The paper studies a mathematical operation called "change-of-numeraire reweighting," which shows up in several important contexts. In finance, it is used when pricing interest rate products like swaptions, where you shift between different ways of measuring value (different "numeraires") to simplify calculations. In statistics, it appears as self-normalized importance sampling, a technique for estimating properties of one probability distribution using samples drawn from a different one. The core question the paper asks is: if your input distribution changes only slightly, does the output of this reweighting operation also change only slightly? In other words, is the operation stable?
The paper gives a precise and complete answer to this stability question using a mathematical framework called Wasserstein convergence, which measures how close two probability distributions are in a way that accounts for the geometry of the underlying space. The main finding is that stability holds if and only if a certain technical condition called uniform integrability is satisfied for a specific combination of the payoff function and the numeraire. Roughly speaking, this condition ensures that no important probability mass is hiding in regions where the numeraire gets very small or vanishes entirely. When the payoff has no mass in such regions, the analysis simplifies dramatically: you only need the payoff itself to be well-behaved, and you do not need to worry about the numeraire blowing up or becoming problematic in reciprocal form. The paper also constructs simple two-point examples that show these conditions cannot be weakened.
The practical consequences are worked out in two directions. First, the results clarify a technical assumption that appears frequently in the mathematical finance literature on optimal transport, essentially explaining why that assumption is needed and exactly when it can be dropped. Second, the results tell practitioners which statistical summaries computed under the so-called annuity measure, a standard tool in interest rate modeling, remain meaningful and well-behaved as market conditions shift. Together, the findings give both theoreticians and quantitative practitioners a cleaner and more complete picture of when this widely used reweighting procedure can be trusted.