Extrapolation is unreliable — this is machine learning's founding anxiety. Within the training distribution, predictions are tested. Outside it, anything can happen. No guarantee extends beyond the data.
Note (2026-11-06): I checked this essay against the raw arXiv abstract of 'Anchor-Based Function Extrapolation with Proven Bounds and Projection Guarantees' (2603.09937). The abstract supports the framework: anchor functions for which one can certify an upper bound on the distance to the unknown target in the extrapolation region, the resulting feasible sets provably containing the true function, projection of a baseline (it names least-squares or regularized regression) onto the set giving a predictor proven not to increase the error there, a tight spectral condition number, probabilistic anchors for less conservative certificates, and numerical experiments on geomagnetic field modelling and nonlinear oscillators. My essay lists a neural network as a baseline, which the abstract does not, says the framework simply 'constructs' the feasible set without noting that it needs certifiable anchors, and says extrapolation is 'provably safe' when the condition number is moderate and 'provably dangerous' when it is large; the abstract states bounds and improvement guarantees, not those two verdicts. I checked the abstract only; the essay names no authors.
Anchor functions provide the guarantee (arXiv:2603.09937). Given any baseline approximation — least squares, regularized regression, neural network — the framework constructs a feasibility set that provably contains the target function, then projects the approximation onto this set. The corrected predictor is proven to never increase the error. The key is a spectral condition number for extrapolation, distinct from the classical interpolation condition number, that quantifies how much information the training data carries about the extrapolation region.
The structural inversion: extrapolation is not inherently unreliable — it is conditionally reliable, and the condition can be computed. The spectral condition number measures the gap between interpolation and extrapolation. When this number is moderate, extrapolation is provably safe. When it is large, extrapolation is provably dangerous. The guarantee is not “extrapolation works” but “the error is bounded by a computable quantity.” The distinction between reliable and unreliable extrapolation is itself measurable.