The Blackwell-Cover Guessing Strategy is a clever mathematical technique originally designed for a seemingly simple problem: you are shown one of two numbers drawn from some unknown distribution, and you want to guess whether it is the larger or smaller of the two. The strategy works by independently generating a random threshold and comparing the observed number to that threshold. Surprisingly, this approach lets you guess correctly more than half the time, even though you never see the other number. This paper extends earlier work by applying a version of this strategy to pairs of coins rather than pairs of numbers, where each coin has some probability of landing heads.
Here is the specific setup: two coins are drawn from a mixture of two probability distributions, one coin is flipped, and based on the result, the unflipped coin is assigned to one of two groups, called A and B. The main finding is that no matter what distributions you use to generate the coins, you can always arrange this process so that the average "heads probability" of the unobserved coins landing in group B is higher than the average for those landing in group A. In other words, the flip of one coin carries genuine statistical information about the unflipped coin's bias, even though you never directly observe it.
The authors highlight why this result is surprising by contrasting coins with marbles. If you weigh one of two marbles and they have fixed known weights, you instantly know the other marble's weight with certainty. But coins with fixed known biases do not work the same way, because knowing the outcome of one flip does not immediately tell you which coin you are looking at. The Blackwell-Cover strategy exploits exactly this uncertainty to extract hidden information, and the paper shows this works across a very broad and general class of distributions, making the result much more widely applicable than previously demonstrated.