Probing Quantum Spin Systems with Kolmogorov-Arnold Neural Network Quantum States

Mahmud Ashraf Shamim, Eric A. F. Reinhardt, Talal Ahmed Chowdhury, Sergei Gleyzer, Paulo T Araujo · 2025

Read Probing Quantum Spin Systems with Kolmogorov-Arnold Neural Network Quantum States by Mahmud Ashraf Shamim, Eric A. F. Reinhardt, Talal Ahmed Chowdhury, Sergei Gleyzer, Paulo T Araujo (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-2506-01891.txt.

Introduction

Shamim and co-authors use a new kind of neural network as a stand-in for the wave function of quantum magnets. Such stand-ins let computers estimate the lowest energy state of many interacting spins. Their 2025 network, SineKAN, builds the wave function from nested single-variable curves with learnable wave shapes.

The team tests it on three standard spin chain models, including a frustrated chain of one hundred sites. The network matched reference energies, overlaps, and correlation patterns well. On the largest chain it beat several earlier network designs, judged against a trusted matrix method. It also trained with modest computer cost. The authors present the design as an efficient tool for probing quantum spin systems.

Catalog record

CREATORMahmud Ashraf Shamim, Eric A. F. Reinhardt, Talal Ahmed Chowdhury, Sergei Gleyzer, Paulo T Araujo
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-01891

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Probing Quantum Spin Systems with Kolmogorov-Arnold Neural Network Quantum States Mahmud Ashraf Shamim, Eric A. F. Reinhardt, Talal Ahmed Chowdhury, Sergei Gleyzer, Paulo T Araujo Abstract Neural Quantum States (NQS) are a class of variational wave functions parametrized by neural networks (NNs) to study quantum many-body systems. In this work, we propose SineKAN , a NQS ansatz based on Kolmogorov-Arnold Networks (KANs), to represent quantum mechanical wave functions as nested univariate functions. We show that SineKAN wavefunction with learnable sinusoidal activation functions can capture the ground state energies, fidelities and various correlation functions of the one dimensional Transverse-Field Ising model, Anisotropic Heisenberg model, and Antiferromagnetic J 1 − J 2 J_{1}-J_{2} model with different chain lengths. In our study of the J 1 − J 2 J_{1}-J_{2} model with L = 100 L=100 sites, we find that the SineKAN model outperforms several previously explored neural quantum state ansätze , including Restricted Boltzmann Machines (RBMs), Long Short-Term Memory models (LSTMs), and Multi-layer Perceptrons (MLP) a.k.a. Feed Forward Neural Networks, when compared to the results obtained from the Density Matrix Renormalization Group (DMRG) algorithm. We find that SineKAN models can be trained to high precisions and accuracies with minimal computational costs. I Introduction Understanding the properties of strongly correlated many-body quantum systems is a major challenge in modern physics. This challenge arises from the complexity of many-body interactions and the exponential growth of the Hilbert space, making exact encoding of the ground state (GS) wave function nearly impossible. Variational methods provide an effective approach to tackle this problem by employing parameterized trial wave functions and minimizing their energy expectation values, thereby enabling an efficient representation of quantum states Becca et al. 2009 ; Becca and Sorella 2017 . Recently, in a seminal work, Carleo and Troyer Carleo and Troyer 2017 introduced a variational ansatz in which a neural network (NN) parametrizes the many-body wave-function. This approach, known as the NN quantum state (NQS), leverages quantum Monte Carlo methods Becca and Sorella 2017 and the expressive power of NNs Lu et al. 2017 to capture complex quantum correlations. This work shows that shallow networks, such as restricted Boltzmann machines ( RBM ), can efficiently approximate the GSs of the quantum Ising and antiferromagnetic Heisenberg models, and are also able to capture volume-law entanglement entropy Deng et al. 2017 , opening a new avenue for representing variational wave functions. As a subsequent development, various NQS architectures, such as Recurrent Neural Networks (RNNs) Hibat-Allah et al. 2020 ; Ahsan Khandoker et al. 2023 ; Roth 2020 ; Dao 2025 , Convolutional Neural Networks (CNNs) Choo et al. 2019 ; Chen et al. 2022 ; Roth and MacDonald 2021 ; Liang et al. 2018 , Autoregressive Models Luo et al. 2023 ; Sharir et al. 2020 , and Vision Transformers (ViTs) Viteritti et al. 2023 ; Loris Viteritti et al. 2023 ; Rende et al. 2023 ; Rende et al. 2025 ; Chen et al. 2025 ; Roca-Jerat et al. 2024 ; Zhang and Di Ventra 2023 , have been proposed, each with varying degrees of effectiveness and expressibility, along with hybrid approaches that combine traditional variational wave functions and NNs Nomura et al. 2017 ; Ferrari et al. 2019 . For a recent discussion and review of the expressibility of the NQS, please see Refs. Yang et al. 2024 ; Lange et al. 2024 . NQSs are generally based on architectures involving feature extraction followed by Multi-Layer Perceptrons (MLPs) Rumelhart et al. 1986 . These classic MLPs are feedforward neural networks with nonlinear activation functions. Figure 1 : Left: NQS based on an MLP with n_{\ell} layers. Spins are encoded as input, combined with learnable weights and biases, and passed through fixed activation functions to produce the wavefunction \Psi_{\texttt{MLP}} . Right: NQS based on SineKAN , where spins are passed through sinusoidal activations with learnable frequencies and amplitudes. A fixed grid of phase shifts is applied across n_{\ell} layers, and outputs are combined to construct \Psi_{\texttt{SineKAN}} . In both cases, the learnable parameters are optimized using VMC ( Appendix A ) to minimize the expectation value of the Hamiltonian H . In a recent work, Liu et al. introduced Kolmogorov-Arnold Networks (KANs) Liu et al. 2024 , offering a promising alternative to traditional MLPs. These networks are motivated by the Kolmogorov-Arnold superposition theorem (KAST) that states that any continuous multivariate function can be represented by a two-layer model composed of univariate functions Kolmogorov 1956 ; Kolmogorov 1957 . The original work explored the use of basis-splines as the learnable, on-edge function. However, a broad class of basis functions are possible, which has motivated the exploration of different KAN architectures beyond the original basis-spline ansatz Liu et al. 2024 . Such work has included rational functions Aghaei 2024a , radial basis functions Ta 2024 , wavelet functions Bozorgasl and Chen 2024 , Jacobi basis functions Aghaei 2024b , Chebyshev polynomials SS et al. 2024 and sinusoidal basis functions Reinhardt et al. 2025 . Numerous follow-up works have shown the viability to various degrees of different KAN architectures across a broad range of tasks Ji et al. 2024 . In this paper, we explore a new class of NQS architecture that utilizes a KAN with sinusoidal activation functions, called the SineKAN ansatz (see fig. 1 ). SineKAN Reinhardt et al. 2025 has already demonstrated superior performance in various tasks compared to other KAN architectures, such as image classification, in terms of numerical accuracy and computational efficiency. The primary motivation for selecting SineKAN as a variational wave function lies in its periodic and smoothly differentiable functional form, which makes it an effective approximator for a variety of functions. We benchmark the SineKAN ansatz by computing the GS energies, fidelity and various correlation functions of the 1D transverse field Ising model (TFIM), anisotropic Heisenberg model (AHM) Lieb et al. 1961 , and antiferromagnetic Heisenberg J_{1}-J_{2} model. In each model for different chain lengths, we find excellent agreement in GS energies and correlation functions determined by Exact Diagonalization (ED) and the state-of-the-art Density Matrix Renormalization Group (DMRG) methods, and SineKAN which required only a modest number of variational parameters. The paper is organized as follows. In section II , we introduce the Kolmogorov Arnold variational wave function and discuss different models we use for benchmarking our architectures in section III . The Hamiltonians we study are discussed in section IV . In Section V we discuss the training methods and the hyperparameters for different Hamiltonians. In section VII , we present the results of our simulations. Section VIII is dedicated to the Marshall Sign Rule (MSR) for SineKAN , and in section IX , we discuss comparative runtime of our model with different models. Section X provides a discussion of the findings, and section XI concludes the paper with a summary and an outlook on future research directions. II Kolmogorov-Arnold Variational Wavefunction For a quantum mechanical system with L sites, where each site is occupied by a single spin- \frac{1}{2} , the dimension of the local Hilbert space is two, corresponding to the two possible spin states. In the computational basis, the spin configuration of the system is represented as \sigma=\{\sigma_{1},\sigma_{2},\cdots,\sigma_{L}\} , where each \sigma_{i} can take the values \uparrow

Read the complete text