Internal correlation in repeated games (Q1814186)

This paper studies two-player, undiscounted, infinitely repeated games. Actions cannot be perfectly monitored. The paper considers an information structure in which, depending on the pair of actions selected, players either perfectly learn the pair of actions after each stage or else they learn nothing. The paper shows that if the information structure is not always trivial, that is, if there exists a pair of actions that are revealed to both players when played, then the set of Nash equilibrium payoffs is equal to the (generally larger) set of correlated equilibrium payoffs. This result is a generalization of the standard folk theorem for repeated games with perfect monitoring; in that setting both the set of Nash equilibrium payoffs and the set of correlated equilibrium payoffs agree with the set of feasible and individually rational payoffs. Under the assumptions of the paper, the first two sets coincide; they may be strictly contained in the third set. The construction used in the paper has three phases. The first phase is a correlation phase, in which players communicate in order to generate an arbitrary correlation device. In the second phase the players coordinate on a particular action pair. The third phase is a test to identify whether any player has deviated.