Iterative regularization methods for nonlinear ill-posed problems (Q927436): Difference between revisions

The book is devoted to the problem of solving nonlinear ill-posed equations \(F(x)=y\) \((F: {\mathcal X} \to {\mathcal Y})\), where \(F\) is a Fréchet-differentiable operator, and \(\mathcal X\), \(\mathcal Y\) are Hilbert spaces. The ill-posedness means especially that solutions \(x^*=x^*(y)\) of the equations are unstable under perturbations of the element \(y\). Numerical methods that cope with this problem are regularization methods. In the last years, much attention was paid to the investigation of iterative regularization methods for ill-posed operator equations. These methods can be considered as an attractive alternative to the classical Tikhonov method. The book consists of eight chapters. In the main part (Chapters 1--4), recent convergence and rate of convergence results related to Landweber iteration and to Newton type methods are presented. It is supposed that instead of the true element \(y \in {\mathcal Y}\) an approximation \(y^{\delta}\) is available, where \(\| y^{\delta}-y\| \leq \delta\). Here, the main objects of investigation are the Landweber iteration \[ x_{k+1}=x_k+F^{\prime *}(x_k)(y^{\delta}-F(x_k)), \] the iteratively regularized Landweber iteration \[ x_{k+1}=x_k+F^{\prime *}(x_k)(y^{\delta}-F(x_k))+\beta_k(x_0-x_k) \] \((\beta_k>0, \lim_{k \to \infty} \beta_k=0)\), the Levenberg-Marquardt method \[ x_{k+1}=x_k+(F^{\prime *}(x_k) F^{\prime}(x_k)+\alpha_k I)^{-1} F^{\prime *}(x_k) (y^{\delta}-F(x_k)), \] and the iteratively regularized Gauss-Newton method \[ x_{k+1}=x_k+(F^{\prime *}(x_k) F^{\prime}(x_k)+\alpha_k I)^{-1} (F^{\prime *}(x_k) (y^{\delta}-F(x_k))+\alpha_k (x_0-x_k)) \] \((\alpha_k>0, \lim_{k \to \infty}\alpha_k=0)\). The convergence is analyzed in the noise free case where \(\delta=0\) and in the presence of errors in \(y\). In the last case, the iterations are equipped with appropriate a priori and a posteriori stopping rules \(k=k(y^{\delta},\delta)\). When studying the convergence of \(\{ x_k \}\) \((\delta=0)\) as \(k \to \infty\) and \(\{x_{k(y^{\delta},\delta)} \}\) \((\delta>0)\) as \(\delta \to 0\), the authors systematically use three types of assumptions on the original equation: the scaling conditions of type \(\| F^{\prime}(x)\| \leq 1\), structural conditions of the form \(\| F(x)-F(\widetilde x)-F^{\prime}(x)(x-{\widetilde x})\| \leq \eta \| F(x)-F(\widetilde x)\|\) \((\eta<1/2)\), both inequalities to be satisfied for all \(x\) and \({\widetilde x}\) from a neighborhood of a solution \(x^*\), and the source conditions \(x^*-x_0 \in R((F^{\prime *}(x^*) F^{\prime} (x^*))^{\mu})\) \((0<\mu \leq 1)\). An approach to numerical implementation of iterative methods on the basis of multilevel (multigrid) technique is discussed in Chapter 5. Chapter 6 is devoted to applications of the iterative regularization schemes to the level set methods. In Chapter 7, some numerical experiments with the iterative methods are presented. Finally, in the last chapter the authors give brief comments on alternative approaches to constructing iterative regularization methods for nonlinear equations.