numpy.arange miscounts when step is large relative to value past stop

[kawochen]

 In [14]: len(numpy.arange(0, 10000000000000001, 100000000000000))
Out[14]: 101

In [15]: len(numpy.arange(0, 100000000000000000001, 1000000000000000000))
Out[15]: 100

[jaimefrio]

To calculate the number of items, a rounded up division of (stop - start) / step is performed. This is done in floating point, and the loss of precision is what caused your missing last entry. If all three values are integers, we could alternatively take advantage of Python's arbitrary precision integers and use the (stop - start - 1) // step + 1 rounded up integer division formula, which should fix your problem.

[kawochen]

Thanks. I suspected it was some kind of floating point arithmetic at work as well, but couldn't find the source code for that function. Can you point me to it? The doc says when all three are integers it does the same thing as the built-in range, but in this case it's different.

[jaimefrio]

It's written in C. The function that ultimately creates the arange array is PyArray_ArangeObj, but the length is calculated by the auxiliary function _calc_length called by it.

[julianeagu]

I am getting a (similar?) bug:

For example, I get an extra number in this case:

x = np.arange(0.75,0.8,0.01) 
print x
[ 0.75  0.76  0.77  0.78  0.79  0.8 ]

but this works fine:

x = np.arange(0.75,1.0,0.01)
print x
[ 0.75  0.76  0.77  0.78  0.79  0.8   0.81  0.82  0.83  0.84  0.85  0.86 0.87  
0.88  0.89  0.9   0.91  0.92  0.93  0.94  0.95  0.96  0.97  0.98 0.99]

I'm using numpy version 1.11.2.

[charris]

The problem is that .8 and .01 are not exactly representable in floating point, so the actual numbers used are not what they look like. For instance

 In [2]: print "%25.20f" % .8
   0.80000000000000004441

The way to do this sort of thing is to use linspace or integers and divide

In [1]: linspace(.75, .8, 5, endpoint=False)
Out[1]: array([ 0.75,  0.76,  0.77,  0.78,  0.79])

In [2]: arange(75, 80)/100.
Out[2]: array([ 0.75,  0.76,  0.77,  0.78,  0.79])