Universal Local Error and Realized Amplification for the First-Order EDM Predictor
Abstract
We analyze the first-order deterministic diffusion sampler of Karras et al. (2022), termed EDM, in 2-Wasserstein distance by separating two sources of error: local discretization error and its amplification by subsequent learned steps. We prove that local error admits a universal bound: for any data distribution with finite second moment, the one-step discretization error is quadratic in the step size, with an explicit constant that does not depend on the data distribution. Error propagation, in contrast, depends on the learned network. At high noise levels, we exploit the network parametrization of EDM to derive an explicit contraction criterion. At low noise levels, we measure propagation through the amplification realized on the distributions transported by the sampler; this realized amplification can be arbitrarily smaller than the worst-case Lipschitz constant. This analysis yields an O(eΛK/K) global discretization error for K sampling steps, where ΛK is the low-noise log-amplification. Experiments on a one-dimensional Gaussian mixture show how measured amplification accounts for slower error decay on finite sampling grids. Diagnostics on a pretrained CIFAR-10 model illustrate related stability mechanisms without certifying the global assumptions.
Related
- A multi-predictor nonparametric local linear generalized Poisson regression model for overdispersed count data.
- Quantifying the Stability of Multi-Step Reasoning via Error Amplification
- Error Amplification in Koopman Predictive Control: Computable Bounds, a Conditional Lyapunov Certificate, and Horizon-Consistent Estimation
- Universal Decoding for Unequal Error Protection
- Universal recovery in approximate quantum error correction
BibTeX
@misc{brosse2026universal,
title = {{Universal Local Error and Realized Amplification for the First-Order EDM Predictor}},
author = {Nicolas Brosse and Arnak S. Dalalyan},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.10190},
eprint = {2610.10190},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.10190},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Universal Local Error and Realized Amplification for the First-Order EDM Predictor
AU - Nicolas Brosse
AU - Arnak S. Dalalyan
PY - 2026
DA - 2026/10/07
PB - arXiv
DO - 10.48550/arxiv.2610.10190
UR - https://arxiv.org/abs/2610.10190
AB - We analyze the first-order deterministic diffusion sampler of Karras et al. (2022), termed EDM, in 2-Wasserstein distance by separating two sources of error: local discretization error and its amplification by subsequent learned steps. We prove that local error admits a universal bound: for any data distribution with finite second moment, the one-step discretization error is quadratic in the step size, with an explicit constant that does not depend on the data distribution. Error propagation, in contrast, depends on the learned network. At high noise levels, we exploit the network parametrization of EDM to derive an explicit contraction criterion. At low noise levels, we measure propagation through the amplification realized on the distributions transported by the sampler; this realized amplification can be arbitrarily smaller than the worst-case Lipschitz constant. This analysis yields an O(e^Λ_K/K) global discretization error for K sampling steps, where Λ_K is the low-noise log-amplification. Experiments on a one-dimensional Gaussian mixture show how measured amplification accounts for slower error decay on finite sampling grids. Diagnostics on a pretrained CIFAR-10 model illustrate related stability mechanisms without certifying the global assumptions.
ER -