Histogramdd returns negative values after bins were not checked for monotonic increase

[HagaiHargil]

A call to np.histogram(2, bins=[1, 3, 1]) will raise a ValueError: bins must increase monotonically. exception. However, arrays generated with a datatype of uint64 or np.uint64 will not be checked (correctly, at least) for monotonicity and will execute without a problem, generating a histogram with a negative value:

hist, edges = np.histogram(2, bins=np.array([1, 3, 1], dtype='uint64'))
# hist == np.array([1, -1]) 

The two problems here are obvious:

1. numpy doesn't check the monotonicity of the bins, or at least does so incorrectly.
2. A histogram with a negative value is returned.

I'm using numpy version 1.12.1 with Python 3.6.1 on Windows 64 bit.
Thanks!

[b-carter]

It seems like the unsigned arithmetic operation should be returning an unsigned result, so the problem here seems to be in the monotonicity check using np.diff.

I guess one solution would be to just convert anything unsigned to a signed type, but this might cause issues for very large unsigned integers that can't fit into a signed integer.

Maybe a better solution (I think also more efficient?) would be just doing np.all(bins[1:] > bins[:-1]) as the check.

If this looks good, I'll write a test and submit a PR.

[b-carter]

Just checking back on this. Does that seem like an acceptable solution to everyone? If so, I can create a PR. Thanks!

[HagaiHargil]

I'd be happy if this issue could be somehow resolved, the proposed solution seems pretty straight-forward IMHO.
I'd also like to point out that this bug still exists in version 1.13.

Thanks.

[seberg]

@b-carter or @HagaiHargil, we always apreciate PRs (no need to ask if it seems acceptable ;), though it does seem so, we can decide on that later). Please remember that we need tests to merge anything new, there is also a development workflow document, one thing we particularly like is our commit messages: https://docs.scipy.org/doc/numpy/dev/gitwash/development_workflow.html#writing-the-commit-message ;)

But these are all details we can hash out once the PR is there, so please just go ahead!

[eric-wieser]

This is fixed for np.histogram in #9294, but is still a problem for histogramdd

[eric-wieser]

Actually, it seems that histogramdd was always ok:

>>> hist, edges = np.histogramdd([[2]], bins=[np.array([1, 3, 1], dtype='uint64')])
ValueError: Found bin edge of size <= 0. Did you specify `bins` withnon-monotonic sequence?