The rate of convergence for the cyclic projections algorithm. III: Regularity of convex sets (Q999284)

This is the third paper by the authors on the rate of convergence for the cyclic projections algorithm (CPA). In the first of these papers [ibid. 142, No. 1, 36--55 (2006; Zbl 1109.41016); ibid. 142, No. 1, 56--82 (2006; Zbl 1109.41017)], the authors showed that the rate could be described in terms of the ``angles'' between the convex sets involved. In the second, they showed that these angles often had a more tractable formulation in terms of the ``norm'' of the product of the (nonlinear) metric projections onto related convex sets. In the present paper, the authors show that the rate of convergence of the cyclic projections algorithm is also intimately related to the ``linear regularity property'' of Bauschke and Borwein, the ``normal property'' of Jameson (as well as Bakan, Deutsch, and Li's generalization of Jameson's normal property), the ``strong conical hull intersection property'' of Deutsch, Li, and Ward, and the rate of convergence of iterated parallel projections. The authors establish some relationships between the strong CHIP (strong conical hull intersection property) and various regularity properties that a finite collection of \(r\) convex sets (with \(r\geq 2\)) may possess. In particular, it is shown that each of the regularity properties for a collection of \(r\) sets is equivalent to the same regularity property for just two related sets, the latter defined in a suitable product space. (In the particular case of linear subspaces, this fact was previously established by \textit{H. H. Bauschke, J. M. Borwein} and \textit{A. S. Lewis} [in: Recent developments in optimization theory and nonlinear analysis. AMS/IMU special session on optimization and nonlinear analysis, May 24--26, 1995, Jerusalem, Israel. Providence, RI: American Mathematical Society. Contemp. Math. 204, 1--38 (1997; Zbl 0874.47029), Proposition 3.7.3]. In Section 4, the authors show that there is a very strong correlation between the speed of convergence of the CPA and the linear regularity property as well as the strong CHIP and various other regularity properties of the sets in question. For example, Theorem 4.5 characterizes linear regularity as the precise property equivalent to ``uniform linear convergence'' of the CPA. Moreover, in the case when all the sets are subspaces, Remark 4.12 shows that there are 44 conditions, each equivalent to the uniform linear convergence of the CPA.