3
$\begingroup$

Let $X \sim \text{N}(0, \sigma_X^2)$ and $Y \sim \text{N}(0, \sigma_Y^2)$ be independent normal random variables with zero mean, but (possibly) different variances. Given some constants $a$ and $b$, I would like to obtain an expression for:

$$\mathbb{E}(\exp( ab X Y)).$$

I am able to solve the univariate case where I only have one of these random variables in the exponential, but I cannot solve this bivariate case.

$\endgroup$
1
  • $\begingroup$ It's not sufficient to know the marginal distributions; you need information about their joint distribution. Are they jointly normal? Are they independent? $\endgroup$
    – Glen_b
    Nov 28 '19 at 23:30
3
$\begingroup$

Without loss of generality, the factor $ab$ can be omitted as including this factor has the same effect as changing $\sigma_X^2\sigma_Y^2$ by a factor of $a^2b^2$.

Using the law of total expectation, the definition of the moment generating function, the mgfs of the normal and chi-square distributions, \begin{align} Ee^{XY} &=E(Ee^{XY}|Y) \\&=EM_X(Y) \\&=Ee^{\frac12\sigma_X^2Y^2} \\&=Ee^{\frac12\sigma_X^2\sigma_Y^2\chi_1^2} \\&=M_{\chi_1^2}(\frac12\sigma_X^2\sigma_Y^2) \\&=(1 - \sigma_X^2\sigma_Y^2)^{-1/2}. \end{align} This only works for $\sigma_X\sigma_Y<1$ which suggest that the expectation is perhaps not finite when $\sigma_X\sigma_Y\ge 1$.

Verifying the result against simulations:

> X <- rnorm(1e+6,sd=.9)
> Y <- rnorm(1e+6,sd=.7)
> mean(exp(X*Y))
[1] 1.284138
> (1-.9^2*.7^2)^(-.5)
[1] 1.287672
$\endgroup$
5
  • $\begingroup$ so if I use the parameters a,b it should be like this?: =(1−a^2*b^2*σX^2*σY^2)^(−1/2) where the parameters are squared (sorry I don't know how to edit equation yet) $\endgroup$ Nov 29 '19 at 3:26
  • $\begingroup$ @Miguel Yes, that's right. $\endgroup$ Nov 29 '19 at 11:29
  • $\begingroup$ one more question. The result above change if some of the parameters is negative? example: E(exp(-abXY) or does not matter because a is squared in the result? $\endgroup$ Nov 29 '19 at 16:43
  • $\begingroup$ sorry for ask again. @Jarlee, do yo know how change the result above when the problem takes the form: E(exp(X/Y)), when X is divided by Y? Thanks very much $\endgroup$ Dec 3 '19 at 2:48
  • $\begingroup$ @MiguelAlegre Neither $X/Y$ nor $\exp(X/Y)$ will have finite expectations, see stats.stackexchange.com/questions/299722/…. $\endgroup$ Dec 3 '19 at 8:55

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.