I am struggling with this problem:

Let $Y \sim gamma(k, \theta)$ with $k\in \mathbb{N}$, that is $Y$ is a continuous random variable with density

$ f(x) = \frac{\theta^k }{(k-1)!} x^{k-1} e^{-\theta x}$

Given $Y$, let $X_1, \dots , X_n \sim Poiss(Y)$ be independent and identically distributed random variables. Find the distirbution of $Y$ given $X_1, \dots , X_n$.

I belive that means that if we have $Y = \lambda$ then $f_{X_1, \dots , X_n |Y}(k_1, \dots, k_n) = \prod_i \frac{\lambda^k_i}{k_i!} e^{-\lambda}$ . Correct?

How do I find $f_{Y | X_1, \dots , X_n}$ ?


Your first step is correct, namely that


Then, apply Bayes' Theorem. In the discrete case, Bayes' Theorem reads

$$\Bbb P(A|B) \Bbb P (B)=\Bbb P(B|A) \Bbb P (A)$$

In the continuous case as we have now, this simply becomes


Thus, applying it to the problem at hand, we get

$$f_{Y|X_1=x_1,\dots ,X_n=x_n}(y)=\frac{f_{X_1,\dots , X_n|Y=y}(x_1,\dots ,x_n)f_Y(y)}{f_{X_1,\dots ,X_n}(x_1,\dots ,x_n)}$$

Now, we know how to calculate the numerator on the RHS, but the problem is the denominator - we do not know the joint distribution of $X_1,\dots,X_n$ without being given information about $Y$.

But, observe that the object $f_{Y|X_1=x_1,\dots ,X_n=x_n}(y)$ we are trying to compute is a function of ONLY $y$ and is a probability density function. $f_{X_1,\dots ,X_n}(x_1,\dots ,x_n)$ is independent of $y$. Thus, we can compute this thing in proportionality, and the normalize at the end to obtain a PDF. So

\begin{align} \ f_{Y|X_1=x_1,\dots ,X_n=x_n}(y) & =\frac{f_{X_1,\dots , X_n|Y=y}(x_1,\dots ,x_n)f_Y(y)}{f_{X_1,\dots ,X_n}(x_1,\dots ,x_n)} \\ \ & \propto f_{X_1,\dots , X_n|Y=y}(x_1,\dots ,x_n)f_Y(y) \\ \ & = \biggl( \prod_{i=1}^n\frac{e^{-y}y^{x_i}}{x_i!} \biggl )\biggl (\frac{\theta ^ k}{(k-1)!}y^{k-1}e^{-\theta y} \biggl) \\ \ & \propto \biggl( \prod_{i=1}^n e^{-y}y^{x_i} \biggl )\biggl (y^{k-1}e^{-\theta y} \biggl) \\ \ & = y^{\sum_{i=1}^n x_i+k-1}e^{-(n+\theta)y} \end{align}

Observe that this is the PDF for a Gamma$ \bigl(\sum_{i=1}^n x_i+k \;,\;n+\theta \bigl)$ distribution, so in normalizing, we get

$$f_{Y|X_1=x_1,\dots ,X_n=x_n}(y) = \frac{(n+\theta)^{\sum_{i=1}^n x_i+k}}{\Gamma \bigl(\sum_{i=1}^n x_i+k \bigl)}y^{\sum_{i=1}^n x_i+k-1}e^{-(n+\theta)y}$$


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.