0
$\begingroup$

Lets say that I have N observations that are poisson and i.i.d. The prior is an exponential with parameter 2. I know that the exponential distribution is given by

$ \lambda e^{(-\lambda x)} $

But how does it work with x when you multiply the prior times the likelihood to get the maximum a posteriori (MAP) estimate? In most places, they seem to just set $x=1$, but I don't understand why.

This would lead to the the MAP estimate to be

$ \operatorname{argmax} \quad \lambda^{\sum x_n}e^{-\lambda N}e^{-2\lambda} $

So my question is, how do you handle $x$ in the exponential distribution when it is a prior?

$\endgroup$
2
$\begingroup$

I'm not clear on the question, so I'm going to write some exposition and then finish my answer when you've clarified.

The Maximum A Posteriori estimate is, as you've said, is

$$ \hat{\lambda} = \underset{\lambda \in \mathbb{R}_+}{\mbox{argmax}} \left\{ \ell(\lambda; x) + \log(p(\lambda)) \right\} $$

Here, $\ell(\lambda;x)$ is the log likelihood and $p(\lambda)$ is the prior density. The prior for $\lambda$ depends on a different parameter which we will call $k$. The likelihood (poisson) is

$$ \dfrac{\lambda ^x e^{\lambda}}{\Gamma(x+1)}$$

The prior is (remember, $\lambda$ is the variable, $k$ is the parameter)

$$ k \exp(-k \lambda) $$

The log posterior is then (if my algebra is correct)

$$ N \bar{x}\log(\lambda) - N \lambda - \sum_i\log(\Gamma(x_i+1)) - \log(k) - k \lambda $$

From here, you can differentuate with respect to lambda and solve.

$$ 0 = \dfrac{N \bar{x}}{\lambda} - N -k \implies \lambda = \dfrac{N}{N+k} \bar{x} $$

Does your question concern $k$, the parameter in the prior?

$\endgroup$
2
  • $\begingroup$ Okay, so x (I guess that would correspond to your k) in my definition of the exponential function has nothing to do with any observations then? $\endgroup$ – user5744148 Mar 20 at 0:33
  • 1
    $\begingroup$ @user5744148 Correct, the prior density is not a function of the observed data. The name implies that it necessarily comes prior. $\endgroup$ – Demetri Pananos Mar 20 at 0:34

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.