You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
composed with LowerCholeskyTransform will create PositiveDefiniteTransform, which transforms a real vector to a positive definite matrix (though I am not interested in Wishart/Inverse-Wishart distributions, this transform will be helpful for other contributors who are interested in implementing those distributions)
composed with CorrCholeskyTransform will create CorrelationTransform, which transforms a real vector to a correlation matrix.
Then we can define LKJ distribution as TransformedDistribution(LKJCholesky, GramianTransform), so that users don't have to work with the scale tril version of a correlation matrix. It is better to for users see directly correlations of variables, rather than getting Cholesky and manually apply Grammian operator.
Inference will happen in LKJCholesky support when the approach to transformed distribution in #241 is available. Then it would be fast and not suffered from Cholesky numerical issue because no Cholesky operator is involved when computing log_prob terms.
The text was updated successfully, but these errors were encountered:
This is a transform from A -> A @ A.T, which if
Then we can define LKJ distribution as TransformedDistribution(LKJCholesky, GramianTransform), so that users don't have to work with the scale tril version of a correlation matrix. It is better to for users see directly correlations of variables, rather than getting Cholesky and manually apply Grammian operator.
Inference will happen in LKJCholesky support when the approach to transformed distribution in #241 is available. Then it would be fast and not suffered from Cholesky numerical issue because no Cholesky operator is involved when computing log_prob terms.
The text was updated successfully, but these errors were encountered: