1: \begin{abstract}

2: We consider the problem of minimizing the sum of two convex functions:

3: one is smooth and given by a gradient oracle, and the other is separable

4: over blocks of coordinates and has a simple known structure over each block.

5: We develop an accelerated randomized proximal coordinate gradient (APCG) 

6: method for minimizing such convex composite functions. 

7: For strongly convex functions, our method achieves faster linear convergence 

8: rates than existing randomized proximal coordinate gradient methods.

9: Without strong convexity, our method 

10: enjoys accelerated sublinear convergence rates.

11: We show how to apply the APCG method to solve the regularized 

12: empirical risk minimization (ERM) problem, and devise efficient implementations

13: that avoid full-dimensional vector operations.

14: For ill-conditioned ERM problems, our method obtains improved convergence rates

15: than the state-of-the-art stochastic dual coordinate ascent (SDCA) method.

16: \end{abstract}