Detecting negative eigenvalues of exact and approximate Hessian matrices in optimization

Warren Hare,Clément W. Royer
2023-06-21
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 access to 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.
Optimization and Control
What problem does this paper attempt to address?
The paper attempts to address the problem of how to effectively detect negative eigenvalues in the Hessian matrix during the optimization process. In non-convex minimization algorithms, utilizing second-order information (such as the Hessian matrix) can improve algorithm performance. When the Hessian matrix at a critical point has negative eigenvalues, the objective function value can be further optimized by moving in the direction of the eigenvectors corresponding to these negative eigenvalues. However, computing such eigenpairs (eigenvalues and their corresponding eigenvectors) is computationally challenging, especially when the Hessian matrix cannot be directly constructed and must be sampled or approximated. The paper proposes a practical method to detect negative eigenvalues without relying on the complete Hessian matrix. Specifically, the paper presents a general framework that starts from the diagonal of the Hessian matrix and incrementally constructs submatrices to detect negative curvature. This method is applicable not only in cases where the Hessian matrix coordinate values can be directly obtained but also in cases where only approximations of the Hessian matrix coordinate values are available. By comparing different instances of this framework on a set of Hessian matrices and their finite difference approximations from popular optimization libraries, the experimental results show the importance of variable ordering in problem description, and that constructing submatrices is generally an effective method for detecting negative curvature. In summary, the paper aims to develop an efficient and practical method for quickly detecting negative eigenvalues in the Hessian matrix in scenarios such as black-box optimization, thereby facilitating the solution of non-convex optimization problems.