1: \begin{abstract}

2: Block-coordinate descent (BCD) is a popular framework for large-scale

3: regularized optimization problems with block-separable structure.

4: Existing methods have several limitations. They often assume that

5: subproblems can be solved exactly at each iteration, {which in

6:   practical terms usually restricts the quadratic term in the

7:   subproblem to be diagonal, thus} losing most of the benefits of

8: higher-order derivative information. Moreover, in contrast to the

9: smooth case, non-uniform sampling of the blocks has not yet been shown

10: to improve the convergence {rate bounds} for regularized

11: problems. This work proposes an inexact randomized BCD method based on

12: a regularized quadratic subproblem, in which the quadratic term can

13: vary from iteration to iteration: a ``variable metric''. We provide a

14: detailed convergence analysis for both convex and nonconvex problems.

15: Our analysis generalizes {to the regularized case Nesterov's

16:   proposal to improve convergence of BCD} by sampling proportional to

17: the blockwise Lipschitz constants. We improve the convergence rate in

18: the convex case by weakening the dependency on the initial objective

19: value. Empirical results also show that significant benefits accrue

20: from the use of a variable metric.

21: \end{abstract}

22: