Paperlane

no. 16 · Tuesday 17 November 2026

Extrinsic Curvature Estimation by Closest Point Projection of Rays

Liangchen Liu and Yen-Hsi Richard Tsai

math.DG · v2 · 33 pp

Abstract

We propose a parametrization-free method that estimates the principal curvatures and principal directions of a submanifold of $\mathbb{R}^d$ from the closest point projections of a few random rays, without coordinate charts or local fitting. The projection of a ray induces a curve on the submanifold, which is a geodesic only in special cases, so its curvature is in general not a curvature of the submanifold. The velocities of these curves, however, relate to the derivative of the closest point projection, whose eigenpairs carry the information of the principal curvatures and directions. Finite differences along $d$ rays can estimate them without relying on the normal information at a cost of $\mathcal{O}(d^3)$ besides the projections, and we provide error bounds. The method extends to submanifolds of any codimension and to point clouds, and experiments confirm the predicted rates in up to $128$ dimensions.

comments
33 pages, 7 figures, refined and polished based upon the previous version
fields
Differential Geometry
pdf ↗ page arXiv:2610.33943 ↗