New Inexact Parallel Variable Distribution Algorithms

We consider the recently proposed parallel variable distribution(PVD) algorithm of Ferris and Mangasarian [4] for solvingoptimization problems in which the variables are distributed amongp processors. Each processor has the primary responsibility forupdating its block of variables while allowing the remaining“secondary” variables tochange in a restricted fashion along some easily computable directions.We propose useful generalizationsthat consist, for the general unconstrained case, of replacing exact global solution ofthe subproblems by a certain natural sufficient descent condition, and,for the convex case, of inexact subproblem solution in thePVD algorithm. These modifications are the key features ofthe algorithm that has not been analyzed before.The proposed modified algorithms are more practical andmake it easier to achieve good load balancing among the parallelprocessors.We present a general framework for the analysis of thisclass of algorithms and derive some new and improved linear convergence resultsfor problems with weak sharp minima of order 2 and strongly convex problems.We also show that nonmonotone synchronization schemesare admissible, which further improves flexibility of PVD approach.