Check for bugs in relativistic MCMC

pints-team / pints

Probabilistic Inference on Noisy Time Series

http://pints.readthedocs.io

Other

228 stars 33 forks source link

Check for bugs in relativistic MCMC #1348

Closed rccreswell closed 3 years ago

rccreswell commented 3 years ago

While writing a functional testing notebook for this sampler, I noticed very poor performance on the correlated Gaussian even with a large number of iterations, indicating a possible bug.

[x] Check equations for relativistic mass and kinetic energy.
[x] Check the distribution for momentum. In the paper they show that it has a hyperbolic distribution, while we are using the multivariate normal from standard Hamiltonian.

ben18785 commented 3 years ago

Thanks @rcw5890 -- that looks like a good spot with the proposal distribution, I must have missed it.