From Score Matching to Diffusion: A Fine-Grained Error Analysis in the Gaussian Setting
Samuel Hurault, Matthieu Terris, Thomas Moreau, Gabriel Peyré · 2025
Read From Score Matching to Diffusion: A Fine-Grained Error Analysis in the Gaussian Setting by Samuel Hurault, Matthieu Terris, Thomas Moreau, Gabriel Peyré (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-11615.txt.
Description
Samuel Hurault and co-authors take apart the error in generative sampling, piece by piece, in the clean case where data follow a Gaussian. Modern samplers first learn the score, the slope of the smoothed log-density, then run a diffusion that turns noise into samples. Four error sources creep in along that path, from learning and from the sampler's own steps and leftover noise.
The paper writes the total sampling error, in Wasserstein distance, as an explicit tally of those sources. The data's power spectrum, which records how lopsided the spread is across directions, enters through a kernel set by the method's knobs. Sample count, step sizes, and noise level all show their price. The formula lays a base for tuning the trade-offs.
Catalog record
| CREATOR | Samuel Hurault, Matthieu Terris, Thomas Moreau, Gabriel Peyré |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 206 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-11615