Let $X_{(1)}\geq X_{(2)}\geq\cdots X_{(K)}$ be the order statistics of the random variables $X_1,\,X_2,\,\ldots,X_K$, which are independent and identically distributed exponential random variables with parameter 1.

Now let $Z_k=\frac{Y_k}{X_{(k)}}$, where $Y_1,\,Y_2,\,\ldots,Y_K$ are also independent and identically distributed exponential random variables with parameter 1.

I wish to find the cumulative distribution function (CDF) of $\max\limits_{k}\{Z_k\}$, which is given by

$$\text{Pr}\left[\max\limits_{k}Z_k\leq z\right]$$

How can I find the above CDF?

The problem is that $\{Z_k\}_{k=1}^K$ are not independent, because the denominator is an order statistics. I tried to condition on the denominator and then average as

$$\begin{split} &\underbrace{\int\int\cdots\int}_{K-\text{Fold}}\text{Pr}\left[\frac{Y_1}{\alpha_1}\leq z, \frac{Y_2}{\alpha_2}\leq z,\cdots, \frac{Y_K}{\alpha_K}\leq z\right]f_{X_{(1)}, X_{(2)}, \ldots, X_{(K)}}(\alpha_1, \alpha_2, \ldots, \alpha_K)\,d\alpha_1\cdots d\alpha_K\\ =& \underbrace{\int\int\cdots\int}_{K-\text{Fold}} \prod_{k=1}^K \text{Pr}\left[\frac{Y_k}{\alpha_k}\leq z\right]f_{X_{(1)}, X_{(2)}, \ldots, X_{(K)}}(\alpha_1, \alpha_2, \ldots, \alpha_K)\,d\alpha_1\cdots d\alpha_K \end{split}$$

but got stuck in evaluating it.

Thanks in advance for any tips.


Define $\tilde{Z}_{k} = \frac{Y_{k}}{X_{k}}$ and $Z_{k} = \frac{Y_{k}}{X_{(k)}}$ with $Y_{k}$ and $X_{k}$ as in the question.

The random vectors $[\tilde{Z}_{1},\tilde{Z}_{2},...,\tilde{Z}_{K}]$ and $[Z_{1},Z_{2},...,Z_{K}]$ will have different distributions, however...

The distribution of the order statistics $\tilde{Z}_{(K)} = \max\limits_{k}\frac{Y_{k}}{X_{k}}$ and $Z_{(K)} = \max\limits_{k}\frac{Y_{k}}{X_{(k)}}$ will be equivalent since the ordering in the denominator is "undone" by introducing new random variables and taking the maximum.

So, to find the density of $Z_{(K)}$, it suffices to find the density of $\tilde{Z}_{(K)}$.

Distribution of $\tilde{Z}_{k}$

The CDF of the ratio of two exponential distributions with rate 1 is $F(z) = (1+z^{-1})^{-1}$, for $z>0$. See relevant question here.

Distribution of $\tilde{Z}_{(K)}$

$\tilde{Z}_{(K)}$ has CDF: $Pr(\tilde{Z}_{(K)} < z) =F(z)^{K}$ by rules for order statistics. So $Pr(\tilde{Z}_{(K)} < z) = (1+z^{-1})^{-K}$.

This is also the CDF of the maximum you are after.

Note: I'm using the standard definition of order statistics where $X_{(i)} \leq X_{(j)}$ for $i < j$.

  • $\begingroup$ Why the distribution of $\tilde{Z}_{(K)}$ and $Z_{(K)}$ are the same again? $\tilde{Z}_{(K)} = \max_k \tilde{Z}_{k}$, while $Z_{(K)} = \frac{Y_{i_K}}{X_{(K)}}$, where $i_K$ is the index of $X_{(K)}$. These are two different random variables, aren't they? I think your definition of $Z_k$ is not correct (I just realized that my definition isn't correct, too). It should be $Z_k=\frac{Y_{i_k}}{X_{(k)}}$. $\endgroup$ – BlackMath Dec 6 '18 at 7:17
  • $\begingroup$ I was defining $\tilde{Z}_{(K)} = \max_k \tilde{Z}_{k}$ and $Z_{(K)} = \max_k Z_{k}$ (to find what you asked for: $\text{Pr}\left[\max\limits_{k}Z_k\leq z\right]$). Right? $\endgroup$ – BonStats Dec 6 '18 at 11:17
  • $\begingroup$ $\tilde{Z}_{(K)}$ is correct, but $Z_{(K)}$ is not. $Z_{(K)}$ is defined as $$Z_{(K)} = \frac{Y_{i_K}}{X_{(K)}}$$ where $$i_K = \text{arg} \max_k X_k$$. $\endgroup$ – BlackMath Dec 6 '18 at 19:56
  • $\begingroup$ Maybe ask a new question then? Regardless of how you index the $Y_{i}$ they are still iid, hence I think that if $S = \frac{Y_{i_K}}{X_{(K)}}$ and $T = \frac{Y}{X_{(K)}}$, with $Y \sim \text{Exp}(1)$, $S$ and $T$ will have the same distribution. $\endgroup$ – BonStats Dec 7 '18 at 4:56
  • $\begingroup$ That's right, but $\frac{Y_{i_{K}}}{X_{(K)}}\neq \max_k \frac{Y_k}{X_{k}}$, does it? $\endgroup$ – BlackMath Dec 7 '18 at 6:15

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.