Developmental Graph Cellular Automata

• DGCA are dynamic systems in which both local states and graph topology evolve under causal, homogeneous, and bounded local rules.
• They leverage a rigorous framework ensuring causality, continuity, and reversibility to support complex, time-varying graph dynamics.
• DGCA find applications in theoretical physics and neural reservoir computing, offering models for adaptive morphogenesis and distributed computation.

Developmental Graph Cellular Automata (DGCA) designate graph-based cellular automata in which both local states and graph structure evolve over discrete time under local, homogeneous update laws. In the most general formal sense supported by the literature, a DGCA is a causal, bounded-degree, time-varying graph dynamics: vertex and edge labels may change, vertices and edges may be created or deleted, and information propagates only within a bounded radius per step. In Arrighi and Dowek’s framework, this class is captured by causal graph dynamics and, equivalently, by localizable dynamics obtained from synchronous application of a bounded, consistent local rule to radius-$r$ neighbourhoods (Arrighi et al., 2012). In the generalized Cayley-graph formalism, the same idea is recast as a shift-invariant, continuous, bounded cellular automaton on pointed graphs modulo isomorphism (Arrighi et al., 2012). More recent work also uses the term DGCA for neural morphogenetic systems that grow directed computational graphs from single-node seeds, notably for reservoir computing (Barandiaran et al., 11 Aug 2025).

1. Historical emergence and scope

The development of DGCA-relevant formalisms proceeds from graph cellular automata on fixed arbitrary topologies to fully developmental, time-varying graph dynamics, and then to application-oriented neural growth systems.

Source	Emphasis	Relation to DGCA
"Outer-totalistic cellular automata on graphs" (0812.2408)	CA rules on arbitrary topologies	Fixed-topology graph CA baseline
"Causal graph dynamics" (Arrighi et al., 2012)	Time-varying labelled graphs under causality	Core axiomatic foundation
"Generalized Cayley Graphs and Cellular Automata over them" (Arrighi et al., 2012)	Compact metric configuration space and CHL-style characterization	Formal CA theory for evolving graphs
"Reversible Causal Graph Dynamics" (Arrighi et al., 2015)	Invertibility, reversibility, block representation	Constraints on reversible developmental dynamics
"Growing Reservoirs with Developmental Graph Cellular Automata" (Barandiaran et al., 11 Aug 2025)	Neural morphogenesis of directed reservoirs	Functional instantiation of DGCA

A common narrowing is to identify DGCA exclusively with single-seed neural growth systems. The formal literature supports a broader meaning. The 2012 graph-dynamics papers describe a class of synchronous, deterministic, local graph evolutions in which structure and state co-evolve under homogeneity and bounded-speed constraints; the 2025 reservoir work is a specific instantiation within that broader conceptual space (Arrighi et al., 2012, Arrighi et al., 2012, Barandiaran et al., 11 Aug 2025).

This broader scope matters because it places developmental graph automata alongside dynamical networks, Boolean networks, generative network automata, graph automata, graph rewriting automata, parallel graph transformations, amalgamated graph transformations, and time-varying graphs. The provided examples in the foundational work suggest applications ranging from complex systems science to theoretical physics (Arrighi et al., 2012).

2. Configuration spaces and graph representation

In the causal-graph-dynamics formalism, the underlying object is a countable, port-labelled, directed graph with labels on vertices and edges. One fixes an uncountable set $V$ of names, a degree bound $\pi$, vertex states in $\Sigma$, and edge states in $\Delta$. A graph $G$ has an at most countable vertex set $V(G)\subseteq V$, ports $1,\ldots,\pi$, a partial vertex-label function $\sigma_G:V(G)\to\Sigma$, and a partial edge-label function

$$\delta_G : (V(G)\times 1..\pi)\times (V(G)\times 1..\pi)\to\Delta,$$

with each port used at most once. Adjacency is defined by the existence of an edge on some pair of ports, ignoring orientation for the induced graph metric. Balls of radius $V$0 and disks $V$1 then provide the local neighbourhoods on which update rules act (Arrighi et al., 2012).

The generalized Cayley-graph formalism reframes the same problem in terms of pointed graphs modulo isomorphism. Vertices are anonymous up to isomorphism and are addressed by path classes relative to a distinguished origin. The configuration space is the set $V$2 of labelled adjacency structures, isomorphic to pointed graphs modulo. This representation preserves three CA-relevant features associated with ordinary Cayley graphs: the ability to name vertices relative to a point, a generalized notion of translation via pointer shift, and a topology suitable for continuity arguments (Arrighi et al., 2012).

A central topological tool is the disk metric

$V$3

which makes two configurations close when their pointed disks agree to large radius. With finite label sets and bounded degree, this space is compact. That compactness underwrites uniform continuity and enables a Curtis–Hedlund–Lyndon-style characterization of graph cellular automata on evolving bounded-degree graphs (Arrighi et al., 2012).

The representational consequence is that DGCA are not defined by absolute vertex identifiers or embedding in a fixed lattice. They are defined relationally: locality is measured by graph distance, symmetry is expressed by renaming invariance or pointer shift, and graph structure itself is part of the state.

3. Causality, local rules, and the localizable form

The decisive technical move in the DGCA literature is to replace ordinary translation invariance by graph isomorphism invariance and to formalize bounded-speed information flow on time-varying graphs. In Arrighi and Dowek’s formulation, a graph dynamics

$V$4

must satisfy conjugacy and freshness. Conjugacy requires that for every isomorphism $V$5 there exists a conjugate isomorphism $V$6 such that $V$7; freshness constrains name creation so that disjoint inputs do not generate a common new vertex (Arrighi et al., 2012).

Local dependence is mediated by antecedent codynamics. For each output vertex name $V$8, the set $V$9 contains those input vertices that must have been present whenever $\pi$0 appears after applying $\pi$1. Causality is then defined by two boundedness conditions: first, there exists a radius $\pi$2 such that the induced neighbourhood around $\pi$3 depends only on the disk $\pi$4 around some antecedent $\pi$5; second, there exists a bound $\pi$6 such that each input vertex can be an antecedent of at most $\pi$7 output vertices (Arrighi et al., 2012).

The constructive counterpart is the notion of a local rule

$\pi$8

where $\pi$9 is the set of disks of radius $\Sigma$0. Such a rule must be bounded and consistent: the graph fragments produced from overlapping disks cannot disagree on shared names, labels, or ports. A dynamics is localizable when

$\Sigma$1

This is the synchronous “apply everywhere in parallel and union the outputs” picture traditionally associated with cellular automata (Arrighi et al., 2012).

The foundational theorem is the equivalence

$\Sigma$2

This is the graph analogue of the Curtis–Hedlund–Lyndon theorem: bounded-speed, homogeneous graph evolution is exactly parallel application of a local rule (Arrighi et al., 2012). The generalized Cayley-graph formalism proves the same equivalence using the compact metric space of pointed graphs modulo, stating that a function is localizable if and only if there exists a vertex-tracking map $\Sigma$3 such that $\Sigma$4 is a causal dynamics, namely shift-invariant, continuous, and bounded (Arrighi et al., 2012).

Several stability theorems follow. Causal graph dynamics are stable under composition; they admit restriction to radius one; in the finite-label case, causality coincides with continuity; and invertible finite-label causal dynamics have causal inverses (Arrighi et al., 2012). These are the core structural reasons DGCA can be treated as a mathematically coherent generalization of classical cellular automata rather than as an ad hoc graph-rewriting scheme.

4. Developmental topology change and reversible constraints

DGCA are developmental because topology is endogenous to the update. The global map may create new vertex names, delete existing vertices, and rewire incident edges, provided freshness and locality are respected. In the foundational examples, graph inflation is explicit: an inflating grid replaces each vertex by a small cluster, producing a finer grid at the next step while preserving causal boundedness (Arrighi et al., 2012). In the generalized Cayley-graph treatment, the corresponding local rule is radius $\Sigma$5, acts uniformly on degree-$\Sigma$6 port-labelled grid vertices, and yields globally consistent inflation through carefully chosen local names and overlaps (Arrighi et al., 2012).

This developmental freedom is not unrestricted when reversibility is imposed. Reversible causal graph dynamics add invertibility as a further physics-like symmetry. The central theorem states that if a causal graph dynamics is invertible, then its inverse is also causal (Arrighi et al., 2015). However, the same work proves an “almost-vertex-preserving” constraint: there exists a bound $\Sigma$7 such that for graphs with more than $\Sigma$8 vertices, the vertex map $\Sigma$9 is bijective. In effect, sufficiently large reversible graph dynamics cannot continue to create or destroy vertices arbitrarily (Arrighi et al., 2015).

A common misconception is that reversible developmental graph automata can support unrestricted persistent growth. The reversible CGD results point in the opposite direction: strict reversibility and unbounded topological growth are in tension within the bounded-degree, causal, shift-invariant setting. The paper’s “turtle dynamics” shows that small symmetric cases can still exhibit reversible changes in vertex count, but the general theorem enforces vertex preservation beyond a finite threshold (Arrighi et al., 2015).

The same paper also gives a block representation theorem. Any reversible causal graph dynamics admits a finite-depth circuit decomposition into local reversible operations, formulated via marked graphs, a local mark operation $\Delta$0, and a conjugate local gate $\Delta$1. This places reversible DGCA conceptually near reversible cellular automata and lattice-gas automata (Arrighi et al., 2015).

5. Fixed-topology graph cellular automata as a neighbouring formalism

An important neighbouring lineage is the study of outer-totalistic cellular automata on arbitrary graphs. Marr and Hütt consider a graph $\Delta$2 with adjacency matrix $\Delta$3, binary node states $\Delta$4, synchronous updates, and neighbourhoods defined by input nodes. The key local statistic is the density

$\Delta$5

with $\Delta$6. Homogeneity requires one rule for all nodes, isotropy requires dependence only on $\Delta$7 and $\Delta$8 rather than neighbour ordering, and functional simplicity makes the rule piecewise constant in $\Delta$9 (0812.2408).

For degree-$G$0 regular graphs, outer-totalistic rules are encoded by triples $G$1 with $G$2, according to whether neighbour density is $G$3, $G$4, or $G$5. There are $G$6 such elementary rules, and a symmetry operation maps $G$7 to $G$8 under state inversion. For regular graphs of degree $G$9, the same idea becomes a threshold rule with $V(G)\subseteq V$0 (0812.2408).

The paper uses Shannon entropy $V(G)\subseteq V$1, word entropy $V(G)\subseteq V$2, and DFA exponents to classify dynamics and identify topologically sensitive rules. Among 34 inequivalent outer-totalistic rules after symmetry reduction, 14 are labelled topologically sensitive by their entropy criterion. The particularly notable cases are $V(G)\subseteq V$3, $V(G)\subseteq V$4, $V(G)\subseteq V$5, and $V(G)\subseteq V$6, corresponding respectively to ECA rules 37, 108, 109, and 133. The authors also find that on regular graphs the absolute number of complex patterns first rises and then falls with degree, reaching a maximum at $V(G)\subseteq V$7 (0812.2408).

These models are not developmental in the strict CGD sense because the formalism described here does not itself update topology. A plausible implication is that they serve as a fixed-topology baseline for DGCA research: they isolate how rule form and graph structure interact before one allows genuine creation, deletion, and rewiring. Their applications to scale-free graphs and metabolic networks also show that graph CA can be used diagnostically to probe how topology stabilizes or destabilizes dynamics (0812.2408).

6. Neural DGCA and functional morphogenesis

A recent application-specific use of the term DGCA appears in a neural, growth-capable graph dynamical system for morphogenesis and reservoir computing. In that model, the evolving object is a directed graph $V(G)\subseteq V$8 grown from a single-node seed over $V(G)\subseteq V$9 developmental steps. Each node is a “cell” with one of three discrete states, and each step consists of a neighbourhood aggregation followed by two neural stages: an action stage and a state stage. The action stage is a 2-layer MLP with 64 hidden units that outputs one of three actions—division, removal, or stasis—plus discrete edge-pattern choices for newborn nodes. The state stage is a single-layer perceptron that updates node type after topology changes. Nodes that would become isolated are removed, and a budget constraint enforces $1,\ldots,\pi$0, with experiments using $1,\ldots,\pi$1 (Barandiaran et al., 11 Aug 2025).

After growth, the directed graph is converted into a reservoir. Edges are bipolarized by node-state agreement:

$1,\ldots,\pi$2

Approximately two-thirds of nodes receive input; input weights are drawn from $1,\ldots,\pi$3 and scaled by $1,\ldots,\pi$4; recurrent weights are scaled by $1,\ldots,\pi$5. Reservoir dynamics follow

$1,\ldots,\pi$6

with per-node activation determined by cell type, using a mixture of linear and $1,\ldots,\pi$7 neurons. Readout uses Bayesian Ridge Regression for 3000 epochs with learning rate $1,\ldots,\pi$8, and performance is measured by normalized root mean square error (Barandiaran et al., 11 Aug 2025).

Training is search-based because division, removal, and edge-choice decisions are discrete. The paper uses a Microbial Genetic Algorithm with two chromosomes, one for the action MLP and one for the state SLP. Population initialization yields 100 genetically complete individuals from 10 MLP chromosomes and 10 SLP chromosomes; loser replication copies winner genes with probability $1,\ldots,\pi$9 per gene; mutation rate is $\sigma_G:V(G)\to\Sigma$0; 1000 MGA iterations are run per experiment; and 150 independent runs are stored as the best individuals for each condition (Barandiaran et al., 11 Aug 2025).

The resulting morphologies are classified as Linear, Loosely Stranded, or Other. With budget 100, Linear structures are most prevalent at about $\sigma_G:V(G)\to\Sigma$1 of best reservoirs; with budget 200, Other becomes more common at about $\sigma_G:V(G)\to\Sigma$2. For NARMA-10, DGCA-grown reservoirs significantly outperform random ESN controls: control versus 100-node DGCA gives $\sigma_G:V(G)\to\Sigma$3, control versus 200-node DGCA gives $\sigma_G:V(G)\to\Sigma$4, and increasing budget from 100 to 200 improves performance with $\sigma_G:V(G)\to\Sigma$5. For the 200-node setting, harder NARMA tasks drive the fittest reservoirs toward higher KR, GR, and LMC: KR increases from $\sigma_G:V(G)\to\Sigma$6 for NARMA-10 to $\sigma_G:V(G)\to\Sigma$7 for NARMA-20 and $\sigma_G:V(G)\to\Sigma$8 for NARMA-30, while GR and LMC also rise markedly (Barandiaran et al., 11 Aug 2025).

The same work also studies task-independent growth via Kernel Rank, Generalization Rank, Linear Memory Capacity, Spectral Radius, and a combined objective $\sigma_G:V(G)\to\Sigma$9. GR-optimized and LMC-optimized reservoirs are the best single-metric proxies for NARMA performance, whereas SR-only optimization performs worst. Reservoirs optimized for the combined “All” objective never see NARMA during training yet still perform strongly across NARMA-10, NARMA-20, and NARMA-30, although task-grown reservoirs remain statistically superior: $$\delta_G : (V(G)\times 1..\pi)\times (V(G)\times 1..\pi)\to\Delta,$$0, $$\delta_G : (V(G)\times 1..\pi)\times (V(G)\times 1..\pi)\to\Delta,$$1, and $$\delta_G : (V(G)\times 1..\pi)\times (V(G)\times 1..\pi)\to\Delta,$$2 for NARMA-10, NARMA-20, and NARMA-30 respectively (Barandiaran et al., 11 Aug 2025).

Within this application literature, DGCA therefore denotes functional morphogenesis rather than merely structural self-organization. The graphs are not optimized as final adjacency matrices; the trainable object is the developmental rule itself. The authors state that these results lay the foundation for DGCA systems that produce plastic reservoirs and for modeling functional, adaptive morphogenesis (Barandiaran et al., 11 Aug 2025).