← All papers
First page of Universality and Convergence of Generative Flows

Universality and Convergence of Generative Flows

Leo Brunswic

cs.LG Oct 4, 2026 · v2 math.PR
Each theorem about generative flow balance losses has a certification status computed from an accompanying Lean 4 development.
Generative flows sample from an unnormalized target by training a flow to be balanced, and the training loss is the signal a practitioner watches. We ask what that signal is worth: whether a small loss certifies an accurate sampler, whether the loss can be driven to zero, and how fast gradient descent does so. The loss decides the first. Losses that compare the two sides of the balance by their difference bound, in total variation, the error of the sampler the flow implies, with explicit constants that do not involve the policy; flow-matching losses that compare them through a ratio admit no such bound, already on a single cycle, whenever their generator is continuous at balance. On graphs, the backward policy decides the other two. Once it is frozen, balance becomes invariance under the backward chain, so that existence is free on finite graphs, and one constant — the norm of that chain's Green operator, which plays the role of an inverse spectral gap — fixes the order of the curvature of the loss around the balanced flow, from above and below, and sets a floor under the rate at which training converges near it. The mechanism is that gradient descent diffuses the flow along the backward policy. For the squared-logarithm generator of detailed and trajectory balance, training the balance loss on states converges globally on every finite path-connected graph, from every positive initialization. The constant can be infinite while backward trajectories are short on average, and exact flow matching can then fail. The bounds and rates are tested by exact computation on enumerable state spaces, and every theorem carries a certification status computed from a Lean 4 development.

Generative flows are trained to be balanced in order to sample from unnormalized targets. It is unclear whether a small training loss certifies an accurate sampler, whether the loss can reach zero, and how fast gradient descent converges.

The paper separates difference-based losses from ratio-based flow-matching losses and derives total-variation bounds for the difference-based ones. On graphs with a frozen backward policy, balance is recast as invariance under the backward chain. The norm of that chain's Green operator is identified as the constant governing loss curvature and convergence rate. The results are checked by exact computation on enumerable state spaces, and each theorem's certification status is computed from a Lean 4 development.

Difference losses bound sampler error in total variation with policy-independent constants. Ratio-based flow-matching losses admit no such bound, already on a single cycle. For the squared-log generator, state-level balance training converges globally on finite path-connected graphs from every positive initialization. Exact flow matching can fail when the Green-operator constant is infinite.