I calculate a threshold T (a real number) for which p * n/2 numbers (0 < p < 1) are expected to fall below T.
Then I count how many numbers are less than T.
That count should have a binomial distribution, right? Hence, its variance should be n/2 * p * (1-p).
But the n/2 numbers are the modulus of the first n/2 complex numbers obtained from a discrete Fourier transform and the variance should be n/2 * p * (1-p) / 2 as explained here (page 10): http://eprint.iacr.org/2004/018.pdf
Unfortunately, the value given in the paper is significantly different from the one obtained by extensive simulations.
Does anybody now a procedure or a formula to calculate the exact variance of the counts?