pipette
ESEspañol

Statistical Guarantees for Covariance Estimation with Adversarial Corruption under Constraints

Samuel Ozminkowski, Hongmei Jiang, Matey Neykov

PreprintClaims a big step

In the authors' words

We obtain a near-minimax rate on covariance estimation of adversarially corrupted Gaussian data in Frobenius norm when the covariance estimation is assumed to lie in a star-shaped set . Assuming a known upper bound of on the fraction of the observations of data are arbitrarily corrupted by an omniscient adversary, we obtain the near-minimax rate (up to constants) of \begin{align*} \left(\varepsilon^2 \vee \delta^{\star2} \right)\wedge d^2 \lesssim \inf_{\hat\Sigma} \sup_{\Sigma \in K} \sup_{C} \mathbb{E} \| \hat\Sigma(C(X)) - \Sigma \|_F^2 \lesssim \left(\varepsilon^2\log^2(1/\varepsilon) \vee \delta^{\star2} \right)\wedge d^2 \end{align*} where is the diameter of the set , , and \begin{align*} \delta^*=\sup\left\{ \delta\geq 0: \frac{N}{\Upsilon^2} \delta^2\leq \log M_K^{\operatorname{loc}}(\delta,c) \right\} \end{align*} for an arbitrary large constant and the local metric entropy of , . We believe this is the first such general result on structured robust covariance estimation in the Frobenius norm.

Main resultLimitation the authors admit

Appeared: Friday, October 9. arXiv. Preprint, not yet peer-reviewed.

Authors' comment: 50 pages