Hamiltonian and Lindbladian Learning

• Hamiltonian and Lindbladian learning is the inference of quantum dynamical generators, distinguishing closed-system unitary evolution from open-system Markovian processes in the GKSL form.
• It employs diverse access models—such as transient trajectories, Pauli probing, and Bell sampling—to estimate or certify both coherent and dissipative parameters.
• Advanced techniques like short-time linearization, Fourier analysis, and neural augmentation provide precise reconstruction and robust error bounds in complex, time-dependent quantum systems.

Hamiltonian and Lindbladian learning is the inference of quantum dynamical generators from measurement data. In the closed-system case, the target is a Hamiltonian $H$ governing unitary evolution; in the open-system Markovian case, the target is a generator $\mathcal L$ in GKSL form, typically written as

$$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$

The subject now spans several distinct tasks: full coefficient estimation, support or structure discovery, certification of whether dissipation is present at all, learning from transient trajectories or Gibbs states, and propagation of learning uncertainty into simulator predictions. Recent work treats both white-box and ansatz-free settings, time-independent and time-dependent generators, and both discrete-variable and continuous-variable systems (Arad et al., 18 Jun 2026, Ivashkov et al., 5 Mar 2026, França et al., 9 Oct 2025, Möbus et al., 31 May 2025).

1. Problem formulations and generator models

A central distinction is between Hamiltonian learning and full Lindbladian learning. In local Lindbladian learning, the unknown generator is often expressed in Pauli-Kossakowski form, with real Hamiltonian coefficients $h_a$ and complex dissipative coefficients $\gamma_{ab}$, and the goal is to estimate all coefficients to $\ell^\infty$ error at most $\varepsilon$ (Arad et al., 18 Jun 2026). In ansatz-free formulations, the support itself is unknown: the learner must discover which Pauli terms appear in the Hamiltonian part and which appear in the dissipator, without prior knowledge of locality structure or interaction graph (Ivashkov et al., 5 Mar 2026). In white-box formulations, by contrast, the operator structure is assumed known and only the coefficients are learned; one example writes

$$H_T=\sum_{j=0}^{4^N-1} c_j P_j,\qquad P_j\in\mathcal P_N,\quad c_j\in\mathbb R,$$

and estimates the coherent parameters and dissipative rates from data while preserving an interpretable GKSL form (Heightman et al., 8 Mar 2026).

The task may also be weaker than full reconstruction. A particularly sharp example is the certification problem for open dynamics: given black-box access to channels $E_t=e^{t\mathcal L}$, distinguish the case $\mathcal D=0$ from the case $\mathcal L$0, where $\mathcal L$1 is the dissipative component and $\mathcal L$2 is the normalized Frobenius norm on superoperators (Cai, 18 Mar 2026). This is explicitly a property-testing or certification problem rather than full learning.

Time dependence introduces a separate formulation. One recent protocol reconstructs unknown coefficient functions $\mathcal L$3 and $\mathcal L$4 of a time-dependent local generator $\mathcal L$5, under the assumption that the coefficients lie in a known finite-dimensional function family admitting stable interpolation (França et al., 9 Oct 2025). Continuous-variable Hamiltonian learning adds another layer of difficulty: the Hilbert space is infinite-dimensional and the generators are generally unbounded, so even defining learning guarantees requires analytic control of unbounded Lindbladian evolutions (Möbus et al., 31 May 2025).

A further structural subtlety is that the decomposition into $\mathcal L$6 and individual jump operators is not unique. The jump operators are gauge-dependent, and only the generator $\mathcal L$7 is physical in a representation-independent sense; this matters whenever “learning the Hamiltonian and the jumps” is interpreted too literally (Albert, 2018).

2. Access models and measurement primitives

Most recent learning protocols operate under restricted experimental access. A standard model is a black-box semigroup oracle $\mathcal L$8: the learner chooses times $\mathcal L$9, prepares inputs, applies the unknown channel, and measures outputs (Arad et al., 18 Jun 2026). Several algorithms are deliberately ancilla-free and non-adaptive, using only random product-state preparation and single-qubit Pauli measurements. One representative scheme prepares each qubit independently in a random Pauli eigenstate from $$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$0, applies $$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$1, and measures each output qubit in a random Pauli basis. This yields unbiased estimators for Pauli transfer matrix entries, with variance bounded by $$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$2 (Arad et al., 18 Jun 2026).

Another recurrent primitive is short-time Pauli probing. In the ansatz-free in situ protocol, each query prepares a product Pauli eigenstate, evolves for a chosen time, and measures each qubit in a single-qubit Pauli basis. The structure-learning stage estimates diagonal $$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$3-matrix entries $$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$4, or Pauli error rates, from random product inputs and Pauli measurements, then uses short-time derivatives to separate dissipative from coherent structure (Ivashkov et al., 5 Mar 2026). Time-dependent local learning uses a similar minimal access model—product Pauli eigenstates, evolution for chosen times, and product Pauli measurements—but combines it with process shadows and interpolation to reconstruct coefficient functions on an interval (França et al., 9 Oct 2025).

Certification protocols use different observables. Dissipation detection can be reduced to Bell sampling on the Choi state

$$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$5

with key statistic

$$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$6

For purely Hamiltonian dynamics $$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$7 for all $$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$8, whereas dissipative modes cause decay in $$\mathcal L(\rho)=-i[H,\rho]+\sum_a\left(L_a\rho L_a^\dagger-\frac12\{L_a^\dagger L_a,\rho\}\right).$$9 (Cai, 18 Mar 2026).

Not all access models rely on continuous-time control. A large-scale experimental protocol assumes access only to repeated applications of an operation $h_a$0 at integer depths, measures selected observables as functions of depth, fits those time series by sums of exponentially damped sinusoids, and differentiates the fitted curves to obtain the gradients needed for learning (Berg et al., 9 Dec 2025). A different line of work uses local expectation values of Gibbs states rather than real-time trajectories, formulating Hamiltonian learning as an SDP based on a free-energy variational principle and an entropy-production bound for Lindblad evolutions (Artymowicz, 2024).

These diverse access models are not interchangeable. A plausible implication is that “Hamiltonian and Lindbladian learning” is best viewed as a family of inverse problems indexed by the available probes—transient dynamics, repeated-depth time series, Choi-state Bell statistics, or thermal data—rather than as a single canonical estimation task.

3. Reconstruction and certification methodologies

A dominant algorithmic pattern is short-time linearization followed by stable inversion. In local Lindbladian learning, one estimates finite-time Pauli transfer matrices $h_a$1 at a small set of short times $h_a$2, differentiates numerically at $h_a$3 using Chebyshev–Lobatto interpolation, and reconstructs Lindbladian coefficients via a local Walsh–Hadamard transform on fixed-support sectors. The inversion is stable because the Hadamard matrix is orthogonal, so the condition number is exactly $h_a$4; support-agnostic recovery is handled by a thresholded peeling recursion (Arad et al., 18 Jun 2026).

Ansatz-free in situ learning uses a two-stage architecture. First, structure learning identifies the support of dissipative and Hamiltonian Pauli terms by exploiting the identities

$h_a$5

Second, coefficient learning turns derivatives of expectation values into a linear system $h_a$6, with “Lindbladian patchwise Pauli tomography” used to restrict probing to patch families determined by the learned supports (Ivashkov et al., 5 Mar 2026).

A more recent structure-learning framework defines local Fourier coefficients of the evolution on few-site regions and studies the obstruction unique to open systems, called “confusing” terms. In the Hamiltonian case the relevant local Fourier coefficients align with the target parameters, but for Lindbladians many Pauli terms contribute to the same local observable. The algorithm therefore uses an iterative update with explicit inversion of the confusion map and repeated rounding of small coordinates to zero, controlled by a smooth notion of approximate degree (Lewis et al., 29 Jun 2026).

Another ansatz-free route uses quantum error correction as a learning primitive. A recursive random stabilizer-code construction suppresses already identified Lindbladian terms while preserving sensitivity to weaker unknown ones. This yields an effective logical generator on the code space, after which Bell sampling and Choi-state low-rank observables isolate the remaining coefficients. The same framework gives a Heisenberg-limited learner for the Hamiltonian component disjoint from the dissipator footprint and an SQL learner for the full sparse Lindbladian (Romanov et al., 16 Jun 2026).

Optimization-based methods remain important when a trusted ansatz exists but the likelihood landscape is highly non-convex. One white-box method learns from Pauli measurement snapshots at multiple transient times by maximum-likelihood, backpropagating through a Lindblad ODE solver, and temporarily augmenting the physical generator with a neural differential equation correction. The neural term is then switched off to distill an interpretable GKSL solution (Heightman et al., 8 Mar 2026). In bounded-error quantum simulation, learning appears as a statistically regularized inverse problem $h_a$7, solved by least squares or semidefinite programming, followed by covariance estimation and uncertainty propagation to observables (Kraft et al., 28 Nov 2025).

Certification methods are algorithmically simpler because they do not reconstruct coefficients. In dissipative-noise detection, random evolution times, Pauli twirling, Bell sampling, and a short-time approximation to the twirled semigroup suffice to distinguish $h_a$8 from $h_a$9 under locality and bounded-degree assumptions (Cai, 18 Mar 2026).

4. Precision regimes, optimality, and complexity

Recent results establish that Hamiltonian and Lindbladian learning do not share a single information-theoretic scaling law. For full local Lindbladian learning with fixed locality and bounded dissipative site degree, one algorithm achieves $\gamma_{ab}$0 channel uses and $\gamma_{ab}$1 total evolution time, with matching lower bounds from a single-qubit dephasing family. These lower bounds hold even for adaptive algorithms with arbitrary ancillas and measurements, and imply that the Heisenberg-limited $\gamma_{ab}$2 scaling available in some Hamiltonian-learning settings is information-theoretically impossible once dissipative coefficients must be estimated (Arad et al., 18 Jun 2026).

Certification can be strictly cheaper than full learning. Detecting whether dissipation is absent or has magnitude at least $\gamma_{ab}$3 can be done with total evolution time $\gamma_{ab}$4, and this scaling is information-theoretically optimal; the lower bound is witnessed by a one-qubit depolarizing dissipator with rate $\gamma_{ab}$5 (Cai, 18 Mar 2026). This sharp contrast formalizes the difference between “Is there dissipation?” and “What are all dissipative coefficients?”

For constant-local Lindbladian structure learning, one recent theorem gives total evolution time

$\gamma_{ab}$6

where $\gamma_{ab}$7 is the single-site energy and $\gamma_{ab}$8 is the approximate degree of the interaction graph. The same work emphasizes a time resolution of only $\gamma_{ab}$9, logarithmic dependence on $\ell^\infty$0, and support-agnostic learning of quasi-local and power-law Lindbladians (Lewis et al., 29 Jun 2026). For arbitrary sparse Lindbladians, the first ansatz-free SQL learner achieves

$\ell^\infty$1

total evolution time, while under the balanced Kossakowski tail condition the Hamiltonian component disjoint from the dissipator footprint can be learned at

$\ell^\infty$2

total evolution time (Romanov et al., 16 Jun 2026).

Time dependence changes the resource law in a different way. Under a Markov-stable function-system assumption of dimension $\ell^\infty$3, time-dependent local Hamiltonian and Lindbladian coefficients can be learned on an interval with

$\ell^\infty$4

samples, avoiding the exponential dependence on $\ell^\infty$5 that would arise from naïve high-order derivative estimation at $\ell^\infty$6 (França et al., 9 Oct 2025).

Continuous-variable Hamiltonian learning provides a contrasting positive result. For bosonic systems with low-intersection polynomial Hamiltonians, strong engineered dissipation projects the dynamics onto a cat-code-like subspace, enabling total evolution time $\ell^\infty$7 in the multi-mode setting and hence Heisenberg-limited scaling in the target precision $\ell^\infty$8 (Möbus et al., 31 May 2025). Taken together, these results suggest that the decisive issue is not merely whether the system is open, but which components are being estimated, how dissipation enters the identifiability structure, and what physical resources the protocol can exploit.

5. Experimental realizations and practical uses

Hamiltonian and Lindbladian learning has moved beyond small proof-of-principle settings. A time-series-based Lindblad learning protocol was demonstrated on IBM’s 156-qubit superconducting processor ibm_pittsburgh, where a local Lindblad model was learned for a full layer of operations including two-qubit $\ell^\infty$9 gates, single-qubit rotations, and idle periods. The learned model included one- and two-local Hamiltonian terms and single-qubit dissipative terms, captured coherent gate terms, idle-qubit errors, and crosstalk, and provided estimates of $\varepsilon$0 and $\varepsilon$1 via the dissipative coefficients (Berg et al., 9 Dec 2025).

In bounded-error quantum simulation, Hamiltonian and Lindbladian learning serves as the characterization stage of a larger prediction-with-uncertainty framework. On trapped-ion simulators implementing long-range Ising interactions, a 10-ion experiment used learned open-system models to verify consistency between predicted and measured long-time dynamics, while a 51-ion experiment used a differential learning method with a $\varepsilon$2 regression matrix and found dephasing matrices with off-diagonal structure indicating collective dephasing (Kraft et al., 28 Nov 2025). The same framework is explicitly formulated to extend to digital quantum simulation through learned effective Floquet Hamiltonians and Lindbladian noise models.

White-box transient-time learning has been benchmarked across neutral-atom and superconducting Hamiltonians, as well as Heisenberg XYZ and PXP chains, with robustness to phase noise, thermal noise, and their combination. The reported numerical studies cover $\varepsilon$3 qubits, noise-to-signal ratios $\varepsilon$4, and fewer than $\varepsilon$5 shots (Heightman et al., 8 Mar 2026). The methodological emphasis there is not only coefficient recovery but also identifying when a neural augmentation improves optimization and when it constitutes overparameterization.

Several papers frame the practical relevance in hardware terms. Ansätze-free in situ learning is motivated by calibration of quantum hardware, design of robust simulation protocols, tailored error-correction methods, and characterization of native device dynamics without extra control overhead (Ivashkov et al., 5 Mar 2026). Dissipation certification is positioned as a lightweight diagnostic for whether observed dynamics are essentially coherent or contaminated by dissipative noise, requiring only chosen evolution times and Bell-basis measurements rather than full process tomography (Cai, 18 Mar 2026).

6. Identifiability, limitations, and broader connections

A recurrent misconception is that accurate prediction of output states is sufficient to identify the generator. Open-system learning results explicitly reject this. Distinct Lindbladians can share the same or similar steady states, so low infidelity at long times does not certify correct recovery of Hamiltonian couplings or dissipative rates; transient snapshots are essential because steady-state data can be almost blind to the Hamiltonian even when it still constrains the dissipator (Heightman et al., 8 Mar 2026). This is consistent with the lower-bound intuition in ansatz-free in situ learning: coarse time resolution can drive the system exponentially close to steady state, at which point structure learning becomes exponentially hard in $\varepsilon$6 for some sparse Lindbladians (Ivashkov et al., 5 Mar 2026).

Another important distinction concerns model adequacy. In weakly dissipative many-body learning based on quench dynamics, the experimentally accessible ratio $\varepsilon$7 of the two smallest singular values is used as a learning-error proxy: $\varepsilon$8 decay indicates a shot-noise-dominated regime, whereas saturation or a plateau indicates that the ansatz is missing operator terms. The same work emphasizes reparametrization of the ansatz as a classical post-processing tool for reducing parameter count and increasing singular-value gaps (Olsacher et al., 2024).

Noise and SPAM effects enter unevenly across protocols. Time-series Lindblad learning is described as in principle insensitive to state-preparation errors because the Ehrenfest relation holds for any initial state, but readout errors do affect performance, and mitigation based on confusion-matrix inversion can inherit bias from state-preparation imperfections (Berg et al., 9 Dec 2025). Many protocols also assume Markovian and time-independent dynamics; when those assumptions fail, standard Lindbladian learners may become model-mismatched. A notable counterexample is a $\varepsilon$9-dependent mixed generator derived from a system coupled to an ancillary continuum sink and a bath, interpolating continuously between non-Hermitian and Lindbladian limits and yielding generally non-Markovian reduced dynamics (Finkelstein-Shapiro, 2022).

Broader structural work sharpens the conceptual relation between Hamiltonian and Lindbladian viewpoints. The asymptotic projection $$H_T=\sum_{j=0}^{4^N-1} c_j P_j,\qquad P_j\in\mathcal P_N,\quad c_j\in\mathbb R,$$0, the dissipative gap, and the distinction between asymptotic and decaying sectors provide a general framework for understanding what can and cannot be inferred from long-time data, especially when multiple steady states are present (Albert, 2018). Related research also uses Hamiltonian–Lindbladian correspondences outside inverse problems: Hamiltonian twirling channels realize certain Lindbladian semigroups and yield ancilla-free fast-forwarding algorithms for a single Hermitian jump operator (Gao et al., 13 Nov 2025), while stochastic quantum Hamiltonian descent treats a Lindbladian dissipator as the quantum analogue of stochastic-gradient noise in open-system quantum optimization (Peng et al., 21 Jul 2025).

Hamiltonian and Lindbladian learning is therefore not a single method but a layered research area at the intersection of system identification, open-quantum-systems theory, measurement design, and complexity theory. Its main organizing distinctions are between certification and reconstruction, white-box and ansatz-free inference, Hamiltonian-only and full dissipative estimation, transient and steady-state data, and Markovian semigroup models versus more general effective dynamics.