Offer someone a choice between two lotteries. An hour later, offer the identical pair again. In the classic experiments of Hey and Orme, roughly a quarter of these repeated choices reverse. The stakes, probabilities, and framing are unchanged, yet the answer flips. For a deterministic theory of preference this is an embarrassment. If preference is a complete, transitive ordering over options, the same question should always return the same answer.
Classical decision theory handled this awkwardly. Von Neumann and Morgenstern axiomatised preference as a fixed relation. Observed inconsistency was pushed into an error term added after the fact, treated as noise around the signal and of no theoretical interest. But the error term turned out to have structure. It varies predictably with the value difference between options, with time pressure, with the allocation of attention, and with the precision of the neural systems that encode value.
This raises a fundamental theoretical question: is preference a point or a distribution? If it is a distribution, variability is not a measurement defect but a constitutive feature of choice. Taking stochastic choice seriously requires three moves. The first is to formalise choice as a draw from latent utility distributions. The second is to decompose that randomness into its mechanistic sources. The third is to ask whether a well-designed decision system might actually want to be noisy.
Random Utility Models: Preference as a Distribution
The formal starting point is Thurstone's 1927 law of comparative judgment, later generalised by Marschak and McFadden. Each option i carries a utility Ui = Vi + εi. Here Vi is a systematic component determined by the option's attributes, and εi is a random disturbance. The decision-maker chooses whichever option has the highest realised utility. The probability of choosing i is therefore the probability that Ui exceeds every competitor, and it depends entirely on the joint distribution of the disturbances.
Distributional assumptions determine everything downstream. If the disturbances are independent and identically Gumbel-distributed, the integral collapses into the multinomial logit: P(i) = exp(Vi/λ) / Σj exp(Vj/λ). The scale parameter λ governs sensitivity to value differences. As λ approaches zero, choice converges on deterministic maximisation. As λ grows, choice approaches uniform randomness. This softmax form appears everywhere, from McFadden's transport econometrics to reinforcement learning models of dopaminergic prediction error.
The logit's tractability comes at a cost. It implies Luce's independence of irrelevant alternatives: the ratio P(i)/P(j) does not change when a third option is added. Debreu's red bus–blue bus counterexample exposes the problem. Suppose a blue bus identical to an existing red bus is introduced. The logit predicts that the new option draws share proportionally from cars and buses alike. In reality it should mostly take riders from the red bus. Real substitution patterns reflect similarity between options, and independent errors cannot represent similarity.
Multinomial probit uses multivariate normal errors with a full covariance matrix. It captures correlated substitution but gives up closed-form choice probabilities. Nested and mixed logit offer intermediate solutions. A distinct family, random preference models, treats the utility function itself as drawn anew from a set of deterministic orderings on each occasion. This is not a notational variant. As Loomes and Sugden showed, random preference models never violate first-order stochastic dominance, whereas Fechnerian strong-utility models predict that such violations will occur.
The deeper theoretical move came from Block and Marschak in 1960, with the characterisation completed by Falmagne in 1978. Their conditions specify what observable choice probabilities must satisfy to be rationalisable by any random utility model. This is a stochastic analogue of revealed preference. Rationality becomes a property of distributions rather than of individual decisions. A single inconsistent choice violates nothing. Only systematic violations of the probabilistic axioms count against the model.
TakeawayUnder random utility theory, rationality stops being a property of individual choices and becomes a constraint on the shape of choice distributions. A single reversal tells you almost nothing.
Sources of Variability: Encoding, Retrieval, and Execution
Random utility theory says nothing about where ε comes from. Neuroeconomics demands a mechanistic answer, and the evidence suggests several sources. At least three stages of processing contribute. The first is encoding, where option attributes are turned into neural representations. The second is retrieval, where value is constructed from memory and prior experience. The third is execution, where a decision variable is translated into a motor action. Each stage leaves a distinct behavioural signature.
Encoding noise arises because magnitudes are represented by finite populations of noisy neurons. Work by Khaw, Li, and Woodford shows that imprecise, roughly logarithmic encoding of payoffs can produce small-stakes risk aversion even when the utility function has no curvature at all. Polanía, Woodford, and Ruff extended the idea to subjective value. They found that the brain appears to allocate coding precision efficiently, matched to the distribution of values it expects to encounter. Some behaviour that looks like preference may partly be perception.
Retrieval variability reflects the fact that many values are not stored but constructed. Stewart, Chater, and Brown's decision by sampling holds that an attribute is evaluated by comparing it with samples drawn from memory. Shadlen and Shohamy make a related argument: value-based choice involves sequential retrieval of episodic evidence, accumulated much as perceptual evidence is accumulated. Because the retrieved sample differs from one occasion to the next, the constructed value differs too, even when the stimulus is identical.
Execution noise operates after the decision variable has been formed. It covers lapses, trembles, and motor slips. It is usually modelled as a mixture: with probability 1−ω the agent follows the softmax, and with probability ω it responds at random. The key difference lies in how errors relate to value. Execution noise ignores value, so it produces errors on easy choices as often as on hard ones. Encoding and retrieval noise concentrate errors near indifference. This dissociation makes the components partially identifiable from psychometric choice curves alone.
Drugowitsch, Wyart, and colleagues decomposed variability in a sequential inference task. They found that the dominant source was neither sensory noise nor response-selection noise. It was imprecision in the inference computation itself, arising at each update rather than at the input or output stages. This reframes the problem. Choice variability is not a single nuisance term but a fingerprint of computation, and its structure reveals which stage of processing is resource-limited.
TakeawayErrors concentrated near indifference point to imprecise valuation, while errors on easy choices point to execution. The pattern of mistakes tells you where the bottleneck lives.
Optimal Noise Levels: When Variability Pays
If noise were purely a cost, evolution and learning should drive it down. Yet the brain sustains substantial variability even in tasks where greater precision seems attainable. This invites a normative question: under what conditions does a stochastic policy outperform a deterministic one? Two classes of environment provide answers. The first involves uncertainty about the world. The second involves strategic opponents.
Consider the exploration–exploitation dilemma first. A deterministic agent that always picks the option with the highest current estimate can lock onto an inferior option permanently, because it never samples alternatives that might be better. Softmax choice is a form of random exploration: it samples options in proportion to their estimated value. Wilson and colleagues used a horizon task to show that people increase random exploration when more choices remain ahead, which is exactly when information is most valuable. The amount of variability is regulated, not fixed.
Findling, Wyart, and colleagues located much behavioural variability in noisy learning updates rather than at the point of choice. They also showed that this noise was adaptive, improving performance when reward contingencies shifted. Imprecise computation can act as regularisation, preventing overconfident commitment to outdated models of the environment. Thompson sampling formalises a related intuition: an agent that acts by drawing from its posterior beliefs explores in proportion to its uncertainty, and this strategy is near-optimal in many bandit problems.
Strategic settings sharpen the argument. Von Neumann's minimax theorem establishes that in zero-sum games without a pure-strategy equilibrium, optimal play must be mixed. Predictability is exploitable. Barraclough, Conroy, and Lee had monkeys play matching pennies against a computer opponent that exploited sequential regularities in their choices. The animals shifted toward near-random play, and neurons in dorsolateral prefrontal cortex encoded the recent history of choices and outcomes. In this setting, stochasticity is a solution rather than a failure.
Rational inattention offers a unifying perspective. Matějka and McKay showed that when acquiring information carries a cost proportional to Shannon entropy, the optimal choice rule is a generalised logit. The softmax temperature then reflects the shadow price of information. On this account, choice noise is the observable residue of an optimal trade-off between precision and its metabolic or cognitive cost. The agent is rational precisely in how it allocates its imprecision.
TakeawayA perfectly consistent agent is a perfectly predictable one. Predictability has a price, both in a changing world and against an opponent who is paying attention.
Stochastic choice dissolves the apparent conflict between rational theory and messy behaviour. Random utility models keep the logic of maximisation while relocating rationality to the level of distributions. The question shifts from whether a single choice was correct to whether the overall pattern of choice probabilities is coherent.
Mechanistic decomposition adds a second layer. Encoding, retrieval, and execution noise leave separable signatures. Identifying them turns variability into a diagnostic instrument for locating computational bottlenecks in the brain's valuation circuitry.
The normative analysis completes the picture. In uncertain or adversarial environments, and wherever information is costly, some noise is optimal. The deterministic agent of classical theory is therefore not an ideal that real agents fall short of. It is a special case: the zero-temperature limit, appropriate only to stable, non-strategic worlds where precision is free. Such worlds rarely exist. Perhaps the deeper question is not why we vary, but why we vary as little as we do.