{"id":"1dcdb2ab-2251-42b3-93c7-796f44c76b44","arxiv_id":"2605.10854","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Superposition relaxation creates separable estimators for factorable functions that are tighter than McCormick relaxations in numerical tests while providing convergence guarantees.","lead":"This paper introduces superposition relaxation, a new arithmetic for creating separable under- and over-estimating functions that bracket multivariate factorable functions on compact domains. It may interest optimization researchers seeking tighter bounds than McCormick relaxations for nonconvex problems including neural network analysis.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest assumption matches the only plausible point of failure, but the abstract plus the described structure give no concrete indication that the rules fail to preserve validity or that the case studies are mis-implemented. The work is therefore internally consistent on its own terms; the low-confidence UNVERDICTED status stems from the prior lack of full text rather than from any detected flaw.","tokens_in":1648,"tokens_out":280,"duration_ms":55055,"concrete_test":"Recompute the ANN relaxation examples from the numerical section using the exact piecewise-linear parameterization and number of pieces reported; verify that the reported gap to McCormick remains positive and that the under/over-estimators still bracket the network output on a dense test grid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on numerical case studies demonstrating tighter bounds than McCormick relaxations for factorable functions and ANNs. The propagation rules for affine and nonlinear compositions are presented as correctly exploiting global monotonicity and convexity on compact domains, with accompanying local convergence analysis (pointwise and Hausdorff). No internal inconsistency, gap in the bracketing argument, or unsupported step in the arithmetic is apparent from the described construction. The higher computational cost is explicitly acknowledged, and the tightness claim is framed as empirical rather than theoretical.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces an arithmetic termed superposition relaxation for bracketing the graph of a multivariate factorable function on a compact domain by a pair of separable under- and over-estimating functions. Propagation rules are derived for affine and nonlinear compositions that exploit global monotonicity and convexity properties of the factors. Local convergence is analyzed in both the pointwise and Hausdorff senses, including conditions for quadratic pointwise convergence to propagate through composition. Practical parameterizations of the univariate summands as piecewise-constant or continuous piecewise-linear functions are presented, and numerical case studies are used to show that the resulting relaxations are consistently tighter than McCormick relaxations for factorable functions and artificial neural networks, at the expense of higher computational cost.","tokens_in":1729,"tokens_out":581,"duration_ms":22060,"significance":"If the propagation rules and numerical evidence hold, the work supplies a concrete alternative to McCormick envelopes that can produce tighter separable relaxations for global optimization, with direct relevance to neural-network relaxations. The explicit local convergence analysis (pointwise and Hausdorff) and the discussion of implementable piecewise parameterizations constitute genuine strengths; the paper also correctly flags the computational trade-off, which is essential for assessing practical utility.","major_comments":[{"comment":"The central empirical claim (tighter bounds than McCormick relaxations, including for ANNs) rests on numerical case studies whose construction details—specific test functions, choice of piecewise breakpoints, and how the superposition is assembled for the network layers—are not fully specified. Without these, it is difficult to judge whether the observed tightness generalizes or depends on favorable choices of the univariate estimators.","section":"Numerical case studies section"},{"comment":"The propagation rules for nonlinear compositions are stated to exploit global monotonicity and convexity on compact domains, yet the manuscript does not provide an explicit verification (e.g., a short proof or counter-example check) that these rules preserve valid bracketing when the outer function is neither monotone nor convex. This step is load-bearing for the arithmetic’s correctness.","section":"Propagation rules for nonlinear compositions"}],"minor_comments":[{"comment":"The abstract mentions “conditions under which quadratic pointwise convergence propagates through composition” but does not reference the corresponding theorem or proposition number; adding the citation would improve readability.","section":"Abstract"},{"comment":"The implementation discussion would benefit from a brief complexity statement (e.g., number of univariate pieces versus McCormick cost) or pseudocode for the propagation step, even if only in an appendix.","section":"Parameterizations and implementation"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript appears to be a self-contained contribution in numerical analysis/global optimization; no obvious citation-pattern or scope issues."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the positive evaluation and recommendation for minor revision. The comments highlight opportunities to strengthen reproducibility and rigor, which we address below. We will incorporate the suggested clarifications into the revised manuscript.","responses":[{"response":"We agree that additional implementation details will improve reproducibility. In the revised manuscript we will expand the numerical case studies section with a new subsection that explicitly lists the multivariate test functions, the ANN architectures considered, the breakpoint selection strategy (including uniform grids and any adaptive criteria), and the precise layer-wise assembly procedure for the separable estimators. These additions will allow independent verification of the reported tightness relative to McCormick relaxations without altering the existing numerical results.","revision_made":"yes","referee_comment":"The numerical case studies lack full specification of test functions, piecewise breakpoints, and superposition assembly for networks, making it hard to assess generalizability of the tightness claims."},{"response":"The derivation of the nonlinear propagation rules begins from the definition of separable under- and over-estimators and applies the monotonicity/convexity properties only when they are globally available on the compact domain; in the absence of these properties the rules fall back to standard interval bounds that are known to be valid. To make this explicit we will insert a short lemma in the revised section on nonlinear compositions that proves bracketing preservation in the general case (including when the outer function is neither monotone nor convex) by direct appeal to the separable estimator definition and the univariate relaxation properties. This addition addresses the load-bearing correctness concern without changing the stated rules.","revision_made":"yes","referee_comment":"The propagation rules for nonlinear compositions lack explicit verification that they preserve valid bracketing when the outer function is neither monotone nor convex."}],"tokens_in":1381,"tokens_out":383,"duration_ms":31431,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core contribution is a new relaxation arithmetic that brackets factorable functions between separable under- and over-estimators. Propagation rules are given for affine and nonlinear compositions that use global monotonicity and convexity on compact domains, plus local convergence results in pointwise and Hausdorff senses, including conditions for quadratic convergence to propagate. Practical parameterizations use piecewise-constant or piecewise-linear univariate summands. Numerical cases, including artificial neural networks, show consistently tighter bounds than McCormick relaxations, though at higher computational cost. The authors flag the cost issue directly. This is genuinely new relative to standard McCormick work and gives a concrete alternative for global optimization. The numerical evidence is the strongest part; it demonstrates practical improvement on the tested problems without obvious cherry-picking. The convergence analysis adds rigor that is often missing in relaxation papers. The main soft spot is that tightness remains an empirical observation rather than a general guarantee, so performance on new problems is not assured in advance. The propagation rules rest on correctly exploiting the monotonicity and convexity properties, and while the construction looks internally consistent, any implementation would need careful verification of those steps. The higher cost is acknowledged but not quantified in detail across problem classes, which leaves the practical trade-off for users to explore. This work is for researchers in global optimization, interval methods, and nonconvex programming who already use relaxation techniques and want options beyond McCormick. A reader focused on bound quality for factorable or neural-network problems will find usable ideas and comparisons. It deserves a serious referee because the arithmetic is distinct, the analysis is present, and the numerical results are substantive enough to merit expert scrutiny even if revisions on efficiency and generality are likely.","headline":"This paper introduces superposition relaxation as a separable estimator arithmetic that empirically produces tighter bounds than McCormick for factorable functions and ANNs, with added convergence analysis but higher cost.","tokens_in":2204,"tokens_out":412,"would_cite":false,"duration_ms":34375,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Superposition relaxations bracket factorable functions with separable estimators that are tighter than McCormick relaxations.","keywords":["superposition relaxation","separable estimators","McCormick relaxation","factorable functions","global optimization","convergence analysis","artificial neural networks"],"falsifier":"A specific factorable function and domain where the superposition relaxation produces a wider bounding interval than the McCormick relaxation, or where quadratic pointwise convergence fails to hold through a composition despite the stated conditions.","tokens_in":2547,"feed_emoji":"📐","tokens_out":605,"duration_ms":52418,"temperature":0.7,"pith_summary":"This paper introduces superposition relaxation as a method to bound the graph of a multivariate factorable function using a pair of separable underestimating and overestimating functions. Propagation rules are given for affine and nonlinear compositions that take advantage of monotonicity and convexity properties. The approach is shown to achieve quadratic local convergence in certain cases and to produce tighter relaxations than the McCormick method in numerical examples, including for artificial neural networks. A sympathetic reader would care because tighter relaxations can lead to more efficient global optimization algorithms by reducing the search space more effectively.","feed_headline":"Separable estimators tighten relaxations beyond McCormick","feed_subtitle":"Superposition relaxations for factorable functions outperform McCormick in tightness tests including neural networks, though at higher cost.","key_machinery":"Superposition relaxation arithmetic, which generates separable under- and over-estimators via composition propagation rules exploiting monotonicity and convexity.","core_discovery":"The paper establishes an arithmetic for superposition relaxations that constructs separable estimators for factorable functions on compact domains. It proves local convergence properties in pointwise and Hausdorff senses, with conditions for quadratic pointwise convergence to propagate through compositions. Numerical case studies demonstrate that these relaxations are consistently tighter than McCormick relaxations for various functions and for artificial neural networks, although they require more computation.","pith_inferences":["Existing global optimization solvers could adopt these relaxations to reduce the number of nodes explored in nonconvex problems.","The separability of the estimators may enable parallel evaluation of the univariate components for high-dimensional functions.","The arithmetic could be extended to other classes of functions or combined with different relaxation techniques for hybrid bounds."],"forward_implications":["Tighter bounds improve the performance of branch-and-bound algorithms in global optimization.","Relaxations can be applied to artificial neural networks with better tightness than McCormick.","Quadratic pointwise convergence propagates through compositions when conditions on monotonicity and convexity are met.","Practical implementations use piecewise-constant or continuous piecewise-linear univariate summands."],"fun_headline_variants":["Superposition relaxations provide tighter separable estimators than McCormick","Convergence analysis for superposition relaxation arithmetic through compositions","Tighter relaxations for factorable functions via superposition separable estimators","Implementation of piecewise constant or linear superposition relaxations"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The propagation rules for affine and nonlinear compositions correctly exploit global monotonicity and convexity properties of the factorable functions.","fun_headline_variants_meta":{"raw":{"variants":["Superposition relaxations provide tighter separable estimators than McCormick","Convergence analysis for superposition relaxation arithmetic through compositions","Tighter relaxations for factorable functions via superposition separable estimators","Implementation of piecewise constant or linear superposition relaxations"]},"model":"grok-4.3","cost_usd":0.008724,"raw_usage":{"total_tokens":3824,"prompt_tokens":615,"num_sources_used":0,"completion_tokens":62,"cost_in_usd_ticks":87240500,"prompt_tokens_details":{"text_tokens":615,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3147,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":615,"tokens_out":62,"duration_ms":21641,"temperature":1.0,"reasoning_tokens":3147,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-12T04:04:02.495943+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A specific factorable function and domain where the superposition relaxation produces a wider bounding interval than the McCormick relaxation, or where quadratic pointwise convergence fails to hold through a composition despite the stated conditions.","supporting_citations":[],"review_version":1}