search

astro-informatics / harmonic

Machine learning assisted marginal likelihood (Bayesian evidence) estimation for Bayesian model selection
https://astro-informatics.github.io/harmonic/
GNU General Public License v3.0
57 stars 7 forks source link

Implement a method for log evidence with log space errors. #281

Open alicjapolanska opened 9 months ago

jasonmcewen commented 9 months ago

@alicjapolanska isn't this done already? Can you elaborate in the description if you meant something different to what we already have?

alicjapolanska commented 9 months ago

@jasonmcewen We have the log evidence with standard deviation here, which is not always numerically stable and inverse evidence with log space errors here but not log evidence with log space errors. It's straightforward to get it from the inverse evidence but I think it would be best to also have a method getting this directly for the user. Or perhaps we can modify the compute_ln_evidence method to give log space erros?

jasonmcewen commented 9 months ago

Thanks for the clarification @alicjapolanska . So the core functionality is there, is just a matter of having a nice wrapper function for the non-inverse values in log space. That would indeed be nice. As we discussed, we should also review the non-log computations and potentially remove them (issue #249). So we could address this issue as part of the work related to that one, or simply do this separately first.