2 ms·
When I woke up and thought about this problem again, I realized that I didn't explain one detail. Why should (pi(n) - li(n))/sqrt(li(n)) follow a standard norm
by btilly 16y ago
When I woke up and thought about this problem again, I realized that I didn't explain one detail. Why should (pi(n) - li(n))/sqrt(li(n)) follow a standard normal distribution?
Here is why. pi(n) is the sum of a bunch of many independent random variables, each of which looks like a flip of a biased coin. If you pick the right version of the central limit theorem (see http://en.wikipedia.org/wiki/Lyapunov_condition http://en.wikipedia.org/wiki/Lyapunov_condition for the right version), it is clearly going to be approximately normal.
Its average will be the sum of the averages, which is 1/log(2) + 1/(log(3) + ... + 1/log(n) which is going to be li(n) plus O(1). If n is large the li(n) term dominates.
Its variance will be the sum of the variances of the individual random variables. The variance of a coin with probability x of turning up 1 is x(1-x). Therefore the sum of the variances is (1/log(2) - (1/log(2)2) + (1/log(3) - (1/log(3)2) + ... + (1/log(n) - (1/log(n)2) which is (modulo a O(1) error) the integral from 2 to n of 1/log(x) - (1/log(x))2 which is li(x) - O(n/log(n)2), which is dominated by the li(x) term.
Therefore the average and variance of pi(n) are approximately li(x). So (pi(n) - li(n))/sqrt(li(n)) should have approximately a standard normal distribution.