I am stuck on the following problem from my research. A version of the problem is as follows.

We observe a binomial random variable X with n=2 and an unknown probability $\theta$ of success on each trial (so X is the number of successes in two trials). Our prior beliefs about p are given by some continuous distribution F with density f.

I am trying to prove or find a counter-example to the following conjecture:

$\mathbb{E}[\theta | X=1]^2 \geq \mathbb{E}[\theta | X=0] \mathbb{E}[\theta | X=2].$

I cannot seem to prove this but I also can't find a counterexample. It works for:

  • a uniform prior: $\frac{1}{4} > \frac{1}{4}\frac{3}{4}$
  • A Beta prior: $\frac{(\alpha +1)^2}{(\alpha +\beta +2)^2}-\frac{\alpha (\alpha +2)}{(\alpha +\beta +2)^2} = \frac{1}{(\alpha +\beta +2)^2} > 0$
  • Every density with support on [0, 1] that I could make up

Some approaches that have not worked so far for proving the statement to be true:

  • I can express all three conditional probabilities as functions of unconditional expectations and the inequality becomes $$ \frac{(\mathbb{E}[\theta^2] - \mathbb{E}[\theta^3])^2}{(\mathbb{E}[\theta] - \mathbb{E}[\theta^2])^2} \geq \frac{\mathbb{E}[\theta^3] - 2\mathbb{E}[\theta^2] + \mathbb{E}[\theta]}{\mathbb{E}[\theta^2] - 2\mathbb{E}[\theta] + 1}\frac{\mathbb{E}[\theta^3]}{\mathbb{E}[\theta^2]}. $$ My idea was that some applications of Jensen's inequality would give me more purchase on this statement, which might be true but I certainly haven't gotten there.
  • I have also tried working from the fact that $\mathbb{E}[\theta]$ is a probability-weighted average of the three terms in my conjecture.

Edit: It turns out another way of stating what I want to prove is that the conditional expectations of p given X are log-concave w.r.t. X. The binomial distribution is log-concave but I am not sure that implies the conditional expectations are log concave given any prior. I added that to the title to help people out.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.