Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

If $X$ and $Y$ are two independent random variables with probability density functions $f$ and $g$, respectively, then the probability density of the difference $Y − X$ is given by the cross-correlation . In contrast, the convolution f * g gives the probability density function of the sum $X + Y$.

what i would like to have is the probability density functions of $|y-x|$ or $(y-x)^2$. Is that possible ?

is that the equivalent as doing the following with two histogram ?

for (bin b1 in Histogram1)
  for (bin b2: Histogram2)
        prob = b1.probability * b2.probability
        distance += prob * abs(b1.value - b2.value)
share|cite|improve this question
Cross-post (simultaneous):… – cardinal Jul 1 '11 at 4:37
up vote 4 down vote accepted

As you said, the density of $X-Y$ is the cross-correlation $h(t) = \int_{-\infty}^\infty f(t+y) g(y) \, dy$. The density of $|X-Y|$ is $f_{|X-Y|}(t) = h(t) + h(-t)$ for $t > 0$, $0$ for $t < 0$. The density of $(X-Y)^2$ is $f_{(X-Y)^2}(t) = (h(\sqrt{t}) + h(-\sqrt{t}))/(2 \sqrt{t})$ for $t > 0$, $0$ for $t < 0$.

share|cite|improve this answer

Expanding on Robert's answer (adapted to $Y-X$, in view of the question).

Let $h(t)=\int_{ - \infty }^\infty {f(x)g(t + x)dx}$, $t \in \mathbb{R}$, be the cross-correlation of the densities $f$ and $g$ (of $X$ and $Y$, respectively), that is the probability density function of $Y-X$. We want to show that, for any $t > 0$, $$ {\rm P}(|Y-X| \leq t) = \int_0^t {[h(u) + h( - u)]du} $$ (implying that the density of $|Y-X|$ is $h(t)+h(-t)$ for $t > 0$; for $t < 0$ the density is trivially $0$). Indeed, $$ \int_0^t {[h(u) + h( - u)]du} = \int_0^t {h(u)du} + \int_0^t {h( - u)du} = \int_0^t {h(u)du} + \int_{ - t}^0 {h(u)du} , $$ and so from $$ \int_0^t {h(u)du} = {\rm P}(0 \leq Y-X \leq t) \;\; {\rm and} \;\; \int_{ - t}^0 {h(u)du} = {\rm P}(-t \leq Y-X \leq 0), $$ it follows that $$ \int_0^t {[h(u) + h( - u)]du} = {\rm P}(-t \leq Y-X \leq t) = {\rm P}(|Y-X| \leq t). $$ Having shown that $h(t)+h(-t)$ is the density of $|Y-X|$ (for $t > 0$), the density of $(Y-X)^2$ can be found easily as follows. For any $t > 0$, $$ {\rm P}((Y-X)^2 \leq t) = {\rm P}(|Y-X| \leq \sqrt t) = \int_0^{\sqrt t } {[h(u) + h( - u)]du} . $$ A change of variable $u \mapsto u^2$ then gives $$ {\rm P}((Y-X)^2 \leq t) = \int_0^t {[h(\sqrt u ) + h( - \sqrt u )]\frac{{du}}{{2\sqrt u }}} , $$ implying that the density of $(Y-X)^2$ is $(h(\sqrt t)+h(-\sqrt t))/(2\sqrt t)$ for $t > 0$; for $t < 0$ the density is trivially $0$.

EDIT: A key point here is that (a nonnegative measurable function) $f_Z$ is a probability density function of $Z$ if and only if $F_Z (z) = \int_{ - \infty }^z {f_Z (} u)du$ $\forall z \in \mathbb{R}$, where $F_Z$ is the distribution function of $Z$. (In this case, $Z$ is said to be absolutely continuous with density function $f_Z$.)

EDIT: In the case of continuously differentiable distribution functions (rather than the general case absolutely continuous distribution functions), you can obtain the above results simply as follows. Let $H$ denote the distribution function of $Y-X$ and, as above, $h$ its probability density function. Then, for any $t > 0$, $$ {\rm P}(|Y-X| \leq t) = {\rm P}(-t \leq Y-X \leq t) = {\rm P}(Y-X \leq t) - {\rm P}(Y-X \leq -t) = H(t) - H(-t). $$ Hence $$ \frac{d}{{dt}}{\rm P}(|Y - X| \le t) = h(t) + h( - t). $$ As for the density function of $(Y-X)^2$, first note that, for any $t > 0$, $$ {\rm P}((Y-X)^2 \leq t) = {\rm P}(|Y-X| \leq \sqrt{t}) = H(\sqrt t ) - H( - \sqrt t ). $$ Hence $$ \frac{d}{{dt}}{\rm P}((Y - X)^2 \le t) = \frac{{h(\sqrt t )}}{{2\sqrt t }} + \frac{{h( - \sqrt t )}}{{2\sqrt t }} = \frac{{h(\sqrt t ) + h( - \sqrt t )}}{{2\sqrt t }}. $$

share|cite|improve this answer
is that the same as doing the average expected difference of two histogram like i do in the snippet of code in the questio? – skyde Jul 1 '11 at 14:45
If $X$ and $Y$ are independent random variables with densities $f$ and $g$, respectively, then ${\rm E}|Y - X| = \int {\int {|y - x|f(x)g(y)dxdy} } $. But you can approximate this expectation simply by using ${\rm E}|Y - X| \approx \frac{{\sum\nolimits_{i = 1}^n {|Y_i - X_i |} }}{n}$, for $n$ sufficiently large, where the $X_i$ and $Y_i$ are independent copies of $X$ and $Y$, respectively. – Shai Covo Jul 3 '11 at 8:05

It is a very big question.

Yes you can find the pdf of any combination of variables, there are many ways.

The one which works most of the time use the characteristic function:

You have many complex and exotic tools to exploit the properties of CF.

Anyhow this process greatly depends from the type of variable. E.g.: It is much easier with exponential variables than with non exponential variables.

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.