Neural Operator Quantum State (NOQS)

• NOQS is a quantum-state modeling framework that learns a reusable map from time-dependent Hamiltonian protocols to many-body wavefunctions.
• It integrates transformer-based autoregressive neural quantum states with Fourier Neural Operators and cross-attention to capture both discrete and continuous dynamics.
• Empirical benchmarks on 2D transverse-field Ising models show high accuracy in observable prediction and zero-shot temporal super-resolution across driving protocols.

Searching arXiv for NOQS and closely related work to ground the article in the latest literature. Neural Operator Quantum State (NOQS) denotes a quantum-state modeling framework in which the learned object is the solution operator of driven many-body quantum dynamics, rather than a single trajectory-specific wavefunction. In its most explicit formulation, NOQS learns a map from an entire time-dependent Hamiltonian protocol $H(t)$ to a time-evolved autoregressive many-body wavefunction, so that previously unseen drivings can be evaluated in a single forward pass without rerunning protocol-specific optimization (Qi et al., 26 Mar 2026). In the literature summarized here, this operator-learning meaning distinguishes NOQS from earlier neural quantum-state constructions in which “operator” referred instead to gate action on neural states, operator-generated RBM circuits, density-operator parameterizations for mixed states, or operator-space Krylov constructions layered on top of a neural ground state.

1. Definition and problem class

NOQS is formulated for the time-dependent Schrödinger equation

$i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$

with formal solution

$$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$

The central claim is that standard solvers operate pointwise in protocol space: exact diagonalization, direct Schrödinger solvers, tensor-network methods such as MPS/tDMRG/PEPS, and time-dependent neural quantum states typically solve one chosen trajectory $H(t)\mapsto |\psi(t)\rangle$ at a time and must be rerun if the drive changes. NOQS instead aims to learn a reusable map over a function space of drivings, so the task is framed as operator learning rather than ordinary regression over a finite-dimensional control vector (Qi et al., 26 Mar 2026).

The learned map is written as

$$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$

where the input is a time-dependent driving protocol, the output is the time-dependent many-body wavefunction amplitude $\langle \boldsymbol{\sigma}|\psi(t)\rangle$, and $\mathcal{N}_\eta[H(t)]\equiv M(t)$ is a neural-operator-generated latent context. The formulation is explicitly tied to a fixed system size, geometry, and initial state, with the protocol varying within a prescribed family (Qi et al., 26 Mar 2026).

This definition makes NOQS a model of amortized quantum dynamics. Once pretrained, inference for an unseen protocol is intended to require only a forward pass. This contrasts with trajectory-by-trajectory variational evolution, in which the network parameters themselves are typically reoptimized for each new $H(t)$. The literature further emphasizes that the novelty lies not merely in using a neural wavefunction, but in learning a map from the entire driving function to the full time-dependent quantum state (Qi et al., 26 Mar 2026).

2. State representation and neural-operator architecture

The NOQS state is represented in the computational basis as

$$\psi(\boldsymbol{\sigma})=\langle \boldsymbol{\sigma}|\psi\rangle$$

for spin configurations $\boldsymbol{\sigma}=(\sigma_1,\dots,\sigma_N)$, $i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$0. Its amplitude-phase decomposition is

$i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$1

with Born probabilities factorized autoregressively as

$i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$2

while “the phase is factorized in an analogous way” (Qi et al., 26 Mar 2026). This places NOQS within the autoregressive NQS tradition on the state side, while introducing operator learning on the conditioning side.

Architecturally, NOQS combines three components: a transformer-based autoregressive neural quantum state, a Fourier Neural Operator (FNO) that acts on the driving protocol, and cross-attention coupling between the protocol context and the wavefunction tokens (Qi et al., 26 Mar 2026). The division of labor is explicit. The transformer handles the discrete Hilbert-space structure of spin configurations, whereas the FNO handles the continuous temporal structure of the drive.

The protocol is represented at discretized times $i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$3 by its vector of time-dependent Hamiltonian coefficients $i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$4. A lifting layer maps these inputs to width $i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$5, after which a stack of Fourier layers performs FFTs along time, truncates to $i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$6 modes, mixes channels, and returns to time domain. A projection then produces $i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$7 context tokens $i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$8, which are corrected to enforce a common initial condition through

$i\partial_t |\psi(t)\rangle = H(t)|\psi(t)\rangle,$9

This offset is not a cosmetic detail: it ensures that protocol dependence enters through $$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$0 while the initial latent context remains fixed across all drivings (Qi et al., 26 Mar 2026).

On the state branch, each spin is embedded into a latent vector, positional encodings are added, and masked multi-head self-attention enforces causality for autoregressive factorization. Cross-attention then conditions each spin token on the protocol-derived context $$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$1. The resulting hidden states are unembedded to produce $$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$2 and $$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$3, and hence the full complex wavefunction $$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$4 (Qi et al., 26 Mar 2026).

This architecture implies that NOQS is neither a recurrent rollout over time nor a trajectory database interpolator. Time dependence enters through a continuously generated latent representation of the protocol. A plausible implication is that the model’s “operator” character resides primarily in the FNO branch $$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$5, while the transformer remains a conditional autoregressive state decoder.

3. Training objective, sampling, and inference

NOQS is trained self-supervised, without supervised labels from exact trajectories, tensor networks, or experiments during pretraining. The central loss is a TDVP-inspired Schrödinger residual,

$$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$6

which at fixed time becomes

$$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$7

with local Schrödinger residual

$$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$8

The local energy estimator is

$$|\psi(t)\rangle = \mathcal{T}\exp\!\left(-i\int_0^t dt'\,H(t')\right)|\psi(0)\rangle.$$9

Because global phase is gauge freedom, the implementation minimizes the variance of $H(t)\mapsto |\psi(t)\rangle$0 over sampled configurations rather than the raw expectation directly (Qi et al., 26 Mar 2026).

The initial state is enforced by two mechanisms: the context-token offset above, and an anchor term

$H(t)\mapsto |\psi(t)\rangle$1

The full loss is

$H(t)\mapsto |\psi(t)\rangle$2

Monte Carlo training averages over three axes simultaneously: a batch of Hamiltonian trajectories, time points sampled along each trajectory, and spin configurations sampled from the Born distribution of the autoregressive wavefunction (Qi et al., 26 Mar 2026).

An important technical feature is the treatment of time derivatives. Since the protocol is encoded by an FNO, differentiation can be performed spectrally: $H(t)\mapsto |\psi(t)\rangle$3 Combined with automatic differentiation through the transformer, this yields $H(t)\mapsto |\psi(t)\rangle$4 without finite differences. The paper presents this as central to temporal-resolution transfer (Qi et al., 26 Mar 2026).

Inference preserves autoregressive sampling. The model outputs $H(t)\mapsto |\psi(t)\rangle$5 and $H(t)\mapsto |\psi(t)\rangle$6, so configurations can be sampled exactly and independently from $H(t)\mapsto |\psi(t)\rangle$7, after which observables are estimated through local estimators. This means that although prediction of the state representation is a forward pass, observable evaluation still generally requires Monte Carlo sampling (Qi et al., 26 Mar 2026).

4. Benchmark system and reported capabilities

The main benchmark is the square-lattice two-dimensional transverse-field Ising model with longitudinal and transverse driving,

$H(t)\mapsto |\psi(t)\rangle$8

with $H(t)\mapsto |\psi(t)\rangle$9, open boundary conditions, and initial state

$$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$0

The main text studies $$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$1 and $$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$2 systems; the former is benchmarked against exact diagonalization and the latter against tDMRG with converged bond dimension $$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$3 and SVD truncation threshold $$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$4 (Qi et al., 26 Mar 2026).

The training and in-distribution protocol family is built from truncated Fourier drives,

$$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$5

with an analogous form for $$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$6, using $$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$7, $$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$8, $$\mathcal{F}_{(\theta,\eta)} : H(t) \longmapsto \psi_\theta(\boldsymbol{\sigma};\mathcal{N}_\eta[H(t)]),$$9, $\langle \boldsymbol{\sigma}|\psi(t)\rangle$0, Gaussian-distributed amplitudes, and uniformly random phases. Out-of-distribution evaluation uses Gaussian pulses and tanh ramps, which the model was never shown during training (Qi et al., 26 Mar 2026).

The reported empirical picture is strongly observable-based. On $\langle \boldsymbol{\sigma}|\psi(t)\rangle$1, NOQS predicts the energy $\langle \boldsymbol{\sigma}|\psi(t)\rangle$2 for unseen in-distribution protocols in a way that “matches almost perfectly” with exact results, and it also predicts $\langle \boldsymbol{\sigma}|\psi(t)\rangle$3 and $\langle \boldsymbol{\sigma}|\psi(t)\rangle$4 accurately for out-of-distribution Gaussian pulses and tanh ramps. On $\langle \boldsymbol{\sigma}|\psi(t)\rangle$5, the same types of agreement are reported relative to tDMRG for energy, transverse magnetization, and nearest-neighbor correlators (Qi et al., 26 Mar 2026).

The paper also reports temporal-resolution transfer. A model trained on $\langle \boldsymbol{\sigma}|\psi(t)\rangle$6 time points is evaluated without retraining on $\langle \boldsymbol{\sigma}|\psi(t)\rangle$7 points, with smooth small absolute errors and no visible grid artifacts in the reported observable curves. This is presented as zero-shot temporal super-resolution and as evidence that the FNO confers discretization-transfer properties across time grids (Qi et al., 26 Mar 2026).

A further capability is fine-tuning with sparse data. For a fixed protocol, the model is adjusted using only four measurement times for $\langle \boldsymbol{\sigma}|\psi(t)\rangle$8 and $\langle \boldsymbol{\sigma}|\psi(t)\rangle$9, through the purely supervised loss

$\mathcal{N}_\eta[H(t)]\equiv M(t)$0

The reported result is that this sparse correction improves not only the measured observables but also the predicted energy over the full time interval, suggesting improvement of the underlying wavefunction rather than merely local regression on the measured channels (Qi et al., 26 Mar 2026).

The evaluation is, however, deliberately not presented as a full-state fidelity benchmark at larger sizes. The paper emphasizes observable agreement, temporal transfer, and protocol generalization, while explicitly not providing aggregate wavefunction-fidelity tables for the larger systems.

5. Relation to earlier neural-state and operator-based methods

The NOQS designation sits within a broader and partly heterogeneous literature on neural quantum states. Several earlier directions are adjacent, but technically distinct.

Approach	Main learned or constructed object	Relation to NOQS
Neural network operations on RBM/UBM states (1803.02118)	Gate action as deterministic network rewriting	Operator action on neural states, not function-space operator learning
Unitary-coupled RBM-NQS circuits (Hsieh et al., 2019)	Operator-generated RBM state preparation in circuits	Circuit/operator realization of NQS, not neural operator over protocols
Neural Density Operator vs POVM-NQS (Zhao et al., 2023)	Mixed-state operator or measurement-distribution representation	Operator-level state encoding, not learned solution operator
Hybrid NQS-PQC ansatz (Zhang et al., 21 Jan 2025)	Product of classical NQS and PQC correction	Fixed-size variational wavefunction, not operator learning across drivings
NQS + operator-Lanczos impurity solver (Rigo et al., 9 Dec 2025)	Operator-Krylov response space built on a neural ground state	NOQS-style operator sector, but symbolic rather than neural-operator
Non-stochastic selected-configuration NQS (Li et al., 2023)	Deterministic optimization/evaluation backend	Training strategy for discrete-state NQS, not a NOQS architecture

The 2018 work on “Neural Network Operations and Susuki-Trotter evolution of Neural Network States” develops explicit rewriting rules for how one- and $\mathcal{N}_\eta[H(t)]\equiv M(t)$1-body quantum operations act on RBM and UBM neural states, showing that exact gate application generally leaves the RBM manifold and enters an unrestricted Boltzmann manifold (1803.02118). This is highly NOQS-like in spirit because it supplies a neural-state manifold plus operator-action rules, but the “operator” there is a finite-dimensional quantum operation acting by graph rewrite, not a learned map over a family of time-dependent protocols.

The 2019 “unitary-coupled RBM-NQS” work similarly interprets an RBM state as a structured sequence of commuting operators acting on a visible register, with ancilla reuse, complex-valued wavefunctions, and avoidance of post-selection under the restriction $\mathcal{N}_\eta[H(t)]\equiv M(t)$2 (Hsieh et al., 2019). Again, the emphasis is operator-generated state preparation rather than learning a solution operator over function space.

For mixed states, “Empirical Sample Complexity of Neural Network Mixed State Reconstruction” compares a direct density-operator parameterization, the Neural Density Operator, with an indirect IC-POVM distribution model, showing that direct physical operator representations can have different shot-complexity behavior from measurement-space representations as mixedness varies (Zhao et al., 2023). This is relevant to NOQS because it isolates operator-level inductive bias, but it addresses tomography and mixed-state reconstruction, not driven dynamics over protocol families.

Two additional neighboring lines further clarify the boundary of the term. “Quantum-enhanced neural networks for quantum many-body simulations” introduces a product ansatz

$\mathcal{N}_\eta[H(t)]\equiv M(t)$3

combining a PQC correction with an autoregressive neural wavefunction, but remains a fixed-size variational ansatz rather than an operator learner across drivings (Zhang et al., 21 Jan 2025). “Operator Lanczos Approach enabling Neural Quantum States as Real-Frequency Impurity Solvers” constructs an operator-Krylov space $\mathcal{N}_\eta[H(t)]\equiv M(t)$4 on top of an autoregressive NQS ground state to compute real-frequency Green’s functions and self-energies, making it strongly NOQS-adjacent in methodology, but the operator sector there is symbolic and commutator-generated rather than neural (Rigo et al., 9 Dec 2025). Finally, deterministic selected-configuration optimization provides an architecture-agnostic route to noise-free NQS training in discrete determinant bases and is relevant mainly as a possible backend rather than as a definition of NOQS itself (Li et al., 2023).

A common misconception is therefore to treat any “operator-flavored” NQS as a NOQS in the modern neural-operator sense. The literature here suggests a sharper distinction: current NOQS refers most specifically to learning a reusable map from protocol functions to quantum states, whereas earlier operator-based NQS work typically concerns how operators act within a chosen neural-state manifold or how operator spaces are built on top of a neural state.

6. Limitations, scope conditions, and open directions

The most explicit limitations come from the 2026 NOQS formulation itself. The model is trained for a fixed Hamiltonian family, fixed coupling $\mathcal{N}_\eta[H(t)]\equiv M(t)$5, fixed lattice size/geometry, and fixed initial state. Generalization is therefore over driving protocols, not over arbitrary changes in system class. The paper states that extension to jointly conditioning on time-dependent protocols, static couplings, disorder, and dissipation rates is future work (Qi et al., 26 Mar 2026).

The empirical evidence is also bounded. The largest demonstrated system is $\mathcal{N}_\eta[H(t)]\equiv M(t)$6, and the evaluation is primarily through observables rather than full-state fidelity. The paper does not deeply probe very long times, and it explicitly notes that robustness beyond the tested protocol families is not guaranteed even though the out-of-distribution Gaussian-pulse and ramp results are favorable (Qi et al., 26 Mar 2026).

From a computational standpoint, amortized inference does not remove all costs. Pretraining uses 60,000 Adam steps with Monte Carlo averaging over protocols, times, and spin configurations in the $\mathcal{N}_\eta[H(t)]\equiv M(t)$7 setup, and observable estimation after inference still requires sampling from the autoregressive state. Fine-tuning with sparse measurements is reported to be cheap, but experimental noise, finite-shot effects, and imperfect observable channels are not analyzed in detail (Qi et al., 26 Mar 2026).

Related literature sharpens the outstanding design questions. Mixed-state tomography results indicate that direct operator-level parameterizations can gain statistical advantages but incur classical overhead in likelihood evaluation and optimization difficulty (Zhao et al., 2023). Operator-Lanczos impurity-solver results suggest that extending neural states with an explicitly constructed operator-response manifold can yield real-frequency observables beyond ground-state estimation (Rigo et al., 9 Dec 2025). Deterministic selected-support methods suggest that some discrete-basis NOQS variants may not need MCMC as the default training or evaluation strategy when the relevant support is sparse (Li et al., 2023).

Taken together, these results suggest that NOQS is best understood as an emerging synthesis of three ideas: autoregressive neural quantum states for exact normalized sampling, neural operators for function-space generalization across drives, and physics-based self-supervision through Schrödinger-residual training. Its distinctive claim is not merely improved representation of a single many-body wavefunction, but the learning of a transferable map

$\mathcal{N}_\eta[H(t)]\equiv M(t)$8

over a family of dynamical tasks (Qi et al., 26 Mar 2026).