1: \begin{abstract}

2: Recently, heuristics based on the Douglas-Rachford splitting algorithm

3: and the alternating direction method of multipliers (ADMM) have found

4: empirical success in minimizing convex functions over nonconvex sets,

5: but not much has been done to improve the theoretical understanding

6: of them. In this paper, we investigate convergence of these heuristics.

7: First, we characterize optimal solutions of minimization problems

8: involving convex cost functions over nonconvex constraint sets. We

9: show that these optimal solutions are related to the fixed point set

10: of the underlying nonconvex Douglas-Rachford operator. Next, we establish

11: sufficient conditions under which the Douglas-Rachford splitting heuristic

12: either converges to a point or its cluster points form a nonempty

13: compact connected set. In the case where the heuristic converges to

14: a point, we establish sufficient conditions for that point to be an

15: optimal solution. Then, we discuss how the ADMM heuristic can be constructed

16: from the Douglas-Rachford splitting algorithm. We show that, unlike

17: in the convex case, the algorithms in our nonconvex setup are not

18: equivalent to each other and have a rather involved relationship between

19: them. Finally, we comment on convergence of the ADMM heuristic and

20: compare it with the Douglas-Rachford splitting heuristic. 

21: \end{abstract}

22: