Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Flow Matching is Adaptive to Manifold Structures

Shivam Kumar, Yixin Wang, Lizhen Lin

TL;DR

A principled explanation is provided for how flow matching adapts to intrinsic data geometry and circumvents the curse of dimensionality by establishing a non-asymptotic convergence guarantee for the learned velocity field and propagating this estimation error through the ODE to obtain statistical consistency of the implicit density estimator induced by the flow-matching objective.

Abstract

Flow matching has emerged as a simulation-free alternative to diffusion-based generative modeling, producing samples by solving an ODE whose time-dependent velocity field is learned along an interpolation between a simple source distribution (e.g., a standard normal) and a target data distribution. Flow-based methods often exhibit greater training stability and have achieved strong empirical performance in high-dimensional settings where data concentrate near a low-dimensional manifold, such as text-to-image synthesis, video generation, and molecular structure generation. Despite this success, existing theoretical analyses of flow matching assume target distributions with smooth, full-dimensional densities, leaving its effectiveness in manifold-supported settings largely unexplained. To this end, we theoretically analyze flow matching with linear interpolation when the target distribution is supported on a smooth manifold. We establish a non-asymptotic convergence guarantee for the learned velocity field, and then propagate this estimation error through the ODE to obtain statistical consistency of the implicit density estimator induced by the flow-matching objective. The resulting convergence rate is near minimax-optimal, depends only on the intrinsic dimension, and reflects the smoothness of both the manifold and the target distribution. Together, these results provide a principled explanation for how flow matching adapts to intrinsic data geometry and circumvents the curse of dimensionality.

Flow Matching is Adaptive to Manifold Structures

TL;DR

A principled explanation is provided for how flow matching adapts to intrinsic data geometry and circumvents the curse of dimensionality by establishing a non-asymptotic convergence guarantee for the learned velocity field and propagating this estimation error through the ODE to obtain statistical consistency of the implicit density estimator induced by the flow-matching objective.

Abstract

Flow matching has emerged as a simulation-free alternative to diffusion-based generative modeling, producing samples by solving an ODE whose time-dependent velocity field is learned along an interpolation between a simple source distribution (e.g., a standard normal) and a target data distribution. Flow-based methods often exhibit greater training stability and have achieved strong empirical performance in high-dimensional settings where data concentrate near a low-dimensional manifold, such as text-to-image synthesis, video generation, and molecular structure generation. Despite this success, existing theoretical analyses of flow matching assume target distributions with smooth, full-dimensional densities, leaving its effectiveness in manifold-supported settings largely unexplained. To this end, we theoretically analyze flow matching with linear interpolation when the target distribution is supported on a smooth manifold. We establish a non-asymptotic convergence guarantee for the learned velocity field, and then propagate this estimation error through the ODE to obtain statistical consistency of the implicit density estimator induced by the flow-matching objective. The resulting convergence rate is near minimax-optimal, depends only on the intrinsic dimension, and reflects the smoothness of both the manifold and the target distribution. Together, these results provide a principled explanation for how flow matching adapts to intrinsic data geometry and circumvents the curse of dimensionality.
Paper Structure (18 sections, 2 theorems, 27 equations, 1 figure, 3 tables)

This paper contains 18 sections, 2 theorems, 27 equations, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. List of contributions
  3. Other relevant literature
  4. Notations
  5. Flow matching
  6. Linear interpolation.
  7. Optimization.
  8. Neural network class
  9. Estimation and sampling
  10. Regularity.
  11. Theoretical results
  12. Smooth manifold.
  13. Numerical results
  14. Example target distributions
  15. Implementation details
  16. ...and 3 more sections

Key Result

Theorem 1

Let $\sd \ge 3$. Suppose $\{t_k\}$ is time grid as follows for $k = 0, 1, \ldots, \sK$. Let $\widehat{v}(\x,t)$ be estimated velocity field obtained with the empirical optimization as in eq: opt emp. Under the Assumptions assume: distribution, assume: Holder, and assume: Lipschitz, we have: Here $C > 0$ is a constant depending on $\sD$, $\sC_\cM$ and $\beta$.

Figures (1)

  • Figure 1: Comparison of generated samples and training data for \ref{['ex: floral']}. The learned flow generates samples that recover the petal geometry and place negligible mass in the regions between segments.

Theorems & Definitions (5)

  • Theorem 1: Velocity field estimation
  • Theorem 2: Main result
  • Example 1: Sphere embedded in high dimension
  • Example 2: Rotated $\sd$-torus embedded in $\bR^\sD$
  • Example 3: Floral segments embedded in $\bR^\sD$