Why do floating point numeric equality/inequality tests fail mysteriously?

Rule #1 of floating-point computing in any language, not just Perl: Never compare two floating-point numbers with == or != (or expect >= et al. to work in the edge cases). Use a tolerance range instead.

The reason for this is that floating-point numbers are represented internally as a binary approximation. A consequence of that is that calculations that would be exact in decimal notation are not exact in binary. For example, a simple test like 36.6 + 0.2 == 36.8 may be return false! The same test could be done safely like this:

$TOL = 1E-8; # or a sufficiently small number
if (abs((36.6 + 0.2) - 38.2) < $TOL) {
    print "They are (almost) equal!\n";
}
[download]

This topic is covered in greater depth in the Q&A QandASection: math section, under the title of : My floating point comparison does not work. Why ?. Please post any further answers there.

Back to Meditations