The Ultimate Upper Bound on the Injectivity Radius of the Stiefel Manifold

Pierre-Antoine Absil, Simon MataigneSIAM J. Matrix Anal. Appl 2025 · DOI

TL;DR

This paper provides the first explicit upper bound for the injectivity radius of the Stiefel manifold under a broad family of Riemannian metrics. This bound is shown to be equal to the injectivity radius itself for the canonical and Euclidean metrics, a conjecture supported by extensive numerical experiments.

What Problem Does This Paper Address?

The injectivity radius of a Riemannian manifold is a fundamental geometric quantity that dictates the region where geodesic normal coordinates are valid and the extent to which geodesics represent shortest paths. While the Stiefel manifold, a crucial object in areas like optimization and statistics, is well-studied, its injectivity radius remained unknown for a significant family of metrics. This gap hindered a complete understanding of its geometric properties and potential computational applications. The challenge lies in the complex "cut locus" of Riemannian manifolds, which can have intricate structures, making it difficult to find conjugate points and thus determine the injectivity radius.

Methods

The authors derive an upper bound on the injectivity radius by first identifying conjugate points on the Stiefel manifold. This involves leveraging linear algebra lemmas and analyzing the derivative of the matrix exponential. They then combine these findings with bounds on the lengths of geodesic loops. A key aspect is the use of a Riemannian submersion framework to define and analyze metrics on the Stiefel manifold, particularly the family of metrics introduced by Hüper et al. The paper also presents a numerical algorithm to test their derived bounds, which provides strong empirical evidence for their conjectures.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Conjugate Points Found: The paper successfully identifies conjugate points on the Stiefel manifold for a wide range of Riemannian metrics, a crucial step in determining the injectivity radius.
  • "Ultimate" Upper Bound: The derived upper bound is not just any bound; it's presented as the "ultimate" one, meaning it's conjectured to be the exact value of the injectivity radius.
  • Metric Family Analysis: The work systematically analyzes a family of Riemannian metrics, including the canonical and Euclidean ones, revealing how the injectivity radius depends on the metric parameter $\beta$.
  • Numerical Validation: Extensive numerical experiments strongly support the theoretical findings and the conjecture that the upper bound is indeed the injectivity radius.
  • Geometric Significance: The findings contribute to a deeper understanding of the geometry of the Stiefel manifold, which has implications for various computational tasks.

Key Results

  • An explicit upper bound for the injectivity radius of the Stiefel manifold $St_\beta(n,p)$ is derived for a family of Riemannian metrics parameterized by $\beta > 0$.
  • For the canonical metric ($\beta=1$) and the Euclidean metric ($\beta=1$), the derived upper bound is shown to be equal to the injectivity radius for $2 < p < n-2$. Specifically, the upper bound is $t\sqrt{2} \approx 0.913\pi$ for the canonical metric and $\pi$ for the Euclidean metric, where $t$ is the smallest positive root of $\sin t + \cos t = 0$.
  • Numerical experiments strongly suggest that these upper bounds are equalities for a broad range of parameters.

Limitations and Future Work

  • Conjecture for $\beta > 1/2$: While the paper proves the conjecture for $\beta=1$ (canonical metric) and $\beta=1$ (Euclidean metric), it relies on numerical evidence and a conjecture for $\beta > 1/2$ (noncanonical metrics).
  • Algorithm Limitations: The numerical algorithm used to test the conjecture has limitations, particularly when the total space $G$ is not Riemannian ($\beta > 1$), where its return probability is not guaranteed.
  • Future Work: The authors suggest that further theoretical work is needed to prove the conjecture for all $\beta > 1/2$. They also hint at ongoing research into the sectional curvature of Stiefel manifolds, which has implications for their injectivity radius. The paper also mentions that the Stiefel manifold still holds many geometric secrets.

Summary updated

BibTeX
@article{absil2025ultimate,
  author = {Pierre-Antoine Absil and Simon Mataigne},
  title = {{The Ultimate Upper Bound on the Injectivity Radius of the Stiefel Manifold}},
  journal = {SIAM J. Matrix Anal. Appl.},
  year = {2025},
  doi = {10.1137/24m1644808},
}

Loading related content...

Figure