Local smoothness and the price of anarchy in splittable congestion games

Congestion games are multi-player games in which players' costs are additive over a set of resources that have anonymous cost functions, with pure strategies corresponding to certain subsets of resources. In a splittable congestion game, each player can choose a convex combination of subsets of resources. We characterize the worst-case price of anarchy — a quantitative measure of the inefficiency of equilibria — in splittable congestion games. Our approximation guarantee is parameterized by the set of allowable resource cost functions, and degrades with the “degree of nonlinearity” of these cost functions. We prove that our guarantee is the best possible for every set of cost functions that satisfies mild technical conditions. We prove our guarantee using a novel “local smoothness” proof framework, and as a consequence the guarantee applies not only to the Nash equilibria of splittable congestion games, but also to all correlated equilibria.