Expected Utility Theory

Abstract

Expected utility theory is the normative account of choice under risk: a rational agent ranks uncertain options by the probability-weighted average of the utilities of their outcomes rather than by expected money. Descended from Bernoulli's eighteenth-century resolution of a famous coin-game paradox and placed on an axiomatic footing by von Neumann and Morgenstern, it derives an agent's attitude toward risk from the curvature of a single utility function and, in Savage's extension, from personal probabilities disciplined by the sure-thing principle. Its scientific value is double-edged, because the systematic ways real people violate it, above all in the Allais and Ellsberg paradoxes, are exactly what descriptive theories such as prospect theory were built to explain. This article develops the axioms, the account of risk, the paradoxes, and the theory's relationship to game theory and prospect theory, alongside three interactive demonstrations.

Keywords: expected utility, decision under risk, risk aversion

Expected utility theory (EUT) is the normative theory of decision making under risk: it specifies how an ideally rational agent should choose among options whose outcomes are uncertain but whose probabilities are known (von Neumann & Morgenstern, 1944). Its central claim is that such an agent ranks risky options, called lotteries or prospects, by their expected utility, the probability-weighted average of the utilities of their outcomes, rather than by their expected monetary value. The idea originated with Daniel Bernoulli's resolution of the St. Petersburg paradox, which introduced the concept of diminishing marginal utility (Bernoulli, 1738/1954); it was placed on an axiomatic footing by John von Neumann and Oskar Morgenstern, whose representation theorem showed that expected utility maximization is equivalent to a short list of consistency requirements on preferences; and it was extended by Leonard Savage to situations where probabilities themselves must be judged, yielding subjective expected utility (Savage, 1954). The theory is also the point of departure for much of cognitive psychology's work on judgment and decision making, because the ways people systematically violate it, catalogued from the Allais and Ellsberg paradoxes through the heuristics-and-biases program of Daniel Kahneman and Amos Tversky, are precisely what descriptive theories such as prospect theory were built to explain (Kahneman & Tversky, 1979). The sections below present the theory, its axioms and account of risk, the evidence against it as a description of human choice, and its relationship to game theory and prospect theory.

What Expected Utility Theory Is

Expected utility theory answers a question about rational choice under risk. A risky option can be described as a lottery: a list of possible outcomes, each with a probability, such as a coin flip that pays 100 dollars on heads and nothing on tails. The most obvious way to value a lottery is by its expected value, the probability-weighted average of its monetary outcomes; the coin flip's expected value is 50 dollars. Expected utility theory says this is the wrong quantity to maximize. What matters to a decision maker is not money itself but what money is worth to that person, its utility, and the rational rule is to maximize the probability-weighted average of utilities instead (Bernoulli, 1738/1954). Two clarifications frame everything that follows. First, the theory is not a claim that people consciously compute expectations; the representation theorem shows only that an agent whose preferences are internally consistent behaves as if maximizing expected utility, so the utility function is a summary of coherent preferences rather than a report of felt pleasure (von Neumann & Morgenstern, 1944). Second, the theory is primarily normative, a standard for how choices ought to hang together, and its career as a descriptive theory of actual human choice is a separate question with a largely negative answer, one whose variants, purposes, evidence, and limitations were already a substantial literature by the early 1980s (Machina, 1987; Schoemaker, 1982). Keeping the normative and descriptive roles distinct is the single most important discipline in this literature.

From the St. Petersburg Paradox to Utility

The theory began as the solution to a puzzle. Consider a game in which a fair coin is tossed until it first lands heads, with a prize that starts small and doubles for every additional toss the game lasts. Because each possible ending contributes the same fixed amount to the total, the game's expected monetary value grows without bound: it is, formally, infinite. Yet no sensible person would pay more than a modest sum to play. This is the St. Petersburg paradox, and Daniel Bernoulli's resolution founded utility theory. Bernoulli argued that the value of money to a person diminishes as they have more of it, so that a fixed gain matters less to the rich than to the poor, and he proposed that people evaluate gambles by the expected utility of wealth rather than its expected amount (Bernoulli, 1738/1954). With a utility function that rises at a diminishing rate, such as the logarithm Bernoulli suggested, the St. Petersburg game's expected utility is finite and its fair price modest, matching intuition. The same idea explains in one stroke why the game is worth little and why insurance is worth buying: a loss hurts more than an equal gain helps, so paying a small certain premium to remove a large risk raises expected utility even when the premium exceeds the expected loss. Diminishing marginal utility, introduced to dissolve a puzzle about a coin game, became one of the most consequential ideas in the study of choice.

The von Neumann and Morgenstern Axioms

For two centuries Bernoulli's proposal remained an insightful hypothesis. In 1944, von Neumann and Morgenstern transformed it into a theorem by asking what assumptions about preferences are equivalent to expected utility maximization. Four axioms do the work. Completeness requires that the agent can compare any two lotteries: one is preferred, or the two are equally good. Transitivity requires that preferences not cycle: an agent who prefers A to B and B to C prefers A to C. Continuity requires that no outcome be infinitely better or worse than another: if A is preferred to B and B to C, there is some probability mixture of A and C that is exactly as good as B. Independence, the axiom with the largest consequences, requires that a preference between two lotteries survive mixing each with the same third lottery in the same proportion: if A is at least as good as B, then a p chance of A combined with a 1 minus p chance of C is at least as good as a p chance of B combined with that same chance of C, for any C and any positive p. The representation theorem then states that an agent's preferences satisfy these four axioms if and only if there exists a utility function over outcomes such that one lottery is preferred to another exactly when its expected utility is higher, and that this function is unique up to a positive affine transformation, meaning its zero point and unit are arbitrary but its curvature is not (von Neumann & Morgenstern, 1944). The theorem's force is philosophical as much as mathematical: it reduces the rationality of a process, maximizing an expectation, to the coherence of a pattern of preferences, so that to reject expected utility one must say which axiom to give up.

Subjective Expected Utility and the Sure-Thing Principle

The von Neumann and Morgenstern theorem assumes the probabilities are given, as at a roulette wheel. Most decisions are not like that: the relevant likelihoods, of a lawsuit succeeding, a technology working, a storm arriving, must be judged. Leonard Savage closed this gap, building on a line of thought opened by Frank Ramsey and Bruno de Finetti, who had earlier proposed treating probability as a coherent personal degree of belief. In his framework a decision maker chooses among acts, whose consequences depend on which state of the world obtains, and Savage showed that if preferences among acts satisfy a set of postulates, the agent behaves as if assigning personal, subjective probabilities to the states and maximizing expected utility computed with them (Savage, 1954). The central postulate is the sure-thing principle: if two acts have identical consequences in some set of states, then the preference between them should depend only on the states where their consequences differ. What happens where they agree is a sure thing either way and should be irrelevant. Savage's construction is the foundation of Bayesian decision theory, because it derives both the utility function and the probability distribution from choice behavior alone, and it extends expected utility from risk, where probabilities are objective, to uncertainty, where they are not. It also sets up the theory's deepest empirical trouble, since the sure-thing principle is the subjective counterpart of the independence axiom, and both, as the paradoxes below show, are exactly where human preferences give way.

Risk Attitudes and the Shape of Utility

In expected utility theory, an agent's attitude toward risk is not a separate personality trait; it is fully determined by the curvature of the utility function. If utility is concave, rising at a diminishing rate, then the utility of a gamble's expected value exceeds the gamble's expected utility, and the agent prefers a sure amount to a fair gamble around it: risk aversion. If utility is convex, the agent is risk seeking, and if it is linear the agent is risk neutral, maximizing expected value. Two quantities make the analysis concrete. The certainty equivalent of a gamble is the sure amount that is exactly as good as the gamble, and the risk premium is the gap between the gamble's expected value and that certainty equivalent, the amount of expected value the agent will sacrifice to escape the risk. Figure 1 shows the geometry. John Pratt, in work paralleled independently by Kenneth Arrow, showed that the local intensity of risk aversion is measured by the negative of the ratio of the utility function's second derivative to its first, the Arrow-Pratt coefficient, which is positive exactly when utility is concave and links the curvature of utility directly to the premiums people will pay, underlying the economics of insurance and portfolio choice (Pratt, 1964). Milton Friedman and Savage, meanwhile, confronted an awkward everyday fact: the same person often buys insurance, a risk-averse act, and lottery tickets, a risk-seeking one. Their explanation kept the theory intact by giving the utility function a shape with both concave and convex regions, so that attitudes toward risk could differ across ranges of wealth (Friedman & Savage, 1948). Harry Markowitz sharpened the account by arguing that utility is better defined over gains and losses measured from present wealth than over total wealth, an early move toward the reference dependence that prospect theory would later place at its center (Markowitz, 1952). That the same person's risk attitude flips with the situation rather than the wealth level would return decades later as central evidence for the descriptive successor.

The Allais and Ellsberg Paradoxes

The case against expected utility as a description of human choice begins with two thought experiments that became real experiments. Maurice Allais constructed a pair of choice problems with a common structure. In the first, people choose between a large sure prize and a gamble offering a slightly larger prize with high probability but a small chance of nothing; most take the sure prize. In the second, the same options are transformed by removing an identical high chance of the sure prize from both, leaving two long-shot gambles; most now take the one with the bigger prize. The pattern feels natural, yet it is flatly inconsistent with the independence axiom, because the two problems differ only in a consequence common to both options, which the axiom says cannot matter (Allais, 1953). Kahneman and Tversky later showed the same reversal with modest stakes and named its source the certainty effect: outcomes that are certain are weighted out of proportion to their probability, so reducing a probability from certainty hurts more than an equal reduction between two merely probable values (Kahneman & Tversky, 1979). Daniel Ellsberg aimed at the subjective side of the theory. Offered bets on draws from an urn with a known fifty-fifty color mix and an urn whose mix is unknown, people prefer to bet on the known urn, whichever color is at stake. No assignment of subjective probabilities can rationalize that pattern, since the preferences imply the unknown urn's two colors each have probability below one half, which is impossible; the pattern violates the sure-thing principle and reveals ambiguity aversion, a dislike of unknown probabilities as such, developed in full on the Ellsberg paradox page (Ellsberg, 1961). What makes both paradoxes profound rather than merely curious is their stubbornness: many people, shown exactly how their choices contradict the axioms, reflect and then decline to change them, which turns a descriptive failure into a live normative argument about whether the axioms deserve their authority (Machina, 1987). The demonstration below presents the Allais choices directly, before the analysis, in the paradox's classic large-prize form.

Why the Axioms Fail

Two lines of explanation account for the violations, one about the limits of the chooser and one about the structure of value. The first begins with the recognition that the theory's ideal agent, who surveys every option, knows every consequence, and computes without error, describes no real mind; actual decision makers operate under bounded rationality, and the probability judgments the theory treats as inputs are generated by heuristics that are efficient but systematically biased, so the numbers entering a real decision are not the coherent probabilities that subjective expected utility requires. The theory also presumes description invariance, that logically equivalent statements of the same options must yield the same choice, yet framing reliably reverses preferences when one and the same outcome is redescribed as a gain rather than a loss. The second line is sharper because it is internal to the theory. Within expected utility, aversion to a small gamble can come only from the curvature of utility, and Matthew Rabin proved that any agent curved enough to turn down a modest favorable fifty-fifty bet at every wealth level would be forced, by that same curvature, to reject enormous favorable bets that no one would actually refuse; realistic small-stakes risk aversion therefore cannot be explained by utility curvature at all (Rabin, 2000). The calibration result closes the last escape route for the standard model as a description, and it points toward the alternative prospect theory would supply: value defined over changes, with a premium on losses, rather than over final wealth. That the failures are lawful rather than random is what makes them usable, and the systematic search for a descriptive theory that captures them without discarding the tractable core of expectation is the subject of a large literature (Starmer, 2000). Even the measurement of risk attitudes proves method-dependent: in incentivized choice lists the same person can appear more or less risk averse depending on how the options are arranged and whether the stakes are real, a sensitivity that blurs the line between a stable preference and the procedure used to elicit it (Holt & Laury, 2002).

Beyond Expected Utility: Prospect Theory and Its Relatives

Prospect theory is the descriptive replacement built directly on this evidence, and it revises expected utility theory in three precise ways (Kahneman & Tversky, 1979). First, the carriers of value are not final states of wealth but changes, gains and losses measured from a reference point, usually the status quo, which is why framing matters: a frame sets the reference point. Second, the value function is concave for gains, convex for losses, and steeper for losses than for gains; this single S-shape produces the reflection effect, risk aversion for gains flipping to risk seeking for losses, and builds in loss aversion. Third, probabilities do not enter linearly but through a weighting function that overweights small probabilities and underweights moderate and large ones, with a sharp change in impact near certainty, which reproduces the Allais pattern and explains the simultaneous demand for lottery tickets and insurance without contorting the utility curve. The 1992 cumulative version applies the weighting to cumulative rather than individual probabilities, repairing a technical flaw by which the original theory could prefer a dominated option, and delivers the fourfold pattern of risk attitudes, risk aversion for probable gains and improbable losses alongside risk seeking for improbable gains and probable losses, together with measured parameter values (Tversky & Kahneman, 1992). Prospect theory is the most prominent member of a large family: rank-dependent utility, regret theory, disappointment theories, and others each relax a different axiom, and the systematic hunt among them for a descriptive successor to expected utility has been surveyed at length (Starmer, 2000). Table 1 sets out the core contrast, and the Prospect Theory page develops the descriptive side in full.

Table 1

Expected Utility Theory and Prospect Theory Compared

Note. Each prospect-theory feature answers a specific expected-utility anomaly: probability weighting absorbs the Allais pattern, the S-shaped value function absorbs the reflection effect and the insurance-and-lottery puzzle, and loss aversion absorbs endowment and status-quo effects (Kahneman & Tversky, 1979; Tversky & Kahneman, 1992).

Expected Utility, Game Theory, and Prospect Theory

Expected utility theory sits at the center of a cluster of theories, and its relationships to the other two members are different in kind. The connection to game theory is not an analogy but an inheritance. Von Neumann and Morgenstern introduced the axioms in the same 1944 book that founded game theory, and for a structural reason: a theory of strategy needs a cardinal measure of what players are trying to get, because the expected payoff of a mixed strategy, a deliberate randomization over moves, is well defined only if utilities are measured on an interval scale, which is exactly what the representation theorem delivers (von Neumann & Morgenstern, 1944). Classical game theory's solution concepts accordingly assume that every player maximizes expected utility given beliefs about the others; game theory extends expected utility theory outward, from a single agent choosing against nature to many agents choosing against each other, while holding the account of rationality fixed. It thereby inherits both the theory's normative force and its descriptive vulnerabilities, since evidence that individuals violate the axioms is equally evidence about the players that equilibrium analysis presumes. The relationship to prospect theory runs the other way. Prospect theory keeps the expectation-like structure of the theory, a sum of weights multiplied by values, but replaces both components with psychologically realistic ones: reference-dependent value in place of final-wealth utility, and transformed decision weights in place of probabilities (Kahneman & Tversky, 1979). It revises expected utility theory downward, substituting a descriptive account of the individual agent while leaving the normative theory standing as the benchmark against which its own departures are defined. The map of the cluster is therefore simple: expected utility theory is the normative core; game theory extends it outward to strategic interaction, retaining the rational agent; prospect theory revises it downward, replacing the rational agent with a real one; and behavioral game theory is what results when prospect-theoretic realism is carried back into strategic settings.

Criticisms and Open Questions

The deepest dispute concerns the theory's normative authority, not its descriptive adequacy. Allais did not present his problems as a report of human error; he argued that the choices most people make are reasonable, that a rational person may care about the certainty of an outcome and about how outcomes are distributed across states, and that it is the independence axiom, not the chooser, that should yield (Allais, 1953). The persistence of paradoxical choices under full reflection keeps this argument alive, and the case is sharpest for ambiguity: many decision theorists hold that preferring known to unknown probabilities in the Ellsberg problem is defensible, which has motivated formal models of ambiguity-sensitive rationality rather than a verdict of irrationality (Ellsberg, 1961; Machina, 1987). Defenders reply that the axioms are backed by pragmatic arguments, an agent with cyclic preferences can be turned into a money pump, and one who violates independence can be led into dominated combinations of choices, so the standard is not arbitrary. A second open question is what the descriptive failures license. Expected utility remains the workhorse of economic modeling, partly because its violations, while systematic, can be small in aggregated market settings, and partly because no descriptive successor has matched its tractability and generality; the non-expected-utility literature contains many models that each fix some anomalies while inheriting or creating others (Starmer, 2000; Schoemaker, 1982). Rabin's theorem sharpened rather than settled this, showing that expected utility cannot be even approximately right about small-stakes risk aversion, which forces modelers to decide, case by case, when the approximation is safe (Rabin, 2000). A third frontier is the input side: the theory takes probabilities and utilities as given, but eliciting them stably is notoriously difficult, and measured risk attitudes shift with the elicitation method itself (Holt & Laury, 2002). These are live disagreements among careful researchers, and the theory's continued centrality alongside them is the field's honest condition.

Worked Example

The utility-curve demonstration makes the theory's core computation concrete. Take the gamble it models: a fifty-fifty chance of 100 against nothing, whose expected value is the probability-weighted average, 0.50 times 100, that is 50. Value the gamble not by that average but by expected utility, using the square-root utility function u(x) equal to the square root of x divided by 100, which is concave and so encodes risk aversion. The gamble's expected utility is 0.50 times u(100) plus 0.50 times u(0), which is 0.50 times 1 plus 0.50 times 0, that is 0.50. The certainty equivalent is the sure amount whose utility equals that expected utility: setting u of CE equal to 0.50 gives CE divided by 100 equal to 0.50 squared, so CE equals 25. The risk premium is the gap between the expected value and the certainty equivalent, 50 minus 25, that is 25; a risk-averse agent with this utility function would give up as much as 25 units of expected value to replace the gamble with a sure thing. Setting the demonstration's curvature exponent to 0.50 reproduces exactly these values; raising it toward 1.00 flattens the curve and shrinks the premium toward zero, while at the default exponent of 0.60 the certainty equivalent rises to about 31.5 and the premium falls to 18.5. The same concavity that produces a positive risk premium is what makes insurance rational: paying a certain premium to shed a risk raises expected utility whenever the utility curve bends enough, which is why a buyer will accept a premium above the actuarially fair price while the insurer still profits (Bernoulli, 1738/1954; Pratt, 1964). Every step here is the theory working as designed, which is precisely why the demonstrations above, where human choices refuse to follow the same logic, carry the force they do.

Discussion

Expected utility theory matters first as a standard. Decision analysis in medicine, engineering, law, and public policy is built on its machinery, structuring hard choices as probabilities and utilities and recommending the option that maximizes the expectation, and Bayesian statistics rests on Savage's foundations, in which coherent belief and coherent action are two faces of the same axioms (Savage, 1954). It matters second as infrastructure: the economics of insurance, portfolio selection, and auctions, and the whole apparatus of game-theoretic modeling, run on von Neumann and Morgenstern utilities, and the Arrow-Pratt measure remains the common language for comparing risk attitudes across people and contexts (von Neumann & Morgenstern, 1944; Pratt, 1964). It matters third, and perhaps most durably for cognitive psychology, as a measuring instrument. Because the theory says exactly what a coherent chooser would do, it converts human behavior into a map of precise, replicable departures, the certainty effect, ambiguity aversion, framing, and loss aversion, and that map is the empirical foundation on which prospect theory and behavioral economics were built (Kahneman & Tversky, 1979; Starmer, 2000). A theory can fail as description and still organize a science; expected utility theory is the standing proof.

Key Terms

Key Researchers

Daniel Bernoulli (1700-1782). Swiss mathematician and physicist who resolved the St. Petersburg paradox by arguing that people value the utility of wealth rather than wealth itself, introducing diminishing marginal utility and the first expected utility analysis.
Wikipedia

John von Neumann (1903-1957). Hungarian-American mathematician who, with Oskar Morgenstern, proved the representation theorem showing that preferences satisfying four consistency axioms are equivalent to expected utility maximization, in the same book that founded game theory.
Wikipedia

Oskar Morgenstern (1902-1977). Austrian-American economist and co-author of Theory of Games and Economic Behavior, which fused axiomatic utility theory with the mathematics of strategic interaction.
Wikipedia

Leonard J. Savage (1917-1971). American mathematician and statistician who extended the theory to situations without objective probabilities, deriving subjective probability and utility jointly from preference postulates centered on the sure-thing principle.
Wikipedia

Maurice Allais (1911-2010). French economist and 1988 Nobel laureate whose choice problems demonstrated systematic violations of the independence axiom and opened the modern debate over the theory's descriptive and normative standing.
Wikipedia

Daniel Kahneman (1934-2024). Israeli-American psychologist and 2002 Nobel laureate who, with Amos Tversky, created prospect theory, the leading descriptive alternative to expected utility theory.
Wikipedia

Amos Tversky (1937-1996). Israeli cognitive psychologist who, with Kahneman, documented the heuristics, framing effects, and loss aversion that explain when and why human choices depart from the axioms.
Wikipedia

Matthew Rabin (living). American behavioral economist at Harvard University whose calibration theorem proved that expected utility cannot explain small-stakes risk aversion through utility curvature, closing a central gap in the case for a descriptive successor.
Faculty Page - Google Scholar - Wikipedia

Frequently Asked Questions

What is expected utility theory?
Expected utility theory is the standard normative theory of decision making under risk. It holds that a rational agent ranks risky options by the probability-weighted average of the utilities of their outcomes rather than by the average of the monetary amounts themselves (von Neumann & Morgenstern, 1944).

What is the difference between expected value and expected utility?
Expected value weights each monetary outcome by its probability, while expected utility weights the utility of each outcome instead. Because the utility of money typically rises at a diminishing rate, a gamble can be worth much less than its expected value, which is how Bernoulli resolved the St. Petersburg paradox (Bernoulli, 1738/1954).

What are the von Neumann and Morgenstern axioms?
They are completeness, transitivity, continuity, and independence. The representation theorem proves that preferences over risky options satisfy these four axioms exactly when they can be represented as maximizing expected utility, with the utility function unique up to a positive affine transformation (von Neumann & Morgenstern, 1944).

What does it mean to be risk averse in expected utility theory?
Risk aversion means preferring a sure amount to a fair gamble around it, and in the theory it corresponds to a concave utility function. The sure amount as good as a gamble is the certainty equivalent, the shortfall from expected value is the risk premium, and the intensity of aversion is measured by the Arrow-Pratt coefficient (Pratt, 1964).

What is the Allais paradox?
It is a pair of choice problems in which most people prefer a sure prize in one problem but a long-shot gamble in the other, even though the two problems differ only in an outcome common to both options. That pattern violates the independence axiom and reveals the certainty effect, the disproportionate weight given to outcomes that are certain (Allais, 1953).

What is the Ellsberg paradox?
It is the finding that people prefer betting on an urn with a known fifty-fifty mix of colors over an urn with an unknown mix, whichever color they bet on. No assignment of subjective probabilities can rationalize this pattern, so it violates the sure-thing principle and demonstrates ambiguity aversion (Ellsberg, 1961).

How does prospect theory differ from expected utility theory?
Prospect theory values gains and losses relative to a reference point instead of final wealth, uses an S-shaped value function that is steeper for losses, and transforms probabilities with a weighting function that overweights small probabilities. These changes describe actual choices, including the Allais pattern and framing reversals, that expected utility theory forbids (Kahneman & Tversky, 1979).

Is expected utility theory still used if people violate it?
Yes. It remains the normative benchmark in decision analysis, the foundation of game theory and Bayesian statistics, and the standard model in much of economics, while descriptive research measures how real choices depart from it. Its violations are systematic but well mapped, and that map is what behavioral theories are built on (Starmer, 2000).

References