Introduction
In computational science, data often resides on non-Euclidean manifolds rather than flat vector spaces. Fields such as numerical relativity and computer vision frequently encounter variables like Lorentzian metrics or Grassmannian elements. Naive component-wise interpolation of such data can lead to physically invalid results, such as a metric losing its required signature. This paper presents a framework for structure-preserving interpolation on symmetric spaces by leveraging the generalized polar decomposition (GPD). The goal is to ensure that interpolated values remain on the manifold, maintain frame invariance, and are computed efficiently.
Methodology: The GPD Framework
The proposed method exploits the mathematical relationship between symmetric spaces—manifolds with inversion symmetry—and Lie triple systems, which are specialized vector spaces. The core of the approach is the Generalized Polar Decomposition (GPD), which factorizes elements of a Lie group into a product representing a point on the symmetric space.
The interpolation algorithm follows a three-step process:
- Mapping to Flat Space: Data points from the manifold are transformed into a Lie triple system using the inverse of a GPD-based map.
- Interpolation: Standard weighted interpolation, such as using Lagrange polynomials, is performed within this vector space.
- Mapping Back: The interpolated vector is mapped back to the manifold via the exponential map.
This framework provides a "one-shot" interpolation formula and an iterative variant. The authors demonstrate that for certain spaces like the Grassmannian, the iterative version is equivalent to finding the Karcher mean but offers significantly better computational efficiency, scaling linearly with ambient dimension.
Results and Evaluation
The framework was evaluated using the Schwarzschild metric, a standard benchmark in numerical relativity representing spacetime around a spherical mass. The tests focused on the performance of "Lorentzian-valued finite elements."
Performance and Robustness
Numerical experiments confirmed that the GPD-based interpolant achieves optimal approximation orders: $O(h^{k+1})$ in the $L^2$-norm and $O(h^k)$ in the $H^1$-norm for degree $k$ polynomials. Crucially, while standard linear interpolation failed by producing positive-definite matrices instead of Lorentzian ones, the GPD method consistently preserved the correct (3, 1) signature.
Frame Invariance
A key property of the GPD operators is their frame invariance. The interpolation commutes with the action of underlying symmetry groups, such as the indefinite orthogonal group $O(1, 3)$, ensuring that results are independent of the observer's coordinate system.
Conclusion and Future Work
The GPD framework offers a robust, efficient, and geometrically consistent method for interpolating manifold-valued data. It unifies existing techniques, such as the log-Euclidean mean, into a broader mathematical context. Future work aims to extend these operators into full-scale dynamic simulations in numerical relativity and to provide formal theoretical proofs for convergence rates across all general GPD cases.
Summary updated