Robust Accelerated Gradient Method

We study the trade-off between rate of convergence and robustness to gradient errors in designing a first-order algorithm. In particular, we focus on gradient descent (GD) and Nesterov's accelerated gradient (AG) method for strongly convex quadratic objectives when the gradient has random errors in the form of additive white noise. To characterize robustness, we consider the asymptotic normalized variance of the centered iterate sequence which measures the asymptotic accuracy of the iterates. Using tools from robust control theory, we develop a tractable algorithm that allows us to set the parameters of each algorithm to achieve a particular trade-off between these two performance objectives. Our results show that there is a fundamental lower bound on the robustness level of an algorithm for any achievable rate. For the same achievable rate, we show that AG with tuned parameters is always more robust than GD to gradient errors. Similarly, for the same robustness level, we show that AG can be tuned to be always faster than GD. Our results show that AG can achieve acceleration while being more robust to random gradient errors. This behavior is quite different than previously reported in the deterministic gradient noise setting.