Title:Adaptive regularization with cubics on manifolds with a first-order analysis

Abstract:We generalize the adaptive regularization with cubics algorithm (ARC) to Riemannian optimization. Building upon recent literature on the Euclidean case, we obtain an upper bound on the iteration complexity of computing an $\varepsilon$-critical point of a cost function on a manifold using this algorithm. Upon introducing appropriate assumptions---notably to account for the effects of a chosen retraction---the complexity is essentially the same as the one obtained in the Euclidean case, that is, it may take up to $O(1/\varepsilon^{1.5})$ iterations at most to produce an $\varepsilon$-critical point, regardless of initialization. A substantial part of the paper is devoted to studying the new assumptions and providing user-friendly sufficient conditions for them. Numerical experiments show that this can become a competitive alternative to other state-of-the-art solvers such as the Riemannian trust-region method.