An iterative multi level algorithm for solving nonlinear ill-posed problems (Q1272821)

The paper deals with the solution of ill-posed problems \( F(x)=y \) where \( F: X \supset {\mathcal D} \to Y \) is a nonlinear Fréchet-differentiable operator in Hilbert spaces \( X \) and \( Y \), and the right-hand side can be perturbed, i.e., only \( y^\delta \in Y \) with \( \| y^\delta - y \| \leq \delta \) is available. For linear finite-dimensional subspaces \( X_1 \subset X_2 \subset \cdots \) of \( X \) with \(\bigcup_{N=1}^\infty X_N \) being dense in \( X \), the following multilevel scheme is considered, \[ x_{k+1,N}^\delta = x_{k,N}^\delta - P^N F'(x_{k,N}^\delta)^* (F(x_{k,N}^\delta) - y^\delta), \qquad k = 0, 1, \ldots, k_*(N,\delta), \quad N = 0, 1, \ldots, \] where the vector \( x_{0,0}^\delta = x_{0,0} \in X_0 \) is fixed, and the starting vector for the other inner iterations is \( x_{0,N}^\delta = x_{k_*(N-1,\delta),N-1}^\delta\in X_{N-1}\) \((N \geq 1) \). Moreover, the numbers \(k_*(N,\delta) \) correspond to a certain termination criterion for the inner iteration, \( P^N \) denotes the orthogonal projection onto \( X_N \), and \( F'(x_{k,N}^\delta)^* \) is the adjoint operator of \( F'(x_{k,N}^\delta) \). Under some additional conditions on \( F \), in the case of exact data (\( \delta = 0 \)) weak convergence of subsequences of \( x_{k_*(N,0),N}^0 \in X_N\), \(N = 0, 1, \ldots \), to solutions of the equation \( F(x) = y \) is established. Moreover, for problems which are mildly ill-posed in some sense, in the perturbed case (\( \delta > 0 \)) monotonicity results are provided for the error \( \| x_{k,N}^\delta - x_* \|\), \(k = 0, 1, \dots, \widehat{k}_*(N,\delta)\), \(N = 0, 1, \ldots, N_*(\delta) \), where \( x_* \in {\mathcal D} \) denotes a solution of the equation \( F(x) = y \), and \( \widehat{k}_*(N,\delta) \) and \( N_*(\delta) \) correspond to specific stopping rules for the inner and outer iteration, respectively.