{"id":"b37ae7b5-69ff-4991-b9e6-8dfb6b72def0","arxiv_id":"2501.01646","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MPS pre-training, neural-network-assisted zero-noise extrapolation, and Pauli grouping combine to give simulated H4 ground-state energies within about 0.02 Hartree of the FCI benchmark under a specific noise model.","lead":"This paper proposes a variational quantum eigensolver that initializes its quantum circuit from a classically pre-trained matrix product state, then applies zero-noise extrapolation with a neural network to reduce simulated hardware noise. It reports better noise tolerance than several baseline VQE circuits on the H4 molecule, though the absolute energy error remains far from chemical accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Neural-network ZNE is unvalidated; the reported noisy MPS-VQE energy of -2.1490 Hartree may be an extrapolation artifact rather than evidence of noise mitigation.","rationale":"The reader's weakest assumption is the neural-network ZNE, and I agree that this is the single most load-bearing point: Table II's headline noisy MPS-VQE result is produced by that untested component, and the paper provides no bracketing, no comparison to polynomial extrapolation, and no error bars; best-of-30 reporting further masks variance. The unphysical HE-VQE noiseless value reinforces that the numerical pipeline has at least one bug, though it is secondary to the ZNE concern. I would keep the CONDITIONAL verdict: the concern does not disprove the method, but it must be resolved by an ablation and an independent reference check before the central claim can be accepted. The public code makes the proposed test directly runnable.","tokens_in":7621,"tokens_out":3713,"duration_ms":37321,"concrete_test":"Re-run the released code for H4/STO-3G with the stated noise model, replacing the NN extrapolator with (i) linear ZNE, (ii) quadratic ZNE, and (iii) no ZNE, reporting mean and standard deviation over the same 30 seeds. If NN-ZNE does not outperform the polynomial fits, or all results are within noise of each other, the headline noisy energy is not attributable to the proposed NN-ZNE. Independently diagonalize the STO-3G H4 Hamiltonian to confirm the FCI value of -2.1664 Hartree, since the noiseless HE-VQE value of -2.1723 Hartree below FCI signals a reference-calculation error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison in Table II rests on the noisy MPS-VQE energy (-2.1490 Hartree), produced by the neural-network zero-noise extrapolation described in Section II-F. The paper specifies only that a three-layer fully connected network fits expectation values at different noise levels and evaluates at zero noise; it omits the number and values of the noise-scaling factors (circuit folding ratios), the training data generation, hyperparameters, and any validation. A flexible NN fit to a small number of noisy points can overfit and extrapolate to an arbitrary value; without an ablation against standard polynomial ZNE or no ZNE, there is no evidence that -2.1490 reflects the noiseless limit rather than the fitting model. This is load-bearing because the central claim of noise errors within O(10^-2) to O(10^-1) is measured relative to this number. A secondary internal inconsistency reinforces the need for caution: in Table II, the noiseless HE-VQE baseline (-2.1723 Hartree) lies below the FCI benchmark (-2.1664 Hartree), which is impossible for a variational ground-state calculation; if the reference energy is miscomputed, all reported errors and comparisons are suspect.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a noise-mitigated variational quantum eigensolver, called MPS-VQE, that combines three ingredients: a shallow hardware-efficient circuit inspired by matrix product states, a classical pre-training step that initializes circuit parameters from an optimized MPS, and a zero-noise extrapolation procedure in which a three-layer neural network fits expectation values at different noise levels and extrapolates to the noiseless limit. The authors also introduce a commuting-grouping measurement strategy for Hamiltonian Pauli strings. Numerical simulations for a linear H4 molecule in the STO-3G basis under a specific depolarizing, thermal-relaxation, and bit-flip noise model are used to claim noise errors in the range O(10^-2) to O(10^-1), outperforming HE-VQE, Qubit UCC, SE Ansatz, and UCCSD baselines.","tokens_in":7846,"tokens_out":4203,"duration_ms":42507,"significance":"If fully substantiated, the paper would offer a useful contribution: a shallow MPS-inspired ansatz with classical MPS pre-training is a legitimate and potentially practical way to initialize VQE circuits, and the combination with ZNE and Pauli grouping addresses two acknowledged bottlenecks in near-term VQE. Strengths include the explicit noise model, the comparison of circuit metrics, and the provision of code repositories. However, the current evidence is only a single small molecule, and the central quantitative claim rests on an internally inconsistent benchmark and an under-specified neural-network extrapolation. The contribution is therefore promising but not yet established at the level claimed.","major_comments":[{"comment":"The noiseless HE-VQE entry (-2.1723 Hartree) lies below the FCI benchmark (-2.1664 Hartree). For any variational trial state, the noiseless energy must be at least the exact ground-state energy in the chosen basis; a value below FCI indicates an inconsistency in the Hamiltonian, the qubit mapping, or the FCI reference. This makes the entire comparison in Table II unreliable and calls into question the stated error ranges relative to the FCI value. The authors must correct the benchmark or the HE-VQE calculation and rerun the comparisons.","section":"Table II"},{"comment":"The neural-network zero-noise extrapolation is described only as a three-layer fully connected network that fits expectation values at different noise levels and evaluates at zero noise. The manuscript omits the number and values of the noise-scaling factors (circuit-fold ratios), the amount of training data, the loss function, hyperparameters, regularization, and any validation against held-out noise levels. It also provides no comparison with standard polynomial ZNE or with no extrapolation. Because the headline noisy MPS-VQE energy (-2.1490 Hartree) is the extrapolated value, the central noise-mitigation claim is not yet supported. Please provide the full protocol and an ablation study.","section":"Section II-F"},{"comment":"Pre-training initializes circuit parameters from an MPS, but the paper does not specify how the MPS tensors A[n] are converted into the rotation angles of the circuit in Eq. (7), nor does it demonstrate that the one-layer circuit can faithfully represent the optimized MPS for the bond dimension used. If the circuit ansatz is not capable of representing the pre-trained state, the benefit of pre-training is not guaranteed. The authors should describe the parameter mapping and report the fidelity or energy difference between the pre-trained MPS and the state prepared by the initialized circuit.","section":"Section II-D / Eq. (7)"},{"comment":"All noisy and noiseless entries in Table II are reported as the best results from 30 independent experiments, with no mean, median, standard deviation, or number of shots. Best-of-30 reporting is sensitive to optimization and sampling noise and can create the appearance of robustness where none exists. The authors should report the distribution of outcomes, the number of shots per expectation value, and multiple noise seeds to support the reproducibility of the results.","section":"Section III"},{"comment":"The experimental evidence consists of a single assumed H4 geometry in a minimal basis under one specific noise model. The conclusion that the method outperforms mainstream variational quantum eigensolvers is therefore overreaching. Additional molecules, different noise strengths, and a corrected HE-VQE baseline are needed before the comparative claim is supported.","section":"Section III"}],"minor_comments":[{"comment":"The text states that the Pauli Z gate has eigenvalues 0 and 1; the correct eigenvalues are +1 and -1. The intended statement is presumably about measuring computational-basis bits 0/1, but the wording is incorrect.","section":"Section II-E"},{"comment":"The axes are not labeled in either panel. The histogram should specify the energy bins, and the lower panel should state what quantity is plotted as a function of what independent variable.","section":"Figure 2"},{"comment":"The qubit encoding of the H4 Hamiltonian (e.g., Jordan-Wigner, parity, or Bravyi-Kitaev) is not stated, even though the encoding affects both the circuit structure and the noise sensitivity.","section":"Section III"},{"comment":"It is unclear why the barrier count is reported as a meaningful metric and whether barriers affect the noise simulation; the authors should clarify this in the table caption or text.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The benchmark inconsistency in Table II is the most urgent concern; it suggests a mistake in either the FCI reference or the HE-VQE implementation. Given that the central claim is measured against FCI, this must be corrected before the paper can be evaluated fairly. I recommend asking the authors to provide the exact Hamiltonian and FCI calculation, to make the ZNE fitting code and raw data available, and to report full statistics rather than best-of-30 results. The underlying idea is plausible and the code availability is a positive sign, but the current manuscript needs substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper combines three known ingredients—MPS-style circuit pre-training, zero-noise extrapolation, and Pauli grouping—into one VQE pipeline and tests it on H4 under a defined noise model. What is genuinely new is the specific combination and the use of a small neural network for the ZNE fit. The code is available, and the gate count for the MPS-VQE circuit is much lower than UCCSD, which is a practical advantage. The pre-training results in Figure 2 do show a tighter energy distribution over many runs, so that part earns its keep.\n\nThe soft spots are real, and the stress-test note points at the load-bearing one. The headline noisy energy of -2.1490 Hartree comes from the neural-network ZNE described only as a three-layer fully connected network. There is no information on the noise-scaling factors, training data, hyperparameters, or validation, and no comparison to standard polynomial extrapolation. With a flexible fit to a small number of noisy points, that number could easily be a fitting artifact rather than a mitigated estimate. That concern is not hypothetical; it is exactly what the paper needs to rule out.\n\nThe secondary issue is also real: in Table II, the noiseless HE-VQE baseline (-2.1723) sits below the stated FCI benchmark (-2.1664). For a variational ground-state calculation this is impossible, which suggests either the reference energy is miscomputed or the HE-VQE calculation used a different Hamiltonian. Either way, it undermines confidence in the error comparisons. The best-of-30 reporting without error bars is a third problem, especially given the authors themselves note sensitivity to initialization.\n\nThese are not fatal flaws in the core idea. The MPS pre-training is legitimate, the circuit is shallow, and neural-network ZNE might work if specified properly. But as written, the central claim about noise robustness is not backed by sufficient evidence. A serious referee could extract a solid paper from this material, but the authors need to add full ZNE details, ablation against polynomial ZNE and no ZNE, error bars or distributions over the 30 runs, and a fix or explanation of the FCI inconsistency.\n\nI would send this to peer review: it is a sensible, code-backed algorithmic study that deserves careful referee time. But I would not cite it in its current form, and I would not trust the -2.1490 number without verification.","headline":"A plausible but under-validated MPS-VQE+ZNE combination; the central noisy energy figure may be an extrapolation artifact.","tokens_in":8382,"tokens_out":1273,"would_cite":false,"duration_ms":14091,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A matrix-product-state circuit plus neural zero-noise extrapolation keeps the noisy H4 ground-state energy within 0.017 Hartree of exact, while conventional VQE variants deviate by up to 1.6 Hartree.","keywords":["variational quantum eigensolver","matrix product states","zero-noise extrapolation","neural network error mitigation","quantum chemistry","H4 molecule","noisy intermediate-scale quantum","Pauli measurement grouping"],"falsifier":"Run the same H4 simulation under the same depolarizing, thermal-relaxation, and bit-flip noise model, but replace the neural-network fit with a standard polynomial or exponential zero-noise extrapolation on the same folded circuits; if the polynomial-extrapolated energy differs from the reported -2.1490 Hartree by more than the claimed $\\mathcal{O}(10^{-2})$ error, then the neural extrapolator, rather than the circuit or pre-training, is the source of the improvement.","tokens_in":7378,"feed_emoji":"⚛️","tokens_out":7587,"duration_ms":66281,"temperature":0.7,"pith_summary":"This paper tries to show that variational quantum eigensolvers can deliver accurate molecular ground-state energies on near-term noisy hardware rather than only in ideal noiseless simulation. It constructs the trial state as a matrix product state, maps that structure into a shallow parameterized circuit, pre-trains the circuit parameters classically, and applies zero-noise extrapolation in which a small neural network fits expectation values at several elevated noise levels and evaluates the fit at zero noise. On an eight-qubit simulation of H4 with depolarizing, thermal relaxation, and bit-flip noise, the method reports a noisy ground-state energy of -2.1490 Hartree against the full-configuration-interaction value of -2.1664 Hartree, an error of about 0.017 Hartree, while four mainstream VQE variants land between -0.5293 and -1.6726 Hartree. The significance of the claim is that a compact circuit plus error mitigation, rather than a very deep circuit, may be enough for useful quantum chemistry on today's devices.","feed_headline":"MPS-VQE keeps H4 ground-state energy within 0.02 Hartree under noise","feed_subtitle":"Shallow matrix-product-state circuit plus neural noise extrapolation beats conventional VQE in noisy H4 simulation.","key_machinery":"The central object is the matrix-product-state ansatz treated as a parameterized quantum circuit: each local MPS tensor is mapped to a two-qubit unitary of the form $(U(\\theta_{n,l_0})\\otimes U(\\theta_{n,l_1}))\\,\\mathrm{CNOT}\\,(U(\\theta_{n,r_0})\\otimes U(\\theta_{n,r_1}))$, with single-qubit $U(\\theta_*)=R_Z(\\theta_0)R_Y(\\theta_1)R_Z(\\theta_2)$, laid out in a brick-wall pattern so the full eight-qubit circuit has 91 gates and 84 parameter gates. The argument is carried by three mechanisms: center-orthogonal gauging of the MPS, which stabilizes local tensor updates and simplifies energy contraction; classical pre-training of the tensor parameters so the circuit starts near the solution and avoids initialization-driven fluctuations; and zero-noise extrapolation, where circuit folding scales the noise and a three-layer fully connected neural network fits the expectation values at nonzero noise levels to extrapolate the value at zero noise. Commuting Pauli strings of the Hamiltonian are grouped so that compatible terms are measured in the same shot, reducing sampling overhead.","core_discovery":"The central claim is that combining an MPS-structured ansatz with classical pre-training and neural-network zero-noise extrapolation makes VQE markedly more noise-robust. In noiseless simulation the 91-gate circuit reaches -2.1609 Hartree, close to the UCCSD value of -2.1615 Hartree and to the FCI benchmark of -2.1664 Hartree, despite using roughly thirty times fewer gates than UCCSD. Under the specified noise model the MPS-VQE noisy energy is -2.1490 Hartree, within 0.017 Hartree of FCI, whereas the hardware-efficient, qubit-UCC, SE, and UCCSD baselines fall to -1.6726, -0.6916, -1.5781, and -0.5293 Hartree respectively. The paper interprets this as evidence that the shallow MPS circuit limits accumulated gate noise, pre-training removes initialization fluctuations, and the neural-network extrapolation removes much of the remaining error.","pith_inferences":["Inference: the method's practical value depends on how many noise-scaling points the neural extrapolator needs; a direct comparison against polynomial ZNE on the same folded circuits would isolate whether the neural fit or the compact circuit is doing the heavy lifting.","Inference: the same MPS-to-circuit construction should extend to longer molecules and to spin-chain Hamiltonians, and the pre-training benefit should grow with the bond dimension of the classical MPS, but neither extension is tested here.","Inference: the reported noisy energy of -2.1490 Hartree still lies 0.0119 Hartree above the noiseless MPS-VQE value, so the extrapolation does not fully remove noise; reporting median and worst-case energies across the 30 runs, rather than only the best, would give a sharper picture of typical performance."],"forward_implications":["If the claim holds, shallow MPS-structured circuits can substitute for much deeper chemistry ansatze in noisy settings, reducing both gate count and error accumulation.","Classical MPS pre-training gives a deterministic, physics-informed starting point for VQE optimization, which should reduce run-to-run variance and speed convergence.","Neural-network zero-noise extrapolation can be added to existing VQE pipelines as a replacement for polynomial extrapolation, as long as enough noise-scaled expectation values are collected.","Grouping commuting Pauli terms cuts the number of measurement rounds, so the same accuracy should be reachable with fewer shots on real hardware.","The reported $\\mathcal{O}(10^{-2})$ to $\\mathcal{O}(10^{-1})$ error range suggests that near-term quantum chemistry on medium-noise devices is plausible for small molecules."],"supporting_citations":[{"why":"Introduces the VQE framework and supplies the UCCSD baseline that MPS-VQE is compared against.","marker":"[11]"},{"why":"Provides the hardware-efficient VQE baseline (HE-VQE) and the noise-related motivation for compact circuits.","marker":"[13]"},{"why":"Defines the qubit-UCC ansatz used as a comparison baseline in the noisy simulations.","marker":"[22]"},{"why":"Provides the SE ansatz baseline used in the comparison table.","marker":"[23]"},{"why":"Supplies the MPS-VQE simulator concept that the circuit construction and pre-training build on.","marker":"[5]"},{"why":"Supplies the differentiable MPS machinery used for classical pre-training of the circuit parameters.","marker":"[6]"},{"why":"Supplies the quantum simulation framework in which all numerical experiments are run.","marker":"[21]"}],"fun_headline_variants":["MPS-style circuit cuts VQE gate count 30x while beating noise","Pre-training plus zero-noise extrapolation shields VQE from hardware noise","Neural zero-noise extrapolation stabilizes VQE on noisy H4","Shallow MPS ansatz with neural denoising keeps VQE accurate under noise","30x fewer gates: MPS-structured VQE withstands noisy H4 hardware"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire accuracy claim rests on the assumption that a small three-layer neural network, trained on expectation values measured at a few artificially boosted noise levels, can reliably extrapolate the noiseless expectation value, yet the paper specifies neither the noise-scaling factors nor the training details and gives no comparison with standard polynomial extrapolation.","fun_headline_variants_meta":{"raw":{"variants":["MPS-style circuit cuts VQE gate count 30x while beating noise","Pre-training plus zero-noise extrapolation shields VQE from hardware noise","Neural zero-noise extrapolation stabilizes VQE on noisy H4","Shallow MPS ansatz with neural denoising keeps VQE accurate under noise","30x fewer gates: MPS-structured VQE withstands noisy H4 hardware"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000932,"raw_usage":{"total_tokens":4062,"prompt_tokens":1087,"completion_tokens":2975,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":703,"completion_tokens_details":{"reasoning_tokens":2874}},"tokens_in":703,"tokens_out":2975,"duration_ms":21319,"temperature":1.0,"reasoning_tokens":2874,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:22:49.786240+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same H4 simulation under the same depolarizing, thermal-relaxation, and bit-flip noise model, but replace the neural-network fit with a standard polynomial or exponential zero-noise extrapolation on the same folded circuits; if the polynomial-extrapolated energy differs from the reported -2.1490 Hartree by more than the claimed $\\mathcal{O}(10^{-2})$ error, then the neural extrapolator, rather than the circuit or pre-training, is the source of the improvement.","supporting_citations":[{"cited_title":"A variational eigenvalue solver on a photonic quantum processor,","cited_arxiv_id":null,"evidence_quote":"Introduces the VQE framework and supplies the UCCSD baseline that MPS-VQE is compared against."},{"cited_title":"Hardware-efficient variational quantum eigensolver for small molecules and quantum magnets,","cited_arxiv_id":null,"evidence_quote":"Provides the hardware-efficient VQE baseline (HE-VQE) and the noise-related motivation for compact circuits."},{"cited_title":"Efficient quantum circuits for quantum computational chemistry,","cited_arxiv_id":null,"evidence_quote":"Defines the qubit-UCC ansatz used as a comparison baseline in the noisy simulations."},{"cited_title":"Circuit-centric quantum classifiers,","cited_arxiv_id":null,"evidence_quote":"Provides the SE ansatz baseline used in the comparison table."},{"cited_title":"Mps-vqe: A variational quantum computational chemistry simulator with matrix product states,","cited_arxiv_id":null,"evidence_quote":"Supplies the MPS-VQE simulator concept that the circuit construction and pre-training build on."},{"cited_title":"Differentiable matrix product states for simulating variational quantum computational chemistry,","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable MPS machinery used for classical pre-training of the circuit parameters."},{"cited_title":"Mindspore quantum: A user-friendly, high-performance, and ai-compatible quantum computing framework,","cited_arxiv_id":null,"evidence_quote":"Supplies the quantum simulation framework in which all numerical experiments are run."}],"review_version":1}