Distributed Sub-gradient Algorithms with Limited Communications

Abstract

We consider the distributed convex optimization scenario in which nodes in a network collectively find the minimum of a function utilizing only local communications and computations. Various sub-gradient algorithms have been developed for this optimization setting for the case in which the global function factorizes as the sum of local functions to be distributed to the nodes in network, including the distributed (i) online gradient descent, (ii) Nesterov gradient descent, and (iii) dual averaging algorithms. Generally speaking, these subgradient algorithms assume that, in each communication round, nodes can exchange messages of size comparable to that of the optimization variable. For many high-dimensional optimization problems, this communication requirement is beyond the capabilities of the communication network supporting the distributed optimization. For this reason, we propose a dimensionality reduction technique based on random projections to adapt these distributed optimization algorithms to networks with communication links of limited capacity. In particular, we analyze the performance of the proposed dimensionality reduction technique for the three algorithms above, i.e. (i)–(iii). Numerical simulations are presented that illustrate the performance of the proposed approach for these three algorithms.