Open
Description
What's the new feature?
The Kernel Thinning algorithm (https://arxiv.org/pdf/2105.05842) recommends using square root kernels. #867 added a square root kernel to SquaredExponentialKernel
, but adding this to other implemented kernels can require a lot of effort.
What value does this add?
According the Kernel Thinning paper, using square root kernels in the KT-split stage is optimal theoretically. It is difficult to quantify how much this really matters in practice.
Is there an alternative you've considered?
No response
Additional context
No response