1: \begin{abstract}

2: In this paper, we present a new stochastic algorithm,

3: namely the stochastic block mirror descent (SBMD) method for

4: solving large-scale nonsmooth and stochastic optimization problems.

5: The basic idea of this algorithm is to incorporate the block-coordinate decomposition

6: and an incremental block averaging scheme into the classic (stochastic)

7: mirror-descent method, in order to significantly reduce the cost per

8: iteration of the latter algorithm. We establish the rate of convergence of the SBMD method

9: along with its associated large-deviation results for solving

10: general nonsmooth and stochastic optimization problems.

11: We also introduce different variants of this method and establish their

12: rate of convergence for solving strongly convex,

13: smooth, and composite optimization

14: problems, as well as certain nonconvex optimization problems.

15: To the best of our knowledge, all these developments related

16: to the SBMD methods are new in the stochastic optimization literature.

17: Moreover, some of our results also seem to be new for block coordinate

18: descent methods for deterministic optimization.

19: 

20: \vspace{.1in}

21: 

22: \noindent {\bf Keywords.} Stochastic Optimization, Mirror Descent,

23: Block Coordinate Descent, Nonsmooth Optimization, Stochastic Composite Optimization,

24: Metric Learning

25: 

26: \vspace{.1in}

27: 

28: \noindent {\bf AMS subject classifications.} 62L20, 90C25, 90C15, 68Q25

29: \end{abstract}