A checkerboard proof that mild preferences can grow an extreme pattern
In the United States of the 1960s residential segregation was an obvious fact, and the default frame for explaining it had exactly one shape: the degree of segregation reflects the degree of prejudice. A pattern this extreme must sit on top of correspondingly strong aversion. The policy implication followed automatically — to undo segregation, first change what people feel.
Schelling asked a purely structural question instead: what is the weakest preference that can still produce a map like this? Note that he is not asking how strong actual preferences are; he is asking for a lower bound. And lower-bound questions have a way of demolishing an entire default picture of causation.
Notice that rule 4 does not say which empty cell. A random one? The nearest one? The best one? Schelling's own version moved agents to the nearest satisfying spot; later reimplementations each chose their own. This apparently trivial detail changes how strong the result is, and under some settings changes the result itself — so any Schelling-style claim should be quoted together with its move rule.
The threshold T is also routinely read backwards. T = 30% is not the strong demand "I insist that three in ten around me are my own kind". It is the rather relaxed tolerance "seven in ten around me are unlike me and that is fine". The entire force of the model lies in that direction.
The start is scattered at random, so the average share of like neighbours is about 50% — which is what "fully mixed" means. Once it runs, unhappy agents begin to move, and every move lowers the like-share of the like neighbours left behind, manufacturing fresh unhappiness. Waves of relocation cross the board.
After somewhere between ten and a few dozen rounds the movement stops. At that point nobody is unhappy, yet the average share of like neighbours is far above where it started — on a 16×16 board with a 30% threshold and about 10% vacancy, it typically lands between 0.65 and 0.85.
Two further properties only show up once it runs. The process ratchets: movers land only where they are satisfied, that is, where more of their own kind already are, so homogeneity on average only rises. And the endpoint locks itself: in the frozen state everyone is satisfied, so no one has any reason to move back alone. The pattern is stable — but "stable" here means only "nobody can do better by acting alone". It is neither optimal nor chosen.
What it delivers is a proof of existence: producing extreme segregation does not require extreme preferences. That refutes a claim of necessity — "segregation this severe must rest on aversion equally severe". Once the lower bound falls, the inference from pattern back to sentiment is broken.
The same structure runs elsewhere: office lunch tables hardening, group chats converging on one topic, academic collaboration networks splitting into blocks, interest communities stratifying on a platform. The three shared preconditions are: agents see only locally, the response is threshold-shaped, and an unhappy agent can move. When all three hold, expect blocks — and do not expect the participants' own accounts to contain the reason.
It also has an under-used constructive side: it names the intervention points. If the pattern is set by thresholds and by the move rule, then the available levers are not exhortations against prejudice but the cost of moving, the distribution of available positions, and who is adjacent to whom. That is why changing the structure of encounter — seating, mixed grouping, shared facilities — so often outperforms appeals.