On the order optimality of the regularization via inexact Newton iterations (Q431306): Difference between revisions

The article deals with iterative methods for approximately solving the ill-posed operator equation \(F(x) = 0\) with a nonlinear Fréchet differentiable operator \(F\) between Banach spaces \({\mathfrak X}\) and \({\mathfrak Y}\). The general scheme of these methods is the following: the method starts with an initial guess \(x_0 \in {\mathcal D}(F)\); if \(x_n\) is a current iterate, one may apply any regularization scheme to the equation \[ F'(x_n)u = y^\delta - F(x_n) \] to produce a family of regularized approximations \(\{u_n(t)\}\). The regularization parameter may be chosen to be the smallest number \(t_n > 0\) such that \[ \|y^\delta - F(x_n) - F'(x_n)u_n(t_n)\| \leq \eta \|y^\delta - F(x_n)\| \] for some preassigned value \(0 < \eta < 1\). The iterate is then defined by \(x_{n+1} = x_n + u_n(t_n)\). In the simplest case it is assumed that the derivative \(F'(x)\) satisfies the Mysovskikh condition \[ \|F'(x)\| \leq \theta < 1, \quad x \in B_r(x_*), \] (\(x_*\) is an exact solution, \(r > 0\)) and a ``smoothness'' condition of type \[ \|[F'(x) - F'(z)]h\| \leq K_0 \|x - z\| \, \|F'(z)h\|, \quad h \in {\mathfrak X}; \] under these conditions the method is well-defined with the regularization schemes by Landweber, Tikhonov and some others (implicit and asymptotic). More complicated and delicate statements are obtained in the case when \({\mathfrak X}\) and \({\mathfrak Y}\) are Hilbert spaces and the operator \(F\) acts in the Hilbert scale generated by a self-adjoint and strictly positive linear operator \(L\). In this case it is assumed that \[ m\|h\|_{-a} \leq \|F'(x)h\| \leq M\|h\|_{-a}, \quad h \in B_\rho(x_*), \] \[ \|F'(x)L^{-s}\|_{{\mathfrak X} \to {\mathfrak Y}} \leq \theta, \quad \in B_\rho(x_*), \] and \[ \|F'(x) - F'(z)\|_{{\mathfrak X}_{-b} \to {\mathfrak Y}} \leq K_0\|x - z\|^\beta, \quad x, z \in B_\rho(x_*) \] where \(\rho > 0\), \(0 < \beta \leq 1\), \(0 \leq b \leq a\), \(K_0 \geq 0\). The author states that his result about the rate of convergence is order optimal.