Rationality: von Neumann-Morgenstern through behavioral

Start one level below the risk apparatus, with plain preference. A rational agent in the spare economic sense has preferences that are complete (any two options can be compared) and transitive (if you prefer $x$ to $y$ and $y$ to $z$, you prefer $x$ to $z$). That is the ordinal substrate — the same preference base the value thread follows in a different direction, into general equilibrium and the Arrow-Debreu model. The two threads share this floor and then diverge: value follows preferences into what things are worth; this thread follows them into how to choose when you do not know what you will get.

Von Neumann and Morgenstern took that floor and built risk onto it. Their objects of choice are not certain outcomes but lotteries — probability distributions over prizes. They asked: what does it take for an agent’s preferences over lotteries to be coherent? Four axioms — completeness, transitivity, continuity, and independence — and one theorem falls out. The theorem is the whole achievement, so read it carefully: an agent who satisfies the four axioms behaves as if assigning a number $u(x)$ to each prize and choosing the lottery with the highest expected value of those numbers.

There is one limit von Neumann-Morgenstern leaves open: their probabilities are objective, the known odds of a roulette wheel. Most real decisions run on odds nobody hands you. Leonard Savage closed that gap in 1954. From preferences over acts — choices whose payoff depends on an uncertain state of the world — plus a coherence axiom called the sure-thing principle, Savage derived both a utility function and a subjective probability measure. The beliefs come out of the choices, not the other way around. That single move turned decision theory into the foundation of Bayesian statistics and of modern econometrics, where it still sits.

Be precise about what this apparatus claims. It is a normative-procedural framework — here is what coherent choice under risk looks like — carried by a representation theorem. It makes no claim about cognition, computation, or what goes on in anyone’s head. That restraint is exactly what makes it testable: the axioms are sharp enough that you can check whether real choices obey them. The next stage is the story of what happened when someone did.

Walk the Allais choice slowly, because the impossibility is exact, not rhetorical. Strip the common 89% chance of $1 million out of both pairs — independence says doing so cannot flip your preference, since it is identical across the two options in each pair. Do the subtraction and Pair A becomes Pair B. So preferring (1) in A but (4) in B is a flat contradiction of the axiom. This is the certainty effect: the guaranteed million in Pair A pulls disproportionately hard precisely because it is certain, and the moment certainty is gone in Pair B that pull vanishes. The axiom says the shared piece should not matter. It does.

Eight years later, Daniel Ellsberg found a second crack — in a different axiom. Imagine an urn with 30 red balls and 60 that are some unknown mix of black and yellow. Most people will bet on red (known odds) over black (unknown odds); offered a different bet, they will then bet on “black or yellow” over “red or yellow.” Those two preferences cannot both come from any coherent set of subjective probabilities — they violate Savage’s sure-thing principle. What Ellsberg exposed is ambiguity aversion: people do not merely dislike bad odds, they dislike not knowing the odds, and they will pay to avoid that ignorance in a way Savage’s framework forbids. Allais attacked the objective-probability apparatus; Ellsberg attacked the subjective-probability foundation underneath it.

A third attack came from a direction the paradoxes did not touch. Herbert Simon, in 1955, granted the axioms entirely and asked a different question: even if your preferences are perfectly coherent, how would you ever find the optimum? The choice set is vast, the computation unbounded, the information incomplete. Real agents, Simon argued, do not optimize — they satisfice: they set an aspiration level, search until they find an option that clears it, and stop. This is not a failure of consistency; it is a refusal of the problem the apparatus poses. The axioms describe a calculation no finite mind could complete, so even a fully rational agent must be doing something else. Bounded rationality is an orthogonal refutation, and it is the intellectual root of a repair that does not arrive until Stage 4.

Prospect theory replaces the heart of the EU apparatus with two pieces. The first is the value function. Where expected utility values final wealth, prospect theory values changes — gains and losses measured from a reference point, usually the status quo. The function is concave over gains (the second $100 thrills less than the first), convex over losses, and — the signature property — steeper for losses than for gains. Losing $100 hurts roughly twice as much as winning $100 pleases. That asymmetry is loss aversion, and it explains a great deal of behavior the EU apparatus had to wave away.

The second piece is the probability-weighting function. People do not use probabilities at face value: they overweight small probabilities (which sells lottery tickets and insurance alike) and underweight moderate-to-large ones. This is the direct descriptive fix for the Stage 2 Allais violation. The certainty effect — the disproportionate pull of a sure thing — falls straight out of a weighting function that treats the jump from 99% to 100% as far larger than the jump from 10% to 11%. The paradox that broke independence becomes a prediction of the new apparatus.

But the descriptive program split internally almost as soon as it formed, and the split matters for the verdict. Kahneman and Tversky read the deviations as biases — systematic errors against the EU norm, the output of fast, heuristic thinking. Gerd Gigerenzer read the very same deviations as ecologically rational: fast-and-frugal heuristics that exploit the structure of real environments and often outperform weighted-additive models when information is scarce and time is short. On this reading the “bias” is not a defect against a norm — the norm is the wrong yardstick for the environment. This is an argument within descriptive economics about how to read the same data; the larger question of whether ecological fitness is “really” rationality is a cross-discipline matter this walkthrough leaves to its sibling.

Cumulative prospect theory (1992) closed the one gap that kept the 1979 version on the outside. The original probability-weighting function, applied outcome by outcome, could in rare constructions prefer a gamble that was strictly dominated — worse in every state — than another. Tversky and Kahneman replaced it with rank-dependent weighting, applied to the cumulative distribution rather than to single outcomes. The repair eliminated the dominance violation while keeping reference-dependence, loss aversion, and probability distortion intact. This is what made prospect theory respectable: it now satisfied the requirement expected utility had never abandoned. The rebel met the establishment’s one non-negotiable standard, and was let in.

Two further developments pushed the descriptive program back toward the apparatus rather than away from it. Neuroeconomics — the brain-imaging and single-neuron work of Glimcher, Camerer, and others — found measurable neural correlates of expected value and of reference-dependent valuation. The variables the apparatus had posited as abstractions turned out to have signatures you could record. This did not displace the models; it grounded them, connecting decision theory to mechanism for the first time.

The deepest reconnection runs through rational inattention. Christopher Sims modeled deviations from full optimization as the optimal response to a real constraint: attention and information processing are costly, so an agent who economizes on them is not failing to optimize — she is optimizing subject to a cost the old apparatus ignored. This is Herbert Simon’s bounded rationality, re-domesticated inside the optimization framework forty years later. The deviation becomes rational-under-constraint. The thread that began by attacking optimization ends by extending it.

So where does the apparatus actually sit? The settlement is a division of labor. Expected utility survives as the normative benchmark — the standard against which choices are judged coherent — and as the applied workhorse, the tractable model that most applied microeconomics still runs on. That survival is licensed by a specific methodological commitment: Milton Friedman’s “as-if” defense, the argument that a model earns its keep by predicting well, not by describing the mind accurately. Alongside it, prospect theory and bounded rationality are the accepted descriptive corrections for actual choice behavior. The field runs both, with a clean split between the job each does.