then transformed to [0,1] using inv_logit.
Nono, I mean the domain is [0, 1] for the Wahba stuff and R for regular RBF GPs. They both can have anything for their outputs.
Maybe exact is the wrong word for a non-parametric model fit on some data that comes from a process I don’t understand :D, but I’m not having to do numerical approximations to implement it and use it on that data. I think that’s what I really meant.
For lack of better citation, it’s exact in the sense here: Fitting GP with noisy observations of time - #9 by anon75146577 . If we try maybe we can summon a @anon75146577 . I had trouble connecting all the math-bits together. Everyone approaches these things differently, but from a random variable perspective (to compliment the optimization approach of Wahba) I really liked Steven Lalley’s “Introduction to Gaussian Processes”. Check eq. 3.3 and the stuff around it for a wild time.