#18 Numerical instabilities in the evaluation of the restricted likelihood

[Julien Bect]

There are numerical instabilities in the evaluation of the restricted likelihood for the attached dataset in a certain region of the parameter space (see figure).

The attached script allows to reproduce the issue (figure produced on default/2142f54192a5).

These instabilities can create difficulties in the parameter optimization process. They should be diagnosed and, if possible, reduced.

[Julien Bect]

The issue seems to be that stk_cholcov doesn't add enough "regularisation noise" to stabilize this computation.

In this case, a noise variance of about var(zi)/1e9 seems to be appropriate (see attached files), but determining such a noise level automatically is a delicate matter that should not delay the 2.2 release => I will assign this ticket to 2.3.

[Julien Bect]

The issue is still here. Won't be fixed in 2.4.

[Julien Bect]

The issue is still here. Won't be fixed in 2.5.

[Julien Bect]

Moved to github: https://github.com/stk-kriging/stk/issues/6. Closing this ticket.