The inverse of the covariance matrix for a multivariate normal random vector can be thought of as holding some measure of conditional dependence between two of the variables in the vector.

Consider we have an inverse covariance matrix $\Sigma^{-1}$ and we want to build approximations to the actual covariance $\Sigma$ based on different length conditional paths between random variables in our random vector. My idea was to write the inverse as $D - X$ giving the following,

$$ \Sigma = (D - X)^{-1} = D^{-1/2}(I - D^{-1/2}XD^{-1/2})^{-1}D^{-1/2}$$

where $D$ is a diagonal matrix and $X$ is a matrix with 0s on the diagonal. Hence we have

$$ D^{1/2} \Sigma D^{1/2} = (I - D^{-1/2}XD^{-1/2})^{-1} = \sum_{k=0}^{\infty} (D^{-1/2}XD^{-1/2})^k $$

by infinite expansion of the series. Now it appeared to me that $(D^{-1/2}XD^{-1/2})^k$ would be some measure of the covariance along paths of size length $k$.

This seemed to intuitively make sense to me, however it seems that $D^{-1/2}XD^{-1/2}$ can have eigenvalues with absolute value greater than 1, hence the series does not converge!

Am I missing something or is my intuition incorrect that $\Sigma$ can be built from $\Sigma^{-1}$ by considering the different length paths between the conditional dependent varaibles.


I realise by symmetry the same argument can be made for building $\Sigma^{-1}$ from $\Sigma$ but this doesn't have the same graphical (conditional dependence) intuition to me.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.