{"id":"dab92a98-65cf-444c-8132-66709e5ef4e9","arxiv_id":"2502.09704","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Iteratively reusing the best measured solutions as a superposed warm-start state improves QAOA results on MaxCut and portfolio optimization in small-scale simulations.","lead":"This paper proposes starting the QAOA quantum optimization algorithm from a quantum state built from the best solutions found in earlier rounds, then repeating the process. In simulations on MaxCut and portfolio optimization, the repeated warm starts improve the results and avoid the stagnation seen with a single warm start.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DGMVP scaling claim relies on fitted exponents whose negative sign implies Pgm grows with problem size, contradicting probability constraints; the reported 'more favourable scaling' needs re-examination.","rationale":"I read the paper in good faith and find the MaxCut core plausible: the simulations show r decreasing over iterations and R remaining positive, and the worst-case graphs in Appendix C are a useful explicit construction. The theoretical Appendices A and B are indeed hand-wavy, but they are presented as explanatory arguments rather than as the sole evidence; the numerical MaxCut results could stand without them. The more damaging issue is in the DGMVP scaling claim, which is part of the abstract's headline and is supported only by the fitted exponents b. The reported negative values are inconsistent with the definition of Pgm as a probability: if Pc decreases with problem size, a negative b makes Pgm increase, which is impossible for a probability and contradicts the expected behavior of any algorithm. The text even labels positive b values as 'much better' for one case, showing the sign matters. This is not merely a missing derivation; it is an internal inconsistency that could reverse the claimed advantage. A concrete refit of the raw data would settle it. The reader's weakest_assumption identified the growth assumptions in Appendices A and B, which I agree are weak, but the DGMVP exponent issue is more directly load-bearing because it affects a headline quantitative claim in the abstract. I therefore recommend keeping the conditional verdict, but the conditions should include correcting and re-fitting the DGMVP scaling exponents, not just strengthening Appendices A and B.","tokens_in":14901,"tokens_out":11370,"duration_ms":117434,"concrete_test":"Re-fit the raw Pgm-versus-Pc data from Fig. 9 on log-log axes without presupposing a sign, i.e., compute \\bar{P_gm} = a · Pc^b with free b, and also plot the average Pgm directly against l and against n. Verify whether the fitted b is positive and less than 1 for the iterative warm-started method, and whether the predicted \\bar{P_gm} stays below 1 across the full range of l and n shown. If b is negative, check whether the predicted Pgm increases with problem size; if so, the reported scaling advantage is invalid and the sign convention must be corrected before the abstract claim can be accepted.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract's claim that the iterative method has 'a more favourable scaling of identifying the global minimal' compared to classical constrained sampling rests on the fitted exponents b in Fig. 9, where the text states \\bar{P_gm} = a · Pc(n,l)^b. The reported b values are negative in most cases: b = -1.36, -0.22, -0.21 (p=1, l-scaling), b = -1.65, -0.78, -0.19 (p=1, n-scaling), and similarly for p=4. Since Pc = 1/B(n,l) is a probability in (0,1) that decreases when l or n increases, a negative b makes \\bar{P_gm} = a / Pc^{|b|} increase as the problem size grows. A probability cannot increase without bound and certainly cannot exceed 1; the reported fits therefore imply an unphysical model. For the n-scaling p=4 case the text reports small positive b values and calls them 'much better', indicating that the sign of b is meant to be positive for advantage, yet most reported b values are negative. This internal inconsistency means the central DGMVP scaling improvement is not established as presented. The MaxCut simulations may still stand, but the abstract's scaling claim depends directly on these exponents and would collapse if the sign convention were corrected.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an iterative warm-started QAOA method in which, after each optimization run, the measured strings are sorted by cost and a 't-order statistic' superposition state is prepared from the best measurements and used as the initial state for the next run. The method is tested numerically on 3-regular MaxCut with standard QAOA and on the DGMVP portfolio-optimization model with a hard nearest-neighbour mixer. The authors report that the iterative scheme avoids the 'stuck issue' of single-string warm-started QAOA, that the MaxCut approximation error decreases across iterations, and that the probability of measuring the DGMVP global minimum scales more favourably than classical constrained sampling. Theoretical appendices attempt to explain the improvement by an energy-injection argument and a state-space-expansion argument.","tokens_in":15226,"tokens_out":3642,"duration_ms":41492,"significance":"If established, the iterative warm-started scheme would be a simple and potentially useful heuristic for improving QAOA beyond the single-string warm-start limitations identified in Ref. [1]. The MaxCut simulations on random and specially constructed 3-regular graphs are internally consistent and show a clear qualitative improvement on small instances; this is a genuine empirical contribution. However, the paper's broader claims are currently conditional: the DGMVP scaling advantage rests on a power-law fit whose reported exponents have an inconsistent sign, and the theoretical explanation in Appendices A and B relies on unproven growth assumptions rather than derivations. The paper does not provide code or data, so exact reproducibility of the numerical results is limited.","major_comments":[{"comment":"The DGMVP scaling claim is not supported by the reported fits. The text fits \\bar{P_gm} = a · Pc(n,l)^b and reports mostly negative b values, e.g. b = -1.36, -0.22, -0.21 for l-scaling at p=1 and b = -1.65, -0.78, -0.19 for n-scaling at p=1. Since Pc(n,l)=1/B(n,l) lies in (0,1) and decreases when l or n increases, a negative exponent makes \\bar{P_gm} increase as the problem size grows, which is impossible for a probability and contradicts the intended 'more favourable scaling' conclusion. The later statement that positive b values for p=4 n-scaling are 'much better' confirms the sign inconsistency. This is a load-bearing issue because the abstract's scaling claim depends directly on these exponents; the authors need to clarify the sign convention, refit with a correct model, or remove the scaling claim.","section":"§IV, Fig. 9 and surrounding text"},{"comment":"The theoretical resolution of the 'stuck issue' is not derived. Equation (A5) assumes the thermality coefficient grows as ε_wK = ε_w0 + D K with an unspecified 'diffusion constant' D, and Eq. (B4) assumes the reachable state count expands as d_K = d_0 e^{λK} via an amplification process governed by unspecified factors g_k. No mechanism in Algorithm 1 is shown to imply D>0 or λ>0, and no bound on K is derived from the algorithm's parameters. The conclusion in Sec. V that establishing theoretical bounds is future work is consistent with this assessment. These appendices should be reframed as heuristic motivation or replaced by a genuine analysis, because the paper currently presents them as explaining the simulation results.","section":"Appendix A and Appendix B"},{"comment":"The claim that the worst-case r 'converges toward' the best classical bound 1-0.9326 and even surpasses it is stronger than what the simulations establish. The data cover only N ≤ 16 and at most four iterations, and no statistical uncertainty or significance testing is reported for the r and R curves. The designed worst-case graphs are specific constructions containing only the g6 subgraph, so they do not establish a worst case for the iterative method itself. The wording should be revised to describe an observed trend on small instances rather than a convergent or benchmark-surpassing guarantee.","section":"§IV, Figs. 2-4"}],"minor_comments":[{"comment":"There are numerous typos and encoding artifacts that should be corrected: 'Hardamard' for Hadamard, 'approxiamtion' for approximation, 'Dual Anealing' for Dual Annealing, and 'ans¨atze' with broken encoding.","section":"§II and §III"},{"comment":"The definition of the percentile state is self-referential: p'_i is defined in terms of p'_i on the right-hand side. It should presumably read p_i or a normalized version of the measured counts.","section":"Eq. (5)"},{"comment":"The definition P = (N_total - N_static)/N_total is the fraction of states that are improved, but the text says 'N_static will increase cumulatively' and interprets low P positively; the naming and interpretation should be made consistent.","section":"Eq. (19) and surrounding text"},{"comment":"The simulation parameter 'M = 218' appears to be a typo for 2^18; please correct it and similarly check the notation for m and M throughout the figure captions.","section":"Fig. 9 caption and §IV"},{"comment":"The sentence 'Thus, the scalings in Fig. 9 demonstrate demonstrate that...' contains a duplicated word and should be edited.","section":"§IV end"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a mixed bag: the MaxCut experiments are a reasonable empirical study, but the DGMVP scaling claim needs a careful reanalysis because the fitted exponents as reported imply an unphysical increasing probability. The theoretical appendices are currently more suggestive than rigorous. If the authors can fix the scaling analysis and reframe the appendices as heuristic, a revised version could be suitable. I would also encourage the editor to ask for error bars and data availability, since the numerical claims are central."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the core idea is genuinely new—using a superposition of the best measured strings as an iterative warm start for QAOA—and the MaxCut simulations look convincing. The DGMVP scaling claim in the abstract, however, is not supported as written.\n\nWhat's new: the t-order statistic state construction is a fresh initialization strategy that directly targets the \"stuck issue\" identified by Cain et al. The MaxCut results show that this iterative method improves the approximation ratio over iterations and, on worst-case 3-regular graphs, pushes r toward 1-0.9326. That's a solid numerical observation on small instances, and the method is simple enough to generalize to other variational algorithms.\n\nWhere it's soft: the DGMVP advantage comes from fitting \\bar{P_gm} = a·Pc^b, and the reported b values are mostly negative (e.g., -1.36, -0.22, -0.21 for l-scaling with p=1). Since Pc = 1/B(n,l) decreases as the problem size grows, a negative b makes P_gm increase with size—which cannot happen for a probability. The authors themselves treat positive b as better in the n-scaling p=4 case, so the negative exponents are either a sign error or a misinterpretation. Either way, the abstract's claim of \"more favourable scaling\" is not established by this fit. The theoretical appendices A and B are also heuristic: they introduce a diffusion constant D and amplification factors g_k without derivation, so they do not rigorously explain why the stuck bound is lifted. No code or data is provided, and several figures lack error bars.\n\nNone of this kills the MaxCut contribution. The iterative superposition method is worth trying and the simulations support it on the tested sizes. But the DGMVP section needs a careful reanalysis, and the appendices should be either made rigorous or clearly labeled as speculation.\n\nWho this is for: anyone working on warm-started variational algorithms or QAOA initialization. It's a useful idea with a promising demonstration, but it is not a finished scaling claim.\n\nMy recommendation: send it to peer review. A referee can check the sign convention and push on the appendices. The central MaxCut observation deserves airtime, and the method is novel enough to warrant careful evaluation.","headline":"Novel iterative warm-start for QAOA with solid MaxCut simulations, but the DGMVP scaling claim rests on a suspicious power-law fit and the theory is hand-wavy.","tokens_in":15751,"tokens_out":3779,"would_cite":true,"duration_ms":37861,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","90C27"],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"An iterative warm-starting scheme builds the next initial state from the best measured strings and reports that the QAOA keeps improving rather than getting stuck.","keywords":["QAOA","warm start","iterative optimisation","MaxCut","3-regular graph","portfolio optimisation","DGMVP","t-order statistic state"],"falsifier":"Take a family of 3-regular graphs built entirely from the g6 subgraph, run p = 1 QAOA with a 20th-order statistic warm start, and record $r$ and $R$ after four iterations. The paper's claimed worst-case behaviour predicts $r$ falls below $1 - 0.6924$ and moves toward $1 - 0.9326$; if for some $N \\ge 12$ the measured $r$ plateaus above $0.3076$ or $R$ turns negative, the central claim is contradicted. A more direct check is to estimate the thermality coefficient $\\epsilon_{wK}$ from the measured output distributions: if it does not grow with $K$, the Appendix A mechanism is absent.","tokens_in":14648,"feed_emoji":"📈","tokens_out":8206,"duration_ms":76454,"temperature":0.7,"pith_summary":"The paper proposes a warm-starting strategy for the quantum approximate optimisation algorithm (QAOA) in which the initial state of each round is a superposition of the best bitstrings measured in the previous round rather than one good classical string. It argues, with numerical simulations, that this 't-order statistic state' keeps the QAOA optimisable across iterations, so the approximation error $r$ decreases and the relative change ratio $R$ stays positive, avoiding the stuck issue that afflicts single-string warm starts. On 3-regular MaxCut, the worst-case $r$ is reported to converge toward $1 - 0.9326$, the best classical bound for cubic graphs; on the discrete portfolio model DGMVP, the probability of sampling the global minimum scales more favourably than classical constrained sampling. The paper interprets each iteration as injecting energy into the optimisation landscape and expanding the reachable state space, which would explain why the original stuck bound no longer applies.","feed_headline":"Warm-started QAOA escapes its stuck bound, iteration by iteration","feed_subtitle":"Superposing the best measured strings pushes 3-regular MaxCut toward the 0.9326 classical bound.","key_machinery":"The central object is the t-order statistic state: after measuring an optimised QAOA circuit, rank the observed bitstrings by cost and form a superposition of the top $t$ of them, with amplitudes proportional to the square roots of their measured frequencies. The state is prepared via a permutation-based sparse-state routine, giving a circuit of $O(tN)$ on $N$ qubits. The argument that iteration escapes the stuck bound rests on two mechanisms: each iteration resets the local thermal equilibrium, so the thermality coefficient grows as $\\epsilon_{wK} = \\epsilon_{w0} + D K$, and the post-selected search space expands as $d_K = d_0 e^{\\lambda K}$; the paper claims these together convert the single-shot $O(1/\\sqrt{m})$ improvement bound into a compounding $O(K/\\sqrt{m})$ one.","core_discovery":"The central claim is that a measurement-derived superposition warm start escapes the stuck bound that holds for single-string warm starts. After each QAOA optimisation the measured bitstrings are sorted by cost, the top $t$ are superposed with probabilities renormalised, and that state is used as the next initial state. Over iterations the approximation error $r$ keeps falling and the relative change ratio $R$ remains positive on simulated 3-regular MaxCut instances, with the worst case reported to approach $1 - 0.9326$; on DGMVP the fitted exponent $b$ in $\\bar{P}_{gm} = a\\,P_c(n,l)^b$ becomes substantially closer to zero than for standalone QAOA or classical constrained sampling, meaning fewer samples are needed to hit the global minimum.","pith_inferences":["The theoretical escape from the stuck bound is argued through two growth laws, $\\epsilon_{wK} = \\epsilon_{w0} + D K$ and $d_K = d_0 e^{\\lambda K}$, that the paper does not derive from the QAOA dynamics; a direct test would be to estimate these quantities from measured distributions on large instances.","A useful control experiment would replace the QAOA with a classical sampler that also returns the top $t$ measured strings; if the classical loop shows the same improvement in $r$, then the advantage is due to post-selection rather than quantum coherence.","If the reported $b$ exponents hold at larger $n$ and $l$, the method gives a concrete route to a sampling advantage for portfolio optimisation on near-term devices, but the plateau observed after four iterations suggests the classical optimiser, not the state preparation, may become the bottleneck."],"forward_implications":["A warm start can be built entirely from measurement outcomes, so no external classical solver or SDP relaxation is needed to initialise QAOA.","The single-shot improvement bound $O(1/\\sqrt{m})$ is replaced by a compounding $O(K/\\sqrt{m})$ bound over $K$ iterations, so the stuck issue is not merely mitigated but escaped.","The same initial-state construction transfers to constrained problems such as DGMVP without changing the QAOA ansatz, improving both mean and minimum-value approximation ratios.","Fitted scaling exponents in Fig. 9 imply the iterative method asymptotically needs fewer samples than classical constrained sampling to find the global minimum of DGMVP instances.","Because the update rule depends only on measured costs, the procedure can be applied to any variational quantum algorithm, not only QAOA."],"supporting_citations":[{"why":"Establishes the stuck issue: a single-string warm-started QAOA cannot improve beyond a thermality bound; this is the baseline the paper claims to escape.","marker":"[1]"},{"why":"Introduces the standard QAOA ansatz with cost and mixing operators used throughout, and the local subgraph analysis for p=1 MaxCut.","marker":"[4]"},{"why":"Supplies the worst-case p=1 QAOA bound $r = 1 - 0.6924$ on 3-regular graphs via the g6 subgraph, the starting point the iterative method surpasses.","marker":"[13]"},{"why":"Gives the best classical approximation bound $0.9326$ for MaxCut on cubic graphs, the target line the worst-case $r$ converges toward.","marker":"[18]"},{"why":"Defines the DGMVP model, the hard nearest-neighbour mixing operator, and the classical constrained sampling baseline $P_c(n,l)$ used for scaling comparisons.","marker":"[8]"},{"why":"Supplies the permutation-based sparse-state preparation routine used to build the t-order statistic state with $O(tN)$ circuit complexity.","marker":"[23]"},{"why":"Supplies the normalised approximation ratio $r$ used to report all MaxCut results.","marker":"[27]"}],"fun_headline_variants":["Iterative measured superpositions escape QAOA's stuck bound","Warm-started superpositions iterate QAOA past stuck limit","Measurement-derived warm start iteratively improves QAOA","Superposed warm start iteratively escapes QAOA stuck state"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each iteration actually expands the accessible search space, modelled by an exponential growth $d_K = d_0 e^{\\lambda K}$ and a linearly growing thermality coefficient $\\epsilon_{wK} = \\epsilon_{w0} + D K$; if these growth laws fail on larger or adversarial instances, the theoretical reason the iterative method escapes the stuck bound collapses, even though the reported simulations could still be correct.","fun_headline_variants_meta":{"raw":{"variants":["Iterative measured superpositions escape QAOA's stuck bound","Warm-started superpositions iterate QAOA past stuck limit","Measurement-derived warm start iteratively improves QAOA","Superposed warm start iteratively escapes QAOA stuck state"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000315,"raw_usage":{"total_tokens":1733,"prompt_tokens":839,"completion_tokens":894,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":828}},"tokens_in":455,"tokens_out":894,"duration_ms":9171,"temperature":1.0,"reasoning_tokens":828,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:45:50.998169+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a family of 3-regular graphs built entirely from the g6 subgraph, run p = 1 QAOA with a 20th-order statistic warm start, and record $r$ and $R$ after four iterations. The paper's claimed worst-case behaviour predicts $r$ falls below $1 - 0.6924$ and moves toward $1 - 0.9326$; if for some $N \\ge 12$ the measured $r$ plateaus above $0.3076$ or $R$ turns negative, the central claim is contradicted. A more direct check is to estimate the thermality coefficient $\\epsilon_{wK}$ from the measured output distributions: if it does not grow with $K$, the Appendix A mechanism is absent.","supporting_citations":[{"cited_title":"The qaoa gets stuck starting from a good classical string, 2023","cited_arxiv_id":null,"evidence_quote":"Establishes the stuck issue: a single-string warm-started QAOA cannot improve beyond a thermality bound; this is the baseline the paper claims to escape."},{"cited_title":"A quantum approximate optimization algorithm, Novem- ber 2014","cited_arxiv_id":null,"evidence_quote":"Introduces the standard QAOA ansatz with cost and mixing operators used throughout, and the local subgraph analysis for p=1 MaxCut."},{"cited_title":"Maxcut quantum ap- proximate optimization algorithm performance guaran- tees for p >1","cited_arxiv_id":null,"evidence_quote":"Supplies the worst-case p=1 QAOA bound $r = 1 - 0.6924$ on 3-regular graphs via the g6 subgraph, the starting point the iterative method surpasses."},{"cited_title":"Max cut in cubic graphs","cited_arxiv_id":null,"evidence_quote":"Gives the best classical approximation bound $0.9326$ for MaxCut on cubic graphs, the target line the worst-case $r$ converges toward."},{"cited_title":"Long, Hugo V","cited_arxiv_id":null,"evidence_quote":"Defines the DGMVP model, the hard nearest-neighbour mixing operator, and the classical constrained sampling baseline $P_c(n,l)$ used for scaling comparisons."},{"cited_title":"Lefterovici, and Anto- nio F","cited_arxiv_id":null,"evidence_quote":"Supplies the permutation-based sparse-state preparation routine used to build the t-order statistic state with $O(tN)$ circuit complexity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the normalised approximation ratio $r$ used to report all MaxCut results."}],"review_version":1}