Numerical precision in summing function

Asked Oct 06 '17 at 14:45

Active Oct 06 '17 at 18:06

Viewed 38 times

When I run

import numpy as np

# Create small random numbers
a = np.random.rand(10)

# Check whether their sum and their flipped sum is the same
print(np.sum(a) == np.sum(a))
print(np.sum(np.flipud(a)) == np.sum(a))

the first statement is always true, whereas the second statement sometimes returns false, depending on the random numbers, and I have no idea why. I see this in both, Python 2.7 and Python 3.6. Interestingly, if I apply flipud twice like so:

print(np.sum(np.flipud(np.flipud(a))) == np.sum(a))

I always get True.

Edit: Some useful info on a related issue from the docs.

edited Oct 06 '17 at 18:06

asked Oct 06 '17 at 14:45

lhcgeneva

1,981
2
21
29

1

floating point operations are not amenable to exact comparison. The result can be different depending on the order of summation and the difference might be slightly larger than the machine float precision – percusse Oct 06 '17 at 15:08
Ok, what's the reason for this? – lhcgeneva Oct 06 '17 at 15:10
You can check the examples here https://scicomp.stackexchange.com/q/27975/21617 – percusse Oct 06 '17 at 15:17
OK, that makes sense, so basically it's a rounding issue? – lhcgeneva Oct 06 '17 at 15:21
Awesome, thanks for that! – lhcgeneva Oct 06 '17 at 17:55

Numerical precision in summing function

0 Answers0