Domany-Kinzel Automaton

• Domany–Kinzel automaton is a 1D probabilistic cellular automaton with binary states that models nonequilibrium transitions between absorbing and active phases using local parent-site rules.
• Its phase diagram features special parameter submanifolds, mapping to directed percolation, Wolfram rule 18, and compact directed percolation, serving as benchmarks for critical phenomena.
• The model underpins diverse methodologies including finite-size scaling, tensor-network constructions, and quantum generalizations, offering practical insights for rigorous analysis and machine-learning applications.

The Domany–Kinzel automaton is a one-dimensional probabilistic cellular automaton on a tilted square $(1+1)$-dimensional space-time lattice, with binary local variables and synchronous discrete-time dynamics. In its standard absorbing-state form, the probability that a site becomes active at the next time step depends only on the two parent sites in the previous time slice, and no activity is created from an empty parent pair. The model therefore exhibits a nonequilibrium transition between an inactive absorbing phase and an active phase with nonzero stationary density. Its special parameter lines realize site directed percolation, bond directed percolation, Wolfram rule 18, and compact directed percolation, and the automaton has become a standard reference system for directed-percolation criticality, quasi-stationary finite-size analysis, tensor-network constructions, and quantum generalizations (Martins, 2012, Tuo et al., 2023, Lee et al., 10 Jun 2026).

1. Local dynamics and geometric formulation

A standard formulation of the Domany–Kinzel cellular automaton (DKCA) uses a one-dimensional lattice of $N$ sites with binary occupation variables $\sigma_i\in\{0,1\}$, where $\sigma_i=1$ denotes an occupied or active site and $\sigma_i=0$ an empty or inactive site. The update is synchronous: the state of site $i$ at time $t+1$ depends only on the two nearest neighbors at time $t$, through the conditional probabilities

$$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$

with

$$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$

A site can therefore become active only if at least one parent was active in the previous layer. The natural absorbing-state order parameter is the density of active sites,

$N$0

In the thermodynamic limit, $N$1 vanishes in the inactive phase and remains finite in the active phase (Martins, 2012).

The automaton is commonly represented on a tilted square lattice, with one axis interpreted as space and the other as discrete time. In that representation, each site is updated from two parent sites immediately above it on the previous time slice. Several later works retain this geometric picture when embedding the model into classical transfer-matrix, tensor-network, or quantum-cellular-automaton constructions (Lesanovsky et al., 2018, Gillman et al., 2020).

A notational subtlety appears across the literature. Some formulations use the direct synchronous PCA rule above, while others use two sublattices evolving on alternating even and odd time steps. The comparison with Stavskaya’s probabilistic cellular automaton shows that, after a complementary-variable transformation, the transition probabilities coincide on the directed site-percolation line even though the microscopic updating conventions differ (Mendonça, 2010).

2. Parameter submanifolds and universality structure

The DKCA has a two-parameter phase diagram in $N$2, with a critical line separating active and absorbing phases. Several special submanifolds recur throughout the literature because they isolate standard universality classes or exact mappings.

Parameter condition	Interpretation	Representative feature
$N$3	Site directed percolation	Diagonal DK submanifold
$N$4	Bond directed percolation	Bond-DP line
$N$5	Wolfram rule 18 / $N$6-XOR line	DP-class transition
$N$7	Compact directed percolation	Compact clusters

Bond directed percolation, site directed percolation, and Wolfram rule 18 are described as belonging to the directed-percolation universality class, whereas compact directed percolation belongs to a distinct universality class with compact clusters (Tuo et al., 2023). On the site-DP line, the critical point quoted in the comparison with Stavskaya’s model is

$N$8

while on the bond-DP line the quasi-stationary MPS study gives

$N$9

For the $\sigma_i\in\{0,1\}$0 line, the order-parameter-distribution method yields

$\sigma_i\in\{0,1\}$1

and for the compact-DP endpoint $\sigma_i\in\{0,1\}$2, the transition occurs at

$\sigma_i\in\{0,1\}$3

These values are used in different contexts as benchmarks for simulation, finite-size scaling, and mappings to other models (Mendonça, 2010, Lee et al., 10 Jun 2026, Martins, 2012, Wada et al., 2020).

For one-dimensional directed percolation, the exponents quoted in the order-parameter-distribution analysis are

$\sigma_i\in\{0,1\}$4

and the Stavskaya study quotes

$\sigma_i\in\{0,1\}$5

These values provide the reference against which DK-related models are compared when establishing directed-percolation scaling (Martins, 2012, Mendonça, 2010).

The compact-DP endpoint is structurally different. At $\sigma_i\in\{0,1\}$6, the dynamics acquires a particle-hole symmetry, the clusters become compact, and the evolution is controlled by interfaces between active and inactive domains rather than by sparse branching structures. This distinction underlies several mappings between DK-type dynamics and random-walk or first-passage problems (Wada et al., 2020).

3. Exact mappings and classical relatives

A central feature of the Domany–Kinzel automaton is that multiple apparently distinct probabilistic cellular automata become exact DK restrictions after a change of variables or a sublattice reinterpretation. This is most explicit for Stavskaya’s probabilistic cellular automaton. Stavskaya’s model is defined by

$\sigma_i\in\{0,1\}$7

with absorbing configuration $\sigma_i\in\{0,1\}$8. Introducing complementary variables

$\sigma_i\in\{0,1\}$9

maps the model to the DK automaton on the directed site-percolation line,

$\sigma_i=1$0

The critical points are then related by

$\sigma_i=1$1

so that

$\sigma_i=1$2

and Monte Carlo gives

$\sigma_i=1$3

in agreement with the known DK site-DP value. The measured exponents,

$\sigma_i=1$4

are consistent with directed percolation within error bars (Mendonça, 2010).

The noisy additive $\sigma_i=1$5-XOR probabilistic cellular automaton provides a second exact identification. Its local rule

$\sigma_i=1$6

coincides with the DK transition table on the line

$\sigma_i=1$7

In this representation, the $\sigma_i=1$8-XOR PCA is exactly the DK PCA restricted to the $\sigma_i=1$9 line, which is also the diluted elementary CA 90 line; the same model is simultaneously a diluted CA 102. The numerical estimate quoted for the transition is

$\sigma_i=0$0

with critical estimates

$\sigma_i=0$1

again in agreement with one-dimensional directed percolation (Mendonça, 2015).

The compact-cluster limit $\sigma_i=0$2 connects DK dynamics to a different class of absorbing-state problems. In the SEI model with correlated temporal disorder on the Bethe lattice, the $\sigma_i=0$3 DK automaton is used as the canonical compact-directed-percolation reference point. Starting from a compact cluster, only the two interfaces fluctuate, so the effective dynamics reduces to a random walk with absorbing boundary at extinction. In the uncorrelated case $\sigma_i=0$4, the critical exponents reduce to the compact-DP values

$\sigma_i=0$5

which is the basis for the equivalence claimed there between compact directed percolation and the corresponding infinite-noise critical point (Wada et al., 2020).

4. Quasi-stationary distributions, finite-size scaling, and rigorous bounds

Because any finite absorbing-state system eventually reaches the absorbing configuration, equilibrium-style stationary sampling is not directly applicable to the DKCA. The relevant finite-system object is instead the quasistationary distribution conditioned on survival. In the order-parameter-distribution approach, one studies the probability density $\sigma_i=0$6 of the active-site density under quasistationary dynamics and assumes the finite-size scaling form

$\sigma_i=0$7

The quasistationary simulation method used there stores a history of previously visited active configurations and replaces attempted absorption events by a randomly chosen saved configuration. In the implementation reported for the DKCA, the history list had

$\sigma_i=0$8

configurations and was updated with probability

$\sigma_i=0$9

The simulations used lattices with up to $i$0 sites, $i$1 samples, and $i$2 Monte Carlo steps per sample, and from distribution matching across sizes extracted the estimate

$i$3

for the $i$4 line (Martins, 2012).

A substantially different quasistationary perspective is obtained by representing the full QSD on the bond-DP line as a matrix product state. In that construction, the QSD is the leading non-negative right eigenvector of the projected transfer matrix

$i$5

and the dominant eigenvalue $i$6 is the survival probability per DK step. This gives direct access to the full conditional distribution of surviving configurations rather than only to moments or Monte Carlo samples. The resulting picture is sharply phase-dependent: in the active phase the QSD is bulk-like with finite density, while in the inactive phase the surviving activity collapses into a single flock occupying a vanishing fraction of the chain. The paper’s central information-theoretic claim is that throughout the inactive phase the bipartite mutual information equals the entropy of a single binary choice—whether the flock lies to the left or right of the cut—so the half-chain mutual information approaches one bit in the thermodynamic limit (Lee et al., 10 Jun 2026).

The DKCA has also served as a testbed for rigorous positivity-based bootstrap methods. In the synchronous and asynchronous $i$7 cases, one combines invariance or master equations with the positivity of finite cylinder-event probabilities,

$i$8

to build hierarchies of linear programs. For the synchronous model, the resulting lower bound is

$i$9

while for the asynchronous model the bound is

$t+1$0

For the asynchronous $t+1$1 model at $t+1$2, the same framework yields the dynamical half-life bound

$t+1$3

These are explicitly presented as rigorous bounds derived without model-specific monotonicity arguments (Cho, 19 May 2025).

5. Quantum generalizations and tensor-network constructions

Several quantum models retain the DK automaton as an exact classical shadow. In one discrete-time construction on a $t+1$4-dimensional spin lattice, the reduced state on each time slice remains separable even though the dynamics is quantum. The classical stochastic counterpart is precisely a DK process with effective activation probability

$t+1$5

Mean-field theory gives the density recursion

$t+1$6

with

$t+1$7

while the exact DK mapping places the numerically observed transition at the DK critical point $t+1$8. Near that transition, both connected density-density correlations and the local quantum uncertainty become long-ranged, even though the reduced state is separable (Lesanovsky et al., 2018).

A more general $t+1$9-dimensional quantum cellular automaton introduces a unitary three-body gate

$t$0

with

$t$1

Here $t$2 controls local entanglement generation. The empty row remains absorbing, the active and absorbing stationary regimes survive, and the phase boundary $t$3 depends on $t$4. Projected entangled pair states provide an exact finite-time representation of the full $t$5 state; the local gate admits an MPO decomposition with maximum bond dimension $t$6, and the evolved state remains a PEPS with $t$7. For $t$8, the critical point is near

$t$9

with decay exponent

$$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$0

consistent with one-dimensional directed percolation. A seed-based infinite-lattice MPO method then exploits the strict light cone of seed dynamics and finds

$$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$1

for $$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$2, and

$$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$3

for $$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$4, again consistent with $$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$5 directed percolation (Gillman et al., 2020, Gillman et al., 2020).

Non-unitary and explicitly quantum-coherent generalizations also take the DKCA as their classical limit. A Lindblad-based QCA with neighbor-conditioned Hamiltonian and jump operators reproduces classical DK dynamics when the coherent terms vanish and shows an absorbing/percolating transition near

$$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$6

while generating nonzero coherence in the active phase (Nigmatullin et al., 2021). In a distinct $$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$7 QCA embedding, the second-order Rényi entropy

$$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$8

acts as a measure of “present-past” entanglement. On the DK critical line

$$P[1|0,0]=0,\qquad P[1|0,1]=P[1|1,0]=p_1,\qquad P[1|1,1]=p_2,$$9

the reported exponents are

$$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$0

so coherence decays much faster than the entropy itself (Gillman et al., 2021). A further development maps the DK automaton to an isometric tensor-network state,

$$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$1

whose parent Hamiltonian has a degenerate ground-state manifold containing the absorbing product state and a second state that undergoes an entanglement-pattern transition from W-like long-range pairwise entanglement to trivial entanglement. At the critical point on the line $$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$2 with $$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$3 in that three-parameter notation, the mapped quantum state shows algebraic correlations in all spatial directions, with quoted exponents

$$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$4

and critical point

$$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$5

for the site-DP case (Boesl et al., 25 May 2026).

6. Learning-theoretic and coding-theoretic uses

The DK automaton has also been used outside conventional statistical-mechanics workflows as a structured source of critical spatiotemporal data. In coding theory, the binary-erasure-channel polarization recursion for conventional polar codes is approximated in the low-erasure regime by a DK automaton on a tilted square lattice. Under that approximation, the polar-code scaling exponent satisfies

$$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$6

where $$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$7 is the directed-percolation critical exponent. Using the best numerical estimate

$$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$8

the resulting prediction is

$$P[0|\sigma_{i-1},\sigma_{i+1}] = 1 - P[1|\sigma_{i-1},\sigma_{i+1}].$$9

On the bond-DP line,

$N$00

the same paper introduces a golden-ratio approximation leading to

$N$01

This is presented as an approximation rather than an exact identity, because the polarization tree is only approximated by the percolating square lattice (Shental, 2019).

Machine-learning studies use the DKCA as a benchmark nonequilibrium system whose raw configurations encode both phase information and universal scaling. A supervised/semi-supervised/unsupervised analysis of the $N$02-dimensional DK model considered bond DP, site DP, Wolfram rule 18, and compact directed percolation. In the supervised setting, CNN outputs were used not only to estimate critical points but also to perform finite-size/time scaling collapse. For the site-DP case, the reported estimates are

$N$03

consistent with the quoted directed-percolation values. The same study reports CNN critical-point estimates

$N$04

and corresponding estimates for Wolfram rule 18 and compact directed percolation (Tuo et al., 2023). A separate transfer-learning study trained a CNN on directed bond percolation images, fixed

$N$05

for the DK model, and extrapolated the critical point in $N$06 to

$N$07

with CNN and

$N$08

with DBSCAN, close to the benchmark value $N$09 used there (Saif et al., 2023).

These uses do not alter the automaton’s underlying dynamics. They show, rather, that DK configurations are sufficiently structured to support transfer learning, finite-size scaling from learned observables, and approximate communication-theoretic reinterpretations. This suggests that the automaton functions as a compact intermediary between nonequilibrium critical phenomena and data-driven inference, although the detailed identifications remain approximation-dependent in the coding-theory setting (Shental, 2019, Tuo et al., 2023).

7. Geometric, operator-theoretic, and quantized extensions

The DK automaton has been generalized to radial growth problems with mutation and selection. In that context, a generalized off-lattice DK model is introduced to minimize persistent lattice artifacts in radial expansions. For a colony with initial radius $N$10 expanding at velocity $N$11, the characteristic time

$N$12

marks the scale beyond which the inflating perimeter causes portions of the colony to become causally disconnected. The paper states that significant genetic demixing occurs only up to that finite time, after which inflation amplifies the effect of selection relative to genetic drift and modifies the underlying directed-percolation transition through new scaling functions and finite-size-like effects (Lavrentovich et al., 2012).

A different line of work treats the DK model as an interacting particle system and studies its spectral and zeta structures. In one formulation the local DK transition operator is

$N$13

with site directed percolation at $N$14, bond directed percolation at $N$15, and $N$16 reproducing Wolfram Rule 90. The same paper defines a quantization of the corresponding Markov chain as a unitary matrix

$N$17

and a zeta function

$N$18

For the $N$19 DK model, both the unitary blocks and the associated absolute zeta functions are computed explicitly (Akahori et al., 2024).

Closely related IPS/zeta work defines, for a finite path graph $N$20, the IPS-type zeta function

$N$21

where $N$22 is the global evolution operator built from the local DK rule. Its logarithm is expressed through traces of powers of the global operator, so the zeta coefficients encode return rates of configurations under the stochastic dynamics. In that setting the DK model is studied not primarily through its absorbing-state transition, but through an operator-theoretic correspondence between probabilistic cellular automata, spectra, and determinant invariants (Kiumi et al., 2022).

Taken together, these extensions show that the Domany–Kinzel automaton is not confined to a single methodological niche. It appears as a radial frontier model, an interacting-particle-system transfer operator, a source of zeta functions, and a quantized Markov chain. A plausible implication is that its continued usefulness derives from the coexistence of a minimal local rule, an absorbing-state transition, and enough algebraic structure to support exact reductions on special parameter lines (Lavrentovich et al., 2012, Akahori et al., 2024, Kiumi et al., 2022).