Towards Quotient Barycentric Subspaces

Anna Calissano, Elodie Maignant, Xavier PennecLecture Notes in Computer Science 2023 · DOI · OpenAlex

Introduction

Dimensionality reduction is a cornerstone of data analysis, yet standard techniques often falter when data possesses inherent symmetries. Objects like unlabeled graphs or 3D shapes reside in quotient spaces—mathematical structures formed by grouping equivalent points under a symmetry action. This paper introduces Quotient Barycentric Subspaces (QBS), extending Barycentric Subspace Analysis (BSA) to these complex, often non-manifold environments.

The Challenge of Quotient Spaces

While statistical tools thrive on smooth manifolds, quotient spaces are frequently "stratified," containing singularities where symmetry actions are not free. For instance, in a reflection space, the origin acts as a fixed point, breaking the manifold structure. Traditional PCA or manifold-based BSA cannot naturally account for these geometric hurdles. BSA is particularly valuable because it defines subspaces through interpretable reference points; however, its application has historically been limited to manifolds.

Defining Quotient Barycentric Subspaces

The authors propose QBS by integrating three key constructs:

  1. Alignment Map: Optimally aligns orbits (sets of equivalent points) before comparison.
  2. Horizontal Logarithm: Adapts the manifold "log map" to linearize space locally within the total space of aligned points.
  3. Implicit Definition: A QBS consists of points expressible as barycentric combinations of reference orbits, provided they are optimally aligned.
Positive Quotient Barycentric Subspaces in a reflection space, illustrating how subspaces can manifest as disjoint segments.

Projection Algorithms

Projecting data onto QBS is computationally demanding due to both combinatorial and geometric complexities. The paper evaluates two primary strategies:

  • Algorithm 1 (Align to Reference Points): Anchors alignment to the closest reference point.
  • Algorithm 2 (Align to Data Point): Anchors alignment to the data point itself.

Experiments in a synthetic reflection space ($\mathbb{R}^2$ under $(x_1, x_2) \sim (-x_1, -x_2)$) reveal that Algorithm 2 is significantly more robust. It successfully identifies "incomplete segments"—parts of the subspace that do not contain the original reference points but are valid representations in the quotient space. Algorithm 1 often fails when data points are closer to these incomplete segments.

Conclusion and Future Work

QBS provides a more accurate "skeleton" for data with symmetries than manifold-only methods. While Algorithm 2 offers a practical path forward, challenges remain in scaling these combinatorial alignments to high-dimensional data, such as large graph spaces. Future research will focus on moving from point-to-point alignment to more sophisticated point-to-subspace procedures.

Summary updated

BibTeX
@inproceedings{calissano2023towards,
  author = {Anna Calissano and Elodie Maignant and Xavier Pennec},
  title = {{Towards Quotient Barycentric Subspaces}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2023},
  doi = {10.1007/978-3-031-38271-0\_36},
}
Figure