Pith. sign in

REVIEW 4 major objections 6 minor 80 references

Heisenberg-Limited Quantum Hamiltonian Learning via Randomly Spread Product-States

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that Heisenberg-limited quadratic Fisher-information scaling can be made practically accessible for Hamiltonian learning using only local random product states, random Pauli measurements, and no dynamical control, with…

desk verdict A practical local-operations protocol that likely achieves finite-window Heisenberg-like Fisher scaling, with convincing numerics but a proof gap in the main theorem and an overreaching scheduling claim. read the letter →

arxiv 2507.21374 v5 pith:3NUHK4AG submitted 2025-07-28 quant-ph

Bora Baran , Timothy Heightman This is my paper
classification quant-ph
keywords HamiltonianlearningHeisenberglimitstandardquantumclassicalFisherinformationspreadstatesrandomPaulimeasurementsparameterestimationmultiparameter
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that Heisenberg-limited quadratic Fisher-information scaling, normally confined to the infinitesimal time limit or to costly control setups, can be made practically accessible for Hamiltonian learning using only local operations. The protocol prepares locally Haar-random product states ('spread states'), lets them evolve under an unknown Hamiltonian, and measures in random Pauli-product bases. These choices activate the full Hamiltonian spectrum in the measurement probabilities, so the classical Fisher information grows as $t^2$ over a finite temporal window and, via the Cramér–Rao bound, the estimation error scales as $1/t$. The paper also claims that averaging over many spread states diagonalizes the Fisher information matrix, allowing all Hamiltonian parameters to be learned simultaneously from the same dataset. Numerically, on five-qubit disordered Heisenberg-type models, the protocol exceeds the standard quantum limit in a black-box maximum-likelihood recovery setting.

What carries the argument

The central object is the spread state: a fixed product state rotated by independent single-qubit Haar-random unitaries. Its role is to ensure that the expansion coefficients $a_k=\langle\lambda_k|\psi_{\text{spread}}\rangle$ of the Hamiltonian eigenbasis are all nonzero, so no spectral gap is suppressed by the initial state. The spectral-activation condition is expressed through the coefficient $c^*_{jk} c_{j\ell} a_k a^*_\ell$ multiplying each oscillatory term $e^{-i(\lambda_k-\lambda_\ell)t}$ in the measurement probability; when these coefficients are nonzero for all pairs $(k,\ell)$, every spectral gap contributes to the time derivative of the probability, giving $\partial_\theta p_j(t)=\Omega(t)$. The paper combines this with $I_C(t)=\sum_j (\partial_\theta p_j(t))^2/p_j(t)\ge \sum_j(\partial_\theta p_j(t))^2$ to obtain $I_C(t)=\Omega(t^2)$. Ensemble diagonalization then follows from the Haar-measure identity that distinct Pauli-string expectations have zero mean and are statistically independent, so cross-parameter correlations vanish while diagonal variances stay positive.

What would settle it

Evaluate $F_j(t)=\partial_\theta p_j(t)/t$ for a specific random $n$-qubit Hamiltonian and see whether the minimum over measurement outcomes $j$ of $|F_j(t)|$ decays with $n$ or crosses zero before $t=\pi/(2\Delta\lambda_{\max})$; either behavior would close the claimed $\Omega(t^2)$ window. A systematic check would scan many disorder realizations and look for any realization where the fitted Fisher exponent falls below 2 inside the window.

Watch

Extended reading notes

Core claim

On its own terms, the paper proves that if a probe-and-measurement ensemble activates all relevant spectral gaps—so that the coefficients $c^*_{jk} c_{j\ell} a_k a^*_\ell$ in the measurement probability expansion are nonzero—then the classical Fisher information obeys $I_C(t)=\Omega(t^2)$ on $t\in(0,\pi/(2\Delta\lambda_{\max}))$. Spread states achieve this activation with probability one because independent local Haar rotations make every eigenstate amplitude nonzero, and random Pauli-product measurements prevent basis-induced suppression. A second theorem states that at leading order in time, the ensemble-averaged Fisher information matrix converges to $\operatorname{diag}(c_1,\ldots,c_d)$ with positive diagonal entries, decoupling parameter estimation. A third result shows that when multiple interrogation times $t_k=\Delta t\,k^\alpha$ are scheduled, cumulative Fisher information scales as $T^p$ with $p=(\alpha\gamma_0+1)/(\alpha+1)$, interpolating from $p=1$ (SQL-like) to $p\to 2$ (Heisenberg-like) as $\alpha$ grows. The paper interprets these results as opening the Heisenberg-limited short-time regime to devices without entanglement, coherent joint measurements, or dynamical control.

Load-bearing premise

The proof of the quadratic window assumes that the oscillatory sum $F_j(t)$ remains bounded away from zero over the entire pre-extremal interval for Haar-random spread states, but the paper does not quantify this lower bound or show that it is independent of system size.

Editorial extensions

If this is right

  • If Theorem 1 holds, any Hamiltonian parameter associated with an activated spectral gap can be estimated with precision scaling as $\Delta\theta=O(t^{-1})$ using only product-state probes and single-shot Pauli measurements.
  • If Theorem 2 holds, the same dataset can be reused for all parameters: ensemble averaging removes the need for parameter isolation, so full Hamiltonian matrices can be learned in parallel.
  • If Proposition 1 holds, the measurement-time schedule $t_k=\Delta t\,k^\alpha$ continuously interpolates between standard-quantum-limit and Heisenberg-like cumulative scaling, with the effective exponent $p\to 2$ as $\alpha\to\infty$.
  • The protocol requires no dynamical control and no entanglement, making it applicable to systems where continuous many-body control or coherent joint measurements are unavailable.
  • Numerical Fisher diagnostics indicate that the quadratic regime appears already with a single spread state given enough measurement bases, and increasing the ensemble size mainly diagonalizes the Fisher matrix and improves conditioning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • One extension suggested by the mechanism, though not explored in the paper, is to replace Haar-random rotations with approximate $t$-designs such as random Clifford circuits; the nonzero-overlap argument should survive while lowering the experimental overhead.
  • Because the quadratic window has length $\pi/(2\Delta\lambda_{\max})$, the practical advantage shrinks for low-gap or gapless Hamiltonians; a natural follow-up would schedule interrogation times adaptively to remain inside the window.
  • The ensemble-diagonalization argument relies only on local Haar moments, so the same averaging trick could plausibly be extended to learning Lindblad generators or other local operator expansions, not just closed-system Hamiltonians.
  • A testable extension would be to verify whether small state-preparation or measurement noise destroys the spread-state overlaps fast enough to close the quadratic window; the paper lists robustness to such noise as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a Hamiltonian learning protocol based on locally Haar-random product states ('spread states') and random Pauli-product measurements. It claims (Theorem 1) that, generically, the classical Fisher information obeys I_C(t) = Ω(t^2) on the window (0, π/(2 Δλ_max)], giving Heisenberg-like sensitivity in the interrogation time without entanglement, coherent measurements, or dynamical control. Theorem 2 claims that the ensemble-averaged Fisher information matrix becomes diagonal, allowing all Hamiltonian parameters to be estimated simultaneously from the same dataset. Section 1.4 and Proposition 1 analyze how cumulative Fisher information scales with total time under power-law schedules t_k = Δt k^α, claiming p = (α γ0 + 1)/(α + 1), which tends to 2 as α → ∞. Numerical recovery experiments on 5-qubit disordered Heisenberg models report error scaling ε ∝ T^{-β} with β ≈ 0.66–0.75, surpassing the SQL, and Fisher diagnostics support the predicted cumulative scaling and diagonalization trend. Open-source code is provided.

Significance. The core idea—making the short-time quadratic Fisher regime practically accessible with only local state preparation and local measurements—is valuable and, if rigorously established, would be a notable step beyond control-based Heisenberg-limited Hamiltonian learning. The paper is clearly written, the numerical system-size scaling diagnostics are a strength, and the open-source implementation is commendable. However, the main theorem's proof rests on an unproven generic lower bound for a trigonometric sum, and the claimed total-time Heisenberg scaling in Sec. 1.4 is inconsistent with the finite window of Theorem 1. The central claims are therefore not yet established at the level claimed, although they may be salvageable with additional rigorous arguments or with appropriately weakened statements.

major comments (4)
  1. [Appendix B, proof of Theorem 1] The proof asserts that the oscillatory sum F_j(t) = -i Σ_{k,ℓ} c*_{jk} c_{jℓ} a_k a*_ℓ e^{-iΔ_{kℓ}t} ∂_θ Δ_{kℓ} is generically Ω(1) on the whole interval (0, π/(2Δλ_max)], because persistent destructive interference would require fine-tuning that occurs only on a measure-zero subset of coefficient space. This is not implied by Lemma C. Lemma C only shows that each amplitude a_k is almost surely nonzero for fixed k; it says nothing about the magnitude of the trigonometric sum uniformly over a continuum of t. A finite sum of exponentials with nonzero coefficients can vanish at an isolated time, or dip arbitrarily close to zero, without any exotic fine-tuning of the coefficients; the coefficient set for which such a dip occurs need not be measure-zero over the interval. Since ∂_θ p_j(t) = T_vec(t) + t F_j(t) with T_vec(t) = O(1), a zero or o(1) dip of |F_j(t)| at any t in the claimed window invalidates the Ω(t^2) Fisher bound at that time. The proof also does not address spectral degeneracies, where distinct pairs (k,ℓ) with equal Δ_{kℓ} can interfere coherently. To establish Theorem 1, the authors need a quantitative uniform lower bound, e.g., inf_{t∈(0,π/(2Δλ_max))} |F_j(t)| ≥ c > 0 with a stated dependence on n, proven for Haar-random spread states; alternatively, the theorem must be restricted to times where such a bound can be shown, or explicitly presented as a conjecture supported by numerics.
  2. [Section 1.4, Proposition 1] The claimed recovery of Heisenberg total-time scaling as α → ∞ is inconsistent with the fixed temporal window of Theorem 1. In the proof in Appendix F, m_t is taken to infinity with Δt and α fixed, so t_k = Δt k^α eventually exceeds π/(2Δλ_max); the hypothesis 'each t_k lies within the regime where F(t_k) = Θ(t_k^{γ0})' is then violated for all sufficiently large k. If, instead, one respects the window by imposing t_k ≤ τ := π/(2Δλ_max), then for m_t measurements the total time satisfies T_tot = Σ t_k ≤ m_t τ and the total Fisher information satisfies I_tot ≤ C Σ t_k^2 ≤ C τ Σ t_k = C τ T_tot. In fact, optimizing the schedule under the constraint t_k ≤ τ gives I_tot ∝ T_tot (p = 1), not p → 2; the superlinear exponent p = (αγ0 + 1)/(α + 1) arises precisely because t_k grows without bound, which is outside the regime where I(t) = Θ(t^2) has been proven. The numerical sweep in Fig. 3 only covers α ≤ 1 and cannot demonstrate p → 2. This total-time scaling claim should be corrected or removed, or an explicit resource analysis within the finite window should be provided.
  3. [Appendix E, proof of Theorem 2] The averaging step in Eq. (70) does not establish that off-diagonal Fisher entries vanish at leading order. The expectation is taken over E_ψ[(1/p_i^{(r)}(θ)) ⟨ψ|[H_j, Π_i]|ψ⟩ ⟨ψ|[H_k, Π_i]|ψ⟩], and the prefactor 1/p_i^{(r)}(θ) is a nontrivial function of the same random state ψ. The argument that E[⟨ψ|Q_α|ψ⟩⟨ψ|Q_β|ψ⟩] = 0 for Q_α ≠ Q_β applies to the bare product of Pauli expectation values, but the denominator correlates with the numerator. A rigorous treatment must either expand 1/p_i around a Haar-averaged value and control the resulting terms, or prove a symmetry of the full integrand. Without this, the claimed limit diag(c_1, …, c_d) is not proven. The numerical diagonalization trend in Fig. 7 is suggestive but does not replace this proof.
  4. [Sections 2.6–2.6.2, Figs. 4–7] The numerical Fisher diagnostics are computed by automatic differentiation of the same model probabilities p^{(rjk)}_x(θ) used in the theory, so they validate internal consistency with the short-time expansion rather than providing an out-of-sample test of Theorem 1. In particular, they do not probe the uniform lower bound F_j(t) = Ω(1) near the boundary of the claimed window, and they do not scan for dips of the Fisher information as a function of t within the window. Moreover, the paper does not report Δλ_max for the simulated Hamiltonians or check that every time stamp t_k satisfies t_k ≤ π/(2Δλ_max); without this check, the observed agreement with the quadratic prediction cannot be unambiguously attributed to Theorem 1. The authors should either restrict the numerical claims to the proven window or provide explicit evidence (and a proof, if possible) that the quadratic regime extends beyond the stated bound.
minor comments (6)
  1. [Notation, Eqs. (3)–(8)] The symbol F_C(t) in Eq. (3) is later replaced by I_C(t) in Eqs. (8) and elsewhere; please use one notation consistently.
  2. [Sec. 2.2, Eq. (26)] The estimator in Eq. (26) includes a shot index s and a product over s, but the protocol is described as one-shot measurements; clarify how repetitions S enter the dataset and the loss function.
  3. [Sec. 1.3, Theorem 2] Theorem 2 states that measurements are performed in a fixed Pauli product basis, while the protocol description in Sec. 1.2 and the numerical sections use random Pauli-product bases; align the theorem statement with the protocol or define the averaging over bases as well.
  4. [Fig. 1 and Sec. 2.3.1] Fig. 1 reports β ≈ 0.66, while Eqs. (29)–(30) with α = 1 and γ0 = 2 predict β = 0.75; the caption and text should comment on this discrepancy, which is only addressed later through the vertical offset in Fig. 3.
  5. [Appendix I, Table 1] The first column entries such as '10.019±0.020' are ambiguous; if they represent R = 1 with value 0.019, the table should be reformatted to avoid this misreading.
  6. [Figs. 2–7] Several figures appear to use multiple markers or colors (e.g., across Hamiltonian families or system sizes) without a legend in the caption; adding legends would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central Fisher-scaling theorems are derived, not fitted, and the numerical diagnostics test the derived formulas against exact simulation rather than encoding them by construction.

full rationale

The central derivation chain is not circular. Theorem 1 derives I_C(t)=Ω(t^2) from a short-time expansion of p_j(t) plus a generic lower bound on the oscillatory sum F_j(t); that lower bound is asserted rather than rigorously proved, but it is not supplied by fitting, by definition, or by a self-citation. Theorem 2 follows from Haar-random local rotations and the vanishing of off-diagonal Pauli expectation values. Proposition 1 is a power-law remainder computation from the schedule t_k=Δt k^α and the assumed single-time scaling F(t)=Θ(t^{γ0}); the cumulative exponent p=(αγ0+1)/(α+1) is derived mathematically, not obtained by fitting. The numerical recovery and Fisher-diagnostic experiments simulate exact unitary dynamics and then fit scaling exponents (β or p) from the resulting data; those fitted values are compared with the derived predictions, not inserted back into the theory. The paper contains self-citations ([24], [69]), but they concern a recovery method and code availability, not the load-bearing theoretical claims, so they do not constitute circular support. The main weakness is the unproven uniform lower bound on F_j(t) in Appendix B, which is a correctness/proof gap rather than a circularity, because the theorem's conclusion is not equivalent to the bound by construction. No step was found in which an input is renamed as a prediction or a fitted parameter is relabeled as an output.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The protocol itself introduces no new physical entities. The main unproven inputs are the genericity lower bound in the Fisher-window proof, the assumption that the short-time expansion remains valid on the whole spectral window, and the resource-accounting rule that ignores preparation and measurement overhead. The numerical predictions fix gamma_0 = 2 and Delta_t = 0.01 by hand.

free parameters (3)
  • Single-time Fisher exponent gamma_0 = 2 (assumed from Theorem 1, not fitted)
    Used in Prop. 1 and all numerical predictions to convert single-time quadratic scaling into cumulative exponent p = (alpha gamma_0 + 1)/(alpha + 1). The theorem meant to justify it has a genericity gap.
  • Time-step scale Delta_t = 0.01 (chosen)
    Sets the evolution-time grid t_k = Delta_t k^alpha in all numerical experiments; the results depend on all stamps lying in the quadratic window.
  • Vertical offset in Fig. 3 dashed curve = estimated from data
    The dashed theoretical curve includes a small vertical offset accounting for finite ensemble and time stamp sampling effects; the offset is not derived.
assumptions (5)
  • standard math Short-time Taylor expansion p_j(t) = p_j(0) + a_j t + O(t^2) and derivative d_theta p_j = (d_theta a_j) t + O(t^2)
    Used in Eqs. (5)-(8) as the basis of quadratic Fisher scaling.
  • standard math Classical Cramer-Rao bound and Fisher information definitions
    Standard statistical estimation background used throughout Sec. 1.1.
  • domain assumption Resource accounting counts only interrogation time; state preparation and measurement costs are treated as linear and excluded
    Stated in the Introduction; this convention is what makes the superlinear total-time scaling interpretation possible.
  • ad hoc to paper Generic lower bound F_j(t) = Omega(1) on the pre-extremal window
    Appendix B asserts destructive interference among activated spectral gaps occurs only on a measure-zero set, but gives no quantitative proof.
  • ad hoc to paper All time stamps remain inside the quadratic window for Proposition 1 asymptotics
    Prop. 1 assumes F(t_k) = Theta(t_k^gamma_0) for every k while taking m_t or alpha large; for a fixed Hamiltonian this is inconsistent with t_k <= pi/(2 Delta_lambda_max).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Heisenberg-Limited Quantum Hamiltonian Learning via Randomly Spread Product-States." pith.science (2026). https://pith.science/paper/3NUHK4AG

@misc{pith2026250721374,
  author       = {Pith},
  title        = {Pith review of: Heisenberg-Limited Quantum Hamiltonian Learning via Randomly Spread Product-States},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3NUHK4AG}},
  note         = {Machine review of arXiv:2507.21374}
}
read the original abstract

We show how the Heisenberg-limited quadratic Fisher-information regime of short-time quantum evolution can be made practically accessible for quantum Hamiltonian learning, using only local quantum operations. Our protocol uses experiments initialized in locally Haar-random product states, accompanied by random one-shot Pauli-product measurements, leading to the activation of the full Hamiltonian spectrum in the measurement statistics. This extends the naturally given quadratic Fisher scaling of short-time dynamics into a practically accessible temporal window without requiring entanglement, globally coherent measurements, or dynamical control. Furthermore, we show that the act of ensemble averaging over these initial states makes unbiased estimation data, meaning all Hamiltonian parameters can be simultaneously estimated from the same data-set, removing the need for parameter isolation. We supplement the theoretical results by showing empirically that, even away from the asymptotic limit, one can surpass the SQL using randomly spread product-state ensembles. We do so numerically by learning a selection of different disordered multi-qubit Hamiltonians in a black-box learning scenario.

Figures

Figures reproduced from arXiv: 2507.21374 by the authors.

Figure 1
Figure 1. Reconstruction error ε versus total experiment time Ttot for four representative Hamiltonian families at α = 1.0. Each data point is given with ten random Hamiltonian realizations. Across all Hamiltonians, the error decays as ε ∝ T −β tot with β ≈ 0.66, surpassing the standard quantum limit and consistent with Heisenberg-limited single-probe scaling. 2.5 Managing the Super-Linear learning rate Degradation in Multi-T… view at source ↗
Figure 2
Figure 2. (a) Reconstruction error ε versus total experiment time Ttot (per spread state) for growing number of spread states R (XYZ model Eq. 24a). (b) Corresponding scaling exponents β obtained from ε ∝ T −β tot across Hamiltonian families defined in Sec. 2.1. As R increases, β converges toward the prediction β ≈ 0.75 (see Eqs. 29 and 30 with α = 1.0, assuming Heisenberg-limited scaling with single evolution time t, γ0 = 2)… view at source ↗
Figure 3
Figure 3. Empirical scaling exponent βTtot as a function of the scheduling parameter α. The solid curve shows the theoretical prediction βTtot (α) = 1 2 αγ0+1 α+1 given based on assuming Heisenberg limited scaling:γ0 = 2, and the dashed curve includes a small vertical offset accounting for finite ensemble and time stamp sampling effects on the diagonalization of the Fisher Matrix, affecting the recovery performance, that vani… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Fisher-information diagnostic for varying numbers of spread states. (a) Growth of the Fisher trace [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Fitted cumulative Fisher-information scaling exponent [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 7
Figure 7. Figure 7: Diagonalization of the Fisher information ma [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

80 extracted references · 65 canonical work pages

  1. [47]

    Zubida, A., Yitzhaki, E., Lindner, N. H. & Bairey, E. Optimal short-time measurements for hamiltonian learning. arXiv preprint arXiv:2108.08824(2021)

    arXiv 2021

  2. [1]

    & Nori, F

    Zhang, J., Liu, Y .-x., Wu, R.-B., Jacobs, K. & Nori, F. Quantum feedback: Theory, experiments, and applications. Physics Reports679, 1–60 (2017)

    work page 2017

  3. [2]

    M., Tagliacozzo, L., Deutsch, I

    Hauke, P., Cucchietti, F. M., Tagliacozzo, L., Deutsch, I. & Lewenstein, M. Can one trust quantum simulators? Reports on Progress in Physics75, 082401 (2012)

    work page 2012

  4. [3]

    URL https://www.science.org/doi/abs/10.1126/sciadv.adt4713

    Guo, S.-A.et al.Hamiltonian learning for 300 trapped ion qubits with long-range couplings.Science Advances11, eadt4713 (2025). URL https://www.science.org/doi/abs/10.1126/sciadv.adt4713. https://www. science.org/doi/pdf/10.1126/sciadv.adt4713

    doi:10.1126/sciadv.adt4713 2025

  5. [4]

    & Cory, D

    Wiebe, N., Granade, C., Ferrie, C. & Cory, D. G. Hamiltonian learning and certification using quantum resources.Phys. Rev. Lett.112, 190501 (2014). URL https://link.aps.org/doi/10.1103/PhysRevLett. 112.190501

    doi:10.1103/physrevlett 2014

  6. [5]

    & Greplova, E

    Valenti, A., van Nieuwenburg, E., Huber, S. & Greplova, E. Hamiltonian learning for quantum error correction. Physical Review Research1, 033092 (2019)

    work page 2019

  7. [6]

    & Liu, D

    Guo, R.-C., Gu, Y . & Liu, D. E. Mitigating errors in analog quantum simulation by hamiltonian reshaping or hamiltonian rescaling.npj Quantum Information11, 14 (2025)

    work page 2025

  8. [7]

    & Datta, A

    Baumgratz, T. & Datta, A. Quantum enhanced estimation of a multidimensional field.Physical review letters116, 030801 (2016)

    work page 2016

Show all 80 references
  1. [8]

    Ferrie, C., Granade, C. E. & Cory, D. G. How to best sample a periodic probability distribution, or on the accuracy of hamiltonian finding strategies.Quantum Information Processing12, 611–623 (2013)

    2013

  2. [9]

    & Motzoi, F

    Singh, J., Zeier, R., Calarco, T. & Motzoi, F. Compensating for nonlinear distortions in controlled quantum systems.Physical Review Applied19, 064067 (2023)

    2023

  3. [10]

    Sergeevich, A., Chandran, A., Combes, J., Bartlett, S. D. & Wiseman, H. M. Characterization of a qubit hamiltonian using adaptive measurements in a fixed basis.Physical Review A—Atomic, Molecular, and Optical Physics84, 052315 (2011)

    2011

  4. [11]

    & Yuan, X

    Endo, S., Zhao, Q., Li, Y ., Benjamin, S. & Yuan, X. Mitigating algorithmic errors in a hamiltonian simulation. Physical Review A99, 012334 (2019)

    2019

  5. [12]

    A., Tamascelli, D

    Fallani, A., Rossi, M. A., Tamascelli, D. & Genoni, M. G. Learning feedback control strategies for quantum metrology.PRX Quantum3, 020310 (2022)

    2022

  6. [13]

    Strikis, A., Qin, D., Chen, Y ., Benjamin, S. C. & Li, Y . Learning-based quantum error mitigation.PRX Quantum 2, 040330 (2021)

    2021

  7. [14]

    & Tokunaga, Y

    Suzuki, Y ., Endo, S., Fujii, K. & Tokunaga, Y . Quantum error mitigation as a universal error reduction technique: Applications from the nisq to the fault-tolerant quantum computing eras.PRX Quantum3, 010345 (2022)

    2022

  8. [15]

    Yang, X.et al.Probe optimization for quantum metrology via closed-loop learning control.npj Quantum Information6, 62 (2020)

    2020

  9. [16]

    Kobrin, B.et al.A universal protocol for quantum-enhanced sensing via information scrambling.arXiv preprint arXiv:2411.12794(2024). 17

    2024 arXiv

  10. [17]

    & Pezzè, L

    Imai, S., Smerzi, A. & Pezzè, L. Metrological usefulness of entanglement and nonlinear hamiltonians.Physical Review A111, L020402 (2025)

    2025

  11. [18]

    Peng, P.et al.Deep reinforcement learning for quantum hamiltonian engineering.Physical Review Applied18, 024033 (2022)

    2022

  12. [19]

    & Lindner, N

    Bairey, E., Arad, I. & Lindner, N. H. Learning a local hamiltonian from local measurements.Physical review letters122, 020504 (2019)

    2019

  13. [20]

    J., Harper, R

    Evans, T. J., Harper, R. & Flammia, S. T. Scalable bayesian hamiltonian learning.arXiv preprint arXiv:1912.07636 (2019)

    2019 arXiv

  14. [21]

    A., Dobrovitski, V

    Stilck França, D., Markovich, L. A., Dobrovitski, V . V ., Werner, A. H. & Borregaard, J. Efficient and robust estimation of many-qubit hamiltonians.Nature Communications15, 311 (2024)

    2024

  15. [22]

    Wilde, F.et al.Scalably learning quantum many-body hamiltonians from dynamical data.arXiv preprint arXiv:2209.14328(2022)

    2022 arXiv

  16. [23]

    & Coles, P

    Gu, A., Cincio, L. & Coles, P. J. Practical black box hamiltonian learning.arXiv preprint arXiv:2206.15464 (2022)

    2022 arXiv

  17. [24]

    & Acín, A

    Heightman, T., Jiang, E. & Acín, A. Solving the quantum many-body hamiltonian learning problem with neural differential equations.Quantum Science and Technology(2024)

    2024

  18. [25]

    Dutkiewicz, A., O’Brien, T. E. & Schuster, T. The advantage of quantum control in many-body hamiltonian learning (2023).2304.07172

    2023 arXiv

  19. [26]

    Hu, H.-Y .et al.Ansatz-free hamiltonian learning with heisenberg-limited scaling.arXiv preprint arXiv:2502.11900 (2025)

    2025

  20. [27]

    T., Preskill, J

    Ma, M., Flammia, S. T., Preskill, J. & Tong, Y . Learningk-body hamiltonians via compressed sensing.arXiv preprint arXiv:2410.18928(2024)

    2024 arXiv

  21. [28]

    Y ., Grossi, M

    Barthe, A., Rad, M. Y ., Grossi, M. & Dunjko, V . Quantum advantage in learning quantum dynamics via fourier coefficient extraction.arXiv preprint arXiv:2506.17089(2025)

    2025 arXiv

  22. [29]

    & Shi, B

    Leng, J. & Shi, B. Quantum optimization via gradient-based hamiltonian descent.arXiv preprint arXiv:2505.14670 (2025)

    2025 arXiv

  23. [30]

    & Zhang, Y

    Zhao, T.-L., Hu, S.-X. & Zhang, Y . Maximum-likelihood-estimate hamiltonian learning via efficient and robust quantum likelihood gradient.Physical Review Research5, 023136 (2023)

    2023

  24. [31]

    & Yuan, X

    Yu, W., Sun, J., Han, Z. & Yuan, X. Robust and efficient hamiltonian learning.Quantum7, 1045 (2023)

    2023

  25. [32]

    & Roushan, P

    Hangleiter, D., Roth, I., Fuksa, J., Eisert, J. & Roushan, P. Robustly learning the hamiltonian dynamics of a superconducting quantum processor.Nature Communications15, 9595 (2024)

    2024

  26. [33]

    & Lucignano, P

    Rattacaso, D., Passarelli, G. & Lucignano, P. High-accuracy hamiltonian learning via delocalized quantum state evolutions.Quantum7, 905 (2023)

    2023

  27. [34]

    & Duan, L.-M

    Wang, S.-T., Deng, D.-L. & Duan, L.-M. Hamiltonian tomography for quantum many-body systems with arbitrary couplings.New Journal of Physics17, 093017 (2015)

    2015

  28. [35]

    & Hsieh, T

    Li, Z., Zou, L. & Hsieh, T. H. Hamiltonian tomography via quantum quench.Physical review letters124, 160502 (2020)

    2020

  29. [36]

    E., Ferrie, C., Wiebe, N

    Granade, C. E., Ferrie, C., Wiebe, N. & Cory, D. G. Robust online hamiltonian learning.New Journal of Physics 14, 103013 (2012)

    2012

  30. [37]

    & Cory, D

    Hincks, I., Alexander, T., Kononenko, M., Soloway, B. & Cory, D. G. Hamiltonian learning with online bayesian experiment design in practice.arXiv preprint arXiv:1806.02427(2018)

    2018 arXiv

  31. [38]

    Dutt, A.et al.Active learning of quantum system hamiltonians yields query advantage.Physical Review Research 5, 033060 (2023)

    2023

  32. [39]

    Ferrie, C., Granade, C. E. & Cory, D. Adaptive hamiltonian estimation using bayesian experimental design. In AIP Conference Proceedings 31st, vol. 1443, 165–173 (American Institute of Physics, 2012)

    2012

  33. [40]

    & Zoller, P

    Olsacher, T., Kraft, T., Kokail, C., Kraus, B. & Zoller, P. Hamiltonian and liouvillian learning in weakly-dissipative quantum many-body systems.Quantum Science and Technology10, 015065 (2025)

    2025

  34. [41]

    Bairey, E., Guo, C., Poletti, D., Lindner, N. H. & Arad, I. Learning the dynamics of open quantum systems from their steady states.New Journal of Physics22, 032001 (2020)

    2020

  35. [42]

    Chen, L. & Wu, Y . Learning quantum dissipation by the neural ordinary differential equation.Physical Review A 106, 022201 (2022). 18

    2022

  36. [43]

    Choi, M., Flam-Shepherd, D., Kyaw, T. H. & Aspuru-Guzik, A. Learning quantum dynamics with latent neural ordinary differential equations.Physical Review A105, 042403 (2022)

    2022

  37. [44]

    Kang, Z.et al.Enhanced hamiltonian learning precision with multistage neural networks.Physical Review A111, 062418 (2025)

    2025

  38. [45]

    & Ying, L

    Li, H., Tong, Y ., Gefen, T., Ni, H. & Ying, L. Heisenberg-limited hamiltonian learning for interacting bosons.npj Quantum Information10, 83 (2024)

    2024

  39. [46]

    & Hayden, P

    Mirani, A. & Hayden, P. Learning interacting fermionic hamiltonians at the heisenberg limit.Physical Review A 110, 062421 (2024)

    2024

  40. [48]

    & Zoller, P

    Carrasco, J., Elben, A., Kokail, C., Kraus, B. & Zoller, P. Theoretical and experimental perspectives of quantum verification.PRX Quantum2, 010102 (2021)

    2021

  41. [49]

    & França, D

    Rouzé, C. & França, D. S. Learning quantum many-body systems from a few copies.Quantum8, 1319 (2024)

    2024

  42. [50]

    & Soleimanifar, M

    Anshu, A., Arunachalam, S., Kuwahara, T. & Soleimanifar, M. Sample-efficient learning of quantum many-body systems. In2020 IEEE 61st Annual Symposium on Foundations of Computer Science (FOCS), 685–691 (IEEE, 2020)

    2020

  43. [51]

    Hu, H.-Y .et al.Demonstration of robust and efficient quantum property learning with shallow shadows.Nature Communications16, 2943 (2025)

    2025

  44. [52]

    & Tang, E

    Haah, J., Kothari, R. & Tang, E. Optimal learning of quantum hamiltonians from high-temperature gibbs states. In2022 IEEE 63rd Annual Symposium on Foundations of Computer Science (FOCS), 135–146 (IEEE, 2022)

    2022

  45. [53]

    Valenti, A., Jin, G., Léonard, J., Huber, S. D. & Greplova, E. Scalable hamiltonian learning for large-scale out-of-equilibrium quantum dynamics.Physical Review A105, 023302 (2022)

    2022

  46. [54]

    & Cory, D

    Wiebe, N., Granade, C., Ferrie, C. & Cory, D. Quantum hamiltonian learning using imperfect quantum resources. Physical Review A89, 042314 (2014)

    2014

  47. [55]

    & Maccone, L

    Giovannetti, V ., Lloyd, S. & Maccone, L. Quantum metrology.Physical Review Letters96(2006). URL http://dx.doi.org/10.1103/PhysRevLett.96.010401

    2006 doi

  48. [56]

    Rao, C. R. Information and the accuracy attainable in the estimation of statistical parameters. InBreakthroughs in Statistics: Foundations and basic theory, 235–247 (Springer, 1992)

    1992

  49. [57]

    Nielsen, M. A. & Chuang, I. L.Quantum computation and quantum information(Cambridge university press, 2010)

    2010

  50. [58]

    Huang, H.-Y ., Tong, Y ., Fang, D. & Su, Y . Learning many-body hamiltonians with heisenberg-limited scaling. Physical Review Letters130, 200403 (2023). URLhttps://doi.org/10.1103/PhysRevLett.130.200403

    2023 doi

  51. [59]

    & Ying, L

    Ni, H., Li, H. & Ying, L. Quantum hamiltonian learning for the fermi-hubbard model.Acta Applicandae Mathematicae191, 2 (2024)

    2024

  52. [60]

    & Fedorov, A

    Luchnikov, I., Gavreev, M. & Fedorov, A. Controlling quantum many-body systems using reduced-order modeling. Physical Review Research6, 013161 (2024)

    2024

  53. [61]

    & Lloyd, S

    Viola, L. & Lloyd, S. Dynamical suppression of decoherence in two-state quantum systems.Physical Review A 58, 2733 (1998)

    1998

  54. [62]

    & Maccone, L

    Giovannetti, V ., Lloyd, S. & Maccone, L. Quantum-enhanced measurements: Beating the standard quantum limit. Science306, 1330–1336 (2004). URL https://www.science.org/doi/abs/10.1126/science.1104149. https://www.science.org/doi/pdf/10.1126/science.1104149

    2004 doi

  55. [63]

    Decoherence, the measurement problem, and interpretations of quantum mechanics.Reviews of Modern physics76, 1267–1305 (2004)

    Schlosshauer, M. Decoherence, the measurement problem, and interpretations of quantum mechanics.Reviews of Modern physics76, 1267–1305 (2004)

    2004

  56. [64]

    & Ghinea, C

    Petz, D. & Ghinea, C. Introduction to quantum fisher information. InQuantum probability and related topics, 261–281 (World Scientific, 2011)

    2011

  57. [65]

    Watrous, J.The theory of quantum information(Cambridge university press, 2018)

    2018

  58. [66]

    Fisher, R. A. Theory of statistical estimation.Mathematical Proceedings of the Cambridge Philosophical Society 22, 700–725 (1925)

    1925

  59. [67]

    & Wang, X

    Liu, J., Yuan, H., Lu, X.-M. & Wang, X. Quantum fisher information matrix and multiparameter estimation. Journal of Physics A: Mathematical and Theoretical53, 023001 (2020). 19

    2020

  60. [68]

    Takahashi, M.et al.Thermodynamics of one-dimensional solvable models (1999)

    1999

  61. [69]

    Github repository for heisenberg-limited quantum hamiltonian learn- ing via randomly spread product-states

    Baran, B. Github repository for heisenberg-limited quantum hamiltonian learn- ing via randomly spread product-states. https://github.com/BoraBaranBora/ heisenberg-limited-quantum-hamiltonian-learning-via-randomly-spread-product-states

  62. [70]

    Learning the structure of any hamiltonian from minimal assumptions.arXiv preprint arXiv:2410.21635 (2024)

    Zhao, A. Learning the structure of any hamiltonian from minimal assumptions.arXiv preprint arXiv:2410.21635 (2024)

    2024 arXiv

  63. [71]

    & Zoller, P

    Pastori, L., Olsacher, T., Kokail, C. & Zoller, P. Characterization and verification of trotterized digital quantum simulation via hamiltonian and liouvillian learning.PRX Quantum3, 030324 (2022)

    2022

  64. [72]

    & Lai, Y .-C

    Han, C.-D., Glaz, B., Haile, M. & Lai, Y .-C. Tomography of time-dependent quantum hamiltonians with machine learning.Physical Review A104, 062404 (2021)

    2021

  65. [73]

    & Yuan, H

    Liu, J. & Yuan, H. Quantum parameter estimation with optimal control.Physical Review A96, 012117 (2017)

    2017

  66. [74]

    Howell, R. R. On asymptotic notation with multiple variables.Dept. Comput. Inf. Sci., Kansas State Univ., Manhattan, KS, USA, Tech. Rep(2008). Acknowledgments We would like to thank José Ramon-Martinez, Marcin Płodzie´n, Carlos Pascual, and Antonio Acìn for their helpful comme...

    2008

  67. [75]

    Reference state: Prepare a fixed reference state, which we chose to be |ψ0⟩=|0⟩ ⊗N without loss of generality |ψ0⟩withρ 0 =|ψ 0⟩⟨ψ0|

  68. [76]

    Initial State Spreading: Apply locally Haar-random rotations Uspread = nO i=1 Rz(ξi)R y(χi)R z(ϕi),(90) whereχ i = arccos(1−2u i),u i ∼U[0,1], andϕ i, ξi ∼U[0,2π], yieldingρ spread =U spread ρ0 U † spread

  69. [77]

    , mt,(91) where∆t >0andα >−1

    Time evolution: Evolveρ spread with the true HamiltonianH true for a sequence of times tk = ∆t kα, k= 1,2, . . . , mt,(91) where∆t >0andα >−1. At eacht k, ρ(tk) =e −iHtruetk ρspread eiHtruetk .(92)

  70. [78]

    Measurement: At each tk, measure in a random product Pauli basis pj ∈ {X, Y, Z}n and record the bit-string outcomess. We use the empirical bit-string distribution Pdata(b) obtained under a chosen configuration, comprising the initial-state ensemble (via state spreading), evolu...

  71. [79]

    , Sshots

    Simulating the experiment under the current estimate ˆH(θ): for each initial state in {|ψr⟩}R r=1, evolve for timest k and measure in the corresponding product Pauli basisp j, repeatings= 1, . . . , Sshots

  72. [80]

    extended parameter embedding

    Collecting simulated bit-string outcomesb rkjd to form the model distributionP ˆH(θ) (b). Based on the dataset D of size |D|=R K mt S, with entries indexed by (r, j, k, s)and outcomes brjks , we define the negative log-likelihood loss: LD(θ) =− 1 RJ mtS RX r=1 JX k=1 mtX k=1 S...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.