It is always argued that the posterior median is the Bayes estimate associated to the absolute loss function. The proofs I have come across rely on differentiating the conditional Bayesian risk and equating this to zero. However, this shows that this is an inflection point, but not that it is a minimum. How can one prove that the posterior median indeed minimizes the Bayesian risk?

enter image description here

Most, if not all, textbooks just go next level and just leave it as an exercise to the poor reader.


2 Answers 2


Second derivative yields

$$2 \pi (\delta | x) \geq 0$$

So the original function is convex and hence the median corresponds to a minimum not an inflection point

  • $\begingroup$ How do you get Zero? Are you using $d/dx \int_a^x f(y)dy =f(x)$ and $. - \int_x^a f(y)dy = \int_a^x f(y)dy $? $\endgroup$
    – Xiaomi
    Nov 14, 2018 at 1:58
  • $\begingroup$ The second derivative as a function of $\delta$ is not zero everywhere. The first derivative is $ 2 \int^\delta \pi(y)dy - 1$. Taking the derivative a second time gives the above result which is clearly not zero everywhere $\endgroup$
    – Xiaomi
    Nov 14, 2018 at 2:13

Adding to Xiaomi answer, here is the derivation, using Leibnitz rule:

$\require{cancel} \frac{\partial \rho}{\partial \delta} = \int_{-\infty}^{\delta}\pi(\theta|x)d\theta - \int_{\delta}^{\infty}\pi(\theta|x)d\theta \\ \frac{\partial^2 \rho}{\partial \delta^2} = \pi(\delta|x)\cdot1 - \cancel{\pi(\delta|x)\cdot0} + \cancel{\int_{-\infty}^{\delta}0d\theta} - [\cancel{\pi(\delta|x)\cdot0} - \pi(\delta|x)\cdot1 + \cancel{\int_{\delta}^{\infty}0d\theta}] = 2\pi(\delta|x) $


Your Answer

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

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