---
title: Hilbert VPM Distance on SPD Matrices
url: https://www.emergentmind.com/topics/hilbert-vpm-distance
type: topic
---

# Hilbert VPM Distance on SPD Matrices

The Hilbert VPM distance is a projective metric defined on the open symmetric positive-definite bicone, also known as the variance–precision manifold (VPM). It provides a closed-form, projectively invariant, and computationally efficient dissimilarity for symmetric positive-definite (SPD) matrices, generalizing the classical Hilbert projective metric from cones and the simplex to matrix-valued settings. Its structure underpins both advanced geometric data analysis of extended Gaussian models and high-dimensional clustering of matrix-valued data [2508.14369, 2603.02483].

## 1. Mathematical Foundation and Domain

The Hilbert VPM distance is developed in the context of convex geometry and projective metrics. For a generic open bounded convex set $C \subset \mathbb{R}^d$, the Hilbert distance between points $x, y \in C$ is defined as the logarithm of a cross-ratio determined by the boundary intersections of the line through $x$ and $y$. Specializing to matrix spaces, the focus is on
$$
\operatorname{VPM}(n) = \{ X \in \operatorname{Sym}(n) : 0 \prec X \prec I \},
$$
where $\operatorname{Sym}(n)$ denotes real symmetric $n \times n$ matrices, and $0 \prec X \prec I$ requires all eigenvalues of $X$ lie in $(0, 1)$. This set arises naturally as the parameter space for the extended Gaussian family, encompassing both covariance and precision degeneracies [2508.14369, 2603.02483].

James' bicone reparametrization provides the diffeomorphic map
$$
\iota: PD(n) \to \operatorname{VPM}(n), \quad \iota(P) = P (I+P)^{-1}
$$
with $PD(n)$ the SPD cone, so any element of the VPM can be interpreted as a normalized blend of covariance and precision structures.

## 2. Closed-Form Expression

The core result is an explicit spectral formula for the Hilbert VPM distance $d_H(A, B)$ between $A, B \in \operatorname{VPM}(n)$. Define
\[
\lambda_{\min} = \lambda_{\min}(B^{-1}A),\quad
\lambda_{\max} = \lambda_{\max}(B^{-1}A)
\]
\[
\mu_{\min} = \lambda_{\min}((I-B)^{-1}(I-A)),\quad
\mu_{\max} = \lambda_{\max}((I-B)^{-1}(I-A))
\]
Then,
\[
d_H(A,B) = \log\left( \frac{\max\{\lambda_{\max},\,\mu_{\max}\}}{\min\{\lambda_{\min},\,\mu_{\min}\}} \right)
\]
Alternatively,
\[
d_H(A,B) = \log\left( \frac{\Lambda_{\max}(A,B)}{\Lambda_{\min}(A,B)} \right)
\]
with
\[
\Lambda_{\max}(A,B) = \max\{\lambda_{\max}(B^{-1}A),\,\lambda_{\max}((I-B)^{-1}(I-A))\}
\]
\[
\Lambda_{\min}(A,B) = \min\{\lambda_{\min}(B^{-1}A),\,\lambda_{\min}((I-B)^{-1}(I-A))\}
\]
This provides rapid evaluation and numerical stability, even near the boundary of the bicone [2508.14369, 2603.02483].

## 3. Metric and Geometric Properties

The Hilbert VPM distance is a true metric, satisfying:

- **Non-negativity**: $d_H(A,B) \geq 0$ and $d_H(A,B) = 0$ iff $A = B$.
- **Symmetry**: $d_H(A,B) = d_H(B,A)$, as the spectral formula is invariant under exchange of $A$ and $B$.
- **Triangle inequality**: Inherited from the cross-ratio construction for convex domains; straight line geodesics in $\operatorname{Sym}(n)$ are isometrically mapped to shortest paths after appropriate reparametrization [2603.02483].
- **Projective invariance**: The metric is invariant under positive congruence transformations and scaling.
- **Geodesic Structure**: Geodesics correspond to straight lines in the affine ambient space, but require non-uniform parameterization in the Finsler geometry given by the Hilbert metric.
- **Finsler Norm**: The infinitesimal norm at $X$ in direction $V$ is
$$
\|V\|^H_X = \frac{1}{t^+} + \frac{1}{t^-}
$$
where $t^+,\ t^-$ are exit times to the boundary along $V$ and $-V$ [2603.02483], admitting closed-form expressions involving extremal eigenvalues.

## 4. Invariance and Isometries

Hilbert VPM distance possesses several notable isometries:

- **Identity–complement invariance**: $d_H(I-A,I-B) = d_H(A,B)$ for all $A, B \in \operatorname{VPM}(n)$.
- **O($n$)-congruence invariance**: For any $U \in O(n)$, $d_H(U^T A U, U^T B U) = d_H(A,B)$.
For $n \geq 2$, these transformations generate the full isometry group on $(\operatorname{VPM}(n), d_H)$ [2508.14369].

## 5. Connections to Classical Metrics and Special Cases

The Hilbert VPM distance generalizes classical projective geometry and is related as follows:

- **Simplex Limit**: When restricted to diagonal matrices with unit trace ($\operatorname{diag}(p)$ with $p \in \Delta^\circ_n$), the Hilbert VPM distance reduces precisely to the Hilbert projective metric on the simplex:
$$
d_H(\operatorname{diag}(p),\operatorname{diag}(q)) = \log\left( \frac{\max_i p_i/q_i}{\min_i p_i/q_i} \right)
$$
- **Scalar Case $n=1$**: On $(0,1)$, $d_H(x,y) = |\log(x/y)|$, recovering the classical cross-ratio metric [2603.02483].
- **Isotropic Case**: $A = \alpha I$, $B = \beta I$ with $\alpha,\beta \in (0,1)$ yields
$$
d_H(A,B) = \log\left( \frac{\max\{\alpha/\beta, (1-\alpha)/(1-\beta)\}}{\min\{\alpha/\beta, (1-\alpha)/(1-\beta)\}} \right)
$$
- **Relation to Fisher–Rao (AIRM)**: The Hilbert VPM distance is a projective-invariant analogue to the affine-invariant Riemannian metric (AIRM), but is defined and continuous even for degenerate limit cases where standard Riemannian geometry fails [2508.14369].

## 6. Computational Aspects

Evaluation of $d_H(A,B)$ requires only the four extremal eigenvalues of $B^{-1}A$ and $(I-B)^{-1}(I-A)$. Computational complexity is $O(n^2)$–$O(n^3)$ per pair, significantly outperforming metrics that require full spectral decompositions, especially in high dimensions. Power and Lanczos iterations provide efficient calculations of extremal eigenvalues. No iterative fitting is required; a single eigenproblem per matrix pair suffices. The metric is highly stable under degeneracy, as the quotient of extremal eigenvalues remains finite unless matrices approach disjoint boundary faces where $d_H \rightarrow \infty$ [2508.14369, 2603.02483].

## 7. Applications and Implications

The Hilbert VPM distance is directly applicable to:

- **Gaussian Distribution Analysis**: As the parameter space of the extended Gaussian family, $\operatorname{VPM}(n)$ encompasses degeneracies in both covariance and precision, and $d_H$ provides a continuous, projective-invariant divergence between $N(0,\Sigma_A)$ and $N(0,\Sigma_B)$ (or their inverses), accommodating nearly singular cases [2508.14369].
- **Clustering and Statistical Learning**: Straight-line geodesics and efficient extremal computations facilitate $k$-means, $k$-centers, and minimum enclosing ball algorithms for SPD-valued data, notably outperforming log-det and Frobenius distance–based approaches in empirical clustering of correlations and covariances [1704.00454, 2603.02483].
- **Boundary-Aware Learning**: The metric’s stability as eigenvalues approach 0 or 1 makes it suitable for modeling covariance/precision degeneracy, a regime where standard metrics become either ill-defined or lose sensitivity.
- **Bridging Simplex Geometry**: The Hilbert VPM metric subsumes and extends the simplex Hilbert geometry, enabling unified clustering and comparison frameworks for discrete (histogram) and continuous (SPD-matrix) data domains [1704.00454, 2603.02483].

Additionally, tight Lipschitz-type inequalities relate the Hilbert VPM distance to AIRM and log-barrier Hessian distances on the bicone, providing comparative control over metric distortion in geometric statistics [2603.02483].

## 8. Summary Table: Key Hilbert VPM Distance Features

| Property                           | Description                                                             | Reference        |
|-------------------------------------|-------------------------------------------------------------------------|------------------|
| Domain                             | $0 \prec X \prec I$, $X \in \operatorname{Sym}(n)$                       | [2508.14369]     |
| Formula                            | $\log \max\{\lambda_{\max},\mu_{\max}\} / \min\{\lambda_{\min},\mu_{\min}\}$ | [2508.14369]     |
| Projective invariance              | Invariant under $U^T X U$, $X \mapsto I - X$                             | [2508.14369]     |
| Computational cost                  | $O(n^2)$–$O(n^3)$ (extremal eigenvalues only)                           | [2508.14369]     |
| Generalizes simplex Hilbert metric | True on diagonal matrices                                              | [2603.02483]     |
| Finsler structure                  | Infinitesimal norm via exit times, explicit in eigenvalues              | [2603.02483]     |
| Applications                       | Gaussian models, clustering SPD matrices, boundary sensitivity           | [2508.14369]     |

The Hilbert VPM distance is thus a natural projective-Finsler metric for symmetric positive-definite bicones, furnishing a unifying, tractable, and theoretically rich tool for extended statistical modeling, convex-geometric analysis, and large-scale SPD data clustering [2508.14369, 1704.00454, 2603.02483].

Source: https://www.emergentmind.com/topics/hilbert-vpm-distance