The bandwidth is selected by a rule of thumb. It approximately minimizes the MISE of the Gaussian copula on the transformed domain. The usual normal reference matrix is multiplied by 1.25 to account for the higher variance on the copula level.
bw_t(udata)
| udata | data. |
|---|
A 2 x 2 bandwidth matrix.
The formula is $$1.25 n^{-1 / 6} \hat{\Sigma}^{1/2},$$ where \(\hat{Sigma}\) is empirical covariance matrix of the transformed random vector.
Nagler, T. (2014). Kernel Methods for Vine Copula Estimation. Master's Thesis, Technische Universitaet Muenchen, https://mediatum.ub.tum.de/node?id=1231221