Research map / Learning theory and algorithms / Sample complexity
Global Well-posedness and Convergence Analysis of Score-based Generative Models via Sharp Lipschitz Estimates
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.
Proves sharp Hessian/Lipschitz bounds on the score potential log p for diffusion models via a nonlinear HJ/heat-equation transform, shows finite-time blow-up example, and derives global well-posedness and KL convergence of the exponential-integrator sampler.
Problem
Well-posedness and convergence of score-based generative models require Lipschitz bounds on the score, which may fail (blow up in finite time or near t=0 for manifold-supported data); the paper asks how long Lipschitz bounds hold and how to get convergence guarantees.
Approach
Transform the forward Fokker-Planck density equation into a viscous Hamilton-Jacobi equation and then a heat equation, yielding pointwise Hessian bounds on log p with optimal time horizon (via convex-envelope/semiconcavity arguments), a sharp counterexample of blow-up, and O(1/t) bounds for manifold-supported data.
Key results
- Hessian bound of log p holds only for t < -log(1-1/M0) (sharp), shown by a 1D counterexample where the Hessian is unbounded at t=log 2 with M0=2.
- For convex -log p0, the Hessian of -log p is bounded globally: 0 <= -D^2 log p <= (e^{-t}L1 + 1 - e^{-t}) I.
- For compactly supported data on a d-dimensional manifold with boundary, ||D^2 qbar(t,x)|| <= C_x / t almost everywhere, and a Y-shaped domain example shows 1/t^2 blow-up on a set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use local Lipschitz estimates rather than requiring a global Hessian bound | Score-based generative models with general smooth p0 · General smooth p0 | Sampling complexity: N=\mathcal{O}(n^3\log^2 n) | Helped |
| Use the near-log-concave analysis | Exponential integrator · Near-log-concave p0 | Sampling complexity: N=\mathcal{O}(n) | Helped |
| Analyze compactly supported data on a compact smooth manifold with boundary | Score-based model score potential · Compact smooth manifold with boundary | Hessian norm bound: {C_x\over t}; {C_x/t^2} | Mixed |
Surprising findings
- The Lipschitz constant of the score can blow up in finite time even when bounded at initial time and the initial data is smooth.
- For manifold-supported data, the gradient of the score generically blows up as O(1/t), so early stopping is needed.
Limitations
- Optimal schedules cannot be determined since the bounds depend on unknown regularity constants.
- For manifold data, early-stopping time guidance is not justifiable given complex geometry.
- Convergence in global case is polynomial but with worse dimension dependence (n^3 log^2 n) and is not optimal for the non-smooth case.
Experimental setup
Baselines: Lipschitz bounds from Lemma C.9 prior work Chen et al. (2023) convergence bounds
Builds on: exponential integrator scheme score-based generative modeling (SDE) Chen et al. 2023 convergence of diffusion models
Closest papers by meaning
- Beyond Log-Concavity and Score Regularity: Improved Convergence Bounds for Score-Based Generative Models in W2-distance (ICML 2025 · similarity 0.90)
- Algorithm- and Data-Dependent Generalization Bounds for Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Score-based generative models are provably robust: an uncertainty quantification perspective (NeurIPS 2024 · similarity 0.88)
- Sampling is as easy as keeping the consistency: convergence guarantee for Consistency Models (ICML 2024 · similarity 0.87)
- Some Fundamental Aspects about Lipschitz Continuity of Neural Networks (ICLR 2024 · similarity 0.86)
- Approximation and Generalization Abilities of Score-based Neural Network Generative Models for Sub-Gaussian Distributions (NeurIPS 2025 · similarity 0.86)
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.