Let $Y_1, Y_2,$ and $Y_3$ be independent, $\mathrm{N}(0,1)$-distributed random variables, and set $$\begin{cases}X_1 = Y_1-Y_3 \\ X_2 = 2Y_1 + Y_2 - 2Y_3 \\X_3 = -2Y_1 + 3Y_3 \end{cases}. $$ Determine the conditional distribution of $X_2$ given that $X_1+X_3 = x$.

So, what I need to compute is $$\frac{f_{X_2, X_1+X_3}(x_2,x_1+x_3)}{f_{X_1+X_3}(x_1+x_3)}. $$

First, I thought I could find the joint distribution $f_{X_2,X_1+X_3}$. When computing $\mathrm{Cov}(X_2,X_1+X_3)$, I find that it is equal to zero. Meaning that $X_2$ and $X_1+X_3$ are independent if they are jointly normal. But if they are not, then how do I compute their joint distribution?

If they were in fact independent, then it would be enough to compute $f_{X_2}(x_2)$ and with $X_2 = 2Y_1+Y_2-2Y_3$ that would have the distribution $\mathrm{N}(B\mu, B\Lambda B')$ with $B= [2, 1, -2]$, but that gives me the result $\mathrm{N}(0,9)$, which is wrong. So $X_2$ and $X_1+X_3$ can't be independent, so they can't be jointly normal, so I have no idea how to compute their joint distribution. I do know how to compute $f_{\bar{Y}}(\bar{y})$ and $f_{\bar{X}}(\bar{x})$, but other than that, I know nothing, it feels like.

  • $\begingroup$ Have you double checked your covariance computation? $\endgroup$ – BGM Mar 6 '18 at 11:05

So I actually - finally - figured it out, and just in case anyone else is ever stuck on this, I'll post what I did.

Throughout I've been using the following formula: That for a random vector $X \in \mathrm{N}(\mu,\Lambda),$ the random vector $Y = BX$ has distribution $\mathrm{N}(B\mu, B\Lambda B')$.

By setting $$B = \begin{pmatrix}1 & 0 &-1 \\2 & 1 & -2\\-2&0&3 \end{pmatrix}$$ I found that $f_{\bar{X}}(\bar{x}) \in \mathrm{N}(0, \Lambda_{\bar{X}}),$ with $$\Lambda_{\bar{X}} = B \underbrace{\begin{pmatrix}1&0&0\\0&1&0\\0&0&1 \end{pmatrix}}_{\Lambda_{\bar{Y}}}B',$$ where the matrix $\Lambda_{\bar{Y}}$ is the covariance matrix for $(Y_1,Y_2,Y_3)'.$ I then did another similar transformation in order to find $f_{X_2, X_1+X_3}(x_2,x_1+x_3) = f_{X_2,U}(x_2,u)$ by setting $$\hat{B} = \begin{pmatrix}0&1&0\\1&0&1 \end{pmatrix}, $$ and this way finding $\Lambda_{(X_2,U)}.$ In other words, I computed $$f_{X_2,U}(x_2,u) = \frac{1}{\sqrt{2\pi}\sqrt{\mathrm{det}(\Lambda_{(X_2,U})}}\mathrm{exp}\left(-\frac{1}{2}\left((x_2,u)\Lambda_{(X_2,U)}^{-1}(x_2,u)' \right) \right). $$

I did the same for $f_{X_1+X_3}(x_1+x_3) = f_U(u)$ by using $\tilde{B} = (1,0,1)$, and then I could compute $$\frac{f_{X_2,U}(x_2,u)}{f_U(u)}. $$


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.