{"id":"c82704d1-714f-4cbf-9196-9307d865a571","arxiv_id":"2608.12711","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"SPLIT-Q solves controlled islanding by sequentially optimizing small regional QUBO subproblems on a bounded-width QAOA circuit and reports Gurobi-optimal cuts on eleven IEEE benchmarks.","lead":"SPLIT-Q is a quantum optimization framework that splits a damaged power grid into safe islands by solving one small region at a time on a fixed-size quantum circuit. On eleven benchmark grids from 9 to 300 buses it reports matching the best classical islanding cuts while using far fewer qubits than a single monolithic quantum circuit would need.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantum attribution is unestablished: with MSSP initialization, retained current assignment, classical fallback candidates, and repair enabled, no experiment shows QAOA samples caused the reported Gurobi-optimal cuts.","rationale":"I read the paper in good faith: the sequential decomposition is coherent, the bounded-register construction is sound, and the resource-scaling argument in Sec. IV.D is a genuine contribution. The strongest claim is an empirical statement about matching Gurobi optima. My concern is not that the numerical values are false or that the method is internally inconsistent; it is that the paper has not established the causal role of the quantum sampler in producing those values. The reader's weakest_assumption identifies exactly this issue: the candidate set always contains the current regional assignment, classical fallback assignments can supplement sparse samples, and the safe-repair map can classically convert infeasible assignments. I agree with that diagnosis. I add precise locations: Algorithm 1 lines 1, 2, and 11, Sec. III.E's classical fallback, and Sec. III.F's safe-repair map. A single ablation experiment, running the same pipeline with QAOA disabled, would settle whether the hardware execution is load-bearing. Because this experiment has not been reported, the empirical quantum-contribution claim is unverified; however, this does not invalidate the resource-reduction contribution or the constructive decomposition. The reader's CONDITIONAL verdict is therefore appropriate, and I would keep it unchanged pending the proposed ablation.","tokens_in":15581,"tokens_out":3492,"duration_ms":49176,"concrete_test":"Rerun Algorithm 1 on all eleven systems with the quantum stage disabled: set Cand(F_m,p,N_s,B) to return only {x_Rm} plus the same classically generated fallback assignments and repairs described in Sec. III.E, so no QUBO coefficients are sent to QAOA. If the best feasible cut still equals the Gurobi optimum for every system, then the hardware columns in Table III are not evidence that QAOA contributed to solution quality; the claim should be revised to describe a hybrid pipeline with classical bookkeeping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim (Sec. IV.A: 'For every reported benchmark, the best cut value obtained by the proposed method on the IBM quantum processor matches the corresponding Gurobi optimum') is not the part I doubt; the eleven reported equalities may well hold. The load-bearing condition is that the quantum stage is causally responsible for reaching those optima, because the claimed noise resilience and the broader quantum-optimization contribution rest on that attribution. Algorithm 1 undermines this condition in four places. First, line 1 initializes the working assignment with MSSP, and line 2 archives it as x_best if feasible; if the MSSP label expansion already achieves the Gurobi cut, the rest of the algorithm is redundant. Second, line 11 always includes the current assignment x_Rm in the candidate set, so the sequential update can only keep or improve the existing classical assignment. Third, Sec. III.E explicitly allows classically generated assignments to complete the candidate pool when sampling produces too few candidates. Fourth, the safe-repair map Rep in Sec. III.F can replace an infeasible raw quantum assignment with a deterministic classical repair and selects the better of the two; final feasibility is therefore established by the DFS test and repair, not by the raw samples. No table or figure reports the provenance of the winning candidate (initial MSSP vs. classical fallback vs. repair vs. raw quantum sample), and no ablation disables the quantum sampler while leaving the classical coordinator, fallback pool, and repair in place. The paper itself does not claim a general optimality guarantee (Sec. III.H), but the empirical claim of 'recovering' optima 'using real quantum hardware' implies the hardware execution contributed. On the evidence presented, the identical CI=CN=CH=C* entries are equally consistent with a purely classical pipeline plus bounded-width QAOA circuits that never affect the answer.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SPLIT-Q, a sequential regional QAOA framework for controlled islanding in power systems. The method fixes coherent generator groups as anchors, encodes only free buses with a symmetry-reduced compact representation, partitions free buses into regions whose qubit width is capped by a user budget Qmax, and solves boundary-conditioned regional QUBOs sequentially while a Lagrangian-style multiplier loop coordinates global resource constraints. Feasibility is enforced by a global DFS connectivity check and a safe-repair map, and the best feasible archive is retained. The authors evaluate the framework on eleven IEEE systems from 9 to 300 buses using ideal simulation, a calibrated-noise simulator, and IBM hardware, reporting that all returned cut values match Gurobi reference optima while the simultaneous quantum register stays between 3 and 18 qubits. The paper also gives a scaling analysis showing per-circuit width independent of N, aggregate quantum workload O(pN) on bounded-degree graphs, and classical feasibility evaluation O(TBN^2/Qmax) as the dominant large-scale cost.","tokens_in":15939,"tokens_out":2862,"duration_ms":40081,"significance":"If the attribution of the reported optimal cuts to the quantum sampling stage is established, the paper makes a useful practical contribution: it provides an explicit construction that decouples QAOA circuit width from network size, a clear coordinate-descent-style coordination loop, and a transparent resource analysis. The regional QUBO derivation in Eqs. (26)-(32) is explicit, the archive-preservation and exhaustive-recovery propositions are correctly stated, and the paper is honest in Section III.H that finite sequential sampling does not guarantee global optimality. The main weakness is that the central empirical claim, that the IBM quantum processor recovers Gurobi-optimal partitions and exhibits noise resilience, is not backed by any experiment isolating the contribution of the quantum samples. Because Algorithm 1 always retains the classical current assignment, admits classically generated fallback candidates, and applies a classical repair map, the reported equality between hardware cut values and Gurobi optima may hold even if the quantum stage contributes nothing to solution quality.","major_comments":[{"comment":"The central claim that the IBM quantum processor recovers Gurobi-optimal partitions is not supported as a statement about the quantum component. Algorithm 1 initializes the working assignment with MSSP (line 1) and archives it if feasible (line 2); line 11 always includes the current regional assignment in the candidate set; Section III.E explicitly allows classically generated assignments to fill the candidate pool; and Section III.F allows the safe-repair map to replace an infeasible raw quantum assignment with a deterministic classical repair. The paper reports no ablation that disables quantum sampling and no table or figure recording the provenance (initial MSSP vs. classical fallback vs. repair vs. raw quantum sample) of the winning candidate. Without such data, the identical entries in Table III across CI, CN, CH, and C* do not demonstrate noise resilience of QAOA; they are equally consistent with the classical bookkeeping alone finding the optimal cut for all eleven systems. This is load-bearing because the abstract and Section IV.C claim that the results 'confirm the resilience of its solution quality' under noise.","section":"Section IV.A, Table III, and Algorithm 1"},{"comment":"The regionalization is a deterministic chunking of buses by increasing bus index, and the paper acknowledges in Section III.B that a graph partitioner could replace it without changing the method. However, the reported optimality on all eleven systems may depend on this regional ordering, since the boundary-conditioned QUBO (Eq. (27)) and the sequential update order in Algorithm 1 are both sensitive to which buses are grouped. The paper gives no sensitivity analysis, no alternative regionalizer experiments, and no discussion of whether the bus-index ordering is itself a tuned choice. As the method is presented as a general scalable framework, the dependence of the main empirical claim on this untested design choice should be characterized.","section":"Section III.B, Eq. (24), and Section IV.A"},{"comment":"The QPU-time comparison is presented as a favorable trade-off, but for the smaller systems the proposed method is slower than the monolithic baseline (e.g., 4 s vs. 2 s at 9 buses, 23 s vs. 6 s at 57 buses), and the comparison for the larger systems is against N/A entries where the monolithic circuit does not fit on the device. The statement in Section IV.B that 'the increase remains moderate in absolute terms' is reasonable, but the asymmetry of the comparison should be stated more carefully: the proposed method is not faster in wall-clock QPU time where both fit; it is the only method that fits at all for the three largest systems. This is a presentation issue rather than a technical error.","section":"Section IV.B, Table IV"}],"minor_comments":[{"comment":"The table header formatting is hard to parse: columns for K, p, Ns, B, nmax, Qmax, Tmin/Tmax, L, Imax, and T are run together with comma-separated values, making it difficult to map entries to systems. A row-per-system layout with explicit column headers would improve readability.","section":"Table I"},{"comment":"The text says 'T denotes the number actually completed,' but Algorithm 1 uses t as the loop index and Tmax as the upper bound; the symbol T is not introduced in the algorithm pseudocode. Please align the notation so the stopping criterion is unambiguous.","section":"Section III.G and Algorithm 1"},{"comment":"The phrase 'the arrow places larger k to the left' is terse; rewriting the product with explicit bounds, e.g., exp(-i beta_p H_M) ... exp(-i beta_1 H_M), would remove ambiguity for readers unfamiliar with the arrow convention.","section":"Eq. (35)"},{"comment":"The caption refers to 'circled numbers' giving the execution order, but the figure as rendered does not clearly show circled numbers in the workflow. Please either add visible numbered markers or revise the caption to describe the ordering in text.","section":"Fig. 1"},{"comment":"The sentence introducing Table II says the complete bus-to-island assignments 'document the island memberships,' but the table caption and text do not state whether these are the hardware results or the Gurobi reference solutions; please clarify that these are the reconstructed partitions returned by the proposed method.","section":"Section IV.A"}],"recommendation":"major_revision","confidential_remarks":"The paper's reported invariance of CI, CN, CH, and C* across all eleven systems is remarkable and, without provenance data, will likely be read by reviewers as evidence that the classical fallback and repair mechanisms, rather than QAOA, determine the final solutions. I recommend requiring the authors to provide an ablation with quantum sampling disabled, as well as per-run provenance of the winning candidate, before acceptance. The reliance on closely related prior work [24] and [25] for the encoding and multiplier steps is disclosed, but the incremental contribution over those papers should be stated more crisply in the final version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the sequential regional decomposition is a legitimate engineering contribution, and the resource-scaling claims hold up as stated. What is not established is the empirical claim that hardware QAOA sampling recovers Gurobi-optimal partitions; nothing in the paper rules out the classical parts of the pipeline doing all the work.\n\nWhat is actually new: a fixed-width regional QAOA formulation with boundary-conditioned regional QUBOs and Lagrangian coordination, plus an aggregate-workload argument that is linear in network size for bounded-degree graphs. The derivation of the regional QUBO in equations (26) through (32) is explicit, and I do not see a missing term. The archive-preservation result is simple but correct, and the width reduction versus monolithic encoding is real. On the smaller systems, the reported gate and depth reductions are consistent with the bounded-width construction.\n\nThe main soft spot is attribution. Algorithm 1 initializes with the MSSP assignment and archives it as best if feasible; it always includes the current assignment in the candidate pool; it allows classically generated candidates when sampling is sparse; and the safe-repair map can replace an infeasible raw assignment with a deterministic classical repair. No table reports the provenance of the winning candidate, and there is no ablation with the QAOA stage disabled. Identical cut values across ideal, noisy, hardware, and Gurobi could indicate noise resilience, but they could equally mean the quantum component never changes the answer. I would not call the central claim circular—the Gurobi benchmark is independent, and the problem formulation is explicit—but the empirical claim that optimal partitions were recovered \"using real quantum hardware\" is currently under-supported.\n\nMinor weaknesses: no run-to-run statistics, no sensitivity analysis for the listed hyperparameters, and the hardware-time comparison against monolithic QAOA is possible only for systems up to 89 buses because larger monolithic circuits do not fit. That limitation is inherent, but it weakens the scalability comparison.\n\nWho this is for: quantum-optimization researchers and power-system people interested in whether fixed-width hybrid methods can handle larger graph-partitioning instances. They should read it with the attribution caveat in mind. The paper deserves a serious referee, not a desk reject. I would ask for a classical-only control—MSSP initialization plus coordinator, fallback candidates, and repair, with no QAOA samples—and for per-run provenance of the winning candidate before accepting.","headline":"The decomposition and scaling analysis are solid, but the paper never shows the QAOA samples—rather than the always-retained classical assignments and repair—are what produce the Gurobi-optimal cuts.","tokens_in":16549,"tokens_out":2854,"would_cite":false,"duration_ms":38790,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","90C27","05C85"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a sequential regional QAOA framework recovers Gurobi-optimal controlled-islanding cuts on all eleven tested IEEE systems while keeping the quantum register to at most 18 qubits.","keywords":["controlled islanding","QAOA","quantum optimization","graph partitioning","power system resilience","QUBO","sequential optimization","NISQ"],"falsifier":"Run the same pipeline with the quantum sampling step disabled, keeping only the MSSP initialization, the always-included current regional assignment, classical fallback candidates, and the safe-repair map, over all eleven IEEE systems; if these quantum-disabled runs also match the Gurobi optimum in every case, the hardware execution contributes nothing to the reported solution quality.","tokens_in":15368,"feed_emoji":"⚡","tokens_out":5456,"duration_ms":64217,"temperature":0.7,"pith_summary":"This paper tries to establish that controlled islanding, the emergency splitting of a damaged power grid into connected and self-sufficient islands, can be handled by a quantum optimization method whose required qubit count does not grow with network size. The method cuts the monolithic islanding problem into boundary-conditioned regional subproblems, solves each as a small QUBO with QAOA, and coordinates the regions classically. Across eleven IEEE networks from 9 to 300 buses, the best cut found on IBM quantum hardware matches the Gurobi optimum in every reported case while satisfying all islanding constraints. A sympathetic reader cares because, if this holds, current noisy quantum hardware could address large combinatorial power-system problems that monolithic quantum formulations cannot even fit onto a device.","feed_headline":"Quantum islanding hits exact optima on grids up to 300 buses","feed_subtitle":"A sequential QAOA splits grids into islands while keeping circuit width to at most 18 qubits.","key_machinery":"The carrying mechanism is the boundary-conditioned regional QUBO: at each update, only the free buses of one region become quantum variables, with all external endpoints fixed by the latest complete assignment, so the register width is bounded by construction. A compact symmetry-reduced encoding uses $K-1$ qubits per free bus, projected Lagrangian multipliers carry island-wide resource-count violations into the regional objective without adding slack qubits, and a safe-repair map together with a depth-first-search connectivity check turns raw assignments into feasible global candidates. This machinery transfers the burden of network growth from circuit width to the number of regional subproblems and to classical feasibility checking.","core_discovery":"The central claim is that decomposing a monolithic quantum search over all bus assignments into a sequence of bounded-width regional QUBO subproblems does not cost solution quality. On all eleven reported IEEE systems, the best cut value obtained by the proposed method on the IBM quantum processor equals the corresponding Gurobi optimum while all islanding constraints are satisfied, and the simultaneous quantum register stays between 3 and 18 qubits. The paper further reports that ideal simulation, IBM-calibrated noisy simulation, and real IBM hardware sampling return identical cut values, which it reads as evidence of noise resilience for the final decoded solutions. The paper is explicit that this is empirical evidence for the tested configurations, not a general convergence guarantee.","pith_inferences":["A direct ablation would isolate the quantum contribution: rerun the same pipeline with QAOA sampling disabled, keeping only the MSSP initialization, the always-included current regional assignment, classical fallback candidates, and the safe-repair map; the paper does not report such a comparison.","The regionalization idea should transfer to other constrained graph-partitioning problems with fixed anchor sets, such as load balancing, community detection, or VLSI partitioning, where a bounded-width QAOA may similarly trade circuit width for more classical feasibility checks.","The reported noise resilience concerns the final decoded cut rather than the quantum state itself; because candidates are embedded into a global assignment and ranked by feasibility before selection, noise can be absorbed as long as the candidate pool contains at least one good assignment, making the framework better described as a noise-tolerant hybrid sampler."],"forward_implications":["On bounded-degree networks, the aggregate quantum workload per coordination sweep scales linearly with network size while each individual circuit stays within the fixed qubit budget, so a fixed-capacity quantum device remains applicable as the grid grows.","For the 118-, 145-, and 300-bus systems, the monolithic symmetry-reduced formulations require 192, 285, and 462 qubits and thus exceed the 156 physical qubits of the device, while the regional method still runs with 18, 18, and 12 qubits respectively.","If the reported hardware results are representative, hybrid quantum-classical pipelines can match an exact classical solver on these islanding benchmarks while using far smaller compiled circuits.","The identical cut values across ideal simulation, calibrated-noise simulation, and IBM hardware imply that, under the tested noise conditions, the sequential coordination and feasibility-first candidate ranking preserve the decoded solution quality.","The method is a heuristic: finite sequential sampling does not guarantee global optimality, although archive preservation ensures the best feasible cut found so far is never lost."],"supporting_citations":[{"why":"Supplies the compact symmetry-reduced free-bus encoding and the Lagrangian multiplier strategy for coupling constraints that the regional QUBO is built on.","marker":"[25]"},{"why":"Supplies the earlier resource-efficient hybrid QAOA framework and the structured postprocessing logic that the sequential method extends.","marker":"[24]"},{"why":"Provides the quantum-annealing approach to power-grid partitioning that motivates the gate-based alternative pursued here.","marker":"[23]"},{"why":"Establishes the weighted-cut formulation of controlled islanding that the disruption objective C(x) is based on.","marker":"[4]"},{"why":"Provides the slow-coherency generator grouping concept used to fix the coherent anchor groups H.","marker":"[5]"},{"why":"Provides the strong mixed-integer formulation baseline that the Gurobi reference solutions are expected to represent.","marker":"[11]"}],"fun_headline_variants":["Sequential QAOA matches Gurobi optimal cuts on 300-bus grids","Qubit-bounded quantum islanding hits exact optima, ≤18 qubits","SPLIT-Q: quantum grid splitting matches exact solutions with limited qubits","Scalable quantum islanding: exact optima, bounded circuit width","Quantum optimizer splits grids into optimal islands within qubit limits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the quantum-generated candidates are responsible for the reported Gurobi-optimal cuts, even though the candidate pool always includes the current regional assignment, can be supplemented by classically generated fallbacks, and passes through a safe-repair map that may replace infeasible raw assignments with classical repairs.","fun_headline_variants_meta":{"raw":{"variants":["Sequential QAOA matches Gurobi optimal cuts on 300-bus grids","Qubit-bounded quantum islanding hits exact optima, ≤18 qubits","SPLIT-Q: quantum grid splitting matches exact solutions with limited qubits","Scalable quantum islanding: exact optima, bounded circuit width","Quantum optimizer splits grids into optimal islands within qubit limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1403,"prompt_tokens":955,"completion_tokens":448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":350}},"tokens_in":571,"tokens_out":448,"duration_ms":5480,"temperature":1.0,"reasoning_tokens":350,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T01:01:35.390754+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pipeline with the quantum sampling step disabled, keeping only the MSSP initialization, the always-included current regional assignment, classical fallback candidates, and the safe-repair map, over all eleven IEEE systems; if these quantum-disabled runs also match the Gurobi optimum in every case, the hardware execution contributes nothing to the reported solution quality.","supporting_citations":[{"cited_title":"PACE-QAOA: Physics-Constrained Quantum Optimization for Qubit-Efficient Power System Islanding","cited_arxiv_id":"2608.02789","evidence_quote":"Supplies the compact symmetry-reduced free-bus encoding and the Lagrangian multiplier strategy for coupling constraints that the regional QUBO is built on."},{"cited_title":"Quantum annealing based power grid partitioning for parallel simulation,","cited_arxiv_id":null,"evidence_quote":"Provides the quantum-annealing approach to power-grid partitioning that motivates the gate-based alternative pursued here."},{"cited_title":"Con- trolled islanding solution for large-scale power systems,","cited_arxiv_id":null,"evidence_quote":"Establishes the weighted-cut formulation of controlled islanding that the disruption objective C(x) is based on."},{"cited_title":"Slow coherency-based islanding,","cited_arxiv_id":null,"evidence_quote":"Provides the slow-coherency generator grouping concept used to fix the coherent anchor groups H."},{"cited_title":"Strong mixed-integer formulations for power system islanding and restoration,","cited_arxiv_id":null,"evidence_quote":"Provides the strong mixed-integer formulation baseline that the Gurobi reference solutions are expected to represent."}],"review_version":1}