Expected Utility Theorem (von Neumann-Morgenstern)

What the theorem actually says

The statement is precise. Let L denote the set of lotteries — probability distributions over a finite set of outcomes. A preference relation ≽ on L satisfies the four vNM axioms:

1. Completeness. For any two lotteries L and M, either L ≽ M, or M ≽ L, or both (indifference).
2. Transitivity. If L ≽ M and M ≽ N, then L ≽ N.
3. Continuity. If L ≽ M ≽ N, there exists a probability α ∈ [0,1] such that αL + (1−α)N is indifferent to M.
4. Independence. For any third lottery N and any α ∈ (0,1], L ≽ M if and only if αL + (1−α)N ≽ αM + (1−α)N.

If — and only if — all four hold, there exists a real-valued function u on outcomes such that L ≽ M if and only if Σ p_i^L·u(x_i) ≥ Σ p_i^M·u(x_i). The function u is unique up to a positive affine transformation — replace u with a·u + b for any a > 0 and the same preferences are represented.

A worked example with three lotteries

Suppose an agent has u(x) = √x — the canonical concave (risk-averse) utility. Consider three lotteries over wealth in dollars:

• L₁: $100 for sure. Expected utility = √100 = 10.
• L₂: 50% chance of $0, 50% chance of $400. Expected utility = 0.5·√0 + 0.5·√400 = 0 + 10 = 10. Expected dollar value = $200.
• L₃: 80% chance of $100, 20% chance of $400. Expected utility = 0.8·10 + 0.2·20 = 12. Expected dollar value = $160.

The agent is indifferent between L₁ and L₂ (same expected utility 10) even though L₂ has expected value $200. The risk premium is $200 − $100 = $100 — the agent gives up $100 of expected money to avoid the variance. L₃ is strictly preferred (EU = 12 > 10), even though its expected value is lower than L₂'s — concavity rewards the lottery with smaller spread.

Now apply the independence axiom. If L₃ ≽ L₁, then mixing both with any third lottery preserves the order. Mix each with a 50% chance of $25 (utility 5):

L₃' = 0.5·L₃ + 0.5·{$25 sure}: EU = 0.5·12 + 0.5·5 = 8.5
L₁' = 0.5·L₁ + 0.5·{$25 sure}: EU = 0.5·10 + 0.5·5 = 7.5

The preference is preserved: 8.5 > 7.5. The independence axiom is exactly what makes this trivially true — the third lottery contributes additively, so the original ranking carries through.

Expected utility vs alternatives

	Expected utility (vNM, 1944)	Prospect theory (KT, 1979)	Mean-variance (Markowitz, 1952)
Object of valuation	Final wealth	Gains/losses vs reference	(Mean, variance) of returns
Axiomatized?	Yes (4 axioms)	Descriptive — not axiomatic until 1992	No — assumes normal returns or quadratic u
Risk aversion source	Concavity of u	Loss aversion ratio λ ≈ 2.25	Variance penalty parameter
Probabilities treated	As is (linear)	Weighted: w(p) overweights small p	As is (in mean computation)
Allais paradox	Violated	Explained	Inconsistent with EU under non-normality
Ellsberg paradox	Violated	Partially explained	Not addressed
Used by	Insurance pricing, finance theory, decision analysis	Behavioral economics, marketing	Portfolio managers, robo-advisors
Empirical fit	Aggregate behavior, large stakes	Individual decisions, small/medium stakes	Asset returns when distributions are stable

From game theory to the Nobel canon

John von Neumann's 1928 minimax theorem founded game theory, but the unified expected-utility framework came only with Oskar Morgenstern's collaboration. The 1944 first edition of Theory of Games and Economic Behavior (Princeton University Press, 625 pages) buried the expected utility theorem in an appendix added during proofs — Morgenstern's economist colleagues found the original mostly-game-theoretic manuscript unintuitive, and Morgenstern insisted on a foundation for individual choice under risk.

The theorem was rediscovered and refined by Marschak (1950), Herstein and Milnor (1953), and Savage (1954) — Savage extending it to subjective probabilities. By the 1960s it was the standard model in microeconomics, and by the 1970s in finance through the Arrow-Debreu state-preference framework. Kenneth Arrow and Gerard Debreu's 1972 Nobel Prize was for general equilibrium built on top of vNM utility. Harry Markowitz, William Sharpe, and Merton Miller's 1990 Nobel was for mean-variance and CAPM — both implicit applications. Daniel Kahneman's 2002 Nobel was, in part, for documenting where the theorem fails.

Variants and extensions

• Subjective expected utility (Savage 1954). Probabilities themselves come from preferences over acts — outcomes contingent on states. Axiomatized via the sure-thing principle. Foundation of Bayesian decision theory.
• State-preference (Arrow 1953; Debreu 1959). Outcomes indexed by states of nature; agents trade state-contingent claims. The foundation of asset pricing.
• Rank-dependent expected utility (Quiggin 1982). Probabilities weighted by rank, not value; preserves stochastic dominance while accommodating Allais-type behavior.
• Choquet expected utility (Schmeidler 1989). Beliefs are non-additive capacities rather than probabilities — handles Ellsberg-style ambiguity.
• Maxmin expected utility (Gilboa-Schmeidler 1989). Agent considers a set of priors and maximizes the worst-case expected utility; the dominant ambiguity-averse model.
• Cumulative prospect theory (Tversky-Kahneman 1992). Kahneman-Tversky with rank-dependent weighting; reconciles prospect theory with stochastic dominance.

Real-world applications

• Insurance pricing. Premiums above expected loss are sustainable only because policyholders are risk-averse (concave u). A 25-year-old buying $500,000 of term life at $300/year — when the expected payout is $200/year — is paying a 50% premium for variance reduction.
• Asset pricing. The Lucas (1978) consumption-CAPM derives the equity premium from CRRA expected utility. Mehra-Prescott (1985) flagged the puzzle — required risk aversion γ ≈ 25 to fit historical premiums vs the empirical γ ≈ 3 from micro data.
• Optimal portfolio. Merton's (1969) continuous-time intertemporal model gives the famous f* = (μ−r)/(γσ²) Merton-share — a direct EU optimization result.
• Decision analysis. Howard Raiffa's 1968 Decision Analysis turned vNM utility into a practitioner toolkit; oil exploration, medical decision-making, and military targeting all run on assessed certainty-equivalents.
• Health economics. Quality-adjusted life years (QALYs) implicitly aggregate health states using EU; cost-effectiveness thresholds (NICE £20-30k/QALY) are EU calibrations.
• Climate policy. Nordhaus's DICE model and Stern Review use EU under uncertainty over climate damages; the choice of utility curvature (γ) explains most of the disagreement between them.

Common pitfalls and critiques

• Treating u as a measurable quantity. u carries cardinal information about this agent's risk attitude — it is not a measure of welfare or happiness across people. Interpersonal sums require an ethical premise (utilitarianism), not anything from the theorem.
• Estimating risk aversion from one decision. Calibration paradox (Rabin 2000): plausible small-stakes risk aversion implies absurd large-stakes risk aversion under EU. Rejecting a 50/50 gamble of −$10/+$11 implies rejecting a 50/50 gamble of −$1,000/+$∞ — a clear empirical falsification.
• Forgetting that independence is empirical. The axiom is normatively appealing but descriptively false in roughly half the population for stakes near choice boundaries.
• Confusing the theorem with utilitarianism. vNM is a representation theorem about individual choice; the social welfare extension requires separate axioms (Harsanyi 1955).
• Assuming u(x) is the same in different contexts. Empirically, risk aversion over money differs by domain (investments vs gambling vs lottery tickets), suggesting reference-dependence and narrow framing — both deviations from EU.