Timeline for How to know if the derivatives exist in Hamiltonian Monte Carlo?

10 events

when toggle format	what		by	license	comment
Mar 7, 2017 at 15:03	comment	added	Astrid		That makes sense. So trouble only happens, which I think is what you were trying to explain, when there is no analytical form for the updates.
Mar 7, 2017 at 14:31	comment	added	conjectures		With HMC you'd probably need to take corrective action to handle the wishart params. NIW prior is very tractable so there are other solutions. E.g. it is conjugate, so unless there's some funky additions, you don't need to sample the posterior mean and variance there's an analytical form to them.
Mar 7, 2017 at 11:30	vote	accept	Astrid
Mar 7, 2017 at 11:30	comment	added	Astrid		Ahh I see, so if I were to a multivariate Gaussian with a NIW prior on the covariance matrix, I would basically be in trouble?
Mar 7, 2017 at 0:23	history	edited	conjectures	CC BY-SA 3.0	grammar
Mar 7, 2017 at 0:04	history	edited	conjectures	CC BY-SA 3.0	grammar
Mar 6, 2017 at 23:05	history	edited	conjectures	CC BY-SA 3.0	added 5 characters in body
Mar 6, 2017 at 23:03	comment	added	conjectures		For example, if your prior is normal with known variance, you're fine. If your prior involves an unknown variance, you may encounter problem II because variances must be non-negative. This is often worked around by using log-scale prior on the variance.
Mar 6, 2017 at 21:54	comment	added	Astrid		So if I am only operating with Gaussians, which are continuous, they always have derivatives, and HMC will thus not have any problems?
Mar 6, 2017 at 18:17	history	answered	conjectures	CC BY-SA 3.0