I know that the product of two independent zero mean Normal random variables follows the normal product distribution.

I want to talk about the random variable $$z = xy$$ where $x \sim \mathcal{N}(0,1)$ and $y\sim \mathcal{N}(\mu,1)$ and they are independent.

In case of $\mu \gg 1$, simulations show that the distribution of $z$ can be approximated by $\mathcal{N}(0,\mu^2)$. I want to prove this fact.

enter image description here

My tries : $z = xy = x(\bar{y} + \mu)=x\bar{y}+x\mu$ where $\bar{y} \sim \mathcal{N}(0,1)$.

It is clear that $x\mu \sim \mathcal{N}(0,\mu^2)$ while $\mathbb{E}[x\bar{y}]=0$ and $\textrm{var}(x\bar{y})=\textrm{var}(x)\textrm{var}(\bar{y})=1$. Thus $z \approx x\bar{y} \sim \mathcal{N}(0,\mu^2)$.

Could anyone provide more rigorous proof ?


Edit to tighten up the idea sketched below:

Consider the scaled random variable $W_{\mu} := XY/\mu$ for $\mu \neq 0$. Then we have that the characteristic function is given by \begin{align} \varphi_{W_\mu}(t) &= \mathbb{E}\left[e^{itW_{\mu}} \right] \\ &= \mathbb{E}\left[e^{i(t/\mu)XY }\right] \\ &= \frac{\exp\left[-\frac{t^2}{2(1+t^2/\mu^2)}\right]}{\sqrt{1+t^2/\mu^2}}. \end{align} Therefore as $\mu \rightarrow \infty$ you have pointwise convergence $\varphi_{W_{\mu}}(t) \rightarrow e^{-\frac{t^2}{2}}$ of the characteristic functions so that $$ \frac{XY}{\mu} \stackrel{\mathscr{D}}{\rightarrow} \mathcal{N}(0, 1). $$

There are details for you to tighten up and prove exactly, but the following should get you well on the way.

If $X \sim \mathcal{N}(0, 1)$ and $Y \sim \mathcal{N}(\mu,1)$ then the random variable $Z=XY$ has moment generating function $$ \begin{align} M_Z(t) &= \mathbb{E}\left[ e^{tXY}\right]\\ &= \frac{\exp\left[ \frac{\mu^2 t^2}{2(1-t^2)}\right]}{\sqrt{1-t^2}}\\ &= 1 + \frac{1}{2}(\mu^2 + 1)t^2 +\frac{1}{8}(\mu^4 + 6\mu^2+3)t^4 +\mathcal{O}(t^6), \end{align} $$ now if $\mu$ is large all of these polynomials, in $\mu$, are well approximated by only their leading degree monomials leading to $$ M_Z(t)\approx 1 + \frac{\mu^2 t^2}{2}+\frac{1}{2}\left(\frac{\mu^2 t^2}{2}\right)^2 + \cdots $$ or $$ M_Z(t) \approx \exp\left(\frac{\mu^2 t^2}{2} \right) $$ which is the moment generating function of a normal random variable with density $\mathcal{N}(0, \mu^2)$.


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.