{"id":"c589fe1c-0cc7-4843-87d7-c00222972a2c","arxiv_id":"2412.12156","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A deep-unfolded distributed ADMM/OSQP solver learns penalty parameters on small QPs and solves much larger ones with large wall-clock speedups, with PAC-Bayes bounds on relative progress.","lead":"This paper unrolls a distributed quadratic programming solver into a neural network, learning penalty and over-relaxation policies that solve large network-structured QPs much faster than OSQP after training only on small problems. A scientist or engineer would read it to see whether learned optimizer hyperparameters can replace hand tuning in large-scale MPC, network flow, and regression tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1 relation (R4) in Appendix C.2 rests on a false polar-cone orthogonality claim; the descent inequality (68) and hence Theorem 1 are not proven.","rationale":"The reader identified the same defect as the weakest assumption: relation (R4) in Lemma 1 is derived from an incorrect definition of polar cones. I agree. This is more fundamental than the PAC-Bayes gap or the baseline-selection issues: the claimed convergence guarantee of DistributedQP, and hence the principled foundation of DeepDistributedQP, rests on the descent inequality (68), which requires (R4). The empirical results may still be valuable and the theorem may be true, but the central theoretical claim is unsupported as written. Since the reader's verdict is REJECT and this concern confirms it, I recommend no change to the verdict.","tokens_in":39398,"tokens_out":7383,"duration_ms":74292,"concrete_test":"Run a one-dimensional counterexample to (R4): take m=1, b_i=0, α=1, ρ=1, z^k=0, s^k=0, λ^k=1. The updates give s^{k+1}=0 and λ^{k+1}=1; with t1=0 and t2=-1, λ^{k+1}^T(t1-t2)=1≠0, so (R4) fails. Then recompute both sides of (68) for this iterate; if the inequality is violated, the claimed descent relation is false. As an analytical check, replace the orthogonality step by the correct polar-cone inequality in the derivation of (74)-(76) and verify whether the sign of (λ^{k+1}-λ^*)^T(s^{k+1}-s^*) can be bounded; if it cannot, Theorem 1 requires a substantially different proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 1's relation (R4) is false. For C_i={s|s≤b_i}, the shifted cone C̄_i={p|p≤0} has polar cone C̄_i^o={p|p≥0}. Moreau decomposition gives λ^{k+1}_i/ρ^k_i ∈ C̄_i^o, hence λ^{k+1}_i∈C̄_i^o, which implies λ^{k+1}_i^T(t-b_i)≤0 for t∈C_i, not equality. The paper's Definition 1 incorrectly states that polar cones satisfy d^T d̄=0; the polar cone is defined by d^T d̄≤0. A one-dimensional counterexample (m=1, b_i=0, λ=1, t1=0, t2=-1) gives λ^T(t1-t2)=1, so (R4) fails. This relation is load-bearing: Lemma 3 uses it to cancel the last term of (R8) in (67); Lemma 4 uses it to assert the signs in (74)-(76) that produce the descent inequality (68); and the proof of Theorem 1 uses it to verify KKT condition (44c). Without (R4), monotonicity of the Lyapunov function is not established, summability (88) does not follow, and the convergence guarantee is unproven. The theorem may be salvageable by a different argument, but the submitted proof does not establish it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes DistributedQP, a distributed ADMM/OSQP hybrid for convex QPs with a consensus structure, and proves a convergence guarantee for iteration-varying local penalties and over-relaxation (Theorem 1, Appendix C). It then unrolls DistributedQP into a deep network, DeepDistributedQP, with learned feed-forward and feedback policies for the penalty parameters, and a centralized variant DeepQP. The authors derive PAC-Bayes bounds on a relative progress metric for the learned policy and report extensive experiments showing that DeepDistributedQP, trained on small problems, generalizes to much larger problems and requires fewer iterations than DistributedQP and far less wall-clock time than OSQP.","tokens_in":39692,"tokens_out":8069,"duration_ms":78959,"significance":"The paper addresses an important and timely problem: making distributed QP solvers adaptive and scalable through deep unfolding. The empirical study is broad (six problem classes, up to  ̃50K variables), and the use of implicit differentiation to backpropagate through the inner linear solver, as well as the PAC-Bayes formulation with a bounded relative-progress loss, are thoughtful components. If the advertised convergence guarantee were proven, this would be a solid contribution to learned distributed optimization. As submitted, however, the central proof of Theorem 1 rests on a false polar-cone identity, so the theoretical foundation of the paper is not established. The empirical results are interesting but do not by themselves compensate for the invalid proof of a headline claim.","major_comments":[{"comment":"Definition 1 incorrectly defines polar cones by d^T dbar = 0; the polar cone is defined by d^T dbar ≤ 0 for all d in the cone. Consequently, relation (R4), i.e., λ_i^{k+1T}(t1 - t2) = 0 for all t1, t2 ∈ C_i, is false. For C_i = {s | s ≤ b_i}, the shifted cone is Cbar_i = {p | p ≤ 0} and its polar is Cbar_i^o = {p | p ≥ 0}; with b_i = 0, λ_i^{k+1} = 1, t1 = 0 and t2 = -1, the left side of (R4) equals 1, not 0. This is not a cosmetic issue: Lemma 3 uses (R4) at equation (67) to cancel the final term of (R8), and Lemma 4 uses (R4) to assert the signs in (74)-(76) that are needed to derive the descent inequality (68). The summability (88) and the convergence conclusions of Theorem 1 consequently do not follow from the submitted proof.","section":"Appendix C.2, Lemma 1 (R4) and Definition 1"},{"comment":"Even if one replaces (R4) with the correct polar-cone inequality, the sign claims (74) and (75) are not valid. With C_i = {s ≤ b_i}, KKT condition (44c) yields λ*_i^T(s_i - s*_i) ≤ 0, and λ_i^{k+1} ∈ Cbar_i^o does not imply λ_i^{k+1T}(s_i - s*_i) ≥ 0; a scalar example with b_i = 0, s*_i = -1 (so λ*_i = 0), s_i = -2 and λ_i^{k+1} = 1 gives (λ_i^{k+1} - λ*_i)^T(s_i - s*_i) = -1 < 0. Thus inequality (76) is not established, and the derivation of (78) and (68) in Lemma 4 is invalid.","section":"Appendix C.2, Lemma 4, inequalities (74)-(76)"},{"comment":"The final step of the proof of Theorem 1 states that relations (R1) and (R4) verify KKT condition (44c)-(44d). Since (R4) is false, the KKT verification of the limit point is also unsupported. The convergence guarantee of DistributedQP is therefore unproven in the submitted manuscript. The authors would need to supply a substantially different proof of convergence, or remove the guarantee and reframe the contribution accordingly.","section":"Appendix C, proof of Theorem 1"}],"minor_comments":[{"comment":"The sample-convergence bound in (16) is misstated: D_KL(\\hat q_S(P;M) \\| M^{-1} log(2/ε)) is not a valid KL divergence because the second argument mixes a probability with a log term; please provide the correct binomial-tail or Bennett-style bound with proper definitions of all quantities.","section":"Equation (16)"},{"comment":"For OSQP with the direct method, the reported time is only for the first iteration under the assumption that the KKT factorization is reused for all subsequent iterations; this assumption should be stated in the main text at the point of comparison, as it materially affects the wall-clock claim.","section":"Table 6 caption and Section 6.2"},{"comment":"The sentence 'both DeepDistributedQP and Distributed significantly outperform OSQP' is missing 'QP' after 'Distributed'.","section":"Section 7, Conclusion"},{"comment":"The text 'performed on an system' should read 'performed on a system'.","section":"Section 6, first paragraph"}],"recommendation":"reject","confidential_remarks":"The empirical study is extensive and the PAC-Bayes machinery is presented carefully, but the main theoretical contribution (Theorem 1) is not proven because of the false polar-cone identity in Lemma 1. This is a central, non-local error; a revision would require a fundamentally new convergence proof. I recommend rejection, while encouraging the authors to resubmit a corrected version if the theorem can be proven by other means."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The central convergence proof is not sound. Lemma 1 in Appendix C states relation (R4): λ_i^{k+1}^T(t1 - t2) = 0 for all t1,t2 in C_i = {s: s ≤ b_i}. This is false. The proof relies on a nonstandard definition of polar cones (Definition 1 says d^T d̄ = 0; the actual defining inequality is d^T d̄ ≤ 0). A one-dimensional counterexample: b_i = 0, λ = 1, t1 = 0, t2 = -1 gives product 1. Since (R4) is used to cancel terms in Lemma 3, to get the signs in Lemma 4, and to verify KKT condition (44c), the descent inequality (68) and Theorem 1 are not established. This is load-bearing, not a gap.\n\nWhat's genuinely valuable here is the empirical architecture. DeepDistributedQP is the first deep-unfolded distributed ADMM for constrained QPs that I know of, with local iteration-varying learned penalty policies. The scaling results – trained on small problems, applied to 50K-variable instances – are impressive, and the iteration and wall-clock gains over OSQP are large. The PAC-Bayes bound is a step in the right direction, though it's on relative progress, not the optimality gap.\n\nOther soft spots: no code or data, no error bars, and the OSQP baselines use oracle-selected fixed penalties. Those are minor compared to the proof issue. The authors do honestly note the need for pre-solved training data and open questions on safety-critical guarantees.\n\nBottom line: send to a serious referee who can push on the proof, but I wouldn't cite it as is. If you want a reading-group lesson, this is a good case of how one bad definition can topple an elaborate proof.","headline":"Promising learning-to-optimize architecture with a broken convergence proof: Lemma 1's polar-cone claim is false, so Theorem 1 is unproven.","tokens_in":40248,"tokens_out":3867,"would_cite":false,"duration_ms":38226,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C20","90C25","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that unrolling a distributed OSQP-style ADMM solver into a deep network, trained on small problems, produces a QP solver that scales to 50K variables and 150K constraints with 1.5-3.5 times fewer iterations than its…","keywords":["quadratic programming","distributed optimization","ADMM","deep unfolding","learning to optimize","PAC-Bayes generalization","OSQP","consensus optimization"],"falsifier":"Run a one-step numerical check of Lemma 1's relation (R4) on a small inequality-constrained QP: after the projection update (6), compute $\\lambda_i^{k+1}$ and evaluate $\\lambda_i^{k+1\\top}(t_1-t_2)$ for two feasible slack vectors $t_1,t_2$ with $t_j \\le b_i$; any nonzero value contradicts (R4) and would falsify the proof of Theorem 1 as written.","tokens_in":39165,"feed_emoji":"🧮","tokens_out":5300,"duration_ms":52929,"temperature":0.7,"pith_summary":"The paper tries to establish that a deep-unrolled version of a newly introduced distributed quadratic programming solver can be trained on small problems and then solve much larger network-structured QPs with fewer iterations and far less wall-clock time than a standard centralized solver. The underlying method, DistributedQP, combines OSQP-style operator splitting with a consensus formulation so that each node solves only local subproblems. Unrolling this optimizer for a fixed number of layers and learning the penalty and over-relaxation parameters as feedback policies yields DeepDistributedQP, which the authors test on random QPs, optimal control, network flow, and distributed LASSO. The authors also claim PAC-Bayes bounds on the expected optimality gap for unseen problems, and they prove convergence of DistributedQP to optimality under iteration-varying local parameters.","feed_headline":"A QP solver trained on small problems scales to 50K variables","feed_subtitle":"Unrolled distributed ADMM with learned feedback cuts iterations 1.5-3.5x and beats OSQP wall-clock by orders of magnitude.","key_machinery":"The load-bearing object is the unrolled iteration graph of DistributedQP: local nodes maintain primal variables $(x_i, z_i, s_i)$ and dual variables $(\\nu_i, \\lambda_i, y_i)$, a consensus step averages local $x_i$ into the global variable $w$ using weights $\\mu_i$, and each layer maps local residuals to new penalty parameters through small fully-connected feedback networks. The local update solves a KKT system that is reduced to a positive-definite linear system and solved by conjugate gradients, with gradients through the CG loop obtained by implicit differentiation rather than by unrolling inner iterations. A separate PAC-Bayes argument trains a Gaussian distribution over the policy weights and bounds the expected progress metric using the inverse-KL form in Theorem 3.","core_discovery":"The central claim is that DeepDistributedQP, trained only on low-dimensional instances, generalizes to problems with up to 50K variables and 150K constraints using the same learned policy, reaching a target accuracy in 1.5-3.5 times fewer iterations than DistributedQP and orders-of-magnitude less wall-clock time than OSQP. The paper further claims that DistributedQP is guaranteed to converge to the optimal solution even when each node uses its own iteration-varying penalty parameters, a property needed for deep unfolding. On top of this, Theorem 3 provides a PAC-Bayes bound on the expected relative progress of the final iterate for unseen problems drawn from the same distribution as the training data, and the experiments report that these bounds are tight relative to actual performance.","pith_inferences":["The same unrolling-and-feedback recipe could be transplanted to other two-block ADMM solvers with projection steps, though the paper does not claim this extension itself.","The reported wall-clock advantage over OSQP depends on the conjugate-gradient indirect solve and GPU parallelism; on very small problems or CPU-only hardware, the comparison curve could shift, which is a testable extension of the paper's scaling claims.","The PAC-Bayes guarantee is distribution-dependent, so applying the trained policy to problem classes far from the training distribution falls outside the certified regime; a natural next experiment would measure how the bound degrades under distribution shift.","Because training uses the known optimal solution as supervision, a self-supervised variant that trains on residual norms instead of optimality gaps would remove the need for pre-solved datasets, which the paper notes only as future work."],"forward_implications":["Practitioners could train a distributed QP policy once on small problems and deploy it on much larger network-structured QPs without retraining at scale.","Closed-loop policies that adapt penalty parameters from local residuals reach target accuracy faster than open-loop or fixed parameters, so learned residual feedback can replace manual ADMM tuning.","The PAC-Bayes guarantee, if valid, makes the average final-iterate progress of the learned solver certifiable with high probability for unseen problems from the same distribution.","Because the underlying method is a two-block ADMM scheme, the same unrolling recipe is a direct candidate for other linearly constrained convex problems beyond quadratic programming."],"supporting_citations":[{"why":"Supplies the OSQP operator-splitting updates that DistributedQP generalizes and the direct/indirect linear-system solvers used for local updates.","marker":"Stellato et al. (2020)"},{"why":"Provides the consensus-ADMM template and the residual-ratio heuristic used to design feedback policies and to compare adaptive parameter rules.","marker":"Boyd et al. (2011)"},{"why":"Supplies the classical two-block ADMM convergence result that covers the simplified constant-parameter version of DistributedQP.","marker":"Deng and Yin (2016)"},{"why":"Gives the adaptive-relaxed ADMM convergence analysis that Theorem 1 extends to node-local iteration-varying penalties.","marker":"Xu et al. (2017)"},{"why":"Provides the PAC-Bayes bound template used to derive Theorem 3.","marker":"Alquier (2024)"},{"why":"Supplies the sample-convergence and bound-optimization practice used to evaluate the PAC-Bayes guarantee.","marker":"Majumdar et al. (2021)"},{"why":"Establishes implicit differentiation of optimization layers, which Theorem 2 adapts to avoid unrolling CG iterations.","marker":"Amos and Kolter (2017)"},{"why":"Provides the differentiable convex optimization layer machinery that the implicit-differentiation result extends.","marker":"Agrawal et al. (2019)"}],"fun_headline_variants":["Train small, solve huge: deep QP scales to 50K variables","DeepDistributedQP: trained on small, solves 50K-variable QPs","Orders-of-magnitude faster than OSQP after scaling to 50K","Small problems, big solutions: QP solver scales to 50K","From small to 50K: deep QP that generalizes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence proof of Theorem 1 relies on Lemma 1's relation (R4), which asserts that the updated dual variable $\\lambda_i$ is orthogonal to the difference of any two feasible slack vectors; if that orthogonality does not actually follow from the polar-cone argument, the descent inequality (68) and the convergence theorem's proof as written collapse.","fun_headline_variants_meta":{"raw":{"variants":["Train small, solve huge: deep QP scales to 50K variables","DeepDistributedQP: trained on small, solves 50K-variable QPs","Orders-of-magnitude faster than OSQP after scaling to 50K","Small problems, big solutions: QP solver scales to 50K","From small to 50K: deep QP that generalizes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3307,"prompt_tokens":986,"completion_tokens":2321,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":2221}},"tokens_in":602,"tokens_out":2321,"duration_ms":19149,"temperature":1.0,"reasoning_tokens":2221,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:03:48.936166+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a one-step numerical check of Lemma 1's relation (R4) on a small inequality-constrained QP: after the projection update (6), compute $\\lambda_i^{k+1}$ and evaluate $\\lambda_i^{k+1\\top}(t_1-t_2)$ for two feasible slack vectors $t_1,t_2$ with $t_j \\le b_i$; any nonzero value contradicts (R4) and would falsify the proof of Theorem 1 as written.","supporting_citations":[],"review_version":1}