{"id":"cbdbcb6b-5bf5-4137-95a3-23f0c15ca6fd","arxiv_id":"2412.11015","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A quantum reservoir processing protocol learns a bosonic cQED device's true measurement dynamics from 36 training states and reconstructs unseen kitten states with fidelity above 91%, outperforming the idealized model map.","lead":"Researchers taught a quantum reservoir on a superconducting circuit to learn the real behavior of a bosonic cavity from just 36 training measurements, then used it to reconstruct unseen quantum states with over 91% fidelity. The result is a practical demonstration that machine learning can replace precise system modeling for continuous-variable quantum tomography on real hardware.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The learnt map is trained and tested against the same GRAPE-simulated state preparation; without independent state characterization, the reported >91% fidelities may be inflated by common-mode simulation error.","rationale":"The reader's weakest assumption already identifies the GRAPE-simulated training labels as the key concern, and my reading agrees. I refine it by noting the common-mode structure: the same simulation supplies both the regression targets and the fidelity reference, so systematic preparation error can be absorbed by the learned map and inflate apparent performance. This is the single most load-bearing issue because the paper's headline number—91% vs 59%—rests on comparing the reconstructed state to a simulated target, and the training procedure has no independent check on that simulation. The manuscript does contain real supporting evidence: the learned map also reduces observable error relative to the idealised map, and it matches the best-case simulated map, which suggests the effect is not purely noise fitting. The GitHub data/code availability is a further positive. But those checks are all model-dependent in the same direction, so they do not settle the concern. An independent characterization of the test states (e.g., Wigner tomography) is necessary and would be decisive. Since the paper's conclusion is otherwise plausible and the requested check is feasible, the existing conditional verdict remains appropriate rather than a rejection.","tokens_in":18989,"tokens_out":6427,"duration_ms":61062,"concrete_test":"Independently characterize the four test kitten states with a method that does not rely on the GRAPE state-preparation simulation—for example, direct Wigner tomography via parity measurements, or the optimized excitation-counting protocol of Ref. [51]—and reconstruct their density matrices rho_ind. Then recompute the reconstruction fidelity F(rho_BME, rho_ind) for the learnt map (and the idealised and simulated maps) for all four states. If the learnt-map fidelities remain above roughly 0.9 while the idealised map stays near 0.59, the common-mode concern is resolved; if they drop substantially, the reported enhancement is partly an artifact of shared GRAPE simulation labels and targets.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is that the learnt map reconstructs previously unseen kitten states with fidelity >91%, versus 59% for the idealised map, because it captures the device's real imperfect dynamics. But both the training labels and the test targets come from the same GRAPE simulation of state preparation. In the Results section the authors state that the training set {Y_n(rho_n), X_n} is built from 'the density matrix simulated from GRAPE pulses taking into account decoherence', and the test states 'simulated from GRAPE pulses are denoted rho_± and rho_y±'. If the actual prepared states differ systematically from these simulations (e.g., due to pulse distortions, calibration errors, or unmodelled Hamiltonian terms), then ridge regression can absorb the preparation discrepancy into the learnt map: the map is fit to send simulated input coordinates to measured outputs, so inverting it on test data returns the simulated coordinates rather than the true physical state. The reported fidelity to the simulated targets would then be inflated, and the claim that the map learns the true dynamics would not be established. The observable-MSE comparison in Fig. 4 uses the same simulated Y vectors, so it also cannot separate preparation errors from dynamical errors. The simulated-map comparison in Fig. 5a is a useful sanity check, but it is likewise model-based and does not break the common-mode dependence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports an experimental demonstration of quantum reservoir processing (QRP) for continuous-variable state reconstruction on a bosonic circuit-QED platform. The method learns a linear process map from measurement outcomes of a training set of states (Fock states and superpositions) prepared by GRAPE-optimized pulses, then inverts the map and applies Bayesian inference to reconstruct four previously unseen kitten states. The authors report reconstruction fidelities greater than 91% with the learnt map, compared to about 59% with an idealized model map, and attribute the improvement to the learnt map's ability to capture physical non-idealities such as dynamical errors, decoherence, and systematic measurement errors. The manuscript also compares the learnt map with a simulated map based on a detailed Hamiltonian model and provides a full description of the experimental setup and analysis procedures.","tokens_in":19236,"tokens_out":5573,"duration_ms":49475,"significance":"If the central claim holds, the work demonstrates a practical and resource-efficient QRP-based tomography scheme for bosonic systems using minimal measurements and a single qubit-cavity device. The manuscript has notable strengths: the learnt map is tested on out-of-sample states not used in training; bootstrap error bars are reported; the data and code are publicly available on GitHub; and the comparison against both an idealized map and a detailed simulated map provides context for the improvement. The protocol uses the theoretical minimum number of observables, which is valuable for scalability. However, the central validation is weakened by the fact that both the training labels and the test targets are obtained from GRAPE simulations of state preparation rather than from an independent experimental characterization, making the reported fidelities potentially sensitive to common-mode simulation error.","major_comments":[{"comment":"The training set {Y_n(rho_n), X_n} and the test targets rho_± and rho_y± are both obtained from GRAPE simulations of the state preparation, rather than from an independent characterization of the actually prepared states. The reported fidelities therefore measure agreement with the simulated states. If the real prepared states differ systematically from the simulation (e.g., due to pulse distortions, calibration errors, or unmodelled Hamiltonian terms), ridge regression can absorb part of that preparation discrepancy into the learnt map, and the claimed >91% fidelity would be inflated relative to the true physical states. The statement in the text that the simulated density matrices are close to the ideal states (average fidelity ≈0.97) does not establish that the simulation is close to the experimental states. The authors should add an independent validation of the prepared states (e.g., Wigner tomography or a second tomographic method), or otherwise demonstrate the simulation's accuracy, before the main claim that the map captures the device's true imperfect dynamics can be accepted.","section":"Results, 'Quantum process reconstruction' and Fig. 5a"},{"comment":"The observable MSE plotted in Fig. 4 is computed with the same simulated Y vectors for both the idealised and learnt maps. Because both the training and test Y are simulation-derived, a systematic preparation error would appear as a common-mode bias in this metric. The lower MSE for the learnt map could reflect the map's ability to reproduce the simulated coordinates rather than the true physical process, so this comparison cannot separate preparation errors from dynamical errors. The text's claim that the learning protocol addresses all imperfections except random measurement errors is therefore not fully supported by the presented data.","section":"Fig. 4 and Eq. (2)"},{"comment":"The simulated-map comparison does not break the common-mode dependence. The blue shaded region in Fig. 5a is generated from a Hamiltonian model and is evaluated against the same GRAPE-simulated kitten-state targets. While the volatility of the simulated map under parameter perturbations is a useful sanity check, it does not validate the learnt map against an independently known ground truth. The conclusion that 'the learnt map achieves reconstruction fidelities comparable to the best case scenarios of the simulated maps' is conditional on the simulation model being accurate; the authors should either provide an independent measurement of the prepared states or clearly state that the reported fidelities are relative to the simulated states in both the main text and the abstract.","section":"Fig. 5a and Appendix 'Simulated map'"}],"minor_comments":[{"comment":"The regularization coefficient ν is described as 'selected to optimize the balance between overfitting and underfitting from noisy data', but no value or selection procedure is given. If ν is tuned on the test states, this would introduce circularity; if it is chosen by cross-validation on the training data, that procedure should be described explicitly.","section":"Appendix 'Learning with ridge regression', Eq. (5)"},{"comment":"The definition of the element-wise MSE is presented ambiguously: the sentence 'averaged over all D^4 - D^2 elements' with the displayed denominator D^2(D^2 - 1) is confusing because these expressions are not obviously equal. Please clarify the normalization used in Fig. 3a.","section":"Fig. 3 and surrounding text"},{"comment":"The abstract and introduction state that the scheme does not require precise control over the quantum system, but the protocol relies on GRAPE-optimized 2 µs state-preparation pulses and precisely timed displacement and Ramsey sequences. This apparent tension should be addressed in the introduction to avoid overstating the relaxation of control requirements.","section":"Main text, 'Implementation in cQED'"},{"comment":"The text states that the simulated density matrices for the training set are close to ideal with fidelity ≈0.97, but the corresponding fidelity for the test kitten states is not reported. Please report the simulated-to-ideal fidelities for the four test states, since they are the targets in Eq. (3).","section":"Results, 'Quantum state reconstruction'"},{"comment":"The term 'true dynamical process' and 'accurate process map' are used in the conclusion, but given the simulation-based validation, these claims are stronger than what the data establish. Please soften the wording or add the caveat about simulation-dependent validation.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the common-mode simulation dependence of the training labels and test targets. This is fixable within the scope of an experimental paper if the authors can provide an independent characterization of the prepared states, or if they clearly reframe the claims as being relative to the simulated model. The heavy reliance on a small set of self-citations for the foundational QRP framework is also noticeable; I would encourage the authors to ground those claims in a broader literature. The paper is otherwise well organized and the experimental data are of interest to the quantum information community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is the first experimental demonstration of quantum reservoir processing for full continuous-variable state reconstruction on a bosonic cQED device, in a six-dimensional truncated Hilbert space. The learned map clearly beats the idealized model map on four out-of-sample kitten states (fidelity >91% vs 59%), and the data and code are public. The core result is real and worth engaging with.\n\nWhat's genuinely new: earlier experimental QRP work only estimated properties of small qubit systems. This paper reconstructs full CV states using the minimum number of measurements (D^2-1 observables, D^2 training states). The training set is minimal, the test states are genuinely unseen, and the comparison to a simulated map with perturbed parameters is a useful sanity check: the learned map is stable and matches the best-case simulation. The paper is also honest about using GRAPE-simulated density matrices as training labels, explicitly saying they are decoupling state-preparation imperfections by simulation.\n\nThe soft spot is the common-mode dependence on that simulation. Both the training labels and the test target states come from the same GRAPE simulation. If the actual prepared states differ systematically from the simulation—pulse distortions, calibration drift, unmodeled Hamiltonian terms—ridge regression can absorb the difference into the learned map. Inverting that map on test data then returns the simulated coordinates, not the true physical state, and the reported fidelities (computed against the simulated targets) could be inflated. The observable-MSE comparison in Fig. 4 uses the same simulated Y vectors, so it can't separate preparation errors from dynamical errors either. The simulated-map comparison in Fig. 5a is model-based and doesn't break this dependence. This is a genuine limitation, but not a deal-breaker: the improvement over the idealized map is large, and the learned map's stability across parameter perturbations is evidence that it captures real dynamics. Still, the absolute fidelity numbers should be treated cautiously until independent state characterization (e.g., Wigner tomography of the prepared kitten states) is done.\n\nWho is this for? Anyone working on quantum reservoir computing, bosonic codes, or practical tomography on cQED. It deserves a serious referee. The method is clearly laid out, the experiment is nontrivial, and the flaw is addressable. I'd send it to peer review with a request for either an independent validation or a more careful statement of what the fidelities mean.","headline":"First QRP-based CV state reconstruction on cQED, with real promise and a simulation-dependence caveat that needs referee attention.","tokens_in":19803,"tokens_out":3909,"would_cite":true,"duration_ms":34411,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By learning a bosonic cavity's real dynamics from a minimal measurement set, quantum reservoir processing reconstructs continuous-variable states at over 91 percent fidelity, versus 59 percent for an idealized model.","keywords":["quantum reservoir processing","quantum state tomography","process tomography","continuous-variable quantum information","circuit quantum electrodynamics","bosonic state reconstruction","Bayesian quantum state estimation","kitten states"],"falsifier":"Perform full Wigner tomography of the four kitten states immediately after preparation and compare the measured density matrices with the simulated training targets; if the average fidelity between them falls well below the reported $\\approx 0.97$, the learnt map is trained on incorrect labels and the claimed reconstruction fidelities are not supported.","tokens_in":18780,"feed_emoji":"⚛️","tokens_out":10770,"duration_ms":90039,"temperature":0.7,"pith_summary":"The paper reports a laboratory demonstration of a machine-learning approach to tomography of a bosonic mode, the continuous-variable quantum system used in many error-correction schemes. The method, quantum reservoir processing, does not require a precise Hamiltonian model; instead it learns the actual map from input states to measurement outcomes from a small training set, then inverts that learnt map to reconstruct unknown states. On four 'kitten' states, small-amplitude superpositions of coherent states that were not part of the training data, the learnt map reconstructed states with fidelity above 91 percent, while a map computed from an idealized model of the device reached 59 percent. The authors claim this closes the gap left by device imperfections such as imperfect pulses, decoherence, and systematic measurement errors, making reliable bosonic state and process reconstruction practical on real hardware.","feed_headline":"Learnt map lifts quantum state fidelity from 59% to over 91%","feed_subtitle":"Learning the cavity's real dynamics, not a perfect model, makes state reconstruction reliable on noisy devices.","key_machinery":"The load-bearing object is the linear map $\\beta = [\\vec{V}, M]$ that connects a vectorised description $\\vec{Y}$ of the input state to the measured observables $\\vec{X}$ through $\\vec{X} = \\beta [1; \\vec{Y}]$. Because the device's real dynamics together with the measurement is still a completely positive, trace-preserving process, this linear relation survives even when the idealised parity model fails; training states provide labelled pairs $(\\vec{Y}, \\vec{X})$ from which ridge regression fixes $\\beta$ using the minimum $D^2$ states and $D^2-1$ observables. Once $\\beta$ is learnt, an unknown state is estimated by inverting the relation and passing the result through Bayesian inference to obtain a physical density matrix. The displacements preceding each parity measurement are chosen by gradient descent to keep the map's condition number small, controlling how measurement noise amplifies into the reconstructed state.","core_discovery":"The central experimental claim is that a learnt process map, fixed by ridge regression from measurement outcomes of 36 known input states, captures the true dynamics of the device's displacement-and-parity measurement better than any parameter-based model the authors could construct. Tested on four kitten states in a six-dimensional truncation, the learnt map gives reconstruction fidelities above 91 percent, while the idealised map, computed by assuming a perfect parity mapping, gives around 59 percent. The gap is traced mainly to coherent errors in the $\\pi/2$-wait-$\\pi/2$ Ramsey parity sequence and to qubit dephasing; because the learnt map absorbs these errors into the linear map $\\beta$, it also outperforms a simulated map whose fidelity is volatile under realistic parameter uncertainty. The paper therefore claims that quantum reservoir processing turns an imperfect, partly unknown device into a usable tomography instrument for continuous-variable states.","pith_inferences":["The element-wise deviation between idealised and learnt maps grows with truncation dimension, so a comparison of the two maps could serve as a diagnostic that identifies which operations drift furthest from the model.","The protocol should transfer to other bosonic platforms, such as mechanical, photonic, or trapped-ion systems, because it assumes only a linear displacement-like operation followed by a parity-type measurement, though the training labels would need independent validation on each platform.","A decisive benchmark would be to retrain the map using states characterized by an independent method, such as full Wigner tomography, in place of simulated labels; if reported fidelities survive that substitution, the method no longer depends on trust in the pulse simulation.","The linearity assumption ties the learnt map to the truncation dimension used in training, so scaling to larger Hilbert spaces would require retraining rather than extrapolating a low-dimensional map."],"forward_implications":["For a fixed truncation dimension $D$, tomography uses only $D^2$ training states and $D^2-1$ observables, the theoretical minimum, so the overhead is set by the state's dimension rather than by detailed assumptions about the device.","The same linear-map machinery yields process tomography of an unknown dynamics, not just state tomography, because the paper shows how to isolate the underlying dynamical map from the learnt $\\beta$.","Because the learnt map absorbs coherent pulse errors, qubit dephasing, and systematic readout offsets, it stays accurate without re-calibrating each imperfection separately.","The method applies in principle to arbitrary states within the truncation dimension, not only to the kitten states used as an out-of-sample test, and the authors project that faster cavity reset could shrink the whole learning run to about a minute."],"supporting_citations":[{"why":"Introduces quantum reservoir processing as learning a single output layer without reservoir control, the conceptual basis of the scheme.","marker":"[18]"},{"why":"Proposes reconstructing quantum states with quantum reservoir networks, the theoretical task this experiment implements.","marker":"[27]"},{"why":"Supplies the idealised-map baseline and identifies coherent parity-mapping errors that motivate learning the real dynamics.","marker":"[51]"},{"why":"Establishes the linear relation between input-state parameters and observables and the minimal-measurement completeness conditions used for $\\beta$ and its inversion.","marker":"[54]"},{"why":"Supplies the GRAPE pulse optimization used to prepare training and test states and the Q-switching speed-up estimate.","marker":"[55]"},{"why":"Provides the direct parity-measurement method on which the displacement-then-parity observables are based.","marker":"[56]"},{"why":"Gives the cavity-QED parity mapping sequence whose imperfections the learnt map accounts for.","marker":"[57]"},{"why":"Provides the efficient Bayesian mean estimator used to turn inverted observables into physical density matrices.","marker":"[63]"}],"fun_headline_variants":["Quantum reservoir processing lifts tomography fidelity to 91%","Learnt map outperforms ideal model in quantum state reconstruction","Learning real cavity dynamics beats perfect-model tomography","Reservoir processing turns noisy device into reliable tomography tool","From 59% to 91%: Learning device's true process improves tomography"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulated density matrices used as training labels match the states actually prepared in the device, because if the simulation misrepresents those states, the learnt map is trained on incorrect ground truth and the reported fidelities could be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Quantum reservoir processing lifts tomography fidelity to 91%","Learnt map outperforms ideal model in quantum state reconstruction","Learning real cavity dynamics beats perfect-model tomography","Reservoir processing turns noisy device into reliable tomography tool","From 59% to 91%: Learning device's true process improves tomography"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000358,"raw_usage":{"total_tokens":1916,"prompt_tokens":898,"completion_tokens":1018,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":936}},"tokens_in":514,"tokens_out":1018,"duration_ms":8090,"temperature":1.0,"reasoning_tokens":936,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:23:34.815099+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Perform full Wigner tomography of the four kitten states immediately after preparation and compare the measured density matrices with the simulated training targets; if the average fidelity between them falls well below the reported $\\approx 0.97$, the learnt map is trained on incorrect labels and the claimed reconstruction fidelities are not supported.","supporting_citations":[{"cited_title":"Quantum reservoir processing,","cited_arxiv_id":null,"evidence_quote":"Introduces quantum reservoir processing as learning a single output layer without reservoir control, the conceptual basis of the scheme."},{"cited_title":"Reconstruct- ing quantum states with quantum reservoir networks,","cited_arxiv_id":null,"evidence_quote":"Proposes reconstructing quantum states with quantum reservoir networks, the theoretical task this experiment implements."},{"cited_title":"Demonstrating efficient and robust bosonic state reconstruction via optimized excitation counting,","cited_arxiv_id":null,"evidence_quote":"Supplies the idealised-map baseline and identifies coherent parity-mapping errors that motivate learning the real dynamics."},{"cited_title":"Tomographic completeness and ro- bustness of quantum reservoir networks,","cited_arxiv_id":null,"evidence_quote":"Establishes the linear relation between input-state parameters and observables and the minimal-measurement completeness conditions used for $\\beta$ and its inversion."},{"cited_title":"Implementing a universal gate set on a log- ical qubit encoded in an oscillator,","cited_arxiv_id":null,"evidence_quote":"Supplies the GRAPE pulse optimization used to prepare training and test states and the Q-switching speed-up estimate."},{"cited_title":"Method for direct measurement of the Wigner function in cavity qed and ion traps,","cited_arxiv_id":null,"evidence_quote":"Provides the direct parity-measurement method on which the displacement-then-parity observables are based."},{"cited_title":"Direct measurement of the Wigner function of a one-photon fock state in a cav- ity,","cited_arxiv_id":null,"evidence_quote":"Gives the cavity-QED parity mapping sequence whose imperfections the learnt map accounts for."},{"cited_title":"A practical and efficient approach for Bayesian quantum state estimation,","cited_arxiv_id":null,"evidence_quote":"Provides the efficient Bayesian mean estimator used to turn inverted observables into physical density matrices."}],"review_version":1}