My view is that we will not add more non-uniform distribution (i.e. "named" statistical probability distributions like Polya-Gamma) methods to `Generator`. I think that we might add a couple more methods to handle some more fundamental issues (like sampling from the unit interval with control over whether each boundary is open or closed, maybe one more variation on shuffling) that helps write randomized algorithms. Now that we have the C and Cython APIs which allow one to implement non-uniform distributions in other packages, we strongly encourage that.

As I commented on the linked PR, `scipy.stats` would be a reasonable place for a Polya-Gamma sampling function, even if it's not feasible to implement an `rv_continuous` class for it. You have convinced me that the nature of the Polya-Gamma distribution warrants this. The only issue is that scipy still depends on a pre-`Generator` version of numpy. So I recommend implementing this function in your own package with an eye towards contributing it to scipy later.

On Sun, Dec 27, 2020 at 6:05 AM Zolisa Bleki <BLKZOL001@myuct.ac.za> wrote:

Hi All,

I would like to know if Numpy accepts addition of new distributions since the implementation of the Generator interface. If so, what is the criteria for a particular distribution to be accepted? The reason why i'm asking is because I would like to propose adding the Polya-gamma distribution to numpy, for the following reasons:

1) Polya-gamma random variables are commonly used as auxiliary variables during data augmentation in Bayesian sampling algorithms, which have wide-spread usage in Statistics and recently, Machine learning.

2) Since this distribution is mostly useful for random sampling, it since appropriate to have it in numpy and not projects like scipy [1].

3) The only python/C++ implementation of the sampler available is licensed under GPLv3 which I believe limits copying into packages that choose to use a different license [2].

4) Numpy's random API makes adding the distribution painless.

I have done preliminary work on this by implementing the distribution sampler as decribed in [3]; see: https://github.com/numpy/numpy/compare/master...zoj613:polyagamma .

There is a more efficient sampling algorithm described in a later paper [4], but I chose not to start with that one unless I know it is worth investing time in.

I would appreciate your thoughts on this proposal.

Regards,

Zolisa

Refs:

[1] https://github.com/scipy/scipy/issues/11009

[2] https://github.com/slinderman/pypolyagamma

[3] https://arxiv.org/pdf/1205.0310v1.pdf

[4] https://arxiv.org/pdf/1405.0506.pdf

Disclaimer - University of Cape Town This email is subject to UCT policies and email disclaimer published on our website at http://www.uct.ac.za/main/email-disclaimer or obtainable from +27 21 650 9111. If this email is not related to the business of UCT, it is sent by the sender in an individual capacity. Please report security incidents or abuse via https://csirt.uct.ac.za/page/report-an-incident.php.
_______________________________________________
NumPy-Discussion mailing list
NumPy-Discussion@python.org
https://mail.python.org/mailman/listinfo/numpy-discussion