The Accommodation

In 1931, Harry Beck was a twenty-nine-year-old engineering draftsman, temporarily laid off from the London Underground Signals Office, when he sketched a new kind of map. The existing Underground map was geographically accurate. It showed the lines as they actually ran beneath London — curving, converging, diverging according to the real distances between stations. It was correct, and riders found it nearly useless. In the center, where most journeys began and ended, stations were packed so tightly that names overlapped. In the outer zones, lines stretched across empty space that no one needed to navigate. The map faithfully represented the territory. The territory was not what riders needed.

Beck straightened the lines. He regularized them to verticals, horizontals, and forty-five-degree diagonals. He expanded central London and compressed the periphery. He eliminated the Thames except as a single stylized curve for orientation. The distances between stations on Beck's map bore no relationship to the distances between stations in London. Oxford Circus to Bond Street — a three-minute walk — occupied the same spacing as Epping to Theydon Bois, which is two miles. The map was, by any cartographic standard, wrong.

London Transport rejected it. Beck persisted. In 1933, a trial printing of 500 pocket maps sold out immediately. Within a year, the design was standard. Within decades, every transit system in the world had adopted the same principle: regularize geometry, equalize spacing, prioritize connections over distances. The template for the Tokyo Metro, the New York subway, the Paris Métro, and the Moscow Metro all descend from a diagram drawn by an unemployed draftsman who noticed that riders do not think in geography.

In 2008, Janet Vertesi at Cornell measured the cognitive effect. Regular Tube users systematically underestimate walking distances between nearby central stations and overestimate distances between outer stations. The map has rewritten their spatial model of London. The distortion is not merely present in the representation — it has migrated into the minds of the people who use it. Beck's map does not depict London. It depicts how people who ride the Underground need London to be organized: topologically, not metrically. The distortion is a portrait of the rider.

In the mid-1980s, Susan Brennan at MIT built a program that generated caricatures from photographs. The system computed the average of a set of faces, measured how each individual face deviated from that average, and then exaggerated the deviations. A slightly large nose became larger. A narrow jaw became narrower. The features that were already close to average remained unchanged.

The result was more recognizable than the photograph it was derived from.

Gillian Rhodes confirmed this experimentally in 1987. Subjects shown caricatures alongside veridical line drawings identified the caricatured faces faster and more accurately. The exaggeration did not add noise — it amplified signal. The brain does not encode faces as coordinates. It encodes them as deviations from a stored prototype. A photograph contains the deviation plus a large quantity of information the recognition system does not use — texture, lighting, absolute proportions. The caricature strips the noise and amplifies the signal. It is a worse picture and a better representation, because it is distorted in the direction the visual system already processes.

This is why skilled caricaturists are recognized as having captured something about a person that a camera cannot. The camera preserves everything. The caricaturist preserves only what differs. The art is in knowing which deviations to exaggerate, and that knowledge — usually implicit, acquired through thousands of drawings — is a model of the viewer's face-recognition architecture. The caricaturist draws the person as the brain sees them, not as the retina receives them.

In 1976, George Box wrote a sentence that became famous in statistics and notorious in the philosophy of science: "All models are wrong, but some are useful." The sentence is usually quoted as an epistemological concession — a shrug toward the inevitable gap between model and reality. But Box's point was sharper than the quotation suggests.

The ideal gas law, PV = nRT, describes a gas composed of point particles with no volume and no intermolecular forces. No such gas exists. Every real gas deviates from this model, and the deviations are well-characterized: van der Waals forces, molecular volume, quantum effects at low temperatures. The ideal gas law is not an approximation that happens to be slightly wrong. It is a deliberate simplification that is wrong in a specific way — it omits exactly the variables that matter least in the operating range where engineers need predictions. At moderate temperatures and pressures, the omitted variables contribute corrections on the order of a few percent. The model is not close to right. It is wrong in the right direction.

The Bohr model of the atom is wrong in a different right direction. It depicts electrons orbiting a nucleus in discrete shells, like planets around a star. No electron orbits. Quantum mechanics replaced this picture within fifteen years of its publication. But the Bohr model predicts the hydrogen emission spectrum to four significant figures, and it does so by being wrong in a way that preserves the quantized energy levels while discarding the wave-particle duality that makes the full theory intractable for quick calculation. It is the caricature of the atom — exaggerating the discrete structure, suppressing the continuous uncertainty. The distortion is tuned to what the user needs to calculate.

In the late 1990s, Gerd Gigerenzer ran a series of prediction tournaments pitting complex statistical models — multiple regression, neural networks, Bayesian classifiers — against simple heuristics that used almost no information.

The simplest was "Take the Best." Given a prediction task (which of two cities is larger? which patient is at higher risk?), the heuristic searches through cues in order of their known validity. As soon as it finds a cue that discriminates between the two options, it stops searching and decides. It ignores all remaining cues. No weighting. No integration. No optimization.

In low-information, high-noise environments — which describes most real-world prediction tasks — Take the Best matched or outperformed the complex models. The result surprised even sympathetic researchers. The mechanism was not that the heuristic was secretly optimal. It was that the complex models, by estimating more parameters, were fitting noise as well as signal. In small or noisy samples, each additional parameter is an opportunity to overfit. The simple heuristic, by ignoring most of the information, was robust to the noise that the complex model tried to model.

Gigerenzer's term for this is the "less-is-more effect." In environments where the signal-to-noise ratio is low, the cost of processing additional information exceeds the benefit. The heuristic's ignorance — its systematic refusal to use available data — is not a limitation. It is the mechanism. The distortion is tuned to the environment's noise structure.

The thread connecting Beck's subway map, Brennan's caricature, Box's models, and Gigerenzer's heuristics is not that distortion is sometimes acceptable. It is that the direction of distortion carries information the accurate version does not.

A perfectly accurate map of the London Underground — showing every curve, every distance, every relationship to the surface geography — tells you everything about London and nothing about how people navigate it. Beck's distortion tells you: riders think in connections, not distances. The direction of simplification is a record of what riders need.

A photograph tells you everything about a face and nothing about how faces are recognized. The caricature tells you: the brain encodes deviation from prototype. The direction of exaggeration is a record of the visual system's encoding scheme.

A complete physical model tells you everything about a system and nothing about what questions you can tractably answer. The ideal gas law tells you: at these temperatures and pressures, intermolecular forces are negligible for engineering purposes. The direction of omission is a record of the calculation's operating range.

A fully parameterized statistical model tells you everything the training data contained and nothing about what will generalize. The simple heuristic tells you: in this noise regime, most cues are unreliable. The direction of exclusion is a record of the environment's information structure.

The accurate representation carries no information about the user because it does not need one. It is complete. It is also, in each of these cases, less useful than the version that has been reshaped to fit the constraints of the thing it serves. The distortion is not a concession. It is a second channel of information — a portrait of the relationship between the representation and its purpose, encoded in the specific ways the representation departs from the truth.

The eye accommodates. The lens changes shape to bring objects at different distances into focus. It does not reproduce the light field faithfully — it bends it, compresses it, inverts it, projects it onto a surface that can only resolve a small fraction of the incoming information. Every stage is a distortion. Every distortion is tuned to the retina. What arrives at the photoreceptors is not the world. It is the world, accommodated.