Representational power of selected neural network quantum states in second quantization
Zhendong Li, Tong Zhao, Bohan Zhang · 2025
Read Representational power of selected neural network quantum states in second quantization by Zhendong Li, Tong Zhao, Bohan Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-04932.txt.
Introduction
Zhendong Li, Tong Zhao, and Bohan Zhang study how well a class of neural network quantum states can express fermion wave states, when written in second quantization. Such states are a hopeful tool for hard quantum many-body tasks, but their limits for fermions were unclear.
The authors widen the Boltzmann machine form into what they call neuron product states, where each unit links sites across the whole system. They prove such products can match any wave state if the response curves meet mild terms. They also give short proofs for the broad reach of product states of this kind.
Catalog record
| CREATOR | Zhendong Li, Tong Zhao, Bohan Zhang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-04932
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Representational power of selected neural network quantum states in second quantization Zhendong Li, Tong Zhao, Bohan Zhang Abstract Neural network quantum states emerge as a promising tool for solving quantum many-body problems. However, its successes and limitations are still not well-understood in particular for Fermions with complex sign structures. Based on our recent work [J. Chem. Theory Comput. 21, 10252-10262 (2025)], we generalizes the restricted Boltzmann machine to a more general class of states for Fermions, formed by product of ‘neurons’ and hence will be referred to as neuron product states (NPS). NPS builds correlation in a very different way, compared with the closely related correlator product states (CPS) [H. J. Changlani, et al. Phys. Rev. B, 80, 245116 (2009)], which use full-rank local correlators. In constrast, each correlator in NPS contains long-range correlations across all the sites, with its representational power constrained by the simple function form. We prove that products of such simple nonlocal correlators can approximate any wavefunction arbitrarily well under certain mild conditions on the form of activation functions. In addition, we also provide elementary proofs for the universal approximation capabilities of feedforward neural network (FNN) and neural network backflow (NNBF) in second quantization. Together, these results provide a deeper insight into the neural network representation of many-body wavefunctions in second quantization. I Introduction Accurate and efficient simulation of quantum many-body problems on classical computers has been a long-standing challenge for computational physics and chemistry due to the exponential growth of the size of the Hilbert space as system size increases. During the past decades, a plethora of methods have been developed with their own advantages and disadvantages. From the early days of quantum chemistry, the configuration interaction (CI) method 47 is the most conceptually simple approach to treating correlated electrons. Later, the hierarchy of coupled cluster (CC) methods 46 becomes more dominant due to their size extensivity. These two methods are ”universal”, in the sense that any wavefunction can be approximated by increasing the excitation rank. Tensor network states 39 (TNS), which include matrix product states (MPS) as a representative, are better choices for strongly correlated systems 7 . They are also universal, as long as the bond dimension can be arbitrarily large. The universal approximation capability of a wavefunction ansatz is important, because it provides a solid theoretical guarantee for its ultimate accuracy. Recently, neural networks (NN) become an emerging technique for simulation of quantum many-body problems. Carleo and Troyer employed restricted Boltzmann machine 6 (RBM), a generative model that can learn a probability distribution over a set of inputs 26 ; 36 , as a variational ansatz for interacting spin problems on lattices, and achieved good accuracy compared with the state-of-the-art TNS. Unlike TNS, which encodes area-law entanglement efficiently, RBM can describe volume-law states 15 . Later, other machine learning architectures have also been exploited for spin systems, including feedforward NN (FNN) 5 ; 9 , convolutional NN (CNN) 29 ; 11 , recurrent NN (RNN) 21 ; 22 , autoregressive NN 44 ; 1 , neural network backflow (NNBF) 33 ; 32 ; 30 ; 31 , etc. The success and practical limitations of these neural network quantum states (NQS) are not fully understood yet, and are being actively studied 49 ; 4 ; 40 ; 48 ; 45 . Compared with spin systems, the application of neural networks for Fermions remains largely unexplored until very recently 38 ; 52 ; 10 ; 53 ; 1 ; 43 ; 50 . In this work, we focus on Fermions on lattices or electrons in molecules described within the second quantization framework, and we refer the readers to Ref. 20 for applications of NN in the first quantization. Figure 1: Illustration of different neural-network wavefunction ansätze for K=6 : (a) restricted Boltzmann machine (RBM) and neuron product states (NPS); (b) feedforward neural networks (FNN); (c) correlator product states 8 (CPS). Each red circle in (a) and (b) represents a hidden neuron. Inspired by RBM, we recently introduce a more general class of states for fermions composed by product of ’neurons’, which will be referred to as neuron product states 51 (NPS). NPS take the following general form \displaystyle\Psi_{\mathrm{NPS}}(n_{1},\cdots,n_{K})=\prod_{\alpha=1}^{N_{h}}\phi\left(b_{\alpha}+\sum_{k=1}^{K}W_{\alpha k}n_{k}\right), (1) where \Psi_{\mathrm{NPS}}(n_{1},\cdots,n_{K}) is the wavefunction in the occupation number representation, n_{k}\in\{0,1\} , K is the number of spin-orbitals, \phi(x) is an activation function, b_{\alpha} and W_{\alpha k} are real parameters. Each factor \phi in Eq. ( 1 ) is referred to as a neuron, and N_{h} is the number of neurons. We will denote \Psi_{\mathrm{NPS}}(n_{1},\cdots,n_{K}) as \Psi_{\mathrm{NPS}}(\vec{n}) and omit the lower/upper limit of summation/product for brevity. This form generalizes the standard RBM for discrete probability, which after tracing out the hidden units can be written as \displaystyle P_{\mathrm{RBM}}(\vec{n})=e^{\sum_{k}a_{k}n_{k}}\prod_{\alpha}(1+e^{b_{\alpha}+\sum_{k}W_{\alpha k}n_{k}}). (2) Note that each factor in RBM is positive. Carleo and Troyer 6 use complex parameters in RBM for wavefunctions that can take negative values. A general \phi(x) is used in NPS, which can take both positive and negative values. In Ref. 51 , we use \phi(x)=\cos(x) and ( 1 ) is used as a correlator multiplied on another wavefunction ansatz. NPS can be pictorially represented in Fig. ( 1 )a. In this work, we focus on NPS itself and ask a fundamental question that whether it can be a universal approximator. Formally, we can also rewrite Eq. ( 1 ) as \displaystyle\Psi_{\mathrm{NPS}}(\vec{n})=\exp\left[{\sum_{\alpha}\kappa\left(b_{\alpha}+\sum_{k}W_{\alpha k}n_{k}\right)}\right]. (3) with \kappa(x)=\ln\phi(x) . The sum \sum_{a}\kappa(b_{\alpha}+\sum_{k}W_{\alpha k}n_{k}) displays an apparent similarity with FNN. Specifically, the FNN with one hidden layer, see Fig. 1 b, can be written in a similar form as \displaystyle f_{\mathrm{FNN}}(\vec{x})=\sum_{\alpha}c_{\alpha}\sigma(b_{\alpha}+\sum_{k}W_{\alpha k}x_{k}), (4) where \sigma is an activation function. However, an important difference is that there is no linear combination coefficients c_{\alpha} in Eq. ( 3 ). Thus, the classical universal approximation theorem (UAT) 16 ; 14 ; 23 ; 24 ; 28 for multilayer FNN cannot be applied. Similarly, we find that the proof RBM as universal approximators of discrete distributions 26 ; 36 ; 25 cannot be applied to NPS ( 1 ), as these proofs 26 ; 25 hinge on the form of factors 1+e^{x} in Eq. ( 2 ), such that one can modify a particular amplitude of wavefunction by choosing appropriate b_{\alpha} and W_{\alpha k} . In this work, we focus on activation functions that are analytic in certain domain, which means that they can be represented by Taylor series locally, i.e., \phi(x)=\sum_{n=0}^{\infty}\frac{\phi^{(n)}(x_{0})}{n!}(x-x_{0})^{n} . Typical examples include elementary functions (polynomials, exponential, trigonometric, etc.) and piecewise defined functions. As in neural networks, they are the most commonly used forms of activation functions 42 , e.g., logistic sigmoid 1/(1+e^{-x}) , hyperbolic tangent \tanh(x) , and ReLU 18 (rectified linear unit). We will show that the necessary and sufficient conditions for NPS ( 1 ) as a universal approximator for quantum states is that (1) \phi(x) can change signs and (2) the logarithm of the activation function \kappa(x)=\ln\phi(x) is not a polynomial with degree less than the number of orbitals K . This result generalizes the universal property of RBM to a more general class of variational ansatz. Before presenting the proof, we discuss the connections of NPS with other variational ansätze. A closely related class of wavefunctions is the correlator product states (CPS) 8