Probabilistic Cellular Automata

• Probabilistic cellular automata are discrete-time Markov systems that update all sites concurrently with local stochastic rules, enabling nuanced stochastic dynamics.
• They offer a flexible framework that bridges cellular automata, interacting particle systems, and Markov chains, facilitating studies in invariant measures, ergodicity, and phase transitions.
• PCA are practically applied in diverse domains such as image processing, epidemic modeling, optimization, and materials science, highlighting their computational and modeling utility.

Probabilistic cellular automata (PCA) are discrete-time Markov chains on spatial configuration spaces such as $X=\mathcal A^E$ or $\{-1,+1\}^{\mathbb Z^d}$ in which all sites are updated simultaneously and, conditional on the present configuration, independently according to local transition laws that depend on a finite neighborhood; deterministic cellular automata are recovered when those local laws are Dirac masses (Busic et al., 2010, Maes, 2016). This parallel-update product-kernel structure makes PCA a distinct class of stochastic lattice systems: they are local like cellular automata, Markovian like interacting particle systems, and sufficiently flexible to support invariant-measure theory, perfect sampling, absorbing-state criticality, metastability, and domain-specific modeling on lattices and graphs (Busic et al., 2010, Mendonça, 2010, Dell'Anna et al., 7 Dec 2025).

1. Formal definition and update architecture

In the standard formulation, a PCA is specified by a finite alphabet $\mathcal A$, a finite or countable set of cells $E$, a finite neighborhood $V\subset E$, and a local rule $f:\mathcal A^V\to \mathcal M(\mathcal A)$. The global configuration space is $X=\mathcal A^E$, and the global update is a Markov operator whose action on cylinder events factorizes over sites because all cells are updated synchronously and conditionally independently (Busic et al., 2010). In equivalent notation used for one-dimensional elementary PCA and graph-based models,

$$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$

with $x_{\mathcal N_i}$ denoting the neighborhood of site $i$ (Mendonça, 2017, Dell'Anna et al., 7 Dec 2025).

For Ising-type PCA on $\{-1,+1\}^{\mathbb Z^d}$0, a convenient parameterization is

$\{-1,+1\}^{\mathbb Z^d}$1

where $\{-1,+1\}^{\mathbb Z^d}$2 is a local, translation-invariant field (Maes, 2016). This makes explicit that PCA are discrete-time, local, parallel-updating Markov systems rather than sequential spin-flip dynamics.

A recurring construction is the probabilistic mixture of deterministic elementary Wolfram rules. The $\{-1,+1\}^{\mathbb Z^d}$3 automaton updates each site according to rule 182 with probability $\{-1,+1\}^{\mathbb Z^d}$4 and rule 200 with probability $\{-1,+1\}^{\mathbb Z^d}$5, thereby generating a one-parameter PCA from two deterministic radius-1 rules (Mendonça et al., 2010). The $\{-1,+1\}^{\mathbb Z^d}$6 model uses the same idea to encode a biologically interpreted population process (Mendonça, 2017). A later two-parameter example interpolates between rules 23, 77, 178, and 232 by using one Bernoulli parameter for majority situations and another for tied neighborhoods (Muñoz et al., 29 Mar 2026). These constructions show that stochasticity in PCA need not be introduced only by specifying transition probabilities directly; it can also arise through random superposition of deterministic local maps.

2. Invariant measures, ergodicity, and exact sampling

A probability measure $\{-1,+1\}^{\mathbb Z^d}$7 is invariant for a PCA when $\{-1,+1\}^{\mathbb Z^d}$8, and the PCA is ergodic when it has a unique invariant measure attracting all initial distributions in the weak topology (Busic et al., 2010). For binary one-dimensional nearest-neighbor PCA depending on the pair $\{-1,+1\}^{\mathbb Z^d}$9, there are exact necessary and sufficient conditions for a Bernoulli product measure $\mathcal A$0 to be invariant. Under these balance conditions, the stationary space-time diagram has unusually weak dependence: horizontal lines are i.i.d., many oblique lines are i.i.d., and in the doubly balanced case every line in every direction is i.i.d., even though the field is generally not globally i.i.d. because dependence persists on upward equilateral triangles (Mairesse et al., 2012).

For finite or infinite cell sets, ergodicity can be approached algorithmically by coupling from the past. An important result is the envelope PCA construction, which enlarges the alphabet to encode uncertainty and yields a perfect-sampling algorithm for the invariant measure of an ergodic PCA without assuming monotonicity of the local rule (Busic et al., 2010). In the deterministic one-dimensional case, ergodicity is equivalent to nilpotency, and is therefore undecidable (Busic et al., 2010). This places even the deterministic boundary of PCA theory outside general algorithmic decidability.

Ergodicity also does not guarantee the existence of a successful local i.i.d. coupling. For the noisy rule-90 PCA,

$\mathcal A$1

with $\mathcal A$2 Bernoulli$\mathcal A$3, the chain is ergodic for every $\mathcal A$4 with invariant Bernoulli$\mathcal A$5 measure, yet for $\mathcal A$6 sufficiently close to $\mathcal A$7 no basic i.i.d. coupling built from nearest-neighbor random local maps is successful (2207.14569). This separates distributional convergence from coalescence-based perfect simulation in a sharp way.

Exact finite-size analysis provides a complementary perspective. For the PCA between rules 23, 77, 178, and 232, explicit stochastic transition matrices for $\mathcal A$8 and $\mathcal A$9 show that when $E$0 the asymptotic distribution is independent of initial condition, whereas on edges and deterministic corners the chain develops absorbing states, split communicating classes, and period-2 behavior (Muñoz et al., 29 Mar 2026). The contrast suggests that interior stochastic mixing can eliminate the basin structure present in deterministic elementary rules.

3. Absorbing states, phase transitions, and universality

One of the central roles of PCA in nonequilibrium statistical mechanics is as a minimal setting for absorbing-state phase transitions. Stavskaya’s PCA is an early one-dimensional example with a single absorbing configuration $E$1. Monte Carlo simulation and finite-size scaling give $E$2, $E$3, $E$4, $E$5, and $E$6, consistent with one-dimensional directed percolation, and the model is explicitly equivalent, after complementation, to the Domany–Kinzel PCA on its directed site-percolation line $E$7 (Mendonça, 2010).

The one-dimensional $E$8 PCA gives a related but structurally different example. It has two absorbing configurations, $E$9 and $V\subset E$0, though the extinction-survival transition is governed by the density of active sites and, for generic random initial conditions, the all-ones state is dynamically irrelevant. Mean-field analysis and Monte Carlo simulations indicate an extinction-survival transition in the directed-percolation universality class, with a characteristic stationary density profile and slow diffusive dynamics close to the pure CA 200 limit (Mendonça et al., 2010).

Ecological PCA make the same mechanism explicit in applied terms. The $V\subset E$1 automaton is a one-dimensional two-state PCA with a single absorbing empty state. Its first-order mean-field approximation yields

$V\subset E$2

which the paper interprets as a cubic map combining logistic limitation and a weak Allee effect; the actual extinction-survival transition of the PCA is again reported to belong to the directed-percolation universality class (Mendonça, 2017). A broader classification of left-right symmetric elementary PCA shows that, under first-order mean-field closure, admissible models fall into two classes: those generating the logistic map and those generating cubic logistic-like growth with a weak Allee factor (Mendonça et al., 2017).

Rigorous phase-transition results also exist in more artificial mixed-range settings. For a noisy majority-voter PCA on $V\subset E$3, an intermediate model with time-dependent candidate-neighbor set size $V\subset E$4 is proved non-ergodic for sufficiently small noise $V\subset E$5 and sufficiently large $V\subset E$6 and $V\subset E$7, thereby establishing a genuine low-noise phase transition between ergodic and non-ergodic behavior (Bricmont et al., 2013). A plausible implication is that mean-field-like mixing can preserve ordered phases while remaining close enough to local PCA dynamics to support a graphical proof.

4. Reversibility, nonequilibrium response, and metastability

PCA admit a developed nonequilibrium formalism. For reversible, time-symmetric PCA with

$V\subset E$8

the stationary law is

$V\subset E$9

which is generally not a standard pair-interaction Gibbs measure because the $f:\mathcal A^V\to \mathcal M(\mathcal A)$0 term induces effective higher-order interactions (Maes, 2016). This yields a fundamental distinction from sequential equilibrium lattice dynamics: not all equilibrium distributions are reachable as stationary laws of parallel-update PCA (Maes, 2016).

Away from detailed balance, the paper develops entropy production and response directly on path space. A local measure of time-reversal breaking is

$f:\mathcal A^V\to \mathcal M(\mathcal A)$1

and the corresponding scaled cumulant generating function satisfies a Gallavotti–Cohen-type symmetry $f:\mathcal A^V\to \mathcal M(\mathcal A)$2 (Maes, 2016). Linear response around equilibrium has a Kubo form, whereas around nonequilibrium stationary states the response contains both an entropic term and a time-symmetric frenetic correction (Maes, 2016). The minimum entropy production principle and McLennan-type corrections are recovered perturbatively near detailed balance (Maes, 2016).

Metastability in PCA requires a further modification of standard energy-barrier intuition. In the reversible tuned cross PCA, the stationary measure has a zero-temperature energy $f:\mathcal A^V\to \mathcal M(\mathcal A)$3, but because updates are parallel, even an energy-lowering transition may be exponentially unlikely; one must supplement $f:\mathcal A^V\to \mathcal M(\mathcal A)$4 by a transition cost $f:\mathcal A^V\to \mathcal M(\mathcal A)$5 derived from the low-temperature asymptotics of the one-step kernel (Cirillo et al., 2016). Path heights are therefore defined by

$f:\mathcal A^V\to \mathcal M(\mathcal A)$6

and metastable states are identified via maximal stability level rather than by energy alone (Cirillo et al., 2016).

For the cross PCA ($f:\mathcal A^V\to \mathcal M(\mathcal A)$7), the unique metastable state is the all-minus configuration $f:\mathcal A^V\to \mathcal M(\mathcal A)$8, and the escape to the stable all-plus state $f:\mathcal A^V\to \mathcal M(\mathcal A)$9 occurs via a critical plus droplet with barrier asymptotic to $X=\mathcal A^E$0. For the nearest-neighbor PCA ($X=\mathcal A^E$1), the metastable set is $X=\mathcal A^E$2, where $X=\mathcal A^E$3 and $X=\mathcal A^E$4 are checkerboards, and the barrier is asymptotic to $X=\mathcal A^E$5 (Cirillo et al., 2016). This suggests that parallel updating can generate metastable pathways through checkerboard structures and effective antiferromagnetic tendencies that are absent from the simplest single-spin-flip ferromagnetic dynamics.

5. Analytical frameworks and structural generalizations

Recent work extends PCA analysis into concentration theory and large deviations. For Ising PCA with update field

$X=\mathcal A^E$6

Gaussian concentration bounds are propagated by the dynamics: if $X=\mathcal A^E$7 satisfies $X=\mathcal A^E$8, then $X=\mathcal A^E$9 satisfies $$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$0, where $$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$1 controls contraction (Chazottes et al., 7 Jul 2025). In the contractive regime $$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$2, the unique stationary measure satisfies $$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$3, and the corresponding space-time measure satisfies a Gaussian concentration bound if and only if its spatial marginal does (Chazottes et al., 7 Jul 2025). The same work shows that GCB is incompatible with several forms of non-uniqueness for stationary PCA measures (Chazottes et al., 7 Jul 2025).

From the viewpoint of occupation-measure large deviations, the Donsker–Varadhan action functional for PCA can be unexpectedly singular. For PCA on $$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$4 satisfying synchronous update, locality, strict positivity, and finite reverse dependency, the functional $$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$5 is finite only on a narrow class of measures; in shift-invariant settings, if $$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$6 is space-shift invariant then $$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$7, and finite action forces full time invariance for the PCA (Eizenberg, 1 Sep 2025). In a one-sided shift setting with a unique invariant measure, finite action further implies convergence of spatial shifts of $$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$8 to that invariant measure (Eizenberg, 1 Sep 2025). A plausible implication is that large-deviation functionals imported from general Markov-chain theory probe only a very thin part of infinite-dimensional PCA measure space.

At an even broader level, cellular automata can be formulated inside operational probabilistic theories. In that framework, infinite composites are constructed from quasi-local effects and states, causal influence is defined as a property stronger than ordinary signalling, and homogeneity is shown to force the underlying system set to be the vertex set of a Cayley graph (Perinotti, 2019). Locality is then imposed through a local-to-global reduction rule, and a general wrapping lemma connects cellular automata on infinite Cayley graphs to finite quotients sharing the same small-scale neighborhood structure (Perinotti, 2019). Since classical probabilistic theory is one of the covered operational theories, classical probabilistic/stochastic automata appear as a special case of this more general construction (Perinotti, 2019).

6. Applied and computational uses

PCA have been adapted to a wide range of applied problems without abandoning their local parallel Markov structure. In Bayesian inversion for image denoising, a synchronous PCA is built from a pair Hamiltonian and an inertial penalty $$P_{t+1}(\mathbf{x}')=\sum_{\mathbf{x}}\Phi(\mathbf{x}'\mid \mathbf{x})P_t(\mathbf{x}), \qquad \Phi(\mathbf{x}'\mid \mathbf{x})=\prod_i \phi(x_i'\mid x_{\mathcal N_i}),$$9, producing fully factorized site updates that approximate the posterior distribution while remaining naturally parallelizable; on synthetic denoising tasks, the reported comparisons with Gibbs sampling use PSNR and SSIM as evaluation metrics (Costarelli et al., 20 Jul 2025). In combinatorial optimization, a graph-based PCA for Maximum Independent Set starts from the empty configuration and updates all vertices synchronously; its absorbing configurations are exactly the maximal independent sets of the graph, and numerical evidence shows that the probability of reaching a maximum independent set increases as the activation probability $x_{\mathcal N_i}$0 tends to $x_{\mathcal N_i}$1, although an explicit counterexample shows that $x_{\mathcal N_i}$2 can occur (Dell'Anna et al., 7 Dec 2025).

In materials modeling, PCA for microstructure evolution are analyzed through discrete probability distribution functions for the number of successful cell transformations in a time step. When all candidate boundary cells share the same transformation probability $x_{\mathcal N_i}$3, the number of successful transformations is binomial with mean $x_{\mathcal N_i}$4 and variance $x_{\mathcal N_i}$5; more generally, the paper derives an upper bound on the variance and uses the coefficient of variation to show that increasing the number of boundary cells, cellular resolution, and model size reduces uncertainty (Seyed-Salehi, 2024). In epidemic modeling, a modified SEIQR PCA on a $x_{\mathcal N_i}$6 square lattice uses local exposure of susceptibles to neighboring exposed and infected individuals, synchronous updating, and a distancing parameter $x_{\mathcal N_i}$7 that changes the effective interaction range; the model is used to study density, lockdown delay, testing efficiency, and quarantine timing in COVID-19 spread (Ghosh et al., 2020).

The same formalism has also been pushed to very coarse-grained scientific scenarios. A two-dimensional four-state PCA for the Galactic Habitable Zone models intrinsic evolution, neighbor-induced forcing, and externally forced catastrophic resets through a $x_{\mathcal N_i}$8 probability tensor $x_{\mathcal N_i}$9, with synchronous update on a square lattice representing an annulus from $i$0 to $i$1 (Vukotić et al., 2012). This suggests that PCA are useful not only when microscopic locality is physically literal, but also when uncertain multiscale dynamics must be encoded in a tractable stochastic spatial architecture.

Across these examples, a common pattern remains visible: PCA trade exact global control for explicit locality, synchronous parallel update, and probabilistic rule design. That combination explains both their mathematical richness and their persistence as models of nonequilibrium dynamics, inference, optimization, and spatially extended complex systems (Maes, 2016, Busic et al., 2010).