Document behavior of min/max/nanmin/nanmax for empty inputs #3287
Conversation
There was a problem hiding this comment.
Scipp returns DBL_MAX for empty inputs, while NumPy returns NaN. For integer inputs, Scipp returns INT_MAX, while NumPy raises.
It is not really correct to say that Numpy returns Nan when the input is empty, or at least it is a bit unclear. Both of the below examples raises ValueError: zero-size array to reduction operation minimum which has no identity independent of the data-type.
np.min(np.ones(0, dtype=np.int64))
np.min(np.ones(0, dtype=np.float64))However, Numpy.min does returns nan if the input contains nan.
As a side note that probably should be the topic of another issue: When looking into this I noticed that sc.nanmin and sc.min does the same thing
# Both return -1
sc.array(dims=['x'], values=[1, float('nan'), -1]).min()
sc.array(dims=['x'], values=[1, float('nan'), -1]).nanmin()while
# Only nanmin returns -1
np.min([1, float('nan'), -1])
np.nanmin([1, float('nan'), -1])I think the reason is that in the implementation of sc.min we use std::min(a, b) equivalent to (b < a) ? b : a; and if b=nan we get a.
|
Hmm, and from import numpy as np
import numpy.ma as ma
ma.masked_array(np.ones(1), mask=[True]).min() # masked |
src/scipp/core/reduction.py
Outdated
| inputs, Scipp returns INT_MAX, while NumPy raises. Note that in the case of | ||
| :py:class:`DataArray`, inputs can also be "empty" if all elements contributing | ||
| to an output element are masked. | ||
| Scipp returns DBL_MAX and INT_MAX for empty inputs of float or int dtype, |
There was a problem hiding this comment.
| Scipp returns DBL_MAX and INT_MAX for empty inputs of float or int dtype, | |
| Scipp returns DBL_MAX or INT_MAX for empty inputs of float or int dtype, |
There was a problem hiding this comment.
And in the other docstrings, too.
src/scipp/core/reduction.py
Outdated
| :py:class:`DataArray`, inputs can also be "empty" if all elements contributing | ||
| to an output element are masked. | ||
| Scipp returns DBL_MAX and INT_MAX for empty inputs of float or int dtype, | ||
| respectively, while NumPy rases. Note that in the case of :py:class:`DataArray`, |
There was a problem hiding this comment.
| respectively, while NumPy rases. Note that in the case of :py:class:`DataArray`, | |
| respectively, while NumPy raises. Note that in the case of :py:class:`DataArray`, |
Fixes #3241.