4
$\begingroup$

I have a sequence of symmetric positive definite matrices in $GL(n)$ (in my case, covariance matrices of some gaussians) and vectors $R^n$ (the mean of these gaussians). I consider this sequence to be the discretization of a curve embedded in $GL(n)\times R^n$.

I would like to compute the embedding curvature of this curve at each of my sample points, using the Wasserstein geometry (see Wasserstein Geometry of Gaussian Measures), not information geometry (like The Riemannian Geometry of the Space of Positive-Definite Matrices ...).

I have some difficulties understanding the parameterization of the space of SPD matrices in the Wasserstein Geometry article above and thus, I am not able to compute the Levi-Civita connection and its projection on my curve with respect to the metric. Could anyone help me understand that (or anything else you feel I might not have understood - I feel I barely understand these things) ? Does anyone know of relevant work ?

Thanks!

$\endgroup$

1 Answer 1

4
$\begingroup$

Your question is not specific enough about what you do not understand in the quoted paper; if you want help on this, you should at least explain what you understand and where the problem appears. Here is a little information about bibliography, that might help (or miss the point, I am not sure).

For an introduction to optimal transport and Wasserstein spaces, you can have a look at Villani's books ("Topics on ..." is more elementary, but the beginning of "... Old and New" is not as difficult to read as the size of the book might lead you to think, and I like it a lot). A more concise introduction can also be found in a nice little book by Nicola Gigli, a version of which seems to be at http://math.unice.fr/~gigli/Site_2/Publications_files/users_guide%20-%20final.pdf (but I am not sure this is exactly the text I read).

You should also now about Lott's paper "Some geometric calculations on Wasserstein space", Comm. Math. Phys. 277, p. 423-437, which computes the curvature of the Wasserstein space of a manifold.

Concerning the notion of curvature of a discretized curve, you might be interested in the concept of Menger's curvature, which applies in a very broad context.

$\endgroup$
4
  • $\begingroup$ I read Villani's topics, and parts of its "Old and New", and Lott's paper. What I don't understand here is specific to matrices that is not addressed in the refs above: what parameterization is used to represent the matrices ? What is the basis to represent the tangent space (ie., symm. matrices) ? To compute the Christoffel symbols of the Levi-Civita connection, I need to be able to differentiate the metric along the vectors of a local basis... but which one ? Can I just take, for example for a 2x2 matrix, the basis consisting of the 3 matrices : e0=[1 0;0 0], e1=[0 1; 1 0], e2=[0 0; 0 1] ? $\endgroup$
    – WhitAngl
    Aug 18, 2012 at 19:33
  • $\begingroup$ I do not know the details of Takatsu's paper, I think you should ask a more precise question about this parametrization. Are you familiar with Lie groups? If not, here is probably the problem. Then a good read is the chapter 0 of Knapp's "Lie groups, beyond an introduction". $\endgroup$ Aug 18, 2012 at 20:57
  • $\begingroup$ I am not very comfortable with Lie groups, so I'll check this book. Thanks! :) $\endgroup$
    – WhitAngl
    Aug 19, 2012 at 6:20
  • 1
    $\begingroup$ The link in the post seems to be dead - but if it's A user’s guide to optimal transport by L. Ambrosio, N. Gigli, it seems to be available on various websites $\endgroup$ Apr 6, 2022 at 14:56

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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