Descent Directions of Quasi-Newton Methods for Symmetric Nonlinear Equations

Abstract

Abstract.In general, when a quasi-Newton method is applied to solve a system of nonlinear equations, the quasi-Newton direction is not necessarily a descent direction for the norm function. In this paper, we show that when applied to solve symmetric nonlinear equations, a quasi-Newton method with positive definite iterative matrices may generate descent directions for the norm function. On the basis of a Gauss-Newton based BFGS method [D. H. Li and M. Fukushima, SIAM J. Numer. Anal., 37 (1999), pp. 152-172], we develop a norm descent BFGS method for solving symmetric nonlinear equations. Under mild conditions, we establish the global and superlinear convergence of the method. The proposed method shares some favorable properties of the BFGS method for solving unconstrained optimization problems: (a) the generated sequence of the quasi-Newton matrices is positive definite; (b) the generated sequence of iterates is norm descent; (c) a global convergence theorem is established without nonsingularity assumption on the Jacobian. Preliminary numerical results are reported, which positively support the method.

Limited memory technique using trust regions for nonlinear equations

“…This inexact backtracking technique was reported by Yuan and Lu [20]. Many TR methods (see [2,18,26]), as well as other methods (see [4,5,9,16,19,24,25,27]), have been developed to solve nonlinear equations. Compared with line search methods, trust region methods possess desirable theoretical features.…”

A globally convergent BFGS method for nonlinear monotone equations without any merit functions

“…By using a nonmonotone line search process, Li and Fukushima [15,16] proposed a Broyden's method for solving nonlinear equations and a Gauss-Newton-based BFGS method for solving symmetric nonlinear equations and prove that these methods converge globally. Quite recently, Gu et al [12] introduced a norm descent line search technique and proposed a norm descent Gauss-Newton-based BFGS method for solving symmetric equations with global convergence.…”

A New Quasi-Newton Method Based on Adjoint Broyden Updates for Symmetric Nonlinear Equations

“…The symmetric nonlinear equations have many practical backgrounds such as in the computation of the stationary points of unconstrained optimization problems, saddle points, large-scale scientific and engineering computing. For symmetric nonlinear equations, there have been many methods [3,6,8,16] proposed for solving them, where BFGS method performs much better. The BFGS update formula takes the following form…”

“…Li and Fukushima [10] proposed a Gauss-Newton-based BFGS method for solving symmetric nonlinear equations and established the global and superlinear convergence. Based on the Gauss-Newton-based BFGS method, Gu et al [8] proposed a norm descent BFGS method for solving symmetric nonlinear equations. The authors in [16] also studied quasi-Newton methods for solving symmetric nonlinear equations.…”