1: \begin{abstract}

2: We consider the sparse regression model where the number of

3: parameters $p$ is larger than the sample size $n$. The difficulty

4: when considering high-dimensional problems is to propose estimators

5: achieving a good compromise between statistical and computational

6: performances. The Lasso is solution of a convex minimization

7: problem, hence computable for large value of $p$. However stringent

8: conditions on the design are required to establish fast rates of

9: convergence for this estimator. Dalalyan and Tsybakov

10: \cite{arnak,Dal-Tsy-2010,Dal-Tsy-2010b} proposed an exponential

11: weights procedure achieving a good compromise between the

12: statistical and computational aspects. This estimator can be

13: computed for reasonably large $p$ and satisfies a sparsity oracle

14: inequality in expectation for the empirical excess risk only under

15: mild assumptions on the design. In this paper, we propose an

16: exponential weights estimator similar to that of \cite{arnak} but

17: with improved statistical performances. Our main result is a

18: sparsity oracle inequality in probability for the true excess

19: risk.

20: \end{abstract}

21: