Let's say $X_1,X_2,...,X_n$ are a sequence of i.i.d random variables from some unknown distribution. If we know that summation of them, $Y$ = $\sum_{i=1}^{n} X_i$, has exponential distribution with parameter $\lambda$, what is distribution of $X_i$?

  • $\begingroup$ Do you know the relationship between $\lambda$ ans $n$? Is this property valid for all $n$ or for one specific fixed one? $\endgroup$ – gt6989b Feb 22 '18 at 23:48
  • $\begingroup$ There is no specific relation between $n$, $\lambda$, but we can assume that we know value of $\lambda$. $n$ could be any arbitrary number.I can find m.g.f of $X_i$. Since it is: $g(X_i)$ = $\sqrt[n]{g(expnential)}$ $\endgroup$ – Yashar Feb 22 '18 at 23:56

Existence. A gamma distribution, denoted by $\operatorname{Gamma}(k, \lambda)$ for parameters $k, \lambda > 0$, is defined as a distribution supported on $[0,\infty)$ with the p.d.f.

$$ f(x) = \frac{\lambda^k x^{k-1} e^{-\lambda x}}{\Gamma(k)} \mathbf{1}_{(0,\infty)}(x). $$

Now if $X \sim \operatorname{Gamma}(k,\lambda)$ and $Y \sim \operatorname{Gamma}(l,\lambda)$ are independent, then $Z = X+Y$ has density $f_Z$ which satisfies

\begin{align*} \forall x > 0 \ : \quad f_Z(x) &= \int_{0}^{x} f_X(t)f_Y(x-t) \, dt \\ &= \int_{0}^{x} \frac{\lambda^k t^{k-1} e^{-\lambda t}}{\Gamma(k)} \cdot \frac{\lambda^l (x-t)^{l-1} e^{-\lambda (x-t)}}{\Gamma(l)} \, dt \\ &= \frac{\lambda^{k+l}e^{-\lambda x}}{\Gamma(k)\Gamma(l)} \int_{0}^{x} t^{k-1}(x-t)^{l-1} \, dt \\ &= \frac{\lambda^{k+l}x^{k+l-1}e^{-\lambda x}}{\Gamma(k+l)} \end{align*}

by the beta function identity. So $Z \sim \operatorname{Gamma}(k+l,\lambda)$. An immediate corollary is

Proposition. If $X_1, \cdots, X_n$ are independent and $X_i \sim \operatorname{Gamma}(k_i, \lambda)$ with $k_i, \lambda > 0$ for each $i$, then $\sum_{i=1}^{n} X_i \sim \operatorname{Gamma}(\sum_{i=1}^{n}k_i, \lambda)$.

Now by noting that $\operatorname{Exp}(\lambda) = \operatorname{Gamma}(1,\lambda)$, we find that $\operatorname{Gamma}(\frac{1}{n}, \lambda)$ is one possible choice.

Uniqueness. Notice that $Y$ has the m.g.f.

$$ \mathsf{E}[e^{sY}] = \frac{\lambda}{\lambda - s} $$

for $s < \lambda$. So it follows that

$$ \mathsf{E}[e^{sX_1}] = \left( \frac{\lambda}{\lambda - s} \right)^{1/n}. $$

Since the m.g.f. determines a distribution as long as it exists on a neighborhood of $0$, there is at most one choice of the distribution of $X_1$.


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.