zbMATH — the first resource for mathematics

Preconditioning of truncated-Newton methods. (English) Zbl 0592.65038
The author discusses the problem of minimizing a real-valued function f(x) of n variables for large n. Truncated Newton methods are used to solve this problem. At each major iteration the Newton equation \(G^{(k)}p=g^{(k)}\) is approximately solved for the direction p using a modified Lanczos method. Here \(g^{(k)}\) and \(G^{(k)}\) are, respectively, the gradient vector and the Hessian matrix of the second derivatives of f evaluated at the point \(x^{(k)}\). A line search procedure is used to find the updated point \(x^{(k+1)}\). Some preconditioning and scaling strategies are investigated. Numerical examples show that a carefully chosen truncated Newton method can perform well in comparison with conjugate gradient algorithms.
Reviewer: I.H.Mufti

65K05 Numerical mathematical programming methods
90C30 Nonlinear programming
65H10 Numerical computation of solutions to systems of equations
Full Text: DOI