1: \begin{abstract}

2: We propose a new first-order optimisation algorithm to solve high-dimensional non-smooth composite minimisation problems. Typical

3: examples of such problems have an objective that decomposes into a non-smooth empirical risk part and a non-smooth regularisation penalty.

4: The proposed algorithm, called Semi-Proximal Mirror-Prox, leverages the Fenchel-type representation of one part of the objective while handling the other part 

5: of the objective via linear minimization over the domain. The algorithm stands in contrast with 

6: more classical proximal gradient algorithms with smoothing, 

7: which require the computation of proximal operators at each iteration and can therefore be impractical for high-dimensional problems. 

8: We establish the theoretical convergence rate of Semi-Proximal Mirror-Prox, which exhibits the optimal complexity bounds, 

9: i.e. $O(1/\epsilon^2)$,

10: for the number of calls to linear minimization oracle. We present promising experimental results showing the interest of the approach

11: in comparison to competing methods.

12: \end{abstract}

13: