Nonmonotone backtracking inexact quasi-Newton algorithms for solving smooth nonlinear equations (Q1764740): Difference between revisions

For the nonlinear equation \(F(x)=0\) with a continuously differentiable mapping \(F: {\mathbb R}^n \to {\mathbb R}^n\) the author discusses a family of inexact quasi-Newton algorithms \(x_{k+1}=x_k+\lambda_k s_k\). Here \(s_k\in {\mathbb R}^n\) is determined in conjunction with some \(\eta_k\in(0,1)\) by the inexact quasi-Newton step \(\| F(x_k)+B_ks_k\| \leq\eta_k\| F(x_k)\| \) for \(B_k\) close to \(F'(x_k)\). The parameter \(\lambda_k=1, \omega, \omega^2, \ldots\) is determined for a given \(\omega \in(0,1)\) by the nonmonotone backtracking condition: \[ f(x_k+\lambda_k s_k)\leq f(x_{l(k)})+\lambda_k \beta \nabla f(x_k)^T s_k. \] Here \(f(x)={1\over 2}\| F(x)\| ^2,\) \(f(x_{l(k)})=\max_{0\leq j\leq m(k)}\{f(x_{k-j})\}\) and \(0\leq m(k)\leq\min\{m(k-1)+1, M\}\), \(M\) is a nonnegative integer. Both local and global convergence properties are established. The order of local convergence is discussed upon conditions on \(B_k\), \(F(x_k)\) and \(\eta_k\).