Pith. sign in

REVIEW 3 major objections 7 minor 30 references

Machine Learning for Ground State Preparation via Measurement and Feedback

T0 review · 3 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read An RNN learns to prepare ground states by first funneling every initial state to a fixed intermediate state, then rotating to the target, with measurement feedback as the learned control signal.

desk verdict An interesting Hamiltonian-conditioned feedback controller for ground-state prep, but the loss in Eq. (9) is not well-defined as written, making the central claim conditional until fixed. read the letter →

arxiv 2502.06517 v2 pith:UQIT57QX submitted 2025-02-10 quant-ph physics.comp-ph

Chuanxin Wang , Yi-Zhuang You This is my paper · ORCID
classification quant-phphysics.comp-ph
keywords groundstatepreparationrecurrentneuralnetworkmeasurementandfeedbackmid-circuitancillaqubitsLSTMvariationalquantumeigensolver
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

The paper tries to establish that a recurrent neural network, trained once on randomly sampled Hamiltonians, can generate feedback-based quantum circuits that prepare the ground states of those Hamiltonians from arbitrary pure or mixed initial states. Its central empirical finding is that the trained controller does not approach the target gradually: it first drives all initial states to a fixed intermediate state, then uses the Hamiltonian information to rotate that state to the ground state. The paper also argues that measurement and feedback are genuinely useful, because preparation fidelity improves as more ancilla qubits are measured and fed back, yet at least one ancilla must be traced out to keep entropy under control. If correct, this points toward reusable, measurement-aware quantum state preparation rather than per-instance optimization.

What carries the argument

The central object is a quantum-classical recurrent loop: an LSTM (long short-term memory) cell whose hidden memory accumulates the measurement record and the Hamiltonian, producing gate parameters for a parametrized quantum circuit at each time step. The circuit applies $U_{\theta_t}$, measures some ancilla qubits, and traces out the remaining ancilla qubits, so the next system state is conditioned on the measurement outcome $m_t$. The training signal is the energy expectation value of the final state with respect to the target Hamiltonian. The mechanism that carries the argument is the emergent two-stage strategy — reset all inputs to an intermediate state, then rotate to the target — together with the role of the traced-out ancilla as an entropy sink that keeps the state-preparation trajectory flexible.

What would settle it

For a fixed test Hamiltonian and many fresh measurement trajectories, evaluate the trained controller and compute the average energy of the final states; if that average lies above the first excited-state energy while individual trajectories look good, the claim that the RNN prepares the ground state in expectation fails. Also check per-step fidelities across many initial states: the claimed two-stage strategy predicts convergence to a common intermediate state before the final rotation, and the absence of that convergence would refute the central observation.

Watch

Extended reading notes

Core claim

The paper claims that a single LSTM-based recurrent controller can learn distinct preparation protocols for different Hamiltonians by dynamically adjusting quantum gates based on mid-circuit measurement outcomes. The controller outputs gate parameters $\theta_t = F_\varphi(m_1,\dots,m_{t-1},H)$ at each step, and the quantum process $C_{\theta_t,m_t}$ maps the system state forward while ancilla qubits are either measured or traced out. Training minimizes the final energy $E(\varphi) = \mathrm{Tr}(\rho_T^{\mathrm{sys}}(\varphi) H)$ over the RNN parameters $\varphi$. Across single-qubit and two-qubit experiments, the optimized strategy separates into two stages: an initial reset phase that concentrates all initial states onto a fixed product state such as $|0\rangle$ or $|00\rangle$, and a short final phase that rotates this intermediate state to the target ground state. The paper further claims that fidelity improves when more ancilla qubits are used for measurement and feedback, but degrades when every ancilla is measured, because projective measurements alone cannot raise entropy and the traced-out ancilla supplies the needed dissipation.

Load-bearing premise

The load-bearing assumption is that the training objective is a well-defined expected energy: the final state depends on random measurement outcomes, yet the paper writes the loss as a single deterministic number and does not state whether it averages over measurement records, so if it does not, minimizing that loss need not prepare the ground state in expectation.

Editorial extensions

If this is right

  • A single trained controller can prepare ground states for a distribution of Hamiltonians without per-instance optimization, acting as a reusable preparation model rather than a solver for one problem.
  • Increasing the number of ancilla qubits devoted to measurement and feedback raises preparation fidelity, so mid-circuit measurement is a resource rather than an overhead.
  • At least one unmeasured, traced-out ancilla is necessary: using all ancillas for projective measurement degrades fidelity, indicating that entropy management is part of the preparation strategy.
  • The two-stage reset-then-rotate structure is a candidate design principle for measurement-based ground-state preparation, at least in the small systems studied.
  • The method is currently limited to noiseless, small-system simulation; the paper itself identifies locality constraints and noise-aware training as the conditions for scaling.

Reading between the lines

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

  • If the two-stage strategy is universal across architectures, it suggests a canonical solution to feedback-based preparation: erase the initial state first, then apply Hamiltonian information. This could be tested by training with different gate sets and circuit depths to see whether the intermediate state changes or disappears.
  • The traced-out-ancilla requirement suggests the essential resource is controllable entropy flow, not extra Hilbert space. Replacing the traced-out ancilla with an explicit reset or dissipation channel and comparing fidelities would isolate whether entropy removal alone explains the benefit.
  • The formally undefined training objective over measurement records could be resolved by defining the loss as an expectation over measurement trajectories; doing so would make the comparison between protocols well-posed and might change which intermediate state is optimal.
  • A sharper generalization test would train on random local Hamiltonians and evaluate on structured or long-range Hamiltonians; the paper's framing predicts partially degraded but nontrivial transfer, whereas a memorizing controller would fail completely.
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

3 major / 7 minor

Summary. The paper proposes an RNN-based controller for ground-state preparation in quantum circuits with mid-circuit measurements and feedback. The RNN receives the Hamiltonian and past measurement outcomes, outputs gate parameters at each step, and is trained by minimizing the energy of the final system state. Numerical simulations on one- and two-qubit systems are used to argue that the RNN learns a two-stage strategy: first steering all initial states to a fixed intermediate product state and then rotating it to the target ground state. An ancilla-scaling study is presented to support the claim that measurement-and-feedback ancillas improve fidelity.

Significance. If the training objective is made well-defined and the results are reproduced with proper statistics, the work would be a useful demonstration that recurrent controllers can learn feedback protocols for state preparation and that such protocols can be discovered by energy minimization rather than by hand. The energy loss is the standard variational objective, so the approach is not circular in principle; the obstacle is that Eq. (9) does not define how the random measurement record is averaged. The authors themselves note that the method is not scalable for large systems and does not incorporate quantum noise, which limits the strength of the "foundational model" wording in the Summary. The paper provides no code, data, or reproducibility details, which matters for an empirical machine-learning paper. The central two-stage claim and the ancilla-scaling claim are plausible but currently rest on averages, best-of-three selection, and visual inspection rather than quantitative evidence.

major comments (3)
  1. [Loss Function, Eqs. (5) and (9)] The training objective is not well defined. Eq. (5) defines ρ_t^sys as the trace over ancillas of P_{m_t} U ρ U† P_{m_t}, which is a subnormalized conditional state unless divided by the outcome probability; Eq. (9) then evaluates Tr(ρ_T^sys H) without stating whether ρ_T^sys is a single-trajectory state, an expectation over measurement records, or a normalized state. Because θ_t in Eq. (8) depends on m_1,...,m_{t-1}, minimizing the energy of one sampled trajectory with the subnormalized state does not minimize the expected energy of the normalized state prepared for the user; the protocol that minimizes Eq. (9) need not prepare the ground state in expectation. The authors should replace Eq. (9) with an explicit expectation over the measurement record and over the sampled initial states, normalize the post-measurement states, and explain how gradients are obtained through the discrete measurement outcomes.
  2. [Results, Figs. 4 and 5] The central claim that the RNN consistently steers all initial states to an intermediate state before transitioning to the target is not quantitatively established. Fig. 5 shows only the average fidelity and average ⟨σ1_z σ2_z⟩; the statement that the RNN "first prepares all initial states to |00⟩" is an interpretation of these averages, and the single-qubit Fig. 4 is a visualization of 100 samples rather than a measured distribution. The fact that the two-stage strategy is an emergent, post-hoc description is not a flaw, but the claim "all initial states" requires distributional evidence, for example the per-sample fidelity to |00⟩ or histograms of Tr(|00⟩⟨00|ρ_t) at each time step. Similarly, the assertion that Hamiltonian information is "only effectively used in the last two steps" needs an explicit probe, such as comparing the early-time policies generated for different Hamiltonians.
  3. [Table I and 'Explore the Role of Ancilla Qubits'] The ancilla-scaling conclusions are not supported by the reported statistics. Each entry in Table I is the best of three training runs, with no error bars or spread, so the nonmonotonic entries may be seed fluctuations. The observed pattern is not monotonic in N_anc^m: N_anc=2,N_m=2 gives 0.853 compared with 0.908 for N_m=1, and N_anc=4,N_m=4 gives 0.888 compared with 0.940 for N_m=2, contradicting the second bullet. Moreover, there is no N_m=0 baseline (all ancillas traced out), so the improvements attributed to measurement and feedback are not separated from the increase in total ancilla resources. The authors should report mean and standard deviation over multiple seeds and include a no-measurement baseline before claiming systematic improvement.
minor comments (7)
  1. [Introduction] The section cross-references are empty ("In Sec. ," appears twice); the section numbers should be filled in.
  2. [Throughout] There are several typos, including "Loss F unction" in the heading, "experiement", and "suggestings"; the manuscript should be proofread.
  3. [Method, Initial State Setting] The notation N is overloaded: N is defined as N_sys+N_anc, but Table I uses N=2,3,4 to denote N_anc; this should be clarified.
  4. [Single-qubit State Preparation] The text says the initial states are mixed states, but Fig. 4 shows only pure-state labels |0⟩, |1⟩, |+⟩, |i⟩; clarify whether the plotted samples are pure or mixed.
  5. [Results] The time-step indexing is inconsistent with the stated number of steps: the single-qubit case says n=5 but discusses T=0 through T=5, and the two-qubit case says n=8 but discusses T=6 through T=8; define whether T labels the number of applied unitaries or the endpoint after each step.
  6. [Related Work] The manuscript cites Refs. [19,20] on learning feedback mechanisms for measurement-based state preparation but does not explain how the present RNN loss, architecture, or results differ from those works; a comparison would clarify the novelty.
  7. [Summary/Reproducibility] No code, random seed, or hyperparameter details are provided; for an empirical machine-learning paper these details are important for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the energy loss is the variational ground-state objective and the two-stage strategy is an emergent observation, not an input.

full rationale

The paper's derivation chain is self-contained. Equation (8) defines the RNN policy θ_t=F_φ(m_1,...,m_{t-1},H); Eq. (5) defines the state transition ρ_t^sys=C_{θ_t,m_t}(ρ_{t-1}^sys); Eq. (9) defines the loss as the energy expectation E(φ)=Tr(ρ_T^sys(φ)H). Minimizing this loss is the standard variational characterization of a ground state, so the objective directly targets the ground state by the definition of ground state; no ground-state answer or protocol is fed into training as supervised data. The claimed two-stage strategy (first steering to |0⟩/|00⟩, then rotating to the target state) is read off from the trained trajectories in Figs. 4 and 5, not imposed by the loss or by the RNN ansatz, so it is an emergent empirical finding rather than a renamed input. There are no self-citations: the relevant prior-work references ([19] Alam & Clark, [20] Puente & Rizzi) are independent groups, and no uniqueness theorem or ansatz is imported from the present authors' prior work. The ambiguities in Eq. (5) (unnormalized conditional state) and Eq. (9) (unspecified averaging over measurement records) are well-posedness and correctness concerns, not circularity. The in-distribution test set is a generalization limitation rather than a definitional reduction. No fitted parameter is relabeled as a prediction, and no load-bearing step reduces to its own input by construction. Hence the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim is supported by a standard variational objective plus a hand-chosen circuit ansatz and time budget; the main unverified premise is that the random training distribution covers the 'wide range' of Hamiltonians the paper claims to handle. No new physical entities are introduced.

free parameters (5)
  • RNN parameters φ = trained via Adam; specific values not reported
    The LSTM weights are optimized to minimize the energy loss; the entire learned protocol is encoded in these fitted parameters.
  • Number of time steps T = T=5 for single qubit, T=8 for two qubits
    Chosen by hand; the observed two-stage strategy and final fidelity depend on having enough steps, and no ablation over T is reported.
  • Gate ansatz per step = Rx/Ry rotations and controlled-X entangling gates
    The circuit architecture is selected ad hoc; the learned protocols are expressed in this specific ansatz.
  • Ancilla configuration (N_anc, N_m, N_t) = swept values listed in Table I
    The claim about measurement-feedback aiding preparation is based on this hand-chosen sweep; the best-of-three selection further biases it.
  • Hamiltonian coefficient sampling range = uniform in [-1,1]
    The training distribution is defined by this choice; generalization to other distributions or structured Hamiltonians is untested.
assumptions (4)
  • standard math Variational principle: the global minimum of E(ρ)=Tr(ρH) is achieved by the ground state of H.
    The loss function Eq. (9) relies on this theorem to turn energy minimization into ground state preparation.
  • domain assumption Eq. (5) with projectors P_mt and partial trace correctly models mid-circuit measurement with feedback.
    The simulation assumes PennyLane's representation of projective measurements and conditioning is faithful.
  • ad hoc to paper Hamiltonians sampled uniformly over Pauli coefficients are representative of generic Hamiltonians.
    Training and test Hamiltonians are drawn from the same distribution; the paper's 'wide range of Hamiltonians' claim extrapolates beyond this without out-of-distribution tests.
  • domain assumption Adam optimization through the quantum-classical recurrent dynamics converges to a good policy.
    The paper reports successful training but provides no convergence analysis or landscape information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Learning for Ground State Preparation via Measurement and Feedback." pith.science (2026). https://pith.science/paper/UQIT57QX

@misc{pith2026250206517,
  author       = {Pith},
  title        = {Pith review of: Machine Learning for Ground State Preparation via Measurement and Feedback},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UQIT57QX}},
  note         = {Machine review of arXiv:2502.06517}
}
read the original abstract

We present a recurrent neural network-based approach for ground state preparation utilizing mid-circuit measurement and feedback. Unlike previous methods that use machine learning solely as an optimizer, our approach dynamically adjusts quantum circuits based on real-time measurement outcomes and learns distinct preparation protocols for different Hamiltonians. Notably, our machine learning algorithm consistently identifies a state preparation strategy wherein all initial states are first steered toward an intermediate state before transitioning to the target ground state. We demonstrate that performance systematically improves as a larger fraction of ancilla qubits are utilized for measurement and feedback, highlighting the efficacy of mid-circuit measurements in state preparation tasks.

Figures

Figures reproduced from arXiv: 2502.06517 by the authors.

Figure 1
Figure 1. FIG. 1. The architecture of the quantum circuit for the 3- [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Structure of an [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. FIG. 4. The single qubit preparation results of 100 random samples. From [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: FIG. 5. The average fidelity and expectation value of [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 17 canonical work pages

  1. [1]

    Peruzzo, J

    A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’Brien, A variational eigenvalue solver on a photonic quantum processor, Nature Communications5, 4213 (2014)

    2014

  2. [2]

    D. Wang, O. Higgott, and S. Brierley, Accelerated varia- tional quantum eigensolver, Phys. Rev. Lett.122, 140504 (2019)

    2019

  3. [3]

    H. R. Grimsley, S. E. Economou, E. Barnes, and N. J. Mayhall, An adaptive variational algorithm for exact molecular simulations on a quantum computer, Nature communications 10, 3007 (2019)

    work page 2019

  4. [4]

    P. G. Anastasiou, Y. Chen, N. J. Mayhall, E. Barnes, andS.E.Economou,Tetris-adapt-vqe: Anadaptivealgo- rithm that yields shallower, denser circuit ansätze, Phys. Rev. Res. 6, 013254 (2024)

    work page 2024

  5. [5]

    Zhang, L

    Y. Zhang, L. Cincio, C. F. Negre, P. Czarnik, P. J. Coles, P. M. Anisimov, S. M. Mniszewski, S. Tretiak, and P. A. Dub, Variational quantum eigensolver with reduced cir- cuit complexity, npj Quantum Information8, 96 (2022)

    work page 2022

  6. [6]

    Miháliková, M

    I. Miháliková, M. Pivoluska, M. Plesch, M. Friák, D. Na- gaj, and M. Šob, The cost of improving the precision of the variational quantum eigensolver for quantum chem- istry, Nanomaterials12, 243 (2022)

    work page 2022

  7. [7]

    Chivilikhin, A

    D. Chivilikhin, A. Y. Samarin, V. I. Ulyantsev, I. V. Iorsh, A. R. Oganov, and O. Kyriienko, Mog-vqe: Multi- objective genetic variational quantum eigensolver, arXiv: Quantum Physics (2020)

    work page 2020

  8. [8]

    R. M. Parrish, E. G. Hohenstein, P. L. McMahon, and T. J. Martínez, Quantum computation of electronic tran- sitions using a variational quantum eigensolver, Phys. Rev. Lett. 122, 230401 (2019)

    work page 2019

Show all 30 references
  1. [9]

    I. G. Ryabinkin, S. N. Genin, and A. F. Izmaylov, Con- strainedvariationalquantumeigensolver: Quantumcom- puter search engine in the fock space, Journal of Chem- ical Theory and Computation 15, 249 (2019), pMID: 30512959, https://doi.org/10.1021/acs.jctc.8b00943

    2019 doi

  2. [10]

    Cerezo, K

    M. Cerezo, K. Sharma, A. Arrasmith, and P. J. Coles, Variational quantum state eigensolver, npj Quantum In- formation 8, 113 (2022)

    2022

  3. [11]

    W. M. Kirby, A. Tranter, and P. J. Love, Contextual subspace variational quantum eigensolver, Quantum 5, 456 (2021)

    2021

  4. [12]

    Y. S. Yordanov, V. Armaos, C. H. Barnes, and D. R. Arvidsson-Shukur, Qubit-excitation-based adaptive vari- ational quantum eigensolver, Communications Physics4, 228 (2021)

    2021

  5. [13]

    T.-C. Lu, L. A. Lessa, I. H. Kim, and T. H. Hsieh, Mea- surement as a shortcut to long-range entangled quantum matter, PRX Quantum3, 040337 (2022)

    2022

  6. [14]

    Tantivasadakarn, R

    N. Tantivasadakarn, R. Thorngren, A. Vishwanath, and R. Verresen, Long-range entanglement from measuring symmetry-protectedtopologicalphases,Phys.Rev.X 14, 021040 (2024)

    2024

  7. [15]

    Piroli, G

    L. Piroli, G. Styliaris, and J. I. Cirac, Quantum circuits assisted by local operations and classical communication: Transformations and phases of matter, Phys. Rev. Lett. 127, 220503 (2021)

    2021

  8. [16]

    K. C. Smith, E. Crane, N. Wiebe, and S. Girvin, Deter- ministic constant-depth preparation of the aklt state on a quantum processor using fusion measurements, PRX Quantum 4, 020315 (2023)

    2023

  9. [17]

    D. Malz, G. Styliaris, Z.-Y. Wei, and J. I. Cirac, Prepa- ration of matrix product states with log-depth quan- tumcircuits,PhysicalReviewLetters 132,10.1103/phys- revlett.132.040404 (2024)

    2024 doi

  10. [18]

    Y. Yan, M. Ma, Y. Zhou, and X. Ma, Variational locc- assisted quantum circuits for long-range entangled states (2024), arXiv:2409.07281 [quant-ph]

    2024 arXiv

  11. [19]

    Alam and B

    F. Alam and B. K. Clark, Learning dynamic quan- tum circuits for efficient state preparation (2024), arXiv:2410.09030 [quant-ph]

    2024 arXiv

  12. [20]

    D. A. Puente and M. Rizzi, Learning feedback mecha- nisms for measurement-based variational quantum state preparation (2025), arXiv:2411.19914 [quant-ph]

    2025 arXiv

  13. [21]

    Ostaszewski, L

    M. Ostaszewski, L. M. Trenkwalder, W. Masarczyk, E. Scerri, and V. Dunjko, Reinforcement learning for optimization of variational quantum circuit architec- 6 tures, inAdvances in Neural Information Processing Sys- tems, Vol. 34, edited by M. Ranzato, A. Beygelzimer, Y. Dauphin...

    2021

  14. [22]

    Uvarov, J

    A. Uvarov, J. D. Biamonte, and D. Yudin, Variational quantum eigensolver for frustrated quantum systems, Phys. Rev. B102, 075104 (2020)

    2020

  15. [23]

    Y. J. Patel, A. Kundu, M. Ostaszewski, X. Bonet- Monroig, V. Dunjko, and O. Danaci, Curriculum rein- forcement learning for quantum architecture search un- der hardware errors (2024), arXiv:2402.03500 [quant-ph]

    2024 arXiv

  16. [24]

    Kundu, A

    A. Kundu, A. Sarkar, and A. Sadhu, Kanqas: Kolmogorov-arnold network for quantum architecture search, EPJ Quantum Technology11, 76 (2024)

    2024

  17. [25]

    Kundu, Improving thermal state preparation of sachdev-ye-kitaev model with reinforcement learning on quantum hardware (2025), arXiv:2501.11454 [quant-ph]

    A. Kundu, Improving thermal state preparation of sachdev-ye-kitaev model with reinforcement learning on quantum hardware (2025), arXiv:2501.11454 [quant-ph]

    2025 arXiv

  18. [26]

    Bergholm, J

    V. Bergholm, J. Izaac, M. Schuld, C. Gogolin, S. Ahmed, V. Ajith, M. S. Alam, G. Alonso-Linaje, B. Akash- Narayanan, A. Asadi, et al., Pennylane: Automatic dif- ferentiation of hybrid quantum-classical computations, arXiv preprint arXiv:1811.04968 (2018)

    2018 arXiv

  19. [27]

    https://www.tensorflow.org/about/bib?hl=en

  20. [28]

    Hochreiter, Long short-term memory, Neural Compu- tation MIT-Press (1997)

    S. Hochreiter, Long short-term memory, Neural Compu- tation MIT-Press (1997)

    1997

  21. [29]

    D. P. Kingma and J. Ba, Adam: A Method for Stochastic Optimization, in 3rd International Conference on Learn- ing Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , edited by Y. Bengio and Y. LeCun (2015) arXiv:1412.6980 [cs.LG]

    2015 arXiv

  22. [30]

    Tilly, H

    J. Tilly, H. Chen, S. Cao, D. Picozzi, K. Setia, Y. Li, E. Grant, L. Wossnig, I. Rungger, G. H. Booth, and J. Tennyson, The variational quantum eigensolver: A re- view of methods and best practices, Physics Reports986, 1 (2022), the Variational Quantum Eigensolver: a review of...

    2022

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.