How to Estimate the Mean of a Random Vector in Polynomial Time

Abstract: We study polynomial time algorithms for estimating the mean of a multivariate random vector under very mild assumptions: we assume only that the random vector X has finite mean and covariance. This allows for X to be heavy-tailed. In this setting, the radius of confidence intervals achieved by the empirical mean are exponentially larger in the case that X is Gaussian or sub-Gaussian. That is, the empirical mean is poorly concentrated.

We offer the first polynomial time algorithm to estimate the mean of X with sub-Gaussian-size confidence intervals under such mild assumptions. That is, our estimators are exponentially better-concentrated than the empirical mean. Our algorithm is based on a new semidefinite programming relaxation of a high-dimensional median. Previous estimators which assumed only existence of finitely-many moments of X either sacrifice sub-Gaussian performance or are only known to be computable via brute-force search procedures requiring time exponential in the dimension.

 Based on to appear in Annals of Statistics