2
$\begingroup$

I understand that for sampling from a univariate Gaussian, we can use $x = g(\epsilon) = \mu + \epsilon \sigma $ and then differentiate this transformation with respect to $\mu, \sigma$. How does this work for sampling a multivariate Gaussian with a covariance matrix $\Sigma$?

In this paper (p19) they say $g(\epsilon) = \mu + L\epsilon$, with $LL^T = \Sigma$. What is this $L$? It can't be a Cholesky factorisation, for the terms to add it must be a vector, but covariance matrices won't decompose this way.

What is the transformation for the reparameterisation trick on a multivariate Gaussian? And then what is its gradient with respect to $\Sigma$?

$\endgroup$

1 Answer 1

3
$\begingroup$

I am being stupid -- it is in fact the Cholesky, $L \epsilon$ is a vector for a vector $\epsilon$. Found my answer here in case it helps anyone else who is confused

$\endgroup$

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.