Title: On Noisy Negative Curvature Descent: Competing with Gradient Descent for Faster Non-convex Optimization

Abstract: Several recent works focus on leveraging the Hessian-vector product to find a second-order stationary solution with strong complexity guarantee (e.g., almost linear time complexity in the problem's dimensionality), which avoids computing the Hessian matrix in conventional algorithms (e.g., cubic regularization). In this paper, we propose to further reduce the number of Hessian-vector products for faster non-convex optimization. In particular, we employ the Hessian-vector product to compute the negative curvature direction (i.e., the eigen-vector corresponding to the smallest eigen-value of the Hessian) for decreasing the objective value similar in spirit to many previous papers. However, previous works need to approximate the smallest eigen-value with a sufficient precision (e.g., $\epsilon_2\ll 1$) in order to achieve a sufficiently accurate second-order stationary solution (i.e., $\lambda_{\min}(\nabla^2 f(\x))\geq -\epsilon_2)$. In contrast, the proposed algorithms only need to compute the smallest eigen-vector approximating the corresponding eigen-value up to a small power of current gradient's norm. As a result, it can dramatically reduce the number of Hessian-vector products during the course of optimization before reaching first-order stationary points (e.g., saddle points). We show that in the worst-case the proposed algorithms with their favorable prescribed accuracy requirements can match the best time complexity in literature for achieving a second-order stationary point but with an arguably smaller per-iteration cost. We also show that the proposed algorithms can benefit from inexact Hessian (e.g., the sub-sampled Hessian for a finite-sum problem) by developing their variants accepting inexact Hessian under a mild condition for achieving the same goal.