Report on scipost_202202_00044v1

Kevin Zhang, Samuel Lederer, Kenny Choo, Titus Neupert, Giuseppe Carleo, Eun-Ah KimOpenalex Unknown 2022 · DOI · OpenAlex

Abstract

Variational approaches are among the most powerful techniques to approximately solve quantum many-body problems. These encompass both variational states based on tensor or neural networks, and parameterized quantum circuits in variational quantum eigensolvers. However, self-consistent evaluation of the quality of variational wavefunctions is a notoriously hard task. Using a recently developed Hamiltonian reconstruction method, we propose a multi-faceted approach to evaluating the quality of neural-network based wavefunctions. Specifically, we consider convolutional neural network (CNN) and restricted Boltzmann machine (RBM) states trained on a square lattice spin-1/2 J1-J2 Heisenberg model. We find that the reconstructed Hamiltonians are typically less frustrated, and have easy-axis anisotropy near the high frustration point. In addition, the reconstructed Hamiltonians suppress quantum fluctuations in the large J2 limit. Our results highlight the critical importance of the wavefunction's symmetry. Moreover, the multi-faceted insight from the Hamiltonian reconstruction reveals that a variational wave function can fail to capture the true ground state through suppression of quantum fluctuations.

Introduction

In quantum many-body physics, finding the ground state of a Hamiltonian is a fundamental challenge. Due to the exponential growth of complexity with system size, researchers frequently employ variational methods, proposing a trial wavefunction (ansatz) and optimizing its parameters to minimize energy. However, evaluating the true quality of these variational states is difficult, as energy alone is often an insufficient metric. This paper introduces "Hamiltonian reconstruction" as a diagnostic tool to reverse-engineer the physical system a trial wavefunction actually describes, revealing hidden biases and physical shortcomings.

Methodology

The authors utilize Hamiltonian reconstruction to evaluate two types of neural network states: Convolutional Neural Networks (CNN) and Restricted Boltzmann Machines (RBM). These are trained to find the ground state of the spin-1/2 $J_1-J_2$ Heisenberg model on a $4 \times 4$ square lattice.

Conceptual framework of Hamiltonian reconstruction.

The process involves constructing a Quantum Covariance Matrix ($Q$) from the wavefunction's correlation functions. By diagonalizing $Q$, a "reconstructed" Hamiltonian ($H_{recon}$) is found for which the trial state is an exact ground state. The study specifically examines deformations in the reconstructed Hamiltonian, such as anisotropy ($\alpha$) and variations in interaction strengths ($\delta J_2, J_3$), to identify where the neural network "cheats" or fails to capture the target physics.

Results and Discussion

The reconstruction reveals that neural network states often break fundamental symmetries to achieve lower energies. A key finding is the emergence of easy-axis anisotropy ($\alpha$) near the point of maximum frustration ($J_2/J_1 = 0.5$). This indicates that the networks, lacking explicit $SU(2)$ symmetry, find a more stable configuration by simplifying the underlying physics.

Quantitative results for reconstructed Hamiltonian parameters.

Furthermore, the reconstructed Hamiltonians are consistently less frustrated than the target model. In the large $J_2$ "stripe" phase, the neural networks tend to suppress quantum fluctuations in favor of classical-like order. While CNNs generally achieve lower variational energies than RBMs, both architectures exhibit similar qualitative biases, such as the suppression of fluctuations and symmetry breaking.

Conclusion

Hamiltonian reconstruction provides a powerful, multi-faceted diagnostic for variational quantum states. The results underscore the critical importance of incorporating physical symmetries, like $SU(2)$, into neural network architectures to avoid systematic biases. While currently limited by sensitivity to statistical noise—requiring exact correlation functions for small systems—future developments could scale this method to larger systems and provide a robust tool for refining Variational Quantum Eigensolvers (VQE) on quantum hardware.

Summary updated

Abstract

Variational approaches are among the most powerful techniques to approximately solve quantum many-body problems. These encompass both variational states based on tensor or neural networks, and parameterized quantum circuits in variational quantum eigensolvers. However, self-consistent evaluation of the quality of variational wavefunctions is a notoriously hard task. Using a recently developed Hamiltonian reconstruction method, we propose a multi-faceted approach to evaluating the quality of neural-network based wavefunctions. Specifically, we consider convolutional neural network (CNN) and restricted Boltzmann machine (RBM) states trained on a square lattice spin-1/2 J1-J2 Heisenberg model. We find that the reconstructed Hamiltonians are typically less frustrated, and have easy-axis anisotropy near the high frustration point. In addition, the reconstructed Hamiltonians suppress quantum fluctuations in the large J2 limit. Our results highlight the critical importance of the wavefunction's symmetry. Moreover, the multi-faceted insight from the Hamiltonian reconstruction reveals that a variational wave function can fail to capture the true ground state through suppression of quantum fluctuations.
BibTeX
@inproceedings{zhang2022report,
  author = {Kevin Zhang and Samuel Lederer and Kenny Choo and Titus Neupert and Giuseppe Carleo and Eun-Ah Kim},
  title = {{Report on scipost\_202202\_00044v1}},
  booktitle = {Openalex Unknown},
  year = {2022},
  doi = {10.21468/scipost.report.4770},
}
Figure