1: \begin{abstract}

2:   We study \emph{distributed big-data nonconvex} optimization in multi-agent

3:   networks.  We consider the (constrained) minimization of the sum of a smooth

4:   (possibly) nonconvex function, i.e., the agents' sum-utility, plus a convex

5:   (possibly) nonsmooth regularizer. Our interest is on big-data problems in

6:   which there is a large number of variables to optimize. If treated by means of

7:   standard distributed optimization algorithms, these large-scale problems may

8:   be intractable due to the prohibitive local computation and communication

9:   burden at each node.  We propose a novel distributed solution method where, at

10:   each iteration, agents update in an uncoordinated fashion only one block of the

11:   entire decision vector. To deal with the nonconvexity of the cost function,

12:   the novel scheme hinges on Successive Convex Approximation (SCA) techniques

13:   combined with a novel \emph{block-wise} perturbed push-sum consensus

14:   protocol, which is instrumental to perform local block-averaging operations

15:   and tracking of gradient averages. Asymptotic convergence to stationary

16:   solutions of the nonconvex problem is established. Finally, numerical results

17:   show the effectiveness of the proposed algorithm and highlight how the block

18:   dimension impacts on the communication overhead and practical convergence

19:   speed.

20: \end{abstract}

21: