An adaptive Newton algorithm based on numerical inversion: Regularization as postconditioner (Q1067360): Difference between revisions

The paper concerns a class of adaptive, approximate Newton methods based on Nash-Moser iteration. For operator equations \(F(u)=0\) the superlinear (or quadratic) convergence of the approximate Newton iterates \(\{u_ k\}\), computed by means of approximate right inverses \(T_ h(u_ k)\) of \(F'(u_ k)\), requires the estimation of \(F'(u_ k)T_ h(u_ k)-I\) in terms of the residual. In the case of nonlinear differential operators and for \(T_ h\) defined by numerical procedures - such as the finite element method - these estimates fail in the sense that a ''loss of derivative'' phenomenon arises, as in the Nash-Moser results. Unlike of \textit{J. Moser}'s basic paper [Ann. Sc. Norm., Sup. Pisa, Sci. Fis., Mat., III. Ser. 20, 265-315 (1966; Zbl 0144.182)] the author separates the required smoothing operators \(S_{t_ k}\) from the inverse approximations \(T_{h_ k}\). In other words, the regularization is applied as postconditioner to the numerical inversion. A superlinearly convergent, approximate Newton process is obtained by adapting the numerical inversion to the magnitude of the current residual in an intermediate norm defined by the defect. In the analysis all norms are of Hölder space type. The design of the algorithm involves multi-parameter selection and hence opens up questions about the complexity of the schemes. For this, results of \textit{A. G. Vitushkin} [Theory of the transmission and processing of information (1961; Zbl 0122.120)] are used.