Decision theory

• 1. Introduction

  ›

  Decision theory asks a deceptively simple question: when you face uncertainty and must choose, what is the rational thing to do? In 1670, Blaise Pascal published his Pensees, which contained the earliest famous application of expected value reasoning. Pascal argued that when multiple actions each lead to different outcomes with different probabilities, the rational move is to calculate the average expectation for each outcome and pick the action with the highest total. That idea, born in the 17th century from thinking about gambling and wagers, eventually grew into a field that touches economics, philosophy, cognitive science, criminal justice, and political science.

• 2. Probability and Expected Utility

  ›

  Blaise Pascal and Pierre de Fermat laid the groundwork in the 17th century by developing probability theory, and Christiaan Huygens refined their framework further. That work gave later thinkers a language for measuring risk, which turned out to be the central problem of rational choice.

  In 1738, Daniel Bernoulli published a paper titled Exposition of a New Theory on the Measurement of Risk. He used a puzzle known as the St. Petersburg paradox to argue that simply maximizing expected monetary value was normatively wrong. To illustrate his alternative, Bernoulli described a Dutch merchant deciding whether to insure a cargo sailing from Amsterdam to St. Petersburg in winter. The merchant's decision could not be reduced to raw financial probabilities. Bernoulli proposed instead that people should maximize expected utility, which is a measure of subjective value rather than raw monetary gain.

  John von Neumann and Oskar Morgenstern formalized that intuition in the 1940s. Their work on game theory and expected utility theory established axiomatic foundations for rational decision-making under uncertainty. Shortly after, in 1950, E. L. Lehmann introduced the phrase "decision theory" itself as a label for this growing body of work.

• 3. The Subjective Turn

  ›

  Abraham Wald's 1939 paper made a striking observation: the two core procedures of classical statistics, hypothesis testing and parameter estimation, are actually special cases of a general decision problem. His paper synthesized concepts including loss functions, risk functions, admissible decision rules, and minimax procedures. That framing pulled statistics and decision theory into closer contact.

  Frank Ramsey, Bruno de Finetti, and Leonard Savage extended the scope of expected utility theory by introducing subjective probability. Their argument was that rational decision-making did not require objectively known probabilities. A person could act on personal degrees of belief, provided those beliefs were internally consistent. Bruno de Finetti developed what became known as the Dutch book paradoxes to show what goes wrong when someone's beliefs are incoherent: a clever opponent can construct a set of bets that guarantees a loss regardless of the outcome.

  Richard Threlkeld Cox provided a separate justification for the probability axioms, and the complete class theorems added another layer of support. Those theorems show that every admissible decision rule is equivalent to a Bayesian decision rule for some utility function and some prior distribution. In other words, any rule that cannot be beaten always has a Bayesian interpretation lurking behind it.

• 4. Cracks in the Rational Model

  ›

  Maurice Allais and Daniel Ellsberg both showed, through paradoxes bearing their names, that human behavior departs from expected-utility maximization in systematic ways. Their work pointed to something the axioms had not captured: real people are not interchangeable with the idealized agent the theory assumed.

  Daniel Kahneman and Amos Tversky pushed that finding further. Their research identified three regularities in actual human decision-making. Losses feel larger than equivalent gains. People focus more on changes in their situation than on absolute levels of wealth or welfare. And people's estimates of subjective probabilities are heavily distorted by a cognitive mechanism called anchoring, where an initial reference point skews subsequent judgments. Their prospect theory modified expected utility theory to account for these psychological factors, and it sparked what became known as behavioral economics.

  Amos Tversky also gave formal structure to an older idea about how people simplify hard choices. His elimination-by-aspects model describes a procedural framework in which a decision-maker rules out options one attribute at a time, rather than computing a full expected-utility score for every possibility. That model sits within descriptive decision theory, which aims not to prescribe but to document the consistent rules people actually seem to follow.

• 5. Time, Complexity, and Heuristics

  ›

  Intertemporal choice confronts a problem that purely expected-value calculations struggle with: outcomes arrive at different moments, and time itself changes how people value them. A windfall could be spent on an immediate pleasure or invested in a pension, but the optimal choice depends on expected interest rates, inflation, life expectancy, and trust in financial institutions. Even after accounting for all those factors, human behavior deviates sharply from what prescriptive models recommend.

  One response to that gap has been to replace objective interest rates with subjective discount rates, which capture how steeply individuals personally devalue future rewards. David Laibson's work on quasi-hyperbolic discounting gave a functional form to the tendency people have to be far more impatient in the short run than in the long run.

  When decisions involve not one person but many, a further complication enters. The emerging field of socio-cognitive engineering focuses specifically on distributed decision-making inside human organizations, including how decisions unfold in emergency and crisis situations where normal procedures break down. Individuals in those settings are limited in time and cognitive resources, which means they are, in the language of the field, boundedly rational. The problem is not only that real behavior departs from optimal behavior. The problem is that determining what optimal behavior even looks like is itself computationally hard.

  Heuristics are the mental shortcuts people use to sidestep that difficulty. They reduce the evaluative work required by focusing on some features of a decision while ignoring others. The gambler's fallacy illustrates the cost. A fair coin that has landed tails several times still has exactly a 0.5 probability of tails on the next flip. Yet people who rely on the heuristic that outcomes should balance out will predict that heads is "due," a prediction the math does not support. A related bias, the compromise effect, leads people to favor moderate options simply because those options share characteristics with both extremes, regardless of whether the moderate choice is actually better.

• 6. Beyond Probability

  ›

  Whether probability is the only legitimate tool for modeling uncertainty is a question decision theorists have not settled. Proponents of fuzzy logic, possibility theory, Dempster-Shafer theory, and info-gap decision theory argue that probability is one option among many. They point to cases where non-probabilistic rules, such as minimax, have been applied with apparent success. Minimax has a practical advantage: it does not require assumptions about the probabilities of events. It is robust in a way that probabilistic models sometimes are not.

  Nassim Nicholas Taleb articulated a related critique through the concept of the ludic fallacy. The argument is that decision theory based on a fixed universe of possibilities only handles the "known unknowns." It cannot address the "unknown unknowns," the events that fall entirely outside the model's frame. Significant events, Taleb's critics of standard decision theory contend, may be precisely those that no model anticipated. Unquestioning reliance on models, that line of argument holds, blinds analysts to the limits of the models themselves. Whether that critique calls for replacing probability or simply for applying it more carefully remains one of the open debates that keeps decision theory alive as a field.