Generative Adversarial Nets

What problem does this paper solve?

Generative modelling without a tractable density. A GAN's generator is a deterministic map $G:\R^k\to\Xcal$ pushing a simple noise distribution forward; its density is not available in closed form, and maximum likelihood is therefore unusable.

What mathematical objects are introduced?

The minimax objective

\min_G\max_D\ \E_{x\sim p_{\text{data}}}\big[\log D(x)\big] + \E_{z\sim p_z}\big[\log\big(1-D(G(z))\big)\big].

What assumptions are made?

That $D$ and $G$ have unlimited capacity and that the inner maximisation is solved exactly — both required for the equilibrium analysis and both false in practice.

What equations matter?

For fixed $G$ with induced distribution $p_G$ , the optimal discriminator is

D^\star(x) = \frac{p_{\text{data}}(x)}{p_{\text{data}}(x)+p_G(x)},

obtained by maximising the integrand pointwise. Substituting back gives the generator's objective as

2\,\mathrm{JSD}(p_{\text{data}}\,\|\,p_G) - \log4 ,

minimised exactly when $p_G = p_{\text{data}}$ .

What theorem or result is important?

The equilibrium characterisation above: the unique global optimum of the game is $p_G = p_{\text{data}}$ with $D^\star\equiv\tfrac12$ . This is a genuine theorem, and its proof is the pointwise maximisation plus the definition of Jensen–Shannon divergence.

What proof techniques are used?

Pointwise optimisation of an integral; convexity of the objective in $p_G$ for the convergence argument.

What prior mathematics is required?

KL and Jensen–Shannon divergences; minimax problems and the gap between $\min\max$ and $\max\min$ ; pushforward measures.

Limitations

The theorem is about the equilibrium of the idealised game. Gradient descent–ascent on a non-convex non-concave objective need not converge to it, and does not: mode collapse, oscillation and discriminator overpowering are the standard failure modes. When $p_G$ and $p_{\text{data}}$ have disjoint support the JSD is constant and the gradient vanishes — the observation that motivated Wasserstein GANs.

Why it mattered

It introduced adversarial training, dominated image generation for five years, and remains the canonical example of a learning objective defined by a game rather than a likelihood. Its displacement by diffusion models was driven precisely by the training instability the theory did not control.