Quote:
are you suggesting I should assume each X_j has it's own probability distribution?
The sequence X_1, ..., X_n, ... is an iid sequence. So each X_j has the same distribution.
Quote:
So then (r - s^4(n - 3)/(n - 1))/n would be relatively straightforward to calculate yes?
Not really, because r depends on the (common) distribution of the X_j's. For example, if they were normal, then r = 3s^4. If they are not normal, then r could be anything. Also, you do not know s. You only know S. I am assuming the whole point of looking at the variance of the sample variance is to get a confidence interval for s. In other words, you want to say something like
P(s in [..., ...]) = 0.95,
where the interval [..., ...] contains expressions involving S. That way, you take what you calculate (S), build an interval, and then claim that the thing you cannot see (s) is likely to be in that interval. If you are going to follow this procedure, then you have to keep s and S separate. You cannot carelessly swap one for the other.
Here is an example. Suppose the X_j's are all normal. Then r = 3s^4. If n is large, then
(r - s^4(n - 3)/(n - 1))/n ~= 2s^4/n.
Hence, S^2 is approximately normal with mean s^2 and variance 2s^4/n. Constructing a 2-standard deviation interval gives
s^2 - 2\sqrt{2/n}s^2 < S^2 < s^2 + 2\sqrt{2/n}s^2,
(1 - 2\sqrt{2/n})s^2 < S^2 < (1 + 2\sqrt{2/n})s^2.
We can rewrite this as
S^2/(1 + 2\sqrt{2/n}) < s^2 < S^2/(1 - 2\sqrt{2/n}).
This gives a 95% confidence interval for s^2 that looks like:
[S^2/(1 + 2\sqrt{2/n}), S^2/(1 - 2\sqrt{2/n})].