2
$\begingroup$

This question regards a previous post, but it is not immediately obvious the two are related, so I ask it anyways: are the eigenvalues of a Wishart matrix $\mathbf{S}$ $=$ $\frac{1}{N}\frac{\mathbf{X'}\mathbf{1}\mathbf{1}'\mathbf{X}}{N}$ independent in the finite size case (i.e., $N$ $\approx$ 3-6, quite small)?

I do realize that the covariance between elements could be computed from the joint PDF, but this seems exceedingly laborious; a reference would be time better spent, or yet, a simple argument.

The implication is that the distribution of the sum of the eigenvalues could be computed from the characteristic function method if that were true.

$\endgroup$
1
  • 2
    $\begingroup$ they are certainly not independent --- search for "eigenvalue repulsion" $\endgroup$ Oct 30, 2013 at 17:54

1 Answer 1

2
$\begingroup$

Perhaps too late for the original questioner's purposes, but in case anyone else has the same question, a couple of relevant results are:

Let $nS\sim W(n,\Sigma)$, and let $\lambda_1,\dots,\lambda_p$ the population eigenvalues.

  • In the case $\lambda_1>\dots>\lambda_p$, the ordered eigenvalues $l_1>\dots>l_p$ of $S$ satisfy $\text{Var}(l_i)=2\lambda_i^2n^{-1}+O(n^{-2})$ and $\text{Cov}(l_i,l_j) = 2\frac{\lambda_i\lambda_j}{(\lambda_i-\lambda_j)^2} n^{-2} + O(n^{-3})$.
  • In the case $\lambda_1=\dots=\lambda_p=1$, the unordered eigenvalues $l_1,\dots,l_p$ of $S$ satisfy $\text{Var}(l_i)=\frac{p+1}{n}$ and $\text{Cov}(l_i,l_j) = -\frac{1}{n}$.

So the sample eigenvalues are not independent, but their covariance decreases asymptotically as $n\rightarrow\infty$.

Source: Khatri and Srivastava, An introduction to multivariate statistics (1979), Section 9.4.

$\endgroup$
1
  • $\begingroup$ I'd like to thank you for the response! It's appreciated, and useful. Additional insight is never too late. $\endgroup$ Sep 9, 2014 at 14:56

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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