Biggar & Papadimitriou disprove key game theory conjecture on AI learning dynamics

In a significant theoretical advance, researchers Oliver Biggar and Christos Papadimitriou have formally disproven a fundamental conjecture about how learning algorithms behave in competitive games. The work, titled 'Sink equilibria and the attractors of learning in games,' tackles the long-standing open question of characterizing the limit behavior, or 'attractors,' of learning dynamics like the replicator dynamic. It was previously conjectured that these attractors correspond perfectly to 'sink equilibria'—stable components of a game's preference graph. Biggar and Papadimitriou demonstrate this one-to-one correspondence is false, presenting three separate counterexample theorems that disprove both stronger and weaker forms of the conjecture across two-player and N-player games.

The core of the disproof hinges on a newly identified object called a 'local source'—a point within a sink equilibrium that is locally repelling, preventing it from being a true learning attractor. The authors prove that the absence of such local sources is necessary but not sufficient for the conjecture to hold. To move the field forward, they introduce a new sufficient condition called 'pseudoconvexity,' a local graph property that generalizes known cases like zero-sum and potential games where the conjecture was known to be true. This work lays out the precise obstacles for a complete theory of learning in games and provides new mathematical tools, which are critical for understanding the stability and convergence of multi-agent AI systems, from trading algorithms to autonomous systems.