l+r+1 l+r+1
necessarily distinct). There are pp = (p0 p1 )(p0 p1 ) one-dimensional deterministic CA rules of
this kind and they can all be classified according to the diagrams shown in Fig. 3. The CA rules with
tree as in a) are graded CA rules. All others are p-decomposable non graded CA rules.

12
FIG. 3: Classification of all l Rpr CA rules on p = p0 p1 symbols. a) Graded rules, b) to d) p-decomposable non-graded CA rules.

The graded rules, diagram in Fig. 3 a), have all the following mathematical model
(1,0),j (1,1),j
xjt+1 = yt+1 + p0 yt+1 (53)
l
!
(1,0),j (1,0),j+k
X
yt+1 = dp0 p0k+r yt , A(1,0)
p0 (54)
k=−r
l
!
(1,1),j (1,1),j+k
X
yt+1 = dp1 pk+r
1 yt , A(1,1)
p1 (55)
k=−r
!
(1,0),j
 
(1,1),j xj
yt = dp0 0, xjt yt = dp 1 0, t (56)
p0

These equations can be more concisely expressed in a single equation as

l l
! ! !
j
X
k+r

j+k

(1,0)
X
k+r xj+k
t (1,1)
xt+1 = dp0 p0 dp0 0, xt , Ap 0 + p0 dp1 p1 dp1 0, , Ap1 (57)
p0
k=−r k=−r

which is equivalent to the above system.

The CAs described by Fig. 3 b) have the following general form
(1,0),j (1,1),j
xjt+1 = yt+1 + p0 yt+1 (58)
l
!
(1,0),j (1,0),j+k
X
yt+1 = dp 0 pk+r
0 yt , A(1,0)
p0 (59)
k=−r
pl+r+1
0 −1 l
! l
!
(1,1),j (1,1),j+k (1,1) (1,0),j+k
X X X
yt+1 = dp1 pk+r
1 yt , An0 ,p1 dp 0 n0 − pk+r
0 yt ,1
n0 =0 k=−r k=−r
(60)
!
(1,0),j
 
(1,1),j xjt
yt = dp0 0, xjt yt = dp 1 0, (61)
p0

13
and it is equally straightforward to write the general form corresponding to the diagram in Fig. 3 c)
(1,0),j (1,1),j
xjt+1 = yt+1 + p0 yt+1 (62)
pl+r+1
1 −1 l
! l
!
(1,0),j (1,0),j+k (1,0) (1,1),j+k
X X X
yt+1 = dp0 pk+r
0 yt , An0 ,p0 dp 1 n0 − pk+r
1 yt ,1
n0 =0 k=−r k=−r
(63)
l
!
(1,1),j (1,1),j+k
X
yt+1 = dp 1 pk+r
1 yt , A(1,1)
p1 (64)
k=−r
!
(1,0),j
 
(1,1),j xj
yt = dp 0 0, xjt yt = dp 1 0, t (65)
p0

The Moufang loop example with p = 12 in [12] belongs to this class (if one constructs a partially
developed tree with p0 = 6 and p1 = 2). Finally Fig. 3 d) is mathematically transposed as
(1,0),j (1,1),j
xjt+1 = yt+1 + p0 yt+1 (66)
pl+r+1
1 −1 l
! l
!
(1,0),j (1,0),j+k (1,0) (1,1),j+k
X X X
0
yt+1 = dp 0 p0k+r yt , An0 ,p0 dp1 n − pk+r
1 yt ,1
n0 =0 k=−r k=−r
(67)
pl+r+1
0 −1 l
! l
!
(1,1),j (1,1),j+k (1,1) (1,0),j+k
X X X
yt+1 = dp 1 p1k+r yt , An00 ,p1 dp0 n00 − pk+r
0 yt ,1
n00 =0 k=−r k=−r
(68)
!
(1,0),j
 
(1,1),j xjt
yt = dp0 0, xjt yt = dp 1 0, (69)
p0

III. EXAMPLES

A. CA rules derived from binary operators

We consider CA rules for which l + r + 1 = 2 and take, for example l = 0, r = 1 below. Thus,
the two inputs xtj−1 ∈ S and xjt ∈ S at time t and neighboring locations j − 1 and j yield an output
xjt+1 ∈ S at time t + 1. Eq. (1) takes in this case the simple form

xjt+1 = dp (xj−1
t + pxjt , R) ≡ xjt ∗ xj−1
t (70)

where we have introduced the notation ∗ to emphasize that dp (xj−1

t +pxjt , R) acts as a binary operator
S 2 → S. Thus, the integer R ∈ [0, p2 − 1] lists all possible binary operators (magmas) on the set S
[15]. If we tabulate the output xjt+1 as a function of the inputs xjt and xj−1t in the rows and in the
columns respectively, we obtain the Cayley table of the binary operator as [15]

14
 ∗ 0 1 ... p−1
0 dp (0, R) dp (1, R) ... dp (p − 1, R)
1 dp (p, R) dp (p + 1, R) ... dp (2p − 1, R)
... ... ... ... ...
p − 2 dp (p(p − 2), R) dp (p(p − 2) + 1, R) ... dp (p(p − 1) − 1, R)
p − 1 dp (p(p − 1), R) dp (p(p − 1) + 1, R) ... dp (p2 − 1, R)
Among these rules, most interesting ones are those with a group structure. Let us consider, for
example, rules with Wolfram code
2 −1
pX
R= pn dp (0, dp (0, n) + dp (1, n)) (71)
n=0
For these rules, from Eq. (1), we have [15]
 2 −1

  pX
xjt+1 = dp xtj−1 + pxjt , R = dp xj−1
t + pxjt , pn dp (0, dp (0, n) + dp (1, n))
n=0
     
= dp 0, dp 0, dp 0, xj−1
t + pxjt + dp 1, xj−1
t + pxjt
      
= dp 0, dp 0, xtj−1 + dp 0, xjt = dp 0, xj−1
t + x j
t (72)

We now observe that the integers xtj−1 and xjt in S have a cyclic group structure Cp under the action
of the operator  
dp 0, xjt + xj−1
t (73)

Indeed, if we tabulate xjt+1 as a function of xjt in the rows of the table and xj−1
t in the columns, we
have [14]
 
dp 0, xtj−1 + xjt 0 1 2 ... p − 2 p − 1
0 0 1 2 ... p−2 p−1
1 1 2 3 ... p−1 0
... ... ... ... ... ... ...
p−2 p−2 p−1 0 ... p−4 p−3
p−1 p−1 0 1 ... p−3 p−2
which corresponds to the Cayley table of the finite cyclic group Cp [18]. The invariance under addition
modulo p of CA with this structure [16] comes from the very fact that the digit function is a group
homomorphism sending the integers Z to the set of residue classes Z/pZ [14]. In a previous work [19],
CA with this structure were suggested as templates for the design of the most complex (Class 4) CA.
If we now take p0 = p1 = 2 we can, e.g, construct the direct sum C2 ⊕ C2 from Eqs. (57) and (71)
(1,0) (1,1) P 2 −1 n
by taking A2 = A2 = 2n=0 2 d2 (0, d2 (0, n) + d2 (1, n)) = 6, r = 0 and l = 1, i.e a CA rule of
range ρ = 2. Eq. (57) reduces in this case to
! !!
j j−1
  x x
xjt+1 = d2 0, xjt + xj−1
t + 2d2 0, d2 0, t + d2 0, t (74)
2 2
And we now obtain the table

15
 ∗ 0 1 2 3
0 0 1 2 3
1 1 0 3 2
2 2 3 0 1
3 3 2 1 0

which corresponds to the Cayley table of the direct sum C2 ⊕ C2 which is isomorphic to Klein’s four
group V4 . The CA so obtained has a diagram given by Fig. 3 a).
In Fig. 4 A the spatiotemporal evolution of xjt obtained from Eq. (74), is plotted for an initial
condition xj0 = d4 (3, Ns − j) on a ring of 200 sites and for 200 time steps. The latter initial condition
is chosen to investigate the CA evolution starting from homogeneous patches that cover the whole
range of possible values for xj0 ∈ [0, 3]. In this way, closed substructures involving subsets of the rule
are revealed: The figure displays Sierpinski-like triangles of different colors that are clearly separated
against the zero value of the background. These structures are the spatiotemporal realization of the
subgroups of the direct sum C2 ⊕ C2 , which correspond to the subsets {0, 1}, {0, 2} and {0, 3} (Fig.
4 A).
It is useful to compare the above CA rule with the one with p = 4 with the structure of the cyclic
group C4 . From Eq. (1) it is given by
 
j j j−1
xt+1 = d4 0, xt + xt (75)

If we again tabulate the possible outputs xjt+1 of the CA by listing xj+1

t in the rows and xjt in the
columns, we now obtain
 
d4 xjt + 4xj+1
t , R 0 1 2 3
0 0 1 2 3
1 1 2 3 0
2 2 3 0 1
3 3 0 1 2

After a little algebra it is seen that Eq. (75) is decomposed, from Eqs. (13) and (14), as
 
(1,0)j (1,0),j−1
xjt+1 = d2 0, yt + yt +
2    
(1,1),j−1 (1,1),j (1,0),j−1 (1,0),j
X
+2 d2 yt + 2yt , 6 d2 n − (yt + 2yt ), 1
n=0
   
(1,1),j−1 (1,1),j (1,0),j−1 (1,0),j
+2d2 yt + 2yt , 9 d2 3 − (yt + 2yt ), 1

(1,1) (1,1)
from which we observe that the rule is of the type of Fig. 3 b), with An,p1 = An,2 = 6 for n = 0, 1, 2
(1,1)
and An,2 = 9 for n = 3.
The spatiotemporal evolution of the CA rule given by Eq. (75), in Fig. 4 B, for the same initial
condition, system size and number of time steps as in Fig. 4 A. We now observe that only certain
triangles involving the elements {0, 2} (the only subgroup of C4 , which is isomorphic to C2 ) can
spatially separated from the others, where all symbols in S are mixed. This is in contrast with (Fig.

16
FIG. 4: Spatiotemporal evolution of xjt for the CA rules given by Eqs. (74) (A) and (75) (B) starting from an initial condition
xj0 = d4 (3, Ns − j) on a ring of 200 sites and for 200 time steps.

4 A) in which, as we have seen above, different triangular structures formed by the subsets {0, 1},
{0, 2} and {0, 3} are separated. From this we see once more that the spatiotemporal evolution of the
CA reveals closed substructures within the algebraic operator that specifies the rule.

B. Decomposable CA models based on the majority rule

An important totalistic rule is the majority rule [4, 20–25], which is given by
 
ξ
1 1
xjt+1 = H − + xj+k
X
t
 (76)
2 2ξ + 1
k=−ξ

where H(x) is the Heaviside function (H(x) = 0 for x < 0, H(0) = 12 and H(x) = 1 for x > 0). This
is a symmetric rule with l = r = ξ with p = 2 which returns, at each location j, the value ’0’ or ’1’
that occurs more frequently within the neighborhood of radius ξ and 2ξ + 1 sites. It can be seen that
the majority rule is equivalently described by Eq. (9) as
 
2ξ+1 ξ
xjt+1 = xj+k
X X
d2 n − t , 1 (77)
n=b 2ξ+3
2 c
k=−ξ

or, alternatively, by Eq. (1) as

 
ξ
xjt+1 = d2  xj+k
X
t , 22(ξ+1) − 2b(2ξ+3)/2c  (78)
k=−ξ

17
Starting from an arbitrary initial condition consisting of ‘0’s and ‘1’s, this rule attains, after a short
transient, a fixed point where domains are formed with no less than ξ + 1 consecutive zeros or ξ + 1
ones. For a system size of Ns sites there are
j k
Ns
2(ξ+1)  
X 2Ns Ns − 2`ξ
F(Ns , ξ) = 2 + (79)
Ns − 2`ξ 2`
`=1

such fixed points [24]. Since there are 2Ns

possible initial conditions for a ring of Ns sites, it is clear
N
that, since, F(Ns , ξ) < 2 , the phase space irreversibly contracts so that the (dimensionless) entropy
s

change involved in attaining any fixed point is

 
F(Ns , ξ)
∆S = ln <0 (80)
2Ns
That the entropy is ’spontaneously’ lowered in the CA evolution is not surprising if one thinks that
the majority rule, as applied locally at each site j, acts as a Maxwell demon [26], by locally setting
the value xjt+1 to the value that most often appears within the neighborhood of radius ξ centered at
j at time t.
We can now generalize the majority rule to domains with site values not restricted to be zero or
one. For, if we consider a graded rule with diagram given by Fig. 2 in which all layers are majority
rules we have
N −1
(1,h),j
xjt+1 =
X
2h yt+1 (81)
h=0

where  
ξh
(1,h),j 1 1 X (1,h),j+k 
yt+1 = H − + yt h = 0, 1, . . . , N − 1 (82)
2 2ξh + 1
k=−ξh

and !
(1,h),j xj
yt ≡ d2 0, ht h = 0, 1, . . . , N − 1 (83)
2
Since the layers are independent, the CA attains again a fixed point involving values in the set
{0, 1, 2, . . . , 2N − 1}. There is now a total number of fixed points given by
 j k 
Ns
N −1 N −1 2(ξh +1)  
Y Y  X 2Ns Ns − 2`ξh 
F(Ns , ξh ) = 2 +  (84)
 Ns − 2`ξh 2` 
h=0 h=0 `=1

and, therefore, since the multiplicity of fixed points factorizes, the entropy change in attaining any of
the fixed points starting from an arbitrary initial condition is the additive sum of the contributions
from all independent layers
QN −1 ! N −1  
h=0 F(Ns , ξh ) F(Ns , ξh )
X
∆S = ln = ln <0 (85)
2Ns 2Ns
h=0

18
In the case N = 1 Eqs. (86) to (89) reduce to Eq. (78). Let us consider the case N = 2 so that the
model has diagram as in Fig. 1 a). Eqs. (81) to (83) reduce in this case to
(1,0),j (1,1),j
xjt+1 = yt+1 + 2yt+1 (86)
 
ξ0
(1,0),j 1 1 X (1,0),j+k 
yt+1 = H − + yt (87)
2 2ξ0 + 1
k=−ξ0
 
ξ1
(1,1),j 1 1 X (1,1),j+k 
yt+1 = H − + yt (88)
2 2ξ1 + 1
k=−ξ1
!
(1,0),j
 
(1,1),j xj
yt = d2 0, xjt yt = d2 0, t (89)
2

In Fig. 5A the spatiotemporal evolution of xjt , obtained from Eqs. (86) to (89) is shown for ξ0 = 3,
ξ1 = 1 for a random initial condition xj0 , and 20 iteration steps. We see that the system reaches
a spatial fixed point where the different possible values of S, ({0, 1, 2, 3} in this case) arrange in
spatial domains. Note that domains with value xjt = 2 are contained within domains of value xjt = 0
while those with value xjt = 3 are contained within domains with value xjt = 1. The reason is that
(1,0)
d2 (0, 2) = 0 and d2 (0, 3) = 1 for a nonzero value yt = 1 of the lifted layer: If the lifted layer is
j j (1,0) (1,0)
nonzero, for xt = 2 (resp. xt = 3), the ground layer has value yt = 0 (resp. yt = 1). Note that
although the layers are independent, if the lifted layer has value zero at some j, the ground layer is,
anyway, spatially correlated with the ground layer of the neighboring sites where the lifted layer is
nonzero. This is reflected in the xjt values.
The graded CA rule can now be exploited to construct a CA model with coupled layers so as to
filter out some specific values of S present in the initial condition. Let us consider again N = 2 as in
the model above, but now let us use the majority rule in the ground layer to influence the lifted layer,
following the diagram of Fig. 1 b). Specifically, we consider now the following model
(1,0),j (1,1),j
xjt+1 = yt+1 + 2yt+1 (90)
 
ξ0
(1,0),j 1 1 X (1,0),j+k 
yt+1 = H − + yt (91)
2 2ξ0 + 1
k=−ξ0
   
ξ1 ξ0
(1,1),j 1 1 X (1,1),j+k  1 1 X (1,0),j+k 
yt+1 = H − + yt H − + yt
2 2ξ1 + 1 2 2ξ0 + 1
k=−ξ1 k=−ξ0
(92)
!
(1,0),j
 
(1,1),j xjt
yt = d2 0, xjt yt = d2 0, (93)
2

The lifted layer is now coupled to the ground layer and the model can be understood
in terms of the simpler model  of Eqs. (86) to (89). Indeed, in those domains where
1 1 Pξ0 (1,0),j+k
H − 2 + 2ξ0 +1 k=−ξ0 yt = 1 the model reduces to Eqs. (86) to (89) and in those domains
 
(1,0),j+k
where H − 21 + 2ξ01+1 ξk=−ξ
P 0
y
0 t
= 0 the model collapses to the simple majority rule, Eq. (76).

19
FIG. 5: A. Spatiotemporal evolution of xjt obtained from Eqs. (86) to (89). B. Spatiotemporal evolution of xjt obtained from
Eqs. (90) to (93) for ξ0 = 3, ξ1 = 1. In both cases the same parameter values ξ0 = 3, ξ1 = 1 and same random initial condition
are used, and 20 iteration steps are shown for a ring with Ns = 200 sites.

As a consequence of this, it is readily observed that the value xjt = 2 cannot occur in the trajectory
of the CA and can only be present in the initial condition.
In Fig. 5B the spatiotemporal evolution of xjt , obtained from Eqs. (90) to (93) is shown for the
same parameter values and initial condition as in In Fig. 5A. We see that the system reaches a spatial
fixed point that is similar to the one of Fig. 5A with the exception that now xjt ∈ {0, 1, 3} since
the value xjt = 2 is not possible in the trajectory. As a consequence of having filtered out the value
xjt = 2 through the coupling among layers, the entropy change from the initial condition till any of
the fixed points is reached is lower than in the graded rule case (i.e. Eq. (85 with N = 2): There is
a smaller number of fixed points because the value xjt = 2 is not possible in these, as a consequence
of the coupling of the lifted layer to the ground layer. Thus, if we evaluate the entropy change with
Eq. (85) without taking into account the coupling, we would overestimate the entropy change. That
entropy is lowered through correlations shows that entropy is no longer equal to the mere addition
of the independent layers. Therefore, this constitutes a dynamical system in which the methods of
nonextensive statistics [27, 28] and superstatistics [29] may be used. In [30, 31] physical examples
(both in and out of equilibrium) where interactions lead to a non-additive entropy (connected to a
reduction in the available phase space) are given.
The dynamical behavior of any of the domains can be specified separately. For example, in the
following model
(1,0),j (1,1),j
xjt+1 = yt+1 + 2yt+1 (94)
 
ξ
(1,0),j 1 1 X (1,0),j+k 
yt+1 = H − + yt (95)
2 2ξ + 1
k=−ξ
 
1 ξ
!
(1,1),j
X (1,1),j+k 1 1 X (1,0),j+k 
yt+1 = d2 2k+r yt , R H − + yt
2 2ξ + 1
k=−1 k=−ξ
(96)
!
(1,0),j
 
(1,1),j xjt
yt = d2 0, xjt yt = d2 0, (97)
2

20
FIG. 6: Spatiotemporal evolution of xjt obtained from Eqs. (94) to (97) and for R = 30 (A), R = 54 (B), R = 110 (C) and
R = 150 (D). In all cases, Ns = 200, ξ = 30 and 400 time steps are shown starting from a random initial condition that is the same
in all cases. Time flows from top to bottom and j increases from right to left in each panel.

(1,0),j
the domains for which yt = 0 (once a fixed point is attained) remain at the quiescent state xjt = 0
(1,0),j (1,1),j
forever. However, in those domains for which  yt = 1 (t → ∞), xjt = 1 + 2yt+1 from Eq. (94)
(1,1),j P1 k+r y (1,1),j+k , R from Eq. (96). The latter corresponds to a Wolfram
and yt+1 = d2 k=−1 2 t
elementary CA with code 0 ≤ R ≤ 255 and, therefore, such Wolfram Boolean CA runs within the
domains where the ground layer has value ’1’ taking values on the set xjt ∈ {1, 3}. In Fig.6 the
spatiotemporal evolution of xjt obtained from Eqs. (94) to (97) and for R = 30 (A), R = 54 (B),
R = 110 (C) and R = 150 (D), is shown. In all cases, Ns = 200, ξ = 30 and 400 time steps are shown
starting from a random initial condition that is the same in all cases. Time flows from top to bottom
and j increases from right to left in each panel. We observe that, because of interaction with the
borders of the domain, the chaos present in Wolfram’s CA rule 30 vanishes after ca. 380 time steps
(panel A). Rules 54 and 110 (panels B and C) display long transients with unpredictable behavior,
and the additive rule 150 yields a chaotic pattern confined to the domain. In all cases, there is a single
domain where the interesting dynamics takes place coexisting with a domain in the quiescent state.
All cases correspond to the class of models given by Fig. 1 b).

21
 FIG. 7: Diagram corresponding to the CA model for chimera states in [13], Eqs. (98) to (102).

The following is a CA model for chimera states that has been introduced in a recent article [13]
(1,0),j (1,1),j (1,2),j
xjt+1 = yt+1 + 2yt+1 + 4yt+1 (98)
 
ξ
(1,0),j 1 1 X (1,0),j+k 
yt+1 = H − + yt (99)
2 2ξ + 1
k=−ξ
  
ξ
(1,1),j 1 1 X (1,1),j+k 
yt+1 = 1 − H − + yt ×
2 2ξ + 1
k=−ξ
  
ξ
1 1 X (1,0),j+k 
× 1 − H − + yt (100)
2 2ξ + 1
k=−ξ

1
!  ξ

(1,2),j
X (1,1),j+k 1 1 X (1,0),j+k 
yt+1 = d2 2k+r yt , 105 H − + yt
2 2ξ + 1
k=−1 k=−ξ
(101)
! !
(1,0),j
 
(1,1),j xjt (1,2),j xjt
yt = d2 0, xjt yt = d2 0, yt = d2 0,
2 4
(102)

It has diagram given by Fig. 7, where the ground layer, which is the majority rule, is a layer CA that
influences the remaining lifted layers. The model works on eight symbols p = 8 = 23 and decomposes
into layers working on two symbols each (Boolean layers).
As we have mentioned above, all above models can be generalized to any dimension and/or neigh-
borhood by introducing more spatial coordinates. Let i be an index that besides j runs on a square
surface under periodic boundary conditions. For example, for two dimensions and square Moore neigh-
borhoods with 2ξh + 1 side (for the layer h = 0, 1 . . . , N − 1) we have that Eqs. (81) to (83) generalize
as
N −1
(1,h),i,j
xi,j
X
t+1 = 2h yt+1 (103)
h=0

22
where
 
ξh ξh
(1,h),i,j 1 1 X X (1,h),i+k,j+m 
yt+1 =H − +
 yt h = 0, 1, . . . , N − 1 (104)
2 (2ξh + 1)2
m=−ξh k=−ξh

and !
(1,h),i,j xi,j
yt ≡ d2 0, th s = 0, 1, . . . , N − 1 (105)
2
If we take N = 2 in these expressions, we obtain the corresponding graded rule in two dimensions.
We obtain
(1,0),i,j (1,1),i,j
xi,j
t+1 = yt+1 + 2yt+1 (106)
 
ξ0 ξ0
(1,0),i,j 1 1 X X (1,0),i+k,j+m 
yt+1 = H − + yt (107)
2 (2ξ0 + 1)2
m=−ξ0 k=−ξ0
 
ξ1 ξ1
(1,1),i,j 1 1 X X (1,1),i+k,j+m 
yt+1 = H − + yt (108)
2 (2ξ1 + 1)2
m=−ξ1 k=−ξ1
!
(1,0),i,j

i,j

(1,1),i,j xi,j
t
yt = d2 0, xt yt = d2 0, (109)
2

These equations generalize Eqs. (86) to (89) to two dimensions. In Fig. 8A the spatiotemporal
evolution of xi,j
t obtained from Eqs. (106) to (109) is shown for a surface with Ns × Ns = 50 × 50 sites,
ξ0 = 3, ξ1 = 1 and a random initial condition. Four iteration steps are shown. After a short transient
a 2D spatial fixed point is reached which is analogous in 2D to the one found for 1D in Fig. 5A.
Eqs. (90) to (93) generalize to 2D as
(1,0),i,j (1,1),i,j
xi,j
t+1 = yt+1 + 2yt+1 (110)
 
ξ0 ξ0
(1,0),i,j 1 1 X X (1,0),i+k,j+m 
yt+1 = H − +

2
yt (111)
2 (2ξ0 + 1)
m=−ξ0 k=−ξ0
 
ξ1 ξ1
(1,1),i,j 1 1 X X (1,1),i+k,j+m 
yt+1 = H − + yt ×
2 (2ξ1 + 1)2
m=−ξ1 k=−ξ1
 
ξ0 ξ0
1 1 X X (1,0),i+k,j+m 
×H − + yt (112)
2 (2ξ0 + 1)2
m=−ξ0 k=−ξ0
!
(1,0),i,j

i,j

(1,1),i,j xi,j
t
yt = d2 0, xt yt = d2 0, (113)
2

In Fig. 8B we show the spatiotemporal evolution of xi,j t obtained from Eqs. (110) to (113) for a surface
with Ns × Ns = 50 × 50 sites, ξ0 = 3, ξ1 = 1 and a random initial condition that is the same as in Fig.
8A. After a short transient a 2D spatial fixed point is reached which is analogous in 2D to the one found
for 1D in Fig. 5B. Note the absence of the value xi,j t = 2 in the trajectory of the CA. Interestingly,

23
FIG. 8: A. Spatiotemporal evolution of xjt obtained from Eqs. (106) to (109). B. Spatiotemporal evolution of xjt obtained from
Eqs. (110) to (113). In both cases the same parameter values ξ0 = 3, ξ1 = 1 and same random initial condition are used, and 4
iteration steps (indicated over the panels) are shown for a surface with Ns × Ns = 50 × 50 sites.

this slightly affects the distribution of the value xi,j

t = 3 when compared to Fig. 8A. This is due to the
j j
fact that values xt = 2 and xt = 3 in the latter figure are correlated in the lifted layer with a value
(1,1),i,j (1,1),i,j
yt+1 = 1. Thus, removing the value xjt = 2 as in Fig. 8B causes that now a value yt+1 = 1 in the
lifted layer is surrounded by more zeros, as it is in Fig. 8A. This means that if regions with value xjt = 2
(green) are absent as in Fig. 8B, regions with value xjt = 3 (orange) are smaller, compared to those of
Fig. 8A. This rather subtle effect is thus explained through the correlations that are clearly revealed
through the construction of the CA model. We note that clustered structures implying formation
of domains arise naturally in many physical experimental systems under global constraints (see e.g.
[32]) and (in an oscillatory fashion) in nonlinear electrochemical systems (see e.g. Fig. 2b in [33]).
Subclustering (’clusters-within-clusters’) has been found ubiquitous in parameter space in a modified
version of the complex Ginzburg-Landau equation under nonlinear global coupling [34, 35]. Note
that with the method presented here, we can generalize this sub-clustering to an arbitrary number of
layers, with the possibility of modeling the individual and differentiated dynamical behavior of each
substructure as well as its interactions and relationships to the rest.

IV. CONCLUSIONS

In this article we have presented a diagrammatic approach for the study and design of sophisticated
CA models, as well as a general framework to mathematically formulate and analyze these models.
In a previous paper [12], we presented the main idea of the p-decomposability of CA rules (whose
consequences have been further explored here) and the way of constructing graded CA rules or of
decomposing complicated non-graded CA rules with coupled layers. The theory presented in that
paper is here completed by establishing the means and the systematic approach to compose non-
graded CA rules, so that the couplings, correlations and interactions of complex systems can be

24
systematically modeled. Here, the traditional tables containing all possible outputs for configurations
of the CAs are no longer necessary, and a few integer parameters suffice to produce complex CA
models with desired properties.
We have presented several examples showing how CA rules can be endowed with useful algebraic
structure and how the spatiotemporal evolution of the CA indeed reveals closed algebraic substructures
hidden within the specification of the rule. We have also shown how the majority rule can be used
as a building block to spatially separate structures and substructures in the form of homogeneous
domains and subdomains and we have thus generalized the majority rule to an arbitrary number of
symbols and layers. With the mathematical method presented in this paper we can, with help of
the diagrams, construct CA models, with as many substructures as layers, systematically including
domains within domains in a ’Matryoshka dolls’ fashion by means of the generalized majority rule
discussed in this manuscript. Since the bulk dynamics of the domains can be systematically designed,
as suggested by Fig. 6 (save for the complicated interaction with the borders and/or surprising
correlations that may appear through the topology), the approach presented constitutes a general
mechanism for the emergence of complex forms (’organisms’) constituted by distinct spatially extended
parts with different dynamical behaviors and potentially different functions.

[1] S. Wolfram, A New Kind of Science, Wolfram Media Inc., Champaign, IL, 2002.
[2] S. Wolfram, Cellular Automata and Complexity: Collected Papers, Addison-Wesley, Reading,
MA, 1994.
[3] L. O. Chua, A Nonlinear Dynamics Perspective of Wolfram’s New Kind of Science, vol. I-VI,
World Scientific, Singapore, 2013.
[4] A. Ilachinski, Cellular Automata: a Discrete Universe, World Scientific, Singapore, 2001.
[5] A. Adamatzky, Identification of Cellular Automata, Taylor and Francis, London, 1994.
[6] H. V. McIntosh, One Dimensional Cellular Automata, Luniver Press, Frome, UK, 2009.
[7] A. Wuensche, M. Lesser, The Global Dynamics of Cellular Automata, Addison-Wesley, Reading,
MA, 1992.
[8] T. Toffoli, Cellular automata as an alternative to (rather than an approximation of) differential
equations in modeling physics, Physica D 10 (1984) 117.
[9] V. Garcı́a-Morales, Universal map for cellular automata, Phys. Lett. A 376 (2012) 2645–2657.
[10] T. Tokihiro, D. Takahashi, J. Matsukidaira, J. Satsuma, From soliton equations to integrable
cellular automata through a limiting procedure, Phys. Rev. Lett. 76 (1996) 3247–3250.
[11] A. Nishiyama, T. Tokihiro, Construction of an isotropic cellular automaton for a reaction-diffusion
equation by means of a random walk, J. Phys. Soc. Jpn. 80 (2011) 800–810.
[12] V. Garcı́a-Morales, Semipredictable dynamical systems, Commun. Nonlinear Sci. Numer. Simulat.
39 (2016) 81–98.
[13] V. Garcı́a-Morales, Cellular automaton for chimera states, EPL 114 (2016) 18002.
[14] V. Garcı́a-Morales, The pλn fractal decomposition: Nontrivial partitions of conserved physical
quantities, Chaos Sol. Fract. 83 (2016) 27.
[15] V. Garcı́a-Morales, Fractal surfaces from simple arithmetic operations, Physica A 447 (2016) 535.
[16] V. Garcı́a-Morales, Symmetry analysis of cellular automata, Phys. Lett. A 377 (2013) 276.

25
[17] G. Cantor, Über einfache Zahlensysteme, Zeitschrift für Math. und Physik 14 (1869) 121–128.
[18] J. A. Gallian, Contemporary abstract algebra, Brooks Cole, Belmont, CA, 2010.
[19] V. Garcı́a-Morales, Origin of complexity and conditional predictability in cellular automata, Phys.
Rev. E 88 (2013) 042814.
[20] G. Vichniac, Simulating physics with cellular automata, Physica D 10 (1984) 96–115.
[21] M. Tchuente, Contribution a l’etude des methodes de calcul pour des systemes de type cooperatif,
Thesis, University of Grenoble, France, 1982.
[22] E. Goles, M. Tchuente, Iterative behavior of one-dimensional threshold automata, Disc. Appl.
Math. 8 (1984) 319–322.
[23] E. Goles, S. Martinez, Neural and Automata Networks, Kluwer, Amsterdam, 1990.
[24] Z. Agur, Fixed points of majority rule cellular automata with application to plasticity and pre-
cision of the immune system, Complex Systems 5 (1991) 351–357.
[25] Z. Agur, Resilience and variability in pathogens and hosts, IMA J. Math. Appl. Med. Biol. 4
(1987) 295–307.
[26] T. Sagawa, Thermodynamics of Information Processing in Small Systems, Springer, Tokyo,
Japan, 2013.
[27] C. Tsallis, Possible generalization of Boltzmann-Gibbs statistics, J. Stat. Phys. 52 (1988) 479.
[28] V. Garcı́a-Morales, J. Pellicer, Microcanonical foundation of nonextensivity and generalized ther-
mostatistics based on the fractality of the phase space, Physica A 361 (2006) 161.
[29] C. Beck, E. G. D. Cohen, Superstatistics, Physica A 322 (2003) 267–275.
[30] V. Garcı́a-Morales, K. Krischer, Superstatistics in nanoscale electrochemical systems, Proc. Natl.
Acad. Sci. USA 108 (49) (2011) 19535–19539.
[31] V. Garcı́a-Morales, J. Cervera, J. Pellicer, Coupling theory for counterion distributions based in
tsallis statistics, Physica A 339 (2004) 482.
[32] M. Conti, B. Meerson, A. Peleg, and P. V. Sasorov, Phase ordering with a global conservation
law: Ostwald ripening and coalescence, Phys. Rev. E 65, 046117 (2002).
[33] L. Schmidt, K. Schönleber, K. Krischer, V. Garcia-Morales, Coexistence of synchrony and inco-
herence in oscillatory media under nonlinear global coupling, Chaos 24 (2014) 013102.
[34] I. Miethe, V. Garcı́a-Morales, K. Krischer, Irregular subharmonic cluster patterns in an au-
tonomous photoelectrochemical oscillator, Phys. Rev. Lett. 102 (2009) 194101.
[35] V. Garcı́a-Morales, A. Orlov, K. Krischer, Subharmonic phase clusters in the complex ginzburg-
landau equation with nonlinear global coupling, Phys. Rev. E 82 (2010) 065202(R).

26