Dice Question Streamline Icon: https://streamlinehq.com

Characterize when profile and normal-form correlated equilibria coincide or separate in polytope games

Determine necessary and sufficient conditions on two-player polytope games—where each player’s action set is a convex polytope and utilities are bilinear—for which a separation exists between profile correlated equilibria (defined as correlated strategy profiles whose players each have zero profile swap regret) and normal-form correlated equilibria (defined as correlated equilibria of the vertex normal-form game and projected into the original space). Specifically, characterize the class of polytope games G for which the projection of the set of normal-form correlated equilibria equals the set of profile correlated equilibria, and the class for which there is a strict gap, beyond the known case that the gap disappears when either player’s action set is a simplex.

Information Square Streamline Icon: https://streamlinehq.com

Background

The paper introduces two distinct notions of correlated equilibrium for polytope games: normal-form correlated equilibria (NFCE), derived from the vertex normal-form game and implementable by a two-sided mediator, and profile correlated equilibria (PCE), defined by the outcomes of minimizing profile swap regret for both players. While these notions coincide in normal-form games, the authors prove a separation in general polytope games: there exist profile equilibria that cannot be implemented by any normal-form correlated equilibrium.

They also provide a partial characterization: the separation disappears when either player’s action set is a simplex, covering important subclasses such as Bayesian games with one-sided private information. Fully characterizing when these notions coincide or separate remains open and is important for understanding the relationship between mediator-implementable outcomes and those achievable by decentralized learning dynamics in structured games.

References

Of course, we can also ask for which polytope games $G$ does a separation exist like that in Theorem~\ref{thm:equilibrium-gap} -- after all, there is no separation for the case of normal-form games, which are special cases of polytope games. Fully characterizing this is an interesting open question, but below we show that this gap disappears whenever either player's action set is a simplex.

Swap Regret and Correlated Equilibria Beyond Normal-Form Games (2502.20229 - Arunachaleswaran et al., 27 Feb 2025) in Section 6, Implications for equilibrium computation