This is an effect of the limit internal precision. DOUBLEs are represented by 64 bits such that the mantissa has a precision of 53 bits. This is about 16 digits. If you add numbers of different size, small differences can dissappear. See:
a = 1.0E17 + 1 - 1.0E17
% >> 0
although 1 would be mathematically correct. But the intermediate result of the first two summands cannot be represented accurately.
Using the standard IEEE arithmetics, there must be a difference between "mathematically" and "numerically". See also: eps and FAQ 6.1.
BTW. Reading the FAQ completely is always a good idea.
