Schubert varieties and distances between subspaces of different dimensions

Ke Ye, Lek‐Heng Lim

Abstract

We resolve a basic problem on subspace distances that often arises in applications: How can the usual Grassmann distance between equidimensional subspaces be extended to subspaces of different dimensions? We show that a natural solution is given by the distance of a point to a Schubert variety within the Grassmannian. This distance reduces to the Grassmann distance when the subspaces are equidimensional and does not depend on any embedding into a larger ambient space. Furthermore, it has a concrete expression involving principal angles, and is efficiently computable in numerically stable ways. Our results are largely independent of the Grassmann distance --- if desired, it may be substituted by any other common distances between subspaces. Our approach depends on a concrete algebraic geometric view of the Grassmannian that parallels the differential geometric perspective that is well-established in applied and computational mathematics.

Introduction

In fields like computer vision, signal processing, and machine learning, data is frequently represented as subspaces—for example, the principal components of a dataset. While the Grassmannian provides a robust framework for comparing subspaces of the same dimension, real-world applications often involve subspaces of differing dimensions. For instance, comparing noise-floor subspaces of signals with varying signal-to-noise ratios requires a distance metric that can handle such discrepancies.

Previous attempts to define distances in these cases were often ad hoc, failing to respect the underlying geometry or depending unnecessarily on the dimension of the ambient space. This paper introduces a mathematically rigorous, intrinsic distance that generalizes the standard Grassmann distance to the inequidimensional case.

Geometric Framework: Schubert Varieties as Targets

The authors bridge the gap between differential geometry and algebraic geometry to solve this problem. Instead of comparing a $k$-dimensional subspace $A$ directly to an $l$-dimensional subspace $B$, they define the distance as the distance from $A$ to the nearest $k$-dimensional subspace contained within $B$.

Geometric representation of the distance from a subspace $A$ to a Schubert variety $\Omega_-(B)$ in the Grassmannian $Gr(1,3)$.

This set of subspaces contained in $B$ forms a Schubert variety, denoted $\Omega_-(B)$. A key theoretical contribution is the proof of symmetry: the distance from $A$ to the nearest $k$-subspace in $B$ is identical to the distance from $B$ to the nearest $l$-subspace containing $A$. To handle these comparisons regardless of the ambient dimension $n$, the authors utilize the concept of a "doubly infinite Grassmannian," ensuring the metric remains intrinsic to the subspaces themselves.

Generalizing Subspace Distances

The proposed distance $\delta(A, B)$ between subspaces of dimension $k$ and $l$ can be computed efficiently using the Singular Value Decomposition (SVD). Specifically, it is defined by the principal angles $\theta_i$ between the spaces: $$\delta(A, B) = \left( \sum_{i=1}^{\min(k,l)} \theta_i^2 \right)^{1/2}$$

One of the paper's most practical results is the extension of eight common Grassmannian distances to the inequidimensional case. These include the Asimov, Binet-Cauchy, Chordal, Fubini-Study, Martin, Procrustes, Projection, and Spectral distances.

Common Grassmannian distances expressed via principal angles and orthonormal bases, which the authors extend to subspaces of different dimensions.

By providing these generalizations, the authors show that their framework is largely independent of the specific choice of Grassmann distance; any standard metric can be substituted into the Schubert variety framework.

Key Insights and Theoretical Implications

  • Intrinsic Nature: The distance between a line and a plane remains constant whether they are embedded in 3D or 1000D space. This independence from the ambient space makes the metric robust for high-dimensional data analysis.
  • Unification of Heuristics: The framework unifies existing heuristics, such as the "containment gap" and "symmetric directional distance," showing they are specific instances (projection and chordal versions) of this broader theory.
  • Probabilistic Duality: The authors derive a volumetric result showing a duality in probabilities: the likelihood of a random $l$-dimensional space containing a fixed $k$-dimensional space is equal to the likelihood of a random $k$-dimensional space being contained within a fixed $l$-dimensional space.

Conclusion and Future Work

While the paper establishes a powerful metric, it notes that the "natural" topology on the space of all subspaces is non-Hausdorff, requiring specific modifications to form a true metric space. Computationally, while SVD is stable, the paper focuses on theoretical foundations rather than large-scale performance benchmarks.

Future research directions include extending these concepts to affine subspaces—subspaces that do not necessarily pass through the origin—which has significant implications for broader geometric data analysis.

Summary updated

Abstract

We resolve a basic problem on subspace distances that often arises in applications: How can the usual Grassmann distance between equidimensional subspaces be extended to subspaces of different dimensions? We show that a natural solution is given by the distance of a point to a Schubert variety within the Grassmannian. This distance reduces to the Grassmann distance when the subspaces are equidimensional and does not depend on any embedding into a larger ambient space. Furthermore, it has a concrete expression involving principal angles, and is efficiently computable in numerically stable ways. Our results are largely independent of the Grassmann distance --- if desired, it may be substituted by any other common distances between subspaces. Our approach depends on a concrete algebraic geometric view of the Grassmannian that parallels the differential geometric perspective that is well-established in applied and computational mathematics.
BibTeX
@misc{ye2014schubert,
  author = {Ke Ye and Lek‐Heng Lim},
  title = {{Schubert varieties and distances between subspaces of different dimensions}},
  year = {2014},
  eprint = {1407.0900},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1407.0900},
}
Figure