{"id":"694bcb3c-b68e-4465-8f65-2000556898d2","arxiv_id":"1909.02196","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"For dephasing, bit-flip, and depolarizing noise on a 7-qubit Max-Cut QAOA, fidelity, cost, and gradients decay like (1-p)^(αN), and fitted optimal parameters stay close to noiseless values for Np<0.5.","lead":"This paper studies how three common types of quantum noise affect the quantum approximate optimization algorithm (QAOA). It reports that noise shrinks the QAOA cost landscape by an exponential factor without moving the best parameters, based on numerical simulations of a small Max-Cut problem.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central claim that noise merely rescales the QAOA cost landscape rests on Eq. (15), which is fitted at one parameter point and never checked across parameter space; the manuscript's own optimization test also excludes all points with Np≥0.5.","rationale":"The reader correctly identifies Eq. (15) as the linchpin. My reading of Secs. III B–III C and IV B–IV D confirms that the multiplicative-rescaling law is assumed, not derived; the only quantitative support is a fit at one parameter point, and Fig. 6 excludes Np≥0.5. The exponential-decay statements for fidelity and cost at fixed parameters are less problematic—for small p they are plausible single-instance numerics—but the parameter-invariance claim, which is the headline practical conclusion, depends entirely on the unverified rescaling. A grid test of R(γ,β) would settle it. Credit is due where the paper does report numerical data with a concrete error analysis and identifies statistical-error limitations; the issue is not the individual data points but the unjustified generalization from one point to the whole landscape. If the grid test confirms constant R, the central claim would be substantially supported; until then it is unverified. I therefore agree with the reader's rejection, with the same principal concern, and recommend no change from their verdict.","tokens_in":10320,"tokens_out":6322,"duration_ms":68417,"concrete_test":"Perform an exact density-matrix simulation of the same 7-qubit Max-Cut QAOA with n=4 and, say, dephasing and depolarizing noise at p=0.005; compute R(γ,β)=f_noise(γ,β)/f_ideal(γ,β) on a grid or random sample of at least 10^4 points covering the full parameter range. If R varies by more than the statistical error (M≥5000, using the 95% confidence intervals from Sec. V B), Eq. (15) is falsified and optimized parameters can shift under noise. As a second arm, rerun the Sec. IV D optimization with p large enough that Np≥0.5 and report the distance defined in Eq. (23).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assertion—that optimized QAOA parameters do not move under noise—requires Eq. (15): f_noise(γ,β) = (1−p)^(αN) f_ideal(γ,β) with a parameter-independent α and with A=0. This is not derived from Eq. (14). Sec. III B posits it from the 'small k' branches, and Sec. IV B fits α using y=f_noise/f_ideal measured at a single parameter setting, the ideal optimum, for one 7-qubit Max-Cut instance. That only tests one point in parameter space. The A=0 step is also unjustified: H_p is traceless, but the states |ψ_k⟩ are noise-evolved versions of one fixed state, not Haar-random states, so their weighted average cost need not vanish. Sec. III C then applies the same scalar factor to every gradient, and Sec. IV D concludes that optimization paths are unchanged, while the Fig. 6 caption restricts the comparison to points satisfying Np<0.5, excluding exactly the high-noise regime where optima could shift. Without a parameter-space check that f_noise/f_ideal is constant, or a derivation from the Kraus expansion, the 'noise only flattens' conclusion is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the effect of three quantum noise channels (dephasing, bit-flip, depolarizing) on the quantum approximate optimization algorithm (QAOA) applied to a 7-qubit Max-Cut instance. The central claims are that the output-state fidelity scales as (1-p)^(δN), the cost function scales as (1-p)^(αN) f_ideal, the gradient scales by the same factor, and as a consequence quantum noise merely flattens the QAOA parameter landscape without shifting the optimal parameters. These claims are supported by numerical simulations with a Monte-Carlo noisy quantum virtual machine, together with a statistical error analysis for the cost and gradient estimators.","tokens_in":10672,"tokens_out":3741,"duration_ms":39484,"significance":"If the central claim were rigorously established, the paper would provide a useful and simple characterization of noise resilience in QAOA, with practical implications for parameter optimization on NISQ devices. The numerical work and the statistical error analysis in Sec. V B are careful and are a genuine strength. However, the central scaling law is introduced as an ansatz rather than derived from the Kraus expansion, and the evidence is limited to a single parameter point and a single problem instance. As it stands, the paper offers a plausibility observation rather than a proven scaling theorem, and the strong conclusion that optimized parameters are unaffected by noise is not supported by the presented data.","major_comments":[{"comment":"The central scaling law f_noise = (1-p)^(αN) f_ideal is asserted, not derived. The text states that for small k the expectation of H_p is close to the ideal value and then \"we use\" the exponential form, with α a free constant. No derivation from the exact Kraus expansion in Eq. (14) is provided. Since Eq. (15) is the foundation for the gradient scaling in Eq. (18) and for the parameter-invariance conclusion in Sec. IV D, the main claim of the paper is currently an unproven ansatz.","section":"Sec. III B, Eq. (15)"},{"comment":"The identification of A with the global average Tr(H_p)/2^m = 0 is unjustified. While H_p is traceless, the states |ψ_k⟩ in the noise expansion are not Haar-random states; they are obtained by applying a fixed set of Kraus operators to one particular product state. The weighted average of ⟨ψ_k|H_p|ψ_k⟩ over the large-k branches need not vanish, so the A=0 simplification can introduce a systematic error in the fitted scaling. This issue is load-bearing because if A is nonzero, Eq. (15) becomes a shift-plus-rescale rather than a pure flattening, and the conclusion that the optimum is unchanged would no longer follow.","section":"Sec. III B, Eqs. (12)-(15)"},{"comment":"The exponent α is fitted from data at a single parameter setting, namely the ideal optimum. The paper does not test whether f_noise/f_ideal is constant across parameter space, which is exactly what Eq. (15) requires. Figures 4 and 5 only show the ratio at that one point. Without a parameter-space scan (e.g., random or grid-sampled γ,β) or an analytic derivation, the claim that noise merely rescales the whole landscape is unsupported.","section":"Sec. IV B, Eqs. (19)-(20)"},{"comment":"The optimization comparison is restricted to points satisfying Np<0.5, as stated in the caption. This restriction excludes the high-noise regime where the optimum would be most likely to shift if the simple rescaling picture breaks down. The text does not discuss this restriction or its implications, yet the conclusion that noisy optimized parameters equal ideal optimized parameters is based on this filtered data. The presented evidence is therefore insufficient for the strong claim made in the abstract and conclusion.","section":"Sec. IV D, Fig. 6"},{"comment":"The gradient scaling factor is inherited from the cost-function ansatz. However, the gradient formulas in Eqs. (16)-(17) involve cost functions evaluated at shifted parameters (γ_k ± π/2 and β_k ± π/2). Even if Eq. (15) held at the nominal point, it would need to hold at all shifted points for Eq. (18) to follow. This is not verified, and the assertion that α is nearly the same for different θ_k is made without numerical evidence in the manuscript.","section":"Sec. III C, Eq. (18)"}],"minor_comments":[{"comment":"The word \"attens\" appears in conclusion (1), which should read \"flattens\".","section":"Conclusion, Sec. V"},{"comment":"\"Sharma at al\" should be \"Sharma et al.\"","section":"Introduction, Ref. [21]"},{"comment":"The caption states that panels (a) and (b) correspond to dephasing, (c) and (d) to bit-flip, and (e) and (f) to depolarizing noise, but the figure panels are arranged by derivative type (∂f/∂γ and ∂f/∂β) for each noise channel; the caption appears inconsistent with the panel labels.","section":"Fig. 5 caption"},{"comment":"The formula for L_costfunction contains a typographical artifact: \"2 vuu√\" should be a single square root symbol.","section":"Appendix B 1, Eq. (34)"},{"comment":"The statement that the fidelity data \"fit well\" with Eq. (9) is not quantified; reporting the fitted δ values and goodness-of-fit measures for all channels and steps would strengthen the presentation.","section":"Sec. IV A"}],"recommendation":"reject","confidential_remarks":"The paper's central claim is an overreach relative to the evidence: the scaling law is an ansatz fitted at one point, and the key conclusion about parameter invariance is tested only in the restricted regime Np<0.5. The conceptual issue with A=0 is not a minor technicality but affects the validity of the main result. A future revision could potentially recast the work as a numerical observation for small noise, but the current version does not support the stated conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the numerical core is honest and the error analysis is more careful than most, but the main conclusion—that QAOA's optimized parameters are unchanged by noise—is not supported by the experiments as reported. The paper should not be cited as evidence for that conclusion until the scaling law is derived or tested across the parameter space.\n\nWhat's new: they study dephasing, bit-flip, and depolarizing channels on a small QAOA Max-Cut instance, and report that output fidelity, cost, and gradient ratios collapse onto (1-p)^(αN) quite well for n=1-4. The appendix gives a proper treatment of measurement shot noise with 95% confidence intervals, which is a genuine strength. If the goal is just to document that shallow QAOA circuits experience exponential decay in these quantities, the data are fine.\n\nThe soft spot is load-bearing. Eq. (15) is not derived from the Kraus expansion; it is an ansatz that says the noisy cost is a rescaled ideal cost plus a term A. They set A=0 because H_p is traceless, but the average over the noise-affected states is not the Haar average—those states are all evolved from one fixed initial state, so their cost expectation need not vanish. The exponent α is fitted to data taken at the ideal optimal parameter setting for one 7-qubit graph, then assumed to be a global constant. The gradient scaling inherits that α, and Fig. 5 only tests the point of largest gradient. Finally, the optimization distance plot in Fig. 6 excludes all points with Np≥0.5, which is exactly the high-noise region where the optimum could shift. So the 'noise merely flattens' conclusion is an extrapolation from one corner of parameter space, not a demonstrated fact.\n\nThat said, I don't think the authors are being sneaky. They label Eq. (15) as a representation and say they 'infer' the fidelity law. The problem is the abstract and conclusion state the strong version without the caveats. A careful revision that restricts the claims to the tested regime, or derives the scaling, could make this a useful paper for the NISQ-algorithm community.\n\nWho this is for: readers working on noise-resilience of variational algorithms, or benchmarking QAOA. It's a reasonable desk-reject as is, but if the editor believes the authors can tighten the claims, a referee would find the statistical work worth engaging with. My vote: send it to review, but brace for a major-revision request.","headline":"A QAOA noise study with solid numerical data and a worrying gap between the plotted exponentials and the 'noise only flattens' claim in the abstract.","tokens_in":11150,"tokens_out":3636,"would_cite":false,"duration_ms":38770,"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":"Under common quantum noise channels, QAOA's output fidelity, cost function, and gradient all decay exponentially with circuit size and noise strength, while the parameter landscape is only rescaled, so optimized parameters do not shift.","keywords":["quantum approximate optimization algorithm","quantum noise","NISQ devices","parameter landscape flattening","exponential fidelity decay","Max-Cut","variational quantum algorithms"],"falsifier":"Take a fixed QAOA instance and a fixed noise channel; measure the ratio $f_{\\mathrm{noise}}(\\vec{\\gamma},\\vec{\\beta})/f_{\\mathrm{ideal}}(\\vec{\\gamma},\\vec{\\beta})$ at many different parameter points, including points away from the optimum. If this ratio is not constant across the landscape, or if it departs from $(1-p)^{\\alpha N}$ by more than sampling error, the claim that noise only rescales the parameter space is refuted. Equivalently, use a target Hamiltonian with $\\mathrm{Tr}(H_p)\\neq 0$ so $A\\neq 0$ and optimize under noise: if the optimal parameters shift compared to the ideal run, the zero-$A$ assumption is doing load-bearing work.","tokens_in":10123,"feed_emoji":"📉","tokens_out":4920,"duration_ms":48640,"temperature":0.7,"pith_summary":"This paper studies what happens to the Quantum Approximate Optimization Algorithm (QAOA) when it runs on a noisy quantum device. For three standard noise channels—dephasing, bit-flip, and depolarizing—it claims that the output-state fidelity, the cost function, and the cost-function gradient all shrink exponentially as the number of gates and the noise strength grow. The central structural claim is stronger: noise does not distort the optimization landscape, it only multiplies it by a single scale factor $(1-p)^{\\alpha N}$. Because a uniform rescaling preserves the location of minima, the paper concludes that the optimal QAOA parameters found without noise remain optimal under noise, so the parameter optimization step can be carried out as if the device were ideal, just more slowly.","feed_headline":"Noise flattens QAOA's landscape without moving its optima","feed_subtitle":"Simulations show QAOA cost, fidelity, and gradient decay exponentially with noise, while optimal parameters stay put.","key_machinery":"The load-bearing object is the rescaling ansatz in Eq. (15): $f_{\\mathrm{noise}}(\\vec{\\gamma},\\vec{\\beta})=(1-p)^{\\alpha N} f_{\\mathrm{ideal}}(\\vec{\\gamma},\\vec{\\beta}) + (1-(1-p)^{\\alpha N})A$, with $A=0$ for the redefined Hamiltonian. The same exponential factor $ (1-p)^{\\alpha N}$ is then carried linearly into the parameter-shift expressions for the gradient, making every partial derivative shrink by the same factor. The constant $\\alpha$ is said to depend on the quantum circuit architecture and the noise model and is obtained by fitting simulation data rather than derived from the Kraus expansion.","core_discovery":"The paper's central claim is that the noisy cost function takes the form $f_{\\mathrm{noise}}(\\vec{\\gamma},\\vec{\\beta})=(1-p)^{\\alpha N} f_{\\mathrm{ideal}}(\\vec{\\gamma},\\vec{\\beta})$ because the average cost term $A=\\mathrm{Tr}(H_p)/2^m$ equals zero for the Max-Cut Hamiltonian $H_p=\\sum_{i,j} C_{ij}Z_iZ_j$. It claims the same exponential factor rescales the cost-function gradient, and that output-state fidelity behaves as $F=(1-p)^{\\delta N}$. Consequently, noise merely flattens the QAOA parameter space without changing its structure, so optimized parameters will not deviate from their ideal values. Numerical fits on a Max-Cut problem with QAOA step numbers $n=1$ through $4$ and noise strengths in $[0.0001,0.02]$ are presented as support.","pith_inferences":["Beyond the paper: the same rescaling logic would predict that if a variational algorithm's Hamiltonian has a nonzero uniform average ($A\\neq 0$), noise would shift the optimum rather than merely flatten it; that case is untested here.","Beyond the paper: the fitted exponent $\\alpha$ may itself grow with noise strength or depth in practice, which would make the simple $(1-p)^{\\alpha N}$ law an approximation valid only at weak noise.","Beyond the paper: a direct test on larger Max-Cut graphs or with correlated noise models would show whether a single fitted $\\alpha$ remains universal enough to be predictive across problem sizes."],"forward_implications":["QAOA parameter search can reuse noise-free optimized parameters on a noisy device, saving classical optimization cost.","Larger QAOA layer number $n$ increases gate count $N$ and hence the noise decay $(1-p)^{\\alpha N}$; beyond a noise-dependent depth, added layers stop helping.","Gradient-based and gradient-free optimizers should follow approximately the same trajectory under noise, only slower.","The same exponential decay implies output fidelity degrades predictably with circuit depth, enabling a pre-silicon noise budget for QAOA instances.","The conclusion extends, as the paper notes, to multi-layer parameterized quantum circuits beyond QAOA."],"supporting_citations":[{"why":"Defines the QAOA circuit and cost function that the paper analyzes.","marker":"[25]"},{"why":"Studied noise effects on variational quantum compiling and introduced optimal parameter resilience, the comparison point for this paper's noise-tolerance claim.","marker":"[21]"},{"why":"Supplies the parameter-shift method used to compute cost-function derivatives.","marker":"[30]"},{"why":"Provides the noisy quantum simulator used for all numerical demonstrations.","marker":"[31]"},{"why":"Supplies the Monte Carlo method for simulating mixed-state evolution under noise.","marker":"[32]"}],"fun_headline_variants":["Noise flattens QAOA landscape, leaves optima unchanged","QAOA noise: costs decay, optima don't budge","Quantum noise shrinks QAOA cost, keeps parameters fixed","Exponential noise decay, but QAOA optima stay put","Noise only scales QAOA cost, structure preserved"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The formula $f_{\\mathrm{noise}}=(1-p)^{\\alpha N} f_{\\mathrm{ideal}}$ with a single fitted exponent $\\alpha$ and zero average term is assumed, not derived; if the actual scaling varies with the parameter or the constant $A$ is not negligible, the flat-landscape conclusion collapses.","fun_headline_variants_meta":{"raw":{"variants":["Noise flattens QAOA landscape, leaves optima unchanged","QAOA noise: costs decay, optima don't budge","Quantum noise shrinks QAOA cost, keeps parameters fixed","Exponential noise decay, but QAOA optima stay put","Noise only scales QAOA cost, structure preserved"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1314,"prompt_tokens":858,"completion_tokens":456,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":371}},"tokens_in":474,"tokens_out":456,"duration_ms":5709,"temperature":1.0,"reasoning_tokens":371,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:57:15.582179+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed QAOA instance and a fixed noise channel; measure the ratio $f_{\\mathrm{noise}}(\\vec{\\gamma},\\vec{\\beta})/f_{\\mathrm{ideal}}(\\vec{\\gamma},\\vec{\\beta})$ at many different parameter points, including points away from the optimum. If this ratio is not constant across the landscape, or if it departs from $(1-p)^{\\alpha N}$ by more than sampling error, the claim that noise only rescales the parameter space is refuted. Equivalently, use a target Hamiltonian with $\\mathrm{Tr}(H_p)\\neq 0$ so $A\\neq 0$ and optimize under noise: if the optimal parameters shift compared to the ideal run, the zero-$A$ assumption is doing load-bearing work.","supporting_citations":[{"cited_title":"Benedetti, D","cited_arxiv_id":null,"evidence_quote":"Defines the QAOA circuit and cost function that the paper analyzes."}],"review_version":1}