Convergence of Beliefs in Bayesian Network Games

In many contexts, players interact only with a subset of the whole population, i.e., players interact on a network. This paper a setting in which players are located on a network and play a fixed game with their neighbors. Players have incomplete information on the network structure. They have a common prior over the network, and in addition, they know the number of connections they have. That is, their type is their degree. We study the sensitivity of game-theoretic predictions to the specification of players’ beliefs. We show that two priors are close in a strategic sense if and only if they assign similar probabilities to all local events, i.e., to all events involving the types of a player and his neighbors. This means that in order to fully explore the range of possible strategic outcomes, it suffices to vary the type distribution and the correlation among player types. On the other hand, it is not enough to vary only the type distribution, which has been the focus of much of the literature so far.