Unconstrained Optimization Techniques for the Acceleration of Alternating Projection Methods
Luis M. Hernández-Ramos, René Escalante, Marcos Raydan · Numerical Functional Analysis and Optimization · 2011
Alternating projection methods have been extensively used to find the closest point, to a given point, in the intersection of several given sets that belong to a Hilbert space. One of the characteristics of these schemes is the slow convergence that can be observed in practical applications. To overcome this difficulty, several techniques, based on different ideas, have been developed to accelerate their convergence. Recently, a successful acceleration scheme was developed specially for Cimmino's method when applied to the solution of large-scale saddle point problems. This specialized acceleration scheme is based on the use of the well-known conjugate gradient method for minimizing a related convex quadratic map. In this work, we extend and further analyze this optimization approach for several alternating projection methods on different scenarios. In particular, we include a specialized analysis and treatment for the acceleration of von Neumann-Halperin's method and Cimmino's method on subspaces, and Kaczmarz method on linear varieties. For some specific applications we illustrate the advantages of our acceleration schemes with encouraging numerical experiments.