{"id":"d13dd4a6-599a-48b9-8271-8efaf18f1a90","arxiv_id":"2607.11593","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new asymmetric regret cost function for scenario reduction provably picks the single best scenario and, with a hybrid pre-selection, matches its accuracy at a fraction of the compute.","lead":"This paper proposes a new scenario-reduction cost function for two-stage stochastic mixed-integer programs and proves that its first pick is the single best scenario. In unit commitment tests it reaches about 2.1% error (small grid) and 0.4% (large grid) with just five scenarios, though the large-grid number is within the solver's own error tolerance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Large-case 0.4% RAE claim is below the 1% MIP gap used to compute reference optima, leaving the headline accuracy numerically unresolved.","rationale":"The reader's weakest assumption precisely identifies the 1% MIP gap as undermining the reported 0.4% RAE for the large case. My independent read agrees: the paper's theoretical contribution (Theorem 1) is a correct and self-contained statement about m=1 optimality, and the small-case results use a 0.01% MIP gap, so they are not affected. However, the abstract's most eye-catching number—'roughly ... 0.4% error' for the large case with five scenarios—depends on solving the reference SAA and reduced MIPs with a 1% gap. Since the claimed error is below that gap, the comparison is numerically unresolved; the true RAE could be several times larger or even negative in the computed interval. This does not automatically invalidate the qualitative conclusion that cPr and the hybrid outperform the other cost functions, but it does mean the precise headline accuracy is not established. The fix is straightforward: either tighten the MIP gap for the reference and reduced problems or, more cheaply, report the RAE interval derived from the incumbent and bound information. Because the issue is addressable and the core theorem and small-case evidence stand, the appropriate verdict remains CONDITIONAL. No change to the reader's verdict is needed.","tokens_in":18260,"tokens_out":8460,"duration_ms":79360,"concrete_test":"For each of the 10 large-case draws, use Gurobi's reported best bound and incumbent for the n=1000 reference SAA and for the reduced m=5 MIP to compute a rigorous interval for the true RAE (using z*(x*(Qm),Pn) evaluated exactly as a sum of LPs). Additionally, re-solve the reference SAA and the reduced MIPs for a random subset of draws with a tighter MIP gap (e.g., 0.1% or 0.01%) and recompute the mean RAE. If the interval half-width exceeds the reported mean RAE (0.4%) or if the mean RAE shifts by more than 0.3 percentage points under the tighter gap, the headline large-case accuracy is not supported as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline empirical claim—that with about five scenarios the proposed cost function reaches ~0.4% relative approximation error on the 300-bus case—rests on numerical results in Section 5.2.1. There, all MIPs (including the n=1000 reference SAA and the reduced m=5 problems) are solved with a 1% MIP gap (Section 5.2). Since the reported mean RAE (0.4%) is smaller than the solver tolerance, the computed z*(Pn) is only known to within about ±1%, and the suboptimality of x*(Qm) from the reduced MIP adds further uncertainty. Consequently, the true RAE could differ from the reported value by more than the claimed error itself; the interval of plausible values spans from near zero to well above 1%. The theorem (Theorem 1) is not affected—it is a mathematical statement about m=1 selection—and the small-case experiments (0.01% MIP gap) are credible. But the large-case quantitative claim is not resolved by the provided numerical evidence. This is a load-bearing weakness because the abstract and conclusions specifically advertise the 0.4% large-case accuracy as a key result, and the paper does not report any bounds or sensitivity analysis that would certify the RAE against the solver gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers scenario reduction for two-stage stochastic mixed-integer programs, revisiting the optimal mass transportation framework and the Forward Selection Algorithm. It reviews input-data-driven and optimization-problem-driven transportation cost functions, proposes a new asymmetric cost function cPr(ξi, ξj) = z(x*(ξj), ξi) − z(x*(ξi), ξi), and proves (Theorem 1, Appendix B) that the first scenario selected by Forward Selection with cPr minimizes the relative approximation error (RAE) among all single-scenario reductions. To reduce the computational burden of evaluating cPr, the paper proposes a hybrid algorithm that first prunes scenarios using the cost function of Morales et al. (2009) and then applies cPr to the pre-selected subset. Numerical experiments on two-stage stochastic unit commitment for a 24-bus system (n=200) and a 300-bus system (n=1000) report that around five scenarios suffice for roughly 2.1% and 0.4% RAE, with the hybrid algorithm cutting wall-clock time by a factor of 18 and solver work by a factor of 66 on the large case.","tokens_in":18561,"tokens_out":5576,"duration_ms":56007,"significance":"If the results hold, the paper offers a practical and scalable scenario-reduction method for two-stage stochastic MIPs, with a first-scenario optimality guarantee and a hybrid algorithm that makes the expensive cPr cost function usable. The strengths of the paper include: the proof of Theorem 1 is correct (though, as discussed below, the result is essentially definitional); the comparison of cost functions is carried out under a common algorithmic framework; the small-case experiments use a 0.01% MIP gap and are credible; and the code and data are publicly available. The main limitation is that the headline large-case accuracy, 0.4% RAE, is below the 1% MIP gap used to obtain the reference solutions, so the quantitative claim is not numerically resolved. The theoretical contribution is also weaker than it appears: Theorem 1 follows directly from the construction of cPr. Overall, the paper has merit, but the advertised large-case performance needs additional numerical support.","major_comments":[{"comment":"The reported large-case mean RAE of 0.4% is smaller than the 1% MIP gap used for all MIPs, including the n=1000 reference SAA and the reduced m=5 problems. Since z*(Pn) is only known to within roughly ±1% and x*(Qm) is itself an approximate solution, the computed RAE is within solver noise; the true RAE could be near zero or well above the reported value. This is load-bearing because the abstract and conclusions advertise the 0.4% figure. The authors should either solve the reference and reduced problems with a much tighter gap (as in the small case) for at least a subset of draws, or provide rigorous bounds/sensitivity analysis that show the RAE is insensitive to the 1% termination tolerance.","section":"§5.2, §5.2.1, Eq. (12)"},{"comment":"Theorem 1 is correct but is a direct algebraic consequence of the definition of cPr. In Eq. (B.3), the term z(x*(ξi), ξi) is subtracted and is independent of j, so the Forward Selection objective collapses to the RAE numerator for m=1. Thus the theorem essentially restates the cost-function construction rather than revealing a new structural property. The paper should explicitly acknowledge this and avoid presenting the result as a deep optimality guarantee; the contribution is better framed as the identification of a cost function that makes the greedy first selection optimal by design.","section":"Appendix B, Eq. (B.3)"},{"comment":"The comparison of cPr with other cost functions for m=1 is partly a tautology: by Theorem 1, cPr is engineered to minimize RAE at m=1, so its superior first-draw performance is expected. The more meaningful claims are for m>1 and for the hybrid algorithm. These empirical claims, however, inherit the numerical tolerance problem described above. Please report results with a common, tighter MIP gap for the small and large cases, or at least demonstrate that the reported RAE differences are not attributable to solver stopping criteria.","section":"§5.2.1, Figures 1–2"}],"minor_comments":[{"comment":"The abstract says “roughly 2.1% and 0.4% error” for five scenarios, while Section 5.1 reports around 2% and Section 6 reports 11.3% and 2.4% for one scenario. Please make the numbers consistent and clarify which m corresponds to each RAE.","section":"Abstract / §6"},{"comment":"The notation Q1 = {Q_j^1 : j ∈ I} and later Q1 ∈ Q1 is confusing. Use a different symbol for the family of distributions, e.g., \\mathcal{Q}_1, and state the assumption z*(Pn)>0 in the theorem, not only in a footnote in the proof.","section":"Theorem 1 statement"},{"comment":"There is a notational inconsistency: the proof uses both ζ_j and ξ_j for the candidate scenarios. Please unify to ξ_j throughout.","section":"Appendix B"},{"comment":"In line 4, the set J is used before any scenario is selected (J = {}). While the expression is correct, a short remark explaining the convention for J = {} would improve readability.","section":"§3.2, Algorithm 1"},{"comment":"The claim that cPr requires n MIPs and n^2−n LPs assumes all entries of the cost matrix are computed. The paper later uses heuristics (pre-reduction with cID and warm-starting, Appendix C). Please state clearly that the reported computational costs in Table 2 include those heuristics, so that the comparison with the exact cost matrix is not misleading.","section":"§4.2.4 / Table 1"},{"comment":"The paper does not discuss the effect of multiple optimal first-stage solutions x*(Qm) on the RAE. If the reduced MIP has several optima, the RAE may depend on which optimum is returned. A brief comment or a check with a different solver seed would strengthen the numerical claims.","section":"§5.2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready for acceptance because the central large-case numerical claim (0.4% RAE) is not resolved at the 1% MIP gap, and the main theorem is essentially an algebraic consequence of the cost function design. Both issues are fixable within the scope of the manuscript: the authors can add a tighter-gap verification or sensitivity analysis, and they can reframe the theoretical contribution more honestly. I encourage the editor to consider a revised version after these points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is better than the skeptic's note might suggest. The new asymmetric cost function cPr is simple and the Theorem 1 proof is correct. It is true that the result follows almost immediately from the construction—subtracting z(x*(ξi), ξi) makes the Forward Selection objective equal to the RAE numerator up to a constant independent of j. That is not circularity, it is a valid and useful design insight. The small-case experiments with 0.01% MIP gap are credible and support the qualitative claims: cPr selects a good first scenario and produces robust reduced sets. The hybrid algorithm is pragmatic and the computational savings are real. Code and data are available, which counts for something.\n\nThe soft spot is exactly where the stress-test lands. In the 300-bus study, all MIPs, including the n=1000 reference SAA, are solved with a 1% MIP gap, and the headline RAE is 0.4%. The reported accuracy is smaller than the solver tolerance, so the true RAE could be anywhere from near zero to well above 1%. That is a load-bearing weakness because the abstract and conclusions advertise 0.4% as a key result. It is addressable: tighter gap or a bound, confidence intervals, or even just reporting the range of objectives across the 10 draws with and without gap. The hand-chosen r=50 in the hybrid algorithm and the lack of sensitivity analysis on r also leave the headline computational claims a bit open, but those are minor in comparison.\n\nThe m=1 theorem is not affected by the numerical issue, and the small-case results stand on their own. For someone working on scenario reduction for stochastic MIPs, this is a genuinely useful piece: a new cost function that provably gets the first scenario right, plus a fast approximation that seems to work well in practice. The reader should be told that the 0.4% number should not be repeated without qualification until it is verified at a tighter gap.\n\nThis deserves peer review. The theory is correct, the experiments are extensive, and the main flaw is a methodological artifact that a careful referee can push the authors to fix. I would not desk reject it.\n\nMy recommendation: send it to review, and in your own work cite the cost function and Theorem 1, not the 0.4% claim.","headline":"New cost function and an m=1 optimality theorem are real contributions; the paper is worth refereeing, but the 0.4% large-case claim is not resolved by experiments run at a 1% MIP gap.","tokens_in":19067,"tokens_out":1713,"would_cite":true,"duration_ms":17189,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C15","90C11","49Q22"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new scenario-reduction cost function provably selects the best single scenario first, and about five scenarios suffice to approximate the full-distribution optimum within a few percent.","keywords":["scenario reduction","two-stage stochastic mixed-integer programming","optimal mass transportation","cost function","Forward Selection","relative approximation error","unit commitment","stochastic programming"],"falsifier":"Solve the large 300-bus case with the same scenario samples but a 0.01% MIP gap for the reference n = 1000 SAA and the m = 5 reduced problem; if the mean RAE is no longer around 0.4% or the ranking of cost functions changes, the headline number is an artifact of solver tolerance. Alternatively, on a different application, enumerate all single-scenario distributions and verify that Forward Selection's first pick under cPr attains the minimal RAE.","tokens_in":18142,"feed_emoji":"⚡","tokens_out":6128,"duration_ms":50166,"temperature":0.7,"pith_summary":"The paper proposes a new cost function, cPr, for the optimal mass transportation problem used in scenario reduction for two-stage stochastic mixed-integer programs. It proves that when this cost function is combined with Forward Selection, the first scenario drawn is the single scenario that minimizes the relative approximation error — the regret of making decisions on the reduced distribution instead of the full one. On stochastic unit commitment test systems, roughly five scenarios selected this way approximate the full-distribution optimum within about 2.1% (small) and 0.4% (large) error, whereas competing cost functions typically need 25 or more. Since solution time in two-stage stochastic MIPs grows sharply with scenario count, a handful of accurate scenarios can be the difference between tractable and intractable. The paper also gives a hybrid algorithm that preserves the quality while cutting computation.","feed_headline":"Five scenarios hit 0.4% of full-distribution accuracy on a 300-bus grid","feed_subtitle":"A provably best first pick and a 66x faster hybrid make few-scenario approximations practical.","key_machinery":"The central object is the transportation cost function cPr(ξi, ξj) = z(x*(ξj), ξi) − z(x*(ξi), ξi), the asymmetric regret of optimizing the first stage for scenario j and then observing scenario i, minus the cost when the decision is tailored to i. The load-bearing mechanism is Theorem 1: with Forward Selection's first draw, minimizing the transported mass under cPr is equivalent to minimizing the relative approximation error, because the subtractive term depends only on i and cancels. The paper also introduces a hybrid of an expected-value-problem-based cost function and cPr, where the former quickly prunes the scenario set and the latter refines it.","core_discovery":"The central claim is that scenario reduction quality is governed by the choice of transportation cost function, and the paper's proposed asymmetric regret cost, cPr(ξi, ξj) = z(x*(ξj), ξi) − z(x*(ξi), ξi), makes the Forward Selection algorithm pick, on the first draw, the scenario whose single-scenario distribution minimizes the relative approximation error among all single-scenario distributions (Theorem 1). The proof exploits the observation that for m = 1 the distribution is a Dirac measure, so the two-stage problem reduces to the single-scenario problem, and the constant term z(x*(ξi), ξi) in cPr cancels in the minimization. Numerically, this first choice alone gives about 11.3% and 2.4%","pith_inferences":["The proof structure suggests a recipe for other asymmetric regret functions: any cost of the form R(ξi, ξj) − const(ξi) with zero diagonal will inherit the m = 1 optimality property, so the result may transfer to other problem classes.","One could test the hybrid's Phase-1 choice r = 50 sensitivity: the paper fixes r for both cases, and a principled stopping rule for pruning could further reduce computation.","The reported 0.4% RAE on the large case is below the 1% MIP gap used for all solves; verifying with a tighter gap would separate the method's accuracy from solver tolerance.","For convex second stages, duality might let cPr be evaluated from dual prices of the n² − n second-stage LPs without solving them all, a route the paper notes as future work."],"forward_implications":["Replacing hundreds or thousands of scenarios with roughly five selected scenarios can bring two-stage stochastic MIPs within reach, since solution time typically scales badly with scenario count.","The first scenario drawn by Forward Selection with cPr is provably the best single-scenario choice, giving a warm start that other cost functions do not guarantee.","The method is consistent across random draws: cPr shows markedly lower variance in RAE across samples than competing cost functions.","The hybrid algorithm delivers the same approximation quality as cPr but at 18x less wall-clock time and 66x less solver work on the 300-bus case, making the approach practical.","Theorem 1's guarantee is limited to m = 1; the good behavior at m > 1 is empirical and not proven."],"fun_headline_variants":["Provably best first scenario pick—cuts scenario count to five","New cost function gives optimal first pick, five scenarios match full accuracy","Stochastic programming: guaranteed first choice, 66x less solver work","Scenario reduction with provable first pick—five scenarios hit 0.4% error","Optimal first scenario choice unlocks 5-scenario solution accuracy"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The headline large-case accuracy (0.4% mean RAE at m = 5) is computed under a 1% MIP gap for every solve, including the reference solution, so the measured error is smaller than the solver's certified optimality tolerance.","fun_headline_variants_meta":{"raw":{"variants":["Provably best first scenario pick—cuts scenario count to five","New cost function gives optimal first pick, five scenarios match full accuracy","Stochastic programming: guaranteed first choice, 66x less solver work","Scenario reduction with provable first pick—five scenarios hit 0.4% error","Optimal first scenario choice unlocks 5-scenario solution accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000461,"raw_usage":{"total_tokens":2185,"prompt_tokens":828,"completion_tokens":1357,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":1263}},"tokens_in":572,"tokens_out":1357,"duration_ms":11782,"temperature":1.0,"reasoning_tokens":1263,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T06:50:20.865293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the large 300-bus case with the same scenario samples but a 0.01% MIP gap for the reference n = 1000 SAA and the m = 5 reduced problem; if the mean RAE is no longer around 0.4% or the ranking of cost functions changes, the headline number is an artifact of solver tolerance. Alternatively, on a different application, enumerate all single-scenario distributions and verify that Forward Selection's first pick under cPr attains the minimal RAE.","supporting_citations":[],"review_version":2}