Title: An accelerated algorithm for minimizing convex compositions

Abstract: We describe a new proximal algorithm for minimizing compositions of finite-valued convex functions with smooth mappings. When applied to convex optimization problems having an additive composite form, the algorithm reduces to FISTA. The method both realizes the best known complexity bound of $\mathcal{O}(1/\epsilon)$ in optimality conditions whenever the objective function has bounded domain, and achieves the accelerated rate $\mathcal{O}(1/\sqrt{\epsilon})$ in function values under standard convexity assumptions. A natural convexity parameter of the composition quantifies the transition between the two modes of convergence.