The parametrization method builds reduced-order models of vibrating structures by computing invariant manifolds — the low-dimensional surfaces in phase space that capture the essential dynamics. These reduced models replace expensive full simulations with cheap, accurate approximations. They work brilliantly within their domain of validity.
The domain of validity has a finite radius in amplitude space, and the model does not announce when you've exceeded it.
Three independent error criteria — invariance equation residuals, homological operator bounds from normal form theory, and Cauchy-d'Alembert series convergence rules — can determine this radius a priori. The finding: the bounds are surprisingly tight. Reduced models fail at amplitudes much lower than practitioners typically assume. And the failure is silent — the reduced model continues to produce smooth, plausible-looking trajectories that happen to be wrong.
This is the generic problem with lossy compression. Any compression scheme — whether in numerical analysis, data science, or conceptual modeling — trades information for tractability. The compressed representation agrees with the full system within some neighborhood of the compression point. Beyond that neighborhood, the agreement degrades. The dangerous case is when the degradation is smooth rather than catastrophic — when the compressed model fails gradually, producing reasonable-looking answers that are quantitatively or even qualitatively incorrect.
The paper's contribution is making the radius computable. You can know, before running a simulation, exactly where the reduced model stops being trustworthy. But the deeper lesson is that every compression has such a boundary, and the boundary's existence is independent of whether anyone computes it. The model that has been validated at one amplitude may already be wrong at twice that amplitude. The error doesn't grow gradually; it detonates — but silently, without raising any flag that a single simulation could detect.