{"id":"9c2de514-bee8-4e69-a4e0-a2870360cc5c","arxiv_id":"2505.18435","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A fixed-depth branch-and-bound algorithm using QC relaxation lower bounds narrows the optimality gap on small PGLib cases, without proof of global optimality.","lead":"This paper wraps a quadratic convex relaxation inside a branch-and-bound search to tighten lower bounds for the AC optimal power flow problem. It shows bound improvements on small test grids, but it does not certify global optimality and is not compared with existing global solvers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fixed-depth termination without a coverage certificate invalidates the claimed BB-QC global lower bound; Algorithm 1's pruning condition and unchanged QC_lower_bound make the reported gaps unreproducible.","rationale":"The reader's rejection is justified. The single most load-bearing issue is lower-bound certification: the paper claims to tighten the QC lower bound and to preserve the potential for global optimality, but Algorithm 1 terminates at a fixed depth without proving that the unexplored or pruned regions cannot contain a better solution. My read agrees with the reader that the coverage assumption is the weakest point. I add two concrete internal inconsistencies that make this assumption especially damaging: line 19 contains a pruning condition that is impossible for valid relaxations, and the returned QC_lower_bound is never updated, so the reported BB-QC gaps cannot be derived from the pseudocode as written. I would not change the reader's verdict: the central claim is unsupported by the algorithm and experiments as presented. My agreement is partial rather than full because I emphasize that the coverage problem is compounded by the algorithm never actually computing the quantity reported in Table I, and because the 'best_solution' update uses infeasible relaxed values, further severing the link between the pseudocode and the claimed global-optimality result.","tokens_in":17809,"tokens_out":3789,"duration_ms":35592,"concrete_test":"Re-run case30_ieee with a faithful implementation of Algorithm 1 in PowerModels.jl, recording every leaf node after level 71. First, check whether any child satisfies fchild < QC_lower_bound; if so, the pruning rule deletes valid regions and the coverage invariant fails. Then compute L* = min over all kept child leaves of their QC objective values. Because the leaf boxes partition the original box when no invalid pruning occurs, L* is a valid global lower bound. Compare L* with the AC upper bound 8208.51: the reported BB-QC gap is 0.01%, so L* must be within 0.01% of 8208.51. If L* is lower, or if the algorithm's returned QC_lower_bound is used instead of L*, the claimed lower-bound certificate in Table I is not reproducible. Optionally, compare L* against an independent SDP relaxation of case30_ieee to check plausibility.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the BB-QC gap is a certified global lower bound after a fixed number of branching levels rests on an unstated coverage invariant: the kept child boxes, together with boxes pruned by comparison with the AC upper bound, must still partition the original QC feasible region. Algorithm 1 does not establish this invariant. It stops at fixed depth n with no optimality certificate, and line 19 prunes a child whenever fchild < QC_lower_bound. For a valid relaxation, restricting a subregion cannot lower the QC objective below the root lower bound, so this condition is impossible in exact arithmetic; if implemented literally, it would delete valid regions and break coverage. The pseudocode also never updates QC_lower_bound after initialization, so the returned lower bound is just the root relaxation value, not the minimum child lower bound that the reported BB-QC gaps in Table I appear to use. In addition, Algorithm 1 updates 'best_solution' with fchild values from the QC relaxation, which are lower bounds and generally not feasible AC-OPF solutions, so the claimed 'Best feasible solution' output is not feasible. Therefore the reported tightening from a 22.96% gap to a 0.01% gap on case30_ieee is not supported as a global optimality certificate; at best it is a heuristic observation about some child relaxation values.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a branch-and-bound assisted Quadratic Convex (QC) relaxation framework for AC-OPF. The idea is to start from the standard QC relaxation, obtain an AC-feasible upper bound from a local solver, and then repeatedly bisect one voltage magnitude or angle-difference variable at each level, solving the QC relaxation on each child subregion and pruning children whose relaxed objective appears too high. The authors report on PGLib-OPF benchmark cases that the method reduces the optimality gap from the root QC relaxation to very small values (e.g., from 22.96% to 0.01% for case30_ieee) while exploring only a modest number of subproblems. The central claim is that the approach preserves the potential to reach the global optimum and provides a much tighter lower bound than the root QC relaxation.","tokens_in":18044,"tokens_out":4421,"duration_ms":34813,"significance":"If the algorithmic guarantees claimed in the abstract were correct, this would be a meaningful contribution: it would show that a QC-based spatial branch-and-bound can systematically tighten loose relaxations on public PGLib benchmarks at far lower cost than SDP-based branch-and-bound. The paper has notable strengths: it is grounded in public benchmark instances, uses the standard PowerModels.jl framework, and the QC relaxation is taken from the literature without fitting parameters to the reported gaps. However, the central lower-bounding claim is not established. The algorithm as written terminates at fixed depth without a coverage certificate, the pruning logic contains an impossible condition, the returned lower bound is never updated, and the reported child objective values are relaxed quantities, not feasible AC-OPF solutions. The empirical gaps are therefore not certified global optimality gaps; at best, the paper reports a heuristic observation about some child relaxation values.","major_comments":[{"comment":"The algorithm terminates after a fixed number of levels n (Algorithm 1, line 7) without any certificate that the union of the kept child regions and the pruned regions covers the original QC-feasible set. The paper never states or proves a coverage invariant. Consequently, the BB-QC gap reported in Table I (e.g., 0.01% for case30_ieee vs. a 22.96% root QC gap) is not a certified global lower bound for the original AC-OPF problem. The abstract's claim of 'preserving the potential to reach the global optimum' is therefore unsupported by the described algorithm.","section":"Algorithm 1 and Section IV-A, Table I"},{"comment":"Line 19 prunes a child when fchild < QC_lower_bound. For a valid relaxation, restricting a subregion can only increase or leave unchanged the optimal value of the relaxed problem relative to the root relaxation; the relaxed objective cannot fall below the root lower bound in exact arithmetic. This condition is therefore impossible for a correctly implemented relaxation, indicating that the pseudocode does not correspond to the implemented algorithm. If the condition is implemented literally, it would discard valid regions and break the coverage property needed for any global lower bound.","section":"Algorithm 1, line 19"},{"comment":"The variable QC_lower_bound is initialized at line 3 and never updated inside the loop. The returned lower bound is therefore the root relaxation value, not the minimum lower bound over active children. Section IV-A reports BB-QC gaps far smaller than the root QC gap (e.g., 0.01% vs. 22.96% for case30_ieee), so those numbers cannot be derived from Algorithm 1 as written. Either the pseudocode omits the essential update of the lower bound to the minimum over child nodes, or the reported results come from a different computation. The empirical central claim is consequently unreproducible.","section":"Algorithm 1, lines 3, 7-30"},{"comment":"The algorithm updates 'best_solution' whenever fchild is lower than the current best, where fchild is the optimal value of the QC relaxation over the child region. The QC relaxation is a lower-bounding problem; its optimal value is not in general attainable by any AC-feasible point, and Section III-B itself states that the subproblem is solved 'to find a local lower bound' on the objective function. Thus the returned 'Best feasible solution' (Algorithm 1, Ensure) is not shown to be feasible for the original AC-OPF. This invalidates the interpretation of child objective values as candidate feasible solutions in Figures 6-12 and Section IV-D.","section":"Algorithm 1, lines 23-24; Section III-B"},{"comment":"The branching variable selection rule is never specified. Algorithm 1 says 'Select variable v_l in V to split' without defining the rule, and the text gives no criterion beyond 'one variable to divide.' The numerical results in Table I and Figures 7-12 depend on this choice, so the experiments are not reproducible. A precise rule, such as largest bound range, most nonconvexity, or a fixed default with tie-breaking, must be stated.","section":"Section III-B, branching strategy"}],"minor_comments":[{"comment":"There are several typos and placeholder references, including 'teh problem', '[]' placeholders, and 'P BA' / 's.t. l=i' formatting issues in Section II. These should be corrected.","section":"Section III-B"},{"comment":"The captions for Figures 8, 9, and 10 describe 'three branching levels' even though the horizontal axes show 14, 30, and 57 levels, respectively. The captions should match the actual number of levels plotted.","section":"Figures 8-10"},{"comment":"The definition of the optimality gap in Equation (8) uses 'Local Solution' and 'Lower Bound' without clarifying which values are used for the 'BB-QC' gaps in Table I. Since the lower bound is not updated in Algorithm 1, the reader cannot tell what quantity the reported 'BB-QC gap' represents.","section":"Equation (8)"},{"comment":"The midpoint computation uses 'lower_bound' and 'upper_bound' without an index, which is ambiguous when multiple variables have different bounds. Clarify that these refer to the bounds of the selected variable v_l.","section":"Algorithm 1, lines 11-14"},{"comment":"The implementation states 'Gurobi 8.0' as the solver; if a newer version was used, that should be corrected. The paper should also report solver tolerances and any feasibility tolerances used when solving the AC-OPF upper-bound subproblem.","section":"Section IV, implementation details"}],"recommendation":"reject","confidential_remarks":"The central claims of this manuscript overreach the algorithm actually described. The fixed-depth termination without a coverage certificate, the impossible pruning condition in Algorithm 1 line 19, the never-updated lower bound, and the use of relaxed child objectives as candidate feasible solutions together invalidate the reported 'BB-QC gap' as a global optimality certificate. These are load-bearing issues, not presentation problems. A substantial rewrite would be needed to either provide a valid lower-bounding algorithm with a termination certificate or to reframe the method as a heuristic and remove all global-optimality language. Given the current wording, rejection is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a standard spatial branch-and-bound with QC relaxation lower bounds, and the paper overstates what is new. The useful part is the empirical observation that child-subregion lower bounds tighten on small PGLib cases; the unsupported part is the global-optimality and scalability story. The cited literature already contains SDP- and SOCP-based spatial B&B [17, 13], and Barati's QC-based global method [14] is essentially the same direction.\n\nTo its credit, the paper uses public benchmarks, reports runtimes and node counts, and the child-objective visualizations are a genuinely nice diagnostic. The three-bus feasible-region plots are also useful for understanding bound tightening. Self-citations are appropriate and the core QC relaxation is standard.\n\nThe soft spots are serious. The algorithm stops at a fixed depth with active nodes left, and no coverage certificate is given, so the reported BB-QC gap is not a proven global lower bound. Algorithm 1 never updates QC_lower_bound, so the Table I gap must be coming from child fchild values—relaxed lower bounds, not feasible AC solutions. The output called 'best feasible solution' is therefore not a feasible solution. Line 19's pruning condition fchild < QC_lower_bound cannot occur for a valid relaxation of a subregion, so either it is dead code or the implementation does something other than what is written. The branching rule is unspecified, and no comparison is made with existing global solvers. Four point nine hours for a 57-bus case does not support the word 'scalable.' Placeholder citations and typos also suggest the manuscript was not cleaned up.\n\nIf the authors fixed the certificate issue—by tracking the minimum leaf lower bound, explicitly stating the coverage invariant, and comparing against [14] and [17]—the empirical lower-bound tightening could be a useful incremental result. As it stands, the central claim of global optimality is not supported.\n\nWho this is for: readers working on QC bound tightening might mine the histograms and candle charts for diagnostic ideas. I would not cite it as a method. It deserves a serious referee because AC-OPF global optimization is important and the experiments are real, but I would expect heavy revision or rejection in current form.","headline":"Standard spatial B&B with QC-relaxation bounds; the empirical lower-bound tightening is real on small cases, but the global-optimality claim and the pseudocode do not hold up.","tokens_in":18596,"tokens_out":5820,"would_cite":false,"duration_ms":55149,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Branching on voltage magnitudes and angle differences turns the loose quadratic convex relaxation of AC-OPF into a near-global certificate on benchmark cases.","keywords":["optimal power flow","AC-OPF","quadratic convex relaxation","branch-and-bound","global optimization","convex envelopes","lower bounds"],"falsifier":"Take a benchmark with a known certified global optimum, run Algorithm 1 as written, and compare its returned QC lower bound against that global optimum: if the returned lower bound exceeds the global optimum, or if a region containing the optimum was pruned, the claim that the method preserves global optimality fails. A direct check is to enumerate the kept child boxes after the final level and verify that their union, together with the pruned boxes, covers the original voltage and angle box.","tokens_in":17584,"feed_emoji":"⚡","tokens_out":8015,"duration_ms":70133,"temperature":0.7,"pith_summary":"This paper proposes a hybrid scheme for the alternating-current optimal power flow problem (AC-OPF): a branch-and-bound search that splits voltage magnitudes and angle differences into smaller intervals, with a quadratic convex (QC) relaxation solved in each subregion to obtain a lower bound. Subregions whose lower bound already exceeds the cost of a feasible AC solution are pruned, and the surviving regions are split again for a fixed number of levels. The paper claims that this shrinks the optimality gap from 22.96 percent to 0.01 percent on case30_ieee and to zero or near zero on several other benchmark cases, while keeping the number of subproblems in the thousands rather than an exponential explosion. If these bounds hold, the method gives power-system operators a practical way to certify that a fast local solution is globally near-optimal.","feed_headline":"Voltage splits shrink OPF optimality gap from 23% to 0.01%","feed_subtitle":"A quadratic convex relaxation plus branch-and-bound nearly certifies global optimality on benchmark power systems.","key_machinery":"The central object is the quadratic convex (QC) relaxation of the AC power flow equations, in which squared voltage magnitudes, bilinear voltage products, and sine and cosine terms are replaced by convex envelopes. The branch-and-bound step splits the interval of one selected variable, either a voltage magnitude or an angle difference, at its midpoint into two child regions, solves the QC relaxation on each child to obtain a valid lower bound, and prunes children whose lower bound exceeds the feasible AC-OPF upper bound. This converts a single loose global relaxation into many smaller, tighter relaxations whose minimum surviving lower bound is reported as the BB-QC bound.","core_discovery":"The paper's central claim is that a branch-and-bound process built on the QC relaxation can reduce the number of subproblems explored while preserving the potential to reach the global optimum. The key reported effect is that branching on voltage-related variables makes the relaxation's lower bound much tighter: the paper's tables show the optimality gap dropping from 22.96 percent to 0.01 percent on case30_ieee, from 23.69 percent to 3.89 percent on case14_ieee_sad, and to zero on several other benchmark cases, while the number of child nodes solved stays between roughly a dozen and thirty thousand.","pith_inferences":["The fixed-depth tree is better interpreted as a lower-bound improvement heuristic than as a complete global-optimality certificate: a rigorous certificate would require either full coverage of unpruned regions or an explicit bound on what was discarded, and the paper's 'potential to reach the global optimum' language leaves this open.","The pseudocode's prune condition that rejects a child whose QC objective is below the root QC lower bound is unreachable for a valid relaxation, since a child relaxation can only be tighter and therefore have a lower bound at least as large as the root's; this suggests the documented pruning logic may not reliably produce the reported gaps as valid global lower bounds.","If the observed lower-bound tightening is real, the same splitting scheme could be tested on other non-convex power-system variables, such as transformer tap ratios or line-switching decisions, where convex relaxations also suffer from loose bounds.","The paper's histogram diagnostics could be turned into an adaptive branching rule: choose the variable whose split produces the largest rightward shift in normalized child costs, which the paper's own analysis suggests would accelerate convergence."],"forward_implications":["The reported BB-QC gaps imply that a local AC-OPF solution can be certified as within 0.01 percent of global optimality on case30_ieee, where the plain QC relaxation leaves a 22.96 percent gap.","Because the tree depth is fixed and only one variable splits per level, the number of subproblems stays manageable, suggesting the approach can scale to systems where full branch-and-bound is impractical.","The improvement relies only on refining voltage-related bounds, which the paper argues concentrates the search on the variables that drive OPF's non-convexity.","The lower-bound tightening can be separated from the upper-bound search, so the method can be combined with any feasible AC-OPF solver to produce a near-global certificate."],"supporting_citations":[{"why":"Defines the QC relaxation with convex envelopes that supplies the lower bounds at every branch-and-bound node.","marker":"[19]"},{"why":"Introduces the spatial branch-and-bound template for OPF that this paper adapts from SDP-based bounding to QC-based bounding.","marker":"[17]"},{"why":"Provides the McCormick envelopes for bilinear products used by the QC relaxation.","marker":"[20]"},{"why":"Shows how convex envelopes and bound tightening strengthen relaxations of AC power flows, informing the QC model.","marker":"[7]"},{"why":"Prior demonstration of global optimization of OPF through a quadratic convex relaxation, the line of work this paper extends.","marker":"[14]"},{"why":"Supplies the benchmark cases whose reported gaps measure the method's performance.","marker":"[25]"},{"why":"Provides the modeling framework used to implement and test the algorithm.","marker":"[27]"}],"fun_headline_variants":["Voltage branching cuts OPF gap from 23% to 0.01%","QC relaxation plus branch-and-bound certifies global OPF on benchmarks","B&B-assisted QC relaxation slashes OPF optimality gap","Prune subregions, keep global optimum: new OPF solver","Branch-and-bound tightens OPF relaxation to near-zero gap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's reported optimality gaps are valid only if, after the fixed number of branching levels, the kept child regions cover the whole feasible set except regions that were safely pruned; the paper does not specify a branching-variable rule or prove this coverage, and its pseudocode includes a pruning condition that is unreachable for a valid relaxation.","fun_headline_variants_meta":{"raw":{"variants":["Voltage branching cuts OPF gap from 23% to 0.01%","QC relaxation plus branch-and-bound certifies global OPF on benchmarks","B&B-assisted QC relaxation slashes OPF optimality gap","Prune subregions, keep global optimum: new OPF solver","Branch-and-bound tightens OPF relaxation to near-zero gap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000558,"raw_usage":{"total_tokens":2667,"prompt_tokens":970,"completion_tokens":1697,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":1613}},"tokens_in":586,"tokens_out":1697,"duration_ms":9490,"temperature":1.0,"reasoning_tokens":1613,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:31:03.951115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a benchmark with a known certified global optimum, run Algorithm 1 as written, and compare its returned QC lower bound against that global optimum: if the returned lower bound exceeds the global optimum, or if a region containing the optimum was pruned, the claim that the method preserves global optimality fails. A direct check is to enumerate the kept child boxes after the final level and verify that their union, together with the pruned boxes, covers the original voltage and angle box.","supporting_citations":[{"cited_title":"The qc relaxation: A theoretical and computational study on optimal power flow,","cited_arxiv_id":null,"evidence_quote":"Defines the QC relaxation with convex envelopes that supplies the lower bounds at every branch-and-bound node."},{"cited_title":"Global optimization of optimal power flow using a branch & bound algorithm,","cited_arxiv_id":null,"evidence_quote":"Introduces the spatial branch-and-bound template for OPF that this paper adapts from SDP-based bounding to QC-based bounding."},{"cited_title":"Computability of Global Solutions to Factorable Non- convex Programs: Part I–Convex Underestimating Problems,","cited_arxiv_id":null,"evidence_quote":"Provides the McCormick envelopes for bilinear products used by the QC relaxation."},{"cited_title":"Strengthening the sdp relaxation of ac power flows with convex envelopes, bound tightening, and valid inequalities,","cited_arxiv_id":null,"evidence_quote":"Shows how convex envelopes and bound tightening strengthen relaxations of AC power flows, informing the QC model."},{"cited_title":"Global optimization of optimal power flow using a quadratic convex relaxation,","cited_arxiv_id":null,"evidence_quote":"Prior demonstration of global optimization of OPF through a quadratic convex relaxation, the line of work this paper extends."},{"cited_title":"PowerModels.jl: An Open-Source Framework for Exploring Power Flow Formulations,","cited_arxiv_id":null,"evidence_quote":"Provides the modeling framework used to implement and test the algorithm."}],"review_version":1}