Title: The Speed-Robustness Trade-Off for First-Order Methods with Additive Gradient Noise

Abstract: We study the trade-off between convergence rate and sensitivity to stochastic additive gradient noise for first-order optimization methods. Ordinary Gradient Descent (GD) can be made fast-and-sensitive or slow-and-robust by increasing or decreasing the stepsize, respectively. However, it is not clear how such a trade-off can be navigated when working with accelerated methods such as Polyak's Heavy Ball (HB) or Nesterov's Fast Gradient (FG) methods, or whether any of these methods can achieve an optimal trade-off. We consider three classes of functions: (1) strongly convex quadratics, (2) smooth strongly convex functions, and (3) nonconvex functions that satisfy a weak notion of strong convexity. For each function class, we present a tractable way to compute convergence rate and sensitivity to additive gradient noise for a broad family of first-order methods, and we present near-Pareto-optimal algorithm designs. Each design consists of a simple analytic update rule with two states of memory, similar to HB and FG. Moreover, each design has a scalar tuning parameter that explicitly trades off convergence rate and sensitivity to additive gradient noise. When tuned as aggressively as possible, our proposed algorithms recover the algorithms with fastest-known convergence rates for each function class. When tuned to be more robust, our algorithms are novel and provide a practical way to control noise sensitivity while maintaining the fastest possible convergence rate. We validate the performance and near-optimality of our designs through numerous numerical simulations.