On the Complexity of the Optimal Correlated Equilibria in Extensive-Form Games

Abstract

A major open question in algorithmic game theory is whether normal-form correlated equilibria (NFCE) can be computed efficiently in succinct games such as extensive-form games. Motivated by this question, we study the associated Threshold problem: deciding whether there exists a correlated equilibrium whose value exceeds a given threshold. We prove that this problem is PSPACE-hard for NFCE in multiplayer extensive-form games with perfect recall, even for fixed thresholds. To contextualize this result, we also establish the complexity of the Threshold problem for Nash equilibria in this setting, showing it is ER-complete. These results uncover a surprising complexity reversal: while optimal correlated equilibria are computationally simpler than optimal Nash in normal-form games, the opposite holds in extensive-form games, where computing optimal correlated equilibria is strictly harder unless ER is equal to PSPACE. We also address a related question by von Stengel and Forges, who introduced the notions of extensive-form correlated equilibrium (EFCE) and agent-form correlated equilibrium (AFCE). They asked how difficult the Threshold problem is for AFCE; we answer this question by proving that it is NP-hard, even in two-player games without chance nodes. Complementing our hardness results, we establish tight complexity classifications for the Threshold problem across several correlated equilibrium concepts - including EFCE, AFCE, normal-form coarse, extensive-form coarse, and agent-form coarse correlated equilibria. For each of these solution concepts in multiplayer stochastic extensive-form games with perfect recall, we prove NP-completeness by providing matching NP upper bounds to the previously known hardness results. Finally, we also place the problem of computing arbitrary Nash Equilibria in extensive-form games in FixP, matching its complexity in normal-form games.