You train a model and measure its error on the same data. The error estimate tells you how well the model learned. But the estimate itself consumed information — the same data that trained the model also judged it. There is a tension between using data to learn and using data to evaluate what was learned.
The paper frames this as a Heisenberg-like uncertainty principle: simultaneous learning and error assessment are fundamentally incompatible beyond a certain precision. Improving the model's fit necessarily degrades the reliability of the error estimate, and improving the error estimate requires withholding data that could have improved the model.
This is not the bias-variance tradeoff, which concerns the model's predictive quality. It is a tradeoff in the observer's epistemic state: you can know how well you learned, or you can learn well, but the product of these accuracies is bounded. The bound is structural — it follows from the information geometry of the estimation problem, not from a particular method's limitations.
Cross-validation and holdout sets are partial workarounds: they separate learning data from evaluation data, reducing the coupling. But they do not eliminate the tradeoff — they sacrifice learning quality (by withholding data) to gain evaluation quality. The bound describes what the workaround costs.
The structural point: the act of assessment interferes with the thing assessed. Not in the quantum-mechanical sense of physical disturbance, but in the information-theoretic sense that the same finite dataset cannot simultaneously serve as both training set and test set without mutual contamination. The uncertainty is not about the model; it is about the knowledge you can have about the model.