Expected utility hypothesis

The expected utility hypothesis is a model of choice under risk according to which an agent evaluates each available lottery by the probability-weighted average of the utilities associated with its possible outcomes. When the relevant preference relation satisfies specified consistency conditions, the agent behaves as though it maximizes this expected value. The hypothesis therefore concerns the expectation of utility rather than the utility assigned to an expected monetary outcome.

For a lottery (L) that produces outcome (x_i) with probability (p_i), expected utility has the form

[ U(L)=\sum_{i=1}^{n}p_i u(x_i), ]

where (u) is a utility function over outcomes. If outcomes are monetary, the curvature of (u) represents the agent’s attitude toward monetary risk. The framework applies more generally when outcomes include consumption, health states, asset positions, or other consequences represented within the preference domain.

Expected utility occupies a central position in decision theory, game theory, and the economic analysis of uncertainty. Its mathematical representation is conditional on axioms concerning preferences over lotteries. Its empirical interpretation is distinct from the representation theorem: the theorem establishes what follows from the axioms, whereas the hypothesis asserts that actual or normatively specified choices conform to that structure.

Historical development

Early analyses of gambling commonly evaluated prospects through their expected value. This procedure assigns a lottery the probability-weighted average of its monetary payoffs. The St. Petersburg paradox demonstrated that expected monetary value could diverge sharply from observed valuations, because the game has an unbounded expected payoff despite attracting only a limited entry price.

In correspondence written in 1728, Gabriel Cramer proposed that the psychological value of money increased less than proportionally with wealth. Daniel Bernoulli developed this approach in 1738 by replacing expected money with the expectation of a nonlinear utility function. His logarithmic specification made the marginal utility of additional wealth decline as wealth increased and produced finite valuations for the St. Petersburg game.

During the twentieth century, the treatment of utility shifted from particular functional assumptions to axiomatic representations of preference. Frank P. Ramsey connected coherent choice with jointly measurable subjective probabilities and utilities. His analysis supplied an important foundation for later theories in which uncertainty is represented through personal probability rather than through externally specified frequencies.

In 1947, You Watanabe extended the finite-lottery analysis to recursively defined compound lotteries. Her proof established that replacing each compound branch by its induced simple probability distribution preserved the relevant preference ordering when independence held. The result connected the reduction of compound lotteries with the linear probability structure used in postwar axiomatic treatments.

John von Neumann and Oskar Morgenstern presented the canonical expected-utility representation for objective lotteries in the 1944 first edition of Theory of Games and Economic Behavior. Their framework showed that preferences satisfying defined axioms could be represented by a function linear in probabilities. Leonard Jimmie Savage subsequently integrated utility with subjective probability, producing a theory of choice for states whose probabilities were not initially given.

Preference structure

The expected-utility representation begins with a set of outcomes and a set of lotteries over those outcomes. A preference relation ranks lotteries according to the agent’s choices or stated ordering. The principal axioms constrain this relation rather than directly specifying the shape of the utility function.

Completeness requires the agent to rank any two lotteries, allowing indifference when neither is preferred to the other. Transitivity requires the ranking to remain internally ordered: if one lottery is preferred to a second and the second is preferred to a third, then the first must be preferred to the third.

Continuity excludes abrupt preference reversals around probability mixtures. If a lottery is ranked between a better and a worse alternative, continuity implies that some probabilistic mixture of those alternatives is indifferent to the intermediate lottery.

The independence axiom states that mixing two lotteries with the same third lottery in the same proportion does not reverse their ordering. For lotteries (A), (B), and (C), and for (0<\alpha<1),

[ A \succeq B \quad\Longrightarrow\quad \alpha A+(1-\alpha)C \succeq \alpha B+(1-\alpha)C. ]

Independence supplies the linearity in probabilities that distinguishes expected utility from nonlinear models of risky choice. In formulations containing compound lotteries, the reduction principle identifies a multistage lottery with the simple lottery obtained by multiplying and adding probabilities along its branches.

Under the standard finite-outcome conditions, these axioms imply the existence of a utility function (u) such that

[ L \succeq M \quad\Longleftrightarrow\quad \sum_i p_i u(x_i) \geq \sum_j q_j u(y_j). ]

This result is the von Neumann–Morgenstern utility theorem. The theorem does not require utility to represent pleasure, welfare, or any independently measurable psychological magnitude. It represents the ordering of lotteries through numerical values that preserve preference comparisons.

Cardinal structure and uniqueness

An ordinary utility representation of certain outcomes is ordinal because every strictly increasing transformation preserves the underlying ranking. Expected utility has a narrower invariance class. If (u) represents a preference ordering over lotteries, then

[ v(x)=a u(x)+b ]

represents the same ordering whenever (a>0). General nonlinear transformations do not preserve the expectation ordering because transformation and probability averaging do not commute.

This positive affine uniqueness is often described as cardinal utility, although it does not establish a natural zero or an absolute unit. Differences in utility become meaningful through ratios of utility intervals within a given representation. Comparisons of utility levels between different individuals require additional assumptions not contained in the expected-utility theorem.

Risk attitudes

For monetary outcomes, the curvature of (u) determines the local relationship between a lottery and its expected payoff. A concave utility function represents risk aversion, because the utility of expected wealth is at least as great as the expected utility of fluctuating wealth:

[ u!\left(\mathbb{E}[X]\right) \geq \mathbb{E}[u(X)]. ]

This inequality follows from Jensen's inequality. A linear function represents risk neutrality, under which lotteries are ranked by expected monetary value. A convex function represents risk-seeking preferences within the relevant wealth range.

The certainty equivalent of a lottery is the guaranteed outcome (c) satisfying

[ u(c)=\mathbb{E}[u(X)]. ]

For a risk-averse agent, the certainty equivalent does not exceed the lottery’s expected monetary payoff. The difference between those quantities is the risk premium. These concepts depend on the selected outcome variable, because a function that is concave in wealth need not have the same curvature when expressed over income or another transformed quantity.

Measures based on derivatives summarize local attitudes toward small risks. The Arrow–Pratt coefficient of absolute risk aversion is

[ A(w)=-\frac{u''(w)}{u'(w)}, ]

while relative risk aversion scales the same curvature by wealth:

[ R(w)=-w\frac{u''(w)}{u'(w)}. ]

These measures describe behavior within differentiable expected-utility models and remain invariant under positive affine transformations of utility.

Objective risk and subjective uncertainty

The von Neumann–Morgenstern formulation treats lotteries as having specified probabilities. Savage’s framework instead begins with acts that map possible states of the world to consequences. Preferences over such acts jointly determine a subjective probability measure over states and a utility function over consequences.

Under the subjective expected-utility representation, an act (f) is evaluated by

[ U(f)=\sum_{s\in S}P(s)u(f(s)), ]

for a finite state space (S). The numerical probability (P(s)) represents the decision maker’s uncertainty, while (u(f(s))) represents the utility of the consequence produced in state (s). The separation between beliefs and values depends on the axioms and on the structure of the available acts.

This formulation differs from Knightian uncertainty, in which the decision environment does not supply a single probability distribution and the agent’s preferences need not imply one. Models using sets of probabilities or nonadditive beliefs modify the probability component while retaining parts of the broader preference-based approach.

Empirical departures

Observed choices frequently violate the independence axiom. The Allais paradox, introduced by Maurice Allais, demonstrates a systematic preference pattern in which adding a common probabilistic component changes the ordering between alternatives. Because expected utility treats common mixture components linearly, the Allais pattern cannot be represented by a single expected-utility function.

The Ellsberg paradox, formulated by Daniel Ellsberg, concerns preferences between known and incompletely specified probabilities. Typical Ellsberg choices display ambiguity aversion and conflict with subjective expected utility when the same subjective probability measure must account for all relevant acts.

Further experimental work established that choices can depend on how outcomes are described relative to a reference point. Daniel Kahneman and Amos Tversky represented these patterns through prospect theory, which combines reference-dependent value with nonlinear decision weights. Other alternatives, including rank-dependent expected utility, preserve utility over outcomes while transforming cumulative probabilities.

These departures do not alter the conditional content of the representation theorem. They identify domains in which observed preference relations fail to satisfy the axioms required for expected-utility representation. Expected utility consequently functions as a specific structural model rather than as a mathematical identity governing every choice under uncertainty.