Convergence Analysis of a Stochastic Projection-free Algorithm

This paper presents and analyzes a stochastic version of the Frank-Wolfe algorithm (a.k.a. conditional gradient method or projection-free algorithm) for constrained convex optimization. We first prove that when the quality of gradient estimate improves as ${\cal O}( \sqrt{ \eta_t^{\Delta} / t } )$, where $t$ is the iteration index and $\eta_t^{\Delta}$ is an increasing sequence, then the objective value of the stochastic Frank-Wolfe algorithm converges in at least the same order. When the optimal solution lies in the interior of the constraint set, the convergence rate is accelerated to ${\cal O}(\eta_t^{\Delta} /t)$. Secondly, we study how the stochastic Frank-Wolfe algorithm can be applied to a few practical machine learning problems. Tight bounds on the gradient estimate errors for these examples are established. Numerical simulations support our findings.