{"id":"2bbeccb7-7d1a-4ef3-8557-685214b3cf0a","arxiv_id":"2607.26407","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Distributed linearized ADMM with exact proximal steps solves dynamic optimal transport stably near vanishing densities and achieves ~6–7× speedup on 12 cores.","lead":"This paper introduces a distributed linearized ADMM algorithm for dynamic optimal transport that stays stable even when initial or terminal densities approach zero. It also cuts per-computer memory by splitting the space-time grid across agents, reaching about a 6–7× speedup with 12 CPU cores.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The discretized problem (2.13) and its projection (2.16) omit the M0≥0 constraint; if the relaxed optimum contains negative staggered densities, LADMM is solving a nonphysical relaxation, not the dynamic OT problem claimed.","rationale":"The reader's weakest assumption correctly identifies the missing M0≥0 constraint. My independent reading confirms this is the most load-bearing issue: it directly affects whether the method solves the stated dynamic OT problem. The paper's objective f is finite even when individual staggered density values are negative, as long as the averaged density is positive, and the projection in (2.16) only enforces the linear continuity constraint. Thus the algorithm can in principle return unphysical negative densities. This is not merely a technicality: the numerical comparisons report lower objective values for LADMM, which could be explained by solving a relaxation rather than the true constrained problem. A second, secondary gap is that Theorem 2.4 assumes constant β_k=β, while experiments use the adaptive schedule (4.1); this means the theoretical convergence guarantee does not cover the actual implementation. However, the nonnegativity issue is more fundamental because it concerns the definition of the problem being solved. I recommend keeping the reader's CONDITIONAL verdict: the paper should be required to either state and justify the relaxation, add the constraint, or demonstrate that the relaxed optimum always has a nonnegative representative. The proposed concrete test would settle the question on a small instance.","tokens_in":21853,"tokens_out":17950,"duration_ms":169709,"concrete_test":"Take the 1D instance of Section 4.2.3 at n0=64, n1=256 (Table 5) and run LADMM to its stated tolerance. Record min_j (M0^K)_j and the reported Obj_proj. Then solve the same discretized problem with the explicit constraint M0≥0 using a general-purpose convex solver (e.g., SCS/CVXPY) on the same grid, or by adding a projection onto R_+ in the M0-update of Algorithm 2.1. Compare the optimal objective and the feasibility of LADMM's final M0. If min M0^K < 0 and the constrained optimum has a strictly larger objective, the relaxation changes the problem and the central claim fails; if the objectives coincide and a nonnegative representative exists, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing issue is the unstated omission of the nonnegativity constraint. In the original Benamou–Brenier problem (1.1), ρ is constrained to R_+ (via F in (1.2)). The discretized problem (2.11)–(2.13) and the projection step (2.16) enforce only the linear continuity constraint C in (2.10); no constraint M0≥0 appears anywhere. The objective (2.6) is finite whenever the averaged density Mbar0 = Avg0(M0)+b_avg is positive, but individual staggered values M0_j may be negative. The proximal update in Lemma 2.1 only guarantees that the auxiliary variable Mbar0 is nonnegative, not that M0 is nonnegative. Consequently, the algorithm can converge to a point with negative entries in M0, which is not an admissible density for (1.1). The paper never states this relaxation, nor proves that every relaxed feasible point can be replaced by a nonnegative staggered density with the same objective. The lower reported objective values for LADMM (e.g., Tables 4–6) could therefore be an artifact of solving a different, nonphysical problem. This bears directly on the central claim of solving dynamic OT.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a linearized ADMM for the discretized Benamou–Brenier dynamic optimal transport problem. The key idea is to introduce auxiliary averaged variables so that the subproblem involving the dynamic OT energy can be solved exactly by a proximal mapping, thereby avoiding gradient-based steps whose Lipschitz constants blow up when the density approaches zero. A distributed variant partitions the space–time variables along the first spatial dimension to reduce per-agent memory and enable parallelism. Numerical experiments in 1D and 2D compare the method with FISTA, ALM, and G-prox, including challenging near-zero density settings, and report speedups of roughly 6–7× when scaling from 1 to 12 agents. A convergence statement is imported from Gao–Zhang and Xu.","tokens_in":22213,"tokens_out":22421,"duration_ms":198733,"significance":"If the technical gaps are repaired, the exact-proximal reformulation is a genuinely useful idea for the vanishing-density regime, and the distributed implementation addresses a real memory bottleneck in dynamic OT. The numerical study is extensive and includes systematically chosen comparisons in one and two spatial dimensions. However, as written, the paper does not yet establish that the algorithm solves the stated dynamic OT problem: the nonnegativity constraint on the discrete density is missing, the proximal lemma contains a false uniqueness claim, the pseudocode has a sign inconsistency, and the convergence theorem is asserted without verifying its hypotheses.","major_comments":[{"comment":"The discretized problem never imposes M0 ≥ 0. In the continuous problem (1.1), F enforces ρ ∈ R_+; the staggered array M0 is the discrete counterpart of ρ, so feasible M0 should be nonnegative. The set C in (2.10) is only the linear continuity equation, and the projection (2.16) is an unconstrained projection onto C. Lemma 2.1 guarantees nonnegativity only of the auxiliary variable Mbar0, not of M0. Thus the algorithm may converge to points with negative entries in M0 that are not densities for (1.1). The paper neither states this relaxation nor proves that every relaxed feasible point can be replaced by a nonnegative staggered density with the same objective. Because the lower objective values reported in Tables 4–6 may arise from this relaxed, nonphysical problem, the central claim that LADMM solves the dynamic OT problem is not established. Relatedly, f in (2.6) is undefined when Avg0","section":"§2.1, Eqs. (2.11)–(2.13); §2.3, Eq. (2.16)"},{"comment":"There is a sign inconsistency in the auxiliary-variable update. With the augmented Lagrangian defined in (2.14) and the dual updates in (2.17), completing the square over Mbar0 with M0 fixed gives an input c + Λ^k/β_k to the proximal map, not c − Λ^k/β_k (and similarly for Mbar). Algorithm 2.1 line 3 writes minus signs. These minus signs are inconsistent with the gradient step (2.18), whose gradient is Avg0^*(Λ^k + β_k(c^k − Mbar0^{k+1})); with the correct plus sign that gradient vanishes at the proximal point, as expected in ADMM. As written, the algorithm is not the LADMM scheme analyzed in Theorem 2.4, and the reported numerical convergence cannot be attributed to the stated updates unless the code uses a different sign convention. The authors should correct the sign or clarify the convention.","section":"§2.3, Algorithm 2.1 line 3 and Eqs. (2.14),(2.17),(2.18)"},{"comment":"The claim that ρ in (2.12) is “the unique real-number solution” of the cubic (A.7) is false. For example, with γ=1, ρ̂=10, m̂=0, the cubic is (ρ+1)^2(ρ−10)=0, which has real roots −1 and 10, while the correct proximal value is ρ=10. Non-uniqueness of real roots can also occur in the “otherwise” branch. Since Algorithm 2.1 requires solving this cubic, the paper must specify a deterministic selection rule (e.g., the unique nonnegative real root, or the largest real root) and prove that this root gives the proximal minimizer. As stated, the lemma is incorrect.","section":"Lemma 2.1 and Appendix A"},{"comment":"The convergence theorem is asserted by direct citation of [18, Thm 4.3] and [47, Thm 2.13] without verifying their hypotheses. In particular, the objective f in (2.6) is not a closed proper convex function unless its domain and extended-value convention are specified; Algorithm 2.1 as written does not match the scheme covered by those references because of the sign issue above; and the numerical schedule (4.1) uses β_k nonconstant, whereas Theorem 2.4 assumes constant β. The O(1/K) guarantee is therefore not established for the implemented method. The authors should either verify the hypotheses of the cited results, or prove convergence directly for the actual algorithm and parameter schedule.","section":"§2.3, Theorem 2.4"}],"minor_comments":[{"comment":"The table formatting is garbled in the manuscript text: column separators are missing and several rows appear to have entries merged or omitted (e.g., Table 1, n1=4096). This makes the numerical claims difficult to verify and should be corrected.","section":"Tables 1–7"},{"comment":"Typo: “egivenvectors” should be “eigenvectors”.","section":"§2.2, near Eq. (2.28)"},{"comment":"Typo: “dentity setting” should be “density setting”.","section":"Figure 9 caption"},{"comment":"The domain/codomain conventions for V_d and V in the definitions of Avg_d and Avg_d^* are terse. A brief statement that bJ_d indexes the staggered grid and how boundary entries are handled would improve readability.","section":"Eqs. (2.2)–(2.3)"}],"recommendation":"major_revision","confidential_remarks":"The convergence theorem is imported from [47], a paper by one of the authors, but its hypotheses are not verified. More importantly, the apparent sign error in Algorithm 2.1 is load-bearing: as written, the pseudocode is not the algorithm whose convergence is cited, and the numerical results may not be reproducible from the printed updates. I recommend asking the authors to provide the actual implementation or a step-by-step derivation of the updates before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a solid algorithmic paper with a genuinely useful idea, and the experiments are far more systematic than most in this area. The main thing to watch is the positivity constraint on the staggered density, which is never imposed and never justified. That is not a minor technicality; it goes to whether the method actually solves the problem being advertised.\n\nWhat is new: the exact proximal mapping of the Benamou–Brenier energy (Lemma 2.1, with proof in the appendix) used inside a linearized ADMM with an exact projection onto the continuity constraint, plus a distributed domain decomposition that really does cut per-agent memory. The near-zero-density regime is handled in a principled way—replacing gradient updates with a proximal step whose local Lipschitz constant doesn't blow up. The 1D and 2D experiments are careful: they compare against FISTA, G-prox, and ALM, report iteration counts, times, objective values and primal residuals, and the speedup numbers (6–7× with 12 agents) are believable. This deserves credit.\n\nThe soft spots. First, the discretized problem (2.13) and the projection (2.16) never include M0≥0. The original problem (1.1) requires ρ≥0. The algorithm only ensures the averaged quantity M0 is nonnegative, not the staggered values themselves. The paper never states this relaxation and never proves that any relaxed feasible point can be replaced by a nonnegative staggered density with the same objective. The lower LADMM objective values in Tables 4–6 could therefore come from a nonphysical solution. This is the one issue that could actually sink the central claim, and it needs a response. It may be fixable in practice—maybe the iterates stay nonnegative in all tested cases—but the paper doesn't show that.\n\nSecond, Lemma 2.1 calls the root of the cubic 'unique real,' but a cubic can have three real roots. The proof doesn't specify which root is the proximal minimizer. This is a presentation gap, not a fatal error, but it should be clarified.\n\nThird, Theorem 2.4 imports a convergence result from [47] without checking its hypotheses, and the experiments use a time-varying βk while the theorem assumes constant β. Likely fine, but the authors should state the match.\n\nThe grid-searched β0 is parameter tuning, not circular reasoning; that criticism doesn't land. The paper also doesn't release code or data, which makes the positivity issue harder to check.\n\nWho it's for: anyone working on numerical OT, mean-field games, or image registration. It deserves a serious referee; the core idea is strong enough that the positivity question should be resolved rather than the paper bounced. I'd send it to review. I would not cite it until the relaxation is addressed.","headline":"The algorithm is genuinely new and the numerical results are strong, but the unstated omission of M0≥0 in the discretized problem is a real gap that should be addressed before the claim of solving dynamic OT is accepted.","tokens_in":22595,"tokens_out":3478,"would_cite":false,"duration_ms":34459,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","49M27","65Y05","90C06"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that dynamic optimal transport with near-zero densities becomes tractable by replacing gradient steps with an exact proximal mapping inside a distributed linearized ADMM.","keywords":["dynamic optimal transport","Benamou–Brenier","linearized ADMM","exact proximal mapping","near-zero density","distributed optimization","domain decomposition","staggered grid"],"falsifier":"Solve a small 1D example with a source density that is exactly zero on some cells, run LADMM to convergence, and inspect the raw density values at half-integer grid points: if any are negative and forcing them nonnegative changes the objective beyond the stopping tolerance, the method is solving a relaxed problem, not the original transport problem.","tokens_in":21807,"feed_emoji":"🚚","tokens_out":7004,"duration_ms":67340,"temperature":0.7,"pith_summary":"The paper tackles two bottlenecks in computing dynamic optimal transport — the Benamou–Brenier problem of moving one density to another at least action: instability when source or target densities approach zero, and the heavy memory bill of storing the whole space–time solution. It proposes to reformulate the discretized problem with auxiliary averaged variables, so the transport energy's proximal mapping is available exactly, as the root of a cubic. Plugging that proximal map into a linearized ADMM removes the need for gradient steps whose step size would otherwise collapse as density vanishes. A second, distributed formulation splits the space–time grid among agents, cutting per-agent memory and running in parallel, with reported speedups above 6× on 12 CPU cores. A sympathetic reader would care because these are exactly the regimes — sparse or nearly empty source or target regions, fine grids, higher dimensions — where standard solvers stall or diverge.","feed_headline":"Exact proximal steps tame zero-density blowup in transport","feed_subtitle":"The method replaces fragile gradient updates with exact cubic-root steps and splits the grid across agents.","key_machinery":"The engine is the exact proximal mapping of the dynamic transport energy F(ρ,m)=||m||²/(2ρ): given an input pair (ρ̂,m̂), it returns (0,0) when ρ̂≤0 and ||m̂||²≤−2γρ̂, and otherwise a unique positive root ρ of the cubic ρ³+(2γ−ρ̂)ρ²+(γ²−2γρ̂)ρ−γ||m̂||²/2−γ²ρ̂=0, with m=m̂ρ/(ρ+γ). This exact map replaces gradient-type updates, so no step size must absorb the local Lipschitz constant 1/ρ³. Around it, LADMM alternates that proximal step with an exact projection onto the linear continuity-equation constraint, which amounts to solving a Poisson (or Sylvester, in 1D) equation via the fast cosine transform; the projection keeps every iterate feasible and makes the primal residual near machine preci","core_discovery":"The central claim is that the discretized dynamic OT problem can be reorganized into a form whose subproblem over the transport energy F(ρ,m)=||m||²/(2ρ) is exactly solvable: the proximal mapping of F at a given pair either returns (0,0) or the unique positive root of a cubic, with momentum rescaled accordingly. Wrapped in a linearized ADMM, this exact-prox step keeps iterates away from the blow-up caused by the energy's Hessian scaling like 1/ρ³, while an exact projection onto the continuity equation — a Poisson or Sylvester step computed with a fast cosine transform — keeps the mass-conservation constraint satisfied at every iteration. The paper reports that in 1D and 2D experiments where","pith_inferences":["The same exact-proximal trick should extend to other homogeneous transport costs and to mean-field planning problems where the same 1/ρ³ gradient blow-up appears; any cost of the form ||a||²/ρ yields a polynomial root-finding proximal step.","The projection step is the main communication bottleneck in the distributed scheme; replacing the cosine-transform solver with a multigrid or basis-free Poisson solver could extend the speedup curve to many more agents.","Stability near zero density is demonstrated empirically rather than by a rate; a quantitative bound relating iteration count to the minimum density would turn the stability claim into a theorem.","The discretized problem is solved without an explicit nonnegativity constraint on the staggered density, so a direct comparison against a nonnegative-constrained solve would show whether the reported transport paths are realizable by true densities."],"forward_implications":["Problems with source or terminal densities that are exactly zero in places no longer force tiny step sizes; the paper's experiments show convergence with smallest densities around 10⁻¹⁰.","Per-agent memory drops because each agent stores only its slice of the space–time variables, making finer grids and higher dimensions more practical.","The distributed version gets a reported 6–7× speedup with 12 CPU cores, with communication limited to shared spatial boundaries.","Because a projection step is performed each iteration, mass conservation is never violated, and reported primal residuals are 10⁻¹¹ to 10⁻¹³.","The ergodic convergence guarantee O(1/K) of LADMM applies, so averaged iterates give a feasible point with a controlled objective gap."],"fun_headline_variants":["Zero-density blowup tamed by exact cubic-root steps","Distributed LADMM cuts memory in dynamic optimal transport","Exact proximal steps stabilize dynamic OT near zero density","Cubic-root steps fix transport blowup, split across agents","Scalable OT: exact proximal steps plus distributed splitting"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that letting the discretized density values take negative values during the computation does not change the final answer; the paper assumes this relaxation is harmless without stating it.","fun_headline_variants_meta":{"raw":{"variants":["Zero-density blowup tamed by exact cubic-root steps","Distributed LADMM cuts memory in dynamic optimal transport","Exact proximal steps stabilize dynamic OT near zero density","Cubic-root steps fix transport blowup, split across agents","Scalable OT: exact proximal steps plus distributed splitting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000486,"raw_usage":{"total_tokens":2268,"prompt_tokens":812,"completion_tokens":1456,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1376}},"tokens_in":556,"tokens_out":1456,"duration_ms":11880,"temperature":1.0,"reasoning_tokens":1376,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:37:38.418215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve a small 1D example with a source density that is exactly zero on some cells, run LADMM to convergence, and inspect the raw density values at half-integer grid points: if any are negative and forcing them nonnegative changes the objective beyond the stopping tolerance, the method is solving a relaxed problem, not the original transport problem.","supporting_citations":[],"review_version":1}