Suppose $X_1,\ldots,X_n$ are i.i.d. standard normal. I'm wondering how to analyze the following quantity: $$\left|\frac{X_{(1)}X_{(n)}+X_{(2)}X_{(n-1)}+\cdots+X_{(n)}X_{(1)}}{n}\right|$$ where $X_{(1)}<X_{(2)}<\cdots<X_{(n)}$ is the order statistics. I suppose it should approach 1 as $n\to\infty$ but do not know how to justify my guess.

  • $\begingroup$ I think the polarization identity $X_{(1)}X_{(n)}+X_{(n)}X_{(1)} = \left(X_{(1)}+X_{(n)}\right)^2 - X_{(1)}^2-X_{(2)}^2$ might be helpful. $\endgroup$ – Lior Silberman Feb 3 '19 at 3:18

$\newcommand{\D}{\overset{\text{D}}=}$ Without loss of generality (wlog), $X_i=\Phi^{-1}(U_i)$ and hence $X_{(i)}=\Phi^{-1}(U_{(i)})$ for $i=1,\dots,n$, where $\Phi$ is the standard normal cdf and the $U_i$'s are iid random variables each uniformly distributed on $[0,1]$. In turn, wlog \begin{equation} U_{(i)}=\frac{S_i}{S_n}, \end{equation} where $S_i:=Y_1+\dots+Y_i$ and the $Y_i$'s are iid random variables each with the standard exponential distribution. So, \begin{equation} R_n:=\frac{X_{(1)}X_{(n)}+X_{(2)}X_{(n-1)}+\cdots+X_{(n)}X_{(1)}}{n} \D\sum_{i=1}^n\Phi^{-1}\Big(\frac{S_i}{S_n}\Big)\Phi^{-1}\Big(\frac{S_{n+1-i}}{S_n}\Big)\frac1n=:J_n, \end{equation} where $\D$ denotes the equality in distribution.

By the law of large numbers, $\frac{S_i}{S_n}\sim\frac in$ almost surely (a.s.) if $n\to\infty$ and $\frac in$ is bounded away from $0$. So, it is plausible that \begin{equation} J_n\to J:=\int_0^1\Phi^{-1}(u)\Phi^{-1}(1-u)\,du. \end{equation} Using the symmetry and the change $u=\Phi(x)$ of variables, we have \begin{equation} J=-\int_0^1\Phi^{-1}(u)^2\,du=-\int_{-\infty}^\infty x^2\,d\Phi(x)=-1. \end{equation} Thus, once "plausible" is replaced here by "proved", your guess that $|R_n|\to1$ (say, in probability) will be confirmed.

| cite | improve this answer | |
  • $\begingroup$ The "plausible" part seems to be crucial. How can we prove that? $\endgroup$ – neverevernever Feb 4 '19 at 1:41
  • $\begingroup$ I think I can do it, will try to find time for this. Can you let me know how this problem arises? $\endgroup$ – Iosif Pinelis Feb 4 '19 at 4:32
  • $\begingroup$ Sure. I'm considering how permuting coordinates of a random vector changes the direction of the original vector. So I start from the worst case, i.e. the permutation that matches the smallest to the largest, etc. $\endgroup$ – neverevernever Feb 4 '19 at 18:43

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.