If $X\sim\mathcal C(0,1)$, find the distribution of $Y=\frac{2X}{1-X^2}$.

We have $F_Y(y)=\mathrm{Pr}(Y\le y)$

$\qquad\qquad\qquad=\mathrm{Pr}\left(\frac{2X}{1-X^2}\le y\right)$

$\qquad\qquad=\begin{cases} \mathrm{Pr}\left(X\in\left(-\infty,\frac{-1-\sqrt{1+y^2}}{y}\right]\right)+\mathrm{Pr}\left(X\in\left(-1,\frac{-1+\sqrt{1+y^2}}{y}\right]\right),\text{if}\quad y>0\\ \mathrm{Pr}\left(X\in\left(-1,\frac{-1+\sqrt{1+y^2}}{y}\right]\right)+\mathrm{Pr}\left(X\in\left(1,\frac{-1-\sqrt{1+y^2}}{y}\right]\right),\text{if}\quad y<0 \end{cases}$

I wonder if the above case distinction is correct or not.

On the other hand, the following seems a simpler method:

We can write $Y=\tan(2\tan^{-1}X)$ using the identity $\frac{2\tan z}{1-\tan^2z}=\tan 2z$

Now, $X\sim\mathcal C(0,1)\implies\tan^{-1}X\sim\mathcal R\left(-\frac{\pi}{2},\frac{\pi}{2}\right)$

$\qquad\qquad\qquad\quad\implies 2\tan^{-1}X\sim\mathcal R(-\pi,\pi)$

$\qquad\qquad\qquad\quad\implies\tan\left(2\tan^{-1}X\right)\sim\mathcal C(0,1)$, the last one being a 2-to-1 transformation.

But if I am asked to derive the distribution of $Y$ from definition, I guess the first method is how I should proceed. The calculation becomes a bit messy, but do I reach the correct conclusion? Any alternate solution is also welcome.

Continuous Univariate Distributions (Vol.1) by Johnson-Kotz-Balakrishnan has highlighted this property of the Cauchy distribution. As it turns out, this is just a special case of a general result.

enter image description here

enter image description here

  • 4
    $\begingroup$ The second solution is completely correct, so there should be no objection to it. $\endgroup$ – Xi'an Dec 17 '17 at 16:40
  • 1
    $\begingroup$ Addendum: since $P(X<x)=\tan^{-1}(x)/\pi+1/2$, the first resolution should end up using this identity on the tangent. $\endgroup$ – Xi'an Dec 17 '17 at 16:56
  • $\begingroup$ @Xi'an Actually I am trying to finish off the argument in the first method. $\endgroup$ – StubbornAtom Dec 17 '17 at 18:30

An alternative, more simplistic, way to look at it:

standard Cauchy distribution: $$f(x) \text{d}x = \frac{\pi^{-1}}{x^2+1} \text{d}x $$

transformations of variables: $$ u(x) = \frac{2x}{1-x^2} \qquad \text{and} \qquad x_1(u) = \frac{-1 - \sqrt{u^2+1}}{u} \, , \, x_2(u) = \frac{-1 + \sqrt{u^2+1}}{u} $$

transformation of distribution: $$g(u)\text{d}u = \sum_{i=1,2} f(x_i(u)) \left|\frac{\text{d}x_i}{\text{d}u}\right|\text{d}u $$

If you work with that, which does not need to become so messy, then you will get

$$g(u) = \frac{\pi^{-1}}{u^2+1}$$

graphical representation

intuitive graphical representation of transformation

This sort of works like the identity $\frac{2\tan z}{1-\tan^2z}=\tan 2z$, but written more explicitly.

Or like your representation with the split cumulative distribution function $F_Y(y) = Pr(Y \leq y)$ but now for a split in $f_Y(y) = Pr(y-\frac{1}{2}dy \leq Y\leq y+\frac{1}{2}dy)$.

  • 2
    $\begingroup$ Actually, the transformation formula, when $x(u)$ has more than one root for any given $u$, say $x_i(u) =u$ for $i=1,2,\ldots n$, is $$g(u) =\sum_{i=1}^n f(x_i(u))\left|\frac{\mathrm dx_i(u)}{\mathrm du}\right|.$$ Thus, the addition that you describe as necessary is actually built into the formula. $\endgroup$ – Dilip Sarwate Dec 18 '17 at 15:21
  • $\begingroup$ @DilipSarwate I will change it. $\endgroup$ – Sextus Empiricus Dec 18 '17 at 15:23

The transformation in the second approach seems lack of motivation (some details in that also need to be filled up with). Here, from the characteristic function calculation, I am trying to back up your "mysterious" transformation.

The characteristic function of $Y$ can be calculated as follows: \begin{align} \varphi_Y(t) = & E[e^{itY}] = \int_{-\infty}^{\infty} e^{it\frac{2x}{1 - x^2}} \frac{1}{\pi(1 + x^2)} dx \\ = & \frac{1}{\pi}\int_{-\infty}^\infty e^{it\frac{2x}{1 - x^2}} d\arctan x, \end{align} which suggests us trying the transformation $u = \arctan x$, which leads to \begin{align} \varphi_Y(t)= \frac{1}{\pi}\int_{-\pi/2}^{\pi/2} e^{it\frac{2\tan u}{1 - \tan^2 u}} du = \frac{1}{\pi}\int_{-\pi/2}^{\pi/2}e^{it\tan(2u)}du. \tag{1} \end{align}

Our goal is to show that the integral in $(1)$ equals to the characteristic function of a standard Cauchy random variable $X$: \begin{align} \varphi_X(t) = & \int_{-\infty}^\infty e^{itx}\frac{1}{\pi(1 + x^2)} dx \\ = & \frac{1}{\pi}\int_{-\pi/2}^{\pi/2} e^{it\tan u} du \tag{2} \end{align}

Why does the integral in $(1)$ equal to the integral in $(2)$? At the first glance, this is a little counter-intuitive. To verify it, we need to treat the monotonicity of the function $\tan(\cdot)$ carefully. Let's continue to work on $(1)$:

\begin{align} \varphi_Y(t) = & \frac{1}{\pi}\int_{-\pi/2}^{\pi/2}e^{it\tan(2u)}du \\ = & \frac{1}{2\pi}\int_{-\pi}^\pi e^{it\tan v} dv \quad (\text{Change of variable } v = 2u) \\ = & \frac{1}{2\pi}\left[\int_{-\pi}^{-\pi/2} + \int_{-\pi/2}^{\pi/2} + \int_{\pi/2}^{\pi}\right]e^{it\tan u} du \\ = & \frac{1}{2}\varphi_X(t) + \frac{1}{2\pi}\int_{-\pi}^{-\pi/2}e^{it\tan v} dv + \frac{1}{2\pi}\int_{\pi/2}^{\pi}e^{it\tan v} dv \quad (3) \\ = & \frac{1}{2}\varphi_X(t) + \frac{1}{2\pi}\int_{-\pi/2}^{0} e^{-it\tan u_1} du_1 + \frac{1}{2\pi}\int_{0}^{\pi/2} e^{-it\tan u_2} du_2 \quad (4) \\ = & \frac{1}{2}\varphi_X(t) + \frac{1}{2\pi}\int_{-\pi/2}^{\pi/2} e^{-it\tan v} dv \\ = & \varphi_X(t) \quad (5) \end{align}

$(3)$: Because the function $u \mapsto \tan(u)$ is not monotone on the interval $(-\pi, \pi)$, I made such division such that each integrand is monotone on the separated interval (which ensures subsequent change of variable formulae valid).

$(4)$: The two change of variable formulae are $u_1 = -\pi - v$ and $u_2 = \pi - v$.

$(5)$: Last change of variable formula $u = -v$.

The steps $(3)$--$(5)$ elaborated the statement "the last one being a 2-to-1 transformation" in OP's question.

  • $\begingroup$ I wonder why the second approach is 'mysterious' or 'lacks motivation'. The fact that $\Theta\sim\text{Rect}(-\pi/2,\pi/2)\Leftrightarrow \tan(\Theta)\sim\mathcal C(0,1)$ is a very standard result which is easily seen using the probability integral transformation. And in the last step where I go from $U\sim\text{Rect}(-\pi,\pi)$ to $V=\tan U\sim\mathcal C(0,1)$ is possibly justified as follows: $\endgroup$ – StubbornAtom Dec 17 '17 at 18:03
  • $\begingroup$ ...$F_V(v)=\mathrm{Pr}(\tan U\le v)=F_U(\tan^{-1}v)$. I differentiate the above wrt $v$ to get $f_V(v)=f_U(\tan^{-1}v)2\frac{d}{dv}(\tan^{-1}v)$, where I multiply the Jacobian by 2 because the transformation is two to one in $(-\pi,\pi)$. All this can be expressed more rigorously I guess. $\endgroup$ – StubbornAtom Dec 17 '17 at 18:03

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.