In this math.se post I described in some detail a certain paradox, which I will summarize:

$A$ writes two distinct numbers on slips of paper. $B$ selects one of the slips at random (equiprobably), examines its number, and then, without having seen the other number, predicts whether the number on her slip is the larger or smaller of the two. $B$ can obviously achieve success with probability $\frac12$ by flipping a coin, and it seems impossible that she could do better. However, there is a strategy $B$ can follow that is guaranteed to produce a correct prediction with probability strictly greater than $\frac12$.

The strategy, in short, is:

- Prior to selecting the slip, $B$ should select some probability distribution $D$ on $\Bbb R$ that is everywhere positive. A normal distribution will suffice.
- $B$ should generate a random number $y\in \Bbb R$ distributed according to $D$.
- Let $x$ be the number on the slip selected by $B$. If $x>y$, then $B$ predicts that $x$ is the larger of the two numbers; if $x<y$ she predicts that $x$ is the smaller of the two numbers. ($y=x$ occurs with probability $0$ and can be disregarded.)

I omit the analysis that shows that this method predicts correctly with probability *strictly* greater than $\frac12$; the details are in the other post.

I ended the other post with “I have heard this paradox attributed to Feller, but I'm afraid I don't have a reference.”

I would like a reference.