Efficient Quadratic Corrections for Frank-Wolfe Algorithms

Abstract: We develop a Frank-Wolfe algorithm with corrective steps, generalizing previous algorithms including blended pairwise conditional gradients and fully-corrective Frank-Wolfe, and propose a highly efficient corrective algorithm in the case of convex quadratic objectives based on linear optimization or linear system solving, akin to Wolfe's minimum-norm point. Beyond optimization problems that are directly quadratic, we revisit two algorithms, split conditional gradient for the intersection of two sets accessible through linear oracles, and second-order conditional gradient sliding, which approximately solves Variable-Metric projection subproblems, proposing improvement of the algorithms and their guarantees that may be of interest beyond our work, and we leverage quadratic corrections to accelerate the quadratic subproblems they solve. We show significant computational acceleration of Frank-Wolfe-based algorithms equipped with the quadratic correction on the considered problem classes.