The univariate Jeffreys prior is defined as any [[prior]] proportional to the square root of the [[Fisher information]] number
$\pi(\theta) \propto \sqrt{I(\theta)}$
Jeffreys prior is invariant to reparameterization and a response to critiques of the [[principle of indifference]]. Jeffreys prior is a type of [[objective prior]] and can be an [[improper prior]].
To find Jeffreys prior
- Find a [[likelihood]] for $\theta$
- Find the log-likelihood
- Then find the [[Fisher information]] by either
- find the first derivative with respect to $\theta$ and square its [[expectation]] with respect to the data
- find the second derivative with respect to $\theta$ and negate its expectation (the second derivative is often easier)
- Finally find a prior that is proportional to the square root of the Fisher information
For multiple [[independent and identically distributed|iid]] data points $Y = Y_1, \dots, Y_n$, the Fisher information is $n$ times the marginal information.
$I(\lambda) = -nE \Big[ \frac{\partial^2}{\partial^2 \theta} \ln f(y_i|\lambda) \Big]$
For example, let $Y | \lambda \sim Poisson(\lambda)$. A likelihood for the [[Poisson distribution]] can be expressed as
$
f(y|\lambda) = \frac{e^{-\lambda}\lambda^y}{y!} \propto e^{-\lambda}\lambda^y
$
The log-likelihood is
$\ln f(y|\lambda) = -\lambda + y \ln \lambda$
The first derivative is
$\frac{\partial}{\partial \lambda} \ln f(y|\lambda) = -1 + \frac{y}{\lambda}$ The second derivative is
$\frac{\partial^2}{\partial^2 \lambda} \ln f(y|\lambda) = -\frac{y}{\lambda^2}$
The expectation of the Poisson distribution is $E[y] = \lambda$. The Fisher information is the negative of the expectation
$I(\lambda) = -E \Big[ \frac{\partial^2}{\partial^2 \theta} \ln f(y|\lambda) \Big] = -E \Big[- \frac{y}{\lambda^2} \Big] = \frac{E[y]}{\lambda^2} = \frac{\lambda}{\lambda^2} = \frac{1}{\lambda}$
Note that in the expectation the parameter $\lambda$ is considered a constant, the expectation is with respect to the data $y$.
Thus, the Jeffreys prior for a Poisson distribution is
$\pi(\lambda) \propto \sqrt{I(\lambda)} = \frac{1}{\lambda^2}$