1
$\begingroup$

If $X$ and $Y$ are independent normally distributed random variables

$$X,Y\sim\mathcal{N}(0,\sigma^2)$$

How are the sum and product, $X+Y$ and $XY$, co-distributed? You can write the moment generating function

$$V=X+Y\qquad Z=XY$$ \begin{align} M_{V,Z}(s,t)&=\int _{-\infty }^{\infty }\int _{-\infty }^{\infty }\exp\left[s (v+z)+t (vz)\right]\psi (v,0,\sigma^2) \psi (z,0,\sigma^2)\mathrm{d}v\mathrm{d}z\\ &=\frac{\exp\left[\dfrac{s^2 \sigma ^2}{1-\sigma ^2 t}\right]}{\sqrt{1-\sigma ^4 t^2}} \end{align}

But I do not know what probability density function this belongs to, besides that it looks like the MGF for a modified Bessel function of the second kind and normal distribution. Attached is a histogram generated from $10^7$ such pairs of random numbers.

Bivariate distribution of V and Z. Projections are Gaussian and the modified Bessel function of the second kind respectively

EDIT: Turns out Mathematica can handle this thing, and spits this monster out $$f(v,z;\sigma^2)=\frac{\sqrt[4]{-1} \left[\sqrt{-i \left(v^2-4 z\right)}-i \sqrt{i \left(v^2-4 z\right)}\right] }{\left| v^2-4 z\right| }\frac{\exp\left(-\dfrac{v^2-2 z}{2 \sigma ^2}\right)}{2 \pi \sigma ^2}$$ Does anyone know if this related to some well-known result I could read around?

f(v,z;\sigma^2)

$\endgroup$
  • $\begingroup$ Ummm ... Your question seeks the joint pdf of $(V, Z)$ where $V=X+Y$ and $Z=X*Y$, where presumably you intend $X$ and $Y$ are assumed independent Normals. You then post some formulae that you claim Mathematica spat out as the solution to your problem, namely $f(x,y) = blah$, which (a) is self-evidently not posed in the form of your question i.e. is not the joint pdf of $(V,Z)$, and (b) in any event, we already know the joint pdf of $(X,Y)$ which is bivariate Normal with zero correlation, because that is your starting assumption ... which is not the equation you post. $\endgroup$ – wolfies Apr 8 '14 at 17:18
  • $\begingroup$ Many apologies, the additional equation introduces confusion by using x and y as the independent variables. It gives the joint PDF of (V,Z). $\endgroup$ – user5751 Apr 8 '14 at 17:53
  • $\begingroup$ Although it just a symbol but my professor always says to me to not ever use symbol $Z$ when dealing with normal distribution but not standard normal distribution. It just a random comment. $\endgroup$ – Tunk-Fey Apr 8 '14 at 18:39
  • $\begingroup$ Thanks for your help Wolfies. I've included a plot of that function. It appears to match the numerical results, although the asymptote at $v^2-4z=0$ is very poorly resolved. The function Mathematica spat at me appears to be the solution, I've simply never encountered this flavour of distribution before. $\endgroup$ – user5751 Apr 8 '14 at 18:42
  • $\begingroup$ @user5751 Perhaps you have entered the Mathematica output incorrectly here? Also, the Mathematica output you have posted appears unnecessarily messy - and methinks is incomplete without giving some thought to defining the domain of support. $\endgroup$ – wolfies Apr 9 '14 at 5:07
3
$\begingroup$

The usual method works. Assume without loss of generality that $\sigma^2=1$, then, for every measurable bounded function $u$, $$ E(u(V,Z))=E(u(X+Y,XY)), $$ hence $$ E(u(V,Z))=\iint u(x+y,xy)\,\mathrm e^{-(x^2+y^2)/2}\,\frac{\mathrm dx\mathrm dy}{2\pi}, $$ and the task is to transform the RHS into $$ \iint u(v,z)\,g(v,z)\,\mathrm dv\mathrm dz, $$ since then the function $g$ shall be the density of $(V,Z)$.

The change of variable $(x,y)\to(v,z)=(x+y,xy)$ is two-to-one, it has Jacobian $\mathrm dv\mathrm dz=|x-y|\,\mathrm dx\mathrm dy$, and it is inverted through the formulas $x,y=\frac12(v\pm\sqrt{v^2-4z})$ for every $(v,z)$ such that $4z\leqslant v^2$. Finally, $|x-y|=\sqrt{v^2-4z}$ and $x^2+y^2=v^2-2z$ hence $$ E(u(V,Z))=\iint u(v,z)\,2\,\mathrm e^{-(v^2-2z)/2}\,\mathbf 1_{v^2\gt4z}\,\frac{\mathrm dv\mathrm dz}{2\pi\sqrt{v^2-4z}}, $$ and the density of $(V,Z)$ is $$ g(v,z)=\frac1\pi\,\mathrm e^{-v^2/2}\,\mathrm e^{z}\,\frac{\mathbf 1_{v^2\gt4z}}{\sqrt{v^2-4z}}. $$ If $\sigma^2\ne1$, $(V,Z)$ is transformed into $(\sigma V,\sigma^2Z)$ hence the density becomes $$ g(v,z)=\frac1\pi\,\frac1{\sigma^2}\,\mathrm e^{-v^2/(2\sigma^2)}\,\mathrm e^{z/\sigma^2}\,\frac{\mathbf 1_{v^2\gt4z}}{\sqrt{v^2-4z}}. $$ In other words, in the formula you suggest at the end of your post, the funny looking factor $$ \frac{\sqrt[4]{-1} \left[\sqrt{-i \left(v^2-4 z\right)}-i \sqrt{i \left(v^2-4 z\right)}\right] }{\left| v^2-4 z\right| }$$ should read $$ \frac{2\,\mathbf 1_{v^2\gt4z}}{\sqrt{v^2-4z}}. $$

$\endgroup$
  • $\begingroup$ Did wrote: $$ g(v,z)=\frac1\pi\,\frac1{\sigma^2}\,\mathrm e^{-v^2/(2\sigma^2)}\,\mathrm e^{2z/\sigma^2}\,\frac{\mathbf 1_{v^2\gt4z}}{\sqrt{v^2-4z}}. $$ Do you perhaps mean: $$ g(v,z)=\frac1\pi\,\frac1{\sigma^2}\,\mathrm e^{-(v^2 - 2 z)/(2\sigma^2)}\,\frac{\mathbf 1_{v^2\gt4z}}{\sqrt{v^2-4z}}.$$ $\endgroup$ – wolfies Apr 8 '14 at 20:05
  • $\begingroup$ @wolfies Yes this is what I mean. $\endgroup$ – Did Apr 9 '14 at 4:34

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.