2
$\begingroup$

Is there a way to calculate quantities of the form $$\mu_k = \mathbf{E}\bigg[ \frac{X_{(k)}}{\sum_{i=1}^n X_i} \bigg]$$ where the $X_i$'s are independent exponentially distributed random variables with mean $\lambda=1$ and $X_{(k)}$ denotes the $k$-th order statistic ?

$\endgroup$
2
$\begingroup$

Your question is actually interesting. If you consider Sukhatme's spacing, then you might get an answer. Since $X_1, \ldots, X_n$ are independent and identically-distributed random variables from $\mathrm{Exp(1)}$, then Sukhatme's spacing are simply $$ S_1 = n X_{(1)}, \quad S_i = (n - i + 1) \left[X_{(i)} - X_{(i - 1)}\right] $$ for $i = 2, \ldots, n$, such that $S_1, \ldots, S_n$ are independent and identically-distributed random variables from $\mathrm{Exp(1)}$, and $$ \sum^n_{i = 1} S_i = \sum^n_{i = 1} X_{(i)} = \sum^n_{i = 1} X_i. $$

Here is an attempt to answer your question, which might be correct or incorrect. In the case of $k = 1$, define $$ Z_1 = \frac{S_1}{\sum^n_{i = 1} S_i} = \frac{S_1}{S_1 + \sum^n_{i = 2} S_i} $$ Since $S_1$ follows $\mathrm{Gamma}(1, 1)$ and $\sum^n_{i = 2} S_i$ follows $\mathrm{Gamma}(n - 1, 1)$, then $Z_1$ follows $\mathrm{Beta}(1, n - 1)$. Therefore, $E(Z_1) = n^{-1}$; hence, $$ \frac{1}{n} = E(Z_1) = E\left(\frac{S_1}{\sum^n_{i = 1} S_i}\right) = n E\left(\frac{X_{(1)}}{\sum^n_{i = 1} X_i}\right) \Rightarrow E\left(\frac{X_{(1)}}{\sum^n_{i = 1} X_i}\right) = \frac{1}{n^2} $$ In the case of $k > 1$, define $$ Z_k = \frac{S_k}{\sum^n_{i = 1} S_i} = \frac{S_k}{S_k + \sum^n_{i = 1, i \neq k} S_i} $$ Clearly, $E(Z_k) = n^{-1}$. Accordingly, $$ n^{-1} = E(Z_k) = E\left(\frac{(n - k + 1) \left[X_{(k)} - X_{(k - 1)}\right]}{\sum^n_{i = 1} X_i}\right) \Rightarrow E\left(\frac{X_{(k)}}{\sum^n_{i = 1} X_i}\right) - E\left(\frac{X_{(k - 1)}}{\sum^n_{i = 1} X_i}\right) = \frac{1}{n (n - k + 1)} $$ Consequently, one can write the following equations $$ \begin{array}{l} E\left(\frac{X_{(k)}}{\sum^n_{i = 1} X_i}\right) - E\left(\frac{X_{(k - 1)}}{\sum^n_{i = 1} X_i}\right) = \frac{1}{n (n - k + 1)} \\ E\left(\frac{X_{(k - 1)}}{\sum^n_{i = 1} X_i}\right) - E\left(\frac{X_{(k - 2)}}{\sum^n_{i = 1} X_i}\right) = \frac{1}{n (n - (k - 1) + 1)} \\ \vdots \\ E\left(\frac{X_{(2)}}{\sum^n_{i = 1} X_i}\right) - E\left(\frac{X_{(1)}}{\sum^n_{i = 1} X_i}\right) = \frac{1}{n (n - 1)} \\ E\left(\frac{X_{(1)}}{\sum^n_{i = 1} X_i}\right) = \frac{1}{n^2} \\ \end{array} $$ By summing the above expressions, one gets $$ E\left(\frac{X_{(k)}}{\sum^n_{i = 1} X_i}\right) = \sum^{k}_{l = 1} \frac{1}{n (n - l + 1)} $$

References: P. V. Sukhatme (1937), Tests of significance for samples of the $\chi^2$ population with two degrees of freedom, Ann. Eugenics 8, 52-56.

$\endgroup$
  • $\begingroup$ Thank you ! This is consistent with the exact integrals I calculated using Mathematica for $n \leq 6$. The quantities are the expected values of a random probability vector chosen uniformly on the $n$-dimensional probability simplex. $\endgroup$ – Nocturne Mar 20 '17 at 12:25

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.