1: \begin{abstract}

2:     We study the problem of estimating a $p$-dimensional

3:     $s$-sparse vector in a linear model with Gaussian design and

4:     additive noise. In the case where the labels are contaminated

5:     by at most $o$ adversarial outliers, we prove that the

6:     $\ell_1$-penalized Huber's $M$-estimator based on $n$ samples

7:     attains the optimal rate of convergence $(s/n)^{1/2} + (o/n)$,

8:     up to a logarithmic factor. For more general design matrices, 

9:     our results highlight the importance of two properties: the 

10:     transfer principle and the incoherence property. These 

11:     properties with suitable constants are shown to yield the 

12:     optimal rates, up to log-factors, of robust estimation with 

13:     adversarial contamination. 

14: \end{abstract}

15: