Expected Utility¶
Core Idea¶
Expected utility is the structural pattern of valuing an uncertain prospect by weighting the value (utility) of each possible outcome by its probability and summing — collapsing a distribution of possible futures into a single comparable scalar that ranks choices under risk. The defining commitment is probability-weighted aggregation of a value function over outcomes, where the value function is generally nonlinear, so that the worth of a gamble is neither its best case nor its naive average payoff but the expectation of utility rather than of money. The pattern was placed on a rigorous axiomatic footing by von Neumann and Morgenstern (1944), who showed that an agent whose preferences over risky prospects satisfy a short list of consistency conditions must behave as if maximizing the expectation of some utility function. [1] The deeper move is the separation of two ingredients that everyday reasoning tends to fuse: how likely an outcome is and how much it matters, recombined by a single multiplicative-then-additive rule. [1]
The nonlinearity of the value function is what gives expected utility its leverage. A concave value function — where each additional unit of the good is worth less than the last — automatically encodes risk aversion: the certain mean of a gamble is preferred to the gamble itself, because the upside is valued less steeply than the downside is penalized. Bernoulli (1738) anticipated this two centuries before the axioms, resolving the St. Petersburg paradox by proposing that people value the logarithm of wealth rather than wealth itself, so that a bet with infinite expected money commands only a finite price. [2] Expected utility thus names not merely an averaging operation but a valuation discipline: it specifies how a rational agent should compress uncertainty into a single ranking, and it makes the agent's attitude toward risk a readable property of the value function's curvature.
How would you explain it like I'm…
Worth of a Gamble
Average Value of Chances
Probability-Weighted Value
Structural Signature¶
Expected utility encodes a structural pattern: outcome space → probability weighting → value-function transform → scalar aggregation → ranking. It separates a prospect (a set of outcomes, each with a probability and a payoff) from its evaluation (one number), and names the rule — expectation taken over a transformed payoff — that maps the first onto the second. The transform is the load-bearing element: take expectation of utility, not of the raw outcome, and the curvature of the utility map carries the agent's posture toward risk. [1]
Recurring features:
- Probability-weighted aggregation of a value function over outcomes
- Collapsing a distribution of futures into one comparable score
- Expectation of utility, not of money
- Nonlinear value transform encoding risk attitude
- Separation of likelihood from significance, recombined multiplicatively
- Ranking uncertain prospects by a single scalar
- Concavity as the signature of risk aversion
The structural insight is robust: a portfolio manager pricing a risky asset, a reinforcement-learning agent choosing among actions by predicted return, a forager selecting between food patches of differing variance, and a clinician weighing treatment outcomes by probability and patient-valued quality of life all run the same machine. Each takes a branching set of uncertain futures, applies a value transform, weights by probability, and reads off a single comparable number. Savage (1954) extended the machinery to the subjective case, showing that the probabilities themselves need not be given by nature but can be the agent's own coherent degrees of belief, which broadens the pattern from objective lotteries to any decision under genuine uncertainty. [3]
What It Is Not¶
Expected utility is not a claim that people actually compute expectations in their heads. It is a normative-structural pattern — a benchmark for coherent choice under risk — and a descriptive idealization, not a literal account of cognition. The axiomatic result says only that an agent whose preferences are consistent in specified ways behaves as if maximizing expected utility; it does not assert that the agent performs the arithmetic. Treating it as a psychological mechanism rather than an as-if characterization is a common misreading. [1]
Nor is expected utility the same as expected value. Expected value averages the raw outcomes (the dollars, the calories, the lives); expected utility averages a transformed outcome (the utility of the dollars). The distinction is the whole point: when the value function is linear, the two coincide and the agent is risk-neutral, but the interesting cases — insurance, diversification, the rejection of fair-but-risky bets — live precisely where the function is nonlinear and the two diverge. Collapsing expected utility into expected value erases the concept's reason for existing.
Expected utility also does not claim that the value function is fixed, observable, or interpersonally comparable. The utility scale is unique only up to positive affine transformation; it is a representation of one agent's preferences, not a cardinal measure of welfare that can be summed across people. Reading utility numbers as if they were a universal currency of happiness, or comparing one person's utils against another's, imports assumptions the pattern itself does not license.
Finally, the prime does not assert that maximizing expected utility is always the right thing to do. It is silent on the ethics of risk, on whether the probabilities are well-founded, and on whether the agent's values are good ones. A coherent expected-utility maximizer can pursue a catastrophic goal flawlessly. The pattern describes how to rank prospects given a value function and a probability assignment; it does not adjudicate whether either input deserves respect. [3]
Broad Use¶
Economics & decision theory: The von Neumann–Morgenstern criterion for rational choice under risk; the foundation of choice theory, insurance economics, and the analysis of risk premia. Utility-of-wealth functions and their curvature underwrite the formal treatment of risk aversion (Pratt 1964; Arrow 1965) through the coefficient of absolute and relative risk aversion. [4][5]
Finance: Pricing and portfolio choice trade expected return against risk via a utility of wealth. Markowitz's mean-variance framework and its extensions can be read as expected-utility maximization under specific assumptions about the value function or the outcome distribution; option pricing and asset valuation rest on expectations taken under appropriately transformed measures.
Artificial intelligence: Expected-value maximization is the organizing principle of decision-theoretic planning and reinforcement learning, where an agent selects the action with the highest expected return (a probability-weighted sum of future rewards). The Bellman optimality equation is an expected-utility recursion; the entire apparatus of Markov decision processes rests on it.
Biology & ecology (non-obvious): Risk-sensitive foraging models treat animals as choosing between variable food patches as if maximizing expected fitness rather than expected calories, with concavity or convexity of the fitness function predicting risk-averse or risk-prone behavior depending on energetic state. [6]
Engineering & reliability: Expected-cost decisions weight failure severities by their probabilities; design choices minimize expected loss across a distribution of operating conditions and fault modes.
Medicine & public policy: Clinical decision analysis weights treatment outcomes by probability and patient-valued utility (quality-adjusted life years); cost-benefit analysis under uncertainty ranks policies by expected social outcome. Behavioral economics measures human deviations from this baseline, with prospect theory (Kahneman & Tversky 1979) reformulating the value transform around a reference point and replacing probabilities with decision weights. [7]
Clarity¶
Naming expected utility lets practitioners see that rational choice under uncertainty decomposes into two ingredients — how likely and how much it matters — and combines them by a single rule. Before the concept, the worth of a gamble is an undifferentiated intuition; after it, the intuition splits into a probability assignment and a value function, each of which can be examined, criticized, and revised independently. This separation is the source of the concept's clarifying power: a disagreement about a risky choice can now be diagnosed as a disagreement about likelihoods, a disagreement about values, or a disagreement about the combination rule itself. [1]
It also clarifies why a rational agent may rationally reject a bet with positive expected money. The everyday verdict — "you're leaving money on the table" — assumes risk neutrality. Expected utility shows that a concave value function makes the certain amount preferable, and that this is not a cognitive error but a coherent expression of how the agent values gains and losses. By the same token, the concept exposes when a decision rule departs from the benchmark: prospect theory's overweighting of small probabilities, the Allais paradox, and the Ellsberg ambiguity aversion are all defined and measured as precise deviations from the expected-utility baseline, which means the baseline supplies the very coordinate system in which irrationality becomes legible. [8][9][7]
Manages Complexity¶
Expected utility reduces a branching tree of uncertain futures, each leaf carrying its own payoff and probability, to one scalar score per option. This is the central complexity-management move: a decision problem that presents as an unmanageable space of contingencies — what if the market falls, what if the patient responds, what if the component fails — becomes a finite list of options each tagged with a single comparable number. The question "what might happen?" is transformed into "which option scores highest?" and the latter has a definite answer. [3]
The bounding move also makes otherwise incomparable alternatives directly rankable. A safe option with a modest sure payoff and a risky option with a wide spread of outcomes are not commensurable on their face; one is steady, the other is a gamble, and there is no obvious sense in which a number describes either. Expected utility supplies exactly that number, on a common scale, by pushing every outcome through the same value transform and the same probability weighting. The complexity of the future is not eliminated — it is absorbed into the structure of the value function and the probability distribution — but the decision is rendered tractable. This is why the pattern recurs wherever an agent must choose among risky options too numerous or too tangled to compare holistically.
Abstract Reasoning¶
Recognizing the structure supports a family of inferences that would otherwise require separate insight in each domain. The curvature of the value function encodes risk attitude, so concavity is risk aversion and convexity is risk seeking — a single geometric fact that explains why the same agent insures against large losses (the steep downside of a concave function) while buying lottery tickets is anomalous under it. The structure explains why diversification and insurance are rational rather than merely prudent: spreading risk across imperfectly correlated outcomes raises expected utility for a risk-averse agent even when it leaves expected money unchanged, because the value function rewards the reduction in variance. [4]
The benchmark is also generative under counterfactual reasoning. "What if this agent were risk-neutral?" collapses the value function to linear and recovers expected-value maximization, isolating exactly what risk attitude contributes. "What if the probabilities were the agent's beliefs rather than nature's frequencies?" recovers the subjective-expected-utility extension and licenses reasoning under genuine uncertainty. "What if small probabilities were systematically overweighted?" recovers prospect theory and predicts the characteristic fourfold pattern of risk attitudes. Each manipulation of one component — value transform, probability assignment, combination rule — yields a precise, transferable prediction, which is the hallmark of a structural pattern whose parts can be reasoned about independently.
Knowledge Transfer¶
The decision-theoretic template transfers cleanly across substrates because its parts are substrate-neutral. From financial portfolio choice, where the value function is a utility of wealth and the outcomes are returns, the same template moves to AI agents weighting action outcomes by predicted reward, where the value function is the reward model and the probabilities are transition dynamics. It moves again to evolutionary models in which genotypes "hedge" against environmental variance, where the value function is fitness and bet-hedging strategies are read as variance-reducing moves up a concave fitness curve. And it moves to medical decision analysis, where treatment outcomes are weighted by clinical probability and by patient-valued utility, so that a choice between a risky surgery and conservative management becomes an expected-utility comparison. [1][6]
A practitioner fluent in one of these domains can recognize the same machine in the others: the portfolio manager and the reinforcement-learning researcher are running the same recursion under different names, and the foraging ecologist and the clinical decision analyst share a value-transform-then-weight structure. The transfer is not metaphorical. It is grounded in the identity of the underlying operation — expectation of a transformed payoff — so that a result proven about one instantiation (the optimality of diversification, say, or the convergence of value iteration) often carries lessons, and sometimes outright theorems, into the others.
Examples¶
Formal/abstract¶
The sure thing versus the fair gamble. An agent is offered a 50/50 chance at $0 or $100 versus a sure $40. The gamble's expected money is $50, which exceeds the sure $40, so a risk-neutral agent (linear value function) takes the gamble. But an agent with a concave value function — say utility equal to the square root of wealth — values the gamble at 0.5·√0 + 0.5·√100 = 5 utils and the sure amount at √40 ≈ 6.32 utils. The certain $40 wins on expected utility even though it loses on expected money. The agent is not making an error; the curvature of the value function makes the avoided downside worth more than the forgone upside. Mapped back: the example isolates the prime's defining move — expectation of utility, not of money — and shows that risk aversion is not an add-on but a direct consequence of a concave value transform. The same calculation, with fitness substituted for wealth, explains why a forager near starvation may prefer a reliable small patch to a variable rich one: when the fitness function is steep at low energy reserves, variance is penalized, and the safe patch maximizes expected fitness.
Subjective probabilities and coherent belief. Suppose the outcomes depend not on a known coin but on an uncertain event — whether a new product will succeed — for which no objective frequency exists. Savage's framework shows that an agent whose preferences over bets on this event are coherent must act as if assigning a personal probability to success and then maximizing expected utility against it. The probability is not read off the world; it is inferred from the agent's own willingness to bet. Mapped back: this generalizes the prime from objective lotteries to any decision under genuine uncertainty, demonstrating that the probability-weighting component need not be supplied by nature. The structural pattern — value transform, probability weight, scalar aggregation — is preserved; only the source of the probabilities changes.
Applied/industry¶
Reinforcement learning and decision-theoretic planning. An autonomous agent at a decision point can take several actions, each leading stochastically to different future states with different rewards. The agent computes, for each action, the expected return — a probability-weighted sum over successor states of immediate reward plus discounted future value — and selects the action with the highest expected return. This is the Bellman optimality recursion, and it is expected-utility maximization with the reward model as the value function and the transition dynamics as the probabilities. Mapped back: the engineering instantiation runs the prime's exact machine: a branching tree of uncertain futures collapsed to one scalar per option, with the option of highest expected (utility) score chosen. The discount factor and the reward shaping play the role of the value-function transform, and the structural identity with financial portfolio choice is why techniques migrate freely between the two fields.
Clinical decision analysis. A clinician must choose between an aggressive treatment with a high chance of cure but a real risk of severe complications and a conservative course with a lower cure rate but fewer harms. A decision-analytic model assigns probabilities to each outcome (cure, complication, no effect) and a patient-valued utility to each health state (often as quality-adjusted life years), then ranks the two strategies by expected utility. A patient who weights the avoidance of severe complications heavily — a steeply concave value function over health states — may rationally prefer the conservative course even when the aggressive option has the higher expected survival. Mapped back: the medical instantiation makes the value function explicitly patient-specific and shows that the prime's "utility" need not be money or fitness; it is whatever the agent values, measured on its own scale. The structure — weight outcomes by probability, transform by a value function, aggregate to a scalar, rank — is identical to the financial and computational cases.
Structural Tensions¶
T1: The value function is doing the real work but is hardest to specify. Expected utility cleanly separates probability from value, yet in practice the value function — the curvature that encodes risk attitude — is rarely known and must be inferred, assumed, or elicited. Probabilities at least have frequencies or coherent-belief constraints to anchor them; the utility map has only the agent's own preferences, which are noisy, unstable, and often discovered only through the very choices the function is meant to explain. The concept's analytical power rests on a component that resists independent measurement, so the apparent precision of an expected-utility ranking can mask deep uncertainty about its central input.
T2: As-if optimality versus literal computation. The axiomatic justification says a coherent agent behaves as if maximizing expected utility, but the concept is constantly pressed into service as a description of how decisions are or should be made, and as a normative demand that they be made that way. Holding all three readings at once — representation theorem, descriptive model, prescriptive rule — creates strain: evidence that people violate the axioms can be read as showing people are irrational, or as showing the model is descriptively wrong, and the same data supports opposite conclusions depending on which reading is foregrounded.
T3: Aggregation to a scalar discards structure that may matter. Collapsing a distribution of futures into one number is the source of the concept's tractability, but it throws away everything about the distribution except what the value function happens to weight. Two prospects with identical expected utility can differ in skewness, in tail risk, in the correlation of their outcomes with other holdings, and in whether their worst cases are survivable. An agent who attends only to the scalar can be led into ruin by a high-expected-utility prospect whose left tail is catastrophic, because the aggregation has already integrated the catastrophe away into an average.
T4: Garbage probabilities, confident rankings. The expected-utility calculation is only as good as the probabilities fed into it, yet the machinery produces a crisp ranking regardless of whether those probabilities are well-founded frequencies, defensible beliefs, or wishful guesses. The formalism confers an aura of rigor on the output that the inputs may not deserve. In domains where probabilities are genuinely unknown rather than merely uncertain — deep ambiguity, one-off events, structural breaks — forcing a probability assignment so the expectation can be computed can be worse than admitting the problem is not yet an expected-utility problem at all.
T5: Risk attitude or value scale — the curvature is overloaded. The concavity of the value function is read as risk aversion, but the same curvature also expresses diminishing marginal value of the underlying good, and the two are not the same thing. An agent can have strongly diminishing marginal utility of wealth and still feel emotionally indifferent to risk, or be risk-averse over outcomes whose marginal value is constant. Expected utility folds both into a single function and cannot, on its own, separate them, which is part of why it struggles with phenomena — like the simultaneous purchase of insurance and lottery tickets — where risk attitude appears to vary with the domain rather than with the value scale.
T6: The benchmark legitimizes whatever it ranks. Because expected utility supplies the coordinate system in which rationality is defined, it tends to confer normative authority on the maximizing choice. But the framework is silent on whether the value function expresses good values or the probabilities express honest beliefs. A coherent maximizer can pursue a destructive end with impeccable consistency, and an analysis that certifies a choice as expected-utility-optimal can launder a bad goal or a self-serving probability assignment as "rational." The very neutrality that makes the pattern transfer across substrates also means it cannot, by itself, tell a well-aimed decision from a well-executed mistake.
Structural–Framed Character¶
Expected Utility sits toward the framed side of the structural–framed spectrum: it is the pattern of valuing an uncertain prospect by weighting the value of each possible outcome by its probability and summing, collapsing a distribution of futures into a single comparable scalar that ranks choices under risk. Its formal core — probability-weighted aggregation of a value function — is clean and structural.
What pulls it toward framing is the decision-theory vocabulary that comes along whenever it is used: "utility," "rational choice," and the apparatus of normative rationality. It carries evaluative weight as a criterion of rational choice — to violate it is to be charged with irrationality — and it presupposes a valuing agent, so an insurer pricing a policy or a gambler ranking bets imports the perspective of a chooser whose preferences the scalar encodes. The aggregation rule is structural; the valuing agent and the rationality standard supply the framing.
Substrate Independence¶
Expected Utility is a highly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its signature — probability-weighted aggregation of a nonlinear value function over outcomes — is fully formal and substrate-agnostic, and it carries real reasoning leverage across the formal-economic von Neumann-Morgenstern criterion, computational reinforcement learning and decision-theoretic planning, biological risk-sensitive foraging behaving as if maximizing fitness, and cognitive-medical decision analysis. Unlike a pure statistics technique it genuinely reaches biology. What keeps it a strong 4 rather than a 5 is that it has no native physical or social-institutional home.
- Composite substrate independence — 4 / 5
- Domain breadth — 4 / 5
- Structural abstraction — 5 / 5
- Transfer evidence — 4 / 5
Relationships to Other Primes¶
Parents (3) — more general patterns this builds on
-
Expected Utility presupposes Preference
Expected utility presupposes preference because the operation collapses an uncertain prospect into a single comparable scalar by probability-weighting the utility of each outcome — and that utility function is the agent's preference ordering over outcomes made cardinal under the von Neumann–Morgenstern axioms. Without preference as the underlying ordering on the choice set, there is no value function to weight, no ranking to maximize, and no meaning to "prefer the certain over the uncertain." Preference supplies the ordering primitive; expected utility supplies the aggregation rule that turns it into choice under risk.
-
Expected Utility is part of Probability
Probability supplies the apparatus of numerical assignments to events obeying coherence rules and supporting reasoning under incomplete information. Expected utility is one of the principal operations built on that apparatus: the probability-weighted summation of a utility function across possible outcomes that collapses a distribution into a single comparable scalar. It is a constituent piece of probability's broader decision-theoretic vocabulary, contributing the specific construction that turns probability distributions over outcomes into rankable choice objects via expectation of a nonlinear value function.
-
Expected Utility is a decomposition of Aggregation
Expected utility is the probability-weighted particularization of aggregation: a distribution of possible outcomes is collapsed into a single comparable scalar by weighting each outcome's utility by its probability and summing. Where aggregation names the deliberate suppression of detail to retain chosen features generally, expected utility specifies that the feature retained is the probability-weighted utility of outcomes while the detail suppressed is the distribution's shape beyond that summary — a particular choice of aggregation function tailored to ranking risky prospects.
Children (1) — more specific cases that build on this
-
Risk Aversion presupposes Expected Utility
Risk aversion presupposes expected utility because the property of preferring a sure outcome to an uncertain prospect of equal expected value is formally captured as concavity of the utility function inside the expected-utility integral. The Arrow-Pratt measures of risk aversion are derivatives of the utility function whose probability-weighted aggregation defines expected utility. Without expected utility's probability-weighted-aggregation machinery, there is no formal way to compare a gamble's certainty-equivalent against its expected value, and no quantitative apparatus for the risk premium that risk aversion describes.
Path to root: Expected Utility → Preference
Neighborhood in Abstraction Space¶
Expected Utility sits among the more crowded primes in the catalog (21st percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.
Family — Risk, Arbitrage & Tail Events (14 primes)
Nearest neighbors
- Risk — 0.85
- Risk–Return Tradeoff — 0.82
- Bias — 0.81
- Learning Curve Effects — 0.81
- Arbitrage (Finance) — 0.80
Computed from structural-signature embeddings · 2026-05-29
Not to Be Confused With¶
Expected utility must be distinguished from Marginal Utility, with which it shares the word "utility" and a concern with how value scales but which addresses an entirely different question. Marginal utility is the change in total utility produced by one more unit of a good — the worth of the next bite, the next dollar, the next hour of leisure — and its defining content is the typically diminishing rate at which value accrues as quantity rises. It is a statement about a value function under certainty: given that you have ten units, how much does the eleventh add? Expected utility, by contrast, says nothing in itself about quantity and everything about uncertainty: it aggregates a value function across mutually exclusive uncertain outcomes using probabilities, collapsing a distribution into a scalar. The two interlock without coinciding, and the interlocking is precisely what makes the pair confusable: the concavity that is diminishing marginal utility is the same concavity that, inside the expectation, produces risk aversion. But the direction of explanation differs. Marginal utility explains why the value function bends; expected utility takes a bent value function as given and asks how to evaluate a gamble against it. An agent can reason about marginal utility with no uncertainty present at all (allocating a fixed budget across goods to equalize marginal utility per dollar), and an agent can reason about expected utility over outcomes whose marginal-utility structure is irrelevant (a one-shot bet on a binary event). The shared curvature is a bridge, not an identity: marginal utility is a property of the value function over quantities; expected utility is an operation — expectation — performed on that value function over probabilities.
Expected utility is also distinct from Optionality, the asymmetric value of holding a right without an obligation. Optionality describes a structural payoff shape — the holder captures the upside while a floor caps the downside — and its value arises from that asymmetry combined with uncertainty and the freedom to act only when favorable. Expected utility, by contrast, is the general scoring rule that evaluates any prospect, asymmetric or not: a symmetric coin flip, a left-skewed insurance contract, and a right-skewed option are all run through the same probability-weighted value-transform machine. The relationship is one of operator and special case rather than rivalry. Optionality is a particular shape of outcome distribution that tends to score well under expected utility for a risk-averse agent precisely because the value function penalizes the truncated downside lightly while rewarding the open upside; expected utility is the evaluation that assigns optionality its number. One can hold optionality without invoking expected utility (a structural observation about asymmetric payoffs), and one can apply expected utility to prospects with no optionality at all (a forced symmetric gamble). The confusion arises because optionality is often motivated by expected-utility reasoning — we value the option because its expected utility is high — but the concepts answer different questions: optionality names a payoff asymmetry that creates value, while expected utility names the rule by which the value of any payoff structure, asymmetric or symmetric, is computed and compared.
A third boundary worth marking, because it is the most common collapse, is the line between expected utility and plain expected value — the probability-weighted average of the untransformed outcomes. Where the value function is linear, expected utility is expected value and the agent is risk-neutral; the two diverge only under nonlinearity, and the entire purpose of the prime lives in that divergence. Expected value is silent about risk attitude because it applies no transform; expected utility makes risk attitude the curvature of the transform it applies. Treating the two as interchangeable — pricing a gamble by its expected money, or assuming a positive expected-value bet is always worth taking — is exactly the error that the concept of expected utility was introduced to correct, from Bernoulli's resolution of the St. Petersburg paradox onward. The neighbor relationship is therefore one of generalization: expected value is the degenerate, risk-neutral special case of expected utility, recovered when the value function is restricted to a straight line.
Solution Archetypes¶
No catalogued solution archetypes reference this prime yet.
Notes¶
The candidate carries an unsure drafting verdict, and the reason is worth recording for the curator. Expected utility sits near the line between a structural pattern — probability-weighted aggregation of a value function, which genuinely recurs across economics, computation, biology, and medicine — and a decision-theoretic criterion or method, a normative rule for how rational agents ought to choose. The substrate-independence assessment scores its structural abstraction high (5) and its composite a strong 4, on the grounds that the signature is fully formal and transfers with real reasoning leverage rather than as loose analogy. The open question for the catalog is whether it admits such normative-formal patterns as primes at all, or whether their prescriptive character (an ought, not merely an is) places them in a different class. The entry is drafted as a structural prime on the strength of the transfer evidence, but the verdict flag should remain until the curator rules on the broader policy.
A second note concerns the relationship to its descriptive successors. Expected utility functions in the modern literature less as the last word on choice than as the baseline against which deviations are measured. Prospect theory, rank-dependent utility, and ambiguity-averse models are all defined by how they bend the probability weighting or the value transform away from the expected-utility form. This gives the prime an unusual role: even where it is descriptively false, it remains the coordinate system in which the truth is expressed, which is itself an argument for its status as a structural fixture rather than a contingent theory.
A third note: the utility scale is unique only up to positive affine transformation, and the probabilities in the subjective version are the agent's own coherent beliefs. Neither component is an objective, interpersonally comparable quantity. Analysts who sum utilities across people, or who treat elicited utility numbers as a cardinal welfare measure, are importing assumptions the von Neumann–Morgenstern and Savage frameworks do not supply.
References¶
[1] von Neumann, J., & Morgenstern, O. (1944). Theory of Games and Economic Behavior. Princeton University Press. First rigorous axiomatization of expected utility: an agent whose preferences over risky prospects satisfy the consistency axioms behaves as if maximizing the expectation of a utility function — the representation-theorem (not psychological-mechanism) reading, the separation of likelihood from value, and the formal core that makes the operation substrate-neutral. ↩
[2] Bernoulli, Daniel. Hydrodynamica, sive de viribus et motibus fluidorum commentarii. Strasbourg: Joannis Reinholdi Dulseckeri, 1738. Establishes the principle of energy conservation in fluid flow: pressure and kinetic energy are inversely related. Bernoulli's equation (P + ½ρv² + ρgh = const along streamline) remains the foundation for steady, incompressible flow analysis across engineering and physics. ↩
[3] Savage, L. J. (1954). The Foundations of Statistics. Wiley. Establishes subjective expected utility: probabilities are the agent's own coherent degrees of belief rather than objective frequencies, extending the pattern to any decision under genuine uncertainty; supplies the scalar-aggregation move that renders contingencies directly rankable while remaining silent on the worth of the values or beliefs supplied. ↩
[4] Pratt, J. W. (1964). Risk aversion in the small and in the large. Econometrica, 32(1–2), 122–136. Derives the coefficient of absolute risk aversion from the curvature of the von Neumann–Morgenstern utility function and links concavity to the risk premium, grounding why risk-averse agents value variance-reducing moves such as diversification and insurance. ↩
[5] Arrow, K. J. (1965). Aspects of the Theory of Risk-Bearing. Yrjö Jahnsson Foundation. Introduces the coefficients of absolute and relative risk aversion from utility-function curvature (independently of and alongside Pratt), formalizing the treatment of risk aversion in choice under uncertainty. ↩
[6] Caraco, T., Martindale, S., & Whittam, T. S. (1980). An empirical demonstration of risk-sensitive foraging preferences. Animal Behaviour, 28(3), 820–830. Shows juncos choosing among variable food patches as if trading mean against variance, with energy-budget state (concave vs. convex effective fitness) predicting risk-averse or risk-prone choice — the canonical biological instance of the expected-utility template. ↩
[7] Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263–291. Foundational behavioral-economics result: outcomes are evaluated as gains and losses relative to a reference point rather than in absolute terms, with diminishing sensitivity and loss aversion — making the choice of baseline (and the contrast it creates with the treatment) constitutive of perceived value and decision behavior. ↩
[8] Allais, M. (1953). Le comportement de l'homme rationnel devant le risque: Critique des postulats et axiomes de l'école américaine. Econometrica, 21(4), 503–546. Presents the Allais paradox: systematic preference reversals that violate the expected-utility axioms, measured as a precise departure from the benchmark. ↩
[9] Ellsberg, D. (1961). Risk, ambiguity, and the Savage axioms. The Quarterly Journal of Economics, 75(4), 643–669. Demonstrates ambiguity aversion: choices over bets with unknown probabilities violate subjective expected utility, a precise deviation from the expected-utility/Savage baseline. ↩