Many contemporary datasets arise on nonlinear manifolds with unignorable outliers or heavy-tailed underlying distributions, where geometric structure must be explicitly respected while considering the robust estimation. In this work, we investigate Ri...
Many contemporary datasets arise on nonlinear manifolds with unignorable outliers or heavy-tailed underlying distributions, where geometric structure must be explicitly respected while considering the robust estimation. In this work, we investigate Riemannian robust M-estimators, combining the robustness of M-estimators with the intrinsic geometry of Riemannian manifolds. This synthesis addresses both the statistical robustness and the geometric necessity of manifold-aware analysis. Our framework is evaluated across manifolds—including Euclidean, sphere and hyperbolic spaces through extensive simulations. In particular, we study the calibration of cutoff parameters for downweighting outliers and propose ranges that achieve approximately 95% of relative efficiency under Gaussian-like distributions, analogous to Euclidean benchmarks. For heavy-tailed distributions, M-estimators are shown to have greater efficiency than the usual Fr´echet mean. We demonstrate the practical utility of our framework through real-world applications such as earthquake epicenter on the Earth, wind direction fields and more. Notably, robust M-estimators yield more reliable estimates than the Fr´echet mean. From a computational perspective, we contribute implementations for computation of manifold-valued M-estimators within the open-source library Geomstats, enabling automatic Riemannian gradient descent and customized loss functions. Our results establish robust M-estimation as a principled, computationally tractable, and empirically effective tool for manifold-valued data analysis.