Research map / Probabilistic and scientific modeling / Variational and generative sampling
An Error Analysis of Flow Matching for Deep Generative Modeling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Provides the first end-to-end statistical error analysis of flow matching (CNFs with least-squares velocity regression), giving Wasserstein-2 consistency for bounded-support targets and an improved rate under a Lipschitz target-score assumption.
Problem
Theoretical guarantees are lacking for end-to-end generative modeling with flow matching: learning the velocity field by deep least-squares regression and then generating samples via ODE integration, with finite samples from the target.
Approach
Shows that the true velocity field of flow matching is Lipschitz under mild target-distribution assumptions (bounded support, or Lipschitz score with a strongly log-concave-type potential), approximates it with a ReLU network with bounded Lipschitz constant, bounds the generalization error of ERM on the regression loss, and propagates the L2 velocity error to W2 via Grönwall and Euler discretizati
Key results
- For bounded-support targets, W2 between generated and target distributions converges to zero in probability (Theorem 1.6, consistency).
- Under an additional Lipschitz target-score assumption, W2 rate is e O(n^{-1/(3(d+5))}) with probability at least 1-1/n (Theorem 1.7).
- Gives the first explicit sample complexity bound for flow matching velocity-field estimation, e O(n^{-2/(d+5)}) for the L2 velocity error (Theorem 4.4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Assume the target score function is Lipschitz, in addition to bounded support | Flow Matching · Target distribution | Wasserstein-2 distance convergence rate: eO(n−1/(3(d+5))) | Helped |
Surprising findings
- Theorem 4.4 becomes vacuous as the stopping time T approaches 1 at fixed sample size, reflecting a blow-up of the velocity field near t=1, which creates a stopping-time tradeoff.
Limitations
- Results are purely theoretical with no experiments.
- Bounds depend on the dimension d and rates degrade exponentially with dimension (exponent 1/(d+5)).
- Analysis requires a stopping time T<1, and the bounds become vacuous as T approaches 1 at fixed n.
Experimental setup
Metrics: Wasserstein-2 distance L2 velocity error sample complexity rate
Baselines: Albergo & Vanden-Eijnden 2023 Chen et al. 2023c Lu et al. 2022 Chen et al. 2023a
Builds on: Flow Matching Continuous Normalizing Flows Chen et al. 2023 score-based ODE analysis Albergo & Vanden-Eijnden 2023 stochastic interpolants Lu et al. 2022
Closest papers by meaning
- Reflected Flow Matching (ICML 2024 · similarity 0.91)
- Markovian Flow Matching: Accelerating MCMC with Continuous Normalizing Flows (NeurIPS 2024 · similarity 0.90)
- On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity (NeurIPS 2025 · similarity 0.89)
- Improving Flow Matching by Aligning Flow Divergence (ICML 2025 · similarity 0.89)
- Optimal Flow Matching: Learning Straight Trajectories in Just One Step (NeurIPS 2024 · similarity 0.89)
- A Theory on Flow Matching with Neural Networks (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.