Discretised prior in Dynesty

[FengqiuAdamDong]

Dynesty version Specify the exact version, and how you installed it (pip/git) 2.1.1 pip Your question Hello, my question centers around discretised prior distributions. In my case, I have a binomial likelihood and I want to sample over N, the number of trials. Is there any way to do this in dynesty?

[segasai]

Yes, I would say so.

At least I did it by basically assuming a continuous parameter with some prior between 0..K which you can then truncate to an integer. I.e. this is implemented here https://github.com/segasai/mdf_modeling_paper (for this paper https://ui.adsabs.harvard.edu/abs/2023MNRAS.520.6091D/abstract )

At least for me it worked fine, but I wouldn't rule out a possibility of some convergence difficulties, depending on the problem.

[segasai]

Specificaly for my case I basically used the prior transform like this
n = $10^{-1 + (MAXLOGN+ 1) * x)}$ where x is the U(0,1) parameter and then truncating n to an integer in the likelihood

[FengqiuAdamDong]

Thank you, I read the paper, and it seems like you added a small level of deterministic noise to the likelihoods. But in the footnote, you say that this is not needed anymore in the new dynesty version. So if I were to recreate what you did with the truncation, I no longer need to add the noise, is that correct?

[segasai]

Yes the 2.1 version does not require any artificial noise as it supports likelihoods with plateaus.

[FengqiuAdamDong]

I implemented a discrete prior transform, specifically this one https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.randint.html And everything seems to work. Is there any reason to prefer one approach over another?

[segasai]

If the PPF of the distribution maps U(0,1) into discrete numbers, that is probably fine as well. I personally wasn't sure if the discreteness after prior transform is an issue or not (maybe not)

[segasai]

I assume this is solved. Feel free to reopen.