What mechanisms prevent interpolation errors from accumulating when wave functions transition between fundamentally different geometric states?

This is a great question about a deep technical issue in quantum dynamics. The preprints address it from several angles, particularly when wave functions must track molecules through conical intersections (points where different electronic states touch).

The Core Problem

When you interpolate a many-body wave function between two different nuclear geometries—say, before and after a chemical bond breaks—you're navigating a high-dimensional space where the "identity" of electronic states can flip or become ambiguous. Naïve interpolation can introduce artificial oscillations or lose phase coherence, causing errors to cascade through a molecular dynamics simulation.

How These Methods Prevent Accumulation

Variational representation with provable convergence: The key insight is to represent interpolated wave functions as *variationally optimal linear combinations* of a small training set, rather than direct geometric interpolation [AG-2024.03-1949]. This means the inferred state at any new geometry is constructed as the best possible combination of your reference states—a constrained minimization that automatically respects quantum structure. The authors show this converges to high accuracy: with just 22 electronic structure calculations at different geometries, they inferred energies and forces at 12,000 geometries along molecular trajectories with "provable convergence to high accuracy" [AG-2024.03-1949].

Gauge-invariant phase tracking: When a trajectory encircles a conical intersection, the wave function picks up a geometric phase—a topological twist like the Aharonov-Bohm effect in magnetism. Rather than letting this accumulate as error, the exact factorization framework tracks instantaneous, gauge-invariant phases separately for electrons and nuclei [AG-2024.02-1323]. This disentangles which part of the phase is physical (tied to the path through configuration space) and which is convention.

Mean-field computational scaling without exponential blowup: The standard problem is that an *exact* many-body wave function has exponentially many components. The interpolation schemes use compressed representations—similarity-transformed Hamiltonians (coupled cluster) or density-matrix renormalization group (DMRG)—that stay numerically tractable [AG-2024.02-1771, AG-2024.05-2039]. As long as entanglement remains modest (true for weakly to moderately correlated systems), errors don't explode.

Active learning for compact training sets: Rather than sample geometries blindly, active learning schemes grow the training set strategically along the actual trajectory, ensuring you only add states where interpolation is genuinely uncertain [AG-2024.03-1949]. This prevents wasted computation and keeps the representation minimal.

The Caveat

These methods work best when electronic states remain well-separated or when you're far from strong-driving regimes. Near a true conical intersection—where two surfaces touch and nonadiabatic effects dominate—trajectory-based approximations "struggle to capture the correct behavior" [AG-2024.01-1942]. In that regime, you may need full quantum wavepacket propagation or quantum simulation [AG-2024.01-1980].

The preprints directly address your question and provide concrete error bounds and numerical evidence.