Population Dynamics: Predator, Prey, and the Grid
The Lotka-Volterra equations have a beautiful, deceptive simplicity.
Alfred Lotka (1925) and Vito Volterra (1926) independently derived the same pair of differential equations describing predator-prey dynamics. Prey grow exponentially in the absence of predators; predators die exponentially in the absence of prey; predators eat prey at a rate proportional to their encounter frequency; and encounter frequency is proportional to the product of both populations. The result is sustained oscillation — rabbits rise, foxes follow, rabbits crash, foxes crash, rabbits recover — cycling indefinitely around an equilibrium that neither population reaches.
It is elegant. It is also wrong in a specific, instructive way.
The Lotka-Volterra model assumes that every predator encounters every prey item with equal probability — that the world is a perfectly mixed soup where spatial location is irrelevant. Real ecosystems are nothing like this. A fox in northern Finland has no encounter with a rabbit in Patagonia. Populations are spatially structured: organisms interact with their neighbors, offspring disperse locally, and the spatial arrangement of organisms across a landscape determines what ecological dynamics are possible.
When you replace the well-mixed assumption with spatial structure — when you turn the Lotka-Volterra world into a cellular automaton — the dynamics change in ways that are not merely quantitative corrections. They are qualitatively different phenomena.
Paulien Hogeweg: CA as an Ecological Paradigm
In 1988, the Dutch computational biologist Paulien Hogeweg published a paper titled “Cellular automata as a paradigm for ecological modeling” in Applied Mathematics and Computation. It was not the first application of CA to ecology, but it was the paper that made the case systematically, demonstrating that spatial CA models of predator-prey systems produced dynamics that were qualitatively distinct from — and in many respects richer than — the well-mixed equations.
Hogeweg’s spatial predator-prey CA works as follows. Each cell in a grid can be in one of three states: empty, occupied by prey, or occupied by predator. At each time step, prey cells reproduce into adjacent empty cells with some probability; predator cells eat adjacent prey cells (removing the prey and surviving) with some probability; predator cells die with some probability. These are local rules, applied simultaneously to every cell.
The dynamics that emerge are not oscillations in the Lotka-Volterra sense. They are traveling waves. Prey spread outward into empty space; predators follow the wave of prey, consuming it from behind; the depleted region in the predators’ wake becomes empty; prey then recolonize the empty space; and the cycle repeats as a spatial wave rather than a temporal oscillation. Different parameter regimes produce qualitatively different dynamics: stable spirals (predator and prey waves rotating around a common center), chaotic patchwork (irregular, aperiodic patterns of prey and predator density), and extinction (one or both populations die out).
The most important result is that coexistence — both predator and prey persisting indefinitely — is far more robust in spatial models than in well-mixed models. In the Lotka-Volterra equations, coexistence requires precise parameter matching; slightly off, and one population wins and the other goes extinct. In Hogeweg’s spatial CA, coexistence occurs over a wide range of parameters, because spatial structure creates local refugia: regions of high prey density that are momentarily beyond the predator front, where prey can rebuild. The spatial structure maintains the diversity that mean-field equations cannot.
Wa-Tor: The Popularization of Ecological CA
A year before Hogeweg’s paper appeared in the scientific literature, Alexander Dewdney published a popularization in the December 1984 issue of Scientific American. Dewdney’s Wa-Tor (short for water-torus, reflecting the toroidal topology) was a simulated ocean populated by fish and sharks on a 2D grid. The rules were simple: fish moved, reproduced, and died; sharks moved, ate fish, reproduced, and starved if they went too long without eating.
Wa-Tor was the first widely accessible spatial predator-prey simulation, and it demonstrated to a broad audience something that the academic literature had been establishing for years: spatial structure produces pattern. In Wa-Tor, you could watch waves of fish being followed by waves of sharks, predator fronts advancing through prey-dense regions, and the spontaneous formation of fish refugia — clusters of fish that had been isolated from predators long enough to build up high densities.
The qualitative lessons of Wa-Tor were the same as those of Hogeweg’s more formal analysis: coexistence is more robust in space, dynamics are richer in space, and the patterns you see are fundamentally spatial phenomena with no analog in the differential equation description.
The Forest Fire Model: Self-Organized Criticality
In 1992, Barbara Drossel and Franz Schwabl published a three-state CA that has become one of the canonical models in complexity science. The Drossel-Schwabl forest fire model appears in Physical Review Letters (volume 69, pages 1629–1632) and works as follows:
Each cell in a grid is in one of three states: empty, tree, or burning. At each step:
- A burning tree becomes empty (it burns down)
- A tree adjacent to a burning tree catches fire (becomes burning)
- An empty cell grows a tree with small probability p (regrowth)
- A tree catches fire spontaneously with very small probability f (lightning)
The model is run in the limit where f/p → 0 — lightning is extremely rare compared to tree growth — which corresponds to the biological reality that forests regrow much faster than catastrophic fires occur.
In this limit, the forest fire model exhibits self-organized criticality: the distribution of fire sizes follows a power law with no characteristic scale. Small fires are common; medium fires are less common; large fires are rare but they do occur, and the relationship between frequency and size follows the same mathematical form across all size scales. There is no “typical” fire size.
This was not a parameter-tuned result. Drossel and Schwabl did not choose their parameters to produce power-law scaling. The power-law behavior emerged from the dynamics of the model without any fine-tuning — hence “self-organized.” The model finds its own critical point and stays there, because the slow regrowth ensures that the forest is always near the percolation threshold where fires can just barely span the grid.
The match to real forest fire data was striking. Actual forest fire size distributions in several ecosystems — chaparral in Southern California, boreal forest in Canada, eucalyptus forests in Australia — follow power laws over several decades of size. The Drossel-Schwabl model provides a mechanistic explanation: real forests are near the critical density for fire percolation, not by coincidence, but because the combination of slow growth and rare ignition naturally drives them there.
The model also connects to Conway’s Life through the concept of percolation, which is closely related to the emergence of infinite structures in CA. Near the percolation threshold, small local events have disproportionate global consequences — a single spark can burn the entire forest. This is the ecological analog of the sensitivity to initial conditions that makes Life’s dynamics unpredictable.
Coral Reef and Patch Dynamics
Marine ecology provided some of the earliest and most influential applications of spatial CA models outside of the predator-prey framework. Coral reef ecosystems are spatially structured at multiple scales: individual polyps compete for space on the reef, coral patches compete with algae for large-scale area, and the distribution of fish grazers (which keep algae in check) follows spatial gradients tied to reef structure.
Models of coral reef dynamics as CA — developed by researchers including researchers working in the tradition of Menge and Sutherland’s spatial ecology work — showed that the spatial patchwork of corals, algae, and grazers could maintain a diversity of states that would collapse to a single dominant state in a well-mixed model. The phenomenon is called the spatial insurance effect: even if algae competitively dominate coral in a locally well-mixed patch, the patchiness of grazer activity means that some coral patches are always recovering from algal overgrowth, and the global system maintains both state.
The degraded state of many real coral reefs — the transition from coral-dominated to algae-dominated systems following bleaching events or fishing pressure — can be understood, in CA terms, as a phase transition: a shift from a heterogeneous attractor (mixed coral-algae patchwork) to a homogeneous attractor (uniformly algae-dominated). Once the transition occurs, recovery is difficult because the spatial feedback that maintained the heterogeneous state has been disrupted.
Spatial Structure and the Fundamental Theorem of Ecology
The central finding from half a century of ecological CA modeling can be stated as a single principle:
Spatial structure changes what ecosystems can do.
This is not a quantitative refinement of mean-field models. It is a qualitative statement about a different category of phenomena. The traveling waves, the spatial refugia, the self-organized criticality, the patch dynamics — none of these exist in the differential equation description of the same systems. They are emergent properties of spatial structure itself.
The reason goes back to Conway’s core insight. When interactions are local, what happens in one region depends on what happened nearby, not on what is happening everywhere. This creates the possibility of spatial heterogeneity — different dynamics in different places — and it is this heterogeneity that is the source of ecological complexity.
Consider extinction. In a well-mixed model, a population goes extinct when its global numbers fall below a critical level. In a spatial model, a population can persist globally even when it is locally eliminated in many places, because local extinction is followed by recolonization from neighboring patches. Conversely, a population can go extinct globally through a cascade of local extinctions — a spatial wave of extinction — even when its global numbers initially appear adequate.
This spatial dynamics of extinction is directly relevant to conservation biology. The fragmentation of habitats — the conversion of continuous natural landscapes to isolated patches — eliminates the recolonization process that spatial CA models identify as the key mechanism of persistence. A population in a fragmented landscape is not just smaller; it is dynamically different. It has lost the spatial coupling that maintains persistence. The CA framework makes this explicit.
What the Grid Reveals
Ecological CA models are not primarily useful because they are more accurate than differential equations — in some contexts, the differential equations are more tractable and equally accurate. They are primarily useful because they make spatial dynamics visible, and spatial dynamics are often the determining factor in ecological outcomes.
The glider in Conway’s Life moves because its rule produces asymmetric dynamics that propagate across space. The predator-prey wave in Hogeweg’s model moves because the local consumption-regrowth dynamics propagate outward from their origin. The forest fire in the Drossel-Schwabl model scales as a power law because the percolation threshold creates a critical state where local ignition events cascade unpredictably across space.
In each case, the interesting phenomenon is a spatial phenomenon. The differential equation misses it entirely. The CA captures it by construction, because local rules applied across space are exactly what produces spatial dynamics.
The ecosystem is not a well-mixed soup. It is a grid — irregular, multilayered, and fantastically complex, but a grid nonetheless. Ecological CA models are not analogies. They are descriptions.