Sample Complexity Bounds for Robust Mean Estimation with Mean-Shift Contamination

TL;DR

This paper investigates the sample complexity of robust mean estimation under mean-shift contamination for general distributions, providing new algorithms and bounds using Fourier analysis.

Contribution

It introduces a spectral condition-based algorithm and matching lower bounds for mean estimation under mean-shift contamination, extending prior results beyond Gaussian and Laplace distributions.

Findings

Existence of a sample-efficient algorithm under mild spectral conditions

Matching lower bounds on sample complexity

Use of Fourier analysis and Fourier witness concept

Abstract

We study the basic task of mean estimation in the presence of mean-shift contamination. In the mean-shift contamination model, an adversary is allowed to replace a small constant fraction of the clean samples by samples drawn from arbitrarily shifted versions of the base distribution. Prior work characterized the sample complexity of this task for the special cases of the Gaussian and Laplace distributions. Specifically, it was shown that consistent estimation is possible in these cases, a property that is provably impossible in Huber's contamination model. An open question posed in earlier work was to determine the sample complexity of mean estimation in the mean-shift contamination model for general base distributions. In this work, we study and essentially resolve this open question. Specifically, we show that, under mild spectral conditions on the characteristic function of the (potentially multivariate) base distribution, there exists a sample-efficient algorithm that estimates the target mean to any desired accuracy. We complement our upper bound with a qualitatively matching sample complexity lower bound. Our techniques make critical use of Fourier analysis, and in particular introduce the notion of a Fourier witness as an essential ingredient of our upper and lower bounds.