The paper studies a basic question about random counting: if you flip a bunch of possibly unfair coins and count the heads, when does the probability of getting exactly some number of heads first become larger than the probability of getting fewer heads? More precisely, imagine adding up many independent yes/no random variables (a Poisson-Binomial distribution). The question is: how large does the average total need to be before the chance of landing on a higher value surpasses the chance of landing on a lower one, depending on how far apart those two values are?
The answer turns out to depend sharply on the gap between the two values being compared. When the gap is just one (neighboring values), a classical 1964 result by Darroch gives the threshold. When the gap is three or more, the threshold is simply a whole number. But a gap of exactly two is special: it has its own unique threshold that falls between those two cases, with a precise formula involving a staircase-shaped function that decreases toward a half-integer limit as the baseline count grows. The proof works by reducing the problem to a clean optimization over a simpler family of distributions (shifted binomial distributions), then solving that discrete optimization exactly. This makes the distance-two case a genuine mathematical anomaly, sitting between the adjacent case and all larger gaps in an unexpected way.
Beyond being a curiosity about probability distributions, the result has a concrete implication for the theory of polynomials. Poisson-Binomial probabilities appear as coefficients of certain polynomials whose roots are all real numbers. The threshold results directly constrain when those coefficients can reverse their size ordering at non-neighboring positions, giving sharp bounds that were not previously known. So the work connects a probabilistic question about coin-flipping to an algebraic question about polynomial coefficients, with the strange exceptionalism of gap-two appearing in both settings.