speeding up thin plate splines?

[Dr. J. Z. <joh...@ze...>]

Hello,

I just had a look at the thin plate splines code. Apparently, this is
pretty slow.

1. Suggestion: I recognized calls like
    splines_kernel(sqrt(dx*dx + dy*dy))
where splines_kernel(double h) essentially returns
    h * h * log(h).

I'd suggest to rewrite this to
    splines_kernel(dx*x + dy*dy)
where splines_kernel(double h) then returns
    h * log(h) * 0.5
This saves one sqrt per call and the factor 0.5 could even be moved to outside
the sum.


2. Are there any other useful (faster) kernels? E.g. a kernel with no
expensive function like log()?

Best regards,

-- 
Dr. Johannes Zellner <joh...@ze...>