Correlated equilibrium

A correlated equilibrium is a generalization of Nash equilibrium introduced by Robert Aumann in 1974. In a Nash equilibrium, each player independently randomizes over their actions according to a mixed strategy. In a correlated equilibrium, players receive a private signal from a common randomizing device — a 'correlation device' — and choose actions based on their signal. The distribution of signals is such that no player can gain by unilaterally deviating from the recommended action. Every Nash equilibrium is a correlated equilibrium, but the converse is false: correlated equilibria can achieve payoffs that no Nash equilibrium can reach. The correlated equilibrium concept captures situations where players can coordinate through shared information — traffic lights, social conventions, market signals — without being able to communicate directly. It is the game-theoretic formalization of coordination through common knowledge of a random event rather than through binding agreement.

Correlated Equilibrium as Distributed Consensus

The correlated equilibrium has a deep structural affinity with problems of distributed coordination. In both settings, autonomous agents must reach coherent collective behavior without centralized control. The correlation device is not a central planner — it does not observe the state of the game or optimize on the agents' behalf. It merely broadcasts signals according to a fixed distribution. The agents, receiving private signals, each decide whether to follow the recommendation. The equilibrium condition is a form of local consistency: no agent, acting on its own information, can improve its payoff by deviation.

This decentralized character is why no-regret learning algorithms converge to correlated equilibrium. Regret matching and its variants do not require agents to know the game structure, the opponents' payoffs, or even the correlation device itself. Agents simply track their own payoffs and adjust. The fact that such myopic, local behavior converges to a correlated equilibrium is remarkable: it means that distributed, boundedly rational agents can achieve the same coordination as agents with full common knowledge of the signal structure. The gap between the bounded rationality of individual learners and the collective rationality of the equilibrium is bridged not by design but by the structure of the learning dynamics themselves.

In multi-agent reinforcement learning, this insight has practical consequences. When independent learning agents interact, their joint behavior may converge to correlated equilibria even without explicit communication protocols. The correlation is implicit in the shared environment — the common random seed, the synchronized training schedule, the shared replay buffer. These are not formally correlation devices, but they functionally approximate them. The field has begun to recognize that coordination in decentralized learning is not a communication problem to be solved; it is a statistical phenomenon to be shaped.

Information Design and the Mediator

A correlated equilibrium can also be understood through the lens of information design — the study of how a mediator, by choosing what information to reveal, can influence the outcomes of strategic interaction. The mediator does not control actions; it controls information. By selecting the signal structure, the designer can implement any correlated equilibrium that is feasible given the game's payoff structure.

This perspective reveals a continuity between correlated equilibrium and Bayesian persuasion. In both frameworks, a sender commits to an information structure, and receivers act on their private signals. The difference is strategic: in Bayesian persuasion, the sender's preferences are exogenous to the game, and the goal is to persuade; in correlated equilibrium, the mediator is neutral or absent, and the goal is coordination. Yet the mathematical structure is identical — both involve optimizing over distributions of signals subject to incentive compatibility constraints.

The information design perspective also clarifies the role of common knowledge. A correlated equilibrium does not require common knowledge of the correlation device — it requires only that each agent knows its own signal and the equilibrium condition. The common randomizing device creates correlated beliefs without requiring that agents share their beliefs explicitly. This is why correlated equilibria are more robust than Nash equilibria in settings with communication constraints: they achieve coordination with less information.

The correlated equilibrium is often treated as a technical refinement of Nash equilibrium — a curious generalization with few applications. This is backwards. Correlated equilibrium is the natural solution concept for decentralized systems: it describes what happens when autonomous agents respond to shared signals without centralized control. Nash equilibrium, by contrast, is the special case where the signals happen to be independent. The history of game theory has privileged the special case over the general one because the special case is mathematically cleaner. But the general case is empirically dominant. Every traffic light, every market signal, every social convention is a correlation device. We do not live in a world of independent randomization. We live in a world of shared noise, and correlated equilibrium is the grammar of that noise.