Quantum Cellular Automata 101

Examples of QCA circuit and thus are computationally universal

Raussendorf proved an explicit construction of (Perez-Delgado and Cheung 2007). Block-
QCA and proved its computational universality partitioned QCA can simulate a quantum computer
(Raussendorf 2005). The QCA lives on a torus with linear overhead in time and space (Brennen and
with a 2  2 Margolus partitioning. The update Williams 2003). Continuous-time QCA are in com-
rule is given by a single four-qubit unitary acting plexity class BQP and thus computationally
on 2  2 blocks of qubits. The four-qubit unitary universal (Vollbrecht and Cirac 2008). The
operation consists of swap operations, the explicit constructions of 2-dimensional QCA by
Hadamard transformation, and a phase gate. Raussendorf are computationally universal; more
The initial state of the QCA is prepared such precisely, it is within a constant as efficient as a
that columns encode alternatingly data and pro- quantum logic network with local and nearest-
gram. When the QCA is running, the data travel neighbor gates (Raussendorf 2005). Shepherd,
in one direction, while the program (encoding Franz, and Werner provided an explicit construc-
classical information in orthogonal states) tion of a 12-state 1-dimensional QCA which is in
travels in the opposite direction. Where the complexity class BQP. It is universally program-
two cross, the computation is carried out mable in the sense that it simulates any quantum-
through nearest-neighbor interaction. After a gate circuit with polynomial overhead (Shepherd
fixed number of steps, the computation is et al. 2006). Arrighi and Fargetton proposed a
done, and the result can be read out of a dedi- 1-dimensional QCA capable of simulating any
cated “data” column. This QCA is computation- other 1-dimensional QCA with linear overhead
ally universal; more precisely, it is within a (Arrighi and Fargetton 2007).
constant as efficient as a quantum logic network Implementations of computationally universal
with local and nearest-neighbor gates. QCA have been suggested by Lloyd (1993) and
Shepherd, Franz, and Werner compared classi- Benjamin (2001).
cally controlled QCA to autonomous QCA
(Shepherd et al. 2006). The former is controlled
by a classical compiler that selects a sequence of Modeling Physical Systems
operations acting on the QCA at each time step.
The latter operates autonomously, performing the One of the goals in developing QCA is to create
same unitary operation at each time step. The only a useful modeling tool for physical systems.
step that is classically controlled is the measure- Physical systems that can be simulated with
ment (and initialization). They show the compu- QCA include Ising and Heisenberg interaction
tational equivalence of the two models. Their spin chains, solid state NMR, and quantum lat-
result implies that a particular quantum simulator tice gases. Spin chains are perhaps the most
may be as powerful as a general one. obvious systems to model with QCA. The sim-
ple cases of such 1-dimensional lattices of spins
are Hamiltonians which commute with their
Computationally Universal QCA own lattice translations. Vollbrecht and Cirac
showed that the computability of the ground
Quite a few models have been shown to be com- state energy of a translationally invariant
putationally universal, i.e., they can simulate any n-neighbor Hamiltonian is in complexity class
quantum Turing machine and any quantum circuit QMA (Vollbrecht and Cirac 2008). For simulat-
efficiently. A Watrous QCA simulates any quan- ing noncommuting Hamiltonians, a block-wise
tum Turing machine with constant slowdown update such as the Margolus partitioning has to
(Watrous 1995). The QCA defined by Van Dam be used (see section on “Reversible Cellular
is a finite version of a Watrous QCA and is com- Automata”). Here the fact is used that any Ham-
putationally universal as well (van Dam 1996). iltonian can be expressed as the sum of two
Local-unitary QCA can simulate any quantum Hamiltonians, H = Ha + Hb. Ha and Hb can
102 Quantum Cellular Automata

then, to a good approximation, be applied which makes QCA a very suitable framework in
sequentially to yield the original Hamiltonian which to study their computational power. Optical
H, even if these do not commute. It has been lattices are artificial crystals of light and consist of
shown that such 1-dimensional spin chains can hundreds of thousands of microtraps. One or more
be simulated efficiently on a classical computer neutral atoms can be trapped in each of the potential
(Vidal 2004). It is not known, however, whether minima. If the potential minima are deep enough,
higher-dimensional spin systems can be simu- any tunneling between the traps is suppressed, and
lated efficiently classically. each site contains the same amount of atoms.
A quantum register – here in form of a so-called
Quantum Lattice Gas Automata Mott insulator – has been created. The biggest chal-
Any numerical evolution of a discretized partial lenge at the moment is to find a way to address the
differential equation can be interpreted as the evo- registers individually to implement quantum gates.
lution of some CA, using the framework of lattice For a QCA, all that is needed is implementing the
gas automata. In the continuous time and space unitary operation(s) acting on the entire lattice
limit, such a CA mimics the behavior of the partial simultaneously. The internal structure of the QCA
differential equation. In quantum mechanical lat- guarantees the locality of the operations. This is a
tice gas automata (QLGA), the continuous limit huge simplification compared to individual manip-
on a set of so-called quantum lattice Boltzmann ulation of the registers. Optical lattices are created
equation recovers the Schrödinger equation routinely by superimposing two or three orthogonal
(Succi and Benzi 1993). The first formulation of standing waves generated from laser beams of a
a linear unitary CA was given in Bialynicki-Birula certain frequency. They are used to study fermionic
(1994). Meyer coined the term quantum lattice and bosonic quantum gases, nonlinear quantum
gas automata (QLGA) and demonstrated the dynamics, and strongly correlated quantum phases,
equivalence of a QLGA and the evolution of a to name a few.
set of quantum lattice Boltzmann equations A type of locally addressed architecture by
(Meyer 1996a, b). Meyer (1997), Boghosian and global control was put forward by Lloyd (1993).
Taylor (1998a), and Love and Boghosian (2005) In this scheme, a 1-dimensional array is built
explored the idea of using QLGA as a model for out of three atomic species, periodically arranged
simulating physical systems. Algorithms for as AℬCAℬCAℬC. Each species encodes a qubit
implementing QLGA on a quantum computer and can be manipulated without affecting the
have been presented in Boghosian and Taylor other species. The operations on any species can
(1998b), Meyer (2002), Ortiz et al. (2001). be controlled by the states of the neighboring
cells. The end-cells are used for readout, since
they are the only individually addressable compo-
Implementations nents. Lloyd showed that such a quantum archi-
tecture is universal. Benjamin investigated the
A large effort is being made in many laboratories minimum physical requirements for such a
around the world to implement a model of a quan- many-species implementation and found a similar
tum computer. So far all of them are confined to a architecture using only two types of species, again
very finite number of elements and are no way arranged periodically AℬAℬAℬ (Benjamin
near to a quantum Turing machine (which in itself 2000, 2001; Benjamin and Bose 2004). By giving
is a purely theoretical construct but can be approx- explicit sequences of operations implementing
imated by a very large number of computational one-qubit and two-qubit (CNOT) operations,
elements). One existing experimental setup that is Benjamin showed computational universality.
very promising for quantum information pro- But the reduction in spin resources comes with
cessing and that does not suffer from this “finite- an increase in logical encoding into four spin sites
ness” is optical lattices (for a review, see Bloch with a buffer space of at least four empty spin sites
2005). They possess a translation symmetry between each logical qubit.
Quantum Cellular Automata 103

A continuation of this multispecies QCA archi- would be a way around practical issues related to
tecture is found in the work by Twamley (2003). the implementation of quantum registers and the
Twamley constructed a proposal for a QCA archi- difficulty of their individual manipulation.
tecture based on Fullerene (C60) molecules doped More concretely, QCA provide a natural
with atomic species 15N and 31P, respectively, framework for describing quantum dynamical
arranged alternatingly in a 1-dimensional array. evolution of optical lattices, a field in which the
Instead of electron spins which would be too sen- experimental physics community has made huge
sitive to stray electric charges, the quantum infor- progress in the last decade.
mation is encoded in the nuclear spins. Twamley The main focus so far has been on reversible
constructed sequences of pulses implementing QCA. Irreversible QCA are closely related to
Benjamin’s scheme for one- and two-qubit opera- measurement-based computation and remain to
tions. The weakest point of the proposal is the be explored further.
readout operation which is not well defined.
A different scheme for implementing QCA was
suggested by Tóth and Lent (2001). Their scheme
Bibliography
is based on the technique of quantum-dot CA. The
term quantum-dot CA is usually used for CA
Primary Literature
implementations in quantum dots (for classical Aoun B, Tarifi M (2004) Introduction to quantum cellular
computation). The authors, therefore, called their automata. http://arxiv.org/abs/quant-ph/0401123
model a coherent quantum-dot CA. They illus- Arrighi P (2006) Algebraic characterizations of unitary
trated the usage of an array of N quantum dots as linear quantum cellular automata. In: Mathematical
foundations of computer science 2006, vol 4162, Lec-
an N-qubit quantum register. However, the setup ture notes in computer science. Springer, Berlin,
and the allowed operations allow for individual pp 122–133
control of each cell. This coherent quantum-dot Arrighi P, Fargetton R (2007) Intrinsically universal one-
CA is more a hybrid of a quantum circuit with dimensional quantum cellular automata. 0704.3961.
http://arxiv.org/abs/0704.3961
individual qubit control and a QCA with constant Arrighi P, Nesme V, Werner R (2007) One-dimensional
nearest-neighbor interaction. The main property quantum cellular automata over finite, unbounded con-
of a QCA, operating under global control only, is figurations. 0711.3517v1. http://arxiv.org/abs/0711.
not taken advantage of. 3517
Arrighi P, Nesme V, Werner R (2007) N-dimensional quan-
tum cellular automata. 0711.3975v1. http://arxiv.org/
abs/arXiv:0711.3975
Future Directions Benioff P (1980) The computer as a physical system: a
microscopic quantum mechanical hamiltonian model
of computers as represented by turing machines. J Stat
The field of QCA is developing rapidly. New Phys 22:563–591
definitions have appeared very recently. Since Benjamin SC (2000) Schemes for parallel quantum com-
QCA are now considered to be one of the standard putation without local control of qubits. Phys Rev
measurement-based models of quantum computa- A 61:020301–020304
Benjamin SC (2001) Quantum computing without local
tion, further work on a consistent and sufficient control of qubit-qubit interactions. Phys Rev Lett
definition of higher-dimensional QCA is to be 88(1):017904
expected. One proposal for such a “final” definition Benjamin SC, Bose S (2004) Quantum computing in arrays
has been put forward in (Arrighi et al. 2007a, b). coupled by “always-on” interactions. Phys Rev
A 70:032314
In the search for robust and easily Bialynicki-Birula I (1994) Weyl, Dirac, and Maxwell equa-
implementable quantum computational architec- tions on a lattice as unitary cellular automata. Phys Rev
tures, QCA are of considerable interest. The main D 49:6920
strength of QCA is global control without the Bloch I (2005) Ultracold quantum gases in optical lattices.
Nat Phys 1:23–30
need to address cells individually (with the possi- Boghosian BM, Taylor W (1998a) Quantum lattice-gas
ble exception of the readout operation). It has model for the many-particle Schrödinger equation in d
become clear that the global update of a QCA dimensions. Phys Rev E 57:54
104 Quantum Cellular Automata

Boghosian BM, Taylor W (1998b) Simulating quantum Perez-Delgado CA, Cheung D (2005) Models of quantum
mechanics on a quantum computer. Phys D Nonlinear cellular automata. http://arxiv.org/abs/quant-ph/0508164
Phenom 120:30–42 Perez-Delgado CA, Cheung D (2007) Local unitary quan-
Brennen GK, Williams JE (2003) Entanglement dynamics tum cellular automata. Phys Rev A (At Mol Opt Phys)
in one-dimensional quantum cellular automata. Phys 76:032320–15
Rev A 68:042311 Raussendorf R (2005) Quantum cellular automaton for
Cook M (2004) Universality in elementary cellular autom- universal quantum computation. Phys Rev A (At Mol
ata. Complex Syst 15:1 Opt Phys) 72:022301–022304
Dürr C, Santha M (2002) A decision procedure for unitary Richter W (1996) Ergodicity of quantum cellular automata.
linear quantum cellular automata. SIAM J Comput J Stat Phys 82:963–998
31:1076–1089 Schumacher B, Werner RF (2004) Reversible quantum
Dürr C, LêThanh H, Santha M (1997) A decision proce- cellular automata. quant-ph/0405174. http://arxiv.org/
dure for well-formed linear quantum cellular automata. abs/quant-ph/0405174
Random Struct Algorithm 11:381–394 Shepherd DJ, Franz T, Werner RF (2006) Universally pro-
Feynman R (1982) Simulating physics with computers. Int grammable quantum cellular automaton. Phys Rev Lett
J Theor Phys 21:467–488 97:020502–020504
Fussy S, Grössing G, Schwabl H, Scrinzi A (1993) Non- Succi S, Benzi R (1993) Lattice Boltzmann equation for
local computation in quantum cellular automata. Phys quantum mechanics. Phys D Nonlinear Phenom
Rev A 48:3470 69:327–332
Grössing G, Zeilinger A (1988) Quantum cellular autom- Toffoli T, Margolus NH (1990) Invertible cellular autom-
ata. Complex Syst 2:197–208 ata: a review. Phys D Nonlinear Phenom 45:229–253
Gruska J (1999) Quantum computing. Osborne/McGraw- Tóth G, Lent CS (2001) Quantum computing with
Hill, New York, QCA are treated in Section 4.3 quantum-dot cellular automata. Phys Rev A 63:052315
Kempe J (2003) Quantum random walks: an introductory Twamley J (2003) Quantum-cellular-automata quantum
overview. Contemp Phys 44:307 computing with endohedral fullerenes. Phys Rev
Lloyd S (1993) A potentially realizable quantum computer. A 67:052318
Science 261:1569–1571 van Dam W (1996) Quantum cellular automata. Master’s
Love P, Boghosian B (2005) From Dirac to diffusion: thesis, University of Nijmegen
decoherence in quantum lattice gases. Quantum Inf Vidal G (2004) Efficient simulation of one-dimensional
Process 4:335–354 quantum many-body systems. Phys Rev Lett
Margolus N (1991) Parallel quantum computation. In: 93(4):040502
Zurek WH (ed) Complexity, entropy, and the physics Vollbrecht KGH, Cirac JI (2008) Quantum simulators,
of information, Santa Fe Institute series. Addison Wes- continuous-time automata, and translationally invariant
ley, Redwood City, pp 273–288 systems. Phys Rev Lett 100:010501
Meyer DA (1996a) From quantum cellular automata to von Neumann J (1966) Theory of self-reproducing autom-
quantum lattice gases. J Stat Phys 85:551–574 ata. University of Illinois Press, Champaign
Meyer DA (1996b) On the absence of homogeneous Watrous J (1995) On one-dimensional quantum cellular
scalar unitary cellular automata. Phys Lett A 223: automata. In: Proceedings of the 36th annual sympo-
337–340 sium on foundations of computer science, Milwaukee,
Meyer DA (1997) Quantum mechanics of lattice gas pp 528–537
automata: one-particle plane waves and potentials. Wolfram S (1983) Statistical mechanics of cellular autom-
Phys Rev E 55:5261 ata. Rev Mod Phys 55:601
Meyer DA (2002) Quantum computing classical physics. Wootters WK, Zurek WH (1982) A single quantum cannot
Philos Trans R Soc A 360:395–405 be cloned. Nature 299:802–803
Nagaj D, Wocjan P (2008) Hamiltonian quantum cellular
automata in 1d. 0802.0886. http://arxiv.org/abs/0802.
0886 Books and Reviews
Ortiz G, Gubernatis JE, Knill E, Laflamme R (2001) Quan- Summaries of the topic of QCA can be found in chapter 4.3
tum algorithms for fermionic simulations. Phys Rev of Gruska (Grössing and Zeilinger 1988), and in Aoun
A 64:022319 and Tarifi (2004) and Ortiz et al. (2001)
 divided into infinitely many blocks of the same
Reversible Cellular Automata size (in the two-dimensional case, e.g., 2
2).
A local transition function consisting of “block
Kenichi Morita rules,” which is a mapping from a block state to
Hiroshima University, Higashi-Hiroshima, Japan a block state, is applied to all the blocks in
parallel. At the next time step, the block divi-
sion pattern is shifted by some fixed amount
Article Outline (e.g., to the north-east direction by one cell),
and the same local function is applied to them.
Glossary This model of CA is convenient to design a
Definition of the Subject reversible CA. This is because if the local
Introduction transition function is injective, then the
Reversible Cellular Automata resulting CA is reversible.
How Can We Find RCAs? Partitioned cellular automaton A partitioned
Simulating Irreversible Cellular Automata by cellular automaton (PCA) is a framework for
Reversible Ones designing a reversible CA. It is a subclass of a
1-D Universal Reversible Cellular Automata usual CA where each cell is divided into sev-
Simulating Cyclic Tag Systems by 1-D RCAs eral parts, whose number is equal to the neigh-
2-D Universal Reversible Cellular Automata That borhood size. Each part of a cell has its own
Can Simulate Reversible Logic Gates state set and can be regarded as an output port
Future Directions to a specified neighboring cell. Depending
Bibliography only on the corresponding parts (not on the
whole states) of the neighboring cells, the
next state of each cell is determined by a local
Glossary function. We can see that if the local function is
injective, then the resulting PCA is reversible.
Cellular automaton A cellular automaton Hence, a PCA makes it feasible to construct a
(CA) is a system consisting of a large reversible CA.
(theoretically, infinite) number of finite autom- Reversible cellular automaton A reversible cel-
ata, called cells, which are connected uni- lular automaton (RCA) is defined as a CA
formly in a space. Each cell changes its state whose global function is injective (i.e.,
depending on the states of itself and the cells in one-to-one). It can be regarded as a kind of a
its neighborhood. Thus, the state transition of a discrete model of reversible physical space. It
cell is specified by a local function. Applying is in general difficult to construct an RCA with
the local function to all the cells in the space a desired property such as computational uni-
synchronously, the transition of a configuration versality. Therefore, the frameworks of a CA
(i.e., a whole state of the cellular space) is with Margolus neighborhood, a partitioned
induced. Such a transition function is called a cellular automaton, and others are often used
global function. A CA is regarded as a kind of to design RCAs.
dynamical system that can deal with various Universal cellular automaton A CA is called
kinds of spatiotemporal phenomena. computationally universal (or Turing univer-
Cellular automaton with block rules A CA sal), if it can simulate a universal Turing
with block rules was proposed by Margolus machine, or equivalently, it can compute
(1984), and it is often called a CA with any recursive function by giving an appropri-
Margolus neighborhood. The cellular space is ate initial configuration. Computational
# Springer Science+Business Media LLC, part of Springer Nature 2018 105
A. Adamatzky (ed.), Cellular Automata,
https://doi.org/10.1007/978-1-4939-8700-9_455
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_455-7
106 Reversible Cellular Automata

universality of RCAs can be proved by simu- following Garden-of-Eden theorem: A CA has a

lating other systems such as arbitrary Garden-of-Eden configuration, if and only if it has
(generally irreversible) CAs, reversible Turing an “erasable configuration.” After that, many
machines, reversible counter machines, and researchers studied on injectivity and surjectivity
reversible logic elements and circuits, which of global functions more generally (Amoroso and
have already been known to be universal. Cooper 1970; Maruoka and Kimura 1976, 1979;
Richardson 1972). In particular, Richardson
Definition of the Subject (1972) showed that if a CA is injective, then it is
surjective.
Reversible cellular automata (RCAs) are defined Toffoli (1977) first studied reversible (i.e.,
as cellular automata (CAs) with an injective injective) CAs from the computational viewpoint.
global function. Every configuration of an RCA He showed that every k-dimensional irreversible
has exactly one previous configuration, and thus CA can be simulated by a (k + 1)-dimensional
RCAs are “backward deterministic” CAs. The RCA. Hence, a two-dimensional RCA has univer-
notion of reversibility originally comes from sal computing capability. Since then, extensive
physics. It is one of the fundamental microscopic studies on RCAs have been done until now.
physical laws of nature. In this sense, an RCA is After the pioneering work of Bennett (1973) on
thought as an abstract model of a physically reversible Turing machines, several models of
reversible space as well as a computing model. It reversible computing were proposed besides
is very important to investigate how computation RCAs. They are, for example, reversible logic
can be carried out efficiently and elegantly in a circuits (Fredkin and Toffoli 1982; Morita 2001;
system having reversibility. This is because future Toffoli 1980), billiard ball model (BBM) of com-
computing devices will surely become those of a puting (Fredkin and Toffoli 1982), and reversible
nanoscale size. counter machines (Morita 1996). Most of these
In this entry, we mainly discuss on the proper- models have a close relation to physical revers-
ties of RCAs from the computational aspects. In ibility. In fact, reversible computing plays an
spite of the strong constraint of reversibility, important role when considering inevitable
RCAs have very high capability of computing. power dissipation in computing (Bennett 1973,
We can see that even very simple RCAs have 1982; Bennett and Landauer 1985; Landauer
universal computing power. We can also recog- 1961; Toffoli and Margolus 1990). It is also one
nize, in some reversible cellular automata, that of the bases of quantum computing (see, e.g.,
computation is carried out in a very different Gruska 1999) because an evolution of a quantum
manner from conventional computing systems; system is a reversible process.
thus, they will give new ways and concepts for In this entry, we discuss how RCAs can have
future computing. universal computing capability and how simple
they can be. Since reversibility is one of the fun-
damental microscopic properties of physical sys-
Introduction tems, it is important to investigate whether we can
use such physical mechanisms directly for com-
Problems related to injectivity and surjectivity of putation. An RCA is a useful framework to for-
global functions of CAs were first studied by malize and investigate these problems. Since this
Moore (1962) and Myhill (1963) in the Garden- entry is not an exhaustive survey, many interest-
of-Eden problem. A Garden-of-Eden configura- ing topics related to RCAs, such as complexity of
tion is such that it can exist only at time zero, i.e., RCA (Sutner 2004), relations to quantum CA
it has no predecessor configuration. Therefore, if a (e.g., Watrous 1995), etc., are omitted.
CA has a Garden-of-Eden configuration, then its An outline of the following sections is as fol-
global function is not surjective. They proved the lows. In section “Reversible Cellular Automata,”
Reversible Cellular Automata 107

we give basic definitions on RCAs and design 8a  Conf ðAÞ, x  Zk :

methods for obtaining RCAs. In section “Simu- FðaÞðxÞ ¼ f ðaðx þ n1 Þ, . . . , aðx þ nm ÞÞ
lating Irreversible Cellular Automata by Revers-
ible Ones,” it is shown how irreversible CAs are Definition 2 Let A = (ℤk, Q, (n1,. . ., nm), f, #) be a
simulated by RCAs. In section “1-D Universal CA. (1) A is called an injective CA if F is injec-
Reversible Cellular Automata,” two kinds of com- tive. (2) A is called an invertible CA if there is a
putationally universal 1-D RCAs are shown. In CA A0 = (ℤk, Q, N0, f 0, #) that satisfies the follow-
section “2-D Universal Reversible Cellular ing condition:
Automata,” several universal 2-D RCAs with
very simple local functions are shown. In section 8a, b  Conf ðAÞ : FðaÞ ¼ b iff F0 ðbÞ ¼ a,
“Future Directions,” we discuss future perspec-
tives and open problems as well as some other where F and F0 are the global functions of A and
problems on RCAs not given in the previous A0 , respectively.
sections. The following theorem can be derived from the
results independently proved by Hedlund (1969)
and Richardson (1972).
Reversible Cellular Automata
Theorem 1 (Hedlund 1969; Richardson 1972) A
We first give definitions on conventional cellular CA A is injective iff it is invertible.
automata (CAs) and then reversible CAs. Next, By the above theorem, we see the notions of
we give design methods of reversible CAs. injectivity and invertibility are equivalent. Hence-
forth, we use the terminology reversible CA (RCA)
for such a CA, instead of injective CA or invertible
Formal Definitions CA, because an RCA is regarded as an analog of
Definition 1 A deterministic k-dimensional (k-D) physically reversible space. (Note that, in some
m-neighbor cellular automaton (CA) is a system other computing models such as Turing machines,
defined by counter machines, and logic circuits, injectivity is

trivially equivalent to invertibility, if they are suit-
A ¼ ℤk , Q, ðn1 , . . . , nm Þ, f ,# , ably defined. Therefore, for these models, we can
directly define reversibility without introducing the
where ℤ is the set of all integers (hence ℤk is the notions of injectivity and invertibility.)
set of all k-dimensional points with integer coor-
dinates at which cells are placed), Q is a non-
empty finite set of states of each cell, (n1,. . ., nm) is How Can We Find RCAs?
an element of (ℤk)m called a neighborhood
(m = 1, 2,. . .), f: Qm ! Q is a local function, The class of RCAs is a special subclass of CAs.
and #  Q is a quiescent state that satisfies Therefore, there arises a problem how we can find
f(#,. . ., #) = #. We also allow a CA such that no or construct RCAs with some desired property. It
quiescent state is specified. is in general hard to do so if we use the conven-
A configuration of A is a mapping a: ℤk ! Q. tional framework of CAs. This is because the
Let Conf(A) denote the set of all configurations of following result is shown by Kari (1994) for the
A, i.e., Conf(A) = {a| a: ℤk ! Q}. We say that a two-dimensional case (hence, it also holds for
configuration a is finite if the set {x| x  ℤk ^ higher-dimensional CAs).
a(x) 6¼ #} is finite. Otherwise, it is called infinite.
The global function F: Conf(A) ! Conf(A) is Theorem 2 (Kari 1994) The problem whether a
defined as the one that satisfies the following given two-dimensional CA is reversible is
formula: undecidable.
108 Reversible Cellular Automata

For the case of one-dimensional CA, it is

known to be decidable.

Theorem 3 (Amoroso and Patt 1972) There is an

algorithm to test whether a given one-dimensional
CA is reversible or not.
There are also several studies on enumerating
all reversible one-dimensional CAs (e.g., Boykett
2004; Mora et al. 2005). But it is generally diffi-
cult to find RCAs with specific properties such as
computational universality, even for the one- Reversible Cellular Automata, Fig. 1 A cellular space
dimensional case. with the Margolus neighborhood
In order to make it feasible to design an RCA,
several methods have been proposed until now.
They are, for example, CAs with block rules
(Margolus 1984; Toffoli and Margolus 1990),
partitioned CAs (Morita and Harao 1989), CAs
with second-order rules (Margolus 1984; Toffoli
et al. 2004; Toffoli and Margolus 1990), and Reversible Cellular Automata, Fig. 2 Block rules for
others (see, e.g., Toffoli and Margolus 1990). the Margolus RCA (1984)
Here, we describe the first two methods in detail.

Partitioned Cellular Automata

Cellular Automata with Block Rules The method of using partitioned cellular automata
Margolus (1984) proposed an interesting variant (PCA) has some similarity to the one that uses
of a CA, by which he composed a computation- block rules. However, resulting reversible CAs
ally universal two-dimensional two-state RCA. are in the framework of conventional CA
In his model, all the cells are grouped into (in other words, a PCA is a special subclass of a
“blocks” of size 2
2 as shown in Fig. 1. A par- CA). In addition, flexibility of neighborhood is
ticular example of a transformation specified by rather high. A shortcoming of PCA is that, in
“block rules” is shown in Fig. 2. This CA general, the number of states per cell becomes
evolves as follows: At time 0, the local transfor- large.
mation is applied to every solid line block, then
at time 1 to every dotted line block, and so on, Definition 3 A deterministic k-dimensional
alternately. Since this local transformation is m-neighbor partitioned cellular automaton
injective, the global function of the CA is also (PCA) is a system defined by
injective. Such a neighborhood is called


Margolus neighborhood. P ¼ ℤk , ðQ1 , . . . , Qm Þ, ðn1 , . . . , nm Þ, f , ð#1 , . . . , #m Þ ,
One can obtain reversible CAs, by giving an
injective block transformation. However, CAs where ℤ is the set of all integers, Q i (i = 1,. . ., m)
with Margolus neighborhood are not conventional is a nonempty finite set of states of the i-th part of
CAs, because each cell should know the relative each cell (thus the state set of each cell is
position in a block and the parity of time besides Q = Q1
  
Qm), (n1,. . ., nm)  (ℤk)m is a
its own state. neighborhood, f: Q ! Q is a local function, and
Related to this topic, Kari (1996) showed that (#1,. . ., #m)  Q is a quiescent state that satisfies f
every one- and two-dimensional RCA can be (#1,. . ., #m) = (#1,. . ., #m).
represented by a block permutations and The notion of a finite (or infinite) configuration
translations. is defined similarly as in CA. Let Conf(P) = {a| a:
Reversible Cellular Automata 109

ℤk ! Q}, and let pi: Q ! Qi be the projection L C R L C R L C R

function such that pi (q1,. . ., qm) = qi for all (q1,. . .,
qm)  Q. The global function F: Conf(P) ! Conf t
(P) of P is defined as the one that satisfies the
following formula: f

8a  Conf ðPÞ, x  ℤk : t +1
FðaÞðxÞ ¼ f ðp1 ðaðx þ n1 ÞÞ, . . . , pm ðaðx þ nm ÞÞÞ
x−1 x x+1
By the above definition, a one-dimensional
PCA P1d with the neighborhood (1, 0, 1) can Reversible Cellular Automata, Fig. 3 One-dimensional
be defined as follows: three-neighbor PCA P1d and its local function f

P1d ¼ ðℤ, ðL, C, RÞ, ð1, 0,  1Þ, f , ð#,#,#ÞÞ

r c l l c r
Each cell is divided into three parts, i.e., left,
center, and right parts, and their state sets are L, C, Reversible Cellular Automata, Fig. 4 A pictorial rep-
resentation of a local rule f(l, c, r) = (l0, c0, r0) of a one-
and R. The next state of a cell is determined by the dimensional three-neighbor PCA P1d
present states of the left part of the right-neighbor
cell, the center part of this cell, and the right part of
the left-neighbor cell (not depending on the whole
three parts of the three cells). Figure 3 shows its
cellular space and how the local function f works.

Let (l, c, r), (l0, c0, r0)  L
C
R. If f(l, c,

r) = (l0, c0, r0), then this equation is called a local
rule (or simply a rule) of the PCA P1d, and it is
sometimes written in a pictorial form as shown in
Fig. 4. Note that, in the pictorial representation,
the arguments of the left-hand side of f(l, c, d
r) = (l0, c0, r0) appear in a reverse order. u
r c l l c r
Similarly, a two-dimensional PCA P2d with d
von Neumann-like neighborhood is defined as u
follows:
Reversible Cellular Automata, Fig. 5 Cellular space of
a two-dimensional five-neighbor PCA P2d and its local rule
P2d ¼ ðℤ2 , ðC, U, R, D, LÞ, ðð0, 0Þ, ð0,  1Þ,
ð1, 0Þ, ð0, 1Þ, ð1, 0ÞÞ, f , ð#, #, #, #, #ÞÞ Proposition 1 The local function f is injective, iff
the global function F is injective.
Figure 5 shows the cellular space of P2d and a It is also easy to see that the class of PCAs is a
pictorial representation of a local rule f(c, u, r, d, subclass of CAs. More precisely, the following
l) = (c0, u0, r0, d0, l0). proposition is derived by extending the domain
Let P = (ℤk, (Q1,. . ., Qm), (n1,. . ., nm), f, of the local function of P.
(#1,. . ., #m)) be a k-dimensional PCA and F be
its global function. It is easy to show the following Proposition 2 For any k-dimensional m-neigh-
proposition (a proof for the one-dimensional case bor PCA P, we have a k-dimensional m-neighbor
given in Morita and Harao (1989) can be extended CA A whose global function is identical with that
to higher dimensions). of P.
110 Reversible Cellular Automata

By the above, if we want to construct an RCA, it As for one-dimensional CA with finite config-
is sufficient to give a PCA whose local function f is urations, reversible simulation is possible without
injective. This makes a design of an RCA feasible. increasing the dimension.

Theorem 5 (Morita 1995) For any one-

Simulating Irreversible Cellular dimensional (irreversible) CA A with finite config-
Automata by Reversible Ones urations, we can construct a one-dimensional
RCA A0 that simulates A (but not in real time).
Toffoli (1977) first showed that for every irrevers-
ible CA, there exists a reversible one that simu-
lates the former by increasing the dimension by 1-D Universal Reversible Cellular
one. From this result, computational universality Automata
of two-dimensional RCA is derived, since it is
easy to embed a Turing machine in a Computational universality of one-dimensional
(irreversible) one-dimensional CA. RCAs can be shown by constructing RCAs that
simulate universal systems such as reversible
Theorem 4 (Toffoli 1977) For any k-dimensional Turing machines or cyclic tag systems.
(irreversible) CA A, we can construct a (k + 1)-
dimensional RCA A0 that simulates A in real time. Simulating Reversible Turing Machines by
Although Toffoli’s proof is rather complex, the 1-D RCAs
idea of the proof is easily implemented by using a It is possible to simulate reversible Turing
PCA. Here we explain it informally. Consider a machines by one-dimensional RCAs. We first
one-dimensional three-neighbor irreversible CA give definitions on reversible Turing machines.
A that evolves as in Fig. 6. Then, we can construct Then, we show how they can be simulated by
a two-dimensional reversible PCA P that simulates RCAs. Bennett (1973) showed a nice construction
A as shown in Fig. 7. The configuration of A is kept method of a reversible Turing machine that simu-
in some row of P. A state of a cell of A is stored in lates a given irreversible Turing machine and never
the left, center, and right parts of a cell in P in leaves garbage signals on its tape at the end of
triplicate. By this, each cell of P can compute the computation. Though TMs of the quadruple for-
next state of the corresponding cell of A correctly. mulation were used in Bennett (1973), here we use
At the same time, the previous states of the cell and TMs of quintuple formulation (Morita 2008).
the left and right neighbor cells (which were used to
compute the next state) are put downward as a Definition 4 A one-tape Turing machine (TM) is
“garbage” signal to keep P reversible. In other defined by
words, the additional dimension is used to record
all the past history of the evolution of A. In this T ¼ ðQ, S, q0 , F, s0 , dÞ,
way, P can simulate A reversibly.
where Q is a nonempty finite set of states, S is a
nonempty finite set of symbols, q0 is an initial state
t =0 q01 q02 q03 q04
(q0  Q), F is a set of final (i.e., accepting) states
1 q11 q12 q13 q14
(F  Q), s0 is a special blank symbol (s0  S), and d
is a move relation which is a subset of
2 q21 q22 q23 q24 (Q
S
S
{, 0, +}
Q). The symbols “,”
“0,” and “+” denote left shift, zero shift, and right
3 q31 q32 q33 q34 shift of the head, respectively. A quintuple [qi, s, s0, d,
qj ]  (Q
S
S
{, 0, +}
Q) means that if
Reversible Cellular Automata, Fig. 6 An example of T reads the symbol s in the state qi, then write s0 , shift
an evolution in an irreversible one-dimensional CA A the head to the direction d, and go to the state qj.
Reversible Cellular Automata 111

t =0 t =1

q01 q01 q01 q02 q02 q02 q03 q03 q03 q04 q04 q04 q11 q11 q11 q12 q12 q12 q13 q13 q13 q14 q14 q14
(q00 ,q01 ,q02 ) (q01 ,q02 ,q03 ) (q02 ,q03 ,q04 ) (q03 ,q04 ,q05 )

t =2 t =3

q21 q21 q21 q22 q22 q22 q23 q23 q23 q24 q24 q24 q31 q31 q31 q32 q32 q32 q33 q33 q33 q34 q34 q34
(q10 ,q11 ,q12 ) (q11 ,q12 ,q13 ) (q12 ,q13 ,q14 ) (q13 ,q14 ,q15 ) (q20 ,q21 ,q22 ) (q21 ,q22 ,q23 ) (q22 ,q23 ,q24 ) (q23 ,q24 ,q25 )

(q00 ,q01 ,q02 ) (q01 ,q02 ,q03 ) (q02 ,q03 ,q04 ) (q03 ,q04 ,q05 ) (q10 ,q11 ,q12 ) (q11 ,q12 ,q13 ) (q12 ,q13 ,q14 ) (q13 ,q14 ,q15 )

(q00 ,q01 ,q02 ) (q01 ,q02 ,q03 ) (q02 ,q03 ,q04 ) (q03 ,q04 ,q05 )

Reversible Cellular Automata, Fig. 7 Simulating the irreversible CA A in Fig. 6 by a two-dimensional reversible PCA P

The TM T is called deterministic if the follow- Theorem 6 (Bennett 1973) For any (irreversible)
ing statement holds for any pair of distinct quin- one-tape Turing machine, there is a reversible three-
tuples [p1, s1, t1, d1, q1] and [p2, s2, t2, d2, q2]: tape Turing machine that simulates the former.
It is also shown in Morita et al. (1989) that for
If p1 ¼ p2 , then s1 6¼ s2 : any irreversible one-tape TM, there is a reversible
one-tape two-symbol TM that simulates the for-
On the other hand, T is called reversible if the mer. To prove computational universality of a
following statement holds for any pair of distinct one-dimensional reversible PCA, it is convenient
quintuples [p1, s1, t1, d1, q1] and [p2, s2, t2, d2, q2]: to simulate a reversible one-tape TM. The follow-
ing theorem was first shown in Morita and Harao
If q1 ¼ q2 , then d1 ¼ d2 ^ t1 6¼ t2 : (1989). Later, the number of states of a reversible
PCA was reduced in Morita (2008) using an RTM
Note that multi-tape reversible TMs can also be of the quintuple formulation.
defined similarly. Hereafter, we consider only
deterministic reversible and deterministic irre- Theorem 7 (Morita and Harao 1989) For any
versible TMs. Hence, the term “deterministic” reversible one-tape TM T, there is a one-
will be omitted. The next theorem shows compu- dimensional three-neighbor reversible PCA
tational universality of a reversible three-tape TM. P that simulates the former.
112 Reversible Cellular Automata

We show how P simulates T. Let T = (Q, S, f ð#, s, pÞ ¼ ðq, t, qÞ if p  Qþ ^ ½p, s, t, 0, q  d

q0, F, s0, d) be a reversible one-tape TM. We f ðp, s,#Þ ¼ ðq, t, qÞ if p  Q ^ ½p, s, t, 0, q  d
assume that q0 does not appear as the fifth
element in any quintuple in d, since we can 4. For each q  Q0 and s  S, define f as follows:
always construct such a reversible TM from an
irreversible one by a method given in Morita f ð#, s, qÞ ¼ ð#, s, qÞ
et al. (1989). We can also assume that for any f ðq, s,#Þ ¼ ðq, s,#Þ
[p, s, t, d, q]  d, if q is a non-halting state,
then d  {, +}, and if q is a halting state, We can see that the right-hand side of each
then d = 0. Now, let Q, Q0, and Q+ be as rule in (1)–(4) differs from that of any other rule,
follows: since T is reversible. Hence, P is reversible.
Assume that the initial computational configura-
Q ¼ fqj ∃p  Q∃s, t  Sð½p, s, t,  , q  dÞg tion of T is
Q0 ¼ fqj ∃p  Q∃s, t  Sð½p, s, t, 0, q  dÞg
  s0 t1   ti1 q0 ti tiþ1   tn s0   
Qþ ¼ fqj ∃p  Q∃s, t  Sð½p, s, t, þ , q  dÞg
where tj  S (j  {1,. . ., n}). Then, set P to the
Note that, since T is an RTM, Q, Q0, and Q+ following configuration:
are mutually disjoint.
  ð#, s0 ,#Þð#, t1 ,#Þ  ð#, ti1 , q0 Þð#, ti ,#Þ
A reversible PCA P that simulates T is as ðq0 , tiþ1 ,#Þ  ð#, tn ,#Þð#, s0 ,#Þ  
follows:
Then, by the rules in (1)–(4), T is simulated
P ¼ ðℤ, ðL, C, RÞ, ð1, 0,  1Þ, f , ð#, s0 ,#ÞÞ step by step. If T becomes a halting state q
L ¼ Q [ Q0 [ fq0 ,#g (  Q0), then two signals qs are created by the
C¼S rules in (3), and then travel leftward and right-
R ¼ Qþ [ Q0 [ fq0 ,#g ward indefinitely by the rules in (4). Note that
P itself cannot halt, because P is reversible. But
The local function f is as below:
the final tape of T is kept unchanged in P. Termi-
nation of a simulation can be sensed from the
1. For each s, t  S, and q  Q  (Q0 [ {q0}),
outside of P by observing if a final state of
define f as follows:
T appears in some cell of P’s configuration as a
kind of a flag.
f ð#, s,#Þ ¼ ð#, s,#Þ
f ð#, s, q0 Þ ¼ ð#, s, q0 Þ
Example 1 Consider a small reversible TM
f ðq0 , s,#Þ ¼ ðq0 , s,#Þ
Tparity = (Q, {0, 1}, q0, {qa}, 0, d) such that
f ðq0 , s, q0 Þ ¼ ð#, t, qÞ if ½q0 , s, t, þ , q  d
Q = {q0, q1, q2, qa, qr}, and d is as follows:
f ðq0 , s, q0 Þ ¼ ðq, t,#Þ if ½q0 , s, t,  , q  d
d ¼ f½q0 , 0, 1, þ , q1 , ½q1 , 0, 1, 0, qa ;
2. For each p, q  Q  (Q0 [ {q0}), and s, t  S,
define f as follows: ½q1 , 1, 0, þ , q2 , ½q2 , 0, 1, 0, qr , ½q2 , 1, 0, þ , q1 g

f ð#, s, pÞ ¼ ð#, t, qÞif p  Qþ ^ ½p, s, t, þ , q  d For a given unary number n on the tape,
f ðp, s,#Þ ¼ ð#, t, qÞif p  Q ^ ½p, s, t, þ , q  d Tparity checks if n is even or odd. If it is even,
f ð#, s, pÞ ¼ ðq, t,#Þif p  Qþ ^ ½p, s, t,  , q  d then Tparity halts in the final (accepting) state qa;
f ðp, s,#Þ ¼ ðq, t,#Þif p  Q ^ ½p, s, t,  , q  d otherwise it halts in the rejecting state qr. All the
symbols read by Tparity are complemented. The
3. For each p  Q  (Q0 [ {q0}), q  Q0, and s, reversible PCA Pparity constructed by the above
t  S, define f as follows: method is as follows:
Reversible Cellular Automata 113

Pparity ¼ ðℤ, ðL, C, RÞ, ð1, 0,  1Þ, f , ð#, 0,#ÞÞ In Morita (2011a), a method of simulating a
L ¼ fq0 , qa , qr g given reversible one-tape TM by a two-neighbor
C ¼ f0, 1g reversible PCA is shown. By this, the total num-
R ¼ fq 0 , q1 , q2 , qa , qr g ber of states of a cell can be reduced.

The local function f is as below: Simulating Cyclic Tag Systems by

1-D RCAs
1. For each s  {0, 1}, define f as follows:
From Theorem 6, we can see the existence of a
f ð#, s,#Þ ¼ ð#, s,#Þ universal reversible TM. In fact, several kinds
f ð#, s, q0 Þ ¼ ð#, s, q0 Þ of small universal reversible TMs have been
f ðq0 , s,#Þ ¼ ðq0 , s,#Þ given (see, e.g., Morita 2017a). Thus, from
f ðq0 , 0, q0 Þ ¼ ð#, 1, q1 Þ Theorem 7, a one-dimensional universal RCA
can be constructed. However, the number of
ðIt simulates ½q0 , 0, 1, þ , q1 Þ
states of an RCA obtained by this method will
become large. To get a universal RCA with a
2. Define f as follows:
small number (say a few dozens) of states, we
need another useful framework of a universal
f ð#, 1, q1 Þ ¼ ð#, 0, q2 ÞðItsimulates ½q1 , 1, 0, þ , q2 Þ system.
f ð#, 1, q2 Þ ¼ ð#, 0, q1 ÞðItsimulates ½q2 , 1, 0, þ , q1 Þ A cyclic tag system (CTAG) is proposed by
Cook (2004) to show universality of the elemen-
3. Define f as follows: tary cellular automaton of rule 110. As we shall
see, it is also useful for composing simple
f ð#, 0, q1 Þ ¼ ðqa , 1, qa Þ ðIt simulates ½q1 , 0, 1, 0, qa Þ universal RCAs.
f ð#, 0, q2 Þ ¼ ðqr , 1, qr Þ ðIt simulates ½q2 , 0, 1, 0, qr Þ
Definition 5 A cyclic tag system (CTAG) is
4. For each s  {0, 1} and q  {qa, qr}, define f defined by C = (k, (p0,. . ., pk  1)), where k
as follows: (k = 1, 2,. . .) is the length of a cycle
(i.e. period) and (p0,. . ., pk  1)  ({Y, N})k is
f ð#, s, qÞ ¼ ð#, s, qÞ a k-tuple of production rules. A pair (v, m) is
f ðq, s,#Þ ¼ ðq, s,#Þ called an instantaneous description (ID) of C,
where v  {Y, N} and m  {0,. . ., k  1}.
Computing process of Tparity for n = 2 is as The nonnegative integer m is called a phase of
follows: q0 0110 ‘ 1 q1 110 ‘ 10 q2 10 ‘ 100 q1 the ID. A transition relation ) on the set of IDs is
0 ‘ 100 qa 1. It is simulated by Pparity as shown in defined as follows. For any (v, m), (v0, m0)  {Y,
Fig. 8. N}
{0,. . ., k  1},

Reversible Cellular
t =0 0 q0 0 q0 1 1 0 0 0
Automata,
Fig. 8 Simulating T parity 1 0 1 q1 1 1 0 0 0
by the reversible PCA
Pparity. The state # is 2 0 1 0 q2 1 0 0 0
indicated by a blank
3 0 1 0 0 q1 0 0 0

4 0 1 0 0 qa 1 q a 0 0

5 0 1 0 qa 0 1 0 qa 0

6 0 1 qa 0 0 1 0 0 qa
114 Reversible Cellular Automata

ðYv, mÞ ) ðv0 , m0 Þ if ½m0 ¼ m þ 1modk ^ ½v0 ¼ vpm , The reversible PCA P24 in Theorem 9 is as
ðNv, mÞ ) ðv0 , m0 Þ if ½m0 ¼ m þ 1modk ^ ½v0 ¼ v: follows:


A sequence of IDs (v0, m0), (v1, m1),. . . is called P24 ¼ Z, ðfY, N,
þ ,g, fy, n, þ ,  ,  ,=gÞ, 0,  1 ,
a computation starting from v  {Y, N} if (v0, f 24 , ðY,Þ
m0) = (v, 0) and (vi, mi) ) (vi + 1, mi + 1) (i = 0,
1,. . .). (In what follows, we write a computation by The state set of each cell is {Y, N, +, }
{y, n,
(v0, m0) ) (v1, m1) )  ). +, , , /}, and thus P24 has 24 states. The local
A CTAG is a variant of a classical tag system function f24 is as below. It is easy to see that f24 is
(see, e.g., Minsky 1967) such that production injective:
rules are applied cyclically. If the first symbol
of a host (i.e., rewritten) string is Y, then it is f 24 ðc, r Þ ¼ ðc, r Þ if c  fY, N g, r  fy, n, þ ,  ,=g
removed, and a specified string at that phase is f 24 ðY,Þ ¼ ðþ,=Þ
attached to the end of the host string. If it is N, f 24 ðN,Þ ¼ ð,=Þ
then it is simply removed and no string is f 24 ð, r Þ ¼ ð, r Þ if r  fy, n,g
attached. f 24 ðc, r Þ ¼ ðr, cÞ if c  fþ,g, r  fþ,g
f 24 ðþ, yÞ ¼ ðY,Þ
Example 2 Let us consider the CTAG C0 = (3, (Y, f 24 ðþ, nÞ ¼ ðN,Þ
NN, YN)). If we give NYY to C0 as an initial string, f 24 ðþ,=Þ ¼ ðþ, yÞ
then f 24 ð,=Þ ¼ ðþ, nÞ
f 24 ðþ,Þ ¼ ðþ,Þ
ðNYY, 0Þ ) ðYY, 1Þ ) ðYNN, 2Þ ) ðNNYN, 0Þ
) ðNYN, 1Þ ) ðYN, 2Þ Consider the CTAG C0 in Example 2. The com-
putation in C0 starting from NYY is simulated in P24
as shown in Fig. 9. The production rules are given
is an initial segment of a computation starting
by a sequence consisting of the states (, y), (, n),
from NYY.
(, ), and (, ) in a reverse order, where the
In Cocke and Minsky (1964) and Minsky
sequence (, )(, ) is used as a delimiter indi-
(1967), it was shown that a two-tag system,
cating the beginning of a rule. Thus, one cycle of the
which is a special class of classical tag systems,
rules (Y, NN, YN) is (, n)(, y)(, )(, )(, n)
is universal. The following theorem shows uni-
(, n)(, )(, )(, y)(, )(, ). We should
versality of CTAG.
give infinite copies of this sequence to the left, since
these rules are applied cyclically. We can see the
Theorem 8 (Cook 2004) For any two-tag system,
right-part states y, n, , and  in this sequence act as
there is a CTAG that simulates the former.
right-moving signals. The initial string NYY is given
It was shown that there are universal one-
to the right of it by the sequence (N, )(Y, )(Y, )
dimensional RCAs that can simulate any CTAG
(, ), where (, ) is a delimiter. All the cells
(Morita 2007, 2011).
right to this sequence are set to (Y, ).
Theorem 9 (Morita 2011) There is a 24-state one-
dimensional two-neighbor reversible PCA P24
that can simulate any CTAG on infinite 2-D Universal Reversible Cellular
(leftward-periodic) configurations. Automata That Can Simulate Reversible
Logic Gates
Theorem 10 (Morita 2007) There is a 98-state
one-dimensional three-neighbor reversible PCA A logic gate is called reversible if its logical
P98 that can simulate any CTAG on finite config- function is injective. Fredkin gate (Fredkin and
urations. (Note: it can also manage halting of a Toffoli 1982) is one of the typical reversible logic
CTAG.) gates, which has three input lines and three output
Reversible Cellular Automata 115

t
0 − n − y − − − ∗ − n − n − − − ∗ − y − − − ∗ N − Y − Y − − − Y − Y − Y − Y − Y − Y − Y −
1 − ∗ − n − y − − − ∗ − n − n − − − ∗ − y − − − / Y − Y − − − Y − Y − Y − Y − Y − Y − Y −
2 − − − ∗ − n − y − − − ∗ − n − n − − − ∗ − y − − Y / Y − − − Y − Y − Y − Y − Y − Y − Y −
3 − y − − − ∗ − n − y − − − ∗ − n − n − − − ∗ − y Y − Y / − − Y − Y − Y − Y − Y − Y − Y −
4 − ∗ − y − − − ∗ − n − y − − − ∗ − n − n − − − ∗ Y y Y − + n Y − Y − Y − Y − Y − Y − Y −
5 − − − ∗ − y − − − ∗ − n − y − − − ∗ − n − n − − + / Y y − + Y n Y − Y − Y − Y − Y − Y −
6 − n − − − ∗ − y − − − ∗ − n − y − − − ∗ − n − n − + Y / − y Y + Y n Y − Y − Y − Y − Y −
7 − n − n − − − ∗ − y − − − ∗ − n − y − − − ∗ − n − n Y + + n Y y Y + Y n Y − Y − Y − Y −
8 − ∗ − n − n − − − ∗ − y − − − ∗ − n − y − − − ∗ − n Y n + + Y n Y y Y + Y n Y − Y − Y −
9 − − − ∗ − n − n − − − ∗ − y − − − ∗ − n − y − − − ∗ Y n N ∗ Y + Y n Y y Y + Y n Y − Y −
10 − y − − − ∗ − n − n − − − ∗ − y − − − ∗ − n − y − − + / N n + / Y + Y n Y y Y + Y n Y −
11 − n − y − − − ∗ − n − n − − − ∗ − y − − − ∗ − n − y − + N / N ∗ Y / Y + Y n Y y Y + Y n
12 − ∗ − n − y − − − ∗ − n − n − − − ∗ − y − − − ∗ − n − y N + N / + / Y / Y + Y n Y y Y +
13 − − − ∗ − n − y − − − ∗ − n − n − − − ∗ − y − − − ∗ − n N y N + + y Y / Y / Y + Y n Y y
14 − y − − − ∗ − n − y − − − ∗ − n − n − − − ∗ − y − − − ∗ N n N y + + Y y Y / Y / Y + Y n
15 − ∗ − y − − − ∗ − n − y − − − ∗ − n − n − − − ∗ − y − − − / N n Y ∗ Y + Y y Y / Y / Y +
16 − − − ∗ − y − − − ∗ − n − y − − − ∗ − n − n − − − ∗ − y − − N / Y n + / Y + Y y Y / Y /
17 − n − − − ∗ − y − − − ∗ − n − y − − − ∗ − n − n − − − ∗ − y N − Y / N ∗ Y / Y + Y y Y /
18 − n − n − − − ∗ − y − − − ∗ − n − y − − − ∗ − n − n − − − ∗ N y Y − N / + / Y / Y + Y y
19 − ∗ − n − n − − − ∗ − y − − − ∗ − n − y − − − ∗ − n − n − − − / Y y N − + y Y / Y / Y +
20 − − − ∗ − n − n − − − ∗ − y − − − ∗ − n − y − − − ∗ − n − n − − Y / N y − + Y y Y / Y /
21 − y − − − ∗ − n − n − − − ∗ − y − − − ∗ − n − y − − − ∗ − n − n Y − N / − y Y + Y y Y /
22 − n − y − − − ∗ − n − n − − − ∗ − y − − − ∗ − n − y − − − ∗ − n Y n N − + n Y y Y + Y y
23 − ∗ − n − y − − − ∗ − n − n − − − ∗ − y − − − ∗ − n − y − − − ∗ Y n N n − + Y n Y y Y +

Reversible Cellular Automata, Fig. 9 Simulating the CTAG C 0 by the reversible PCA P24 (Morita 2011)

Reversible Cellular c x=c c c y1 c

Automata, p y = cp + cq cx y2
Fig. 10 Fredkin gate x x
q z = cp + cq
cx y3
(a) (b)

Reversible Cellular Automata, Fig. 11 (a) Switch gate.

lines (Fig. 10). A reversible logic gate is called (b) Inverse switch gate, where c = y1 and x = y2 + y3 under
logically universal, if any reversible sequential the assumption ððy2 ! y1 Þ ^ ðy3 ! y1 ÞÞ
machine (i.e., finite automaton with output ports
We show several two-dimensional RCAs in
as well as input ports) can be realized using only
which switch gate and its inverse are embeddable.
copies of it and delay elements. Since a finite
By this, we obtain very simple computationally
control and tape cells of a reversible Turing
universal 2-D RCAs.
machine are formulated as reversible sequential
machines, we can construct any reversible Turing
machine by them. Therefore, if an RCA can sim- 2-D Universal RCAs on a Square Tessellation
ulate any circuit composed of a universal revers- Here, three models of simple universal RCAs on a
ible gate, we can say it is computationally square tessellation are explained. The first one is a
universal. It is known that Fredkin gate is a uni- two-state RCA with Margolus neighborhood. The
versal reversible gate (Fredkin and Toffoli 1982). second and the third are 16-state reversible PCAs.
In Morita (1990) a construction method of a
reversible sequential machine out of Fredkin Two-State RCA with Margolus Neighborhood
gates is given. It is also known that Fredkin gate Margolus (1984) first showed a two-dimensional
can be composed of switch gate and inverse RCA in which any circuit composed of Fredkin
switch gate (Fig. 11) (Fredkin and Toffoli 1982). gates can be simulated. His model is an RCA with
Hence, the set of switch gate and its inverse is the Margolus neighborhood (Fig. 1) and has the
logically universal. block rules given in Fig. 2. He showed that the
116 Reversible Cellular Automata

Reversible Cellular • •
Automata, Fig. 12 The • •
• • • • • •
local function of the
16-state rotation-symmetric
reversible PCA S1 (Morita • • •
and Ueno 1992) •
• • • • • •
• •
• •

Reversible Cellular cx c
Automata, ↑ ↑
Fig. 13 Switch gate ••
realized in the reversible • •
•• • • •• • •
PCA S1 (Morita and Ueno • • • •
1992). The moving • • •• • • •• ••
direction of a signal is • •
•• •• •• → cx
changed by reflectors. • •
Small circles show virtual • ••• ••
• •
• ••• ••
collision points of signals
• •
• • •• • •
• •
•• •• ••
• •
• • • ••
•
• •
•• ••• •
• •
•• ••
c→ • •
•• ••
• • • •
• • ••• ••• • • ••• ••
• • • • • •
•• •• •• •• •• ••
•

↑
x

billiard ball model (BBM) (Fredkin and Toffoli S1 (Morita and Ueno 1992) is a four-neighbor
1982) of computation is simulated in its cellular rotation and reflection-symmetric reversible
space. The BBM is a kind of physical model of PCA. A cell is divided into four parts, and each
computation where a signal is represented by an part has the state set {0, 1}. Its local transition
ideal ball, and logical operations and signal rules are shown in Fig. 12. Rotated rules are
routing can be performed by their elastic colli- omitted since it is rotation-symmetric. The states
sions and reflections by reflectors. Since switch 0 and 1 are represented by a blank and a dot,
gate and its inverse can be realized in the BBM respectively. The set of these rules has some sim-
(Fredkin and Toffoli 1982), computational univer- ilarity with that of Margolus’ RCA, and in fact, it
sality of the Margolus’ RCA is concluded. can simulate the BBM in a similar manner. In S1, a
signal is represented by two particles. Figure 13
A 16-State Reversible PCA Model S1 gives a configuration of a switch gate module in
If we use the framework of a PCA to simulate S1. The moving direction of a signal is controlled
reversible logic gates, we can obtain a standard by a reflector pattern, and the switch gate opera-
type of an RCA (see Proposition 2), though the tion is realized by two collisions of signals. It is
total number of states of a cell is larger than that of also possible to realize an inverse switch gate.
the RCA of Margolus neighborhood. The model Thus, S1 is computationally universal.
Reversible Cellular Automata 117

A 16-State Reversible PCA Model S2 three parts, each of which has its own state set
The second computationally universal model S2 (Fig. 16). The next state of a cell is determined by
(Morita and Ueno 1992) is also a four-neighbor the present states of the three edge-adjacent parts
reversible PCA having the set of local rules shown of the neighboring cells. The reason we use
in Fig. 14. It is rotation-symmetric but not TPCAs here is that their local functions can be
reflection-symmetric. In S2, reflection of a signal simpler than those of PCAs on a square lattice,
by a reflector is different from that in S1, i.e., only since the number of edge-adjacent cells are only
left turn is possible. Hence, right turn should be three. Hence, it is convenient to study how com-
realized by three left turns. The other features are putational universality emerges from a simple
similar to S1. Figure 15 shows a configuration of a reversible local function.
switch gate. An elementary triangular partitioned cellular
automaton (ETPCA) is a TPCA such that each
part of a cell has the state set {0, 1}, and it is
2-D Universal RCAs on a Triangular rotation-symmetric (i.e., isotropic) (Morita
Tessellation 2016a). A local function of an ETPCA is specified
Next, we give three models of computationally by only four local transition rules. Figure 17
universal reversible triangular partitioned cellu- shows an example of local rules of an ETPCA,
lar automata (TPCAs). In a TPCA, the shape of a by which a local function is completely deter-
cell is an equilateral triangle, and it is divided into mined. Each ETPCA is referred by a four-digit

Reversible Cellular • •
Automata, Fig. 14 The • •
• • • • •
local function of the •
16-state rotation-symmetric
reversible PCA S2 (Morita • • •
and Ueno 1992) •
• • • • • •
• •
• •

Reversible Cellular c
Automata, ↑
Fig. 15 Switch gate •• ••
realized in the reversible • • • •
•• •• •• ••
PCA S2 (Morita and Ueno
1992) ••
• •
•• •• •• •• → cx
• • • •
•• •• •• • •

•• •• ••
• • • • • •
•• •• •• •• •• ••

•• ••
• • • •
•• •• •• ••

c→ • • ••
• •
•• •• • • • ••
•
• • • • • •
• •• •• •• •• •• ••

•• → cx
• •
•• •• •

↑
x
118 Reversible Cellular Automata

number that is obtained by reading the bit patterns signal routing. It is composed of stable blocks,
of the right-hand sides of the four local rules as each of which consists of 12 particles as in
binary numbers. The identification number of the Fig. 20. Crossing of two signals and delay of a
ETPCA shown in Fig. 17 is 0457. There are signal are implemented using auxiliary control
256 ETPCAs in total, and 36 ETPCAs among particles as well as blocks (see Imai and Morita
them are reversible (Morita 2016a). For example, 2000 for details). Figure 20 shows a switch gate
ETPCA 0457 is reversible, since there is no pair of module implemented in the ETPCA 0157 (the
rules that have the same right-hand side. original pattern of a switch gate module in Imai
In the following, we investigate three revers- and Morita 2000 is reduced in its size here). Com-
ible ETPCAs 0157, 0137, and 0347. In spite of the bining two switch gates and two inverse switch
simplicity of their local functions, they are com- gates, a Fredkin gate module is obtained (Fig. 21).
putationally universal, since universal reversible By the above, we can conclude ETPCA 0157 is
logic gates can be realized in their cellular space. computationally universal. It should be noted that
the local rules of ETPCA 0457 (Fig. 17) are the
ETPCA 0157
mirror images of those of ETPCA 0157. Hence,
This model was first studied in Imai and Morita configurations of ETPCA 0157 are directly simu-
(2000). Its local function is shown in Fig. 18. It is lated by their mirror images in ETPCA 0457.
easy to see that the local function is injective. Likewise, the local rules of ETPCAs 0267 and
Thus, it is a reversible ETPCA. In ETPCA 0157, 0237 are obtained from ETPCAs 0157 and 0457,
a signal is represented by a single particle, and respectively, by the complementation (i.e., 0-1
switch gate is realized by one cell as shown in exchange) of each rule. Therefore, ETPCAs
Fig. 19. However, signal routing, crossing, and 0267 and 0237 are also computationally univer-
delay are very complex to realize. Since a single sal. A similar argument can be applied to ETPCAs
particle simply rotates around some point, a 0137 and 0347 below.
“wall,” along which a particle goes, is used for
ETPCA 0137
The second ETPCA is the one whose local func-
tion is shown in Fig. 22 (Morita 2016b). Again, it
is easy to see the local function is injective. Sim-
ilar to the case of ETPCA 0157, a signal is
represented by a single particle, and the switch
gate operation is realized by one cell (Fig. 23). In
this case, a stable block consists of six particles,
from which transmission wire is composed.
Reversible Cellular Automata, Fig. 16 Cellular space Crossing of signals and signal delay is also real-
of a triangular partitioned cellular automaton (TPCA) ized using auxiliary control particles. Figure 24

, • , • • •• , • • ••
•
0 • 4 5 • 7

Reversible Cellular Automata, Fig. 17 The set of four local rules of ETPCA 0457, which defines its local function

, • , • • •• , • • ••
•
0 • 1 5 • 7

Reversible Cellular Automata, Fig. 18 The local function of the reversible ETPCA 0157
Reversible Cellular Automata 119

shows a switch gate module in the ETPCA 0137. ETPCA 0347

We can construct inverse switch gate in a similar The third ETPCA has the local function shown in
manner. Hence, ETPCA 0137 is computationally Fig. 25. This ETPCA was proposed in Morita
universal. (2016a). It is easy to see ETPCA 0347 is revers-
c ible. It should be noted that ETPCAs 0157 and
cx 0137 are conservative in the sense that the number
x 0 c cx of particles is conserved in each local rule. On the
other hand, ETPCA 0347 is nonconservative.
ETPCA 0347 shows quite complex and inter-
Reversible Cellular Automata, Fig. 19 Switch gate esting behavior. In particular, a moving object
operation is realized by one cell of the reversible
ETPCA 0157
called a glider exists (Fig. 26), which can be

• •• •
• •
• • •
• •• • • • •• • • •• • • •• • • •• • 75
58 62 66 70
52 53
• • • • •
51
5 49
c→ • • • • • •
• •• • • •• • • •• • • •• • • •• →c
• •• • 5
50 78
6 7 7 6 8 49 50 •
• • 9 • •
• • • •
• •• • • •• •
10
• •• •
• •
43 15
• • 38
38• ••• ••
• • 19 • •• • 43
33
• • • •
• • • •
34• •• • 24 27 • • •••• ••
28 29 •• • • 47
• • •
→ cx
29 25
26 • • • 78
28
27 25 26 • • ••
• ••• •• • • 52 73 • •• •
• • • •
20 24 • • • •
15 • •• • • •• • • •• • 56 68 •
•• •
• • •
x→ • •
10 •
• • 58
• •• • 5 •• • • •• • • •• • 63 65• •• •
• • • • • •
• • 8 • • • •
• •• • • •• • • •• • • •• • • •• •
• • •
• • •
• •• • • •• • •

• •• •
• •
•
• •• •
• → cx
• •• • • •• • • •• • • •• • • •• • • •• • • •• •
• • • • • • • •
• • • • • • • •
• •• • • •• • • •• • • •• • • •• • • •• • • •• •

Reversible Cellular Automata, Fig. 20 Switch gate module realized in the reversible ETPCA 0157. The cell that
performs the switch gate operation is indicated by bold lines
120 Reversible Cellular Automata

Reversible Cellular Automata, Fig. 21 Fredkin gate module realized in the reversible ETPCA 0157

, • , • • • , • • ••
• •
0 • 1 3 • 7

Reversible Cellular Automata, Fig. 22 The local function of reversible ETPCA 0137

Inverse switch gate is constructed likewise, and

c x cx cx thus ETPCA 0347 is computationally universal.
c
0
Simulating Reversible Counter Machines by
2-D RCAs
Reversible Cellular Automata, Fig. 23 Switch gate
operation is realized by one cell of the reversible Besides reversible logic gates like switch gate and
ETPCA 0137 Fredkin gate, there are also reversible logic ele-
ments with memory that have universality.
used to represent a signal. The moving direction A rotary element (RE) (Morita 2001) is typical
of a glider is controlled by appropriately placing one of such elements. An RE has four input lines
copies of the pattern block, which consists of nine {n, e, s, w} and four output lines {n0, e0, s0, w0} and
particles in this ETPCA. As shown in Fig. 27, if a has two states called state H and state V shown in
glider collides with a sequence of two blocks, it Fig. 30 (hence it has a one-bit memory). All the
first splits into two small objects (t = 56). But they values of inputs and outputs are either 0 or
are finally combined to form a glider again, and it 1. Here, the input (and the output) is restricted as
goes to the south-west direction (t = 334). By this, follows: at most one signal “1” appears as an input
a 120 right turn is realized. Sequences of three (output) at a time. The operation of an RE is
and five blocks also act as right-turn modules. It undefined for the cases that signals 1’s are given
has been shown that left turn, backward turn, and to two or more input lines.
U-turn are possible by the patterns composed of We employ the following intuitive interpreta-
several blocks (Morita 2016a). Furthermore, it is tion for the operation of an RE. Signals 1 and 0 are
known that the phase of a glider is adjusted by interpreted as existence and nonexistence of a
these turn modules. particle. An RE has a “rotatable bar” to control
Colliding two gliders appropriately, switch the moving direction of a particle. When no par-
gate operation is realized as shown in Fig. 28. ticle exists, nothing happens on the RE. If a par-
Figure 29 is a switch gate module in the ETPCA ticle comes from a direction parallel to the
0347, where many turn modules are placed to rotatable bar, then it goes out from the output
control the move directions and phases of gliders. line of the opposite side (i.e., it goes straight
Reversible Cellular Automata 121

•
•
x→ •
• • •5 7• • •19
12, 14
• •• • •• • ••
→ cx
• • • • •
• • • • 24 25
• • • • • •
• • •39
34
26
25 27 26 28 29
27
• •
23 28
• • 44 •
c→ •
• • •5 • • •47 121• ••
126 → cx
• • •
• • 10 • • 52 • • •75 116 • •
70
65
53
•
• • •13 • •• • • 80 • ••
62
54
53 55 54 56 57 113
55 85
• • •
• • 18 • • •57 56 • • 88 108 • •
52
47

•
• • •21 • • 62 • • •91 • • •103 • ••
98
42
93 105

• • • •
• • 26 • • •65 39 • • • • • •
•
• • •29 31 • • • 70
36

• •
• • •
• • • • • • •73 • • •109
104
99

•
• • 78
•
94
• •114 • •• 119
124,126
→c
•
• • •85 • ••
92
• •• • •
80 116
87

• • •
• • • • • •

Reversible Cellular Automata, Fig. 24 Switch gate module in the reversible ETPCA 0137 (Morita 2016b)

, • , • • • , • • ••
• •
0 • 3 4 • 7

Reversible Cellular Automata, Fig. 25 The local function of the reversible ETPCA 0347

ahead) without affecting the direction of the bar It has been shown that any reversible two-
(Fig. 31a). If a particle comes from a direction counter machine can be implemented in a quite
orthogonal to the bar, then it makes a right turn simple way by using REs and some additional
and rotates the bar by 90 (Fig. 31b). It is clear its elements (Morita et al. 2002). Since a reversible
operation is reversible. two-counter machine is known to be universal
122 Reversible Cellular Automata

t =0 t =1 t =2 t =3

• •• • ••
• •• •
• •• •• • • ••
•• •• •

t =4 t =5 t =6

• •• • ••
•• • • ••
• ••
• ••

Reversible Cellular Automata, Fig. 26 Glider in the ETPCA 0347

t =0 t = 56 t = 334

• • •• ••• • ••• • ••
••
•• •• ••• • ••• • ••• •
•• •• •• •• • • •• •• •• •• •• ••
•• • • •• •• •••• • • • •
•• •• •

•
••

Reversible Cellular Automata, Fig. 27 Right turn of a glider by a sequence of two blocks in the ETPCA 0347

t =0 c t = 24 c t = 48 c
•
•• •
••

•
•• •
• • ••
x • •• ••
x • •• ••
x cx
• • •
•• •• •
•• ••

cx c
Reversible Cellular Automata, Fig. 28 Switch gate operation in the ETPCA 0347

(Morita 1996), such a reversible PCA is also uni- In Morita (1996) a CM is defined as a kind of
versal. A counter machine (CM) is a simple com- multi-tape Turing machine whose heads are read-
putation model consisting of a finite number of only ones and whose tapes are all blank except the
counters and a finite-state control (Minsky 1967). leftmost squares as shown in Fig. 32 (P is a blank
Reversible Cellular Automata 123

t = 2232
• ••• • ••• • ••• •
•• ••
• • •• •• •• •• •• •• •
•• • • • •• •• ••• •
• • • •• ••
•• •• •• •
• • •• •
•• •• • •• ••
• • • • • •
•• •• •• •• •• •• ••• • •• •• ••• ••• •
•• ••
•
•• ••
•
•• •• ••• •
• • • • • • •• •• • • • • • • • • •• ••
•• •• •• • •• •• •• •• •
••• • •
•• ••
•• •• • •
• ••

••• • •
•• ••
•
•• •• ••• •
• • • • • • •• ••
• •• ••• • •• • ••• • •• •• • • ••
c •
••
•• ••
• •
•• ••
•• ••
•
•
••
c
• • • • •
•• •• •• •• •• •• ••• • •• •
•• ••
• • • • • • •• •• •• • •
•• •• •• • • • ••
•• ••
• ••• •
•• •• •
• •• ••
• •
••• • •• •
•• •• ••• •
• •• •• • • •• ••
•• •• ••• • ••• • ••• • • •
•• •• •• •
• • •• •• •• •• •• •• • •
•• • • • ••• • •• ••• • •
•• ••
•• •• •• •• • •
••• • • •
•• •• • ••
•• •• • • •
• •• •• •• • ••
• •
••
•
••
cx
• •• ••• • ••• • •
•• •• ••• •
x •
••
•• ••
•
•• ••
• ••• •
• •
••
•• ••
•
• • •
•• •• ••• • •• • ••• • ••• • •
•• •• ••• • •
•• ••
• • •• •• •• •• •• •• • • •• •• • •
•• • • • •• • ••
• • • ••• • • ••• •
•• •• •• •• •• •• •• ••
• • • • • • •• •• • • •• ••
•• •• •• • •• •
• •
•• •• •• ••
• • • • •
•• •• •• •• ••• •
• • • •• ••
•• •• •• •
• •
••
cx
•
•• ••
• •
••
•
•• ••
• •
••
••• • •
•• ••
•• •• • •
• ••
••• • •
•• ••
•• •• • •
• ••
••• • ••• • •
•• ••
• • •• •• • •
••• • •• • ••• • • ••
•• •• •• ••
• •
• ••• • •
•• •• •• ••
• • •• •• • •
•• • ••

Reversible Cellular Automata, Fig. 29 Switch gate module in the ETPCA 0347 (Morita 2016a). Switch gate operation
(Fig. 28) is performed in the middle of this pattern

n n n n t t +1
w w n n n n
e e
w e w e w e w e
w e
⇒ w e
s s s s
State V State H s s
(a)
s s

Reversible Cellular Automata, Fig. 30 Two states of a n n n n

rotary element (RE)
w e w e
symbol). This definition is convenient for giving w e
⇒ w e
the notion of reversibility on a CM.
It is known that a CM with two counters is s s s s
(b)
computationally universal (Minsky 1967). This
result also holds even if the reversibility constraint Reversible Cellular Automata, Fig. 31 Operations of
is added as shown in the next theorem. an RE: (a) the parallel case and (b) the orthogonal case
124 Reversible Cellular Automata

Element name Pattern

q

LR-turn element ◦◦ ◦◦
2 ◦◦ ◦◦
Z P P P P P P P · · ·

◦◦ ◦◦ ◦◦ ◦◦
0
Z P P P P P P P · · · ◦◦ ◦◦• • ◦◦ ◦◦
• •
R-turn element •• ••• •
• •
• •
◦◦ ◦◦ • •◦◦ ◦◦
Reversible Cellular Automata, Fig. 32 A counter
machine with two counters ◦◦ ◦◦ ◦◦ ◦◦

◦ •
◦ ◦ ◦ ◦ ◦
• ◦◦ ◦◦ ◦◦ ◦◦
◦◦ ◦◦ • •◦◦ ◦◦
Reflector •
•
• • ◦◦ ◦◦• • ◦◦ ◦◦
◦◦ ◦◦ ◦◦ ◦◦
◦ • ◦
• ◦
◦

• • ◦◦ ◦• •◦ ◦◦
• ◦• ••◦ ◦ ◦ ◦•• •◦
• • ◦ ◦ • ◦ ◦
◦ ◦ ◦ ◦
◦ Rotary element ◦ ◦ ◦
◦ ◦
• ◦ ◦• ••◦ ◦ ◦ ◦•• •◦
• • ◦◦ ◦• •◦ ◦◦
• • • • • ◦ •
• •
•

◦ • ◦
Position marker
◦ • ◦ ◦ • ◦
•
• ◦

◦ w Reversible Cellular Automata, Fig. 34 Basic elements

w
• ◦ • z x z x realized in the reversible cellular space of P3 (Morita et al.
• y
• y 2002)

◦
◦
• ◦ • function shown in Fig. 33 (Morita et al. 2002).
◦
• In P3, five kinds of signal processing elements
shown in Fig. 34 can be realized. Here, a single
Reversible Cellular Automata, Fig. 33 The local func- ● acts as a signal.
tion of the 81-state rotation-symmetric reversible PCA P3 An LR-turn element, an R-turn element, and a
(Morita et al. 2002). The last rule scheme represents
33 rules not specified by the others, where w, x, y, reflector in Fig. 34 are used for signal routing.
z  {blank, ○, ●} Figure 35 shows the operations of an RE in the
P3 space. A position marker is used to keep a head
Theorem 11 (Morita 1996) For any Turing position of a CM, and realized by a single ○,
machine T, there is a deterministic reversible which rotates clockwise at a certain position by
CM with two counters M that simulates T. the first rule in Fig. 33. Figure 36 shows the
pushing and pulling operations of a position
An 81-State Reversible PCA Model P3 marker. Figure 37 shows an example of a whole
Any reversible CM with two counters is configuration for a reversible CM with two coun-
embeddable in the model P3 with the local ters embedded in the P3 space. In this model, no
Reversible Cellular Automata 125

t =0 t =1 t =2 t =3 t =4 t =5
• ◦◦
◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• • ◦
◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦•◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦◦ ◦◦◦ ◦ ◦ ◦ ◦ ◦◦ ◦◦◦ ◦ ◦•◦ ◦ ◦◦ ◦◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ • ◦ ◦ ◦ ◦ ◦ ◦
◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦
◦◦ ◦• •◦ ◦◦ ◦◦ ◦• • •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦
•

(a)
t =0 t =1 t =2 t =3 t =4 t =5
◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦
◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ • ◦ ◦◦ ◦◦ ◦ ◦ ◦ ◦ •
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦• ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ • ◦ ◦ • ◦ ◦ ◦ ◦
◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ • ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦ ◦• ••◦ ◦ ◦ ◦•• •◦
◦◦ ◦• •◦ ◦◦ ◦◦ ◦• • •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦ ◦◦ ◦• •◦ ◦◦
•

(b)
Reversible Cellular Automata, Fig. 35 Operations of RE in P3: (a) the parallel case and (b) the orthogonal case

t =0 t =1 t =2 t =3 t =4 How Simple Can Universal RCAs Be?

◦ We have seen that there are many kinds of simple
◦ ◦• ◦ RCAs having computational universality. These
◦ • RCAs with least number of states known so far are
•
• summed up as follows:
•
(a)
One-dimensional case:
t =0 t =1 t =2 t =3 t =4 Finite configuration:
◦ 98-state reversible PCA (Morita 2007)
◦ •◦ ◦ Infinite configuration:
• ◦ 24-state reversible PCA (Morita 2011)
•
• Two-dimensional case:
•
(b) Finite configuration:
81-state reversible PCA (Morita et al. 2002.
Reversible Cellular Automata, Fig. 36 Pushing and Infinite configuration:
pulling operations to a position marker in P3 Two-state RCA with Margolus neighbor-
hood (Margolus 1984)
conventional logic elements like AND, OR, and
Eight-state reversible triangular PCAs (Imai
NOT are used. Computation is simply carried out
and Morita 2000; Morita 2016a, b)
by a single signal that interacts with REs and
16-state reversible square PCAs (Morita
position markers.
and Ueno 1992)

Future Directions We think the number of states for universal

RCA can be reduced much more for each case of
In this section, we discuss future directions and the above. Although the framework of PCA is
open problems as well as topics not dealt with in useful for designing an RCA of a standard type,
the previous sections. the number of states becomes relatively large
126 Reversible Cellular Automata

Begin
→

End
←

Reversible Cellular Automata, Fig. 37 An example of a reversible counter machine, which computes the function
2 x + 2, embedded in P3 (Morita et al. 2002)

because the state set is the direct product of the Furthermore, if we want to implement a CA in a
sets of the states of the parts. Hence, we shall need real physical system, the following problem arises.
some other technique to find a universal RCA with In a CA, both time and space are discrete, and all
a small number of states. the cells operate synchronously. On the other hand,
in a real system, time and space are continuous, and
How Can We Realize RCAs in Reversible Physical no synchronizing clock is assumed beforehand.
Systems? Hence, we need some novel theoretical framework
This is a very difficult problem. At present, there for dealing with such problems.
is no good solution. The billiard ball model
(Fredkin and Toffoli 1982) is an interesting idea, Self-reproduction in RCAs
but it is practically impossible to implement it von Neumann first invented a self-reproducing
perfectly. Instead of using a mechanical collision cellular automata by using his famous 29-state
of balls, at least some quantum mechanical revers- CA (von Neumann 1966). In his model, the size
ible phenomena should be used. of a self-reproducing pattern is quite huge,
Reversible Cellular Automata 127

Reversible Cellular
Automata, Fig. 38 Self-
reproduction of a pattern in
a 3-D RCA

because the pattern has both computing and self- Bibliography

reproducing abilities. Later, Langton (1984) cre-
ated a very simple self-reproducing CA by remov- Primary Literature
ing the condition that the pattern need not have Amoroso S, Cooper G (1970) The Garden of Eden theorem
for finite configurations. Proc Am Math Soc 26:158–164
computational universality.
Amoroso S, Patt Y-N (1972) Decision procedures for
It was shown that self-reproduction of the surjectivity and injectivity of parallel maps for tessel-
Langton’s type is possible in two- or three- lation structures. J Comput Syst Sci 6:448–464
dimensional reversible PCA (Imai et al. 2002; Bennett C-H (1973) Logical reversibility of computation.
IBM J Res Dev 17:525–532
Morita and Imai 1996). Figure 38 shows a self-
Bennett C-H (1982) The thermodynamics of computation –
reproducing pattern in a three-dimensional revers- a review. Int J Theor Phys 21:905–940
ible PCA (Imai et al. 2002). But it is left for the Bennett C-H, Landauer R (1985) The fundamental physi-
future study to design a simple and elegant RCA cal limits of computation. Sci Am 253:38–46
Boykett T (2004) Efficient exhaustive listings of reversible
in which objects with computational universality
one dimensional cellular automata. Theor Comput Sci
can reproduce themselves. 325:215–247
Cocke J, Minsky M (1964) Universality of tag systems
with P = 2. J ACM 11:15–20
Firing Squad Synchronization in RCAs Cook M (2004) Universality in elementary cellular autom-
ata. Complex Syst 15:1–40
It is also possible to solve the firing squad syn-
Fredkin E, Toffoli T (1982) Conservative logic. Int J Theor
chronization problem using RCAs. Imai and Phys 21:219–253
Morita (1996) gave a 99-state reversible PCA Gruska J (1999) Quantum computing. McGraw-Hill, London
that synchronize an array of n cells in three Hedlund G-A (1969) Endomorphisms and automorphisms of
the shift dynamical system. Math Syst Theory 3:320–375
n time steps. Though it seems possible to give an
Imai K, Morita K (1996) Firing squad synchronization
optimal time solution, i.e., a (2 n  2)-step solu- problem in reversible cellular automata. Theor Comput
tion, its concrete design has not yet done. Sci 165:475–482
128 Reversible Cellular Automata

Imai K, Morita K (2000) A computation-universal Morita K (2017a) Two small universal reversible Turing
two-dimensional 8-state triangular reversible cellular machines. In: Adamatzky A (ed) Advances in unconven-
automaton. Theor Comput Sci 231:181–191 tional computing. Vol. 1: Theory. Springer, Cham,
Imai K, Hori T, Morita K (2002) Self-reproduction in three- pp 221–237
dimensional reversible cellular space. Artif Life Morita K, Harao M (1989) Computation universality of
8:155–174 one-dimensional reversible (injective) cellular autom-
Kari J (1994) Reversibility and surjectivity problems of ata. Trans IEICE Jpn E72:758–762
cellular automata. J Comput Syst Sci 48:149–182 Morita K, Imai K (1996) Self-reproduction in a reversible
Kari J (1996) Representation of reversible cellular automata cellular space. Theor Comput Sci 168:337–366
with block permutations. Math Syst Theory 29:47–61 Morita K, Ueno S (1992) Computation-universal models
Landauer R (1961) Irreversibility and heat generation in of two-dimensional 16-state reversible cellular autom-
the computing process. IBM J Res Dev 5:183–191 ata. IEICE Trans Inf Syst E75-D:141–147
Langton C-G (1984) Self-reproduction in cellular autom- Morita K, Shirasaki A, Gono Y (1989) A 1-tape 2-symbol
ata. Phys D 10:135–144 reversible Turing machine. Trans IEICE Jpn
Margolus N (1984) Physics-like model of computation. E72:223–228
Phys D 10:81–95 Morita K, Tojima Y, Imai K, Ogiro T (2002) Universal
Maruoka A, Kimura M (1976) Condition for injectivity of computing in reversible and number-conserving
global maps for tessellation automata. Inf Control two-dimensional cellular spaces. In: Adamatzky
32:158–162 A (ed) Collision-based computing. Springer, London,
Maruoka A, Kimura M (1979) Injectivity and surjectivity pp 161–199
of parallel maps for cellular automata. J Comput Syst Myhill J (1963) The converse of Moore’s Garden-of-Eden
Sci 18:47–64 theorem. Proc Am Math Soc 14:658–686
Minsky M-L (1967) Computation: finite and infinite von Neumann J (1966) In: Burks AW (ed) Theory of self-
machines. Prentice-Hall, Englewood Cliffs reproducing automata. Urbana, The University of Illi-
Moore E-F (1962) Machine models of self-reproduction. nois Press
Proc Symp Appl Math Am Math Soc 14:17–33 Richardson D (1972) Tessellations with local transforma-
Mora JCST, Vergara SVC, Martinez GJ, McIntosh HV tions. J Comput Syst Sci 6:373–388
(2005) Procedures for calculating reversible one- Sutner K (2004) The complexity of reversible cellular
dimensional cellular automata. Phys D 202:134–141 automata. Theor Comput Sci 325:317–328
Morita K (1990) A simple construction method of a revers- Toffoli T (1977) Computation and construction universal-
ible finite automaton out of Fredkin gates, and its ity of reversible cellular automata. J Comput Syst Sci
related problem. Trans IEICE Jpn E73:978–984 15:213–231
Morita K (1995) Reversible simulation of one-dimensional Toffoli T (1980) Reversible computing. In: de Bakker JW,
irreversible cellular automata. Theor Comput Sci van Leeuwen J (eds) Automata, languages and pro-
148:157–163 gramming. LNCS 85, Springer, Berlin, Heiderberg,
Morita K (1996) Universality of a reversible two-counter pp 632–644
machine. Theor Comput Sci 168:303–320 Toffoli T, Margolus N (1990) Invertible cellular automata:
Morita K (2001) A simple reversible logic element and cellu- a review. Phys D 45:229–253
lar automata for reversible computing. In: Margenstern M, Toffoli T, Capobianco S, Mentrasti P (2004) How to turn a
Rogozhin Y (eds) Proceedings of the MCU 2001. LNCS second-order cellular automaton into a lattice gas: a
2055, Springer, Berlin, Heidelberg, pp 102–113 new inversion scheme. Theor Comput Sci
Morita K (2007) Simple universal one-dimensional revers- 325:329–344
ible cellular automata. J Cell Autom 2:159–166 Watrous J (1995) On one-dimensional quantum cellular
Morita K (2008) Reversible computing and cellular automata. In: Proceedings of the FOCS, IEEE Com-
automata – a survey. Theor Comput Sci 395:101–131 puter Society Press, pp 528–537
Morita K (2011) Simulating reversible Turing machines
and cyclic tag systems by one-dimensional reversible
cellular automata. Theor Comput Sci 412:3856–3865 Books and Reviews
Morita K (2016a) An 8-state simple reversible triangular Adamatzky A (ed) (2002) Collision-based computing.
cellular automaton that exhibits complex behavior. In: Springer, London
Cook M, Neary T (eds) AUTOMATA 2016. LNCS Bennett CH (1988) Notes on the history of reversible
9664, Springer, Cham, pp 170–184. Slides with movies computation. IBM J Res Dev 32:16–23
of computer simulation: Hiroshima University Institu- Burks A (ed) (1970) Essays on cellular automata. Univer-
tional Repository. http://ir.lib.hiroshima-u.ac.jp/00039321 sity of Illinois Press, Urbana
Morita K (2016b) Universality of 8-state reversible and Kari J (2005) Theory of cellular automata: a survey. Theor
conservative triangular partitioned cellular automaton. Comput Sci 334:3–33
In: El Yacoubi S et al (eds) ACRI 2016. LNCS 9863, Morita K (2017b) Theory of reversible computing.
Springer, Cham, pp 45–54. Slides with movies of com- Springer, Tokyo
puter simulation: Hiroshima University Institutional Wolfram S (2001) A new kind of science. Wolfram Media,
Repository. http://ir.lib.hiroshima-u.ac.jp/00039997 Champaign
 local neighborhood mapping, or equivalently
Additive Cellular Automata as a global update mapping.
Cellular automata Cellular automata are
Burton Voorhees dynamical systems that are discrete in space,
Center for Science, Athabasca University, time, and value. A state of a cellular automaton
Athabasca, Canada is a spatial array of discrete cells, each
containing a value chosen from a finite alpha-
bet. The state space for a cellular automaton is
Article Outline the set of all such configurations.
Cyclic states A cyclic state of a cellular autom-
Glossary aton is a state lying on a cycle of the automaton
Definition of the Subject update rule, hence it is periodically revisited in
Introduction the evolution of the rule.
Notation and Formal Definitions Garden-of-Eden A Garden-of-Eden state is a
Additive Cellular Automata in One Dimension state that has no predecessor. It can be present
d-Dimensional Rules only as an initial condition.
Future Directions Injectivity A mapping is injective (one-to-one)
Bibliography if every state in its domain maps to a unique
state in its range. That is, if states x and y both
Glossary map to a state z then x = y.
Linear cellular automata A linear cellular
Additive cellular automata An additive cellular automaton is a cellular automaton whose update
automaton is a cellular automaton whose rule satisfies the condition that its action on the
update rule satisfies the condition that its action sum of two states separately equals action on the
on the sum of two states is equal to the sum of sum of the two states plus its action on the state
its actions on the two states separately. in which all cells contain the quiescent symbol.
Alphabet of a cellular automaton The alphabet Note that some researchers reverse the defini-
of a cellular automaton is the set of symbols or tions of additivity and linearity.
values that can appear in each cell. The alphabet Local maps of a cellular automaton The local
contains a distinguished symbol called the null or mapping for a cellular automaton is a map from
quiescent symbol, usually indicated by 0, which the set of all neighborhoods of a cell to the
satisfies the condition of an additive identity: automaton alphabet.
0 + x = x. Neighborhood The neighborhood of a given cell
Basins of attraction The basins of attraction of a is the set of cells that contribute to the update of
cellular automaton are the equivalences classes value in that cell under the specified update rule.
of cyclic states together with their associated Predecessor state A state x is the predecessor
transient states, with two states being equivalent of a state y if and only if x maps to
if they lie on the same cycle of the update rule. y under application of the cellular automaton
Cellular automata rule The rule, or update rule update rule. More specifically, a state x is
of a cellular automaton describes how any an nth order predecessor of a state y if it maps
given state is transformed into its successor to y under n applications of the update rule.
state. The update rule of a cellular automaton Reversibility A mapping X is reversible if and
is described by a rule table, which defines a only if a second mapping X 1 exists such that

# Springer-Verlag 2009 129

A. Adamatzky (ed.), Cellular Automata,
https://doi.org/10.1007/978-1-4939-8700-9_4
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_4
130 Additive Cellular Automata

if X ðxÞ ¼ y then X 1 ðyÞ ¼ x: For finite state Introduction

spaces reversibility and injectivity are
identical. A wide variety of cellular automata applications,
Rule table The rule table of a cellular automaton in a number of differing disciplines, has appeared
is a listing of all neighborhoods together with in the past 50 years, see, e.g. (Codd 1968; Duff
the symbol that each neighborhood maps to and Preston 1984; Sarkar 2000; Chopard and
under the local update rule. Droz 1998). Among other things, cellular autom-
State transition diagram The state transition ata have been used to model growth and aggrega-
diagram (STD) of a cellular automaton is a tion processes (Lindenmayer and Rozenberg
directed graph with each vertex labeled by a 1976; Mackay 1976; Langer 1980; Lin and
possible state and an edge directed from a Goldenfeld 1990); discrete reaction-diffusion sys-
vertex x to a vertex y if and only if the state tems (Greenberg and Hastings 1978; Greenberg
labeling vertex x maps to the state labeling et al. 1978; Madore and Freedman 1983;
vertex y under application of the automaton Adamatzky et al. 2005; Oono and Kohmoto
update rule. 1985); spin exchange systems (Falk 1986; Can-
Surjectivity A mapping is surjective (or onto) if ning and Droz 1991); biological pattern formation
every state has a predecessor. (Vitanni 1973; Young 1984); disease processes
Transient states A transient state of a cellular and transmission (Dutching and Vogelsaenger
automaton is a state that can at most appear 1985; Moreira and Deutsch 2002; Sieburg et al.
only once in the evolution of the 1991; Santos and Continho 2001; Beauchemin
automaton rule. et al. 2005); DNA sequences, and gene interac-
tions (Burks and Farmer 1984; Moore and Hahn
2002); spiral galaxies (Gerola and Seiden 1978);
Definition of the Subject social interaction networks (Flache and
Hegselmann 1998); and forest fires (Chen et al.
Cellular automata are discrete dynamical systems 1990; Drossel and Schwabl 1992). They have
in which an extended array of symbols from a finite been used for language and pattern recognition
alphabet is iteratively updated according to a spec- (Smith 1972; Sommerhalder and van Westrhenen
ified local rule. Originally developed by John von 1983; Ibarra et al. 1985; Morita and Ueno 1994;
Neumann (von Neumann 1963; von Neumann and Jen 1986a; Raghavan 1993; Chattopadhyay et al.
Burk 1966) in 1948, following suggestions from 2000); image processing (Rosenfeld 1979;
Stanislaw Ulam, for the purpose of showing that Sternberg 1980); as parallel computers (Hopcroft
self-replicating automata could be constructed. and Ullman 1972; Cole 1969; Benjamin and
Von Neumann’s construction followed a compli- Johnson 1997; Carter 1984; Hillis 1984; Manning
cated set of reproduction rules but later work 1977); parallel multipliers (Atrubin 1965); sorters
showed that self-reproducing automata could be (Nishio 1981); and prime number sieves (Fischer
constructed with only simple update rules, e.- 1965).
g. (Arbib 1966). More generally, cellular automata In recent years, cellular automata have become
are of interest because they show that highly com- important for VLSI logic circuit design (Pries
plex patterns can arise from the application of very et al. 1986). Circuit designers need “simple, reg-
simple update rules. While conceptually simple, ular, modular, and cascadable logic circuit struc-
they provide a robust modeling class for applica- ture to realize a complex function” and cellular
tion in a variety of disciplines, e.g. (Toffoli and automata, which show a significant advantage
Margolis 1987), as well as fertile grounds for the- over linear feedback shift registers, the traditional
oretical research. Additive cellular automata are the circuit building block, satisfy this need (see
simplest class of cellular automata. They have been Chaudhuri et al. 1997 for an extensive survey).
extensively studied from both theoretical and prac- Cellular automata, in particular additive cellular
tical perspectives. automata, are of value for producing high-quality
Additive Cellular Automata 131

pseudorandom sequences (Bardell and McAnney Notation and Formal Definitions

1986; Hortensius et al. 1989; Tsalides et al. 1991;
Matsumoto 1998; Tomassini et al. 2000); for Let S ðLÞ ¼ fsi g be the set of lattice sites of a
pseudoexhaustive and deterministic pattern gen- d-dimensional lattice L with nr equal to the num-
eration (Das and Chaudhuri 1989, 1993; Serra ber of lattice sites on dimension r. Denote by A a
1990; Tziones et al. 1994; Mrugalski et al. 2000; finite symbols set with jA j ¼ p (usually prime).
Sikdar et al. 2002); for signature analysis An A -configuration on L is a surjective map v
(Hortensius et al. 1990; Serra et al. 1990; : A 7! S ðLÞ that assigns a symbol from A to each
Dasgupta et al. 2001); error correcting codes site in S ðLÞ. In this way, every A-configuration
(Chowdhury et al. 1994, 1995a); pseudo- defines a size n1      nd, d-dimensional matrix
associative memory (Chowdhury et al. 1995b); m of symbols drawn from A. Denote the set of all
and cryptography (Nandi et al. 1994). A-configurations on L by E ðA,LÞ.
In this discussion, attention focuses on the Each si  S ðLÞ is labeled by an integer vector
!
subclass of additive cellular automata. These are i ¼ ði1 , . . . , id Þ where ir is the number of
the simplest cellular automata, characterized by sites along the rth dimension separating s i from
the property that the action of the update rule on the assigned origin in L . The shift operator on
the sum of two states is equal to the sum of the rule the rth dimension of L is the map sr : L 7! L
acting on each state separately. Hybrid additive defined by
rules (i.e., with different cells evolving according
to different additive rules) have proved particu- !
larly useful for generation of pseudorandom and sr ðsi Þ ¼ sj , j ¼ ði1 , . . . , ir  1, . . . , id Þ: (1)
pseudoexhaustive sequences, signature analysis,
!
and other circuit design applications, e.- Equivalently, the shift maps the value at site i
!
g. (Chaudhuri et al. 1997; Cattell and Muzio to the value at site j .
1996; Cattell et al. 1999). Let mðsi ;t Þ ¼ mði1 , . . . , id ;t Þ  A be the entry of
The remainder of this article is organized as m corresponding to site si at iteration t for any
follows: section “Notation and Formal Definitions” discrete dynamical system having E ðA,LÞ as
introduces definitions and notational conventions. state space. Given a finite set of integer d-tuples
In section “Additive Cellular Automata in One N ¼ fðk 1 , . . . , k d Þg define the N -neighborhood
Dimension”, consideration is restricted to one- of a site si  S ðLÞ as
dimensional rules. The influence of boundary con-
n
! ! ! ! o
ditions on the evolution of one-dimensional rules, N ðsi Þ ¼ sj
j ¼ i þ k , k  N (2)
conditions for rule additivity, generation of fractal
space-time outputs, equivalent forms of rule repre-
sentation, injectivity and reversibility, transient A neighborhood configuration is a surjective
lengths, and cycle periods are discussed using sev- map v : A 7! N ðs0 Þ. Denote the set of all neigh-
eral approaches. TakingX as the global operator for borhood configurations by E N ðA Þ.
an additive cellular automata, a method for analytic The rule table for a cellular automata
solution of equations of the form X ðmÞ ¼ b is acting on the state space E ðA,LÞ with standard
described. Section “d-Dimensional Rules” neighborhood N ðs0 Þ is defined by a map x : E N
describes work on d-dimensional rules defined on ðA Þ 7! A (note that this map need not be
tori. The discrete baker transformation is defined surjective or injective). The value of x for
and used to generalize one-dimensional results on a given neighborhood configuration is called
transient lengths, cycle periods, and similarity of the (value of the) rule component of that config-
state transition diagrams. Extensive references to uration. The map x : E N ðA Þ 7! A induces
the literature are provided throughout and a set of a global map X : E ðA,LÞ 7! E ðA,LÞ as
general references is provided at the end of the follows: For any given 
mðt Þ E ðA,LÞ,
 element
bibliography. the set Cðs i Þ ¼ m sj ;t
sj  N ðsi Þ is a
132 Additive Cellular Automata

neighborhood configuration for the site si , hence Example 1 (Continuous Transformations of

the map mðsi ;t Þ 7! xðCðsi ÞÞ for all si produces a the Shift Dynamical System)
new symbol m(si; t + 1). The site si is called the Let L be isomorphic to the set of integers Z. Then
mapping site. When taken over all mapping sites, E ðA,Z Þ is the set of infinite sequences with entries
this produces a matrix m( t + 1) that is the repre- from A. With the product topology induced by the
sentation of X ðmðt ÞÞ: A cellular automaton discrete topology on A and s as the left shift map,
is indicated by reference to its rule table or to the system ðE ðA,Z Þ,sÞ is the shift dynamical
the global map defined by this rule table. system on A. The set of cellular automata maps X
A cellular automaton with global map X : E ðA,Z Þ 7! E ðA,Z Þ constitutes the class of con-
is additive if and only if, for all pairs of states m tinuous shift-commuting transformations of
and b, ðE ðA,Z Þ,sÞ , a fundamental result of Hedlund
(1969).
X ðm þ bÞ ¼ X ðmÞ þ X ðbÞ (3)
Example 2 (Elementary Cellular Automata)
Let L be isomorphic to L with A ¼ f0,1g and N
Addition of states is carried out site-wise ¼ f1,0,1g . The neighborhood of site si is
mod(p) on the matrix representations of m and b; fsi1 , si , siþ1 g andE N ðA Þ ¼ f000,001,010 ,011,
for example, for a one-dimensional six-site lattice 100,101,110,111g: In this one-dimensional case,
with p = 3 the sum of 120112 and 021212 is the rule table can be written as xi = x(i0i1i2) where
111021. i0i1i2 is the binary form of the index i. Listing this
The definition for additivity given in gives the standard form for the rule table of an
Chaudhuri et al. (1997) differs slightly from elementary cellular automata.
this standard definition. There, a binary valued
cellular automaton is called “linear” if its
local rule only involves the XOR operation 000 001 010 011 100 101 110 111
and “additive” if it involves XOR x0 x1 x2 x3 x4 x5 x6 x7
and/or XNOR. A rule involving XNOR can be
written as the binary complement of a The standard labeling scheme for elementary
rule involving only XOR. In terms of the cellular automata was introduced by Wolfram
global operator of the rule, this means that it (1983), who observed that the rule table for elemen-
P
has the form 1 þ X where X satisfies Eq. (1) tary rules defines the binary number 71¼0 xi 2i and
and 1 represents the rule that maps every site to 1. used this number to label the corresponding rule.
Thus, ð1 þ X Þðm þ bÞ equals 1 . . . 1 þ X ðm þ bÞ
while Example 3 (The Game of Life)
This simple 2-dimensional cellular automata was
ð1 þ X ÞðmÞ þ ð1 þ X ÞðbÞ invented by John Conway to illustrate a self-
¼ 1 . . . 1 þ 1 . . . 1 þ X ðmÞ þ X ðbÞ reproducing system. It was first presented in 1970
¼ X ðmÞ þ X ðbÞ modð2Þ: by Martin Gardner (1970, 1971). The game takes
place on a square lattice, either infinite or toridal.
In what follows, an additive rule is defined The neighborhood of a cell consists of the eight
strictly as one obeying Eq. (3), corresponding cells surrounding it. The alphabet is {0, 1}: a 1 in a
to rules that are “linear” in Chaudhuri cell indicates that cell is alive, a 0 indicates that it is
et al. (1997). dead. The update rules are: (a) If a cell contains a 0,
Much of the formal study of cellular automata it remains 0 unless exactly three of its neighbors
has focused on the properties and forms of repre- contain a 1; (b) If a cell contains a 1 then it remains
sentation of the map X : E ðA,LÞ 7! E ðA,LÞ . a 1 if and only if two or three of its neighbors are 1.
The structure of the state transition diagram This cellular automata produces a number of inter-
ðSTDðX ÞÞ of this map is of particular interest. esting patterns including a variety of fixed points
Additive Cellular Automata 133

(still life); oscillators (period 2); and moving pat- where all indices are taken mod(n) in the case of
terns (gliders, spaceships); as well as more exotic Z n . In the remaining cases,
patterns such as glider guns which generate a
8
stream of glider patterns. >
< m1 i¼0
½dðmÞi ¼ mi1 þ miþ1 0<i<n1
>
:
mn2 i¼n1
Additive Cellular Automata in One null conditions (4)
Dimension 
m1 i¼0
½dðmÞi ¼
mi1 þ miþ1 0<i
Much of the work on cellular automata has
focused on rules in one dimension (d = 1). This half  infinite conditions
section reviews some of this work.

Note that L and Z n are representations of the

Boundary Conditions and Additivity
intervals [1, 1] and [0, 1] respectively. Cellu-
In the case of one-dimensional cellular automata,
lar automata rules are not quite functions on
the lattice L can be isomorphic to the integers; to
these intervals, however, since they are generally
the non-negative integers; to the finite set
double valued on rational points having distinct
f0, . . . , n  1g  Z; or to the integers modulo an
representations as binary strings (Voorhees
integer n. In the first case, there are no boundary
1996). For example, both 01 and 10 in Z þ ,
conditions; in the remaining three cases, different
where underlining indicates infinite repetition,
boundary conditions apply. If L is isomorphic to
are numerically 1/2 but dð01Þ ¼ 110 ¼ 3=4
Z n , the integers mod(n), the boundary conditions
while dð10Þ ¼ 010 ¼ 1=4.
are periodic and the lattice is circular (it is a p-adic
The state space E ðf0,1g,LÞ for binary valued
necklace). This is called a cylindrical cellular
one-dimensional cellular automata is just the set of
automata (Jen 1988a) because evolution of the
binary sequences over the specified one-
rule can be represented as taking place on a cyl-
dimensional lattice. For the cases of Z and Z n all
inder. If the lattice is isomorphic to
such rules commute with the shift operator s.
{0, . . ., n  1}, null, or Dirchlet boundary condi-
When null boundary conditions are involved, how-
tions are set (Tadaki and Matsufuji 1993; Nandi
ever, commutativity fails at the boundary sites. For
and Pal Chaudhuri 1996; Chin et al. 2001). That
example, let X be the global operator for an ele-
is, the symbol assigned to all sites in L outside of
mentary cellular automata operating on strings
this set is the null symbol. When the lattice is
m = m0. . .mn1 of length n with null boundary
isomorphic to the non-negative integers Z þ , null
conditions. Noting that 1 = 1 mod (2), the com-
boundary conditions are set at the left boundary. In
mutator ½X ,s has components
these latter two cases, the neighborhood structure
assumed may influence the need for null conditions.
½½X ,sðmÞi ¼ ½X sðmÞþsX ðmÞi
8
Example 4 (Elementary Rule 90) < X ð0m1 m2 ÞþX ðm0 m1 m2 Þmodð2Þ i ¼ 0
Let d represent the global map for the elementary ¼ 0 0 < i < n1
:
cellular automata rule 90, with rule table X ðmn1 00Þ i ¼ n1
(5)
000 001 010 011 100 101 110 111
0 1 0 1 1 0 1 0 The relation ½X ,s ¼ 0 can be preserved for rules
defined on Z þ by altering the neighborhood struc-
For a binary string m in Z or Z n the action of ture. The case N ¼ f1,0,1g is called nearest neigh-
rule 90 is defined by [d(m)]i = mi  1 + mi + 1 mod (2), bor because the value introduced at site i at the next
134 Additive Cellular Automata

iteration of a rule is determined by the values at site Additive Cellular Automata and Fractals
i and its immediately neighboring sites i  1 and There is a direct connection between the space-
i + 1. Taking N ¼ f0,1,2g yields left justified neigh- time output patterns of additive cellular automata
borhoods. This eliminates the need for null boundary and self-similar fractal patterns (Willson 1984a, b,
conditions at the left boundary in Z þ . 1987a, b, 1992; Peitgen and Richter 1986; Culik
Changes in the neighborhood structure of and Dube 1989). The simplest examples are ele-
this sort are equivalent to changes in the mapping mentary rules 102 and 90. When acting on a
site. It is important to recognize that such changes doubly infinite sequence with the initial state 01
can significantly alter the topological structure of the 0 iteration of these rules yields the space-time
state transition diagram (STD) for a rule. If X output indicated in Fig. 1. In the case of rule
represents the global map for a rule with nearest 60, this output is the mod(2) Pascal triangle
neighbor neighborhoods, then the global map for while for rule 90 it consists of every other row of
the same rule with left justified neighborhoods is s this triangle (Voorhees 1988).
X and the presence of the shift operator can change The pattern generated by rule 60 (or, equiva-
cycle periods. For example, the maximum lently, by rule 102) rescales to yield the fractal
cycle period for nearest neighbor rule 90 acting on known as the Sirpinski gasket (von Haeseler et al.
Z 6 is 2 while the maximum cycle period for this 1992a; Allouche et al. 1996). Direct connections
same rule with left justified neighborhoods is 3. between cellular automata outputs and the fractal
generation schemes of matrix substitution systems
100010 7! 010100 7! 100010 and hierarchical iterated function systems are shown
nearest neighbor case in (von Haeseler et al. 1992b, 1993, 1995, 2001a, b;
100010 7! 101000 7! 001010 7! 100010 Barbé et al. 1995, 2003; Nagler and Claussen 2005).
left justified case In Takahashi (1990, 1992) the dimension spectrum
associated to the space-time output of additive cel-
The additivity condition of Eq. (3) has an lular automata is shown to be equal to the singularity
expression in terms of rule table components. spectrum of an associated multifractal.
For A ¼ f0,1g this is given by:
Forms of Representation
Theorem 1 (Voorhees 1996) Unless otherwise noted, the lattice in this section
A k-site rule X : E ðf0,1g,LÞ 7! E ðf0,1g,LÞ with will be Z n with periodic boundary conditions.
rule components xi is additive if and only if for Several forms of representation for additive
i = i0. . .ik1 and j = j0. . .jk1, with rules with periodic boundary conditions appear
(i + j)r = ir + jr mod (2) it is true that in the literature. Rules have been represented as
xi + xj = xi+j mod (2). dipolynomials over finite fields (Martin et al.
1984); as recursion relations (Jen 1986b,
Corollary 1 (Voorhees 1996) 1988b); as circulant matrices (Guan and He
A k-site rule X : E ðf0,1g,LÞ 7! E ðf0,1g,LÞ is 1986; Das et al. 1992); and as polynomials in
additive if and only if for all i = i0. . .ik1 roots of unity or in powers of the left shift
(Voorhees 1996).
X
k1 The global operator for a k-site additive rule with
xi ¼ ikr1 x2r mod ð2Þ: (6) neighborhood structure N ¼ fr, . . . , k  r  1g
r¼0 can be written as
Xk1
X ¼ sr as ss as  A: (7)
By Eq. (6), the k rule components x2r , s¼0
0  r  k  1, determine the set of k-site additive In the dipolynomial representation, a state
k
rules. Hence only 2k of the possible 22 k-site rules m = m0. . .mn1 defines a dipolynomial m(x) while
are additive, including the 0 rule that maps all the rule X is represented by the dipolynomial
sites to 0. equivalent of Eq. (7):
Additive Cellular Automata 135

Additive Cellular
Automata, Fig. 1 Space-
time output of rules 60 and
90 from initial state with a
single 1

X
n1 and by s + s5 in the form of Eq. (9)
mðxÞ ¼ ms x , as  A
s
because 1 = 5 mod (6).
s¼0
(8) When a rule X acting on Z n is represented as in
X
k1
Eq. (9) the string (a0, . . ., an1) is directly
X¼ s
as x , as  A
s¼0
connected to the representation of X as a circulant
matrix.
A right circulant matrix is a matrix in which
The action of X on a state m is obtained by each successive row is obtained from the row
multiplication of the corresponding immediatly above by shifting that row one unit
dipolynomials: X ðxÞmðxÞ , with all products to the right, with the final row entry shifted to the
reduced mod(n) (Martin et al. 1984). front. Thus,
The rule X is also representable in terms
of powers of the shift operator, with the circða0 , a1 , a2 , . . . , an2 , an1 Þ
expression
0 1
a0 a1 a2  an2 an1
X
k1 B an1 a0 a1  an3 an2 C
X¼ as ss as  A (9) ¼B
@
C
A
⋮
s¼0
a1 a2 a3  an1 a0

where the coefficients in this equation differ from If m  E ðA, Z n Þ is written as a column
those in Eq. (7). For example, rule 90 acting on vector, the operation of X n m is given as
Z n is represented by s1 + s in the form of Eq. (7), multiplication by the right circulant matrix circ
136 Additive Cellular Automata

(a0, . . ., an1) with all terms reduced mod(p). Lemma 2 (Davis 1979)
The value of this representation is that properties An n  n matrix A is circulant if and only
of circulant matrices are well known and if A = PA(s) for some polynomial PA of degree
this provides significant information about the less than or equal to n. Further, if A = circ
cellular automata rule. This approach has a nat- (a0, . . ., an1) then
ural extension to the case of null boundary
conditions (Tadaki 1994; Voorhees 2008), X
n1
although the matrix involved is no longer a PA ðsÞ ¼ as ss (11)
complete circulant. For example, the nearest s¼0
neighbor form of rule 90 acting on Z 6 with
periodic boundary conditions has matrix
representation Other properties of circulant matrices
provide links to the representation of rules over
0 1 Z n by polynomials in the nth roots of unity.
0 1 0 0 0 1
B1 0 1 0 0 0C Let o = e2pi/n be the first complex nth root of unity.
B C
B0 1 0 1 0 0C
B C (10a)
B0 0 1 0 1 0C
B C
@0 0 0 1 0 1A Definition 1
1 0 0 0 1 0 The Fourier matrix of order n is the matrix

while the matrix representation for null boundary 0 1

1 1 1 ... 1 1
conditions is B1 on1 on2 ... o2 o C
1 B
B1 on2 on4 ... o4
C
o2 C
0 1 F n ¼ pffiffiffi B C
nB ⋮ C
0 1 0 0 0 0 @1 o2
o 4
... on4
on2 A
B1 0 1 0 0 0C
B C 1 o o2 ... on2 on1
B0 1 0 1 0 0C
B C (10b)
B0 0 1 0 1 0C (12)
B C
@0 0 0 1 0 1A
0 0 0 0 1 0
Denote the Hermitian conjugate (transpose,
The circulant representation of the left shift s complex conjugate) of this by F n.
on Z n is

0 1 Lemma 3 (Davis 1979)

0 1 0 0 0  0 0
B0 a) The Fourier matrix is unitary: F n F n ¼ F n F n
B 0 1 0 0  0 0CC
s¼B ⋮ C (10c) ¼ I.
B C
@0 0 0 0 0  0 1A b) The eigenvalues of Fn are 1 and i with
1 0 0 0 0  0 0 multiplicities depending on n.
c) The characteristic polynomials of F n are
The following lemmas (Davis 1979) show the
direct connection to Eq. (9) and to Hedlund’s  n1
condition: ðl  1 Þ2 ðl  i Þðl þ 1 Þ l 4  1 4 n ¼ 0 mod ð4Þ
 ðn1Þ=4
ðl  1 Þ l 4  1 n ¼ 1 mod ð4Þ
 2  ðn2Þ=4
l  1 l4  1 n ¼ 2 mod ð4Þ
Lemma 1 (Davis 1979)   ðn3Þ=4
An n  n matrix A is circulant if and only if ðl  iÞ l2  1 l4  1 n ¼ 3 mod ð4Þ
[A, s] = 0. (13)
Additive Cellular Automata 137

Every n  n circulant matrix A is diagonalized period cycles provide an effective means of generat-
by Fn. Further, if PA is the polynomial defined by ing pseudorandom sequences. For a binary valued
Eq. (11) then rule operating on strings of length n and all divisors
of n, the maximal possible cycle period is 2n  1 (i.e.,
F n AF n ¼ LðAÞ every state but the 0 state is on the cycle). If a rule
  operating on strings of length n has periodic bound-
¼ diag PA ð1Þ, PA ðoÞ, . . . , PA on1 ary conditions, the shift operator produces cycles of
(14) length n. As a result, no additive rule acting on
strings with periodic boundary conditions can pro-
hence the rth eigenvalue of A is PA(or). Define the duce cycles of maximal period. If null boundary
n  n matrix pr = diag (0, . . ., 0, 1, 0, . . ., 0) conditions are used, however, maximal period cycles
with the 1 in the rth position and set Pr ¼ F n pr F n. can appear (Nandi and Pal Chaudhuri 1996).
The matrices Pr are Hermitian and satisfy the For a given cellular automata rule, the set
conditions of states having no predecessor is called the
 Garden-of-Eden (GoE). All additive rules
0 r 6¼ s
Pr Ps ¼ acting on E ðA,Z Þ are surjective but this is not
Ps r ¼ s
true of rules acting on E ðA, Z n Þ: For
X
n1
Pr ¼ I additive rules acting on strings of length n
r¼0 with periodic boundary conditions there will be
configurations having no predecessors. Neverthe-
Thus, they are orthogonal, idempotent, and less, these Garden-of-Eden states are not intrinsic
form a resolution of unity. Hence, they are a to the rule since they do have predecessors when the
complete set of projection matrices. state space E ðA, Z n Þ is embedded in E ðA,Z Þ, as
the following example shows.
Lemma 4 (Davis 1979)
Let A = circ(a0, . . ., an1). Then Example 5
Let X : E ðf0,1g,Z Þ 7! E ðf0,1g,Z Þ be defined by
X
n1
the rule table
A¼ PA ðor ÞPr (15)
r¼0
000 001 010 011 100 101 110 111
Representation of an additive rule in terms of 0 0 1 1 1 1 0 0
complex polynomials yields an interesting result on
injectivity. A rule X : E ðA,Z Þ 7! E ðA,Z Þ is sur- For X : E ðf0,1g, Z 3 Þ 7! E ðf0,1g, Z 3 Þ the
jective if every configuration has a predecessor. If the states {001, 010, 100, 111} have no predeces-
predecessor of a configuration having a predecessor sors. It is easy to show, however, that these states
is unique, the rule is injective. A rule that is both do have predecessors in Z 6 when Z 3 and Z 6 are
surjective and injective is bijective. For cellular embedded in Z:
automata, injectivity is equivalent to reversibility. If
a ruleX is reversible, there is an inverse rule; that is, a f000111, 111000g 7! 001 ;
reversible rule X 1 such that if x  E ðA,Z Þ and
f001110, 110001g 7! 010;
X ðÞ ¼ x, then X 1 ðxÞ ¼  (Culik 1987; Toffoli
and Margolis 1990; Kari 1990; Sutner 1991; Morita f011100, 100011g !7 100;
1994; Moraal 2000). The question of whether or not f01, 10g 7! 1
a cellular automata rule is surjective or injective is
decidable only in dimension one (Kari 1990). Injec-
tive additive rules are also called group rules While all additive rules are surjective in
(Chaudhuri et al. 1997). Those with maximum this sense, not all additive rules are injective.
138 Additive Cellular Automata

The condition for injectivity is easily seen using Additive Cellular Automata, Table 1 Injective rules
the representation in terms of nth roots of unity. for k  5
From Eq. (14) a rule will be injective on X k Injectivity condition
E ðA, Z n Þ if and only if its diagonalized matrix s4 5 None
representation is invertible, hence there can be s3 4 None
no zero eigenvalues. s2 3 None
s 2 None
I 1 None
Lemma 5 (Voorhees 1994, 1996)
s2 + s3 + s4 5 n 6¼ 3m
Let X : E ðA, Z n Þ 7! E ðA, Z n Þ be an additive
s + s3 + s4 5 None
rule represented by A = circ(a0, . . ., an1). X s + s2 + s4 5 None
is injective on Z n Z if and only if I + s3 + s4 5 None
PA(os) 6¼ 0 mod (p) for 0  s  n  1. I + s2 + s4 5 n 6¼ 3m
I + s + s4 5 None
If the rule in Lemma 5 is to be injective on Z s + s2 + s3 4 n 6¼ 3m
then it must be so on Z n for all n. Thus I + s2 + s3 4 None
the complex polynomial PA(z) must be I + s + s3 4 None
irreducible with respect to all nth roots of I + s + s2 3 n 6¼ 3m
unity, for all n. This leads to the next theorem: I + s + s2 + s3 + s4 5 n 6¼ 5m

Theorem 2 (Voorhees 1994, 1996)

Let X : E ðA,Z Þ 7! E ðA,Z Þ be an additive rule Transient Lengths and Cycle Periods
represented as in Eq. (9). X will be injective if and For any cellular automata acting on a finite state
only if space, every state eventually maps to a fixed
þ 0 point or cycle. If a rule is injective, it is revers-
1 P A ðzÞ
lim dz ¼ 0 (16) ible and every state is a fixed point, or is on a
e!0 2pi P A ðzÞ
cðeÞ cycle. If not injective, there will be states with-
out predecessors, Garden-of-Eden states. As
where P0A ðzÞ is the derivative of PA(z) and c(e) is a indicate, however, if a rule is additive its
contour consisting of circles of radius 1 + e and Garden-of-Eden states are spurious in the sense
1  e. that they do have predecessors if the state space
is enlarged.
This follows since the integral in Eq. (16) The following theorem lists several signifi-
counts the number of zeros minus the number of cant properties of cellular automata rules acting
poles of PA(z) contained within the contour. Since on E ðA, Z Þ or E ðA, Z þ Þ with left justified
there are no poles, in the limit this counts the neighborhoods.
number of zeros on the unit circle.
Theorem 3 (Voorhees 1996)
Corollary 2 (Voorhees 1996) Let X be a k-site cellular automata rule acting on
Let X : E ðf0,1g, Z n Þ 7! E ðf0,1g, Z n Þ be defined E ðA, Z Þ or on E ðA, Z þ Þ with left justified
P
by X ¼ k1 s
s¼0 s . neighborhoods. Then the following
statements are equivalent: (a) X is surjective,
1. If k is even, X is not injective. (b) X has an empty Garden-of-Eden,
2. If k is odd, X will be injective for all n such that (c) Every finite sequence m0. . .mn1 as exactly
n 6¼ mk for any m. pk1 pre-images and every state m has at
most pk1 predecessors, (d) X maps eventually
Table 1 lists all rules that are injective on E periodic states to eventually periodic states and
ðf0,1g, Z n Þ for at least some n, with k  5. non-periodic states to non-periodic states, (e) as
Additive Cellular Automata 139

a map of the interval [0, 1] X maps rationals to 4. Let n = 2km, m odd. The maximum
rationals and irrationals to irrationals. cycle period c(n) for an additive rule
acting on Z n satisfies c(n) j 2ord(n, m)  2m
If X : E ðA, Z n Þ 7! E ðA, Z n Þ is a k-site rule where ord(n, m) = min {r| 2r = 2m mod (m)}
with jA j ¼ p and either periodic or null boundary
conditions, the state transition diagram, STDðX Þ is In most cases, the maximum cycle
a graph with pn vertices labeled by the set of p-adic period equals 2ord(n, m)  2m or, if the rule is
numbers {i0, . . ., in1| 0  ir  p  1}. An edge is symmetric, 2sord(n, m)  2m. In Jen (1988b,
directed from the vertex i0, . . ., in1 to the vertex j0, 1989), Jen shows that when this is not the
. . ., jn1 if and only if X ði0 , . . . , in1 Þ ¼ j0 , . . . , case, it is a number theoretic consequence of
jn1 . Each state m maps to a unique state X ðmÞ so an anomalous shift that reduces the maximum
STDðX Þ consists of a set of trees rooted on fixed cycle period. As indicated, the choice of
points or cycles. States at the top of trees are mapping site can influence cycle periods
Garden-of-Eden states. and the effect of anomalous shifts is
If hðX ,nÞ is the maximum tree height, states analogous. This is an immediate result of the
at heights h  hðX ,nÞ cannot appear after hðX ,nÞ next theorem.
h þ 1 iterations and after hðX ,nÞ iterations
only fixed points and states on cycles remain. Theorem 5 (Jen 1988a, 1989)
Thus, iteration of a non-injective rule on Let X : E ðA, Z n Þ 7! E ðA, Z n Þ be the global
E ðA, Z n Þ decreases the number of available- mapping for a cellular automaton rule. A state m
states with a corresponding reduction in is on a cycle of X if and only if there exist
entropy. On the other hand, non-injective integers r and s such that X r ðmÞ ¼ ss ðmÞ.
additive rules acting on E ðA,Z þ Þ do not In some cases r = c(n) and s = n so the theorem
reduce entropy (Lind 1984) even though the do is not as strong as it might first appear. If X r ðmÞ
so on E ðA, Z n Þ for all n. The explanation for this ¼ ss ðmÞ for all states on cycles of maximum
apparent paradox is that the Garden-of-Eden states period, however, then c(n) = rt where t = min
that appear in E ðA, Z n Þ are artifacts of the finite {j| js = 0, mod(n)}. A change of mapping site is
length of states in this space. When embedded inZ þ, equivalent to multiplication by a power of the shift
states in Z n correspond to periodic configurations, s.
 k If mr is on akscycle of maximum period then
hence to rational numbers in [0,1] and the set of all s X ðmÞ ¼ s ðmÞ and the cycle period is rq
rationals has measure 0 in the reals. with q = min {j| j(k  s) = 0 mod (n)}. In general,
Parameters of interest for characterizing state there is no requirement that q = t.
transition diagrams of rules acting on E ðA, Z n Þ Comprehensive results on cycle periods
are the maximum tree height hðX ,nÞ and the cycle and maximum transient lengths in the case of
periods cs ðX ,nÞ. null or periodic boundary conditions were
first obtained by Elspasa (1959) in work charac-
Theorem 4 (Martin et al. 1984) terizing the cycle sets in the state transition
1. Trees rooted at all vertices on cycles or at fixed diagrams of linear machines. Since then a
points of the STD for additive cellular autom- number of researchers have independently
ata are isomorphic to the tree rooted at the fixed derived similar results (Stevens et al. 1993;
point 0. Stevens 1999; Thomas et al. 2006; Sutner 1988a,
2. The periods of all cycles of an additive rule b, 2000, 2001). Let A = circ(a0, . . ., an1)
acting on Z n are divisors of the period for and let m be a state in E ðA, Z n Þ. The minimal
cycles obtained by starting from an initial annihilating polynomial of m is the monic
state containing only a single 1. polynomial Pm(z) such that
3. Let c(m) be the maximum cycle period for an Pm(A)m = 0 mod (p). This polynomial exists
additive cellular automaton acting on Z m and since A always satisfies its characteristic equation.
take n = 2m. Then c(n) divides 2c(m). Let Pm ðzÞ ¼ zk Pm ðzÞ with Pm ð0Þ 6¼ 0. The order of
140 Additive Cellular Automata

Pm(z), ord(Pm(z)), is defined as the smallest natural A if and only if L(O) j m(o). The fraction
number c such that Pm ðzÞ j ðzc  1Þ. The following of reachable configurations is 2n. Further, if
theorem is given in (Stevens 1999; Thomas et al. d(m) is the in degree of the STD vertex labeled
2006). Alternate versions appear in (Sutner by m then
1988a, b, 2001).

0 m is a Garden  of  Eden state
d ðmÞ ¼
Theorem 6 2n otherwise
Let m  E ðA, Z n Þ with minimal annihilating (18)
polynomial Pm ðzÞ ¼ zk Pm ðzÞ. Then Akm belongs
to a cycle with period c = ord(Pm(z)). Example 6a (Rule 90 acting on E(f0,1g, Z n )
with periodic boundary conditions)
Since the minimal annihilating polynomial The circulant matrix for nearest neighbor
always divides the minimal polynomial, rule 90 acting on Z n with periodic boundary
which, for additive cellular automata conditions is A(d) = circ(0, 1, 0, . . ., 0, 1).
represented by circulant matrices, or by the The characteristic polynomial of this
corresponding null boundary condition matrix, is matrix satisfies the recurrence relation
the same as the characteristic polynomial of that Qn + 1(x) = xQn(x) + Qn  1(x) mod (2). For
matrix, all cycle periods and maximum transient n = 2sm with m odd, the characteristic polynomial
lengths can be found from the characteristic poly- of A(d) has the form
nomial. Hence, the maximum cycle period is the
order of the characteristic polynomial since there s sþ1

always exists a state whose minimal Qn ðxÞ ¼ x2 r2m ðxÞ

(19)
annihilating polynomial is the minimal polynomial. r2kþ1 ðxÞ ¼ xr2k1 ðxÞ þ r2k3 ðxÞ
The questions of reachability, and conditions
for states to be on cycles is also addressed using and the minimal annihilating polynomial is
the formulation in terms of roots of unity. If an Voorhees (2008)
additive rule acting on E ðA, Z n Þ is represented
by a circulant matrix A and a state m is represented xrm ðxÞ s¼0
P (20)
by m ¼ n1 s
s¼0 ms o , write
s1 s
x2 r2m ðxÞ s>0

Y
r Y
s where rm(0) 6¼ 0. For n = 6 (s = 1, m = 3),
PA ðoÞ ¼ rj ðoÞ Ok ðoÞ ¼ rðoÞOðoÞ the circulant matrix representing rule
j¼1 k¼1 90 is given by Eq. (10a). The characteristic
(17) polynomial of this matrix, with
coefficients reduced mod(2), is x2-
where r(o) is a product of the irreducible factors of (x + 1) = x (x + 1) . Thus, the minimal
4 2 4

PA(o) representing injective rules, and O(o) is a polynomial is x(x + 1)2 = x(x2 + 1) and r3-
product of the irreducible factors representing non- (x) = (x + 1). The minimum integer c such that
injective rules. Diagonalization of A yields L- (x2 + 1) divides (xc + 1) is just c = 2, showing that
(A) = L(r)L(O) and L1(r) exists since r(o) the maximum cycle period is 2.
represents injective (hence reversible) rules. Let
v denote the nullity of L(O).
Example 6b (Rule 90 Represented by Roots of
Unity, Acting on Z 6 )
Theorem 7 (Voorhees 1996; Lidl and Pilz 1984) The Theorem 5 conditionX r ðmÞ ¼ ss ðmÞfor a state
A state m(o) is reachable by an additive cellular m to be on a cycle can be written as Lr ðX ÞF n ðmÞ
automata rule with circulant matrix representation ¼ Lns ðsÞF n ðmÞ. This gives the conditions for m to
Additive Cellular Automata 141

be on a cycle as a set of linear equations (Voorhees 1. The initial 0 in 0110 can arise from either 00 or 11.
1996). For n = 6, elementary rule 90 has the form 2. Starting with a 00, the next symbol in 0110 is a
d = s + s5, or in terms of the sixth root of unity, 1 and this can arise from a 01 or a 10, but this
d(o) = o + o5. Using o3 =  1 with all sums must also connect to the original 00 so only 01 is
taken mod(2) allowed, giving 001. Starting from a 11, on the
  other hand, the same reasoning requires 110.
LðdÞ ¼ diag 0, o þ o5 , o2 þ o4 ,0, o2 þ o4 , o þ o5
  3. The third symbol in 0110 is also a 1. To be
L2 ðdÞ ¼ diag 0, o2 þ o4 , o2 þ o4 ,0, o2 þ o4 , o2 þ o4
  consistent with 001 requires that 10 be
L3 ðdÞ ¼ diag 0, o þ o5 , o2 þ o4 ,0, o2 þ o4 , o þ o5
selected, and to be consistent with 110 requires
(21) that 01 be selected, thus giving the two partially
constructed possibilities as 0010 and 1101.
hence L3(d) = L(d) or equivalently, L(d)(L2- 4. Finally, the fourth symbol must be a 0. This
(d) + I) = 0 mod (2). Thus, the maximum tree height requires that the predecessor string conclude
is one and the maximum cycle period is two. In with either 00 or 11. Since the strings are in Z 4
addition, n = 2 so that 1/4 of the total of 64 states with periodic boundary conditions, the final
will be on cycles. Further, for n = 6 (observing that symbol in the predecessor string must also be
1 = 1 mod (2)) the first symbol in that string. Thus, both 0010
0 1 and 1101 are seen to be predecessors of 0110.
m0 þ m1 þ m2 þ m3 þ m4 þ m5
B m0 þ m1 o þ m2 o þ m3 o þ m4 o þ m5 o C
5 4 3 2
B C
1 B m0 þ m1 o4 þ m2 o2 þ m3 þ m4 o4 þ m5 o2 C Other ways of computing predecessor states for
F 6 ðmÞ ¼ pffiffiffi B C
6B þ m2 þ m3 o3 þ m4 þ m5 o3 C finite strings is through the construction of a rule
3
B m0 þ m1 o C
@ m0 þ m1 o2 þ m2 o4 þ m3 þ m4 o2 þ m5 o4 A matrix (Voorhees 1996) or the use of de Bruijn
m0 þ m1 o þ m2 o2 þ m3 o3 þ m4 o4 þ m5 o5 diagrams (Voorhees 1996; Sutner 1991). Backward
(22) reconstruction, the rule matrix, and use of a de
Bruijn diagram are valid methods for computing
Since o2 + o4 = o3 =  1 = 1 mod (2), L2- predecessor states for all one-dimensional rules.
(d) = diag (0, 1, 1, 0, 1, 1) and the condition For additive rules, however, there is an analytic
L2(d)F6(m) = F6(m) requires that [F6(m)]0 = [F6(- means for computing predecessor states, starting
m)]3 = 0 which reduces to from left justified neighborhoods defined on
m0 + m2 + m4 = m1 + m3 + m5 = 0 or equivalently, E ðA, Z n Þ or E ðA,Z þ Þ (Voorhees 1993, 1996).
m4 = m0 + m2 and m5 = m1 + m3. Hence a state m will This can be illustrated for rules defined on
be on a cycle if and only if it has the form E ðf0,1g,Z þ Þ . This method also works for rules
m = (m0, m1, m2, m3, m0 + m2, m1 + m3). defined on E ðf0,1g, Z n Þ if it is embedded in E
ðf0,1g,Z þ Þ as the subset of half-infinite periodic
Computing Predecessor States sequences with periods that divide n. Define oper-
A problem of general interest for cellular automata ators B : E ðf0,1g,Z þ Þ 7! E ðf0,1g,Z þ Þ and s1
is computation ofpredecessor states. For a rule X : E ðf0,1g,Z þ Þ 7! E ðf0,1g,Z þ Þ by
: E ðA,LÞ 7! E ðA,LÞ with a state b given this
requires solution of the equation X ðmÞ ¼ b. It is X
s
always possible to construct solutions for this ½BðmÞs ¼ mi modð2Þ
equation, or to show that none exist by a method 
i¼0 (23)
0 s¼0
of backward reconstruction based on the rule table. ½s1 ðmÞs ¼
ms1 s>0
Example 7 (Rule 60 Acting on Z 4 With Periodic
Boundary Conditions) Theorem 8 (Voorhees 1993, 1996)
Rule 60 is a 2-site rule, defined by (00, 11) 7! 0, Let D = I + s be the global operator for elemen-
(01, 10) 7! 1. Given the state 0110 the predeces- tary rule 60 acting on E ðf0,1g,Z þ Þ and define the
sors of this state can be computed as follows: state a(s) in Z þ by
142 Additive Cellular Automata

h i  ð0Þ
ðsÞ 0 i 6¼ s absorbing the term m0 að0Þ into the constant
a ¼ parameter, results in the equation
i 1 i¼s
     
I þ B2 s1 mð0Þ ¼ B a0 að0Þ þ a1 B að1Þ
1. The general solution of D(m) = b is
þBs1 bð0Þ þ B2 s2 bð1Þ
(28)
m ¼ a0 B að0Þ þ Bs1 ðbÞ:
Theorem 9 (Voorhees 1993, 1996)
2. The general solution of D (m) = b is k
The operator (I + B2s1) is invertible with
(I + B2s1)1 = I + C(2, 1) with
X
k1
C ð2,1Þ ðxÞ i
m¼ as Bsþ1 aðsÞ þ Bk sk ðbÞ, 
0 i¼0
s¼0 ¼ Pdði1 Þ
3 e
s¼0 ðxi3s2 þ xi3s1 Þ i > 0
where the coefficients as provide initial conditions. (29)

The general technique for computing predeces- where all sums are taken mod(2), xr = 0 for r < 0,
sors can be illustrated with the case of rule and dxe indicates the greatest integer less than or
150, expressed in left justified form as X ¼ I þ s equal to x.
þs2 . To solve (I + s + s2)(m) = b define four
sequences: The solution for m(0) is substituted into the
second equation of (27) yielding a solution for
ð0Þ
mi ¼ m2i ,
ð1Þ
mi ¼ m2iþ1 m(1). These are recombined to get the general
ð0Þ ð1Þ (24) solution for m. This technique of reducing a single
bi ¼ b2i , bi ¼ b2iþ1 i ¼ 0,1,2, . . .
equation to a set of coupled equations involving
simpler additive rules works in general although
The equation (I + s + s2)(m) = b reduces to the
the form for partitioning of sequences is specific
pair of coupled equations
to the particular case. Computation of predeces-
ð0Þ ð0Þ ð1Þ ð0Þ sors involves inversion of operators of the form
mi þ miþ1 ¼ mi þ bi
ð1Þ ð1Þ ð0Þ ð0Þ (25) I + Brss. The general form for the inverse of this
mi þ miþ1 ¼ miþ1 þ bi
operator is I + C(r, s) where C(r, s) is the lower
triangular matrix that is the solution of the
and these can be written as
equation
 
D mð0Þ ¼ mð1Þ þ bð0Þ X 
    (26) jþr
rþ1
D mð1Þ ¼ s mð0Þ þ bð1Þ C ðr,sÞ þ C ðr,sÞ i,jþs
m¼j
mjþ1 im

From Theorem 8, Eq. (26) can be formally ¼ di,jþs (30)

solved to obtain
 
mð0Þ ¼ a0 B að0Þ þ Bs1 mð1Þ þ bð0Þ
  (27) d-Dimensional Rules
mð1Þ ¼ a1 B að0Þ þ Bs1 smð0Þ þ bð1Þ
Both (Martin et al. 1984; Guan and He 1986) dis-
Substituting the second equation of (27) into cuss the extension from one-dimensional to
 
the first, making use of the identity s2 s mð0Þ d-dimensional rules defined on tori. In Martin et al.
ð0Þ (1984) this discussion uses a formalism of multino-
¼ s1 mð0Þ þ m0 að0Þ , rearranging terms and
mials defined over finite fields. In Guan and He
Additive Cellular Automata 143

(1986), the one-dimensional analysis based on j ...j

½CðS ÞC
X ðT Þi11 ...idd
circulant matrices is generalized. The matrix ...k d j ...j (33)
¼ ½CðS Þki11...i d
½CðT Þk11 ...kd d mod ðpÞ
formulism of state transitions is retained by defining k 1 ...k d
a d-fold “circulant of circulants,” which is not, of
itself, necessarily a circulant. Computation of the is also a generalized circulant and CðS ÞCðT Þ ¼ C
non-zero eigenvalues of this matrix yields results ðcðS,T ÞÞ.
on transient lengths and cycle periods.
More recently, an extensive analysis of additive An important tool for analysis of additive rules
rules defined on multi-dimensional tori has on multi-dimensional tori is the discrete baker
appeared (Bulitko et al. 2006). A d-dimensional transformation. This is a discrete version of
!
integer vector n ¼ ðn1 , . . . , nd Þ defines a discrete the baker transformation (Bernoulli shift) for con-
!
toridal lattice L n . Every d-dimensional matrix tinuous one dimensional dynamical systems
!
of size n with entries in A , jA j ¼ p (prime), (Arnold and Aviz 1968). The discrete baker trans-
!
defines an additive rule acting on E A, L n formation is the operation Bp : E A, L n
!
7!
as follows: Let T and m(t) be elements of
!
E A, L n
!
with X the rule defined by T and E A, L n defined by

m(t) a state at time t. The state transition defined by X

X is mðt þ 1Þ ¼ X ðmðt ÞÞ and this is given by Bp  T i1 ...id
¼ T k 1 ...k d modðpÞ (34)
is ¼pk s modðns Þ

½mðt þ 1Þi1 ...id

X ...k d ...k d
with the empty sum set to 0. In the one dimen-
¼ ½CðT Þki11...id
½mðt Þk 1 ...k d ½CðT Þki11...id sional case with p = 2 and T ¼ ða0 , . . . , an1 Þ this
k 1 ,...,k d
becomes
¼ T j1 ...jd js ¼ k s  is modðns Þ
(31) X
½B2  T i ¼ ak modð2Þ (35)
The matrix CðT Þ is the d-dimensional general- k:i¼2k modðnÞ
ization of a circulant matrix with T as the equiv-
alent of its first row. For example, if d = 1 and Example 8 If d = 1, p = 2 and n = 7 then
p = 2 with T ¼ ð0,1,0,0,0,1Þ this defines the B2(a0, a1, a2, a3, a4, a5, a6) = (a0, a4, a1, a5,
additive rule s + s5 (rule 90) and the matrix C a2, a6, a3). On the other hand, if n = 6 then
ðT Þ is given in Eq. (10a). B2(a0, a1, a2,
!
Let S and T be elements of E A, L n and a3, a4, a5) = (a0 + a3, 0, a1 + a4, 0, a2 + a5, 0).
!
define the binary operation c : E A, L n  In one dimension, if X is the rule defined by
E A, L n
!
7! E A, L n
!
by T ¼ ða0 , . . . , an1 Þ then Bp X ¼ X 2 .

X Example 9
½cðS,T Þi1 ...id ¼ S k 1 ...k d T i1 k 1 ...id k d
If d = 2 and p = 3 with n1 = 5, n2 = 6 and
k 1 ,...,k d
0k s <ns 2 3
a00 a01 a02 a03 a04
(32) 6 a10
6 a11 a12 a13 a14 7
7
6 a20 a21 a22 a23 a24 7
with all sums taken mod(p). 6 7
T ¼6
6 a30 a31 a32 a33 a34 7
7
6 a40 a41 a42 a43 a44 7
6 7
Lemma 6 (Bulitko et al. 2006) 4 a50 a51 a52 a53 a54 5
Let S and T be as above, with generalized
circulant matricesCðS Þ, CðT Þ. The product CðS ÞC
ðT Þ defined by then
144 Additive Cellular Automata

2 3
a00 þ a20 þ a40 a02 þ a22 þ a42 a04 þ a24 þ a44 a01 þ a21 þ a41 a03 þ a23 þ a43
6 0 0 0 0 0 7
6 7
6 0 0 0 0 0 7
B2 T ¼ 6

6 a30 þ a10 þ a50
7
6 a32 þ a12 þ a52 a34 þ a14 þ a54 a31 þ a11 þ a51 a33 þ a13 þ a53 7
7
4 0 0 0 0 0 5
0 0 0 0 0

with all sums mod(3). The baker transformation is a linear transfor-

The discrete baker transformation exponen- mation on the space of d-dimensional matrices
tially speeds up rule evolution. with entries from A . Since each element of this
space defines an additive cellular automata rule,
Theorem 10 (Bulitko et al. 2006) the vertices of the state transition diagram for the
r
If p is prime then ½CðT Þp ¼ C Brp T . baker transformation can be labeled by these
rules, and this is exhaustive.
Definitions:
Thus, if the matrix T defines a rule
! ! r
X : E A, L n 7! E A, L n then X p 1. An oriented graph G = (V, E) is a set V of
is obtained from the matrix .Brp T vertices together with an edge set E V  V.
The operator Bp is a permutation if every sum If (v, w)  E then there is an edge directed
in Eq. (34) contains at most one non-zero sum- from vertex v to vertex w.
mand. In Example 7, B2 is a permutation for n = 7 2. An oriented graph G1 = (V1, E1) reduces to an
but is not for n = 6. oriented graph G2 = (V2, E2) modulo
p(G1<pG2) if (a) V1 = V2 = V, and
Lemma 7 (Bulitko et al. 2006) (b) (v, w)  E if and only if there is an oriented
Set ι = max {expns| 1  s  d}. That is, ι is the chain of length p from v to w in G2.
largest integer k such that for some s, pk divides ns.
!
Then for all r > ι and any T  E A, L n , Bp Example 10
is a permutation on Brp T . Figure 2 shows the reduction modulo 3 of a
connected graph with 12 vertices to a graph with
Theorem 11 (Bulitko et al. 2006) three connected components.
Let q be prime and let ordmq be the order of q in m
The baker diagram for the space
when this is defined and 1 otherwise. Write ns !
¼ pk s ms and set c ¼ lcmðord m1 p, . . . , ord md pÞ . E A, L n has an edge directed from rule
Then, for any rule X defined by a matrix T  E X to rule Y if and only if Y ¼ Bp X .
!
A, L n the following are true:
Lemma 8 (Bulitko et al. 2006)
If ðX ,Y Þ is an edge in STD(Bp) then STDðY Þ<p
Biþc i
p T ¼ Bp T STDðX Þ.

1. The period of any cycle in STDðX Þ divides !

Since E A, L n is finite, iteration of Bp
pι(pc  1) !
2. The height of trees in STDðX Þ does not on E A, L n must eventually reach a fixed
!
exceed pι. point or cycle. If T  E A, L n set tðT Þ
Additive Cellular Automata 145

Additive Cellular
Automata,
Fig. 2 Example of Graph
Reduction Modulo 3 for a
Graph with 12 Vertices
(from (Bulitko et al. 2006))

equal to the sum mod(p) of all entries in T . If X is theoretical questions revolve around the issue of
the rule defined by T take ðX Þ as the maximum surjectivity. While all additive cellular automata
height of trees in STDðX Þ, pðX Þ as the number of are surjective (bracketing the spurious Garden-of-
vertices in STDðX Þ lying on cycles, and ys ðX Þ as Eden states that exist for rules defined on finite
the number of cycles of length s contained in state spaces) and for general rules surjectivity in
STDðX Þ. dimensions higher than one is undecidable, it is
desirable to have a simple general criteria for
surjectivity in one dimension. It would also be
Theorem 12 (Bulitko et al. 2006) useful to know if any surjective cellular automa-
1. If STDðX Þ<p STDðY Þ then ton (additive or non-additive) is capable of uni-
a) ðX Þ  ðpY Þ versal computation. If there are surjective rules
b) If j is the largest number such that jypj ðY Þ that are universal computers this would connect
to work in mathematical logic on the computa-
> 0 then ypj ðX Þ ¼ 0
tional limits of formal systems (Chaitin 2006).
c) pðX Þ ¼ pðY Þ.
This connection arises because surjective rules
2. Let X, Y be two rules belonging to a cycle R of
only exhibit Garden-of-Eden states when acting
STD(Bp) with period lR and let {c1, . . ., ck} be
on finite state spaces. The appearance of prede-
the set of all cycle periods in STDðX Þ. Then
cessor states when the state space is enlarged
a) STDðX Þ and STDðY Þ are isomorphic as
seems analogous to the increase in computational
graphs
power of a formal system when it is enlarged by
b) ðX Þ  1 and p ∤ cs for all s
  the addition of a new axiom.
c) For all s, c s j pl R  1 ,
Section “d-Dimensional Rules” introduced the
l R ¼ l cmðord c1 p, . . . , ord ck pÞ, lR j c discrete baker transformation of additive cellular
d) In addition, if X is defined by the matrix T automata rules. This transformation is a linear
and tðT Þ ¼ 0 then the number of connected operator on the space of additive rules and further
components of STDðX Þ is divisible by p. research into its properties can contribute to a
deeper understanding of the structure of additive
Future Directions rules. In addition, there appears to be a connection
between certain universal numbers that can be
Much work remains on both the theoretical anal- defined from this transformation and the well-
ysis and the applications of cellular automata. known Mersenne and Fermat numbers. Elucida-
While much of this work will utilize non-additive tion of this connection would provide a significant
automata, there are still many open questions on link to number-theoretic aspects of cellular
additive cellular automata as well. Several automata.
146 Additive Cellular Automata

When applications are considered, et al. (2000), Rosenfeld (1979), Sternberg (1980),
(Chaudhuri et al. 1997) indicates many avenues Hopcroft and Ullman (1972), Cole (1969), Benjamin
and Johnson (1997), Carter (1984), Hillis (1984),
of continued development for additive cellular Manning (1977), Atrubin (1965), Nishio (1981),
automata. In addition, many of the references in Fischer (1965), Pries et al. (1986), Chaudhuri et al.
the general bibliography point to directions of (1997), Bardell and McAnney (1986), Hortensius
current cellular automata research in a number et al. (1989, 1990), Tsalides et al. (1991), Matsumoto
(1998), Tomassini et al. (2000), Das and Chaudhuri
of areas. As indicated in the introduction, appli- (1989, 1993), Serra (1990), Tziones et al. (1994),
cations in physics (crystal growth, hydrodynam- Mrugalski et al. (2000), Sikdar et al. (2002), Serra
ics, reaction-diffusion systems, astronomy), et al. (1990), Dasgupta et al. (2001), Chowdhury et al.
medicine and biology (pattern formation, genetic (1994, 1995a, b), Nandi et al. (1994), Cattell and
Muzio (1996), Cattell et al. (1999). Chaudhuri et al.
interaction networks, disease modeling, ecosys- (1997), which deals extensively with the use of addi-
tem modeling), pattern recognition and image tive cellular automata in computing applications and
processing, and computation (random number VLSI chip design, is of particular value. Willson
generation, language and pattern recognition, (1984a, b, 1987a, b, 1992), Peitgen and Richter
(1986), Culik and Dube (1989), Voorhees (1988),
test pattern generation for VLSI chips, signature von Haeseler et al. (1992a, b, 1993, 1995, 2001a, b),
analysis, error correcting codes, cryptography) Allouche et al. (1996), Barbé et al. (1995, 2003),
are only a small portion of the cellular automata Nagler and Claussen (2005), Takahashi (1990,
applications that continue to be studied. 1992) provide a good background in the relation
between cellular automata and fractal patterns.
Voorhees (1993, 1994, 1996, 2008), Martin et al.
(1984), Guan and He (1986), Das et al. (1992), Tadaki
(1994), Davis (1979), Culik (1987), Toffoli and
Bibliography Margolis (1990), Kari (1990), Morita (1994), Moraal
(2000), Lind (1984), Elspas (1959), Stevens et al.
This article provides a brief survey of some of the signif- (1993, 1999), Thomas et al. (2006), Sutner (1988a,
icant theoretical results on additive cellular automata, b, 1991, 2000, 2001), Lidl and Pilz (1984), Bulitko
together with references to applications of both cel- et al. (2006) deal with the theoretical analysis of
lular automata in general and additive cellular autom- additive cellular automata. A good survey of work
ata in particular. For historical information, in cellular automata up to the mid-1990s is Illichinsky
references von Neumann (1963), von Neumann and (2001). An important computational survey of cellu-
Burk (1966), Arbib (1966), Codd (1968), Sarkar lar automata dynamics is given in Wuensche and
(2000) are recommended. References Toffoli and Lesser (1992).
Margolis (1987), Duff and Preston (1984), Chopard
and Droz (1998), Lindenmayer and Rozenberg
(1976) provide a general background in the use of
cellular automata in modeling, as well as a number of
examples. Specific exemplary cases of applications Primary Literature
are found in Mackay (1976), Langer (1980), Lin and Adamatzky A, Costello BDL, Asai T (2005) Reaction-
Goldenfeld (1990), Greenberg and Hastings (1978), diffusion computers. Elsevier, London
Greenberg et al. (1978), Madore and Freedman Allouche JP, von Haeseler F, Peitgen H-O, Skordev
(1983), Adamatzky et al. (2005), Oono and Kohmoto G (1996) Linear cellular automata, finite automata and
(1985), Falk (1986), Canning and Droz (1991), Pascal’s triangle. Discret Appl Math 66:1–22
Vitanni (1973), Young (1984), Dutching and Arbib M (1966) Simple self-reproducing universal autom-
Vogelsaenger (1985), Moreira and Deutsch (2002), ata. Inf Control 9:177–189
Sieburg et al. (1991), Santos and Continho (2001), Arnold V, Aviz A (1968) Ergodic problems of classical
Beauchemin et al. (2005), Burks and Farmer (1984), mechanics. Benjamin, New York
Moore and Hahn (2002), Gerola and Seiden (1978), Atrubin AJ (1965) A one-dimensional real time iterative
Flache and Hegselmann (1998), Chen et al. (1990), multiplier. IEEE Trans Comput EC-14:394
Drossel and Schwabl (1992), Smith (1972), Barbé A, von Haeseler F, Peitgen H-O, Skordev G (1995)
Sommerhalder and van Westrhenen (1983), Ibarra Course-graining invariant patterns of one-dimensional
et al. (1985), Morita and Ueno (1994), Jen (1986a, two-state linear cellular automata. Int J Bifurc Chaos
b, 1988b, 1989), Raghavan (1993), Chattopadhyay 5:1611–1631
Additive Cellular Automata 147

Barbé A, von Haeseler F, Peitgen H-O, Skordev G (2003) Chowdhury DR, Gupta IS, Chaudhuri PP (1995b) A low-cost
Rescaled evolution sets of linear cellular automata on a high-capacity associative memory design using cellular
cylinder. Int J Bifurc Chaos 13(4):815–842 automata. IEEE Trans Comput 44:1260–1264
Bardell PH, McAnney WH (1986) Pseudo-random arrays Codd EF (1968) Cellular automata. Academic, New York
for built-in tests. IEEE Trans Comput C-35(7): Cole SN (1969) Real time computation by n-dimensional
653–658 iterative arrays of finite state machines. IEEE Trans
Beauchemin C, Samuel J, Tuszynski J (2005) A simple Comput C-18:349
cellular automaton model for influenza A viral infec- Culik IIK (1987) On invertible cellular automata. Complex
tion. J Theor Biol 232(2):223–234 Syst 1:1035–1044
Benjamin SC, Johnson NF (1997) A possible nanometer- Culik IIK, Dube S (1989) Fractals and recurrent behavior
scale computing device based on an additive cellular of cellular automata. Complex Syst 3:253–267
automata. Appl Phys Lett 70(17):2321–2323 Das AK, Chaudhuri PP (1989) An efficient on-chip deter-
Bulitko V, Voorhees B, Bulitko V (2006) Discrete baker ministic test pattern generation scheme. Euromicro
transformation for linear cellular automata analysis. J Microprocess Microprogramm 26:195–204
J Cell Autom 1:40–70 Das AK, Chaudhuri PP (1993) Vector space theoretical
Burks C, Farmer D (1984) Towards modeling DNA analysis of additive cellular automata and its applica-
sequences as automata. Physica D 10:157–167 tions for pseudo-exhaustive test pattern generation.
Canning A, Droz M (1991) A comparison of spin exchange IEEE Trans Comput 42:340–352
and cellular automata models for diffusion-controlled Das AK, Sanyal A, Chaudhuri PP (1992) On characteriza-
reactions. In: Gutowitz HA (ed) Cellular automata: tion of cellular automata with matrix algebra. Inf Sci
theory and experiment. MIT Press, Cambridge, 61:251–277
pp 285–292 Dasgupta P, Chattopadhyay S, Sengupta I (2001) Theory
Carter F (1984) The molecular device computer: point of and application of nongroup cellular automata for mes-
departure for large scale cellular automata. Physica sage authentification. J Syst Architecture
D 10:175–194 47(7):383–404
Cattell K, Muzio JC (1996) Synthesis of one-dimensional Davis PJ (1979) Circulant matrices. Wiley-Interscience,
linear hybrid cellular automata. IEEE Trans Comput- New York
Aided Des 15:325–335 Drossel B, Schwabl F (1992) Self-organized critical forest-
Cattell K, Zhang S, Serra M, Zmuzio JC (1999) 2-by-n fire model. Phys Rev Lett 69:1629–1632
hybrid cellular automata with regular configuration: Duff MJB, Preston K Jr (1984) Modern cellular automata:
Theory and applications. IEEE Trans Comput theory and applications. Plenum, New York
48(3):285–295 Dutching W, Vogelsaenger T (1985) Recent progress in
Chaitin G (2006) Meta Math! Vintage, New York modeling and simulation of three dimensional tumor
Chattopadhyay S, Adhikari S, Sengupta S, Pal M (2000) growth and treatment. Biosystems 18(1):79–104
Highly regular, modular, and cascadable design of cel- Elspas B (1959) The theory of autonomous linear
lular automata-based pattern classifier. IEEE Trans sequential networks. TRE Trans Circuit Theory
VLSI Syst 8(6):724–735 CT-6:45–60
Chaudhuri PP, Chowdhury DR, Nandi S, Chattopadhyay Falk H (1986) Comments on a simple cellular automata in
S (1997) Additive cellular automata: theory and appli- spin representation. Physica D 20:447–449
cations, vol 1. IEEE Computer Society Press, Los Fischer PC (1965) Generation of primes by a one-
Alamitos dimensional real time iterative array. J Assoc Comput
Chen K, Bak P, Jensen MH (1990) A deterministic critical Machin 12:388
forest-fire model. Phys Lett A 149:207–210 Flache A, Hegselmann R (1998) Understanding complex
Chin W, Cortzen B, Goldman J (2001) Linear cellular social dynamics: a plead for cellular automata based
automata with boundary conditions. Linear Algebra modeling. J Artif Soc Social Simul 1(3):1
Appl 322:193–206 Gardner M (1970) The fantastic combinations of John
Chopard B, Droz M (1998) Cellular automata modelling Conway’s new solitaire game ‘life. Sci Am
of physical systems. Cambridge University Press, 223:120–123
Cambridge Gardner M (1971) On cellular automata self-reproduction,
Chowdhury DR, Basu S, Gupta IS, Chaudhuri PP the Garden of Eden and the game of ‘life. Sci Am
(1994) Design of CAECC cellular automata based 224:112–117
error correcting code. IEEE Trans Comput Gerola H, Seiden P (1978) Stochastic star formation and
43:759–764 spiral structure of galaxies. Astrophys J 223:129–135
Chowdhury DR, Gupta IS, Chaudhuri PP (1995a) Cellular Greenberg JM, Hastings SP (1978) Spatial patterns for
automata based byte error correcting code. IEEE Trans discrete models of diffusion in excitable media. SIAM
Comput 44:371–382 J Appl Math 34(3):515–523
148 Additive Cellular Automata

Greenberg JM, Hassard BD, Hastings SP (1978) Pattern Manning FB (1977) An approach to highly integrated
formation and periodic structures in systems modeled computer-maintained cellular arrays. IEEE Trans
by reaction-diffusion equations. Bull Am Math Soc Comput C-26:536
84:1296–1327 Martin O, Odlyzko A, Wolfram S (1984) Algebraic prop-
Guan P, He Y (1986) Exact results for deterministic cellular erties of cellular automata. Commun Math Phys
automata with additive rules. J Stat Phys 93:219–258
43(3/4):463–478 Matsumoto M (1998) Simple cellular automata as
Hedlund GA (1969) Endomorphisms and automorphisms pseudorandom m-sequence generators for built-in
of the shift dynamical system. Math Syst Theory self-test. ACM Trans Modeling Comput Simul
4:320–375 8(1):31–42
Hillis WD (1984) The connection machine: a computer Moore JH, Hahn LW (2002) Cellular automata and genetic
architecture based on cellular automata. Physica algorithms for parallel problem solving in human
D 10:213–228 genetics. In: Merelo JJ, Panagiotis A, Beyer H-G
Hopcroft JE, Ullman JD (1972) Introduction to automata (eds) Lecture notes in computer science. Springer, Ber-
theory, language, and computation. Addison-Wesley, lin, pp 821–830
Reading Moraal H (2000) Graph-theoretical characterization of
Hortensius PD, McLeod RD, Card HC (1989) Parallel invertible cellular automata. Physica D 141:1–18
random number generation for VLSI systems using Moreira J, Deutsch A (2002) Cellular automata models of
cellular automata. IEEE Trans Comput tumor development: A critical review. Adv Complex
38(10):1466–1473 Syst 5(2&3):247–269
Hortensius PD, McLeod RD, Card HC (1990) Cellular Morita K (1994) Reversible cellular automata. J Inf Pro-
automata based signature analysis for built-in self-test. cess Soc Japan 35:315–321
IEEE Trans Comput C-39:1273–1283 Morita K, Ueno S (1994) Parallel generation and parsing of
Ibarra OH, Palis MA, Kim SM (1985) Fast parallel lan- array languages using reversible cellular automata. Int
guage recognition by cellular automata. Theor Comput J Pattern Recognit Artif Intell 8:543–561
Sci 41:231–246 Mrugalski G, Rajski J, Tyszer J (2000) Cellular automata-
Illichinsky A (2001) Cellular automata: a discrete universe. based test pattern generators with phase shifters. IEEE
World Scientific, Singapore Trans Comput-Aided Des 19(8):878–893
Jen E (1986a) Invariant strings and pattern recognizing Nagler J, Claussen JC (2005) 1/faspectra in elementary cel-
properties of 1D CA. J Stat Phys 43:243–265 lular automata and fractal signals. Phys Rev E 71:067103
Jen E (1986b) Global properties of cellular automata. J Stat Nandi S, Pal Chaudhuri P (1996) Analysis of periodic and
Phys 43(1/2):219–242 intermediate boundary 90/150 cellular automata. IEEE
Jen E (1988a) Cylindrical cellular automata. Commun Trans Comput 45(1):1–12
Math Phys 118:569–590 Nandi S, Kar BK, Chaudhuri PP (1994) Theory and appli-
Jen E (1988b) Linear cellular automata and recurring cation of cellular automata in cryptography. IEEE Trans
sequences in finite fields. Commun Math Phys Comput 43(12):1346–1357
119:13–28 Nishio H (1981) Real time sorting of binary numbers by a
Jen E (1989) Limit cycles in one-dimensional cellular 1-dimensional cellular automata. Kyoto University
automata. In: Stein DL (ed) Lectures in the sciences Technical Report
of complexity. Addison Wesley, Reading Oono Y, Kohmoto M (1985) A discrete model for chemical
Kari J (1990) Reversibility of 2D cellular is undecidable. turbulance. Phys Rev Lett 55:2927–2931
Physica D 45:379–385 Peitgen H-O, Richter PH (1986) The beauty of fractals:
Langer JS (1980) Instabilities and pattern formation in images of complex dynamical systems. Springer,
crystal growth. Rev Mod Phys 52:1 Berlin
Lidl R, Pilz G (1984) Applied abstract algebra. Springer, Pries W, Thanailakis A, Card HC (1986) Group properties
New York of cellular automata and VLSI applications. IEEE Trans
Lin F, Goldenfeld N (1990) Generic features of late-stage Comput C-35:1013–1024
crystal growth. Phys Rev A 42:895–903 Raghavan R (1993) Cellular automata in pattern recogni-
Lind DA (1984) Applications of ergodic theory and tion. Inf Sci 70:145–177
sofic systems to cellular automata. Physica D 10:36–44 Rosenfeld A (1979) Picture languages. Academic Press, New
Lindenmayer A, Rozenberg G (1976) Automata, lan- York
guages, development. North Holland, Amsterdam Santos RMZD, Continho S (2001) Dynamics of HIAV
Mackay AL (1976) Crystal Symmetry. Phys Bull approach: a cellular automata approach. Phys Rev
27:495–497 Lett 87(16):102–104
Madore BF, Freedman WL (1983) Computer simulations of Sarkar P (2000) A brief history of cellular automata. ACM
the Belousov-Zhabotinsky reaction. Science 222:615–616 Comput Syst 32(1):80–107
Additive Cellular Automata 149

Serra M (1990) Algebraic analysis and algorithms for Toffoli T, Margolis N (1987) Cellular automata machines: a
linear cellular automata over GF(2) and their applica- new environment for modeling. MIT Press, Cambridge
tions to digital circuit testing. Congressus Toffoli T, Margolis N (1990) Invertible cellular automata: a
Numerantium 75:127–139 review. Physica D 45:229–253
Serra M, Slater T, Muzio JC, Miller DM (1990) Analysis Tomassini M, Sipper M, Perrenoud M (2000) On the gen-
of one dimensional cellular automata and their eration of high-quality random numbers by two-
aliasing probabilities. IEEE Trans Comput-Aided Des dimensional cellular automata. IEEE Trans Comput
9:767–778 49(10):1146–1151
Sieburg HB, McCutchan JA, Clay OK, Cabalerro L, Tsalides P, York TA, Thanailakis A (1991) Pseudorandom
Ostlund JJ (1991) Simulation on HIV infection in number generation for VLSI systems using cellular
artificial immune systems. In: Gutowitz HA automata. IEEE Proc E: Comput Digit Technol
(ed) Cellular automata: theory and experiment. MIT 138:241–249
Press, Cambridge, pp 208–227 Tziones P, Tsalides P, Thanailakis A (1994) A new cellular
Sikdar BK, Ganguly N, Chaudhuri PP (2002) Design of automaton-based nearest neighbor pattern classifier and
hierarchical cellular automata for on-chip test pattern its VLSI implementation. IEEE Trans VLSI Implement
generator. IEEE Trans Comput Assist Des 2(3):343–353
21(12):1530–1539 Vitanni P (1973) Sexually reproducing cellular automata.
Smith AR III (1972) Real-time language recognition by one- Math Biosci 18:23–54
dimensional cellular automata. J Comput Syst Sci von Haeseler F, Peitgen H-O, Skordev G (1992a) Pascal’s
6:233–253 triangle, dynamical systems and attractors. Ergod The-
Sommerhalder R, van Westrhenen SC (1983) Parallel lan- ory Dyn Syst 12:479–486
guage recognition in constant time by cellular autom- von Haeseler F, Peitgen H-O, Skordev G (1992b) Linear
ata. Acta Inform 19:397–407 cellular automata, substitutions, hierarchical iterated
Sternberg SR (1980) Language and architecture for parallel function systems. In: Encarnacao JL, Peitgen H-O,
image processing. In: Gelesma ES, Kanal LN (eds) Sakas G, Englert G (eds) Fractal geometry and com-
Pattern recognition in practice. North-Holland, Amster- puter graphics. Springer, Berlin
dam, p 35 von Haeseler F, Peitgen H-O, Skordev G (1993) Cellular
Stevens JG (1999) On the construction of state diagrams automata, matrix substitution, and fractals. Ann Math
for cellular automata with additive rules. Inf Sci Artif Intell 8(3,4):345–362
115:43–59 von Haeseler F, Peitgen H-O, Skordev G (1995) Global
Stevens JG, Rosensweig RE, Cerkanowicz AE analysis of self-similar features of cellular automata:
(1993) Transient and cyclic behavior of cellular autom- selected examples. Physica D 86:64–80
ata with null boundary conditions. J Stat Phys von Haeseler F, Peitgen H-O, Skordev G (2001a)
73(1,2):159–174 Self-similar structures of rescaled evolution sets of
Sutner K (1988a) On s-automata. Complex Syst 2(1):1–28 cellular automata I. Int J Bifurc Chaos 11(4):913–926
Sutner K (1988b) Additive automata on graphs. Complex von Haeseler F, Peitgen H-O, Skordev G (2001b) Self-
Syst 2:649–661 similar structures of rescaled evolution sets of cellular
Sutner K (1991) De Bruijn graphs and linear cellular automata II. Int J Bifurc Chaos 11(4):927–941
automata. Complex Syst 5(1):19–30 von Neumann J (1963) The general and logical theory of
Sutner K (2000) Sigma-automata and Chebyschev poly- automata. In: Taub A (ed) J. von Neumann collected
nomials. Theor Comput Sci 230:49–73 works, vol 5. Pergamon, Oxford. pp 288–328
Sutner K (2001) Decomposition of additive CA. Complex von Neumann J, Burk AW (eds) (1966) Theory of self-
Syst 13(2):245–270 reproducing automata. University of Illinois Press,
Tadaki S (1994) Orbits in one-dimensional finite linear Urbana
cellular automata. Phys Rev E 49(2):1168–1173 Voorhees B (1988) Cellular automata, Pascal’s triangle,
Tadaki S, Matsufuji S (1993) Periodicity in one- and generation of order. Physica D 31:135–140
dimensional finite linear cellular automata. Prog Voorhees B (1993) Predecessors of cellular automata
Theor Phys 89(2):325–331 states: I. Additive automata. Physica D 68:283–292
Takahashi S (1990) Cellular automata and multifractals: Voorhees B (1994) A note on injectivity of additive cellular
dimension spectra of linear cellular automata. Physica automata. Complex Syst 8(3):151–159
D 45:36–48 Voorhees BH (1996) Computational analysis of one
Takahashi S (1992) Self-similarity of linear cellular autom- dimensional cellular automata. World Scientific,
ata. J Comput Syst Sci 44(1):114–140 Singapore
Thomas DM, Stevens JG, Letteiri S (2006) Characteristic Voorhees B (2008) Representations of rule 90 and related
and minimal polynomials of linear cellular automata. rules for periodic, null, and half-infinite boundary con-
Rocky Mountain J Math 36(3):1077–1092 ditions. J Cell Autom 3(1):1–25
150 Additive Cellular Automata

Willson S (1984a) Cellular automata can generate fractals. Goles E, Martenez S (1996) Dynamics of complex
Discret Appl Math 8:91–99 interacting systems. Kluwer, Dordrecht
Willson S (1984b) Growth rates and fractional dimension Goles E, Martinez S (eds) (1992) Statistical physics,
in cellular automata. Physica D 10:69–74 automata networks and dynamical systems. Kluwer,
Willson S (1987a) The equality of fractional dimension for Dordrecht
certain cellular automata. Physica D 24:179–189 Goles E, Martinez S (1994) Cellular automata, dynamical
Willson S (1987b) Computing fractional dimension of systems and neural networks. Kluwer, Dordrecht
additive cellular automata. Physica D 24:190–206 Goles E, Martinez S (eds) (1999) Cellular Automata and
Willson S (1992) Calculating growth rates and moments Complex Systems. Kluwer, Dordrecht
for additive cellular automata. Discret Appl Math Griffeath D, Moore C (eds) (2003) New constructions in
35:47–65 cellular automata. Oxford University Press, Oxford
Wolfram S (1983) Statistical mechanics of cellular autom- Gros C (2007) Complex and adaptive dynamical systems: a
ata. Rev Mod Phys 55:601–644 primer. Springer, Berlin
Wuensche A, Lesser M (1992) The global dynamics of Kier LB, Seybold PG, Cheng C-K (2005) Cellular autom-
cellular automata. Addison Wesley, Reading ata modeling of chemical systems: a testbook and lab-
Young D (1984) A local activator-inhibitor model of ver- oratory manual. Springer, Dordrecht
tebrate skin patterns. Math Biosci 72:51–58 Lafe O (2000) Cellular automata transformations: theory
and applications in multimedia compression, encryp-
tion and modeling. Kluwer, Dordrecht
Legendi T (1987) Parallel processing by cellular automata
Books and Reviews and arrays. Kluwer, Dordrecht
Adamatzky A (1994) Identification of cellular automata. Manneville P, Boccara N, Vishniac GY, Bidaux R (1990)
Taylor & Francis, London Cellular automata and modeling of complex physical
Amos M (2004) Cellular computing. Oxford University systems. Springer, New York
Press, Oxford/New York Perdang JM, Lejeune A (eds) (1994) Cellular automata:
Bandini S, Moroni L (eds) (1996) ACRI `96: proceedings prospects in astronomy and astrophysics. World Scien-
of the 2nd international conference on cellular autom- tific, Singapore
ata for research and industry. Springer, Berlin Rietman E (1989) Exploring the geometry of nature:
Bandini S, Worsch T (eds) (2000) Theory and practical computer modeling of chaos, fractals, cellular autom-
issues on cellular automata: proceedings of the 4th ata, and neural networks. McGraw-Hill, New York
international conference on cellular automata for Rothman DH, Zaleski S (2004) Lattice-gas cellular autom-
research and industry. Springer, Berlin ata: simple models of complex hydrodynamics. Cam-
Bandini S, Serra R, Liverani FS (eds) (1998) Cellular bridge University Press, Cambridge
automata: research towards industry, ACRI `98: pro- Schiff JL (2007) Cellular automata: a discrete view. Wiley-
ceedings of the 3rd international conference on cellular Interscience, New York
automata for research and industry. Springer, Sipper M (1997) Evolution of parallel cellular machines: the
London/New York cellular programming approach. Springer,
Batty M (2005) Cities and complexity: understanding cit- Berlin/New York
ies with cellular automata, agent-based models, and Sloot PMA, Chopard B, Hoekstra AG (eds) (2004) Cellular
fractals. MIT Press, Cambridge automata: proceedings of the 6th international confer-
Boccara N, Goles E, Martenez S (1993) Cellular automata ence on cellular automata for research and industry.
and cooperative systems. Kluwer, Dordrecht Springer, Berlin/New York
Crutchfield JP, Hanson JE (1999) Computational mechan- Tomassini M, Chopard B (eds) (2002) Cellular automata:
ics of cellular processes. University of California Press, proceedings of the 5th international conference on cel-
Berkeley lular automata for research and industry. Springer,
Delorme M, Mazoyer J (eds) (1998) Cellular automata: a Berlin/New York
parallel model. Kluwer, Dordrecht Wolf-Gladrow DA (2000) Lattice-gas cellular automata
Deutsch A, Dormann S (2004) Cellular automata modeling and lattice Boltzmann models. Springer, Berlin
of biological pattern formation: characterization, appli- Wolfram S (1994) Cellular automata and complexity.
cations, and analysis. Springer, Berlin Addison Wesley, Reading
El Yacoubi S, Chopard B, Bandini S (eds) (2006) Cellular Wolfram S (2002) A new kind of science. Wolfram Media,
automata: proceedings of the 7th international confer- Champaign
ence on cellular automata for research and industry. Yang T (2001) Cellular image processing. Nova,
Springer, Berlin/Heidelberg Huntington
Additive Cellular Automata 151

Websites http://www.ddlab.com. An excellent site; it provides access to

http://cell-auto.com/links/. Gives many links to other sites the Discrete Dynamics Lab program, a valuable asset in
on cellular automata work on cellular automata and random Boolean networks
http://cellular.ci.ulsa.mx. Provides access to a number of http://www.theory.org/complexity/cdpt/html/node4.html.
worthwhile unpublished papers and a number of useful Provides reviews of theoretical aspects of cellular
references automata
 application of the CA rules by implementing
Cellular Automata with memory capabilities in cells and links when
Memory topology is dynamic.

Ramón Alonso-Sanz Definition

Technical University of Madrid, ETSIAAB
(Estadística, GSC), Madrid, Spain Cellular Automata (CA) are discrete, spatially
explicit extended dynamic systems. A CA system
is composed of adjacent cells characterized by an
internal state whose value belongs to a finite set.
The updating of these states is made simulta-
Article Outline neously according to a common local transition
rule involving only a neighborhood of each cell.
ðT Þ
Glossary Thus, if si is taken to denote the value of cell i at
Definition time step T, the site values nevolve o by iteration
 of
ðT þ1Þ ðT Þ
Introduction the mapping: si ¼ f sj  N i , where
Average Memory f is an arbitrary function which specifies the
Other Memories cellular automaton rule operating on N i , i.e. the
CA with Three Statesk set of cells in the neighborhood of the cell i. Last
Reversible CA but not least, standard CA are ahistoric
Heterogeneous CA (memoryless): i.e., the new state of a cell depends
Structurally Dynamic CA on the neighborhood configuration only at the
Memory in Other Discrete Time Contexts preceding time step.
Future Directions The standard framework of CA can be
Bibliography extended by the consideration of all past states
(history) in the application of the CA
Glossary rules by implementing n omemory  capabilities
ðTþ1Þ ðT Þ ðT Þ
in cells: si ¼ f sj  N i , with sj ¼ s
Cellular automata Cellular Automata (CA) are  
ð1Þ ðT 1Þ ðT Þ
discrete, spatially explicit extended dynamic sj , . . . , sj , sj being a state function of
systems composed of adjacent cells character- the series of states of the cell j up to time-step T.
ized by an internal state whose value belongs to Thus, in CA with memory here: while the transition
a finite set. The updating of these states is made functions ’ of the CA remain unaltered, historic
simultaneously according to a common local memory of all past iterations is retained by featur-
transition rule involving only a neighborhood ing each cell by a summary of its past states.
of each cell. The memory mechanism considered here is
Memory Standard CA are ahistoric different from that of other CA with memory
(memoryless): i.e., the new state of a cell reported in the literature (e.g., p. 7 in Adamatzky
depends on the neighborhood configuration (1994), p. 43 in Ilachinski (2000), p. 118 in Wol-
only at the preceding time step. The standard fram (1984)). Typically, these are higher-order-in-
framework of CA can be extended by the con- time rules that incorporate memory into the
sideration of all past states (history) in the transition rule, determining the new configuration

# Springer-Verlag 2009 153

A. Adamatzky (ed.), Cellular Automata,
https://doi.org/10.1007/978-1-4939-8700-9_55
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_55
154 Cellular Automata with Memory

in terms of the configurations at previous mean of the states up to T is computed for every
time-steps. Thus,
n in
o secondnorder o in time rules: cell i:
ðTþ1Þ ðT Þ ðT1Þ
si ¼ F sj  N i , sj  N i . CA
  sðT Þ þ PT 1 aTt sðtÞ
with memory in cells are cited in Wuensche and ðT Þ ð1Þ ðT Þ
si , . . . , si ¼ i i
mi Pt¼1
1 T t
Lesser (1992), but just to state that “CA with 1 þ Tt¼1 a
memory in cells would result in a qualitatively ðT Þ
oi
different behavior”. Some authors (Wolf-Gladrow 
2000), define rules with memory as those OðT Þ
with dependence in f on the state of the cell
to be updated. So, one-dimensional rules with and then, the featuring states s are obtained by
no memory,
ðTþ1Þ
si ¼ rounding the m ones to 1 for m > 0.5 and to 0 for
 take the form: m < 0.5. If m is exactly 0.5, then the last state is
ðT Þ ðT Þ
f si1 , siþ1 . Our use of the term memory is ðT Þ ðT Þ
assigned (si ¼ si ). This memory mechanism
not any of these. is accumulative in their demand of knowledge of
past history: to calculate the memory charge
ðT Þ
oi oit is not necessary to know the whole
n
Introduction ðt Þ
si series, while it can be sequentially calcu-
ðT Þ ðT 1Þ ðT Þ
lated as: oi ¼ aoi þ si . It is, O(T) =
As a simple example, in the two-dimensional,
(aT  1)/(a  1).
two-state automaton labeled ahistoric in Fig. 1,
The choice of the memory factor a simulates
a cell becomes (or remains) alive if any cell in its
the long-term or remnant memory effect:
nearest neighborhood is alive, but becomes
the limit case a = 1 corresponds to memory
(or remains) dead on the contrary case. The ini-
with equally weighted records (full memory,
tial single perturbation in Fig. 1 spreads as fast as
equivalent to mode if k = 2), whereas a  1
possible, i.e. at the speed of light.
intensifies the contribution of the most
The lower series of patterns in Fig. 1 shows
recent states and diminishes the contribution
the effect of featuring cells by their most
ðT Þ of the past ones (short term working
frequent state,i.e. mode memory: si ¼ mode memory). The choice a = 0 leads to the ahistoric
ð 1Þ ðT Þ ðT Þ ðT Þ
si , . . . , si (with si ¼ si in case of a tie) model.
on the speed of light. Memory has a character- In the most unbalanced scenario up to T, i.e.:
ð1Þ ðT1Þ ðT Þ
istic inertial effect. si ¼ . . . ¼ si 6¼ si , it is:

1 a1 1
Average Memory mð0,0, . . . ,0,1Þ ) T
¼ ¼
2 a 1 2
1 aT  a 1
mð1,1, . . . ,1,0Þ ¼ ) T ¼ :
Cells can be featured by a weighted mean value 2 a 1 2
of all their previous states through powers of
some parameter a  [0, 1] acting as a memory Thus, memory is only operative if a is greater
factor. Thus, at every time-step T, the weighted than a critical aT that verifies:

Cellular Automata with

Memory, Fig. 1 The
speed of light starting from
a single active cell
Cellular Automata with Memory 155

aTT  2aT þ 1 ¼ 0, (1) the initial one; afterward both evolutions are fairly
similar up to T = 18, but at this time step both
in which case cells will be featured at T with state kinds of patterns notably differs, and since then
values different to the last one. Initial operative the evolution patterns in Fig. 3 notably diverge
values are: a3 = 0.61805, a4 = 0.5437. When from the ahistoric ones.
T ! 1, Eq. 1 becomes: 2a1 + 1 = 0, thus, in To give consideration to previous states
the k = 2 scenario, a-memory is not effective if (historic memory) in two-dimensional CA tends
a  0.5. to confine the disruption generated by a single live
cell. As a rule, full memory tends to generate
oscillators, and less historic information retained,
A Worked Example: The Parity Rule i.e. smaller a value, implies an approach to the
The so-called parity rule states: cell alive if the ahistoric model in a rather smooth form. But the
number of neighbors is odd, dead on the contrary transition which decreases the memory factor
case. Figure 2 shows the effect of memory on the from a = 1.0 (full memory) to a = 0.5 (ahistoric
parity rule starting from a single live cell in the model), is not always regular, and some kind of
Moore neighborhood. In accordance with the erratic effect of memory can be traced.
above given values of a3 and a4: (i) The pattern The inertial (or conservating) effect of memory
at T = 4 is the ahistoric one if a  0.6, altered dramatically changes the dynamics of the semi-
when a  0.7, and (ii) the patterns at T = 5 for totalistic LIFE rule. Thus, (i) the vividness that
a = 0.54 and a = 0.55 differ. some small clusters exhibit in LIFE, has not
Not low levels of memory tend to freeze the been detected in LIFE with memory. In particular,
dynamics since the early time-steps, e.g. over 0.54 the glider in LIFE does not glide with memory,
in Fig. 2. In the particular case of full memory but stabilizes very close to its initial position as the
small oscillators of short range in time are fre-
tub , (ii) as the size of a configuration
quently generated, such as the period-two oscilla-
tor that appears as soon as at T = 2 in Fig. 2. The increases, often live clusters tend to persist with
group of evolution patterns shown in the a higher number of live cells in LIFE with mem-
[0.503,0.54] interval of a variation of Fig. 2, is ory than in the ahistoric formulation, (iii) a single
rather unexpected to be generated by the parity mutant appearing in a stable agar can lead to its
rule, because they are too sophisticated for this destruction in the ahistoric model, whereas its
simple rule. On the contrary, the evolution pat- effect tends to be restricted to its proximity with
terns with very small memory, a = 0.501, resem- memory (Alonso-Sanz et al. 2001a).
ble those of the ahistoric model in Fig. 2. But this
similitude breaks later on, as Fig. 3 reveals: from One-Dimensional CA
T = 19, the parity rule with minimal memory Elementary rules are one-dimensional, two-state
evolves producing patterns notably different to rules operating on nearest neighbors. Following
the ahistoric ones. These patterns tend to be Wolfram’s notation, these rules are characterized
framed in squares of size not over T  T, whereas by a sequence of binary values (b) associated with
in the ahistoric case, the patterns tend to be framed each of the eight possible triplets
in 2T  2T square regions, so even minimal mem-
ory induces a very notable reduction in the ðT Þ ðT Þ ðT Þ 
affected cell area in the scenario of Fig. 2. The ðsi1 ,si ,siþ1 :
patterns of the featured cells tend not to be far to 111 110 101 100 011 010 001 000
the actual ones, albeit examples of notable diver- b1 b2 b3 b4 b5 b6 b7 b8
gence can be traced in Fig. 2. In the particular case
of the minimal memory scenario of Fig. 2, that of The rules are conveniently specified by their
P
a = 0.501, memory has no effect up to T = 9, rule number R ¼ 8i¼1 bi 28i . Legal rules are
when the pattern of featured live cells reduces to reflection symmetric (b5 = b2, b7 = b4), and
156 Cellular Automata with Memory

Cellular Automata with Memory, Fig. 2 The 2D parity rule with memory up to T = 15

quiescent (b8 = 0), restrictions that leave 32 pos- Patterns are shown up to T = 63, with the memory
sible legal rules. factor varying from 0.6 to 1.0 by 0.1 intervals, and
Figure 4 shows the spatio-temporal patterns of adopting also values close to the limit of its effec-
legal rules affected by memory when starting from tivity: 0.5. As a rule, the transition from the
a single live cell (Alonso-Sanz and Martin 2002a). a = 1.0 (fully historic) to the ahistoric scenario
Cellular Automata with Memory 157

Cellular Automata with Memory, Fig. 3 The 2D parity rule with a = 0.501 memory starting from a single site live cell
up to T = 55

is fairly gradual, so that the patterns become more decrease in the width of evolving patterns as a is
expanded as less historic memory is retained higher, but they do not reach extinction.
(smaller a). Rules 50, 122, 178,250, 94, and Figure 5 shows the effect of memory on
222,254 are paradigmatic of this smooth evolu- legal rules when starting at random: the values
tion. Rules 222 and 254 are not included in Fig. 4 of sites are initially uncorrelated and chosen at
as they evolve as rule 94 but with the inside of random to be 0 (blank) or 1 (gray) with probabil-
patterns full of active cells. Rules 126 and ity 0.5. Differences in patterns resulting from
182 also present a gradual evolution, although reversing the center site value are shown as
their patterns with high levels of memory models black pixels. Patterns are shown up to T = 60, in
hardly resemble the historic ones. Examples with- a line of size 129 with periodic boundary condi-
out a smooth effect of memory are also present in tions imposed on the edges. Only the nine legal
Fig. 4: (i) rule 150 is sharply restrained at a = 0.6, rules which generate non-periodic patterns in the
(ii) the important rule 54 extinguish in [0.8,0.9], ahistoric scenario are significantly affected by
but not with full memory, (iii) the rules in the memory. The patterns with inverted triangles
group {18,90,146,218} become extinct from dominate the scene in the ahistoric patterns of
a = 0.501. Memory kills the evolution for these Fig. 5, a common appearance that memory tends
rules already at T = 4 for a values over a3 (thus to eliminate.
over 0.6 in Fig. 4): after T = 3 all the cells, even History has a dramatic effect on Rule 18. Even
the two outer cells alive at T = 3, are featured as at the low value of a = 0.6, the appearance of its
dead, and (iv) rule 22 becomes extinct for spatio-temporal pattern fully changes: a number
a = 0.501, not in 0.507, 0.6, and 0.7, again extin- of isolated periodic structures are generated, far
guish at 0.8 and 0.9, and finally generate an oscil- from the distinctive inverted triangle world of the
lator with full memory. It has been argued that ahistoric pattern. For a = 0.7, the live structures
rules 18, 22, 122, 146 and 182 simulate Rule 90 in are fewer, advancing the extinction found in
that their behavior coincides when restricted to [0.8,0.9]. In the fully historic model, simple peri-
certain spatial subsequences. Starting with a sin- odic patterns survive.
gle site live cell, the coincidence fully applies in Rule 146 is affected by memory in much the
the historic model for rules 90, 18 and 146. Rule same way as Rule 18 because their binary codes
22 shares with these rules the extinction for high a differ only in their b1 value. The spatio-temporal
values, with the notable exception of no extinction of rule 182 and its equivalent Rule 146 are remi-
in the fully historic model. Rules 122 and niscent, though those of Rule 182 look like a
182 diverge in their behavior: there is a gradual negatives photogram of those of Rule 146.
158 Cellular Automata with Memory

Fig. 4 (continued)
Cellular Automata with Memory 159

Cellular Automata with Memory, Fig. 4 Elementary, legal rules with memory from a single site live cell

The effect of memory on rule 22 and the com- sophisticated in their basis. A notable discontinu-
plex rule 54 is similar. Their spatio-temporal pat- ity is found for both rules ascending in the value of
terns in a = 0.6 and a = 0.7 keep the essential of the memory factor: in a = 0.8 and a = 0.9 only a
the ahistoric, although the inverted triangles few simple structures survive. But unexpectedly,
become enlarged and tend to be more the patterns of the fully historic scenario differ
160 Cellular Automata with Memory

markedly from the others, showing a high degree persistent propagating structures (solitons).
of synchronization. Rules 122 and 126, evolve in a similar form,
The four remaining chaotic legal rules showing a high degree of synchronization in the
(90, 122, 126 and 150) show a much smoother fully historic model.
evolution from the ahistoric to the historic sce- As a rule, the effect of memory on the differences
nario: no pattern evolves either to full extinction in patterns (DP) resulting from reversing the value
or to the preservation of only a few isolated of its initial center site is reminiscent of that on the

Fig. 5 (continued)
Cellular Automata with Memory 161

spatio-temporal patterns, albeit this very much (r0 = 0.5). The simulation is implemented for the
depends on the actual simulation run. In the case same rules as in Fig. 4, but with notably wider
of rule 18 for example, damage is not present in the lattice: N = 500. A visual inspection of the plots in
simulation of Fig. 5. The group of rules 90, 122, Fig. 6, ratifies the general features observed in the
126 and 150 shows a, let us say canonical, fairly patterns in Fig. 5 regarding density. That also stands
gradual evolution from the ahistoric to the historic for damage spreading: as a rule, memory depletes
scenario, so that the DP appear more constrained as the damaged region.
more historic memory is retained, with no extinction In one-dimensional r = 2 CA, the value of
for any a value. Figure 6 shows the evolution of the a given site depends on values of the nearest
fraction rT of sites with value 1, starting at random and next-nearest neighbors. Totalistic r = 2

Fig. 5 (continued)
162 Cellular Automata with Memory

Cellular Automata with Memory, Fig. 5 Elementary, legal rules with memory starting at random

ðTþ1Þ
rules with memory have the form:  si ¼f to another, it is natural to generalize the determin-
ðT Þ ðT Þ ðT Þ ðT Þ ðT Þ istic CA framework to the probabilistic scenario. In
si2 þ si1 þ si þ siþ1 þ siþ2 . The effect of
the elementary scenario, the b are replaced by
memory on these rules follows the way traced probabilities
in the r = 1 context, albeit with a rich casuistic
studied in Alonso-Sanz (2007b).

Probabilistic CA
So far the CA considered are deterministic. In order
to study perturbations to deterministic CA as well As in the deterministic scenario, memory can
as transitional changes from one deterministic CA be embedded in probabilistic CA (PCA) by featuring
Cellular Automata with Memory 163

Cellular Automata with Memory, Fig. 6 Evolution of the density starting at random in elementary legal rules. Color
code: blue ! full memory, black!a = 0.8, red ! ahistoric model

cells by a summary ofpast states si instead of by their

 which the larger the value of c, the more heavily is
ðT þ1Þ ðT Þ ðT Þ ðT Þ
last state si: p ¼ P si ¼ 1=si1 , si , siþ1 . the recent past taken into account, and
consequently the closer the scenario to the ahistoric
Again, memory is embedded into the cha-
model (Alonso-Sanz and Martin 2003, 2004b). Both
racterization of cells but not in the construction of
weights allow for integer-based arithmetics (à la
the stochastic transition rules, as done in the canon-
CA) comparing 2o(T) to 2O(T) to get the featuring
ical approach to PCA. We have explored the effect
states s (a clear computational advantage over the
of memory on three different subsets (0, p2, 0, p4,
a-based model), and are accumulative in respect to
p2, 0, p4, 0), (0, 0, p3, 1, 0, p6, 1, 0), and (p1, p2, ðT Þ ðT 1Þ ðT Þ ðT Þ ðT1Þ
p1, p2, p2, 0, p2, 0) in Alonso-Sanz (2005a). charge: oi ¼ oi þ T c si , oi ¼ oi þ
T ðT Þ
c si . Nevertheless, they share the same drawback:
powers explode, at high values of t, even for c = 2.
Limited trailing memory would keep mem-
Other Memories
ory of only the last t states. This is implemented
ðT Þ
A number of average-like memory mechanisms can in the context of average memory as: oi ¼
PT ðt Þ
readily be proposed by using weights different to t¼⊤ dðt Þsi , with ⊤ = max (1, T  t + 1). Lim-
that implemented in the a-memory mechanism: iting the trailing memory would approach the
d(t) = aTt. Among the plausible choices of d, we model to the ahistoric model (t = 1). In the geo-
mention the weights d(t) = tc and d(t) = ct, c  ℕ, in metrically discounted method, such an effect is
164 Cellular Automata with Memory

Cellular Automata with Memory, Fig. 7 Legal (first row of patterns) and quiescent asymmetric elementary rules
significantly affected by the mode of the three last states of memory

more appreciable when the value of a is high, 184 and 226. The patterns generated by rules
whereas at low a values (already close to the 6 and 14 are rectified (in the sense of having
ahistoric model when memory is not limited) the the lines in the spatio-temporal pattern
effect of limiting the trailing memory is not so slower slope) by memory in such a way that the
important. In the k = 2 context, if t = 3, provided total number of live cells in the historic and
that a > a3 = 0.61805, the memory mechanism ahistoric spatio-temporal patterns is the same.
turns out to be that of selecting
 the mode of the Again, the historic patterns of the remaining
ðT Þ ðT 2Þ ðT Þ ðT 1Þ
last three states:si ¼ mode si , s i , si , rules in Fig. 7 seem, as a rule, like the
ahistoric ones compressed (Alonso-Sanz and
i.e. the elementary rule 232.
Martin 2005).
Figure 7 shows the effect of this kind of
Elementary rules (ER, noted f) can in turn act
memory on legal rules. As is known, history has
as memory rules:
a dramatic effect on Rules 18, 90, 146 and
218 as their pattern dies out as early as at T = 4.  
ðT Þ ðT 2Þ ðT Þ ðT1Þ
The case of Rule 22 is particular: two branches si ¼ f si , s i , si
are generated at T = 17 in the historic model;
the patterns of the remaining rules in the Figure 8 shows the effect of ER memories up to
historic model are much reminiscent of the ahis- R = 125 on rule 150 starting from a single site live
toric ones, but, let us say, compressed. Figure 7 cell up to T = 13. The effect of ER memories with
shows also the effect of memory on some relevant R > 125 on rule 150 as well as on rule 90 is shown
quiescent asymmetric rules. Rule 2 shifts in Alonso-Sanz and Martin (2006a). In the latter
a single site live cell one space at every time- case, complementary memory rules (rules whose
step in the ahistoric model; with the pattern rule number adds 255) have the same effect on rule
dying at T = 4. This evolution is common to all 90 (regardless of the role played by the three last
rules that just shift a single site cell without states in f and the initial configuration). In the
increasing the number of living cells at T = 2, ahistoric scenario, Rules 90 and 150 are linear
this is the case of the important rules (or additive): i.e., any initial pattern can be
Cellular Automata with Memory 165

Cellular Automata with Memory, Fig. 8 The Rule 150 with elementary rules up to R = 125 as memory
166 Cellular Automata with Memory

Cellular Automata with Memory, Fig. 9 The parity rule with elementary rules as memory. Evolution from T = 4  15
in the Neumann neighborhood starting from a singe site live cell

decomposed into the superposition of patterns from CA with Three Statesk

a single site seed. Each of these configurations can
be evolved independently and the results super- This section deals with CAwith three possible values
posed (module two) to obtain the final at each site (k = 3), noted {0, 1, 2}, so the rounding
complete pattern. The additivity of rules 90 and mechanism is implemented by comparing the
150 remains in the historic model with linear mem- unrounded weighted mean m to the hallmarks 0.5
ory rules. and 1.5, assigning the last state in case on an equality
Figure 9 shows the effect of elementary rules on to any of these values. Thus,
the 2D parity rule with von Neumann neighborhood
from a singe site live cell. This figure shows patterns
sT ¼ 0 if mT < 0:5, sT ¼ 1 if 0:5 < mT < 1:5,
from T = 4, being the three first patterns: .
The consideration of CA rules as memory induces a sT ¼ 2 if mT > 1:5,
fairly unexplored explosion of new patterns. and sT ¼ sT if mT ¼ 0:5 or mT ¼ 1:5:
Cellular Automata with Memory 167

Cellular Automata with Memory, Fig. 10 Parity k = 3 rules starting from a single s = 1 seed. The red cells are at state
1, the blue ones at state

In the most unbalanced cell dynamics, historic retained (smaller a). This characteristic inhibition of
memory takes effect after time step T only if a > aT, growth effect of memory is traced on rules 300 and
with 3aTT  4aT þ 1 ¼ 0 , which in the temporal 543 in Fig. 10, (ii) the transition from the fully
limit becomes 4a + 1 = 0 , a = 0.25. historic to the ahistoric scenario tends to be gradual
In general, in CA with k states (termed from in regard to the amplitude of the spatio-temporal
0 to k  1), the characteristic equation at T is patterns, although their composition can differ nota-
ð2k  3ÞaTT  ð2k  1ÞaT þ 1 ¼ 0, which becomes bly, even at close a values, (iii) in contrast to the
2(k  1)a + 1 = 0 in the temporal limit. It is then two-state scenario, memory fires the pattern of some
concluded that memory does not affect the scenario three-state rules that die out in the ahistoric model,
if a  a(k) = 1/(2(k  1)).  and no rule with memory dies out. Thus, the effect
ðT þ1Þ ðT Þ of memory on rules 276, 519, 303 and 546 is
We study first totalistic rules:si ¼ f si1 þ
ðT Þ ðT Þ somewhat unexpected: they die out at a  0.3 but
si þ siþ1 Þ , characterized by a sequence of ter- at a = 0.4 the pattern expands, the expansion being
nary values (bs) associated with each of the seven inhibited (in Fig. 10) only at a  0.8. This activation
possible values of the sum (s) of the neighbors: under memory of rules that die at T = 3 in the
(b6, b5, b4, b3P, b2, b1, b0), with associated rule ahistoric model is unfeasible in the k = 2 scenario.
number R ¼ 6s¼0 bs 3s  ½0,2186 : The features in the evolving patterns starting
Figure 10 shows the effect of memory on quies- from a single seed in Fig. 10 are qualitatively
cent (b0 = 0) parity rules, i.e. rules with b1, b3 and reflected starting at random as shown with rule
b5 non null, and b2 = b4 = b6 = 0. Patterns are 276 in Fig. 11, which is also activated (even at
shown up to T = 26. The pattern for a = 0.3 is a = 0.3) when starting at random. The effect of
shown to test its proximity to the ahistoric one average memory (a and integer-based models,
(recall that if a  0.25 memory takes no effect). unlimited and limited trailing memory, even t = 2)
Starting with a single site seed it can be concluded, and that of the mode of the last three states has been
regarding proper three-state rules such as those in studied in Alonso-Sanz and Martin (2004b).
Fig. 10, that: (i) as an overall rule the patterns When working with more than three states, it
become more expanded as less historic memory is is an inherent consequence of averaging the
168 Cellular Automata with Memory

Cellular Automata with Memory, Fig. 11 The k ¼ 3,R ¼ 276 rule starting at random

tendency to bias the featuring state to the mean Figure 12 shows the effect of mode of the last
value: 1. That explains the redshift in the previ- three time steps memory on the defensive-
ous figures. This led us to focus on a much more inhibition CA rule with the Moore neighborhood,
fair memory mechanism: the mode, in what fol- starting from a simple configuration. At T = 3 the
lows. Mode memory allows for manipulation of outer excited cells in the actual pattern are not
pure symbols, avoiding any computing/ featured as excited but as resting cells (twice
arithmetics. resting versus one excited), and the series of evolv-
In excitable CA, the three states are featured: ing patterns with memory diverges from the ahis-
resting 0, excited 1 and refractory 2. State transi- toric evolution at T = 4, becoming less expanded.
tions from excited to refractory and from refrac- Again, memory tends to restrain the evolution.
tory to resting are unconditional, they take place The effect of memory on the beehive rule, a
independently on a cell’s neighborhood state: totalistic two-dimensional CA rule with three
ðT Þ ðT þ1Þ ðT Þ ðTþ1Þ
si ¼ 1 ! si ¼ 2, si ¼ 2 ! si ¼ 0. states implemented in the hexagonal tessellation
In Alonso-Sanz and Adamatzky (2008) the exci- (Wuensche 2005) has been explored in Alonso-
tation rule adopts a Pavlovian phenomenon of Sanz (2006b).
defensive inhibition: when strength of stimulus
applied exceeds a certain limit the system ‘shuts Reversible CA
down’, this can be naively interpreted as an inbuilt
protection of energy loss and exhaustion. To simu- The second-order in time implementation based on
late the phenomenon of defensive inhibition we the subtraction modulo
adopt interval excitation rules (Adamatzky 2001),  of the number of states
ðT þ1Þ ðT Þ ðT 1Þ
(noted ): si ¼ f sj  N i si , read-
and a resting cell becomes excited only if one or two  
ðT 1Þ ðT Þ ðT þ1Þ
of its neighbors
ðT Þ
si ¼ 0 ! si ¼ 1
ðT Þ ily reverses as: si ¼ f sj  N i si .
P  are excited:

ðT Þ To preserve the reversible feature, memory has to be
if j  N i sj ¼ 1  f1,2g (Adamatzky and
endowed only in the pivotal  component  of the rule
Holland 1998). ðT 1Þ ðT Þ ðTþ1Þ
transition, so: si ¼ f sj  N i si .
Cellular Automata with Memory 169

Cellular Automata with

Memory, Fig. 12 Effect
of mode memory on the
defensive inhibition
CA rule

For reversing from T it is necessary to know speed of diffusion of the area affected are notably
ðT Þ ðT þ1Þ ðT Þ
not only si and si but also oi to be reduced, even by minimal memory (a = 0.501),
compared to O(T), to obtain: (iii) high levels of memory tend to freeze the
dynamics since the early time-steps.
8
ðT Þ We have studied the effect of memory in the
>
<0 if 2oi < OðT Þ
ðT Þ reversible formulation of CA in many scenarios,
si ¼ sði Tþ1Þ if
ðT Þ
2oi ¼ OðT Þ
>
: ðT Þ e.g., totalistic, k = r = 2 rules (Alonso-Sanz
1 if 2oi > OðT Þ: 2004a), or rules with three states (Alonso-Sanz
and Martin 2004b).
Then to progress
 in the reversing, to obtain Reversible systems are of interest since they
ðT 1Þ ðT 1Þ
si ¼ round oi =OðT  1Þ , it is neces- preserve information and energy and allow unam-
 
ðT 1Þ ðT Þ ðT Þ
sary to calculate oi ¼ oi  si =a. But biguous backtracking. They are studied in com-
in order to avoid dividing by the memory factor puter science in order to design computers which
(recall that operations with real numbers are not would consume less energy (Toffoli and Margolus
exact in computer arithmetic), it is preferable 1987). Reversibility is also an important issue in
ðT1Þ ðT Þ ðT Þ
to work with gi ¼ oi  si , and to com- fundamental physics (Fredkin 1990; Margolus
P 1 Tt
pare these values to GðT  1Þ ¼ Tt¼1 a . This 1984; Toffoli and Margolus 1990; Vichniac 1984).
leads to: Geraldt’t Hooft, in a speculative paper (Hooft
8 1988), suggests that a suitably defined determinis-
ðT1Þ
>
<0 if 2gi < GðT  1Þ tic, local reversible CA might provide a viable
ðT 1Þ ðT Þ ðT1Þ
si ¼ si if 2gi ¼ GðT  1Þ formalism for constructing field theories on a
>
: ðT1Þ
1 if 2gi > GðT  1Þ: Planck scale. Svozil (1986) also asks for changes
in the underlying assumptions of current field the-
ðT tÞ ðT tþ1Þ ðT tþ1Þ ories in order to make their discretization appear
In general: gi ¼ gi  at1 si ,
more CA-like. Applications of reversible CA with
GðT  tÞ ¼ GðT  t þ 1Þ  a . t1
memory in cryptography are being scrutinized
Figure 13 shows the effect of memory on the
(Alvarez et al. 2005; Martin del Rey et al. 2005).
reversible parity rule starting from a single site live
cell, so the scenario of Figs. 2 and 3, with the
reversible qualification. As expected, the simula-
tions corresponding to a = 0.6 or below shows the Heterogeneous CA
ahistoric pattern at T = 4, whereas memory leads to
a pattern different from a = 0.7, and the pattern at CA on networks have arbitrary connections, but, as
T = 5 for a = 0.54 and a = 0.55 differ. Again, in the proper CA, the transition rule is identical for all
reversible formulation with memory, (i) the config- cells. This generalization of the CA paradigm
uration of the patterns is notably altered, (ii) the addresses the intermediate class between CA and
170 Cellular Automata with Memory

Cellular Automata with Memory, Fig. 13 The reversible parity rule with memory

Boolean networks (BN, considered in the following lower clustering coefficients and shorter average
section) in which, rules may be different at each site. path length between nodes commonly known as
In networks two topological ends exist, random small world property. On the other hand, regular
and regular networks, both display totally opposite graphs, have a large average path length between
geometric properties. Random networks have nodes and high clustering coefficients.
Cellular Automata with Memory 171

Cellular Automata with Memory, Fig. 14 The parity the ahistoric model (red) and memory models of a levels:
rule with four inputs: effect of memory and random 0.6,0.7.0.8, 0.9 (dotted) and 1.0 (blue)
rewiring. Distance between two consecutive patterns in

In an attempt to build a network with character- particularly when the degree of rewiring is
istics observed in real networks, a large clustering high. With full memory, quasi-oscillators tend to
coefficient and a small world property, Watts and appear. As a rule, the higher the curve the lower the
Strogatz (WS, (Watts and Strogatz 1998)) proposed memory factor a, but in the particular case of a
a model built by randomly rewiring a regular lat- regular lattice (and lattice with 10% of rewiring),
tice. Thus, the WS model interpolates between the evolution of the distance in the full
regular and random networks, taking a single new memory model turns out rather atypical, as it is
parameter, the random rewiring degree, i.e.: the maintained over some memory models with lower
probability that any node redirects a connection, a parameters.
randomly, to any other. The WS model displays Figure 15 shows the evolution of the damage
the high clustering coefficient common to regular spread when reversing the initial state of the 3  3
lattices as well as the small world property (the central cells in the initial scenario of Fig. 14. The
small world property has been related to faster fraction of cells with the state reversed is plotted
flow in the information transmission). The long- in the regular and 10% of rewiring scenarios. The
range links introduced by the randomization proce- plots corresponding to higher rates of rewiring are
dure dramatically reduce the diameter of the net- very similar to that of the 10% case in Fig. 15.
work, even when very few links are rewired. Damage spreads fast very soon as rewiring is
Figure 14 shows the effect of memory and topol- present, even in a short extent.
ogy on the parity rule with four inputs in a lattice of
size 65  65 with periodic boundary conditions, Boolean Networks
starting at random. As expected, memory depletes In Boolean Networks (BN, (Kauffman 1993)),
the Hamming distance between two consecutive instead of what happens in canonical CA, cells
patterns in relation to the ahistoric model, may have arbitrary connections and rules may be
172 Cellular Automata with Memory

Cellular Automata with Memory, Fig. 15 Damage up to T = 100 in the parity CA of Fig. 14

Cellular Automata with Memory, Fig. 16 Relative Hamming distance between two consecutive patterns. Boolean
network with totalistic, K = 4 rules in the scenario of Fig. 14

different at each
P site. Working  with totalistic rules: over some of those memory models with lower a
ðTþ1Þ ðT Þ
si ¼ fi j  N i sj . parameters. The relative Hamming distance
between the ahistoric patterns and those of historic
The main features on the effect of memory in
rewiring tends to be fairly constant around 0.3, after
Fig. 14 are preserved in Fig. 16: (i) the ordering of
a very short initial transition period.
the historic networks tends to be stronger with a
Figure 17 shows the evolution of the damage
high memory factor, (ii) with full memory, quasi-
when reversing the initial state of the 3  3 central
oscillators appear (it seems that full memory tends to
cells. As a rule in every frame, corresponding to
induce oscillation), (iii) in the particular case of the
increasing rates of random rewiring, the higher the
regular graph (and a lesser extent in the networks
curve the lower the memory factor a. The damage
with low rewiring), the evolution of the full memory
vanishing effect induced by memory does result
model turns out rather atypical, as it is maintained
apparently in the regular scenario of Fig. 17, but
Cellular Automata with Memory 173

Cellular Automata with Memory, Fig. 17 Evolution of the damage when reversing the initial state of the 3  3 central
cells in the scenario of Fig. 16

only full memory controls the damage spreading conventional CA. This means that given certain
when the rewiring degree is not high, the dynamics conditions, specified by the link transition rules,
with the remaining a levels tend to the damage links between rules may be created and destroyed;
propagation that characterizes the ahistoric model. the neighborhood of each cell is dynamic, so, state
Thus, with up to 10% of connections rewired, full and link configurations of an SDCA are both
memory notably controls the spreading, but this dynamic and continually interacting.
control capacity tends to disappear with a higher If cells are numbered 1 to N, their connectivity
percentage of rewiring connections. In fact, with is specified by an N  N connectivity matrix in
rewiring of 50% or higher, neither full memory which li j = 1 if cells i and j are n connected;o
seems to be very effective in altering the final rate ðT Þ
0 otherwise. So, now: N i ¼ j=lij ¼ 1
ðT Þ

of damage, which tends to reach a plateau around  

ðT þ1Þ ðT Þ ðT Þ
30% regardless of scenario. A level notably coinci- and si ¼ f sj  N i . The geodesic dis-
dent with the percolation threshold in site percola- tance between two cells i and j, dij, is defined as
tion in the simple cubic lattice, and the critical point the number of links in the shortest path between i
for the nearest neighbor Kaufmann model on the and j. Thus, i and j are direct neighbors if dij = 1,
square lattice (Stauffer and Aharony 1994): 0.31. and are next-nearest
n neighbors
o if dij = 2, so
ðT Þ ðT Þ
NN i ¼ j=dij ¼ 2 . There are two types of
link transition functions in an SDCA: couplers
Structurally Dynamic CA and decouplers, the former add new links, the
latter remove links. The coupler and decoupler
Structurally dynamic cellular automata (SDCA) ðT þ1Þ
 link transition rule: lij
set determines the ¼
were suggested by Ilachinski and Halpern (1987). ðT Þ ðT Þ ðT Þ
The essential new feature of this model was that the c l ij , si , sj .
connections between the cells are allowed to Instead of introducing the formalism of the
change according to rules similar in nature to the SDCA, we deal here with just one example in
state transition rules associated with the which the decoupler rule removes all links
174 Cellular Automata with Memory

Cellular Automata with Memory, Fig. 18 The SDCA described in text up to T = 6

connected to cells in which both values are zero A plausible wiring dynamics when dealing with
ðT Þ ðT þ1Þ ðT Þ ðT Þ
(lij ¼ 1 ! lij ¼ 0 iff si þ sj ¼ 0) and excitable CA is that in which the decoupler rule
the coupler rule adds links between all next- removes all links connected to cells in which both
nearest neighbor sites in which both values are values are at refractory state (lðijT Þ ¼ 1 ! lðijT þ1Þ ¼ 0
ðT Þ ðT þ1Þ ðT Þ ðT Þ ðT Þ ðT Þ
one (lij ¼ 0 ! li j ¼ 1 iff si þ sj ¼ 2 iff si ¼ sj =2) and the coupler rule adds links
ðT Þ
and j  NN i ). The SDCA with these transition between all next-nearest neighbor sites in which
ðT Þ ðT þ1Þ
rules for connections, together with the parity rule both values are excited (lij ¼ 0 ! lij ¼ 1 iff
ðT Þ ðT Þ ðT Þ
for mass states, is implemented in Fig. 18, in si ¼ sj ¼ 1 and j  NN i ).
which the initial Euclidean lattice with four neigh- In the SDCA in Fig. 20, the transition rule for cell
bors (so the generic cell ☐ has eight next-nearest states is that of the generalized defensive inhibition
neighbors: ) is seeded with a 3  3 block of rule: resting cell is excited if its ratio of excited and
ones. After the first iteration, most of the lattice connected to the cell neighbors to total number of
structure has decayed as an effect of the decoupler connected neighbors lies in the interval [1/8,2/8].
rule, so that the active value cells and links are The initial scenario of Fig. 20 is that of Fig. 12 with
confined to a small region. After T = 6, the link the wiring network revealed, that of an Euclidean
and value structures become periodic, with a peri- lattice with eight neighbors, in which, the generic
odicity of two. cell ☐ has 16 next-nearest neighbors: . No
Memory can be embedded in links in a similar decoupling is verified at the first iteration in
manner as in state values, so the link between Fig. 20, but the excited cells generate new connec-
any two cells is featured by a mapping  of tions, most of them lost, together with some of the
ðT Þ ð1Þ ðT Þ
its previous link values: l ij ¼ l lij , . . . , lij . initial ones, at T = 3. The excited cells at T = 3
The distance between two cells in the historic generate a crown of new connections at T = 4.
model (dij), is defined in terms of l instead Figure 21 shows the ahistoric and mode memory
of l values, so that i and j are direct patterns at T = 20. The figure makes apparent the
neighbors if dij = 1, and are nnext-nearest o
preserving effect of memory.
ðT Þ ðT Þ The Fredkin’s reversible construction is feasible
neighbors if dij n= 2. Now: N oi ¼ j=d ij ¼ 1 ,
ðT Þ ðT Þ in the SDCA scenario extending the  operation
and NN i ¼ j=d ij ¼ 2 . Generalizing the 
ðT 1Þ
approach to embedded memory applied to states, also to links: lðijT þ1Þ ¼ c lðijT Þ , sði T Þ , sðj T Þ lij .
the unchanged transition rules (f and c) operate These automata  may be endowed  with memory
ðT þ1Þ ðTþ1Þ ðT Þ ðT Þ ðT 1Þ ðT þ1Þ
on the featured link and cell state values: si as: si ¼ f sj  N i si ,lij ¼c
     
ðT Þ ðT þ1Þ ðT Þ ðT Þ ðT Þ ðT Þ ðT Þ ðT Þ ðT 1Þ
¼ f sj  N i , lij ¼ c l i j, si , sj . l ij , si , sj lij (Alonso-Sanz 2007a).
Figure 19 shows the effect of a-memory on the The SDCA seems to be particularly appropri-
cellular automaton above introduced starting as in ate for modeling the human brain function –
Fig. 18. The effect of memory on SDCA in the updating links between cells imitates variation
hexagonal and triangular tessellations is scruti- of synaptic connections between neurons
nized in Alonso-Sanz (2006a). represented by the cells – in which the relevant
Cellular Automata with Memory 175

Cellular Automata with Memory, Fig. 19 The SD cellular automaton introduced in text with weighted memory of
factor a. Evolution from T = 4 up to T = 9 starting as in Fig. 18

Cellular Automata with Memory, Fig. 20 The k = 3 SD cellular automaton described in text, up to T = 4

role of memory is apparent. Models similar to Memory in Other Discrete Time Contexts
SDCA have been adopted to build a dynamical
network approach to quantum space-time physics Continuous-Valued CA
(Requardt 1998, 2006b). Reversibility is an The mechanism of implementation of memory
important issue at such a fundamental physics adopted here, keeping the transition rule unaltered
level. Technical applications of SDCA may also but applying it to a function of previous states, can
be traced (Ros et al. 1994). Anyway, besides their be adopted in any spatialized dynamical system.
potential applications, SDCA with memory have Thus, historic memory can be embedded in:
an aesthetic and mathematical interest on their
own (Adamatzky 1994; Ilachinski 2000). Never- • Continuous-valued CA (or Coupled Map Lat-
theless, it seems plausible that further study on tices in which the state variable ranges in R,
SDCA (and Lattice Gas Automata with dynamical and the transition rule ’ is a continuous func-
geometry (Love et al. 2004)) with memory should tion (Kaneko 1986)), just by considering
turn out to be profitable. m instead of s in the application of the updating
176 Cellular Automata with Memory

Cellular Automata with Memory, Fig. 21 The SD cellular automaton starting as in Fig. 20 at T = 20, with no memory
(left) and mode memory in both cell states and links
 
ðTþ1Þ ðT Þ ðT Þ
rule: si ¼ ’ mj  N i . An elemen- Spatial Prisoner’s Dilemma
tary CA of this kind with memory would The Prisoner’s Dilemma (PD) is a game played by
ðT þ1Þ
be (Alonso-Sanz and Martin 2004a): si ¼ two players (A and B), who may choose either to
 
ðT Þ ðT Þ ðT Þ cooperate (C or 1) or to defect (D or 0). Mutual
3 mi1 þ mi þ miþ1 .
1
cooperators each score the reward R, mutual
• Fuzzy CA, a sort of continuous CA with states defectors score the punishment P; D scores the
ranging in the real [0,1] interval. An illustra- temptation T against C, who scores S (sucker’s
tion of the effect of memory in fuzzy CA is payoff) in such an encounter. Provided that T >
given in Alonso-Sanz and Martin (2002a). R > P > S, mutual defection is the only equilib-
The illustration operates on the elementary rule
      rium strategy pair. Thus, in a single round both
ðT þ1Þ ðT Þ ðT Þ ðT Þ ðT Þ
90 : si ¼ si1 ^ :siþ1 _ :si1 ^ siþ1 , players are to be penalized instead of both
which after fuzzification (a _ b ! min (1, a + b), rewarded, but cooperation may be rewarded in
ðT þ1Þ
a ^ b ! ab, : a ! 1  a) yields: si ¼ an iterated (or spatial) formulation. The game is
ðT Þ ðT Þ ðT Þ ðT Þ simplified (while preserving its essentials) if
si1 þ siþ1  2si1 siþ1 ; thus incorporating
ðT þ1Þ ðT Þ ðT Þ ðT Þ ðT Þ
P = S = 0. Choosing R = 1, the model will have
memory: si ¼ mi1 þ miþ1  2mi1 miþ1 . only one parameter: the temptation T=b.
• Quantum CA, such, for example, as the simple In the spatial version of the PD, each player
1D quantum CA models introduced in occupies at a site (i, j) in a 2D lattice. In each
Grössing and Zeilinger (1988): ðT Þ
generation the payoff of a given individual (pi,j ),
is the sum over all interactions with the eight
ðTþ1Þ 1  ðT Þ ðT Þ  ðT Þ

nearest neighbors and with its own site. In the
sj ¼ 1=2
ids j1 þ s j þ id sjþ1 ,
N next generation, an individual cell is assigned
ðT Þ
the decision (d i,j ) that received the highest payoff
which would become with memory (Alonso-Sanz among all the cells of its Moore’s neighborhood.
and Martin 2004a): In case of a tie, the cell retains its choice. The
spatialized PD (SPD for short) has proved to be a
ðTþ1Þ 1  ðT Þ ðT Þ ðT Þ

sj ¼ idmj1 þ mj þ id mjþ1 : promising tool to explain how cooperation can
N 1=2 hold out against the ever-present threat of
Cellular Automata with Memory 177

exploitation (Nowak and May 1992). This is a conditions when b = 1.85. When starting from
task that presents problems in the classic struggle a single defector, f at time step T is computed as
for survival Darwinian framework. the frequency of cooperators within the square of
When dealing with the SPD, memory can be size (2(T  1) + 1)2 centered on the initial D site.
embedded not only in choices but also in rewards. The ahistoric plot reveals the convergence of f to
Thus, in the historic model we dealt with, at 0.318, (which seems to be the same value regard-
T: (i) the payoffs coming from previous rounds less of the initial conditions (Nowak and May
ðT Þ
are accumulated ( pi,j ), and (ii) players are fea- 1992)). Starting from a single defector (a), the
tured by a summary of past decisions ( di,j ).
ðT Þ model with small memory (a = 0.1) seems to
Again, in each round or generation, a given cell reach a similar f value, but sooner and in a
plays with each of the eight neighbors and itself, smoother way. The plot corresponding to
the decision d in the cell of the neighborhood with a = 0.2 still shows an early decay in f that leads
the highest p being adopted. This approach to it to about 0.6, but higher memory factor values
modeling memory has been rather neglected, the lead f close to or over 0.9 very soon. Starting at
usual being that of designing strategies that spec- random (b), the curves corresponding to
ify the choice for every possible outcome in the 0.1  a  0.6 (thus with no memory of choices)
sequence of historic choices recalled (Hauert and do mimic the ahistoric curve but with higher f, as
Schuster 1997; Lindgren and Nordahl 1994). a  0.7 (also memory of choices) the frequency
Table 1 shows the initial scenario starting from of cooperators grows monotonically to reach
a single defector if 8b > 9 , b > 1.125, which almost full cooperation: D persists as scattered
means that neighbors of the initial defector unconnected small oscillators (D-blinkers), as
become defectors at T = 2. shown in Fig. 23. Similar results are found for
Nowak and May paid particular attention in any temptation value in the parameter region
their seminal papers to b = 1.85, a high but not 0.8 < b < 2.0, in which spatial chaos is charac-
excessive temptation value which leads to complex teristic in the ahistoric model. It is then con-
dynamics. After T = 2, defection can advance to a cluded that short-type memory supports
5  5 square or be restrained as a 3  3 square, cooperation.
depending on the comparison of 8a + 5  1.85 (the As a natural extension of the described binary
maximum p value of the recent defectors) with model, the 0-1 assumption underlying the model
9a + 9 (the p value of the non-affected players). can be relaxed by allowing for degrees of cooper-
As 8a + 5  1.85 = 9a + 9 ! a = 0.25, i.e., if ation in a continuous-valued scenario. Denoting
a > 0.25, defection remains confined to a 3  3 by x the degree of cooperation of player A and by
square at T = 3. Here we see the paradigmatic effect y the degree of cooperation of the player B, a
of memory: it tends to avoid the spread of consistent way to specify the pay-off for values
defection. of x and y other than zero or one is to simply
If memory is limited to the last three iterations: interpolate between the extreme payoffs of the
ðT Þ ðT2Þ ðT 1Þ ðT Þ ðT Þ ðT 2Þ binary case. Thus, the payoff that the player
pi,j ¼ a2 pi,j þ api,j þ pi,j ,mi,j ¼ ða2 d i,j
ðT 1Þ ðT Þ  ðT Þ A receives is:
þad i,j þ d i,j =ða2 þ a þ 1Þ, ) di,j ¼ round
    
ðT Þ ð2Þ ð1Þ
mi,j , with assignations at T = 2: pi,j ¼ api,j R S y
GA ðx,yÞ ¼ ðx, 1  xÞ :
ð2Þ ð2Þ ð2Þ
þpi,j ,di,j ¼ d i,j . T P 1y
Memory has a dramatic restrictive effect on
the advance of defection as shown in Fig. 22. In the continuous-valued historic
 formulation 
ð2Þ ð 1Þ ð2Þ
This figure shows the frequency of cooperators it is d  m, including di,j ¼ ad i,j þ d i,j =
(f) starting from a single defector and from a ða þ 1Þ . Table 2 illustrates the initial scenario
random configuration of defectors in a lattice of starting from a single (full) defector. Unlike in
size 400  400 with periodic boundary the binary model, in which the initial defector
178 Cellular Automata with Memory

Cellular Automata with d(1) = d(1)

Memory,
1 1 1 1 1
Table 1 Choices at T = 1
and T = 2; accumulated 1 1 1 1 1
payoffs after T = 1 and 1 1 0 1 1
T = 2 starting from a single 1 1 1 1 1
defector in the SPD. b > 9/8 1 1 1 1 1
p(1) = p(1)
9 9 9 9 9
9 8 8 8 9
9 8 8b 8 9
9 8 8 8 9
9 9 9 9 9
d(2) = d(2)
1 1 1 1 1 1 1
1 1 1 1 1 1 1
1 1 0 0 0 1 1
1 1 0 0 0 1 1
1 1 0 0 0 1 1
1 1 1 1 1 1 1
1 1 1 1 1 1 1
p(2) = ap(1) + p(2)
9a + 9 9a + 9 9a + 9 9a + 9 9a + 9 9a + 9 9a + 9
9a + 9 9a + 8 9a + 7 9a + 6 9a + 7 9a + 8 9a + 9
9a + 9 9a + 9 8a + 5b 8a + 3b 8a + 5b 9a + 9 9a + 9
9a + 9 9a + 9 8a + 3b 8a 8a + 3b 9a + 9 9a + 9
9a + 9 9a + 9 8a + 5b 8a + 3b 8a + 5b 9a + 9 9a+ 9
9a + 9 9a + 8 9a + 7 9a + 6 9a + 7 9a + 8 9a + 9
9a + 9 9a + 9 9a + 9 9a + 9 9a + 9 9a + 9 9a + 9

never becomes a cooperator, the initial defector Discrete-Time Dynamical Systems

cooperates with degree a/(1 + a) at T = 3: Memory can be embedded in any model in which
its neighbors which received the highest accumu- time plays a dynamical role. Thus, Markov chains
lated payoff (those in the corners with p(2) = p0T þ1 ¼ p0T M become with memory: p0T þ1 ¼ ı0T M
8a + 5b > 8ba), achieved this mean degree of with ıT being a weighted mean of the probability
cooperation after T = 2. Memory dramatically distributions up to T: ıT = p(p1, . . ., pT). In such
constrains the advance of defection in a smooth scenery, even a minimal incorporation of memory
way, even for the low level a = 0.1. The effect notably alters the evolution of p (Alonso-Sanz and
appears much more homogeneous compared Martin 2006a). Last but not least, conventional,
to the binary model, with no special case for non-spatialized, discrete dynamical systems
high values of a, as memory on decisions is become with memory: xT + 1 = f(mT) with mT
always operative in the continuous-valued model being an average of past values. As an overall
(Alonso-Sanz and Martin 2006b). The effect of rule, memory leads the dynamics a fixed point of
unlimited trailing memory on the SPD has been the map f (Aicardi and Invernizzi 1982).
studied in Alonso-Sanz (1999, 2003, 2004a, b, We will introduce an example of this in the
2005a, b). context of the PD game in which players follow
Cellular Automata with Memory 179

Cellular Automata with Memory, Fig. 22 Frequency model, the blue ones to the full memory model, the
of cooperators (f) with memory of the last three iterations. remaining curves to values of a from 0.1 to 0.9 by 0.1
a starting from a single defector, b starting at random intervals, in which, as a rule, the higher the a the higher the
(f(1) = 0.5). The red curves correspond to the ahistoric f for any given T

Cellular Automata with

Memory, Fig. 23 Patterns
at T = 200 starting at
random in the scenario of
Fig. 22b
180 Cellular Automata with Memory

Cellular Automata with d(2)

Memory,
1 1 1 1 1
Table 2 Weighted mean a a a
degrees of cooperation after 1 1þa 1þa 1þa 1
T = 2 and degree of 1 a
1þa 0 a
1þa 1
cooperation at T = 3 a a a
1 1þa 1þa 1þa 1
starting with a single
defector in the continuous- 1 1 1 1 1
valued SPD with b = 1.85 d(3)(a < 0.25)
1 1 1 1 1 1 1
a a a a a
1 1þa 1þa 1þa 1þa 1þa 1
a a a a a
1 1þa 1þa 1þa 1þa 1þa 1
a a a a a
1 1þa 1þa 1þa 1þa 1þa 1
a a a a a
1 1þa 1þa 1þa 1þa 1þa 1
a a a a a
1 1þa 1þa 1þa 1þa 1þa 1
1 1 1 1 1 1 1
d(3)(a < 0.25)
1 1 1 1 1 1 1
1 1 1 1 1 1 1
a a a
1 1 1þa 1þa 1þa 1
a a a
1 1 1þa 1þa 1þa 1
a a a
1 1 1þa 1þa 1þa 1
1 1 1 1 1 1 1
1 1 1 1 1 1 1

the so-called Paulov strategy: a Paulov player immutable. Therefore, in an iterated PAP contest,
cooperates if and only if both players opted for Paulov will always defect, and Anti-Paulov will
the same alternative in the previous move. The always cooperate. Relaxing the 0-1 assumption in
name Paulov stems from the fact that this strategy the standard formulation of the PAP contest,
embodies an almost reflex-like response to the degrees of cooperation can be considered in a
payoff: it repeats its former move if it was continuous-valued scenario. Now x and y will
rewarded by T or R, but switches behavior if it denote the degrees of cooperation of players
was punished by receiving only P or S. By coding A and B respectively, with both x and y lying in
cooperation as 1 and defection as 0, this strategy [0,1].
can be formulated in terms of the choices x of In this scenario, not only (0,1) is a fixed point,
Player A (Paulov) and y of Player B as: but also T(0.8, 0.6) = (0.8, 0.6). Computer imple-
x(T+1) = 1  j x(T)  y(T)j. The Paulov strategy mentation of the iterated PAP tournament turns
has proved to be very successful in its contests out to be fully disrupting of the theoretical dynam-
with other strategies (Nowak and Sigmund 1993). ics. The errors caused by the finite precision of the
Let us give a simple example of this: suppose that computer floating point arithmetics (a common
Player B adopts an Anti-Paulov strategy (which problem in dynamical systems working modulo 1)
cooperates to the extent Paulov defects) with make the final fate of every point to be (0,1). With
y(T+1) = 1  j 1  x(T)  y(T)j. Thus, in an no exceptions: even the theoretically fixed point
iterated Paulov-Anti-Paulov (PAP) contest, with (0.8,0.6) ends up as (0,1) in the computerized
T(x, y) = (1  |x  y|, 1  |1  x  y|), it implementation.
is T(0, 0) = T(1, 1) = (1, 0), T(1, 0) = (0, 1), A natural way to incorporate older choices in
and T(0, 1) = (0, 1), so that (0,1) turns out to be the strategies of decision is to feature players by a
Cellular Automata with Memory 181

Cellular Automata with Memory, Fig. 24 Dynamics of the mean values of x (red) and y (blue) starting from any of the
points of the 1  1 square

summary (m) of their own choices farther back in or perform a particular function. Average memory
time. The PAP contest becomes in this way: xðT þ1Þ in CA tends to inhibit complexity, inhibition that
¼ 1 j mðxT Þ  mðyT Þ j , yðT þ1Þ ¼ 1 j 1  mðxT Þ  can be modulated by varying the depth of memory,
mðyT Þ j. The simplest historic extension results in but memory not of average type opens a notable
new perspective in CA. This could mean a potential
considering only the two last choices: m(z(T  1),
advantage of CAwith memory over standard CA as
z(T)) = (az(T  1) + z(T))/(a + 1) (z stands for both
a tool for modeling. Anyway, besides their potential
x and y) (Alonso-Sanz 2005b).
applications, CA with memory (CAM) have an
Figure 24 shows the dynamics of the mean
aesthetic and mathematical interest on their own.
values of x and y starting from any of the
Thus, it seems plausible that further study on
101  101 lattice points of the 1  1 square with
CA with memory should turn out profitable, and,
sides divided by 0.01 intervals. The dynamics in
maybe that as a result of a further rigorous study of
the ahistoric context are rather striking: immedi-
CAM it will be possible to paraphrase T. Toffoli in
ately, at T = 2, both x and y increase from 0.5 up
presenting CAM – as an alternative to (rather than
to app. 0.66( ’ 2/3), a value which remains stable
an approximation of) integro-differential equa-
up to app. T = 100, but soon after Paulov cooper-
tions in modeling – phenomena with memory.
ation plummets, with the corresponding firing of
cooperation of Anti-Paulov: finite precision arith-
metics leads every point to (0,1). With memory,
Paulov not only keeps a permanent mean degree Bibliography
cooperation but it is higher than that of Anti-
Paulov; memory tends to lead the overall dynamics Primary Literature
to the ahistoric (theoretically) fixed point (0.8, 0.6). Adamatzky A (1994) Identification of cellular automata.
Taylor and Francis, London
Adamatzky A (2001) Computing in nonlinear media and
automata collectives. IoP Publishing, London
Future Directions Adamatzky A, Holland O (1998) Phenomenology of exci-
tation in 2-D cellular automata and swarm systems.
Chaos, Solitons Fractals 9:1233–1265
Embedding memory in states (and in links if wiring Aicardi F, Invernizzi S (1982) Memory effects in discrete
is also dynamic) broadens the spectrum of CA as a dynamical systems. Int J Bifurc Chaos 2(4):815–830
tool for modeling, in a fairly natural way of easy Alonso-Sanz R (1999) The historic prisoner’s dilemma. Int
J Bifurc Chaos 9(6):1197–1210
computer implementation. It is likely that in some
Alonso-Sanz R (2003) Reversible cellular automata with
contexts, a transition rule with memory could memory. Phys D 175:1–30
match the correct behavior of the CA system of a Alonso-Sanz R (2004a) One-dimensional, r = 2 cellular
given complex system (physical, biological, social automata with memory. Int J Bifurc Chaos
14:3217–3248
and so on). A major impediment in modeling with
Alonso-Sanz R (2004b) One-dimensional, r = 2 cellular
CA stems from the difficulty of utilizing the CA automata with memory. Int J BifurcChaos
complex behavior to exhibit a particular behavior 14:3217–3248
182 Cellular Automata with Memory

Alonso-Sanz R (2005a) Phase transitions in an elementary Alonso-Sanz R, Martin MC, Martin M (2001b) The
probabilistic cellular automaton with memory. Phys effect of memory in the spatial continuous-valued pris-
A 347:383–401 oner’s dilemma. Int J Bifurc Chaos 11(8):2061–2083
Alonso-Sanz R (2005b) The Paulov versus Anti-Paulov Alonso-Sanz R, Martin MC, Martin M (2001c) The his-
contest with memory. Int J Bifurc Chaos toric strategist. Int J Bifurc Chaos 11(4):943–966
15(10):3395–3407 Alonso-Sanz R, Martin MC, Martin M (2001d) The
Alonso-Sanz R (2006a) A structurally dynamic cellular historic-stochastic strategist. Int J Bifurc Chaos
automaton with memory in the triangular tessellation. 11(7):2037–2050
Complex Syst 17(1):1–15 Alvarez G, Hernandez A, Hernandez L, Martin A (2005)
Alonso-Sanz R (2006b) The beehive cellular automaton A secure scheme to share secret color images. Comput
with memory. J Cell Autom 1:195–211 Phys Commun 173:9–16
Alonso-Sanz R (2007a) Reversible structurally dynamic Fredkin E (1990) Digital mechanics. An informal process
cellular automata with memory: a simple example. based on reversible universal cellular automata.
J Cell Autom 2:197–201 Physica D 45:254–270
Alonso-Sanz R (2007b) A structurally dynamic cellular Grössing G, Zeilinger A (1988) Structures in quantum
automaton with memory. Chaos, Solitons Fractals cellular automata. Physica B 15:366
32:1285–1295 Hauert C, Schuster HG (1997) Effects of increasing the
Alonso-Sanz R, Adamatzky A (2008) On memory and number of players and memory steps in the iterated
structurally dynamism in excitable cellular automata prisoner’s dilemma, a numerical approach. Proc
with defensive inhibition. Int J Bifurc Chaos R Soc Lond B 264:513–519
18(2):527–539 Hooft G (1988) Equivalence relations between determin-
Alonso-Sanz R, Cardenas JP (2007) On the effect of mem- istic and quantum mechanical systems. J Stat Phys
ory in Boolean networks with disordered dynamics: the 53(1/2):323–344
K = 4 case. Int J Modrn Phys C 18:1313–1327 Ilachinski A (2000) Cellular automata. World Scientific,
Alonso-Sanz R, Martin M (2002a) One-dimensional cellu- Singapore
lar automata with memory: patterns starting with a Ilachinsky A, Halpern P (1987) Structurally dynamic cel-
single site seed. Int J Bifurc Chaos 12:205–226 lular automata. Complex Syst 1:503–527
Alonso-Sanz R, Martin M (2002b) Two-dimensional cel- Kaneko K (1986) Phenomenology and characterization of
lular automata with memory: patterns starting with a coupled map lattices. In: Dynamical systems and
single site seed. Int J Mod Phys C 13:49–65 sigular phenomena. World Scientific, Singapore
Alonso-Sanz R, Martin M (2003) Cellular automata with Kauffman SA (1993) The origins of order: self-
accumulative memory: legal rules starting from a single organization and selection in evolution. Oxford Uni-
site seed. Int J Mod Phys C 14:695–719 versity Press, Oxford
Alonso-Sanz R, Martin M (2004a) Elementary probabilis- Lindgren K, Nordahl MG (1994) Evolutionary dynamics
tic cellular automata with memory in cells. In: Sloot of spatial games. Physica D 75:292–309
PMA et al (eds) LNCS, vol 3305. Springer, Berlin, Love PJ, Boghosian BM, Meyer DA (2004) Lattice gas
pp 11–20 simulations of dynamical geometry in one dimension.
Alonso-Sanz R, Martin M (2004b) Elementary cellular Phil Trans R Soc Lond A 362:1667
automata with memory. Complex Syst 14:99–126 Margolus N (1984) Physics-like models of computation.
Alonso-Sanz R, Martin M (2004c) Three-state Physica D 10:81–95
one-dimensional cellular automata with memory. Martin del Rey A, Pereira Mateus J, Rodriguez
Chaos, Solitons Fractals 21:809–834 Sanchez G (2005) A secret sharing scheme based on
Alonso-Sanz R, Martin M (2005) One-dimensional cellu- cellular automata. Appl Math Comput
lar automata with memory in cells of the most frequent 170(2):1356–1364
recent value. Complex Syst 15:203–236 Nowak MA, May RM (1992) Evolutionary games and
Alonso-Sanz R, Martin, M (2006a) A structurally dynamic spatial chaos. Nature 359:826
cellular automaton with memory in the hexagonal tes- Nowak MA, Sigmund K (1993) A strategy of win-stay,
sellation. In: El Yacoubi S, Chopard B, Bandini S (eds) lose-shift that outperforms tit-for-tat in the Prisoner’s
LNCS, vol 4774. Springer, Berlin, pp 30–40 Dilemma game. Nature 364:56–58
Alonso-Sanz R, Martin M (2006b) Elementary cellular Requardt M (1998) Cellular networks as models for Plank-
automata with elementary memory rules in cells: the scale physics. J Phys A 31:7797
case of linear rules. J Cell Autom 1:70–86 Requardt M (2006a) The continuum limit to discrete geom-
Alonso-Sanz R, Martin M (2006c) Memory boosts coop- etries. arxiv.org/abs/math-ps/0507017
eration. Int J Mod Phys C 17(6):841–852 Requardt M (2006b) Emergent properties in structurally
Alonso-Sanz R, Martin MC, Martin M (2000) Discounting dynamic disordered cellular networks. J Cell Autom
in the historic prisoner’s dilemma. Int J Bifurc Chaos 2:273
10(1):87–102 Ros H, Hempel H, Schimansky-Geier L (1994) Stochastic
Alonso-Sanz R, Martin MC, Martin M (2001a) Historic dynamics of catalytic CO oxidation on Pt(100). Pysica
life. Int J Bifurc Chaos 11(6):1665–1682 A 206:421–440
Cellular Automata with Memory 183

Sanchez JR, Alonso-Sanz R (2004) Multifractal properties Wolf-Gladrow DA (2000) Lattice-gas cellular automata
of R90 cellular automaton with memory. Int J Mod and lattice Boltzmann models. Springer, Berlin
Phys C 15:1461 Wolfram S (1984) Universality and complexity in cellular
Stauffer D, Aharony A (1994) Introduction to percolation automata. Physica D 10:1–35
theory. CRC Press, London Wuensche A (2005) Glider dynamics in 3-value hexagonal
Svozil K (1986) Are quantum fields cellular automata? cellular automata: the beehive rule. Int J Unconv
Phys Lett A 119(41):153–156 Comput 1:375–398
Toffoli T, Margolus M (1987) Cellular automata machines. Wuensche A, Lesser M (1992) The global dynamics of
MIT Press, Cambridge, MA cellular automata. Addison-Wesley, Reading
Toffoli T, Margolus N (1990) Invertible cellular automata:
a review. Physica D 45:229–253
Vichniac G (1984) Simulating physics with cellular autom- Books and Reviews
ata. Physica D 10:96–115 Alonso-Sanz R (2008) Cellular automata with memory.
Watts DJ, Strogatz SH (1998) Collective dynamics of Old City Publising, Philadelphia
small-world networks. Nature 393:440–442
 no loss of information: the state at time t can be
Classification of Cellular recovered from the state at time t + 1. For CAs
Automata this means that the global map is injective.
Semi-decidability A problem is said to be semi-
Klaus Sutner decidable or computably enumerable if it
Carnegie Mellon University, Pittsburgh, PA, USA admits an algorithm that returns “yes” after
finitely many steps if this is indeed the correct
answer. Otherwise the algorithm never termi-
nates. The Halting Problem is the standard
example for a semi-decidable problem.
Article Outline A problem is decidable if, and only if, the
problem itself and its negation are semi-
Glossary decidable.
Definition of the Subject Surjectivity The global map of a CA is surjec-
Introduction tive if every configuration appears as the image
Reversibility and Surjectivity of another. By contrast, a configuration that
Definability and Computability fails to have a predecessor is often referred to
Computational Equivalence as a Garden-of-Eden.
Conclusion Undecidability It was recognized by logicians
Bibliography and mathematicians in the first half of the
20th century that there is an abundance of
Glossary well-defined problems that cannot be solved
by means of an algorithm, a mechanical proce-
Cellular automaton For our purposes, a dure that is guaranteed to terminate after
(one-dimensional) cellular automaton (CA) is finitely many steps and produce the appropri-
given by a local map r : Sw ! S where S is the ate answer. The best known example of an
underlying alphabet of the automaton and w is its undecidable problem is Turing’s Halting Prob-
width. As a data structure, suitable as input to a lem: there is no algorithm to determine whether
decision algorithm, a CA can thus be specified by a given Turing machine halts when run on an
a simple lookup table. We abuse notation and empty tape.
write r(x) for the result of applying the global Universality A computational device is univer-
map of the CA to configuration x  Sℤ. sal it is capable of simulating any other com-
Finite configurations One often considers CA putational device. The existence of universal
with a special quiescent state: the homogeneous computers was another central insight of the
configuration where all cells are in the quiescent early days of computability theory and is
state is required to be fixed point under the global closely related to undecidability.
map. Infinite configurations where all but finitely Wolfram classes Wolfram proposed a heuristic
many cells are in the quiescent state are often classification of cellular automata based on
called finite configurations. This is somewhat of observations of typical behaviors. The classifi-
a misnomer; we prefer to speak about configura- cation comprises four classes: evolution leads
tions with finite support. to trivial configurations, to periodic configura-
Reversibility A discrete dynamical system is tions, evolution is chaotic, evolution leads to
reversible if the evolution of the system incurs complicated, persistent structures.

# Springer-Verlag 2009 185

A. Adamatzky (ed.), Cellular Automata,
https://doi.org/10.1007/978-1-4939-8700-9_50
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_50
186 Classification of Cellular Automata

Definition of the Subject • W1: Evolution leads to homogeneous fixed

points.
Cellular automata display a large variety of behav- • W2: Evolution leads to periodic configurations.
iors. This was recognized clearly when extensive • W3: Evolution leads to chaotic, aperiodic
simulations of cellular automata, and in particular patterns.
one-dimensional CA, became computationally • W4: Evolution produces persistent, complex
feasible around 1980. Surprisingly, even when patterns of localized structures.
one considers only elementary CA, which are
constrained to a binary alphabet and local maps Thus, Wolfram’s first three classes follow
involving only nearest neighbors, complicated closely concepts from continuous dynamics:
behaviors are observed in some cases. In fact, fixed point attractors, periodic attractors and
it appears that most behaviors observed in strange attractors, respectively. They corre-
automata with more states and larger neighbor- spond roughly to systems with zero temporal
hoods already have qualitative analogues in the and spatial entropy, zero temporal entropy but
realm of elementary CA. Careful empirical stud- positive spatial entropy, and positive temporal
ies lead Wolfram to suggest a phenomenological and spatial entropy, respectively. W4 is more
classification of CA based on the long-term evo- difficult to associate with a continuous analogue
lution of configurations, see Wolfram (1984b, except to say that transients are typically very
2002b) and section “Introduction”. While Wol- long. To understand this class it is preferable to
fram’s four classes clearly capture some of the consider CA as models of massively parallel
behavior of CA it turns out that any attempt at computation rather than as particular discrete
formalizing this taxonomy meets with consider- dynamical systems. It was conjectured by Wol-
able difficulties. Even apparently simple ques- fram that W4 automata are capable of performing
tions about the behavior of CA turn out to be complicated computations and may often be com-
algorithmically undecidable and it is highly chal- putationally universal. Four examples of elemen-
lenging to provide a detailed mathematical analy- tary CA that are typical of the four classes are
sis of these systems. shown in Fig. 1. Li and Packard (1990), Li et al.
(1990) proposed a slightly modified version of
this hierarchy by refining the low classes and in
Introduction particular Wolfram’s W2. Much like Wolfram’s
classification, the Li–Packard classification is
In the early 1980s Wolfram published a collec- concerned with the asymptotic behavior of the
tion of 20 open problems in the theory of CA, automaton, the structure and behavior of the lim-
see Wolfram (1985). The first problem on his list iting configurations. Here is one version of the
is “What overall classification of cellular Li–Packard classification, see Li et al. (1990).
automata behavior can be given?” As Wolfram
points out, experimental mathematics provides • LP1: Evolution leads to homogeneous fixed
a first answer to this problem: one performs a points.
large number of explicit simulations and • LP2: Evolution leads to non-homogeneous
observes the patterns associated with the long fixed points, perhaps up a to a shift.
term evolution of a configuration, see Wolfram • LP3: Evolution leads to ultimately periodic
(1984a, 2002b). Wolfram proposed a classifica- configurations. Regions with periodic behavior
tion that is based on extensive simulations in are separated by domain walls, possibly up to a
particular of one-dimensional cellular automata shift.
where the evolution of a configuration can be • LP4: Configurations produce locally chaotic
visualized naturally as a two-dimensional behavior. Regions with chaotic behavior are
image. The classification involves four classes separated by domain walls, possibly up to a
that can be described as follows: shift.
Classification of Cellular Automata 187

Classification of Cellular Automata, Fig. 1 Typical examples of the behavior described by Wolfram’s classes among
elementary cellular automata

• LP5: Evolution leads to chaotic patterns that reached after finitely many steps. The map is
are spatially unbounded. sensitive (to initial conditions) if
• LP6: Evolution is complex. Transients are long
8x,e > 0∃d
and lead to complicated space-time patterns
> 08y  Bd ðxÞ∃n  ℕðd ðf n ðxÞ, f n ðyÞÞ  eÞ:
which may be non-monotonic in their
behavior.
Lastly, the map is positively expansive if
By contrast, a classification closer to traditional
dynamical systems theory was introduced by ∃e > 08x 6¼ y∃n  ℕðd ðf n ðxÞ, f n ðyÞÞ  eÞ:
Kůrka, see Kůrka (1997, 2003). The classification
rests on the notions of equicontinuity, sensitivity Kůrka’sclassification then takes the following
to initial conditions and expansivity. Suppose x is form.
a point in some metric space and f a map on that
space. Then f is equicontinuous at x if • K1: All points are equicontinuous under the
global map.
• K2: Some but not all points are equicontinuous
8e > 0∃d under the global map.
> 08y  Bd ðxÞ, n  ℕðd ðf n ðxÞ, f n ðyÞÞ < eÞ • K3: The global map is sensitive but not posi-
tively expansive.
where d(., .) denotes a metric. Thus, all points in a • K4: The global map is positively expansive.
sufficiently small neighborhood of x remain close
to the iterates of x for the whole orbit. Global This type of classification is perfectly suited to
equicontinuity is a fairly strong condition, it the analysis of uncountable spaces such as the
implies that the limit set of the automaton is Cantor space {0, 1}ℕ or the full shift space Sℤ
188 Classification of Cellular Automata

which carry a natural metric structure. For the properties starting with injectivity and surjectivity
most part we will not pursue the analysis of CA as well as topological properties such as
by topological and measure theoretic means here equicontinuity and sensitivity can be expressed
and refer to ▶ “Topological Dynamics of Cellular in terms of simple arithmetic conditions on the
Automata” in this volume for a discussion of these rule coefficients. For example, equicontinuity is
methods. See section “Definability and Comput- equivalent to all prime divisors of the modulus
ability” for the connections between topology and m dividing all coefficients ci, i > 1, see Manzini
computability. and Margara (1999) and the references therein. It
Given the apparent complexity of observable is also noteworthy that in the linear case methods
CA behavior one might suspect that it is difficult tend to carry over to arbitrary dimensions; in gen-
to pinpoint the location of an arbitrary given CA eral there is a significant step in complexity from
in any particular classification scheme with any dimension one to dimension two.
precision. This is in contrast to simple parameter- No claim is made that the given classifications
izations of the space of CA rules such as are complete; in fact, one should think of them as
Langton’s l parameter that are inherently easy to prototypes rather than definitive taxonomies. For
compute. Briefly, the l value of a local map is the example, one might add the class of nilpotent CA
fraction of local configurations that map to a non- at the bottom. A CA is nilpotent if all configura-
zero value, see Langton (1990), Li et al. (1990). tions evolve to a particular fixed point after
Small l values result in short transients leading to finitely many steps. Equivalently, by compact-
fixed points or simple periodic configurations. As ness, there is a bound n such that all configura-
l increases the transients grow longer and the tions evolve to the fixed point in no more than
orbits become more and more complex until, at n steps. Likewise, we could add the class of
last, the dynamics become chaotic. Informally, intrinsically universal CA at the top. A CA is
sweeping the l value from 0 to 1 will produce CA intrinsically universal if it is capable of simulating
in W1, then W2, then W4 and lastly in W3. The last all other CA of the same dimension in some rea-
transition appears to be associated with a threshold sonable sense. For a fairly natural notion of sim-
phenomenon. It is unclear what the connection ulation see Ollinger (2003). At any rate,
between Langton’s l-value and computational prop- considerable effort is made in the references to
erties of a CA is, see Mitchell et al. (1994), Packard elaborate the characteristics of the various classes.
(1988). Other numerical measures that appear to be For many concrete CA visual inspection of the
loosely connected to classifications are the mean orbits of a suitable sample of configurations read-
field parameters of Gutowitz (1996a, b), the ily suggests membership in one of the classes.
Z-parameter by Wuensche (1999), see also Oliveira
et al. (2001). It seems doubtful that a structured
Reversibility and Surjectivity
taxonomy along the lines of Wolfram or Li–Packard
can be derived from a simple numerical measure
A first tentative step towards the classification of a
such as the l value alone, or even from a combina-
dynamical systems is to determine its reversibility
tion of several such values. However, they may be
or lack thereof. Thus we are trying to determine
useful as empirical evidence for membership in a
whether the evolution of the system is associated
particular class.
with loss of information, or whether it is possible
Classification also becomes significantly easier
to reconstruct the state of the system at time t from
when one restricts one’s attention to a limited class
its state at time t + 1. In terms of the global map of
of CA such as additive CA, see ▶ “Additive
the system we have to decide injectivity. Closely
Cellular Automata”. In this context, additive means
related is the question whether the global map is
that

the
 local rule of the automaton has the form
! P surjective, i. e., whether there is no Garden-of-
r x ¼ i ci xi where the coefficients as well as Eden: every configuration has a predecessor under
the states are modular numbers. A number of the global map. As a consequence, the limit set of
Classification of Cellular Automata 189

the automaton is the whole space. It was shown of Thus one should expect reversible CA to exhibit
Hedlund that for CA the two notions are fairly complicated behavior in general.
connected: every reversible CA is also surjective, For infinite, one-dimensional CA it was shown
see Hedlund (1969), ▶ “Reversible Cellular by Amoroso and Patt (1972) that reversibility is
Automata”. As a matter of fact, reversibility of decidable. Moreover, it is decidable if the global
the global map of a CA implies openness of the map is surjective. An efficient practical algorithm
global map, and openness implies surjectivity. using concepts of automata theory can be found in
The converse implications are both false. By a Sutner (1991), see also Culik (1987), Delorme and
well-known theorem by Hedlund (1969) the Mazoyer (1999), Head (1989). The fast algorithm
global maps of CA are precisely the continuous is based on interpreting a one-dimensional CA as
maps that commute with the shift. It follows from a deterministic transducer, see Beal and Perrin
basic topology that the inverse global map of a (1997), Rozenberg and Salomaa (1997) for back-
reversible CA is again the global map of a suitable ground. The underlying semi-automaton of the
CA. Hence, the predecessor configuration of a transducer is a de Bruijn automaton B whose
given configuration can be reconstructed by states are words in Sw  1 where S is the alphabet
another suitably chosen CA. For results of the CA and w is its width. The transitions are
c
concerning reversibility on the limit set of the given by ax! xb where a, b, c  S, x  Sw2
automaton see Taati (2007). and c = r(axb), r being the local map of the
From the perspective of complexity the key CA. Since B is strongly connected, the product
result concerning reversible systems is the work automaton of B will contain a strongly connected
by Lecerf (1963) and Bennett (1973). They show component C that contains the diagonal D, an
that reversible Turing machines can compute any isomorphic copy of B . The global map of the
partial recursive function, modulo a minor tech- CA is reversible if, and only if, C = D is the only
nical problem: In a reversible Turing machine non-trivial component. It was shown by Hedlund
there is no loss of information; on the other (1969) that surjectivity of the global map is equiv-
hand even simple computable functions are alent with local injectivity: the restriction of the
clearly irreversible in the sense that, say, the map to configurations with finite support must be
sum of two natural numbers does not determine injective. The latter property holds if, and only if,
these numbers uniquely. To address this issue C = D and is thus easily decidable. Automata
one has to adjust the notion of computability theory does not readily generalize to words of
slightly in the context of reversible computation: dimensions higher than one. Indeed, reversibility
given a partial recursive function f : ℕ ! ℕ the and surjectivity in dimensions higher than one are
function f^ðxÞ ¼ hx, f ðxÞi can be computed by a undecidable, see Kari (1990) and ▶ “Tiling Prob-
reversible Turing machine where h., .i is any lem and Undecidability in Cellular Automata” in
effective pairing function. If f itself happens to this volume for the rather intricate argument
be injective then there is no need for the coding needed to establish this fact.
device and f can be computed by a reversible While the structure of reversible one-
Turing machine directly. For example, we can dimensional CA is well-understood, see ▶ “Tiling
compute the product of two primes reversibly. Problem and Undecidability in Cellular Autom-
Morita demonstrated that the same holds true for ata”, (Durand-Lose 2001), and while there is an
one-dimensional cellular automata (Morita efficient algorithm to check reversibility, few
1994; Morita and Harao 1989; Toffoli and methods are known that allow for the construction
Margolus 1990), ▶ “Tiling Problem and of interesting reversible CA. There is a noteworthy
Undecidability in Cellular Automata”: reversibil- trick due to Fredkin that exploits the reversibility of
ity is no obstruction to computational universality. the Fibonacci equation Xn+1 = Xn + Xn1. When
As a matter of fact, any irreversible cellular addition is interpreted as exclusive or this can be
automaton can be simulated by a reversible used to construct a second-order CA from any
one, at least on configurations with finite support. given binary CA; the former can then be recoded
190 Classification of Cellular Automata

as a first-order CA over a 4-letter alphabet. For Definability and Computability

example, for the open but irreversible elementary
CA number 90 we obtain the CA shown in Fig. 2. Formalizing Wolfram’s Classes
Another interesting class of reversible one- Wolfram’s classification is an attempt to catego-
dimensional CA, the so-called partitioned cellu- rize the complexity of the CA by studying the
lar automata (PCA), is due to Morita and Harao, patterns observed during the long-term evolution
see Morita (1994, 1995), Morita and Harao of all configurations. The first two classes are
(1989). One can think of a PCA as a cellular relatively easy to observe, but it is difficult to
automaton whose cells are divided into multiple distinguish between the last two classes. In partic-
tracks; specifically Morita uses an alphabet of the ular W4 is closely related to the kind of behavior
form S = S1  S2  S3. The configurations of the that would be expected in connection with sys-
automaton can be written as (X, Y, Z) where tems that are capable of performing complicated
X  S1ℤ, Y  S2ℤ and Z  S3ℤ. Now consider computations, including the ability to perform
the shearing map s defined by s(X, Y, Z) = (RS universal computation; a property that is notori-
(X), Y, LS(Z)) where RS and LS denote the right ously difficult to check, see Soare (1987). The
and left shift, respectively. Given any function focus on the full configuration space rather than
f : S ! S we can define a global map f ∘ s a significant subset thereof corresponds to the
where f is assumed to be applied point-wise. worst-case approach well-known in complexity
Since the shearing map is bijective, the CA will theory and is somewhat inferior to an average
be reversible if, and only if, the map f is bijective. case analysis. Indeed, Baldwin and Shelah point
It is relatively easy to construct bijections f that out that a product construction can be used to
cause the CA to perform particular computational design a CA whose behavior is an amalgamation
tasks, even when a direct construction appears to of the behavior of two given CA, see Baldwin
be entirely intractable. (2002), Baldwin and Shelah (2000). By

Classification of Cellular
Automata, Fig. 2 A
reversible automaton
obtained by applying
Fredkin’s construction to
the irreversible elementary
CA 77
Classification of Cellular Automata 191

combining CA in different classes one obtains 8 x,y,zðx ! z and y ! z implies x ¼ yÞ,

striking examples of the weakness of the worst- 8 x∃yðy ! xÞ:
case approach. A natural example of this mixed
type of behavior is elementary CA 184 which
As we have seen, both properties are easily
displays class II or class III behavior, depending
decidable in the one-dimensional case. In fact,
on the initial configuration. Another basic exam-
one can express the basic predicate x ! y (as well
ple for this type of behavior is the well-studied
as equality) in terms of finite state machines on
elementary CA 30, see section “Conclusion”.
infinite words. These machines are defined like
Still, for many CA a worst-case classification
ordinary finite state machines but the acceptance
seems to provide useful information about the
condition requires that certain states are reached
structural properties of the automaton. The first
infinitely and co-infinitely often, see Börger et al.
attempt at formalizing Wolfram’s class was made
(2001), Grädel et al. (2002). The emptiness prob-
by Culik and Yu who proposed the following
lem for these automata is easily decidable using
hierarchy, given here in cumulative form, see
graph theoretic algorithms. Since regular languages
Culik and Sheng (1988):
on infinite words are closed under union, comple-
• CY1: All configurations evolve to a fixed point. mentation and projection, much like their finite
• CY2: All configurations evolve to a periodic counterparts, and all the corresponding operations
configuration. on automata are effective, it follows that one can
• CY3: The orbits of all configurations are decide the validity of first order sentences over
decidable. A r such as the two examples above: the model-
• CY4: No constraints. checking problem for these structures and first
order logic is decidable, see Libkin (2004). For
The Culik–Yu classification employs two example, we can decide whether there is a config-
rather different methods. The first two classes uration that has a certain number of predecessors.
can be defined by a simple formula in a suitable Alternatively, one can translate these sentences
logic whereas the third (and the fourth in the into monadic second order logic of one successor,
disjoint version of the hierarchy) rely on notions and use well-known automata-based decision algo-
of computability theory. As a general framework rithms there directly, see Börger et al. (2001). Sim-
for both approaches we consider discrete dynam- ilar methods can be used to handle configurations
ical systems, structures of the form A ¼ hC, !i with finite support, corresponding to weak
where C  Sℤ is the space of configurations of the monadic second order logic. Since the complexity
system and ! is the “next configuration” relation of the decision procedure is non-elementary one
on C. We will only consider the deterministic case should not expect to be able to handle complicated
where for each configuration x there exists pre- assertions. On the other hand, at least for weak
cisely one configuration y such that x ! y. Hence monadic second order logic practical
we are really dealing with algebras with one unary implementations of the decision method exist, see
function, but iteration is slightly easier to deal Elgaard et al. (1998). There is no hope of general-
with in the relational setting. The structures most izing this approach as the undecidability of, say,
important in this context are the ones arising from reversibility in higher dimensions demonstrates.
t
a CA. For any local map r we consider the struc- Write x! t if x evolves to y in exactly t steps,
þ
ture A r ¼ hC, !i where the next configuration x! y if x evolves to y in any positive number of

relation is determined by x ! r(x). steps and x! y if x evolves to y in any number of
t
Using the standard language of first order logic steps. Note that ! is definable for each fixed t,

we can readily express properties of the CA in but ! fails to be so definable in first order
terms of the system A r . For example, the system logic. This is in analogy to the undefinability of
is reversible, respectively surjective, if the follow- path existence problems in the first order theory of
ing assertions are valid over A: graphs, see Libkin (2004). Hence it is natural to
192 Classification of Cellular Automata

extend our language so we can express iterations 30 into decision algorithms for Reachability.
of the global map, either by adding transitive There is another, somewhat more technical prob-
closures or by moving to some limited system of lem to overcome in formalizing classifications.

higher order logic over A r where ! is definable, Recall that the full configuration space is
see Börger et al. (2001). C = Sℤ. Intuitively, given x  C we can effec-
Arguably the most basic decision problem tively determine the next configuration y = r(x).
associated with a system A that requires iteration However, classical computability theory does not
of the global map is the Reachability Problem: deal with infinitary objects such as arbitrary con-
given two configurations x and y, does the evolu- figuration so a bit of care is needed here. The key
tion of x lead to y? A closely related but different insight is that we can determine arbitrary finite
question is the Confluence Problem: will two con- segments of r(x) using only finite segments of
figurations x and y evolve to the same limit cycle? x (and, of course, the lookup table for the local
Confluence is an equivalence relation and allows map). There are several ways to model comput-
for the decomposition of configuration space into ability on Sℤ based on this idea of finite approx-
limit cycles together with their basins of attrac- imations, we refer to Weihrauch (2000) for a
tion. The Reachability and Confluence Problem particularly appealing model based on so-called
amount to determining, given configurations type-2 Turing machines; the reference also con-
x and y, whether tains many pointers to the literature as well as a
comparison between the different approaches. It is
 easy to see that for any CA the global map r as

 x! y,   well as all its iterates rt are computable, the latter
∃z x! z and y! z ,
uniformly in t. However, due to the finitary nature
of all computations, equality is not decidable in
respectively. As another example, the first two type-2 computability: the unequal operator
Culik–Yu class can be defined like so: U0(x, y) = 0 if x 6¼ y, U0(x, y) undefined other-

  wise, is computable and thus unequality is semi-
8x∃z x! z and z ! z , decidable, but the stronger U0(x, y) = 0 if x 6¼ y,

 þ
 U0(x, y) = 1, otherwise, is not computable. The
8x∃z x! z and z! z : last result is perhaps somewhat counterintuitive,
but it is inevitable if we strictly adhere to the finite
It is not difficult to give similar definitions for approximation principle.
the lower Li–Packard classes if one extends the In order to avoid problems of this kind it has
language by a function symbol denoting the shift become customary to consider certain subspaces
operator. of the full configuration space, in particular Cfin,
The third Culik–Yu class is somewhat more the collection of configurations with finite sup-
involved. By definition, a CA lies in the third port, Cper, the collection of spatially periodic con-
class if it admits a global decision algorithm to figurations and Cap, the collection of almost
determine whether a given configuration periodic configurations of the form . . .uuuwvvv. . .
x evolves to another given configuration y in a where u, v and w are all finite words over the
finite number of steps. In other words, we are alphabet of the automaton. Thus, an almost peri-
looking for automata where the Reachability odic configuration differs from a configuration of
Problem is algorithmically solvable. While one the form ouvo in only finitely many places. Con-
can agree that W4 roughly translates into figurations with finite support correspond to the
undecidability and is thus properly situated in special case where u = v = 0 is a special quiescent
the hierarchy, it is unclear how chaotic patterns symbol and spatially periodic configurations cor-
in W3 relate to decidability. No method is known respond to u = v, w = e. The most general type of
to translate the apparent lack of tangible, persis- configuration that admits a finitary description is
tent patterns in rules such as elementary CA the class Crec of recursive configurations, where
Classification of Cellular Automata 193

the assignment of state to a cell is given by a any given semi-decidable set is decidable. It is not
computable function. difficult to adjust these undecidability results to
It is clear that all these subspaces are closed similar classes such as the lower levels of the
under the application of a global map. Except for Li–Packard hierarchy that takes into account spa-
Cfin there are also closed under inverse maps in the tial displacements of patterns.
following sense: given a configuration y in some
subspace that has a predecessor x in Call there Effective Dynamical Systems and Universality
already exists a predecessor in the same subspace, The key property of CA that is responsible for
see Sutner (1991, 2003a). This is obvious except all these undecidability results is the fact that
in the case of recursive configurations. The refer- CA are capable of performing arbitrary compu-
ence also shows that the recursive predecessor tations. This is unsurprising when one defines
cannot be computed effectively from the target computability in terms of Turing machines, the
configuration. Thus, for computational purposes devices introduced by Turing in the 1930s, see
the dynamics of the cellular automaton are best Rogers (1967), Turing (1936). Unlike the
reflected in Cap: it includes all configuration with Gödel–Herbrand approach using general recur-
finite support and we can effectively trace an orbit sive functions or Church’s l-calculus, Turing’s
in both directions. It is not hard to see that Cap is devices are naturally closely related to discrete
the least such class. Alas, it is standard procedure dynamical systems. For example, we can
to avoid minor technical difficulties arising from express an instantaneous description of a Turing
the infinitely repeated spatial patterns and estab- machine as a finite sequence
lish classifications over the subspace Cfin. There is
a arguably not much harm in this simplification al alþ1 . . . a1 p a1 a2 . . . ar
since Cfin is a dense subspace of Call and compact-
ness can be used to lift properties from Cfin to the where the ai are tape symbols and p is a state
full configuration space. of the machine, with the understanding that the
The Culik–Yu hierarchy is correspondingly head is positioned at a1 and that all unspecified
defined over Cfin, the class of all configurations tape cells contain the blank symbol. Needless to
of finite support. In this setting, the first three say, these Turing machine configurations can also
classes of this hierarchy are undecidable and the be construed as finite support configurations of
fourth is undecidable in the disjunctive version: a one-dimensional CA. It follows that a one-
there is no algorithm to test whether a CA admits dimensional CA can be used to simulate an arbi-
undecidable orbits. As it turns out, the CA classes trary Turing machine, hence CA are computational
are complete in their natural complexity classes universal: any computable function whatsoever
within the arithmetical hierarchy (Shoenfield can already be computed by a CA.
1967; Soare 1987). Checking membership in the Note, though, that the simulation is not entirely
first two classes comes down to performing an trivial. First, we have to rely on input/output con-
infinite number of potentially unbounded searches ventions. For example, we may insist that objects
and can be described logically by a P2 expression, in the input domain, typically tuples of natural
a formula of type 8 x ∃ y R(x, y) where R is a numbers, are translated into a configuration of
decidable predicate. Indeed, CY1 and CY2 are the CA by a primitive recursive coding function.
both P2-complete. Thus, deciding whether all Second, we need to adopt some convention that
configurations on a CA evolve to a fixed point is determines when the desired output has occurred:
equivalent to the classical problem of determining we follow the evolution of the input configuration
whether a semi-decidable set is infinite. The third until some “halting” condition applies. Again, this
class is even less amenable to algorithmic attack; condition must be primitive recursively decidable
one can show that CY3 is S3-complete, see Sutner though there is considerable leeway as to how the
(1989). Thus, deciding whether all orbits are end of a computation should be signaled by the
decidable is as difficult as determining whether CA. For example, we could insist that a particular
194 Classification of Cellular Automata

cell reaches a special state, that an arbitrary cell configurations. The behavior of a Turing machine
reaches a special state, that the configuration be a on all instantaneous descriptions, rather than just
fixed point and so forth. Lastly, if and when a the ones that can occur during a legitimate com-
halting configuration is reached, we a apply a putation on some actual input, was first studied by
primitive recursive decoding function to obtain Davis, see Davis (1956, 1957), and also Hooper
the desired output. (1966). Call a Turing machine stable if it halts on
Restricting the space to configurations that have any instantaneous description whatsoever. With
finite support, that are spatially periodic, and so some extra care one can then construct a CA that
forth, produces an effective dynamical system: the lies in the first Culik–Yu class, yet has the same
configurations can be coded as integers in some computational power as the Turing machine.
natural way, and the next configuration relation is Davis showed that every total recursive func-
primitive recursive in the sense that the tion can already be computed by a stable Turing
corresponding relation on code numbers is so prim- machine, so membership in CY1 is not an
itive recursive. A classical example for an effective impediment to considerable computational
dynamical system is given by selecting the instan- power. The argument rests on a particular
taneous descriptions of a Turing machine M as decomposition of recursive functions. Alterna-
configurations, and one-step relation of the Turing tively, one directly manipulate Turing machines
machine as the operation of C. Thus we obtain a to obtain a similar result, see Shepherdson
system A M whose orbits represent the computa- (1965), Sutner (1989). On the other hand, unsta-
tions of the Turing machine. Likewise, given the ble Turing machines yield a natural and coding-
local map r of a CA we obtain a system A r whose free definition of universality: a Turing machine
operation is the induced global map. While the full is Davis-universal if the set of all instantaneous
configuration space Call violates the effectiveness description on which the machine halts is S1-
condition, any of the spaces Cper, Cfin, Cap and Crec complete.
will give rise to an effective dynamical system. The mathematical theory of infinite CA is argu-
Closure properties as well as recent work on the ably more elegant than the actually observable finite
universality of elementary CA 110, see section case. As a consequence, classifications are typically
“Conclusion”, suggests that the class of almost concerned with CA operating on infinite grids, so
periodic configurations, also known as back- that even a configuration with finite support can
grounds or wallpapers, see Cook (2004), Sutner carry arbitrarily much information. If we restrict
(2003a), is perhaps the most natural setting. Both our attention to the space of configurations on a
Cfin and Cap provide a suitable setting for a CA that finite grid a more fine-grained analysis is required.
simulates a Turing machine: we can interpret A M For a finite grid of size n the configuration space has
as a subspace of A r for some suitably constructed the form Cn = [n] ! S and is itself finite, hence any
one-dimensional CA r; the orbits of the subspace orbit is ultimately periodic and the Reachability
encode computations of the Turing machine. It Problem is trivially decidable. However, in practice
follows from the undecidability of the Halting there is little difference between the finite and infi-
Problem for Turing machines that the Reachability nite case. First, computational complexity issues
Problem for these particular CA is undecidable. make it practically impossible to analyze even sys-
Note, though, that orbits in A M may well be tems of modest size. The Reachability Problem for
finite, so some care must be taken in setting up the finite CA, while decidable, is PSPACE-complete
simulation. For example, one can translate halting even in the one-dimensional case. Computational
configurations into fixed points. Another problem hardness appears in many other places. For exam-
is caused by the worst-case nature of our classifi- ple, if we try to determine whether a given config-
cation schemes: in Turing machines and their uration on a finite grid is a Garden-of-Eden the
associated systems A M it is only behavior on problem turns out to be NLOG-complete in dimen-
specially prepared initial configurations that mat- sion one and ℕℙ-complete in all higher dimensions,
ters, whereas the behavior of a CA depends on all see Sutner (1995).
Classification of Cellular Automata 195

Second, it stands to reason that the more inter- p(x) = i iff x  Ci. Hence the projection translates
esting classification problem in the finite case orbits in the full space Call into a class W of
takes the following parameterized form: given a o-words over Sn, the symbolic orbits of the sys-
local map together with boundary conditions, tem. The Cantor space Sℤn together with the shift
determine the behavior of r on all finite grids. describes all logically possible orbits with respect
Under periodic boundary conditions this comes to the given partition and W describes the sym-
down to the study of Cper and it seems that there is bolic orbits that actually occur in the given
little difference between this and the fixed bound- CA. The shift operator corresponds to an applica-
ary case. Since all orbits on a finite grid are ulti- tion of the global map of the CA. The finite factors
mately periodic one needs to apply a more fine- of W provide information about possible finite
grained classification that takes into account tran- traces of an orbit when filtered through the given
sient lengths. It is undecidable whether all config- partition. Whole orbits, again filtered through the
urations on all finite grids evolve to a fixed point partition, can be described by o-words. To tackle
under a given local map, see Sutner (1990). Thus, the classification of the CA in terms of W it was
there is no algorithm to determine whether suggested by Delvenne et al., see Delvenne et al.

  (2006), to refer to the CA as decidable if there it is
hC n , !i8x∃z x ! z and z ! z decidable whether W has non-empty intersection
with a o-regular language. Alas, decidability
in this sense is very difficult, its complexity
for all grid sizes n. The transient lengths are triv- being S11 -complete and thus outside of the arith-
ially bounded by kn where k is the size of the metical hierarchy. Likewise it is suggested to call
alphabet of the automaton. It is undecidable a CA universal if the problem of deciding whether
whether the transient lengths grow according to the cover of W, the collection of all finite factors, is
some polynomial bound, even when the polyno- S1-complete, in analogy to Davis-universality.
mial in question is constant.
Restrictions of the configuration space are one
way to obtain an effective dynamical system. Computational Equivalence
Another is to interpret the approximation-based
notion of computability on the full space in terms In recent work, Wolfram suggests a so-called
of topology. It is well-known that computable Principle of Computational Equivalence, or PCE
maps Call ! Call are continuous in the standard for short, see Wolfram (2002b, p. 717). PCE states
product topology. The clopen sets in this topology that most computational processes come in only
are the finite unions of cylinder sets where a two flavors: they are either of a very simple kind
cylinder set is determined by the values of a con- and avoid undecidability, or they represent a uni-
figuration in finitely many places. By a celebrated versal computation and are therefore no less com-
result of Hedlund the global maps of a CA on the plicated than the Halting Problem. Thus, Wolfram
full space are characterized by being continuous proposes a zero-one law: almost all computational
and shift-invariant. Perhaps somewhat counter- systems, and thus in particular all CA, are either as
intuitively, the decidable subsets of Call are quite complicated as a universal Turing machine or are
weak, they consist precisely of the clopen sets. computationally simple. As evidence for PCE
Now consider a partition of Call into finitely many Wolfram adduces a very large collection of simu-
clopen sets C0, C2, . . ., Cn1. Thus, it is decidable lations of various effective dynamical systems
which block of the partition a given point in the such as Turing machines, register machines, tag
space belongs to. Moreover, Boolean operations systems, rewrite systems, combinators, and cellu-
on clopen sets as well as application of the global lar automata. It is pointed out in Chap. 3 of
map and the inverse global map are all comput- Wolfram (2002b), that in all these classes of sys-
able. The partition affords a natural projection tems there are surprisingly small examples that
p : Call ! Sn where Sn = {0, 1, . . ., n  1} and exhibit exceedingly complicated behavior – and
196 Classification of Cellular Automata

presumably are capable of universal computation. reals (decidable) or of Diophantine equations

Thus it is conceivable that universality is a rather (equivalent to the Halting Problem). No natural
common property, a property that is indeed shared examples of intermediate semi-decidable sets are
by all systems that are not obviously simple. Of known to date.
course, it is often very difficult to give a complete Nonetheless, given an intermediate set A one
proof of the computational universality of a natu- can construct a one-dimensional CA whose
ral system, as opposed to carefully constructed Reachability Problem has the same degree as A.
one, so it is not entirely clear how many of Wol- This suggests a degree-based classification: given
fram’s examples are in fact universal. As a case in any computably enumerable degree d, define the
point consider the universality proof of Conway’s class ℂd to consist of all CA whose Reachability
Game of Life, or the argument for elementary CA Problem has degree exactly d, see Sutner (2002,
110. If Wolfram’s PCE can be formally 2003b). The degree classification is non-trivial in
established in some form it stands to reason that the sense that every class is non-empty. Note that
it will apply to all effective dynamical systems and the first three Culik–Yu classes are all contained in
in particular to CA. Hence, classifications of CA ℂ0 whereas ℂ00 comprises all computationally
would be rather straightforward: at the top there universal CA. Unsurprisingly, it is again
would be the class of universal CA, directly pre- undecidable whether a CA belongs to any partic-
ceded by a class similar to the third Culik–Yu ular class. At the bottom end of the hierarchy it is
class, plus a variety of subclasses along the lines S3-complete to determine membership in ℂ0; at
of the lower Li–Packard classes. the top end it is S4-complete to determine mem-
The corresponding problem in classical com- bership in ℂ00 . Thus, it is easier to determine
putability theory was first considered in the 1930s decidability than universality. In general, deciding
by Post and is now known as Post’s Problem: is membership in ℂd is Sd3 -complete for any semi-
there a semi-decidable set that fails to be decid- decidable degree d. Similar results hold for the
able, yet is not as complicated as the Halting Set? analogous cumulative classes ℂd = [e  dℂe.
In terms of Turing degrees the problem thus is to Unlike the Culik–Yu classification, the struc-
construct a semi-decidable set A such that 0<T A ture of the degree classification between ℂ0 and


<T 00 , or to rule out the existence of any such set, ℂ00 is exceedingly complicated. For example, the


see Lerman (1983), Rogers (1967), Soare (1987) proof of the Friedberg–Muchnik theorem shows
for background on Turing degrees in general and that there are incomparable semi-decidable
semi-decidable degrees in particular. Post’s Prob- degrees d1 and d2. Hence there is are CA whose
lem resisted all attempts at resolution until orbits are undecidable but not as complicated as
Friedberg and Muchnik independently and almost the Halting Problem. Indeed, complete knowl-
simultaneously discovered a way to construct a set edge of the orbits of one of the two CA will not
of intermediate complexity, see Friedberg (1957), help in deciding membership in the orbits of the
Muchnik (1956). The construction is based on the other. Another surprising result in the theory of
idea of a so-called priority argument and is signif- computably enumerable degrees is Sack’s Density
icantly more complicated than any construction of Theorem, see Soare (1987): between any two
semi-decidable sets previously known (Soare computably enumerable degrees d1 < d2 there
1987). Indeed, priority arguments have since lies a third: d1 < d < d2. Thus, between any two
become the hallmark of computability theory CA of strictly increasing complexity there is an
and have even engendered some criticism as infinite and dense hierarchy of other CA. The
being so very technical that, occasionally, the computably enumerable degrees form a semi-
proofs seem to attract more attention than the lattice, so it is natural to try to understand the
theorems being established, see Wang (1993). complexity of the structure by analyzing its first
Be that as it may, it is striking how much more order theory. It is well-known that the S1-theory
artificial and ad hoc intermediate sets are, as com- of this semi-lattice is decidable. However, the
pared to natural examples such as the theory of the reason for this decidability result lies in the fact
Classification of Cellular Automata 197

that any countable partial order can be embedded considering only the reachability relation instead
into the semi-lattice so that the relative computa- of a whole segment of the orbit we also achieve
tional strength of cellular automata is indeed arbi- information-hiding, much as in the classical
trarily complicated. On the other hand, the full Friedberg–Muchnik construction.
theory of the semi-lattice of semi-decidable degrees
is known to be highly undecidable, see Harrington
and Shelah (1982); its degree is 0ðoÞ . One might Conclusion


hope that restriction to reversible CAwould simplify
the situation somewhat. Somewhat surprisingly it Classification schemes of cellular automata based
turns out that each class ℂd already contains an on the long-term evolution of pattern are typically
irreversible CA, see Sutner (2004), so the same undecidable, even if the property in question can be
difficulties arise in the classification of reversible expressed in a fairly week system. While it is easy
CA as in the classification of ordinary CA. to construct examples of CA in particular classes it
While reachability is arguably the most basic is usually very difficult to establish the position of
relation between configurations, similar difficul- a given CA in a particular classification. An excel-
ties also arise with confluence. As a matter of fact, lent example for the difficulty of analyzing a given
one can construct a CA whose Reachability Prob- CA is Cook’s proof of the universality of elemen-
lem has complexity some arbitrarily chosen tary CA number 110 whose local rule is given by
computably enumerable degree d1 while the Con- rðx,y,zÞ ¼ ðx ^ y ^ zÞ y z where denotes
fluence Problem for the same CA has degree d2, exclusive or, see Cook (2004), ▶ “Universality of
another arbitrarily chosen computably enumera- Cellular Automata”. The argument shows that
ble degree. Thus, a classification according to cyclic tag systems, which are known to be com-
reachability is entirely independent of a plete, can be simulated by elementary CA 110 pro-
confluence-based classification. vided one allows an almost periodic background.
How do these results relate to PCE? Wolfram Recent work by Turlough and Woods has shown
would not accept any of the intermediate classes that the whole simulation can be effected with only
of CA as a counterexample to PCE. The argument a polynomial slow-down, see Neary and Woods
is that though intermediate degrees exist, their (2006a, b). This result suggests that the appropriate
construction is critically linked to universal com- setting for classifications is the space of almost
putation. While the universal computation is periodic configurations rather than finite ones.
invisible when only the output of the system is In light of the successful analysis of elementary
observed, the associated computational process CA 110 it is tempting to ask about the classifica-
includes the whole computation and is thus uni- tion of elementary CA 30. Figure 3 shows a seg-
versal. As a case in point, consider the standard ment of the orbit of a one-point seed configuration
Friedberg–Muchnik construction for an interme- under rule 30. It is striking how chaotic and appar-
diate semi-decidable set A. The construction actu- ently random the image is. As a matter of fact, rule
ally builds two semi-decidable sets A and B that 30 has been used for many years as the default
are mutually incomparable with respect to Turing random number generator in the commercial
reducibility. Only A is output and B remains hid- computer algebra system Mathematica, see Wolfram
den. However, even ignoring all the intricate tech- (2002a). The underlying local map is simply
nical details of the whole construction, if we r(x, y, z) = x (y _ z). Alas, there appear to be
consider both A and B as output then the compu- no structures in the evolution of configurations
tation is indeed universal: the disjoint union A B under rule 30 such as “moving particles” that
is S1-complete, see Soare (1972). It remains to be might be exploited in a universality argument
seen if similar arguments can be put forth in along the lines of rule 110. On the other hand, it
connection with priority-free constructions of is unclear how a decision procedure for
intermediate degrees or if natural examples of reachability could be developed. This makes it
intermediate sets can be found. At any rate, by tempting to conjecture that rule 30 in Cap might
198 Classification of Cellular Automata

be a member of one of the intermediate classes ℂd, computations; see Fig. 4 (Mazoyer 1987).
though at present there seems to be no way to A more recent example is Cook’s ingenious
either establish or refute this conjecture. method of using natural gliders in elementary CA
While undecidability results rule out the pos- 110 to implement a cyclic tag system in Cap, thereby
sibility of automatic classification mechanisms establishing computational universality of rule
there is still ample room for the development of 110, see Cook (2004). Notable here is the fact that
sufficient criteria for membership in certain clas- the automaton was fixed from the start and the
ses, see Adamatzky (1994), Vorhees (1996), appropriate coding mechanisms had to be devel-
Wuensche (1999). For example, a proof of com- oped in a very constrained environment. This is in
putational universality in a CA that has not been stark contrast to other hardness arguments where
artificially constructed to simulate some other the CA is carefully constructed to display the
device often rests on the presence of “particles” desired behavior. Careful visual inspection of rule
or “gliders” that can be used to send “signals” 110 orbits was a crucial component in Cook’s proof,
between spatially separated locations. More- it is difficult to imagine that the result could have
over, one has to be able to process these signals been established in a purely combinatorial or alge-
much in the way of Boolean logic gates, to store braic fashion. One can envision an interactive
state and so forth. A good example for compli- software system that helps to tackle some algorith-
cate interactions between signals are the various mically unsolvable classification problems in spe-
solutions to the firing squad problem, albeit cial cases, much as Baumslag’s Magnus project in
not in the context of simulating arbitrary group theory, see Baumslag (2007).

Classification of Cellular
Automata, Fig. 3 A
pseudo-random pattern
generated by elementary
CA 30

Classification of Cellular
Automata,
Fig. 4 Interacting signals
in Mazoyer’s optimal
solution to the firing squad
problem
Classification of Cellular Automata 199

Bibliography Head T (1989) Linear CA: injectivity from ambiguity.

Complex Syst 3(4):343–348
Adamatzky A (1994) Identification of cellular automata. Hedlund GA (1969) Endomorphisms and automorphisms
Taylor & Francis, London of the shift dynamical system. Math Syst Theory
Amoroso S, Patt YN (1972) Decision procedures for 3:320–375
surjectivity and injectivity of parallel maps for Hooper PK (1966) The undecidability of the Turing
tesselation structures. J Comput Syst Sci 6:448–464 machine immortality problem. J Symb Log
Baldwin JT (2002) Computation versus simulation. 31(2):219–234
http://www.math.uic.edu/%7Ejbaldwin/pub/cafom.ps. Kari J (1990) Reversibility of 2D cellular automata is
Accessed May 2007 undecidable. Physica D 45:397–385
Baldwin JT, Shelah S (2000) On the classifiability of cel- Kůrka P (1997) Languages, equicontinuity and attractors in
lular automata. Theor Comput Sci 230(1–2):117–129 cellular automata. Ergod Th Dyn Syst 17:417–433
Baumslag G (2007) Magnus. http://caissny.org/. Accessed Kůrka P (2003) Topological and symbolic dynamics,
May 2007 Cours Spécialisés, vol 11. Societe Mathematique de
Beal M-P, Perrin D (1997) Symbolic dynamics and finite France, Paris
automata. In: Rozenberg G, Salomaa A (eds) Hand- Langton CG (1990) Computation at the edge of chaos.
book of formal languages. Springer, Berlin Physica D 42:12–37
Bennett CH (1973) Logical reversibility of computation. Lecerf Y (1963) Machine de Turing réversible. Insolubilité
IBM J Res Dev 17:525–532 récursive en n  N de l’équation u = ynu, où y est un
Börger E, Grädel E, Gurevich Y (2001) The classical “isomorphisme de codes”. C R Acad Sci Paris
decision problem. Springer, Berlin 257:2597–2600
Cook M (2004) Universality in elementary cellular autom- Lerman M (1983) Degrees of unsolvability. Perspectives in
ata. Complex Syst 15(1):1–40 mathematical logic. Springer, Berlin
Culik K (1987) On invertible cellular automata. Complex Li W, Packard N (1990) The structure of the
Syst 1(6):1035–1044 elementary cellular automata rule space. Complex Syst
Culik K, Sheng Y (1988) Undecidability of CA classifica- 4(3):281–297
tion schemes. Complex Syst 2(2):177–190 Li W, Packard N, Langton CG (1990) Transition phenom-
Davis M (1956) A note on universal Turing machines. In: ena in CA rule space. Physica D 45(1–3):77–94
Shannon CE, McCarthy J (eds) Automata studies. Libkin L (2004) Elements of finite model theory. Springer,
Annals of mathematics studies, vol 34. Princeton Uni- Berlin
versity Press, Princeton, pp 167–175 Manzini G, Margara L (1999) A complete and efficiently
Davis M (1957) The definition of universal Turing computable topological classification of D-dimensional
machines. Proc Am Math Soc 8:1125–1126 linear cellular automata over Zm. Theor Comput Sci
Delorme M, Mazoyer J (1999) Cellular automata: a parallel 221(1–2):157–177
model. Mathematics and its applications, vol 460. Mazoyer J (1987) A six state minimal time solution to the
Kluwer, Dordrecht firing squad synchronization problem. Theor Comput
Delvenne J-C, Kůrka P, Blondel V (2006) Decidability and Sci 50:183–238
universality in symbolic dynamical systems. Mitchell M, Crutchfield JP, Hraber PT (1994) Evolving
Fundamenta Informaticae 74(4):463–490 cellular automata to perform computations: mecha-
Durand-Lose J (2001) Representing reversible cellular nisms and impediments. Physica D (75):361–369
automata with reversible block automata. Disc Math Morita K (1994) Reversible cellular automata. J Inf Pro-
Theor Comp Sci Proc AA:145–154 cess Soc Japan 35:315–321
Elgaard J, Klarlund N, Møller A (1998) MONA 1.x: new Morita K (1995) Reversible simulation of one-dimensional
techniques for WS1S and WS2S. In: Proceeding of the irreversible cellular automata. Theor Comput Sci
10th international conference on computer-aided verifica- 148:157–163
tion, CAV ’98. LNCS, vol 1427. Springer, Berlin, Morita K, Harao M (1989) Computation universality of
pp 516–520 1 dimensional reversible (injective) cellular automata.
Friedberg RM (1957) Two recursively enumerable sets of Trans Inst Electron, Inf Commun Eng E 72:758–762
incomparable degrees of unsolvability. Proc Natl Acad Muchnik AA (1956) On the unsolvability of the problem of
Sci U S A 43:236–238 reducibility in the theory of algorithms. Dokl Acad
Grädel E, Thomas W, Wilke T (eds) (2002) Automata, Nauk SSSR 108:194–197
logics, and infinite games, LNCS, vol 2500. Springer, Neary R, Woods D (2006a) On the time complexity of
Berlin 2-tag systems and small universal turing machines. In:
Gutowitz H (1996a) Cellular automata and the sciences of FOCS. IEEE Comput Soc, Berkeley, pp 439–448
complexity, part I. Complexity 1(5):16–22 Neary R, Woods D (2006b) Small fast universal turing
Gutowitz H (1996b) Cellular automata and the sciences of machines. Theor Comput Sci 362(1–3):171–195
complexity, part II. Complexity 1(6) Oliveira G, Oliveira P, Omar N (2001) Definition and
Harrington L, Shelah S (1982) The undecidability of the application of a five-parameter characterization of
recursively enumerable degrees. Bull Am Math Soc one-dimensional cellular automata rule space. Artif
6:79–80 Life 7(3):277–301
200 Classification of Cellular Automata

Ollinger N (2003) The intrinisic universality problem of Sutner K (2003a) Almost periodic configurations on linear
one-dimensional cellular automata. In: Alt H, Habib cellular automata. Fundamentae Informaticae 58(3,4):
M (eds) Proceeings STACS. LNCS, vol 2607. Springer, 223–240
Berlin, pp 632–641 Sutner K (2003b) Cellular automata and intermediate
Packard NH (1988) Adaptation towards the edge of chaos. degrees. Theor Comput Sci 296:365–375
In: Dynamic patterns in complex systems. World Sci- Sutner K (2004) The complexity of reversible cellular
entific, Singapore, pp 29–301 automata. Theor Comput Sci 325(2):317–328
Rogers H (1967) Theory of recursive functions and effec- Taati S (2007) Cellular automata reversible over limit set.
tive computability. McGraw Hill, New York J Cell Autom 2(2):167–177
Rozenberg G, Salomaa A (1997) Handbook of formal Toffoli T, Margolus N (1990) Invertible cellular automata:
languages. Springer, Berlin a review. Physica D 45:229–253
Shepherdson JC (1965) Machine configuration and word Turing AM (1936) On computable numbers, with an appli-
problems of given degree of unsolvability. Z Math cation to the Entscheidungsproblem. P Lond Math Soc
Logik Grundl Math 11:149–175 42:230–265
Shoenfield JR (1967) Mathematical logic. Addison Wesley, Vorhees B (1996) Computational analysis of one-dimensional
Reading cellular automata. World Scientific, Singapore
Soare RI (1972) The Friedberg-Muchnik theorem Wang H (1993) Popular lectures on mathematical
re-examined. Can J Math 24:1070–1078 logic. Dover Publications, Dover/New York
Soare RI (1987) Recursively enumerable sets and Weihrauch K (2000) Computable analysis. EATCS mono-
degrees. Perspectives in mathematical logic. graphs. Springer, Berlin
Springer, Berlin Wolfram S (1984a) Computation theory of cellular autom-
Sutner K (1989) A note on Culik-Yu classes. Complex Syst ata. Commun Math Phys 96(1):15–57
3(1):107–115 Wolfram S (1984b) Universality and complexity in cellular
Sutner K (1990) Classifying circular cellular automata. automata. Physica D 10:1–35
Physica D 45(1–3):386–395 Wolfram S (1985) Twenty problems in the theory of cellu-
Sutner K (1991) De Bruijn graphs and linear cellular lar automata. Phys Scr T9:170–183
automata. Complex Syst 5(1):19–30 Wolfram S (2002a) The mathematica book. Cambridge
Sutner K (1995) The complexity of finite cellular automata. University Press, Cambridge
J Comput Syst Sci 50(1):87–97 Wolfram S (2002b) A new kind of science. Wolfram
Sutner K (2002) Cellular automata and intermediate Media, Champaign
reachability problems. Fundamentae Informaticae Wuensche A (1999) Classifying cellular automata auto-
52(1–3):249–256 matically. Complexity 4(3):47–66
 sequence of m distinct elements of ℤd speci-
Tiling Problem and fying the relative locations of the neighbors of
!
Undecidability in Cellular the cells: A cell located at x  ℤd has m
! ! ! !
Automata neighbors, in positions x þ n 1 , x þ n 2 ,
! !
. . . , x þ n m . Finally, the local update rule f:
Jarkko Kari Sm ! S specifies the new state of a cell, based on
Department of Mathematics, University of the old states of its neighbors. In one step, con-
Turku, Turku, Finland figuration c is transformed into configuration e
!
where, for all x  ℤd ;

Article Outline

 h

! ! !
e x ¼f c x þ n 1 ,
Glossary

i
! ! ! !
Definition of the Subject c x þ n 2 , ...,c x þ n m :
Introduction
The Tiling Problem and Its Variants
Definition of Tiles The mapping c ! e is the global transition
d d

Computations and Tilings function, or the CA function, G : Sℤ ! Sℤ

The Tiling Problem specified by the CA A = (d,S,N,f).
Variants of the Tiling Problem Decision Problem A decision problem is an
Deterministic Tiles algorithmic question with a yes/no answer.
Plane-Filling Directed Tiles The problem has an input (called the instance
Undecidability in Cellular Automata of the problem) and a well-defined answer
Undecidable Properties of One-Dimensional CA “yes” or “no” associated to each instance.
Other Undecidability Results Equicontinuous Cellular Automaton Cellular
Future Directions automaton G is called equicontinuous if, for
Bibliography every finite A  ℤd , there exists a finite
B  ℤd such that any two initial configura-
tions that agree inside B will agree inside A for
Glossary all subsequent steps. In other words


! !
Aperiodic Tile Set A two-dimensional tile set 8 x B : c x ¼


that admits a valid tiling of the plane, but does !
e x ) 8tℕ and
not admit any valid periodic tilings. The

! !
smallest known aperiodic set of Wang tiles 8 x A : Gt ðcÞ x ¼
contains 11 tiles (Jeandel and Rao 2015). See

!
also (Culik 1996; Kari 1996). Gt ðeÞ x :
Cellular Automaton (CA) A d-dimensional cel-
lular automaton consists of an infinite d-dimen- This means that equicontinuous CA can be reli-
sional grid of cells, indexed by ℤd. Each cell ably simulated in finite windows. It is known
stores an element of a finite state set S. Config- that a CA is equicontinuous if and only if it is
uration c: ℤd ! S specifies the states of all cells. ultimately periodic ½17 : ∃n,p  ℕ : Gn ¼ Gnþp .


The set of all configurations is Sℤd. Theneigh- In this sense, the dynamics of equicontinuous
! ! !
borhood vector N ¼ n 1 , n 2 , . . . , n m is a CA is trivial.

# Springer Science+Business Media LLC, part of Springer Nature 2018 201

A. Adamatzky (ed.), Cellular Automata,
https://doi.org/10.1007/978-1-4939-8700-9_552
Originally published in
R. A. Meyers (ed.), Encyclopedia of Complexity and Systems Science, # Springer Science+Business Media New York 2015
https://doi.org/10.1007/978-3-642-27737-5_552-5
202 Tiling Problem and Undecidability in Cellular Automata

Injective (Reversible) Cellular Automaton A decidable whether a given word belongs to L.

CA is injective if every configuration has at The language is called recursively enumerable
most one pre-image, that is, the global transition (RE for short) if this membership problem is
function is one-to-one. It is well known that a semi-decidable.
CA is injective if and only if it is bijective (every Semi-algorithm An algorithm-like procedure
configuration has a unique pre-image), which, in for a decision problem that correctly returns
turn, is equivalent to reversibility. (There exists a positive answer on positive input instances,
an inverse CA that traces the CA back in time). but on negative instances runs forever without
Limit Set The limit set of a CA is its maximal ever returning an answer.
attractor. In other words, it is the compact and Semi-decidability A decision problem is called
translation invariant set semi-decidable if there is a semi-algorithm for
it. For example, the decision problem Turing
1
d machine halting on blank tape is semi-decid-
\ G sℤ able since one can simulate any given Turing
i¼0
machine step-by-step until (if ever) it halts.
where G is the global transition function and S Sensitive Cellular Automata Cellular automa-
is the state set. ton G is called sensitive to initial conditions if
Nilpotent Cellular Automaton Nilpotent CA there exists a finite set B  ℤd of cells such that
has trivial dynamics. A CA is called nilpotent for every configuration c and every finite set A
if its limit set contains only one configuration.  ℤ of cells there exists



a configuration e and
! ! !
time t 0 such that e x ¼ d x for all x
The unique element of the limit set is the


! !
quiescent configuration. This is equivalent to  A but 6¼ Gt ðeÞ x 6¼ Gt ðcÞ x for some
every initial configuration eventually becom- !
x  B . This means that arbitrarily distant
ing the quiescent configuration.
modifications to any configuration c may
Periodic Tiling A tiling that is invariant under
propagate to a fixed observation window B.
some nonzero translation. A two-dimen-
Surjective Cellular Automaton A cellular
sional? tiling is called totally periodic if it is
automaton (CA) is called surjective if every
invariant under two linearly independent
configuration has a pre-image, that is, if its
translations. A totally periodic tiling is auto-
global transition function G : Sℤ ! Sℤ is
d d

matically periodic in horizontal and vertical

surjective.
directions, which means that it consists of a
The Finite Tiling Problem A decision problem
rectangular pattern that is repeated horizon-
where we are given a set T of Wang tiles and a
tally and vertically to fill the plane. A two-
specific blank tile B  T whose all four edges
dimensional tile set that admits a valid peri-
are colored with the same color. A tiling t: ℤ
odic tiling automatically admits also totally
! T is called finite if
periodic tiling.
More generally, a d-dimensional tiling t: ℤ!d ! T
! ! n
 o
is periodic with period p  ℤd , p 6¼ 0 , if ! !
x  ℤ2 jt x 6¼ B



! ! ! !
t x ¼ t x þ p for all x  ℤd. It is totally

! !
periodic if it is periodic with d linearly inde- is a finite set. If t x ¼ B for all x  ℤ2 ;
! ! !
pendent periods p 1 , p 2 , . . . p d . Note that then t is called trivial. The finite tiling prob-
when d > 2, it is possible that a tile set admits lem asks whether there exist nontrivial valid
a periodic tiling but does not admit any totally finite tilings. The problem is undecidable but
periodic tiling. semi-decidable.
Recursive and Recursively Enumerable (RE) The Halting Problem Turing machine halting
A formal language L is called recursive if it is on blank tape is the decision problem whose
Tiling Problem and Undecidability in Cellular Automata 203

input is a Turing machine M = (Q, G, d, q0, qh, b) follows: For all c,e  Sℤd
and the answer is positive if and only if the
Turing machine eventually enters its halting c  A, n e
!
state qh when started in the initial state q0 on a , 8 x  A, 0  t < n :



totally blank tape, i. e., initially every tape ! !
G t ðc Þ x ¼ G t ðe Þ x :
location has the blank symbol b.
The Periodic Tiling Problem The decision
In other words, two configurations are equiv-
problem to determine if a given set of Wang
alent if we cannot observe any difference in
tiles admits a periodic tiling. The problem is
their orbits in region A within the first n time
undecidable, but it is semi-decidable
instances. Let us denote by NG (A,n) the num-
(Gurevich and Koryakov 1972).
ber of equivalence classes of A,n. Then the
The Tiling Problem (Also Known as the Dom-
topological entropy is
ino Problem) The decision problem that gets
as input a finite tile set T and asks whether
hðGÞ ¼ sup lim
there exists a valid tiling by T: The tiling A n!1
problem was proved undecidable for Wang logN G ðA, nÞ
tiles by R.Berger (1966). Its complement n
(i.e., “Does there not exist a valid tiling”) is
semi-decidable. where the supremum is over all finite A  ℤd.
The Tiling Problem with a Seed Tile The deci- The entropy always exists. In the one-dimen-
sion problem that gets as input a tile set T and sional case, the entropy is always a finite,
one tile five and asks whether T admits a tiling nonnegative number. If d2, then the entropy
that contains tiles at least once. The problem is can also be infinite.
undecidable for Wang tiles, but its comple- Turing Machine (TM) Turing machines are
ment is semi-decidable (Wang 1961). computation devices commonly used to for-
Tiling Tiling is a covering of the plane using mally define the concept of an algorithm. They
tiles. A valid Wang tiling by a Wang tile set also provide us with the most basic
T is an assignment t: ℤ2 ! T of tiles to cells undecidable decision problems. A Turing
such that the local matching rule is satisfied machine consists of a finite state control unit
between all adjacent tiles. We say that T that moves along an infinite tape. The tape has
admits tiling t. symbols written in cells that are indexed by ℤ.
A more general definition: a d-dimensional Depending on the state of the control unit and
tiling using tile set T = (d,T,N,R) is a mapping the symbol currently scanned on the tape, the
! machine may overwrite the tape symbol,
t: ℤ2 ! T. Tiling t is valid at cell x  ℤd if
change the internal state, and move along the

tape one cell to the left or right. We formally
! ! !
t x þ n 1, x þ
 define a Turing machine as a six-tuple
!
n 2,
! !
. . . , x þ n m  R: M ¼ ðQ, G, d, q0 , qh , bÞ where Q and G are
finite sets (the state alphabet and the tape
alphabet, respectively), q0,qh  Q are the ini-
Tiling t is called valid if it is valid at every cell tial and the halting states, respectively, b  G
!
x  ℤd . is the blank symbol, and d : Q G ! Q  G
Topological Entropy The topological entropy f1, 1g is the transition function that specifies
h(G) of a CA G measures the complexity the moves of the machine.
of its dynamics. For any finite A  ℤd and A configuration (or instantaneous description)
positive integer n, we define the equivalence of the machine is a triplet (q, i, t) where q  Q is
relation A,n among initial configurations as the current state, i  ℤ is the position of the
204 Tiling Problem and Undecidability in Cellular Automata

machine on the tape, and t: ℤ ! G describes the problems are decision problems, that is, the
content of the tape. In one time step, configu- answer for each input instance is either yes or
ration (q, i, t) becomes ðq0 , i þ d, t0 Þ if no. All problems considered are undecidable,
dðq, tðiÞÞ ¼ ðq0 , y, dÞ and t0 ðiÞ ¼ y and i.e., no algorithm can solve them. We only
0
t ðjÞ ¼ tðjÞ for all j 6¼ i. We denote this move by consider problems whose undecidability is
proved using a reduction from the tiling prob-
ðq, i, tÞ ‘ ðq0 , i þ d, t0 Þ lem or its variant.

The reflexive, transitive closure of is denoted INJECTIVITY

by ‘ , that is
Input: Cellular automaton A.
ðq, i, tÞðq0 , i0 , t0 Þ Question: Is A injective (i.e., reversible)?

if and only if (q0 , i0 , t0 ) can be reached from (q,

i, t) by executing zero or more Turing machine There is an algorithm that solves
moves. INJECTIVITY for one-dimensional CA
(Un)Decidability Some decision problems can- (Amoroso and Patt 1972), but the problem is
not be solved by any algorithm. Such prob- undecidable among two-dimensional CA (Kari
lems are called undecidable. In contrast, 1990, Kari 1994a). The problem is semi-
decidable decision problems are solved by decidable in any dimension.
some algorithm. An example of an
undecidable problem is the Turing machine
SURJECTIVITY
halting on blank tape.
Wang Tile A Wang tile is a unit square tile with
Input: Cellular automaton A.
colored edges. Tiles have an orientation, i.e.,
Question: Is A surjective?
they may not be rotated or reflected. The
colors give a local matching rule that spec-
ifies which tiles may be placed next to each
other: Two adjacent tiles must have identical Also, SURJECTIVITY is decidable among
colors on the abutting edges. A Wang tile set one-dimensional CA. The two-dimensional ques-
consists of a finite number of Wang tiles.A tion is, however, undecidable (Kari 1994a). The
more general definition: a d-dimensional tile complement of the problem (i.e., non-surjectivity)
set is a quadruple T = (d, T,N,R) where T is a is semi-decidable in any dimension.
finite set whose elements are called tiles,


! ! !
N ¼ n 1 , n 2 , . . . , n m is a neighborhood NILPOTENCY
vector of m distinct elements of ℤ , and R d

Tm is a relation of allowed patterns. The Input: Cellular automaton A.

neighborhood vector has the same interpreta- Question: Is A nilpotent?
tion as in the definition of cellular automata:
it gives the relative locations of the neighbors
of cells. Nilpotency is undecidable even among
one-dimensional CA (Kari 1992). It is
undecidable even if A has a spreading state, i.e.,
Definition of the Subject a state q such that any cell whose neighborhood
contains q becomes q. However, NILPOTENCY
We consider the following algorithmic ques- is semi-decidable in any dimension. Based on
tions concerning cellular automata. All problem NILPOTENCY, we can prove Rice’s
Tiling Problem and Undecidability in Cellular Automata 205

theorem for cellular automaton limit sets: any is surjective or injective, whether a
nontrivial decision problem concerning limit one-dimensional CA is nilpotent or equicontinuous,
sets is undecidable (Kari 1994b). and whether the topological entropy is less than
some constant.
TOPOLOGICAL ENTROPY The fact that the tiling problem is closely related
to cellular automata is not surprising considering
Input: Cellular automaton A. their apparent similarity: both involve assignments
Question: Is the topological entropy of A less of symbols over a finite alphabet onto integer lattice
than constant c > 0? points. The difference is that tilings are static while
cellular automata change the assignments dynami-
cally according to the local rule. Undecidability of
the tiling problem on the two-dimensional plane
Problem TOPOLOGICAL ENTROPY is naturally leads to undecidability results concerning
undecidable for every constant c > 0, even in the single-step properties of two- and higher-
one-dimensional case. This can be proven using a dimensional cellular automata. But, also, asymp-
direct reduction from NILPOTENCY (Hurd totic properties of one-dimensional cellular autom-
et al. 1992). Also, direct reductions from ata can be related to tiling problems by viewing the
NILPOTENCY prove the undecidability of the space-time diagram of the CA as a tiling. This
following two problems (Durand et al. 2003; Kari naturally leads to the definition of deterministic
2008): tile sets: Wang tiles where tiles are uniquely deter-
mined by some of their neighbors.
EQUICONTINUITY We start by discussing the tiling problem and
its variants. We do not prove the undecidability of
Input: Cellular automaton A. all the variants. Rather, literature references for
Question: Is A equicontinuous? the proofs are provided. We then define a partic-
ular tile set that has an interesting plane-filling
property. This tile set is a useful tool in the reduc-
tion to prove that it is undecidable to tell whether
SENSITIVITY TO INITIAL CONDITIONS a given two-dimensional CA is injective
(reversible). We then provide reductions that
Input: Cellular automaton A. show several questions concerning cellular
Question: Is A sensitive to initial conditions? automata undecidable.

The Tiling Problem and Its Variants

Introduction
In this section, we discuss the tiling problem and
several of its variants.
Several decision problems concerning cellular
automata are known to be undecidable, that is, no
algorithm exists that solves them. Some
undecidability results easily follow from the uni- Definition of Tiles
versal computation capabilities of cellular autom-
ata, while others require more elaborate proofs. For our purposes, it is convenient to define tiles
Reductions from the tiling problem and its variants in a way that most closely resembles cellular
turn out to be useful in proving various questions automata. In the d-dimensional cellular space,
concerning CA undecidable. We consider the prob- the cells are indexed by ℤd. A neighborhood
lems of determining if a given two-dimensional CA vector
206 Tiling Problem and Undecidability in Cellular Automata

a a b a′ c a′ d qa e qa

q′ q′ q q

a qa qa a a

Tiling Problem and Undecidability in Cellular Automata, Fig. 1 Machine tiles associated to a Turing machine


N¼
! ! !
n 1, n 2, . . . , n m Computations and Tilings

! ! The basic observation in establishing undecid-

consists of m distinct elements n i  ℤd . Each n i ability results concerning tilings is the fact that
specifies the relative location of a neighbor of valid tilings can be forced to contain a complete
each cell. simulation of a computation by a given Turing
More precisely, the ith neighbor of the cell in machine. To any given Turing machine
! ! !
position x  ℤd is located at x þ n i . M ¼ ðQ, G, d, q0 , qh , bÞ , we associate the Wang
A tile set is a finite set T whose elements are tiles shown in Fig. 1, and we call these tiles the
called tiles. A local matching rule tells which machine tiles of M . Note that in the illustrations,
patterns of tiles are allowed in valid tilings. The instead of colors, we use labeled arrows on the sides
matching rule is given as an m-ary relation R  Tm of the tiles. Two adjacent tiles match if and only if
where m is the size of the neighborhood. Tilings an arrowhead meets an arrow tail with the same
are assignments label. Such arrow representation can be converted
into the usual coloring representation of Wang tiles
t : ℤd ! T by assigning to each arrow direction and label a
!
unique color.
of tiles into cells. Tiling t is valid at x  ℤd if The machine tiles of M contain the following

tiles:
! ! ! !
t x þ n 1, x þ n 2,
 1. For every tape letter a  G a tape tile of Fig. 1a.
! !
. . . , x þ x m  R: 2. For every tape letter a  G and every state q  Q
an action tile of Fig. 1b or c. Tile (b) is used if
Tiling t is called valid if it is valid at every
!
position x  ℤd . dðq, aÞ ¼ ðq0 , a0 ,  1Þ;
A convenient – and historically earlier – way
of defining tiles is in terms of edge labelings. and tile (c) is used if
A Wang tile is a two-dimensional unit square
with colored edges. The local matching rule is dðq, aÞ ¼ ðq0 , a0 þ 1Þ:
determined by these colors: A tiling is valid at
!
position x  ℤ2 if each of the four edges of 3. For every tape letter a  G and non-halting state
!
the tile in position x has the same color as q  Qnfqh g, two merging tiles shown in Fig. 1d.
the abutting edge in the adjacent tile. Clearly,
this is a two-dimensional tile set with the neigh- The idea of the tiles is that a configuration of
borhood vector ½ð1, 0Þ, ð1, 0Þ, ð0,  1Þ, ð0, 1Þ the Turing machine M is represented as a row of
and a particular way of defining the local tiles in such a way that the cell currently scanned
relation R. by M is represented by an action tile, its neighbor
Tiling Problem and Undecidability in Cellular Automata 207

where the machine moves into has a merging tile until (if ever) a region is found that cannot be
and all other tiles on the row are tape tiles. If this properly tiled. Note also that a semi-algorithm
row is part of a valid tiling, then it is clear that the exists for those tile sets that admit a valid,
rows above must be similar representations of totally periodic tiling: All totally periodic til-
subsequent configurations in the Turing machine ings can be effectively enumerated, and it is a
computation, until the machine halts. simple matter to test each for validity of the
The machine tiles above are the basic tiles tiling constraint. Combining the two semi-
associated to Turing machine M: Additional algorithms above yields a semi-algorithm that
tiles will be added depending on the actual vari- correctly identifies tile sets that (i) do not
ant of the tiling problem. admit any valid tiling or (ii) admit a valid
periodic tiling. Only aperiodic tile sets fail to
satisfy either (i) or (ii), so we see that the
The Tiling Problem existence of aperiodic tile sets is implied by
Theorem 1.
The tiling problem is the decision problem of In the following sections, we consider some
determining if at least one valid tiling is admitted variants of the tiling problem whose
by the given set of tiles. undecidability is easier to establish.

TILING PROBLEM
Variants of the Tiling Problem
Input: Tile set T.
Question: Does T admit a valid tiling?
TILING PROBLEM WITH A SEED TILE

Input: Tile set T and one tile s.

The tiling problem is easily seen decidable if Question: Does T admit a valid tiling such that
the input is restricted to one-dimensional tile sets. tile 5 is used at least once?
It is a classical result by R. Berger that the tiling
problem of two-dimensional tiles is undecidable,
even if the input consists of Wang tiles (Berger
1966; Robinson 1971). The seeded version was shown undecidable by
H. Wang (1961). We present the proof here
Theorem 1 because the proof is quite simple and shows the
TILING PROBLEM is undecidable for Wang tile general idea of how Turing machine halting prob-
set T. The complement problem (nonexistence of lem can be reduced to problems concerning tiles.
valid tilings) is semi-decidable.
We do not prove this result here. The Theorem 2
undecidability proofs in (Berger 1966; Robinson TILING PROBLEM WITH A SEED TILE is
1971) are based on an explicit construction of an undecidable for Wang tile sets. The complement
aperiodic tile set such that additional tiles problem is semi-decidable.
implementing Turing machine simulations can
be embedded in valid tilings. The aperiodic set Proof
is needed to force the presence of tiles that initiate The semi-decidability of the complement problem
Turing machine simulation in arbitrarily large follows from the following semi-algorithm:
regions. For r = 1,2,3,. . ., try all tilings of the radius
Note that semi-decidability of the comple- r square around the origin to see if there is a
ment problem is apparent: a semi-algorithm valid tiling of the square such that the origin
simply tries to tile larger and larger regions contains the seed tile s. If for some r such a tiling
208 Tiling Problem and Undecidability in Cellular Automata

a b b q b b
0

Tiling Problem and Undecidability in Cellular Automata, Fig. 2 (a) The blank tile and (b) three initialization tiles

is not found, then halt and report that there is no the seed row to contain the consecutive config-
tiling containing the seed tile. urations of the Turing machine. There is no
Consider then undecidability. We reduce the merge tile containing a halting state, so the
decision problem TURING MACHINE HALT- Turing machine does not halt – otherwise, a
ING ON BLANK TAPE, a problem that is well valid tiling could not be formed.
known to be undecidable. For any given Turing
machine M, we can effectively construct a tile set Conclusion: Suppose we had an algorithm
and a seed tile in such a way that they form a that solves TILING PROBLEM WITH
positive instance of TILING PROBLEM WITH A SEED TILE. Then we also have an algo-
A SEED TILE if and only if M is a negative rithm (which simply constructs the tile set as
instance of TURING MACHINE HALTING ON above and determines if a tiling with seed tile
BLANK TAPE. For the given Turing machine M, exists) that solves TURING MACHINE
we construct the machine tiles of Fig. 1 as well as HALTING ON BLANK TAPE. This contra-
the four tiles shown in Fig. 2. These are the blank dicts the fact that this problem is known to be
tile and three initialization tiles. They initialize all undecidable.
tape symbols to be equal to blank b and the Turing In the following tiling problem variant, we are
machine to be in the initial state q0. The middle given a Wang tile set T and specify one tile B  T
initialization tile is chosen as the seed tile s. as the blank tile. The blank tile has all four sides
Let us prove that a valid tiling containing a copy colored by the same color. A finite tiling is a tiling
of the seed tile exists if and only if the Turing where only a finite number of tiles are non-blank.
machine M does not halt when started on the A finite tiling where all tiles are blank is called
blank tape: trivial.

“(”: Suppose that the Turing machine M does FINITE TILING PROBLEM
not halt on the blank tape. Then a valid tiling
exists where one horizontal row is formed Instance: A finite set T of Wang tiles and a
with the initialization tiles, all tiles below blank tile B  T.
this row are blank, and the rows above the Problem: Does there exist a valid finite tiling
initialization row contain consecutive config- that is not trivial?
urations of the Turing machine.
“)”: Suppose that a valid tiling containing the
middle initialization tile exists. The seed tile
forces its row to be formed by the initialization Theorem 3
tiles, representing the initial configuration of the The FINITE TILING PROBLEM is undecidable.
Turing machine on the blank tape. The machine It is semi-decidable while its complement is not
tiles force the following horizontal rows above semi-decidable.
Tiling Problem and Undecidability in Cellular Automata 209

L a R
b a L R

a qh

a L R
a L b q b b R
0

qh

Tiling Problem and Undecidability in Cellular Automata, Fig. 3 (a) The blank tile B, (b) halting tiles, and (c)
border tiles

Proof Let us prove that the tile set admits a valid,

For semi-decidability, notice that we can try all finite, nontrivial tiling if and only if the Turing
valid tilings of larger and larger squares until we machine halts on the empty tape.
find a tiling of a square where all tiles on the
boundary are blank, while some interior tile is “(”: Suppose that the Turing machine halts on the
different from the blank tile. If such a tiling is blank tape. Then a tiling exists where the bound-
found, then the semi-algorithm halts, indicating ary tiles isolate a finite portion of the plane
that a valid, finite, nontrivial tiling exists. (a “board”) for the simulation of the Turing
To prove the undecidability, we reduce the machine, the bottom tiles of the board initialize
problem TURING MACHINE HALTING ON the Turing machine on the blank tape, and inside
BLANK TAPE. For any given Turing machine the board the Turing machine is simulated until
M, we construct the machine tiles of Fig. 1 as well it halts. After halting, only tape tiles are used
as the blank tile, boundary tiles, and the halting until they are absorbed by the topmost row of the
tiles shown in Fig. 3. board. If the board is made sufficiently large, the
The halting tiles of Fig. 3b are constructed for entire computation fits inside the board, so the
all tape letters a  G and the halting state qh. The tiling is valid. All tiles outside the board are
purpose of the halting tiles is to erase the Turing blank, so the tiling is finite.
machine from the configuration once it halts. The “)”: Suppose then that a finite, nontrivial tiling
lower border tiles of Fig. 3c initialize the config- exists. The only non-blank tiles with a blank
uration to consist of the blank tape symbol b and bottom edge are the lower border tiles of
the initial state q0. The top border tiles are made Fig. 3c, so the tiling must contain a lower
for every tape symbol a  G. They allow the border tile. Horizontal neighbors of lower bor-
absorption of the configuration as long as the der tiles are lower border tiles, so we see that
Turing machine has been erased. The border the only way to have a finite tiling is to have a
tiles on the sides are labeled with symbols L and contiguous lower border that ends at both
R to identify the left and the right border of the sides in a corner tile where the border turns
computation area. upwards. The vertical borders must
210 Tiling Problem and Undecidability in Cellular Automata

Tiling Problem and b

Undecidability
in Cellular Automata,
Fig. 4 NW-deterministic
sets of Wang tiles: (a) there
is at most one matching tile a
z for any x and y and (b)
diagonals of
NW-deterministic tilings y
interpreted as
configurations of x z
one-dimensional CA

again – due to the finiteness of the tiling – end Deterministic Tiles

at corners where the top border starts. All in
all, we see that the boundary tiles are forced to The tiling problem of one-dimensional tiles is
form a rectangular board. decidable. However, tiles can provide
undecidability results for one-dimensional CA
The lower boundary of the board initializes the when we use the trick that we view space-time
Turing machine configuration on the blank tape, diagrams as two-dimensional tilings. But not
and the rows above it are forced by the machine every tiling can be a space-time diagram of a
tiles to simulate consecutive configurations of the CA: the tiling must be locally deterministic in the
Turing machine. Because the Turing machine direction that corresponds to time. This leads to the
state symbol is not allowed to touch the side or consideration of determinism in Wang tiles.
the upper boundary of the board, the Turing Consider a set T of Wang tiles, i.e., squares with
machine must be erased by a halting tile, i.e., colored edges. We say that T is NW-deterministic if
the Turing machine must halt. for all a, b  T, a 6¼ b; either the upper (northern)
The third variation of the tiling problem we edges of a and b or the left (western) edges of a and
consider is the PERIODIC TILING PROBLEM b have different colors. See Fig. 4a for an
where we ask whether a given set of tiles admits a illustration.
valid periodic tiling. Consider now a valid tiling of the plane by
NW-deterministic tiles. Each tile is uniquely
PERIODIC TILING PROBLEM determined by its left and upper neighbors.
Then, tiles on each diagonal in the NE-SW direc-
Input: Tile set T. tion locally determine the tiles on the next
Question: Does T admit a valid periodic tiling? diagonal below it. If we interpret these diagonals
as configurations of a CA, then there is a local
rule such that valid tilings are space-time dia-
Theorem 4 grams of the CA; see Fig. 4b.
The PERIODIC TILING PROBLEM is We define analogously NE-, SW-, and SE-deter-
undecidable for Wang tile sets. It is semi-decidable, ministic tile sets. Finally, we call a tile set four-way
while its complement is not semi-decidable. deterministic if it is deterministic in all four direc-
For a proof, see Gurevich and Koryakov (1972). tions simultaneously.
Tiling Problem and Undecidability in Cellular Automata 211

The tiling problem is undecidable among A set of two-dimensional directed tiles is said
NW-deterministic tile sets (Kari 1992), even to have the plane-filling property if it satisfies the
among four-way deterministic tile sets (Lukkarila following two conditions:
2009).
1. There exists t  T ℤ and a one-way infinite path
2

! ! !
Theorem 5 p 1 , p 2 , p 3 , . . . such that the tiling in t is valid
!
The decision problem tiling problem is undecidable at p i for all i ¼ 1, 2, 3, ::::.
! ! !
among four-way deterministic sets of Wang tiles. 2. For all t and p 1 , p 2 , p 3 , . . . as in (a), there are
As discussed at the end of section “The Tiling arbitrarily large n  n squares of cells such
Problem,” the theorem also means that four-way that all cells of the squares are on the path.
deterministic aperiodic tile sets exist. In fact, the
proof of Theorem 5 in Lukkarila (2009) uses one Intuitively, the plane-filling property means
such aperiodic set that was reported in Kari and that the simple device that moves over tiling
Papasoglu (1999). t repeatedly verifies the correctness of the tiling
in its present location and moves on to the fol-
lower, necessarily eventually either finds a tiling
Plane-Filling Directed Tiles error or covers arbitrarily large squares. Note that
the plane-filling property does not assume that the
A d-dimensional directed tile is a tile that is tiling t is correct everywhere: as long as it is
associated with a follower vector f  ℤd : Let correct along a path, the path must snake through
T ¼ ðd, T, N, RÞ be a tile set, and let F : T ! ℤd larger and larger squares.
be a function that assigns tiles their follower Note that conditions (a) and (b) imply that the
vectors. We call D = (d,T,N,R,F) a set of directed tile set is aperiodic. There exist tile sets that
tiles. Let t  T ℤ be an assignment of
tiles
d

to cells. satisfy the plane-filling property, as proved in
! ! ! Kari (1994a).
For every p  ℤd ; we call p þF t p the
!
follower of p in t. In other words, the follower of
! ! Theorem 6
p is the cell whose position relative to p is given
! There exists a set of directed Wang tiles that has
by the follower vector of the tile in cell p .
! ! ! ! the plane-filling property.
Sequence p 1 , p 2 , . . . , p k where all p i  ℤd is
The proof of Theorem 6 in Kari (1994a) con-
a (finite) path in t if
structs a set of Wang tiles such that the path that


 does not find any tiling errors is forced to follow
! ! !
p iþ1 ¼ p i þ F t pi the well-known Hilbert curve shown in Fig. 5.

for all 1  i < k: In other words, a path is a

sequence of cells such that the next cell is always Undecidability in Cellular Automata
the follower of the previous cell. One-way infinite
and two-way infinite paths are defined analogously. Let us begin with one-step properties of
In the following, we only discuss the two-dimensional CA.
two-dimensional case (d = 2), and the
follower of each tile is one of the four adjacent Theorem 7
positions: Injectivity is undecidable among
two-dimensional CA. It is semi-decidable in any
FðaÞ  fð 1, 0Þ, ð0 1Þ g dimension. Proof The semi-decidability follows
for all a  T: from the fact that injective CA has an inverse
CA. One can effectively enumerate all CA and
In this case, the follower is indicated in drawings check them one by one until (if ever) the inverse
as a horizontal or vertical arrow over the tile. CA is found.
212 Tiling Problem and Undecidability in Cellular Automata

Tiling Problem and

Undecidability
in Cellular Automata,
Fig. 5 Fractions of the
plane-filling Hilbert curve
through 4  4 and 16  16
squares

Let us next prove injectivity undecidable by “(”: Suppose a valid tiling exists. Construct
reducing the tiling problem into it. In the reduc- two configurations c0 and c1 where the
tion, a set D of directed tiles that has the plane- T and D components form the same valid
tilings t  T ℤ and d  Dℤ : respectively. In
2 2
filling property is used. The existence of such
D was stated in Theorem 6. c0, all bits are 0 while in c1 they are all
Let T be a given set of Wang tiles that is an 1. Since the tilings are everywhere valid,
instance of the tiling problem. One can effec- every cell performs modulo 2 addition of
tively construct a two-dimensional CA whose two bits, which means that every bit
state set is becomes 0. Hence, Gðc0 Þ ¼ Gðc1 Þ ¼ c0 ,
S ¼ T  D  f0, 1g and G is not injective.
“)”: Suppose then that G is not injective. There
and the local rule updates the bit component of a are two different configurations c0 and c1 such
cell as follows: that Gðc0 Þ ¼ Gðc1 Þ. Tile components are not
modified by the CA, so they are identical in c0
!
• If either the T or the D components contain a and c1. There is a cell p 1 such that c0 and c1
!
tiling error at the cell, then the state of the cell have different bits at cell p 1 . Since these bits
is not changed. become identical in the next configuration, the
!
• If the tilings according to both T and D tiling must be correct at p 1 , and c0 and c1
D components are valid at the cell, then the must have different bits in the follower posi-
!
bit of the follower cell (according to the direc- tion p 2. We repeat the reasoning and obtain an
! ! !
tion in the D component) is added to the pre- infinite sequence of positions p 1 , p 2 , p 3 , . . .
! !
sent bit value modulo 2. such that each p iþ1 is the follower of p i and
!
the D tiling is correct at each p i . It
follows from the plane-filling property of
! ! !
The tile components are not changed. Let us D that path p 1 , p 2 , p 3 , . . . covers arbitrarily
prove that this CA G is not injective if and only if large squares. Also, the tiling according to the
T admits a valid tiling. T components must be valid at each cell of the
Tiling Problem and Undecidability in Cellular Automata 213

Tiling Problem and

Undecidability
in Cellular Automata,
Fig. 6 Tiles used in the
proof of the undecidability
of surjectivity A A B B A A

NW NE
NW NE

NW NE
NW NE

SW SE
SW SE

SW C C SE
SW C SE

path. Hence, tile set T admits valid tilings of Proof

arbitrarily large squares, and, therefore, it A semi-algorithm for non-surjectivity enumerates
admits a valid tiling of the entire plane. all finite patterns one by one until a pattern is
found that cannot appear in G(c) for any
Analogously, we can prove the configuration c.
undecidability of surjectivity. It is convenient To prove undecidability, we reduce the finite
to use the well-known Garden of Eden theo- tiling problem, using the set D of 23 directed tiles
rem of Moore and Myhill to convert the shown in Fig. 6. These directed tiles are used in
surjectivity property into injectivity on finite an analogous way as in the proof of Theorem 7.
configurations. The topmost tile in Fig. 6 is called blank. All
other tiles have a unique incoming and outgoing
Theorem 8 (Garden of Eden Theorem) arrow. In valid tilings, arrows and labels must
A cellular automaton is non-surjective if and only match. The non-blank tiles are considered
if there are two distinct configurations that differ directed: the follower of a tile is the neighbor
in a finite number of cells and that have the same directed to by the outgoing arrow on the tile.
successor (Moore 1962; Myhill 1963). Since each non-blank tile has exactly one incom-
ing arrow, it is clear that if the tiling is valid at a
tile, then the tile is the follower of exactly one of
Theorem 9 its four neighbors.
Surjectivity is undecidable among two-dimensional The tile at the center in Fig. 6 where the dark
CA. Its complement is semi-decidable in any and light thick horizontal lines meet is called the
dimension. cross. It has a special role in the forthcoming
214 Tiling Problem and Undecidability in Cellular Automata

Tiling Problem and Undecidability in Cellular Automata, Fig. 7 A rectangular loop of size 12  7

proof. A rectangular loop is a valid tiling of a ðd, t, xÞ  D  T  f0, 1g

rectangle using tiles in D where the follower path
forms a loop that visits every tile of the rectangle
and the outside border of the rectangle is colored under the following constraints:
blank. See Fig. 7 for an example of a rectangular
loop through a rectangle of size 12  7. (The • If d = c, then t 6¼ B.
edge labels are not shown for the sake of clarity • If d = b or d is any tile containing label SW,
of the figure.) It is easy to see that a rectangular SE, NW, NE, A, B, or C, then t = B.
loop of size 2n  m exists for all n 2 and m 3.
Any tile in an even column in the interior of the In other words, the cross must be associ-
rectangle can be made to contain the unique cross ated with a non-blank tile in T, while the
of the rectangular loop. blank of D and all the tiles on the boundary
It is easy to see that the tile set D has the of a rectangular loop are forced to be associated
following property: with the blank tile of T. The triplet (b, B, 0)
where both tile components are blank and the
Finite plane-filling property. Let t  Dℤ be a
2
bit is 0 is the quiescent state of the CA. The
! ! !
tiling and p 1 , p 2 , p 3 , . . . a path in t such local rule is as follows: Let (d, t, x) be the
!
that the tiling t is valid at p i for all current state of a cell.
i ¼ 1, 2, 3, . . .. If the path covers only a finite
number of different cells, then the cells on the • If d = b, then the state is not changed.
path form a rectangular loop. • If d 6¼ b, then the cell verifies the validity of
the tilings according to both D and T at the
Let b and c be the blank and the cross of set D. cell. If either tile component has a tiling error,
For any given tile set T with blank tile B, we then the state is not changed. If both tilings are
construct the following two-dimensional cellular valid, then the cell modifies its bit component
automaton. The state set S contains triplets by adding the bit of its follower modulo 2.
Tiling Problem and Undecidability in Cellular Automata 215

Let us prove that this CA is not surjective if in a finite number of cells, we see that the
and only if T admits a valid, finite, nontrivial path can only contain a finite number of
tiling. distinct cells. It follows then from the finite
plane-filling property of D that the path must
“(”: Suppose a valid, finite, nontrivial tiling form a valid rectangular loop.
t  T ℤ exists. Consider a configuration of
2

the CA whose T components form the valid Also, the tiling according to the T components
tiling t and the D components form a rectan- must be valid at each cell of the path. Because of
gular loop whose interior covers all the constraints on the allowed triplets, the
non-blank elements of t. Tiles outside the T components on the boundary of the rectangle
rectangle are all blank and have bit 0. The are the blank B, while the cross in the interior
cross can be positioned so that it is in the contains a non-blank element of T. Hence, there
same cell as some non-blank tile in t. In is a valid tiling of a rectangle according to T that
such a configuration, both T and D tilings contains a non-blank tile and has a blank bound-
are everywhere valid. The CA updates the ary. This means that a finite, valid, and nontrivial
bits of all tiles in the rectangular loop by tiling is possible.
performing modulo 2 addition with their fol-
lowers, while the bits outside the rectangle
remain 0. We get two different configurations
Undecidable Properties of One-
that have the same image: In c0, all bits in the
Dimensional CA
rectangle are equal to 0, while, in c1, they are
all equal to 1. The local rule updates the bits
Using deterministic Wang tiles and interpreting
so that Gðc0 Þ ¼ Gðc1 Þ ¼ c0 . Configurations
space-time diagrams as tilings, one obtains
c0 and c1 only differ in a finite number of
undecidability results for long-term behavior of
cells, so it follows from the Garden of Eden
one-dimensional CA.
theorem that G is not surjective.
“)”: Suppose then that the CA is not surjec-
tive. According to the Garden of Eden theo- Theorem 10
rem, there are two finitely different Nilpotency is undecidable among
configurations c0 and c1 such that G(c0) = one-dimensional CA. It is undecidable even
G(c1). Since only bit components of states among one-dimensional CA that have a spread-
are changed, the tilings in c0 and c1 ing state q, i.e., a state that spreads to all neigh-
according to D and T components of the bors. Nilpotency is semi-decidable in any
!
states are identical. There is a cell p 1 such dimension.
!
that c0 and c1 have different bits at cell p 1 :
Since these bits become identical in the next Proof
configuration, the D tiling must be correct at For semi-decidability, notice that, for
!
p 1 , and c0 and c1 must have different bits in n = 1,2,3. . ., we can effectively construct a cel-
!
the follower position p 2 .We repeat the rea- lular automaton whose global function is Gn and
soning and obtain an infinite sequence of check whether the local rule of the CA maps
! ! !
positions p 1 , p 2 , p 3 , . . . such that each everything into the same state. If that happens
! !
p iþ 1 is the follower of p i and the D tiling for some n, then we halt and report that the CA
!
is correct at each p i . Moreover, c0 and c1 is nilpotent.
!
have different bits in each position p i : To prove undecidability, we reduce the tiling
Because configurations c0 and c1 only differ problem of NW-deterministic Wang tiles. Let
216 Tiling Problem and Undecidability in Cellular Automata

T be a given NW-deterministic tile set. One can G is not nilpotent, then there is a configuration
effectively construct a one-dimensional CA c  Sℤ such that no cell ever turns into the spread-
whose state set is S ¼ T [ fqg , and the local ing state q. But then the second components form
rule turns a cell into the quiescent state q except a left shift over the alphabet {1,2,. . .,n}, so the
in the case that the cell and its right neighbor are topological entropy is at least log2n > c.
in states x, y  T, respectively, and tile z  T exists It also follows that it is undecidable to deter-
so that tiles x□y, □z match as in Fig. 4a. In this mine if a given one-dimensional CA is ultimately
case, z is the new state of the cell. Note that state periodic (Durand et al. 2003).
q is a spreading state.
Let us prove that the CA is not nilpotent if and
only if T admits a valid tiling. Theorem 12
Equicontinuity is undecidable among
“(”: Suppose a valid tiling exists. If c  T ℤ is a one-dimensional CA.
diagonal of this tiling, then the configurations
Gn (c) in its orbit are subsequent diagonals of
the same tiling, for all n ¼ 1, 2, . . . . This Proof
means that c never becomes quiescent, and Among one-dimensional CA with a spreading
the CA is not nilpotent. state, equicontinuity is equivalent to nilpotency.
“)”: Suppose no valid tiling exists. Then there is
number n such that no valid tiling of an n  n Theorem 13
square exists. This means that for every initial Sensitivity to initial conditions is undecidable
configuration c  Sℤ, the configuration G2n (c) is among one-dimensional CA.
quiescent: If it is not quiescent, then a valid tiling
of an n  n square can be read from the space-
time diagram of configurations c, G(c),. . .,G2n Proof
(c) . We conclude that the CA is nilpotent. Originally, the result was proved in Durand
et al. (2003) using an elaborate reduction of the
Undecidability of nilpotency has some inter- Turing machine halting problem. However,
esting corollaries. First, it implies that the topo- undecidability of nilpotency provides the result
logical entropy of a one-dimensional CA cannot directly, as pointed out in Kari (2008). Namely, a
be calculated, not even approximated (Hurd one-dimensional cellular automaton whose
et al. 1992). neighborhood vector contains only strictly posi-
tive numbers is either nilpotent or sensitive.
Theorem 11 Adding a constant to all elements of the neigh-
Topological entropy is undecidable. borhood vector does not affect the nilpotency
status of a CA. So, for any given
Proof one-dimensional CA, we proceed as follows:
Let us reduce nilpotency. Let c > 0 be any con- add a positive constant to the elements of the
stant, and let n > 2C be an integer. For any given neighborhood vector so that they all become pos-
one-dimensional CA G with state set S and a itive. The new CA is sensitive if and only if the
spreading state q  S, construct a new CA whose original CA was not nilpotent. The result then
state set is S  {1,2,. . .,n}, and the local rule follows from Theorem 10.
applies G in the first components of the states As a final application of undecidability of
and shifts the second components one cell to the nilpotency consider other questions concerning
left. In addition, state (q, i) is turned into state the limit set (maximal attractor) of
(q,1). one-dimensional CA. One can show that
If G is nilpotent, then also the new CA is nilpotency can be reduced to any nontrivial ques-
nilpotent, and its topological entropy is 0. If tion (Kari 1994b). More precisely, let PROB be a
Tiling Problem and Undecidability in Cellular Automata 217



decision problem that takes arbitrary ! ! !
time t  0 such that e x ¼ d x for all x  A
one-dimensional CA as input. Suppose that


! ! !
PROB always has the same answer for any two but Gt ðeÞ x 6¼ Gt ðcÞ x for some x  B.
CA that have the same limit set. Then, we say that
PROB is a decision problem concerning the limit SENSITIVITY TO INITIAL CONDITIONS
sets of CA. We call PROB nontrivial if there exist
both positive and negative instances. Input: Cellular automaton A.
Question: Is A sensitive to initial conditions?

Theorem 14
Let PROB be any nontrivial decision problem
Theorem 16
concerning the limit sets of CA. Then, PROB is
Sensitivity to initial conditions is undecidable
undecidable (Kari 1994b).
among one-dimensional CA (Durand et al. 2003).
The following problems deal with dynamics
on finite configuration. We, hence, suppose that
Other Undecidability Results the given CA has a quiescent state, i.e., a state
q such that f (q,q,. . .,q) = q where f is the local
update rule of the CA. A configuration c  Sℤ is
d
In the previous sections, we only considered deci-
sion problems that have been proved undecidable called finite (w.r.t. q) if all but a finite number of
using reductions from the tiling problem or its var- cells are in state q. Questions similar to
iant. There are many other decision problems that nilpotency and equicontinuity can be asked in
have been proved undecidable using other tech- the space of finite configurations:
niques. Below are a few, with literature references.
We call a CA G periodic if there is number NILPOTENCY ON FINITE CONFIGURATIONS
n such Gn is the identity function. This is equiv-
alent to saying that every configuration is peri- Input: Cellular automaton A with a
odic, that is, every configuration returns back to quiescent state.
itself. Clearly, a periodic CA is necessarily injec- Question: Does every finite configuration
tive. In fact, periodic CA are exactly those CA evolve into the quiescent
that are injective and equicontinuous. configuration?

PERIODICITY
EVENTUAL PERIODICITY ON FINITE
Input: Cellular automaton A CONFIGURATIONS
Question: Is A periodic?
Input: Cellular automaton A with a
The question is undecidable among quiescent state.
two-dimensional CA (the construction in the Question: Does every finite configuration
proof of Theorem 7 shows it) but also with evolve into a temporally periodic
one-dimensional inputs. configuration?

Theorem 15
Periodicity is undecidable among
one-dimensional CA (Kari and Ollinger in press).
A CA is called sensitive to initial conditions Theorem 17 Nilpotency on finite configurations
if there exists a finite set B  ℤd of cells such that, and eventual periodicity on finite configurations
for every configuration c and every finite set are undecidable for one-dimensional CA (Culik
A  ℤ of cells, there exists a configuration e and and Yu 1988; Sutner 1989).
218 Tiling Problem and Undecidability in Cellular Automata

Future Directions Berger R (1966) The undecidability of the domino prob-

lem. Mem Am Math Soc 66:1–72
Culik K II (1996) An aperiodic set of 13 Wang tiles.
Some interesting and challenging open questions Discret Math 160:245–251
remain. In particular, the decidability status of the Culik K II, Yu S (1988) Undecidability of CA classifica-
decision problem concerning expansivity tion schemes. Complex Syst 2:177–190
(a strong type of sensitivity) is unknown. Durand B, Formenti E, Varouchas G (2003) On
undecidability of equicontinuity classification for
We call a one-dimensional CA G positively cellular automata. In: Proceedings of discrete
expansive if there exists a finite set A  ℤ of cells models for complex systems, Lyon, 16–19 June
such that for any two distinct configurations 2003, pp 117–128
c and e there exists t 0 such that Gt(c) and Finelli M, Manzini G, Margara L (1998) Lyapunov expo-
nents versus expansivity and sensitivity in cellular
Gt(e) differ in some cell in A. We call an injec- automata. J Complex 14:210–233
tive CA G expansive for some finite A if it holds Gurevich YS, Koryakov IO (1972) Remarks on Berger’s
the following: for any two distinct configura- paper on the domino problem. Sib Math J 13:319–321
tions c and e, there exists t  ℤ such that Gt(c) Hurd LP, Kari J, Culik K (1992) The topological entropy
of cellular automata is uncomputable. Ercodic Theor
and Gt(e) differ in some cell in A. It is known that Dyn Syst 12:255–265
a two- and higher-dimensional CA cannot be Jeandel E, Rao M (2015) An aperiodic set of 11 Wang
expansive or positively expansive (Finelli tiles. Preprint arXiv 1506.06492
et al. 1998; Shereshevsky 1993), so the follow- Kari J (1990) Reversibility of 2D cellular automata is
undecidable. Phys D 45:379–385
ing decision problems are only asked for Kari J (1992) The nilpotency problem of one-dimensional
one-dimensional CA: cellular automata. SIAM J Comput 21:571–586
Kari J (1994a) Reversibility and surjectivity problems of
Positive Expansivity cellular automata. J Comput Syst Sci 48:149–182
Kari J (1994b) Rice’s theorem for the limit sets of cellular
automata. Theor Comput Sci 127:229–254
Input: One-dimensional cellular Kari J (1996) A small aperiodic set of Wang tiles. Discret
automaton A. Math 160:259–264
Question: Is A positively expansive? Kari J (2008) Undecidable properties on the dynamics of
reversible one-dimensional cellular automata. In: Pro-
ceedings of Journées Automates Cellulaires, Uzés,
21–25 April 2008
EXPANSIVITY Kari J, Ollinger N (2008) Periodicity and immortality in
reversible computing. In: Mathematical Foundations
of Computer Science, Lecture Notes in Computer
Input: One-dimensional cellular Science 5162, pp 419–430
automaton A. Kari J, Papasoglu P (1999) Deterministic aperiodic tile
Question: Is A expansive? sets. J Geom Funct Anal 9:353–369
Kurka P (1997) Languages, equicontinuity and attractors
in cellular automata. Ergod Theory Dyn Syst
17:417–433
Lukkarila V (2009) The 4-way deterministic tiling prob-
The decidability status of both positive expan- lem is undecidable. Theoretical Computer Science
sivity and expansivity is unknown. 410(16):1516–1533
Moore EF (1962) Machine models of self-reproduction.
Proc Symp Appl Math 14:17–33
Acknowledgments This research is supported by the Myhill J (1963) The converse to Moore’s garden-of-Eden
Academy of Finland grants 211967 and 131558. theorem. Proc Am Math Soc 14:685–686
Robinson RM (1971) Undecidability and non-
periodicity for tilings of the plane. Invent Math
12:177209
Bibliography
Shereshevsky MA (1993) Expansiveness, entropy
and polynomial growth for groups acting on
Primary Literature subshifts by automorphisms. Indag Math 4:
Amoroso S, Patt Y (1972) Decision procedures for 203–210
surjectivity and injectivity of parallel maps for tessel- Sutner K (1989) A note on the Culik-Yu classes. Compl
lation structures. J Comput Syst Sci 6:448–464 Syst 3:107–115
Tiling Problem and Undecidability in Cellular Automata 219

Wang H (1961) Proving theorems by pattern Kari J (2005) Theory of cellular automata: a survey. Theor
recognition – II. Bell Syst Techn J 40:1–42 Comput Sci 334:3–33
Toffoli T, Margolus N (1987) Cellular automata
machines. MIT Press, Cambridge
Books and Reviews
Wolfram S (ed) (1986) Theory and applications of cellular
Codd EF (1968) Cellular automata. Academic, New York
automata. World Scientific Press, Singapore
Garzon M (1995) Models of massive parallelism: analysis
Wolfram S (2002) A new kind of science. Wolfram
of cellular automata and neural networks. Springer,
Media, Champaign
New York
Hedlund G (1969) Endomorphisms and automorphisms of
shift dynamical systems. Math Syst Theory 3:320–375
 si 6¼ si+1 for 1  i  n  1. Such a set S is called
Cellular Automata and Groups a free basis for F. The cardinality of S is an
invariant of the group F and it is called the rank
Tullio Ceccherini-Silberstein1 and of F.
Michel Coornaert2 A group G is finitely generated if there
1
Dipartimento di Ingegneria, Università del exists a finite subset S  G such that every
Sannio, Benevento, Italy element g  G can be expressed as a product
2
Institut de Recherche Mathématique Avancée, of elements of S and their inverses, that is, g
Université Louis Pasteur et CNRS, Strasbourg, ¼ sϵ11 sϵ22   sϵnn, where n  0 and si  S, ϵ i =  1
France for 1  i  n. The minimal n for which such an
expression exists is called the word length of
g with respect to S and it is denoted by ‘(g).
Article Outline The group G is a (discrete) metric space with
the distance function d : G
G ! ℝ+ defined
Glossary by setting d(g, g0) = ‘(g1g0) for all g, g0  G.
Definition of the Subject The set S is called a finite generating subset for
Introduction G and one says that S is symmetric provided that
Cellular Automata s  S implies s1  S.
Cellular Automata with a Finite Alphabet The Cayley graph of a finitely generated
Linear Cellular Automata group G w.r. to a symmetric finite generating
Group Rings and Kaplansky Conjectures subset S  G is the (undirected) graph Cay(G, S)
Future Directions with vertex set G and where two elements g,
Bibliography g0  G are joined by an edge if and only if
g1g0  S.
Glossary A group G is residually finite if the intersec-
tion of all subgroups of G of finite index is
Groups A group is a set G endowed with a trivial.
binary operation G
G 3 (g, h) 7! gh  G, A group G is amenable if it admits a right-
called the multiplication, that satisfies the fol- invariant mean, that is, a map
lowing properties: (i) for all g, h and k in G, m : P ðGÞ ! ½0,1 , where P ðGÞ denotes the
(gh)k = g(hk) (associativity); (ii) there exists an set of all subsets of G, satisfying the following
element 1G  G (necessarily unique) such conditions: (i) m(G) = 1 (normalization);
that, for all g in G, 1Gg = g1G = g (existence (ii) m(A [ B) = m(A) + m(B) for all
of the identity element); (iii) for each g in G, A,B  P ðGÞsuch that A \ B = ∅ (finite additivity);
there exists an element g1  G (necessarily (iii) m(Ag) = m(A) for all g  G and A  P(G)
unique) such that gg1 = g1g = 1G (existence (right-invariance).
of the inverses). Rings A ring is a set R equipped with two binary
A group G is said to be Abelian operations R
R 3 (a, b) 7! a + b  R and
(or commutative) if the operation is commuta- R
R 3 (a, b) 7! ab  R, called the addition
tive, that is, for all g, h  G one has gh = hg. and the multiplication, respectively, such that
A group F is called free if there is a subset the following properties are satisfied: (i) R,
S  F such that any element g of F can be with the addition operation, is an Abelian
uniquely written as a reduced word on S, i.e. in group with identity element 0, called the zero
the form g ¼ sa11 sa22   sann , where n  0, si  S element, (the inverse of an element a  R is
and ai  ℤ ∖ {0} for 1  i  n, and such that denoted by a); (ii) the multiplication is
# Springer-Verlag 2009 221
A. Adamatzky (ed.), Cellular Automata,
https://doi.org/10.1007/978-1-4939-8700-9_52
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_52
222 Cellular Automata and Groups

associative and admits an identity element 1, following condition: there exists a finite subset
called the unit element; (iii) multiplication is M  G and a map m : AM ! A such that t(x)(-
distributive with respect to addition, that is, a- g) = m(xg|M) for all x  AG, g  G, where xg|M
(b + c) = ab + ac and (b + c)a = ba + ca for all a, denotes the restriction of xg to M. Such a set M is
b and c  R. called a memory set and m is called a local
A ring R is commutative if ab = ba for all a, defining map for t.
b  R. If A = V is a vector space over a field ,
A field is a commutative ring  6¼ f0g where then a cellular automaton t : VG ! VG, with
every non-zero element a   is invertible, that is memory set M  G and local defining map
there exists a1   such that aa1 = 1. m : VM ! V, is saito be linear provided that m is
In a ring R a non-trivial element a is called a linear.
zero-divisor if there exists a non-zero element Two configurations x, x0  AG are said to
b  R such that either ab = 0 or ba = 0. be almost equal if the set {g  G; x(g) 6¼ x0(g)}
A ring R is directly finite if whenever ab = 1 at which they differ is finite. A cellular autom-
then necessarily ba = 1, for all a, b  R. If the aton is called pre-injective if whenever t-
ring Md(R) of d
d matrices with coefficients (x) = t(x0) for two almost equal
in R is directly finite for all d  1 one says that configurations x, x0  AG one necessarily
R is stably finite. has x = x0.
Let R be a ring and let G be a group. Denote A Garden of Eden configuration is a con-
by R[G] the set of all formal sums g  Gagg figuration x  AG ∖ t(AG). Clearly, GOE
where ag  R and ag = 0 except for finitely configurations exist if and only if t is not
many elements g  G. We define two binary surjective.
operations on R[G], namely the addition, by
setting Definition of the Subject
! !
X X X
ag g þ bh h ¼ ag þbg Þg, A cellular automaton is a self-mapping of the set
gG hG gG of configurations of a group defined from local
and invariant rules. Cellular automata were first
and the multiplication, by setting only considered on the n-dimensional lattice
group ℤn and for configurations taking values in
! !
X X X a finite alphabet set but they may be formally
ag g bh h ¼ ag bh gh defined on any group and for any alphabet. How-
gG hG g ,h  G
X ever, it is usually assumed that the alphabet set is
k ¼ gh ag bg1 k k: endowed with some mathematical structure and
g ,h  G
that the local defining rules are related to this
structure in some way. It turns out that general
Then, with these two operations, R[G]
properties of cellular automata often reflect prop-
becomes a ring; it is called the group ring of
erties of the underlying group. As an example, the
G with coefficients in R.
Garden of Eden theorem asserts that if the group is
Cellular automata Let G be a group, called the
amenable and the alphabet is finite, then the
universe, and let A be a set, called the alphabet.
surjectivity of a cellular automaton is equivalent to
A configuration is a map x : G ! A. The set AG
its pre-injectivity (a weak form of injectivity). There
of all configurations is equipped with the
is also a linear version of the Garden of Eden theo-
right action of G defined by AG
G 3 (x,
rem for linear cellular automata and finite-
g) 7! xg  AG, where xg(g0) = x(gg0) for all
dimensional vector spaces as alphabets. It is an
g0  G.
amazing fact that famous conjectures of Kaplansky
A cellular automaton over G with coefficients
about the structure of group rings can be
in A is a map t : AG ! AG satisfying the
reformulated in terms of linear cellular automata.
Cellular Automata and Groups 223

Introduction Shortly later, John Myhill proved the converse

to Moore’s theorem. The equivalence of surjectivity
The goal of this paper is to survey results related to and pre-injectivity of cellular automata is referred to
the Garden of Eden theorem and the surjunctivity as the Garden of Eden theorem (briefly GOE theo-
problem for cellular automata. rem), this biblical terminology being motivated by
The notion of a cellular automaton goes back to the fact that it gives necessary and sufficient condi-
John von Neumann (1966) and Stan Ulam (1952). tions for the existence of configurations
 2 x that are
Although cellular automata were firstly considered not in the image of t, i.e. x  A ∖t Aℤ , so that,
ℤ2

only in theoretical computer science, nowadays they  

thinking of t,Aℤ as a discrete dynamical sys-
2

play a prominent role also in physics and biology,

where they serve as models for several phenomena tem, with t being the time, they can appear only
(▶ “Cellular Automata Modeling of Physical Sys- as “initial configurations”.
tems” and ▶ “Chaotic Behavior of Cellular Autom- It was immediately realized that the GOE the-
ata”), and in mathematics. In particular, cellular orem was holding also in higher dimension,
automata are studied in ergodic theory (“Entropy namely for cellular automata with universe
in Ergodic Theory”, and “Ergodic Theory: U = ℤd, the lattice of integer points in the
Basic Examples and Constructions, ▶ “Ergodic d-dimensional space. Then, Machì and Mignosi
Theory of Cellular Automata”, and “Ergodicity (1993) gave the definition of a cellular automaton
and Mixing Properties”) and in the theory of over a finitely generated group and extended the
dynamical systems (▶ “Topological Dynamics GOE theorem to the class of groups G having sub-
of Cellular Automata”, and “Symbolic Dynam- exponential growth, that is for which the growth
ics”), in functional and harmonic analysis function gG(n), which counts the elements g  G
(“Spectral Theory of Dynamical Systems”), and in at “distance” at most n from the unit element 1G of
group theory. G, has a growth weaker
p than the exponential, in
ffiffiffiffiffiffiffiffiffiffiffi
In the classical framework, the universe U is formulæ, limn!1 gG ðnÞ¼ 1 . Finally, in 1999
n

the lattice ℤ2 of integer points in Euclidean plane Ceccherini-Silberstein et al. (1999) extended the
and the alphabet A is a finite set, typically GOE theorem to the class of amenable groups. It
A = {0, 1}. The set AU = {x : U ! A} is the is interesting to note that the notion of an amena-
configuration space, a map x : U ! A is a config- ble group was also introduced by von Neumann
uration and a point (n, m)  U is called a cell. (1930). This class of groups contains all finite
One is given a neighborhood M of the origin groups, all Abelian groups, and in fact all solv-
(0, 0)  U, typically, for some r > 0, M = able groups, all groups of sub-exponential
{(n, m)  ℤ2 : |n| + |m|  r} (von Neumann growth and it is closed under the operation of
r-ball) or M = {(n, m)  ℤ2 : |n|, |m|  r} taking subgroups, quotients, directed limits and
(Moore’s r-ball) and a local map m : AM ! A. extensions. In Machì and Mignosi (1993) two
One then “extends” m to the whole universe examples of cellular automata with universe the
obtaining a map t : AU ! AU, called a cellular free group F2 of rank two, the prototype of a
automaton, by setting t(x)(n, m) = m(x(n + s, m + t) non-amenable group, which are surjective but
(s, t)  M). This way, the value t(x)(n, m)  A of the not pre-injective and, conversely, pre-injective
configuration x at the cell (n, m)  U only but not surjective, thus providing an instance of
depends on the values x(n + s, m + s) of x at the failure of the theorems of Moore and Myhill
its neighboring cells (x + s, y + t) (x, y) + and so of the GOE theorem. In Ceccherini-
(s, t)  (x, y) + M, in other words, t is Silberstein et al. (1999) it is shown that this
ℤ2-equivariant. M is called a memory set for t examples can be extended to the class of groups,
and m a local defining map. thus necessarily non-amenable, containing the
In 1963 E.F. Moore proved that if a cellular free group F2. We do not know whether the
automaton t : Aℤ ! Aℤ is surjective then it is
2 2
GOE theorem only holds for amenable groups
also pre-injective, a weak form of injectivity. or there are examples of groups which are non-