Detecting negative eigenvalues of exact and approximate Hessian matrices in optimization

Abstract

Nonconvex minimization algorithms often benefit from the use of second-order information as represented by the Hessian matrix. When the Hessian at a critical point possesses negative eigenvalues, the corresponding eigenvectors can be used to search for further improvement in the objective function value. Computing such eigenpairs can be computationally challenging, particularly if the Hessian matrix itself cannot be built directly but must rather be sampled or approximated. In blackbox optimization, such derivative approximations are built at a significant cost in terms of function values. In this paper, we investigate practical approaches to detect negative eigenvalues in Hessian matrices without accessing the full matrix. We propose a general framework that begins with the diagonal and gradually builds submatrices to detect negative curvature. Crucially, our approach works both when exact Hessian coordinate values are available and when Hessian coordinate values are approximated. We compare several instances of our framework on a test set of Hessian matrices from a popular optimization library, and finite-differences approximations thereof. Our experiments highlight the importance of the variable order in the problem description, and show that forming submatrices is often an efficient approach to detect negative curvature.