Can somebody bring a light why pig tails occur for nicely mixing chains and is there a way to link the occurrence of pig tails to some event (such as NAN in normal_log or divergence or …). I am using cmdstan 2.14.0. There are four chains and one of them mixes nicely and then suddenly a pigtail appears. Remaining three mix nicely for 2000 iter.
The geometry in some direction is constraining movement to be below the current step size tolerances. This can happen in the neck of the funnel in hierarchical models with centered parameterizations.
It can lead to biases and you want to reparameterize.
Michael Betancourt has a case study on divergences you should check out (web site >> users >> documentation >> case studies).