The Projected Newton Method Has Order $1 + \sqrt 2 $ for the Symmetric Eigenvalue Problem
暂无分享,去创建一个
In their study of the classical inverse iteration algorithm, Peters and Wilkinson considered the closely related algorithm that consists of applying Newton’s method, followed by a 2-norm normalization, to the nonlinear system of equations consisting of the eigenvalue-eigenvector equation and an equation requiring the eigenvector to have the square of its 2-norm equal to one. They argue that in practice the $\infty $-norm is easier to work with, and they therefore replace the 2-norm normalization equation with a linear equation requiring that a particular component of the eigenvector be equal to one (effectively an $\infty $-norm normalization). Next, they observe that, because of the linearity of the normalization equation, the normalization step is no longer necessary; the algorithm thus reduces to Newton’s method and quadratic convergence follows from standard theory. Peters and Wilkinson choose to dismiss the 2-norm formulation in favor of the $\infty $-norm formulation; one factor in their choice seem...
[1] J. H. Wilkinson,et al. IMPROVING THE ACCURACY OF COMPUTED EIGENVALUES AND EIGENVECTORS , 1983 .
[2] J. H. Wilkinson,et al. Inverse Iteration, Ill-Conditioned Equations and Newton’s Method , 1979 .
[3] James M. Ortega,et al. Iterative solution of nonlinear equations in several variables , 2014, Computer science and applied mathematics.