Hermite interpolation with retractions on manifolds (Q6651933)

The authors propose an extension of the De Casterjau algorithm to approximate curves on a manifold interpolating both points and tangents of the curve. It generalizes previous ideas of \textit{K. A. Krakowski} et al. [J. Comput. Appl. Math. 311, 84--99 (2017; Zbl 1357.53044)] and \textit{E. Nava-Yazdani} and \textit{K. Polthier} [Comput. Aided Geom. Des. 30, No. 7, 722--732 (2013; Zbl 1286.65023)]. The basic ideas of the De Casteljau algorithm are maintained.\N\NA classical Euclidean setting consists of recursive linear interpolation between successive control points. In the case of Riemannian manifolds, this requires an invertible retraction map \(R\) that can map the data from the tangential bundle to the manifold and back. This is called retraction based Hermite (RH) interpolation, which results in a piecewise defined interpolating curve \(H(t)\) that matches the prescribed data (position and velocity) in the points \(\{t_k\}_{k=1}^N\).\N\NUnder certain convexity conditions, it is proved that the method is well posed and \(O(h^4)\) convergence is proved where \(h\) is the maximal distance between successive \(t_k\). This is illustrated with numerical examples for the manifolds of rectangular matrices: the Stiefel manifold of column orthogonal matrices and the manifold of matrices with fixed rank and the results are compared with other methods.\N\NThe paper concludes with two examples of practical applications: the prediction-correction continuation method for Riemannian optimization [\textit{A. Séguin} and \textit{D. Kressner}, SIAM J. Optim. 32, No. 2, 1069--1093 (2022; Zbl 1509.65053)] and dynamic low-rank approximation to integrate matrix differential equations [\textit{O. Koch} and \textit{C. Lubich}, SIAM J. Matrix Anal. Appl. 29, No. 2, 434--454 (2007; Zbl 1145.65031)].