Statistical Equilibrium of Optimistic Beliefs

We study finite normal-form games in which payoffs are subject to random perturbations and players face uncertainty about how these shocks co-move across actions, an ambiguity that naturally arises when only realized (not counterfactual) payoffs are observed. We introduce the Statistical Equilibrium of Optimistic Beliefs (SE-OB), inspired by discrete choice theory. We model players as \textit{optimistic better responders}: they face ambiguity about the dependence structure (copula) of payoff perturbations across actions and resolve this ambiguity by selecting, from a belief set, the joint distribution that maximizes the expected value of the best perturbed payoff. Given this optimistic belief, players choose actions according to the induced random-utility choice rule. We define SE-OB as a fixed point of this two-step response mapping. SE-OB generalizes the Nash equilibrium and the structural quantal response equilibrium. We establish existence under standard regularity conditions on belief sets. For the economically important class of marginal belief sets, that is, the set of all joint distributions with fixed action-wise marginals, optimistic belief selection reduces to an optimal coupling problem, and SE-OB admits a characterization via Nash equilibrium of a smooth regularized game, yielding tractability and enabling computation. We characterize the relationship between SE-OB and existing equilibrium notions and illustrate its empirical relevance in simulations, where it captures systematic violations of independence of irrelevant alternatives that standard logit-based models fail to explain.

💡 Research Summary

The paper introduces a novel equilibrium concept for finite normal‑form games in which payoffs are subject to random shocks whose joint distribution across actions is unknown. While the marginal distribution of shocks for each action can often be estimated from data, the dependence structure (copula) cannot be identified because counterfactual payoffs are never observed. The authors formalize this “dependence ambiguity” and propose that each player is an “optimistic better responder”: given opponents’ mixed strategies, a player selects a joint distribution from a belief set that maximizes the expected value of the best‑perturbed payoff (i.e., the expected maximum across actions). This selection rule is called optimistic belief selection.

Once the joint distribution is chosen, the player follows the induced random‑utility choice rule: each action is chosen with the probability that it is optimal under the selected distribution. An equilibrium, called the Statistical Equilibrium of Optimistic Beliefs (SE‑OB), is a fixed point of the two‑step mapping (optimistic belief selection → stochastic choice). SE‑OB nests the Nash equilibrium (when belief sets collapse to a single distribution) and the structural Quantal Response Equilibrium (QRE) (when the belief set contains a single i.i.d. extreme‑value shock distribution).

The authors prove existence of SE‑OB under standard convexity, compactness, and moment conditions on belief sets, thereby extending the existence results for QRE to much broader specifications, including arbitrary joint laws of payoff perturbations.

A central contribution is the analysis of “marginal belief sets,” which fix the action‑wise marginal distributions while leaving the dependence unrestricted. In this case, optimistic belief selection reduces to an optimal coupling problem: among all joint distributions consistent with the given marginals, choose the one that maximizes the expected maximum payoff. The authors show that this problem is equivalent to solving a smooth regularized game whose regularizer is determined by the marginal quantile functions. Consequently, SE‑OB coincides with the Nash equilibrium of that regularized game, yielding a tractable smooth fixed‑point problem. This characterization recovers familiar choice maps such as softmax (logit) and sparsemax as special cases, and it enables efficient computation via a convergent iterative algorithm.

From an informational perspective, the paper argues that in bandit‑type environments where only realized payoffs are observed, the marginal distributions are point‑identified while the copula is not; the identified set of joint laws is precisely the Fréchet class of all couplings consistent with the marginals. SE‑OB evaluates the upper envelope of the expected maximum over this class, providing the most optimistic prediction compatible with the data.

Empirically, the authors conduct simulations showing that SE‑OB can capture systematic violations of Luce’s Independence of Irrelevant Alternatives (IIA), “clone” effects, and sparse support patterns that standard logit‑based QRE cannot explain. They also discuss falsifiability: with unrestricted belief sets SE‑OB inherits the non‑falsifiable nature of unrestricted QRE, but restricting to marginal belief sets restores testable restrictions on behavior.

Overall, the paper makes several important advances: (1) it introduces a unified framework that incorporates payoff‑perturbation ambiguity into strategic interaction; (2) it provides rigorous existence and tractability results; (3) it connects the equilibrium to optimal transport theory via optimal coupling; (4) it offers a clear informational foundation for the belief sets; and (5) it demonstrates empirical relevance by explaining phenomena that elude existing models. The work bridges behavioral game theory, econometrics of discrete choice, and optimal transport, opening new avenues for both theoretical analysis and applied modeling of strategic behavior under uncertainty.