8
$\begingroup$

Consider a random variable $X\sim_X(x)$ with $\operatorname{supp}{p_X}=[a,b]\subset{\mathbb R}$, and let $X_1, X_2, \ldots, X_n$ be $n$ $i.i.d.$ samples from $p_X(x)$, then we have their associated order statistics $X_{(1:n)}, X_{(2:n)}, \ldots, X_{(n:n)}$ such that $X_{(1:n)}\leqslant X_{(2:n)}\leqslant \cdots\leqslant X_{(n:n)}$. Define the spacing between any two consecutive order statistics as $W_{(r:n)}=X_{(r+1:n)}-X_{(r:n)} (1\leqslant r\leqslant n-1)$, then in "Note on Francis Galton's problem" by K. Pearson (1902), it has been shown that $$ {\mathbb E}\left[W_{(r:n)}\right]=\frac{n!}{(n-r)!r!}\int_{-\infty}^{\infty}F(x)^{n-r}[1-F(x)]^r \, dx, \hspace{2mm} 1\leqslant r\leqslant n-1, $$ where $F(x)=\int_{-\infty}^xp_X(y) \, dy$ is the distribution function of $X$.

So, $\forall r, n$ such that $1\leqslant r\leqslant n-1$, $2\leqslant n$, do we have $$ {\mathbb E}\left[W_{(r:n)}\right] \geqslant {\mathbb E}\left[W_{(r:n+1)}\right], $$ i.e., the expected gap between any two consecutive order statistics will decrease or remain the same as $n$ increases?

Intuitively, this is correct since the more samples one draws, the denser the order statistics will be filled in the support. I also checked several specific distributions (e.g., uniform, Gaussian, etc), and the statement holds, but I have not figured out a good way to prove this for the general case. Hope someone can provide some inspiration.

$\endgroup$
  • $\begingroup$ Nice question. For counterexamples, I would look at the asymptotics when $n\to\infty$ through a Taylor expansion of $F$ around the point $x^*$ such that $F(x^*)=\frac12$. Differentiable densities $f$ such that $f(x^*)^2f'(x^*)$ is large might contradict the claim. $\endgroup$ – Did May 29 '13 at 7:09
  • $\begingroup$ Instead of saying "Let $X_1,\ldots,X_n$ be i.i.d. samples", you should say "Let $X_1,\ldots,X_n$ be an i.i.d. sample". The sample is the whole thing. $X_i$ is an observation, not a sample. $\endgroup$ – Michael Hardy May 29 '13 at 14:40
  • 1
    $\begingroup$ If the cumulative distribution function $F$ of the distribution of each observation $X_i$ is continuous, then $F(X_{(1:n)}), \ldots, F(X_{(n:n)})$ are the order statistics from the uniform distribution on the interval $(0,1)$. I'm not sure whether that can be readily exploited here. $\endgroup$ – Michael Hardy May 29 '13 at 14:48
  • $\begingroup$ Similar subsequent question. $\endgroup$ – BruceET Feb 5 '18 at 18:58
1
$\begingroup$

This is not true. Consider a Bernoulli distribution with $p=\frac12$. (Technically, this doesn't have support on an interval, but we can approximate it arbitrarily closely with support on an interval.)

Then with $S_n=\sum_{k=1}^nX_n$ we have

$$ \mathsf E\left[W_{(r:n)}\right]=\mathsf P\left(X_{(r:n)}=0\land X_{(r+1:n)}=1\right)=\mathsf P(S_n=n-r)=2^{-n}\binom nr\;. $$

But

$$ 2^{-(n+1)}\binom{n+1}r\gt2^{-n}\binom nr $$

whenever

$$ \frac{n+1}{n-r+1}\gt2\;, $$

that is, for all $r\gt\frac{n+1}2$.

In this case, the support cannot be filled in; increasing $n$ merely changes the probability that these particular order statistics are separated by the constant gap. It could be that this effect can't occur if the distribution is unimodal.

$\endgroup$

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.