ResearchTrend.AI
  • Papers
  • Communities
  • Events
  • Blog
  • Pricing
Papers
Communities
Social Events
Terms and Conditions
Pricing
Parameter LabParameter LabTwitterGitHubLinkedInBlueskyYoutube

© 2025 ResearchTrend.AI, All rights reserved.

  1. Home
  2. Papers
  3. 2402.07747
22
14

Optimal score estimation via empirical Bayes smoothing

12 February 2024
Andre Wibisono
Yihong Wu
Kaylee Yingxi Yang
ArXivPDFHTML
Abstract

We study the problem of estimating the score function of an unknown probability distribution ρ∗\rho^*ρ∗ from nnn independent and identically distributed observations in ddd dimensions. Assuming that ρ∗\rho^*ρ∗ is subgaussian and has a Lipschitz-continuous score function s∗s^*s∗, we establish the optimal rate of Θ~(n−2d+4)\tilde \Theta(n^{-\frac{2}{d+4}})Θ~(n−d+42​) for this estimation problem under the loss function ∥s^−s∗∥L2(ρ∗)2\|\hat s - s^*\|^2_{L^2(\rho^*)}∥s^−s∗∥L2(ρ∗)2​ 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 ddd. 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\betaβ-H\"older continuous scores with β≤1\beta \leq 1β≤1, as well as the implication of our theory on the sample complexity of score-based generative models.

View on arXiv
Comments on this paper