Generalizations of the limited-memory BFGS method based on the quasi-product form of update (Q1931787)

The paper deals with the limited-memory variable metric line search methods for unconstrained minimization that generalize the limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) method. These methods are based on updates of matrices \(H\) approximating the Hessian matrix among which the Broyden class updates belong to the most efficient ones. The authors propose five new variants of methods based on the quasi-product form of an update of the matrix \(H\). Four methods are obtained using transformations that convert the updates from the Broyden class to the formal BFGS update. The last method uses vectors from the previous iteration step and has some common features with the multi-step quasi-Newton methods. Two variants do not require additional matrix-vector multiplication and certain types of problems are more efficient than the L-BFGS method. Complete algorithms are given, global convergence is proved (except for the least efficient method), and numerical experiments are performed.