Source-linked AI summary
Approximating quantum many-body wave-functions using artificial neural networks
Zi Cai, Jinguo Liu
TL;DR
The paper asks whether compact artificial neural networks can efficiently represent ground states of quantum many-body systems, including states with difficult sign structures. It trains feed-forward ANNs on wave-function representations, modifies the architecture for sign-changing states, and applies the approach variationally. The reported results show high-precision approximations across several systems, including frustrated magnetism and an antiferromagnetic Heisenberg model.
Problem
Efficiently representing quantum many-body ground states is difficult because generic Hilbert spaces grow exponentially and sign-changing wavefunctions challenge standard methods.
Method
The paper trains small feed-forward ANNs to approximate wave-function coefficients, using fidelity optimization and separate or modified architectures for difficult sign structures.
Results
The ANNs approximate several many-body ground states with high precision, capture frustrated-system sign rules after architectural modification, and reach ground-state-energy precision of ~O(10^-3) variationally.
Takeaways & Limitations
Neural networks can express selected quantum many-body ground states with feasible architectures, while architecture choices can determine whether complex sign structures are learnable.
Takeaways & Limitations
The fully connected architecture may contain redundant connections, and the method's efficiency depends substantially on designer experience and intuition.
Abstract
from arXiv · showhide
In this paper, we demonstrate the expressibility of artificial neural networks (ANNs) in quantum many-body physics by showing that a feed-forward neural network with a small number of hidden layers can be trained to approximate with high precision the ground states of some notable quantum many-body systems. We consider the one-dimensional free bosons and fermions, spinless fermions on a square lattice away from half-filling, as well as frustrated quantum magnetism with a rapidly oscillating ground-state characteristic function. In the latter case, an ANN with a standard architecture fails, while that with a slightly modified one successfully learns the frustration-induced complex sign rule in the ground state and approximates the ground states with high precisions. As an example of practical use of our method, we also perform the variational method to explore the ground state of an anti-ferromagnetic $J_1-J_2$ Heisenberg model.
I. INTRODUCTION
The paper frames neural networks as a way to compress and approximate structured quantum many-body ground states despite exponentially large Hilbert spaces. It tests when simple architectures succeed and modifies them to handle rapidly changing sign structures.
- Quantum many-body Hilbert spaces grow exponentially with system size, making efficient numerical tools necessary for strongly correlated systems.
- Ground states can have simpler structure than generic states, allowing approximation with far fewer parameters in systems with limited entanglement or positive-definite wavefunctions.
- The paper investigates whether small feed-forward ANNs can represent ground states of several physically relevant many-body systems and learn them in reasonable time.
- A regular ANN can struggle with sign problems, so the paper separates amplitude and sign learning and designs corresponding architectures.
- For frustrated quantum magnetism, a slight neuron modification captures the sign rule even at the phase transition point, followed by a variational-method application.
II. METHODS
The method uses a fully connected feed-forward ANN to map Fock-basis inputs to wave-function coefficients, then trains its parameters by minimizing a fidelity-based objective. The architecture and activation function are treated as choices that affect approximation efficiency.
- The ANN is a layered network whose adaptive interlayer weights encode its expressibility.
- The four-layer architecture has two hidden layers of Nb neurons between an input layer accepting Fock-basis configurations and an output layer.
- Activation functions may be chosen from smooth nonlinear forms, and the paper reports that the choice can significantly improve efficiency for some target functions.
- Training minimizes a fidelity function measuring the difference between target coefficients CT[σ] and ANN-predicted coefficients CP[σ] over the ANN weights and biases.
III. FREE BOSONS/FERMIONS SYSTEMS
The paper tests ANNs on analytically characterized free-boson and free-fermion ground states, including cases with large Hilbert spaces and difficult fermionic sign structures. Small networks approximate one-dimensional states accurately, while two-dimensional fermions require targeting amplitudes and evaluating observables rather than directly fitting signed coefficients.
- Targets and models: Free-boson and free-fermion ground states provide analytic targets and accuracy benchmarks for ANN expressibility.The study includes one-dimensional systems and two-dimensional free fermions away from half-filling.
- Network size: ANNs use O(N_b^2) parameters, about 10^3 here, versus Hilbert spaces of about 10^6 states.This parameter choice is intended to avoid over-fitting while testing substantial compression.
- One-dimensional results: With only O(10^1) neurons, ANNs approximate one-dimensional free-boson and free-fermion target functions with extraordinarily high precision.The comparison uses exact ground-state wave functions as training and evaluation targets.
- Two-dimensional fermions: For two-dimensional free fermions, direct fitting of the signed characteristic function fails because fermionic statistics produce dramatically changing signs.The wave-function sign structure is therefore treated separately from its magnitude.
- Observable evaluation: Using |C_FF[n]| as the target enables diagonal observables, with O(10^2) neurons achieving precisions of O(10^-3) for density correlations and local density.The reported observables are the nearest-neighbor density correlation O_nn and local density O_n1.
- Scaling beyond exact comparison: For larger systems, the study shifts from wave-function overlaps to physical quantities because the full predicted wave function cannot be stored or compared exactly.Importance sampling is used to calculate ground-state physical quantities instead.
IV. NON POSITIVE DEFINITE WAVE FUNCTIONS
The paper addresses non-positive wave functions whose rapidly changing signs challenge standard neural networks. Separating amplitude and sign representations, and modifying the sign network with cosine activations, enables accurate learning of frustrated quantum-magnet ground states.
- Challenge: Standard ANNs fail to approximate some non-positive wave functions because their characteristic functions can change sign sharply with small input changes.This limitation is tied to the sign problem encountered in quantum many-body wave-function approximation.
- Sign representation: A modified sign ANN using cosine activations learns the Marshall sign rule with 100% accuracy, whereas a standard ANN fails to extract it.The cosine activation captures even/odd features in the input data more effectively.
- Frustrated magnetism: For the J1-J2 model, sign-ANN accuracy remains high across 0 ≤ α ≤ 0.5, reaching 99%, with its minimum at the phase transition.The sign function for general α is obtained from finite-size Lanczos calculations and compared with ANN predictions.
- Computational boundary: The transition-time scaling with system size remains unresolved because the studied systems are too small to distinguish polynomial from exponential behavior.The transition time and hidden-neuron count characterize computational resources needed to capture the sign rule.
- Combined approximation: Combining amplitude and sign ANNs produces frustrated-magnet ground states that agree very well with exact results.The two-part architecture is proposed as a strategy for complex quantum many-body systems with sign problems.
V. VARIATIONAL RESULTS
The paper uses ANN trial wave functions with variational Monte Carlo to explore an unknown ground state, optimizing the variational energy rather than fitting a known target. For the 1D L = 30 J1-J2 antiferromagnetic Heisenberg model, a simple architecture reaches ground-state energies with precision ∼O(10^-3).
- V. VARIATIONAL RESULTS: ANNs combined with variational Monte Carlo are used to explore systems whose ground-state wave function is unknown.The variational energy of the ANN trial wave function is minimized during training.
- V. VARIATIONAL RESULTS: The 1D L = 30 J1-J2 antiferromagnetic Heisenberg model is solved using an ANN divided into amplitude and sign parts.The sign output is continuous, and each ANN has one hidden layer.
- V. VARIATIONAL RESULTS: ∼O(10^-3) precision is achieved for the ground-state energy with the simple ANN architecture.The excess energy per site is evaluated relative to the exact energy obtained by density matrix renormalization group.
- V. VARIATIONAL RESULTS: The variational ANN uses only one hidden layer yet performs better than the deeper ANN used for separately trained amplitude and sign functions.The paper attributes this comparison to joint training, simultaneous optimization, and the variational-energy objective.
VI. CONCLUSION AND OUTLOOK
The paper demonstrates that simple ANNs can approximate several notable quantum many-body ground states, but emphasizes that substantial efficiency improvements remain necessary. It identifies symmetry incorporation, deeper or more advanced architectures, and a systematic understanding of validity and limitations as future directions.
- VI. CONCLUSION AND OUTLOOK: Simple ANNs approximate ground-state wave functions of several notable quantum many-body systems.The conclusion presents this as the paper’s demonstrated applicability.
- VI. CONCLUSION AND OUTLOOK: The method remains far from solving problems inaccessible to established numerical methods, making efficiency improvement a key priority.The paper frames this as a remaining challenge rather than a demonstrated capability.
- VI. CONCLUSION AND OUTLOOK: Imposing Hamiltonian symmetries can reduce variational parameters and training data, while two hidden layers outperform one in simulations.Further increasing depth does not significantly improve performance because optimization becomes more costly in a more complex parameter landscape.
- VI. CONCLUSION AND OUTLOOK: Fully connected ANNs may contain redundant connections, and systematic understanding of their validity and limitations is still needed.The paper suggests convolutional architectures as a possible efficiency improvement, especially for short-range entanglement.
Appendix A: Details of the optimization algorithm
The optimization appendix describes stochastic gradient descent and Adam-based training for fitting ANN-predicted wave-function coefficients. It rescales small coefficients, uses training and testing splits with minibatches, and updates parameters through loss gradients with a decreasing learning rate.
- Appendix A: Details of the optimization algorithm: Training adjusts ANN parameters by minimizing a multivariable distance function using SGD or Adam.The appendix explains SGD while referring to another source for Adam.
- Appendix A: Details of the optimization algorithm: For large Hilbert spaces, coefficients scale as C[σ] ∼O(1/D), so the target function is multiplied by a factor of √D to improve fitting efficiency.The renormalization does not change physical observables.
- Appendix A: Details of the optimization algorithm: 80% of the data are used for training and the remainder for testing, with training examples randomly reshuffled into N batches of M data.ANN parameters are initialized randomly before optimization.
- Appendix A: Details of the optimization algorithm: Each SGD step computes a gradient of the loss L(Cp[σ], C[σ]) = (C[σ] − Cp[σ])^2 and updates the parameter vector W.The learning rate ϵ gradually decreases over time.
- Appendix A: Details of the optimization algorithm: Typical training lasts 10^2 ∼10^3 T0, where T0 is the time for one training iteration.The duration depends on problem convergence.
Appendix B: Details of the variational method
For an unknown ground state, the variational method minimizes the ANN trial wave function’s Hamiltonian expectation value using variational Monte Carlo. Monte Carlo importance sampling replaces exponentially large Hilbert-space sums with representative configurations, including amplitude ratios for off-diagonal terms.
- Appendix B: Details of the variational method: When the target function is unknown, training minimizes the variational energy—the Hamiltonian expectation value of the ANN trial wave function.The variational energy is evaluated using variational Monte Carlo.
- Appendix B: Details of the variational method: The Hamiltonian is split into diagonal and non-diagonal parts, Ĥ = T̂ + V̂, for evaluating the trial wave function’s energy.The expectation value is formulated in the basis |σ⟩.
- Appendix B: Details of the variational method: The ANN defines configuration probabilities through pσ{W} = |CANN[σ]|^2 for sampling and evaluating diagonal contributions.The exact sum otherwise spans the exponentially growing Hilbert space.
- Appendix B: Details of the variational method: Importance sampling with the Metropolis algorithm replaces the full Hilbert-space sum by an exponentially small set of representative configurations.Configurations are selected according to their ANN-predicted probabilities.
- Appendix B: Details of the variational method: For off-diagonal terms, sampling evaluates the operator matrix element multiplied by the coefficient ratio CANN[σ̄′]/CANN[σ̄].This ratio supplements the sampled transition matrix element.