What if you could predict how strategic decisions emerge in games—not just by watching what players do together, but by only seeing how often each action is chosen, in isolation? The idea sounds almost magical: could you really detect subtle forms of coordination, like correlated equilibria, from such limited information? In competitive situations from economics to sports, understanding these hidden patterns can reveal much about underlying incentives and behavior.
Short answer: In general, correlated equilibria cannot be uniquely identified using only marginal action frequencies. Marginal frequencies tell us how often each action is chosen by each player, but correlated equilibria depend on the joint distribution of actions, which contains richer information about how players coordinate or randomize their choices together. While marginal frequencies can sometimes indicate that certain equilibria are possible or rule out some options, they do not provide enough detail to uniquely determine whether play is consistent with correlated equilibrium or to reconstruct the exact form of any correlated strategy.
To see why marginal frequencies are insufficient, it helps to recall what a correlated equilibrium is. Game theory, as described by econlib.org, studies the strategic interdependence of players’ choices, where each person’s outcome depends not just on their own actions but on those of others. In classical Nash equilibrium, each player chooses their strategy independently, and the equilibrium is defined by mutual best responses. In contrast, a correlated equilibrium allows for the possibility that players’ actions can be correlated—perhaps by signals from a mediator or by shared randomization rules—so long as no player has an incentive to deviate given the advice or correlation signal.
This broader concept captures real-world situations where players might condition their actions on shared information, leading to outcomes that could not arise from independent randomization alone. For example, in a traffic intersection, a signal light can coordinate drivers’ actions so that they don’t crash; their moves are correlated, even though each only cares about their own outcome. The key technical point is that a correlated equilibrium is a joint probability distribution over action profiles that satisfies certain incentive constraints for all players.
Marginal Frequencies vs. Joint Distributions
Now, what do we mean by marginal action frequencies? Simply put, these are the observed probabilities with which each player chooses each of their possible actions, disregarding what the other players do. For example, if Alice flips a coin and chooses heads half the time and tails half the time, and Bob does the same, their marginal frequencies for heads and tails are both 0.5. But the joint distribution would tell us whether they tend to pick the same side together, or if their choices are independent.
Correlated equilibrium, however, is about the joint distribution—how often each possible combination of actions occurs, and how those combinations relate to players’ incentives. Marginal frequencies are just the “shadows” of the full joint distribution: they’re projections that ignore how actions are paired. As a result, many different joint distributions (and thus many different correlated equilibria) can have exactly the same marginal frequencies.
To put it another way: Two very different joint strategies—one where players coordinate perfectly, and one where they randomize independently—could both lead to the same marginal action frequencies. But their strategic content is totally different. This is why, as implied by the logic in econlib.org, you cannot back out all the information about players’ coordination just by looking at marginals.
Indeterminacy and Examples
Let’s illustrate this with a concrete case. Imagine a simple two-player, two-action game where each player can choose either “A” or “B”. Suppose you observe that Player 1 chooses “A” 60% of the time and “B” 40% of the time; Player 2 does the same. These are the marginal frequencies.
But what about the joint frequencies? It could be that both players always choose the same action—either both “A” or both “B”—so the joint probabilities are 0.6 for (“A”, “A”), 0.4 for (“B”, “B”), and zero for the mismatched pairs. Alternatively, perhaps their choices are completely independent: in this case, the joint probability for each pair is just the product of the marginals (0.36 for (“A”, “A”), 0.24 for (“A”, “B”), and so on). Both joint distributions are compatible with the same marginals, but only one involves perfect coordination.
The upshot is that, from marginals alone, you cannot tell whether the players are coordinating or acting independently. The same marginal frequencies could arise from a Nash equilibrium (where choices are made independently) or from a correlated equilibrium (where a mediator or signal coordinates actions), or even from other forms of joint randomization. As a result, the set of possible correlated equilibria consistent with a set of marginal frequencies is typically very large.
Limits on What Marginals Reveal
Does this mean marginal frequencies are useless? Not quite. In some cases, marginals can be informative. For example, if a player never chooses a particular action (marginal frequency zero), any correlated equilibrium must assign zero probability to that action as well. Similarly, if a player always chooses the same action, there can be no correlation with the other player’s choices. These are restrictive, but only in extreme cases.
More generally, though, the information loss is substantial. As noted in the econlib.org discussion, the “logical circle” of simultaneous-move games is squared by equilibrium concepts that rely on full knowledge of how strategies interact. Marginals, by definition, erase the details of those interactions.
Even if one could observe the marginals over many repetitions of the game, unless additional assumptions are made (such as independence of actions, or structural constraints on the form of equilibrium), it is generally impossible to reconstruct the joint distribution. Therefore, it is impossible to uniquely identify the correlated equilibrium from marginals alone.
Broader Implications and Real-World Contexts
This distinction is not just academic. In real-world strategic settings—such as auctions, traffic systems, or sports tactics—observers often have access only to aggregate statistics about individuals’ choices, not the full record of joint plays. For example, a sports analyst might know how often a quarterback chooses to run or pass, but not how those choices are paired with the defense’s strategies on each play. Without the joint data, the analyst cannot determine whether the teams are correlating their actions in subtle ways, or simply playing mixed strategies independently.
This is why, as econlib.org emphasizes, the essence of game theory is the interdependence of strategies—something that can only be fully captured by examining how choices are paired, not just how often each is made in isolation.
Technical Summary
To sum up, correlated equilibria are defined by joint probability distributions over action profiles that satisfy incentive constraints for each player. Marginal action frequencies are projections of these joint distributions onto individual players’ action spaces. While marginals can sometimes rule out certain equilibria (for instance, if some actions are never observed), they cannot, in general, identify or reconstruct the joint structure of a correlated equilibrium. The same marginal frequencies can be generated by many different joint distributions, some of which may correspond to correlated equilibria, Nash equilibria, or other forms of play.
Therefore, identifying correlated equilibria requires access to the full joint distribution of action profiles, not just the marginals. Marginals alone provide too little information to reveal the presence or precise structure of correlated strategies.
In closing, while marginal action frequencies can give hints about strategic behavior, they are only part of the picture. To truly understand how players might be coordinating—or failing to coordinate—in complex games, one must look beyond the marginals to the richer tapestry of joint actions and the incentives that bind them together. As game theory has shown, the most interesting forms of cooperation and competition are often hidden in the details of how choices intertwine, not just in the summary statistics of individual play.
