Time for action – comparing with assert_array_almost_equal_nulp

Let's see the assert_array_almost_equal_nulp function in action:

1. Determine the machine epsilon with the finfo function:

   eps = np.finfo(float).eps
   print "EPS", eps

   The epsilon would be:

   EPS 2.22044604925e-16

2. Compare two almost equal floats: Compare 1.0 with 1 + epsilon (eps) using the assert_almost_equal_nulp function. Do the same for 1 + 2 * epsilon (eps):

   print "1",
   np.testing.assert_array_almost_equal_nulp(1.0, 1.0 + eps)
   print "2",
   np.testing.assert_array_almost_equal_nulp(1.0, 1.0 + 2 * eps)

   The result:

   1 None
   2
   Traceback (most recent call last):
     …
   assert_array_almost_equal_nulp
   raiseAssertionError(msg)
   AssertionError: X and Y are not equal to 1 ULP (max is 2)

What just happened?

We determined the machine epsilon with the finfo function. We then compared 1.0 with 1 + epsilon (eps) with the assert_almost_equal_nulp function. This test passed, however, adding another epsilon resulted in an exception.