Theorem

If and are floating-point numbers in a format with parameters and , and if subtraction is done with digits (i.e. one guard digit), then the relative rounding error in the result is less than .

Proof