English

Optimal score estimation via empirical Bayes smoothing

Statistics Theory 2024-06-13 v2 Machine Learning Statistics Theory

Abstract

We study the problem of estimating the score function of an unknown probability distribution ρ\rho^* from nn independent and identically distributed observations in dd dimensions. Assuming that ρ\rho^* is subgaussian and has a Lipschitz-continuous score function ss^*, we establish the optimal rate of Θ~(n2d+4)\tilde \Theta(n^{-\frac{2}{d+4}}) for this estimation problem under the loss function s^sL2(ρ)2\|\hat s - s^*\|^2_{L^2(\rho^*)} that is commonly used in the score matching literature, highlighting the curse of dimensionality where sample complexity for accurate score estimation grows exponentially with the dimension dd. Leveraging key insights in empirical Bayes theory as well as a new convergence rate of smoothed empirical distribution in Hellinger distance, we show that a regularized score estimator based on a Gaussian kernel attains this rate, shown optimal by a matching minimax lower bound. We also discuss extensions to estimating β\beta-H\"older continuous scores with β1\beta \leq 1, as well as the implication of our theory on the sample complexity of score-based generative models.

Keywords

Cite

@article{arxiv.2402.07747,
  title  = {Optimal score estimation via empirical Bayes smoothing},
  author = {Andre Wibisono and Yihong Wu and Kaylee Yingxi Yang},
  journal= {arXiv preprint arXiv:2402.07747},
  year   = {2024}
}

Comments

COLT 2024; added the new results on extending to beta-Holder scores with beta <= 1

R2 v1 2026-06-28T14:46:08.079Z