betaFeedback

WorksPhysics and Astronomy

LaQX: A JAX Toolkit for Neural Quantum State Simulations of Lattice Quantum Many-Body Models

Abstract

Neural quantum states provide accurate variational representations for interacting lattice fermions and spins. Turning a new ansatz into a reproducible calculation, however, requires coordinated implementations of the Hamiltonian, sampling, optimization, observables, and accelerator parallelism. We present LaQX (Lattice Quantum simulation based on jaX), an open-source neural quantum state library based on JAX that organizes these components through common functional interfaces. Beyond ground-state variational Monte Carlo, LaQX supports excited-state calculations, Green's Function Monte Carlo refinement, adiabatic evolution, and real-time evolution with time-dependent variational Monte Carlo. It consolidates methods developed in recent high-accuracy neural quantum state studies of the Hubbard model and frustrated magnets, and the package has been used to investigate the triangular-lattice Hofstadter-Hubbard model. We document these capabilities with reproducible calculations for Hubbard, J1J_1-J2J_2 Heisenberg, altermagnetic Hubbard, and interacting Hofstadter models. Function-oriented modules separate the scientific components and limit the source context needed for extensions by human developers and large language models. Source code, documentation, and executable examples are available at https://github.com/bytedance/laqx.

Introduction

Lattice models connect microscopic Hamiltonians to collective quantum phenomena. Their numerical solution is nevertheless difficult because the Hilbert-space dimension grows exponentially with system size. Exact diagonalization (ED) is consequently restricted to finite clusters [5], density-matrix renormalization group [6] and matrix product state (MPS) methods [7] are most effective in one dimension and quasi-one-dimensional geometries, and higher-dimensional tensor-network contractions require additional approximations [8, 9]. Quantum Monte Carlo [5, 10, 11, 12, 13] can reach much larger systems when a non-negative sampling weight is available, but frustration and fermionic statistics can produce a sign problem whose generic resolution is computationally hard [14, 15]. Comparisons on paradigmatic models such as the Hubbard model [16] therefore continue to rely on complementary numerical methods with different controlled limits [17, 18, 19].

Neural quantum states (NQS) add a flexible class of variational wavefunctions to this landscape. In variational Monte Carlo (VMC) [10, 12], a neural network represents the many-body amplitude, and its parameters are optimized using stochastic estimates over the Born distribution [20]. The field now encompasses a broad range of architectures and applications [21, 22]. Early work combined restricted Boltzmann machines with conventional fermionic pair-product states [23], while convolutional networks addressed frustrated magnets [24] and autoregressive recurrent networks enabled direct sampling [25]. Subsequent developments introduced symmetry-aware deep networks for frustrated spin systems [26], neural backflow and hidden-fermion constructions for lattice fermions [27, 28, 29], and variational protocols for excited states and dynamics [30, 31].

Recent studies have used hidden-fermion Pfaffian states [32], Transformer backflow [1, 33], convolutional backflow [2], and symmetry-preserving pairing states [34] to investigate the ground states of the square-lattice tt-t′t' Hubbard model. Large-scale NQS calculations have also examined topological superconductivity and quantum criticality in the triangular-lattice Hofstadter-Hubbard model [4, 35]. Combining neural networks with tensor networks provides another route to correlated variational states [3]. These studies compare energies and order-specific correlations across variational families to distinguish competing phases. The range of fermionic, spin, excited-state, and dynamical applications motivates software that can accommodate new methods without duplicating the full VMC workflow.

Reusable software has developed alongside these methods. NetKet and jVMC provide general NQS and VMC infrastructure [36, 37], while Quantax is a JAX package for flexible lattice many-body and NQS calculations [38]. In the complementary setting of first-quantized, real-space electronic structure, FermiNet established the accuracy of antisymmetric neural wavefunctions [39, 40], and JaQMC provides an open JAX framework for neural-network VMC and diffusion Monte Carlo [41, 42]. Together with recent GPU-vectorized tensor-network VMC [43], these efforts show that accelerator performance depends not only on the cost of a single wavefunction evaluation, but also on batched sampling, differentiation, and linear algebra. They also establish the value of independent codebases with different physical scope and extension patterns.

Built on JAX [44], LaQX targets interacting lattice fermions and spins. It provides neural backflow and tensor-network states, symmetry projection, Metropolis-Hastings sampling, first-order and natural-gradient optimization, and observable evaluation. The same interfaces support ground-state VMC with optional Green’s Function Monte Carlo (GFMC) [45, 46] refinement, joint optimization of low-lying states with Natural Excited States (NES) [30], measurements on trained states, adiabatic evolution by variational continuation, and real-time evolution with time-dependent VMC (tVMC).

LaQX consolidates the computational framework developed for neural quantum state calculations of the doped Hubbard model, including Transformer-backflow states and the MARCH optimizer [1], followed by the SCALE and ACE convolutional backflow architectures [2]. It also incorporates architecture that combine convolutional neural networks (CNNs) with MPS, denoted CNN-MPS, for frustrated magnets [3]. The convolutional-backflow and CNN-MPS studies report variational energies among the best available for their respective fermionic and spin benchmarks. LaQX has subsequently been applied to the triangular-lattice Hofstadter-Hubbard model, where NQS and independent projected entangled pair state (PEPS) calculations provided complementary real-space evidence of spin-charge separation [4]. The present release brings these developments together in a documented package that supports multiple ansatz families and physical workflows.

The implementation is deliberately function-oriented. Wavefunctions, physical operators, samplers, and parameter updates are connected through small interfaces, with numerical state passed explicitly between them. This separation makes JAX transformations directly applicable and allows many scientific extensions to remain local to one component. It also supports programming with large language models (LLMs), including the conversational workflow known as vibe coding [47]. Limiting cross-component dependencies reduces the source context needed for an extension. § 3.2 illustrates these extension paths.

The remainder of this paper is organized as follows. § 2 defines the numerical methods and notation. § 3 describes the software architecture. § 4 presents calculations for Hubbard, J1J_1-J2J_2 Heisenberg, altermagnetic, and interacting Hofstadter models.

Algorithms

Variational Monte Carlo

Let n=(n1,…,nL)\mathbf{n}=(n_1,\ldots,n_L) contain the LL local occupation variables of a fermionic or spin configuration, and let ψθ(n)\psi_\theta(\mathbf{n}) be its amplitude under a variational state with parameters θ\theta. The variational total energy is

E(θ)=⟨ψθ∣H∣ψθ⟩⟨ψθ∣ψθ⟩.(1)E(\theta)=\frac{\langle\psi_\theta|H|\psi_\theta\rangle}{\langle\psi_\theta|\psi_\theta\rangle}. \tag*{(1)}

Introducing the normalized Born distribution

pθ(n)=∣ψθ(n)∣2∑n′∣ψθ(n′)∣2(2)p_\theta(\mathbf{n})=\frac{|\psi_\theta(\mathbf{n})|^2}{\sum_{\mathbf{n}'}|\psi_\theta(\mathbf{n}')|^2} \tag*{(2)}

with ψθ(n)=⟨n∣ψθ⟩\psi_\theta(\mathbf{n})=\langle\mathbf{n}|\psi_\theta\rangle gives E(θ)=⟨Eloc⟩pθE(\theta)=\langle E_{\mathrm{loc}}\rangle_{p_\theta}, where

Eloc(n)=⟨n∣H∣ψθ⟩ψθ(n).(3)E_{\mathrm{loc}}(\mathbf{n})=\frac{\langle\mathbf{n}|H|\psi_\theta\rangle}{\psi_\theta(\mathbf{n})}. \tag*{(3)}

For real variational coordinates θa\theta_a (including separate real and imaginary coordinates when needed), define the logarithmic derivative Oa(n)=∂θalog⁡ψθ(n)O_a(\mathbf{n})=\partial_{\theta_a}\log\psi_\theta(\mathbf{n}). For a Hermitian Hamiltonian independent of θ\theta, differentiating the total energy E(θ)E(\theta) gives

∂θaE=2Re⁡[⟨Oa∗Eloc⟩−⟨Oa∗⟩⟨Eloc⟩].(4)\partial_{\theta_a}E=2\operatorname{Re}\left[\langle O_a^*E_{\mathrm{loc}}\rangle-\langle O_a^*\rangle\langle E_{\mathrm{loc}}\rangle\right]. \tag*{(4)}

Neural wavefunction architectures

SCALE and ACE are convolutional neural backflow states for lattice fermions [2], while CNN-MPS combines convolutional features with a tensor-network output for spins [3]. Their feature extractors use the same elementary operations: occupation embeddings, spatial convolutions, and site-wise multilayer perceptrons (MLPs) [48]. We define these operations first, then specify how each architecture constructs the many-body amplitude.

Occupation embeddings and convolutional features. Let Ns=LxLyN_s=L_xL_y denote the number of spatial sites, and let qiq_i label the local basis state at site ii. For spinful fermions, qi=ni↑+2ni↓q_i=n_{i\uparrow}+2n_{i\downarrow} distinguishes the four states ∣0⟩|0\rangle, ∣↑⟩|\uparrow\rangle, ∣↓⟩|\downarrow\rangle, and ∣↑↓⟩|\uparrow\downarrow\rangle; for a spin-1/2 or spin-polarized occupation basis, qi∈{0,1}q_i\in\{0,1\}. A trainable, site-dependent lookup table E\mathbf{E} maps this discrete label to a feature vector with dhd_h channels,

hi(0)=Ei,qi∈Rdh.(5)\mathbf{h}_i^{(0)}=\mathbf{E}_{i,q_i}\in\mathbb{R}^{d_h}. \tag*{(5)}

A convolution then mixes features at nearby sites. For a k×kk\times k kernel with odd kk, define Nk={−(k−1)/2,…,(k−1)/2}2\mathcal{N}_k=\{-(k-1)/2,\ldots,(k-1)/2\}^2 and

[C(ℓ)(h)i]β=bβ(ℓ)+∑δ∈Nk∑α=1dhKδαβ(ℓ)hi+δ,α.(6)[\mathcal{C}^{(\ell)}(\mathbf{h})_i]_\beta=b_\beta^{(\ell)}+\sum_{\delta\in\mathcal{N}_k}\sum_{\alpha=1}^{d_h}K_{\delta\alpha\beta}^{(\ell)}h_{i+\delta,\alpha}. \tag*{(6)}

Here ℓ\ell labels the layer, α\alpha and β\beta label input and output channels, and K(ℓ)K^{(\ell)} and b(ℓ)b^{(\ell)} are the trainable kernel and bias shared across sites. Indices wrap in a periodic direction, while features outside an open boundary are set to zero. The examples use the element-wise sigmoid linear unit (SiLU) activation [49], σ(x)=x/(1+e−x)\sigma(x)=x/(1+e^{-x}). A residual block adds the transformed features to its input, hi+σ[C(h)i]\mathbf{h}_i+\sigma[\mathcal{C}(\mathbf{h})_i]. Although convolutional weights are shared, the site-dependent embeddings do not by themselves impose translation symmetry on the state.

After the spatial blocks, an MLP acts separately on each site’s features:

ui(0)=hi(LC),gi=ui(LM),ui(m+1)=σ(W(m)ui(m)+b(m)).(7)\begin{aligned} \mathbf{u}_i^{(0)} &= \mathbf{h}_i^{(L_{\mathrm{C}})}, \qquad\mathbf{g}_i = \mathbf{u}_i^{(L_{\mathrm{M}})},\\ \mathbf{u}_i^{(m+1)} &= \sigma\left(W^{(m)}\mathbf{u}_i^{(m)}+\mathbf{b}^{(m)}\right). \tag*{(7)} \end{aligned}

LCL_{\mathrm{C}} and LML_{\mathrm{M}} count the convolutional and MLP layers. The MLP widths may differ from dhd_h. Its weights are shared across sites, but it introduces no additional spatial mixing: the receptive field, namely the input sites on which gi\mathbf{g}_i depends, is set by the convolutional stack.

Determinantal backflow. For NN fermions, an affine output map produces an orbital table Mθ(n)M_\theta(\mathbf{n}) with NN columns and nsNsn_sN_s rows, where ns=2n_s=2 for spinful and ns=1n_s=1 for spin-polarized fermions:

M(i,s),j(n)=∑μWsjμoutgiμ(n)+bsjout.(8)M_{(i,s),j}(\mathbf{n})=\sum_{\mu}W^{\mathrm{out}}_{sj\mu}g_{i\mu}(\mathbf{n})+b^{\mathrm{out}}_{sj}. \tag*{(8)}

The indices s,j=1,…,Ns,j=1,\ldots,N, and μ=1,…,dg\mu=1,\ldots,d_g label the spin channel, orbital, and output feature of the MLP, respectively; dgd_g is its output width. If Ia(n)I_a(\mathbf{n}) is the aath occupied spin orbital, the single-determinant setting used here is

Φaj(n)=MIa(n),j(n),ψθ(n)=det⁡Φθ(n).(9)\Phi_{aj}(\mathbf{n})=M_{I_a(\mathbf{n}),j}(\mathbf{n}),\qquad\psi_\theta(\mathbf{n})=\det\Phi_\theta(\mathbf{n}). \tag*{(9)}

Thus an N×NN\times N matrix is obtained by selecting occupied rows from the orbital table. Unlike a fixed-orbital Slater determinant, each row depends on the surrounding occupations through the neural features, introducing backflow correlations [27]. ACE can produce complex orbitals by combining separate real and imaginary affine outputs. When requested, lattice-symmetry projection forms a character-weighted sum of transformed amplitudes, including the fermionic reordering signs [26].

SCALE and sparse updates. SCALE (Sparse Convolutional Ansatz for Lattice Electrons) uses one residual convolution without layer normalization [50],

hi(1)=hi(0)+σ[C(0)(h(0))i],(10)\mathbf{h}_i^{(1)}=\mathbf{h}_i^{(0)}+\sigma\left[\mathcal{C}^{(0)}(\mathbf{h}^{(0)})_i\right], \tag*{(10)}

followed by Eqs. (7)–(9). Changing a few occupations modifies only the features within their convolutional neighborhoods; the remaining orbital rows can be reused from a cache. With consistent occupied-row labels, the updated Slater matrix is Φ′=Φ+UVT\Phi'=\Phi+UV^{\mathrm{T}}, where U,V∈CN×rU,V\in\mathbb{C}^{N\times r} and rr counts the affected rows. UU selects these rows and VTV^{\mathrm{T}} contains their changes. The matrix determinant lemma and Sherman-Morrison-Woodbury identity give

Br=Ir+VTΦ−1U,det⁡Φ′det⁡Φ=det⁡Br,(Φ′)−1=Φ−1−Φ−1UBr−1VTΦ−1.(11)\begin{aligned} B_r &= I_r+V^{\mathrm{T}}\Phi^{-1}U,\qquad\frac{\det\Phi'}{\det\Phi}=\det B_r,\\ (\Phi')^{-1} &= \Phi^{-1}-\Phi^{-1}UB_r^{-1}V^{\mathrm{T}}\Phi^{-1}. \tag*{(11)} \end{aligned}

Here IrI_r is the r×rr\times r identity, and T\mathrm{T} denotes the algebraic transpose. Only an r×rr\times r matrix must be factorized; updating the cached inverse costs O(N2r+r3)O(N^2r+r^3) rather than a fresh O(N3)O(N^3) factorization. At fixed density, kernel size, and number of changed occupations, rr stays bounded as the lattice grows. Together with local feature recomputation, this reduces the determinant-update cost of a sweep of O(N)O(N) local moves from O(N4)O(N^4) to O(N3)O(N^3) [2].

ACE and deep convolutional features. ACE (Accurate Convolutional ansatz for lattice Electrons) replaces Eq. (10) by a deep stack,

hi(ℓ+1)=LNℓ(hi(ℓ)+σ[C(ℓ)(h(ℓ))i]),ℓ=0,…,LC−1.(12)\mathbf{h}_i^{(\ell+1)}=\mathrm{LN}_{\ell}\left(\mathbf{h}_i^{(\ell)}+\sigma\left[\mathcal{C}^{(\ell)}\left(\mathbf{h}^{(\ell)}\right)_i\right]\right),\qquad\ell=0,\ldots,L_C-1. \tag*{(12)}

Layer normalization [50] is applied after the residual addition, independently at each site:

[LNℓ(zi)]α=γℓαziα−zˉivi+ϵLN+βℓα,(13)[\mathrm{LN}_{\ell}(z_i)]_{\alpha}=\gamma_{\ell\alpha}\frac{z_{i\alpha}-\bar{z}_i}{\sqrt{v_i+\epsilon_{\mathrm{LN}}}}+\beta_{\ell\alpha}, \tag*{(13)}
zˉi=1dh∑αziα,vi=1dh∑α(ziα−zˉi)2.\bar{z}_i=\frac{1}{d_h}\sum_{\alpha}z_{i\alpha},\qquad v_i=\frac{1}{d_h}\sum_{\alpha}(z_{i\alpha}-\bar{z}_i)^2.

Here zˉi\bar{z}_i and viv_i are the mean and variance across channels at site ii; γℓα\gamma_{\ell\alpha} and βℓα\beta_{\ell\alpha} are the trainable scale and offset. ACE uses ϵLN=10−4\epsilon_{\mathrm{LN}}=10^{-4}. Each layer enlarges the receptive field before the same MLP and orbital output are applied. A sufficiently deep stack allows a local occupation change to affect features throughout the lattice, and the ACE implementation reevaluates the full network rather than using SCALE’s sparse updates. SCALE and ACE therefore trade strict update locality for deeper, more expressive backflow features. The Hubbard benchmarks in Ref. [2] report a more favorable energy-runtime trade-off for ACE than for the Transformer-backflow architectures of Ref. [1] in the systems and configurations compared.

CNN-MPS for spins. For spins, the local-state label is qi=ni∈{0,1}q_i=n_i\in\{0,1\}, with ni=0n_i=0 and 11 denoting spin down and spin up, respectively; n=(n1,…,nNs)\mathbf{n}=(n_1,\ldots,n_{N_s}) specifies the spin configuration. The CNN-MPS implementation initializes its features with trainable vectors eq,pi∈Rdh\mathbf{e}_{q},\mathbf{p}_i\in\mathbb{R}^{d_h}: eq\mathbf{e}_q embeds local state qq and is shared across sites, while pi\mathbf{p}_i encodes the position of site ii,

hi(0)=eqi+pi.(14)\mathbf{h}_i^{(0)}=\mathbf{e}_{q_i}+\mathbf{p}_i. \tag*{(14)}

It then uses the residual convolutional stack in Eq. (12), with layer normalization at every site, and the site-wise MLP in Eq. (7). In this architecture, each layer’s normalization scale and offset are scalars shared across channels, and ϵLN=10−5\epsilon_{\mathrm{LN}}=10^{-5}. Let dgd_g be the number of output features of the MLP, so giμ(n)g_{i\mu}(\mathbf{n}) is component μ=1,…,dg\mu=1,\ldots,d_g of gi(n)\mathbf{g}_i(\mathbf{n}). A tensor-valued linear output replaces the orbital map, producing a matrix Ai(r),q(n)∈RD×DA_i^{(r),q}(\mathbf{n})\in\mathbb{R}^{D\times D} for each local spin state qq and head rr,

[Ai(r),q(n)]ab=∑μ=1dgTabμ(r),qgiμ(n),a,b=1,…,D.(15)\left[A_i^{(r),q}(\mathbf{n})\right]_{ab}=\sum_{\mu=1}^{d_g}T_{ab\mu}^{(r),q}g_{i\mu}(\mathbf{n}),\qquad a,b=1,\ldots,D. \tag*{(15)}

Here aa and bb are the left and right MPS bond indices, and DD is the bond dimension. The index rr labels one of the NheadN_{\mathrm{head}} MPS heads. The trainable output weights Tabμ(r),qT_{ab\mu}^{(r),q} are shared across sites. Selecting q=niq=n_i and contracting the bond indices with periodic MPS boundary conditions gives, before lattice-symmetry projection,

ψθ(n)=∑r=1NheadTr⁡[∏i=1NsAi(r),ni(n)].(16)\psi_{\theta}(\mathbf{n})=\sum_{r=1}^{N_{\mathrm{head}}}\operatorname{Tr}\left[\prod_{i=1}^{N_s}A_i^{(r),n_i}(\mathbf{n})\right]. \tag*{(16)}

The ordered matrix product follows i=xLy+y+1i=xL_y+y+1, with lattice coordinates x=0,…,Lx−1x=0,\ldots,L_x-1 and y=0,…,Ly−1y=0,\ldots,L_y-1; Tr contracts the last MPS bond with the first. Each head supplies a separate set of output weights while sharing the neural features, and the contracted amplitudes are summed. Because these tensors depend on occupations beyond their own site, this is a backflow MPS, not an ordinary MPS with fixed local tensors. In the neural tensor-network interpretation of Ref. [3], the convolutional map acts as a generally nonunitary disentangler and the bond dimension DD controls the tensor-network representation of the remaining correlations.

Sampling and ground-state optimization

Metropolis-Hastings sampling. A Markov chain samples configurations from pθp_\theta [51, 52]. LaQX constructs proposals from the off-diagonal terms of the Hamiltonian: an allowed hopping or spin-exchange term generates a candidate configuration n′\mathbf{n}' from n\mathbf{n}. A proposal n′∼q(n′∣n)\mathbf{n}' \sim q(\mathbf{n}'|\mathbf{n}) is accepted with probability

a(n→n′)=min⁡{1,∣ψθ(n′)∣2q(n∣n′)∣ψθ(n)∣2q(n′∣n)}.(17)a(\mathbf{n} \to\mathbf{n}') = \min\left\{1, \frac{|\psi_\theta(\mathbf{n}')|^2 q(\mathbf{n}|\mathbf{n}')}{|\psi_\theta(\mathbf{n})|^2 q(\mathbf{n}'|\mathbf{n})}\right\}. \tag*{(17)}

Otherwise the chain remains at n\mathbf{n}. Including the forward-reverse proposal ratio accommodates asymmetric proposals while satisfying detailed balance with the Born distribution. Local occupation moves preserve the sampled particle-number or magnetization sector.

Stochastic reconfiguration. LaQX provides Adam [53], SPRING [54], and MARCH [1]. Adam uses the energy-gradient covariance in Eq. (4) to construct adaptive first- and second-moment updates. The latter two methods are derived from stochastic reconfiguration (SR) [55]: instead of descending in the Euclidean geometry of the parameters, SR projects an imaginary-time step onto the tangent space of the variational state.

Let θ=(θ1,…,θP)\theta= (\theta_1,\ldots,\theta_P) contain the PP real trainable coordinates, with indices a,b=1,…,Pa,b=1,\ldots,P. Brackets denote averages over the Born distribution pθp_\theta, and ∗^* denotes complex conjugation. Using the logarithmic derivatives OaO_a defined above, introduce the centered quantities ΔOa=Oa−⟨Oa⟩\Delta O_a = O_a - \langle O_a\rangle and ΔEloc=Eloc−⟨Eloc⟩\Delta E_{\mathrm{loc}} = E_{\mathrm{loc}} - \langle E_{\mathrm{loc}}\rangle. The derivative covariance matrix SS and energy-covariance vector FF have entries

Sab=⟨ΔOa∗ΔOb⟩,Fa=⟨ΔOa∗ΔEloc⟩.(18)S_{ab} = \langle\Delta O_a^* \Delta O_b\rangle,\qquad F_a = \langle\Delta O_a^* \Delta E_{\mathrm{loc}}\rangle. \tag*{(18)}

At optimization step kk, let d=(d1,…,dP)∈RPd=(d_1,\ldots,d_P)\in\mathbb{R}^P be a trial parameter-update direction, with dad_a specifying the change in coordinate θa\theta_a before step-size scaling. All averages are evaluated at the current parameters θk\theta_k. Fitting the normalized imaginary-time derivative −(H−E)∣ψθ⟩-(H-E)|\psi_\theta\rangle gives the minimizing direction dSRd_{\mathrm{SR}} and parameter update

dSR=arg⁡min⁡d∈RP{⟨∣∑a=1PΔOada+ΔEloc∣2⟩+λ∥d∥22},d_{\mathrm{SR}} = \underset{d\in\mathbb{R}^P}{\arg\min}\left\{\left\langle\left|\sum_{a=1}^{P}\Delta O_a d_a+\Delta E_{\mathrm{loc}}\right|^2\right\rangle+\lambda\|d\|_2^2\right\},
(Re⁡S+λIP)dSR=−Re⁡F,θk+1=θk+ηkdSR.(19)(\operatorname{Re}S+\lambda I_P)d_{\mathrm{SR}}=-\operatorname{Re}F,\qquad\theta_{k+1}=\theta_k+\eta_kd_{\mathrm{SR}}. \tag*{(19)}

The real parts arise from minimizing over real parameter changes dd, even when the wavefunction and its logarithmic derivatives are complex. Here λ>0\lambda>0 is the damping parameter, IPI_P the P×PP\times P identity matrix, and ∥⋅∥2\|\cdot\|_2 is the Euclidean norm. The step size is ηk\eta_k.

For a batch of BB configurations nν\mathbf{n}_\nu, with sample index ν=1,…,B\nu=1,\ldots,B, let O‾a\overline{O}_a and E‾\overline{E} denote the sample means of OaO_a and ElocE_{\mathrm{loc}}, respectively. Define the centered derivative matrix JJ and target vector ϵ\epsilon by

Jνa=Oa(nν)−O‾aB,ϵν=−Eloc(nν)−E‾B.(20)J_{\nu a}=\frac{O_a(\mathbf{n}_\nu)-\overline{O}_a}{\sqrt{B}},\qquad\epsilon_\nu=-\frac{E_{\mathrm{loc}}(\mathbf{n}_\nu)-\overline{E}}{\sqrt{B}}. \tag*{(20)}

For a real wavefunction set X=JX=J and e=ϵe=\epsilon. For a complex wavefunction with real trainable coordinates, stack the two components, X=(Re⁡J;Im⁡J)X=(\operatorname{Re}J;\operatorname{Im}J) and e=(Re⁡ϵ;Im⁡ϵ)e=(\operatorname{Re}\epsilon;\operatorname{Im}\epsilon). Then XTXX^{\mathsf{T}}X estimates Re⁡S\operatorname{Re}S and XTeX^{\mathsf{T}}e estimates −Re⁡F-\operatorname{Re}F. The damped SR direction has the equivalent forms

dSR=(XTX+λIP)−1XTe=XT(XXT+λI)−1e.(21)d_{\mathrm{SR}}=(X^{\mathsf{T}}X+\lambda I_P)^{-1}X^{\mathsf{T}}e=X^{\mathsf{T}}(XX^{\mathsf{T}}+\lambda I)^{-1}e. \tag*{(21)}

Here II is the identity in sample space, whose dimension is BB for real states and 2B2B for complex states. The matrix XX has PP columns and one row per sample-space component, and ee has the corresponding length. The second form reduces the solve to sample space, as in minimum-step stochastic reconfiguration (MinSR) and related formulations [56, 57].

In the implementation, local energies are clipped before centering at the sample median plus or minus a chosen multiple of the mean absolute deviation from that median. Real and imaginary parts are clipped separately. Reported energies and energy variances per site use the unclipped local energies.

SPRING. Subsampled Projected-Increment Natural Gradient Descent (SPRING) retains the preceding unscaled direction rather than selecting a minimum-norm solution independently for each batch [54]. Subscripts kk on XkX_k and eke_k denote the sample arrays at optimization step kk. Let dk−1d_{k-1} be the stored direction, μ\mu the momentum decay coefficient, and mk=μdk−1m_k = \mu d_{k-1} the decayed direction. SPRING obtains the new unscaled direction dkd_k by solving

dk=arg min⁡d∈RP{∥Xkd−ek∥22+λ∥d−mk∥22}(22)d_k = \underset{d \in\mathbb{R}^{P}}{\operatorname{arg\,min}}\left\{\lVert X_kd-e_k\rVert_2^2+\lambda\lVert d-m_k\rVert_2^2\right\} \tag*{(22)}

Its sample-space expression is

dk=mk+XkT(XkXkT+λI+Rk)−1(ek−Xkmk).(23)d_k=m_k+X_k^{\mathsf{T}}\left(X_kX_k^{\mathsf{T}}+\lambda I+R_k\right)^{-1}\left(e_k-X_km_k\right). \tag*{(23)}

Thus the new batch corrects the part of the previous direction inconsistent with the current imaginary-time target; this differs from simply adding momentum to an independent SR update. The extra stabilizer is Rk=∥Tk∥FΠ0R_k=\lVert T_k\rVert_{\mathrm{F}}\Pi_0, where Tk=XkXkTT_k=X_kX_k^{\mathsf{T}} is the sample-space Gram matrix and ∥⋅∥F\lVert\cdot\rVert_{\mathrm{F}} is the Frobenius norm. For real states, Π0=1B1BT/B\Pi_0=\mathbf{1}_B\mathbf{1}_B^{\mathsf{T}}/B projects onto the constant sample vector, with 1B\mathbf{1}_B the length-BB vector of ones. For complex states it projects separately onto the constant vectors of the real and imaginary sample blocks. Centering gives XkTΠ0=0X_k^{\mathsf{T}}\Pi_0=0 and Π0ek=0\Pi_0e_k=0, so this term does not change the exact-arithmetic solution of Eq. (22); it stabilizes the numerical solve in these null directions.

MARCH. The Moment-Adaptive ReConfiguration Heuristic (MARCH) extends this construction by weighting parameter directions according to a second-moment accumulator [1]. In the notation of Eq. (20), the minimization objective is

dk=arg min⁡d∈RP{1λ∥Xkd−ek∥22+∥diag⁡(vk−1)1/4(d−mk)∥22}.(24)d_k=\underset{d\in\mathbb{R}^{P}}{\operatorname{arg\,min}}\left\{\frac{1}{\lambda}\lVert X_kd-e_k\rVert_2^2+\left\lVert\operatorname{diag}\left(v_{k-1}\right)^{1/4}(d-m_k)\right\rVert_2^2\right\}. \tag*{(24)}

Here mk=μdk−1m_k=\mu d_{k-1} is the decayed previous direction and vk−1v_{k-1} is the second-moment accumulator, with one positive component per parameter. The operation diag⁡(v)\operatorname{diag}(v) places the components of vv on a diagonal matrix, and its fourth root acts on those entries. Large accumulator components increase the penalty for changing the corresponding parameters. The implementation optionally rescales the accumulator to v~k−1\widetilde{v}_{k-1} and adds a small positive ϵ\epsilon to form the diagonal penalty metric DkD_k:

v~k−1={P vk−1/∥vk−1∥2,with accumulator normalization,vk−1,without accumulator normalization,\widetilde{v}_{k-1}= \begin{cases} \sqrt{P}\,v_{k-1}/\lVert v_{k-1}\rVert_2, & \text{with accumulator normalization},\\ v_{k-1}, & \text{without accumulator normalization}, \end{cases}
Dk=diag⁡(v~k−1+ϵ).(25)D_k=\operatorname{diag}\left(\sqrt{\widetilde{v}_{k-1}}+\epsilon\right). \tag*{(25)}

Square roots act component-wise, and accumulator normalization is enabled by default. With the penalty in Eq. (24) evaluated as (d−mk)TDk(d−mk)(d-m_k)^{\mathsf{T}}D_k(d-m_k), the update is

δk=Dk−1XkT(XkDk−1XkT+λI+Rk)−1(ek−Xkmk),dk=mk+δk,vk=β2v~k−1+δk⊙δk.(26)\begin{aligned} \delta_k&=D_k^{-1}X_k^{\mathsf{T}}\left(X_kD_k^{-1}X_k^{\mathsf{T}}+\lambda I+R_k\right)^{-1}\left(e_k-X_km_k\right),\\ d_k&=m_k+\delta_k,\qquad v_k=\beta_2\widetilde{v}_{k-1}+\delta_k\odot\delta_k. \tag*{(26)} \end{aligned}

Here δk\delta_k is the correction to the decayed direction, β2\beta_2 is the accumulator decay coefficient, and ⊙\odot denotes component-wise multiplication. The stabilizer is Rk=∥XkDk−1XkT∥Π0Π0R_k = \lVert X_kD_k^{-1}X_k^{\mathsf{T}}\rVert_{\Pi_0} \Pi_0. The initial direction is zero and the initial accumulator has equal positive entries.

Both implementations store dkd_k before step-length normalization. With ℓk\ell_k denoting the scheduled Euclidean length of the parameter change, they apply

θk+1=θk+ℓkdk∥dk∥2.(27)\theta_{k+1} = \theta_k + \ell_k \frac{d_k}{\lVert d_k\rVert_2}. \tag*{(27)}

Let ℓbase\ell_{\mathrm{base}} be the base step length. The schedules are

ℓkSPRING=ℓbase1+k/k0,\ell_k^{\mathrm{SPRING}} = \frac{\ell_{\mathrm{base}}}{1+k/k_0},
ℓkMARCH=ℓbasemin⁡{kkw,11+max⁡(k−ks,0)/k0}.(28)\ell_k^{\mathrm{MARCH}} = \ell_{\mathrm{base}} \min\left\{\frac{k}{k_w},\frac{1}{1+\max(k-k_s,0)/k_0}\right\}. \tag*{(28)}

Here k0k_0 is the decay timescale, kwk_w the warmup duration, and ksk_s the decay-start step.

Natural Excited States

The natural excited states (NES) method turns the calculation of the lowest KK states into a ground-state variational problem on the tensor product of KK replicas of the physical Hilbert space [30]. Let N=(n1,…,nK)\mathcal{N}=(n_1,\ldots,n_K) contain one many-body configuration per replica. Given KK trainable state functions ψa\psi_a, define the matrix with replica index bb labeling rows and state index aa labeling columns, where a,b=1,…,Ka,b=1,\ldots,K,

Mba(N)=ψa(nb),Ψ(N)=det⁡M(N).(29)M_{ba}(\mathcal{N})=\psi_a(n_b),\qquad\Psi(\mathcal{N})=\det M(\mathcal{N}). \tag*{(29)}

The expanded Hamiltonian is the Kronecker sum of KK copies of the physical Hamiltonian HH,

H=⨁b=1KH=∑b=1KH(b),\mathcal{H}=\bigoplus_{b=1}^{K}H=\sum_{b=1}^{K}H^{(b)},
H(b)=I⊗(b−1)⊗H⊗I⊗(K−b).(30)H^{(b)}=I^{\otimes(b-1)}\otimes H\otimes I^{\otimes(K-b)}. \tag*{(30)}

Here ⨁\bigoplus denotes the Kronecker sum, ⊗\otimes the tensor product, and II the identity on one replica, with I⊗0=1I^{\otimes0}=1. Thus H(b)H^{(b)} acts only on replica bb. The determinant makes Ψ\Psi antisymmetric under exchange of replicas. If E1≤E2≤⋯E_1\le E_2\le\cdots are the eigenvalues of HH, the lowest eigenvalue of H\mathcal{H} in this antisymmetric sector is the sum of the lowest KK physical energies. The variational objective is therefore

E=⟨Ψ∣H∣Ψ⟩⟨Ψ∣Ψ⟩≥∑a=1KEa.(31)\mathcal{E}=\frac{\langle\Psi|\mathcal{H}|\Psi\rangle}{\langle\Psi|\Psi\rangle}\ge\sum_{a=1}^{K}E_a. \tag*{(31)}

Linearly dependent state functions make the determinant identically zero, excluding collapse onto fewer than KK independent states without pairwise orthogonality penalties.

LaQX samples pNES(N)=∣det⁡M∣2∑N′∣det⁡M(N′)∣2p_{\mathrm{NES}}(\mathcal{N})=\frac{|\det M|^2}{\sum_{\mathcal{N}'}|\det M(\mathcal{N}')|^2}. A Metropolis-Hastings proposal changes one replica and uses the ratio of these determinants in Eq. (17). Define the matrix Bba(N)=(Hψa)(nb)\mathcal{B}_{ba}(\mathcal{N})=(H\psi_a)(n_b). Acting with H(b)H^{(b)} replaces row bb of MM by the corresponding row of B\mathcal{B}; summing these row replacements gives

Eloc(N)=M−1B,E=⟨Tr⁡Eloc⟩pNES,E_{\mathrm{loc}}(\mathcal{N})=M^{-1}\mathcal{B},\qquad\mathcal{E}=\langle\operatorname{Tr}E_{\mathrm{loc}}\rangle_{p_{\mathrm{NES}}},
(HΨ)(N)Ψ(N)=Tr⁡Eloc(N).(32)\frac{(\mathcal{H}\Psi)(\mathcal{N})}{\Psi(\mathcal{N})}=\operatorname{Tr}E_{\mathrm{loc}}(\mathcal{N}). \tag*{(32)}

Consequently, the same energy-gradient and optimizer formulas apply to log⁡Ψ\log\Psi and the scalar trace. ACE-NES supplies the individual functions using a shared ACE feature extractor and KK state-resolved orbital outputs. Each ψa\psi_a is itself a fermionic determinant over occupied orbitals; the additional determinant in Eq. (29) acts on whole many-body states, not on individual particles.

After optimization, the full energy matrix is accumulated rather than only its trace. Let Gab=⟨ψa∣ψb⟩G_{ab}=\langle\psi_a|\psi_b\rangle and Hab=⟨ψa∣H∣ψb⟩H_{ab}=\langle\psi_a|H|\psi_b\rangle. For linearly independent functions, its exact expectation satisfies [30]

E‾=⟨Eloc⟩pNES=G−1H,E‾ca=Eaca⟺Hca=EaGca.(33)\overline{E}=\langle E_{\mathrm{loc}}\rangle_{p_{\mathrm{NES}}}=G^{-1}H, \qquad \overline{E}c_a=\mathcal{E}_a c_a \Longleftrightarrow Hc_a=\mathcal{E}_aGc_a. \tag*{(33)}

Diagonalizing the mean matrix gives the Ritz energies Ea\mathcal{E}_a and coefficient vectors cac_a for the corresponding states ∑b(ca)bψb\sum_b(c_a)_b\psi_b.

Green's Function Monte Carlo projection

Lattice fixed-node Green’s Function Monte Carlo (GFMC) uses imaginary-time propagation to find the ground state of an effective Hamiltonian HFNH^{\mathrm{FN}} constructed from a guiding state [45, 46]. The neural parameters remain fixed and a weighted walker population represents the evolving state. On configurations where ψG\psi_G is nonzero, introduce the importance-sampled matrix

hn′n=ψG(n′)ψG(n)Hn′n,Hn′n=⟨n′∣H∣n⟩.(34)h_{n'n}=\frac{\psi_G(n')}{\psi_G(n)}H_{n'n}, \qquad H_{n'n}=\langle n'|H|n\rangle. \tag*{(34)}

The imaginary-time distribution is f(n,τ)=ψG(n)ψ(n,τ)f(n,\tau)=\psi_G(n)\psi(n,\tau). Off-diagonal hn′n>0h_{n'n}>0 would produce negative transition weights in a short-time propagator. The fixed-node construction removes these transitions and transfers their contributions to the diagonal [58]:

H~n′nFN={hn′n,n′≠n, hn′n<0,0,n′≠n, hn′n≥0,Hnn+∑m≠nmax⁡(hmn,0),n′=n.(35)\widetilde{H}^{\mathrm{FN}}_{n'n}= \begin{cases} h_{n'n}, & n'\ne n,\ h_{n'n}<0,\\ 0, & n'\ne n,\ h_{n'n}\ge0,\\ H_{nn}+\sum_{m\ne n}\max(h_{mn},0), & n'=n. \end{cases} \tag*{(35)}

The matrix H~FN\widetilde{H}^{\mathrm{FN}} represents HFNH^{\mathrm{FN}} after importance sampling. The fixed-node correction preserves the guiding-state local energy, so the walker weights and mixed energy estimate use

ElocG(n)=(HFNψG)(n)ψG(n)=(HψG)(n)ψG(n).(36)E_{\mathrm{loc}}^G(n)=\frac{(H^{\mathrm{FN}}\psi_G)(n)}{\psi_G(n)}=\frac{(H\psi_G)(n)}{\psi_G(n)}. \tag*{(36)}

Together with the positive semidefiniteness of HFN−HH^{\mathrm{FN}}-H, this gives the variational bound [58]

E0≤E0FN≤E[ψG],E[ψG]=⟨ψG∣H∣ψG⟩⟨ψG∣ψG⟩.(37)E_0\le E_0^{\mathrm{FN}}\le E[\psi_G], \qquad E[\psi_G]=\frac{\langle\psi_G|H|\psi_G\rangle}{\langle\psi_G|\psi_G\rangle}. \tag*{(37)}

Here E0E_0 and E0FNE_0^{\mathrm{FN}} are the ground-state energies of HH and HFNH^{\mathrm{FN}}, respectively. The fixed-node bias vanishes when the guiding function is an exact ground state.

For a sufficiently small uniform step Δτ\Delta\tau, the nonnegative linear propagator is

G~=I−Δτ(H~FN−ETI),(38)\widetilde{G}=I-\Delta\tau(\widetilde{H}^{\mathrm{FN}}-E_T I), \tag*{(38)}

where ETE_T is a reference energy and II is the identity in configuration space. LaQX limits the step at each walker to avoid a negative diagonal entry. Writing dn=H~nnFNd_n=\widetilde{H}^{\mathrm{FN}}_{nn} and denoting the requested step by Δτ0\Delta\tau_0, the implemented rule is

Δτn={(dn−ET)−1,dn−ET>1/Δτ0,Δτ0,otherwise.(39)\Delta\tau_n= \begin{cases} (d_n-E_T)^{-1}, & d_n-E_T>1/\Delta\tau_0,\\ \Delta\tau_0, & \text{otherwise}. \end{cases} \tag*{(39)}

For each walker, the transition and weight update are then

G~n′n=δn′n−Δτn(H~n′nFN−ETδn′n),b(n)=∑n′G~n′n=1+Δτn[ET−ElocG(n)],p(n′∣n)=G~n′n/b(n),w′=wb(n).(40)\begin{aligned} \widetilde{G}_{\mathbf{n}'\mathbf{n}} &= \delta_{\mathbf{n}'\mathbf{n}}-\Delta\tau_{\mathbf{n}}\left(\widetilde{H}^{\mathrm{FN}}_{\mathbf{n}'\mathbf{n}}-E_T\delta_{\mathbf{n}'\mathbf{n}}\right),\\ b(\mathbf{n}) &= \sum_{\mathbf{n}'}\widetilde{G}_{\mathbf{n}'\mathbf{n}}=1+\Delta\tau_{\mathbf{n}}\left[E_T-E_{\mathrm{loc}}^{G}(\mathbf{n})\right],\\ p(\mathbf{n}'\mid\mathbf{n}) &= \widetilde{G}_{\mathbf{n}'\mathbf{n}}/b(\mathbf{n}),\qquad w'=wb(\mathbf{n}). \tag*{(40)} \end{aligned}

Here δ\delta is the Kronecker delta, bb is the column sum, and ww and w′w' are the walker weights before and after propagation. This update samples a propagator column directly, including its diagonal stay-put contribution. For NwN_w walkers with configurations np\mathbf{n}_p and weights wpw_p, the mixed energy estimate is

Emix=∑p=1NwwpElocG(np)∑p=1Nwwp.(41)E_{\mathrm{mix}}=\frac{\sum_{p=1}^{N_w}w_pE_{\mathrm{loc}}^{G}(\mathbf{n}_p)}{\sum_{p=1}^{N_w}w_p}. \tag*{(41)}

Adiabatic evolution

Adiabatic evolution is approximated by a sequence of variational optimizations for nearby Hamiltonians H(φj)H(\varphi_j), where jj labels successive values of a control parameter φ\varphi. Starting from a prepared state, the optimized parameters at φj\varphi_j initialize the calculation at φj+1\varphi_{j+1}:

θj+1=arg min⁡θ←θj⟨ψθ∣H(φj+1)∣ψθ⟩⟨ψθ∣ψθ⟩.(42)\theta_{j+1}=\underset{\theta\leftarrow\theta_j}{\operatorname{arg\,min}}\frac{\langle\psi_\theta|H(\varphi_{j+1})|\psi_\theta\rangle}{\langle\psi_\theta|\psi_\theta\rangle}. \tag*{(42)}

The arrow denotes initialization at θj\theta_j. Small parameter increments and a fixed symmetry sector allow the optimization to follow a continuous low-energy branch.

In the charge-pumping application [59], φ\varphi is a boundary twist increased from zero to 2π2\pi. In the Hofstadter implementation, the xx-direction hopping takes the uniform-gauge form

Hx(φ)=−∑x,y(eiφ/Lxcx+1,y†cx,y+H.c.).(43)H_x(\varphi)=-\sum_{x,y}\left(e^{i\varphi/L_x}c^\dagger_{x+1,y}c_{x,y}+\mathrm{H.c.}\right). \tag*{(43)}

Thus hopping in the positive xx direction carries eiφ/Lxe^{i\varphi/L_x}, and reverse hopping carries its complex conjugate. The accumulated phase around the periodic xx direction is eiφe^{i\varphi}, so increasing φ\varphi by 2π2\pi inserts one flux quantum through the torus. Each step is followed by a polarization measurement. Continuing the low-energy momentum sectors separately gives the charge-pumping response described in § 4.5.

Time-dependent VMC and parameter-efficient evolution

ACE-PEFT. Parameter-efficient fine-tuning (PEFT) restricts time evolution to the final affine orbital map of a prepared ACE state, adapting the neural-basis tVMC principle of Ref. [60]. Let dgd_g be the number of features in gig_i from Eq. (7). Define g~i0=1\widetilde{g}_{i0}=1 to include the output bias and g~iμ=giμ\widetilde{g}_{i\mu}=g_{i\mu} for μ=1,…,dg\mu=1,\ldots,d_g, so that

M(i,s),j(n,t)=∑μ=0dgWsjμ(t)g~iμ(n;η0),M_{(i,s),j}(\mathbf{n},t)=\sum_{\mu=0}^{d_g}W_{s j\mu}(t)\widetilde{g}_{i\mu}(\mathbf{n};\eta_0),
ψW(t)(n)=det⁡[MIa(n),j(n,t)]a,j=1N.(44)\psi_{W(t)}(\mathbf{n})=\det\left[M_{I_a(\mathbf{n}),j}(\mathbf{n},t)\right]_{a,j=1}^{N}. \tag*{(44)}

Here ii, ss, and jj label the site, spin channel, and orbital, and Ia(n)I_a(\mathbf{n}) selects the occupied spin orbitals as in Eq. (9). The embedding, convolutional, layer-normalization, and MLP parameters η0\eta_0 remain fixed. Only the complex coefficients W(t)W(t), including the bias coefficients Wsj0W_{s j0}, evolve. Freezing the nonlinear feature extractor reduces the number of dynamical parameters while retaining its configuration-dependent correlations.

Projected real-time dynamics. Time-dependent VMC approximates i∂t∣ψ⟩=H∣ψ⟩i\partial_t|\psi\rangle= H|\psi\rangle, with ℏ=1\hbar= 1, by evolution within a variational manifold [31, 61]. Collect the PP evolved coefficients of WW into θ∈CP\theta\in\mathbb{C}^{P}. The ACE-PEFT amplitude is holomorphic in these coefficients, and its logarithmic derivatives Oa=∂θalog⁡ψθO_a = \partial_{\theta_a}\log\psi_\theta, a=1,…,Pa = 1,\ldots,P, are evaluated by automatic differentiation. With the centered quantities and Born-distribution averages of § 2.3, the time-dependent variational principle (TDVP) determines the parameter velocity θ˙=dθ/dt\dot{\theta} = d\theta/dt by

θ˙=arg⁡min⁡v∈CP⟨∣∑a=1PΔOava+iΔEloc∣2⟩.(45)\dot{\theta} = \underset{v\in\mathbb{C}^{P}}{\arg\min}\left\langle\left|\sum_{a=1}^{P}\Delta O_a v_a + i\Delta E_{\mathrm{loc}}\right|^2\right\rangle. \tag*{(45)}

Here vv is a trial complex parameter velocity. Minimization gives

∑b=1PSabθ˙b=−iFa,(46)\sum_{b=1}^{P} S_{ab}\dot{\theta}_b = -iF_a, \tag*{(46)}

where SS and FF have the covariance form of Eq. (18), evaluated with the complex-coordinate derivatives.

Regularization and integration. To regularize small eigenvalues of the sampled covariance matrix, LaQX either solves (S+λIP)θ˙=−iF(S+\lambda I_P)\dot{\theta}=-iF by Cholesky factorization, with damping λ>0\lambda>0 and P×PP\times P identity IPI_P, or diagonalizes S=Vdiag⁡(sν)V†S=V\operatorname{diag}(s_\nu)V^\dagger and uses

θ˙=−iVdiag⁡[f(sν)]V†F,f(sν)={1sν11+(c/∣sν/smax⁡∣)6,∣sν/smax⁡∣>ϵeig,0,otherwise.(47)\dot{\theta}=-iV\operatorname{diag}[f(s_\nu)]V^\dagger F, \qquad f(s_\nu)= \begin{cases} \frac{1}{s_\nu}\frac{1}{1+(c/\lvert s_\nu/s_{\max}\rvert)^6}, & \lvert s_\nu/s_{\max}\rvert>\epsilon_{\mathrm{eig}},\\ 0, & \text{otherwise}. \end{cases} \tag*{(47)}

Here sνs_\nu are the eigenvalues, ν=1,…,P\nu=1,\ldots,P, the columns of VV are the corresponding orthonormal eigenvectors, and †\dagger denotes the Hermitian conjugate. The largest eigenvalue is smax⁡s_{\max}, cc is the smooth relative cutoff, and ϵeig=10−14\epsilon_{\mathrm{eig}}=10^{-14} is the numerical threshold. Writing the regularized velocity as V(θ,t)\mathcal{V}(\theta,t) and θt=θ(t)\theta_t=\theta(t), the fourth-order Runge-Kutta (RK4) update [62] over a time step Δt\Delta t uses four stage velocities k1,…,k4k_1,\ldots,k_4:

k1=V(θt,t),k2=V(θt+Δt2k1,t+Δt2),k3=V(θt+Δt2k2,t+Δt2),k4=V(θt+Δtk3,t+Δt),θt+Δt=θt+Δt6(k1+2k2+2k3+k4).(48)\begin{aligned} k_1 &= \mathcal{V}(\theta_t,t),\\ k_2 &= \mathcal{V}\left(\theta_t+\frac{\Delta t}{2}k_1,t+\frac{\Delta t}{2}\right),\\ k_3 &= \mathcal{V}\left(\theta_t+\frac{\Delta t}{2}k_2,t+\frac{\Delta t}{2}\right),\\ k_4 &= \mathcal{V}(\theta_t+\Delta t k_3,t+\Delta t),\\ \theta_{t+\Delta t} &= \theta_t+\frac{\Delta t}{6}(k_1+2k_2+2k_3+k_4). \tag*{(48)} \end{aligned}

At each stage, walkers sample the intermediate wavefunction and SS and FF are reestimated. Observables are averaged over the stages with weights (1,2,2,1)/6(1,2,2,1)/6.

Function-oriented software design

Composition of the numerical workflow

In LaQX, a complete calculation is divided into four scientific components: the variational state, the Hamiltonian and observables, the sampling procedure, and the parameter-update rule. Each component is exposed through a small functional interface, so a complete workflow is assembled by composition rather than through a hierarchy of stateful solver objects. Numerical state is passed explicitly between components. This design permits compilation, vectorization, automatic differentiation, and device parallelism to be applied to the assembled numerical update [44].

The state layer implements the architectures described in § 2.2, together with the ACE-NES and ACE-PEFT specializations of § 2.4 and § 2.7. Independently, the operator layer represents Hamiltonians and observables through the configurations and matrix elements connected to each sampled occupation vector. Sampling and optimization then act on these common representations without depending on the internal construction of a particular ansatz or model.

This separation also determines the parallelization strategy. Batches of sampled configurations are distributed across accelerator devices, while the variational parameters and update state are shared. Ground-state optimization methods reuse the same outer VMC workflow, and measurement, projection, excited-state, and dynamical calculations reuse the same state and operator interfaces. The architecture therefore separates physical definitions from sampling, optimization, and distributed execution while preserving a common numerical core.

Locality of changes and LLM-assisted development

Small functional interfaces also support LLM-assisted programming. Reconstructing a large inheritance hierarchy or tracing implicit mutable state across many files consumes additional context and tokens, increasing the cost of a change. In LaQX, many extensions require only the relevant scientific component and its interface to be included in the context.

Table 1 summarizes this locality at the level of scientific concepts. A new Hamiltonian defines its connected configurations and matrix elements, after which the existing local-energy and sampling machinery can be reused. A new ansatz supplies initialization and amplitude evaluation without changing the Hamiltonian or optimization methods. Likewise, a new update rule can operate on the common stochastic estimators without depending on the construction of the state.

ExtensionRequired scientific definitionReused components
Variational stateInitialization and amplitudesHamiltonians, sampling, and optimization
Hamiltonian or observableConnected configurations and matrix elementsVariational states and update methods
Optimization methodUpdate from stochastic estimatorsState and operator definitions
Physical workflowComposition of established componentsStates, operators, and numerical kernels

Table 1. Locality of the principal scientific extension paths in LaQX.

The example of the altermagnetic Hubbard model in § 4.3 provides a concrete measure of this locality.

Physical applications

We now specify the example physical problems used to exercise the code. They cover fermionic and spin ground states, one-body observables, many-body topological responses, low-energy manifolds, and real-time dynamics. Reproducible input specifications and documentation accompany each example in the library.

The square-lattice Hubbard Hamiltonian [16] is defined as

HHub=−t∑⟨i,j⟩,σ(ciσ†cjσ+cjσ†ciσ)+U∑ini↑ni↓.(49)H_{\mathrm{Hub}}=-t\sum_{\langle i,j\rangle,\sigma}\left(c_{i\sigma}^{\dagger}c_{j\sigma}+c_{j\sigma}^{\dagger}c_{i\sigma}\right)+U\sum_i n_{i\uparrow}n_{i\downarrow}. \tag*{(49)}

Here ciσ†c_{i\sigma}^{\dagger} creates a fermion of spin σ\sigma at site ii, niσ=ciσ†ciσn_{i\sigma}=c_{i\sigma}^{\dagger}c_{i\sigma}, ni=ni↑+ni↓n_i=n_{i\uparrow}+n_{i\downarrow}, and ⟨i,j⟩\langle i,j\rangle denotes nearest-neighbor bonds. We set t=1t=1 and consider only nearest-neighbor hopping in this benchmark. Extensions with next-nearest-neighbor hopping t′t' are used in studies of competing stripe and superconducting orders [32, 33, 34].

The frustrated spin benchmark uses the square-lattice spin-1/2 J1J_1-J2J_2 Heisenberg Hamiltonian

HJ1J2=J1∑⟨i,j⟩Si⋅Sj+J2∑⟨⟨i,j⟩⟩Si⋅Sj.(50)H_{J_1J_2}=J_1\sum_{\langle i,j\rangle}\mathbf{S}_i\cdot\mathbf{S}_j+J_2\sum_{\langle\langle i,j\rangle\rangle}\mathbf{S}_i\cdot\mathbf{S}_j. \tag*{(50)}

Here Si\mathbf{S}_i is the spin operator at site ii, and ⟨⟨i,j⟩⟩\langle\langle i,j\rangle\rangle denotes next-nearest-neighbor bonds. The calculation uses a fixed-magnetization sector.

The interacting Hofstadter examples [59, 63] use

HHof=−∑x,y(cx+1,y†cx,y+e2πiαxcx,y+1†cx,y+h.c.)+V∑⟨i,j⟩ninj.(51)H_{\mathrm{Hof}}=-\sum_{x,y}\left(c_{x+1,y}^{\dagger}c_{x,y}+e^{2\pi i\alpha x}c_{x,y+1}^{\dagger}c_{x,y}+\mathrm{h.c.}\right)+V\sum_{\langle i,j\rangle}n_i n_j. \tag*{(51)}

Here cx,yc_{x,y} annihilates a spin-polarized fermion at site (x,y)(x,y), ni=ci†cin_i=c_i^{\dagger}c_i, and h.c. denotes the Hermitian conjugate. The magnetic flux per plaquette is α\alpha in units of the flux quantum, and VV is the nearest-neighbor density interaction; the hopping amplitude is set to unity. A boundary twist φ\varphi through the periodic xx direction replaces the xx-direction hopping by Eq. (43).

The altermagnetic Hubbard model adds sublattice- and direction-dependent diagonal hopping to Eq. (49) [64]:

Halt=HHub−∑⟨⟨i,j⟩⟩,σtij(2)(ciσ†cjσ+H.c.).(52)H_{\mathrm{alt}}=H_{\mathrm{Hub}}-\sum_{\langle\langle i,j\rangle\rangle,\sigma}t_{ij}^{(2)}\left(c_{i\sigma}^{\dagger}c_{j\sigma}+\mathrm{H.c.}\right). \tag*{(52)}

Let ri=(xi,yi)\mathbf{r}_i=(x_i,y_i) be the integer coordinates of site ii, with A and B denoting sites with odd and even xi+yix_i+y_i, respectively. The real hopping amplitude tij(2)t_{ij}^{(2)} depends on the sublattice of site ii and the bond displacement rj−ri\mathbf{r}_j-\mathbf{r}_i:

(1,1)(1,1)(1,−1)(1,-1)
At−t_-t+t_+
Bt+t_+t−t_-

Table 53.

Opposite bond orientations have the same amplitude, tji(2)=tij(2)t_{ji}^{(2)}=t_{ij}^{(2)}, and each undirected diagonal bond is counted once in Eq. (52).

The resulting set of calculations is summarized in Table 2.

Doped Hubbard ground states and GFMC

Large fermionic ansatzes are initialized through a sequence of variational states and Hamiltonians [1, 65]. The reference state is a determinant constructed directly from trainable local

ProtocolPhysics targetVariational stateMain output
HubbardGround-state VMC and GFMCACE, SCALEEnergy traces and projected energy
SpinFrustrated J1J_1-J2J_2 modelCNN-MPSEnergy and variance per site
AltermagnetSpin-resolved momentum structureACEDensity matrix and ⟨nkσ⟩\langle n_{k\sigma}\rangle
NESLow-lying Hofstadter statesACE-NESEnergy matrix and spectrum
Charge pumpingHofstadter topological responseACEMomentum sectors and polarization
tVMCReal-time density dynamicsACE-PEFTDensity time series

caption: Physical systems, variational states, and measured quantities in the documented LaQX calculations.

:::

orbital embeddings,

M(i,s),jref(n)=Ai,qi,s,j,ψref(n;A)=det⁡[MIa(n),jref(n)]a,j=1N.(53)M^{\mathrm{ref}}_{(i,s),j}(\mathbf{n}) = A_{i,q_i,s,j}, \\ \psi_{\mathrm{ref}}(\mathbf{n};A) = \det\left[M^{\mathrm{ref}}_{I_a(\mathbf{n}),j}(\mathbf{n})\right]_{a,j=1}^{N}. \tag*{(53)}

Here qi=ni↑+2ni↓q_i = n_{i\uparrow}+2n_{i\downarrow} labels the local occupation, ss is the spin channel, jj is the orbital index, and Ia(n)I_a(\mathbf{n}) labels the occupied spin orbitals as in Eq. (9). The entries Ai,qi,s,jA_{i,q_i,s,j} are optimized directly, with no convolutional or MLP layers.

Both Hubbard calculations use 1/8 hole doping. For the periodic 16×416 \times4 lattice with N=56N = 56 electrons, this reference state is first optimized at U=4U = 4 with Adam. It then initializes an ACE state, optimized for 30000 MARCH steps at U=8U = 8. For the periodic 16×816 \times8 lattice with N=112N = 112, the reference stage includes a stripe-pinning field of strength 0.05 and charge period eight, targeting a filled stripe [2]. The field is removed for the final 20000-step SCALE optimization at U=8U = 8. Both optimizations use 4096 walkers. The ACE state has 12 convolutional layers of width 128; SCALE uses width 256 and four site-wise nonlinear layers of width 512, following the architecture families of Ref. [2].

The local observables in Figure 2 show the hole density hi=1−⟨ni↑+ni↓⟩h_i = 1 - \langle n_{i\uparrow} + n_{i\downarrow}\rangle through site colors and the spin polarization ⟨Siz⟩=⟨ni↑−ni↓⟩/2\langle S_i^z\rangle= \langle n_{i\uparrow} - n_{i\downarrow}\rangle/2 through vertical arrows. Both states exhibit filled stripes [66] with a charge-density-wave wavelength of eight sites along the long direction and an antiphase modulation of the staggered magnetization mi=(−1)xi+yi⟨Siz⟩m_i = (-1)^{x_i+y_i}\langle S_i^z\rangle.

Figure 2. Local stripe profiles in the final Hubbard variational states at U=8U = 8: ACE on 16×416 \times4 (left) and SCALE on 16×816 \times8 (right). Circle colors encode the hole density hi=1−⟨ni⟩h_i = 1 - \langle n_i \rangle. Vertical arrows encode the unstaggered spin polarization ⟨Siz⟩\langle S_i^z \rangle, with direction indicating its sign and length proportional to its magnitude within each panel; very small moments are rendered as dots. Both panels use the same color scale, but arrow lengths are scaled separately so that the maximum absolute spin polarization in each panel has the same displayed arrow length. Each reference arrow is labeled with the corresponding maximum absolute spin polarization. No spatial smoothing or symmetry averaging is applied.

Evaluation of the final states gives energies per site of −0.762822(5)-0.762822(5) for ACE on 16×416 \times4 and −0.753059(7)-0.753059(7) for SCALE on 16×816 \times8. Both states are subsequently used as importance-sampling and fixed-node guiding functions for the GFMC algorithm of § 2.5, with the Hamiltonians and boundary conditions held fixed. Discarding the first 1000 of 5000 projection steps gives energies per site of −0.76378(2)-0.76378(2) for ACE-guided GFMC and −0.75628(3)-0.75628(3) for SCALE-guided GFMC. These estimates lower the respective VMC energies by approximately 9.6×10−49.6 \times10^{-4} and 3.2×10−33.2 \times10^{-3} per site.

Figure 1. VMC energy traces for the doped square-lattice Hubbard model. Left: ACE on a periodic 16×416 \times4 lattice with N=56N = 56 electrons. Right: SCALE on a periodic 16×816 \times8 lattice with N=112N = 112 electrons. Both final optimizations use U=8U = 8 and the staged initialization described in the text. Total energies are divided by 64 and 128, respectively, and the curves show trailing 300-step averages.

Frustrated J1J_1-J2J_2 spin model

For the spin benchmark we consider the periodic 6×66 \times6 J1J_1-J2J_2 Heisenberg model defined in Eq. (50) at J1=1J_1 = 1 and J2=0.5J_2 = 0.5, in the zero-magnetization sector. The variational state is CNN-MPS with MPS bond dimension D=20D = 20 [3]: the convolutional feature map first reorganizes the correlations of the sampled spin configuration, after which a periodic backflow MPS contracts the configuration-dependent tensors as in Eq. (16). The calculation applies the Marshall sign transformation [67] and projection onto the D4D_4 lattice-symmetry sector. The optimization uses 15000 MARCH steps without second-moment accumulator normalization, corresponding to v~k−1=vk−1\tilde{v}_{k-1} = v_{k-1} in Eq. (25). The step-length normalization in Eq. (27) remains active. In addition to the energy, we report the local-energy variance per site,

σE2=Var⁡(Eloc)Ns=⟨∣Eloc∣2⟩−∣⟨Eloc⟩∣2Ns,Ns=36.(54)\sigma_E^2 = \frac{\operatorname{Var}(E_{\mathrm{loc}})}{N_s} = \frac{\langle|E_{\mathrm{loc}}|^2\rangle- |\langle E_{\mathrm{loc}}\rangle|^2}{N_s}, \qquad N_s = 36. \tag*{(54)}

Independent evaluation of the final state gives an energy per site of e=−0.5038070(2)e = -0.5038070(2) and a local-energy variance per site of σE2=1.7586×10−5\sigma_E^2 = 1.7586 \times10^{-5} in units of J12J_1^2. The measured energy lies approximately 3×10−63 \times10^{-6} per site above the rounded ED reference in Figure 3.

Figure 3. CNN-MPS optimization of the periodic 6×66 \times6 J1J_1-J2J_2 model at J2/J1=0.5J_2/J_1 = 0.5, using MARCH without second-moment accumulator normalization. Left: late-stage convergence of the energy per site e=⟨HJ1J2⟩/36e = \langle H_{J_1J_2}\rangle/36, with the rounded ED reference eED=−0.503810e_{\mathrm{ED}} = -0.503810 [3, 68]. Right: local-energy variance per site σE2=Var⁡(Eloc)/36\sigma_E^2 = \operatorname{Var}(E_{\mathrm{loc}})/36. Training curves are trailing 200-step averages.

Altermagnetic Hubbard model

The altermagnetic calculation tests both a nonstandard kinetic term and an observable beyond the total energy. The diagonal hopping amplitudes in Eq. (53) are t+=0.4t_+ = 0.4 and t−=0t_- = 0. A half-filled periodic 12×1212 \times12 system is first optimized at U=3U = 3 using the reference determinant ansatz in Eq. (53). This occupation-dependent reference state initializes the subsequent ACE optimization at U=8U = 8. The trained state is then sampled to obtain the fermionic one-body density matrix.

Let ckσ=Ns−1/2∑je−ik⋅rjcjσc_{\mathbf{k}\sigma} = N_s^{-1/2} \sum_j e^{-i\mathbf{k}\cdot\mathbf{r}_j}c_{j\sigma} annihilate a fermion of momentum k\mathbf{k} and spin σ\sigma, where rj\mathbf{r}_j is the position of site jj. The measured momentum distribution is

⟨nkσ⟩=⟨ckσ†ckσ⟩=1Ns∑i,jeik⋅(ri−rj)⟨ciσ†cjσ⟩.(55)\langle n_{\mathbf{k}\sigma}\rangle= \langle c_{\mathbf{k}\sigma}^{\dagger}c_{\mathbf{k}\sigma}\rangle= \frac{1}{N_s}\sum_{i,j}e^{i\mathbf{k}\cdot(\mathbf{r}_i-\mathbf{r}_j)}\langle c_{i\sigma}^{\dagger}c_{j\sigma}\rangle. \tag*{(55)}

Figure 4 shows the difference ⟨nk↑⟩−⟨nk↓⟩\langle n_{\mathbf{k}\uparrow}\rangle- \langle n_{\mathbf{k}\downarrow}\rangle, highlighting the anisotropic spin-dependent momentum structure obtained from this estimator.

Figure 4. Spin-channel difference ⟨nk↑⟩−⟨nk↓⟩\langle n_{\mathbf{k}\uparrow}\rangle-\langle n_{\mathbf{k}\downarrow}\rangle for the half-filled periodic 12×1212 \times12 altermagnetic Hubbard model at U=8U=8. Both momentum distributions are obtained from the sampled one-body density matrix.

Hofstadter low-energy spectrum with NES

The NES calculation uses a 12×412 \times4 interacting Hofstadter system with N=4N = 4 spin-polarized particles, α=0.25\alpha= 0.25, and V=2V = 2. Four state-resolved determinant outputs share the nonlinear ACE feature extractor and are optimized jointly through the auxiliary determinant in Eq. (29), under either open or periodic boundary conditions. Sampling after optimization gives the 4×44 \times4 energy matrix; its Monte Carlo mean is diagonalized to resolve the energies represented in the learned subspace. Figure 5 compares these eigenvalues with ED.

Figure 5. Low-energy spectrum obtained with four-state ACE-NES for a 12×412 \times4 interacting Hofstadter model with four spin-polarized particles, α=0.25\alpha= 0.25, and V=2V = 2, compared with ED. Left: open boundaries. Right: periodic boundaries. The NQS energies are the eigenvalues of the Monte Carlo-averaged energy matrix; eigenvalue error bars are omitted because the stored data do not include the matrix covariances required for uncertainty propagation.

The open-boundary NES energies are approximately −10.5473-10.5473, −10.4485-10.4485, −10.4265-10.4265, and −10.4066-10.4066, with deviations from ED ranging from 4.6×10−54.6 \times10^{-5} to 1.0×10−41.0 \times10^{-4}. The finite cluster has a nondegenerate ground state separated from the first excitation by approximately 0.0990.099.

For periodic boundaries, the energies are approximately −11.1322-11.1322, −11.1295-11.1295, −11.1295-11.1295, and −11.0665-11.0665. The lowest three states form a closely spaced group, while the fourth lies about 0.0630.063 above the third. The first three deviations from ED are below 6.1×10−46.1 \times10^{-4}, whereas the fourth is about 4.5×10−34.5 \times10^{-3}, indicating a less accurate representation of that state. At filling ν=N/(αNs)=1/3\nu= N/(\alpha N_s) = 1/3, this three-state manifold is consistent with the torus ground-state structure of a fractional quantum Hall phase [59]. § 4.5 examines the associated fractional charge-pumping response.

Hofstadter charge pumping

Following the charge-pumping formulation of Ref. [59], Eq. (51) is defined on a periodic 12×812 \times8 lattice with N=8N = 8 spin-polarized particles, α=0.25\alpha= 0.25, and V=2V = 2. In this gauge, the magnetic unit cell spans four sites along xx. The integers Kx=0,1,2K_x = 0,1,2 and Ky=0,…,7K_y = 0,\ldots,7 label momentum sectors for translations by one magnetic unit cell along xx and one site along yy, respectively. For a translation by (4m,n)(4m,n) sites, the symmetry-projection phase is

χKx,Ky(m,n)=exp⁡[2πi(Kxm3+Kyn8)],m=0,1,2,n=0,…,7.(56)\chi_{K_x,K_y}(m,n) = \exp\left[2\pi i\left(\frac{K_x m}{3} + \frac{K_y n}{8}\right)\right], \qquad m = 0,1,2,\quad n = 0,\ldots,7. \tag*{(56)}

The sector scan identifies three low-energy states at (Kx,Ky)=(0,4)(K_x,K_y) = (0,4), (1,4)(1,4), and (2,4), as shown in Figure 6 (left). Their measured total energies are −21.85505(9)-21.85505(9), −21.84695(9)-21.84695(9), and −21.84713(9)-21.84713(9), respectively; the next level in the scanned sectors lies at −21.81715(8)-21.81715(8), about 0.0300.030 above the top of this manifold. The finite variational splitting within the manifold is approximately 0.00810.0081. Separate adiabatic-continuation calculations follow these three states using the protocol of § 2.6. The boundary twist φ\varphi increases from 00 to 2π2\pi, inserting one flux quantum.

Figure 6. Charge pumping in the interacting Hofstadter model on a 12×812 \times8 torus. Left: spectrum resolved by the integer momentum-sector labels (Kx,Ky)(K_x,K_y) defined in Eq. (56), with the three lowest states in (0,4)(0,4), (1,4)(1,4), and (2,4)(2,4). Energy bars show nominal standard errors without an autocorrelation correction. Right: unwrapped dimensionless polarization PyP_y and separate linear fits versus φ/(2π)\varphi/(2\pi) for each continued state; cc denotes the fitted slope.

At each flux value, the yy-direction Resta polarization is measured [69],

Z^y=exp⁡(2πiLy∑x,yyn^x,y),Py(φ)=12πIm⁡log⁡⟨Z^y⟩φ.(57)\hat{Z}_y = \exp\left(\frac{2\pi i}{L_y}\sum_{x,y} y\hat{n}_{x,y}\right), \qquad P_y(\varphi) = \frac{1}{2\pi}\operatorname{Im}\log\langle\hat{Z}_y\rangle_{\varphi}. \tag*{(57)}

The measured phase is unwrapped before division by 2π2\pi. Separate linear fits of PyP_y against φ/(2π)\varphi/(2\pi) give slopes 0.3340.334, 0.3350.335, and 0.3300.330 for the three continued states (Figure 6, right). Each branch therefore pumps approximately one third of a particle per inserted flux quantum, with a combined response of approximately one particle. This fractional response is consistent with flux insertion connecting states within a topological ground-state manifold [59].

LaQX was also used in a recent study of the triangular-lattice Hofstadter-Hubbard model [4]. There, LaQX-based NQS and independent PEPS calculations provided complementary evidence of spin-charge separation from spin pumping and real-time spin and charge dynamics. This application is distinct from the spin-polarized charge-pumping example documented here, but it demonstrates that the Hofstadter workflows extend to interacting spinful fermions and cross-method validation.

Real-time tVMC

The real-time calculation considers the density response to the removal of a local potential. An ACE reference state is first optimized for an open 8×68 \times6 Hofstadter lattice with N=4N = 4 spin-polarized particles, V=2V = 2, α=0.25\alpha= 0.25, and the pinning term hνn0h_\nu n_0 with hν=−4h_\nu= -4. At t=0t = 0, the pinning term is removed and the optimized state initializes the ACE-PFT evolution.

The ACE-PFT variant applies Eq. (44); the convolutional and site-wise nonlinear layers of the optimized reference state define a fixed, configuration-dependent neural basis, and only its complex final orbital map evolves. The calculation uses RK4 integration with time step 2×10−32 \times10^{-3}, 40960 walkers, and the eigendecomposition solver with smooth cutoff 10−810^{-8}. The density snapshots and time series in Figure 7 track the redistribution of the initially pinned charge and reproduce the main features of the ED response.

Figure 7. Density dynamics following a local quench in the 8×68 \times6, four-particle Hofstadter model. Left: local-density snapshots from ED and tVMC at the nearest recorded times to t=0,2.5,5,7.5,10t = 0, 2.5, 5, 7.5, 10. Right: ED and tVMC density time series at sites (0,0)(0, 0) and (0,1)(0, 1). No additional temporal smoothing is applied. The stored tVMC observables are weighted averages over the four RK4 stages and are plotted at the corresponding step midpoints. Only the final affine orbital map of ACE-PEFT is evolved.

Conclusion

We have introduced LaQX, a JAX implementation of neural quantum state methods for interacting lattice fermions and spins. The code combines Metropolis-Hastings sampling, several variational ansatz families, first-order and natural-gradient optimization, observables, stochastic projection, excited states, adiabatic continuation, and real-time evolution through a common set of functional interfaces. The doped Hubbard and frustrated J1J_1-J2J_2 Heisenberg model calculations connect the release to methods that have achieved high variational accuracy [1, 2, 3]. The Hofstadter-Hubbard applications encompass many-body charge pumping [59] and NQS calculations of spin-charge separation [4]. The altermagnetic, excited-state, and dynamical examples test the same interfaces on one-body observables, low-energy states, and real-time evolution.

The library separates scientific components through explicit functional interfaces. Wavefunctions, physical operators, and update rules can therefore be extended through local source changes. This locality is useful for conventional development and for LLM-assisted programming, where unnecessary cross-file dependencies consume a finite context window. Source code, documentation, and the scripts used for the calculations in this paper are available at https://github.com/bytedance/laqx.

Acknowledgements

We thank ByteDance Seed for inspiration and encouragement.

References

  1. [1]Y. Gu, W. Li, H. Lin, B. Zhan, R. Li, Y. Huang, D. He, Y. Wu, T. Xiang, M. Qin, L. Wang and D. Lv, Solving the Hubbard model with neural quantum states, Nature Communications 17(1), 7838 (2026), doi:10.1038/s41467-026-74028-6.DOI
  2. [2]Y. Gu, Z. Han, W. Li, Z. Xiao, T. Xiang, M. Qin, L. Wang and D. Lv, Pareto frontier of neural quantum states: Scalable, affordable, and accurate convolutional backflow for strongly correlated lattice fermions (2026), 2604.25775.arxiv.org/abs/2604.25775
  3. [3]C. Fan, B. Zhan, Y. Gu, T. Liu, Y. Wu, M. Qin, D. Lv and T. Xiang, Disentangling tensor network states with deep neural network (2026), 2603.14425.arxiv.org/abs/2603.14425
  4. [4]Y. Gu, H. Yang, Z. Dai and Y. Wu, Spin-charge separation in the triangular-lattice Hofstadter–Hubbard model, doi:10.48550/arXiv.2608.02727 (2026), 2608.02727.DOI
  5. [5]A. W. Sandvik, Computational studies of quantum spin systems, In Lectures on the Physics of Strongly Correlated Systems XIV, vol. 1297 of AIP Conference Proceedings, pp. 135–338. AIP, doi:10.1063/1.3518900 (2010).DOI
  6. [6]S. R. White, Density matrix formulation for quantum renormalization groups, Physical Review Letters 69(19), 2863 (1992), doi:10.1103/PhysRevLett.69.2863.DOI
  7. [7]U. Schollwoeck, The density-matrix renormalization group in the age of matrix product states, Annals of Physics 326(1), 96 (2011), doi:10.1016/j.aop.2010.09.012.DOI
  8. [8]J. I. Cirac, D. Perez-Garcia, N. Schuch and F. Verstraete, Matrix product states and projected entangled pair states: Concepts, symmetries, theorems, Reviews of Modern Physics 93(4), 045003 (2021), doi:10.1103/RevModPhys.93.045003.DOI
  9. [9]T. Xiang, Density Matrix and Tensor Network Renormalization, Cambridge University Press (2023).DOI
  10. [10]W. M. C. Foulkes, L. Mitas, R. J. Needs and G. Rajagopal, Quantum monte carlo simulations of solids, Reviews of Modern Physics 73(1), 33 (2001), doi:10.1103/RevModPhys.73.33.DOI
  11. [11]F. Assaad and H. Evertz, World-line and Determinantal Quantum Monte Carlo Methods for Spins, Phonons and Electrons, In H. Fehske, R. Schneider and A. Weiße, eds., Computational Many-Particle Physics, vol. 739, pp. 277–356. Springer Berlin Heidelberg, ISBN 978-3-540-74685-0, doi:10.1007/978-3-540-74686-7_10 (2008).DOI
  12. [12]F. Becca and S. Sorella, Quantum Monte Carlo Approaches for Correlated Systems, Cambridge University Press (2017).DOI
  13. [13]S. Zhang, Auxiliary-field quantum monte carlo for correlated electron systems, In E. Pavarini, E. Koch and U. Schollwöck, eds., Emergent Phenomena in Correlated Matter: Modeling and Simulation, vol. 3. Verlag des Forschungszentrums Jülich, Jülich (2013).
  14. [14]E. Y. Loh, J. E. Gubernatis, R. T. Scalettar, S. R. White, D. J. Scalapino and R. L. Sugar, Sign problem in the numerical simulation of many-electron systems, Physical Review B 41(13), 9301 (1990), doi:10.1103/PhysRevB.41.9301.DOI
  15. [15]M. Troyer and U.-J. Wiese, Computational complexity and fundamental limitations to fermionic quantum Monte Carlo simulations, Physical Review Letters 94(17), 170201 (2005), doi:10.1103/PhysRevLett.94.170201.DOI
  16. [16]J. Hubbard, Electron correlations in narrow energy bands, Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences 276(1365), 238 (1963), doi:10.1098/rspa.1963.0204.DOI
  17. [17]J. P. F. LeBlanc, A. E. Antipov, F. Becca, I. W. Bulik, G. K.-L. Chan, C.-M. Chung, Y. Deng, M. Ferrero, T. M. Henderson, C. A. Jimenez-Hoyos, E. Kozik, X.-W. Liu et al., Solutions of the two-dimensional Hubbard model: Benchmarks and results from a wide range of numerical algorithms, Physical Review X 5(4), 041041 (2015), doi:10.1103/PhysRevX.5.041041.DOI
  18. [18]M. Qin, T. Schäfer, S. Andergassen, P. Corboz and E. Gull, The Hubbard Model: A Computational Perspective, Annual Review of Condensed Matter Physics 13(1), 275 (2022), doi:10.1146/annurev-conmatphys-090921-033948.DOI
  19. [19]D. P. Arovas, E. Berg, S. A. Kivelson and S. Raghu, The Hubbard Model, Annual Review of Condensed Matter Physics 13, 239 (2022), doi:10.1146/annurev-conmatphys-031620-102024.DOI
  20. [20]G. Carleo and M. Troyer, Solving the quantum many-body problem with artificial neural networks, Science 355(6325), 602 (2017), doi:10.1126/science.aag2302.DOI
  21. [21]H. Lange, A. Van de Walle, A. Abedinnia and A. Bohrdt, From architectures to applications: A review of neural quantum states, Quantum Science and Technology 9(4), 040501 (2024), doi:10.1088/2058-9565/ad7168.DOI
  22. [22]M. Medvidovic and J. Robledo Moreno, Neural-network quantum states for many-body physics, The European Physical Journal Plus 139(7), 631 (2024), doi:10.1140/epjp/s13360-024-05311-y.DOI
  23. [23]Y. Nomura, A. S. Darmawan, Y. Yamaji and M. Imada, Restricted-boltzmann-machine learning for solving strongly correlated quantum systems, Physical Review B 96(20), 205152 (2017), doi:10.1103/PhysRevB.96.205152.DOI
  24. [24]K. Choo, T. Neupert and G. Carleo, Two-dimensional frustrated j₁–j₂ model studied with neural network quantum states, Physical Review B 100(12), 125124 (2019), doi:10.1103/PhysRevB.100.125124.DOI
  25. [25]M. Hibat-Allah, M. Ganahl, L. E. Hayward, R. G. Melko and J. Carrasquilla, Recurrent neural network wave functions, Physical Review Research 2(2), 023358 (2020), doi:10.1103/PhysRevResearch.2.023358.
  26. [26]C. Roth, A. Szabo and A. H. MacDonald, High-accuracy variational Monte Carlo for frustrated magnets with deep neural networks, Physical Review B 108(5), 054410 (2023), doi:10.1103/PhysRevB.108.054410.DOI
  27. [27]D. Luo and B. K. Clark, Backflow transformations via neural networks for quantum many-body wave functions, Physical Review Letters 122(22), 226401 (2019), doi:10.1103/PhysRevLett.122.226401.DOI
  28. [28]J. Robledo Moreno, G. Carleo, A. Georges and J. Stokes, Fermionic wave functions from neural-network constrained hidden states, Proceedings of the National Academy of Sciences 119(32), e2122059119 (2022), doi:10.1073/pnas.2122059119.DOI
  29. [29]Z. Liu and B. K. Clark, Unifying view of fermionic neural network quantum states: From neural network backflow to hidden fermion determinant states, Physical Review B 110(11), 115124 (2024), doi:10.1103/PhysRevB.110.115124.DOI
  30. [30]D. Pfau, S. Axelrod, H. Sutterud, I. von Glehn and J. S. Spencer, Accurate computation of quantum excited states with neural networks, Science 385(6711), eadn0137 (2024), doi:10.1126/science.adn0137.DOI
  31. [31]M. Schmitt and M. Heyl, Quantum many-body dynamics in two dimensions with artificial neural networks, Physical Review Letters 125(10), 100503 (2020), doi:10.1103/PhysRevLett.125.100503.DOI
  32. [32]C. Roth, A. Chen, A. Sengupta and A. Georges, Superconductivity in the two-dimensional Hubbard model revealed by neural quantum states, doi:10.48550/arXiv.2511.07566 (2025), 2511.07566.DOI
  33. [33]L. L. Viteritti, R. Rende, C. Roth, A. Sengupta, G. Carleo and A. Georges, Beyond variational bias: Resolving intertwined orders in the Hubbard model, doi:10.48550/arXiv.2604.21978 (2026), 2604.21978.DOI
  34. [34]R. Rende, L. L. Viteritti and A. Georges, Superconductivity in the t−t′ Hubbard model from symmetry-preserving neural-network quantum states, doi:10.48550/arXiv.2608.12465 (2026), 2608.12465.DOI
  35. [35]C. Roth, A. Millis and T. Soejima, Large-scale neural quantum states reveal the interplay between superconductivity and quantum criticality in the Hofstadter–Hubbard model, doi:10.48550/arXiv.2608.02753 (2026), 2608.02753.DOI
  36. [36]F. Vicentini, D. Hofmann, A. Szabó, D. Wu, C. Roth, C. Giuliani, G. Pescia, J. Nys, V. Vargas-Calderón, N. Astrakhantsev and G. Carleo, NetKet 3: Machine learning toolbox for many-body quantum systems, SciPost Physics Codebases p. 7 (2022), doi:10.21468/SciPostPhysCodeb.7.DOI
  37. [37]M. Schmitt and M. Reh, jVMC: Versatile and performant variational Monte Carlo leveraging automated differentiation and GPU acceleration, SciPost Physics Codebases p. 2 (2022), doi:10.21468/SciPostPhysCodeb.2.DOI
  38. [38]A. Chen and C. Roth, Quantax: Flexible neural quantum states in JAX (2024).
  39. [39]D. Pfau, J. S. Spencer, A. G. D. G. Matthews and W. M. C. Foulkes, Ab initio solution of the many-electron Schrödinger equation with deep neural networks, Physical Review Research 2(3), 033429 (2020), doi:10.1103/PhysRevResearch.2.033429.DOI
  40. [40]J. S. Spencer, D. Pfau, A. Botev and W. M. C. Foulkes, Better, faster fermionic neural networks (2020), 2011.07125.arxiv.org/abs/2011.07125
  41. [41]W. Ren, W. Fu, X. Wu and J. Chen, Towards the ground state of molecules via diffusion Monte Carlo on neural networks, Nature Communications 14(1), 1860 (2023), doi:10.1038/s41467-023-37609-3.DOI
  42. [42]ByteDance Seed, JaQMC: JAX-accelerated quantum Monte Carlo (2026).
  43. [43]S.-J. Du, J. Gray, A. Chen and G. K.-L. Chan, Vectorized symmetric and fermionic tensor network implementations for GPU-accelerated variational Monte Carlo, doi:10.48550/arXiv.2608.27861 (2026), 2608.27861.DOI
  44. [44]J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman-Milne and Q. Zhang, JAX: Composable transformations of Python+NumPy programs (2018).
  45. [45]N. Trivedi and D. M. Ceperley, Ground-state correlations of quantum antiferromagnets: A green-function monte carlo study, Physical Review B 41(7), 4552 (1990), doi:10.1103/PhysRevB.41.4552.DOI
  46. [46]S. Sorella, Green function Monte Carlo with stochastic reconfiguration, Physical Review Letters 80(20), 4558 (1998), doi:10.1103/PhysRevLett.80.4558.DOI
  47. [47]A. Sarkar and I. Drosos, Vibe coding: Programming through conversation with artificial intelligence, In Proceedings of the 36th Annual Conference of the Psychology of Programming Interest Group, doi:10.48550/arXiv.2506.23253 (2025), 2506.23253.DOI
  48. [48]D. E. Rumelhart, G. E. Hinton and R. J. Williams, Learning representations by back-propagating errors, Nature 323, 533 (1986), doi:10.1038/323533a0.DOI
  49. [49]S. Elfwing, E. Uchibe and K. Doya, Sigmoid-weighted linear units for neural network function approximation in reinforcement learning, Neural Networks 107, 3 (2018), doi:10.1016/j.neunet.2017.12.012.DOI
  50. [50]J. L. Ba, J. R. Kiros and G. E. Hinton, Layer normalization (2016), 1607.06450.arxiv.org/abs/1607.06450
  51. [51]N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller and E. Teller, Equation of state calculations by fast computing machines, The Journal of Chemical Physics 21(6), 1087 (1953), doi:10.1063/1.1699114.DOI
  52. [52]W. K. Hastings, Monte carlo sampling methods using Markov chains and their applications, Biometrika 57(1), 97 (1970), doi:10.1093/biomet/57.1.97.DOI
  53. [53]D. P. Kingma and J. Ba, Adam: A method for stochastic optimization, In 3rd International Conference on Learning Representations (2015), 1412.6980.arxiv.org/abs/1412.6980
  54. [54]G. Goldshlager, N. Abrahamsen and L. Lin, A Kaczmarz-inspired approach to accelerate the optimization of neural network wavefunctions, Journal of Computational Physics 516, 113351 (2024), doi:10.1016/j.jcp.2024.113351, 2401.10190.DOI
  55. [55]S. Sorella, Generalized lanczos algorithm for variational quantum monte carlo, Physical Review B 64(2), 024512 (2001), doi:10.1103/PhysRevB.64.024512.DOI
  56. [56]A. Chen and M. Heyl, Empowering deep neural quantum states through efficient optimization, Nature Physics 20(9), 1476 (2024), doi:10.1038/s41567-024-02566-1.DOI
  57. [57]R. Rende, L. L. Viteritti, L. Bardone, F. Becca and S. Goldt, A simple linear algebra identity to optimize large-scale neural network quantum states, Communications Physics 7, 260 (2024), doi:10.1038/s42005-024-01732-4.DOI
  58. [58]D. F. B. ten Haaf, H. J. M. van Bemmel, J. M. J. van Leeuwen, W. van Saarloos and D. M. Ceperley, Proof for an upper bound in fixed-node monte carlo for lattice fermions, Physical Review B 51(19), 13039 (1995), doi:10.1103/PhysRevB.51.13039.DOI
  59. [59]H. Chen, Y. Qian, W. Ren, X. Li and J. Chen, Topological invariant of periodic many body wavefunction from charge pumping simulation (2026), 2604.08702.arxiv.org/abs/2604.08702
  60. [60]W. Fu, Z. Li, Y. Qian, R. Li, W. Ren and J. Chen, Towards stable and accurate electron dynamics via neural network based time-dependent variational monte carlo (2026), 2606.05850.arxiv.org/abs/2606.05850
  61. [61]G. Carleo, F. Becca, M. Schiró and M. Fabrizio, Localization and glassy dynamics of many-body quantum systems, Scientific Reports 2, 243 (2012), doi:10.1038/srep00243.DOI
  62. [62]J. C. Butcher, A history of Runge–Kutta methods, Applied Numerical Mathematics 20(3), 247 (1996), doi:10.1016/0168-9274(95)000108-5.DOI
  63. [63]D. R. Hofstadter, Energy levels and wave functions of Bloch electrons in rational and irrational magnetic fields, Physical Review B 14(6), 2239 (1976), doi:10.1103/PhysRevB.14.2239.DOI
  64. [64]P. Das, V. Leeb, J. Knolle and M. Knap, Realizing altermagnetism in fermi-hubbard models with ultracold atoms, Physical Review Letters 132(26), 263402 (2024), doi:10.1103/PhysRevLett.132.263402.DOI
  65. [65]A. Chen, Z.-Q. Wan, A. Sengupta, A. Georges and C. Roth, Neural network-augmented pfaffian wave-functions for scalable simulations of interacting fermions (2025), 2507.10705.DOI
  66. [66]B.-X. Zheng, C.-M. Chung, P. Corboz, G. Ehlers, M.-P. Qin, R. M. Noack, H. Shi, S. R. White, S. Zhang and G. K.-L. Chan, Stripe order in the underdoped region of the two-dimensional Hubbard model, Science 358(6367), 1155 (2017), doi:10.1126/science.aam7127.DOI
  67. [67]W. Marshall, Antiferromagnetism, Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences 232(1188), 48 (1955), doi:10.1098/rspa.1955.0200.DOI
  68. [68]H. J. Schulz, T. A. L. Ziman and D. Poilblanc, Magnetic order and disorder in the frustrated quantum Heisenberg antiferromagnet in two dimensions, Journal de Physique I 6(5), 675 (1996), doi:10.1051/jp1:1996236.DOI
  69. [69]R. Resta, Quantum-mechanical position operator in extended systems, Physical Review Letters 80(9), 1800 (1998), doi:10.1103/PhysRevLett.80.1800.DOI

Paper details

vlnc:W_qdz9vdwp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents