I have two 2-dimensional Gaussian distributions:
$$ D_1 := \mu_1=\pmatrix{x \cr y}, \quad \Sigma = \pmatrix{{\rm var}(x) &{\rm cov}(xy) \cr {\rm cov}(yx) &{\rm var}(y)} \\ D_2 := \mu_2=\pmatrix{x \cr y}, \quad \Sigma = \pmatrix{{\rm var}(x) &{\rm cov}(xy) \cr {\rm cov}(yx) &{\rm var}(y)} $$ and I want to characterize their 'intersection space' either analytically or with computationally generated samples. My ultimate goal is to be able to say something about the collisions of D_1 with D_2. Both are probabilistic representations of moving objects in 2-space.

One sampling idea was to sample each space (using a normal random number generator) and find the joint probability for each point. This seems to require me to pick a sensible threshold, which doesn't seem like the right direction.

I came across this CV thread, which yields a percentage. Can this be extended to 2 dimensions?

Any help would be appreciated by way of solutions, articles or avenues.

  • Is your question to find the proportion of area in overlap (as with the link you give), or to find out something else? – Glen_b Feb 8 '14 at 23:44
  • @Glen_b It looks like something like the Bhattacharyya coefficient gives the proportion of overlap with not too much difficulty in 1-d. I am wondering 1) how to compute this for bivariate gaussians and/or 2) some other characterization of the space other than a proportion. ie I would be willing to normalize the overlap space and I would like to know mean and variance of that new gaussian. – djh Feb 9 '14 at 1:10
  • Now I am very confused. What new Gaussian? – Glen_b Feb 9 '14 at 2:25

Your Answer

 
discard

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.