3
$\begingroup$

Suppose that $X_1, \ldots, X_n$ are mutlivariate normal with equal correlation $\rho$ and each of them are marginally distributed as $N(0,1)$. Let $X_{(1)}, \ldots, X_{(n)}$ be the corresponding order statistics. The distribution of $X_{(1)}$ and $X_{(n)}$ are easily found. What about the distribution of the other order statistics? Can anyone give some reference on this?

Thank you. Hanna

$\endgroup$
  • $\begingroup$ C.W.Helstrom's book Statistical Theory of Signal Detection (1967?) has some discussion of the equal correlation case and $X_{(n)}$ though not of the order statistics as I recall. Just remember that for any (not necessarily multivariate normal) equicorrelated equivariance random variables, $$\operatorname{var}\left(\sum_{i=1}^n X_i\right)=n\sigma^2+n(n-1)\rho\sigma^2 \geq 0$$ shows that $\rho \geq -\frac{1}{n-1}$. $\endgroup$ – Dilip Sarwate Mar 27 '13 at 2:38
2
$\begingroup$

As shown in Tong, Y. L. (1990). Multivariate normal distribution. Springer-Verlag., ch. 6, for the setup described in the question and for non-negative correlation coefficient $\rho\in [0,\;1)$, the distribution function (cdf) and the density of an order statistic $X_{(i)}$ are (where $\phi()$ and $\Phi()$ are the standard normal pdf and cdf)

$$G_{(i)}(x) = \int_{-\infty}^{\infty}F_{(i)}\left(\frac{x+\sqrt{\rho}z}{\sqrt{1-\rho}}\right)\phi(z)dz$$

and differentiating,

$$g_{(i)}(x) = \int_{-\infty}^{\infty}\frac 1{\sqrt{1-\rho}}f_{(i)}\left(\frac{x+\sqrt{\rho}z}{\sqrt{1-\rho}}\right)\phi(z)dz$$

where

$$f_{(i)}(y) = \frac{n!}{(i-1)!(n-i)!}[\Phi(y)]^{i-1}[\Phi(-y)]^{n-i}\phi(y)$$ and $$F_{(i)}(y) = \sum_{j=i}^n {n \choose j}[\Phi(y)]^{j}[\Phi(-y)]^{n-j}$$

i.e $f_{(i)}(y)$ and $F_{(i)}(y)$ are the pdf and cdf of the order statistic $(i)$ from an i.i.d. standard normal random sample.

For the corresponding results when the correlation coefficient is negative, the author refers to the book "Order Statistics", by H.A. David & H.N. Nagaraja ch. 5 (now in its 3d edition, 2003).

$\endgroup$
  • $\begingroup$ Does $G_{(i)}(x)$ or $g_{(i)}(x)$ simplify to a different form if $i=n$ or $i=1$? $\endgroup$ – Frey Apr 24 '17 at 0:30
  • $\begingroup$ @Frey Obviously the expressions are simplified to a degree, but I am not aware of "closed-form expressions" even in these cases (we are talking about correlated variable here), although the OP argues that "they are easily found". $\endgroup$ – Alecos Papadopoulos Apr 25 '17 at 1:30

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.