On mesh independence and Newton-type methods (Q1316213)

The author considers a two-level solution procedure for nonlinear strongly monotone operator equations of the form: seek \(u\in V\) satisfying \(\langle F(u),v\rangle= \langle g,v\rangle\) for all \(v\in V\) where \hbox{\(F: V\to V'\)} and \(g\in V'\) is given, \(V\) is a Hilbert space with dual \(V'\). \(F\) is assumed to be strongly monotone, locally Lipschitz and its Fréchet derivative satisfies some additional smoothness condition. Typical examples are quasilinear equations occurring in the mathematical modelling of torsion of a bar, electromagnetic field equations etc. It is shown by properly relating the mesh parameters \(H\) and \(h\) for a coarse and a fine discretization mesh, that it suffices to compute the solution of the nonlinear equation on the coarse mesh and subsequently correct it once using a linearized (Newton) equation on the fine mesh. This assumes that a certain relation, \(H= h^{1/\alpha}\), holds between the coarse and fine mesh parameters for some positive \(\alpha\). In the ideal case, for piecewise linear finite element approximations the relation is \(H= h^{1/4}\) which means, that in practice the coarse mesh contains very few node-points. The iteration error of the so computed solution has the same order as the discretization error. For the solution on the coarse mesh a global convergence result is shown to hold if the nonlinear function \(F\) is Hölder continuous. To achieve the global convergence, a truncated generalized conjugate gradient type method is used to compute the successive approximations and the search directions which are then used are updated from a vector which is an approximate solution of the Newton (linearized) equation. Finally it is shown that the number of steps before the superlinear rate of convergence sets in is bounded independent of the mesh parameter.