I am sampling covariance matrix from a Inverse Wishart distribution. In one dimensional case, after doing sufficient iterations I am taking the mode value for variance (after removing the burn-in values). How to do the same in a multivariate case?

  • $\begingroup$ 1) Do you mean Wishart or Inverse Wishart? $\endgroup$ – Simon Byrne Dec 6 '10 at 11:07
  • $\begingroup$ 2) What do you mean by "maximum occurrence"? Are you trying to find the mode (the point of highest density)? $\endgroup$ – Simon Byrne Dec 6 '10 at 11:10
  • $\begingroup$ Sorry for being vague, I have modified the question. $\endgroup$ – Saurabh Saxena Dec 7 '10 at 4:26

You shouldn't need a Gibbs sampler: the mode of an inverse-Wishart has a closed form.

Also, independent random samples from the Cholesky factor of a Wishart can be obtained from the Bartlett decomposition: as it is triangular, it can be inverted easily by forward subsitution to get the Cholesky factor of an inverse-Wishart.


If I understand your question correctly:

Covariance matrix for 1-dim case reduces to the variance. Wishart Distribution (or Inv wishart distribution depending on your formulation) is a prior of covariance matrices, which for dimensions $\geq$ 2 correspond to multivariate case.

However, I may have misunderstood you. Please correct me if that is the case.

  • $\begingroup$ You got it correctly, it is very easy to sample a variance from the distribution (by taking out the mode value). But I am not sure how to do that in multivariate case. I am using an Inverse Wishart distribution for sampling Covariance matrix. $\endgroup$ – Saurabh Saxena Dec 7 '10 at 4:31

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.