Recovering linear operators from inaccurate data (Q1893074)

Let \((F, |\cdot |_F)\) and \((G, |\cdot |_G)\) be a seminormed respectively a normed space (over \(\mathbb{R}\) or \(\mathbb{C}\)) and \(S : F \to G\) a given linear operator. One supposes that for \(f \in B := \{x \in F : |x|_F \leq 1\}\) only an inaccurate information is available, given by a vector \(y \in \mathbb{C}^n\), \(y = [y_1 \dots y_n]^T\) and a linear operator \(N : F \to \mathbb{C}^n\) satisfying \(|||y - N (f) |||\leq \delta\), for a real number \(\delta > 0\) and a norm \(|||\cdot |||\) on \(\mathbb{C}^n\). Admitting as an approximation algorithm any mapping \(\Phi: \mathbb{C}^n \to G\), the error is defined by \(e(\Phi, S, N, \delta) := \sup \{|S(f) - \Phi(y)|_G : f \in B, |||y - N(f) |||\leq \delta\}\). The aim of the present paper is to compare the quantity \(d(S, N, \delta) := 2 \cdot \sup \{|S(f) |_G : f\in B, |||N(f)|||\leq \delta\}\) to \(d(S,N,0)\), i.e. to study the influence of data perturbations on the best approximation error. It is well known that \(2^{-1} \cdot d(S,N,\delta) \leq \inf \{e (\Phi, S, N, \delta) : \mathbb{C}^n @>\Phi >> G \} \leq d(S, N, \delta)\). The general results obtained by the authors are applied to an inner product space \(F\) and a linear functional \(S(f) = <f, u_0>\) on \(F\). The case of the Sobolev space \(F = W^\infty_n\) and a function \(f\) known only from inaccurate samples at distinct points \(t_1, \dots, t_n\) in \([a,b]\) is also considered.