1
$\begingroup$

Suppose that r.v $X$ and $Y$ are normally distributed and independent each other. Under which conditions $(X,Y)$ are bivariate normally distributed?
Now I want to calculate $\text{Prob}(X>a, Y<a, bX+cY<d)$, where $a,b,c,d$ are constant. Should I treat the third element as another r.v (let say $Z$) whose parameters are easily calculated and as a consequence, consider $Pr(X>a, Y<a, bX+cY<d)$ under a trivariate normal distribution perspective?

$\endgroup$
2
$\begingroup$

Well if $X$ and $Y$ are independent and each one is normally distributed, then $(X,Y)$ is also bivariate normal (with $Cov(X,Y)=0$). And no, I do not suggest using a trivariate normal distribution because the distribution of $(X,Y, bX+cY)$ is degenerate, in the sense that its support lies in a 2 dimensional subspace of $\mathbb R^3$.

You just need to integrate over the pdf of the bivariate normal (which due to Fubini's theorem and independence is really simple). For example if $b>0$ then $bX+cY < d$ is equivalent to $X < \frac{d-cY}{b}$. Then you just integrate:

$$\Pr[X>a, Y<a, bX+cY<d] = \int_{-\infty}^a\int_{a}^{\frac{d-cy}{b}} \mathbf{1}_{\left\{\frac{d-cy}{b}\geq a\right\}} f_X(x)f_Y(y) dx dy$$

Here $f_X$ (resp. $f_Y$) is the density of the distribution of $X$ ($Y$) and $\mathbf{1}_{\{\}}$ is the indicator function.

$\endgroup$
  • $\begingroup$ thank you! what about if b or c are negative? I take advantage of your expertise for another question: as you don't suggest to use a trivariate normal distribution,which package of R do you suggest in this case (I've thought of mvtnorm in case of trivariate distribution).Many thanks $\endgroup$ – Andrea Oct 23 '15 at 17:37
  • $\begingroup$ Part1) Well if b is negative then you just flip the inequality around and it holds that $X < \frac{d-cY}{b}$. I will let you figure out on your own what this means regarding the integration region! Also to be honest I would probably derive an analytic expression for this case (at least get rid of the inner integral) and then just write my own R function. But I am sure there are also numeric integration packages you can use in R (but not sure which or how you would specify the information). $\endgroup$ – air Oct 23 '15 at 17:45
  • $\begingroup$ Part 2) But for such a low dimension you can actually do it via Monte Carlo if the integrals scare you! So you just generate independent random variables $X_i$, $Y_i$ according to the distribution of $X$ and $Y$. And you do that many times (e.g. $m$ times) so that you get pairs $(X_i,Y_i)$ for $i=1,\dotsc,m$. Now for each of these pairs you check whether all 3 conditions ($X_i >a$, $Y_i < a$ and $bX+cY<d$) are fullfilled. Now the probability you are interested in is just the proportion of pairs which fullfill this property. $\endgroup$ – air Oct 23 '15 at 17:47

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.