Chad Scherrer
Chad Scherrer
The simplex I have in mind is https://en.wikipedia.org/wiki/Simplex
A few nice things about that definition: - It's more symmetric, in that all edges have the same length. If you include the origin, edges can either be length one...
By "here" do you mean in the current implementation? I think this is very easy for the definition I'm used to
Yes, rounding errors are unavoidable if you try to do `in` directly on embeddings. Have you ever done any work with the Stan language? It focuses on Hamiltonian Monte Carlo...
Update: I just realized this is definitely wrong, hopefully something along these lines might still be possible :)
Thanks for the link :) I guess it's a matter of finding a good RNG that only uses reversible operations, is that right? So xor is ok, but no bit...
Definitely the latter. The idea here is based on rejection sampling. Suppose you have a prior, and some observations drawn from a conditional distribution based on that prior. Now imagine...
Getting PPL to work in this way could be challenging, but having a reversible RNG at all will make it now easy to try out some different ideas. We can...
Sounds good, I'll wait for your go-ahead to give it another shot. Thanks.
Very nice!! I'll find a couple of examples to try out. @mschauer this might also be useful for you, if your PDMP sampler needs a warm-up