Statistical Properties of Rectified Flow
2511.03193v2
math.ST, cs.LG, stat.ME, stat.ML, stat.TH
2025-11-07
Авторы:
Gonzalo Mena, Arun Kumar Kuchibhotla, Larry Wasserman
Abstract
Rectified flow (Liu et al., 2022; Liu, 2022; Wu et al., 2023) is a method for
defining a transport map between two distributions, and enjoys popularity in
machine learning, although theoretical results supporting the validity of these
methods are scant. The rectified flow can be regarded as an approximation to
optimal transport, but in contrast to other transport methods that require
optimization over a function space, computing the rectified flow only requires
standard statistical tools such as regression or density estimation. Because of
this, one can leverage standard data analysis tools for regression and density
estimation to develop empirical versions of transport maps. We study some
structural properties of the rectified flow, including existence, uniqueness,
and regularity, as well as the related statistical properties, such as rates of
convergence and central limit theorems, for some selected estimators. To do so,
we analyze separately the bounded and unbounded cases as each presents unique
challenges. In both cases, we are able to establish convergence at faster rates
than the ones for the usual nonparametric regression and density estimation.