3
$\begingroup$

I tried to make $T$ close to

$$T_1=\left(\frac{(x_1 - \mu_1)^2}{\sigma_1^2}+\frac{(x_2 - \mu_2)^2}{\sigma_2^2}-2 \frac{\rho}{\sigma_1 \sigma_2} \frac{x_1 - \mu_1}{\sigma_1}\frac{x_2 - \mu_2}{\sigma_2}\right)\frac{1}{1-(\frac{\rho}{\sigma_1\sigma_2})^2}$$

to solve the following problem.

Problem: let $x=(X_1,X_2)'$ follow $N_2(\mu,\Sigma)$, where $\mu=(\mu_1,\mu_2)$, and $\Sigma=\begin{bmatrix} \sigma_1^2 & \rho\\ \rho & \sigma_2^2 \end{bmatrix}$.

Show that

$$T=\frac{1}{2(1- \rho^2)}\left(\frac{(x_1 - \mu_1)^2}{\sigma_1^2}+\frac{(x_2 - \mu_2)^2}{\sigma_2^2}-2 \rho \frac{x_1 - \mu_1}{\sigma_1}\frac{x_2 - \mu_2}{\sigma_2}\right)$$

has a $\chi^2$ distribution. And give the its parameters? Since $T_1$ follows as $\chi^2$, I tried to make $T$ close to $T_1$, but I failed since

$$2 \frac{\rho}{\sigma_1 \sigma_2} \frac{x_1 - \mu_1}{\sigma_1}\frac{x_2 - \mu_2}{\sigma_2}\text{ and }2 \rho \frac{x_1 - \mu_1}{\sigma_1}\frac{x_2 - \mu_2}{\sigma_2}$$

are different.

$\endgroup$
  • $\begingroup$ I was trying use theorem, that is "Let $y$ be distributed as $N_p (\mu,\Sigma)$, let $A$ be a symmetric matrix of constants of rank $r$, and let $\lambda=\frac{1}{2} \mu^{'}A\mu$. Then $y^{'}Ay$ is $\chi^2 (r,\lambda)$ if and only if $A\Sigma$ is idempotent". But I cannot find $A$. $\endgroup$ – 81235 Oct 4 '14 at 21:40
  • $\begingroup$ Now you know what $A $is, since it appears in my answer (under the same name...). But there is absolutely no reason to assume that $A\Sigma$ is idempotent. You have to realize that this theorem you cite has the form "If ever some matrix has a very specific form, then the answer is so-and-so", not "The answer is so-and-so in the general case". $\endgroup$ – Did Oct 5 '14 at 9:57
  • $\begingroup$ Unaccepted? Why? $\endgroup$ – Did Mar 12 '15 at 7:25
3
$\begingroup$

If $X=(X_1,X_2)'$ is normal with mean $\mu=(\mu_1,\mu_2)$, then every quadratic form $$T=a_1(X_1-\mu_1)^2+a_2(X_2-\mu_2)^2+2b(X_1-\mu_1)(X_2-\mu_2),$$ with $a_1\gt0$ and $a_2\gt0$, follows a generalized $\chi^2$ distribution. Only, the parameters depend on $(a_1,a_2,b)$ and on $\Sigma$ the covariance matrix of $X$.

Edit: Assume without loss of generality that $\mu=(0,0)$, consider the matrices $A=\begin{pmatrix}a_1&b\\b&a_2\end{pmatrix}$ and $L$ such that $\Sigma=LL'$, then $X=LY$ where $Y$ is standard normal and $T=X'AX=Y'MY$ where $M=L'AL$ is symmetric. Thus $M=PDP'$ where $P$ is orthogonal and $D$ diagonal, and $T=Z'DZ$ where $Z=P'Y$ is (again) standard normal, that is, $$T=d_1Z_1^2+d_2Z_2^2,$$ where $d_1$ and $d_2$ are the diagonal entries of $D$. If $d_1$ or $d_2$ is zero, then $T$ is a multiple of a $\chi^2(1)$ random variable. If $d_1$ and $d_2$ are both positive, then a change of variables shows that $T$ is a generalized $\chi^2(2)$ random variable, since the density of $T$ is $$f_T(t)=\int_0^{2\pi}\mathrm e^{-((\cos^2 s)/d_1+(\sin^2s)/d_2)t/2}\frac{\mathrm ds}{2\pi\sqrt{d_1d_2}},$$ which can be simplified into $$f_T(t)=\sqrt{\alpha^2-\beta^2}\,\mathrm e^{-\alpha t}\,I_0\left(\beta t\right),$$ where $I_0$ is the modified Bessel function of the first kind and $$\alpha=\frac14\left(\frac1{d_1}+\frac1{d_2}\right),\qquad\beta=\frac14\left(\frac1{d_1}-\frac1{d_2}\right).$$ In the end:

The key parameters of the problem are the eigenvalues of the matrix $M=L'AL$, where $A$ depends on $(a_1,a_2,b)$, and $\Sigma=LL'$.

The case in your post reads $$T=\frac{1}{2(1- \rho^2)}\left(\frac{(x_1 - \mu_1)^2}{\sigma_1^2}+\frac{(x_2 - \mu_2)^2}{\sigma_2^2}-2 \rho \frac{x_1 - \mu_1}{\sigma_1}\frac{x_2 - \mu_2}{\sigma_2}\right)$$ and is equivalent to $\Sigma=\begin{pmatrix}1&\rho\\\rho&1\end{pmatrix}$ and $A=\frac{1}{2(1- \rho^2)}\begin{pmatrix}1&-\rho\\-\rho&1\end{pmatrix}$.

(Note that your $\Sigma$, before normalization, should read $\Sigma=\begin{pmatrix}\sigma_1^2&\rho\sigma_1\sigma_2\\\rho\sigma_1\sigma_2&\sigma_2^2\end{pmatrix}$, otherwise $\Sigma$ is not a covariance matrix when $\sigma_1^2\sigma_2^2\lt\rho^2$.)

One can use $L=\begin{pmatrix}1&0\\\rho&\tau\end{pmatrix}$ with $\tau^2=1-\rho^2$, then a simple computation yields $M=\frac12I$ hence $T=\frac12(Y_1^2+Y_2^2)$ where $(Y_1,Y_2)$ is i.i.d. standard normal. Finally, $d_1=d_2=\frac12$ hence $\alpha=1$ and $\beta=0$, thus, the distribution of $T$ is standard exponential, with $$f_T(t)=\mathrm e^{-t}.$$

$\endgroup$
  • $\begingroup$ Could you please provide more details about how to calculate the parameters? $\endgroup$ – 81235 Oct 4 '14 at 20:43
  • $\begingroup$ If you want, you can refer a book or a link. Thanks $\endgroup$ – 81235 Oct 4 '14 at 21:19
  • $\begingroup$ Can I calculate $\lambda$ and degrees of freedom from $\alpha$ and $\beta$? $\endgroup$ – 81235 Oct 5 '14 at 0:34
  • $\begingroup$ See my comment to your question. (Reading yours, I am not sure you really digested the Edit in my post... For example, in general $T$ is NOT $\chi^2$, only generalized $\chi^2$...) $\endgroup$ – Did Oct 5 '14 at 9:59

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.