- The paper develops an axiomatic foundation that uniquely characterizes correlated equilibrium using consistency, consequentialism, and rationality.
- It demonstrates that with strong consequentialism and weak rationality, the coarse correlated equilibrium emerges, bridging behavioral and operational approaches.
- The work offers a rigorous framework integrating Bayesian updating and invariance principles, clarifying equilibrium selection in game theory.
Introduction and Motivation
The paper "Axioms for Correlated Equilibrium" (2607.06282) addresses the axiomatic characterization of correlated equilibrium (CE) and coarse correlated equilibrium (CCE) in finite normal-form games. Unlike the classical Nash equilibrium, correlated strategies allow for outcome distributions that leverage public randomization, expanding the set of solution concepts and potential equilibrium outcomes. The work systematically develops and analyzes axioms—consistency, consequentialism, rationality, and their variants—to uniquely determine CE and CCE among all solution concepts that assign sets of correlated strategies to games.
This axiomatic program is not only conceptually clarifying but also operational: selecting a solution concept is framed as a commitment to a bundle of behavioral and informational assumptions. The analysis uncovers precise coherence criteria that distinguish standard equilibrium notions from their refinements and weakenings.
Game-Theoretic Framework and Solution Concepts
The analysis assumes finite normal-form games with fixed player set and variable finite action sets. Correlated strategies are interpreted as recommendations from a correlation device; a solution concept is a map from games to convex sets of correlated strategies. Three solution concepts are pivotal:
- Nash equilibrium: Product distributions where each player's action is a best response.
- Correlated equilibrium (CE): Distributions where no player benefits by unilaterally deviating from their recommendation, conditioning on the received signal.
- Coarse correlated equilibrium (CCE): Distributions where no player benefits from committing ex ante to a fixed action (before seeing a recommendation).
The paper also formalizes operational details for cloning actions (blow-ups) and relates solution concept properties to deeper invariance and rationality structures.
The Axioms
Four main axioms and their transformations underlie the characterization:
- Consistency: If a correlated strategy is selected for each of two games and these are combined (convexly), the same strategy should be feasible in the mixture. This encapsulates the sure-thing principle under payoff uncertainty and implies expected utility maximization under Bayesian updating.
- Consequentialism: Invariance under cloning payoff-equivalent actions. If an action is duplicated (as a new label), solution concept outcomes should be unchanged except for redistribution over the clones, up to fixed ratios.
- Strong Consequentialism: Allows arbitrary splitting of probability among clones, even if this modifies the information conveyed by recommendations. This is tailored to the CCE concept, where post-recommendation conditional beliefs are ignored.
- (Weak) Rationality: Rules out recommending strictly dominated pure actions. Weak rationality only requires that dominant actions must be recommended with positive probability.
Each axiom has operational and epistemic interpretations, connecting to Bayesian rationality, information monotonicity, labeling invariance, and implementability via mediators.
Main Results
The primary result is that correlated equilibrium is the unique total, continuous, convex-valued solution concept satisfying consistency, consequentialism, and rationality. In other words, whenever a selection rule over correlated strategies respects the sure-thing principle for both payoff and information equivalence, and never recommends dominated actions, it can select only the set of CEs in each game.
Formally:
Theorem 1: The correlated equilibrium correspondence is characterized as the unique continuous, convex-valued, and total solution concept that satisfies (1) consistency, (2) consequentialism, and (3) rationality.
If consequentialism is replaced by strong consequentialism, and rationality is weakened:
Theorem 2: The unique continuous, convex-valued, and total solution concept that satisfies (1) consistency, (2) strong consequentialism, and (3) weak rationality is coarse correlated equilibrium.
This emphasizes that CCE models players who ignore the informational content of their recommendation, committing either to always follow it or to a fixed action ex ante.
Proof Structure and Noteworthy Claims
Both theorems employ reduction steps:
- Decomposition: Games are written as convex combinations of simpler games, leveraging consistency to extend the validity of strategies.
- The characterizations proceed by verifying that any solution concept respecting the axioms must, in two-player essentially zero-sum games, select all Nash equilibria and, by extension, all correlated equilibria in general games.
- For CCE, the stronger form of consequentialism and relaxed rationality allow for selection rules that ignore information revelations in the recommendation.
- The work shows that any proper coarsening or refinement will violate at least one axiom; for instance, Nash equilibrium violates convexity and strong consequentialism, while CCE violates strong rationality.
Connections to Existing Literature
The work provides a rigorous extension of previous axiomatic characterizations developed for Nash equilibrium or maximin strategies ([BrBr23a], [BrBr17c], [KoMe86a]), showing how the transition from independent strategies to arbitrary correlation demands new or modified coherence principles. The behavioral content of these axioms ties together epistemic game theory, decision under uncertainty, invariance principles, and the implementability paradigm.
Implications and Future Directions
The characterizations reveal that correlated equilibrium corresponds exactly to solution concepts built upon fully Bayesian, expected utility-maximizing agents who update conditional on recommendations and who distinguish actions only by their informational and payoff consequences. The CCE characterization shows how equilibrium notions transform when agents are assumed to disregard information extracted from recommendations.
Practically, these results provide clarity on the interpretative content of equilibrium selection in economic, computational, and mechanism design settings. They justify the robustness of CE as a behavioral prediction tool under minimal but compelling rationality and consistency assumptions.
Theoretically, the characterizations indicate what variants of the axioms force refinements or coarsenings. The open questions raised in the discussion, such as obtaining analogous results for equilibrium refinements (e.g., perfect CE), or weakening convex-valuedness and consistency to enable further solution concepts, suggest fertile ground for future research. Further, computational applications in algorithmic game theory and distributed decision processes stand to benefit from these axiomatically robust definitions.
Conclusion
This paper elucidates the logical structure and necessity of foundational equilibrium concepts in game theory by providing tight axiomatic characterizations. The main results establish correlated and coarse correlated equilibrium as uniquely and operationally justified solution concepts under sets of natural and behaviorally interpretable axioms. The analysis not only provides insight into the conceptual underpinnings of equilibrium reasoning but also points to future avenues for refining solution concepts to match diverse strategic and informational models.