Optimal numerical approximation of a linear operator (Q792068)

Let B and \(\Sigma\) be two Hilbert spaces and let \(S\in {\mathcal L}(B,\Sigma)\). The problem is to find an optimal approximation of Su by finitely many computations when an incomplete linear information of u is given in the form \(Nu=\nu\) and \(\| u\|_ B\leq 1\) where \(N:B\to E_ n\), \(n<\infty\). The author accepts the following definition of optimality depending on given linear interpolation M from \(E_ m\) into \(\Sigma\): an element \(M\hat a(\nu)\) is the optimal approximation of Su if \(\hat a\) minimizes the maximum error \(d(u,\nu)=\sup \{\| Su- Ma\|_{\Sigma}:u\in B\), \(Nu=\nu\), \(\| u\|_ B\leq 1\}\). A characterization of \(M\hat a(\nu)\) is presented. It is proved that \(M\hat a(\nu)\) is uniquely defined. In general the optimal approximation is not a linear transformation of \(\nu\). The author gives a necessary and sufficient condition for the optimal approximation to be linear. The relations with other concepts of optimality are investigated. The results are applied to a two-point boundary value problem.