Higher-Order Cellular Automata (HOCA)

• Higher-Order Cellular Automata (HOCA) are models where the evolution of the system depends on multiple past configurations, integrating explicit temporal memory.
• They employ algebraic techniques, such as linear updates and operator representations, to analyze decisiveness, sensitivity, and phase transitions.
• HOCA offer practical applications in quantum codes, topological phases, subsystem symmetries, and even deep learning via high-order cross-modal attention.

Higher-Order Cellular Automata (HOCA) are cellular automata in which the evolution of the system at time step $t$ depends not only on the configuration at time $t-1$ but on configurations at several previous times, introducing explicit temporal memory or allowing the update law itself to dynamically depend on past states or the current configuration. This generalization encompasses a diverse set of formalisms, with significant theoretical and applicative consequences across discrete dynamical systems, theoretical computer science, statistical physics, quantum information, and artificial intelligence.

1. Formal Definitions and Mathematical Framework

The canonical definition of a higher-order cellular automaton extends a standard one-dimensional CA, which evolves via $c^{(t+1)} = F\left( c^{(t)} \right)$, to an evolution process where

$$c^{(t+1)} = F\left( c^{(t)}, c^{(t-1)}, \ldots, c^{(t-k+1)} \right)$$

for a memory size $k$ and alphabet $S$ (commonly $S = \mathbb{Z}_m$ or finite fields). The local rule $h$ determining each cell's update at location $i$ admits the form

$$c^{(t+1)}_i = h\left( c^{(t)}_{i-r}, \ldots, c^{(t)}_{i+r}, c^{(t-1)}_{i-r}, \ldots, c^{(t-1)}_{i+r}, \ldots, c^{(t-k+1)}_{i-r}, \ldots, c^{(t-k+1)}_{i+r} \right)$$

where $t-1$0 is the CA radius. This explicit temporal depth distinguishes HOCA from standard first-order CA.

A highly influential subclass is linear HOCA over the ring $t-1$1, where the total update is an $t-1$2-linear function of the $t-1$3 previous configurations. The global HOCA dynamical system acts on $t-1$4 by shifting the tuple and computing the new state using a linear local rule. This formalism naturally embeds HOCA into the theory of additive (vector-valued) cellular automata and permits reductions to algebraic dynamical systems, notably those involving companion or Frobenius matrices (Dennunzio et al., 2019).

An alternative, operator-level generalization is to encode the CA update as an operator $t-1$5 acting on a space of histories, or to decompose the local update rule into a polynomial in shift and delay operators, giving rise to "rule polynomials" such as those used in algebraic constructions of fracton topological phases (Zhang et al., 2023, Zhang et al., 19 Aug 2025).

2. Dynamical Properties: Sensitivity, Equicontinuity, and Decidability

The dynamical classification of HOCA is shaped by their sensitivity to initial conditions, equicontinuity, injectivity, and surjectivity, with linear HOCA being analytically tractable. For these, key results include:

• Sensitivity and Equicontinuity: For linear HOCA of memory $t-1$6 over $t-1$7, sensitivity to initial conditions and equicontinuity are decidable and equivalent properties. Using the algebraic normal form of the associated vector-valued CA (Frobenius LCA), sensitivity is characterized by the existence of polynomials in the update matrix with positive or negative Laurent degree (Dennunzio et al., 2019).
• Injectivity and Surjectivity: These properties for linear HOCA also reduce to computing the determinant of the associated matrix Laurent polynomial, leveraging known decidability results for linear CA over $t-1$8 (Dennunzio et al., 2019).
• Reduction to Vector CA: Any linear HOCA of memory $t-1$9 is topologically conjugate to a linear CA over $c^{(t+1)} = F\left( c^{(t)} \right)$0 with a Frobenius (companion) form, and vice versa, yielding algebraic methods for analyzing dynamical complexity and phase structure.

This algebraic machinery enables, for example, systematic construction of HOCA-engineered subsystem symmetries for quantum codes and symmetry-protected phases (Zhang et al., 2023, Zhang et al., 19 Aug 2025).

3. Rule Self-Reference and State-Dependent Dynamics

A distinctive variant arises when the CA rule itself is made a dynamical variable, dependent on the current or past global state—a "state-to-rule" feedback architecture. In the PICARD formalism, the CA has

• a microstate $c^{(t+1)} = F\left( c^{(t)} \right)$1 (the current configuration),
• a macrostate $c^{(t+1)} = F\left( c^{(t)} \right)$2 (the rule induced from $c^{(t+1)} = F\left( c^{(t)} \right)$3), with evolution

$c^{(t+1)} = F\left( c^{(t)} \right)$4

where $c^{(t+1)} = F\left( c^{(t)} \right)$5 is a mapping from configuration to rule (Pavlic et al., 2014). This genuinely higher-order feature allows the automaton to partition its state space into disconnected macro-execution regions. Invariant regions act like standard CA with fixed rules, but the global architecture encodes many possible CA behaviors in a coupled dynamical system, making the model self-referential and structurally adaptive.

Quantitative analysis demonstrates that the space of invariant macrostates ("macroexecutions") can be measured, revealing a Zipfian distribution reminiscent of language frequency distributions, and supporting the biological analogy with cell differentiation and evolutionary selection (Pavlic et al., 2014).

4. HOCA in Quantum Many-Body Physics and Subsystem Symmetries

Recent advances have realized HOCA as foundational organizers for symmetry-enriched and subsystem-protected topological phases (SPT and SET) in quantum matter. In the algebraic approach, a linear HOCA rule is encoded as a polynomial

$c^{(t+1)} = F\left( c^{(t)} \right)$6

where $c^{(t+1)} = F\left( c^{(t)} \right)$7 indexes spatial translations and $c^{(t+1)} = F\left( c^{(t)} \right)$8 encodes discrete time or vertical spatial layers. The configuration polynomial defines the allowed symmetric operators and encodes the recursive evolution

$c^{(t+1)} = F\left( c^{(t)} \right)$9

for $$c^{(t+1)} = F\left( c^{(t)}, c^{(t-1)}, \ldots, c^{(t-k+1)} \right)$$0-th order systems (Zhang et al., 2023, Zhang et al., 19 Aug 2025).

HOCA-generated symmetry operators act as subsystem symmetries (line-like, fractal, chaotic) on square or cubic lattices, and the universality theorem guarantees that all locally generated symmetry supports (finite, line-like, fractal, mixed, or chaotic) can be simulated by HOCA (Zhang et al., 19 Aug 2025, Zhang et al., 2023). In topological code constructions, the HOCA polynomial serves as a programmatic mobility code, fully classifying anyon mobility (fractons, lineons, mobile anyons) and symmetry-enriched fusion rules. Fusion outcomes for mobility classes realize non-Abelian fusion rules for Abelian anyon content, a direct consequence of the mobility polynomial structure (Zhang et al., 19 Aug 2025).

5. Hierarchical, Layered, and Diagrammatic HOCA Generalizations

Hierarchical and layered CA frameworks, as developed diagrammatically and algebraically, decompose the alphabet and rule space into nested layers and sublayers based on the factorization structure of the alphabet size (García-Morales, 2016). While not temporal higher-order in the strict sense, this yields a recursive spatial or algebraic order, with each digit or factor contributing an interacting subsystem, inducing coupled dynamics that can generate nested spatial patterns, domains, or chimera states.

Rules can be universally expressed via mixed-radix decompositions; graded rules yield independent CA in each layer, while $$c^{(t+1)} = F\left( c^{(t)}, c^{(t-1)}, \ldots, c^{(t-k+1)} \right)$$1-decomposable rules capture nonlinear coupling across layers. For instance, the majority rule can serve as a building block for constructing CA with nested domain structures and emergent subdynamics (García-Morales, 2016).

6. HOCA in Machine Learning: High-Order Cross-Modal Attention

The acronym HOCA has also been used for "High-Order Cross-Modal Attention" in the context of video captioning, where the attention mechanism employs a high-order correlation tensor to explicitly model multi-way frame-level interactions across modalities (image, motion, audio) (Jin et al., 2019). While not related to CA with temporal memory, this approach structurally leverages high-order interactions, and low-rank tensor factorization renders the method computationally efficient. Such models have demonstrated state-of-the-art results on large-scale multimodal video captioning benchmarks, illustrating the practical utility of higher-order architectures in deep learning.

7. Complexity, Entropy, and Information-Theoretic Properties

For higher-dimensional or higher-order CA, classical entropy measures often diverge or become uninformative. The entropy rate, defined as the normalized boundary growth of partition entropy, provides a finite, shift-commuting, and conjugacy-invariant complexity measure for multidimensional systems (Blanchard et al., 2012). The entropy rate scales with the number of "active" or permutative directions of the local rule and captures the complexity of information propagation at the relevant geometric scale in high-dimensional or higher-order automata.

---

Table: Principal HOCA Variants and Applications

Variant / Context	Formalism	Key Applications
Linear HOCA over $$c^{(t+1)} = F\left( c^{(t)}, c^{(t-1)}, \ldots, c^{(t-k+1)} \right)$$2	$$c^{(t+1)} = F\left( c^{(t)}, c^{(t-1)}, \ldots, c^{(t-k+1)} \right)$$3, linear $$c^{(t+1)} = F\left( c^{(t)}, c^{(t-1)}, \ldots, c^{(t-k+1)} \right)$$4	Decidable dynamics, compression, cryptography (Dennunzio et al., 2019)
Self-referential/state-to-rule HOCA (PICARD)	$$c^{(t+1)} = F\left( c^{(t)}, c^{(t-1)}, \ldots, c^{(t-k+1)} \right)$$5, $$c^{(t+1)} = F\left( c^{(t)}, c^{(t-1)}, \ldots, c^{(t-k+1)} \right)$$6	Differentiation, evolutionary modeling, language (Pavlic et al., 2014)
Polynomial HOCA for subsystem symmetries	$$c^{(t+1)} = F\left( c^{(t)}, c^{(t-1)}, \ldots, c^{(t-k+1)} \right)$$7, spatial-temporal polynomial	SET/SPT phases, anyon mobility, quantum codes (Zhang et al., 2023, Zhang et al., 19 Aug 2025)
Hierarchical/layered CA (spatial HOCA)	Mixed-radix decomposition, layered update rules	Pattern emergence, chimera states, spatial substructure (García-Morales, 2016)
High-Order Cross-Modal Attention (deep learning)	High-order correlation tensors, low-rank factorization	Video captioning, multimodal fusion (Jin et al., 2019)

---

Higher-Order Cellular Automata thus subsume a spectrum of temporally, spatially, and hierarchically extended cellular automaton models, with concrete mathematical structure and wide-ranging practical impact across discrete mathematics, statistical mechanics, quantum information, and machine learning.