5
$\begingroup$

I'm studying properties of the manifold of symmetric positive-definite (SPD) matrices and I've learnt about the following connection to the hyperbolic space [1, Section 2.2], $$\mathcal{P}(n) = \mathcal{SP}(n) \times \mathbb{R}^+,\qquad \mathcal{SP}(n) \cong \mathbb{H}^p,$$ where $\mathcal{P}(n)$ is the space of SPD matrices, $\mathcal{SP}(n) = \{ A \in \mathcal{P}(n) : \lvert A \rvert = 1 \}$, and $p = n (n + 1) / 2 - 1$. In words,

$\mathcal{P}(n)$ is a foliated manifold whose codimension-one leaves are isomorphic to the hyperbolic space $\mathbb{H}^p$.

I wanted to see a proof of this isomorphism (and what it means more precisely) and I found [2] and its follow up [3] which focus on $2 \times 2$ SPD matrices. They mention certain connections between the standard metrics in these spaces, e.g. [3, p4], \begin{align}\tag{1}\label{eq:1} d_{\mathcal{P}}(X_1,X_2) = \sqrt{\frac{1}{2} \big(\log \lvert X_1 \rvert - \log \lvert X_2 \rvert \big)^2 + d_{\mathbb{D}}^2(y_1,y_2)}, \end{align} where $d_{\mathcal{P}}$ is the canonical metric on $\mathcal{P}(n)$ (see [2, eq. (6)]), $d_{\mathbb{D}}(y_1,y_2)$ is the distance in the Poincaré disk between two points that "correspond" to the SPD matrices $X_1, X_2$ (more precisely, their scaled versions $\tilde{X}_1,\tilde{X}_2 \in \mathcal{SP}(n)$; see [2, p4-6]).

That being said, I couldn't find a proof for the general case of $n \times n$ SPD matrices. Is anyone aware of other relevant resources? Or is it really obvious? Does a generalized form of \eqref{eq:1} still hold?

Thank you.


[1]: Moakher, M. (2005). A differential geometric approach to the geometric mean of symmetric positive-definite matrices. SIAM Journal on Matrix Analysis and Applications, 26(3), 735-747.

[2]: Chossat, P., & Faugeras, O. (2009). Hyperbolic planforms in relation to visual edges and textures perception. PLoS Computational Biology, 5(12), e1000625.

[3]: Faye, G., Chossat, P., & Faugeras, O. (2011). Analysis of a hyperbolic geometric model for visual texture perception. The Journal of Mathematical Neuroscience, 1(1), 4.

$\endgroup$
2
  • 2
    $\begingroup$ It is just false for $n\ge 3$; what you get is a different symmetric space. $\endgroup$
    – Misha
    May 28, 2019 at 13:43
  • $\begingroup$ @Misha You're saying $\mathcal{SP}(n) \cong \mathbb{H}^{n (n + 1) / 2 - 1}$ is false for $n \ge 3$? Could you please elaborate? $\endgroup$
    – Călin
    May 28, 2019 at 13:52

1 Answer 1

7
$\begingroup$

This 'isomorphism' does not hold for $n>2$, in the sense that the 'natural' $\mathrm{SL}(n,\mathbb{R})$-invariant metric on what you are calling $\mathcal{SP}(n)$ and the constant sectional curvature hyperbolic metric on $\mathbb{H}^p$ for $p = n(n{+}1)/2-1$ are only isometric when $n=1$ (the trivial case) and $n=2$.

The reason is that $\mathcal{SP}(n) = \mathrm{SL}(n,\mathbb{R})/\mathrm{SO}(n)$ as symmetric spaces while $\mathbb{H}^p = \mathrm{SO}(p,1)/\mathrm{SO}(p)$ as symmetric spaces. In fact, for $n>2$, the symmetric space $\mathrm{SL}(n,\mathbb{R})/\mathrm{SO}(n)$ does not have constant sectional curvature, while hyperbolic space $\mathrm{SO}(p,1)/\mathrm{SO}(p)$ has constant sectional curvature for all $p$.

When $n=2$, it's one of the 'accidental' isomorphisms that $\mathrm{SO}(2,1)$ is double-covered by $\mathrm{SL}(2,\mathbb{R})$. For $p>2$, the group $\mathrm{SO}(p,1)$ is not evenly-covered by any $\mathrm{SL}(n,\mathbb{R})$ for any $n$.

$\endgroup$
3
  • $\begingroup$ I see. Thanks a lot, Robert. It's funny that the author states the property in a pretty generic way in several places, though. $\endgroup$
    – Călin
    May 28, 2019 at 14:20
  • 1
    $\begingroup$ @CălinCruceru: What's true is that both spaces are diffeomorphic to $\mathbb{R}^p$ as smooth manifolds, but that is a very weak statement (and essentially obvious to boot). When you asked about 'isomorphism', I assumed that you were asking about 'isometry'. Because they are both spaces of non-positive sectional curvature (and simply-connected) there is a 'canonical' diffeomorphism between them (once base points have been chosen in each) got by comparing their exponential maps at their respective base points, but, when $n>2$, that map is very far from being an isometry. $\endgroup$ May 28, 2019 at 16:58
  • $\begingroup$ The reason I used "isomorphism" is because that's what the author of the first reference uses and I wasn't sure myself what it means. After I found the 2nd and 3rd references, it was clear that there is an isometry for the case $n=2$, but I was still wondering what he means for $n > 2$. $\endgroup$
    – Călin
    May 29, 2019 at 9:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.