New study proves temporal graph drift is unfixable from observations

A new paper from Tianpeng Li, Xuan Guo, and colleagues proves a fundamental limitation in temporal graph generation: the distribution shift between training and deployment is inherently uncorrectable through observation. The masked flow-matching loss decomposes exactly (no independence assumptions) into an irreducible entropy term plus a divergence whose derivative along the training path is positive precisely for structures rare during training but common at deployment—the divergence goes to infinity as training probability approaches zero. Empirically, this trade-off follows a power law with exponent –0.605 (R²=0.9977). Across seven well-powered conditions, the drift-period marginal error varies by at most 6% over a 50× range of sampling budgets, while the error floor sits 2.2× to 34.3× above the in-period floor.

Because the deployment period is observed, correction seems like a measurement problem—but it is not. The authors prove that any corrector measurable with respect to past observations leaves at least the conditional variance of the statistic it tracks. Trend extrapolation only beats trusting the last observation when μ² > v(1 – 2ρ), but both premises go the wrong way: the drift is trendless and mean-reverting, with a one-step innovation as large as the drift itself. An oracle removes 60% of the error; the best observation-based corrector recovers only 5.7% of that oracle improvement. Extrapolation is strictly worse than doing nothing clever. This result has immediate implications for any system that generates evolving network structures, from social media dynamics to financial transaction graphs.