Least favorable prior - Find the distribution that maximizes the Bayes risk

Asked Jan 16 '21 at 15:03

Active Sep 20 '21 at 00:39

Viewed 65 times

Suppose I've found that the Bayes risk is of the form $$r(\theta) = \int_{-a}^a \theta^2 \pi(\theta)d\theta $$

I want to show that the following distribution, $\pi(a)=\pi(-a)=0.5$, maximizes this quantity, i.e. that it's a least favorable prior distribution.

This seems intuitively right to me, as any distribution over a range of values, will surely reduce the variance, compared to a distribution that allocates only the extreme values - but I'm not sure how to prove this rigorously.

edited Sep 20 '21 at 00:39

kjetil b halvorsen

63,378
26
142
467

asked Jan 16 '21 at 15:03

Maverick Meerkat

2,147
14
27

Notation correction: $r(\theta)$ should be $r(\pi)$ since $\theta$ is integrated out. Any prior fully supported by $\{-a,a\}$ is an acceptable solution. Any prior that gives positive weight to $\{-a,a\}^c$ has a lower risk. – Xi'an Jan 16 '21 at 18:07
@Xi'an what is $\{-a,a\}^c$? – Maverick Meerkat Jan 29 '21 at 08:47
The complementary set of $\{-a,a\}^c$, i.e., $(-a,a)$. – Xi'an Jan 29 '21 at 09:20
Thanks. Any formal proof of what you wrote? i.e. how to answer the original question – Maverick Meerkat Jan 29 '21 at 09:41

Least favorable prior - Find the distribution that maximizes the Bayes risk

0 Answers0