Suppose $X_1$ and $X_2$ are iid with a common standard normal distribution. Find the joint pdf of $Y_1 = X_1^2 + X_2^2$ and $Y_2=X_2$ and the marginal pdf of $Y_1$.

It's easy to see that $Y_1$ has a $\chi^2(2)$ distribution, but this problem asks to show it using change of variables technique. Here, I can find the Jacobian $\partial(x_1,x_2) / \partial(y_1,y_2) = \frac{1}{2x_1}$, where $x_1^2 = y_1 - y_2^2$. Hence we must have

$$f_{Y_1,Y_2}(y_1,y_2) = \frac{1}{2x_1} \frac{1}{2\pi} e^{-y_1^2}.$$

However, in this case we have $x_1 = \pm \sqrt{y_1 - y_2^2}$. How can we integrate $y_2$ out from this pdf to get the marginal pdf of $Y_1$?

  • $\begingroup$ If $X_1,X_2\sim\operatorname{{i.}{i.}{d.}} N(0,1),$ then $X_1^2+X_2^2 \sim \chi^2_2,$ not $\chi^2_1. \qquad$ $\endgroup$ – Michael Hardy Aug 5 '18 at 15:28
  • $\begingroup$ @MichaelHardy Thanks that was my mistake. $\endgroup$ – takecare Aug 5 '18 at 15:29
  • $\begingroup$ When the transformation is not injective (in your post you noted that there are two values of $x_1$ to consider), you need to add up all the cases separately. Also, you forgot to take the absolute value of your Jacobian determinant. \begin{align}f_{Y_1, Y_2}(y_1, y_2) = f_{X_1, X_2}(\sqrt{y_1 - y_2^2}, y_2) \cdot \frac{1}{2 \sqrt{y_1 - y_2^2}} + f_{X_1, X_2}(-\sqrt{y_1 - y_2^2}, y_2) \cdot \frac{1}{2 \sqrt{y_1 - y_2^2}} = \frac{1}{\sqrt{y_1 - y_2^2}} \frac{1}{2 \pi} e^{-y_1/2} \end{align} But this is not integrable with respect to $y_2$, so I must have made a mistake somewhere... $\endgroup$ – angryavian Aug 5 '18 at 15:53
  • $\begingroup$ Ah, I see from StubbronAtom's answer below that I should have integrated over $y_2$ in $[-\sqrt{y_1}, \sqrt{y_1}]$ rather than over all real numbers. $\endgroup$ – angryavian Aug 5 '18 at 15:59
  • $\begingroup$ If the purpose was to find the distribution of $X_1^2+X_2^2$ using change of variables, I would have proposed the one-to-one polar transformation $(x_1,x_2)\to(r,\theta)$ such that $x_1=r\cos\theta$ and $x_2=r\sin\theta$. $\endgroup$ – StubbornAtom Aug 5 '18 at 16:36

Joint density of $(X_1,X_2)$ is $$f_{X_1,X_2}(x_1,x_2)=\frac{1}{2\pi}e^{-\frac{1}{2}(x_1^2+x_2^2)}\quad,(x_1,x_2)\in\mathbb R^2$$

We change variables $(X_1,X_2)\to(Y_1,Y_2)$ such that $Y_1=X_1^2+X_2^2$ and $Y_2=X_2$.

The inverse solutions of the transform are

$$x_1=\begin{cases}\sqrt{y_1-y_2^2}&,\text{ if }x_1\ge0\\-\sqrt{y_1-y_2^2}&,\text{ if }x_1<0\end{cases}\qquad\text{ and }\quad x_2=y_2$$

We have, $(x_1,x_2)\in\mathbb R^2\implies y_1\ge0\,,y_2\in\mathbb R$.

But for $x_1$ to be defined, $y_1-y_2^2\ge0\implies -\sqrt{y_1}\le y_2\le \sqrt{y_1}$.

So the joint support of $(Y_1,Y_2)$ is $$S=\{(y_1,y_2):y_1\ge0\,,\, -\sqrt{y_1}\le y_2\le\sqrt{y_1}\}$$

Clearly, this is not a one-to-one transformation.

Jacobian of the transformation is given by

$$J_1=\det J\left(\frac{x_1,x_2}{y_2,y_2}\right)=\frac{1}{2\sqrt{y_1-y_2^2}}\quad\text{ if }x_1\ge0$$

and $$J_2=\frac{-1}{2\sqrt{y_1-y_2^2}}\quad\text{ if }x_1<0$$

Hence joint density of $(Y_1,Y_2)$ is

\begin{align} f_{Y_1,Y_2}(y_1,y_2)&=\frac{1}{2\pi}e^{-y_1/2}|J_1|\mathbf1_S+\frac{1}{2\pi}e^{-y_1/2}|J_2|\mathbf1_S \\&=2\times\frac{1}{2\pi}e^{-y_1/2}\frac{1}{2\sqrt{y_1-y_2^2}}\mathbf1_S \\&=\frac{1}{2\pi}e^{-y_1/2}\frac{1}{\sqrt{y_1-y_2^2}}\mathbf1_S \end{align}

For $y_1\ge 0$, marginal pdf of $Y_1$ is thus

\begin{align} f_{Y_1}(y_1)&=\int_{-\sqrt {y_1}}^{\sqrt {y_1}}f_{Y_1,Y_2}(y_1,v)\,dv \\&=\frac{e^{-y_1/2}}{2\pi}\sin^{-1}\left(\frac{v}{\sqrt{y_1}}\right)\big|_{-\sqrt {y_1}}^{\sqrt {y_1}} \\&=\frac{1}{2}e^{-y_1/2} \end{align}

That is, $$f_{Y_1}(y_1)=\frac{1}{2}e^{-y_1/2}\mathbf1_{y_1\ge 0}$$

So we have shown that $Y_1\sim\text{Exp}$ with mean $2$ or simply, $Y_1\sim\chi^2_2$.


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.