REVIEW 1 major objections 5 minor 35 references
Structure-Agnostic Unitary Learning from Quantum Observable Dynamics with Application to Hamiltonian Identification
T0 review · 1 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash
Pith's one-line read This paper claims that an unknown quantum unitary can be learned from time-series observable measurements with no assumption about its generator, and that any Hamiltonian behind it is then recovered classically by a matrix logarithm and a P
desk verdict Promising two-phase variational scheme for small-unitrary learning, but the identifiability proof is broken as stated and the non-Hamiltonian claim is overstated; deserves a proper review after fixes. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the multi-time observable-matching loss, which compares circuit-computed expectation values with measured ones at several times nτ and drives U(θ)^n to match e^{-iHnτ}. The companion identity is the matrix-log map Ĥ = (i/τ) log U(θ*), restricted to the principal branch, followed by the Pauli projection cα = Tr(ĤPα)/2^n. This turns Hamiltonian identification into a one-line classical computation once U is learned. The SPSA-Adam optimiser, which estimates gradients with two circuit evaluations per step and carries Adam moments across curriculum stages, is the mechanism that makes the non-convex search tractable in practice.
What would settle it
Take a two-qubit target unitary U_target whose distance from the reachable set of the fixed L=3 RY/RZ/CNOT ansatz is nonzero — for example, compute inf_W ||U_target − W||_F over the 28 parameters and check whether the infimum is above ~1e-6. Run the same curriculum on it and see whether the loss reaches the shot-noise floor and whether the matrix-log coefficients stay within the reported error bounds; if it fails, 'structure-agnostic' holds only for targets the fixed ansatz already contains.
Extended reading notes
Core claim
On its own terms, the paper's discovery is a decoupling: a hardware-efficient parameterised circuit matches observable expectation values at multiple time steps to learn U directly, while Hamiltonian identification is deferred to optional classical post-processing. The paper states an identifiability condition — initial states spanning the operator algebra and observables spanning the Hermitian matrices fix U up to global phase — and uses Pauli orthogonality to assert that the coefficients cα = Tr(ĤPα)/2^n are unique once Ĥ is known. It then demonstrates three regimes: exact noiseless recovery, gate learning of targets with no generating Hamiltonian, and noisy recovery of all three Ising ter
Load-bearing premise
The load-bearing premise is that the fixed 28-parameter (two-qubit) or 102-parameter (three-qubit) RY/RZ/CNOT ansatz can represent the target unitary closely enough for observable matching to reach the noise floor; no universality or expressibility proof is given.
Editorial extensions
If this is right
- Because U itself is the training target, the same algorithm applies to gate learning without modification; CNOT, iSWAP, and a Haar-random SU(4) gate are all fit to process fidelity 1.000000.
- Hamiltonian coefficients are recovered without any prior assumption about which Pauli terms are present; the Pauli expansion after the matrix logarithm supplies the structure automatically.
- The principal-branch condition ∥H∥τ < π sets a hard bound on the timestep: for Hamiltonians with large spectral norm, τ must be reduced, which in turn raises shot-noise sensitivity.
- The curriculum is claimed to be necessary: training on all timesteps simultaneously stalls and recovers every coefficient with the wrong sign, while the staged schedule converges near the shot-noise floor and recovers all target terms.
- For more than three or four qubits the exponential Pauli basis is a bottleneck, and the paper points to compressed sensing or structure-aware priors as the needed extension.
Reading between the lines
- Left implicit is that the identifiability condition could be read as a measurement-design rule: minimal informationally complete sets should in principle suffice, so the method's data overhead could be cut well below the 4 states × 9–63 observables used in the demonstrations.
- The same two-phase separation suggests a diagnostic use the paper does not develop: because the matrix-log step only makes sense for Hamiltonian-generated targets, the residual of the logarithm and the Hermiticity error could serve as a statistical test for whether observed dynamics are generated by a time-independent Hamiltonian at all.
- If the fixed ansatz family were proven universal, the method would amount to model-free process tomography from time-series data; the absence of such a proof is currently the main gap between 'demonstrated on small systems' and 'structure-agnostic in general'.
- The curriculum's staged time horizons resemble a continuation method on the loss landscape; an explicit study of how stage boundaries trade convergence against total circuit depth could make the schedule transferable to larger systems.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a variational two-phase method for learning an unknown unitary from time-series observable measurements. In the first phase, a hardware-efficient parametrized circuit U(θ) is trained to match the measured expectation values ⟨P⟩(t) at multiple times; in the second phase, a Hamiltonian is recovered classically by computing the principal matrix logarithm Θ = (i/τ) log U(θ) and projecting onto Pauli operators. The authors claim that this separates unitary learning from Hamiltonian identification, requires no structural assumption about the target Hamiltonian, and can also learn non-Hamiltonian gates. Supporting experiments include a noiseless 2-qubit proof of correctness, a 3-qubit transverse-field Ising model, gate learning for CNOT, iSWAP, and a Haar-random SU(4) element, and a noisy Qiskit Aer simulation with SPSA-Adam and a four-stage curriculum. The paper reports high-fidelity fits and recovery of all Pauli coefficients in the tested cases.
Significance. If the claims were established, the proposed separation—learning a unitary from dynamics data and then extracting a Hamiltonian as classical post-processing—would be a conceptually useful contribution to variational Hamiltonian learning. The paper contains several genuinely positive elements: the explicit use of multiple timesteps with a single parameterized unitary, out-of-sample checks on unseen timesteps and states, a curriculum schedule that is shown to be necessary in the comparison of Section 8, and a low per-iteration circuit evaluation count for SPSA-Adam. The noiseless 2-qubit results and the 3-qubit TFIM recovery are internally plausible and provide useful evidence for the method. However, the theoretical identifiability guarantee is stated incorrectly, the claim that CNOT/iSWAP are not Hamiltonian-generated is false, and the paper overstates its structure-agnosticism. These issues are load-bearing for the paper's central claims and must be corrected.
major comments (1)
- [§2.2, Proposition 1] The statement 'For n=2 qubits, |S|=4 informationally complete states and |P|=9 Pauli observables suffice' is mathematically false. Four rank-one projectors in B(C^4) have linear span of dimension at most 4, not 16; nine Pauli operators span at most a 9-dimensional subspace of the 16-dimensional Hermitian space. Thus the hypothesis of Proposition 1 is not satisfied by the paper's own experimental configurations in §4 (4 states, 15 Paulis) or §7 (4 states, 9 Paulis). The proof's step 'fixes the operator U†PαU' from expectation values on S fails: with only four projectors one obtains diagonal elements in one basis, leaving off-diagonal information undetermined. Since §3.2 states that the loss is zero iff U=e^{iφ}U_true, this missing identifiability result is load-bearing. If multi-timestep data are intended to compensate for the insufficiency of the n=1 data, a rigorous argument is needed.
minor comments (5)
- [§2.2] The symbol n is used both for the number of qubits and for the discrete time step in Eq. (3). Please disambiguate.
- [§6.2.2] The ansatz in §3.1 contains RY gates, so the sentence 'the circuit uses only X and Z rotations internally' is inconsistent with the ansatz description. This affects the interpretation of the 'wrong operator family' experiment.
- [Table 8] The baseline row and the YY+XX row report exactly the same MSE (3.22×10−14) and same ||Herr|| (7.96×10−4). This suggests a copy-paste error or a duplicate run. Please verify and report distinct results.
- [§2.4] The branch condition is stated in terms of eigenvalues |λk|τ<π, but the experimental text says '||H_true||_F τ ≈ 0.8 < π'. Since the Frobenius norm is not the spectral norm, please clarify that this inequality is only a sufficient condition for the eigenvalue condition, and justify it.
- [§2.5] The statement that Adam with β1=0.9 'boosts effective gradient SNR by sqrt(10)' is heuristic; SPSA gradient estimates are not independent across updates, and bias correction changes the effective averaging. Please soften or justify this quantitative claim.
Circularity Check
No significant circularity: matrix-log Hamiltonian extraction is explicit post-processing, and the only self-citation is non-load-bearing.
full rationale
Walking the derivation chain: the algorithm fits a single parameterized unitary U(θ) to observable expectation values over a staged set of timesteps (loss Eq. 7), then forms Ĥ = (i/τ) log U(θ*) and projects onto Paulis via Eq. 2. The Hamiltonian coefficients are therefore deterministic functions of the learned unitary, not independently fitted parameters; the paper explicitly frames this as 'free classical post-processing' (Section 1) and 'no additional quantum measurements are required' (Section 2.4). This is a transparent mathematical identity (when the branch condition holds), not a disguised prediction. The multi-step coupling of the same U across n=1,...,20 provides genuine out-of-sample checks on unseen n and states (Sections 4, 7.5). The only self-citation [35] is a provenance statement ('extends and generalises ... Master's dissertation') and no theorem or numerical result is imported from it. The identifiability proposition's dimension claim (4 states + 9 Paulis for 2 qubits) appears mathematically wrong, and the ansatz expressibility is unproved; these are correctness/falsifiability risks, not circularity. Hence no significant circularity.
Assumptions & free parameters
free parameters (5)
- Ansatz repetitions L =
3 (Ntheta=28 for 2 qubits; 102 for 3 qubits)
- Time step tau =
0.05 (noiseless), 0.5 (quantum)
- SPSA-Adam hyperparameters =
alpha=0.01, c0=0.05, gamma=0.101, beta1=0.9, beta2=0.999, eps=1e-8
- Curriculum active-step schedule =
{1,2}, {1,2,4,6}, {1,2,4,6,8,10}, {1,...,20}
- Pauli coefficient threshold =
delta=0.05
assumptions (5)
- standard math Pauli operators form an orthogonal basis of Herm(2^n).
- standard math Conjugation by a unitary is determined up to global phase by its action on a spanning set of operators.
- ad hoc to paper The fixed hardware-efficient ansatz can represent the target unitaries to sufficient accuracy.
- domain assumption The principal matrix logarithm branch condition ||H|| tau < pi holds for the true generator.
- domain assumption The depolarizing noise model (p1=0.001, p2=0.01) is representative of near-term hardware.
Cite this review
Pith. "Pith review of Structure-Agnostic Unitary Learning from Quantum Observable Dynamics with Application to Hamiltonian Identification." pith.science (2026). https://pith.science/paper/WQGGIAGL
@misc{pith2026260715316,
author = {Pith},
title = {Pith review of: Structure-Agnostic Unitary Learning from Quantum Observable Dynamics with Application to Hamiltonian Identification},
year = {2026},
howpublished = {\url{https://pith.science/paper/WQGGIAGL}},
note = {Machine review of arXiv:2607.15316}
}
read the original abstract
We present a variational algorithm for learning an unknown quantum unitary from time-series observable measurements, with no structural assumption about the target. The core separation: a hardware-efficient parametrised circuit learns the evolution operator U via observable matching; Hamiltonian identification follows as classical post-processing via matrix logarithm, when the target happens to be exp(-iH*tau). Three experiments establish the method's generality. First, a noiseless proof of correctness with exact gradients (L-BFGS-B) achieves MSE 1.61e-14 and recovers all Hamiltonian coefficients to six decimal places. Second, a gate-learning experiment fits CNOT, iSWAP, and a Haar-random SU(4) element -- none generated by any fixed Hamiltonian -- all to process fidelity 1.000000, confirming the method does not rely on Trotterisation structure. Third, quantum deployment via SPSA-Adam under Qiskit Aer depolarising noise (p1=0.001, p2=0.01, Nshots=1024) recovers all three Ising Hamiltonian terms with errors below 8%. The optimiser, SPSA-Adam, combines SPSA's hardware-efficient two-point gradient estimation with Adam's adaptive moment updates. A four-stage moment-warm curriculum progressively extends the training horizon, converting a global non-convex problem into a sequence of well-posed local ones.
Reference graph
Works this paper leans on
-
[1]
Robust and effi- cient Hamiltonian learning,
W. Yu, J. Sun, Z. Han, and X. Yuan, “Robust and effi- cient Hamiltonian learning,”Quantum, vol. 7, p. 1045, 2023
2023
-
[2]
Learning many-body Hamiltonians with Heisenberg-limited scaling,
H. Huang, Y. Tong, D. Fang, and Y. Su, “Learning many-body Hamiltonians with Heisenberg-limited scaling,”Phys. Rev. Lett., vol. 130, no. 20, p. 200403, 2022
2022
-
[3]
Practical Hamil- tonian learning with unitary dynamics and Gibbs states,
A. Gu, L. Cincio, and P. J. Coles, “Practical Hamil- tonian learning with unitary dynamics and Gibbs states,”Nat. Commun., vol. 15, p. 44008, 2024
2024
-
[4]
Solving the quantum many-body Hamiltonian learning problem with neural differential equations,
T. Heightman, E. Jiang, and A. Acín, “Solving the quantum many-body Hamiltonian learning problem with neural differential equations,”Quantum Sci. Technol., vol. 10, p. 014001, 2024
2024
-
[5]
Hamiltonian learning via quantum Zeno effect,
G. Franceschetto et al., “Hamiltonian learning via quantum Zeno effect,” 2025
2025
-
[6]
Hamiltonian and Liouvillian learn- ing in weakly-dissipative quantum many-body sys- tems,
T. Olsacher et al., “Hamiltonian and Liouvillian learn- ing in weakly-dissipative quantum many-body sys- tems,”Quantum Sci. Technol., vol. 10, p. 024002, 2024
2024
-
[7]
Scalably learning quantum many- body Hamiltonians from dynamical data,
F. Wilde et al., “Scalably learning quantum many- body Hamiltonians from dynamical data,”arXiv, 2209.14328, 2022
arXiv 2022
-
[8]
Optimal and robust in-situ quantum Hamiltonian learning through paral- lelization,
S. Liu, X. Wu, and M. Niu, “Optimal and robust in-situ quantum Hamiltonian learning through paral- lelization,” 2025
2025
Show all 35 references
-
[9]
Sample-efficient learning of inter- actingquantumsystems,
A. Anshu, S. Arunachalam, T. Kuwahara, and M. Soleimanifar, “Sample-efficient learning of inter- actingquantumsystems,”Nat. Phys., vol.17, pp.931– 935, 2021
2021
-
[10]
Hamiltonian learning using machine-learning models trained with continu- ous measurements,
K. Tucker et al., “Hamiltonian learning using machine-learning models trained with continu- ous measurements,”Phys. Rev. Applied, vol. 22, p. 044080, 2024
2024
-
[11]
Unified quantum state tomography and Hamiltonian learning,
Z. An et al., “Unified quantum state tomography and Hamiltonian learning,”Phys. Rev. Applied, vol. 21, p. 014037, 2023
2023
-
[12]
A hybrid quantum- classical Hamiltonian learning algorithm,
Y. Wang, G. Li, and X. Wang, “A hybrid quantum- classical Hamiltonian learning algorithm,”Sci. China Inf. Sci., vol. 66, 2023
2023
-
[13]
Parameterized Hamiltonian learning with quantum circuit,
J. Shi et al., “Parameterized Hamiltonian learning with quantum circuit,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, 2023. 6
2023
-
[14]
Variational quantum algorithms,
M. Cerezo et al., “Variational quantum algorithms,” Nat. Rev. Phys., vol. 3, pp. 625–644, 2021
2021
-
[15]
Hamiltonian learning from time dy- namics using variational algorithms,
R. Gupta et al., “Hamiltonian learning from time dy- namics using variational algorithms,”J. Phys. Chem. A, 2022
2022
-
[16]
Multivariate stochastic approximation using a simultaneous perturbation gradient approx- imation,
J. C. Spall, “Multivariate stochastic approximation using a simultaneous perturbation gradient approx- imation,”IEEE Trans. Autom. Control, 37(3):332– 341, 1992
1992
-
[17]
Adam: A method for stochastic optimization,
D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”ICLR, 2015
2015
-
[18]
Learning a local Hamiltonian from local measurements,
E. Bairey, I. Arad, and N. H. Lindner, “Learning a local Hamiltonian from local measurements,”Phys. Rev. Lett.122, 020504, 2019
2019
-
[19]
Optimal learn- ing of quantum Hamiltonians from high-temperature Gibbs states,
J. Haah, R. Kothari, and E. Tang, “Optimal learn- ing of quantum Hamiltonians from high-temperature Gibbs states,”FOCS, 2022
2022
-
[20]
Hamiltonian learning and certification using quan- tum resources,
N. Wiebe, C. Granade, C. Ferrie, and D. Cory, “Hamiltonian learning and certification using quan- tum resources,”Phys. Rev. Lett.112, 190501, 2014
2014
-
[21]
Self-verifying variational quantum simulation of lattice models,
C. Kokail et al., “Self-verifying variational quantum simulation of lattice models,”Nature569, 355–360, 2019
2019
-
[22]
Predicting many properties of a quantum system from very few measurements,
H.-Y. Huang, R. Kueng, and J. Preskill, “Predicting many properties of a quantum system from very few measurements,”Nature Physics16, 1050–1057, 2020
2020
-
[23]
Hardware-efficient variational quantum eigensolver,
A. Kandala et al., “Hardware-efficient variational quantum eigensolver,”Nature549, 242–246, 2017
2017
-
[24]
Exploiting symmetry reduces the cost of training QAOA,
R. Shaydulin and S. M. Wild, “Exploiting symmetry reduces the cost of training QAOA,”IEEE Trans. Quantum Eng.2, 1–9, 2021
2021
-
[25]
Sachdev,Quantum Phase Transitions, 2nd ed., Cambridge University Press, 2011
S. Sachdev,Quantum Phase Transitions, 2nd ed., Cambridge University Press, 2011
2011
-
[26]
N. J. Higham,Functions of Matrices: Theory and Computation, Other Titles in Applied Mathematics, Vol. 104, Society for Industrial and Applied Mathe- matics (SIAM), Philadelphia, PA, 2008
2008
-
[27]
A thermodynamic theory of weak ferromagnetism of antiferromagnetics,
I. Dzyaloshinskii, “A thermodynamic theory of weak ferromagnetism of antiferromagnetics,”J. Phys. Chem. Solids4, 241–255, 1958
1958
-
[28]
Nocedal and S
J. Nocedal and S. J. Wright,Numerical Optimization, 2nd ed., Springer, 2006
2006
-
[29]
Curriculum Learning,
Y. Bengio, J. Louradour, R. Collobert, and J. We- ston, “Curriculum Learning,” inProceedings of the 26th Annual International Conference on Machine Learning (ICML), pp. 41–48, 2009
2009
-
[30]
Barren Plateaus in Quantum Neural Network Training Landscapes,
J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Bab- bush, and H. Neven, “Barren Plateaus in Quantum Neural Network Training Landscapes,”Nature Com- munications, vol. 9, p. 4812, 2018
2018
-
[31]
Barren Plateaus Are Swamped with Traps,
N. A. Nemkov, E. O. Kiktenko, and A. K. Fedorov, “Barren Plateaus Are Swamped with Traps,”Physical Review A, vol. 108, no. 3, p. 032406, 2023
2023
-
[32]
Logic Gates at the Surface Code Threshold: Superconducting Qubits Poised for Fault-Tolerant Quantum Computing,
R. Barends, J. Kelly, A. Megrant, A. Veitia, D. Sank, E. Jeffrey, T. C. White, J. Mutus, A. G. Fowler, B. Campbell, Y. Chen, Z. Chen, B. Chiaro, A. Dunsworth, C. Neill, P. O’Malley, P. Roushan, A. Vainsencher, J. Wenner, A. N. Korotkov, A. N. Cle- land, and J. M. Martinis, “Lo...
2014
-
[33]
Quantum Computing in the NISQ Era and Beyond,
J. Preskill, “Quantum Computing in the NISQ Era and Beyond,”Quantum, vol. 2, p. 79, 2018
2018
-
[34]
Qiskit: An open-source frame- work for quantum computing,
Qiskit contributors, “Qiskit: An open-source frame- work for quantum computing,” 2023
2023
-
[35]
Berkani,Variational Learning of Unknown Two-Qubit Hamiltonians from Quantum Observ- able Dynamics, Master’s dissertation, Univer- sité Ferhat Abbas Sétif 1, 2026
M. Berkani,Variational Learning of Unknown Two-Qubit Hamiltonians from Quantum Observ- able Dynamics, Master’s dissertation, Univer- sité Ferhat Abbas Sétif 1, 2026. Available(DOI): 10.13140/RG.2.2.29146.09924 A. Pauli Conventions Pauli stringsAB denote A⊗B with A on qubit 1 a...
2026
Reviewed August 2, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.