Let $X_1,X_2,\ldots,X_n$ be i.i.d variables having a Pareto distribution with density $$f(x)=\frac{a\theta^a}{x^{a+1}}1_{x>\theta}\,,$$ where $a,\theta>0$. What is the distribution of $\sum\limits_{j=1}^n \ln\left(\frac{X_{(j)}}{X_{(1)}}\right)$ ?

Suppose $\mathsf{Gamma}(p,\alpha)$ denotes the density $g(t)\propto e^{-\alpha t}t^{p-1}1_{t>0}$.

We have $$T=\sum_{j=1}^n\ln\left(\frac{X_{(j)}}{X_{(1)}}\right)=\sum_{j=1}^n\ln(X_{(j)})-n\ln(X_{(1)})=\sum_{j=1}^n \ln X_j-n\ln X_{(1)}$$


\begin{align}&\ln(X_j/\theta)\stackrel{\text{i.i.d}}{\sim}\mathsf{Exp}\text{ with mean }1/a\qquad,\,j=1,\ldots,n \\&\implies\sum_{j=1}^n\ln(X_j/\theta)=\sum_{j=1}^n\ln X_j-n\ln \theta\sim\mathsf{Gamma}(n,a) \end{align}

I could show that $X_{(1)}$ has another Pareto density, so that $$\ln\left(\frac{X_{(1)}}{\theta}\right)=\ln X_{(1)}-\ln \theta \sim \mathsf{Exp}\text{ with mean }1/(na)$$

Not sure if the last two facts help me get the exact distribution of $T$.


Turns out this was rather simple had I simply rewritten $T$ as

$$T=\sum_{j=1}^n \ln\left(\frac{X_j}{X_{(1)}}\right)=\sum_{j=1}^n(\ln X_j-\ln X_{(1)})=\sum_{j=1}^n (Y_j-Y_{(1)})\,,$$

where $Y_j=\ln (X_j/\theta)$. Since $aY_j\sim\mathsf{Exp}(1)$, using this result I have $a T\sim \mathsf{Gamma}(n-1,1)$.

This is equivalent to $T\sim \mathsf{Gamma}(n-1,a)$ or $2aT\sim \chi^2_{2n-2}$.

  • 1
    $\begingroup$ The first approach ignores the correlation between $\sum_{j=1}^n\ln X_j$ and $\ln X_{(1)}$ $\endgroup$ – Xi'an May 12 '18 at 20:51
  • $\begingroup$ Yes, I noticed that they are not independent. $\endgroup$ – StubbornAtom May 12 '18 at 20:55

A simpler approach might be to use the fact that if $x \sim \text{Pareto}(\theta,a)$, then conditioning upon $x \geq b$ results in $x \sim \text{Pareto}(b,a)$. Consequently, $x | x_{(1)} \sim \text{Pareto}(x_{(1)}, a)$, except for the single observation corresponding to $x_{(1)}$. When we then take the ratio $x/x_{(1)}$, we are rescaling $x$ by its minimum value, and the resulting variate has a $\text{Pareto}(1, a)$ distribution, independent of $x_{(1)}$.
Therefore, if we don't pay attention to the rank of the $x_i$ in the sample, the ratios $x_i/x_{(1)} \sim \text{Pareto}(1,a)$ and are independent (except for the observation corresponding to $x_{(1)}$, which is equal to 1.)

This, combined with the fact that the log of a $\text{Pareto}(1,a)$ variate is distributed $\text{Exponential}(a)$, and the sum of $n-1$ i.i.d. variates $\sim \text{Exponential}(a)$ is $\sim \text{Gamma}(n-1,a)$, leads directly to the result that the sum

$$\sum_{j=1}^n \ln\left(\frac{X_{(j)}}{X_{(1)}}\right) \sim \text{Gamma}(n-1,a)$$

where the $n-1$ comes from the fact that exactly one of the ratios will have value $1$, hence $\log(\cdot) = 0$, leaving $n-1$ nonzero terms in the sum.

| cite | improve this answer | |
  • $\begingroup$ I think you missed saying $x_{(1)}$ is independent of $x_i/x_{(1)}$ in the last sentence of the first paragraph. $\endgroup$ – StubbornAtom May 13 '18 at 6:05
  • 2
    $\begingroup$ That's implied by the point of the next to last sentence, admittedly not nearly as clearly stated as it might have been - the ratio and $x_{(1)}$ are independent. $\endgroup$ – jbowman May 13 '18 at 13:57

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.