When Can You Correct Distribution Drift in Temporal Graph Generation? A Sharpening--Drift Tension and an Impossibility for Observation-Based Correction

Researchers at Tianjin University and Hangzhou Dianzi University analyze why temporal graph generators trained on one period degrade on the next. For masked flow matching, they prove an exact decomposition of loss into irreducible conditional entropy plus a divergence from the deployment distribution, without assuming independent graph slots. Along a source-training sharpening path, the divergence increases specifically for structures rare in training but common later, with the empirical trade-off following a power law of exponent −0.605 and R2 = 0.9977. Across seven conditions spanning Enron, MOOC, Reddit, and Wikipedia, changing sampling from 1 to 50 steps alters drift-period marginal error by at most 6.0%, while the converged joint-error floor is 2.2× to 34.3× higher than in-period performance, showing that drift changes the destination rather than sampling speed. The central impossibility theorem states that any correction based only on past unlabeled observations retains at least the target statistic’s conditional variance. An oracle knowing current marginals removes 60% of marginal error, but the best observation-based correction recovers only 5.7% of that gain. Because measured drift is trendless and mean-reverting, extrapolation is worse than persistence, with error ratios of 1.27× to 1.51×. DiGress shows the same directional degradation, though its weaker sharpening does not reproduce the power law. The paper therefore directs future remedies toward exogenous side information or deliberately less-confident generators, rather than more test-time adaptation.