{"id":"20976bc5-d3f2-499b-9de5-54418aa22082","arxiv_id":"2607.21286","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"For separable mixed-binary quadratic programs, the continuous variables are integrated out analytically at fixed Lagrange multipliers, yielding a QUBO sampling formulation that avoids binarization and outperforms penalty-based QUBO on the quadratic p-median problem.","lead":"The authors show that a class of mixed-binary quadratic programs can be converted into a QUBO by analytically integrating out the continuous variables at fixed Lagrange multipliers, avoiding the usual binary discretization. Tested on a quadratic p-median relaxation, the method beats penalty-based QUBO on feasibility and is competitive with local search and a commercial solver.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Analytic integration may not leave a QUBO in the original binary variables when a continuous variable couples bilinearly to three or more binaries; the needed structural restriction is unstated.","rationale":"The reader correctly identified separability of the continuous sector as load-bearing. My concern refines this: even when the continuous variables factor into independent 1-D integrals, the resulting binary objective may not be quadratic. In particular, bilinear coupling between one continuous variable and multiple binary variables is allowed in an MBQP and yields a log-integral whose binary free energy generically has degree greater than two. This would break the abstract's explicit claim of 'leaving a QUBO over the original binary variables' — not merely an empirical inconvenience but a failure of the construction's mathematical core. The concrete test would settle it on a minimal instance. I still mark the verdict UNCHANGED because the full text may contain the needed class restriction or an approximation omitted from the abstract; the concern is strong enough that the full derivation and class definition must be checked, but it does not by itself justify rejecting a paper I have not seen. Thus the reader's UNVERDICTED low-confidence disposition remains appropriate.","tokens_in":908,"tokens_out":8370,"duration_ms":97654,"concrete_test":"Instantiate the minimal case: y ∈ [0,1], x1,x2,x3 ∈ {0,1}, fixed β = 1, and effective Lagrangean y(1 + x1 + x2 + x3) with no other terms. Compute H_eff(x) = -log ∫_0^1 e^{-β y(1+x1+x2+x3)} dy for all eight x. If the coefficient of x1 x2 x3 in the multilinear expansion of H_eff is nonzero (equivalently, the third finite difference y0 - 3y1 + 3y2 - y3 is nonzero, where y_s is H_eff when x1+x2+x3 = s), then the result is not a QUBO over the original binary variables. If nonzero, ask the authors to exhibit the class restriction or approximation that excludes this instance; if none is provided, the central construction fails for the claimed separable class.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction claims that at fixed Lagrange multipliers the continuous sector is integrated out analytically, leaving a QUBO over the original binary variables. This is not implied by separability of the continuous variables. Consider a single continuous variable y ∈ [0,1] whose fixed-multiplier contribution is y(a0 + Σ_{j=1}^k a_j x_j), which arises naturally from bilinear MBQP terms x_j y. The marginal Boltzmann weight is ∫_0^1 exp[-β y(a0 + Σ a_j x_j)] dy = (1 - exp[-β(a0 + Σ a_j x_j)]) / (β(a0 + Σ a_j x_j)), so the binary effective Hamiltonian is the negative logarithm of this expression. For k = 3, this is a nonlinear function of the linear form a0 + a1 x1 + a2 x2 + a3 x3. Its multilinear expansion generically contains the third-order term x1 x2 x3: the coefficient is the third finite difference of F(a) = -log((1 - e^{-β a})/(β a)) and is nonzero for typical β. Such a term is not representable as a QUBO on the original binary variables. The abstract states neither a restriction excluding this case (e.g., each continuous variable couples to at most two binary variables) nor an approximation that restores a quadratic form. Since the QUBO claim is the basis for the entire method and all downstream empirical claims, this missing condition is the most load-bearing concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a finite-temperature formulation for a separable class of mixed-binary quadratic programs (MBQPs) that avoids discretizing continuous variables. The central claim is that, at fixed Lagrange multipliers, the continuous sector can be integrated out analytically and enters only the multiplier update, leaving a QUBO over the original binary variables. The method is evaluated on the continuous relaxation of the quadratic p-median problem, with reported improvements in feasible-solution reliability over penalty-based QUBO, comparable or better conditional relative error relative to local search, and faster time-to-target than a commercial MIQP solver at the larger tested sizes. The supplied manuscript consists only of the abstract; no derivation, algorithmic details, or experimental statistics are included.","tokens_in":1171,"tokens_out":2993,"duration_ms":34613,"significance":"If the central construction is correct, the paper addresses a real bottleneck: standard QUBO conversions of continuous variables inflate binary dimension and can degrade feasibility. Avoiding discretization while retaining a QUBO at each multiplier step would be a practically valuable and falsifiable contribution, especially since the claims are benchmarked against external methods rather than fit to own data. The main significance hinges on whether the 'integrate out analytically' step genuinely preserves QUBO structure for the intended class of MBQPs; the abstract alone does not establish this.","major_comments":[{"comment":"The statement 'At fixed Lagrange multipliers, the continuous sector is integrated out analytically ... leaving a QUBO over the original binary variables' is not implied by separability of the continuous variables. For a single continuous variable y with fixed-multiplier contribution y(a0 + sum_j a_j x_j), the marginal weight is integral_0^1 exp[-beta y(a0+sum a_j x_j)] dy = (1 - exp[-beta(a0+sum a_j x_j)]) / (beta(a0+sum a_j x_j)). The effective binary Hamiltonian is the negative logarithm of this expression. When three or more binary variables couple to the same y, the multilinear expansion generically contains a third-order term x1 x2 x3, which is not representable as a QUBO on the original binary variables. The abstract must either explicitly restrict the class (e.g., each continuous variable couples to at most two binary variables) or describe an approximation that restores a quadrat","section":"Abstract, central construction"},{"comment":"The reported comparisons are conditioned on 'an appropriate inverse temperature,' but no selection rule for beta, Lagrange multiplier update schedule, or step-size protocol is provided. Without these, the feasibility and conditional relative error results are not reproducible. Moreover, the empirical statements contain no sample sizes, error bars, or statistical tests, so the strength of the improvements cannot be assessed from the abstract.","section":"Abstract, empirical methodology"},{"comment":"The time-to-target claim ('reaches the target faster than a commercial mixed-integer optimization solver toward the upper end of the tested range') lacks a precise definition of the target, instance sizes, number of runs, hardware/software configuration, and the solver name and version. Because this is one of the three headline empirical claims, the missing experimental detail prevents verification.","section":"Abstract, time-to-target comparison"}],"minor_comments":[{"comment":"The phrase 'a separable class of MBQPs' is undefined. Please state the exact structural conditions on the objective and constraints, especially the allowed coupling between continuous and binary variables.","section":"Abstract, terminology"},{"comment":"Define 'conditional relative error' and 'feasible solutions' precisely. For a continuous relaxation, feasibility should specify which constraints are checked and to what tolerance.","section":"Abstract, metrics"},{"comment":"The commercial solver should be named with version, and the quantum annealer or QUBO sampler should be identified with parameters, so readers can reproduce the comparison.","section":"Abstract, reproducibility"}],"recommendation":"major_revision","confidential_remarks":"I reviewed only the abstract, since no full text was supplied. The central QUBO-reduction claim is plausible for a suitably restricted class, but the abstract's unqualified wording is vulnerable to the demonstrated higher-order coupling counterexample. If the full manuscript contains the required structural restriction and a derivation, the major comments are likely addressable; if not, the method as stated would not deliver a QUBO at fixed multipliers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is genuinely non-standard: instead of binarizing continuous variables, they integrate them out at fixed Lagrange multipliers, leaving a QUBO over the original binaries and shunting the continuous sector into the multiplier update. For a separable class of MBQPs this is a neat way to avoid the dimension blowup of standard binarization. The empirical benchmarks against penalty QUBO, local search, and a commercial MIQP solver are the right kind of external check, and the feasibility improvement over penalty QUBO is plausible.\n\nThe soft spot is load-bearing. The stress-test counterexample is real: take one continuous variable y with fixed-multiplier weight exp[-β y(a0 + a1 x1 + a2 x2 + a3 x3)]. Integrating out y gives a marginal whose negative logarithm generically contains a nonzero x1 x2 x3 term. That is not quadratic. So the QUBO claim cannot hold for every separable MBQP; it needs an explicit restriction (e.g., each continuous variable couples to at most two binaries) or an approximation that restores quadraticity. The abstract says \"separable class\" but doesn't state that restriction, and separability alone doesn't buy you quadraticity. Everything downstream depends on this, so it has to be checked in the full text.\n\nTwo smaller issues: the performance is conditioned on \"an appropriate inverse temperature\" with no selection rule stated, and the error/time comparisons appear without error bars or instance details. Those are secondary, but they do limit what the abstract alone can support.\n\nI can't tell from the abstract whether the full paper supplies the missing condition. If it does, this is a solid practical contribution. If it doesn't, the empirical results are demonstrations of an approximation, not the claimed exact reduction. The paper deserves a serious referee — the idea is fresh and the failure mode is exactly what a careful reviewer should probe. I wouldn't cite it until the QUBO claim is pinned down.","headline":"Clever finite-temperature QUBO reformulation for separable MBQPs, but the abstract's central construction likely fails when a continuous variable couples to three or more binaries — needs a stated restriction or approximation.","tokens_in":1705,"tokens_out":2477,"would_cite":false,"duration_ms":25633,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C11","90C20","81P68"],"pacs":[],"model":"deepseek-v4-flash","headline":"A finite-temperature QUBO formulation that integrates out continuous variables analytically, rather than binarizing them, can solve a separable class of mixed-binary quadratic programs without discretization.","keywords":["QUBO","mixed-binary quadratic programming","continuous variable integration","quantum annealing","finite-temperature sampling","quadratic p-median","Lagrange multipliers"],"falsifier":"Run the method on a mixed-binary quadratic program whose continuous variables appear in products with each other (violating separability) and show there is no closed-form integration, forcing binarization again; or, within the separable class, find an instance where the reported feasibility and time-to-target advantages over penalty QUBO and a commercial MIQP solver reverse at a moderate size.","tokens_in":719,"feed_emoji":"⚛️","tokens_out":4434,"duration_ms":38121,"temperature":0.7,"pith_summary":"The paper aims to show that a class of mixed-binary quadratic programs can be solved by QUBO sampling without discretizing continuous variables. Standard conversions enlarge the binary dimension and make feasible states harder to sample; the authors avoid this by integrating out the continuous sector analytically at fixed Lagrange multipliers, leaving a QUBO over the original binary variables. They test the idea on the continuous relaxation of the quadratic p-median problem and report more reliable feasibility than penalty-based QUBO, conditional relative error comparable to or better than local search on larger instances, and faster time-to-target than a commercial MIQP solver at the upper end of tested sizes. If the construction holds, it widens the applicability of quantum-annealing and other QUBO samplers to practical mixed-integer problems that currently require discretization.","feed_headline":"Integrate out, don't binarize: QUBO sampling for mixed-binary programs","feed_subtitle":"Continuous variables are folded into the multiplier update, leaving a small QUBO; p-median tests show better feasibility and speed.","key_machinery":"The central mechanism is the finite-temperature formulation in which the continuous sector is integrated out at fixed Lagrange multipliers, reducing the mixed-binary problem to a QUBO over the original binary variables with multipliers updated in the outer loop. The analytic integration requires the continuous part of the objective and constraints to factorize into independent one-dimensional (or Gaussian-type) integrals—the 'separable class' condition—and this is what carries the dimension reduction.","core_discovery":"For a separable class of MBQPs, the continuous variables can be integrated out analytically at fixed Lagrange multipliers, so they appear only in the multiplier update and the binary part remains an exact QUBO over the original variables. The paper argues this avoids the dimension blowup and feasibility degradation caused by binarizing continuous variables, and backs it with numerical evidence on the quadratic p-median relaxation: feasible solutions more reliable than a penalty-based QUBO, relative error comparable to or better than local search, and faster time-to-target than a commercial solver at larger sizes.","pith_inferences":["The separable condition is the real boundary: any problem whose continuous variables couple in the objective or constraints cannot use this integration and would need a change of variables or a hybrid scheme, which the paper does not address.","The role of the inverse temperature β is under-specified; a principled schedule or adaptive rule for β could be a direct extension and would determine how robust the feasibility gains are across problem classes.","The approach could plausibly combine with other QUBO-based constraint-handling tools (e.g., disciplined convex QUBO conversion) to cover broader mixed-integer classes, with the same integration trick applied per separable block.","The p-median test case suggests facility-location problems, where continuous demand allocations are standard, are a natural first application; other logistics or network-design MBQPs with separable continuous allocation variables could inherit these gains."],"forward_implications":["QUBO samplers (quantum annealing, simulated annealing) can attack separable MBQPs without inflating the binary variable count.","Feasible solutions become easier to reach than with penalty-based QUBO conversions, since continuous feasibility is handled analytically rather than by penalty-weight tuning.","Solution quality, measured by conditional relative error, is comparable to or better than local search on the tested p-median instances, especially as problem size grows.","Time-to-target can beat a commercial MIQP solver for the larger instances tested, suggesting the method scales better in practice at the upper end.","The method supplies a template for other mixed-integer problems whose continuous blocks are separable at fixed multipliers."],"fun_headline_variants":["No binarization: integrate continuous variables in QUBO","Skip the binary conversion: QUBO with continuous integrated out","Better feasibility, speed: integrate continuous variables in QUBO","Mixed-binary QUBO: avoid discretizing continuous variables","Faster than MIP at scale by integrating continuous variables"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The continuous variables' contribution must factorize into independent one-dimensional or Gaussian integrals at fixed multipliers, so the method only applies to a separable class of MBQPs; if the continuous block couples variables, the analytic integration and all downstream claims collapse.","fun_headline_variants_meta":{"raw":{"variants":["No binarization: integrate continuous variables in QUBO","Skip the binary conversion: QUBO with continuous integrated out","Better feasibility, speed: integrate continuous variables in QUBO","Mixed-binary QUBO: avoid discretizing continuous variables","Faster than MIP at scale by integrating continuous variables"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000411,"raw_usage":{"total_tokens":1940,"prompt_tokens":694,"completion_tokens":1246,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":1161}},"tokens_in":438,"tokens_out":1246,"duration_ms":11425,"temperature":1.0,"reasoning_tokens":1161,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T07:54:17.896995+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a mixed-binary quadratic program whose continuous variables appear in products with each other (violating separability) and show there is no closed-form integration, forcing binarization again; or, within the separable class, find an instance where the reported feasibility and time-to-target advantages over penalty QUBO and a commercial MIQP solver reverse at a moderate size.","supporting_citations":[],"review_version":1}