Fast multilevel augmentation methods with compression technique for solving ill-posed integral equations (Q548792)

The subject of this paper is the regularization of linear ill-posed problems \( \mathcal{A} u = f \) by multilevel augmentation methods combined with compression techniques. Here, \( \mathcal{A}: \mathbf{X} \to \mathbf{X} \) is a linear, positive semidefinite integral operator in the Hilbert space \( \mathbf{X} = L^2(E) \), where \( E \subset \mathbf{R}^d \) is a bounded domain, and \( f \in \mathbf{X} \). The basic regularization scheme is Lavrentiev's method \( (\alpha \mathcal{I} + A)u_\alpha^\delta = f^\delta \), where \( f^\delta \in \mathbf{X} \) is an approximation of the right-hand side \( f \) with \( \| f^\delta - f \| \leq \delta \). In addition, \( \alpha > 0 \) is a regularization parameter, and \( \mathcal{I} \) denotes the identity operator. For the multiscale decomposition, piecewise polynomial spaces \( \mathbf{X}_n \) with \( \mathbf{X}_{n} \subset \mathbf{X}_{n+1} \) (\( n \in \mathbf{N}_0\)) are considered, and \( \bigcup_{n \in \mathbf{N}_0} \mathbf{X}_{n} \) is supposed to be dense in \( \mathbf{X} \). The basis for the multilevel augmentation algorithm is the approximative equation \( (\alpha \mathcal{I} + \widetilde{\mathcal{A}}_n)\widetilde{u}_{\alpha,n}^\delta = f_n^\delta \), where \( \widetilde{\mathcal{A}}_n \) is some compression of \( \mathcal{A}_n = \mathcal{P}_n \mathcal{A} \mathcal{P}_n \), with \( \mathcal{P}_n : \mathbf{X} \to \mathbf{X}_n \) denoting the orthogonal projection onto \( \mathbf{X}_n \), and \( f_n^\delta = \mathcal{P}_n f^\delta \). The authors present an estimate of the error \( \| \widehat{u} - \widetilde{u}_{\alpha,n}^\delta \| \), where \( \widehat{u} \in \mathbf{X} \) denotes the minimum solution of \( \mathcal{A} u = f \), and \( \alpha > 0 \) is arbitrary here. For this estimate, smoothness assumptions on the kernel of the integral operator \( \mathcal{A} \) are made, and a source condition for the minimum solution \( \widehat{u} \in \mathbf{X} \) of \( \mathcal{A} u = f \) is needed. In addition, approximation properties of the spaces \( \mathbf{X}_n \) are required. The sparsity of \( \widetilde{\mathcal{A}}_n \) is also considered in this paper. Then the truncated multilevel augmentation algorithm is considered. It is based on a decomposition of the form \( [\widetilde{\mathcal{B}}_{\ell,m}(\alpha) + \widetilde{\mathcal{C}}_{\ell,m}(\alpha)] \widetilde{u}_{\alpha,\ell+m}^\delta = \alpha^{-1} f_{\ell+m}^\delta \). Here, \( n = \ell+m \), \( \widetilde{\mathcal{B}}_{\ell,m}(\alpha) = \mathcal{I} + \alpha^{-1} \widetilde{\mathcal{A}}_{\ell,m}^L \) and \( \widetilde{\mathcal{C}}_{\ell,m}(\alpha) = \alpha^{-1} \widetilde{\mathcal{A}}_{\ell,m}^H \), where \( \widetilde{\mathcal{A}}_{\ell,m}^L = \mathcal{P}_\ell \widetilde{\mathcal{A}}_{\ell+m} \mathcal{P}_{\ell+m} \) and \( \widetilde{\mathcal{A}}_{\ell,m}^H = (\mathcal{P}_{\ell+m}- \mathcal{P}_{\ell}) \widetilde{\mathcal{A}}_{\ell+m} \mathcal{P}_{\ell+m} \). An a posteriori parameter choice of the parameter \( \alpha \) is proposed, and order optimal error estimates are presented under the assumptions considered above, and for \( \ell \) sufficiently large. Finally some numerical results are presented.