Matricial Wasserstein-1 Distance

Yongxin Chen, Tryphon T. Georgiou, Lipeng Ning, Allen R. TannenbaumIEEE Control. Syst. Lett 2017 · DOI

TL;DR

This paper introduces a novel extension of the Wasserstein-1 distance to handle matrix-valued probability densities. This new metric offers a computationally efficient and flexible way to compare complex data structures like power spectra or quantum states, generalizing existing methods for scalar distributions.

What Problem Does This Paper Address?

Optimal Mass Transport (OMT) and its associated Wasserstein distances are powerful tools for comparing probability distributions. While the scalar Wasserstein-1 distance has been well-studied and applied in various fields, extending it to matrix-valued densities – which arise in areas like multivariate spectral analysis, quantum mechanics, and systems theory – has been a significant challenge. Existing approaches often rely on quadratic costs or are computationally intensive. This paper aims to bridge this gap by developing a new, computationally tractable Wasserstein-1 distance for matrix-valued densities, offering a more general and practical framework for comparing these complex data types.

Methods

The authors propose an extension of the Wasserstein-1 metric by leveraging duality theory, specifically a "dual of the dual" formulation. This approach draws inspiration from the Kantorovich-Rubinstein duality for scalar distributions. A key innovation is the introduction of a "quantum gradient operator" and its dual (the divergence operator) acting on spaces of Hermitian matrices. This allows them to define the Wasserstein-1 distance for density matrices (where trace is normalized to 1) and then generalize it to matrix-valued densities and measures, even when they have unequal traces (unbalanced case). The unbalanced case is handled by introducing a source term and considering a trade-off between the transport cost and the difference between the original and transported densities.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and illustrates its concepts with a synthetic example of power spectra.

Key Insights and Ideas

  • Matrix-Valued OMT is Crucial: The paper highlights the growing importance of comparing matrix-valued distributions in diverse scientific domains, motivating the need for robust and efficient metrics.
  • Duality is Key: The authors demonstrate that leveraging duality theory, particularly the "dual of the dual" formulation, provides a computationally advantageous path to extending Wasserstein distances to matrices.
  • Gradient and Divergence on Matrix Spaces: The introduction of specialized gradient and divergence operators for Hermitian matrices is a core technical contribution, enabling the generalization of the metric.
  • Unbalanced Transport for Matrices: The extension to unbalanced matrix-valued densities is a significant step, allowing comparisons even when the total "mass" or trace differs, which is common in real-world applications.
  • Tunable Metric: The unbalanced Wasserstein-1 distance includes a parameter that allows users to tune the relative importance of different features (e.g., translation vs. rotation in spectra), making the metric adaptable to specific application needs.

Key Results

The paper successfully defines and proves properties of the matricial Wasserstein-1 distance. Specifically:

  • It establishes a Wasserstein-1 distance for density matrices (normalized trace) and proves it forms a metric.
  • It extends this to unbalanced matrix-valued densities, showing the equivalence of two different formulations for the unbalanced case.
  • The paper provides a concrete example comparing three matrix-valued power spectra, demonstrating how the tunable parameter in the unbalanced metric can reflect intuitive notions of similarity based on spectral characteristics.

Limitations and Future Work

  • Computational Complexity: While the proposed "dual of the dual" formulation is computationally more attractive than some alternatives, especially for sparse graphs, the paper doesn't provide explicit complexity bounds for the general matrix-valued case.
  • Applications: The paper presents a theoretical framework and a synthetic example. Further research is needed to explore its application to real-world datasets in areas like diffusion tensor imaging, quantum systems, and multivariate time series analysis.
  • Curvature and Robustness: The authors suggest that their framework could lead to new notions of curvature and robustness for networks with matrix-valued data, which is a promising avenue for future research.
  • Comparison with Other Metrics: While the paper contrasts its approach with existing Wasserstein-2 generalizations, a more extensive comparison with other matrix-valued distance metrics would be beneficial.

Summary updated

BibTeX
@article{chen2017matricial,
  author = {Yongxin Chen and Tryphon T. Georgiou and Lipeng Ning and Allen R. Tannenbaum},
  title = {{Matricial Wasserstein-1 Distance}},
  journal = {IEEE Control. Syst. Lett},
  year = {2017},
  doi = {10.1109/lcsys.2017.2699319},
}
Figure