Emergent Behavior: When Simple Rules Make Complex Societies
In 1972, Philip W. Anderson published a two-page essay in Science (volume 177, issue 4047, August 4, 1972) with the title “More is Different.” Anderson was a condensed matter physicist who had just won the Nobel Prize in Physics (shared in 1977). He was not writing about sociology, or cities, or cellular automata. He was writing about superfluidity and symmetry breaking. But the argument he made applies to all of them, and it is the argument that underlies everything this cluster is about.
Anderson’s claim was an attack on reductionism — not the scientific method, but the philosophical assumption that the behavior of complex systems can be understood as a simple extension of the behavior of their components. “The ability to reduce everything to simple fundamental laws,” he wrote, “does not imply the ability to start from those laws and reconstruct the universe.” A collection of water molecules does not “contain,” in any computable sense, the phenomenon of superfluidity. A collection of neurons does not “contain” the phenomenon of consciousness. A collection of individuals making locally rational residential choices does not “contain” the pattern of residential segregation.
Each level of complexity, Anderson argued, has its own phenomena, its own regularities, its own concepts. These phenomena are not predicted by the laws of the level below; they are genuinely new. He called this emergence.
What Emergence Actually Means
“Emergence” is one of the most overused words in popular science writing. It often appears as a kind of intellectual hand-wave: complexity emerges from simplicity, consciousness emerges from neurons, cities emerge from individuals. The word explains nothing; it labels a phenomenon and moves on.
Anderson was more precise. John H. Holland, in his 1998 book Emergence: From Chaos to Order (Addison-Wesley), was more precise still. Holland defined emergence as the appearance of system-level properties that are not present in, and cannot be straightforwardly predicted from, the properties of the system’s components. The emphasis on predicted is important: it is not enough to say that a property was not anticipated. Emergence requires that the property cannot be derived from the component-level description by any finite computational procedure within a relevant timescale.
In this sense, Conway’s Life is the canonical model of emergence. The component-level description is simple: a binary grid, four rules. The system-level properties — the glider, the Gosper glider gun, the Turing-complete universal constructor — are not derivable from the component description without running the simulation. You cannot look at the rules and predict the glider. You have to watch what happens.
Schelling’s segregation model is emergence in this strong sense. The component-level description: agents with a mild preference to not be entirely isolated from their own type. The system-level property: large-scale, stable residential segregation that exceeds what any individual preferred. You cannot predict the segregation from the preference alone. You have to run the model.
Schelling’s Model in Precise Terms
Thomas Schelling’s “Dynamic Models of Segregation,” published in the Journal of Mathematical Sociology in 1971 (volume 1, pages 143–186), is worth reading in its original form. The paper is remarkably careful about what it is and is not claiming.
The model is this: a grid (or a line) of positions, each occupied by one agent of type A or type B, or empty. Each agent has a threshold — a minimum fraction of same-type neighbors it requires to be satisfied. An agent whose actual fraction of same-type neighbors falls below its threshold is dissatisfied and will move to a randomly chosen empty cell. The dynamics: repeatedly select dissatisfied agents at random; move them to the nearest satisfying empty location. Stop when all agents are satisfied, or the system has reached a stable state.
The initial configuration matters: Schelling considered both random initial distributions and structured ones. In both cases, with preference thresholds set as low as one-third (agents satisfied in a neighborhood where 2/3 of neighbors are different type), the equilibrium state was dramatically segregated — typically more than 70 to 80 percent of each agent’s neighbors were their own type, far exceeding the one-third preference.
Schelling used real coins on a real checkerboard. He reported his results qualitatively (this was 1971, before desktop computers made extensive numerical simulation feasible). The quantitative relationship between threshold and equilibrium segregation was worked out in subsequent decades of computational simulation. The broad result is robust: there is a phase transition in the system around a threshold of roughly 30 percent. Below this, the system typically stays integrated. Above this, it converges to high segregation. The transition is not gradual — it is abrupt, characteristic of a phase transition.
This is the same phase transition that appears in the Ising model of ferromagnetism (where the transition is between disordered and ordered magnetic states), in directed percolation (between absorbing and active phases), and in the spread of epidemics (between sub-threshold decay and epidemic growth). The Schelling model is, mathematically, a member of the same universality class as these physical systems. The sociological phenomenon and the physical phenomenon are the same phenomenon.
The Surprising Results and What They Tell Us
The core finding — that mild preferences produce extreme outcomes — is sometimes described as Schelling’s “tipping point” result, a phrase Malcolm Gladwell later borrowed for a different purpose. But Schelling’s result is sharper than the tipping point metaphor suggests.
The result is not merely that preferences above some threshold produce segregation. It is that the dynamics of local interaction produce outcomes that systematically exceed individual preferences. An agent with a one-third same-type preference does not want to live in a 90-percent same-type neighborhood — it has no preference about that. But that is where it ends up, because the dynamics of a dissatisfied neighbor’s move shifts the composition of its neighborhood, potentially making it dissatisfied, causing it to move, shifting compositions elsewhere in a cascading process that continues until the system freezes in a maximally segregated state.
The gap between preference and outcome is the fundamental fact. This gap appears in every CA model of social dynamics, and it is the reason these models are illuminating rather than merely formal exercises. The gap tells you that individual-level interventions — changing preferences — may not produce the expected system-level changes. If the preference threshold is above the critical value, even shifting individual preferences downward will not change the equilibrium state until the threshold moves below the critical value. Policy interventions need to target the dynamics, not just the inputs.
Holland’s formulation helps here: emergence is precisely the existence of this gap. An emergent property is one that cannot be reduced to the components’ properties because the system-level dynamics create phenomena that are absent at the component level. Schelling segregation is emergent because segregation is a property of the spatial pattern across the whole grid, not of any individual agent’s preference or location.
Extensions and Variations
The three decades since Schelling’s paper have produced extensive computational investigation of how the model’s results vary with its assumptions.
Threshold distribution. Schelling assumed uniform thresholds. When thresholds vary across agents — some highly tolerant, some highly intolerant — the dynamics change. A small number of highly intolerant agents (say, 10 percent of the population with thresholds above 50 percent) can drive the system to high segregation even when the remaining 90 percent would be perfectly satisfied in an integrated neighborhood. The intolerant minority forces the majority into a segregated equilibrium neither wanted.
Neighborhood structure. The standard model uses a Moore neighborhood (8 nearest cells on a square grid). Larger neighborhoods reduce segregation at equilibrium, because agents can be satisfied in globally heterogeneous environments if their local neighborhood is sufficiently large. Different network topologies produce qualitatively different dynamics.
Multiple types. Extending the model to three or more types increases complexity dramatically. The three-type model can produce rotating spatial patterns — a type-A region pushing against a type-B region, which pushes against type-C, which pushes against type-A — that have no two-type analogue. It also produces more varied equilibrium configurations, from nearly complete segregation to partial integration, depending on the balance of preferences.
Utility functions. Schelling’s binary satisfaction model (satisfied or not) can be replaced with a continuous utility function. When agents prefer a neighborhood with some specific fraction of same-type (say, 50 percent, neither minority nor majority), the model produces integration. When the optimal fraction is high (above 50 percent), it produces segregation. The richness of the utility function dramatically affects the outcome.
These extensions have not overturned Schelling’s core finding. They have established its robustness and specified its limits: the result holds across a wide range of parameter settings and model variations. What changes with the extensions is the quantitative relationship between preference and segregation, and the range of equilibrium configurations. The qualitative phenomenon — local rules producing global structure that exceeds individual intentions — persists throughout.
The Relationship to Conway’s Life
The formal parallel between Schelling’s model and Conway’s Life is straightforward: both are CA on a two-dimensional grid, both use a neighborhood-based update rule, both produce complex global patterns from simple local rules. The parallel is worth dwelling on for a moment, because it illuminates something about Life that is easy to overlook.
Conway’s Life has no sociology — the cells have no intentions, no preferences, no social structure. Schelling’s model has sociology but simple dynamics — the update rule is less rich than Life’s. And yet the dynamics that matter most — the emergence of macroscopic pattern from microscopic interaction, the sensitivity to initial conditions, the irreversibility of certain state transitions, the existence of phase transitions in parameter space — appear in both systems.
This convergence suggests something about emergence in general: these phenomena are not specific to biology or sociology or physics. They are features of a class of dynamical systems — any system with local interactions, parallel update, and a sufficiently complex rule space. Conway discovered this in 1970 by designing a toy universe. Schelling discovered it in 1971 by modeling residential behavior. Anderson articulated it in 1972 by analyzing superconductors. They were all describing the same thing.
The Life universe and the social world are both instances of what happens when local rules generate global patterns. The difference is that the Life universe has no human cost, and the social world does. Schelling’s segregation is not just mathematically interesting. The mechanism he identified — local preference thresholds producing citywide segregation — describes, in stylized form, the spatial structure of American cities that has produced and sustained enormous inequality. Understanding the mechanism does not make the inequality disappear. But it is a prerequisite for knowing what kinds of interventions could.
The Limits of the Metaphor
Schelling’s model is powerful and has been enormously influential. It is also, in obvious ways, a simplification of the real world, and honest engagement with it requires naming what it leaves out.
It treats preference and behavior as identical. People may prefer integrated neighborhoods but be constrained from living in them by housing prices, discriminatory lending practices, or landlord behavior. The constraint is not modeled.
It treats households as atomic agents. Real households have internal dynamics, negotiate preferences, and respond to economic incentives that are absent from the model.
It ignores path dependence. The initial conditions of American residential segregation were established by legally enforced discrimination — racially restrictive covenants, discriminatory mortgage guarantees, explicit redlining. These created the initial configuration from which the Schelling dynamics then operated. A model that starts from a random initial configuration misses this historical dimension entirely.
It has no market. Housing prices, neighborhood amenities, and economic sorting all operate in addition to (and often independently of) racial or ethnic preference. Economic segregation and racial segregation are deeply entangled and Schelling’s model cannot disentangle them.
Schelling was clear about these limits. He was not claiming to explain American residential segregation. He was claiming to have identified one mechanism — the amplification of individual preferences by local dynamics — that is necessary for a complete explanation. The CA model does not replace the historical and sociological analysis. It clarifies one component of a complex causal story.
This is the appropriate use of CA in social science: not as a complete explanation, but as a formal tool for isolating specific mechanisms, testing their sufficiency, and identifying their parameters. When the mechanism is confirmed to operate and the parameters are identified, the model becomes a tool for intervention design. When the model fails to match reality, the gap reveals what additional mechanisms are needed. In either case, the formalism is productive.