Divergent transitions

performance

#1

Dear experts,

  1. How small should the number of divergent transitions must be ?
  2. If there are any divergent transitions, does that mean we should not trust the output?
  3. with adapt_delta =0.99, my code still gives 2 divergent transitions, does that mean there is a problem with the stan code?

Thanks


#2

Hey!

No big expert here, but I’ve seen quite a few answers by experts here. To give short and conservative answers:

  1. 0
  2. Generally, no
  3. Probably, yes

Often divergences can be taken by re-parameterizing your model. But without additional information it’s hard to give more specific advice. You might want to check the relevant chapter (on re-parameterization) in the manual.


#3

Michael’s given several talks and written this up quite clearly -

https://betanalpha.github.io/assets/case_studies/divergences_and_bias.html


#4

Thanks @mitzimorris and @Max_Mantei.
These are very useful information for me.