Strange behaviour for emcee.autocorr.integratedtime

[Dr Hogarth]

Hi,

I'm running a custom mcmc modelling code and using emcee.autocorr.integratedtime as a way to test whether the parameters in the models have successfully converged (i.e. models with chains with integratedtime * 50/(length of chain)>1 are discarded).

However, I'm finding that it's often throwing out what look like good models in favour of models that look less well converged. I've attached two corner plots from two models fit to the same data. The two_exp model fails but the exp_gauss doesn't, which seems wrong.

Thanks,

Lucy

[Dr Hogarth]

Here are the chains for the two models. The auto correlation time finds that the rturn and vmax parameters fail for the two_exp model,  but says that all the chains for the exp_gauss model have converged successfully.

Lucy

[Dan Foreman-Mackey]

Hi Lucy,

I'll need more details, so I'll email you offlist to try to diagnose the situation.

Best,

Dan

--
You received this message because you are subscribed to the Google Groups "emcee users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to emcee-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/emcee-users/b88897c8-b5b4-4e19-946a-af0afcef52a1n%40googlegroups.com.

--

Dan Foreman-Mackey

Associate Research Scientist

Flatiron Institute

https://dfm.io