Currently, the stick-breaking setup uses tanh(x) to map the real line to (-1, 1) in the correlation Cholesky factor.
This is not always ideal, as it is easy to get numerically near-singular matrices, as in
using TransformVariables, LinearAlgebra
t = corr_cholesky_factor(SMatrix{9,9})
C = transform(t, -10 * ones(dimension(t)))A milder mapping, such as x / hypot(1, x), may be able to mitigate this.
We should also consider reifying the mapping as an alternative to TVLogistic (suitably scaled).
Currently, the stick-breaking setup uses
tanh(x)to map the real line to(-1, 1)in the correlation Cholesky factor.This is not always ideal, as it is easy to get numerically near-singular matrices, as in
A milder mapping, such as
x / hypot(1, x), may be able to mitigate this.We should also consider reifying the mapping as an alternative to
TVLogistic(suitably scaled).