Perturbation Methods for Saddle Point Computation

A general class of iterative methods for saddle point seeking is developed. The directions used are subgradients evaluated at perturbed points. Convergence of the methods is proved and alternative strategies for implementation are discussed. The procedure suggests scalable algorithms for solving large-scale linear programs via saddle points. For illustration, some encouraging tests with the standard Lagrangian of linear programs from the Netlib library are reported.