No label defined (Q4492855)

The problem of solving a system of nonlinear equations \(F(x) = 0\), where \(F:\mathbb{R}^n \rightarrow \mathbb{R}^n\) is a nonlinear mapping of a special form \(F(x) = Ax + G(x)\) with linear part \(A \in \mathbb{R}^{n,n}\) and nonlinear part \(G:\mathbb{R}^n \rightarrow \mathbb{R}^n\), is considered. Newton's method solves this system by generating an iterative sequence \(\{x^k\}\) defined by \[ J(x^k)s^k = - F(x^k), \quad x^{k+1} = x^k+s^k \quad (k=0,1,\dots), \] where \(J(x)\) is the Jacobian matrix of \(F\). This procedure requires the calculation of \(J(x^k)\) and solving the linear system in each iteration, which is rather expensive from the practical point of view. Restricting to the cases when the matrix \(A\) has some special properties like symmetry, diagonal dominance, good sparsity, the authors propose a new quasi-Newton method with local \(q\)-linear convergence and a high computational efficiency. Although the new method converges slower than the classical Newton method, its computational cost is decreased since the same matrix of the Newtonian system is used in every iterative step, which is a significant advantage in solving systems with large dimension. Practical implementation of the presented quasi-Newton method is considered in detail. Special attention is devoted to the choice of the relaxation parameter involved in the new algorithm.