RESEARCH DOSSIER
Revealed Preference and Rationalizability of Choice
Abstract. Revealed preference turns the theory of consumer optimization into a finite-data empirical question: whether one stable ordering of consumption bundles can account for choices made under different feasible budgets. This monograph develops the historical and mathematical foundations of that question; specifies the weak, strong, and generalized axioms; establishes the conditions under which Afriat's inequalities are feasible; and constructs a continuous, strictly increasing, concave utility rationalization. It supplies exact worked counterexamples, a constructive argument using an ordered graph of affordability relations, and an illustration of observationally equivalent utility functions with different counterfactual choices. It then examines demand integrability, measurement and identification, the critical cost efficiency index, statistical power, random utility, and the interpretation of rationalizability in welfare economics. Published research on experimental choices, household expenditures, and repeated cross-sections supplies the empirical context. The central analytical finding is that consistency, recoverability, predictive precision, and welfare interpretation address distinct inferential questions, each requiring its own evidence. Finite expenditure records establish compatibility with optimization under a maintained specification of opportunities; informative reconstruction of preferences depends on the coverage and credibility of those opportunities. The numerical examples are original pedagogical constructions, not estimates from household microdata.
Keywords: Revealed preference; rationalizability; consumer choice; WARP; SARP; GARP; Afriat inequalities; utility maximization; demand integrability; nonparametric identification; critical cost efficiency index; stochastic choice; welfare economics
Object and historical lineage
The observable content of economic choice
The scientific object of revealed preference is a relation inferred from choices made among feasible alternatives. In consumer economics, the observational unit is a price vector, a consumption vector, and the opportunity set that allows the latter to be compared with other bundles. A preference ordering supplies a possible account of these choices. Rationalizability is the existence of an ordering, satisfying explicitly stated properties, for which every observed choice is optimal on its corresponding feasible set.
An expenditure record is therefore evidence of a comparison only when the alternatives are simultaneously feasible. A person who buys fruit instead of a television has revealed a meaningful economic comparison only if the particular alternatives at issue were available and affordable. A list of purchases detached from their opportunities records realized outcomes. The inferential role of price and budget information is to reconstruct the opportunities against which the outcomes can be evaluated.
The question is deliberately narrower than a general theory of individual behavior. It concerns optimization of a stable ordering within specified opportunities. Psychology, information acquisition, institutional restrictions, strategic interactions, and temporal changes may all influence the decision under observation. Their relevance enters through the definition of alternatives, budgets, information, and stability assumptions. The method becomes empirically informative once these elements have been stated and held sufficiently fixed to support comparison.
Historical development
Samuelson's 1938 theory of consumers' behaviour placed directly observable demand comparisons at the center of consumer theory. His formulation anticipated the empirical discipline imposed by pairwise consistency. Its historical significance lies in the transition from an assumed psychological interpretation of utility to restrictions on choice behavior. Houthakker's (1950) strong revealed preference formulation extended the argument to chains of comparisons, connecting consistency across several decisions with an underlying ordering. The historical account in Tipoe, Adams, and Crawford's (2022) survey of revealed preference provides the context for the distinction between complete demand functions and finite observations.
Afriat's (1967) construction from expenditure data established the finite-observation problem as a constructive mathematical question. Varian's (1982) nonparametric demand analysis developed its practical use for testing, preference recovery, and forecasting. Those contributions transformed the program into an empirical technique whose basic restrictions can be checked without selecting a parametric demand equation in advance.
The historical sequence matters because an integrable demand system observed over a continuum and a finite collection of purchases contain different information. The first can in principle support restrictions on derivatives and global behavior when suitable smoothness and domain conditions hold. The second supports finite comparisons, feasibility systems, and partial identification. A finite-data theorem does not require observing the full demand function.
Budgets, observables, and preference relations
Commodity space and expenditure
Let there be K commodities, with feasible nonnegative consumption vectors x in R^K_+. At observation t, prices p_t belong to R^K_{++}, income available for this choice is m_t, and the linear budget is B_t = {x ≥ 0 : p_t · x ≤ m_t}. The observed vector x_t is feasible. Strictly positive finite prices and finite income give a compact budget. A continuous utility function consequently attains a maximum on it.
For locally nonsatiated preferences, an optimum exhausts this linear budget. Local nonsatiation means that every neighborhood of any feasible commodity vector within the commodity domain contains a point strictly preferred to that vector. If x_t has positive budget slack, a sufficiently nearby strictly preferred point remains affordable by continuity of expenditure, contradicting optimality. It follows that m_t = p_t · x_t for the maintained model. Recovering income from expenditure therefore relies on budget exhaustion, and the resulting test concerns the specified commodity domain.
When a household retains income as savings, a model of all consumption and saving should include savings or formulate a conditional expenditure allocation. Using observed food expenditure as a food budget can test a conditional food-demand model. It does not test optimization over all household resources unless the assumptions linking the conditional allocation to overall preferences are justified. Quantity rationing, indivisible goods, credit restrictions, and stockouts can similarly invalidate the linear-budget interpretation.
Prices must have comparable units. If quantities are measured in kilograms one month and grams the next without price conversion, a cross-budget comparison has no economic meaning. Multiplying every price and the expenditure budget in one observation by the same positive constant leaves that observation's feasible set unchanged. Rescaling a commodity requires the corresponding inverse transformation of prices across all observations. Inflation adjustment should preserve within-budget real relative prices and the same choice set.
Revealed relations
Define the cost matrix C by C_ts = p_t · x_s and the expenditure difference A_ts = C_ts − C_tt. A weak direct edge t → s exists when A_ts ≤ 0. A strict direct edge exists when A_ts < 0. Rows correspond to the environment in which choices were made; columns correspond to the bundles evaluated in that environment. The diagonal entries of A are zero.
Write x_t R^D x_s when x_s was affordable at t, and x_t P^D x_s when it was strictly cheaper than x_t. Under optimization, the first implies u(x_t) ≥ u(x_s). Under local nonsatiation over the continuous commodity domain, the second implies u(x_t) > u(x_s): the cheaper bundle admits a nearby superior affordable alternative. The notation P^D refers to a strict expenditure comparison, and does not label every revealed weak preference as psychologically strict.
The transitive closure R of R^D includes all ordered pairs connected by a directed path. Observed bundles need not be directly comparable at their own two budgets to be comparable through an intermediate bundle. The strength of indirect comparisons is derived from the transitivity of an optimizing preference order, together with stability of the order over the observations.
Preference representations and ties
A complete and transitive weak order ranks every pair of alternatives. A real-valued utility representation preserves this ranking when such a representation exists under appropriate domain and regularity conditions. Its numerical scale is ordinal: every strictly increasing transformation preserves the same comparisons. Utility levels and marginal utility of income are therefore not directly measured physical quantities.
Observed choices can select one of several equally desirable bundles. An empirical test that requires unique maximizers is stronger than a test allowing a utility-maximizing correspondence. This distinction becomes essential on equal-cost budget intersections and when alternative-specific tie-breaking rules select different elements of a common maximizing set. Strictly increasing preferences rule out satiation but can retain indifference between bundles with different commodity compositions. Concavity likewise permits flat segments of indifference surfaces. Strict concavity imposes additional uniqueness conditions.
Axioms and the structure of consistency
Weak and strong revealed preference
Terminology differs across treatments of WARP and SARP. For this monograph, classical unique-choice WARP means that for two distinct selected bundles x_t and x_s, mutual weak affordability is excluded: x_t R^D x_s implies that x_s R^D x_t cannot also hold. A tie-permitting pairwise condition excludes only mutual affordability with at least one strict expenditure inequality. These two conventions coincide for many datasets without cross-budget equalities. They differ when observations lie on each other's budget boundaries.
Classical SARP for distinct bundles extends the unique-choice restriction to paths: after x_t is indirectly revealed preferred to x_s, no return path is allowed if the bundles are distinct. Duplicate observations of the same commodity vector are not contradictions. Strong consistency concerns cycles involving several choices rather than merely pairs. Under single-valued demand assumptions it is the natural extension of the pairwise restriction.
GARP uses a weaker treatment of ties while imposing global consistency. In Varian's (1982) terminology, whenever t reaches s through weak affordability edges, the return edge s → t cannot be strict. Equivalently, the weak graph contains no directed cycle with a strictly negative A entry. A cycle consisting exclusively of equal-expenditure comparisons can be rationalized by indifference; a cycle containing a strict comparison would require an ordering to be both weakly and strictly greater than itself.
These conventions make the relationship between the axioms precise. SARP under the stated distinct-bundle convention implies GARP. GARP implies the tie-permitting pairwise condition. Neither pairwise convention by itself generally implies GARP in arbitrary commodity dimension. GARP can allow ties that classical unique-choice SARP excludes. These are statements about precise definitions rather than interchangeable labels of rationality.
Graph interpretation
The weak affordability graph reveals where the data impose comparisons on a candidate preference relation. A strongly connected component is a set of observations mutually reachable in this graph. All utility levels within such a component must coincide if the observed choices can be rationalized. Hence GARP is equivalent to the absence of strict edges inside any strongly connected component.
This is useful computationally. A transitive closure constructed from a T-by-T Boolean adjacency matrix takes O(T^3) operations using a standard triple-loop algorithm. A strongly connected component algorithm can directly detect the forbidden edges in O(T+E) graph operations after the edges have been constructed, where E is the number of weak edges. Computing the cost matrix itself requires O(T^2 K) arithmetic operations in a dense implementation. Dense storage requires O(T^2) memory.
For auditability, a rejection should provide an explicit cycle rather than a Boolean result alone. A strict edge i → j within a component is a witness when a path j → i also exists. The cycle identifies the particular observations, goods, and price comparisons that deserve inspection. In a statistical application it permits one to compare the rejection with the magnitude of price and quantity measurement error.
Afriat's theorem and its constructive proof
Statement and necessity
For a finite set of observations with p_t > 0 componentwise and nonnegative x_t, take m_t = p_t · x_t and the full linear budgets defined above. The following statements are equivalent under the maintained commodity domain:
- A locally nonsatiated utility function rationalizes the observed choices.
- The observations satisfy GARP.
- There exist real numbers U_t and strictly positive λ_t satisfying the Afriat inequalities for all ordered pairs t,s: U_s ≤ U_t + λ_t A_ts.
- A continuous, strictly increasing, concave utility function rationalizes the observed choices.
The original research is Afriat (1967), with modern formulations in Varian (1982) and Fostel, Scarf, and Todd's two proofs of Afriat's theorem (2004). The equivalence concerns existence, not uniqueness. It rules out all locally nonsatiated rationalizations when GARP is violated and guarantees a particularly regular rationalizer when it passes.
Necessity follows immediately from utility comparisons along a cycle. Weak edges require weakly declining utility levels, and a strict edge makes at least one decline strict. Returning to the starting point is impossible. This argument uses local nonsatiation; allowing a constant utility function would remove the empirical content of maximization for ordinary nonempty budgets.
The linear inequalities themselves imply the same necessity. Along any weak cycle U_s ≤ U_t, with a strict inequality wherever A_ts < 0 because λ_t > 0. Summing inequalities around such a cycle produces 0 ≤ a strictly negative sum, a contradiction.
Existence of the inequalities
The following finite constructive argument explains sufficiency directly. Collapse the weak graph into strongly connected components. GARP ensures that all edges internal to a component have A_ts = 0. Order the resulting acyclic component graph so that every edge between components goes from an earlier rank to a later one. Assign each component an integer rank r, starting at zero.
Let d be the smallest strictly positive entry of A, or 1 if there is no positive entry. Let b be the maximum of zero and the absolute values of all negative entries of A. Take M > 1 + b/d and choose
U_t = −M^{r_t}, λ_t = M^{r_t}/d.
These values satisfy every inequality. If A_ts = 0, the corresponding edge implies r_s ≥ r_t, hence U_s ≤ U_t. If A_ts < 0, GARP rules out an internal strict edge, so r_s ≥ r_t + 1. Then U_t − U_s ≥ (M−1)M^{r_t}, which exceeds λ_t|A_ts| by the choice of M. If A_ts > 0, λ_t A_ts ≥ M^{r_t}; also U_s − U_t ≤ M^{r_t}, independent of the relative ranks. Thus U_s ≤ U_t + λ_t A_ts in every case.
This argument is an explanatory reconstruction of the established theorem. Exponential rank weights can produce poor numerical conditioning for long chains; they are useful as an existence proof, not as a recommendation for floating-point estimation. The proof separates the economic condition, absence of strictly marked cycles, from the choice of numerical utility and slope witnesses. A feasible linear-programming solution is usually better scaled.
Constructing the rationalizer
Given any solution of the Afriat inequalities, define
u(x) = min_t {U_t + λ_t p_t · (x − x_t)}.
For each x_s, every affine term is at least U_s by the inequalities, and the s-th term equals U_s. Therefore u(x_s) = U_s. For any x in B_t, the t-th affine term is at most U_t, so u(x) ≤ U_t = u(x_t). The selected bundle maximizes the constructed utility function on the entire budget.
The finite minimum of affine functions is continuous and concave. Every affine function has strictly positive coordinates in its slope vector, because λ_t and all components of p_t are positive. If y ≥ x coordinatewise and y differs from x in at least one coordinate, every affine term increases strictly; the minimum also increases strictly. The rationalizer is consequently strictly monotone on the commodity space.
The utility representation is ordinarily piecewise affine and may be nondifferentiable. This does not diminish the existence claim. Smoothness, curvature beyond concavity, and unique demand generally require additional restrictions and cannot be inferred merely from the theorem.
Computational feasibility
A linear-programming implementation may impose λ_t ≥ 1 and U_1 = 0 without changing the existence question. If a strictly positive feasible vector λ exists, its minimum over finitely many observations is positive; multiply both U and λ by a common factor and shift all U by a common constant. The normalized system then satisfies these constraints. Using λ_t ≥ 0 alone is inadequate because zero multipliers would allow a trivial constant solution. For numerical work, choose sensible price and quantity units, report feasibility tolerances, and independently evaluate the residual U_s − U_t − λ_t A_ts.
The characterization is nonparametric with respect to the form of ordinal utility. Linear budgets, stable preferences, observed prices, opportunity-set completeness, and local nonsatiation remain substantive maintained assumptions. This explains why the method is simultaneously flexible about functional form and demanding about the empirical specification of choices.
Numerical experiments and counterfactual identification
A cycle invisible to the pairwise axiom
Consider three commodities with x_1=(1,0,0), x_2=(0,1,0), and x_3=(0,0,1). Prices are p_1=(2,1,3), p_2=(3,2,1), and p_3=(1,3,2). Each chosen bundle costs 2. Cross-expenditures are:
| Budget | x_1 | x_2 | x_3 |
|---|---|---|---|
| p_1 | 2 | 1 | 3 |
| p_2 | 3 | 2 | 1 |
| p_3 | 1 | 3 | 2 |
There is a strict affordability cycle x_1 → x_2 → x_3 → x_1. No distinct pair has mutual weak affordability, so both pairwise conventions used above pass. GARP fails because every edge in the three-cycle is strict. The Afriat system would require U_2 ≤ U_1 − λ_1, U_3 ≤ U_2 − λ_2, and U_1 ≤ U_3 − λ_3. Adding yields 0 ≤ −(λ_1 + λ_2 + λ_3), impossible.
The counterexample demonstrates the additional empirical content of transitivity without relying on measurement error, unobserved tastes, or a change in income. It also shows why checking only pairwise reversals is inadequate for general finite demand data.
A rationalizable sample and two feasible utilities
Take two commodities with x_1=(2,1), x_2=(1,2), p_1=(1,2), and p_2=(2,1). Both chosen vectors cost 4. Each alternative observed bundle costs 5 under the other observation's prices.
| Budget | x_1 | x_2 |
|---|---|---|
| p_1 | 4 | 5 |
| p_2 | 5 | 4 |
No distinct bundle is affordable at the other observed budget. GARP passes. U_1=U_2=0 with λ_1=λ_2=1 satisfies the inequalities. The resulting rationalizer is
u_A(x) = min {x_1 + 2x_2 − 4, 2x_1 + x_2 − 4},
where subscripts within a utility argument now indicate commodity coordinates. A second feasible assignment U_1=0, U_2=1, λ_1=λ_2=1 yields
u_B(x) = min {x_1 + 2x_2 − 4, 2x_1 + x_2 − 3}.
The two utility functions rank the observed bundles differently: the first assigns them equal utility, the second strictly prefers the second observation's bundle. Since the budgets do not cross at affordable observed bundles, the recorded choices allow both rankings.
At the new price vector (1,1) and income 3, strict monotonicity implies x_2=3−x_1. The first utility reduces to min {2−x_1, x_1−1}, with a unique maximum at x_1=1.5. The second reduces to min {2−x_1, x_1}, with a unique maximum at x_1=1. The corresponding choices are (1.5,1.5) and (1,2). Both predict a different bundle at the same new opportunity set while exactly rationalizing the past.
This example does not compute the entire identified set. It proves that the identified set contains at least two distinct counterfactual quantities. Even a dataset satisfying the strongest distinct-bundle conditions in this example can be uninformative about preferences across its observed choices. More informative experimental designs deliberately create overlapping affordability regions.
Recoverability and prediction
For an unobserved candidate y in a new budget, append (p_new,y) to the original dataset and check GARP. If GARP holds, at least one utility in the admissible class rationalizes the augmented observations. If it fails, the candidate is incompatible with every utility in that class. The resulting feasible choice set is therefore a nonparametric prediction set under the maintained model.
The interpretation requires the same goods, opportunity model, and stable preferences for both historical observations and the hypothetical budget. When the price change introduces a new product, quality, technology, or constraint, the old data alone may not identify the new commodity or opportunity structure.
An internally defensible counterfactual can remain a set rather than a point. A known functional family can narrow this set, but that narrowing is attributable to the extra assumptions. Reporting a convenient rationalizer as the recovered true utility confuses an existence witness with an identified structural object.
Demand integrability and economic interpretation
The continuum problem
Let x(p,m) denote a differentiable Marshallian demand function on a suitable price-income domain. Budget exhaustion gives p · x(p,m)=m. Homogeneity of degree zero in prices and income follows because a common positive rescaling preserves the budget. If the demand arises from differentiable utility optimization under standard regularity conditions, its Slutsky substitution matrix S has entries
S_ij = ∂x_i/∂p_j + x_j ∂x_i/∂m.
The expression follows from compensating an infinitesimal price change by its first-order expenditure effect x_j dp_j. If Hicksian demand is differentiable and the expenditure function twice differentiable, the substitution matrix is the Hessian of the expenditure function with respect to prices. Symmetry follows from equality of cross-partials, and negative semidefiniteness follows from concavity of expenditure in prices. The latter property is a consequence of expenditure being the infimum of price-linear costs of bundles attaining the target utility level.
These conditions describe local regularity of a sufficiently smooth demand system. Global integrability statements also depend on the domain and other hypotheses. Finite observations cannot directly establish a derivative at unobserved price-income points. Using a fitted demand equation to estimate a Slutsky matrix introduces statistical and specification assumptions beyond the raw finite-data test.
Afriat's construction instead yields finite inequalities and a possibly nonsmooth utility rationalizer. Its existence demonstrates compatibility with a concave ordering but does not identify the derivatives of the true utility, the elasticity matrix, or the compensated response to a policy change. The continuum and finite approaches are complementary: one relates differential structures, the other imposes global consistency on available observations without estimating unobserved derivatives.
The budget multiplier and demand interpretation
In a differentiable utility maximization problem with an interior solution, the first-order condition has the form ∇u(x_t)=λ_t p_t for a multiplier λ_t associated with the budget. For a concave utility representation, the supporting hyperplane condition gives u(x_s) ≤ u(x_t)+λ_t p_t·(x_s−x_t). The Afriat system is analogous to this condition but remains valid as a finite existence characterization without interior or differentiability assumptions.
A feasible multiplier vector need not reveal the marginal welfare effect of actual income. Transforming utility by a suitable increasing transformation changes numerical slopes while preserving choices. On nonsmooth regions, supporting slopes may also be nonunique. Positive multipliers certify economically ordered constraints in a constructed representation. Interpersonal comparison of utility or money-metric aggregation demands additional assumptions.
Empirical identification, measurement, and international comparisons
Constructing an observation panel
The practical unit is a sequence of comparable opportunities for a decision-making unit. Repeated household purchases, experimental allocations, and repeated choices from explicitly designed menus can all be informative. Repeated cross-sections of different households cannot simply be treated as a single person's time series. A pooled GARP rejection can reflect heterogeneous preferences even if every individual is a consistent optimizer.
For household budget microdata, the researcher must identify the products, quantities, paid prices, available prices for unchosen alternatives, resource measure, timing, household identity, and relevant constraints. Quantity and expenditure are often recorded more reliably than the unchosen offers; unit values derived by dividing expenditure by quantity may reflect quality selection or discounts rather than exogenous market prices. A common price index can obscure local variation. Substituting average product prices may manufacture affordability edges at some points and erase them at others.
A survey that observes a single household in one expenditure period does not yield an individual panel revealed-preference test by itself. Longitudinal panels have different selection and attrition problems. Cross-sectional nonparametric demand analysis can estimate conditional distributions or Engel curves under additional assumptions, but the constructed object and the maintained restrictions must be named.
Durable-goods purchases introduce another issue: the purchase date and service-consumption period need not coincide. Inventory accumulation can explain why a household purchases an unusually large quantity at a discount and little later. Treating every purchase as contemporaneous consumption can generate artificial instability. The model should either measure the relevant service flows or describe a dynamic choice problem.
Measurement and institutional heterogeneity
Let observed costs be uncertain within intervals. A strict edge should be described as robust only if the maximum plausible cross-expenditure is below the minimum plausible chosen expenditure, with joint measurement constraints respected. A weak edge near equality should be labeled fragile when reasonable perturbations can change its sign. If uncertainty includes correlated prices or quantities, independent per-entry perturbations can create impossible cost matrices; the sensitivity calculation should operate on the underlying measurement process.
Institutional feasibility varies with credit eligibility, rationing, minimum purchase sizes, delivery restrictions, contractual obligations, and local product availability. These restrictions affect which bundles are attainable even at identical money prices. Comparisons of revealed preference consistency across countries should use comparable product definitions, household composition, recall periods, and feasible opportunities. An apparent difference in consistency may be caused by market access or survey design.
The mathematically universal theorem applies to data satisfying its assumptions regardless of national context. Empirical estimates remain specific to their samples, institutional settings, and measurement systems. This distinction supplies a disciplined basis for international comparison without imposing an unverified universal distribution of consumer preferences.
Selected evidence from published applications
Echenique, Lee, and Shum (2011), in their money-pump analysis, study household grocery expenditure panels. Their introduction reports that 396 of 494 households encounter at least one GARP violation, yet the authors find small violations according to their severity metric and do not reject their rationality-with-measurement-error null. The evidence illustrates how binary consistency, violation severity, and statistical significance can yield different conclusions from the same observations. It does not establish that all households or markets behave in the same manner.
Choi, Kariv, Müller, and Silverman (2014), in Who Is (More) Rational?, connect experimental consistency scores to household-economic characteristics. Their published abstract reports an association between a one-standard-deviation increase in consistency and 15–19 percent more household wealth. This is an association conditional on the design and reported controls, not a causal estimate of the effect of an intervention that improves consistency. The paper's abstract and bibliography support the reported association; this monograph does not independently reproduce the regression or infer a worldwide effect.
Kitamura and Stoye's (2018) random utility analysis uses British household expenditure data to illustrate stochastic rationalizability tests. The authors report that the estimated distributions do not exactly satisfy their population-model restriction but that the statistical rejection is insignificant. This result concerns a mixture model and uncertainty in estimated choice probabilities, rather than individual deterministic GARP.
The studies are deliberately contrasted because they investigate different observational objects: scanner panel purchases, designed individual choices linked to household outcomes, and repeated expenditure distributions. A common conceptual interest in rationality does not make their statistics directly poolable. A global meta-analysis would require harmonized definitions and raw data; none is asserted here.
Approximate rationality, statistical power, and competing models
Critical cost efficiency
Exact GARP treats any strict forbidden cycle as a contradiction. Near-boundary observations motivate the critical cost efficiency index (CCEI), defined here using a common efficiency parameter e in (0,1]. Declare a weak e-affordability edge t → s if C_ts ≤ e C_tt and a strict edge if C_ts < e C_tt. The associated GARP test uses the transitive closure of the weak edges and excludes a strict returning edge. At e=1 it is exact GARP. Reducing e contracts the affordability comparisons available to the test, so violations cannot be introduced by such a contraction.
Define the index as the supremum of passing e values, including zero as a limiting lower bound when needed. Equality conventions matter at thresholds, and an empirical implementation should report whether a supremum is actually attained. With positive chosen expenditures, breakpoints are ratios C_ts/C_tt; exact rational arithmetic or a carefully documented tolerance avoids treating a threshold tie as strict.
For the three-good cycle constructed above, the forward cross-cost ratio is 1/2. If e > 1/2, all three edges are strict and the test fails. At e = 1/2 the same cycle consists of weak equalities and passes the modified test. At e < 1/2 those edges disappear. Its CCEI is consequently 0.5. For the two-good rationalizable sample the CCEI is 1.
An efficiency of 0.95 is a cost-based consistency measure. It does not mean a person is 95 percent rational, nor does it automatically measure five percent of realized income lost, five percent lower utility, or the probability of optimization. Its interpretation depends on the contraction convention and the observed opportunity sets.
Economic severity and measurement error
The money-pump approach assesses cycles in terms of the monetary opportunity for exploitation under specified exchanges. The measure and the CCEI address related but distinct aspects of violation severity. A count of violating pairs or cycles measures frequency; a cost-efficiency threshold measures how much affordability must be contracted; a money-pump measure attaches a monetary scale to inconsistent exchange opportunities under the model.
Theoretical exploitability should not be described as realized financial loss. Actual exploitation requires repeated implementation, compatible trading opportunities, and a decision rule that continues to produce the relevant inconsistent choices. Transaction costs, learning, limits on quantities, and refusal to repeat the sequence can prevent it.
A statistical test requires a data-generating error model. The minimum adjustment to prices or quantities needed to reach the rationalizable set can be meaningful, but comparison with a sampling distribution requires assumptions about error magnitudes, dependence, heteroskedasticity, or experimental randomization. A raw GARP rejection with exact measurements has a logical interpretation; a rejection with noisy measurements requires inferential calibration. Tipoe et al. (2022) discuss how measurement and behavioral departures change this problem.
Power, alternative hypotheses, and model flexibility
A dataset can pass because it contains rich evidence of rational ordering, or because few choices were mutually affordable. These situations are logically equivalent for an existence test but evidentially different for model assessment. Bronars's (1987) analysis of test power addresses the importance of evaluating the ability of nonparametric tests to detect deviations under a specified alternative.
A transparent power exercise simulates alternative choices on the actual experimental budgets and calculates the share of simulated datasets that the test rejects. The alternative distribution must be defined: uniform allocation of expenditure shares, uniform random selection from an enumerated menu, or another specified process. These distributions are generally different. Rejection frequencies under a simulated alternative are design-specific estimates, not universal error rates.
Budget intersections, relative price variation, and observation count influence power. Adding observations cannot restore exact GARP once an incompatible subset is fixed; it can expose additional restrictions. Repeating nearly identical budgets may add little informational variation. Experimental designs should balance power, interpretability, participant burden, and coverage of economically relevant alternatives.
A model permitting preference changes at every observation can rationalize essentially any feasible choice by fitting the order separately. Conversely, a restrictive parametric family can be rejected even where general utility maximization survives. Model comparison should consider empirical fit together with restrictiveness. A behavioral model that passes more often may simply admit a larger set of observable patterns.
Cherchye, Demuynck, De Rock, and Lanier (2025; examined 2023 manuscript) propose a reversal of the testing burden, treating random behavior as the null and approximate utility maximization as the alternative. This approach asks for affirmative evidence exceeding a calibrated chance benchmark. Its error guarantees depend on the specified randomization null and procedure; a non-rejection of exact optimizing consistency has a different meaning. The two questions should be reported separately when distinguishing compatibility from evidential strength.
Stochastic choice, households, and extensions
Population mixtures of rational types
Random utility represents observed choice frequencies by heterogeneity in underlying individual preference orderings. Consider finitely many choice environments and a finite partition of their observable alternatives into relevant patches or types. Each admissible deterministic choice type selects one patch per environment and satisfies the model's individual consistency restrictions. Stack the probabilities into a vector π and form a matrix A_R whose columns encode admissible deterministic types. If w is a nonnegative vector of population weights summing to one, then
π = A_R w.
The probability vectors consistent with this finite representation lie in the convex hull of the admissible type vectors. Under the patch construction and population assumptions, nonnegative feasibility enforces the stochastic restrictions. A statistical test examines deviations of estimated probabilities from this set, accounting for sampling error. Kitamura and Stoye (2018) develop operational algorithms and inference for this type of problem without imposing a finite-dimensional parametric distribution of preferences.
The same population mixture across environments is essential. If the distribution of preference types changes arbitrarily with prices or survey composition, distinct weights by environment can remove the linking restrictions. Endogenous expenditure and sample selection accordingly require attention. A pooled average consumption vector is neither a revealed preference ordering nor necessarily the demand of a representative rational consumer.
Household and dynamic choice
Households have several members, potentially distinct preferences, and goods that may be private or shared. A unitary household model treats observed aggregate choices as those of a single objective. Stable aggregate consistency is a test of that specification, not direct proof that household members have identical preferences. A collective model may represent bargaining and intra-household allocations under additional restrictions; the data required to identify these allocations differ from the data required for an individual GARP test.
Time introduces expectations, state dependence, habit formation, inventories, and changes in income prospects. A stable ordering over dated consumption streams need not generate the same static ordering over contemporaneous goods at every date. Observing current purchases alone generally omits the future alternatives being compared. Revealed preference analysis can be extended to suitable dynamic models by explicitly describing intertemporal resources and restrictions.
Likewise a portfolio choice can be represented as a vector of contingent payoffs. Rationality of the payoff ranking and expected-utility structure are separate restrictions: the former need not establish a single state-independent Bernoulli utility combined with a particular probability model. A higher-level model must add and test its own observable implications. These extensions illustrate the method's generality while preserving the scope of the classical theorem.
Welfare, institutional interpretation, and analytical synthesis
Preference comparisons and welfare
In the stable optimizing model, a chosen bundle is weakly preferred to all affordable alternatives, and a strictly cheaper affordable bundle is strictly inferior under local nonsatiation. For two policy environments, observed affordability relations can establish bounds on preference comparisons. Money-metric welfare comparisons require expenditure functions and a reference environment. With finite data, admissible expenditure functions can vary, so compensating and equivalent variations may be set-identified rather than uniquely determined.
If choices are affected by mistaken beliefs, contextual framing, self-control problems, or restricted attention, an adequate welfare evaluation must articulate which observed choices are normatively authoritative. Bernheim and Rangel's (2009) choice-theoretic welfare framework develops an unambiguous choice relation for behavioral settings. Its motivation is precisely the partial nature of welfare rankings when choices vary across contexts.
The predictive task is to account for choices. The welfare task is to evaluate feasible outcomes according to a specified normative criterion. The two can coincide in the classical maintained model, but behavioral applications require the relationship to be established. Observational consistency cannot justify interpersonal utility comparisons or aggregate justice judgments on its own.
Analytical synthesis
Four inferential objects organize the conclusions of this research. First, compatibility asks whether the dataset admits a specified optimizing model. Second, preference recovery asks which comparisons are shared by all compatible orderings. Third, prediction asks which choices are shared by or possible under those orderings at new opportunities. Fourth, welfare asks what can be evaluated from the recovered comparisons under an explicitly stated normative interpretation.
The mathematical examples demonstrate the logical independence of these questions. In the three-good cycle, compatibility fails even though every pair passes. In the two-good example, compatibility holds, yet observed preferences and future demand are set-identified. Neither example supports a cardinal utility measure. Empirical observations that satisfy GARP therefore deserve a result describing the comparison network and the identified set rather than an unqualified assertion of recovered preferences.
Opportunity reconstruction has a decisive role in the interpretation. A price is part of a comparative opportunity set; incomplete opportunity records can generate false comparisons. Data richness should be judged by the number and robustness of economically meaningful comparisons, not by the observation count alone. A shorter dataset containing well-designed intersections may be more discriminating than a long series of nearly identical budgets.
A rationalization result has value because it establishes what a model can explain without additional functional-form restrictions. Its strength increases when opportunities are credible, the graph contains substantial cross-observation reachability, rejection power is high against meaningful alternatives, and conclusions remain stable under plausible measurement perturbations. A model that passes under weakly informative opportunities deserves a narrower statement of evidential achievement.
These conclusions provide an analytical synthesis of existing revealed preference theory and explicit constructed examples. They do not establish a newly named theory or doctrine. They explain how to preserve the distinction between formal existence and empirical identification when classical results are applied to institutional economics, consumer welfare, financial choices, or policy evaluation.
Research frontiers and reproducible practice
A research workflow
A reproducible application begins by specifying the decision unit, commodity mapping, period, feasible-set model, stability assumptions, and treatment of ties. It preserves the underlying price and quantity data in documented units and records how expenditure or income was determined. It computes a cross-cost matrix before estimating utility and checks whether actual chosen bundles are feasible.
The deterministic test should return the weak graph, strict comparisons, GARP result, and a concrete violating path where relevant. If feasible, the Afriat inequalities can supply a witness whose residuals are checked independently. Near equality deserves a documented numerical tolerance, not silent rounding. For approximate consistency, the efficiency convention, thresholds, and missing observations belong in the method description.
Next, the researcher should report graph density, variation in opportunities, and sensitivity to measurement assumptions. Where the objective is inference, use a declared sampling or randomization model and report its restrictions. Where the objective is counterfactual prediction, append candidate budgets and compute identified sets before introducing extra preference assumptions.
Empirical comparisons require sampling weights and representativeness checks appropriate to the dataset. Panel attrition, missing prices, and household-level clustering must be handled according to the design. The monograph's own numerical examples satisfy an exact finite-data specification; they contain no inferred population estimates or simulated evidence passed off as observations.
Open questions
Several directions remain scientifically important. Transaction-level data can reveal detailed prices and product availability, but personalized offers, changing assortments, and unobserved alternatives complicate feasibility reconstruction. Larger datasets improve coverage yet make consistent commodity mapping and computational representation more demanding.
Experimental systems can select budgets that distinguish rationalization from chance, but require credible incentives and manageable cognitive complexity. Behavioral approaches can explain systematic violations; meaningful comparisons require accounting for their added flexibility and their assumptions about attention, memory, or preference change. Stochastic models can use cross-sections, while longitudinal data offer different opportunities for testing stability directly.
International research can investigate whether measured differences in consistency survive harmonization of budgets, prices, commodities, and economic institutions. Such work would require access to appropriate datasets, codebooks, and sampling designs. This monograph identifies the methodological prerequisites and evaluates selected published findings; it does not claim a new microdata study or exhaustive geographical coverage.
Conclusion
Revealed preference establishes a rigorous empirical connection between observable choice and optimization. Under finite linear budgets, strictly positive prices, stable preferences, and local nonsatiation, GARP provides the exact compatibility criterion, Afriat's inequalities turn it into linear feasibility, and their solutions construct continuous, strictly increasing, concave rationalizers. Graph methods expose the underlying cycle restrictions, including violations hidden from pairwise comparisons.
The existence of a rationalizer is an economically substantive and mathematically precise conclusion. Recovery and prediction require additional comparison coverage. The explicit two-budget example shows that two admissible utility functions may rank observed bundles differently and yield distinct optimal choices at a new budget. As a result, finite rationalization is naturally accompanied by set identification.
Applications gain force when they reconstruct actual opportunities, distinguish measurement uncertainty from behavioral instability, assess the severity and power of tests, and identify the unit of optimization. Experimental choice scores, grocery panels, and cross-sectional stochastic tests address different objects; each supports conclusions within its own assumptions and samples. Welfare interpretation requires a defensible bridge from choice and feasible opportunities to normative comparison.
The enduring contribution of revealed preference is to make optimization empirically disciplined without requiring a prespecified utility formula. The discipline is strongest when the researcher states precisely what the data establish and preserves the legitimate range of conclusions consistent with them.
References
- Afriat, S. N. (1967). The Construction of Utility Functions from Expenditure Data. International Economic Review, 8(1), 67–77. DOI: 10.2307/2525382. Source
- Bernheim, B. D., & Rangel, A. (2009). Beyond Revealed Preference: Choice-Theoretic Foundations for Behavioral Welfare Economics. Quarterly Journal of Economics, 124(1), 51–104. DOI: 10.1162/qjec.2009.124.1.51. Source
- Bronars, S. G. (1987). The Power of Nonparametric Tests of Preference Maximization. Econometrica, 55(3), 693–698. DOI: 10.2307/1913608. Source
- Cherchye, L., Demuynck, T., De Rock, B., & Lanier, J. (2025). Are Consumers (Approximately) Rational? Shifting the Burden of Proof. Review of Economics and Statistics, 107(6), 1652–1666. Earlier manuscript dated March 30, 2023. Source
- Choi, S., Kariv, S., Müller, W., & Silverman, D. (2014). Who Is (More) Rational? American Economic Review, 104(6), 1518–1550. DOI: 10.1257/aer.104.6.1518. Source
- Echenique, F., Lee, S., & Shum, M. (2011). The Money Pump as a Measure of Revealed Preference Violations. Journal of Political Economy, 119(6), 1201–1223. DOI: 10.1086/665011. Source
- Fostel, A., Scarf, H. E., & Todd, M. J. (2004). Two New Proofs of Afriat's Theorem. Economic Theory, 24(1), 211–219. DOI: 10.1007/s00199-003-0438-4. Earlier version: Cowles Foundation Discussion Paper 1415 (2003). Source
- Houthakker, H. S. (1950). Revealed Preference and the Utility Function. Economica, 17(66), 159–179. DOI: 10.2307/2549382. Source
- Kitamura, Y., & Stoye, J. (2018). Nonparametric Analysis of Random Utility Models. Econometrica, 86(6), 1883–1909. DOI: 10.3982/ECTA14478. Source
- Samuelson, P. A. (1938). A Note on the Pure Theory of Consumer's Behaviour. Economica, 5(17), 61–71. DOI: 10.2307/2548836. Source
- Tipoe, E., Adams, A., & Crawford, I. (2022). Revealed Preference Analysis and Bounded Rationality. Oxford Economic Papers, 74(2), 313–332. DOI: 10.1093/oep/gpab018. Source
- Varian, H. R. (1982). The Nonparametric Approach to Demand Analysis. Econometrica, 50(4), 945–973. DOI: 10.2307/1912771. Source