Variational Adaptive Gaussian Decomposition: Scalable Quadrature-Free Time-Sliced Thawed Gaussian Dynamics

Time-slicing has emerged as a strategy for incorporating semiclassical propagation into real-time path integral formulation and recovering full quantum dynamics. A central step is the decomposition of a time-evolved wave function into a superposition of Gaussian wave packets (GWPs). Here we introduce a quadrature-free variational framework for GWP decomposition, reformulating it as an optimization problem in which the GWP parameters are chosen to maximize the overlap with the time-evolving wave function. An autoencoder-decoder neural network is used for this optimization, with the representation being adaptively reoptimized during propagation. Each wave packet in this decomposition represents a localized patch of the underlying semiclassical manifold, while retaining full correlations between all degrees of freedom. This variational adaptive Gaussian decomposition (VAGD) approach yields a compact Gaussian expansion, providing a scalable route to time-sliced semiclassical quantum dynamics. While general, applying VAGD to facilitate time-slicing of thawed Gaussian approximation (TGA) allows a route to improving the semiclassical treatment to the full quantum mechanical result in a systematic manner.

💡 Research Summary

The paper introduces a novel variational framework called Variational Adaptive Gaussian Decomposition (VAGD) that enables scalable, quadrature‑free time‑sliced semiclassical dynamics. In traditional time‑sliced approaches, a wavefunction is periodically re‑expanded into a large set of Gaussian wave packets (GWPs) and each packet is propagated independently. This requires high‑dimensional integrals, suffers from the Monte‑Carlo sign problem, and the number of Gaussians grows exponentially with system dimensionality, limiting practical applications.

VAGD reformulates the re‑expansion step as a variational optimization problem. Given an input wavefunction expressed as a sum of N_in GWPs, the method seeks a compact representation using at most K GWPs (N_out ≤ K) that maximizes the overlap (fidelity) with the original state. The objective function is L = –log(F) + (1 – F), where F is the overlap magnitude. An auto‑encoder‑decoder neural network is employed not as a predictive model but as a numerical optimizer: the encoder compresses the high‑dimensional GWP parameters (positions, momenta, complex width matrix, phase) into a low‑dimensional latent space, and the decoder reconstructs a new set of GWP parameters. By directly minimizing the loss, the network yields the optimal Gaussian parameters for the decomposition.

Key technical details ensure numerical stability. The complex width matrix A is split into a real symmetric part and an imaginary part A_I that must be positive‑definite. Instead of optimizing A_I directly, the algorithm optimizes a lower‑triangular matrix L with positive diagonal entries and reconstructs A_I = L Lᵀ via a Cholesky‑like factorization, guaranteeing positive definiteness. To accelerate convergence, a “warm‑start” strategy reuses the optimized parameters from the previous time‑slice as the initial guess for the next slice. Because the wavefunction’s shape changes slowly but its centroid may shift, a preprocessing step recenters the wavefunction and rescales its coordinate axes to match the variance of the initial state before feeding it to the network; the inverse transformation is applied after optimization.

The VAGD procedure is combined with the thawed Gaussian approximation (TGA) to form VAGD‑TGA, a time‑sliced semiclassical propagation scheme. At each slicing boundary, the propagated wavefunction is decomposed using VAGD to meet a fidelity threshold (e.g., F_thresh = 0.9995). The resulting compact Gaussian set is then propagated forward with TGA until the next slicing point.

Numerical experiments demonstrate the method’s effectiveness. In one‑dimensional Morse potentials, traditional TGA’s overlap with exact split‑operator Fourier transform (SOFT) dynamics deteriorates rapidly, especially for larger anharmonicity χ. VAGD‑TGA maintains high fidelity across long times while using a modest number of Gaussians (often far fewer than the maximum K). In multidimensional tests with independent Morse oscillators, conventional time‑sliced TGA (TSTG) requires a number of Gaussians that scales exponentially with dimensionality, whereas VAGD‑TGA achieves comparable accuracy with a number of Gaussians that grows only modestly with dimension, confirming its scalability.

The paper’s contributions are threefold: (1) it eliminates the need for high‑dimensional quadrature by casting Gaussian decomposition as a variational optimization; (2) it leverages an auto‑encoder‑decoder architecture as a flexible optimizer that automatically enforces physical constraints (symmetry, positive‑definiteness) and yields minimal‑size Gaussian bases; (3) it introduces practical stabilization techniques (warm‑start, variance‑matching preprocessing) that make the approach robust for realistic molecular dynamics. Moreover, the authors argue that VAGD is not limited to TGA and can be integrated with other semiclassical propagators or multilevel schemes, offering a general platform for accurate, efficient quantum dynamics.

In summary, VAGD‑TGA provides a quadrature‑free, variationally optimal Gaussian expansion that overcomes the sign problem and exponential scaling of traditional time‑sliced semiclassical methods, enabling accurate long‑time quantum dynamics for high‑dimensional molecular systems.