1: \begin{abstract}

2: We consider the problem of decentralized composite optimization

3: over a symmetric connected graph, in which each node holds its own

4: agent-specific private convex functions, and communications are

5: only allowed between nodes with direct links. A variety of

6: algorithms have been proposed to solve such a problem in an

7: alternating direction method of multiplier (ADMM) framework. Many

8: of these algorithms, however, need to include some proximal term

9: in the augmented Lagrangian function such that the resulting

10: algorithm can be implemented in a decentralized manner. The use of

11: the proximal term slows down the convergence speed because it

12: forces the current solution to stay close to the solution obtained

13: in the previous iteration. To address this issue, in this paper,

14: we first introduce the notion of \emph{simplest bipartite graph},

15: which is defined as a bipartite graph that has a minimum number of

16: edges to keep the graph connected. A simple two-step message

17: passing-based procedure is proposed to find a simplest bipartite

18: graph associated with the original graph. We show that the

19: simplest bipartite graph has some interesting properties. By

20: utilizing these properties, an ADMM without involving any proximal

21: terms can be developed to perform decentralized composite

22: optimization over the simplest bipartite graph. Simulation results

23: show that our proposed method achieves a much faster convergence

24: speed than existing state-of-the-art decentralized algorithms.

25: \end{abstract}

26: