Statistical Equilibrium of Optimistic Beliefs

1. Problem Statement

The paper addresses the limitations of standard game-theoretic equilibrium concepts (Nash Equilibrium and Quantal Response Equilibrium) in environments where players face dependence ambiguity regarding payoff perturbations.

• Context: In many strategic settings, agents can estimate the marginal distribution of noise (uncertainty) for each action individually (e.g., via historical data), but they cannot observe the joint dependence structure (copula) of these noise terms across different actions. This is because counterfactual payoffs (what would have happened had a different action been chosen) are unobserved.
• The Gap:
  • Nash Equilibrium (NE): Assumes perfect foresight and no noise, failing to capture bounded rationality and systematic deviations observed in experiments.
  • Quantal Response Equilibrium (QRE): Introduces noise but typically assumes specific, restrictive joint distributions (e.g., i.i.d. Gumbel shocks in Logit-QRE). This imposes the Independence of Irrelevant Alternatives (IIA) axiom, which often contradicts empirical data (e.g., "clone" effects where adding a similar option disproportionately shifts choice probabilities).
• Core Question: How can we define an equilibrium that accounts for the fact that players know the marginals of payoff shocks but are ambiguous about their dependence, and how do players resolve this ambiguity?

2. Methodology: Statistical Equilibrium of Optimistic Beliefs (SE-OB)

The authors propose SE-OB, a new equilibrium concept where players are modeled as "optimistic better responders."

A. The Two-Step Response Mapping

For a given profile of opponents' strategies, a player $j$ performs the following:

1. Optimistic Belief Selection: The player possesses a belief set $\mathcal{B}_j$ representing all plausible joint distributions of payoff perturbations $\xi$ consistent with their information (typically fixing marginals). The player selects the distribution $\mu^* \in \mathcal{B}_j$ that maximizes the expected value of the best perturbed payoff:
   $$\mu^* \in \arg\max_{\mu \in \mathcal{B}_j} \mathbb{E}_{\xi \sim \mu} \left[ \max_{i} (u_i + \xi_i) \right]$$
   This resolves ambiguity via an "optimistic" criterion (maximizing the inclusive value), akin to a "maximax" rule over the set of possible dependence structures.
2. Stochastic Choice: Given the selected $\mu^*$, the player chooses actions according to the induced random-utility choice rule. The probability of choosing action $i$ is the probability that $u_i + \xi_i$ is the maximum under $\mu^*$.

B. Equilibrium Definition

An SE-OB is a strategy profile $P^*$ where every player's mixed strategy is consistent with this two-step procedure given the opponents' strategies. It is a fixed point of the mapping from opponents' strategies to the player's optimistic best response.

C. Marginal Belief Sets (MBS) and Tractability

A key contribution is the focus on Marginal Belief Sets, where the belief set $\mathcal{B}$ contains all joint distributions with fixed marginal distributions $F_i$ but unrestricted dependence (the Fréchet class).

• Optimal Coupling: For MBS, the optimistic belief selection problem reduces to an optimal coupling problem.
• Smooth Regularized Game: The authors prove that an SE-OB with MBS is equivalent to the Nash Equilibrium of a smooth, regularized game. The expected payoff function is modified by a regularization term derived from the marginal quantile functions:
  $$\bar{u}_j(p) = u_j(p) + \sum_{i=1}^N \int_{1-p_i}^1 F^{-1}_{j,i}(t) dt$$
  This transforms the problem into finding a fixed point of a smooth best-response mapping, enabling computation and comparative statics.

3. Key Contributions

1. Theoretical Unification: SE-OB generalizes both Nash Equilibrium (when belief sets collapse to a point mass) and structural QRE (when belief sets are singletons). It provides a unified framework for equilibrium under ambiguity.
2. Existence Proof: The authors establish the existence of SE-OB under standard regularity conditions (convexity, compactness, and moment bounds on belief sets) using Kakutani's Fixed Point Theorem.
3. Computational Tractability: By characterizing SE-OB with Marginal Belief Sets as the NE of a smooth regularized game, the paper provides a tractable method for computation. The regularization term allows for the use of standard optimization algorithms.
4. Informational Foundations: The paper justifies MBS using bandit feedback models. It proves that under realized-payoff feedback (where only the chosen action's payoff is observed), the copula of payoff shocks is non-identified, while marginals are point-identified. Thus, MBS represents the sharp nonparametric identified set of feasible joint laws.
5. Empirical Flexibility: The framework naturally generates diverse choice behaviors:
   • Exponential Marginals: Recover standard Logit-QRE (Softmax).
   • Uniform Marginals: Generate Sparsemax responses, where players assign zero probability to sufficiently suboptimal actions (sparse support).
   • Pareto Marginals: Provide intermediate behaviors.

4. Key Results

• Falsifiability: While unrestricted SE-OB (like unrestricted QRE) is non-falsifiable (can rationalize any behavior), restricting players to Marginal Belief Sets restores falsifiability. The model imposes testable restrictions on behavior, particularly in constant-payoff games where the equilibrium is unique.
• Violation of IIA: SE-OB with MBS (specifically Uniform or Pareto marginals) captures systematic violations of Luce's IIA.
  • Example: In a "clone" scenario (adding identical alternatives), Logit-QRE predicts that the probability of the original superior option collapses to zero as the number of clones grows. In contrast, SE-OB with Uniform marginals predicts that the superior option retains a strictly positive probability, converging to a constant rather than zero. This aligns better with empirical observations of context effects.
• Sparsity: Unlike Logit-QRE, which always assigns positive probability to all actions (full support), SE-OB with Uniform marginals can produce sparse equilibria (zero probability for some actions) even at finite noise levels. This connects the model to bounded rationality models where agents ignore low-value alternatives.

5. Significance and Implications

• Behavioral Realism: SE-OB offers a more realistic model of decision-making under ambiguity. It acknowledges that agents may not know how uncertainties co-move but act optimistically to maximize their potential upside.
• Computational Advantage: By linking SE-OB to smooth regularized games, the paper opens the door to efficient algorithms for computing equilibria in complex strategic environments, overcoming the PPAD-completeness of general Nash Equilibrium.
• Empirical Relevance: The model explains phenomena that standard Logit models fail to capture, such as the "decoy effect" or "similarity effect" (violations of IIA) and sparse choice patterns. This makes SE-OB a powerful tool for structural econometrics and policy analysis in markets with complex substitution patterns.
• Bridge to Learning: The paper connects the equilibrium concept to learning dynamics (Algorithm A in the appendix), showing that repeated play with optimistic belief updates converges to the SE-OB, providing a dynamic justification for the static equilibrium.

In summary, this paper introduces a robust, computationally tractable, and empirically relevant equilibrium concept that bridges the gap between perfect rationality and bounded rationality by explicitly modeling how agents resolve ambiguity about the dependence of payoff shocks.