This document discusses generative adversarial nets (GANs). GANs use an adversarial modeling framework where a generator and discriminator are trained against each other. The generator learns to generate fake samples from noise to match the real data distribution, while the discriminator learns to distinguish real from fake samples. They are trained together through a minimax game, with the generator trying to maximize the discriminator's errors. The document proves that the global minimum of the GAN's training criterion is achieved when the generator's distribution pg matches the real data distribution pdata, with the criterion value reaching -log4.