Let $(X_1,X_2$) have bivariate normal distribution with means 0, variance $\sigma_1^2, \sigma_2^2$ respectively, and with correlation coefficient $-1<\rho <1$.

(i) Determine the distribution of $aX_1 + bX_2$, where $a,b \in \mathbb R $, such that $a^2 + b^2> 0$.

(ii) Find constants $b$ such that $X_1 + bX_2$ is independent of $X_1$.

(iii) Find the probability that the following equation has real roots: $$X_1 x^2 -2X_1x - bX_2=0 $$ where b is the constant found in part (ii)


(i) Using the transformation $U=aX_1, V = bX_2$, we have the joint distribution $$f_{U,V}= \frac{1}{2\pi \sigma_1 \sigma_2 a b \sqrt{1-\rho^2}}exp(\frac{-1}{2(1-\rho^2)}[\frac{u^2}{\sigma_1^2}+\frac{v^2}{\sigma_2^2}-\frac{2\rho uv}{ab\sigma_1 \sigma_2}]) $$

(ii) Not sure how to approach this, $b=0$ is a trivial solution, but are there any other solutions?

edit: $b=0$ is not allowed since $a^2 + b^2> 0$

(iii) Equation has real roots if $4X_1^2 + 4bX_1X_2\ge 0 $, which simplifies to $X_1(X_1+bX_2) \ge 0$,so we are interested in $P(X_1(X_1+bX_2) \ge 0 )$ which can be evaluated with a double integral.


Part (i). You can avoid working with the density by looking at properties of the Gaussian distribution.

  • Hint: $aX_1 + bX_2$ is (univariate) Gaussian. (Why?)

    This is an important property of the bivariate Gaussian distribution, and more generally of the multivariate Gaussian distribution.

  • Given the previous hint, it suffices to find the mean and variance of $aX_1 + bX_2$, call them $\tilde{\mu}$ and $\tilde{\sigma}^2$. Then $aX_1 + bX_2 \sim N(\tilde{\mu}, \tilde{\sigma}^2)$.

Part (ii).

  • $(X_1+b X_2, X_1)$ is also bivariate Gaussian (why?), so it suffices to compute $\text{Cov}(X_1 + bX_2, X_1)$ and check which values of $b$ make it zero.

    $\text{Cov}(X_1 + bX_2, X_1) = \sigma_1^2 + b \rho \sigma_1 \sigma_2$

Part (iii).

  • If you want to show $X_1(X_1 + bX_2) \ge 0$, then note that by part (b), $X_1$ and $X_1 + bX_2$ are independent. \begin{align} P(X_1(X_1 + bX_2) \ge 0) &= P(X_1 \ge 0) P(X_1+bX_2 \ge 0) + P(X_1 \le 0)P(X_1+bX_2 \le 0)\\ &= \frac{1}{2} \cdot \frac{1}{2} + \frac{1}{2} \cdot \frac{1}{2} = \frac{1}{2}. \end{align}
  • $\begingroup$ Im not aware of the properties of univariate gaussian, will look into them. There was a typo in my deductions, apologies for that! Thanks for the hints, really appreciate them :) $\endgroup$ – Timothy May 2 '17 at 3:53
  • $\begingroup$ In order to leverage part (ii) when solving part (iii), the discriminant should look like the product of the random variables in part (ii). I suspect there's a typo somewhere, or maybe I have made some mistake... $\endgroup$ – angryavian May 2 '17 at 4:15
  • $\begingroup$ You are right, there was indeed a typo in the question $\endgroup$ – Timothy May 2 '17 at 10:57

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.