Gradient Methods With Dynamic Inexact Oracles

We present a framework for generalizing the primal-dual gradient method, also known as descent ascent solving convex-concave minimax problems. The is based on observation that method can be viewed an inexact applied to primal problem. Unlike setting of traditional methods, computed by dynamic oracle, which discrete-time dynamical system whose output asymptotically approaches exact gradient. For problems, oracles are capable modeling range first-order methods computing objective, relies inner maximization provide unified convergence analysis with and demonstrate its use in creating new accelerated algorithms.