1: \begin{abstract}

2: The last several years has witness the huge success on the stochastic variance reduction method  in the finite sum problem.  However assumption on strong convexity to have linear rate limits its applicability. In particular, it does not include several important formulations such as Lasso, group Lasso, logistic regression,  and some non-convex models including corrected Lasso and SCAD. In this paper, we prove that, for a class of statistical M-estimators covering examples  mentioned above,  SVRG solves the formulation with {\em a linear convergence rate}  without strong convexity or even convexity.

3: Our analysis makes use of {\em restricted strong convexity}, under which we show that SVRG converges linearly to   the fundamental statistical precision of the model, i.e., the difference between true unknown parameter $\theta^*$ and the optimal solution $\hat{\theta}$ of the model.

4: \end{abstract}