{"id":"a3215d6f-c5b0-4b65-b2dd-9027ae30add1","arxiv_id":"2412.07634","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An enhanced inertial projected gradient descent algorithm with Schur-complement projection and constraint-aware step splitting matches or exceeds MMA on a 3D heat-sink topology optimization problem with an overhang constraint.","lead":"This paper upgrades the projected gradient descent optimizer so it can handle difficult nonlinear constraints in topology optimization. On a 3D heat-sink design problem, the upgraded method matches or beats the standard MMA optimizer while needing fewer user settings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's convergence guarantee does not cover the implemented algorithm: Eq. (42) is used outside the feasible regime, and the Section 2.3 bounds require α_n = γ_n, which the authors state does not hold in practice.","rationale":"The reader's weakest assumption identifies exactly the mismatch between the convergence proof and the implemented algorithm: the Section 2.3 bounds require α_n = γ_n and feasible iterates, while the implementation uses γ_n ≠ α_n and the infeasible branch of Eq. (42). The manuscript itself acknowledges this after Eq. (54), which makes the concern concrete rather than speculative. This is the most load-bearing issue because the paper's central formal claim — that the modified PGD converges — is only demonstrated for a restricted regime, while the empirical demonstration of robustness relies on the unrestricted algorithm. That said, the concern does not by itself invalidate the empirical comparison with MMA on the heat-sink benchmark; it weakens the theoretical support rather than producing observed failure. The paper also supplies a Python implementation of the algorithmic core, which is a partial independent support, but it is not used for the topology optimization results. The reader's CONDITIONAL verdict is appropriate: the proof gap and proprietary main code are addressable, and the empirical claim is plausible. No adjustment to the reader's verdict is needed.","tokens_in":25002,"tokens_out":5929,"duration_ms":59442,"concrete_test":"Re-run the overhang heat-sink case with two variants: (A) a proof-covered variant with γ_n = α_n and no infeasible branch, re-projecting any infeasible iterate via the full Section 2.1 active-set projection before updating, and (B) the default algorithm of Eq. (42). Record final cost, final constraint values, and the number of iterations with φ_n outside Q. If B matches A within 2% on final cost and constraint violation, the gap is a documentation issue; if B outperforms A by more than that, the reported gains rely on an unproved branch. Additionally, check Inequality (44) on logged default-parameter iterates to measure how often the sufficient descent condition actually fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing concern is that the formal convergence guarantee does not cover the algorithm as implemented. Theorem 1 establishes convergence of the active-set projection subproblem in Section 2.1, not of the outer inertial PGD iteration. The outer-iteration analysis in Section 2.3 explicitly assumes the iterates remain feasible (the sentence preceding Eq. (43)) and derives sufficient conditions for the feasible branch with min(1, γ_n/α_n) applied to both components of the update. Immediately after Eq. (54), the authors write that these bounds 'only guarantee convergence when using α_n = γ_n' and that 'in practice, γ_n ≠ α_n in most cases.' The implemented update, Eq. (42), also contains an infeasible branch in which γ_n is not applied to the parallel component and the orthogonal component is damped by μ^h. No descent inequality is proved for this branch, yet Figures 6–8 show constraint breakage precisely in the loops where the reported speedups occur. The paper itself also flags the related risk of indefinite step-size reduction in Section 2.1 when too many constraints are broken. Thus the central claim that the enhanced PGD is a robust, provably convergent alternative is supported only in a parameter regime the implementation does not follow. The empirical comparison may still be valid, but the formal claim is overstated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two enhancements to an inertial projected gradient descent (PGD) method for large-scale constrained optimization, targeting topology optimization. First, univariate bound constraints are incorporated directly into the projection step through a Schur complement within an active-set method that manipulates constraints in bulk. Second, the update step is decomposed into components parallel and orthogonal to the constraint gradients, with the orthogonal component scaled by a factor derived from an approximation of the Lagrangian and by a relaxation factor activated when constraints are broken. The authors prove convergence of the active-set projection subproblem (Theorem 1), give sufficient descent conditions for the modified outer iteration, and report numerical results on a 3D heat-sink topology optimization problem with a volume constraint and a nonlinear overhang constraint, comparing against MMA and GCMMA. The paper includes a Python reference implementation in the supplementary material and a parameter-sensitivity study.","tokens_in":25326,"tokens_out":3602,"duration_ms":35335,"significance":"If the claims are substantiated, the work is a useful contribution to first-order methods for topology optimization: it offers a practical way to handle many univariate constraints inside the projection and improves robustness for nonlinear constraints while retaining the scalability of PGD. The paper's strengths include detailed algorithmic presentation, a supplementary Python implementation, a careful ablation study (traditional, intermediary, and proposed variants), and an explicit discussion of parameter sensitivity and limitations. The main weakness is that the formal convergence analysis does not cover the algorithm as actually implemented: the descent conditions require feasible iterates and α_n = γ_n, while the implemented update admits an infeasible branch and the paper states that γ_n ≠ α_n in most cases. The numerical evidence is also limited to one proprietary-solver test case, so the 'comparable to or exceeding MMA' claim is suggestive rather than definitive.","major_comments":[{"comment":"The convergence analysis does not cover the implemented algorithm. The sufficient descent conditions in Eqs. (49)–(51) are derived under the assumptions that the iterates remain feasible and that min(1, γ_n/α_n) is applied to both components of the update. Immediately after Eq. (54) the paper states that these bounds 'only guarantee convergence when using α_n = γ_n' and that 'in practice, γ_n ≠ α_n in most cases.' The implemented update, Eq. (42), also contains an infeasible branch in which the parallel component is not scaled by γ_n and the orthogonal component is damped by μ^h. No descent inequality is proved for this branch, yet Figures 6 and 8 show constraint breakage precisely in the loops where the reported speedups occur. Therefore Theorem 1 establishes convergence of the active-set projection subproblem, not of the outer inertial PGD iteration, and the abstract's claim that the algorithm 'converges to a stable active set that satisfies the KKT criteria' is overstated for the full method. This is a load-bearing gap that should be fixed either by extending the analysis to the implemented branch or by explicitly presenting the method as heuristic with empirical support.","section":"Section 2.3, Eqs. (42)–(54)"},{"comment":"The practical step-size rules are not shown to satisfy the sufficient descent conditions. The paper defines α_n = 1/L using the local estimate (55), γ_n by (57), and β_n by (58) with a user-chosen β_hat, and then states that 'in practice, the choice of Lipschitz constant does not guarantee convergence, as it is only a local approximation that can carry a non-negligible error.' Since the descent conditions (49)–(51) require an accurate global or sufficiently large local Lipschitz constant, and the implementation uses a heuristic estimate, the formal guarantee does not apply to the parameter settings used in the numerical experiments. The authors should either add a safeguard (for example, a backtracking or restart rule that provably restores descent when the cost increases or constraints are violated) or clearly label the step-size selection as heuristic and separate it from the formal convergence statements.","section":"Section 2.3, Eqs. (55)–(58)"},{"comment":"The proof of Theorem 1 has a gap in the fallback mechanism. Step 6.d.ii reintroduces the 'most binding' broken constraint and the proof asserts that repeated single-constraint manipulation converges by citing classical active-set literature, but no direct argument is given that the fallback loop terminates without cycling, particularly after multiple constraints have been removed from the stored copy of J_{o-1}. The lemmas rely on convexity and self-compatibility, but the proof does not show that the algorithm cannot cycle between the reintroduction and removal steps. Since Theorem 1 is a central formal claim, the proof should be made rigorous for this fallback path, or the theorem should be restricted to the cases where the fallback is not needed.","section":"Section 2.1.2, Theorem 1 proof"}],"minor_comments":[{"comment":"The active-set description is difficult to follow because the internal steps 6, 6a–6d and the flowchart in Figure 1 use inconsistent notation for the 'cost function': in Lemmas 1 and 2 and in the proof of Theorem 1, the cost is the distance ||Δφ_tilde - Δφ||, but step 6 is described as checking whether the 'cost function increased' while Eq. (25) is written as a distance inequality. Please clarify that the merit function is the projection distance, not the original objective C.","section":"Section 2.1.2"},{"comment":"The numerical comparison is based on a single heat-sink test case with a proprietary finite element solver (DFEM). The claim that the method performs 'comparable to or exceeding MMA' would be strengthened by testing on at least one additional problem class or by reporting results with an open-source solver, especially since the MMA move limit was tuned for this problem (values 0.1 and 0.05 are used in the two cases).","section":"Section 3.3"},{"comment":"Default parameter values (β_hat = 0.2, μ = 0.95, ζ = 0.5, ε_i = 0.02 a_i) appear to be selected using the same test cases that are later used to demonstrate robustness. This is not circular in a mathematical sense, but the paper should acknowledge that the parameter-sensitivity study does not fully validate the 'minimal parameter tuning' claim on independent problems.","section":"Section 3.4"},{"comment":"Tables A1–A4 report total iteration counts and counts of fallback events, but they do not directly report the 'frequency' of step 6 or step 6c failures as claimed in the text. Please state the fallback rates as fractions of total iterations, which would make the empirical claim about the rarity of the fallback easier to assess.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the mismatch between the formal convergence claims and the algorithm as implemented. I do not see this as an irreparable flaw: the numerical results and the Python implementation provide valuable empirical support, and the paper already contains candid statements about the gaps (e.g., after Eq. (54) and in the discussion of the Lipschitz estimate). A revision that either extends the analysis to the implemented branch or explicitly reframes the convergence claims as applying to the idealised version, while presenting the implemented version as heuristic, would be publishable in principle. The empirical scope (one test case, proprietary solver) should also be qualified more carefully in the abstract and conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is a real engineering contribution: the Schur-complement handling of univariate constraints inside an infeasible active-set projection, the bulk constraint manipulation with a backtracking fallback, and the post-projection step splitting are all sensible, and the authors back the core algorithm with a Python implementation. Second, the formal convergence guarantee does not cover the algorithm as actually implemented. The paper admits that the Section 2.3 bounds only hold when α_n = γ_n, and that in practice γ_n ≠ α_n; the infeasible branch of Eq. (42) has no analysis. Theorem 1 proves convergence of the active-set subproblem, not of the outer inertial PGD iteration. So the claim that the method is provably convergent is overstated, even though the empirical comparison may still be valid.\n\nWhat the paper does well: the algorithmic combination is genuinely new, the parameter sensitivity study is honest and informative, and the comparison against MMA on a heat-sink problem shows competitive performance with minimal tuning. The Python supplement is a plus, even though it only covers a generic convex problem, not the actual topology-optimization test cases.\n\nThe soft spots are proportionate. The missing ablation is real: the \"traditional PGD\" baseline still uses the new projection machinery, so you cannot isolate the benefit of the Schur complement or bulk active-set manipulation. The evidence is one heat-sink case at about 26k degrees of freedom, which is modest for topology optimization. The convergence gap is the most serious issue, but it is addressable: either adapt the proof to the actual update rules or weaken the claimed guarantee in the abstract and conclusion. The paper's own acknowledgment after Eq. (54) suggests the authors know the limitation, which counts in their favor.\n\nThis paper is for practitioners and method developers in topology optimization who want a first-order method that handles nonlinear constraints with little tuning. It deserves a serious referee: the ideas are worth engaging with, and the issues I listed can be fixed. I would send it to peer review, but I would ask for a revised version that clearly separates proved properties from empirical observations and either ablate the projection contribution or explicitly state that the current experiments do not isolate it.\n\nMy recommendation: engage with it, but do not take the convergence theorem at face value.","headline":"A useful engineering contribution to PGD for topology optimization, but the formal convergence claim is narrower than the abstract suggests.","tokens_in":25784,"tokens_out":2024,"would_cite":false,"duration_ms":20223,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C30","90C06","65K05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A projected gradient descent method with Schur-complement bound handling and a constraint-aware update split matches or beats MMA on 3D heat-sink topology problems with minimal tuning, and its active-set projection provably reaches a KKT…","keywords":["inertial projected gradient descent","infeasible active set","bulk constraints manipulation","Schur complement decomposition","step decomposition","nonlinear constraints","topology optimization","heat sink design"],"falsifier":"Solve a small nonconvex test problem in which the local Lipschitz estimate of Equation (55) is systematically too small, for instance a quadratic cost whose curvature jumps sharply when a continuation parameter changes, and check whether the implemented algorithm, which uses $\\gamma_n \\neq \\alpha_n$ and the infeasible branch of Equation (42), ever produces an iteration with increased cost or a constraint violation that persists instead of decaying; the paper reports that no such backtracking was needed in its heat-sink tests, so a single instance that requires it would show the robustness claim is case-dependent.","tokens_in":24820,"feed_emoji":"🔥","tokens_out":10918,"duration_ms":87471,"temperature":0.7,"pith_summary":"This paper is trying to establish that projected gradient descent, a simple first-order method that scales well to large design spaces, can be made robust enough for topology optimization problems with nonlinear constraints, a regime where the method is often considered fragile. Two mechanisms are added: bound constraints are folded directly into the projection step through a Schur complement inside an active-set solver that adds and removes constraints in bulk, and the update step is split into a component lying in the constraint directions and a component orthogonal to them, so the orthogonal part can be rescaled using an approximation of the Lagrangian and damped when constraints are broken. Tested on a three-dimensional heat-sink design with a volume constraint and a nonlinear overhang constraint, the enhanced PGD reaches final cost values comparable to or slightly better than the Method of Moving Asymptotes while requiring almost no parameter tuning. The interest, if the claim holds, is that a robust first-order optimizer could replace tuning-heavy workhorse methods like MMA on very large problems where second-order methods become impractical.","feed_headline":"Gradient descent upgrade keeps pace with MMA on heat sinks","feed_subtitle":"Two changes — Schur-complement projections and step splitting — tame nonlinear constraints with minimal tuning.","key_machinery":"The machinery has three pieces. The first is a Schur complement decomposition of the projection system, separating a small block of active global constraints from a diagonal block of univariate constraints, so the projection requires solving only an $m^* \\times m^*$ linear system rather than a system involving every bound. The second is an infeasible active-set algorithm that adds all violated constraints in bulk, removes constraints with negative Lagrange multipliers in bulk, and reverts to one-at-a-time constraint manipulation when the cost-monotonicity check fails; Theorem 1 proves it converges to a stable active set satisfying the KKT conditions of the projection subproblem. The third is a post-projection split of the update vector, computed by a Gram-Schmidt process, into a component $\\Delta\\boldsymbol{\\phi}^\\parallel$ within the span of the constraint gradients and a component $\\Delta\\boldsymbol{\\phi}^\\perp$ in the null space of those gradients; only the null-space component is scaled by $\\gamma_n/\\alpha_n$ (with $\\gamma_n$ built from the variation of an approximate Lagrangian gradient) and damped by $\\mu^h$ when constraints have been broken for $h$ consecutive steps, which keeps the cost function moving while actively correcting constraint violations.","core_discovery":"In the authors' terms, the central claim is that two modifications to the inertial projected gradient descent method make it reliable for nonlinear constrained topology optimization: first, univariate constraints such as design-variable bounds are incorporated directly into the projection step using a Schur complement that separates the diagonal block of univariate constraints from the block of global constraints, inside an active-set algorithm that manipulates constraints in bulk and only falls back to single-constraint handling when a monotonicity check fails; second, the projected update is decomposed, relative to the constraint gradients, into a component in the constraint variation space and a component in its orthogonal complement, and only the orthogonal component is scaled by a factor derived from the variation of an approximate Lagrangian, with an additional relaxation factor applied when constraints have been broken for several consecutive steps. Theorem 1 states that the active-set iteration converges to a stable active set satisfying the KKT conditions of the linearized projection subproblem. Empirically, on a 3D heat-conduction topology optimization problem, with and without the overhang constraint, the default-parameter PGD converges to final cost values comparable to or lower than those of the Method of Moving Asymptotes, converges faster after the nonlinear constraint is broken, and shows only mild sensitivity to its tunable parameters.","pith_inferences":["The infeasible branch of the update acts like an implicit penalty-recovery mechanism, and a natural extension is to test whether the same robustness carries over to other nonlinear constraint families, such as stress or buckling constraints in structural design.","Replacing the fixed relaxation factor $\\mu$ with one that scales with the actual magnitude of constraint violation, which the authors list as future work, would likely remove the remaining sensitivity to constraint jumps when the continuation parameters change between optimization loops.","The Schur-complement active-set projection is a general device for quadratic-programming projections with many separable bound constraints and few global constraints, so other first-order methods could borrow it independently of the PGD update scheme."],"forward_implications":["With default settings ($\\hat{\\beta}=0.2$, $\\mu=0.95$) the proposed PGD reaches final cost values comparable to or lower than MMA on both tested heat-sink cases, while MMA required its asymptote move limit to be tuned for the problem.","The overhang constraint, which is strongly nonlinear, is recovered faster after being broken when the Lagrangian-based $\\gamma$ correction is active, shortening the intervals during which the constraint is violated inside continuation loops.","Because the Schur complement reduces the projection to a small $m^* \\times m^*$ linear solve, the cost per iteration stays near-constant as the number of design variables grows, which is what makes the method practical for large design spaces.","Theorem 1 guarantees that every projection subproblem terminates at a stable active set satisfying the KKT conditions of the linearized problem, so the inner iteration is well-defined before the outer step is taken."],"supporting_citations":[{"why":"Svanberg's method of moving asymptotes is the baseline the proposed PGD is compared against in both heat-sink test cases.","marker":"[5]"},{"why":"Nishioka and Kanno's inertial projected gradient method is the starting point whose efficiency this work extends.","marker":"[11]"},{"why":"Bartlett and Biegler's QPSchur supplies the Schur-complement strategy for structured convex quadratic programming used in the projection step.","marker":"[27]"},{"why":"Wong's active-set theory for convex quadratic programming supports the fallback to single-constraint manipulation in the convergence argument.","marker":"[28]"},{"why":"Schittkowski's bulk constraint manipulation with activity duration motivates the bulk add-and-remove strategy of the active set.","marker":"[19]"},{"why":"Lamarche-Gagnon and colleagues supply the overhang constraint formulation and the adjoint machinery inside DFEMwork used for the numerical tests.","marker":"[23]"},{"why":"Qian's density-gradient integral formulation is the origin of the nonlinear overhang constraint exercised in the experiments.","marker":"[22]"}],"fun_headline_variants":["PGD with Schur-complement projection rivals MMA on heat sinks","Two modifications make PGD robust for nonlinear topology constraints","Projected gradient descent gains active-set bulk constraints handling","PGD steps split by constraint space to match MMA performance","Inertial PGD with Schur complement tames nonlinear constraints"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The algorithm's convergence is proven only for the case where its two step-size parameters are equal and every iterate stays inside the feasible region; the version actually run uses unequal parameters and deliberately steps outside feasibility, so the reliability of the results rests on empirical behavior rather than on the proof.","fun_headline_variants_meta":{"raw":{"variants":["PGD with Schur-complement projection rivals MMA on heat sinks","Two modifications make PGD robust for nonlinear topology constraints","Projected gradient descent gains active-set bulk constraints handling","PGD steps split by constraint space to match MMA performance","Inertial PGD with Schur complement tames nonlinear constraints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000704,"raw_usage":{"total_tokens":3210,"prompt_tokens":1013,"completion_tokens":2197,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":2114}},"tokens_in":629,"tokens_out":2197,"duration_ms":13951,"temperature":1.0,"reasoning_tokens":2114,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:38:31.879246+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve a small nonconvex test problem in which the local Lipschitz estimate of Equation (55) is systematically too small, for instance a quadratic cost whose curvature jumps sharply when a continuation parameter changes, and check whether the implemented algorithm, which uses $\\gamma_n \\neq \\alpha_n$ and the infeasible branch of Equation (42), ever produces an iteration with increased cost or a constraint violation that persists instead of decaying; the paper reports that no such backtracking was needed in its heat-sink tests, so a single instance that requires it would show the robustness claim is case-dependent.","supporting_citations":[{"cited_title":"Inertial projected gradient method for large-scale topology optimization.Japan Journal of Industrial and Applied Mathematics","cited_arxiv_id":null,"evidence_quote":"Nishioka and Kanno's inertial projected gradient method is the starting point whose efficiency this work extends."},{"cited_title":"QPSchur: a dual, active-set, Schur-complement method for large-scale and structured convex quadratic programming","cited_arxiv_id":null,"evidence_quote":"Bartlett and Biegler's QPSchur supplies the Schur-complement strategy for structured convex quadratic programming used in the projection step."},{"cited_title":"Active-set methods for quadratic programming","cited_arxiv_id":null,"evidence_quote":"Wong's active-set theory for convex quadratic programming supports the fallback to single-constraint manipulation in the convergence argument."},{"cited_title":"An active set strategy for solving optimization problems with up to 200,000,000 nonlinear constraints","cited_arxiv_id":null,"evidence_quote":"Schittkowski's bulk constraint manipulation with activity duration motivates the bulk add-and-remove strategy of the active set."},{"cited_title":"Additively manufactured conformal cooling channels through topology optimization","cited_arxiv_id":null,"evidence_quote":"Lamarche-Gagnon and colleagues supply the overhang constraint formulation and the adjoint machinery inside DFEMwork used for the numerical tests."},{"cited_title":"Undercut and overhang angle control in topology optimization: A density gradient based integral approach","cited_arxiv_id":null,"evidence_quote":"Qian's density-gradient integral formulation is the origin of the nonlinear overhang constraint exercised in the experiments."}],"review_version":1}