Abstract:Positive semidefinite (PSD) matrices are indispensable in many fields of science. A similarity measurement for such matrices is usually an essential ingredient in the mathematical modelling of a scientific problem. This paper proposes a unified framework to construct similarity measurements for PSD matrices. The framework is obtained by exploring the fiber bundle structure of the cone of PSD matrices and generalizing the idea of the point-set distance previously developed for linear subsapces and positive definite (PD) matrices. The framework demonstrates both theoretical advantages and computational convenience: (1) We prove that the similarity measurement constructed by the framework can be recognized either as the cost of a parallel transport or as the length of a quasi-geodesic curve. (2) We extend commonly used divergences for equidimensional PD matrices to the non-equidimensional case. Examples include Kullback-Leibler divergence, Bhattacharyya divergence and Rényi divergence. We prove that these extensions enjoy the same consistency property as their counterpart for geodesic distance. (3) We apply our geometric framework to further extend those in (2) to similarity measurements for arbitrary PSD matrices. We also provide simple formulae to compute these similarity measurements in most situations.
What problem does this paper attempt to address?
The problems that this paper attempts to solve are: How to construct a unified framework for positive semi - definite (PSD) matrices to measure the similarity between them. Specifically, the paper addresses the following key issues:
1. **Limitations of Existing Methods**:
- When applied to PSD matrices, existing distance functions fail to reflect their intrinsic geometric structures. For example, geodesics under the Frobenius distance do not preserve rank, which can lead to adverse effects in applications.
- For PSD matrices of different dimensions, commonly used divergences (such as Kullback - Leibler divergence, Bhattacharyya divergence, and Rényi divergence) cannot be directly applied.
2. **Utilization of Geometric Structures**:
- The paper proposes a method based on fiber bundle structures. By exploring the geometric properties of the PSD matrix cone, a new similarity measure is constructed. This method not only has theoretical advantages but is also more computationally convenient.
- The author generalizes the concept of point - set distance to linear subspaces and positive definite matrices and further extends it to PSD matrices of arbitrary dimensions.
3. **Specific Objectives**:
- Propose a unified framework that can construct similarity measures applicable to all PSD matrices.
- Prove that the newly proposed similarity measure can be interpreted as the cost of parallel transport or the length of a quasi - geodesic.
- Extend common divergences from equal - dimensional positive definite matrices to non - equal - dimensional cases and prove that these extensions are consistent.
4. **Practical Applications**:
- Provide simple formulas to calculate these similarity measures, making them easy to implement in most cases.
### Formula Representation
To understand the mathematical concepts in the paper more clearly, here are the Markdown representations of several key formulas:
- **Definition of PSD Matrix**:
\[
A\in\text{Herm}_n^r = \{A\in\text{Herm}_n:\text{rank}(A) = r, A\succeq0\}
\]
- **Definition of Geometric Distance**:
\[
GD_{d,\delta}(A, B)=\left(d^2(\pi_{n,r}(A),\iota_{m,n}\circ\pi_{m,s}(B))+\delta_H^2(Z_{A|\pi_{n,r}(A)},\iota_m^*(B)|\iota_{m,n}\circ\pi_{m,s}(B))\right)^{\frac{1}{2}}
\]
where \(Z_{A|\pi_{n,r}(A)},\iota_m^*(B)|\iota_{m,n}\circ\pi_{m,s}(B)\) are sets defined according to principal vectors, and \(\delta_H\) is the generalized Hausdorff function.
- **Principal Angles and Principal Vectors**:
\[
\theta_i=\arccos(a_i^*b_i)
\]
where \(a_i\) and \(b_i\) are the principal vectors of subspaces \(U\) and \(V\), respectively.
Through these methods, the paper successfully solves the problem of how to measure the similarity of PSD matrices of different dimensions while maintaining geometric properties.