1: \begin{abstract}

2: Variants of the coordinate descent approach for minimizing a nonlinear

3: function are distinguished in part by the order in which coordinates

4: are considered for relaxation. Three common orderings are cyclic

5: (CCD), in which we cycle through the components of $x$ in order;

6: randomized (RCD), in which the component to update is selected

7: randomly and independently at each iteration; and random-permutations

8: cyclic (RPCD), which differs from CCD only in that a random

9: permutation is applied to the variables at the start of each cycle.

10: Known convergence guarantees are weaker for CCD and RPCD than for RCD,

11: though in most practical cases, computational performance is similar

12: among all these variants. There is a certain family of quadratic

13: functions for which CCD is significantly slower than for RCD; a recent

14: paper of Sun and Ye has explored the poor behavior of CCD on this

15: family.  The RPCD approach performs well on this family, and this

16: paper explains this good behavior with a tight analysis.

17: \end{abstract}

18: