Revisiting Frank-Wolfe for Structured Nonconvex Optimization

We introduce a new projection-free (Frank-Wolfe) method for optimizing structured nonconvex functions that are expressed as a difference of two convex functions. This problem class subsumes smooth nonconvex minimization, positioning our method as a promising alternative to the classical Frank-Wolfe algorithm. DC decompositions are not unique; by carefully selecting a decomposition, we can better exploit the problem structure, improve computational efficiency, and adapt to the underlying problem geometry to find better local solutions. We prove that the proposed method achieves a first-order stationary point in $\mathcal{O}(1/\epsilon^2)$ iterations, matching the complexity of the standard Frank-Wolfe algorithm for smooth nonconvex minimization in general. Specific decompositions can, for instance, yield a gradient-efficient variant that requires only $\mathcal{O}(1/\epsilon)$ calls to the gradient oracle by reusing computed gradients over multiple iterations. Finally, we present numerical experiments demonstrating the effectiveness of the proposed method compared to other projection-free algorithms.