cs.DSMay 26, 2026

Smoothed Score Queries and the Complexity of Sampling

Authors: Jingbo Liu

Organizations: Department of Statistics, University of Illinois Urbana–Champaign

Abstract

We study the query complexity of sampling from high-dimensional Gaussian distributions using gradient information. In the standard oracle model, exact gradients expose only matrix-vector products with the precision matrix, leading to polynomial approximation barriers and a characteristic κ\sqrtκ dependence on the condition number. We show that this barrier disappears when the sampler is allowed to query \emph{smoothed scores}, namely gradients of the logarithms of the Gaussian-convolved densities. For a Gaussian target with precision matrix ΛΛ, a smoothed-score query at noise level ττ gives access to the resolvent (Λ+τ1I)1(Λ+τ^{-1}I)^{-1}. Combining geometrically spaced noise levels with sinc-quadrature rational approximation, we obtain a sampler with q=O ⁣((logκ+log(ed/δTV))log(ed/δTV))q=O\!\left(\bigl(\logκ+\log(e\sqrt d/δ_{\rm TV})\bigr)\log(e\sqrt d/δ_{\rm TV})\right) smoothed-score queries for total variation error δTVδ_{\rm TV}, improving the condition-number dependence from κ\sqrtκ to logarithmic. We also study finite-bit gradient oracles. Using coordinatewise quantization of the transformed smoothed-score answers and a final dithering step, we obtain a sampling scheme whose total communicated gradient information is polylogarithmic in κκ; in particular, for fixed dimension and accuracy, the bit complexity is O(log2κ)O(\log^2κ). To complement these upper bounds, we introduce a channel-synthesis, or reverse-Shannon, converse technique for sampling lower bounds. This converts total-variation simulation guarantees into communication requirements and yields an Ω(logκ)Ω(\logκ) lower bound on the required gradient information. Together, these results identify smoothed scores as a provably more informative oracle for sampling and give nearly matching upper and lower bounds for its finite-bit complexity.

Explore similar work

CardsList