[MATH-866] New method to compute relative deviation between two floating numbers

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.1
• Fix Version/s: 3.1
• Assignee: Unassigned
• Reporter: Yannick TANGUY
• Created: 20/Sep/12 08:13
• Updated: 04/Mar/13 18:53
• Resolved: 20/Sep/12 10:10

Description

A new method to compute a relative deviation would be very useful.

In order to handle properly odd values (O+,0-,NaN,Infinity), we suggest to call first the classical equals() method (strict equality, using one ulp as a threshold) and then the relative tolerance.

Here is the code we use :

  public static boolean equalsWithRelativeTolerance(final double x, final double y, final double eps) {

        // Initialisation
        boolean isEqual = false;

        // very close (including both equals 0.0) case
        if (equals(x, y)) {
            isEqual = true;
        }
        // common case
        else {
            // Relative difference computation
            final double absoluteMax = FastMath.max(FastMath.abs(x), FastMath.abs(y));
            final double relativeDifference = FastMath.abs((x - y) / absoluteMax);

            //test
            if (relativeDifference < eps) {
                isEqual = true;
            }
        }
        return isEqual;
    }