Results from GNU Octave, version 2.9.9 on Mac OS X 10.4.8, PowerPC G5:
u = eps/2 = 1.1102e-16
eps = 2.2204e-16
u is the unit roundoff of Golub & Van Loan, Matrix Computations:
1 + u == 1 is true
1 + eps > 1 is true
precision for base = 2 (Golub & Van Loan):
precision_round = 1 - log(eps)/log(2) = 53
precision_trunc = 1 - log(u)/log(2) = 54
Does this mean that my Mac uses 53 bits for the "precise" part of a floating
point number and 11 bits for its sign and exponent? I guess I am using 64-bit arithmetic.