[Python] Why does bottleneck move_mean give different answers for different size arrays but...

This may be similar to bottleneck.move_mean get different result for the same ndarray, but that 'question' was ill-posed and so attracted no answers.

I have found that for a single array, bottleneck move_mean gives different answers depending on whether an average value is extracted from a larger numpy array following averaging, or the average is taken over a subset of the same larger array. This should not be the case. The differences are relatively small but are not zero, as (I think) they should be.

Test code:

import numpy as np
import bottleneck as bn

howManyIters = 10

smallDifferences = np.zeros([howManyIters])

for whichIter in range(howManyIters):

testArray = np.random.random_sample([512, 353, 53])

smallDifferences[whichIter] = (bn.move_mean(testArray[:, :, 46:49],
window=3, axis=2)[0, 41, 2]
- bn.move_mean(testArray, window=3,
axis=2)[0, 41, 48])
print(smallDifferences)


I get answers like

[ 1.11022302e-16 4.44089210e-16 2.22044605e-16 2.77555756e-16
8.32667268e-17 -3.33066907e-16 1.11022302e-16 -5.55111512e-17
1.66533454e-16 -1.11022302e-16]


But I expect these to be 0.0. To my understanding, for a window of length 3, the bottleneck windowed averaging is trailing over the previous 2 and the current value, so the slice indexing here should (I hope) match up exactly.

Continue reading...