The rate of convergence for the cyclic projections algorithm. II: Norms of nonlinear operators (Q855484)

The study of algorithms to identify points in the intersection of convex sets is of classical importance and has applications to physical problems. Let \(C_1, \dots , C_n\) be nonempty closed convex subset of a Hilbert space \(H\). For \(x \in H\) let \(P_k(x)\) be the best approximation to \(x\) from \(C_{k \bmod n}\). Let \(Q_1(x) = P_1(x)\), and \(Q_{m+1}(x) = P_{m+1}(Q_m(x))\) for \(m=2, 3, 4, \dots \). The sequence \(Q_m(x)\) is the cyclic projection algorithm. If the intersection of the sets \(C_i\) is not empty then the cyclic projection algorithm converges weakly to some point in the intersection. If the \(C_i\) are linear spaces the cyclic projection algorithm is known to converge strongly to the best approximation to \(x\) from the intersection of the spaces \(C_i\) (von Neumann-Helperin theorem). If the \(C_i\) are not linear, the limit of the cyclic projection algorithm might not be the best approximation to \(x\). This work studies the rate of convergence of the cyclic projection algorithm. The bounds depend on a definition of an angle between convex sets. The angle is defined in terms of a localized dual cone of a convex set. [For Part I, cf. the authors, ibid. 142, No. 1, 36--55 (2006; Zbl 1109.41016).]