In differential geometry, tangent vectors are considered operators. At point p, the local tangent space is defined as $$ T_p(M)=\{X^i\partial_i|X\in R^n\} $$ This is quite easy to understand for me.

However, I study information geometry recently and get stuck with the tangent vectors defined on statistical manifold.

As http://en.wikipedia.org/wiki/Information_geometry points out, the tangent vectors defined at point $p_\xi$ are $\partial_ip_\xi$ in mixture representation. I really cannot understand why tangent spaces can be defined like this! It does not make any sense to me. And I cannot digest the explanations on wikipedia.

Can anyone help me understand it? Thanks in advance!

I can understand the previous problem now. Is there anyone who has studied information geometry before. I have a new question.

on http://en.wikipedia.org/wiki/Information_geometry $D[\partial_i\partial_j||\cdot]= D[\cdot||\partial_i\partial_j]=-D[\partial_i||\partial_j]$. I think they should all equal to 0. Here is my reason:

because $D[\partial_j||\cdot]=0$, we have $$ 0=\partial_iD[\partial_j||\cdot]=\partial_iD((\partial_j)_p||p)=\partial_i\partial_jD(p||p)=D((\partial_i\partial_j)_p||p)=D[\partial_i\partial_j||\cdot] $$

But the true result seems to support such equation: $$ \partial_iD((\partial_j)_p||p)=D((\partial_i\partial_j)_p||p)+D((\partial_i)_p||(\partial_j)_p) $$


  • 2
    $\begingroup$ The wikipedia article needs work. Perhaps the book referenced there could help. But I think it is clear that if you have a family of probability distributions, which you think of as functions, and they depend smoothly on a parameter, and you differentiate in that parameter, you get a function. A tangent vector is the derivative of a 1-parameter family of points. $\endgroup$ – Ben McKay Apr 13 '14 at 7:27
  • 1
    $\begingroup$ The book Methods of Information Geometry seems to be well written (after a glance) and looks much easier to digest than the wikipedia article. $\endgroup$ – Ben McKay Apr 13 '14 at 7:29

There are situations where you think of a function space as an infinite-dimensional manifold. Then the tangent space is also infinite-dimensional, and also naturally a function space. The tangent space to a product manifold is the direct sum of the tangent spaces, and a function space is like a (contiuum) product, so its tangent space should be a product over the same index set, i.e. consist of functions on the same domain.

  • $\begingroup$ Yeah! I get it. $\endgroup$ – user49550 Apr 16 '14 at 7:11

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.