The Schelling Segregation Model

In 1969, Thomas Schelling placed pennies and dimes on a checkerboard in his office at Harvard. He was an economist studying racial housing patterns, and the most powerful computer at MIT still required a room. He ran the simulation by hand.

The rule was mild: each coin is satisfied if at least one-third of its neighbors are the same type. If unsatisfied, move to a random empty cell. Repeat until no coin wants to move.

Schelling expected moderate integration. He got near-total separation. The outcome looked like the product of agents demanding 80 or 90 percent homogeneity. His agents had demanded only 33 percent. He published the result in the Journal of Mathematical Sociology in 1971, and the gap between individual preference and collective outcome became one of the foundational puzzles of complex systems research.

Setup

Two agent types occupy cells on a two-dimensional grid. A fraction of cells are left empty --- typically 20 to 30 percent. Empty cells are the resource that makes movement possible; without vacancy, no agent can relocate.

Each agent has a Moore neighborhood of eight surrounding cells (or fewer at grid boundaries). The only state each agent carries is its type and its position. The only information available to each agent is the composition of its immediate neighborhood.

The grid may be finite with boundaries (agents at edges have fewer than eight neighbors) or wrapped as a torus (no edge effects). Both produce qualitatively similar behavior. Extensions place agents on random graphs, scale-free networks, or empirically calibrated street maps, with neighborhood defined by network adjacency rather than spatial proximity.

The Rule

Each agent computes the fraction of its occupied neighboring cells that share its type. If that fraction meets or exceeds the tolerance threshold f, the agent is satisfied and stays. If the fraction falls below f, the agent is dissatisfied and moves to a randomly selected empty cell.

Update order: agents are selected one at a time (asynchronous random sequential update). An agent moves, the grid updates, and the next dissatisfied agent is selected. Synchronous update produces qualitatively similar segregation but different transient dynamics.

The simulation terminates when no agent is dissatisfied (equilibrium) or when no acceptable empty cell exists for any remaining dissatisfied agent.

Tunable parameters:

Tolerance threshold f: the minimum same-type fraction for satisfaction. At f = 1/3, an agent accepts being outnumbered 2:1 --- a preference that would be described as tolerant by any ordinary standard.
Vacancy rate: the fraction of empty cells. Higher vacancy enables faster sorting; lower vacancy constrains movement and slows convergence.
Agent density ratio: the relative proportion of the two types. Asymmetric populations produce asymmetric equilibria.

Emergent Behavior

Macro segregation from mild micro preference (observed). At f = 1/3, simulations consistently produce neighborhoods where 70 to 90 percent of each agent’s neighbors are its own type. The aggregate outcome is three times more extreme than the individual preference. This result has been replicated across thousands of simulation studies since 1971.

A sharp phase transition (proven). Jun Zhang (2004) proved rigorously that below a critical threshold f_c, stable integration is an equilibrium; above f_c, segregation is the unique stable outcome. The relationship between f and segregation is discontinuous --- a small increase in f near the critical value produces a qualitative jump in outcome. This is a bifurcation, mathematically analogous to phase transitions in physical systems.

Path dependence (observed). Two runs with identical parameters but different random initial distributions produce different final configurations. The system locks into whichever cluster structure the initial fluctuations favor. The current pattern reflects not only current preferences but the historical path.

The Mechanism

The mechanism is boundary erosion through positive feedback.

Consider a neighborhood that is 40 percent type A and 60 percent type B, with f = 0.4 for type A. Type A agents at the cluster interior are surrounded by other A agents and are satisfied. Type A agents at the boundary are exposed to more B neighbors. When a boundary A agent falls below threshold and moves, it leaves the boundary more B-dominated. The A agents just behind the former boundary are now exposed to a worse composition. Some of them become dissatisfied and move.

Each departure worsens conditions for the remaining agents of the departing type. The disturbance propagates inward from the boundary. The feedback is positive and self-amplifying: small initial fluctuations in local composition push boundary agents below threshold, their departure amplifies the imbalance, and the cascade continues until the cluster is consumed or the boundary stabilizes at a sharp interface between homogeneous regions.

The system overshoots. The macro outcome is more segregated than any individual agent wanted. This is not a failure of the agents to express their preferences --- it is the dynamical consequence of the interaction between those preferences and the spatial structure.

Transferable Principle

When agents with mild local composition preferences can relocate, the positive feedback of boundary erosion drives the system to an equilibrium far more sorted than any individual prefers, regardless of whether the agents are households, social media users, or political partisans.

Formal Properties

Proven:

Zhang (2004) established the existence of a critical threshold f_c below which integration is stable and above which segregation is the unique equilibrium, for a continuous-space version of the model.
Pancs and Vriend (2007) proved that for f = 1/2 on a one-dimensional ring, the only stable states are fully segregated, even though agents would prefer integration.
Brandt et al. (2012) proved convergence guarantees for specific grid topologies and update rules.

Observed / conjectured:

The critical threshold f_c for the standard 2D grid with Moore neighborhoods falls near 0.30 to 0.35 across simulation studies. No exact analytical value exists for the discrete grid model.
Segregation levels (measured by isolation index) at f = 1/3 consistently reach 0.70 to 0.90 in simulations, regardless of grid size above approximately 30 x 30.
On scale-free networks, segregation is amplified: high-degree nodes act as anchors that accelerate sorting (Fagiolo, Valente, Vriend, 2007).

Cross-Domain Analogues

Online filter bubbles. Users are agents; the “neighborhood” is the content feed; the tolerance threshold is the fraction of ideologically congruent content that sustains engagement. Platform algorithms reinforce engagement, accelerating the sorting. Transfer is structural: the feedback loop is qualitatively the same (composition-dependent departure amplifies homogeneity), but movement is costless online, neighborhoods are algorithmically constructed rather than spatial, and users occupy multiple communities simultaneously --- all of which change the dynamics quantitatively. The analogy breaks if users do not preferentially disengage from diverse feeds, which some empirical studies contest.

Political geographic sorting. Households are agents; counties or neighborhoods are cells; partisan identity is the type. Bishop’s The Big Sort (2008) documented increasing geographic concentration of partisan voting in the U.S. Transfer is structural: the mechanism of preference-driven migration producing more-sorted-than-preferred outcomes is the same, but income sorting, housing costs, and employment geography independently drive location choice, confounding the identification of Schelling dynamics specifically. The analogy would be falsified if geographic sorting were explained entirely by economic factors with no residual preference effect.

Workplace composition. Employees are agents; departments or teams are neighborhoods; demographic categories are types. Mild representation thresholds (wanting at least some colleagues of the same group) can produce homogeneous departments from heterogeneous labor pools through differential attrition. Transfer is structural. The analogy breaks when organizational hierarchy constrains mobility --- employees cannot freely relocate to another team the way a household can move to another neighborhood.

Academic discipline clustering. Researchers are agents; citation networks or department affiliations are neighborhoods; methodological or topical orientation is the type. Mild preferences for co-citation with similar work produce methodological silos. Transfer is structural. Breaks when disciplinary boundaries are enforced institutionally rather than emerging from preference-driven sorting.

Limits

Scope conditions. The model requires agents with the ability to relocate. When mobility is constrained --- by poverty, immigration status, lease contracts, employer location --- the mechanism operates more slowly or not at all. The model assumes a single dimension of type; real segregation involves the intersection of race, income, language, and religion, which the two-type model does not capture.

Known failures. The model predicts that reducing preference intensity (lowering f) should reduce segregation. Empirically, U.S. residential segregation declined only modestly between 1970 and 2020 despite documented increases in racial tolerance, suggesting that Schelling dynamics alone are insufficient to explain the persistence of segregation. Institutional mechanisms --- zoning, lending discrimination, school district boundaries --- sustain segregation independent of preference dynamics.

Common misapplications. Using the model to argue that segregation is “just preferences” and therefore not a policy problem. The model demonstrates that mild preferences are sufficient for segregation; it does not demonstrate that they are the operative cause in any specific real case. Inferring individual preferences from aggregate patterns is exactly the ecological inference the model warns against.

Connections

Methods: Agent-Based Modeling --- Schelling’s model is the historical prototype of spatial agent-based modeling, demonstrating how simple local rules produce macro patterns inaccessible to equation-based analysis.

Critiques: The Limits of Simple Models --- The strongest objection is that two-type, single-threshold models omit the institutional and economic forces that dominate real segregation. The best response is that the model identifies a mechanism, not a complete explanation; the mechanism’s contribution must be estimated empirically, not assumed.

Related Models: Ising Model --- Both exhibit phase transitions driven by local interaction preferences; the Schelling tolerance threshold is analogous to the Ising coupling constant, and the vacancy rate plays a role similar to temperature. Conway’s Game of Life --- Both are spatial models where local rules produce macro patterns through boundary dynamics, though Life is deterministic and Schelling is stochastic.

References

Schelling, T. C., “Dynamic Models of Segregation,” Journal of Mathematical Sociology (1971). The foundational paper defining the model and demonstrating the preference-segregation amplification.
Zhang, J., “Residential Segregation in an All-Integrationist World,” Journal of Urban Economics (2004). The first rigorous proof of tipping-point dynamics and critical threshold existence.
Card, D., Mas, A., and Rothstein, J., “Tipping and the Dynamics of Segregation,” Quarterly Journal of Economics (2008). The most cited empirical test of Schelling tipping dynamics using U.S. Census data.
Bruch, E. E. and Mare, R. D., “Neighborhood Choice and Neighborhood Change,” American Journal of Sociology (2006). Tests the model against residential mobility data from Los Angeles County.
Pancs, R. and Vriend, N. J., “Schelling’s Spatial Proximity Model of Segregation Revisited,” Journal of Public Economics (2007). Proves that mild preferences produce complete segregation on the 1D ring.