{"id":"6685fa71-d121-4fed-b163-372be71ba454","arxiv_id":"2603.20411","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The dual rotated second-order cone constraints of the Jabr ACOPF relaxation are always active at optimality, which lets the paper replace them with equality constraints and produce a certified lower bound.","lead":"This paper reformulates the dual of a conic AC optimal power flow relaxation so that the rotated second-order cone inequalities disappear, leaving only non-negativity constraints, and shows numerically that the reformulated problem matches a standard conic solver's lower bounds on test grids up to 1354 buses. Cheaper, simpler dual problems could speed up GPU-accelerated first-order methods for large-scale power-grid optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 2's proof of d_t2 = 1 relies on unproven bound t̄ ≥ ||d_t||^2/4; the argument also mishandles the one-sided box on t.","rationale":"The central claim of the paper is that all dual RSOC constraints are active at optimality. This claim supports the variable-elimination step (29) that produces Model 4 (ATD). The elimination of the generator cost epigraph uses d_t2 = 1, which is the content of Lemma 2. Without a rigorous proof of d_t2 = 1, the ATD formulation does not follow from the preceding lemmas; the numerical agreement with MOSEK would be unexplained or at best a conjecture. The reader's weakest_assumption correctly identifies the unproven bound t̄ ≥ ||d_t||^2/4. I agree with this assessment. The additional issue that the box on t is one-sided ([0,t̄]) rather than symmetric ([−t̄,t̄]) strengthens the concern: the function h(d_t2) used in the proof is not the correct objective term. The condition for the maximum to occur at d_t2=1 remains exactly the same unproven inequality on the dual norm. This is a genuine gap in the proof of the central theorem. However, the conclusion d_t2 = 1 is likely true by a simpler argument: since t̄ is a strict upper bound on the optimal epigraph cost, the box constraint on t is inactive, the corresponding multiplier μ_t is zero, and stationarity (18b) for t gives 1 - d_t2 = 0. Thus the paper's proof is incomplete but the claim is probably salvageable. For this reason, a conditional verdict is appropriate rather than rejection. The concrete test of computing ||d_t||^2/(4 t̄) on the paper's own benchmark set would either expose counterexamples or support the empirical validity of the bound, and it would be a useful addition regardless.","tokens_in":14066,"tokens_out":14899,"duration_ms":130289,"concrete_test":"For every PGLib instance in Table VIII, solve Model 3 (All Conic Dual) to a high-accuracy optimum with MOSEK, extract the dual variables (d_t1, d_t2, d_t) for the generator cost epigraph, compute t̄ = Σ_{k∈G} c_q^k (P_max^k)^2 and test whether ||d_t||^2/(4 t̄) ≤ 1. If any case violates this, Lemma 2's key hypothesis is false at the computed optimum and the proof of d_t2 = 1 is refuted for that instance. If all pass, the inequality holds empirically but the missing analytical link remains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 2 is the linchpin of the 'All Tight Dual' equivalence: it justifies setting d_t2 = 1 and replacing d_t1 by ||d_t||^2/2 in (29). The proof analyzes h(d_t2) = -t̄|1-d_t2| - ||d_t||^2/(4 d_t2) and claims that for t̄ ≥ ||d_t||^2/4 the maximum is at d_t2 = 1. But t̄ is a bound on the primal epigraph variable t, not on the dual variable d_t; no argument links the two. The footnote that t̄ can be set arbitrarily large is not an answer: increasing t̄ changes the dual objective via the term -t̄|1-d_t2|, so the optimum may shift, and the proof's condition is on the solution, not on the problem data. Moreover, (24) minimizes over |t| ≤ t̄, while t is nonnegative and boxed as 0 ≤ t ≤ t̄; with the correct box, the t-dependent objective term is -t̄ max(0,d_t2-1), and the threshold for the maximizer to be at d_t2=1 is again t̄ ≥ ||d_t||^2/4 (with a different subgradient analysis). Unless this bound is proven, the elimination of the cost epigraph is not justified. Lemma 1 also ignores the coupling of ds2/dr2 with the stationarity equality (18b), though that gap may be repairable; Lemma 2's gap is central.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an \"All Tight Dual\" (ATD) reformulation of the conic dual of the Jabr second-order cone relaxation of AC optimal power flow. The central claim is that all dual rotated second-order cone (RSOC) constraints are tight at optimality (Section III, Lemmas 1-3), so the dual cone inequalities can be replaced by equalities and the corresponding dual variables eliminated, yielding a non-conic maximization problem (Model 4). The authors also propose an epsilon-stabilized variant with a postprocessing projection step that produces a certified lower bound (Eq. (37)). Numerical experiments on PGLib cases from 3 to 1354 buses compare the ATD formulation against the standard conic dual solved with Ipopt/Knitro and against MOSEK.","tokens_in":14457,"tokens_out":11721,"duration_ms":122093,"significance":"If the tightness claim is correct, the ATD formulation is a genuinely useful structural simplification: it removes the dual RSOC constraints, reduces the size of the KKT system, and is a natural target for first-order/GPU-accelerated methods. The certified-lower-bound postprocessing is also valuable. The paper is mostly constructive, with machine-checkable algebraic derivations for the elimination step and a clear numerical demonstration on standard benchmarks. The main risk is that the proof of the key elimination formula rests on an unproven bound, so the central equivalence is not yet established with the required rigor.","major_comments":[{"comment":"The proof that d_t2 = 1 at optimality depends on the condition t̄ ≥ ||d_t||^2/4, but t̄ is a bound on the primal epigraph variable t, not on the dual vector d_t. No chain of inequalities links t̄ to the dual variables. The footnote claiming t̄ can be set arbitrarily large is not responsive: t̄ enters the dual objective h(d_t2) through the term -t̄|1-d_t2|, so increasing t̄ changes the maximizer. This is load-bearing because the replacement d_t1 = ||d_t||^2/2 in Eq. (29) and Model 4 is justified only by Lemma 2. Please supply a bound on ||d_t|| in terms of problem data (e.g., via stationarity (18b) and cone geometry) or replace the argument with a complementary-slackness proof using the positive constant term in the cost cone.","section":"Section III, Lemma 2 and Eq. (27)"},{"comment":"The proof of Lemma 1 argues that decreasing a dual scalar s strictly improves the objective if the RSOC constraint is slack. This ignores that s also appears in the stationarity equality (18b) through F^T d̃, so decreasing s may violate stationarity unless other variables are adjusted, and the objective also depends on λ. The claimed result is likely true by the standard complementary slackness argument for conic programs (if a dual cone variable is interior, the corresponding primal cone component must be zero; for the flow and cost cones this component includes a positive constant such as Smax or 1/2, contradicting primal feasibility). The proof should be replaced by this rigorous argument.","section":"Section III, Lemma 1 and Eq. (18b)-(22)"},{"comment":"The dualization convention for the line-flow and cost-epigraph RSOC constraints is inconsistent as written. Model 2 writes constraints in terms of variables t_s and t, while Appendix B and Eq. (22) treat the corresponding cone components as the constants S_max and 1/2. This ambiguity determines which dual variables enter stationarity (18b) and therefore affects Lemma 1 and the derivation of Eq. (22). Please clarify whether t_s and t are eliminated through the linear equality block or remain cone variables, and make the F and g definitions consistent throughout.","section":"Section II.C-II.D and Appendix B, Eqs. (13e)-(13g), (22)"}],"minor_comments":[{"comment":"The primal box is stated as 0 ≤ t ≤ t̄, but Eq. (24) and the surrounding text use |t| ≤ t̄, and Lemma 2 asserts ω_1 = 0 without explanation. The correct term for 0 ≤ t ≤ t̄ is -t̄ max(0, d_t2-1). Please reconcile the box convention.","section":"Section III, Lemma 2, Eq. (24)"},{"comment":"The abstract says the tightness is 'observed' and then the paper proves it; consider rephrasing to 'established' to avoid implying an empirical observation is the basis.","section":"Abstract and Section I"},{"comment":"Table II reports all gaps as 0.00 while the text notes an average gap of 2.31e-3%; clarify significant digits or report more precisely. In Table I, the Ipopt 'Primal RSOC' objective of -88723153.12 for the 500-bus case appears to be a solver failure; label it as such.","section":"Section IV, Tables I-II"},{"comment":"Minor typos: 'prposed' in Section I, 'coice' in Section V, and 'Preformance' in Table heading. Also, the reference [21] is a prior DCOPF result by the same authors; in Lemma 2 please state explicitly which part of the result is new for ACOPF.","section":"Throughout"},{"comment":"After projecting the dual variables, the paper does not explicitly restore stationarity (18b); since Eq. (20) is a Lagrangian dual over the box, dual-cone feasibility may suffice for a lower bound, but this should be stated clearly so readers do not assume the projected point is feasible for Model 3.","section":"Section III.B, Eq. (37)"}],"recommendation":"major_revision","confidential_remarks":"The paper contains a promising idea and solid numerical evidence, but the proof of Lemma 2 is currently incomplete in a way that directly threatens the central equivalence. If the authors can provide a rigorous bound on ||d_t|| or an alternative proof of d_t2 = 1, the result is likely publishable. The Lemma 1 gap is repairable by a standard complementarity argument, but it must be fixed. I would not accept the manuscript in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. The paper does something genuinely useful: it gives a clean, non-conic reformulation of the dual of the Jabr ACOPF relaxation by proving (or at least strongly arguing) that all dual RSOC constraints are tight at optimum, and it backs the reformulation with solid numerical results on PGLib cases up to 1354 buses. The certified lower bound procedure—project the epsilon-stabilized dual point back onto the RSOC faces, then evaluate the dual objective—is a legitimate and practical contribution, and the experiments show ATD matching MOSEK's dual values closely, with big speedups on larger cases.\n\nNow the soft spots. Lemma 3 is a correct KKT/complementarity argument. Lemma 1's proof is incomplete: the local improvement argument ignores the stationarity equality (18b), so changing a dual scalar can break feasibility unless you adjust lambda or mu. I think that is repairable, but as written it does not go through. Lemma 2 is the load-bearing one, and the stress-test note is right on both counts. The bound t-bar >= ||d_t||^2/4 is asserted without proof; t-bar is a bound on the primal variable t, not on the dual vector, and the footnote about choosing t-bar large does not work because t-bar appears in the dual objective and shifts the optimum. Also, the primal box is one-sided (0 <= t <= t-bar), so the minimization leading to (24) gives -t-bar max(0,d_t2-1), not -t-bar|1-d_t2|. With that correction the threshold condition is still t-bar >= ||d_t||^2/4, but it remains unproven. So the equivalence between Model 4 and Model 3 is not yet rigorous. That is a genuine gap in the paper's central theoretical claim.\n\nNone of that makes me doubt the empirical conclusion. The numerical evidence is strong and consistent, and the gap looks fixable—likely by deriving a bound on ||d_t|| from the stationarity equations, or by a different argument about the maximizer of h. The 1354-bus result is not quite the 'same objective value' the abstract advertises (about 0.25% gap to MOSEK's value), but it's still close and the trend is clear.\n\nWho should read this: anyone working on first-order or GPU methods for OPF, and anyone who wants a dual formulation that avoids conic projections. I would bring it to a reading group. It deserves peer review—send it out, but the referee should require a complete proof of Lemma 2 before accepting the equivalence claim.","headline":"Solid and useful reformulation with strong numerics, but Lemma 2's proof has a real gap that must be closed before the equivalence claim is rigorous.","tokens_in":14955,"tokens_out":6151,"would_cite":true,"duration_ms":61949,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","90C46","90C90"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that every dual rotated-second-order-cone constraint in the Jabr relaxation of AC optimal power flow must be tight (active) at optimality, which lets the conic dual be rewritten as a non-conic problem with only non-negativi","keywords":["ACOPF","Jabr relaxation","rotated second-order cone","dual tightness","certified lower bounds","conic optimization","first-order methods"],"falsifier":"Find any PGLib instance (or constructed network) where the optimal dual RSOC constraint for the generator cost epigraph is not tight — i.e., where the ATD objective strictly exceeds the conic dual objective, or where an interior-point solution to the conic dual has 2 d_t1 d_t2 > ||d_t||^2. Alternatively, compute the certified lower bound (Eq. 37) on a large set of instances and check whether it ever exceeds the primal RSOC optimal value.","tokens_in":13920,"feed_emoji":"⚡","tokens_out":3439,"duration_ms":32112,"temperature":0.7,"pith_summary":"The paper tries to establish that, in the dual of the Jabr second-order-cone relaxation of AC optimal power flow, every rotated second-order cone inequality is tight (active) at an optimal solution. Given that, the dual can be rewritten with the conic constraints replaced by equalities, eliminating the cone entirely and leaving only non-negativity constraints. If correct, this yields a simpler dual that interior-point solvers handle faster on large networks, and a cheap post-processing projection turns any near-feasible solution into a certified lower bound on the ACOPF cost. The authors verify numerically that their tight dual matches a mature conic solver on PGLib cases up to 1354 buses.","feed_headline":"ACOPF dual cones are always tight: drop the conic constraints","feed_subtitle":"Reformulating the dual of the Jabr relaxation as a non-conic problem yields the same lower bounds and faster solves.","key_machinery":"The rotated second-order cone (RSOC) and its dual cone; the paper uses self-duality of RSOC and the fact that dual variables sit in the dual cone. The load-bearing identity is 2 d1 d2 = ||d_vec||^2 (tightness), which allows eliminating d1 as a function of the others. The epsilon-stabilized replacement function r(z, eps) = ||z_vec||^2 / (2 z2 + eps) for the eliminated variable, plus a post-processing projection onto the cone boundary, converts a slightly infeasible dual solution into a certified lower bound (Eq. 37).","core_discovery":"The central claim is that at any optimal solution of the dual of the Jabr RSOC relaxation, the inequalities 2d1*d2 >= ||d||^2 hold with equality for every dual RSOC tuple (voltage relaxation, line flow limits, and generator cost epigraph). Lemmas 1 and 2 show this for the line-flow and cost-epigraph cones by arguing that any slack would allow decreasing a dual variable with a strictly negative objective coefficient, and that the optimal cost-cone scalar d_t2 equals 1. Lemma 3 shows it for the voltage cone via KKT complementarity, since a strictly interior dual cone would force the primal variables to zero, contradicting feasibility. Together they justify the 'All Tight Dual' model (Model 4),","pith_inferences":["If the tightness claim holds, it implies that the dual optimal face of the Jabr relaxation is entirely contained in the boundary of the dual RSOC cones, which is a kind of strict-complementarity-like behavior; one could test whether similar tightness holds for other RSOC-based relaxations (e.g., those with added cycle constraints or tighter cuts).","The proof of Lemma 2 relies on t being bounded, but t is a primal variable; a rigorous chain linking the dual norm to that bound is missing. A natural extension is to check whether the result can be proven without that assumption, perhaps via complementary slackness on the cost epigraph.","The epsilon-stabilized dual could be used as a warm-start for the exact conic dual, or as a Lagrangian dual for first-order methods; the paper's certified lower bound postprocessing already makes it safe for bounding.","A testable extension is to apply the same elimination to the SDP relaxation of ACOPF — if dual SDP constraints also tend to be tight, the same argument could simplify the dual SDP."],"forward_implications":["The dual of the Jabr ACOPF relaxation can be solved as a non-conic problem with only box/non-negativity constraints, enabling projected-gradient and other first-order methods without conic projections.","A certified lower bound on the relaxed ACOPF cost can be computed from any near-feasible dual point via the projection formula, independent of the stabilization parameter.","On large PGLib cases, the simplified dual solves faster and more reliably with interior-point solvers than the original conic dual.","Eliminating conic constraints shrinks the KKT system, reducing the bottleneck in IPM-based approaches."],"fun_headline_variants":["Dual cones are always tight: ACOPF without conic constraints","Why dual cones in ACOPF are always active","Dropping conic constraints in ACOPF dual: proven tight","Always-tight dual cones: a simpler ACOPF dual"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Lemma 2 assumes the primal upper bound t_bar on the epigraph variable is large enough to dominate the squared norm of the dual cost-cone vector at optimality, but no argument connects the primal bound to the dual variable's size; if ||d_t|| can be large, the maximum of the reduced objective need not occur at d_t2 = 1.","fun_headline_variants_meta":{"raw":{"variants":["Dual cones are always tight: ACOPF without conic constraints","Why dual cones in ACOPF are always active","Dropping conic constraints in ACOPF dual: proven tight","Always-tight dual cones: a simpler ACOPF dual"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000805,"raw_usage":{"total_tokens":3386,"prompt_tokens":773,"completion_tokens":2613,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":2552}},"tokens_in":517,"tokens_out":2613,"duration_ms":17050,"temperature":1.0,"reasoning_tokens":2552,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:46:40.416615+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find any PGLib instance (or constructed network) where the optimal dual RSOC constraint for the generator cost epigraph is not tight — i.e., where the ATD objective strictly exceeds the conic dual objective, or where an interior-point solution to the conic dual has 2 d_t1 d_t2 > ||d_t||^2. Alternatively, compute the certified lower bound (Eq. 37) on a large set of instances and check whether it ever exceeds the primal RSOC optimal value.","supporting_citations":[],"review_version":1}