Euclidean Embedding of Co-occurr ence Data under Section 2, the problem formulation is given. In the paper link, the Authors find the distribution for lower embedding for the squared distances between 2 points : $d^2_{x,y} = ||\psi(x) - \xi(y)||^2$

I am having difficulty in understanding why the probability is considered exponential. In this answer What is the PDF when X and Y are non standard distribution is different. In my case, I need to find the pdf of the square root of the distances of the error points X,Y in higher embedding and they can be correlated. How do I begin? I am new to statistics and have limited knowledge in joint distribution and having a tough time in understanding these complicated joint distribution. Please help.


It seems that they choose the exponential distribution. That is, it's a modeling decision, rather than the consequence of some rule or property. However, they state in section 2 that "this reflects the intuition that closer objects should co-occur more frequently than distant objects"

Recall that the exponential distribution has very high density at low values, and the density decays exponentially. It's just a smooth and mathematically convenient way to model "small [distance] is more likely than large" which is what the authors are asserting when objects co-occur.

The rest of the section is dedicated to working around the issue that this implies: if x is unlikely, then x and y together must also be unlikely. But then x and y must also be far apart, which is not the intended model.

As far as the square root the distance goes: if distance has an exponential distribution, root distance has a Rayleigh distribution.

  • $\begingroup$ The point of section 2 is that there's more than one way to define the joint distribution. The last paragraph is a brief comparison which is a summary of the rest of the paper. Unfortunately this isn't something I know anything about -- I got my answer right out of the paper -- so I wouldn't know how to help you further. $\endgroup$ – shadowtalker Oct 26 '14 at 13:01
  • $\begingroup$ If you could go into more detail in the question about your application, maybe I could help $\endgroup$ – shadowtalker Oct 26 '14 at 13:03

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.