numpy histogram precision bug

When dealing with weights with a large dynamic range, the result is problematic. The problem can be reproduced with this code in numpy 1.9.1 and 1.9.2:

sample_size=1e5
xmin,xmax=-10.,0.
data=np.random.rand(sample_size)*(xmax-xmin)+xmin #some uniform data
weight=np.exp(-4*data) #weights

y,x=np.histogram(data,50,weights=weight)
print y[-1]
print np.sum(weight[(data>x[-2])&(data<x[-1])])

This should print two identical numbers (which is the case if, for example, we use xmin=-1.) but instead it produces:

0.0
2991.44627547

This is not a datatype issue because every variable is float64. Replacing the bin number parameter with double precision bins does not help.

[Kambrian]

Below is a quick work around:

bin_index=np.digitize(data, x)
yy=np.bincount(bin_index, weights=weight, minlength=len(bins)+1)[1:-1]
print yy[-1]

Not sure about its efficiency but hopefully not much slower. Maybe this should replace the histogram algorithm.

[seberg]

There is no issue with sample_size, it is a long standing change that numpy expects you to put integers and 1e5 is a float not an integer, so now you have to use 10**5 instead in those instances, we know it is a bit annoying, but it is identical to range(1e5), etc. in python.