A group of researchers recently conjectured about the best way to measure a specific kind of "distance" between two probability distributions that are both bell-curve shaped (Gaussian) but centered at different points. The distance measure in question combines ideas from information theory (Renyi divergence, a generalization of a classical quantity called Kullback-Leibler divergence) and probability (Wasserstein distance, which roughly measures how much "work" it takes to move one distribution to match another). The conjecture claimed that the worst-case or most expensive way to shift one Gaussian to match another is simply a straight, deterministic slide of the center point, which would have produced a clean, simple formula.
This paper resolves that conjecture, but with a split verdict. When the Renyi parameter equals one, the divergence reduces to the classical Kullback-Leibler divergence, and in that special case the conjecture is confirmed to be correct. The deterministic shift is indeed optimal there, and the clean formula holds.
However, for every other value of the Renyi parameter greater than one, the conjecture turns out to be false. This means that in the general case, the optimal way to "shift" between two Gaussians under this combined measure is not a simple deterministic translation but something more complex. The practical implication is that the neat closed-form budget accounting hoped for in earlier theoretical work on sampling algorithms does not carry over beyond the classical special case, and understanding the true optimizer in those settings will require additional work.