The alternating algorithm in a uniformly convex and uniformly smooth Banach space (Q401401)

Let \(X\) be a uniformly convex Banach space and \(C\) a closed convex subset of \(X\). It is very well known that there exists a unique best approximation from \(C\), denoted by \(P_{C}(x),\) of every vector \(x\in X\). If you assume that \(M_{k}\), \(k=1, 2, \dots, r,\) are closed linear subspaces of a uniformly convex and uniformly smooth Banach space \(X,\) then a natural question is how and when you can find a best approximation to \(x\in X\) from \(M= M_{1} + M_{2} + \dots + M_{r}\) in terms of \(P_{M_{k}}\), \(k=1, 2, \dots, r\). In the paper under review the author proves that the so called alternating algorithm gives the best approximation to \(x\) from \(M\) assuming \(M\) is a closed linear subspace. Indeed, if \(E= (I-P_{M_{r}})(I-P_{M_{r-1}}) \dots (I-P_{M_{1}})\), then \(\lim_{n\rightarrow\infty} E^{n} x = x- P_{M}(x).\) In particular, if the linear subspaces \(M_{k}\) are finite-dimensional, then \(M\) is closed and the result can be more easily proven. The alternating method was considered in [\textit{J. von Neumann}, Functional operators. Vol. II. The geometry of orthogonal spaces. Princeton: Princeton University Press (1950; Zbl 0039.11701)], in the Hilbert space setting when \(r=2\) without the requirement that the sum is closed. For a finite number of subspaces the result was deduced in [\textit{I. Halperin}, Acta Sci. Math. 23, 96--99 (1962; Zbl 0143.16102)]. Notice that the linearity and orthogonality properties of the orthogonal projections in linear subspaces of a Hilbert space play a key role in the analysis of the alternating method in such a situation.