{"id":"88fe84b9-d550-4820-a456-74428e8b579a","arxiv_id":"2412.01051","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A deep-unrolled PDQP network trained with an unsupervised KKT-residual loss predicts near-optimal QP solutions and accelerates the PDQP solver by up to 45%.","lead":"This paper introduces PDQP-net, a neural network that unrolls the PDQP solver for convex quadratic programs, trained without solver-generated labels using KKT residuals. Warm-starting PDQP with these predictions speeds up solving by up to 45% in-distribution and 14-31% on out-of-distribution instances.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Acceleration claim rests on an unvalidated proxy: KKT loss is not shown to predict PDQP warm-start iteration reduction, and the cited convergence theory does not cover warm-started inequality-constrained PDQP.","rationale":"The reader's weakest assumption correctly identifies the surrogate loss-to-speedup link as unproven. I agree that this is the most load-bearing empirical risk. My read is partial rather than full agreement because I also flag that the theoretical support for warm-starting is absent: the cited PDQP linear convergence theorem (Appendix B, Theorem B.1) is for equality-constrained problems and zero initialization, so even the mechanism by which a warm start should reduce iterations is not established for the inequality-constrained setting used in the experiments. The theoretical claims of alignment (Theorem 3.1) and O(log 1/epsilon) neurons (Proposition 3.1) also have correctness issues: the proof of Theorem 3.1 equates a matrix X^k in R^{n x d_x} with a vector x^k in R^n, ignoring the channel-expansion dimension and the role of the MLPs fx, fy, gx, gy; the proof is fixable by taking d=1 and identity maps, but as written it is not rigorous. Proposition 3.1 inherits the equality-constraint limitation and asserts a neuron count without specifying the constant width of the MLPs. None of these theoretical issues directly falsify the acceleration numbers, but they mean the theory cannot be used to compensate for the missing empirical validation of the loss-to-speedup proxy. The paper does have independent support: the algorithm-unrolling architecture is well motivated, the unsupervised loss avoids solver-generated labels, and the reported residual improvements over supervised learning and GNN baselines are concrete and internally consistent. The empirical comparison, however, lacks error bars, code release, and a direct test of the proxy relationship. Given the reader's CONDITIONAL verdict, my concern does not change the verdict: it sharpens the conditions that should be imposed, namely a correlation analysis between loss components and iteration reduction, and a clearer theoretical statement for warm-started inequality-constrained PDQP. Thus the verdict remains CONDITIONAL, and I mark it UNCHANGED relative to the reader's assessment.","tokens_in":18162,"tokens_out":7763,"duration_ms":74263,"concrete_test":"On the 50 QPLIB-8559 test instances, generate a pool of candidate warm starts from PDQP-net checkpoints with different loss values (e.g., early, middle, and final training epochs, or checkpoints trained with modified loss weights that keep total loss similar). For each candidate, record the three normalized loss components and the number of PDQP iterations to the solver's default tolerance. Compute the Spearman rank correlation between the total loss and the iteration reduction relative to the zero-start baseline, and also report the same correlation for each loss component separately. If |rho| < 0.5 or the relationship is non-monotone, the loss is not a faithful proxy for warm-start quality, and the reported 14-45% accelerations cannot be attributed to the KKT-informed training objective.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that PDQP-net predictions, trained with the unsupervised KKT loss of Section 3.3/Appendix G, reliably accelerate PDQP by 14-45%. This requires that smaller normalized residuals (r_primal_hat, r_dual_hat, r_gap_hat) translate into fewer PDQP iterations when used as a warm start. The paper provides no such quantitative link. The loss is an ad hoc sum of terms normalized by non-gradient denominators such as max(||b||_inf, ||Ax||_inf), and r_gap_hat includes an RCC term whose derivation (Appendix H) assumes stationarity while stationarity is penalized separately with arbitrary weighting. Nothing in the paper shows that minimizing this surrogate moves the iterate closer to the PDQP trajectory or to the optimal set in the sense required by warm-start theory. Moreover, the only convergence result invoked for PDQP (Theorem B.1, quoted from Lu & Yang) is stated for equality-constrained QPs and for zero initialization with a fixed restart schedule; it does not cover warm-started inequality-constrained PDQP with arbitrary initial points. Table 4 already shows a negative acceleration (-2.12%) on DUAL4, and the out-of-distribution evidence is only five instances. If the loss is not a faithful proxy, the claimed speedups may be artifacts of the specific six in-distribution and five out-of-distribution instances rather than a property of the framework. This is the single most load-bearing weakness: the framework's practical value is the acceleration, and the mechanism connecting the trained loss to that acceleration is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PDQP-net, a deep-unrolling network that mimics the PDQP primal-dual algorithm for convex quadratic programs, and trains it with an unsupervised loss built from normalized KKT residuals (primal feasibility, dual stationarity, and primal-dual gap). The authors prove (Theorem 3.1) that with a specific parameter assignment the network can exactly reproduce the PDQP iterate sequence, argue (Proposition 3.1) that a K-layer network inherits PDQP's linear convergence with O(log 1/epsilon) neurons, and use Proposition 3.2 to motivate why supervised training can produce large primal-dual gaps. Empirically they report lower residuals for the unsupervised PDQP-net relative to supervised and GNN baselines on QPLIB and synthetic instances, and report 14-45% warm-start acceleration of PDQP, including out-of-distribution Maros-Meszaros instances.","tokens_in":18542,"tokens_out":8816,"duration_ms":82283,"significance":"If the results hold, this is a practically useful extension of learning-to-optimize to convex QPs, and the unsupervised KKT-informed loss is an appealing idea that avoids solver-generated labels. The work is original and clearly within the journal's scope. Strengths include the explicit unrolling construction, the exact replication argument in Theorem 3.1, and the focus on primal-dual gap quality rather than only distance to the optimum. However, the theoretical support is substantially weaker than claimed: Proposition 3.1 relies on a convergence theorem for equality-constrained QPs that does not cover the inequality-constrained problem actually solved, and the proof of Proposition 3.2 contains an invalid Taylor bound. The central acceleration claim also lacks statistical support and a demonstrated link between the surrogate loss and warm-start performance. These issues are fixable but currently block acceptance.","major_comments":[{"comment":"The stated bound is not proved. The Taylor expansion step uses |P(x0)-P(x*)| <= ||c|| ||x0-x*|| + (1/2)||Q|| ||x0-x*||^2, which omits the Qx* term in the gradient of P at x*. For a constrained optimum, Qx*+c need not vanish, and the inequality can fail; for example, with Q=I, c=0, the constraint x1 >= 1 gives x*=e1 and x0=e1+delta e2, for which the left side is delta + 0.5 delta^2 while the right side is 0.5 delta^2. Consequently Proposition 3.2's conclusion is not established, and the motivation for preferring the KKT-residual loss over a distance-based loss needs a corrected argument.","section":"Appendix C, Proposition 3.2"},{"comment":"The linear convergence result quoted as Theorem B.1 is for equality-constrained QPs (min 1/2 x^T Q x + c^T x subject to Ax = b), whereas the paper's problem (1) has inequality constraints Ax >= b and bound constraints l <= x <= u. The paper does not show that the fixed-restart schedule and parameter choices required by Theorem B.1 apply to the PDQP form used in Algorithm 1. Therefore the claim that PDQP-net 'exhibits at least the same convergence properties' and the O(log 1/epsilon) neuron bound are not established for the problem class under study; Theorem 3.1 shows only exact replication of finitely many iterates, not convergence.","section":"Appendix B, Proposition 3.1"},{"comment":"The central speedup claim assumes that smaller normalized KKT residuals produce larger PDQP warm-start gains, but no experiment or analysis connects the value of the loss L to the number of PDQP iterations. The paper reports residual values and speedups in separate tables, but does not show their relationship; there is no ablation varying the relative weights of r_hat_primal, r_hat_dual, and r_hat_gap, and no comparison against randomized or supervised warm starts matched by residual value. Because the acceleration is the paper's main practical payoff, a direct scatter plot or rank-correlation analysis between residual components and iteration reduction, together with repeated-run averages, is needed to support the causal interpretation.","section":"Sections 3.3, 4.3, and Appendix G"},{"comment":"The empirical evidence as reported is not reproducible: no seeds, no error bars, no hyperparameter details (network depth, channel widths, MLP sizes, number of training instances, loss weights), and no code. In addition, Table 2 contains an internal inconsistency: for QPLIB-8845 the reported improvement is 30.64%, but (104.51 - 76.17)/104.51 = 27.1%. Section 4.5's statement that the framework is 'nearly 20 times faster' is also contradicted by the 31.61% speedup shown in Table 4. These discrepancies must be corrected, and the experiments repeated across multiple seeds, before the acceleration claims can be assessed fairly.","section":"Tables 2 and 4"}],"minor_comments":[{"comment":"The caption says 'two QPLIB instances' but three panels (a)-(c) are shown, and the text says 'for both problems' while discussing three datasets; this should be corrected.","section":"Section 4.6, Figure 2"},{"comment":"There are several typos, including 'Supervised-Learning vs Unsupervised-Learning' in the section title, 'trianing' in Figure 3, and 'intances' in Section 4.5; a careful proofread is needed.","section":"Section 4.2 and Appendix G"},{"comment":"The improvement metric is written as 'Improv. := PDQP - ours / PDQP' without clarifying whether time or iterations are used for computation; a formal definition with the exact numerator and denominator would remove ambiguity.","section":"Section 4.1"},{"comment":"The reported number of nonzeros for QSHIP04L is '6,33', which appears to be a typo for 633, and the formatting of the QISREAL entry is inconsistent; these should be checked against the dataset documentation.","section":"Table 6"},{"comment":"The sentence 'It is necessary for zeta = 0' should be replaced by a statement about the projected gradient vanishing; the sign conventions in the RCV expression should also be verified for consistency.","section":"Appendix H"},{"comment":"The remark states as established that PDQP-net approximates optimal solutions with O(log 1/epsilon) neurons, but this depends on Proposition 3.1, whose proof currently relies on an inapplicable theorem; the remark should be conditioned on a corrected convergence argument.","section":"Remark 3.1"}],"recommendation":"major_revision","confidential_remarks":"I am sympathetic to the framework and the reported qualitative trends, but the theoretical claims in Propositions 3.1 and 3.2 are not currently supportable, and the empirical core requires seeds, error bars, and a direct analysis connecting the surrogate loss to warm-start gains. These are substantial but addressable issues; with a corrected theory and a more rigorous empirical section, the paper could be a good contribution to the learning-to-optimize literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid, incremental extension of the unrolling work from LPs to convex QPs. The unsupervised KKT loss is a sensible way to train without solver labels, and the warm-start speedups look real on the tested instances. The main claims mostly hold up, but there are a few soft spots that need fixing before I'd trust the details.\n\nWhat's new and good: PDQP-net unrolls PDQP, a matrix-free first-order QP solver, and trains with a normalized KKT residual loss rather than supervised labels. That's a genuine new application of the unrolling idea, done reasonably carefully. The experiments show consistent improvements over both supervised training and generic GNNs, with up to 45% time savings in-distribution and mostly positive results out-of-distribution. Inference overhead is below 1% of solver time, which makes the method practical. The alignment theorem (3.1) is straightforward but valid.\n\nSoft spots, in order of severity:\n- Proposition 3.2's proof uses |a| - |b| = |a-b|. That's wrong. The bound itself can be recovered by applying the triangle inequality to the difference (since P* - D* - RCC* = 0), so the result is salvageable, but the proof as written is mathematically invalid.\n- Proposition 3.1 claims a PDQP-net with O(log 1/ε) neurons. That's overstated: the argument treats the per-layer width as constant, but the network input and output scale with n and m. The correct statement is O((n+m) log 1/ε), or just 'polynomial size' as in the abstract. As written, the claim hides the problem-size dependence.\n- No error bars, no code, and minimal hyperparameter disclosure. The QPLIB-3547 row shows 40% time improvement with only 6% iteration reduction, which is odd; I'd want standard deviations before trusting the timing numbers.\n- Out-of-distribution evidence is only five instances, one of which (DUAL4) is slower. That's thin, and the paper admits the failure but doesn't analyze it.\n- The KKT loss is a surrogate for warm-start quality, and the paper doesn't demonstrate a direct link between loss value and iteration count. That's a fair concern, though not fatal: the aggregate results are plausible. A correlation plot would strengthen the claim.\n\nThe stress-test note is a bit harsh in demanding a proof of the proxy; for an empirical paper it's enough to show the method works across datasets. But the missing error bars and the small OOD set are legitimate concerns.\n\nRecommendation: send to peer review. The method is useful, the paper is readable, and the flaws are fixable. Ask for code/data, corrected proofs, and error bars. It should be accepted only after those are addressed.","headline":"A useful extension of unrolled LPs to QPs with a sensible unsupervised KKT loss and promising speedups, but the proof of Prop 3.2 is wrong, the neuron bound is overstated, and the experiments need error bars and code.","tokens_in":19070,"tokens_out":6219,"would_cite":true,"duration_ms":56212,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C20","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Unrolling PDQP into a network and training it with an unsupervised KKT-condition loss yields near-optimal warm starts that accelerate solving by up to 45% in-distribution and 14-31% out-of-distribution.","keywords":["convex quadratic programming","deep unrolling","primal-dual hybrid gradient","unsupervised learning","KKT conditions","warm-starting","learning to optimize","graph neural networks"],"falsifier":"On a held-out family of convex QPs, train PDQP-net with the proposed loss and compare PDQP iteration counts for low-loss predictions against randomly perturbed starting points with controlled larger loss; if low-loss starts do not consistently reduce iterations, or if equal-loss starts produce widely different speedups, the residual-to-speedup link underlying the framework fails.","tokens_in":1670,"feed_emoji":"⚡","tokens_out":4928,"duration_ms":105968,"temperature":0.7,"pith_summary":"This paper tries to make convex quadratic programming faster by learning good starting points. It unrolls the PDQP solver, a matrix-free primal-dual hybrid gradient algorithm, into a neural network called PDQP-net, then trains it with an unsupervised loss built from KKT conditions: primal residual, dual residual, and primal-dual gap. The central claim is that the resulting predictions, used as warm starts, cut PDQP solving time by up to 45% on in-distribution instances and by 14-31% on out-of-distribution instances.","feed_headline":"Unrolled PDQP network speeds QP solving by up to 45%","feed_subtitle":"Trained with a KKT-informed loss and no solver labels, it still speeds up out-of-distribution instances by 14-31%.","key_machinery":"The load-bearing object is the unrolled PDQP layer: each layer of PDQP-net mirrors one PDQP iteration, with learnable step sizes $\\beta_k,\\eta_k,\\theta_k,\\tau_k$, learnable channel-expansion matrices $W^k_{\\bar{x}},W^k_y,W^k_\\theta$, and piecewise-linear projection operators expressed through ReLU and shifts. Unrolling guarantees that PDQP-net can align exactly with PDQP when the extra matrices are set to the identity. The training signal is the unsupervised KKT residual loss $\\mathcal{L}=\\hat{r}_{\\mathrm{primal}}+\\hat{r}_{\\mathrm{dual}}+\\hat{r}_{\\mathrm{gap}}$, where each residual is normalized, so the network directly minimizes feasibility violations and the optimality gap instead of distance to collected labels.","core_discovery":"The central discovery is that a network obtained by unrolling the PDQP iteration, with learnable step sizes and ReLU-based projections, can replicate the PDQP sequence exactly for a suitable parameter assignment, and that training it to minimize normalized KKT residuals produces primal-dual predictions with much smaller primal-dual gaps than supervised training. The theory states that a K-layer PDQP-net of polynomial size can align with PDQP (Theorem 3.1), that $O(\\log(1/\\epsilon))$ neurons suffice for an $\\epsilon$-optimal solution while preserving linear convergence (Proposition 3.1), and that supervised predictions close in distance can nonetheless have large duality gaps (Proposition 3.2). Empirically, warm-started PDQP achieves up to 45% acceleration and up to 49% iteration reduction, with 14-31% acceleration on the out-of-distribution Maros-Meszaros instances.","pith_inferences":["The same unrolling-plus-KKT-loss recipe could transfer to other first-order solvers whose updates are piecewise-linear and differentiable with respect to step sizes.","Because the loss is fully unsupervised and differentiable, it could be embedded inside larger end-to-end systems such as model-predictive control where optimal QP labels are not available at train time.","The empirical link between normalized KKT residuals and solver speedup is not proven, so a quantitative residual-to-iteration bound would make the acceleration claim robust for unseen distributions.","The logarithmic neuron bound suggests that QP families with fixed sparsity need only mild network growth as required accuracy increases, encouraging tests on much larger instances."],"forward_implications":["PDQP-net warm starts reduce PDQP solving time by about 30-45% and iteration count by up to 49% on the tested in-distribution QP instances.","The framework generalizes to out-of-distribution instances, achieving 14-31% acceleration on most Maros-Meszaros test problems.","Unsupervised training eliminates the need for solver-generated optimal solutions as labels, while producing lower primal-dual gaps than supervised training.","The theoretical results tie network size to solver convergence: a PDQP-net with $O(\\log(1/\\epsilon))$ neurons reaches $\\epsilon$-accuracy with linear convergence.","The unrolled architecture outperforms a conventional GNN baseline on solution-quality residuals across real-world and synthetic benchmarks."],"supporting_citations":[{"why":"Supplies the unrolled PDLP framework and channel-expansion technique that PDQP-net adapts.","marker":"Li et al. (2024)"},{"why":"Defines the PDQP algorithm and the KKT residual metric used to construct the unsupervised loss.","marker":"Lu & Yang (2024)"},{"why":"Provides the theoretical GNN expressive-power result and the conventional GNN baseline in comparisons.","marker":"Chen et al. (2024b)"},{"why":"Supplies the real-world QPLIB instances used for training and testing.","marker":"Furini et al. (2018)"},{"why":"Supplies the out-of-distribution QP instances used to evaluate generalization.","marker":"Maros & Mészáros (1999)"}],"fun_headline_variants":["PDQP-net: unsupervised unrolling speeds QP solving by 45%","No solver labels? PDQP-net still accelerates QPs by up to 45%","Deep unrolling for QPs: KKT loss, no solvers, 45% faster","Unsupervised PDQP-net: learning optimal QP solutions without solvers","Warm-start PDQP with PDQP-net: up to 45% acceleration on QPs"],"cache_read_input_tokens":21120,"weakest_assumption_plain":"The acceleration result rests on the assumption that a prediction with smaller normalized KKT residuals will reliably make PDQP converge in fewer iterations; the paper does not prove a quantitative link between residual values and solver speedup.","fun_headline_variants_meta":{"raw":{"variants":["PDQP-net: unsupervised unrolling speeds QP solving by 45%","No solver labels? PDQP-net still accelerates QPs by up to 45%","Deep unrolling for QPs: KKT loss, no solvers, 45% faster","Unsupervised PDQP-net: learning optimal QP solutions without solvers","Warm-start PDQP with PDQP-net: up to 45% acceleration on QPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000331,"raw_usage":{"total_tokens":1872,"prompt_tokens":1001,"completion_tokens":871,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":760}},"tokens_in":617,"tokens_out":871,"duration_ms":6697,"temperature":1.0,"reasoning_tokens":760,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:44:28.669826+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out family of convex QPs, train PDQP-net with the proposed loss and compare PDQP iteration counts for low-loss predictions against randomly perturbed starting points with controlled larger loss; if low-loss starts do not consistently reduce iterations, or if equal-loss starts produce widely different speedups, the residual-to-speedup link underlying the framework fails.","supporting_citations":[{"cited_title":"A repository of convex quadratic programming problems","cited_arxiv_id":null,"evidence_quote":"Supplies the out-of-distribution QP instances used to evaluate generalization."}],"review_version":1}