Why does math.isclose() fail to detect minor differences between very large values?

user_7605367 0
Viewed 1122

math.isclose() exists to determine if the difference between two values is within a tolerance.
As far as I understand it, the default values for this tolerance are:

  • rel_tol = 1e-09
  • abs_tol = 0.0


If I set both values to 0.0, then it should be testing if the two values are identical.

This seems to work fine for smaller values:

import math

math.isclose(1000, 1000, rel_tol=0.0, abs_tol=0.0)                       # True
math.isclose(1000, 1001, rel_tol=0.0, abs_tol=0.0)                       # False

But it fails with very large values:

import math
import sys

math.isclose(sys.maxsize, sys.maxsize,      rel_tol=0.0, abs_tol=0.0)    # True
math.isclose(sys.maxsize, sys.maxsize-1,    rel_tol=0.0, abs_tol=0.0)    # True
math.isclose(sys.maxsize, sys.maxsize-100,  rel_tol=0.0, abs_tol=0.0)    # True
math.isclose(sys.maxsize, sys.maxsize-1000, rel_tol=0.0, abs_tol=0.0)    # False

It seems that there is still a relative tolerance?


Why does this behavior occur?

The above code was run with Python 3.5.2.

UPDATE 1:

It seems that a similar behavior occurs when using very large float values:

import math
import sys

m = sys.float_info.max                                                    # type 'float'

math.isclose(m, m)                                                        # True
math.isclose(m, m-1.0)                                                    # True
math.isclose(m, m-1e100)                                                  # True
math.isclose(m, m-1e300)                                                  # False

math.isclose(m, m,       rel_tol=0.0, abs_tol=0.0)                        # True
math.isclose(m, m-1.0,   rel_tol=0.0, abs_tol=0.0)                        # True
math.isclose(m, m-1e100, rel_tol=0.0, abs_tol=0.0)                        # True
math.isclose(m, m-1e300, rel_tol=0.0, abs_tol=0.0)                        # False

On the other hand, the comparison operators do not work here either.

import math
import sys

m = sys.float_info.max                                                    # type 'float'

m == m                                                                    # True
m < m                                                                     # False
m > m                                                                     # False

m == m-1.0                                                                # True
m < m-1.0                                                                 # False
m > m-1.0                                                                 # False

m == m-1e100                                                              # True
m < m-1e100                                                               # False
m > m-1e100                                                               # False

m == m-1e300                                                              # False
m > m-1e300                                                               # True
m < m-1e300                                                               # False

UPDATE 2:

The answer to 'UPDATE 1' was given here.

precision compare python-3.x python equality
1 Answers

sys.maxsize is an integer, but math.isclose() works on floating-point values. On a 64-bit box, sys.maxsize has more bits of precision than a float can represent, so float(sys.maxsize - N) == float(sys.maxsize) for all sufficiently small postive integer N.

>>> from sys import maxsize as m
>>> m
9223372036854775807
>>> float(m)
9.223372036854776e+18
>>> float(m - 1)  # identical
9.223372036854776e+18
>>> float(m - 100) # identical 
9.223372036854776e+18
>>> float(m - 1000)  # finally subtracting enough to make a tiny difference
9.223372036854775e+18

Short course: when you want to compare integers for equality, converting to float first simply doesn't make sense.

user_27055420
Related