0
$\begingroup$

I am very new to Bayesian statistics and I wanted to know how could I show that Uniform($0,A$) ,as $A \to \infty$, is an improper density?

I mean the Uniform($0,A$) ,as $A \to \infty$ density is given by $ \lim_{A \to \infty}f(x)=\frac{1}{A} \mathbb{1}_{[0,A]}=0$ and hence it doesnt integrate to 1 and therefore improper based on the definition I found here Does "improper" posterior or prior refer to a density function that does not integrate to 1 or to one that does not integrate to a finite value?

Am I wrong?

$\endgroup$

1 Answer 1

3
$\begingroup$

Basically, yes.

More formally, what you want to know is: what is the value of

$$ \int_{-∞}^∞ \lim_{A \to ∞} \tfrac{1}{A}1_{[0, A]}(x) \; dx \;\;\; ?$$

For all $x$, the limit converges to 0. To see this, notice that for any $x$ and $A$, $\tfrac{1}{A}1_{[0, A]}(x)$ is either $0$ or $\frac{1}{A}$, the latter of which is less than any $ε > 0$ for all $A > \frac{1}{ε}$. Then we have just $\int_{-∞}^∞ 0 \; dx$, which is $0$.

Because the function's integral is $0$, it isn't a proper density.

$\endgroup$
7
  • $\begingroup$ Hi if you read from the link I attached , I see that any finite number is considered to be proper but somehow i gather that this finite number should be non zero for it to be proper since we can scale it to make it equal to 1 . Right ? $\endgroup$ Jun 24, 2017 at 15:40
  • 1
    $\begingroup$ If the density approaches 0 everywhere it can't integrate to 1. That is why it is improper. $\endgroup$ Jun 24, 2017 at 16:23
  • $\begingroup$ @user3503589 That's right. There can't be a normalizing constant when the integral is 0. $\endgroup$ Jun 24, 2017 at 16:56
  • $\begingroup$ @Kodiologist But does it even make sense for a density to be zero everywhere? I was reading a paper by Prior distribution for variance parameter in hierarchical models by Andrew Gelman, where they talk about approximating the improper distribution $U(0, \infty)$ as a limit of proper distributions i.e $U(0,A), A \to \infty$ $\endgroup$ Jun 26, 2017 at 19:59
  • $\begingroup$ @user3503589 No, it does not make sense, because it means that the function has integral 0. The improper prior $U(0, ∞)$ indeed has no corresponding random variable or probability density function. $\endgroup$ Jun 26, 2017 at 20:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.