1: \begin{abstract}

2:  This paper deals with large-scale composite  optimization problems having  the  objective function formed as a sum of two  terms, one has Lipschitz continuous gradient along random subspaces and may be nonconvex and the second  term is simple, but  possibly nonconvex and \textit{nonseparable}.  Under these settings we design a stochastic  coordinate proximal gradient method which takes into account the nonseparable composite form of the objective function. This algorithm achieves scalability by constructing at each iteration a local approximation model of the nonseparable  objective function along  a random  subspace with user-determined dimension.  We outline efficient techniques for selecting the random subspace, yielding an implementation that has  low cost per-iteration  while also achieving fast convergence rates.  We present a probabilistic worst-case complexity analysis for our stochastic  coordinate proximal gradient method in convex and nonconvex settings, in particular we prove high-probability bounds on the number of iterations before a given optimality is achieved.  To the best of our knowledge, this work is the first proposing a pure stochastic  coordinate descent algorithm which is supported by global efficiency estimates for general  nonseparable  composite optimization problems.   Preliminary numerical results also  confirm the  efficiency of our  algorithm.  

3: 

4: \keywords{composite minimization \and nonseparable objective function  \and stochastic coordinate descent \and convergence rates}

5: \end{abstract}

6: