Learning Hamiltonian Densities

• Operator learning of Hamiltonian densities is a framework that infers the operator structure and space-dependent coupling constants from experimental or simulation data.
• It applies methodologies such as maximum likelihood, tensor parametric models, and operator neural networks to reconstruct local density terms with quantitative accuracy.
• Practical implementations on quantum hardware, cold atom simulators, and molecular systems demonstrate robust error bounds and scalability.

Operator learning of Hamiltonian densities refers to the data-driven inference of the operator structure and spatially resolved parameters governing quantum and classical dynamical systems. This field synthesizes advances in quantum tomography, machine learning, variational methods, and reduced-order modeling to reconstruct the functional form and coefficients of Hamiltonian densities or generators directly from measurement data, time-series, or quantum states. Applications encompass quantum many-body systems, molecular electron dynamics, field theories, and classical Hamiltonian PDEs, with the primary objective of extracting not only the operator basis (i.e., the density terms) but also the scale- and space-dependent coupling constants with quantitative accuracy.

1. Mathematical Foundations of Hamiltonian Density Learning

Given a system described by a Hamiltonian of the form

$H[\phi,\pi] = \int d^d x\, \mathcal{H}(x)$

where $\mathcal{H}(x)$ is the local Hamiltonian density, operator learning pursues the identification of $\mathcal{H}(x)$ from observed data. In quantum systems, this frequently involves expansions in a basis of local operators,

$$\mathcal{H}(x) = \sum_i g_i\, O_i(\phi(x), \pi(x), \nabla\phi(x), ...)$$

where $O_i$ are field or operator monomials and $g_i$ are the unknown parameters. In discrete quantum systems such as spin chains or lattices, the Hamiltonian is written as

$H = \sum_\alpha h_\alpha \Pi_\alpha$

with each $\Pi_\alpha$ a $k$-local density operator. For molecular systems, the density matrix evolution is governed by the quantum Liouville–von Neumann equation,

$$i\,\frac{d\mathbf{P}'(t)}{dt} = [\mathbf{H}'(t),\,\mathbf{P}'(t)]$$

and operator learning seeks a map $\mathcal{H}(x)$0 that closes the dynamics (Bhat et al., 2020).

Gibbs state or equilibrium data introduces the variational free-energy principle: $\mathcal{H}(x)$1 with $\mathcal{H}(x)$2 parameterized as a sum over local densities. The gradient of this objective with respect to the parameters informs efficient learning strategies (Artymowicz, 2024, Zhao et al., 2022, Wang et al., 2021).

2. Core Learning Methodologies

Several rigorous frameworks have been developed for Hamiltonian density operator learning, tailored to distinct data access and system settings:

• Maximum Likelihood and Free-Energy Descent: These approaches utilize observed measurement statistics or local expectation values in Gibbs or thermal states. The parameter vector $\mathcal{H}(x)$3 corresponding to the Hamiltonian expansion $\mathcal{H}(x)$4 is optimized via gradient descent on the negative log-likelihood (Zhao et al., 2022) or directly via the free-energy functional (Artymowicz, 2024, Wang et al., 2021). The update rule typically takes the form

$\mathcal{H}(x)$5

where the first term is estimated for the model and the second from data.

• Operator Inference and Tensor Parametric Models: Structure-preserving reduced-order models for Hamiltonian PDEs leverage affine parameterizations of operators, e.g., $\mathcal{H}(x)$6 contracted as $\mathcal{H}(x)$7. Learning, then, reduces to a convex least-squares fit with imposed symmetries for energy conservation (Vijaywargiya et al., 15 Feb 2025).
• Direct Time-Series Regression: For closed quantum systems, evolution data (e.g., electronic densities in small molecules) is used to fit a linear map between the time-discrete density matrices and the Hamiltonian, using least-squares optimization. The learning problem thus becomes a statistical regression constrained by the discrete-time Liouville equation (Bhat et al., 2020).
• Operator Neural Networks (DeepONets): For PDEs and continuous systems, operator learning is realized via architectures such as DeepONets, which approximate nonlinear maps from function spaces (e.g., spatial profiles) to the Hamiltonian density, with variational derivatives recovered via automatic differentiation (Xu et al., 27 Feb 2025).
• Quench-Constraint and Tomography-Based Protocols: In analog quantum simulators, time-series of expectation values after quenches allow for extraction of the generator via linear constraint equations based on the Ehrenfest theorem or generalized energy conservation. In large-scale quantum devices, spatial localization via the quantum Zeno effect enables patchwise process tomography, reconstructing local operator coefficients with explicit error controllability (Franceschetto et al., 19 Sep 2025, Olsacher et al., 2024).

3. Algorithmic and Statistical Considerations

Method

Data Regime / Approach	Key Algorithmic Step	Comments
Gibbs/Thermal state, local measurements	Free-energy minimization via SGD or SDP	Scalable to $\mathcal{H}(x)$8100 qubits (Artymowicz, 2024)
Projective measurement outcomes	Negative log-likelihood gradient descent	Robust for $\mathcal{H}(x)$9 (Zhao et al., 2022)
Analog quench/nonequilibrium dynamics	Linear constraint fit from time-resolved observables	Detects missing operator terms (Olsacher et al., 2024)
Patchwise tomography (quantum hardware)	Zeno effect localization + process tomography	Demonstrated up to 109 qubits (Franceschetto et al., 19 Sep 2025)
PDE/smooth function input	DeepONet/MLP operator architecture, automatic (variational) differentiation	No handcoding of differential operators
Parametric PDE ROMs	Tensor contraction plus symmetry-constrained least-squares	Energy conservation enforced

In all approaches, the selection of a physically motivated operator basis ("ansatz") is critical. Statistical error typically scales with sample size as $\mathcal{H}(x)$0 in the regime dominated by shot noise, with error plateaux indicating model mismatch or incomplete ansatz (Olsacher et al., 2024).

4. Hamiltonian Learning in Quantum Field Theories and PDEs

For effective field theories, Hamiltonian density learning explicitly targets the operator content and scale-dependent couplings: $\mathcal{H}(x)$1 The identification of $\mathcal{H}(x)$2 at a given spatial resolution $\mathcal{H}(x)$3 is accomplished by matching energy-conservation constraints or Schwinger–Dyson equations, constructed from experimentally measured correlation functions. This process is iterated across spatial scales, permitting the extraction of RG-like flows of the effective couplings, e.g.,

$\mathcal{H}(x)$4

and their associated $\mathcal{H}(x)$5-functions. Demonstrations include the extraction of sine-Gordon model parameters from Bose gas experiments with spatially coarse-grained measurements, revealing both the operator structure and scale-dependence of the Hamiltonian density (Ott et al., 2024).

In the setting of Hamiltonian PDEs, operator learning via DeepONet circumvents the manual discretization of derivatives. By training on data generated from random initial profiles and solution pairs, DeepONet learns to map function samples to the Hamiltonian density, and variational derivatives are computed purely via automatic differentiation, enabling symplectic integration of the resulting learned PDE (Xu et al., 27 Feb 2025).

5. Experimental Realizations and Validation

Representative experimental advances include:

• Quantum Hardware Verification: A Zeno-based protocol was implemented on an IBM 127-qubit device ("ibm_brisbane") to individually identify 109 local coefficients in a nearest-neighbor spin Hamiltonian. The protocol yielded median relative errors on local terms of $\mathcal{H}(x)$6 and was executed in under five minutes of wall-clock time (Franceschetto et al., 19 Sep 2025).
• Cold Atom Simulators: Field-theoretic Hamiltonian learning was applied to data from tunnel-coupled 1D Bose gases, enabling the extraction of both quadratic (free) and interacting (cosine, sine-Gordon) tensor densities, as well as their parameter flow with measurement scale (Ott et al., 2024).
• Molecular Electron Dynamics: For small molecules, learning a reduced-dimensional linear operator mapping densities to Fock matrices enabled accurate electron density propagation for thousands of time steps, even extrapolating to regimes of strong external fields not present in training data (Bhat et al., 2020).

In each scenario, explicit separation of scheme (integration) error and Hamiltonian model error is performed via propagation and reference trajectories, with error quantified by Frobenius norms or angle metrics in coefficient space.

6. Theoretical Guarantees and Scalability

Hamiltonian density operator learning methods provide quantifiable guarantees on parameter reconstruction:

• Statistical Error Bounds: Finite-sample errors in protocol schemes such as Zeno+tomography or maximum-likelihood inference scale as $\mathcal{H}(x)$7, where $\mathcal{H}(x)$8 is the number of projective measurement copies. Explicit error bounds in operator norm or diamond norm are established, and union-bound arguments ensure that reconstruction complexity grows only logarithmically in patch count above local scaling factors (Franceschetto et al., 19 Sep 2025, Zhao et al., 2022, Olsacher et al., 2024).
• Convexity and Stability: Many optimization objectives are convex (e.g., negative log-likelihood, free energy), with positive semi-definite Hessians ensuring robust convergence. Shot-noise models and model-mismatch detection via the singular value gap provide empirical means for diagnosing learning sufficiency (Olsacher et al., 2024).
• Scalability: Methods exploiting locality, parallelization (patchwise learning), and tensor-network machinery enable operator learning on large-scale systems (order 100 qubits) with no exponential scaling in total system size (Artymowicz, 2024, Franceschetto et al., 19 Sep 2025).

7. Outlook, Extensions, and Limitations

Operator learning of Hamiltonian densities is applicable across many system classes, including quantum channels, field theories, dissipative (Liouvillian) systems, and parametric PDEs. Limitations include the need for:

• A sufficient and physically consistent operator basis. Plateaux in error scaling signal model incompleteness.
• Positive-definiteness and faithfulness of correlation or estimator matrices, especially as $\mathcal{H}(x)$9 or with high noise (Artymowicz, 2024).
• Engineering of input states (e.g., in quantum simulators) and feasible measurement protocols for high-point correlators or nonclassical observables.

Extensions incorporate learning of time-dependent, spatially inhomogeneous, or open-system dynamics, probabilistic and Bayesian inference, and hybrid quantum-classical algorithmic structures (Wang et al., 2021, Olsacher et al., 2024).

Operator learning of Hamiltonian densities thus forms a unified framework enabling the rigorous, data-driven reconstruction of the fundamental generators of quantum and classical dynamics, leveraging advances in convex optimization, machine learning, and experimental control across domains (Franceschetto et al., 19 Sep 2025, Zhao et al., 2022, Vijaywargiya et al., 15 Feb 2025, Ott et al., 2024, Xu et al., 27 Feb 2025, Bhat et al., 2020, Artymowicz, 2024, Olsacher et al., 2024, Wang et al., 2021).