Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

From Wikipedia

The Fisher information matrix is a N x N positive semidefinite symmetric matrix, defining a Riemannian metric on the N-dimensional parameter space,

But a Riemannian metric is defined to be a family of (positive definite) inner products.

So I was wondering whether a Riemannian metric is positive definite or positive semidefinite?

Thanks and regards!

share|improve this question
    
They need not refer to the same "metric." Take the Minkowski 3-space with form $x^2 + y^2 - z^2$ and the surface $z = \sqrt{1 + x^2 + y^2}.$ The form in three variables, restricted to (vectors tangent to) the indicated surface, actually is positive definite, the result being the hyperbolic plane. en.wikipedia.org/wiki/Hyperboloid_model You could be looking at an analogous situation. –  Will Jagy May 4 '13 at 22:08
add comment

2 Answers

A Riemannian metric is positive definite and produces a Riemannian manifold. If you take the metric tensor to be positive semidefinite then you get what is called a semi-Riemannian manifold or pseudo-Riemannian manifold. With regards to the Fisher information metric on a statistical manifold it is usually assumed that the metric is positive definite. I would not trust Wikipedia. The standard reference on information geometry is Amari's Methods of Information Geometry.

share|improve this answer
    
Thanks! "With regards to the Fisher information metric on a statistical manifold it is usually assumed that the metric is positive definite." Is Fisher information matrix positive definite or positive semidefinite then? –  Tim May 4 '13 at 19:35
    
Tim, I posted a second answer in response to your comment so that I can fit in what I have to say. –  mtiano May 5 '13 at 15:03
add comment

It is usually assumed to be positive definite but you can have semi positive definite if you like. This second condition will change the manifold structure as I stated above. For example on the normal distribution the metric is given by $$\begin{pmatrix}\frac{1}{\sigma^2}&0\\ 0&\frac{2}{\sigma^2}\\ \end{pmatrix}$$ Since you have the restriction $\sigma > 0$ this matrix is always positive definite. In the case of the Weibull distribution the metric is given by $$\begin{pmatrix}\frac{k^2}{\lambda^2}&\frac{\gamma-1}{\lambda}\\ \frac{\gamma-1}{\lambda} & \frac{\gamma^2 -2\gamma +\frac{\pi^2}{6}+1}{\lambda^2}\\ \end{pmatrix}$$ Here is is possible to pick $\gamma$ so that the bottom right hand corner is 0. This will make the metric tensor semi positive definite. People will usually place a resriction so that you can't choose such a $\gamma$.

share|improve this answer
    
Here is is possible to pick γ so that the bottom right hand corner is 0... Sure about that? –  Did May 5 '13 at 15:18
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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