Indefinite knowledge about an object and accuracy of its recovery methods (Q2487081)

A very general scheme of optimal recovery is introduced and illustrated by examples. It is assumed that the unknown object \(x\) belongs to a given set (``class'') \(C\). The information about \(x\) is given by an element \(y\) of a certain set \(Y\) so that \(y=F(x)\), where \(F: C \to Y\) is some given information operator. The information may be inaccurate, in which case \(F\) may be multivalued. Given \(F(x)\), \(x \in C\), one has to recover \(x\) or, more generally, some feature \(z=f(x)\), in the best possible way. It is assumed that \(z\) is an element of a given metric space \((Z,d)\) with \(f: C \to Z\) being some known map. Any map \(m: F(C) \to Z\) is called a recovery method. The error of such a method is the quantity \[ e(f,C,F,m)=\sup d(f(x), m(y)), \] where the supremum is taken over all \(x \in C\), \(y \in F(x)\). The error of optimal recovery, denoted by \(E(f,C,F)\), is a solution of the extremal problem \[ \inf_m e(f,C,F,m). \] The authors provide several examples in which the latter problem is solved. As the first example, they consider recovery of a function \(x(t)\) at a point \(\tau \in [-1, 1]\) from its values at other points \((t_1, \ldots , t_n)\), assuming that \(x\) belongs to the Sobolev class \(W_\infty^1 [-1, 1]\). Here \(C=W_\infty^1\), \(Z={\mathbb R}\), \(f(x(\cdot))=x(\tau)\), \(Y={\mathbb R}^n\), and \(F(x(\cdot))=(x(t_1), \ldots , x(t_n))\). The other examples treated in the paper are: recovery of an integral of a function from its values at several points; recovery of the derivative of a function at a point from inaccurate values at other points; recovery of a function from its inaccurate Fourier coefficients; recovery of a function at a point from the function itself given with an error in the \(L_2\) norm. The paper is concluded with a general result concerning optimal recovery of linear functionals.