{"id":"0774da1d-59cd-4e32-a7cb-3c44518a9230","arxiv_id":"2509.02849","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors prove that arrow decomposition can be applied after forming moment-SOS relaxations of polynomial matrix inequalities, yielding convergent lower bounds and significant computational speedups in structural optimization.","lead":"This paper extends the arrow decomposition technique, previously used to break up large linear matrix inequalities, to polynomial optimization problems with polynomial matrix inequalities, and combines it with moment-sum-of-squares hierarchies. The authors report large speedups on structural optimization problems such as frame design, with up to 100x faster solves in some cases.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Variable reduction in §4.3 assumes constant null spaces of A_k(x); zero-area elements in topology optimization violate this, so the reported speedups may not be valid for the target problems.","rationale":"The reader's conditional verdict is reasonable. Theorem 5 is the theoretical centerpiece and its proof appears sound: the posterior hierarchy adds variables at the localizing-matrix level, sums to Md(Gy), and the sandwich argument gives convergence. The strongest claim about convergence is therefore not endangered. The weak point is the practical reduction used in all reported speedups. Section 4.3 requires constant range/null space; topology optimization with x ≥ 0 permits zero areas, where element stiffness terms disappear and null spaces grow. The paper tries to prove this away in Appendix A.2, but the argument is a pointwise-to-identity fallacy. Since the numerical examples are the main evidence of computational advantage, this is load-bearing for the paper's applied claims, even if not for the abstract convergence theorem. The same concern was identified by the reader. I also note the unproven statement pprior_AD,r = pr in §4.1, but the posterior results do not depend on it, so it is secondary. No change to the conditional verdict is needed: the paper should either prove the reduction under degenerate designs, add a positive lower bound x_e ≥ ε in the numerical study, or clearly restrict the reduction to problems where null spaces are constant.","tokens_in":39785,"tokens_out":10342,"duration_ms":125730,"concrete_test":"Recompute the double-hinged beam example at relaxation r = 2 (Section 6.1) twice: (i) solve (ADmSOS2) with full bD, bC variables and no null-space elimination; (ii) solve the same instance with the Section 4.3 reduction using W_k from rigid-body modes. Record the optimal x and check whether any x_e = 0. If the optimal x contains a zero-area element and the two lower bounds differ (or only one is feasible), the reduction is not equivalent and the reported speedups are unsupported. As a smaller analytical check, evaluate Null(K_k(x)) for any x with a zero element in this instance; if it is larger than the rigid-body space, the constant-null-space assumption fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 4 and Section 4.3 reduce (ADmSOS2) by fixing matrices P_k with Span(A_k(x)) ⊆ Span(P_k) and, for elimination, assuming A_k has a constant null space N_k on K so that Span(B_k + Π_kD) ⊆ Span(P_k) can be imposed by linear equations with W_k spanning N_k. In the frame examples, K_k(x) = K0 + Σ_{e∈D_k}(K_e^(1)x_e + K_e^(2)x_e^2 + K_e^(3)x_e^3) and the constraints only enforce x_e ≥ 0. If an optimal or intermediate design has x_e = 0, the term vanishes and Null(K_k(x)) can enlarge beyond the rigid-body modes used to define W_k. The Appendix A.2 justification is not valid: from K_k(x)v = 0 at a single x it concludes each monomial coefficient annihilates v, but linear independence of monomials only supports this if the identity holds for all x. A simple counterexample is K(x) = x_1 diag(1,0) + x_2 diag(0,1), where Null(K(0,0)) = R^2 while Null(K(1,1)) = {0}. Consequently the reduced problem solved in the numerical section is not proven equivalent to (ADmSOS2) when zero-stiffness elements are allowed; the numerical speedups and the 'zero-stiffness design' claims rest on an unverified assumption. This does not invalidate Theorem 5, which needs no such assumption, but it undermines the practical central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the Arrow Decomposition (AD) method for linear matrix inequalities to polynomial optimization problems with polynomial matrix inequalities, and combines it with moment-SOS hierarchies. The main theoretical results are: (i) Theorem 3, which generalizes the AD theorem of Kocvara (2020) from positive definite to positive semidefinite blocks; (ii) a posterior-to-mSOS AD hierarchy (ADmSOS2) whose optimal values lie between the standard mSOS bound and the true optimum and converge to the true optimum as the relaxation order goes to infinity (Theorem 5); and (iii) a variable-reduction procedure (Proposition 4 and Section 4.3) for matrices with constant range and null space, which is then used in the numerical experiments on topology optimization of frame structures. The paper also discusses an AD-prior-to-mSOS formulation, physical interpretations of the additional variables, and reports substantial computational speedups on beam and frame examples, with code available online.","tokens_in":40154,"tokens_out":11049,"duration_ms":130135,"significance":"If the results hold as stated, this is a useful contribution to the sparse polynomial optimization literature. Theorems 3 and 5 are clean, self-contained, and appear correct; Theorem 5 provides a convergent hierarchy with smaller semidefinite blocks than the standard mSOS relaxation, and the numerical evidence of speedups is striking. The paper also ships reproducible code (GitLab), which is a strength. However, the practical central claim—that the reduced problem solved in the numerical section is equivalent to the ADmSOS2 relaxation—rests on a constant-null-space assumption that is not justified for the targeted topology-optimization problems when zero-stiffness elements are allowed. This does not invalidate the core theoretical convergence results, but it materially weakens the paper's main computational claim as currently written.","major_comments":[{"comment":"The variable-reduction procedure used in the numerical section assumes that each A_k(x) has constant null space on K. The justification in Appendix A.2 is not valid. From K_k(x)v=0 at a single point x one cannot conclude that K_0 v=0 and K_e^(i)v=0 for all e,i; linear independence of monomials only supports such a conclusion if the identity holds for all x. A simple counterexample is K(x)=x_1 diag(1,0)+x_2 diag(0,1), where Null(K(0,0))=R^2 but Null(K(x))={0} for x_1,x_2>0. In the frame problems, K includes x_e>=0, so zero-area elements can enlarge the null space beyond rigid-body modes. Consequently, the reduced problem solved in Section 6 is not proven equivalent to (ADmSOS2), and the reported speedups and zero-stiffness design claims rest on an unverified assumption. Theorem 5 does not depend on this assumption, but the practical central claim does. Please either prove the required con","section":"Section 4.3 / Appendix A.2"},{"comment":"The statement 'Under Assumptions 1 and 3, and according to Proposition 2 and Theorem 2, we can show that pprior_AD,r = pr for any relaxation order r' is asserted without proof. Proposition 2 gives p_AD=p*, but equality of the degree-r mSOS truncations of two equivalent POPs is not automatic: adding variables and constraints can change the relaxation bound. This claim is used to position the prior approach as merely a bridge to the posterior approach. Please supply a proof or explicitly weaken the statement to an inequality or a limiting equality.","section":"Section 4.1, after (ADmSOS1)"},{"comment":"The proof of the case u∉Span(P_k) contains an incorrect implication: from u∉Span(A_k) it does not follow that u^T A_k u=0. For example, with A=diag(1,0,0), P=[e1 e2], and u=e1+e3, we have Span(A)⊆Span(P), u∉Span(P), yet u^T A u=1. The equivalence can still be recovered by decomposing u into its projection onto Span(P_k) and the orthogonal complement, using Span(A_k), Span(B_k+D_k)⊆Span(P_k), but the proof must be rewritten. Since this proposition underpins Proposition 4, the gap should not be left as is.","section":"Proposition 1, proof of reverse implication"}],"minor_comments":[{"comment":"Typo: 'wwhich' should be 'which' in the first paragraph.","section":"Introduction"},{"comment":"The identity matrix in 'Π_k ⊗ I_d' should be I_{L_d} (the L_d-dimensional identity), not the d-dimensional identity, to match the dimensions of the matrices bD and bC.","section":"Corollary 1 / Eq. (24)"},{"comment":"In the displayed decomposed compliance relaxation, the last block appears to write 'K_p(x)' where it should write the corresponding localizing matrix term M_{r-r_K}(Z_p y) (with the correct 2x2 block structure). Please check the typesetting.","section":"Section 5.2"},{"comment":"The proof uses 'f post_AD,r' and 'f*' for values that are elsewhere denoted p post_AD,r and p*. Please unify notation.","section":"Theorem 5 proof"},{"comment":"There is a typo in the expression for K_3: 'K3(x(3)' should be 'K3(x3)'. Also, the null-space basis vectors for K_2 should be checked for consistency with the stated dimensions.","section":"Appendix A.2, Example 6"},{"comment":"The paper explicitly lists 'extend the results of Section 4.3 to cases where the range and null space of the matrices A_k(x) may depend on x' as future work. This is appropriate, but it should be acknowledged earlier in the paper that the current numerical results do not cover that case.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"The core convergence theorems (Theorem 3 and Theorem 5) appear sound and are likely of interest to the community. The main problem is the gulf between the theoretical results and the numerical claims: the variable-reduction step used in Section 6 relies on the constant-null-space assumption, and the appendix proof of that assumption is incorrect for the stated feasible set. This is fixable in principle—for example, by proving that the reduced problem gives the same optimal value for the specific instances tested, or by adding safeguards when zero-area elements appear—but it requires real work. The unproved equality in Section 4.1 is a lesser but still important issue. I would not recommend rejection, because the main theorems are valuable and reproducible code is provided, but the paper should not be accepted without addressing these concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid paper that deserves a serious referee, but it needs revision. The genuinely new thing is the posterior approach: instead of adding D,C to the monomial basis, they decompose the localizing matrix in the mSOS relaxation. Theorem 5, which gives pr <= ppost_AD,r <= p* and convergence, is proven from Putinar's Positivstellensatz and looks correct. The weakening of Assumption 2 from positive definite to PSD in Theorem 3 is also a real extension, with a self-contained proof. Numerical work comes with public code and data, and the reported speedups are impressive when they hold.\n\nSoft spots:\n\n1. Section 4.1 asserts pprior_AD,r = pr without proof. This is not obvious and can fail in general, because the relaxation over (x,D,C) is tighter than the marginal relaxation over x. The paper even says it's not practical, so the claim should either be proven or removed.\n\n2. The bigger issue is Section 4.3/Appendix A.2. The variable reduction relies on A_k(x) having constant null space on K. For the frame problems, K_k(x) = K0 + sum x_e K_e^(i), and the feasible set allows x_e=0. If an element vanishes, the null space can enlarge beyond the rigid-body modes. The appendix tries to justify constancy by arguing from K(x)v=0 at a single x that each monomial coefficient annihilates v. That inference is invalid—linear independence of monomials only helps if the identity holds for all x. So the reduced problem solved numerically is not proven equivalent to (ADmSOS2) when zero-stiffness elements are allowed, and the speedups rest on an unverified assumption. This does not invalidate Theorem 5, but it does undermine the practical central claim.\n\nThe paper is honest about this being an extension, not a paradigm shift, and the citations are appropriate. A referee should ask for: proof or removal of the Section 4.1 equality, and a clear statement of the null-space assumption with either a proof that it holds for the examples or a modification of the procedure that handles rank changes. I'd send it to review.","headline":"Solid extension of arrow decomposition to PMIs with a correct posterior convergence theorem, but the variable-reduction step and an unproven equality need fixing before the speedup claims are reliable.","tokens_in":40644,"tokens_out":3363,"would_cite":true,"duration_ms":36234,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["74P05","90C23","90C22","65F50"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that arrow decomposition—a sparsity-based way to break large semidefinite matrices into smaller ones—works inside moment-SOS hierarchies for polynomial matrix inequalities without losing global convergence, and shows large","keywords":["arrow decomposition","polynomial matrix inequalities","moment-SOS hierarchy","semidefinite programming","topology optimization","frame structures","rank-deficient matrices","structural optimization"],"falsifier":"Take a two-element frame compliance problem at relaxation order r=2 and force one element cross-sectional area to zero at a feasible point. Solve the full localizing-matrix relaxation (mSOS) and the reduced arrow-decomposed relaxation (ADmSOS2 with the Proposition 4 projection). If the reduced problem gives a different lower bound, becomes infeasible, or fails to certify the same global optimum as the full relaxation, the constant-range/null-space assumption has been violated.","tokens_in":39686,"feed_emoji":"🏗️","tokens_out":8281,"duration_ms":89702,"temperature":0.7,"pith_summary":"The paper extends the Arrow Decomposition technique—originally developed for linear matrix inequalities—to polynomial matrix inequalities inside moment-SOS hierarchies. Its central claim is Theorem 5: for an arrow-type PMI satisfying the paper's assumptions, the decomposed posterior hierarchy is at least as tight as the standard relaxation (pr ≤ ppost_AD,r ≤ p*) and still converges to the true optimum as the relaxation order grows. This matters because large polynomial matrix inequalities arise naturally in structural optimization, where stiffness matrices have the required arrow structure, and standard hierarchies become computationally or numerically intractable as the number of elements grows. The paper also weakens the original AD theorem to allow positive semidefinite blocks, so degenerate and rank-deficient cases—such as zero-stiffness elements—can be handled, and it adds a projection post-processing step that shrinks the matrix inequalities and can restore strict feasibility.","feed_headline":"Splitting matrix inequalities cuts optimization time 5,500x","feed_subtitle":"Arrow-shaped blocks stay smaller in moment-SOS relaxations, so global-optimality proofs get cheaper without weakening.","key_machinery":"Arrow decomposition: a way of rewriting an arrow-type matrix G = Σ_k G_k + diag(0, Γ) as a sum of smaller PSD matrices G_k plus correcting interface variables Dk,ℓ and blocks Ck that satisfy consistency equations. The proof rests on the generalized Schur complement (Lemma 1) using the Moore-Penrose inverse, which characterizes positive semidefiniteness even when the blocks are singular. In the posterior hierarchy, the correction matrices enter through the Kronecker factor Πk ⊗ IL, so the extra variables live outside the monomial basis; Proposition 4 then projects each localizing block onto the constant range space of Ak(x), reducing matrix sizes and variable counts.","core_discovery":"The manuscript's main result is that applying arrow decomposition at the level of the moment-SOS relaxation, rather than to the original polynomial problem, preserves convergence while avoiding the explosion of monomial variables. The decomposed localizing-matrix inequalities in (ADmSOS2) use matrices (Πk ⊗ IL)bD and bCk as linearized surrogates for the polynomial auxiliary variables, keeping the monomial basis anchored in the original variables x. Theorem 5 shows the resulting lower bounds lie between the classical mSOS bound and p*, with limr→∞ ppost_AD,r = p*. In the LMI case, Theorem 3 generalizes the original AD theorem by replacing A ≻ 0 and Γ ≻ 0 with A ⪰ 0 and Γ ⪰ 0, and Proposition","pith_inferences":["The paper's variable-reduction shortcut assumes each block Ak(x) keeps the same range and null space for all feasible x; this is a real sensitivity point, because in topology optimization an element with zero cross-section can change the null space of the stiffness matrix, and an adaptive or differently-partitioned decomposition may be needed there.","The physical interpretation of the auxiliary variables as interface forces suggests a natural connection to domain-decomposition solvers: one could use a coarse equilibrium solve to initialize or precondition the SDP, an idea the paper notes but does not implement.","The same linearization trick that derives ADmSOS2 from ADmSOS1 could be applied to other structured sparsity patterns, such as chordal or correlative sparsity, wherever localizing matrices are already block-lifted by Kronecker factors.","The reported speedups pair arrow decomposition with the non-mixed-term basis; the paper's own data show AD alone can be slower at high relaxation orders when the added variables outnumber the savings, so choosing when to combine the two is itself a practical question."],"forward_implications":["For every relaxation order r, the arrow-decomposed bound is between the standard mSOS bound and the true optimum, and asymptotically it reaches p*.","One can replace a single large localizing-matrix inequality by smaller decomposed ones without enlarging the monomial basis, enabling higher relaxation orders on finer structural discretizations.","Degenerate or rank-deficient blocks no longer need to be avoided: projection onto range spaces shrinks the inequalities and can restore the interior-point condition (strict feasibility).","Combined with a non-mixed-term monomial basis, the method reports large practical speedups on frame topology optimization, including a 5,500× reduction in solution time on one fourth-order relaxation example.","The same framework covers both compliance minimization and weight minimization of frame structures, with the auxiliary variables interpreted physically as interface forces between subdomains."],"supporting_citations":[{"why":"Introduces the arrow decomposition method for LMIs and the domain-decomposition link; the paper extends its theorem and weakens its assumptions.","marker":"[1]"},{"why":"Putinar's Positivstellensatz underlies the moment-SOS representation of positive polynomials used throughout the hierarchy.","marker":"[2]"},{"why":"Lasserre's moment-SOS hierarchy for global polynomial optimization supplies the base relaxations that the paper decomposes.","marker":"[3]"},{"why":"Convergent relaxations for polynomial matrix inequalities, the PMI counterpart on which the paper builds.","marker":"[16]"},{"why":"Provides the compliance-minimization PMI formulation for frame structures and the upper-bound/gap strategy used in the numerical examples.","marker":"[14]"},{"why":"Provides the weight-minimization frame formulation and the 24-element modular frame benchmark used in the experiments.","marker":"[27]"},{"why":"Supplies the generalized Schur complement via the Moore-Penrose inverse, the key tool in Theorem 3 and Lemma 1.","marker":"[22]"},{"why":"Introduces the non-mixed-term (NMT) basis whose combination with arrow decomposition produces the largest reported speedups.","marker":"[19]"}],"fun_headline_variants":["Arrow decomposition extends to polynomial matrix inequalities","Faster global optimization via arrow-shaped matrix splits","Moment-SOS hierarchies get a boost from arrow decomposition","Shrinking matrix blocks speeds up structural optimization proofs","New method cuts variables in polynomial optimization relaxations"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The variable-reduction shortcut assumes that each sub-block Ak(x) loses rank in exactly the same directions for every feasible x; in topology optimization, elements with zero cross-sectional area change the directions in which the stiffness matrix is degenerate, so the shortcut—and with it the reported speedups—may break down on the very designs it targets.","fun_headline_variants_meta":{"raw":{"variants":["Arrow decomposition extends to polynomial matrix inequalities","Faster global optimization via arrow-shaped matrix splits","Moment-SOS hierarchies get a boost from arrow decomposition","Shrinking matrix blocks speeds up structural optimization proofs","New method cuts variables in polynomial optimization relaxations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1338,"prompt_tokens":823,"completion_tokens":515,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":443}},"tokens_in":567,"tokens_out":515,"duration_ms":5874,"temperature":1.0,"reasoning_tokens":443,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:21:35.000154+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-element frame compliance problem at relaxation order r=2 and force one element cross-sectional area to zero at a feasible point. Solve the full localizing-matrix relaxation (mSOS) and the reduced arrow-decomposed relaxation (ADmSOS2 with the Proposition 4 projection). If the reduced problem gives a different lower bound, becomes infeasible, or fails to certify the same global optimum as the full relaxation, the constant-range/null-space assumption has been violated.","supporting_citations":[{"cited_title":"Mathematical Programming190(1–2), 105– 134 (2020) https://doi.org/10.1007/s10107-020-01526-w","cited_arxiv_id":null,"evidence_quote":"Introduces the arrow decomposition method for LMIs and the domain-decomposition link; the paper extends its theorem and weakens its assumptions."},{"cited_title":"IEEE Transactions on Automatic Control 51(2), 192–202 (2006) https://doi.org/10.1109/TAC.2005.863494","cited_arxiv_id":null,"evidence_quote":"Convergent relaxations for polynomial matrix inequalities, the PMI counterpart on which the paper builds."},{"cited_title":"Structural and Multidisciplinary Optimization64(4), 1963–1981 (2021) https://doi.org/10.1007/s00158-021-02957-5","cited_arxiv_id":null,"evidence_quote":"Provides the compliance-minimization PMI formulation for frame structures and the upper-bound/gap strategy used in the numerical examples."},{"cited_title":"Structural and Multidisciplinary Opti- mization 66(12), 257 (2023) https://doi.org/10.1007/s00158-023-03715-5","cited_arxiv_id":null,"evidence_quote":"Provides the weight-minimization frame formulation and the 24-element modular frame benchmark used in the experiments."},{"cited_title":"Texts in Applied Mathematics","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized Schur complement via the Moore-Penrose inverse, the key tool in Theorem 3 and Lemma 1."},{"cited_title":"Optimization and Engineering, 1–44 (2025) https: //doi.org/10.1007/s11081-025-10000-5","cited_arxiv_id":null,"evidence_quote":"Introduces the non-mixed-term (NMT) basis whose combination with arrow decomposition produces the largest reported speedups."}],"review_version":1}