{"id":"96fe8a63-27e6-4300-a810-849f5676df03","arxiv_id":"2506.09527","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Noise, especially decoherent gate errors, systematically reduces Fourier coefficient magnitudes, expressibility, and entangling capability of quantum Fourier models, with circuit architecture and encoding modulating the effect.","lead":"This paper measures how hardware noise changes the Fourier spectrum, expressibility, and entangling capability of variational quantum circuits used as quantum Fourier models. It finds that decoherent gate errors uniformly shrink Fourier coefficients and training accuracy, while the circuit architecture and input encoding determine how severe the damage is.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Expressibility metric (Sec. 4.3, Eq. 11) is applied to noisy mixed states without defining the fidelity distribution; unless an unstated pure-state sampling convention is used, the expressibility results in Fig. 11 and the Result Summary are unsupported.","rationale":"The reader identified the same weakest link, and I agree: the Fourier coefficient measurements are well-defined for noisy expectation values, and the entangling capability section at least acknowledges an upper-bound interpretation. Expressibility is the only metric where a central quantity, the pure-state fidelity, is used without any stated extension to mixed states. If the implementation uses density matrices and a direct trace formula, the KL values in Fig. 11 do not measure what Sec. 4.3 defines; if it uses stochastic pure-state samples, the metric is a different expressibility notion and the paper should say so. Either way, the current text does not support the expressibility component of the Result Summary. I considered the uniformity claim contradicted by Appendix C, where RX encoding shows increasing high-frequency coefficients for HEA and C19 under decoherent gate errors while RY shows decreases, but that is an overclaim the authors partially explain as a finite-sampling artifact and it is less central than an undefined metric. The verdict stays CONDITIONAL: the paper is a useful benchmark, but the expressibility-on-mixed-states premise must be resolved before the strongest conclusions are taken at face value.","tokens_in":22485,"tokens_out":8759,"duration_ms":102962,"concrete_test":"In the Zenodo reproduction package (doi:10.5281/zenodo.15211317), locate the expressibility computation for noisy circuits and determine whether states are kept as density matrices or sampled as pure states via Kraus-operator sampling. Then rerun Fig. 11 for six qubits (SEA, HEA, C15, C19) with BF, PF, DP at 0-3% noise under the appropriate metric: (a) if density matrices, compute Uhlmann fidelity F(rho_theta1, rho_theta2) = Tr^2(sqrt(sqrt(rho_theta1) rho_theta2 sqrt(rho_theta1))) and recompute the KL divergence to the t=2 Haar distribution; (b) if sampled pure states, state that convention explicitly and verify the pure-state overlap distribution. If the direction or magnitude of the KL shift changes materially, the expressibility conclusions in the Result Summary must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 4.3 defines expressibility via Eq. (11) using the fidelity F = |<psi_phi|psi_phi>|^2 between pure states sampled from a Haar t-design and the model, following Ref. [42]. Sec. 5.2 then applies this same KL divergence to circuits subject to decoherent noise (BF, PF, DP, AD, PD, SP, ME), whose outputs are mixed states. The manuscript nowhere specifies the mixed-state generalization: it does not state whether noisy states are represented as density matrices (requiring, e.g., Uhlmann fidelity or a purification of the t-design) or whether each noisy run is stochastically collapsed to a pure state via Kraus-operator sampling. These two choices define different distributions. For a density matrix rho, the distribution of overlaps of two independently sampled pure states is not the same as the distribution of Uhlmann fidelities between rho and Haar-random pure states, so the KL divergence to P_Haar measures different objects. Without a stated convention, the conclusion that decoherent gate errors reduce expressibility (Fig. 11, Sec. 5.2) is not derivable from the metric as defined. Since expressibility is one of the three properties named in the Result Summary, this ambiguity is load-bearing for the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a numerical study of how various noise models affect three properties of variational quantum circuits used as quantum Fourier models (QFMs): the Fourier spectrum (coefficients), expressibility, and entangling capability. The authors simulate four ansätze (SEA, HEA, C15, C19) with 3–6 qubits, one or two encoding features, and seven noise levels up to 3%, and they also train the QFMs on synthetic regression targets. The central claims are that noise effects are predictable and qualitatively uniform across circuits, and that ansatz structure and input encoding dominate the impact of these properties, including in the noiseless setting. The paper includes a software reproduction package and uses both FFT-based and analytical coefficient computations, with cross-validation.","tokens_in":22617,"tokens_out":8049,"duration_ms":87793,"significance":"If the central claims hold, the work would provide a useful empirical map of how decoherent and coherent noise degrade the Fourier spectrum, expressibility, and entanglement of QFMs, with potential value for ansatz selection and error-mitigation design. The study is broad in its coverage of noise channels and ansätze, and the public code package is a strength. However, several load-bearing methodological gaps currently prevent the results from fully supporting the stated conclusions. In particular, the expressibility metric is applied to mixed states without a defined generalization, the synthetic target definition in Eq. (14) appears internally inconsistent, and the expressibility/entanglement circuits omit the encoding that is central to QFM properties. These issues need to be resolved before the paper's main claims can be accepted.","major_comments":[{"comment":"The expressibility metric is defined via the pure-state overlap F = |<ψ_φ|ψ_ϕ>|^2 and the KL divergence between the resulting fidelity distribution and that of a Haar t-design. In Sec. 5.2 this same metric is applied to circuits subjected to decoherent noise, whose outputs are mixed states. The manuscript nowhere specifies the mixed-state generalization: it does not state whether noisy states are represented as density matrices (requiring, e.g., Uhlmann fidelity or a purification of the Haar states) or whether each noisy run is stochastically collapsed to a pure state via Kraus-operator sampling. These choices define different distributions, so the KL divergence plotted in Fig. 11 is not well-defined as written. Since Fig. 11 is used to conclude that decoherent noise reduces expressibility, and expressibility is one of the three properties named in the Result Summary, this ambiguity is load-bearing. The authors should either define and justify a mixed-state fidelity distribution or restrict the expressibility claim to a clearly stated operational convention.","section":"Sec. 4.3, Eq. (11); Sec. 5.2, Fig. 11"},{"comment":"The target coefficients for the synthetic regression task are defined as c'_ω = a r_ω |∑_{ω∈Ω} r_ω e^{iω^T x}|^{-1}. Because the denominator depends on the input x, c'_ω is not a constant Fourier coefficient; consequently f'(x) is not a Fourier series with the stated spectrum. In fact, substituting this definition into f'(x) yields f'(x) = a times the phase factor of the random sum, whose Fourier coefficients are not the c'_ω. This makes the target a pathological function rather than a random Fourier series with a controlled spectrum. The coefficient-difference metric Δc_ω in Eq. (15) is therefore ill-defined as written, and the training results in Sec. 5.1.5 and Fig. 10 cannot support conclusions about noise-induced degradation of coefficient learning. The authors should correct Eq. (14) (e.g., normalizing by a constant such as ∑|r_ω|) and verify that the implemented code matches the corrected formula, or clearly report the exact target used in the experiments.","section":"Sec. 4.5, Eq. (14)"},{"comment":"For the expressibility and entanglement measurements, the authors state that they 'discard encoding gates and the second trainable layer to make the results more consistent with Ref. [42]'. Thus the circuits analyzed in Figs. 11–13 do not include the data-encoding unitaries that are essential to the QFM representation. This creates a mismatch: the Result Summary claims that 'the structure of a VQC and the input encoding have a crucial impact on these properties', but the expressibility experiments never vary the input encoding, and the entanglement experiments likewise omit it. At minimum, the claim should be restricted to the trainable ansatz structure, or separate experiments should be performed to test the influence of the encoding on expressibility and entanglement. As written, the connection between these metrics and the QFM properties used in the coefficient and training experiments is indirect.","section":"Sec. 5 (third paragraph); Sec. 5.2"},{"comment":"The paper reports that the Entanglement of Formation (EF) measure gives different values from the Meyer-Wallach (MW) measure even for noiseless circuits, and attributes this to non-uniqueness of the eigendecomposition used in the EF computation. For a pure state, the eigendecomposition of the density matrix is unique up to a global phase, and the MW measure is invariant under local unitary transformations; the explanation as written is therefore not convincing. The discrepancy might indicate a numerical artifact (e.g., mixing of nearly degenerate eigenvalues) or a different computation than described. Since EF is used to support the central claim that noise reduces entangling capability, the authors should validate the zero-noise limit of EF against the MW measure and clarify the computation, or explicitly restrict EF claims to mixed-state regimes where the non-uniqueness genuinely matters.","section":"Sec. 4.4.2; Fig. 13"}],"minor_comments":[{"comment":"The text says expressibility is the inverse of the KL divergence, but Fig. 11 plots the KL divergence itself with an axis label showing 'more expr.← KL-Divergence [log] → less expr.'; please align the terminology with the plotted quantity.","section":"Sec. 4.3; Fig. 11"},{"comment":"The claim that decoherent gate errors cause an 'exponential decay' of the coefficient mean is based on visual inspection across only seven noise levels up to 3%; consider fitting a decay law or stating this as a qualitative trend rather than a functional form.","section":"Sec. 5.1.3, Fig. 6"},{"comment":"The statement that 'the effects of each noise type are uniform throughout all tested circuits' is qualified by the reported exception for the SEA with bit-flip noise, where σ_c(ω=0) increases with noise level; please make this exception explicit in the summary.","section":"Result Summary; Sec. 5.1.3"},{"comment":"The sentence 'In a noiseless environment, the SEA achieves a higher expressibility compared to the other ansätze. However, when coherent noise is applied, the expressibility is quickly equalised' appears to refer to decoherent noise, which contradicts the earlier statement that coherent noise has no measurable effect on expressibility; please correct this apparent wording error.","section":"Sec. 5.2, final paragraph"},{"comment":"The y-axis label '# Qubits – Limit' is unclear; please define what this limit represents, for instance in relation to the maximum possible number of frequencies for the given encoding.","section":"Fig. 8 caption"}],"recommendation":"major_revision","confidential_remarks":"The two most serious issues are (i) the undefined mixed-state expressibility metric and (ii) the internally inconsistent target-coefficient definition in Eq. (14), which affects the entire training analysis. If the code implements a different (correct) formula for the target, the authors should state this explicitly; otherwise the training results need to be redone. The expressibility issue is fixable by adding a precise definition, but it is central to the paper's claims. I would suggest that the editor require the authors to clarify these points before further consideration; the paper's extensive numerical sweep and open code are valuable, but the current presentation does not yet support the headline conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the noise-effects QFM paper. Bottom line: it is a useful, honestly-scoped numerical benchmark, and the core qualitative finding—decoherent gate errors contract the Fourier coefficients roughly exponentially and degrade training—holds up across seeds and ansätze. The thing I'd want you to know before citing it: the expressibility results under decoherent noise are built on a metric that is only defined for pure states, and the paper never says how it handled mixed states.\n\nWhat's actually new is the breadth. Four ansätze, eight noise types, three properties, plus training. No prior work I know of benchmarks Fourier coefficients, expressibility, and entangling capability side by side under noise. The observation that coherent gate errors on the encoding gates generate non-integer frequency components and inflate the apparent spectrum without making those frequencies tunable is genuinely new and the most interesting part. The paper ships code and data (Zenodo, QML-Essentials), uses direct simulation with no fitted parameters, and is transparent about the small circuit sizes. Credit where due: the coefficient decay, MSE degradation, and entanglement decrease under decoherent errors are consistently observed, and the HEA's missing high-frequency spectrum is a concrete, useful observation.\n\nSoft spots. The expressibility metric (Sec. 4.3) is defined via the fidelity distribution of pure states sampled from the model and a Haar t-design. Sec. 5.2 applies that same KL divergence to circuits with BF, PF, DP, AD, PD, SP, ME—all of which output mixed states. The manuscript nowhere states whether it purifies, collapses via Kraus sampling, or uses a mixed-state fidelity. Those choices give different distributions, so Fig. 11 and the claim that decoherent noise reduces expressibility are not actually derivable from the metric as written. This is a real gap, and it is load-bearing for one of the three properties named in the Result Summary. It is not fatal to the whole paper, because the Fourier-coefficient and training results stand on direct measurements, but the summary overreaches.\n\nTwo smaller things. The 'exponential' decay is asserted, never fitted or quantified. And the Result Summary says noise effects 'can be predicted,' but the paper offers empirical regularities, not a predictive model. Tone those down.\n\nWho it's for: QML practitioners choosing ansätze and encodings under noise, and researchers working on QFM spectral properties. It deserves a serious referee; the expressibility gap should be fixed (or its scope explicitly narrowed) before the strongest conclusions are taken at face value. I'd accept it for review and push for that revision.","headline":"A broad, honestly-scoped numerical benchmark of noise effects on quantum Fourier models; the coefficient and training results are credible, but the expressibility metric is applied to mixed states without a stated convention, which is a load-bearing gap.","tokens_in":23264,"tokens_out":2898,"would_cite":true,"duration_ms":31610,"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":"The paper claims that noise effects on quantum Fourier models—on their Fourier spectrum, expressibility, and entangling capability—are predictable per noise type, uniform across the tested circuits, and shaped more by circuit architecture…","keywords":["quantum machine learning","quantum Fourier models","variational quantum circuits","noise modelling","Fourier spectrum","expressibility","entangling capability","error mitigation"],"falsifier":"Compute expressibility for the same noisy circuits using a mixed-state fidelity (for example the Uhlmann fidelity or a purification-based overlap) instead of the pure-state overlap the paper uses, and compare the resulting KL divergences to the uniform-state distribution; if the ranking of ansätze or the direction of noise dependence changes, the expressibility findings are an artefact of the metric rather than a property of the circuits.","tokens_in":1613,"feed_emoji":"⚛️","tokens_out":1675,"duration_ms":79097,"temperature":0.7,"pith_summary":"The paper tries to establish that the influence of hardware noise on quantum Fourier models can be anticipated: for a given circuit ansatz and input encoding, noise affects the Fourier spectrum, expressibility, and entangling capability in systematic ways, with each noise type acting uniformly across all tested circuits. If true, this means noise-induced degradation of a quantum model can be predicted before running noisy hardware, and architectural choices can be selected for robustness. The authors support the claim with numerical simulations on circuits of three to six qubits, four standard ansätze, seven noise types, and noise levels up to three percent, linking these spectral and entanglement metrics to training performance on synthetic regression problems. A sympathetic reader would take the core message as: the structure of the variational circuit and the way data is encoded matter at least as much as the noise level for what a noisy quantum model can learn.","feed_headline":"Noise damage to quantum Fourier models is predictable","feed_subtitle":"Architecture and input encoding, not just noise level, set which Fourier features survive.","key_machinery":"The central object is the representation of a variational quantum circuit as a quantum Fourier model: the observable expectation value is written as a truncated Fourier series $f(x,\\theta)=\\sum_{\\omega\\in\\Omega} c_\\omega(\\theta)e^{i\\omega^T x}$, where the frequency set $\\Omega$ is determined by differences of eigenvalues of the Hermitian generators in the input-encoding unitaries, and the coefficients $c_\\omega(\\theta)$ depend on the trainable parameters and circuit structure. This Fourier lens turns questions about noise into questions about how noise alters coefficients and frequencies. The paper measures coefficients numerically with the fast Fourier transform after cross-validating against analytical expansions, quantifies expressibility as the inverse KL divergence between the model's fidelity distribution and the uniform distribution over quantum states, and quantifies entangling capability with a pure-state global-entanglement measure for noiseless circuits and a mixed-state entanglement measure for noisy ones. Noise is modelled with standard channels: bit flip, phase flip, depolarisation, amplitude damping, phase damping, state-preparation and measurement errors, and coherent gate errors. The mechanism behind the coherent-error result is that an error $\\epsilon_x$ on an encoding gate shifts the frequency exponent from $\\omega^T x$ to $(\\omega+\\epsilon_x)^T x$, which produces additional non-integer frequencies and reduces the redundancy of the spectrum.","core_discovery":"The central claim, stated as the paper's Result Summary, is that the influence of noise on properties of quantum Fourier models—particularly the Fourier spectrum, expressibility, and entangling capability—can be predicted, that the effects of each noise type are uniform throughout all tested circuits, and that the structure of the variational quantum circuit and the input encoding have a crucial impact on these properties independently of the noise type, also in a noiseless setting. Concretely, decoherent gate errors such as bit flip, phase flip, and depolarisation exponentially damp the mean magnitude of Fourier coefficients across all four ansätze, while SPAM and damping noise produce weaker or no observable contraction. Expressibility, measured by how far the circuit's state-overlap distribution sits from the uniform distribution over quantum states, drops under decoherent gate errors and more mildly under SPAM and damping noise, while coherent gate errors leave it essentially unchanged. Entangling capability falls most clearly under amplitude damping. In training on synthetic Fourier series, decoherent gate errors push all ansätze to a similar elevated mean-squared error, whereas the heavily entangled ansatz that achieves a full noiseless spectrum trains best without noise and the ansätze with incomplete spectra fail to reach the optimum even noiselessly. The paper also finds that coherent gate error acting on the encoding gates changes the frequency set itself, adding new non-integer frequencies whose coefficients are not individually tunable.","pith_inferences":["If the reported uniformity of noise-type effects holds at larger qubit counts and depths, noise-aware architecture search could rank candidate ansätze by their predicted spectral robustness before any hardware deployment; the paper does not itself propose such a search.","The expressibility findings rest on applying a pure-state fidelity metric to mixed states, so a mixed-state formulation of the metric could change quantitative conclusions while perhaps preserving the qualitative ordering of ansätze; this is an open question the paper does not address.","Because coherent gate errors add non-integer frequencies that are not individually tunable, a testable extension would be to check whether deliberately introduced coherent shifts on encoding gates can be exploited as a controlled spectral-design tool, for example to reach off-grid frequencies helpful for a specific learning task.","A natural next experiment would be to measure the same spectrum, expressibility, and entangling-capability quantities on real quantum hardware under calibrated noise, to see whether the simulated uniform patterns survive device-specific crosstalk and drift."],"forward_implications":["For a chosen ansatz and input encoding, the severity and type of noise-induced degradation can be anticipated without running the noisy circuit: which coefficients decay, how expressibility drops, and how much entanglement is lost are all predictable from the architecture and the noise channel.","Decoherent gate errors are the dominant threat to quantum Fourier models, exponentially shrinking Fourier coefficient magnitudes, raising training error, and reducing expressibility, whereas SPAM and damping errors have milder effects.","Architecture choice matters even in a noiseless setting: some ansätze have full Fourier spectra and train well but lose expressibility quickly under noise, while others lack full spectra entirely and therefore cannot learn certain functions regardless of noise.","Coherent gate errors on encoding gates are qualitatively different from other noise because they alter the frequency set itself; error mitigation must therefore target encoding gates, not only trainable layers, to preserve the intended spectrum.","The observed patterns suggest that noise-aware ansatz selection and tailored error mitigation can be informed by spectral properties measured on a few small circuits, since the uniform behaviour across tested circuits indicates broader applicability beyond the limited sizes studied."],"supporting_citations":[],"fun_headline_variants":["Quantum Fourier noise damage is predictable and design-dependent","Noise effects on quantum Fourier models vary with architecture","Fourier model noise impact hinges on circuit and encoding","Decoherent errors uniformly damp quantum Fourier coefficients","Structure, not just noise, controls quantum Fourier resilience"],"cache_read_input_tokens":25344,"weakest_assumption_plain":"The expressibility conclusions for noisy circuits rest on applying a fidelity-based overlap measure designed for pure quantum states to the mixed states that noise actually produces, and the paper supplies no mixed-state generalization; if that step is invalid, the reported noise-induced loss of expressibility is unsupported.","fun_headline_variants_meta":{"raw":{"variants":["Quantum Fourier noise damage is predictable and design-dependent","Noise effects on quantum Fourier models vary with architecture","Fourier model noise impact hinges on circuit and encoding","Decoherent errors uniformly damp quantum Fourier coefficients","Structure, not just noise, controls quantum Fourier resilience"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1615,"prompt_tokens":1071,"completion_tokens":544,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":687,"completion_tokens_details":{"reasoning_tokens":469}},"tokens_in":687,"tokens_out":544,"duration_ms":5967,"temperature":1.0,"reasoning_tokens":469,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:47:32.278474+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute expressibility for the same noisy circuits using a mixed-state fidelity (for example the Uhlmann fidelity or a purification-based overlap) instead of the pure-state overlap the paper uses, and compare the resulting KL divergences to the uniform-state distribution; if the ranking of ansätze or the direction of noise dependence changes, the expressibility findings are an artefact of the metric rather than a property of the circuits.","supporting_citations":[],"review_version":1}