Optimal residual algorithms for linear operator equations (Q1179027): Difference between revisions

The paper is concerned with an optimal algorithm to solve a linear problem \(Lu=f\) where \(L\) maps a Banach space \(U\) into another one \(V\). The main concern is an ill-posed problem of that type. As an example, the Laplace transform is considered. A linear algorithm to solve the problem is a solution in the form \[ \varphi(Nf)=\sum^ n_{j=1}\alpha_ j\lambda_ j(f)\;\forall f\in F\cap D.\tag{1} \] Here, \(\lambda_ j\) is a linear continuous real functional and \(\alpha_ j\) are appropriate elements in the Banach space \(U\). \(D\) is the range of \(L\), which is supposed to be dense in \(V\). Moreover, \(F\) is the set of right-hand sides in \(V\) which are considered. Finally, the abbreviation \(Nf=[\lambda_ 1(f),\dots,\lambda_ n(f)]\) \(\forall f\in V\) is used. This mapping contains the information of the right-hand side \(f\) which is used to produce the approximation (1). The error of the approximation is measured in terms of the defect \(Lu-f\). More precisely, an imbedding \(E\) of \(V\) into a Banach space \(W\) is used to minimize the expression \(E(Lu-f)\) in the norm of the Banach space \(W\). For this purpose, the expression \[ r(\varphi,N)=\sup_{f\in F\cap D}\| E(L\varphi(Nf)-f\| \] is introduced and the author is interested in the number \(r(N)=\inf_ \varphi(\varphi N)\), where \(\varphi\) ranges over all algorithms using the information \(N\). To formulate the results of the paper, the representation \(F=\{f\in V:\| Tf\|\leq 1\}\) of the set \(F\) of problem elements is assumed where \(T\) is a continuous linear restriction operator of \(V\) into the Banach space \(X\). There are two main results of the paper: One states that the kernel of the operator \(T\) is in \(D\) if and only if there exists a linear algorithm (1) such that the corresponding residual \(r(\varphi,N)\) is finite. The other result states that there is an almost optimal algorithm in the sense that the number \(r(N)\) defined above is almost assumed. This algorithm can be chosen in the form (1) if \(\text{ker}T\subseteq D\) holds. In general, an algorithm is constructed which is almost but not quite of the form (1). In a closing section, applications to the Laplace transform are given.