1: \begin{abstract}

2: Decentralized optimization algorithms have received much attention due to the

3: recent advances in

4: network information processing.

5: However, conventional decentralized algorithms based on projected gradient descent are

6: incapable of handling high dimensional constrained problems,

7: as the projection step becomes computationally prohibitive to compute.

8: To address this problem, this paper adopts a projection-free optimization approach,

9: a.k.a.~the Frank-Wolfe (FW) or conditional gradient algorithm.

10: We first develop a decentralized FW (DeFW) algorithm from the classical

11: FW algorithm. The convergence of the proposed algorithm is studied by viewing the

12: decentralized algorithm as an \emph{inexact} FW algorithm.

13: Using a diminishing step size rule and letting $t$ be the iteration number,

14: we show that the DeFW algorithm's convergence rate is ${\cal O}(1/t)$

15: for convex objectives; is ${\cal O}(1/t^2)$ for strongly convex objectives with the optimal solution

16: in the interior of the constraint set; and is ${\cal O}(1/\sqrt{t})$ towards

17: a stationary point for smooth but non-convex objectives.

18: We then show that a {  consensus}-based

19: DeFW algorithm meets the above guarantees

20: with two communication rounds per iteration.

21: Furthermore, we demonstrate the advantages of the proposed DeFW algorithm

22: on low-complexity robust matrix completion

23: and communication efficient sparse learning.

24: Numerical results on synthetic and real data

25: are presented to support our findings.\end{abstract}

26: