The multi-dimensional von Neumann alternating direction search algorithm in \(C(B)\) and \(L_1\) (Q1188107)

An \(n\)-dimensional subspace \(U\) of a normed linear space \(X\) is given, where \(U\) is spanned by \(m\) elements \(u_1,\ldots,u_m\). To find the best approximation to a given element \(f\in X\) from \(U\), this paper gives an algorithm to solve this problem by cyclically searching the one- dimensional spaces spanned by each \(u_i\) (\(i=1,\ldots,m\)). The best approximation \(u^{(k)}\) from the one-dimensional spaces is subtracted from \(f\) before searching in the next direction. This algorithm is essentially the cyclic coordinate algorithm described by J. von Neumann or by \textit{W. I. Zangwill} [Nonlinear programming. A unified approach. Englewood Cliffs, NJ: Prentice-Hall (1969; Zbl 0195.20804)], in special cases the Diliberto-Strauss algorithm is obtained. Two questions are studied: Does the sequence \(u^{(k)}\) converge and does it converge to a best approximation to \(f\) from \(U\)? There are some new results in so called smooth linear spaces \(X\) to characterize convergence. Detailed considerations of the authors' method are given in the spaces \(C(B)\) and \(L_1\), respectively.