gan-stable-convergence-guaranteed

Antecedents (all must be IN):

• IN gan-equilibrium-generator-matches-data — At the unique GAN equilibrium, the generator matches the data distribution exactly (μ̂_G = μ_ref), the discriminator outputs 1/2 everywhere, and the objective value is −2 ln 2
• IN gan-ttur-convergence — The Two Time-Scale Update Rule (TTUR) sets the generator's learning rate lower than the discriminator's and provably converges to a stationary local Nash equilibrium, but does not prevent mode collapse
• IN gan-asymptotically-consistent — GANs are asymptotically consistent estimators of the data distribution due to the universal approximation theorem for neural networks

Unless (any of these IN defeats this justification):

• 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