1: \begin{abstract}

2: In this paper we {\bl consider} a parallel version of  a randomized

3:  (block) coordinate descent method for minimizing the sum of a

4: partially separable smooth convex  function and a fully separable

5: non-smooth convex function. Under the assumption of  Lipschitz

6: continuity of the gradient of the  smooth function, {\bl this method

7: has} a sublinear convergence rate. Linear convergence rate of the

8: method is obtained for the newly introduced class of

9: \textit{generalized error bound functions}. We prove that the new

10: class of generalized error bound functions encompasses both

11: global/local error bound functions and smooth strongly convex

12: functions.  We also show that the theoretical estimates on the

13: convergence rate depend on the number of blocks chosen randomly and

14: a natural measure  of  separability of the objective function.

15: Numerical simulations are also provided to confirm our theory.

16: \end{abstract}