Research map / Probabilistic and scientific modeling / Variational and generative sampling
Flow matching achieves almost minimax optimal convergence
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.
Shows that conditional flow matching with Gaussian-smoothed paths attains an almost minimax optimal W_p (1≤p≤2) convergence rate for Besov-class densities, using time-partitioned neural networks and a generalization analysis.
Problem
Convergence rates of flow matching (ODE-based generative models) in Wasserstein distance were unknown; it was unclear whether FM is statistically as strong as diffusion models in minimax rate terms.
Approach
Analyze a general class of Gaussian conditional paths with mean m_t and std sigma_t (sigma_t ~ t^kappa), bound W_2 by the L2 risk of the learned vector field via an Alekseev-Gröbner argument, and combine B-spline/ReLU network approximation with time-division of [T0,1] into adaptively sized networks to reach the almost minimax rate n^{-(s+1/(2kappa)-delta)/(2s+d)}.
Key results
- Establishes that FM with sigma_t ~ t^kappa, kappa>=1/2, attains E[W_2] = O(n^{-(s+(2kappa)^{-1}-delta)/(2s+d)}), which for kappa=1/2 matches the minimax lower bound n^{-(s+1)/(2s+d)} up to arbitrarily small delta and polylog factors.
- Main result holds for all 1<=r<=2 in W_r.
- Shows that only kappa=1/2 (sigma_t ~ sqrt(t), as in diffusion models) achieves the almost optimal rate among the considered class.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use time-partitioned neural networks with variance decay σ_t ∼ √t | Flow matching · Besov density estimation | W_r convergence rate: n^{-\frac{s+1-\delta}{2s+d}}; minimax lower bound: n^{-\frac{s+1}{2s+d}} | Helped |
| increase the variance decay exponent above κ = 1/2 (ablation) | Flow matching · Besov density estimation | W_r convergence rate optimality: O\biggl(n^{-\frac{s+(2\kappa)^{-1}-\delta}{2s+d}}\biggr) | Hurt |
| remove time division and use one neural network across the time interval (ablation) | Flow matching · Besov density estimation | W_2 convergence rate: \tilde{O}(n^{-\frac{s}{2s+d}}) | Hurt |
Limitations
- Optimal rate requires partitioning time into intervals with separate networks, which is impractical and an artifact of the analysis.
- Without time division the bound is only O(n^{-s/(2s+d)}), not optimal for W_2.
- Analysis restricted to Gaussian conditional paths and Besov-space densities with assumptions A1-A5.
Experimental setup
Metrics: Wasserstein-p convergence rate
Baselines: diffusion models (Oko et al. 2023) kernel density estimation
Builds on: flow matching diffusion model convergence analysis (Oko et al.) Alekseev-Grobner lemma
Closest papers by meaning
- Theoretical guarantees in KL for Diffusion Flow Matching (NeurIPS 2024 · similarity 0.92)
- Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers (ICML 2025 · similarity 0.90)
- Optimal Flow Matching: Learning Straight Trajectories in Just One Step (NeurIPS 2024 · similarity 0.90)
- Discrete Flow Matching (NeurIPS 2024 · similarity 0.89)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.89)
- Primal-Dual Flow Matching for Sample-Wise Constrained Generation (NeurIPS 2026 · similarity 0.89)
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.