Does emcee reuse likelihood values for a repeated parameters vector?

This is not an issue, rather a question.

My likelihood is particularly costly to obtain. Since the sampler necessarily visits the same point in parameter space more than once it would be very convenient if the sampler, instead of processing the full likelihood every time, took the value from the stored array of values.

Does emcee do this already? If not, is there a specific reason for not doing it, or it just isn’t implemented?

[dfm]

Andy is right here. The probability of ever visiting exactly the same coordinates is vanishingly small and when a value is repeated in the chain (e.g. if the proposal is rejected) the log prob function is not called again. Adding caching is unlikely to help performance!

[andyfaff]

In fact it may be something you can implement yourself without having to do it in emcee:

https://pythonhosted.org/joblib/memory.html

You could use the memory.cache decorator on your log-likelihood function.