0

Description of the issue:

Here is a (I hope reproducible) issue I encountered with a simple Python 3.6.7 function (on Ubuntu 18.04 4.15.0-43-generic x86_64 GNU/Linux):

def myfunc(x):
    while x < 0.2:
        y = x
        print("x: %.2f, y: %.2f: " % (x,y))
        print("Test: {}".format((x>=0 and x<=1 and y>=0 and y<=1)))
        x += 0.1

Then, when starting from -0.4: 

myfunc(-0.4)

it returns:

x: -0.40, y: -0.40: 
Test: False
x: -0.30, y: -0.30: 
Test: False
x: -0.20, y: -0.20: 
Test: False
x: -0.10, y: -0.10: 
Test: False
x: -0.00, y: -0.00: 
Test: False
x: 0.10, y: 0.10: 
Test: True
x: 0.20, y: 0.20: 
Test: True

And when starting with -0.2: 

myfunc(-0.2)

it returns:

x: -0.20, y: -0.20: 
Test: False
x: -0.10, y: -0.10: 
Test: False
x: 0.00, y: 0.00: 
Test: True
x: 0.10, y: 0.10: 
Test: True

The behavior at 0.00 is not the same (False when starting with x=-0.4, True when starting with x=-0.2).

This is obscure to me.
Do you know what phenomena happens here?

swiss_knight
  • 5,787
  • 8
  • 50
  • 92
  • 1
    The problem is that neither 0.2 nor 0.1 can be exactly stored in memory as a float, they are approximated, and as you add them the tiny errors add up. There are many Stack Overflow questions on similar problems. – RemcoGerlich Jan 25 '19 at 09:21
  • 1
    Check this https://stackoverflow.com/questions/7545015/can-someone-explain-this-0-2-0-1-0-30000000000000004 – Chien Nguyen Jan 25 '19 at 09:23
  • 1
    If you want to compare floating point numbers, you should check out numpy. i.e. 'numpy.isclose(.1+.2,.3)'. – fabianegli Jan 25 '19 at 10:19
  • Thanks for pointing me out to the other Q/A. I then solved my issue by adding or subtracting a small epsilon value to my tests. :) – swiss_knight Jan 25 '19 at 13:26

0 Answers0