The Dual Role of Low-Weight Pauli Propagation: A Flawed Simulator but a Powerful Initializer for Variational Quantum Algorithms

TL;DR

Variational quantum algorithms (VQAs) face rugged optimization landscapes that hinder reliable energy estimation. The authors show that the low-weight Pauli propagation (LWPP) algorithm, while a flawed simulator for single-instance energy evaluation, acts as a spectral filter via Pauli-weight truncation, smoothing high-frequency local minima and aligning the search with the global basin. This landscape-alignment insight enables a practical two-stage workflow: a classical LWPP pre-optimization to generate high-quality initial parameters, followed by exact quantum optimization, yielding order-of-magnitude improvements in final accuracy and convergence speed across spin models, hierarchical architectures such as MERA, varied lattice topologies, and molecular ground states, even under noise. By validating universality across domains and contrasting with MPS initialization, the work demonstrates that LWPP provides a robust navigator that can substantially reduce quantum-resource requirements and accelerate progress toward practical quantum advantage in VQAs.

Abstract

Variational quantum algorithms are often hindered by rugged optimization landscapes. In this Letter, we investigate the low-weight Pauli propagation (LWPP) algorithm and find that it serves as an unreliable energy estimator for variational circuits. However, we reveal a counterintuitive insight: the Pauli-weight truncation acts as a spectral filter, effectively smoothing out high-frequency local minima while preserving the global basin of attraction in the landscape. We identify this mechanism as landscape alignment, where the approximate landscape becomes a superior navigator compared to the rugged exact landscape. Benchmarks across diverse spin models and molecular systems demonstrate that LWPP-initialized optimization yields order-of-magnitude improvements in accuracy, often finding solutions inaccessible to direct exact optimization. This work reframes LWPP from a flawed simulator into a vital pre-optimizer that serves not only as a cheap classical substitute but also as an essential tool for addressing quantum optimization challenges.

Figures (18)

• Figure 1: Parameter-dependent accuracy of the LWPP estimation. Performance analysis on a $3 \times 4$ antiferromagnetic Heisenberg XYZ model with circuit depth $d=4$ and varying truncation cutoffs $k$. The top row (a, b) starts from random initialization, while the bottom row (c, d) starts from near-identity initialization. (a, c) Exact optimization + LWPP evaluation: The optimizer is driven by the exact energy gradient $\nabla_\theta E$ (solid black line), while the LWPP energy estimate $E_{\text{LWPP}}$ (blue dashed lines) is passively recorded at each iteration to benchmark its accuracy. The inset in (a) provides a magnified view of the LWPP evaluations, which collapse to near-zero values. (b, d) LWPP optimization + Exact evaluation: The optimizer is driven by the approximate LWPP gradient $\nabla_\theta E_{\text{LWPP}}$ (solid colored lines) as the cost function. The true exact energy $E$ (dashed colored lines) is passively recorded to verify the quality of the state. Despite the numerical inaccuracy, minimizing the LWPP cost robustly guides the true energy down, suggesting its potential utility as a navigator.
• Figure 2: An illustrative example of LWPP-initialized optimization from a random start. VQA optimization of a $d=6$ VQA circuit on a $3\times4$ lattice, starting from the same random parameters. The figure contains 12 independent optimization runs, each initiated from a randomly sampled parameter set and represented by a different color. (a) Direct optimization consistently fails, converging to a high-energy local minimum. (b) LWPP-initialized optimization succeeds in finding a low-energy state. (c,d) The corresponding LWPP pre-optimization dynamics, with (d) showing the energy deviation from zero on a log scale.
• Figure 3: A comparative example for LWPP from near-identity parameters. Comparison of optimization trajectories of a $d=6$ circuit on a $3 \times 4$ lattice, both starting from the same near-identity parameter set. Direct optimization performs well (a), but the LWPP-initialized run (b), which includes a pre-optimization phase, achieves a significantly better final accuracy.
• Figure 4: Statistical comparison of initialization strategies with near-identity heuristic. Results for antiferromagnetic models on lattices ranging from $3 \times 3$ to $3 \times 6$. (a-d) Final optimization accuracy after 1500 optimization steps: The final accuracy achieved with LWPP initialization with $k=3$ (green line) is consistently superior to direct near-identity optimization, with the advantage often increasing with circuit depth. (e-h) Optimization speedup: The relative steps required for LWPP-initialized runs ($k=3$) to reach the target accuracy, where the target is defined as the final accuracy reached by 1500 steps of direct near-identity optimization. The relative steps are calculated as the number of required iterations normalized by 1500. Lower values indicate faster convergence; LWPP-initialized optimization consistently achieves the target accuracy in a small fraction of the iterations required by the direct method, often providing a 10-fold speedup.
• Figure 5: Schematic of landscape alignment. (a) Direct optimization: This optimization typically traps parameters (colored dots) in high-energy local minima. (b) LWPP-initialized optimization: The exact landscape (black) is rugged due to high-frequency operator spreading. The LWPP approximation (grey) acts as a smoothed surrogate by filtering out high-weight terms. Optimization on this coarse-grained landscape ($\nabla_\theta E_{\text{LWPP}}$) avoids narrow traps and robustly guides parameters into a promising basin of attraction.