gan-game-theoretic-foundations-fragile-beyond-original

IN derived (depth 1)

Created 2026-06-21T11:59:56+00:00 · Reviewed 2026-06-21T15:37:01+00:00

GAN game-theoretic foundations are fragile beyond the original formulation — equilibrium equivalence (minimax, maximin, Nash) holds only for the original game and not its variants, Nash equilibria are not guaranteed to exist in general (Farnia & Ozdaglar 2020), and the two dominant failure modes (mode collapse and vanishing gradients) represent opposed destabilizing forces that the equilibrium theory does not resolve.

Justifications

SL — three independent game-theoretic limitations compound to undermine GAN foundations beyond the original game

Antecedents (all must be IN):

IN gan-equilibria-coincide-original-only — Sequential (minimax/maximin) and Nash equilibria all coincide for the original GAN game, but this equivalence is not guaranteed for general GAN variants
IN gans-no-guaranteed-nash-equilibrium — Farnia & Ozdaglar (ICML 2020) proved that GANs do not always have Nash equilibria, establishing a theoretical limitation of adversarial training
IN gan-mode-collapse-vs-vanishing-gradient — Mode collapse (generator produces limited modes) and vanishing gradient (generator cannot learn) are opposite GAN failure modes caused by the discriminator being too weak vs. too strong, respectively

Dependents

These beliefs depend on this one:

IN optimization-landscape-determines-theoretical-robustness — Optimization landscape topology appears to influence how well ML theory generalizes beyond its original formulation — SVMs' convex objective guarantees global optimality and contributes to mathematical elegance, while GANs' minimax game-theoretic foundations are fragile beyond the original formulation (equilibrium equivalence breaks, Nash equilibria not guaranteed). This contrast suggests that convexity may be an important factor in theoretical robustness, though the evidence from two cases is insufficient to establish it as a necessary condition.