Let $X_1,...,X_n$ be a iid random sample having pdf $\theta x^{\theta-1}1(0 < x \le 1)$

Find a sufficient statistic $Y$ for $\theta$ then find the Bayes estimator $w(Y)$ based on this statistic using the loss function $L(\theta,a) = (a-\theta)^2$ where the prior distribution is exponential with mean $\frac{1}{\beta}$.

First sufficiency:

The likelihood function is $\displaystyle L(\theta) = \Pi_{i = 1}^n\theta x_i^{\theta -1} = \theta^n(x_1\cdots x_n)^\theta(x_1\cdots x_n)^{-1}$ thus by the factorization theorem we can take $Y = (x_1\cdots x_n)^{-1}$.

Bayes Estimator:

For square-error loss the estimator $w(Y) = \hat{\theta} = E[\theta \mid Y\,]$ i.e. the mean of the posterior.

For the posterior need to first solve $m(y) = \displaystyle \int_0^\infty \beta e^{-\beta \theta}y^{1-\theta}d\theta$ Is this a well known integral? I was trying to solve by u-substitution but I am making a mistake somewhere. I am trying $u = y^{-\theta}, du = -y^{-\theta}\log(y)d\theta$ but for some reason I can't see how to take care of $e^{-\beta\theta}$.

Before continuing would appreciate to know if this is correct:

EDIT: $y^{-\theta} = e^{-\theta \log(y)}$ so re-write as $\displaystyle \beta y \int_{0^\infty}e^{-\theta(\beta + \log(y))}d\theta$ and set $u = -\theta(\beta + \log(y)) $

Then we will have $\displaystyle -\frac{\beta y}{\beta + y}e^{-\theta(\beta + \log(y))} \bigg \vert_{\theta = 0}^\infty= \frac{\beta y}{\beta + y}$

Would still like to know if this a well known integral.

Now the next step is to solve $\displaystyle E[\theta \mid Y] = \int_0^\infty \theta \frac{y+\beta}{\beta y}\beta e^{-\beta \theta}\theta^ny^{1-\theta}d\theta$ correct? and this will give use the estimator we seek.


Using factorization theorem the sufficient statistic for $\theta$ is $y=\prod_i X_i$. This because the function $g(\theta,t(\mathbf{x}))$ depends on the data only through the statistic "t=product".

The function $\frac{1}{\prod_{i}X_{i}}$ you wrongly identfied as the sufficient statistic is the function of "x alone".

Then the posterior is the following (hint: when calculating the posterior discard any quantity that does not depend on $\theta$)

$$\pi(\theta|y) \propto e^{-\beta \theta}\theta^n y^{\theta-1}$$

$$\propto e^{-\beta \theta}\theta^n e^{(\theta-1) log y}$$

$$\propto \theta^ne^{-(\beta-logy)\theta}$$

...we immediately recognize in this posterior a Gamma distribution...

now you can kill the problem by yourself without solving the integral analitically

| cite | improve this answer | |
  • $\begingroup$ The expectation will then give $\displaystyle \frac{n+1}{\beta-\log(Y)}$ correct? Based off of being a Gamma distribution. $\endgroup$ – oliverjones Aug 19 at 23:20
  • 1
    $\begingroup$ @oliverjones : correct! $\endgroup$ – tommik Aug 20 at 6:27

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.