Convergence analysis of modified Newton-HSS method for solving systems of nonlinear equations (Q2636951)

Inexact Newton methods are often employed to solve nonlinear systems of the form \(F(x) = 0\), such that \(F: \mathbf{D} \subset \mathbb C^n \rightarrow \mathbb C^n\) is continuously differentiable, and the generated Jacobian matrices are large, sparse, non-Hermitian and positive definite. Such methods can be characterized as inner-outer iterative methods, as linear iterative methods are applied to solve approximately the Newton equation \(F'(x_k)s_k = -F(x_k)\) for \(k \geq 0\) (this is the inner iteration), generated by Newton's method (the outer iteration). \textit{Z. Bai} and \textit{X. Guo} [J. Comput. Math. 28, No. 2, 235--260 (2010; Zbl 1224.65133)] introduced a Newton-HSS method to solve these systems, employing the Hermitian and skew-Hermitian splitting method for the inner iteration. \textit{M. T. Darvishi} and \textit{A. Barati} [Appl. Math. Comput. 187, No. 2, 630--635 (2007; Zbl 1116.65060)] introduced the modified Newton method to solve the simple roots of nonlinear equations, the main advantage of their method over Newton's method being that it has \(R\)-order of convergence at least three, while Newton's method converges quadratically. In the present paper, the authors combine these two methods to obtain a modified Newton-HSS (MN-HSS) method. Specifically, each inner iteration step of the modified Newton-HSS method requires that two linear systems are solved: \(F'(x_k)d_k = -F(x_k)\), \(y_k = x_k+d_k\), and \(F'(x_k)h_k = -F(y_k)\), \(x_{k+1} = y_k+h_k\), and these are solved approximately by the HSS-method. The authors prove local and semilocal convergence. The necessary hypotheses for and properties of both types of convergence are similar to those for the Newton HSS-method. The authors also introduce a global modified HSS-Newton method which has global convergence properties. The authors compare their method to the Newton HSS-method on the two-dimensional nonlinear convection-diffusion equation given in Bai and Guo's paper [loc. cit.], as well as a nonlinear problem they refer to as BROYDN3D, which contains \(n\) coupled quadratic equations. The modified Newton HSS-method takes less time and fewer iterations than the Newton HSS-method, but the error estimate is generally larger.