Open WuShichao opened 2 days ago
Hi @WuShichao, I'm afraid I don't think lintsampler
doesn't quite do what you're looking for here. The key input for lintsampler
is the PDF (i.e., position input, density output), not PDF samples.
You're finding that KDE oversmooths: I assume you've tried a smaller bandwidth? In that case, I suppose you're working on a distribution with a large dynamic range. What can work quite well here is a variable-bandwidth KDE, or something like the Delaunay tessellation field estimator.
@aneeshnaik Thanks a lot! I will learn that Delaunay tessellation field estimator.
Sometimes, KDE tends to oversmooth the PDF, which introduces bias in the estimated PDF. Inverse-transform sampling doesn't have this issue, but it is usually implemented in 1D. This package extends it to at least 6D. I know this package requires grid/tree-like data, but I can provide the marginalized 1D posterior distribution for each parameter. Besides, will
lintsampler
keep the covariance?