> For the complete documentation index, see [llms.txt](https://laurence-wilse-samson.gitbook.io/textbooks/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://laurence-wilse-samson.gitbook.io/textbooks/philosophy-of-economics/part-iv-rationality-and-behavior/ch08_rational_choice.md).

# Chapter 8: Rational Choice Theory

***

> "The theory of games of strategy is not a psychological theory. It does not try to say how people *do* behave, but rather how they *should* behave if they want to win." — John von Neumann and Oskar Morgenstern, *Theory of Games and Economic Behavior* (1944)

***

Expected utility maximization is the foundation of modern economics. The rational agent—introduced in Chapter 1 as an entity with preferences, constraints, and an optimization procedure—finds its fullest expression in expected utility theory. Here preferences extend to uncertain prospects, and rationality acquires precise mathematical content through a set of axioms.

But what is this theory actually claiming? Is it a description of how people behave, a prescription for how they should behave, or merely a useful fiction for generating predictions? The answer matters enormously. If expected utility theory is descriptive, then systematic violations are anomalies requiring explanation. If it is normative, violations might be rational responses that the theory fails to accommodate. If it is merely a modeling convenience, the whole apparatus loses its grip on real choice behavior.

This chapter examines the foundations, interpretations, and limits of rational choice theory. We begin with the von Neumann-Morgenstern axioms that underpin expected utility, consider competing interpretations of what the theory claims, extend the analysis to strategic interaction in games, and confront experimental evidence that challenges the axioms.

***

## 8.1 The Axioms of Rational Choice

### From Ordinal to Cardinal

Chapter 1 introduced the basic requirements for rational preferences: completeness (any two alternatives can be compared) and transitivity (if A is preferred to B and B to C, then A is preferred to C). These conditions suffice for ordinal utility—a ranking of outcomes where only the order matters, not the magnitude of differences.

But economics requires more. Whenever decisions involve risk, we need to compare gambles: would you prefer a certain $50, or a 50% chance at $100? Answering such questions requires not just ranking outcomes but assessing how much better one outcome is than another. We need cardinal utility—preferences with meaningful intensity.

John von Neumann and Oskar Morgenstern provided the foundation in their monumental 1944 work, *Theory of Games and Economic Behavior*. They showed that if preferences over lotteries satisfy certain axioms, those preferences can be represented by an expected utility function: the agent ranks lotteries by their probability-weighted average utilities.

***

**Figure 8.1: The von Neumann-Morgenstern Axioms**

![ch08\_fig01\_vnm\_axioms](/files/or5gY2u3K7h388oTtzXR)

**Box 8.1: The vNM Axioms (Formal Statement)**

Let $$\mathcal{L}$$ denote the set of lotteries (probability distributions) over a finite set of outcomes $$X$$. A preference relation $$\succsim$$ on $$\mathcal{L}$$ satisfies the vNM axioms if:

1. **Completeness.** For any $$L, M \in \mathcal{L}$$: either $$L \succsim M$$ or $$M \succsim L$$ (or both).
2. **Transitivity.** For any $$L, M, N \in \mathcal{L}$$: if $$L \succsim M$$ and $$M \succsim N$$, then $$L \succsim N$$.
3. **Continuity (Archimedean axiom).** For any $$L, M, N \in \mathcal{L}$$ with $$L \succ M \succ N$$, there exist $$\alpha, \beta \in (0,1)$$ such that $$\alpha L + (1-\alpha)N \succ M \succ \beta L + (1-\beta)N$$.
4. **Independence.** For any $$L, M, N \in \mathcal{L}$$ and $$\alpha \in (0,1]$$: $$L \succsim M$$ if and only if $$\alpha L + (1-\alpha)N \succsim \alpha M + (1-\alpha)N$$.

**Representation Theorem (von Neumann and Morgenstern 1944).** $$\succsim$$ satisfies axioms 1–4 if and only if there exists a function $$u: X \rightarrow \mathbb{R}$$ such that for all lotteries $$L, M$$: $$L \succsim M \iff \sum\_{x} p\_L(x) \cdot u(x) \geq \sum\_{x} p\_M(x) \cdot u(x)$$ Moreover, $$u$$ is unique up to positive affine transformation: if $$u$$ represents $$\succsim$$, so does $$v = au + b$$ for any $$a > 0$$.

### Understanding the Axioms

Completeness and transitivity are familiar from ordinal theory. The new axioms are continuity and independence.

*Continuity* rules out lexicographic preferences—preferences where one dimension always dominates, regardless of how extreme the tradeoffs become. If you prefer life to death and death to being tortured, continuity says there exists some probability p such that you are indifferent between certain death and a lottery giving life with probability p and torture with probability (1-p). In other words, for any three outcomes ranked A > B > C, there is always some mixture of the best and worst that is exactly as good as the middle option. Put starkly: everything has its price.

This is controversial. Many people insist there are things they would never trade for any amount of money or any small risk. Sacred values, as psychologists call them, seem to violate continuity. Whether such absolutism is rational or merely a failure of imagination is contested.

*Independence* is the workhorse axiom, doing most of the theoretical work. It says that preferences between two lotteries should not change when both are mixed in the same proportion with a common third lottery. If you prefer $100 for certain over a 50/50 gamble between $200 and $0, you should maintain that preference even when both options are mixed with, say, a 10% chance of winning a car. The car is irrelevant to the comparison—it appears in both options equally.

Independence captures a kind of "consequentialist" reasoning: what matters in each lottery is just the probabilities of each outcome, not how those probabilities are generated or what other outcomes might have occurred. If you prefer A to B, then in a lottery that gives you A with some probability, you should also prefer to have a higher chance of getting A.

But independence is empirically fragile. As we shall see, the Allais paradox demonstrates systematic violations by reasonable people. Whether these violations are mistakes to be corrected or reasonable responses to features the axioms ignore remains disputed.

### The Representation Theorem

The power of the vNM framework lies in its representation theorem: any preferences satisfying the four axioms can be represented by expected utility maximization. The agent acts as if assigning utilities to outcomes and choosing the lottery with highest probability-weighted utility sum.

Crucially, the utility function is unique up to positive affine transformation. If u(x) represents preferences, so does v(x) = 2u(x) + 5. This means utility differences have meaning—we can say that the difference between outcomes A and B is twice as great as between C and D—but utility levels do not. There is no absolute zero of utility.

This is cardinal utility, but of a restricted kind. We can make ratio comparisons of utility *differences*, but not of utility *levels*. This suffices for expected utility calculation while avoiding claims about absolute magnitudes of welfare. Note the contrast with the ordinal utility of Chapter 2: ordinal utility is unique up to any *monotonic* transformation (any order-preserving relabeling), so only the ranking matters. VNM utility is unique up to *positive affine* transformation ($$v = au + b$$, $$a > 0$$), which preserves both ranking and the ratios of differences. The additional structure is precisely what makes expected utility calculation possible—and it is precisely what the independence axiom buys.

### Savage's Extension: Subjective Expected Utility

The vNM framework assumes that probabilities are given—the coin is known to be fair, the lottery has stated odds. But many economic decisions involve uncertainty where probabilities are not objectively given. What is the probability that a new product will succeed? That a currency will depreciate?

Leonard Savage's *The Foundations of Statistics* (1954) extended expected utility to this domain. Savage showed that if an agent's preferences over acts (functions from states of the world to outcomes) satisfy a richer set of axioms, then the agent acts *as if* she holds a unique subjective probability distribution over states and maximizes expected utility with respect to it. The agent's behavior simultaneously reveals both her utility function and her probability beliefs.

This is a remarkable result. It means that coherent preferences under uncertainty imply probabilistic beliefs, even when no objective probabilities exist. The agent need not consciously assign probabilities; the structure of her choices implies them. Savage's framework thus provides the theoretical foundation for Bayesian decision theory: all uncertainty is reducible to subjective probability, and all rational choice is expected utility maximization.

But Savage's axioms are demanding. His "sure-thing principle"—a generalization of vNM independence—requires that preferences between acts depend only on states where the acts differ. And his framework assumes that agents *can* assign probabilities to all events, however uncertain. The Ellsberg paradox (§8.5) directly challenges this assumption, showing that people systematically distinguish situations with known probabilities from those with unknown probabilities—precisely the distinction between risk and Knightian uncertainty explored in Chapter 15.

***

## 8.2 Interpretations of Expected Utility

{% tabs %}
{% tab title="Descriptive" %}
**How People Actually Decide**

* Claims that expected utility theory describes real decision-making processes
* Homo economicus really does calculate expected utilities and choose accordingly
* Invites empirical testing: do choices satisfy the axioms?
* Problem: Systematic violations documented since Allais (1953)
* Defense: Violations may be artifacts of laboratory settings or errors correctable by reflection
  {% endtab %}

{% tab title="Normative" %}
**How People Should Decide**

* Expected utility prescribes how rational agents *should* behave
* Violations are mistakes—failures of rationality to be corrected
* Requires defending axioms as requirements of rationality
* Transitivity: Strong support via "money pump" arguments
* Independence: Most contested—Allais violators often do not feel irrational
  {% endtab %}

{% tab title="As-If" %}
**Predictive Instrument**

* Agents behave *as if* maximizing expected utility, regardless of mental processes
* Sidesteps debates about cognition while retaining predictive power
* Allows diverse mechanisms (habit, heuristics, learning) to produce EU-consistent behavior
* Problem: When does the approximation hold? Novel situations may deviate substantially
* Risk: Theory becomes unfalsifiable if any violation is dismissed as "outside the domain"
  {% endtab %}
  {% endtabs %}

### Descriptive Interpretation

The most ambitious interpretation holds that expected utility theory describes how people actually make decisions under risk. Homo economicus really does calculate expected utilities and choose accordingly.

This interpretation invites empirical testing. If people are expected utility maximizers, their choices should satisfy the axioms. Researchers can present subjects with carefully constructed choice problems and check for consistency.

The results are not encouraging. Beginning with Allais (1953) and continuing through decades of experimental economics and psychology, researchers have documented systematic violations of the axioms—particularly independence. People do not behave like expected utility maximizers.

Defenders of the descriptive interpretation offer several responses. Perhaps violations occur only in artificial laboratory settings, not in high-stakes real-world decisions where people think carefully. Perhaps experience and feedback lead people to converge on EU-consistent behavior. Perhaps the violations are errors that people would correct upon reflection.

These defenses have merit but are not fully convincing. Some violations persist in high-stakes settings. Experience sometimes increases rather than decreases violations. And whether reflection changes choices depends on how the choice is framed—itself a violation of EU's predictions.

### Normative Interpretation

A more defensible interpretation holds that expected utility theory prescribes how rational agents *should* behave. It is a normative standard against which actual behavior can be evaluated. Violations are mistakes—failures of rationality that people would ideally avoid.

This interpretation requires defense of the axioms as requirements of rationality. Why should rational agents have complete, transitive, continuous, independent preferences?

*Completeness* is perhaps hardest to defend. Why must a rational agent be able to compare any two alternatives? Some choices might be genuinely incomparable—not merely close, but incommensurable. The world may not force us to rank everything, and rationality may not require it.

*Transitivity* has stronger intuitive support. Intransitive preferences seem to license self-defeating behavior: if you prefer A to B, B to C, and C to A, a clever trader can extract unlimited money from you through a sequence of trades you accept at each step. This "money pump" argument provides practical grounds for transitivity.

*Independence* is most contested. Many people, when confronted with their independence-violating choices, do not feel they have made a mistake. The Allais paradox (discussed below) seems to reveal reasonable risk-sensitivity that independence rules out. If violating independence is not obviously irrational, the normative interpretation weakens.

### The "As If" Interpretation

Milton Friedman's influential methodology suggests a third approach: expected utility theory need not be literally true to be useful. Agents behave *as if* they maximize expected utility, even if they do not consciously calculate utilities and probabilities. The theory is a predictive instrument, not a description of mental processes.

This instrumentalist interpretation has attractive features. It sidesteps debates about what people "really" think while retaining EU's predictive power. It allows that the same behavior might be produced by various psychological mechanisms—habit, heuristics, learning—while modeling it as if produced by optimization.

But the as-if interpretation faces challenges. If the theory is merely predictive, its domain of applicability becomes crucial: when does the as-if approximation hold? In novel situations, extreme stakes, or complex decisions, people might deviate substantially from EU predictions. Without understanding the actual mechanisms producing behavior, we cannot know when the approximation fails.

Moreover, the as-if defense threatens to make the theory untestable. Any violation can be dismissed as outside the theory's domain; any confirmation can be cited as evidence for the approximation. Scientific theories should make risky predictions, not insulate themselves from falsification.

### Dutch Book Arguments

A different approach grounds expected utility in pragmatic considerations. Frank Ramsey and Bruno de Finetti showed that agents with non-EU-consistent preferences can be exploited through "Dutch books"—sets of bets that guarantee a loss.

Suppose you violate the probability laws implied by EU. Perhaps you assign probability 0.6 to rain and 0.7 to no rain, summing to more than 1. A cunning bookie can construct a set of bets, each of which you find acceptable, but which together guarantee your loss regardless of whether it rains.

Similarly, violations of the vNM axioms can generate dynamic inconsistency—preference reversals over time that make you worse off by your own standards. Independence violations can lead you to make choices at sequential decision nodes that you would have rejected if you had considered the entire decision tree at the outset.

Dutch book arguments establish a tight connection between EU and coherence. But they have limitations. The exploitation requires a bookie with perfect knowledge of your preferences and the ability to offer arbitrary bets. In the real world, such exploitation may be rare. And even if incoherent preferences create vulnerability in principle, it does not follow that maximizing expected utility is the only way to avoid exploitation.

***

## 8.3 Game-Theoretic Rationality

### From Individual to Strategic Choice

Expected utility theory analyzes individual choice under risk—decisions whose outcomes depend on chance, not on the choices of other agents. But much economic behavior is strategic: the outcome of my choice depends on your choice, and vice versa. Game theory extends rational choice to strategic interaction.

The conceptual framework shifts in important ways. In individual decision theory, uncertainty is about states of nature that do not respond to my choices. In game theory, I face uncertainty about your choices, but your choices themselves depend on your beliefs about my choices. The interdependence creates conceptual complications that plague game theory to this day.

### Nash Equilibrium

The central solution concept in game theory is Nash equilibrium: a profile of strategies, one for each player, such that no player can improve their outcome by unilaterally changing their strategy. Each player is playing a best response to the strategies of others.

***

**Figure 8.2: The Prisoner's Dilemma**

![ch08\_fig02\_prisoners\_dilemma](/files/9zqBXdK5IIWQdElDvNLC)

Nash equilibrium has proved remarkably fruitful as a modeling tool. It organizes analysis of strategic situations from oligopoly competition to arms races to evolutionary biology. But its status as a solution concept rests on assumptions about rationality that are more demanding than they first appear.

### The Epistemic Foundations Problem

For Nash equilibrium to be compelling, players must believe that they are in equilibrium. But what justifies this belief?

The standard story invokes *common knowledge of rationality*: each player is rational, each player knows that each player is rational, each player knows that each player knows that each player is rational, and so on ad infinitum. Given common knowledge of rationality, each player can deduce what others will do, leading (in some games) to Nash equilibrium.

But this justification is problematic. Common knowledge of rationality is a stringent requirement—stronger than any real-world situation can guarantee. And in many games, common knowledge of rationality does not uniquely determine Nash equilibrium: there may be multiple equilibria, or none that common knowledge alone selects.

Alternative justifications appeal to learning, evolution, or focal points. Perhaps players converge on equilibrium through repeated interaction, or equilibrium strategies spread through evolutionary selection, or shared cultural knowledge makes certain equilibria salient. These justifications are more realistic but less tightly connected to rationality per se.

The epistemic foundations of game theory remain contested. Nash equilibrium is a useful tool, but its relationship to rational choice is less straightforward than often assumed.

### The Equilibrium Selection Problem

Many games have multiple Nash equilibria. Consider the coordination game "Battle of the Sexes": two players prefer to meet at the same location, but disagree about which location is better. Both meeting at location A and both meeting at location B are Nash equilibria. So is a mixed strategy equilibrium where each randomizes. Common knowledge of rationality does not select among them.

This creates a philosophical puzzle. If rational players can reach *any* equilibrium, what does game theory actually predict? The answer must appeal to something beyond individual rationality—focal points (Schelling 1960), social conventions, communication, or shared history. But these are precisely the social and institutional factors that microeconomic theory typically treats as exogenous.

### Mixed Strategy Interpretation

When no pure-strategy equilibrium exists (as in Matching Pennies), Nash equilibrium requires mixed strategies: players randomize over actions with specific probabilities. But what does it mean for a rational agent to randomize deliberately?

Three interpretations have been proposed:

1. **Deliberate randomization**: The agent consciously flips coins or uses random devices. But this seems psychologically implausible and strategically questionable—why would an agent *want* to be unpredictable?
2. **Epistemic interpretation**: The probabilities represent opponents' *beliefs* about what I will do, not actual randomization. My equilibrium strategy is the one that makes my opponent indifferent, which requires them to *believe* I'm mixing with the right probabilities. This interpretation raises questions about how such beliefs are coordinated.
3. **Population interpretation**: Each "player" is actually a large population, and the probabilities represent the fraction of the population playing each strategy. Evolution or learning drives population frequencies toward equilibrium. This interpretation abandons individual rationality as the foundation.

Each interpretation raises distinct philosophical issues. The choice among them affects what game theory claims to explain: individual psychology, rational belief, or population dynamics.

### Game Theory as Explanation

What does game theory *explain*? Consider the Prisoner's Dilemma. Game theory predicts that rational players defect. Does this explain *why* people defect in prisoner's-dilemma-like situations?

The explanatory claim is ambiguous. It might mean:

* **Psychological explanation**: People reason through the logic and conclude they should defect. The prediction explains by describing the mental process.
* **Optimization explanation**: People choose optimally given incentives, regardless of whether they consciously reason through game theory. The theory captures the logic of the situation.
* **Equilibrium explanation**: Defection is stable because it is a best response to defection. The explanation is about *stability*, not process.

These differ in their empirical commitments. Psychological explanation requires that people actually think strategically. Optimization explanation requires that behavior matches what strategic thinking would produce (as-if reasoning). Equilibrium explanation requires only that we observe stable patterns—which might arise from evolution, learning, or institutional constraints rather than individual calculation.

The distinction matters because it affects how we interpret cooperation in prisoner's dilemmas. If explanation is psychological, cooperation is *irrational*—a failure of reasoning. If explanation is equilibrium-based, cooperation might be rational in repeated games, or in games with different informational assumptions, or when players have other-regarding preferences. Game theory becomes a framework for characterizing situations, not a universal prediction of defection.

***

## 8.4 Revealed Preference and Behavioral Consistency

### Samuelson's Program

Paul Samuelson's revealed preference theory, introduced briefly in Chapter 1, attempted to place consumer theory on behaviorist foundations. Rather than starting with unobservable preferences and deriving choice behavior, Samuelson proposed to define rationality directly in terms of choice consistency.

The key axiom is the *Weak Axiom of Revealed Preference* (WARP): if bundle x is chosen when y is affordable, then y should never be chosen when x is affordable and no cheaper. More formally: if x is revealed preferred to y (x chosen when y was available), y cannot be revealed preferred to x.

WARP captures a minimal consistency requirement. Strengthened versions—the Strong Axiom (SARP) and Generalized Axiom (GARP)—provide necessary and sufficient conditions for the existence of a utility function rationalizing observed choices.

It is important to distinguish GARP from expected utility. GARP describes *consistency patterns in market behavior*—how choices respond to budget constraints and prices. It is silent on probability, risk, or the psychological process of decision-making. Expected utility, by contrast, is a *normative theory of coherent choice under uncertainty*—it specifies how a rational agent should evaluate risky prospects. GARP can be satisfied by agents who violate EU (e.g., loss-averse agents in riskless choice), and EU-conforming agents can violate GARP if their choices are inconsistent across contexts. The two frameworks address different questions: GARP asks whether choice patterns are internally consistent; EU asks whether attitudes toward risk are coherent.

The revealed preference approach has significant attractions. It avoids introspective reports about mental states. It provides a clear operational criterion for rationality: consistency of choice behavior. It grounds economics in observable data—prices, quantities, choices—rather than psychological speculation.

### Sen's Critique: The Poverty of Consistency

Amartya Sen mounted an influential challenge to the revealed preference program in his 1993 paper "Internal Consistency of Choice." Sen's basic point is that consistency alone cannot capture rationality—we need to know *what* the agent is trying to achieve.

Consider an example. You face a choice between two bundles: one containing apples and oranges, another containing only apples. You choose the second bundle. Later, you face a choice between the second bundle and a third containing only oranges. You choose the third.

This violates WARP: the second bundle was revealed preferred to the first (which contained all the options in the third), yet the third was chosen over the second. But is this irrational? Not if your goal was to choose the smallest bundle, or to avoid variety, or to select whatever comes last alphabetically. Consistency depends on what you're trying to do.

Sen's critique exposes a tension in the revealed preference program. On one hand, revealed preference was meant to avoid claims about mental states—we observe choices, not preferences. On the other hand, the consistency conditions make sense only if we assume the agent is trying to maximize *something*. WARP assumes that choices reveal an underlying preference ordering that the agent is maximizing. But if we grant that assumption, we have reintroduced the mental states we sought to avoid.

The deeper point is that rationality is not just formal consistency but appropriate response to reasons. An agent who consistently pursues a goal is rational only if the goal itself makes sense. This is rationality of ends as well as means—something the revealed preference framework cannot capture.

### The Neo-Samuelsonian Response

Sen's critique targets what we might call the *psychological* reading of revealed preference: the view that choices reveal mental states called preferences, which the agent seeks to satisfy. On this reading, WARP and GARP are consistency conditions on the mind—requirements that the psychological entity "preference" be well-ordered.

But there is another interpretation. Don Ross and others have developed what might be called a *neo-Samuelsonian* position: revealed preference theory characterizes patterns in *market behavior*, not cognitive processes. On this view, GARP is not a claim about what happens inside the agent's head; it is a description of how choice behavior relates to changes in incentives within institutional settings.

The key move is to distinguish *choice behavior* from *cognitive deliberation*. Ross defines choice behavior as "any behavior that is systematically (but typically stochastically) related to changes in incentives." This is broader than conscious decision-making. It includes habitual responses, institutionally channeled actions, and behaviors whose causes the agent cannot articulate. What makes something a choice, on this view, is not that it emerges from mental optimization but that it responds systematically to prices and constraints.

This reframing changes how we interpret the axioms. GARP does not require that agents have consistent mental states. It requires that their behavior, in market contexts, exhibit certain patterns. These patterns might emerge from explicit calculation, but they might equally arise from learning, habit, social pressure, or institutional channeling. The axioms describe the output, not the process.

Consider an analogy. We might observe that a thermostat "chooses" temperature settings in ways that satisfy consistency conditions—it never "prefers" 70°F to 68°F in one context while preferring 68°F to 70°F in another with the same constraints. This consistency does not reveal anything about the thermostat's mental states (it has none). It reveals the structure of its response to environmental inputs.

Neo-Samuelsonians suggest something similar for human choice in markets. The consistency is real, but it need not be grounded in consistent mental preferences. It may be grounded in the institutional structure of markets, which channels behavior in GARP-consistent ways. Vernon Smith's experimental economics has repeatedly shown that markets generate efficient, consistent outcomes even when individual participants are cognitively unsophisticated. The rationality, in such cases, is in the institution, not the individual.

This interpretation sidesteps Sen's critique in a particular way. Sen objected that consistency is meaningless without knowing what the agent is trying to achieve. The neo-Samuelsonian response: we need not know what the agent is "trying" to achieve in any psychological sense. GARP describes how market behavior responds to incentives, regardless of the mental processes—if any—that produce it.

But this interpretation also has costs. If revealed preference does not reveal psychological preferences, what does welfare economics measure? Standard welfare analysis uses revealed preference to infer what makes people better off. If choices reveal only market behavior patterns, not underlying well-being, the normative authority of choice-based welfare analysis is undermined. We return to this problem in Chapter 3.

The neo-Samuelsonian position also changes what counts as a "violation" of rationality. If GARP characterizes market behavior, then violations of GARP in laboratory settings—where there are no market institutions—may simply show that the axioms describe market behavior, not behavior in general. The biases documented by behavioral economists might be artifacts of studying individuals stripped of the institutional scaffolding that normally structures their choices. This connects to the "ecological rationality" discussed in Chapter 9.

### What Does Revealed Preference Reveal?

The revealed preference framework faces difficulties in specifying what exactly choice reveals. Consider several complications:

*Context effects*: Choices depend on how alternatives are presented. Adding a dominated option can change the choice between remaining options—a violation of basic revealed preference logic. If you prefer A to B when those are the only options, but B to A when a clearly inferior option C is added, what are your "true" preferences?

*Menu dependence*: People value having options. The same outcome may be valued differently depending on what alternatives were available. Choosing an option from a large menu may feel different—and be valued differently—than having no choice. Revealed preference cannot capture this.

*Framing effects*: Equivalent descriptions of the same choice can elicit different responses. If a medical treatment is described as having a "90% survival rate" versus a "10% mortality rate," people respond differently, though the information is identical. Which response reveals the true preference?

These phenomena suggest that preferences are not simply "revealed" by choices but are partly *constructed* in the act of choosing. If so, the revealed preference program's behaviorist aspirations are undermined. There may be no fact of the matter about preferences independent of the specific choice situation.

The neo-Samuelsonian interpretation offers a different response to these challenges. Context effects, menu dependence, and framing effects are problems for the *psychological* reading of revealed preference—they show that choice does not reliably reveal stable mental states. But on the market-behavior reading, these effects are simply facts about how behavior varies across contexts. The question becomes: in which contexts does GARP-consistent behavior emerge?

The answer, according to neo-Samuelsonians, is: in market contexts with appropriate institutional structure. Laboratory experiments that elicit framing effects typically strip away the price signals, competitive pressures, and learning opportunities that characterize actual markets. It is unsurprising that behavior in such impoverished environments differs from behavior in rich market environments. The "anomalies" may reveal less about human psychology than about what happens when markets are absent.

This shifts the research question. Instead of asking "Do people have consistent preferences?" we ask "Under what institutional conditions does behavior exhibit GARP-consistency?" The answer has implications for institutional design: if markets generate consistent behavior while other contexts do not, we might design institutions to capture market-like properties—or we might conclude that choice-based welfare analysis is valid only for market-like contexts.

***

## 8.5 Challenges to the Axioms

### The Allais Paradox

**Figure 8.3: The Allais Paradox**

![ch08\_fig03\_allais](/files/fK8AEi2BLlCmDjyHaDfK)

Maurice Allais, accepting the Nobel Prize in 1988, remained a fierce critic of expected utility theory. His 1953 paradox demonstrated that intelligent, reflective agents systematically violate the independence axiom—and do not consider themselves irrational for doing so.

Consider two pairs of choices:

**Choice 1**:

* Option A: $1 million with certainty
* Option B: 89% chance of $1 million, 10% chance of $5 million, 1% chance of nothing

**Choice 2**:

* Option C: 11% chance of $1 million, 89% chance of nothing
* Option D: 10% chance of $5 million, 90% chance of nothing

Most people choose A over B (preferring certainty to a slightly better gamble) and D over C (preferring the higher prize when both are unlikely). But this combination violates independence.

To see why, note that options A and B can be decomposed: A = 89%($1M) + 11%($1M), while B = 89%($1M) + 11%(lottery yielding 10/11 chance of $5M, 1/11 chance of nothing). The 89% common component should be irrelevant by independence. What matters is the 11% component: certainty of $1M versus the gamble.

Similarly, C = 89%(nothing) + 11%($$1M), while D = 89%(nothing) + 11%(lottery yielding 10/11 chance of$$5M, 1/11 chance of nothing). Again, the 89% common component (now nothing) should be irrelevant.

But the 11% components are *identical* across the two choice pairs. If you prefer the certain $1M component in Choice 1, independence says you should prefer it in Choice 2 as well—which means choosing C over D. The A-and-D pattern violates independence.

### The Significance of Allais

Allais's paradox is not merely a laboratory curiosity. It reveals something important about risk attitudes that expected utility theory cannot capture.

The typical A-over-B choice reflects what Allais called the "certainty effect": people place special weight on outcomes that are certain, beyond what probability weighting in EU would suggest. The difference between certain and almost-certain feels psychologically larger than the same probability difference when both outcomes are uncertain. This is intuitively compelling—a 1% risk of losing everything feels different depending on whether the alternative is certain success or merely probable success.

Expected utility theory cannot accommodate this. Independence requires that the common 89% component be treated the same whether it yields $1 million (in Choice 1) or nothing (in Choice 2). But intuitively, the certain-million context of Choice 1 makes the 1% risk of getting nothing feel more salient.

Whether Allais choices are irrational is contested. Some argue that respondents, upon understanding the independence axiom, should correct their choices. Others—Allais included—maintain that the choices reflect a reasonable risk sensitivity that independence artificially excludes. The axiom, not the behavior, is flawed.

### The Ellsberg Paradox

Daniel Ellsberg (later famous for the Pentagon Papers) identified a different kind of violation: ambiguity aversion. Expected utility theory assumes agents assign probabilities to all uncertain events. Ellsberg showed that people distinguish between risk (known probabilities) and uncertainty (unknown probabilities) in ways EU cannot capture.

Consider an urn containing 90 balls: 30 red and 60 that are either black or yellow in unknown proportions. You can bet on the color of a randomly drawn ball.

**Bet 1**: Win $100 if red vs. Win $100 if black Most people prefer red (known 1/3 probability) to black (unknown probability).

**Bet 2**: Win $100 if red or yellow vs. Win $100 if black or yellow Most people prefer black or yellow (known 2/3 probability) to red or yellow (unknown).

These choices are jointly inconsistent with EU. If you prefer red to black in Bet 1, you must believe P(red) > P(black). Since P(red) = 30/90 = 1/3, this implies P(black) < 1/3. But if you prefer black-or-yellow in Bet 2, you believe P(black) + P(yellow) > P(red) + P(yellow), which simplifies to P(black) > P(red) = 1/3—contradicting the inference from Bet 1 that P(black) < 1/3.

Ellsberg choices reflect ambiguity aversion: a preference for known over unknown probabilities, even holding expected value constant. This is not accommodated by EU, which treats all uncertainty as reducible to precise probabilities.

### Beyond Expected Utility?

The Allais and Ellsberg paradoxes, along with other systematic violations, have spawned alternatives to expected utility theory. Prospect theory (Kahneman and Tversky), rank-dependent utility, and models incorporating ambiguity aversion attempt to capture observed behavior more accurately.

These alternatives face their own challenges. They typically require more parameters, reducing parsimony. They may predict violations that do not occur as frequently as the violations they were designed to explain. And they raise the question of whether descriptive accuracy should be the primary criterion for a theory of rationality.

The relationship between expected utility and its alternatives is not simple replacement. EU remains the benchmark—the theory against which alternatives are measured. Its normative appeal persists even as its descriptive adequacy is questioned. Whether any single theory can serve both descriptive and normative purposes, or whether we need different theories for different purposes, remains an open question.

***

## Summary

Rational choice theory provides the formal foundation for economic analysis of decision-making under risk and uncertainty. This chapter has examined that foundation, its interpretations, and its limits.

*Expected utility theory* derives from the von Neumann-Morgenstern axioms: completeness, transitivity, continuity, and independence. These axioms, if satisfied, guarantee the existence of a utility function such that agents rank lotteries by expected utility. The independence axiom does most of the theoretical work and is most vulnerable to empirical challenge.

*Interpreting EU* requires deciding whether it is descriptive, normative, or instrumental. Descriptive interpretations face experimental falsification. Normative interpretations require defending the axioms as requirements of rationality—possible for some axioms, contested for independence. Instrumental "as if" interpretations sacrifice depth of understanding for predictive convenience.

*Game theory* extends rational choice to strategic interaction. Nash equilibrium provides a powerful organizing concept, but its epistemic foundations—requiring common knowledge of rationality—are demanding. What rational players should do in games remains contested.

*Revealed preference* attempts to ground rationality in choice consistency alone, avoiding claims about mental states. Sen's critique shows that consistency is not enough: we must know what the agent is trying to achieve. Context effects, menu dependence, and framing effects further complicate the relationship between choice and preference. The *neo-Samuelsonian* response reinterprets GARP as characterizing market behavior patterns, not cognitive processes—shifting the question from "Do people have consistent preferences?" to "Under what institutional conditions does consistent behavior emerge?"

*Empirical challenges*—the Allais and Ellsberg paradoxes—demonstrate systematic violations of expected utility by reasonable agents. Whether these violations are errors or reveal limitations in the theory itself remains debated. Alternative theories capture observed behavior better but sacrifice parsimony and normative clarity.

Rational choice theory thus occupies an ambiguous position: too successful as a modeling framework to abandon, too flawed as a description or prescription to accept uncritically. Understanding both its power and its limits is essential for evaluating its applications throughout economics.

***

## Discussion Questions

1. Is expected utility theory best understood as descriptive, normative, or merely a useful modeling device? Can a single theory serve all three purposes?
2. The Allais paradox can be explained by the "certainty effect"—people overweight certain outcomes relative to probable ones. Is this irrational, or does it reveal something important that expected utility theory misses?
3. Sen argues that "consistency of choice" cannot capture rationality because we need to know what the agent is trying to achieve. How damaging is this critique to the revealed preference program?
4. Dutch book arguments show that EU-inconsistent agents can be exploited. Does vulnerability to exploitation make behavior irrational, or is real-world exploitation rare enough that this consideration lacks force?
5. Game theory assumes players have common knowledge of rationality. Is this assumption ever satisfied? Does game theory's usefulness depend on its being satisfied?
6. The neo-Samuelsonian interpretation claims GARP describes market behavior, not mental states. If so, what (if anything) does revealed preference tell us about well-being? Can we do welfare economics without psychological preferences?
7. If behavioral anomalies disappear in market contexts (as Vernon Smith argues), should we conclude that (a) markets make people rational, (b) markets select for rational behavior, or (c) rationality is a property of institutions rather than individuals? Do these interpretations differ practically?

***

## Key Terms

* **Expected utility**: Probability-weighted average of utilities across outcomes
* **von Neumann-Morgenstern axioms**: Conditions (completeness, transitivity, continuity, independence) sufficient for expected utility representation
* **Independence axiom**: Preferences between lotteries should not change when both are mixed with a common third lottery
* **Nash equilibrium**: Strategy profile where each player best-responds to others
* **Common knowledge of rationality**: Each player is rational, knows others are rational, knows others know, etc.
* **Revealed preference**: The view that preferences are defined by or inferred from observable choices
* **WARP (Weak Axiom of Revealed Preference)**: If x is chosen when y is available, y should not be chosen when x is available
* **GARP (Generalized Axiom of Revealed Preference)**: Consistency condition allowing for budget constraint changes
* **Neo-Samuelsonian**: Interpretation of revealed preference as characterizing market behavior patterns, not cognitive processes
* **Allais paradox**: Systematic violation of independence involving certainty effects
* **Ellsberg paradox**: Preference for known over unknown probabilities (ambiguity aversion)
* **Dutch book**: Set of bets guaranteeing loss for agent with inconsistent preferences
* **Equilibrium selection**: The problem of choosing among multiple Nash equilibria
* **Focal point**: Salient equilibrium selected by shared expectations or conventions

## Connections

* **← Ch 1 (Economic Agents)**: The homo economicus framework that this chapter axiomatizes.
* **← Ch 2 (Theories of Value)**: Revealed preference and ordinal utility as the foundations for expected utility theory.
* **→ Ch 9 (Behavioral Economics)**: Systematic empirical challenges to the rational choice axioms developed here.
* **→ Ch 10 (Learning and Evolution)**: Evolutionary and learning-theoretic justifications for equilibrium play.
* **→ Ch 15 (Uncertainty and Knowledge)**: The Savage/Ellsberg framework and its implications for deep uncertainty.

***

## Notes

***

## Further Reading

### Essential

* Mas-Colell, A., M. Whinston, and J. Green (1995). *Microeconomic Theory*. Oxford University Press. \[Chapters 1, 6]
* Sen, A. (1993). "Internal Consistency of Choice." *Econometrica* 61(3): 495-521.
* Allais, M. (1953). "Le comportement de l'homme rationnel devant le risque." *Econometrica* 21(4): 503-546.

### Background

* Binmore, K. (2009). *Rational Decisions*. Princeton University Press.
* Gilboa, I. (2009). *Theory of Decision under Uncertainty*. Cambridge University Press.
* Schelling, T. (1960). *The Strategy of Conflict*. Harvard University Press. \[On focal points and coordination]

### Advanced

* Rubinstein, A. (1998). *Modeling Bounded Rationality*. MIT Press.
* Machina, M. (1987). "Choice Under Uncertainty: Problems Solved and Unsolved." *Journal of Economic Perspectives* 1(1): 121-154.
* Ellsberg, D. (1961). "Risk, Ambiguity, and the Savage Axioms." *Quarterly Journal of Economics* 75(4): 643-669.

### On Neo-Samuelsonian Revealed Preference

* Ross, D. (2014). *Philosophy of Economics*. Palgrave Macmillan. \[Chapters 2, 4]
* Binmore, K. (1994). *Game Theory and the Social Contract, Vol. 1: Playing Fair*. MIT Press. \[Chapter 3]
* Gul, F. and W. Pesendorfer (2008). "The Case for Mindless Economics." In *The Foundations of Positive and Normative Economics*, ed. A. Caplin and A. Schotter. Oxford University Press.
* Harrison, G. (2008). "Neuroeconomics: A Critical Reconsideration." *Economics and Philosophy* 24(3): 303-344.

***

*Chapter 8 Draft | Version 1.1 | December 2025*