Note, I’m pretty sure Stan would do this sample exactly, and in way less time. Stan is awesome. Most likely it wouldn’t diverge even in the corner region of the tail. I’ll run it and see what happens and post the picture and timing.
Is there even a case for my algorithm given Stan exists? It’s actually quite questionable, the main issue is whether the fuzzy nature of the diffusion HMC de-constrains the system enough to get through divergence inducing regions better, at the expense of giving you a sample from a not-quite-right but not too bad distribution which you could potentially then equilibriate down to correct.
Also, it’s pretty cool that you can get something HMC like without needing exact gradients by embracing noise. Anyway, I learned some Julia in the process ;-)