Schelling Model Overview

• Schelling Model is an agent-based model that explains how individual relocation decisions based on neighborhood composition lead to urban segregation.
• It encompasses classical, heterogeneous, game-theoretic, and continuum formulations to capture a range of dynamics from integrated to segregated states.
• Quantitative analyses reveal tipping points, phase transitions, and the significant influence of intervention strategies on residential patterns.

The Schelling model is a simple agent based model that demonstrates how individuals' relocation decisions generate residential segregation in cities. In its canonical form, agents belong to one of two groups and occupy cells of rectangular space; each agent reacts to the fraction of agents of their own group within the neighborhood around their cell, stays put when this fraction is above a given tolerance threshold, and seeks a new location if the fraction is below the threshold (Hatna et al., 2014). A unified mathematical framework represents the model on a graph $G=(V,E)$ with site states $\sigma_i\in\{1,-1,0\}$, allowing the classical grid model and many later variants to be treated as Markov chains defined by a network, an initial condition, a satisfaction function, and transfer probabilities (Rogers et al., 2011).

1. Classical formulation and local decision rules

In the standard two-dimensional formulation, space is a rectangular lattice of cells, each cell is either vacant or occupied by a single agent, and agents belong to one of two “color” groups. A common neighborhood specification is the Moore neighborhood. In the heterogeneous-population formulation, the neighborhood of a cell $c$ is the $n\times n$ block around $c$, excluding $c$ itself; a typical choice is a $5\times 5$ Moore neighborhood, giving up to 24 neighbors (Hatna et al., 2014). If $N_a^{\mathrm{tot}}(c)$ is the number of occupied neighbors around $c$ and $N_a^{\mathrm{friends}}(c)$ is the number of same-color neighbors, the fraction of friends is

$\sigma_i\in\{1,-1,0\}$0

Vacant cells are ignored in this fraction (Hatna et al., 2014).

The homogeneous model assigns a common tolerance threshold $\sigma_i\in\{1,-1,0\}$1, so that an agent is satisfied when $\sigma_i\in\{1,-1,0\}$2 and dissatisfied otherwise. A generalized formulation assigns each agent a personal tolerance threshold $\sigma_i\in\{1,-1,0\}$3 and evaluates a location through the utility

$\sigma_i\in\{1,-1,0\}$4

with the convention that the utility of a cell whose neighborhood is entirely empty is 0 (Hatna et al., 2014). This utility preserves the classical logic—dissatisfied agents relocate—but makes heterogeneity in tolerance explicit.

A more general mathematical formulation encodes occupancy and type through

$\sigma_i\in\{1,-1,0\}$5

and treats the dynamics as a Markov chain on configurations $\sigma_i\in\{1,-1,0\}$6 (Rogers et al., 2011). Within that framework, classical nearest-neighbor threshold models, best-response variants, and stochastic relocation rules differ mainly in the choice of satisfaction function and transfer probabilities.

2. Thresholds, phases, and segregation metrics

The model is well known for tipping point behavior. In the classical narrative, an initial random pattern remains integrated when the tolerance threshold is below $\sigma_i\in\{1,-1,0\}$7 but becomes segregated when the tolerance threshold is above $\sigma_i\in\{1,-1,0\}$8 (Hatna et al., 2014). In one specific implementation—$\sigma_i\in\{1,-1,0\}$9 torus, density $c$0, $c$1 Moore neighborhood, equal numbers of blue and green agents, and random relocation rate $c$2—the effective tipping point is $c$3. For $c$4 the system remains integrated; for $c$5 it converges to segregated patterns with two large patches and a minimal boundary. The transient typically proceeds through formation of small homogeneous clusters, coarsening, and then a final state with two large homogeneous regions; the coarsening process is slower near the tipping point and much faster for high tolerance demands (Hatna et al., 2014).

Different literatures quantify segregation with different order parameters. In heterogeneous-threshold models, Moran’s $c$6 is computed separately for color, $c$7, and for tolerance, $c$8, while the $c$9 index detects the simultaneous presence of a pure blue region, a pure green region, and an integrated region through

$n\times n$0

Integrated patterns have $n\times n$1 and $n\times n$2; fully segregated patterns have high $n\times n$3 and $n\times n$4; mixed patterns have intermediate $n\times n$5 and $n\times n$6 (Hatna et al., 2014). Other formulations use contact density $n\times n$7, adjusted perimeter per agent $n\times n$8, seclusiveness $n\times n$9, scale of aggregation $c$0, number of clusters $c$1, segregation coefficient $c$2, and interface density $c$3 to distinguish mixed, segregated, quasi-stationary, and fragmented states (Hazan et al., 2012, 0711.2212, Abella et al., 2022).

Quantitative scaling analyses show that aggregation depends strongly on city size, density, and threshold. In one systematic study, the “striking global aggregation” observed on Schelling’s $c$4 board is “strictly a small city phenomenon,” and the number of clusters obeys approximate power laws $c$5 with $c$6 for $c$7 and $c$8 with $c$9 for $c$0 (0711.2212). A common misconception is therefore that the original small-board phenomenology transfers unchanged to larger lattices.

3. Heterogeneity, mixed patterns, and alternative utility structures

Allowing agents to differ in tolerance produces a richer repertoire of steady states than the segregation–integration dichotomy. In a heterogeneous model with personal thresholds $c$1, the system generates stable configurations that contain segregated patches for each group alongside patches where both groups are spatially integrated. These mixed patterns arise robustly when one subgroup has tolerance $c$2 and another has tolerance $c$3, that is, when part of the population is tolerant enough to accept diversity while another part is demanding enough to self-segregate. With beta-binomial threshold distributions, U-shaped distributions with many agents at very low or very high tolerance produce especially pronounced mixed patterns (Hatna et al., 2014).

Another extension introduces a fraction $c$4 of spatially fixed switching agents. These agents never move but can change displayed type, and the same mechanism admits two interpretations stated explicitly in the model: random, non-preferential allocation of fixed sites in an open residential system, and superimposition of social and spatial mobility in a closed residential system. The presence of switching agents can lead to “intermediate patterns” such as “mixture of patches” and “fuzzy interfaces,” and increasing $c$5 expands the mixed phase while shrinking the segregated region (Hazan et al., 2012). A plausible implication is that type mobility and random allocation alter segregation even when movement rules for ordinary agents are unchanged.

A different Schelling-like formulation replaces same-type preference by density preference. In that model, each resident in block $c$6 has individual utility

$c$7

the global utility is $c$8, and relocation is evaluated through the mixed utility change

$c$9

The parameter $5\times 5$0 is the cooperation degree, ranging from laissez-faire at $5\times 5$1 to fully cooperative dynamics at $5\times 5$2. A dynamically varying cooperation degree,

$5\times 5$3

enlarges the high-utility region of the phase diagram and, at $5\times 5$4, reduces the critical average cooperation needed for $5\times 5$5 from about $5\times 5$6 to about $5\times 5$7, which the authors interpret as a 76% reduction (Okada et al., 2022).

A further dynamic extension treats intolerance itself as a state variable. In that model, each agent has $5\times 5$8, representing the desired percentage of neighbors from the same group, and contact plus media effects update intolerance according to local neighborhood composition and a conformity parameter. The reported result is that a society composed of individuals who do not easily conform to their surroundings, together with positive examples of both groups in media, promotes integration within society (Johnson et al., 2024).

4. Game-theoretic, dynamical-systems, and continuum formulations

A game-theoretic line of work turns Schelling segregation into a strategic location game on a graph. In the “Schelling Segregation with Strategic Agents” formulation, the local happiness ratio is

$5\times 5$9

and agents minimize a lexicographic cost vector whose first component is unhappiness and whose second component is distance to a favorite node. The model distinguishes Swap Schelling Games and Jump Schelling Games, proves that the uniform swap game with $N_a^{\mathrm{tot}}(c)$0 is a potential game for any window size $N_a^{\mathrm{tot}}(c)$1, and gives tight asymptotic bounds $N_a^{\mathrm{tot}}(c)$2 for some network $N_a^{\mathrm{tot}}(c)$3 (Chauhan et al., 2018). This replaces stochastic relocation with equilibrium analysis, potential functions, and efficiency bounds.

The bounded neighborhood model has also been recast as a continuous dynamical system. With densities $N_a^{\mathrm{tot}}(c)$4 and $N_a^{\mathrm{tot}}(c)$5, linear tolerance schedules produce the ODE system

$N_a^{\mathrm{tot}}(c)$6

In rescaled variables $N_a^{\mathrm{tot}}(c)$7 and $N_a^{\mathrm{tot}}(c)$8, stable integrated population mixes exist exactly in the parameter region $N_a^{\mathrm{tot}}(c)$9 and $c$0, and neighborhood tipping is explained in terms of basins of attraction rather than by the tolerance parabolas themselves (Haw et al., 2017). This supplies an explicit phase-plane description of integration, segregation, and tipping.

Rigorous continuum limits connect the model to PDEs and Gaussian fields. For a perturbed Schelling model with spontaneous Glauber and Kawasaki dynamics, the hydrodynamic limit is a reaction-diffusion equation with a discontinuous non-linear reaction term,

$c$1

and the resulting potential structure motivates a phase diagram with a mixed phase, a segregated phase, and a metastable segregation phase, together with an interpretation in terms of relevant and irrelevant disorder (Barret et al., 2023). In a separate scaling-limit analysis, the normalized bias field $c$2 of the Schelling model converges to an integro-differential equation with Gaussian initial data; almost sure existence and uniqueness are proved even though the initial conditions are described by white noise, and in the one-dimensional case the scaling limit of the limiting clusters at time infinity is characterized explicitly (Holden et al., 2018).

5. Temporal memory, locality, and universality classes

Temporal memory can qualitatively alter both the phase diagram and the kinetics. In a noisy constrained Schelling model with long-range satisfying moves, aging is introduced through an internal clock $c$3 and a movement probability for satisfied agents

$c$4

The longer an agent has been satisfied, the less likely it is to move. This removes the segregated–mixed transition of the original noisy constrained model, yields segregated states with high levels of agent satisfaction even for high values of the tolerance, and produces slow power-law coarsening $c$5 together with glassy-like dynamics in which asymptotic time translational invariance is broken (Abella et al., 2022).

Allowing the global tolerance threshold to vary in time creates another class of nonstationary dynamics. In a closed-city model with open boundaries and vacancy density $c$6, tolerance can decrease continuously according to

$c$7

so that the change rate $c$8 controls whether the final state contains many small clusters or large segregated clusters (Ortega et al., 2021). Under sudden drops in tolerance, vacancies aggregate on the interface between large red and blue clusters; for $c$9, a vacancy border tends to appear and stabilize, and the roughness of this frontier falls in the Edwards–Wilkinson universality class, with measured roughness exponents close to $N_a^{\mathrm{friends}}(c)$0 and Gaussian height fluctuations (Ortega et al., 2021).

Rigorous one-dimensional analyses show that locality and movement rules are decisive. In a noiseless circular model with minority proportion $N_a^{\mathrm{friends}}(c)$1, there is a phase transition at $N_a^{\mathrm{friends}}(c)$2: for $N_a^{\mathrm{friends}}(c)$3 the process converges to complete segregation, whereas for $N_a^{\mathrm{friends}}(c)$4 it is static with high probability (Barmpalias et al., 2015). On $N_a^{\mathrm{friends}}(c)$5, a swap-based Schelling model exhibits different asymptotic behavior depending on whether the moving distribution has bounded or unbounded support, and the behavior changes again when agents are lazy in the sense that they only swap location if this strictly improves their situation; bounded support implies fixation, while unbounded support yields local homogenization without almost sure convergence at a fixed site (Deijfen et al., 2019).

A metapopulation version further couples segregation to highly heterogeneous occupancies. In that model, low tolerance produces a long quasi-stationary transient during which the population remains in a well mixed phase, after which the system stabilizes on heterogeneous “urban skylines”; varying tolerance identifies three regimes: microscopic clusters with local coexistence, macroscopic clusters with local coexistence (“soft segregation”), and macroscopic clusters with local segregation but homogeneous densities (“hard segregation”) (Gargiulo et al., 2015).

6. Interpretive scope, interventions, and recurring misconceptions

The most common compressed reading of the Schelling model is that “mild local preferences” imply a universal march from integration to segregation. The record in later variants is more specific. One extension states explicitly that “the variety of the Schelling model steady patterns is richer than the segregation-integration dichotomy,” because heterogeneity in tolerance produces mixed patterns with segregated patches alongside integrated neighborhoods (Hatna et al., 2014). Another shows that “the striking global aggregation Schelling observed is strictly a small city phenomenon,” so extrapolation from small boards to large lattices is not automatic (0711.2212).

Intervention models also shift the classical interpretation. Random allocation and social mobility, represented by switching agents, can create intermediate patterns and enlarge the mixed phase (Hazan et al., 2012). In the density-based Schelling-like model, the cooperation degree can be interpreted as the magnitude with which Pigouvian tax is enforced, and a dynamic, state-contingent allocation of cooperation produces homogenized phases with globally high utility at much lower average cooperation than a static policy (Okada et al., 2022). Contact-based intolerance updating combined with media effects further suggests that integration depends not only on neighborhood composition, but also on how intolerance evolves in response to contact and information flows (Johnson et al., 2024).

Taken together, these results place the Schelling model less as a single threshold automaton than as a family of local-interaction systems spanning Markov chains, potential games, nonlocal ODEs, reaction-diffusion limits, and kinetic-roughening problems. Across that family, the recurring technical question is not whether segregation can occur, but under which assumptions on tolerance, mobility, heterogeneity, memory, and noise integration, segregation, mixed mosaics, metastability, or dynamic coexistence are the stable large-scale outcomes.