Closed ntoong closed 2 months ago
I think I found it after doing some tracing - the lengthscales are applied in the "_WeightedSqEuclidean_ijk" function in kernels/distance.jl. It takes in the inverse lengthscales squared from the kernel as the "weights" argument and uses it to scale the Euclidean distance between points.
I noticed in the isotropic Matern kernels, the lengthscale is used to normalize r in the cov functions. However, this is not the case in the anisotropic versions. Can someone point me to where the lengthscale is applied to normalize r in the anisotropic Matern kernels? Thank you!