1: \begin{abstract}

2: 	Recently, the decentralized optimization problem is attracting growing attention.

3: 	Most existing methods are deterministic with high per-iteration cost and have a convergence rate quadratically depending on the problem condition number.

4: 	Besides, the dense communication is necessary to ensure the convergence even if the dataset is sparse.

5: 	In this paper, we generalize the decentralized optimization problem to a monotone operator root finding problem, and propose a stochastic algorithm named DSBA  that

6: 	(i) converges geometrically with a rate linearly depending on the problem condition number, and (ii) can be implemented using sparse communication only.

7: 	Additionally, DSBA handles learning problems like AUC-maximization which cannot be tackled efficiently in the decentralized setting.

8: 	Experiments on convex minimization and AUC-maximization validate the efficiency of our method.

9: \end{abstract}

10: