Mechanism behind t-NQS’s enhanced sample efficiency

Determine whether the ability of the time-dependent neural quantum state (t-NQS) trained via a global variational objective to share information between different time points is indeed the primary mechanism responsible for its substantially enhanced sample efficiency compared to forward-integration time-dependent neural quantum state algorithms such as time-dependent variational Monte Carlo, which typically require many more Monte Carlo samples per simulation time step.

Background

The paper introduces time-dependent neural quantum states (t-NQS), which embed time explicitly into a neural network ansatz and optimize a single, time-independent parameter set across an entire interval using a global variational objective. This departs from traditional step-by-step integration approaches (e.g., TDVP/tVMC) and enables simultaneous optimization at multiple time points.

In the conclusion, the authors report strong generalization of t-NQS to unseen times and posit that this global training across time points may allow information sharing that reduces the number of samples needed per time step. They explicitly conjecture that this generalization capability is responsible for the observed sample efficiency gains over other time-dependent NQS methods.