Data assimilation and sampling in Banach spaces (Q1675437)

The paper studies the following propblem in a general Banach space \(X\): Let \(V\subset X\) be a finite-dimensional subspace and \(K=K(V,\varepsilon) = \{f\in X: \mathrm{dist}(f,V)\leq \varepsilon\}\). The problem is to reconstruct an element \(f\in K\) from \(m\) measurements, i.e., for a given linear operator \(M: X\to \mathbb{R}^m\), the vector \(w=M(f)\) is known. An algorithm for reconstructing \(f\) is any mapping \(A: \mathbb{R}^m \to X\), and it is admissible if \(M\circ A= \mathrm{Id}\). The problem is to find ``good'' algorithms. One way to measure the performance of \(A\) is to compare \(\sup_{f\in K_w} \|f-A(w)\|\) (where \(K_w= K\cap \{M=w\}\)) to the Chebyshev radius \(\mathrm{rad}(K_w)= \inf_{a\in X} \inf\{r: K_w\subset B(a,r)\}\). The algorithm \(A\) is near optimal if there is a constant \(C\) such that \[ \sup_{f\in K_w} \|f-A(w)\| \leq C \,\mathrm{rad}(K_w) \quad \text{for all }w\in \mathbb{R}^m. \] The authors make it clear that such problems can be solved using traditional tools of Banach space theory (e.g., duality of quotients and subspaces, the Bartle-Graves theorem, minimal projections, the Kadets-Snobar theorem); in fact, one might say they have been solved for decades, but it is only now that modern approximation theory asks for the solutions. In particular, the authors give an estimate of \(\mathrm{rad}(K_w)\) using the angle between \(\ker M\) and \(V\), and they improve this estimate for uniformly convex spaces. Further, they present near optimal algorithms based on liftings of \(M\), discuss several examples and deal with relations with sampling theory.