3
$\begingroup$

I am fighting to understand a perhaps simple question, described below.

Let $X$ be a positive continuous random variable. Now, define another random variable $$ Y = \begin{cases} \text{undefined} & \text{ if } X \leq d \\ X - d & \text{ if } X > d \end{cases} \> . $$

Now my job is to calculate the pdf of $Y$.

My friend calculates: $$ \frac{f(y+d)}{1 - F(d)} ,\quad y > 0, $$ where, $f$ and $F$ are the pdf and cdf of $X$, respectively.

What is bothering me is: Should it not be, instead, $$ \frac{f(y+d)}{1 - F(d)} ,\quad y > d \> ? $$

What I am missing here? I would really appreciate if somebody explain me in some easy way.

$\endgroup$

3 Answers 3

4
$\begingroup$

You also have to define $Y$ when $X\leq d$ otherwise $Y$ is not a well-defined random variable. For example set $Y=\infty$ when $X\leq d$. Whatever your choice is, $Y$ no longer has a density because it has an atom at $\infty$ (or at $a$ if you set $Y=a$ when $X\leq d$, whatever the choice of $a$). You can write the distribution of $Y$ as a weighted sum of a Dirac distribution at $\infty$ and a continuous distribution on $[0, +\infty[$ : the distribution of $Y$ is $$p\delta_\infty + (1-p) g(y)dy$$ where $p=\Pr(X\leq d)=F(d)$, $\delta_\infty$ is the Dirac mass at $\infty$ and $g(y)dy=\Pr(Y \in dy \mid X>d)=\Pr(X-d \in dy)/\Pr(X>d)=\frac{f(y+d)}{1-F(d)}dy$.

$\endgroup$
4
  • 1
    $\begingroup$ Wellcome, Stéphane. There are some osbcure points in your answer, I think by $Y \in dy$ you mean $Y \in [y, y+ dy]$. Moreover $g(y)$ have to be null when $y < 0$. Finally, there is a simplification between $1-p$ and $1-F(d)$, which should be explicited... $\endgroup$
    – Elvis
    Commented Jan 8, 2012 at 13:58
  • 1
    $\begingroup$ With $Y = \infty$ if $X \leq d$, the cumulative distribution function $F_Y(y)$ does not have the property $$\lim_{y \to \infty} F_Y(y) = 1$$ which is commonly attributed to distribution functions. Also, $Y$ does not have an expectation or any other moments etc. So, is it worthwhile using $\infty$ as an example of a possible definition of the value of $Y$ when $X \leq d$ when a finite value (as Elvis has used) is so much more convenient (and has common properties that $Y = \infty$ does not share)? $\endgroup$ Commented Jan 8, 2012 at 15:43
  • $\begingroup$ Yes, your comments about the choice $Y=\infty$ make sense. But I really use the notation $Y\in dy$, see my answer in this post stats.stackexchange.com/questions/18222/… $\endgroup$ Commented Jan 8, 2012 at 16:52
  • $\begingroup$ Is this notation used anywhere else? $\endgroup$
    – Elvis
    Commented Jan 8, 2012 at 18:14
5
$\begingroup$

Stéphane is right to say that $Y$ cannot stay undefined, you have to give a value to $Y(\omega)$ for all $\omega\in\Omega$.

Stéphane’s choice of $Y = \infty$ when $X \le d$ is a bit difficult to handle even if it is natural. Let’s take $Y = y_0\in\mathbb R$ when $X \le d$. You have, $$\begin{align*} \mathbb P( Y \le y ) &= \mathbb P( y_0 \le y \cap X \le d \cup X-d \le y \cap X > d ) \\ &= \mathbb P( y_0 \le y \cap X \le d) + \mathbb P(d < X \le y+d) \\ &= \mathbb P(X\le d) \cdot 1_{y_0\le y} + \left\{\begin{array}{l} 0 \text{ if } y < 0 \\ F(y+d)-F(d) \text{ if } y \ge 0. \end{array}\right. \end{align*}$$ This cdf has no derivative in $y = y_0$, hence strictly speaking no density.

However defining $g(y)$ by $$g_0(y) =\left\{\begin{array}{l} 0 \text{ if } y < 0 \\ f(y+d) \text{ if } y \ge 0, \end{array}\right. $$ (the derivative of the pdf "ignoring" the step in $y_0$) and $p = \mathbb P(X>d) = 1 - F(d)$, we can write the density of $Y$ by the mixture $$g(y) = g_0(y) + (1-p) \cdot \delta(y-y0)$$ where $\delta$ is the Dirac function.

Note that the integral $\int_{-\infty}^\infty g_0(y) \mathrm dy$ is equal to $p$ and by convention the integral of the $\delta$ function is 1, so $\int_{-\infty}^\infty g(y) \mathrm dy = 1$.

Edit 1 you can chose $y_0 = +\infty$ is you wish, in that case there is a Dirac mass in $+\infty$. I don’t know if there is some standard notation for this.

Edit 2 I suspect that the original question was to give the density of $Y$ conditional to $X>d$.

In that case you don’t need bother with the values taken by $Y$ when this condition is not fulfilled. You can write

$$\begin{array}{rcl} \mathbb P( Y \le y | X > d) &=& { \mathbb P(X \le y+d \cap X > d) \over \mathbb P(X>d)} \\ &=& \left\{\begin{array}{l} 0 \text{ if } y < 0 \\ {F(y+d)-F(d)\over 1 - F(d)} \text{ if } y \ge 0, \end{array}\right. \end{array}$$ and the density of $Y$ conditional to $X>d$ is $$g(y|X>d) = \left\{\begin{array}{l} 0 \text{ if } y < 0 \\ {1\over 1-p}f(y+d) \text{ if } y \ge 0. \end{array}\right. $$

As whuber pointed out in the comments, a natural interpretation of this is to change the sample space $\Omega$ to $\Omega' = \{ \omega \in \Omega \>:\> X(\omega) > d \}$.

$\endgroup$
7
  • $\begingroup$ I agree that $Y = \infty$ as Stéphane has chosen is a little trickier to handle. Also, some people like to define random variables as mappings from the sample space to the real line instead of the extended real line, and in that sense, Stéphane's $Y$ is not a random variable at all. While arbitrary $y_0$ as you have chosen has no problems, a very natural choice is $Y = 0$ if $X \leq d$ and $Y = X-d$ if $X > d$ so that $Y$ takes on values in $[0,\infty)$. $\endgroup$ Commented Jan 8, 2012 at 13:51
  • $\begingroup$ Yes, my opinion is that Bogaso forgot to write a part of the question... my intuition is that it is "the pdf of $Y$ conditional to $X > d$". $\endgroup$
    – Elvis
    Commented Jan 8, 2012 at 13:56
  • $\begingroup$ Another approach is to view $Y$ as a random variable on a different but related sample space, $\Omega'=\{\omega\in\Omega\ |\ X(\omega)\gt d\}$. (This would be a good model for rejection sampling for instance.) This avoids all the considerations mentioned here and in Stephane's reply, which are incidental and distracting anyway. $\endgroup$
    – whuber
    Commented Jan 8, 2012 at 16:55
  • 1
    $\begingroup$ @whuber Thanks, I added a second edit... for me this is the pdf conditional to $X>d$! $\endgroup$
    – Elvis
    Commented Jan 8, 2012 at 18:35
  • $\begingroup$ Edit 2 clearly answers the question (+1). $\endgroup$
    – whuber
    Commented Jan 8, 2012 at 22:34
3
$\begingroup$

A simple explanation for why your friend is correct - $y > 0$ - is to see what happens to $y$ as $x \to d$ from above. Since $y = x - d$, as $x \to d$, by subtracting $d$ from both sides of $x \to d$, we get $x - d \to 0$. Since $y = x - d$, $y$ can get arbitrarily close to $0$, so the lower bound on $y$ is 0, not $d$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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