Heisenberg-Limited Hamiltonian Learning

• Heisenberg-limited Hamiltonian learning is a quantum metrology problem that reconstructs Hamiltonian parameters with error scaling as O(1/ε), surpassing the standard quantum limit.
• Techniques couple robust frequency/phase estimation with engineered controls and randomized operations to convert many-body dynamics into tractable subsystems.
• Practical implementations span few-qubit Pauli systems, hybrid spin-boson models, and fermionic lattices, illustrating trade-offs in resource allocation and noise sensitivity.

Heisenberg-limited Hamiltonian learning is the task of reconstructing the parameters, support, or selected functionals of a quantum Hamiltonian with precision that scales inversely with the total interrogation resource, rather than with its square root. In the most common formulation, the resource is the total evolution time under the unknown generator, and the target is to achieve error $\epsilon$ with $T_{\mathrm{tot}}=O(1/\epsilon)$ up to logarithmic factors instead of the standard-quantum-limit scaling $T_{\mathrm{tot}}=O(1/\epsilon^2)$. Within that umbrella, the literature now includes few-qubit Pauli Hamiltonians learned with static single-qubit fields (Brahmachari et al., 15 Jan 2026), hybrid spin-boson models learned with $O(\epsilon^{-1})$ total time and $O(\mathrm{polylog}(\epsilon^{-1}))$ measurements (Zhang et al., 27 Feb 2025), bosonic and fermionic lattice models with bounded-degree interactions (Li et al., 2023, Ni et al., 2023, Mirani et al., 2024), ansatz-free structure learning from real-time evolution (Bakshi et al., 2024), sparse nonlocal Hamiltonian learning via pseudo-Choi states (Zhao, 2024), and open-system extensions in which only part of the Hamiltonian remains Heisenberg-limited under dissipation (Romanov et al., 16 Jun 2026).

1. Metrological meaning and problem formulations

Across the literature, Hamiltonian learning begins from a parameterized generator, typically expanded in a Pauli or mode-operator basis, and asks for an estimator whose output lies within a prescribed norm ball around the true parameters. A representative few-qubit formulation considers a traceless $n$-qubit Hamiltonian

$$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$

with $\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$, and seeks

$$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$

using total evolution time $T=O(\epsilon^{-1}\log(1/\delta))$ in the Heisenberg-limited regime (Brahmachari et al., 15 Jan 2026). In that setting, the contrast class is the standard quantum limit, where naive prepare-evolve-measure procedures typically require $T_{\mathrm{tot}}=O(1/\epsilon)$0 for multi-parameter learning (Brahmachari et al., 15 Jan 2026).

The precise error metric varies with the model class. Hybrid spin-boson learning uses per-parameter RMSE guarantees for $T_{\mathrm{tot}}=O(1/\epsilon)$1, $T_{\mathrm{tot}}=O(1/\epsilon)$2, and $T_{\mathrm{tot}}=O(1/\epsilon)$3, again with total evolution time $T_{\mathrm{tot}}=O(1/\epsilon)$4 and $T_{\mathrm{tot}}=O(1/\epsilon)$5 measurements (Zhang et al., 27 Feb 2025). Structure-learning results instead target support recovery and coefficient estimation over an unknown Pauli dictionary, often in $T_{\mathrm{tot}}=O(1/\epsilon)$6 norm or thresholded support form (Bakshi et al., 2024, Zhao, 2024). Other works reinterpret Hamiltonian learning more broadly as learning a nontrivial functional of $T_{\mathrm{tot}}=O(1/\epsilon)$7, such as fidelity susceptibility, with Heisenberg-limited oracle complexity $T_{\mathrm{tot}}=O(1/\epsilon)$8 (Zhang et al., 1 Sep 2025).

A second source of variation is the resource notion itself. Most Hamiltonian-learning papers discussed here count total evolution time under the unknown generator (Brahmachari et al., 15 Jan 2026, Zhang et al., 27 Feb 2025, Li et al., 2023, Ni et al., 2023, Mirani et al., 2024, Bakshi et al., 2024). By contrast, the fidelity-susceptibility algorithm counts block-encoding queries to $T_{\mathrm{tot}}=O(1/\epsilon)$9 and $T_{\mathrm{tot}}=O(1/\epsilon^2)$0 (Zhang et al., 1 Sep 2025), while an in-situ Rydberg protocol formulates Heisenberg-like scaling in terms of coherent repetition depth $T_{\mathrm{tot}}=O(1/\epsilon^2)$1 and shot number $T_{\mathrm{tot}}=O(1/\epsilon^2)$2, with Cramér-Rao-saturating variances $T_{\mathrm{tot}}=O(1/\epsilon^2)$3 or $T_{\mathrm{tot}}=O(1/\epsilon^2)$4 depending on the parallelization mode (Liu et al., 9 Oct 2025). The shared theme is inverse-linear precision in a coherent metrological resource, but the exact metric is model dependent.

2. Common algorithmic architecture

A recurrent pattern is to reduce many-body Hamiltonian learning to frequency or phase estimation on engineered low-dimensional sectors. In the static-field qubit protocol, each control configuration produces an energy gap

$T_{\mathrm{tot}}=O(1/\epsilon^2)$5

of a controlled Hamiltonian, and the measured observables approximate $T_{\mathrm{tot}}=O(1/\epsilon^2)$6 and $T_{\mathrm{tot}}=O(1/\epsilon^2)$7 with bounded absolute error; robust frequency estimation then yields $T_{\mathrm{tot}}=O(1/\epsilon^2)$8-accurate gap estimates using total time $T_{\mathrm{tot}}=O(1/\epsilon^2)$9 (Brahmachari et al., 15 Jan 2026). In the hybrid spin-boson setting, random unitary transformations reduce the problem to commuting spin blocks or driven harmonic-oscillator signals, after which robust phase estimation and robust frequency estimation recover the relevant eigenvalue differences and displacement slopes with the same Heisenberg time scaling (Zhang et al., 27 Feb 2025).

A second recurring ingredient is Hamiltonian reshaping. Bosonic and fermionic protocols use random local phase rotations, beam splitters, or fermionic linear-optics gates to enforce effective symmetries, average away undesired couplings, and decompose the global problem into independent one- or two-body subsystems (Li et al., 2023, Ni et al., 2023, Mirani et al., 2024). In structure-learning results, the same idea appears in a more abstract form: one iteratively learns a residual Hamiltonian, rescales it, and simulates long effective evolution by constant-time increments, thereby bootstrapping constant-accuracy local estimators into overall $O(\epsilon^{-1})$0 scaling (Bakshi et al., 2024).

The final ingredient is an inverse map from the engineered observables back to the original coefficients. Sometimes this inverse map is explicit linear algebra, as in Hadamard-type systems for commuting spin blocks (Zhang et al., 27 Feb 2025). Sometimes it is local nonlinear inversion, as in the map from many controlled energy gaps to $O(\epsilon^{-1})$1 in few-qubit Pauli learning (Brahmachari et al., 15 Jan 2026). In all cases, Heisenberg-limited learning requires more than frequency estimation alone: it also requires that the coefficient-recovery map be sufficiently well conditioned that $O(\epsilon^{-1})$2 error in the signal domain induces only $O(\epsilon^{-1})$3 error in the Hamiltonian domain.

3. Static single-qubit fields and few-qubit Pauli Hamiltonians

A concrete realization of the subject is the protocol for learning

$O(\epsilon^{-1})$4

using only static single-qubit control fields of precision-independent strength (Brahmachari et al., 15 Jan 2026). The controlled Hamiltonian is

$O(\epsilon^{-1})$5

with

$O(\epsilon^{-1})$6

Here $O(\epsilon^{-1})$7 is a distinguished qubit, $O(\epsilon^{-1})$8 fixes signs, and $O(\epsilon^{-1})$9 fixes Pauli directions. The $O(\mathrm{polylog}(\epsilon^{-1}))$0 “defect” on qubit $O(\mathrm{polylog}(\epsilon^{-1}))$1 breaks the low-energy degeneracy and makes the lowest two eigenstates nondegenerate product states of the control part. For $O(\mathrm{polylog}(\epsilon^{-1}))$2, the ground and first excited states $O(\mathrm{polylog}(\epsilon^{-1}))$3 of $O(\mathrm{polylog}(\epsilon^{-1}))$4 obey

$O(\mathrm{polylog}(\epsilon^{-1}))$5

where $O(\mathrm{polylog}(\epsilon^{-1}))$6 are known unentangled product states of $O(\mathrm{polylog}(\epsilon^{-1}))$7 (Brahmachari et al., 15 Jan 2026).

The experiment prepares the product state

$O(\mathrm{polylog}(\epsilon^{-1}))$8

lets it evolve for time $O(\mathrm{polylog}(\epsilon^{-1}))$9, and measures one of two single-qubit Paulis $n$0 on qubit $n$1, chosen so that in the $n$2 subspace they act as $n$3 and $n$4. The resulting expectation values satisfy

$n$5

so the device behaves as an effective two-level system with gap $n$6 up to bounded distortion (Brahmachari et al., 15 Jan 2026). A robust frequency-estimation theorem then implies that if random variables approximating $n$7 and $n$8 are available to within $n$9 with probability at least $$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$0, $$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$1 can be estimated to RMS error $$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$2 using total time $$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$3 and only $$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$4 measurements. In the protocol, choosing

$$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$5

suffices to meet that bounded-error condition, and both constants are independent of $$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$6 (Brahmachari et al., 15 Jan 2026).

The passage from gap estimates to Hamiltonian coefficients is a nonlinear inverse problem. Writing

$$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$7

the reconstruction is posed as the least-squares minimization

$$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$8

The Jacobian satisfies

$$H(\boldsymbol{\lambda})=\sum_{\mathbf a\in\{0,1,2,3\}^n\setminus 0^n}\lambda_{\mathbf a}\,\sigma^{\mathbf a},$$9

with $\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$0 diagonal and

$\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$1

so $\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$2 is bounded below by a positive constant for sufficiently large but fixed $\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$3. The Hessian is locally strongly convex, and an implicit-map theorem yields a unique smooth local inverse obeying

$\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$4

Accordingly, $\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$5 error in the gap vector produces $\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$6 error in $\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$7, preserving the Heisenberg scaling established at the frequency-estimation stage (Brahmachari et al., 15 Jan 2026).

The same work also proves an information-theoretic lower bound: for single-qubit Hamiltonian learning with static controls of strength $\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$8, if the number $\boldsymbol{\lambda}\in\mathbb{R}^{4^n-1}$9 of discrete control operations satisfies $$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$0, then

$$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$1

This implies that a non-vanishing static field strength is necessary for Heisenberg-limited scaling unless one allows an extensive number of discrete control operations (Brahmachari et al., 15 Jan 2026).

4. Model classes and representative realizations

Heisenberg-limited Hamiltonian learning is no longer confined to a single physical setting. The current literature covers qubit, bosonic, fermionic, hybrid, structure-learning, and even certain open-system regimes.

Setting	Representative result	Characteristic resources
Few-qubit Pauli Hamiltonians	Static single-qubit fields and local measurements (Brahmachari et al., 15 Jan 2026)	$$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$2
Hybrid spin-boson systems	RUT + RPE/RFE for $$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$3 (Zhang et al., 27 Feb 2025)	$$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$4, $$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$5
Interacting bosons	Random unitaries enforce effective symmetries (Li et al., 2023)	$$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$6
Fermionic Hubbard models	FLO, Gaussian probes, local number measurements (Ni et al., 2023, Mirani et al., 2024)	$$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$7 total time
Hamiltonian structure learning	Unknown Pauli support learned from dynamics (Bakshi et al., 2024, Zhao, 2024)	$$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$8 or $$\Pr\!\big[\|\hat{\boldsymbol{\lambda}}-\boldsymbol{\lambda}\|_2\le \epsilon\big]\ge 1-\delta$$9
Open-system generators	HDD Hamiltonian sector remains HL (Romanov et al., 16 Jun 2026)	HDD: $T=O(\epsilon^{-1}\log(1/\delta))$0; full Lindbladian: SQL

In interacting bosons, the central device-level idea is symmetry enforcement by random unitaries. Random phase rotations average hopping terms to zero and reduce many-mode learning to single- or two-mode subproblems, which are then solved by robust frequency estimation using coherent states, beam splitters, phase shifters, and homodyne measurements (Li et al., 2023). Continuous-variable generalizations go further: the displacement-random-unitary-transformation protocol learns arbitrary finite-order bosonic Hamiltonians with $T=O(\epsilon^{-1}\log(1/\delta))$1 total evolution time, and extends the same strategy to first-quantized Hamiltonians in $T=O(\epsilon^{-1}\log(1/\delta))$2 and $T=O(\epsilon^{-1}\log(1/\delta))$3 under additional assumptions on the reference frame and a known vanishing coefficient (Huang et al., 9 Oct 2025).

For fermionic systems, the main technical difficulty is parity superselection. Heisenberg-limited learning of Fermi-Hubbard models therefore uses fermionic Gaussian states, fermionic linear optics, local number measurements, and $T=O(\epsilon^{-1}\log(1/\delta))$4 ancillary fermionic modes to remain in a fixed parity sector while still engineering interferometric signals. A simplified Hubbard class with real hoppings and zero chemical potentials was learned with $T=O(\epsilon^{-1}\log(1/\delta))$5 total time (Ni et al., 2023); a more general Hubbard model with complex hoppings and nonzero chemical potentials retains the same $T=O(\epsilon^{-1}\log(1/\delta))$6 total-time scaling and only $T=O(\epsilon^{-1}\log(1/\delta))$7 experiments, independent of system size on bounded-degree graphs (Mirani et al., 2024).

Hybrid spin-boson results show that Heisenberg-limited learning also extends to systems with both finite- and infinite-dimensional sectors. There, random unitary transformations cancel interaction terms or compress the spin Hamiltonian into commuting blocks, while robust phase and frequency estimation recover spin couplings, bosonic frequencies, and spin-boson couplings with $T=O(\epsilon^{-1}\log(1/\delta))$8 and $T=O(\epsilon^{-1}\log(1/\delta))$9 (Zhang et al., 27 Feb 2025). In open systems, the situation is more subtle: only Hamiltonian terms outside the dissipator footprint remain Heisenberg-limited, while the remaining Lindbladian parameters are SQL-limited (Romanov et al., 16 Jun 2026).

5. Structure learning, practical variants, and resource notions

A major development has been the transition from coefficient learning with known support to true structure learning. For $T_{\mathrm{tot}}=O(1/\epsilon)$00-local Pauli Hamiltonians with unknown interaction terms but bounded local $T_{\mathrm{tot}}=O(1/\epsilon)$01-norm,

$T_{\mathrm{tot}}=O(1/\epsilon)$02

structure learning from real-time evolution achieves

$T_{\mathrm{tot}}=O(1/\epsilon)$03

and for low-intersection local Hamiltonians this becomes $T_{\mathrm{tot}}=O(1/\epsilon)$04, with constant time resolution (Bakshi et al., 2024). The method combines bootstrap refinement, Trotterized residual evolution in constant time increments, and a Goldreich-Levin-type search over the Pauli spectrum. A separate pseudo-Choi-state approach learns any Pauli-sparse Hamiltonian with $T_{\mathrm{tot}}=O(1/\epsilon)$05 total evolution time when time reversal is available, but only $T_{\mathrm{tot}}=O(1/\epsilon)$06 in a forward-time-only model (Zhao, 2024). This sharp contrast has become one of the field’s clearest access-model trade-offs.

Practically oriented variants often emphasize deployability rather than a single universal resource metric. An in-situ parallelized algorithm for parallel-learnable many-body Hamiltonians reduces the number of experiment rounds from $T_{\mathrm{tot}}=O(1/\epsilon)$07 to $T_{\mathrm{tot}}=O(1/\epsilon)$08 for fully connected Ising couplings by learning many invariant two-dimensional subspaces simultaneously, and its estimators saturate the classical Cramér-Rao bound for the adopted measurement model (Liu et al., 9 Oct 2025). Another line shows that randomized product probes and single-shot Pauli measurements can exhibit a transient Heisenberg-limited regime in interrogation time, and sub-SQL scaling in total experiment time through appropriate time scheduling, without entanglement resources or dynamical control (Baran et al., 28 Jul 2025). These protocols use “Heisenberg-like” in a sense tied to the chosen coherent-depth or scheduling resource, rather than the total-time definition most common in the Hamiltonian-learning literature.

A distinct but related direction replaces explicit control design with model-based inference. Inverse physics-informed neural networks for Hamiltonian learning fit Schrödinger dynamics directly to measurement data and, in several noiseless spin-chain benchmarks, numerically achieve

$T_{\mathrm{tot}}=O(1/\epsilon)$09

which the authors describe as approaching the Heisenberg limit (Liu et al., 12 Jun 2025). Because those claims are empirical scaling fits rather than information-theoretic proofs, they occupy a different evidentiary category from the rigorous protocols above.

6. Limits, trade-offs, and unresolved questions

The literature now makes clear that “Heisenberg-limited Hamiltonian learning” is not a single theorem but a family of results conditioned on model class, control model, and noise assumptions. One common misconception is that Heisenberg scaling necessarily requires entangled probes or entangling gates. Several rigorous counterexamples are now available: static single-qubit fields suffice for few-qubit Pauli Hamiltonians (Brahmachari et al., 15 Jan 2026), coherent states and Gaussian controls suffice for interacting bosons (Li et al., 2023), and fermionic Gaussian states plus fermionic linear optics suffice for bounded-degree Hubbard models (Mirani et al., 2024). Another misconception is that Heisenberg scaling is universal once one can estimate a phase; in fact, access restrictions matter sharply. Time reversal enables $T_{\mathrm{tot}}=O(1/\epsilon)$10 sparse-Hamiltonian learning, whereas the forward-time-only model currently yields only $T_{\mathrm{tot}}=O(1/\epsilon)$11 (Zhao, 2024).

Noise introduces a further hierarchy. In ansatz-free Lindbladian learning, the Hamiltonian sector disjoint from the dissipator can remain Heisenberg-limited, but the dissipator itself and the Hamiltonian terms overlapping the dissipator are fundamentally SQL-limited (Romanov et al., 16 Jun 2026). A trajectory-based product-state protocol below the SQL proves only a transient Heisenberg-limited regime in short interrogation times (Baran et al., 28 Jul 2025). Many unitary-learning papers assume closed-system dynamics during interrogation and treat SPAM robustness, but not a full open-system error model (Brahmachari et al., 15 Jan 2026, Zhang et al., 27 Feb 2025). This suggests that the operational domain of rigorous Heisenberg scaling is currently broad but still significantly narrower than the set of experimentally relevant noisy settings.

Several open problems recur across the field. One is the removal of structural assumptions: structure learning is now possible without known supports (Bakshi et al., 2024, Zhao, 2024), but the strongest results still rely on sparsity, bounded local norm, or time reversal. Another is the extension from Hamiltonians to more general generators. Open-system learning already splits into Heisenberg- and SQL-limited sectors (Romanov et al., 16 Jun 2026), and continuous-variable first-quantized learning still requires a sufficiently accurate reference frame and a known vanishing coefficient (Huang et al., 9 Oct 2025). A further open question is how much of the Heisenberg-limited theory survives on realistic NISQ hardware with finite coherence, restricted connectivity, and imperfect controls. The emerging body of work indicates that the answer is not simply negative, but it is also not yet universal.