1
$\begingroup$

This is a Gaussian bell (aka normal distribution).

enter image description here

Its square, I belive looks the same. Yet, I see that chi-square distribution, which is a sum of k such bell squares, looks like

enter image description here

Take a look at yellow chi-squared, with k=1. It should be just single bell squared. But, it looks like a hyperbola instead of Gaussian. Why?

PS I read that normally-distributed signal generates a uniformly distributed power. Is it the same kind of magic?

$\endgroup$
17
  • 1
    $\begingroup$ This is a psychological question: "Why is my intuition contradicted by mathematical facts?" It's better to do the algebra first and then adjust your graphical intuition accordingly. $\endgroup$ May 26 '13 at 15:15
  • $\begingroup$ Here is a simple algebra. Let's normal distribution is 1,2,1 -- a pretty nice Gaussian curve. Its square is 1,4,1 -- exactly the same Gaussian curve. Ok? $\endgroup$
    – Val
    May 26 '13 at 15:17
  • $\begingroup$ The normal is already very concentrated in a neighbourhood of $0$. Squaring magnifies that concentration a great deal. $\endgroup$ May 26 '13 at 15:17
  • 1
    $\begingroup$ @Val - the square of the standard normal distribution (mean = 0, standard deviation = 1) has a $\chi^2_1$ distribution. If you square an arbitrary Gaussian distribution, you'll get something more complicated. $\endgroup$ May 26 '13 at 15:22
  • 1
    $\begingroup$ @Val. Read what I wrote. "Gaussian" is the same as "normal". But "Standard normal" is a special case. Did's correct comment has nothing to do with this and is possibly the clue to your confusion. $\endgroup$ May 26 '13 at 15:28
1
$\begingroup$

Consider a random variable $X$ with density $f$, say $X$ gaussian, and $Y=X^2$. Then $Y\geqslant0$ almost surely and, for every $y\geqslant0$, $[Y\leqslant y]=[-\sqrt{y}\leqslant X\leqslant\sqrt{y}]$ hence $$ P[Y\leqslant y]=\int_{-\sqrt{y}}^{\sqrt{y}}f(x)\mathrm dx. $$ Differentiating this yields the density of $Y$, namely, for every $y\geqslant0$, $$ g(y)=\frac{f(\sqrt{y})+f(-\sqrt{y})}{2\sqrt{y}}. $$ When $X$ is normal (centered or not), $f$ is continuous at $0$ hence the behaviour of $g$ at $0$ is $$ g(y)\sim\frac{f(0)}{\sqrt{y}}. $$ Note finally that $f^2$ is rarely a density function.

Edit: Recall that an integral depending on a parameter, say $$ J(y)=\int_{u(y)}^{v(y)}w(x)\mathrm dx, $$ has derivative $$ J'(y)=w(v(y))v'(y)-w(u(y))u'(y). $$

$\endgroup$
7
  • $\begingroup$ I am not ready to consume this. But, thanks anyway. I see that you based your proof on the standard normal whereas enemy in comments suggested that discrepancy of X^2 from from normal distribution stems from non-standard Gaussians. $\endgroup$
    – Val
    May 26 '13 at 16:19
  • $\begingroup$ The centering or decentering of gaussians has little to do with the $c/\sqrt{y}$ behaviour, when $y\to0$, that my answer exhibits. $\endgroup$
    – Did
    May 26 '13 at 16:21
  • 1
    $\begingroup$ Thanks confirming that. But, unfortunately, I am too stupid to comprehend your answer. The $\sqrt y$ appears out of blue. I look at it like illusion, when a thing materializes out of nothing in the hands of magician. I guess that if we had power of 1 instead of two, i.e. y=x^1, you'd get g(y) = f(x)/y. $\endgroup$
    – Val
    May 26 '13 at 17:42
  • $\begingroup$ Re the appearance of $\sqrt{y}$, see my Edit. To sum up, $y\mapsto\pm\sqrt{y}$ is the inverse of $x\mapsto x^2$ one started with and the final $c/\sqrt{y}$ is more or less its derivative. $\endgroup$
    – Did
    May 26 '13 at 17:48
  • $\begingroup$ @Val - I am upvoting your last comment because there is some truth in it. $\endgroup$ May 26 '13 at 20:15
0
$\begingroup$

Let $Z$ be standard normal, or more generally a normal with mean $0$. But in fact we will use almost no properties of the normal.

Let $Y=Z^2$. We want to show that the density function $f_Y(y)$ of $Y$ is very large for $y$ near $0$. This density function is approximately $\frac{F_Y(y)}{y}$.

We have $F_Y(y)=2\Pr(Z\le \sqrt{y})$. For $y$ close to $0$, the density function of $Z$ is close to a non-zero constant $c$.

Thus $F_Y(y)$ is approximately $2c\sqrt{y}$. It follows that near $0$, $\frac{F_Y(y)}{y}$ is very large.

$\endgroup$

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.