Abstract
We derive a deterministic, non-asymptotic upper bound on the Kullback-Leibler(KL) divergence of the flow-matching distribution approximation. In particular,if the $L_2$ flow-matching loss is bounded by $\epsilon^2 > 0$, then the KLdivergence between the true data distribution and the estimated distribution isbounded by $A_1 \epsilon + A_2 \epsilon^2$. Here, the constants $A_1$ and $A_2$depend only on the regularities of the data and velocity fields. Consequently,this bound implies statistical convergence rates of Flow Matching Transformersunder the Total Variation (TV) distance. We show that, flow matching achievesnearly minimax-optimal efficiency in estimating smooth distributions. Ourresults make the statistical efficiency of flow matching comparable to that ofdiffusion models under the TV distance. Numerical studies on synthetic andlearned velocities corroborate our theory.