Quote:
Originally Posted by imjosh
I have Z = (X1* ----* Xn)^n, where X1 thru Xn all have mean u and sd o, all are SIID.
I need to find P(Z > a) for a given a, n, u, and o.
What should I do? I had to get the distribution of Z^n earlier, which I got was normal with mean u^n and variance (o^2 + u^2)^n - u^(2*n).
Im thinking distribution of Z is lognormal, and i think i may need to use a t-distribution similar to normal - thats where im stuck
This is messed up. The mean and variance you got u^n and (o^2 + u^2)^n - u^(2*n) are the mean and variance of X1*---*Xn, not Z^n. That's because for independent r.v.s, the expected value of a product is a product of expected values since the covariance is 0. So the mean is u^n. Then for the variance, we know that
E[(X1*---*Xn)^2] = E(X1^2) *---*E(Xn^2) = (o^2+u^2)^n
so the variance is
(o^2+u^2)^n - u^(2*n).
So I think you mean those are the mean and variance of Z, and Z = (X1*---*Xn). But this distribution is not normal. In the limit for large n, it will become lognormal. For small values of n, it is complicated and depends on n and the distribution of X. If X is normal, for n=2, the distribution is a modified Bessel function of the 2nd kind. For n=3 it is a Meijer G-function. Now assuming you have large enough n that this is lognormal, raising a lognormal variable to the nth power will produce another lognormal random variable, and this will multiply the mean by n and the variance by n^2. So whether you want P(Z<a) or P(Z^n) < a, you would use the cumulative lognormal distribution with the appropriate mean and variance. Clearly if you know the distribution for Z^n, then P(Z<a) = P(Z^n < a^n). The cummulative lognormal distribution can be expressed in terms of the error function as
1/2 + 1/2 * erf[(ln(x) - u) / sqrt(2*sigma^2)]
where u and sigma^2 are the mean and variance of ln(x) which you computed, not the original u and o^2. This is like using the cumulative normal distribution when the r.v.s are normal which is equivalent to the t-distribution when sigma is known. You only need the t-distribution when sigma is not known and you are using an estimate of sigma from your data.
Last edited by BruceZ; 04-25-2013 at 12:40 PM.