Axiom of Revealed Stochastic Preference

• ARSP is a formal criterion that defines when observed stochastic choice probabilities can be rationalized by mixtures over deterministic utility-maximizing types.
• It translates behavioral rationalizability into a comprehensive set of linear inequalities, enabling rigorous empirical tests via convex hull analysis.
• Extensions like GARSP and WARSP broaden ARSP to accommodate set-valued and non-transitive preferences, expanding its applicability in economic and behavioral research.

The Axiom of Revealed Stochastic Preference (ARSP) provides a necessary and sufficient condition for rationalizing observed stochastic choice probabilities on finite choice problems via random utility models (RUMs). ARSP characterizes exactly when such choice probabilities can be interpreted as arising from a population mixture over deterministic, utility-maximizing types on a finite universal choice space. In stochastic choice settings, ARSP generalizes the classical revealed preference approach by translating behavioral rationalizability into an exhaustive set of linear inequalities, each corresponding to potential sub-collections of choice-problem outcomes. This axiom, and its generalizations, underpin empirical and theoretical analysis of random utility models and have been further extended to accommodate set-valued choice correspondences and non-transitive preference frameworks.

1. Formal Setup and Notational Framework

Consider a finite universal choice set $X = \{x_1, \dots, x_K\}$ and a set of observed choice problems $\{C_1, \dots, C_J\}$, where each $C_j \subset X$. For each $C_j$ is observed a probability vector $\pi(C_j) \in \Delta^{|C_j|-1}$, where $\Delta^{|C_j|-1}$ denotes the simplex of dimension $|C_j|-1$. All choice probabilities are concatenated into a vector $\Pi \in [0, 1]^I$ where $I = \sum_{j=1}^J |C_j|$, and each entry $\Pi_i$ gives the observed probability that a particular alternative $x \in C_j$ is chosen from problem $C_j$, with the normalization $\sum_{i=1}^I \Pi_i = J$.

Deterministic rationalizable types (choice functions) consistent with utility maximization correspond to binary vectors $R \in \{0,1\}^I$ with exactly one 1 in each block of indices for each $C_j$. The finite set of such types is denoted $\mathcal{R} \subset \{0,1\}^I$. Under a random utility model, the set of all stochastically rationalizable choice probabilities is precisely $\text{co}(\mathcal{R})$, the convex hull of $\mathcal{R}$, i.e., mixtures over deterministic types (Stoye, 2018).

2. Statement and Interpretation of the ARSP

To express ARSP, a "trial" $T \in \{0,1\}^I$ is defined as a vector with support confined to a single choice problem $C_j$ and can indicate any subset of alternatives within that $C_j$. Let $\mathcal{T}$ denote the set of all such trials. ARSP states:

Axiom of Revealed Stochastic Preference (ARSP): For every finite sequence of trials $(T_1, \dots, T_M) \in \mathcal{T}^M$,

$$\sum_{m=1}^M T_m \cdot \Pi \le \max_{R \in \mathcal{R}} \sum_{m=1}^M T_m \cdot R.$$

Here, $T \cdot \Pi$ is the standard inner product in $\mathbb{R}^I$. Equivalently, the total weight on any combined set of trial outcomes in the observed probabilistic data never exceeds that placed on these outcomes by any deterministic type $R \in \mathcal{R}$. This system of inequalities exhaustively characterizes membership of $\Pi$ in the convex hull $\text{co}(\mathcal{R})$ (Stoye, 2018).

3. Necessity and Sufficiency: Hyperplane Separation Argument

The proof of necessity and sufficiency of ARSP is a direct application of the hyperplane separation theorem for convex hulls. Since any nonnegative integer linear combination of trials can be written as a single vector $T \in \mathbb{N}^I \cap \mathbb{R}_{+}^I$, and all such $T$ can be covered by repeated trials, ARSP is equivalent to

$T \cdot \Pi \le \max_{R \in \mathcal{R}} T \cdot R$

for all $T \in \mathbb{N}^I \cap \mathbb{R}_{+}^I$. Scaling extends this equivalence to all $T \in \mathbb{R}_{+}^I$, and a translation argument (by adding a positive multiple of the all-ones vector) extends it further to all $T \in \mathbb{R}^I$. Thus, if ARSP holds for all these $T$, then $\Pi$ cannot be separated from $\text{co}(\mathcal{R})$ by any hyperplane, which implies $\Pi \in \text{co}(\mathcal{R})$ (Stoye, 2018).

4. Implications for Utility Representation and Empirical Methods

Carathéodory’s theorem ensures that any $\Pi \in \text{co}(\mathcal{R})$ admits a finite convex combination

$$\Pi = \sum_{k=1}^{K'} \lambda_k R^k, \quad \lambda_k \ge 0,~\sum_{k} \lambda_k = 1,$$

with each $R^k \in \mathcal{R}$ representing a deterministic utility maximizer. Thus, ARSP is both necessary and sufficient for representing observed stochastic choice data as a random utility model—a population (or sequence) of utility maximizers—with an empirical mixture over discrete types (Stoye, 2018).

Empirically, this means that validating ARSP amounts to verifying a finite (exponentially many in the worst case) set of linear inequalities. In small cases, these reduce to classical “triangle” or “block–Marschak” inequalities. Failure of any single ARSP inequality provides a certificate that observed data are not rationalizable by a random utility model (Stoye, 2018).

5. Extensions: GARSP for Set-Valued Choice and Non-Transitive Preferences

The framework generalizes to set-valued choice correspondences via the Generalized Axiom of Revealed Stochastic Preference (GARSP). Here, at each choice problem $C_j$, the decision-maker may select any subset $S \subseteq C_j$, not restricted to singletons. The lifted choice universe $\tilde{X} = 2^X$ and corresponding lifted trials $\tilde{T} \in \{0,1\}^{\tilde{I}}$ yield the GARSP axiom: $$\sum_{m=1}^M \tilde{T}_m \cdot \tilde{\Pi} \le \max_{\tilde{R} \in \tilde{\mathcal{R}}} \sum_{m=1}^M \tilde{T}_m \cdot \tilde{R}$$ where $\tilde{R}$ runs over deterministic set-valued choice correspondences. The collection of allowable trials $\tilde{T}$ must be the full set of binary vectors supported within each lifted choice problem to guarantee a complete half-space description of the polytope of rationalizable probabilities (Stoye, 2018).

Recent contributions extend the methodology to random preference models (RPM) capable of accommodating non-transitive preferences, as detailed in Youmbi (2024) (Youmbi, 2024). In this generalization, called the Weak Axiom of Revealed Stochastic Preference (WARSP), the set of deterministic types underlying the mixture expands from SARP-consistent utility maximizers to WARP-consistent but possibly non-transitive monotone preferences. The analogous system of inequalities exactly characterizes when observed choice data are rationalizable under such a generalized RPM.

The following table summarizes model classes and their associated stochastic revealed preference axioms:

Model Class	Deterministic Types	Stochastic Rationalizability Criterion
RUM (singly-valued, transitive)	SARP-consistent utility maximizers	ARSP (Stoye, 2018)
RPM (possibly non-transitive)	Monotone, WARP-consistent preferences	WARSP (Youmbi, 2024)
Set-valued stochastic choice	Set-valued rational maximizers	GARSP (Stoye, 2018)

6. Statistical Testing and Empirical Relevance

In empirical applications, ARSP and its extensions determine the feasibility of a random utility representation via finite systems of inequalities. Statistical tests such as the cone-projection test of Kitamura–Stoye (2018) are employed to assess hypotheses of stochastic rationalizability by minimizing distances from observed choice frequency vectors to the relevant convex hull defined by deterministic types. In analyses of UK Family Expenditure Survey data, the classical RUM (ARSP/SARP) is frequently rejected (with p-values below 5%), while the non-transitive RPM (WARSP) is not rejected, evidencing empirical relevance for the non-transitive extension (Youmbi, 2024).

A plausible implication is that market data often demonstrate patterns (preference cycles or violations of stochastic transitivity) inconsistent with standard RUMs but consistent with more flexible RPMs governed by WARSP. The ARSP/GARSP/WARSP family thus provides unified, computationally explicit criteria for stochastic rationalizability, encompassing both classical and modern relaxations of economic rationality (Stoye, 2018, Youmbi, 2024).

7. Theoretical and Practical Significance

The ARSP (and its generalizations) unifies and simplifies the foundational analysis of stochastic choice, reducing it to constructive polyhedral characterization. This enables detailed structural and inferential tests for rationalizability in microeconomic demand analysis, market design, and behavioral economics. Furthermore, extensions such as WARSP establish a formal bridge to emerging models of bounded and non-transitive rationality, broadening the empirical applicability of stochastic revealed preference analysis without sacrificing computational tractability (Stoye, 2018, Youmbi, 2024).