Population Dynamics: Birth, Death, and the Grid
In 1971, Thomas Schelling — who would later win the Nobel Prize in Economics — published a paper with a disquieting argument. He placed two groups of agents on a checkerboard. Each agent had a modest preference: it was happy if at least a third of its neighbors were the same type as itself; unhappy otherwise. Unhappy agents moved to the nearest empty cell that satisfied their preference. Schelling ran the simulation on the checkerboard by hand, with pennies and dimes.
The result was total segregation. A mild preference for partial similarity — not a preference for segregation, just a tolerance threshold of roughly 33% similar neighbors — produced almost complete geographic sorting. The two groups clustered into separate territories, with nearly no mixed cells remaining.
The lesson was not about racism specifically (though Schelling intended it partly as a model of residential segregation). It was about the relationship between local preferences and global outcomes: when you allow spatial dynamics to operate, you get spatial outcomes that bear no simple relationship to the intentions of individual agents. The global pattern is not the intent of any actor; it is an emergent consequence of local rules operating on a grid.
This is the founding insight of spatial population modeling, and it is the same insight that drives cellular automata across all their applications. Population dynamics, seen through the CA lens, produces phenomena that population equations — which treat populations as aggregate numbers — cannot even formulate.
What Differential Equation Models Miss
Classical population demography operates with remarkable success in a narrow domain. Malthusian exponential growth describes unconstrained population increase correctly. The logistic model, which adds a carrying capacity K that slows growth as population approaches it, captures the basic dynamics of resource-limited growth. The Leslie matrix model correctly propagates age-structured populations forward in time, capturing the effects of age-dependent birth rates and death rates.
What all these models sacrifice is spatial heterogeneity. They describe a population as a number — or at most, a vector of numbers stratified by age — and they describe how that number changes. They cannot describe where in a country population grows, where it contracts, how a rural-urban gradient forms and steepens, how a wave of immigration propagates through a region, or how a local epidemic in one city spreads to neighboring cities at a specific rate tied to their geographic proximity.
These are not edge cases. They are often the central dynamics of interest. The 20th century’s great demographic shifts — the urbanization of previously rural countries, the depopulation of agricultural regions, the concentrated growth of megacities — are fundamentally spatial processes. Understanding them requires models that are spatial from the ground up.
CA models provide this.
Spatial Lotka-Volterra: Human Demography as Ecology
The Lotka-Volterra equations, originally developed to describe predator-prey dynamics in animal populations, have a natural extension to human population dynamics. In a spatial Lotka-Volterra model of regional population:
Each cell in a grid represents a geographic region — a county, a district, a grid square of land. The cell’s state encodes the population density (or, in a simplified model, whether the cell is populated, depopulated, or in transition). At each step, populated cells can “reproduce” into adjacent underpopulated cells (the spatial analog of population growth driving migration), and cells can “die” (become depopulated) if their population falls below a viability threshold due to resource depletion or economic collapse.
The signature result of spatial Lotka-Volterra models is that spatial structure produces heterogeneous equilibria — and it does so robustly. Durrett and Levin demonstrated this with mathematical precision in their 1994 paper “The Importance of Being Discrete (and Spatial)” (Theoretical Population Biology): move from a well-mixed model to a spatially explicit one, and the qualitative behavior changes entirely. Well-mixed models predict that all regions converge to the same equilibrium density — or to extinction, or to maximal density, depending on parameters. Put space back in, and you get persistent heterogeneity instead: dense patches and sparse patches, with stable boundaries maintained by the local dynamics even when global averages suggest they should not exist.
This matches what we observe. The world’s population is not uniformly distributed. It is clustered into dense urban nodes and sparse rural areas, with a persistent gradient between them that has existed for millennia and shows no sign of converging to uniformity. The spatial CA framework provides a mechanistic explanation: the local feedback between population density, resource availability, and migration creates and maintains this heterogeneity. The mean-field equation cannot produce it by construction.
The SIRS Model: Population Dynamics with Memory
The SIR (Susceptible-Infected-Recovered) model of epidemic spread is a population dynamics model in the demographic tradition — it tracks aggregate numbers of individuals in each compartment. But when it is extended to the SIRS model — where recovered individuals eventually lose immunity and become susceptible again — and when it is implemented spatially as a CA, it produces dynamics that reveal a fundamental truth about population heterogeneity.
In a spatial SIRS CA, each cell is in one of three states: Susceptible (can be infected), Infected (can infect neighbors), or Recovered/Resistant (cannot be infected for a period). The state transition rules are:
- A Susceptible cell becomes Infected if it has at least one Infected neighbor, with probability p (infection rate)
- An Infected cell becomes Recovered after a fixed or probabilistic number of steps (recovery rate)
- A Recovered cell becomes Susceptible again after a fixed or probabilistic number of steps (immunity waning rate)
In a well-mixed (mean-field) SIRS model, the dynamics converge to a steady state where all three compartments are permanently present at fixed proportions — an endemic equilibrium. The epidemic never ends; it persists as a stable fraction of infections in the population.
In a spatial SIRS CA, the dynamics are qualitatively different. The infection spreads as a traveling wave through the susceptible population, leaving recovered (temporarily immune) cells behind. When the wave reaches a region where most cells are recovered, it falters. The susceptibles rebuild over time as immunity wanes; then a new wave can propagate. The global dynamics are not a steady state but a spatial oscillation: waves of infection sweeping through waves of recovery.
The heterogeneous outcome is the one that matches real epidemic patterns. The repeated waves of influenza, the cycling of measles in pre-vaccine populations, the spatial propagation of COVID-19 through metropolitan regions — these are spatial wave phenomena, produced by the local interaction dynamics of the SIRS process. The connection to real data was established explicitly by Fuentes and Kuperman (1999), whose paper “Cellular Automata and Epidemiological Models with Spatial Dependence” (Physica A, 267:471–486) showed that spatial CA dynamics reproduced the characteristic wave patterns of measles outbreaks in ways that the mean-field SIRS model could not. The steady state of the mean-field model is not wrong; it is an average over a heterogeneous process, and the average hides the dynamics that matter.
Migration and the Speed of Colonization Fronts
A specific prediction of spatial population CA models that has been validated against real demographic data is the speed of colonization fronts.
When an agricultural or herding population expands into sparsely populated territory — the spread of farming through Neolithic Europe, the settlement of North American prairie in the 19th century, the advance of Brazilian agricultural settlements into Amazonia in the 20th century — the expansion occurs as a spatial wave: a front of demographic growth moving through space at a characteristic speed.
The speed of this front is not determined by the global population size or the global growth rate. It is determined by the local birth rate, the local carrying capacity, and the dispersal distance — the typical distance over which offspring move from their parents’ location. CA models capture all of these local parameters explicitly. The predicted front speed scales as the square root of the product of the local growth rate and the dispersal distance squared — a result that was derived analytically for the related reaction-diffusion equations and is reproduced in discrete CA models.
This prediction has been tested against archaeological and historical data, with reasonable success. The spread of farming from the Near East into Europe appears to have proceeded at roughly 1 kilometer per year — consistent with a model of demic diffusion (population spread) with local growth rates and dispersal distances typical of early agricultural societies. The CA framework makes this a precise, testable prediction rather than a narrative description.
Urban-Rural Dynamics: The CA of Concentration
The demographic transition from rural to urban society — one of the defining processes of the last two centuries — is structurally a CA process.
In a spatial model of urbanization, each cell represents a location that can host population, jobs, and amenities. The local rule captures a feedback: locations with high population density attract more jobs and amenities (economies of scale, labor market thickness, infrastructure investment), which in turn attract more population. This is a positive feedback — a birth rule in which a cell’s attractiveness increases with its neighbors’ density — combined with a negative feedback at very high densities (congestion, housing costs, disease).
The result of this simple rule, applied across a landscape with a historical pattern of initial population distribution, is the spontaneous concentration of population into a small number of large urban nodes. The cities that exist today are not, in most cases, the result of deliberate planning or natural advantage (rivers and harbors matter, but they do not determine city size uniquely). They are the result of local positive feedbacks operating over historical time on an initially heterogeneous landscape.
The CA framework makes an additional prediction: once a city has established itself as a node, the local positive feedbacks that created it continue to reinforce it, and displacing it through policy is extraordinarily difficult. The obverse is equally telling: cities built in deliberate isolation — without an existing positive feedback network to attach to — tend to stagnate regardless of investment. Naypyidaw, Myanmar’s purpose-built capital relocated to an empty plateau in 2005, remains a largely vacant administrative shell despite massive state expenditure; Egypt’s New Administrative Capital, begun in 2015 to relieve pressure on Cairo, has struggled to draw residents and businesses away from the established node it was meant to replace. The positive feedbacks that made Cairo Cairo do not transfer by decree. Attempts that succeed — Canberra, Brasília — are partial exceptions that tend to confirm the rule: both remain significantly smaller than the established urban centers they nominally parallel, and both required decades of mandated government presence to reach even modest critical mass.
This is the spatial CA of demography. It is not a metaphor; it is the actual mechanism.
The Result Schelling Could Not Prove
Schelling’s 1971 checkerboard model established the founding result of spatial population dynamics: local rules that are individually mild in their implications can produce global patterns of dramatic intensity.
But Schelling worked with a checkerboard and pennies. He could not explore parameter space systematically. He could not ask: at what threshold of local preference does segregation appear? Is the transition gradual or sharp? Does noise — random movement — change the outcome?
Subsequent CA models provided the answers. The transition from mixed to segregated outcomes in Schelling-type models is sharp: it has the character of a phase transition. Below a critical tolerance threshold (roughly one-third similar neighbors required for satisfaction), the system remains mixed with small perturbations. Above this threshold, the system undergoes a rapid reorganization into segregated clusters. This result is robust across a wide range of model specifications and has been confirmed in many independent implementations.
The policy implication is direct. If the segregation phase transition occurs at a low tolerance threshold — if only mild in-group preference is needed to produce strong segregation — then reducing segregation requires either reducing that preference (a long social project) or increasing the mixing probability (affirmative housing policies that directly modify the local rules). The CA model specifies the intervention point precisely: you need to change the local rule, because the global pattern is entirely determined by the local rule.
This is the power of spatial population modeling. It does not just describe demographic patterns; it identifies the mechanisms that produce them and the leverage points at which those mechanisms can be changed.