Correlation parameter / length scale transformation from the auxiliary space

Dear UQ-community,

I would like to transform the correlation parameters (myKriging.Kriging.theta) from the auxilary space U back into the original space of MetaOpts.ExpDesign.X.
Just to reassure myself after doing some quick Algebra: Would that work as follows?

theta = myKriging.Kriging.theta.*std(MetaOpts.ExpDesign.X)

Thank you and best regards


Hi @Rob,

It depends on which kernel and scaling scheme are used. If you only provide samples (no input object is defined, see Page 40 of the Kriging manual) and you do not use isotropic kernels (Page 11 of the Kriging manual), the formula you derived is correct to transform the correlation parameters from the auxiliary space to the physical space. If you used some specific settings, please provide more details and we will check them. Note that you can actually disable the scaling operation by setting Metaopts.Scaling = false.


1 Like

Thank you and right, I forgot to mention I was working with a squared exponential kernel.

I tried this already, but disabling the scaling option will still deliver presumably scaled correlation parameters for me (somewhere in the range of 1)