2
$\begingroup$

First of all, I apologize for not being rigorous, but I am not a statistitian by background.

Imagine you have $N$ i.i.d. positive random variables $X_1...X_N$ and you are trying to compute a logarithmic sum of expectation values: $$S = \ln\textrm{E}[X_1]+...+\ln\textrm{E}[X_N]$$

This is of course equal to: $$S = \ln\Big(\prod_{n=1}^{N}\textrm{E}[X_n]\Big)$$

Since the random variables are independent: $$S = \ln\Big(\textrm{E}\Big[\prod_{n=1}^{N}X_n\Big]\Big) \equiv \ln\textrm{E}[Z]$$

where $Z$ is a product of $N$ random variables.

For very large finite $N$ the distribution of $Z$ approaches the lognormal distribution and this $S$ can be estimated by:

$$S \approx \mu_{\ln Z} + \frac{1}{2}\sigma^2_{\ln Z} = N\Big(\mu_{\ln X_1} + \frac{1}{2}\sigma^2_{\ln X_1}\Big)$$

But we also know that, since the the random variables have identical probability distributions:

$$S = N\ln\textrm{E}[X_1]$$

However, since the PDF of $X$ need not be lognormal, the following is not in general true:

$$N\ln\textrm{E}[X_1] \neq N\Big(\mu_{\ln X_1} + \frac{1}{2}\sigma^2_{\ln X_1}\Big)$$

It seems that we have reached a paradox, which stems from a limiting behaviour I am most probably not handling correctly. On one hand, the higher-order cumulants in the limiting case approach zero but in the other case they always contribute the same amount to the logarithmic sum. Any suggestions where this paradox comes from?

$\endgroup$
  • $\begingroup$ I just assume that $X\sim N(\mu,\sigma)$ then $log(X)$ is not defined! you are missing an assumption in your definition ... $\endgroup$ – TPArrow Jun 13 '19 at 9:51
  • $\begingroup$ Thanks for that, I completely missed that one out, I will edit it. $\endgroup$ – miroslav351 Jun 13 '19 at 9:54
1
$\begingroup$

It appears that you are confusing the random variables $X_1,X_2,...,X_N$ with their expected values. As presently defined, the value $S$ is not a random variable at all; it is a constant (and so the notation of using an upper-case letter is not really appropriate). If we let $\mu \equiv \mathbb{E}(X_i)$ be the common mean of your IID values, then you have the constant value:

$$S = \ln \mathbb{E}(X_1) + \cdots + \ln \mathbb{E}(X_N) = N \ln \mu.$$

It follows that you also have the constant value $Z = e^S = \mu e^N$. This is a constant value, so it is not true that $Z$ is normally distributed for large $N$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.