Algorithms for Heavy-Tailed Statistics: Regression, Covariance Estimation, and Beyond

TL;DR

This paper develops polynomial-time algorithms for heavy-tailed statistical problems like covariance estimation and linear regression, achieving near-optimal error bounds under minimal distributional assumptions.

Contribution

It introduces new polynomial-time estimators that match information-theoretic error bounds for heavy-tailed data, improving upon previous computationally inefficient methods.

Findings

Spectral norm covariance estimation error: O(d^{3/4}/\u221a n)

Linear regression loss: O(d/n) with high probability

Algorithms based on degree-8 sum-of-squares semidefinite programs

Abstract

We study efficient algorithms for linear regression and covariance estimation in the absence of Gaussian assumptions on the underlying distributions of samples, making assumptions instead about only finitely-many moments. We focus on how many samples are needed to do estimation and regression with high accuracy and exponentially-good success probability. For covariance estimation, linear regression, and several other problems, estimators have recently been constructed with sample complexities and rates of error matching what is possible when the underlying distribution is Gaussian, but algorithms for these estimators require exponential time. We narrow the gap between the Gaussian and heavy-tailed settings for polynomial-time estimators with: 1. A polynomial-time estimator which takes $n$ samples from a random vector $X \in R^d$ with covariance $\Sigma$ and produces $\hat{\Sigma}$ such that in spectral norm $\|\hat{\Sigma} - \Sigma \|_2 \leq \tilde{O}(d^{3/4}/\sqrt{n})$ w.p. $1-2^{-d}$. The information-theoretically optimal error bound is $\tilde{O}(\sqrt{d/n})$; previous approaches to polynomial-time algorithms were stuck at $\tilde{O}(d/\sqrt{n})$. 2. A polynomial-time algorithm which takes $n$ samples $(X_i,Y_i)$ where $Y_i = \langle u,X_i \rangle + \varepsilon_i$ and produces $\hat{u}$ such that the loss $\|u - \hat{u}\|^2 \leq O(d/n)$ w.p. $1-2^{-d}$ for any $n \geq d^{3/2} \log(d)^{O(1)}$. This (information-theoretically optimal) error is achieved by inefficient algorithms for any $n \gg d$; previous polynomial-time algorithms suffer loss $\Omega(d^2/n)$ and require $n \gg d^2$. Our algorithms use degree-$8$ sum-of-squares semidefinite programs. We offer preliminary evidence that improving these rates of error in polynomial time is not possible in the median of means framework our algorithms employ.