numpy-discussion

[Numpy-discussion] Puzzling numpy results?

[Andrew M. <an...@ob...>]

The behavior I'm seeing with zero length Numeric arrays is not what I
would have expected:

    >>> from Numeric import *
    >>> array([5]) != array([])
    zeros((0,), 'l')
    >>> array([]) == array([])
    zeros((0,), 'l')
    >>> allclose(array([5]), array([]))
    1

This is with Numeric-20.3 (and Numeric-20.2.1) - is this behavior correct,
or have I stumbled across a bug?

If both sides of the comparison are arrays with a length greater than
zero, the comparisons work as expected:

    >>> array([5]) != array([6])
    array([1])
    >>> array([5, 5]) != array([6])
    array([1, 1])
    >>> array([5]) != array([5])
    array([0])

The problem came up when I was writing unittests for some Numpy code:
under some circumstances, the code under test is expected to return a
zero length array: I was somewhat surprised when I couldn't make the
test fail! 8-)

-- 
Andrew McNamara, Senior Developer, Object Craft
http://www.object-craft.com.au/

Re: [Numpy-discussion] Puzzling numpy results?

[Tim C. <tc...@op...>]

Andrew McNamara wrote:
> 
> The behavior I'm seeing with zero length Numeric arrays is not what I
> would have expected:
> 
>     >>> from Numeric import *
>     >>> array([5]) != array([])
>     zeros((0,), 'l')
>     >>> array([]) == array([])
>     zeros((0,), 'l')
>     >>> allclose(array([5]), array([]))
>     1

The Numpy docs point out that == and != are implemented via the logical
ufuncs, and that:

 "The ``logical'' ufuncs also perform their operations on arrays in
elementwise fashion, just like the ``mathematical'' ones."

I think this explains the results you are seeing: if you do an
element-wise comparison of a length-one array with a zero-length array,
the Numpy
recycling rule means that you should always get a zero-length result.
Note that zeros((0,),'l') is not zero, it is zero zeros. So although the
results are surprising (at least to me, and you), I think the observed
results are logically correct, although surprising.

But, if that is the case, why does this hold (which I suspect reflects
what you originally expected)?:

>>> from Numeric import *
>>> array([5,6]) != array([])
1
>>> array([5,6]) == array([])
0

Tim C