How do many-body wavefunction interpolation methods decide which quantum configurations to sample during training?

# Many-Body Wavefunction Interpolation: Sampling Strategies

The key insight is that modern wavefunction interpolation methods use active learning rather than random sampling—they strategically select which configurations to compute during training based on where accuracy matters most.

The Core Strategy: Variational Basis Interpolation

The foundational approach treats quantum configurations like points in a high-dimensional chemical space. Instead of sampling uniformly, these methods [AG-2024.03-1949] build training sets by computing accurate correlated wavefunctions (using methods like DMRG—density matrix renormalization group) at *selected* nuclear geometries, then interpolate through intervening configurations using a variationally optimal linear combination of the training states.

This matters because it transforms an exponentially hard problem into a tractable one: with just 22 expensive DMRG calculations at different molecular geometries, the method inferred energies, forces, and nonadiabatic couplings across 12,000 geometries [AG-2024.03-1949]—a computational gain of ~500×.

Active Learning: Closing the Loop

Rather than pre-selecting a fixed training set, the best schemes use active learning to grow the training set iteratively. The method monitors where interpolation errors are large along actual molecular dynamics trajectories, then requests new calculations at those problematic configurations [AG-2024.03-1949]. This ensures the training set remains "compact and systematically improvable"—you only compute what you actually need.

The sampling criterion is *predictive uncertainty*: configurations where the interpolation model is least confident get flagged for inclusion.

Connection to Optimization

Once training configurations are chosen, the wavefunction inference itself becomes a constrained optimization problem. The training states are combined as linear combinations whose coefficients vary smoothly with nuclear geometry [AG-2024.03-1949]. Modern optimizers—including second-order methods with manifold geometry [AG-2024.02-1564] and decision-theoretic approaches [AG-2024.01-2092]—refine these coefficients to minimize variational energy while maintaining orthogonality between electronic states.

The preprints focus on *which geometries* to sample (active learning) and *how to fit* the resulting data (optimization), rather than detailed Monte Carlo sampling strategies within each configuration.