The mechanics behind exponent bias in floating point