The Composition

Bradley Efron constructed a set of dice in the early 1970s with a property that should not exist. Die A beats die B more than half the time. Die B beats die C more than half the time. Die C beats die A more than half the time. Each comparison is straightforward — roll both, higher number wins, count the outcomes. No individual result is wrong. The dominance relation simply refuses to form a hierarchy. It cycles.

The instinct is to look for a trick. There is no trick. The numbers are chosen so that each die exploits a different vulnerability in its opponent. A has high values that overwhelm B's middle values. B has consistent values that outlast C's variable ones. C has a few very high values that catch A's gaps. Each pairwise contest is a fair measurement of a real advantage. The impossibility is not in any pair. It is in the assumption that pairwise superiority composes into a total ranking.

Kenneth Arrow proved the general case in 1951. Given three or more alternatives, no ranked-choice voting system can simultaneously satisfy four conditions: every possible ordering of preferences is allowed, if every voter prefers A to B then the group prefers A to B, adding or removing a third option does not change the relative ranking of the first two, and no single voter dictates the outcome. Each axiom is individually unobjectionable. Arrow proved their conjunction is impossible. The following year, Kenneth May showed that with exactly two alternatives, simple majority rule uniquely satisfies anonymity, neutrality, and positive responsiveness. The system works perfectly. Add a third option and it breaks. The failure is not gradual. It is a phase transition at three.

Edward Simpson described the statistical version in 1951 — the same year as Arrow, though Udny Yule had noticed the phenomenon in 1903. A trend present in every subgroup reverses when the subgroups are combined. The canonical example is UC Berkeley's 1973 graduate admissions: each department admitted women at equal or higher rates than men, but the aggregate data showed apparent bias against women. The resolution was that women applied disproportionately to more competitive departments. Every department was fair. The university looked unfair. Both statements were correct. The contradiction was in the assumption that the property of the parts — equal treatment — would persist in their sum.

Dietrich Braess found the network version in 1968. In a traffic network where each driver selects the fastest route, adding a new road can increase everyone's travel time. Each driver still optimizes individually. The new road creates a shortcut that attracts traffic away from two longer but previously adequate routes, overloading a shared segment. In 2003, Seoul demolished the Cheonggyecheon Expressway — a six-lane elevated highway through the city center — and replaced it with a public park. Traffic improved. The road had been the problem, not the solution. Braess's paradox has since been demonstrated in electrical networks, spring-mass systems, and basketball strategy. The mechanism is always the same: individually rational choices, aggregated through shared resources, produce collectively irrational outcomes.

Roger Shepard synthesized the perceptual version in 1964. Layer sine waves at octave intervals — each frequency exactly double the one below — with a bell-curve amplitude envelope fixed in spectral space. Now shift all tones upward. Each tone rises until it fades at the top of the envelope, while a new tone enters at the bottom. No individual component ascends forever. But the percept is of endless, continuous ascent. Each local moment — this tone is higher than the last — is true. The global conclusion — the pitch is always rising — is impossible. Jean-Claude Risset extended this to continuous glissandos in 1968, creating sounds that appear to descend perpetually. The ear confirms what the mathematics forbids.

Five domains — probability, social choice, statistics, network theory, psychoacoustics. Each demonstrates the same structural principle: properties that hold locally do not necessarily compose into global properties. Every die wins its matchup. Every axiom is reasonable. Every department is fair. Every driver chooses optimally. Every moment ascends. The whole cycles, contradicts, congests, or loops.

The assumption that breaks is so deep it barely registers as an assumption. If each part has a property, the whole has that property. Transitivity. Additivity. Composition. These are not laws of nature. They are habits of inference that happen to hold in the domains where we first learned to reason — counting, measuring, stacking. They fail precisely where interaction between parts creates structure that no part contains. The die's strength against one opponent is not the same property as its strength against another. The department's admission rate is not independent of who applies. The road's capacity is not independent of which drivers use it. The tone's direction is not independent of its spectral context.

May's theorem draws the boundary precisely. With two options, everything composes. Majority rule is fair, strategy-proof, and unique. At three, the space of possible relationships exceeds what any linear ordering can represent. The additional option does not merely add complexity. It creates structure — cycles, confounds, strategic possibilities — that did not exist in the pairwise world. The third option is not a complication. It is a phase transition.