SoS Certifiability of Subgaussian Distributions and Its Algorithmic Applications


Loading...

Date

2025-06

Publication Type

Conference Paper

ETH Bibliography

yes

Citations

Altmetric

Data

Abstract

We prove that there is a universal constant C > 0 so that for every d ϵ $\mathbb{N}$, every centered subgaussian distribution D on $\mathbb{R}$$^d$, and every even p ϵ $\mathbb{N}$, the d-variate polynomial (Cp)$^{p/2}$ · ||v||$_{2}^{p}$ - $\mathbb{E}$$_{X∼D}$$\langle$v,X$\rangle$$^p$ is a sum of square polynomials. This establishes that every subgaussian distribution is SoS-certifiably subgaussian - a condition that yields efficient learning algorithms for a wide variety of high-dimensional statistical tasks. As a direct corollary, we obtain computationally efficient algorithms with near-optimal guarantees for the following tasks, when given samples from an arbitrary subgaussian distribution: robust mean estimation, list-decodable mean estimation, clustering mean-separated mixture models, robust covariance-aware mean estimation, robust covariance estimation, and robust linear regression. Our proof makes essential use of Talagrand's generic chaining/majorizing measures theorem.

Publication status

published

Book title

STOC '25: Proceedings of the 57th Annual ACM Symposium on Theory of Computing

Journal / series

Volume

Pages / Article No.

1689 - 1700

Publisher

Association for Computing Machinery

Event

57th Annual ACM Symposium on Theory of Computing (STOC 2025)

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

sum-of-squares; robust statistics; subgaussian distributions; injective tensor norms

Organisational unit

Notes

Funding

815464 - Unified Theory of Efficient Optimization and Estimation (EC)

Related publications and datasets