{"id":"957df284-389c-4f56-a605-c261960ea911","arxiv_id":"2502.04277","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Non-variational QAOA with fixed angles solves QRAO's relaxed MaxCut Hamiltonian with performance close to optimized parameters and about three times fewer qubits than standard QAOA.","lead":"This paper tests a quantum optimization shortcut that encodes three problem variables into each qubit and uses one fixed set of circuit angles, so no per-instance parameter training is needed. It finds the shortcut works reasonably on MaxCut instances, using about three times fewer qubits than standard QAOA but giving worse solutions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fixed-parameter universality rests on a small, single-family fit with undisclosed angles; mixer comparison is confounded by shared parameters.","rationale":"The reader's weakest assumption—parameter concentration for the relaxed Hamiltonian—is indeed the most load-bearing condition for the central claim. I agree that the evidence is limited to one graph family and small sizes, and that the paper honestly acknowledges the need for future validation. However, I sharpen this with two additional, concrete problems. First, the training-set size is ambiguous: Section III-B says 30 instances, while Fig. 4's caption reports 120 instances for the same data; this undermines confidence in the averaged parameters and needs clarification. Second, the fixed parameter values are not given anywhere in the paper, and code is only 'available upon reasonable request,' so no independent verification of the key numerical claim is possible without direct communication with the authors. Third, the mixer selection is methodologically flawed: all mixers are benchmarked in Fig. 3 using the same fixed parameters that were optimized for the Z mixer, so the conclusion that the Z mixer performs best is not supported by a fair comparison. These issues do not prove the central claim false, but they make the evidence weaker than the text suggests. The paper is otherwise honest and clearly written; it does not overclaim quantum advantage and explicitly frames the result as a resource trade-off. Given that the reader already assigned CONDITIONAL, my concern reinforces that verdict without moving it: the paper should either provide the missing data and fair mixer comparison, or the claims should be narrowed to the empirical setting actually tested.","tokens_in":13581,"tokens_out":9286,"duration_ms":89920,"concrete_test":"The authors should publish the exact fixed parameter schedules (β_l, γ_l for p=1..6) and the QRAC mapping heuristic. An independent group should then: (1) reproduce Fig. 5A on 3-regular graphs with N=10–16 using the published parameters; (2) evaluate the same fixed parameters on 4-regular and Erdős–Rényi MaxCut instances with N=50–100, comparing αc against per-instance optimized parameters; and (3) rerun the mixer comparison of Fig. 3 with parameters optimized separately for each mixer. If (1) fails to match the reported performance, or (2) yields a performance ratio below 0.9 on any family, or (3) reverses the mixer ranking, then the universal fixed-parameter claim or the chosen design strategy is not validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that instance-independent fixed parameters make QRAO non-variational—depends entirely on the parameter concentration assumption for the relaxed, non-diagonal Hamiltonian. In Section III-B, the fixed parameters are defined as the average of optimized angles over 30 randomly generated 3-regular MaxCut instances with N=10–16 (the text states 30, but Fig. 4's caption says 120, an inconsistency that must be resolved). This average is then tested only on the same graph family (up to N=26) and on a (2,1)-QRAC appendix using the same family. No theoretical argument is given for why concentration should hold for non-diagonal Hamiltonians, where the objective landscape and the role of Trotterization differ from standard QAOA. Moreover, the exact fixed parameter values are not disclosed, and code/data are not shipped, so the central quantitative claim is not independently reproducible. In addition, the mixer comparison in Section III-A/Fig. 3 evaluates X, Y, and Z mixers all with the same fixed parameters that were optimized for the Z mixer, so the reported superiority of the Z mixer is confounded. If parameter concentration fails outside the training distribution, or if the undisclosed parameters cannot be reproduced, the paper's headline finding collapses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes QAOA-for-QRAO, a fixed-parameter (non-variational) approach to Quantum Random Access Optimization, and benchmarks it for MaxCut on 3-regular graphs. The authors compare several mixers and initial states, define a parameter-setting protocol that averages optimized angles over training instances, approximate the non-diagonal cost Hamiltonian with (grouped) Trotterization, and compare the resulting approximation ratios and qubit counts with standard QAOA in noiseless and noisy simulations. The headline claims are that instance-independent fixed parameters achieve good performance without per-instance variational training and that the QRAO encoding uses up to three times fewer qubits than standard QAOA.","tokens_in":13776,"tokens_out":14574,"duration_ms":139591,"significance":"If the parameter-concentration behavior holds beyond the tested family, this is a useful step toward executing QRAO on early fault-tolerant devices without variational optimization overhead, with a clear resource-vs-quality trade-off. The paper is honest in acknowledging that standard QAOA gives better classical approximation ratios, and the observation that only a few Trotter steps suffice is of independent interest. However, the evidence is currently limited to random 3-regular MaxCut instances, and the exact fixed parameters are not disclosed, which limits near-term reproducibility.","major_comments":[{"comment":"The text in Section III-B states that the fixed parameters are obtained from 30 randomly generated instances, while Fig. 4 (left) and several other figure captions (Figs. 3, 5, 6, 8) report 120 random instances. This inconsistency must be resolved. Please clarify whether the optimization/training set contains 30 or 120 instances, and whether the evaluation in Fig. 5 is performed on the same instances used for training or on a disjoint held-out set.","section":"III-B and Fig. 4"},{"comment":"The exact fixed parameter values (gamma_l, beta_l for p=1 through 6) are never tabulated. Since the central claim is that these fixed parameters perform well without per-instance optimization, the omission prevents independent verification. The authors should include a table of the fixed angles (or release code/data) so that the numerical results can be reproduced.","section":"III-B"},{"comment":"The mixer comparison in Fig. 3 uses the parameter set \"as described in Section III-B\". Because those fixed parameters are obtained by averaging optimized parameters under a particular mixer (apparently the Z mixer, given the later design choice), the comparison is confounded: the X and Y mixers are evaluated with parameters that were not optimized for them. To support the claim that the Z mixer is the best choice, the authors should either optimize parameters separately for each mixer or explicitly frame Fig. 3 as a comparison within their specific fixed-parameter protocol rather than a general mixer comparison.","section":"III-A, Fig. 3"},{"comment":"The claim that optimized parameters are \"spread in a small range\" is supported only by visual inspection of a scatter plot at p=1. No quantitative measure (variance, interquartile range, or similar) is provided for p=1 through 6. Moreover, the fixed parameters are tested only on random 3-regular MaxCut graphs (up to N=26) and on the same family in the (2,1)-QRAC appendix. This validates the protocol for this specific distribution but does not substantiate the broader abstract statement that fixed parameters \"remove the need for variational parameter optimization\" for QRAO in general. Please qualify the claim and provide quantitative concentration statistics.","section":"III-B, Fig. 4 (left)"},{"comment":"The QRAO encoding requires that adjacent vertices be placed on distinct qubits. The paper states that the number of qubits is \"a third\" of the standard QAOA encoding, but it does not state when this is achievable with three variables per qubit. For some 3-regular graphs, no independent set of size three exists (e.g., the triangular prism), so the idealized N/3 qubit count is not universal. The \"up to three times fewer\" phrasing in the conclusion is safe, but the earlier statement in Section II-B should be qualified with the conditions under which the (3,1)-QRAC encoding achieves exactly one qubit per three variables.","section":"II-B and Fig. 7C"}],"minor_comments":[{"comment":"The sentence \"the ground state of the Z mixing Hamiltonian is |0>⊗N\" is inconsistent with the standard convention Z|0>=+|0>, for which the ground state of Σ_i Z_i is |1...1>. If the authors instead mean the ground state of −Σ_i Z_i, the sign convention should be stated explicitly.","section":"III-A"},{"comment":"The phrase \"grouped two-qubit Pauli terms commute if they do not share the same Pauli string\" is imprecise; such terms commute when they act on disjoint qubits (or differ by a scalar). Please rephrase for accuracy.","section":"III-C"},{"comment":"The notation |\\hat{\\psi}_p> in the Ensure line is not defined. Please state explicitly that it denotes the state produced by the Trotterized approximation of the cost evolution.","section":"Algorithm 1"},{"comment":"The comparison in Fig. 7 uses Grouped Trotter with T=2, but the choice of T=2 is not justified. Please explain why this value was selected, especially since Fig. 6 shows a small performance gain from T=3.","section":"IV"},{"comment":"The x-axis label for the noise scale appears to be missing or truncated in the figure. Please ensure that all axes are clearly labeled.","section":"Fig. 7B"},{"comment":"The statements that code and data are \"available upon reasonable request\" are insufficient for reproducibility of the reported fixed parameters. Please provide an open-access repository or, at minimum, the exact fixed angle values.","section":"Data/Code Availability"},{"comment":"Reference [9] is listed without a journal or arXiv identifier. Please complete the citation so readers can locate the source.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The inconsistency between the 30 instances in the text and the 120 instances in the figure captions, together with the confounded mixer comparison and the undisclosed fixed parameters, makes the current version unsuitable for acceptance. However, the core idea is viable and the authors are appropriately modest about the performance gap with standard QAOA. With the parameter values disclosed and the comparison protocol clarified, the paper could become publishable. I see no citation-ethics concerns beyond the usual need to resolve the instance-count discrepancy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read of arXiv:2502.04277. The paper proposes QAOA-for-QRAO: run QAOA on QRAO's non-diagonal relaxed Hamiltonian instead of variational training. That's genuinely new, and they report a real design finding — a Z mixer with |0> initial state beats the usual X mixer — plus evidence that fixed, instance-independent parameters transfer within 3-regular MaxCut. The resource claim (about 3x fewer qubits) is real, and they honestly note that standard QAOA still gives better solution quality. The Trotter step result (T<=3 suffices) is a useful practical data point.\n\nWhat I like: the experiments have error bars, the parameter concentration plots are convincing, and they don't oversell. The comparison to QAOA with fixed angles from Wurtz-Lykov is a fair baseline.\n\nThe soft spots are mostly addressable. First, the fixed parameters are averages of optimized angles from training instances, but the actual numbers are never tabulated and code/data are only 'available on request.' So the central quantitative claim isn't independently reproducible. Second, the mixer comparison in Fig. 3 uses the same parameters that were tuned for the Z mixer for all three mixers. That confounds the comparison; the X and Y mixers might do better with their own optimized parameters. Third, the test set is the same graph family (3-regular MaxCut) used for the fit, plus a (2,1)-QRAC appendix on the same family. That's some generalization evidence, but it's a narrow base for the claim that the fixed-parameter heuristic transfers broadly. Also, there's a small inconsistency: the text says 30 training instances, while Figs. 4 and 5 captions say 120. That needs fixing.\n\nNone of this sinks the paper. The central idea — that QAOA parameter heuristics survive the move to a non-diagonal relaxed Hamiltonian — holds up for the tested family. I'd want to see the angles published and the mixer comparison redone before I'd trust the design conclusions, but the work deserves a serious referee. It's a solid empirical contribution for the QRAO and early-FT QAOA community.\n\nRecommendation: send to peer review, with requests for code, data, tabulated fixed angles, and a deconfounded mixer comparison.\n\nBest","headline":"A useful empirical study of QAOA on QRAO's relaxed Hamiltonian, with fixed parameters that appear to transfer within one graph family; the main soft spots are undisclosed fitted angles and a confounded mixer comparison.","tokens_in":14353,"tokens_out":2229,"would_cite":true,"duration_ms":22147,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","68Q12"],"pacs":["03.67.Ac","03.67.Lx"],"model":"deepseek-v4-flash","headline":"This paper proposes QAOA-for-QRAO, a non-variational scheme that runs quantum random access optimization with one fixed set of instance-independent parameters and needs up to three times fewer qubits than standard QAOA.","keywords":["quantum random access optimization","QAOA","non-variational quantum optimization","fixed parameters","MaxCut","quantum random access codes","Trotterization","entanglement entropy"],"falsifier":"Take the fixed angles obtained from 30 random 3-regular MaxCut instances and run QAOA-for-QRAO on a different graph family, such as 4-regular graphs or random graphs with a different degree distribution at N≈30, comparing the rounded approximation ratio αc against per-instance optimized angles; if the fixed-angle αc falls noticeably below the optimized value, or if the optimized β and γ values are not tightly clustered, parameter concentration fails and the central non-variational claim collapses.","tokens_in":13338,"feed_emoji":"⚛️","tokens_out":6601,"duration_ms":63917,"temperature":0.7,"pith_summary":"Quantum Random Access Optimization (QRAO) encodes several optimization variables into each qubit, but every implementation so far required a variational loop to tune parameters for each instance. This paper proposes running QRAO with the Quantum Alternating Operator Ansatz (QAOA) using a single fixed set of instance-independent parameters, and benchmarks the idea on MaxCut. It reports that on random 3-regular graphs up to 26 nodes and depth 6, the fixed parameters match per-instance optimized parameters in approximation quality, and that only a couple of Trotter steps are needed to implement the non-diagonal cost layer. If the results hold, space-efficient quantum optimization could run on early fault-tolerant hardware without the overhead of variational training.","feed_headline":"Fixed angles let QRAO skip training and use a third of the qubits","feed_subtitle":"Non-variational QAOA keeps MaxCut quality near optimized parameters while removing per-instance training overhead.","key_machinery":"The central object is the relaxed, non-diagonal Hamiltonian \\(\\tilde{H}_C\\) built from a (3,1)-quantum random access code, which encodes three binary variables into one qubit through Pauli X, Y, and Z assignments. The paper drives this Hamiltonian with alternating cost and mixer unitaries, selecting the Z mixer with the |0⟩ initial state rather than the usual X mixer. Because the Pauli terms in \\(\\tilde{H}_C\\) do not commute, the cost layer is implemented by first-order Trotter or Grouped Trotter decomposition, with only a few Trotter steps sufficient in practice. The load-bearing mechanism is parameter concentration: averaging optimized QAOA angles over random instances yields a fixed schedule that behaves like the fully optimized angles.","core_discovery":"The paper's central claim is that QRAO's space savings do not have to come with variational training. On MaxCut over random 3-regular graphs with N up to 26 and p up to 6, a fixed parameter schedule obtained by averaging optimized angles over 30 small instances performs comparably to fully optimized, per-instance parameters, in both the relaxed approximation ratio αr and the classical rounded ratio αc. A Z mixer with the |0⟩ initial state outperforms the standard X mixer and |+⟩ state for the relaxed Hamiltonian, and a Grouped Trotter implementation with T=2 or 3 Trotter steps recovers near-exact performance. The paper concludes that QRAO can be executed as a non-variational algorithm, using up to three times fewer qubits than standard QAOA while keeping good solution quality.","pith_inferences":["Inference: If parameter concentration extends beyond 3-regular MaxCut, the same averaging protocol could produce fixed schedules for other NP-hard optimization families, making QRAO a drop-in space-saver for industrial solvers.","Inference: The observed approximation-ratio gap to standard QAOA means the threefold qubit saving is bought at some quality cost; on noisy hardware, the extra two-qubit gates from Trotterization may widen that gap, so the trade-off should be re-evaluated at realistic error rates.","Inference: A direct testable extension is to train fixed parameters on one graph family and benchmark on another; if concentration fails, a small per-instance fine-tuning step could recover most of the performance while keeping most of the space savings."],"forward_implications":["QAOA-for-QRAO can be executed without per-instance variational training, removing the shot overhead and barren-plateau risk associated with variational parameter search.","The relaxed encoding cuts the qubit count by up to a factor of three for MaxCut, allowing a fixed hardware register to address larger problem instances.","A small number of Trotter steps (T=2 or 3) is enough for the non-diagonal cost layer, so the circuit overhead of the relaxed Hamiltonian is modest in practice.","The fixed parameter set transfers to larger instances than those used to choose it, and the same averaging protocol works for the (2,1)-QRAC variant.","Entanglement entropy grows with QAOA depth and correlates with approximation quality, suggesting that the relaxed target state uses entanglement in a way that standard QAOA on classical Hamiltonians does not."],"supporting_citations":[{"why":"Introduces QRAO, the relaxed Hamiltonian construction, and the Pauli and magic rounding procedures that this paper builds on.","marker":"[8]"},{"why":"Defines quantum random access codes, the encoding family that gives QRAO its space savings.","marker":"[10]"},{"why":"Introduces the quantum approximate optimization algorithm whose alternating unitary structure is reused here.","marker":"[24]"},{"why":"Extends QAOA to an alternating operator ansatz with flexible mixers, providing the basis for the QRAO-specific mixer design.","marker":"[25]"},{"why":"Supplies the fixed-angle conjecture for standard QAOA on regular graphs that this paper transfers to the non-diagonal setting.","marker":"[40]"},{"why":"Provides the parameter-setting heuristics that justify using a fixed, instance-independent schedule.","marker":"[35]"}],"fun_headline_variants":["QRAO without variational training: fixed angles still get good MaxCut","Fixed QAOA angles cut qubit count by 3x, skip optimization","Non-variational QRAO: three-fold qubit savings with fixed parameters","Instance-free QAOA angles for MaxCut: near-optimal quality, fewer qubits","Skip the training: fixed angles make QRAO practical for MaxCut"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that optimal QAOA angles for the relaxed, non-diagonal QRAO Hamiltonian are tightly concentrated across random instances of a problem class, so the average over 30 small 3-regular MaxCut instances serves as a universal schedule for all instances of that class and beyond.","fun_headline_variants_meta":{"raw":{"variants":["QRAO without variational training: fixed angles still get good MaxCut","Fixed QAOA angles cut qubit count by 3x, skip optimization","Non-variational QRAO: three-fold qubit savings with fixed parameters","Instance-free QAOA angles for MaxCut: near-optimal quality, fewer qubits","Skip the training: fixed angles make QRAO practical for MaxCut"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00079,"raw_usage":{"total_tokens":3471,"prompt_tokens":920,"completion_tokens":2551,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":2449}},"tokens_in":536,"tokens_out":2551,"duration_ms":18470,"temperature":1.0,"reasoning_tokens":2449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T22:55:23.660229+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the fixed angles obtained from 30 random 3-regular MaxCut instances and run QAOA-for-QRAO on a different graph family, such as 4-regular graphs or random graphs with a different degree distribution at N≈30, comparing the rounded approximation ratio αc against per-instance optimized angles; if the fixed-angle αc falls noticeably below the optimized value, or if the optimized β and γ values are not tightly clustered, parameter concentration fails and the central non-variational claim collapses.","supporting_citations":[{"cited_title":"Approximate solutions of combinatorial problems via quantum relaxations,","cited_arxiv_id":null,"evidence_quote":"Introduces QRAO, the relaxed Hamiltonian construction, and the Pauli and magic rounding procedures that this paper builds on."},{"cited_title":"Dense quantum coding and quantum finite automata,","cited_arxiv_id":null,"evidence_quote":"Defines quantum random access codes, the encoding family that gives QRAO its space savings."},{"cited_title":"From the quantum approximate optimization algorithm to a quantum alternating operator ansatz,","cited_arxiv_id":null,"evidence_quote":"Extends QAOA to an alternating operator ansatz with flexible mixers, providing the basis for the QRAO-specific mixer design."},{"cited_title":"Fixed-angle conjectures for the quantum approximate optimization algorithm on regular maxcut graphs,","cited_arxiv_id":null,"evidence_quote":"Supplies the fixed-angle conjecture for standard QAOA on regular graphs that this paper transfers to the non-diagonal setting."},{"cited_title":"Parameter setting heuristics make the quantum approximate optimization algorithm suitable for the early fault-tolerant era,","cited_arxiv_id":null,"evidence_quote":"Provides the parameter-setting heuristics that justify using a fixed, instance-independent schedule."}],"review_version":1}