An extension of the frank and Wolfe method of feasible directions

The Frank and Wolfe method of feasible directions is shown to be a case of the more general computational approach of inner linearization followed by restriction. An extension is proposed based on this observation. The extended procedure converges, and under certain conditions the asymptotic convergence rate is geometric. Limited computational experience comparing the two procedures is reported.The author is grateful to Mr. Peter Friesen for help with the computational work and to the Graduate School of Business, Stanford University for their support of this research.