Bug in y-minimum for weighted, log, stepped `Axes.hist`

[duncanmmacleod]

I think there is a bug in Axes.hist for the specific combination of log=True, histtype='step{filled}', weights=[something small], where the y-minimum is set to 1/base despite the possibility of weights << 1. Please see the following example:

from numpy import (ones_like, random)
from matplotlib import pyplot
data = random.normal(size=10000)
fig = pyplot.figure()
ax = fig.gca()
ax.hist(data, log=True, histtype='step', weights=ones_like(data) * 1e-5)
fig.show()

with log=False, the figure looks normal, while with log=True the patch minimum is set to 0.1 (> 1e-5) meaning the patches are all 1-D lines at 0.1.

I believe this patch should fix the problem.

[tacaswell]

What version did you test this on?

[duncanmmacleod]

1.3.1 (macports) and 1.4.x (HEAD of master).

[tacaswell]

@duncanmmacleod Could you put that patch in a PR please?

@neggert Do you have time to look at this (sorry to keep pestering you, but you are in my brain as the current histogram expert)?

[neggert]

Yup, looks like that patch should do the trick. Just need a PR for it. If you could modify one of the existing histogram tests to pick this up, that would be good.

@tacaswell No problem.

[tacaswell]

closed by #3089

[pelson]

Do you have time to look at this (sorry to keep pestering you, but you are in my brain as the current histogram expert)

P.S. That is a good thing @tacaswell - @neggert has done some great work on this aspect of mpl and is rightfully the "mpl histogram expert". 👍 😄