An approximate folk theorem with imperfect private information (Q1178808)

This paper proves a partial folk theorem for repeated games with imperfect information. It studies games with complete but imperfect information; it requires that information be linked in the sense that no individual player can prevent any other pair of players from communicating indirectly through their action choices. An individually rational payoff vector \((v_ 1,\dots,v_ n)\) is one in which there is a mutual punishment action such that for each \(i\), player \(i\) cannot do better than \(v_ i\) if all other agents use their punishment action. A payoff is enforceable if there is an action profile that yields the payoff and such that any unilateral deviation that yields a higher payoff of at least one other player. The folk theorem is an approximate one: It states that in informationally linked games any individually rational and enforceable payoff is the limit of \(\varepsilon\)-equilibria of discounted games as \(\varepsilon\) goes to zero and the discount factor goes to one. The folk theorem is partial in the sense that the set of enforceable payoffs may be strictly smaller than the set of feasible payoffs. The proof consists of a multi-part construction. In the first part, players play to obtain the target payoff. In the second part, players test to see if there have been statistically important deviations from the target. In the third part, the results of the second part are communicated to all of the players. If a deviation from the target has been detected, then players follow a punishment strategy for the rest of the game.