Decision Theory Decoded

A decision is a mapping from possible actions to possible outcomes under uncertainty. Decision theory is the study of how to make that mapping rational. The field contains exactly one core tension: the normative question (what should a rational agent do?) versus the descriptive question (what do actual humans do?). Expected value theory answers the first. Prospect theory answers the second. The practical question—what should you actually do when facing a real decision—requires synthesizing both, plus a set of heuristics and techniques that outperform either theory in specific domains. This essay decodes the entire structure.

Expected Value: The Starting Point

Expected value (EV) is the simplest decision framework: for each possible action, multiply each outcome's probability by its payoff, sum the products, and choose the action with the highest sum. EV = Σ(pᵢ × vᵢ), where pᵢ is the probability of outcome i and vᵢ is the value of that outcome. This is the foundation of actuarial science, gambling theory, and elementary microeconomics. It works when three conditions hold: probabilities are known or reliably estimable, payoffs are linear in value (doubling money means doubling utility), and the decision is repeated enough times for the law of large numbers to smooth variance.

EV fails catastrophically in single-shot decisions with extreme downside. The St. Petersburg paradox (Bernoulli, 1738) demonstrated this: a game where expected value is infinite but no rational person would pay more than a modest sum to play. The problem is that EV treats the difference between $0 and $1,000 as identical to the difference between $1,000,000 and $1,001,000. Psychologically and economically, these are not equivalent. The first thousand dollars prevents starvation. The last thousand is rounding error. EV is a useful first approximation that breaks down precisely where decisions matter most: when stakes are high, when outcomes are asymmetric, and when you only get one shot.

Expected Utility: Making Value Nonlinear

Daniel Bernoulli's solution (1738), formalized by von Neumann and Morgenstern (1944), was to replace value with utility—a concave function of wealth that captures diminishing marginal returns. The utility of an additional dollar decreases as total wealth increases. A person with $1,000 gains more utility from receiving $1,000 than a person with $1,000,000 does from the same windfall. Expected utility theory (EU) says: maximize Σ(pᵢ × u(vᵢ)), where u is a concave utility function.

This resolves the St. Petersburg paradox and explains risk aversion as rational behavior. If your utility function is concave (diminishing marginal utility), you will prefer a certain $500 over a 50/50 gamble for $0 or $1,000, even though the expected values are identical. This isn't irrational—it reflects the genuine asymmetry between gains and losses when you have finite resources and non-repeatable decisions. Risk aversion scales with stakes relative to wealth: the same person who is risk-neutral on a $10 bet is rationally risk-averse on a $100,000 bet. EU dominated decision theory for two centuries and remains the normative standard in economics and finance.

But EU makes predictions that systematic human behavior violates. People don't just have concave utility over wealth. They exhibit patterns that no smooth utility function can explain.

Prospect Theory: How Humans Actually Decide

Daniel Kahneman and Amos Tversky's prospect theory (1979) replaced EU with a descriptive model grounded in experimental observation. Three features define it. Reference dependence: people evaluate outcomes as gains or losses relative to a reference point (usually the status quo), not as final states of wealth. A $50,000 salary feels different depending on whether your previous salary was $40,000 or $60,000—same outcome, different reference frame, different psychological experience. Loss aversion: losses loom larger than equivalent gains. Empirically, the pain of losing $100 is roughly twice the pleasure of gaining $100. The loss aversion coefficient λ ≈ 2.0–2.5 across diverse experimental populations. This is not risk aversion—it is a distinct asymmetry in the value function itself. The S-shaped value function: concave for gains (diminishing sensitivity to larger gains), convex for losses (diminishing sensitivity to larger losses). The function is steeper for losses than for gains. This generates the fourfold pattern: risk aversion for high-probability gains, risk seeking for low-probability gains (hence lottery tickets), risk seeking for high-probability losses (hence doubling down on losing investments), and risk aversion for low-probability losses (hence insurance against rare catastrophes).

Additionally, prospect theory replaces probabilities with decision weights: people overweight small probabilities and underweight large ones. A 1% chance of catastrophe gets treated as though it were 5–10%. A 95% chance of success gets treated as though it were 80–85%. This probability distortion, combined with loss aversion and reference dependence, explains a vast range of empirical "irrationalities"—the endowment effect, sunk cost persistence, status quo bias, the equity premium puzzle—that EU cannot accommodate.

Decision Under Uncertainty: Beyond Risk

Frank Knight (1921) distinguished risk (known probability distributions) from uncertainty (unknown distributions). Most consequential real-world decisions involve Knightian uncertainty—you don't know the probabilities, and you may not even know the full set of possible outcomes. EU assumes known probabilities. When probabilities are unknown, expected utility calculation becomes impossible.

Daniel Ellsberg (1961) demonstrated this with a simple experiment. Two urns: Urn A contains 50 red and 50 black balls. Urn B contains 100 balls, red and black in unknown proportions. Most people prefer betting on Urn A—known risk over unknown risk—even though no probability assignment to Urn B makes this preference consistent with EU. This is ambiguity aversion: a preference for known risks over unknown ones, independent of risk aversion. It is rational in environments where adversarial dynamics or model uncertainty make worst-case reasoning prudent. It is irrational if you accept EU's axioms. The tension between these positions remains unresolved in formal decision theory.

Practical implication: in domains of genuine uncertainty (geopolitics, entrepreneurship, novel technology, personal life decisions), EU-based optimization is not just imprecise—it requires inputs (probabilities, exhaustive outcome enumeration) that don't exist. Alternative frameworks are needed.

Heuristics That Work: Less Can Be More

Gerd Gigerenzer and the ABC Research Group (1999) challenged the assumption that more information and computation always produce better decisions. Their research program on "fast and frugal heuristics" demonstrated that simple decision rules often outperform complex optimization in uncertain environments—not despite their simplicity but because of it. Complex models overfit to noise in the training data. Simple heuristics, by ignoring most available information, achieve robustness.

The recognition heuristic: if you recognize one option but not the other, infer the recognized option has higher value on the criterion of interest. In experiments, less-knowledgeable decision makers using recognition alone outperformed more-knowledgeable ones who tried to integrate multiple cues—because the additional information introduced noise without proportional signal. Satisficing (Herbert Simon, 1956): set an aspiration threshold; search sequentially; choose the first option that exceeds the threshold. Satisficing outperforms optimization when search costs are non-trivial, options arrive sequentially, and the environment is uncertain. The optimal stopping problem (the "secretary problem") shows that even mathematically, the optimal strategy involves setting a threshold after sampling ~37% of options, not evaluating all of them. The 1/N rule: allocate resources equally across N options. In portfolio allocation, 1/N outperformed mean-variance optimization (Markowitz's Nobel Prize-winning framework) across the majority of empirical datasets tested by DeMiguel, Garlappi, and Uppal (2009). The reason: mean-variance optimization requires estimating expected returns and covariances, which introduces estimation error that swamps the theoretical gains from optimization.

The meta-principle: match the complexity of your decision strategy to the uncertainty of your environment. In well-characterized, low-noise domains (actuarial science, blackjack), optimize. In uncertain, high-noise domains (venture capital, hiring, life decisions), use simple heuristics.

Pre-Mortem and Reversibility

Gary Klein's pre-mortem technique (1998) is the single highest-impact practical intervention in decision quality. The procedure: before executing a decision, assume the decision has failed spectacularly, then generate explanations for why. This exploits prospective hindsight—the documented cognitive finding that people generate 30% more reasons for an outcome when told it has already occurred than when asked to predict whether it will. The pre-mortem circumvents two failure modes that plague group decision-making: groupthink (suppression of dissent after a decision direction is established) and overconfidence (systematic underestimation of failure probability).

The reversibility heuristic provides a complementary framework: classify decisions as reversible (Type 2, in Jeff Bezos's terminology) or irreversible (Type 1). Reversible decisions should be made quickly with minimal analysis—the cost of being wrong is low because you can correct course. Irreversible decisions warrant extensive deliberation, pre-mortems, and scenario analysis. Most organizations over-analyze reversible decisions and under-analyze irreversible ones. The asymmetry correction: speed up where the cost of reversal is low; slow down where it is high.

Decision Hygiene: Reducing Noise

Kahneman, Sibony, and Sunstein's Noise (2021) identified a category of decision error distinct from bias. Bias is systematic deviation in one direction. Noise is random scatter—variability in decisions that should be identical. When judges sentence the same crime profile differently based on time of day, temperature, or whether the local football team won, that's noise. When insurance adjusters produce wildly different valuations of the same claim, that's noise. Noise is at least as large as bias in most organizational decision-making and significantly harder to detect because it is invisible in individual cases—it appears only in statistical analysis of decision patterns.

Decision hygiene reduces noise without requiring diagnosis of specific biases. Key techniques: independent judgment aggregation—have multiple evaluators assess independently before discussion, then aggregate (structurally similar to the wisdom of crowds). Structured decision protocols—decompose complex judgments into independent components, evaluate each separately, then combine. Reference class forecasting—before estimating inside-view specifics, anchor on the base rate of the reference class ("what fraction of projects like this one succeed?"). Mediating assessments—replace holistic judgments with scores on predefined dimensions, reducing the influence of irrelevant case-specific features on the final assessment.

The Meta-Decision: How Much to Analyze

Every decision framework generates a recursive problem: how do you decide how much time and effort to spend deciding? This is the meta-decision, and it has a pragmatic resolution. The value of additional analysis is bounded by the value difference between the best available option and the option you would choose without analysis, multiplied by the probability that analysis changes your choice. For most decisions, this product is small—either because the options are close in value (so choosing "wrong" is low-cost) or because your intuitive judgment is already well-calibrated (so analysis rarely changes the outcome).

The practical heuristic: decisions involving irreversible consequences above a meaningful value threshold warrant deliberate analysis. Decisions that are reversible, low-stakes, or where your experience provides well-calibrated intuition should be made fast. Time spent analyzing is itself a cost—opportunity cost of action not taken, cognitive depletion, and the very real possibility that more information introduces noise rather than signal. The mark of a good decision-maker is not thoroughness on every decision—it's accurate triage of which decisions deserve thoroughness and which deserve speed.

How I Decoded This

Synthesized the normative tradition (von Neumann–Morgenstern expected utility theory) with the behavioral tradition (Kahneman–Tversky prospect theory) and the ecological rationality program (Gigerenzer's fast-and-frugal heuristics). Knight's risk-uncertainty distinction and Ellsberg's ambiguity aversion experiments provided the framework for decisions under genuine uncertainty. Klein's pre-mortem technique and the reversibility heuristic represent the highest-impact practical interventions identified across the applied decision science literature. Kahneman, Sibony, and Sunstein's noise framework addresses the systematic variability that contaminates organizational decisions independent of bias. The meta-decision framework follows from the observation that decision effort is itself a resource with diminishing marginal returns—the same structural insight that drives satisficing theory. The core pattern: normative theory tells you what perfect rationality looks like, behavioral research tells you where humans systematically deviate, and ecological rationality tells you when simple rules outperform complex optimization. A complete decision framework requires all three.

— Decoded by DECODER.