{"id":"fb649b54-a256-4ba3-ab12-425188fca985","arxiv_id":"2602.21689","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DO-QAOA shows that the 2^m subproblems in frozen-qubit divide-and-conquer QAOA share near-identical variational landscapes, so training one representative and transferring its parameters cuts training cost from exponential to constant with competitive solution quality.","lead":"A divide-and-conquer quantum optimization method reuses one trained set of angles for every sub-problem, based on the observation that freezing qubits barely changes the energy landscape. It claims to cut training overhead from exponential to constant while keeping solution quality, on 6,000 simulated circuits.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Landscape similarity as measured (L∞ bound, correlation) does not by itself guarantee near-optimal transferred parameters; the paper never measures the per-target transfer gap, so the K=1 collapse remains unverified.","rationale":"The reader's weakest_assumption identifies exactly the gap between landscape similarity and parameter-transfer optimality. My review sharpens this with specifics: Theorem 1 (Appendix B) bounds only pointwise energy differences and gives no control on the optimization gap; Table II's 'With Coeffs' correlations drop to ≈0.8 for m=2,3, while the 'No Coeffs' rows are tautological identities; and the paper never directly measures per-target transfer error. The reported ARG comparisons may also be confounded by differing initialization protocols between DO-QAOA and the FrozenQubits baseline, since Appendix D shows initialization strongly affects convergence. However, these are not internal contradictions; they are missing evidence. The paper's practical idea is plausible and the shot-count reductions are real if the transfer works. The correct verdict remains CONDITIONAL: the central claim needs a direct per-subproblem transfer-gap measurement, matched baselines, and error bars. No change to the reader's verdict is warranted.","tokens_in":24568,"tokens_out":6998,"duration_ms":69408,"concrete_test":"For each benchmark graph and m=1,2,3, compute for every target subproblem i: Δ_i = ARG(θ*_rep applied to i) − ARG(θ*_i), where θ*_i is obtained by optimizing subproblem i alone with the same optimizer, same initialization strategy, and matched shot budget. Report mean and worst-case Δ_i across all targets, with error bars. If mean Δ_i is not small relative to the DO-QAOA-vs-FrozenQubits ARG differences in Table IV, the K=1 collapse fails and the reported gains are not attributable to landscape transfer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that one representative parameter set suffices for all 2^m frozen subproblems—rests on the unstated inference that landscape similarity (small pointwise deviation, Theorem 1, Appendix B; high correlation, Table II) implies near-optimality of transferred parameters. This inference is not established. Theorem 1 bounds |E^{(z')}(θ)-E^{(z)}(θ)| by the L1 norm of field differences; this pointwise closeness does not bound the optimization gap E^{(z')}(θ*_rep) - min_θ E^{(z')}(θ) unless the landscape is, e.g., strongly convex with known curvature, which is not shown. Table II shows the actual full-landscape correlations for m=2,3 are r≈0.80 (MSE 0.176, L∞≈1.04) in the 'With Coeffs' rows—not 'near 1.0' as the caption claims; the 'No Coeffs' rows are identities because H_quad is literally the same operator. The paper never reports the per-target transfer gap (ARG or energy of transferred parameters vs. target-optimized parameters). Without this measurement, the K=1 collapse is an empirical assertion, and the reported ARG improvements over FrozenQubits may be confounded by initialization differences: DO-QAOA uses the 'Shortcut' init (Appendix D) while no equivalent init is specified for the FrozenQubits baseline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DO-QAOA, a divide-and-conquer strategy for QAOA in which a subset of high-degree qubits is frozen, generating 2^m reduced subproblems. The authors argue that the variational energy landscapes of these subproblems are highly similar—quantified by a replica-overlap order parameter and by pairwise correlations—so that a single representative subproblem can be optimized and its parameters transferred to all others. This collapses the exponential training cost to O(K), typically K=1. The method includes a bias-aware transfer rule with optional fine-tuning, and is benchmarked on synthetic and real-world graphs under a realistic noise model, reporting large reductions in quantum shot count and runtime while maintaining or improving approximation ratio gap versus the FrozenQubits baseline.","tokens_in":24914,"tokens_out":3593,"duration_ms":34941,"significance":"If the central claim were established, the paper would offer a practically important result: replacing the exponential training overhead of divide-and-conquer QAOA with a constant cost while preserving approximation quality. The authors provide a simple, correct pointwise landscape-stability bound (Appendix B) and a broad benchmark suite of over 6,000 noisy circuits, which is a strength. However, the key inference—that landscape similarity implies near-optimal parameter transfer—is not directly measured. The reported correlations are modest (r≈0.8 for m=2,3 in the meaningful 'With Coeffs' rows), the 'No Coeffs' rows are tautological, and no per-target transfer gap is reported. The paper is therefore best viewed as an interesting empirical proposal whose main efficiency claim is not yet supported by the evidence presented.","major_comments":[{"comment":"The load-bearing claim is that one representative parameter set is near-optimal for all subproblems. Theorem 1 (Eq. B2) bounds only the pointwise L∞ distance between energy surfaces |E^{(z')}(θ) - E^{(z)}(θ)|; it does not bound the distance between their optima or the transfer gap E^{(z')}(θ*_rep) - min_θ E^{(z')}(θ). The measured correlations in Table II (With Coeffs) are r≈0.80 for m=2,3 with L∞≈1.04 and 0.77, which do not imply close optima or near-optimal transfer. The paper never reports the per-subproblem ARG or energy of transferred parameters versus fully optimized parameters. This is a required measurement; without it, the K=1 collapse remains an empirical assertion.","section":"§II.C, Appendix B, Table II"},{"comment":"The 'No Coeffs' rows are tautological: setting the induced linear fields to zero makes every subproblem Hamiltonian exactly equal to H_quad, so the subproblems are identical as operators and r≈1 by construction. These rows cannot serve as empirical support for landscape similarity. The caption's claim that 'even with induced linear coefficients the correlation remains near 1.0' is contradicted by the table's own 'With Coeffs' rows for m=2 and m=3 (r=0.796 and r=0.802). Either the table, the caption, or the verification protocol should be corrected, and the empirical claim should rest on the With Coeffs numbers only.","section":"Table II, §IV.A"},{"comment":"The Bias-Aware Transfer Rule uses a threshold ΔB=0.3, a 10-epoch fine-tune, and the Shortcut initialization angles (γ≈−π/6, β≈−π/8). These appear to be fitted on the same benchmark set. Since the reported ARG improvements over FrozenQubits could plausibly arise from the Shortcut initialization rather than from landscape-aware transfer, the paper should provide a sensitivity analysis and/or a train/test split, and should specify the initialization used for the FrozenQubits baseline. Without this, the comparison in Table IV is not clean.","section":"§III.B, Appendix D"},{"comment":"The efficiency claims compare DO-QAOA's total shots to FrozenQubits' 2^m independent optimizations. This comparison is valid only if the transferred parameters actually achieve comparable per-subproblem solution quality. Since the paper does not report the distribution of per-subproblem ARGs for transferred parameters, the reader cannot tell whether the method is 'collapsing the landscape classes' or simply ignoring poorly solved subproblems. Please report, for each target subproblem, the ARG obtained with θ*_rep and, if available, the ARG after full target-specific optimization, across all instances.","section":"§IV.C, Table IV"}],"minor_comments":[{"comment":"The notation '2 m' and '2 m distinct reduced problems' should be typeset as 2^m throughout.","section":"Abstract"},{"comment":"The caption is confusing: 'MSE denotes the mean squared error in evaluating the Correlation' is unclear. It should say MSE of the raw energy surfaces and define it explicitly.","section":"Table II caption"},{"comment":"The sentence 'the landscape is defined over the single-component parameter space (γ1, β1)' is contradictory; the landscape is two-dimensional in (γ,β) for p=1.","section":"Fig. 2 caption"},{"comment":"The Shortcut initialization angles are stated without derivation or reference. If they are empirically chosen, say so and provide the evidence for the cluster location.","section":"Appendix D"},{"comment":"The noise model is described only as 'derived from FakeBrisbane' and compiled with 'a popular quantum software stack'; specify the simulator backend, noise-model version, and compilation settings so the results can be reproduced.","section":"Appendix E1"},{"comment":"The sentence 'In the rare cases where its framework naturally extends to the fragmented phase (s < sc)' is grammatically incomplete; it appears to mean that the framework can be extended to the fragmented phase, but the current wording obscures the intended meaning.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper's practical claim is attractive, but the central evidence is incomplete in a way that is fixable. The authors need to add a direct measurement of the transfer gap on target subproblems and to remove the tautological 'No Coeffs' verification from the main evidence. The threshold and initialization hyperparameters also need robustness checks and a fairer baseline. I would not reject at this stage, but the revision must address these load-bearing points before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on QAOA parameter transfer or divide-and-conquer variational methods. The genuinely new piece is the idea of treating the 2^m frozen-subproblem family as a set of landscapes and collapsing them via a similarity measure; the overlap order parameter q and the connectivity-driven transition are a reasonable spin-glass-inspired addition, though the transition is inferred from unscaled crossings, so treat it as suggestive. The empirical gains are real: DO-QAOA cuts shots by two to three orders of magnitude versus FrozenQubits on the tested graphs and matches or improves ARG. That is a practically relevant result for NISQ-style simulation.\n\nThe soft spots are where the stress test lands. Theorem 1 bounds the pointwise difference between energy surfaces, not the distance between their optima, so it does not by itself justify direct parameter transfer. The paper's own Table II shows the 'With Coeffs' correlations for m=2,3 are around 0.80, not 'near 1.0' as the caption claims. The 'No Coeffs' rows are close to tautologies: removing the only term that varies across subproblems makes them identical, so r≈0.999 is definitional. The paper never reports the per-target transfer gap—the energy of transferred parameters versus target-specific optimized parameters—which is the direct evidence needed for K=1 collapse. The bias threshold, Shortcut initialization angles, and fine-tuning epochs are all fitted on the same benchmarks with no error bars and no released code or data. That is a reproducibility problem, and the comparison with FrozenQubits may be confounded by initialization differences.\n\nAll that said, the central pragmatic claim is plausible. For sparse graphs with a few high-degree nodes, frozen subproblems share most of their quadratic term, and parameter transfer has already been shown to work in QAOA for similar instances. The paper extends that to the divide-and-conquer setting, and the reported gains are large enough to take seriously. The missing per-target gap measurement is the key thing a referee should demand.\n\nThis paper deserves a serious referee, but with a clear request: report the transfer gap directly, give error bars, release the code, and fix the tautological or overstated parts. If the authors do that, this becomes a solid contribution. As it stands, it is a promising result with scaffolding too weak for the central claim.","headline":"A plausible method-level win for divide-and-conquer QAOA, but the paper's central claim—that one representative parameter set suffices for all frozen subproblems—is asserted rather than directly measured.","tokens_in":25436,"tokens_out":3715,"would_cite":true,"duration_ms":35647,"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":"Freezing qubits in a QAOA problem produces exponentially many reduced instances, but their energy landscapes are near-identical; one optimized parameter set can serve them all.","keywords":["QAOA","divide-and-conquer","parameter transfer","energy landscape similarity","landscape overlap order parameter","qubit freezing","NISQ","Ising model"],"falsifier":"For a fixed graph in the self-averaging regime, compute the exact optimal parameters (gamma*, beta*) for all 2^m frozen subproblems and measure the spread of these optima and the energy lost by using a single representative's parameters on each subproblem. If that energy gap grows with m or with system size L, or if the optimal parameter spread does not shrink as connectivity increases, the claimed collapse to K=1 is falsified.","tokens_in":24427,"feed_emoji":"⚛️","tokens_out":7719,"duration_ms":61909,"temperature":0.7,"pith_summary":"This paper tries to establish a universality principle for the Quantum Approximate Optimization Algorithm: when a QAOA problem is split by freezing a few high-degree qubits, the exponentially many reduced subproblems (up to 2^m, one per frozen configuration) have variational energy landscapes that are near-copies of one another. If that holds, a single representative subproblem can be fully optimized and its parameters reused for all the others, converting the training cost of divide-and-conquer QAOA from exponential in m to constant. The paper measures landscape similarity with a spin-glass-style overlap order parameter, finds a sharp similarity transition as graph connectivity is tuned, and still observes aligned basin centers on both sides. On the strength of that, DO-QAOA is shown to cut total quantum shots by 280x to 385x and wall-clock time by 10x to 15x on more than 6,000 circuits while matching or improving the approximation ratio gap of independent retraining. The reason to care: it removes the main classical bottleneck that made divide-and-conquer QAOA impractical on near-term hardware.","feed_headline":"One QAOA training run replaces 2^m frozen-subproblem optimizations","feed_subtitle":"Similar energy landscapes let one optimized parameter set serve every reduced instance, cutting shots by up to 385x.","key_machinery":"The load-bearing structure is the decomposition of a frozen QAOA instance into an invariant quadratic part (edges inside the active subgraph) and a configuration-dependent linear part (induced local fields from frozen neighbors). The similarity measure is the landscape-overlap order parameter q(s), the mean pairwise cosine similarity of standardized energy surfaces over a (gamma,beta) grid, adapted from replica-overlap ideas in spin-glass physics; it detects the sharp transition between fragmented and self-averaging regimes as the connectivity decay exponent s is tuned. Theorem 1 (Landscape Stability) supplies the theoretical bound: the L-infinity pointwise distance between any two subproble","core_discovery":"The paper's central claim is that the QAOA variational landscape of a decimated Ising problem is governed by the quadratic interaction backbone that is invariant across frozen configurations, so the configuration-dependent linear biases act only as bounded perturbations. Concretely, freezing m qubits produces a Hamiltonian split into an invariant quadratic part and a configuration-dependent linear part; Theorem 1 bounds the pointwise difference between any two subproblem energy surfaces by the sum of induced-field differences, independent of circuit depth. Empirically the full-landscape correlation between subproblems stays above 0.999 for m=1 and near 0.8 for m=2,3, while the isolated quadr","pith_inferences":["The K=1 collapse is an empirical observation, not a derived theorem: Theorem 1 bounds pointwise surface separation but not the displacement of the optima. A direct measurement of the per-subproblem energy gap of transferred parameters would test whether K=1 persists as m and L grow, or whether K>1 clustering eventually becomes necessary.","The sharp transition in q(s) suggests a practical diagnostic: a graph's effective connectivity could be estimated before deciding whether direct transfer is safe, potentially replacing the hand-set bias threshold (0.3) with a physics-derived criterion.","The same 'invariant quadratic backbone plus local linear perturbations' argument is not specific to QAOA; any variational ansatz whose cost is dominated by a fixed quadratic term and whose light cone limits perturbation spread could inherit the same transferability, which the paper gestures at but does not demonstrate.","For dense graphs in the fragmented phase, the paper's own phase diagram predicts that subproblems will eventually fracture into a small number of distinct landscape families rather than one; the natural extension is to cluster the 2^m instances into K groups and train one representative per cluster, exactly the route the paper outlines for future work."],"forward_implications":["Divide-and-conquer QAOA training cost drops from O(2^m x N_shots x N_iter) to O(K x N_shots x N_iter); on every benchmark the measured K is 1, so the exponential overhead disappears entirely.","On more than 6,000 noisy circuits, transferred parameters match or beat independent full optimization in approximation ratio gap, cutting total quantum shots by 280x to 385x and wall-clock time by 10x to 15x relative to the standard divide-and-conquer baseline.","The method is most effective on sparse, locally connected graphs (power-law, 3-regular, and structured real-world networks) where the graph diameter exceeds the QAOA light cone; dense globally connected problems still benefit at finite sizes but show smaller gains.","The empirical operating rule m less than or equal to 3 gives the best trade-off; freezing additional high-degree nodes beyond three yields diminishing returns and mostly adds bookkeeping overhead.","Because the circuit structure is identical to standard freezing, all gains come from removing redundant optimization loops, so DO-QAOA is compatible with existing noise-mitigation and circuit-lowering techniques."],"fun_headline_variants":["One QAOA run covers 2^m frozen subproblems via landscape overlap","DO-QAOA: share one landscape across all frozen subproblems","Frozen QAOA landscapes align; train once, skip 2^m runs","Landscape similarity collapses 2^m QAOA subproblems to K classes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire speedup rests on the assumption that two energy surfaces that look similar also have their best parameter settings in nearly the same place; the paper shows the surfaces are close pointwise, but it does not directly measure how far their optima shift.","fun_headline_variants_meta":{"raw":{"variants":["One QAOA run covers 2^m frozen subproblems via landscape overlap","DO-QAOA: share one landscape across all frozen subproblems","Frozen QAOA landscapes align; train once, skip 2^m runs","Landscape similarity collapses 2^m QAOA subproblems to K classes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1289,"prompt_tokens":714,"completion_tokens":575,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":493}},"tokens_in":458,"tokens_out":575,"duration_ms":31423,"temperature":1.0,"reasoning_tokens":493,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:56:12.912698+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed graph in the self-averaging regime, compute the exact optimal parameters (gamma*, beta*) for all 2^m frozen subproblems and measure the spread of these optima and the energy lost by using a single representative's parameters on each subproblem. If that energy gap grows with m or with system size L, or if the optimal parameter spread does not shrink as connectivity increases, the claimed collapse to K=1 is falsified.","supporting_citations":[],"review_version":1}