This paper presents a new, probabilistic model of learning in games which investigates the often stated intuition that common knowledge of strategic intent may arise from repeated interaction. The model is set in the usual repeated game framework, but the two key assumptions are framed in terms of the likelihood of beliefs and actions conditional on the history of play. The first assumption formalizes the basic intuition of the learning approach; the second, the indeterminacy that inspired resort to learning models in the first place. Together the assumptions imply that, almost surely, play will remain almost always within one of the stage game's minimal inclusive sets. In important classes of games, including those with strategic complementarities, potential functions, and bandwagon effects, all such sets are singleton Nash.

We consider an infinitely repeated Bertrand game, in which prices are publicly observed and each firm receives a privately observed, i.i.d. cost shock in each period. We focus on symmetric perfect public equilibria, wherein any “punishments” are borne equally by all firms. We identify a tradeoff that is associated with collusive pricing schemes in which the price to be charged by each firm is strictly increasing in its cost level: such “fully sorting” schemes offer efficiency benefits, as they ensure that the lowest-cost firm makes the current sale, but they also imply an informational cost (distorted pricing and/or equilibrium-path price wars), since a higher-cost firm must be deterred from mimicking a lower-cost firm by charging a lower price. A rigid-pricing scheme, where a firm's collusive price is independent of its current cost position, sacrifices efficiency benefits but also diminishes the informational cost. For a wide range of settings, the optimal symmetric collusive scheme requires (i) the absence of equilibrium-path price wars and (ii) a rigid price. If firms are sufficiently impatient, however, the rigid-pricing scheme cannot be enforced, and the collusive price of lower-cost firms may be distorted downward in order to diminish the incentive to cheat. When the model is modified to include i.i.d. public demand shocks, the downward pricing distortion that accompanies a firm's lower-cost realization may occur only when current demand is high.