1: \begin{abstract}

2: We obtain estimation error rates for estimators obtained by aggregation of regularized median-of-means tests, following a construction of Le Cam. The results hold with exponentially large probability,

3: % -- as in the gaussian framework with independent noise-- 

4:  under only weak moments assumptions on data.

5: %  and without assuming independence between noise and design.

6: Any norm may be used for regularization. When it has some sparsity inducing power we recover sparse rates of convergence. 

7: %

8: The procedure is robust since a large part of data may be corrupted, these outliers have nothing to do with the oracle we want to reconstruct. Our general risk bound is of order

9: \begin{equation*}

10: \max\left(\mbox{minimax rate in the i.i.d. setup}, \frac{\mbox{number of outliers}}{\mbox{number of observations}}\right)  \enspace.

11: \end{equation*}In particular, the number of outliers  may be as large as \textit{(number of data) $\times$(minimax rate)} without affecting this rate. The other data do not have to be identically distributed but should only have equivalent $L^1$ and $L^2$ moments.

12: %

13: %

14: For example, the minimax rate $s \log(ed/s)/N$ of recovery of a $s$-sparse vector in $\R^d$ is achieved with exponentially large probability by a median-of-means version of the LASSO when the noise has $q_0$ moments for some $q_0>2$, the entries of the design matrix should have  $C_0\log(ed)$ moments and the dataset can be corrupted up to $C_1 s \log(ed/s)$ outliers.

15: % and the result holds . 

16: %This result holds with exponentially large probability as if the noise and the design were i.i.d. Gaussian random variables. 

17: \end{abstract}

18: