Implement incomplete beta function #2678
Conversation
This implementation is based on the Cephes library's implementation by Steve Moshier.
|
For some context, the implementation of all log CDF methods is in the |
| '''Evaluates the continued fraction form of the incomplete Beta function. | ||
| Derived from implementation by Ali Shoaib (https://goo.gl/HxjIJx). | ||
| def incomplete_beta_cfe(a, b, x, small): | ||
| '''Incomplete beta continued fraction expansions |
There was a problem hiding this comment.
Maybe it is better to implement cfunc with custom op, scan is slow. Is it possible to have closed form gradient?
There was a problem hiding this comment.
It only matters when there are too few steps for the continued fraction expansion to be close enough for the given float precision. Allowing for more steps is not useful.
Sorry, I'm not sure I understand what you mean by the gradient here.
In any case, the main issue with the previous (partial) implementation was that the numerical convergence was not very good.
This implementation is much better in that sense while being just as slow.
Of course it would be much faster to not do this symbolically...
|
@domenzain Can you add this to the release-notes? Would be great to get this merged. |
|
Sorry, I should not have opened so many pull requests for these. Point me where you'd like an explanation and I'll write a synthesis. |
|
Closing in favor of #2688 |
This implementation is based on the Cephes library's implementation by Steve
Moshier.
This completes the work started in #2048 and continued in #2073 . With it all continuous distributions have working implementations of log CDFs.
These can then be used to more adequately implement censored distributions as described in #1867 and #1864 .