The expected number of events for a Poisson distribution under a Gamma prior with parameters $\alpha$ and $\beta$ (with mean $\alpha/\beta$) is:

$$ \newcommand{\paren}[1]{\left(#1\right)} \begin{aligned} \sum_{k=1}^\infty k & \cdot P(k \mid \alpha, \beta) \\ = \sum_{k=1}^\infty k & \int_0^\infty P(k \mid \lambda) p(\lambda \mid \alpha, \beta) \, d\lambda \\ = \sum_{k=1}^\infty k & \int_0^\infty \paren{\frac{\lambda^k e^{-\lambda}}{k!}} \paren{\frac{\beta^\alpha}{\Gamma(\alpha)} \lambda^{\alpha-1} e^{-\beta \lambda}} \> d\lambda \\ = \sum_{k=1}^\infty k & \> \frac{\beta^\alpha}{k! \> \Gamma(\alpha)} \int_0^\infty \lambda^{k+\alpha-1} e^{-(\beta+1)\lambda} \> d\lambda \\ = \sum_{k=1}^\infty k & \> \frac{\beta^\alpha}{(\beta+1)^{\alpha+k}} \frac{\Gamma(k+\alpha)}{\Gamma(\alpha) \Gamma(k+1)} \\ = \sum_{k=1}^\infty & \frac{\beta^\alpha}{(\beta+1)^{\alpha+k}} \frac{\Gamma(k+\alpha)}{\Gamma(\alpha) \Gamma(k)} \\ = \sum_{k=1}^\infty & \frac{\beta^\alpha}{(\beta+1)^{\alpha+k}} \frac{1}{B(\alpha, k)} \\ = \frac{\beta^\alpha}{(\beta+1)^\alpha} & \sum_{k=1}^\infty \frac{(\beta+1)^{-k}}{B(\alpha, k)} \\ = \paren{\frac{\beta}{\beta+1}}^\alpha & \sum_{k=1}^\infty \frac{(\beta+1)^{-k}}{B(\alpha, k)} \end{aligned} $$ where $\Gamma$ is the Gamma function and $B$ is the Beta function.

Is there any way to further simplify this expression?

Intuitively, I would expect it to be at or near the expected number of events for a Poission distribution with mean equal to the expected $\lambda$ from the prior,

$$ \begin{aligned} & \mathbb{E}_k \text{Poisson}\left(\lambda = \mathbb{E}_\lambda(\text{Gamma}(\alpha, \beta))\right) \\ = {} & \> \mathbb{E}_k \text{Poisson}\left(\lambda = \alpha/\beta\right) \\ = {} & \> \alpha/\beta \end{aligned} $$


I found a description of the Gamma-Poisson distribution, which confirms my intuition that the expected value is in fact $\alpha/\beta$.

  • 1
    $\begingroup$ Hint: what is the posterior predictive distribution of the Poisson-Gamma model? I think you're making this a bit more difficult by cranking directly at the definition of expectation. $\endgroup$ – Sycorax Mar 13 '17 at 18:27
  • 1
    $\begingroup$ Try to use double expectation formula $\DeclareMathOperator{\E}{\mathbb{E}} \E \E (X \mid Y) = \E X$ $\endgroup$ – kjetil b halvorsen Mar 13 '17 at 18:28
  • 1
    $\begingroup$ stats.stackexchange.com/questions/163236/… $\endgroup$ – Sycorax Mar 13 '17 at 19:57

From the link given by Sycorax above it was pointed out that

$$ \frac{d^k \beta^\alpha}{(d+\beta)^{k+\alpha}} \frac{\Gamma(k+\alpha)}{\Gamma(k+1)\Gamma(\alpha)} = \binom{k+\alpha-1}{\alpha-1} (1-p)^\alpha p^k $$

is the PDF of a Negative Binomial Distribution with parameters $(\alpha,p=d/(d+\beta))$.

We can insert this halfway through the derivation in my original question (with $d = 1$) to get:

$$ \sum_{k=0}^\infty k \cdot P(k \mid \alpha, 1/(1+\beta)) = \mathbb{E}\left[k\right] = \alpha/\beta $$

Very cool!

  • 1
    $\begingroup$ This answer would be more clear if you cleaned up the notation to match that of your question. For example, you don't have $d$ in the question. $\endgroup$ – Sycorax Mar 14 '17 at 21:34
  • $\begingroup$ edited to make more clear $\endgroup$ – Mageek Mar 14 '17 at 21:54

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.