Sub-Gaussian estimators of the mean of a random vector (Q1731055)

The mean $\mu$ of a random vector $X$ distributed in $\mathbb{R}^d$ is estimated based on an i.i.d. sample of $N$ points. If $X$ has a multivariate normal distribution, then the sample mean $\overline{\mu}_N$ is also multivariate normal, hence with probability at least $1-\delta,$ \[ \|\overline{\mu}_N - \mu\|\le \sqrt{\frac{\mathrm{Tr}(\Sigma)}{N}}+\sqrt{\frac{2\lambda_{\max}\log(\frac{1}{\delta})}{N}}, \] where $\Sigma$ is the covariance matrix and $\lambda_{\max}$ stands for its largest eigenvalue; see [\textit{D. L. Hanson} and \textit{F. T. Wright}, Ann. Math. Stat. 42, 1079--1083 (1971; Zbl 0216.22203)]. Similar bounds hold true if $X$ has a sub-Gaussian distribution. \par In the paper under review, a new mean estimator $\hat{\mu}_N^{(\delta)}$ is constructed that achieves purely sub-Gaussian performance under the minimal condition that the second moments of $X$ are finite. Namely, for all $N,$ with probability at least $1-\delta,$ it holds \[ \|\hat{\mu}_N^{(\delta)} - \mu\| \le C\left(\sqrt{\frac{\mathrm{Tr}(\Sigma)}{N}}+\sqrt{\frac{\lambda_{\max}\log(\frac{2}{\delta})}{N}}\right) \] for an explicit universal constant $C.$ Since the bound does not depend on the dimension $d$ explicitly, the same estimator may be defined for Hilbert-space valued random vectors and the bound remains valid as long as the strong second moment of $X$ is finite.