On the convergence of some Procrustean averaging algorithms

From MaRDI portal





The author deals with algorithms acting on Euclidean shape space \(\Sigma_N^k\) or size-and-shape space \(\widetilde{\Sigma}_N^k\). These are the spaces of configurations of \(k\) non-identical labeled points in \(\mathbb{R}^N\) modulo equivalence under translations, rotations and dilatations (rescalings) in \(\Sigma_N^k\), or just under translations and rotations in \(\widetilde{\Sigma}_N^k\). The aim of these algorithms is to average a sample of shapes, or size-and-shapes, thought of as points in one of these parameter spaces. Two different definitions of ``average are explored, namely the Procrustean mean and the Riemannian average. On size-and-shape spaces the Procrustean mean coincides with the Riemannian average, but not on shape spaces. Several algorithms based on generalized Procrustean analysis (GPA) are considered and compared with a general Riemannian averaging algorithm (RA). If \(\mathcal{Q}\) is a list of points in a Riemannian manifold and \(\text{diam}(\mathcal{Q})\) is not too large, then the map generating this RA algorithm is a contraction with a constant that is quadratic in \(\text{diam}(\mathcal{Q})\) as \(\text{diam}(\mathcal{Q})\to 0\). In particular, two proofs of the convergence of GPA-shape (for suitable initial conditions) are given. The first one applies only to the geometrically special case \(N=2\), but establishes convergence for a larger set of initial conditions than does the second one. On the other hand, the second proof applies to all \(N\), and if \(N=2\), it gives a better bound on the convergence rate for small values of \(\text{diam}(\mathcal{Q})\). Moreover, if \(\text{diam}(\mathcal{Q})\to 0\), the distance between the Procrustean mean and the Riemannian average is asymptotically computed, and an interesting comparison of the convergence rates of the Procrustean and Riemannian algorithm is given.











This page was built for publication: On the convergence of some Procrustean averaging algorithms

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q5460724)