{"id":"88db49ce-f7ba-4699-8a60-18eed3ab027e","arxiv_id":"2607.19599","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Tseng-style forward-backward-forward splitting solves the exact reduced Coulomb law by alternating a strongly convex cone QP with an explicit De Saxcé-Feng coupling correction.","lead":"A new solver for rigid-body physics splits exact Coulomb friction into a convex cone problem plus a small non-convex coupling term, iterating until the unmodified friction law is satisfied. It keeps stacks, arches, and stick-slip objects stable without smoothing or relaxing Coulomb's law.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3's outer-budget sweep contradicts the claimed empirical convergence: raising the cap from 200 to 2000 worsens the house-of-cards residual from 6.3e-7 to 1.5e-5, so the reported evidence does not support the claim that the iteration converges to exact Coulomb.","rationale":"The strongest claim is that the converged FBF iterate satisfies the exact reduced Coulomb law. This requires both that fixed points of the update are zeros of (7) and that the iteration actually reaches such a point. The paper has no proof of the second requirement for the non-monotone coupling B, and instead substitutes residual-based empirical convergence. My concern targets that evidence directly: the same scene, same residual metric, same solver, with only the outer cap enlarged, goes from 93% substeps at rc≤1e-6 to 3%, with a median residual that worsens. If the stopping rule stops at k=5 because rc<ε, but continuing to k=40 yields rc=1.5e-5, then the residual is not a monotone or even eventually-decreasing measure of distance to the exact law. The paper's own Table 3 exposes this. This is more actionable than the abstract lack of a proof: it is an internal contradiction that can be resolved by a focused experiment. I agree with the reader's weakest assumption and recommend the verdict stay conditional: the architecture is plausible and the physical benchmarks are encouraging, but the central claim is not yet supported by the reported convergence data. The concrete test above would determine whether the anomaly is a reporting artifact or a genuine failure of convergence; if real, the 'exact law at convergence' narrative must be weakened or the algorithm modified.","tokens_in":19958,"tokens_out":8358,"duration_ms":73481,"concrete_test":"Re-run the house-of-cards benchmark with the Table 3 settings and, for a fixed set of substeps, disable early stopping and record rc(k) for k=1..2000. Isolate the confound by comparing cap=200 vs cap=2000 with warm-start disabled and γ fixed at the std run's final value. If rc crosses below 1e-6 near k=5 and then rises above 1e-5 by k=40, the empirical convergence claim is falsified; if rc remains ≤1e-6 throughout, Table 3 is a reporting artifact and the central claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Algorithm 1's outer iteration converges to the exact reduced Coulomb law, with rc in §4.5 as faithful certificate. §4.4 and §6 concede there is no convergence proof for the non-monotone B; convergence is asserted 'empirically'. The empirical record, however, contains a direct contradiction. Table 3 / Appendix A: house-of-cards std (200-iteration cap) has median rc=6.3e-7 and 93% substeps below 1e-6; cards10x (2000 cap) has median rc=1.5e-5, only 3% below 1e-6, and median iterations rise from 5 to 40. Running longer makes the monitored residual an order of magnitude worse on the same scene. This is not 'the extra budget does nothing' (App. A/§6); it says the iteration is not converging in rc and that the early stop at k=5 was a transient crossing, not a fixed point. Consequently, line 8 of Algorithm 1 is not a certificate: if the residual can fall below ε at k=5 and rise above it by k=40, the reported near-floor residuals do not establish that the converged iterate satisfies exact complementarity. The arch row (47%→61%) shows the expected improvement, making the card row unambiguously anomalous. Unless this is a typo or an artifact of warm-starting/γ-restoration between substeps, the empirical support for the central claim fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a splitting solver for the exact reduced Coulomb friction law. It rewrites the reduced inclusion (7) as the sum of a cone-constrained linear response A(λ) and a scalar non-associated De Saxcé–Feng coupling B(λ), then applies a Tseng-style forward–backward–forward (FBF) outer iteration: evaluate B at the current iterate, solve a strongly convex cone QP (12), apply the explicit correction (14), and project onto the friction cone (15). The inner cone QP is modular and can be solved by different SOCP solvers. The authors implement the method on NVIDIA Warp/Newton and evaluate it on cube-on-incline, backspin ball, turntable, Painlevé box, house-of-cards, and masonry-arch benchmarks, reporting that the dimensionless Coulomb residual r_c of §4.5 approaches the single-precision floor and that the method reproduces exact complementarity without smoothing or relaxation.","tokens_in":20427,"tokens_out":5102,"duration_ms":48675,"significance":"If the convergence claim held, the paper would make a useful contribution: it cleanly isolates the non-monotone De Saxcé coupling from the cone QP, gives a modular inner-solve architecture, provides a matrix-free implementation, and targets the exact reduced law rather than a relaxed or smoothed surrogate. The algebraic derivation from (7) to the cone QP (12) and the correction (14)–(15) is sound, and the comparison with MuJoCo and Kamino on the frictional stacking benchmarks is well motivated. However, the central claim that the outer iteration converges to the exact reduced law is currently supported only by empirical evidence, and the empirical record in Appendix A/Table 3 contains a direct contradiction. The paper also honestly acknowledges that Tseng’s theorem does not apply because B is non-monotone (§4.4, §6), so the missing formal convergence guarantee is not an omission but a stated limitation.","major_comments":[{"comment":"The empirical convergence claim is contradicted by the outer-budget sweep. For the house of cards, raising the outer-iteration cap from 200 to 2000 changes the median r_c from 6.3e-7 to 1.5e-5, the share of substeps with r_c ≤ 1e-6 from 93% to 3%, and the median outer-iteration count from 5 to 40. This is not 'the extra budget does nothing' as §6 states; running longer makes the monitored residual an order of magnitude worse on the same scene. Since §4.4 and §6 establish convergence only empirically, this anomaly directly undermines the central claim that the converged iterate satisfies the exact reduced law.","section":"Appendix A / Table 3"},{"comment":"The paper explicitly concedes that Tseng’s convergence theorem cannot be applied because B(λ) is not monotone, and that 'convergence here is therefore established empirically' (§6). This is a load-bearing gap because the title and abstract claim convergence to the exact law. The contradiction in Table 3 makes the empirical support unreliable. A formal convergence result for a restricted class (e.g., when B is contractive relative to W+γ^{-1}I), or a clearly stated local/global condition under which the FBF iteration converges, would be needed to support the exact-law claim; otherwise the claims should be softened to a residual-based stopping heuristic.","section":"§4.4 / §6"},{"comment":"The stopping criterion r_c < ε is the same quantity used to report success, and Algorithm 1 stops as soon as this condition is met. Reporting that 93% of substeps reach r_c ≤ 1e-6 on the house of cards is therefore partly tautological: the solver stops precisely when that condition holds, and the arch’s 47% share simply reflects the 200-iteration cap. The more informative evidence would be a demonstration that the iteration actually settles at a fixed point, e.g., monitoring r_c over continued iterations after first crossing ε, or comparing against a known exact solution on a small problem. The Table 3 cards10x row shows that the residual can rise after an early low crossing, so a single low r_c at the stopping time is not a certificate of convergence.","section":"Algorithm 1, line 8 / §4.5"},{"comment":"The safeguarded adaptive rule introduces thresholds θ=0.9, β=0.7, and cγ=0.5 without a sensitivity study. Appendix A sweeps fixed γ but does not vary θ, β, or the acceptance policy. Since the step-size rule is a central algorithmic contribution and is claimed to be 'tuned to the stick-slip boundary', the robustness of the method to these choices should be documented, especially because the convergence claim is empirical.","section":"§4.4, Eq. (18)"}],"minor_comments":[{"comment":"The text says 'on the house of cards, which already converges in a few iterations, the extra budget does nothing', but Table 3 reports cards10x with median 40 iterations and a worse residual. This inconsistency should be corrected regardless of the explanation.","section":"§6 / Table 3"},{"comment":"The scale choices s_r and s_u are not fully described: the floor ϵ_0=1e-12 is introduced without motivation, and the use of ∥λ_0∥ in s_r makes the residual scale depend on the initial guess. Clarifying the initialization-dependence would make the reported r_c values more reproducible.","section":"§4.5, Eq. (21)"},{"comment":"The notation 'MuJoCo' and 'Kamino' columns as relative factors is clear, but the kamino factor 0.38× on the arch should be explicitly labeled as 'faster than FBF' to avoid ambiguity in the table.","section":"Table 5 / Appendix B"},{"comment":"The GPU timing table is useful, but the statement that FBF is slower on GPU than on CPU, due to an unoptimized port, should be flagged more prominently in the main text; currently it appears only in the appendix and could be misread as a performance claim against the architecture.","section":"§4.6 / Table 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and the algebraic core is sound, but the central empirical convergence claim has a direct contradiction in Table 3, and the formal convergence gap is explicitly acknowledged. The issues are fixable: either provide additional experiments that resolve the cards10x anomaly and demonstrate residual non-increase or fixed-point behavior, or substantially soften the convergence claims and frame the method as a residual-based stopping heuristic. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: this is a clean, honest solver paper with a real architectural idea, but the empirical convergence story is undercut by the authors' own Table 3. I'd send it to reviewers, but the authors need to explain that table before I'd trust the central claim.\n\nThe paper adopts the Acary–Cadoux–Lemaréchal–Malick decomposition of the reduced Coulomb law into a cone QP plus a scalar non-associated De Saxcé coupling, and replaces their Picard iteration with a Tseng forward-backward-forward scheme plus a safeguarded adaptive step size. The derivation from the inclusion to the cone QP and correction is algebraically clean, and the rc residual in §4.5 is a sensible, dimensionless convergence surrogate. The modularity of the inner solve is a genuine practical advantage, and the benchmark comparisons (backspin ball, turntable, house of cards, arch) show the expected qualitative failures of the convex-relaxed MuJoCo baseline. The paper doesn't fit anything: the exact law and decomposition come from prior work.\n\nThe soft spots are real and concentrated in convergence. Section 4.4 explicitly says Tseng's theorem does not apply because B is non-monotone; §6 says convergence is empirical. That's an honest limitation, but the empirical record is self-contradictory. Table 3 / Appendix A reports that on the house of cards, raising the outer-iteration cap from 200 to 2000 makes the median rc worse by more than an order of magnitude (6.3e-7 → 1.5e-5), drops the fraction of substeps below 1e-6 from 93% to 3%, and raises the median iteration count from 5 to 40. The text claims the extra budget \"does nothing\" on the already-converged scene; the table shows it actively hurts. That is not a harmless finding — it means the residual is not behaving like a fixed-point certificate, and the near-floor residual at k=5 may be a transient crossing, not a converged solution. The arch row shows the expected improvement, so the card row looks like a genuine anomaly rather than a general pattern. Maybe it's a warm-start artifact or a step-size restoration issue between substeps, but the paper doesn't say. Until this is explained, the central claim that the converged iterate satisfies exact Coulomb is not supported by the data. The code is also only promised, not shipped, so the reader can't easily check.\n\nThe float32 floor analysis and the step-size sweeps are careful, and the comparison against Kamino and MuJoCo is useful even if the timings are acknowledged as unoptimized.\n\nWho is this for: people building rigid-body contact solvers for graphics or robotics, especially anyone thinking about splitting nonsmooth laws rather than smoothing or relaxing them. I'd bring it to a reading group to argue about Table 3. My recommendation: send it to peer review — the architecture is worth refereeing, and the anomaly is addressable. But treat the convergence claim as unproven until the budget sweep is explained and the code is out.","headline":"A clean splitting architecture for the exact reduced Coulomb law, but the empirical convergence claim is undercut by the paper's own budget sweep; referee it, but require an explanation of Table 3.","tokens_in":20878,"tokens_out":2803,"would_cite":true,"duration_ms":24691,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A splitting iteration solves the exact reduced Coulomb law without smoothing, reproducing stick, slip, and separation behavior in rigid-body simulation.","keywords":["Coulomb friction","cone complementarity","non-associated friction","operator splitting","forward-backward-forward","cone-constrained quadratic program","rigid-body dynamics","contact simulation"],"falsifier":"One concrete test: run the algorithm on a contact set with strong global coupling (for example, a 200-stone arch or a cyclic chain of contacts) in double precision, and check whether the dimensionless residual r_c drops below a tolerance of, say, 1e-12. If the residual stalls at the old single-precision floor or fails to decrease with more outer iterations, then the observed convergence was an artifact of the float32 floor rather than true convergence of the exact law. A second falsifier is a constructed two-contact configuration where the explicit correction is known to be non-contractive; if","tokens_in":19881,"feed_emoji":"🧱","tokens_out":4974,"duration_ms":43428,"temperature":0.7,"pith_summary":"This paper tries to establish that the exact reduced Coulomb friction law — the one that distinguishes sticking, sliding, and separation — can be solved by a fixed-point iteration that splits the law into a convex cone-constrained quadratic program and a scalar non-associated coupling updated explicitly. The iteration converges to the exact law, not an approximation of it, so no smoothing parameter or convex-relaxation step is needed. If true, this gives simulation engines a way to produce physically faithful friction behavior — stable stacks, arches that stand, clean stick-slip — at a cost comparable to convex solvers. The paper demonstrates the claim on rigid-body benchmarks, with Coulomb residuals that reach the single-precision floor and structures that stand where relaxed or smoothed solvers drift or collapse.","feed_headline":"A split iteration solves pure Coulomb friction without smoothing","feed_subtitle":"Benchmarks show stick-slip and stacking match exact physics where relaxed solvers drift.","key_machinery":"The load-bearing mechanism is the forward-backward-forward (FBF) outer iteration, a three-step update: evaluate the non-associated coupling B at the current reaction, solve a strongly convex cone-constrained quadratic program with that coupling frozen, and apply an explicit correction that compensates for the change in B between the current and intermediate iterates, followed by a projection onto the friction cone. The inner cone QP is modular — any second-order cone solver can be used — and its proximal regularization makes it well-conditioned even when the Delassus operator is rank-deficient. The outer iteration is safeguarded by an adaptive step-size rule that shrinks the step when a loca","core_discovery":"The central discovery is that the reduced inclusion for exact Coulomb friction decomposes as A(λ)+B(λ)=0, where A is a linear response plus the normal cone of the friction cone (a strongly convex cone QP when regularized) and B is a scalar non-associated coupling that adds the tangential sliding speed into the normal velocity. The paper's solver freezes B, solves the cone QP, then corrects the reaction by subtracting the change in B and projecting back onto the cone. At convergence, this correction reinstates the full coupling, so the fixed point satisfies the De Saxcé–Feng cone complementarity condition exactly. The paper validates this by measuring a dimensionless residual that scores prim","pith_inferences":["The paper's empirical convergence results suggest a testable hypothesis: if the contact graph is globally coupled (force transmitted across the whole structure), the explicit correction loses contractivity and the outer iteration becomes iteration-bound; a partial implicit treatment of B — suggested by the paper's own discussion — is a natural next step and could restore a formal guarantee.","The dimensionless residual r_c is defined with scale floors; the paper treats the single-precision floor as a convergence ceiling. In a float64 pipeline the same residual would be smaller, but the distinction between 'converged' and 'at the precision floor' would need a tighter tolerance or a different convergence criterion.","The decomposition A+B is not restricted to the rigid-body velocity-level setting; it may extend to deformable contact if paired with a position-level treatment, which the paper lists as future work. A reader could test that extension by pairing the FBF outer loop with a position-level contact solver.","The paper's comparison against an 'industrial-strength' exact-law solver indicates the method is competitive, but the comparison is on a few benchmarks; a systematic benchmark across random contact configurations would quantify how often the empirical convergence holds."],"forward_implications":["The method produces the exact Coulomb complementarity — separation, stick, and slip — without introducing a smoothing parameter, so the user does not have to trade accuracy against solver cost.","The inner solve is a standard strongly convex cone QP, so the outer loop can be combined with different cone solvers chosen for a particular hardware or contact count.","The algorithm is matrix-free and parallel over contacts, so it scales to contact-rich scenes; the benchmarks include a 101-stone arch and a ten-level house of cards.","Structures that depend on frictional support (arches, card houses) remain standing under the exact law, whereas convex-relaxed solvers exhibit creep or premature collapse in the same scenes.","Since the outer update is explicit in B, the method's cost per iteration is low; the paper reports wall times competitive with or faster than a leading industrial solver on the tested scenes."],"fun_headline_variants":["Split solver reproduces exact Coulomb friction","Exact Coulomb friction via split cone QP","No smoothing: split scheme solves pure Coulomb","Split iteration matches exact Coulomb complementarity","Cone split cracks stick-slip without relaxation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's load-bearing premise is that the FBF outer iteration converges for the non-monotone coupling B — a property it states it establishes empirically, using a single-precision Coulomb residual as the convergence signal — so if convergence fails on larger or more tightly coupled contact graphs, the claim that the iterate satisfies the exact Coulomb law is unsupported.","fun_headline_variants_meta":{"raw":{"variants":["Split solver reproduces exact Coulomb friction","Exact Coulomb friction via split cone QP","No smoothing: split scheme solves pure Coulomb","Split iteration matches exact Coulomb complementarity","Cone split cracks stick-slip without relaxation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1224,"prompt_tokens":660,"completion_tokens":564,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":404,"completion_tokens_details":{"reasoning_tokens":498}},"tokens_in":404,"tokens_out":564,"duration_ms":5459,"temperature":1.0,"reasoning_tokens":498,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T12:16:15.346844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete test: run the algorithm on a contact set with strong global coupling (for example, a 200-stone arch or a cyclic chain of contacts) in double precision, and check whether the dimensionless residual r_c drops below a tolerance of, say, 1e-12. If the residual stalls at the old single-precision floor or fails to decrease with more outer iterations, then the observed convergence was an artifact of the float32 floor rather than true convergence of the exact law. A second falsifier is a constructed two-contact configuration where the explicit correction is known to be non-contractive; if","supporting_citations":[],"review_version":1}