Distribution of a point farther away from a certain point

33 Views Asked by At

Related to this question,

Suppose that two points are independently distributed according to some CDF $F$ (and pdf $f$) over $[0,1]$.

What would be the distribution of the point farther away from a certain point $a$?

If I follow the derivation of the distribution of the closest point, let $f(x|a)$ denote the distribution of the point which is farther than the other point from $a$.

We have $$P[\textrm{the other point is inside of the interval }[a-|a-x|,a+|a-x|]]\\=F[a+|a-x|]-F[a-|a-x|].$$

There are two possible cases: point 1 is closer or point 2 is closer, so we multiply the probability by 2.

So, the distribution of the point farther away from $a$ is given by

$$2(F[a+|a-x|]-F[a-|a-x|])f(x).$$

I'm wondering whether this is a right derivation of the distribution or not.. I mechanically followed the derivation which is shown in the above link, but I don't understand why I need to multiply the probability to the baseline pdf $f$. Can anyone confirm the answer or explain the behind idea of the multiplication?

1

There are 1 best solutions below

0
On BEST ANSWER

It will become easier to understand when you master the common technique in deriving the joint pdf of order statistics. Let me call this a multinomial argument.

Some common results can be founded here:

https://en.wikipedia.org/wiki/Order_statistic#The_joint_distribution_of_the_order_statistics_of_an_absolutely_continuous_distribution

Now in your question,

Let $U_1, U_2$ be the original points which is unordered, and thus i.i.d. Assume they are absolutely continuous with common pdf $f_U$ and CDF $F_U$. Let $X$ be the point which is farther than the other point from $a$ and we want to derive its pdf $f_X$.

Consider the realization of $X$ being $x$. Next we partition the whole real line into $4$ sets accordingly:

The first set contains all the points which is closer to $a$ than $x$, that is $(a - |a - x|, a + |a - x|)$. And $$p_1 \triangleq \Pr\{U \in (a - |a - x|, a + |a - x|)\} = F_U(a + |a - x|) - F_U(a - |a - x|)$$

The second set is the point $x$. Since $U$ is absolutely continuous, the probability that it fall into an infinitesimal small interval $(x, x+dx)$ is $p_2 \triangleq f_U(x)dx$, and the probability that more than one point fall into the identical point is $0$.

The third set is $\{a - |a - x|\} \cup \{a + |a - x|\} \backslash \{x\}$, the point of reflection which has the identical distance to $a$ as $x$. Similarly the probability is $p_3 \triangleq f_U(x^*)dx$

The fourth set is the complement, which contains all the points which farther away: $(-\infty, a - |a - x|) \cup (a + |a - x|, +\infty)$. And $$ \begin{align} p_4 &\triangleq \Pr\{U \in (-\infty, a - |a - x|) \cup (a + |a - x|, +\infty)\} \\ &= 1 - F_U(a + |a - x|) + F_U(a - |a - x|) \end{align} $$

By multinomial argument, the number of points fall into these $4$ sets jointly follows a multinomial distribution. For $X$ to be the farther point, we require exactly $(1, 1, 0, 0)$ of points in the $4$ sets respectively. The probability is

$$ \frac {2!} {1!1!0!0!}p_1^1p_2^1p_3^0p_4^0 = 2[F_U(a + |a - x|) - F_U(a - |a - x|)]f_U(x)dx $$

Note that as $X$ is also continuous, $\Pr\{X \in (x, x+dx)\} = f_X(x)dx$ and this is equivalent to the above probability. Comparing both sides, by eliminating the $dx$ we obtain the desired result.

The good point of this argument is that it can easily extend to cases with general sample size $n$, and you can derive the joint pdf in general.