5
$\begingroup$

Let $X$ and $Y$ be two discrete random variables. I know the joint probability distribution of the vector $(X,Y)$, namely $P(X = x, Y = y)$ for all $x$ and $y$ in the sample spaces $\Omega_X$ and $\Omega_Y$, respectively. Using the joint distribution, I discovered that $X$ and $Y$ are conditionally dependent.

Now let $U := X + Y$ and $V := X \cdot Y$ be two random variables. I need to calculate the distribution of both $U$ and $V$. Is it correct to calculate the distribution of $U$ as \begin{align} P(U = u) = \sum_{x \in \Omega_X, y \in \Omega_Y \text{ such that } x \cdot y = u} P(X = x, Y = y), \end{align} even though $X$ and $Y$ are conditionally dependent?

Furthermore I need to calculate the joint distribution of the probabilty vector $(U,V)$. I know that the distribution can be calculated as \begin{align} P(U = u, V = v) = P(U = u) \cdot P(V = v), \end{align} if $U$ and $V$ are conditionally independent. However, to show that $U$ and $V$ are conditionally independent, I would look at the joint distribution, which I have to calculate. How does one calculate the joint distribution of $(U,V)$, if the variables are conditionally dependent? Furthermore, is there any way to find out whether $U$ and $V$ are conditionally dependent or not, without using the joint probability distribution? Since $U$ and $V$ are defined as sum and product of two random variables (which are conditionally dependent), I have the feeling that $U$ and $V$ are conditionally dependent.

$\endgroup$
  • $\begingroup$ Thanks for asking. I was wondering this myself! $\endgroup$ – amcalde Dec 10 '14 at 23:13
  • $\begingroup$ Implicit in Graham Kemp's answer is that it doesn't matter whether $X$ and $Y$ are dependent. (You speak of them being "conditionally dependent" without specifying the event with respect to which they are conditionally dependent.) $\endgroup$ – joriki Jul 1 '18 at 7:11
1
$\begingroup$

Yes, if you know the joint probability distribution of $X,Y$ then the distributions of $U$ and $V$ are, respectively:

$$\begin{align} \mathsf P(U=u) & = \sum_\overbrace{x\in\Omega_X, u-x\in\Omega_Y} \mathsf P(X=x, Y=u-x) \\ \mathsf P(V=v) & = \sum_\overbrace{x\in\Omega_X, v/x\in\Omega_Y} \mathsf P(X=x, Y=v/x) \end{align}$$


The joint distribution on $U,V$, would similarly be $$\mathsf P(U=u, V=v) = \sum_\overbrace{x\in\Omega_X, u-x\in\Omega_Y, v/x\in\Omega_Y} \mathsf P(X=x, Y=u-x, Y=v/x)$$

However, for any given pair of $U=u,V=v$ there is in fact only two corresponding pair of $X=\Box,Y=\Box$.

$\begin{align} (U=X+Y) \wedge (V= XY) & \iff \left(X= \dfrac{U\pm\sqrt{U^2-4V}}{2}\right) \wedge \left(Y=\dfrac{U\mp\sqrt{U^2-4V}}{2}\right) \\[4ex] \therefore \mathsf P(U=u, V=v) & = \mathsf P\left(X= \dfrac{u\pm\sqrt{u^2-4v}}{2}, Y=\dfrac{u\mp\sqrt{u^2-4v}}{2}\right) \\[1ex] & = {\mathsf P\left(X= \tfrac{u+\sqrt{u^2-4v}}{2}, Y=\tfrac{u-\sqrt{u^2-4v}}{2}\right) + \mathsf P\left(X= \tfrac{u-\sqrt{u^2-4v}}{2}, Y=\tfrac{u+\sqrt{u^2-4v}}{2}\right)} \end{align}$

$\endgroup$

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.