1: \begin{abstract}

2: We consider coordinate descent methods on convex quadratic problems,

3: in which exact line searches are performed at each iteration. (This

4: algorithm is identical to Gauss-Seidel on the equivalent symmetric

5: positive definite linear system.) We describe a class of convex

6: quadratic problems for which the random-permutations version of cyclic

7: coordinate descent (RPCD) outperforms the standard cyclic coordinate

8: descent (CCD) approach, yielding convergence behavior similar to the

9: fully-random variant (RCD). A convergence analysis is developed to

10: explain the empirical observations.

11: \end{abstract}

12: