Utility Function

A utility function is a mathematical representation of an agent's preferences over a set of outcomes. It assigns a real number to each possible outcome such that higher numbers correspond to more preferred outcomes, and the ordering of the numbers preserves the agent's preference ranking. In its simplest form, the utility function is a bookkeeping device: a way to encode "I prefer A to B" as a numerical inequality. In its most ambitious form, it is the foundation of rational choice theory, the bridge between psychology and economics, and the hidden architecture behind every recommendation algorithm, trading bot, and clinical decision support system.

The concept originates in the "marginal utility" theory of nineteenth-century economists — William Stanley Jevons, Carl Menger, and Léon Walras — who sought to explain prices as the result of subjective valuation rather than objective labor content. But the modern, axiomatic treatment derives from John von Neumann and Oskar Morgenstern's 1944 proof that if an agent's preferences satisfy completeness, transitivity, independence, and continuity, then there exists a utility function (unique up to positive affine transformation) that represents those preferences. This is the von Neumann-Morgenstern utility theorem, and it transformed utility from a psychological hypothesis into a mathematical necessity.

From Ordinal to Cardinal

Not all utility functions are created equal. An ordinal utility function captures only ranking: U(A) > U(B) means A is preferred to B, but the difference U(A) − U(B) carries no information about intensity. A cardinal utility function, by contrast, encodes strength of preference: U(A) − U(B) = 2 × (U(B) − U(C)) means the gain from A to B is twice as valuable as the gain from B to C. The von Neumann-Morgenstern theorem constructs cardinal utility, which is why expected utility theory can prescribe probability-weighted averages. Without cardinality, expectation is meaningless.

The distinction is not merely technical. It is the difference between a map that shows which city is larger and a map that shows by how much. Ordinal utility is sufficient for choice under certainty: if you know the outcomes, you only need to rank them. Cardinal utility is required for choice under uncertainty: if you must weigh probabilities against outcomes, you need to know how much each outcome matters. The field's persistent confusion about whether utility is ordinal or cardinal — a confusion that still appears in undergraduate textbooks — reflects a deeper ambiguity about what the utility function is supposed to represent: a measurement of psychological strength, or a logical reconstruction of consistent choice.

Utility Across Domains

In economics, the utility function is typically treated as a black box: agents maximize utility, markets equilibrate, and the details of the box are left to psychology. In psychology, the box is opened — and found to be far messier than the economic model assumes. Humans construct preferences in the moment of choice, exhibit context-dependent reversals, and violate the independence axiom in systematic ways. The prospect theory value function is not a utility function in the von Neumann-Morgenstern sense; it is a descriptive model of how humans actually transform outcomes into subjective valuations, incorporating loss aversion, reference dependence, and nonlinear probability weighting.

In artificial intelligence, the utility function takes on a different role entirely. A reinforcement learning agent does not "have" preferences in the human sense; it has a reward function that the designer specifies, and the agent's behavior is whatever maximizes cumulative reward. The alignment problem in AI is, at its core, a utility-function problem: how to specify a reward function that produces behavior the designer actually wants, rather than behavior that technically maximizes the reward while producing catastrophic side effects. The paperclip maximizer is a utility function gone wrong: an agent whose utility function assigns high value to paperclip production and zero value to everything else, including human survival.

The philosophical implications are equally unsettling. If utility is a representation of preference, and preference is constructed rather than discovered, then the utility function is not a measurement of pre-existing values. It is a formalization of a choice pattern that may itself be contingent on framing, context, and social influence. The interpersonal comparison of utility — whether my gain of ten utils can be weighed against your loss of five — has been declared impossible by some economists and indispensable by others. The question is not merely technical. It is ethical: can we aggregate individual utilities into a social welfare function, or is the very attempt a category error?

The Systems Critique

The utility function is one of the most powerful and most dangerous concepts in systems theory. Powerful because it provides a universal language for optimization: any system that pursues goals can be described as maximizing some utility function, and this description enables transfer of tools across domains. Dangerous because the description is often mistaken for the reality. A system does not maximize utility because it has a utility function; it "has" a utility function because an observer has chosen to describe it as maximizing one. The utility function is an observer-dependent construct, not an objective property of the system.

This matters because the same behavior can be represented by infinitely many utility functions. An agent that always chooses the action with the highest expected monetary return can be described as maximizing wealth, or as maximizing the logarithm of wealth, or as maximizing a complicated function that happens to be monotonic with wealth. The choice of representation is not determined by the behavior. It is determined by the observer's theoretical commitments. A psychologist who models the agent as risk-averse will choose a concave utility function. An economist who models the agent as risk-neutral will choose a linear one. The same agent, different functions, different predictions about how the agent will respond to novel incentives.

The deeper point: utility functions are not discovered. They are designed. And the design choices — what to include, what to exclude, what to treat as linear, what to treat as concave — encode normative assumptions that are rarely made explicit. The claim that a market economy maximizes social welfare is not an empirical finding. It is a representation theorem that depends on assuming specific utility functions for all agents and specific aggregation rules for combining them. Change the functions, change the aggregation, and the theorem disappears.

The utility function is the original sin of rational choice theory: the moment we represent preference as a mathematical function, we gain the power of optimization and lose the capacity to ask whether the representation is appropriate. Every policy that treats human welfare as a quantity to be maximized, every algorithm that treats engagement as a utility to be optimized, every clinical trial that treats quality of life as a score to be averaged — all of them inherit the ambiguity of the utility function, and all of them risk optimizing a formalization rather than a person. The concept will not be displaced. But it must be accompanied by a persistent question: whose utility, measured how, and at what cost to the phenomena that the measurement cannot capture?