15
$\begingroup$

I have a sequence $T_1,T_2,\ldots$ of independent exponential random variables with paramter $\lambda$. I take the sum $S=\sum_{i=1}^n T_i$ and now I would like to calculate the probability density function.

Well, I know that $P(T_i>t)=e^{-\lambda t}$ and therefore $f_{T_i}(t)=\lambda e^{-\lambda t}$ so I need to find $P(T_1+\cdots+T_n>t)$ and take the derivative. But I cannot expand the probability term, you have any ideas?

$\endgroup$
  • $\begingroup$ it's called Erlang distribution $\endgroup$ – Alex Jan 29 '14 at 0:14
  • $\begingroup$ $P(T_1+...+T_n>t)$ is $1-F_S(t)$ i.e. it relates to the cumulative distribution function, not to the density. Different degrees of computational difficulty. $\endgroup$ – Alecos Papadopoulos Jan 29 '14 at 0:26
26
$\begingroup$

The usual way to do this is to consider the moment generating function, noting that if $S = \sum_{i=1}^n X_i$ is the sum of IID random variables $X_i$, each with MGF $M_X(t)$, then the MGF of $S$ is $M_S(t) = (M_X(t))^n$. Applied to the exponential distribution, we can get the gamma distribution as a result.

If you don't go the MGF route, then you can prove it by induction, using the simple case of the sum of the sum of a gamma random variable and an exponential random variable with the same rate parameter. Let's actually do this. Suppose $Y \sim {\rm Gamma}(a,b)$ and $X \sim {\rm Exponential}(b)$ are independent, so that $$f_Y(y) = \frac{b^a y^{a-1} e^{-by}}{\Gamma(a)} \mathbb 1(y > 0), \quad f_X(x) = be^{-bx} \mathbb 1(x > 0), \quad a, b > 0.$$ Then, we notice that if $a = 1$, $Y$ would also be exponential (i.e., the exponential distribution is a special case of the Gamma with $a = 1$). Now consider $Z = X+Y$. The PDF is $$\begin{align*} f_Z(z) &= \int_{y=0}^z f_Y(y) f_X(z-y) \, dy \\ &= \int_{y=0}^z \frac{b^{a+1} y^{a-1} e^{-by} e^{-b(z-y)}}{\Gamma(a)} \, dy \\ &= \frac{b^{a+1} e^{-bz}}{\Gamma(a)} \int_{y=0}^z y^{a-1} \, dy \\ &= \frac{b^{a+1} e^{-bz}}{\Gamma(a)} \cdot \frac{z^a}{a} = \frac{b^{a+1} z^a e^{-bz}}{\Gamma(a+1)}. \end{align*}$$ But this is just a gamma PDF with new shape parameter $a^* = a+1$. So, it is easy to see by induction that the sum of $n$ IID exponential variables with common rate parameter $\lambda$ is gamma with shape parameter $a = n$, and rate parameter $b = \lambda$.

$\endgroup$
  • $\begingroup$ Thanks for your answer, according to your first part concerning the MGF, I have some difficuties with this, the MGF for each of the $T_i$ is $M_T(x)=(1-\frac{x}{\lambda})^{-1}$, therefore $M_S(x)=(1-\frac{x}{\lambda})^{-n}$, but how to derive $f_S(x)$ now? $\endgroup$ – TI Jones Jan 29 '14 at 21:01
  • $\begingroup$ Take the PDF of a gamma distribution, and calculate its MGF. When you see it has the same form, it follows from the uniqueness of MGFs that the sum of exponential RVs is therefore gamma distributed. $\endgroup$ – heropup Jan 29 '14 at 21:10
  • $\begingroup$ @heropup Could you please let me know how to find the PDF of the sum of $n$ (independent non-identically distributed) Exponential random variables? $\endgroup$ – sky-light Sep 7 '16 at 21:49

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.