3
$\begingroup$

Let $X$ be a real valued random variable.

Let $\operatorname{sgn}(x)$ be $1$ when $x>0$, $-1$ when $x<0$ and $0$ when $x=0$.

Why are $|X|$ and $\operatorname{sgn}(X)$ independent, when the density function of $X$ is symmetric with respect to $0$?

Are $|X|$ and $\operatorname{sgn}(X)$ independent, when the density function of $X$ is not necessarily symmetric with respect to $0$?

Thanks!

$\endgroup$
6
$\begingroup$

We write out part of the argument. Assume that $X$ has continuous symmetric distribution. Then $\Pr(X=0)=0$, and, for positive $x$, $$Pr(0\lt X\le x)=\frac{1}{2}\Pr(-x\le X \le x)=\frac{1}{2}\Pr(|X| \le x).$$ Thus $$\Pr((|X|\le x)\cap (\text{sgn}(X)=1))=\Pr((-x \le X \le x))\cap (\text{sgn}(X)=1))=\Pr(0\lt X \le x)=\frac{1}{2}\Pr(|X|\le x)=\Pr(|X|\le x)\Pr((\text{sgn}(X)=1)).$$ Almost exactly the same idea works when we replace $\text{sgn}(X)=1$ by $\text{sgn}(X)=-1$. The case $\text{sgn}(X)=0$ is trivial.

For the second part, we give an informal example and argument that we do not necessarily have independence when the distribution is not symmetric.

Suppose that $X$ has density function $\frac{1}{2}$ on $[-1,0]$, also $\frac{1}{2}$ on $[99,100]$, and $0$ elsewhere. If we know that $\text{sgn}(X)=1$, then we know that $|X|$ is big. So $|X|$ and $\text{sgn}(X)$ are not independent. We can write this out more formally in terms of conditional probabilities.

$\endgroup$
  • 4
    $\begingroup$ Conditional probabilities are not needed in the second part: if there is independence, then $P(X\in A)=\frac12P(X\in (A\cup-A))$ for every Borel set $A\subseteq\mathbb R_+$ hence (i) $P(X=0)=0$ and (ii) $P(X\in A)=P(X\in -A)$ for every Borel set $A\subseteq\mathbb R_+$. Thus, if there is a density, it is symmetric. $\endgroup$ – Did Jun 30 '12 at 6:08
5
$\begingroup$

If $X$ is a continuous random variables (absolutely continuous with respect to Lebesgue),

$$P(|X| \leq x|sgn(X)=1) = P(|X| \leq x|X > 0) = \frac{P(0 < X \leq x)}{P(X > 0)} = \frac{\int_{0}^{x}{f(x)dx}}{\int_{0}^{\infty}f(x)dx}$$

$$P(|X| \leq x|sgn(X)=-1) = P(|X| \leq x|X < 0) = \frac{P(-x \leq X < 0)}{P(X < 0)} = \frac{\int_{-x}^{0}{f(x)dx}}{\int_{-\infty}^{0}f(x)dx}$$

and $P(sgn(X)=0)$. If $f(x)$ is symmetric with respect to $0$, observe that $P(|X| \leq x|sgn(X)=1) = P(|X| \leq x|sgn(X)=-1)$ for any $x \geq 0$. Hence, $|X|$ is independent of sgn(x).

Now consider $X$ to be uniform in $(-1,2)$. Observe that $P(sgn(X)=1)=2/3$ and $P(sgn(X)=1||X|>1)=1$. Hence, $|X|$ and $sgn(X)$ are not independent.

Also observe that it was important for $X$ to be continuous. For example, consider $X $ uniform in $\{-1,0,1\}$. Its mass function is symmetric with respect to $0$ but $P(sgn(X)=0) = 1/3$ and $P(sgn(X)=0||X|=0)=1$ and, thus sgn(X) and |X| are not independent.

$\endgroup$
  • $\begingroup$ Isn't it only important for $X$ to be symmetric and have $P(X=0)=0$? (e.g. for $X$ uniform on $\{-1,1\}$, $\mathrm{sgn}(X)$ and $|X|$ are independent.) $\endgroup$ – huon Jun 30 '12 at 6:43
  • $\begingroup$ That's what did proved in his comment on André Nicola's answer, right? $\endgroup$ – madprob Jun 30 '12 at 6:46
  • $\begingroup$ yep seems so, I didn't notice that. :) $\endgroup$ – huon Jun 30 '12 at 6:48

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.