I am reading the 2008 annals of statistics paper "Ranking and empirical minimisation of U-statistics" by Clémençon et. al, and read a statement which I do not know why is true. In order to accurately frame my question, I will give a brief account of the context. It is not necessary to read this to understand the question, so if you trust that I have reasoned correctly, you can skip to the bottom.

Assume that we observe two independent pairs $(Y_1, X_1),$ $(Y_2, X_2)$ from the same distribution, with support $\{-1, 1\}\times\mathcal{X}.$ We seek a ranking rule $r:\mathcal{X}\times\mathcal{X}\rightarrow \{-1, 1\}$ that minimises the ranking risk $$ R(r) = Pr((Y_1 - Y_2)\cdot r(X_1, X_2) < 0 ). $$ One can show that the minimum ranking risk in this case is $$ R^* = \mathbb{E}(\min(\eta(X_1), \eta(X_2)) - (\mathbb{E}(\eta(X)))^2, $$ where $\eta(x) = Pr(Y=1\vert X=x).$

So far so good. The authors then state, and this is where the basis of my question lies, that we also have $$ R^* = \text{Var}\left(\frac{Y+1}{2}\right) - \frac{1}{2}\mathbb{E}\vert \eta(X_1) - \eta(X_2)\vert. $$ We obviously have that for $W = (Y+1)/2,$ $W\vert X\sim\text{Bernoulli}(\eta(X)),$ so by using the double rule of variance we can see that $$ \text{Var}((Y+1)/2) = \mathbb{E}\{\eta(X)\} - (\mathbb{E}\{\eta(X)\})^2. $$ By inserting this, and equating the two supposed equivalent expressions for $R^*,$ we can see that both these statements are true if $$ \mathbb{E}\{\min(\eta(X_1), \eta(X_2))\} = \mathbb{E}\{\eta(X)\} - \mathbb{E}\{\vert\eta(X_1) - \eta(X_2)\vert\}. $$

My question:

The above statement may be rephrased in the following way. Given $X_1, X_2$ independent and identically distributed according to a distribution function $F$ whose support is a subset of $[0, 1],$ we have that $$ \mathbb{E}(\min(X_1, X_2)) = \mathbb{E}(X_1) - \frac{1}{2}\mathbb{E}(\vert X_1 - X_2\vert), $$ or equivalently that $$ \mathbb{E}(\min(X_1, X_2)) = 2\mathbb{E}(X_1) - \mathbb{E}(\max(X_1, X_2)). $$ This seems quite reasonable, and I have checked some examples to verify that it is true for a couple of distributions. I have tried searching for a result that states something along the lines of this, asked around if anyone knew, but have so far been unsuccessful. Does anyone know of such a result, what the proof is, and what the most general conditions for when it holds are?

  • 3
    $\begingroup$ The last result follows easily from the iid assumption and the observation $X_1+X_2=\min(X_1,X_2) + \max(X_1,X_2).$ $\endgroup$
    – whuber
    Feb 1, 2019 at 17:51

1 Answer 1


You want to show that $ \mathbb{E}(\min(X_1, X_2)) = \mathbb{E}(X_1) - \frac{1}{2}\mathbb{E}(\vert X_1 - X_2\vert), $ following the comment from @whuber, we have that $\DeclareMathOperator{\E}{\mathbb{E}} X_1 + X_2 = \min(X_1, X_2) + \max(X_1, X_2)$ as our starting point. $$ X_1 + X_2 = \min(X_1, X_2) + \max(X_1, X_2) \\ = \min(X_1, X_2) + \left(\min(X_1, X_2) + \mid X_1 - X_2 \mid \right) \\ = 2 \min(X_1, X_2) + \mid X_1-X_2 \mid $$ reorganizing $$ \min(X_1, X_2) = \frac12 \left( X_1 + X_2 - \mid X_1 - X_2 \mid \right) $$ taking expectations and using that $X_1$ and $X_2$ have the same expectation $$ \E \min(X_1, X_2) = \E X_1 - \frac12 \E \mid X_1-X_2 \mid $$


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.