Cascadic multilevel methods for ill-posed problems (Q1044860)

This paper is concerned with a CGNR-based multilevel method for solving linear ill-posed problems in Hilbert spaces. Regularisation is achieved by restricting the number of iterations on each level using the discrepancy principle. In order to give some details, consider an ill-posed linear equation \(Ax=b\), where \( A: L_2(\Omega) \to Y \) is a compact linear operator into a Hilbert space \( Y\), and \( b \in Y \) is an approximately given right- hand side, i.e., \( b^\delta \in Y \) with \( \| b^\delta - b \| \leq \delta \) is available only. Moreover let \( S_1 \subset S_2 \subset \cdots \subset S_\ell \) be given nested linear subspaces of \( L_2(\Omega) \) with \( S_\ell = L_2(\Omega) \), and consider restriction operators \(R_i: L_2(\Omega) \to S_i \) and prolongation operators \(Q_i: S_i \to L_2(\Omega) \) (\( 1 \leq i \leq \ell \)), with \( R_\ell \) and \( Q_\ell \) identity operators. The CGNR-based multilevel method considered here determines on each level \(S_i\) an approximate solution of \( R_iAQ_i x=R_ib^{\delta} \) by CGNR (the conjugate gradient method applied to the normal equations), and the iterations on this level are terminated by a stopping rule based on the discrepancy principle. The resulting iterate is mapped from \( S_i \) into \( S_{i+1} \) by a prolongation operator and is used then as initial vector for CNGR on the next level \( S_{i+1} \). The authors discuss the regularizing properties for each level \( S_i \) and thus in particular for the final level \( S_\ell = L_2(\Omega) \). Some numerical results are presented, and a similar scheme is considered for self-adjoint problems.