ENH: stats: Crystalball distribution #6798
Conversation
|
One issue with tests which fail is that higher moments do not exist for n=3. |
|
I'm working on the _munp implementation, Calculating the formulas using mathematica was easy, but there are numerical instabilities and odd definitions of the functions which are required to calculate the n-th moment (e.g. hyp2f1). So if munp is required I will need some more time (maybe I find some time next weekend to dive deeper into the mathematics behind hyp2f1). |
|
From a 10000 feet: Mathematica likes to express things in terms of hyp2f1, even if there's a simpler expression available (or at least it did in version 3 or something). Moments can be left for later I think. |
thomaskeck
force-pushed
the
Crystalball
branch
from
832ec2c
to
3a38729
Compare
|
Fixed merge conflicts. @ev-br |
scipy/stats/_continuous_distns.py
Outdated
|
|
||
| %(example)s | ||
| """ | ||
| def _pdf(self, x, alpha, n): |
There was a problem hiding this comment.
Maybe better use some other variable name, #5982
There was a problem hiding this comment.
I renamed alpha to beta and n to m.
scipy/stats/_continuous_distns.py
Outdated
| N = 1.0 / (n/alpha / (n-1) * np.exp(-alpha**2 / 2.0) + _norm_pdf_C * _norm_cdf(alpha)) | ||
| return N * _lazywhere(np.atleast_1d(x > -alpha), (x, alpha, n), | ||
| f = lambda x, alpha, n: np.exp(-x**2 / 2), | ||
| f2 = lambda x, alpha, n: (n/alpha)**n * np.exp(-alpha**2 / 2.0) * (n/alpha - alpha - x)**(-n)) |
There was a problem hiding this comment.
Line way too long, here and elsewhere.
There was a problem hiding this comment.
I refactored the lazywhere and introduced a right-hand-side (rhs) and left-hand-side (lhs)
lambda function which is passed to f and f2 of _lazywhere.
The lines are now shorter.
…nflicts with other scipy functions. Additionally, shortened some lines
|
Ping. I fixed the latest merge conflicts, are there further suggestions or comments to this PR? |
|
Pong? |
|
Hello, this is my bi-monthly reminder for this PR. |
|
Looks good to me now. Thanks, merging! |
|
Some issues remaining: gh-7687 |
The crystalball function is used in high-energy physics to parametrize data.
It is basically a Gaussian with a one-sided power-law tail.
https://en.wikipedia.org/wiki/Crystal_Ball_function
This is a split of, of another PR which contained several other distributions #6795