Divergence is Uncertainty: A Closed-Form Posterior Covariance for Flow Matching

2026-05-01

AI Summary

This work addresses the challenge of efficiently and accurately quantifying sample uncertainty in flow-matching generative models. The authors propose a closed-form posterior covariance estimation method that requires neither model retraining nor architectural modifications. By discovering and rigorously proving an exact identity—termed the divergence–uncertainty identity—between the divergence of the velocity field and the posterior covariance, the method enables end-to-end uncertainty estimation via a single forward pass. Leveraging differential equation theory, the approach derives a closed-form solution based on the velocity field’s divergence and Jacobian, applicable to any pretrained flow-matching model. On MNIST, the resulting pixel-level uncertainty maps exhibit semantically meaningful patterns, precisely capturing variations along digit boundaries, while incurring only about one ten-thousandth of the computational cost of ensemble or Monte Carlo Dropout methods and achieving comparable error prediction performance.

Abstract

Flow matching has become a leading framework for generative modeling, but quantifying the uncertainty of its samples remains an open problem. Existing approaches retrain the model with auxiliary variance heads, maintain costly ensembles, or propagate approximate covariance through many integration steps, trading off training cost, inference cost, or accuracy. We show that none of these trade-offs is necessary. We prove that, for any pre-trained flow matching velocity field, the trace of the posterior covariance over the clean data given the current state equals, in closed form, the divergence of the velocity field, up to a known time-dependent prefactor and an additive constant. We call this the \emph{divergence-uncertainty identity} for flow matching. The matrix-level form of the identity is similarly closed-form, depending solely on the velocity Jacobian. Because the identity is exact and post-hoc, it is computable on any pre-trained flow matching model, with no retraining and no architectural modification. For one-step generators such as MeanFlow, the same identity yields the exact end-to-end generation uncertainty in a single forward pass, eliminating the multi-step variance propagation required by all prior methods. Experiments on MNIST confirm that the resulting per-pixel uncertainty maps are semantically meaningful, concentrating on digit boundaries where inter-sample variation is highest, and that the scalar uncertainty score tracks actual prediction error, all at roughly 10,000$\times$ less total compute than ensembling or Monte Carlo dropout.

Problem

• flow matching
• uncertainty quantification
• posterior covariance
• generative modeling
• divergence

Innovation

• flow matching
• posterior covariance
• divergence-uncertainty identity
• uncertainty quantification
• closed-form solution