{"id":"94679421-e303-472d-9d1b-9bbab816eae5","arxiv_id":"2509.15385","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"VPTPD is a variable-preconditioned transformed primal-dual algorithm for dynamic JKO schemes that reports large speedups over prior primal-dual methods in 1D to 3D numerical tests.","lead":"A new solver, VPTPD, combines a transformed primal-dual scheme with proximal splitting to compute Wasserstein gradient flows with nonsmooth energies and nonlinear mobilities. Tests from 1D to 3D report up to 20x speedups over existing primal-dual JKO solvers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Adaptive step-size rule is self-contradictory: Table 2 sets γ−=16 while §4.3 requires γ−∈(0,1), so the 'decrease' branch multiplies λ by 16; reported adaptive speedups may not be reproducible.","rationale":"The reader's weakest assumption was the lack of an outer convergence proof and the heuristic adaptive step-size rule. I agree that this is the central structural gap. But in reviewing the algorithm I found a more concrete, internal inconsistency: the adaptive rule (4.11) is described with γ−∈(0,1) as a decrement factor, yet the experimental setup in Table 2 prescribes γ−=16. Interpreting the algorithm literally, the branch that is supposed to shrink λ under worsening monitors instead grows it by 16×. This directly threatens the reproducibility of the adaptive speedups, which are the basis for the 'up to 20× speedup' headline. The proposed test settles whether this is merely a typo in the caption or a real algorithmic flaw. If the literal run diverges, the claims need immediate correction; if it converges, the manuscript should still be revised because the stated parameters violate the rule. In either case, the paper's conditional status as an efficient but unproven heuristic is appropriate: the constant-step-size VPTPD results and the proximal-operator theorem stand as solid contributions, but the central convergence guarantee remains absent. Hence I do not move the reader's CONDITIONAL verdict; I would keep it CONDITIONAL pending the correction and repro check.","tokens_in":16957,"tokens_out":5757,"duration_ms":66389,"concrete_test":"Run the 1D saturation experiment of Table 2 using Algorithm 4.1 literally as specified, with adaptive parameters λ0=1, λmax=100, λmin=0.0001, γ+=10, γ−=16. If the iteration diverges, stalls, or violates energy dissipation, then the reported VPTPD(s) speedups are not reproducible without correction. Then repeat with the one-character fix γ−=1/16 and compare iteration counts and CPU times to the table.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 4.1's adaptive strategy is load-bearing for the headline speedups, but as written it is internally inconsistent. Section 4.3 defines γ−∈(0,1) as a decrement factor and requires γ+·γ−<1, yet the 1D saturation table (Table 2) specifies γ+=10, γ−=16 (so γ+·γ−=160). Under (4.11), if a convergence monitor worsens, λ_n = max(λ*_n γ−, λ_min), which would multiply λ by 16 exactly when errors grow—the opposite of the stated intent. This makes the 10x+ CPU-time gains reported for VPTPD(s) in Tables 2–4 unrepeatable from the text. Separately, the outer iteration (4.1) has no convergence proof: Theorem 4.2 only covers the component-wise proximal subproblem, and §4.3 concedes that no rigorous λ,σ stability estimate is available. The γ− inconsistency is the sharpest concrete place where the missing theory surfaces as an algorithmic defect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VPTPD, a variable-preconditioned transformed primal-dual method for the fully discrete dynamic JKO scheme (2.3) of generalized Wasserstein gradient flows. The method combines proximal-gradient splitting of the nonsmooth transport action with explicit treatment of the smooth energy, variable-dependent diagonal preconditioners based on the Hessian of a regularized objective, a component-wise proximal-operator Newton solver, and an adaptive step-size rule. Numerical experiments in 1D, 2D, and 3D report large reductions in iteration counts and CPU time relative to PDJKO and PrePDJKO, including speedups up to about 20x, while preserving energy dissipation, mass conservation, and density bounds.","tokens_in":17304,"tokens_out":5017,"duration_ms":56689,"significance":"If the algorithmic claims hold, the paper is a useful contribution to structure-preserving computation of Wasserstein-type gradient flows, with potential practical impact on large-scale simulation. The paper gives the algorithm in detail, specifies default parameters and stopping criteria, and reports thorough benchmark comparisons against existing methods rather than deriving the central efficiency claim from those methods; the comparative claims are therefore not circular. The inner proximal subproblem is treated seriously: Theorem 4.2 addresses existence, uniqueness, bounds, and Newton convergence for two families of concave mobilities. The main weaknesses are the lack of a convergence theory for the outer iteration and an internal inconsistency in the adaptive step-size parameters used in the headline speedup table.","major_comments":[{"comment":"The adaptive rule is internally inconsistent with the reported parameters. Section 4.3 defines γ−∈(0,1) as a decrement factor and requires γ+·γ−<1, but the caption of Table 2 sets γ−=16 (with γ+=10), so γ+·γ−=160. Under (4.11), the 'decrease' branch would multiply λ by 16 whenever a convergence monitor worsens, which is the opposite of the stated decrement mechanism. Since the VPTPD(s) speedups in Tables 2–4 rely on this adaptive rule, the reported results are not reproducible from the text. Please correct the table or the rule and re-run the affected experiments; this is not a typographical nit, as it changes the algorithm actually executed.","section":"§4.3, Eq. (4.11) and Table 2"},{"comment":"The outer iteration (4.1) has no convergence proof. Theorem 4.2 concerns only the component-wise proximal subproblem (4.6), and Section 4.3 explicitly states that no rigorous upper bound on λ and σ is currently available. The paper's central numerical claim—that VPTPD converges in far fewer iterations than PrePDJKO—therefore rests on an unproved algorithmic step. In a numerical analysis journal this needs to be addressed: either provide a convergence/stability theorem for (4.1) under stated assumptions, or clearly state that the outer iteration is heuristic and temper the corresponding claims in the abstract and introduction.","section":"§4.1 / §4.3"},{"comment":"The proof of Theorem 4.2 is sketched rather than complete. After deriving L(ρ) in (4.8), the proof states that L'(ρ)>0 is 'easy to check' and then refers to Table 1 for root intervals, signs of L'', initial Newton values, and endpoint solutions. No derivation is given for the interval classification or for the Newton convergence conclusion, even though this theorem is one of the four main contributions claimed in the abstract. Please supply a complete argument, or a reference to a prior result that contains it, for each case in Table 1.","section":"§4.2, Theorem 4.2"}],"minor_comments":[{"comment":"There are several typos and grammatical slips: 'trasport' (p.2), 'contuity' (p.4), 'taclking' (p.6), 'inditates' (p.11), 'ojec-tive' (p.8), and 'conditioner number' should be 'condition number' (p.9).","section":"Throughout"},{"comment":"The control flow is confusing: the algorithm uses an outer 'while' and inner 'repeat' with the stopping criteria placed at line 9. The nesting should be clarified, and the algorithm should specify how the outer JKO iterations and inner fixed-point iterations terminate.","section":"Algorithm 4.1"},{"comment":"The notation '∇F_h = ∇J_h(u^n) + 1/(2τ)∂Φ_h(u^{n+1})' is dimensionally/notationally imprecise: since F_h=Φ_h/(2τ)+J_h, the subgradient term should be written as (1/(2τ))∂Φ_h(u^{n+1}) with an explicit understanding of where it is evaluated. Please rewrite to avoid confusion.","section":"§4.1, near Eq. (4.1)"},{"comment":"The text says the full discrete scheme is 'elaborated elsewhere' and 'we refer the readers to the supplementary document,' but no supplementary document is included in the arXiv listing. If the derivation is not repeated, a complete reference to the prior publication is needed.","section":"§2"},{"comment":"Reference [11] is listed with volume '0 (0)' and pages 'S386–S413'; the publication details should be completed or marked as accepted/in press.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The main concerns are the unrepeatable adaptive-parameter setting in Table 2 and the missing convergence theory for the outer iteration. I would ask the authors to fix the γ− inconsistency and re-run the experiments, and to either add a convergence theorem for (4.1) or explicitly label the adaptive strategy as heuristic. If those points are addressed, the paper could be suitable for publication in this journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely useful: it extends the transformed primal-dual framework to the nonsmooth dynamic JKO scheme by adding a proximal gradient step, and it backs the inner solver with a real theorem (existence, uniqueness, boundedness, Newton convergence for two concave mobility families). The construction of variable diagonal preconditioners from the regularized Hessian is a sensible way to keep the proximal subproblems separable while improving conditioning. The numerical comparisons against PDJKO and PrePDJKO are consistent and the reported speedups are large, especially in 3D. That is the honest core of the contribution.\n\nThe soft spots are real but not fatal. The outer VPTPD iteration has no convergence proof; Section 4.3 explicitly says there is no rigorous step-size estimate. So the method's efficiency is currently an empirical claim, and the adaptive rule is load-bearing for the best numbers. That alone would make the paper conditional. But there is a sharper problem: the adaptive step-size rule is internally inconsistent. Section 4.3 defines γ− as a decrement factor in (0,1) and requires γ+·γ−<1, yet Table 2 (1D saturation) reports γ+=10 and γ−=16, so γ+·γ−=160. Under (4.11), the “decrease” branch would multiply λ by 16 when the convergence monitors worsen. That is the opposite of the stated intent, and it makes the reported 10x+ CPU-time gains for VPTPD(s) in that table unreproducible from the text. This is not a minor typo; it is exactly where the missing outer convergence theory surfaces as an algorithmic defect. The stress-test note is right, and I would not have caught it without the numbers.\n\nI also second the reader's worry about reproducibility: no code or data are shipped, and the tables only report aggregates, so the iteration counts and CPU times cannot be independently checked. The mobility restriction is stated clearly, and the paper cites its own prior work appropriately; self-citation here is legitimate building, not circularity.\n\nWho is this for? People working on structure-preserving schemes for generalized Wasserstein gradient flows, especially those who need to run large 2D/3D JKO simulations. They will want to try this method. But before the efficiency claims are reliable, the authors need to either supply a convergence analysis for the outer iteration or at least fix the adaptive rule and justify the parameter choices.\n\nRecommendation: serious referee, yes. The paper deserves a real review round, and the γ− inconsistency should be caught and fixed in revision.","headline":"Useful numerical extension of TPD to nonsmooth JKO, with a real internal inconsistency in the adaptive step-size table that needs fixing before the speedups are reproducible.","tokens_in":17797,"tokens_out":1056,"would_cite":true,"duration_ms":12216,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35A15","47J25","47J35","49M29","65K10","76M30","90C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A variable-preconditioned transformed primal-dual method solves generalized Wasserstein gradient flows up to 20 times faster than existing methods while preserving structure.","keywords":["Wasserstein gradient flows","JKO scheme","transformed primal-dual method","variable preconditioning","proximal splitting","nonlinear mobility","structure-preserving schemes","adaptive step-size"],"falsifier":"Run Algorithm 4.1 on the 1D saturation test with λ0 set 100 times above the paper's value (λ0=5 instead of 0.05). If iteration counts still drop by 99% and energy remains monotone, the hidden step-size bound is not load-bearing; if the iteration count or energy error blows up, the adaptive rule is genuinely fixing an otherwise unstable regime.","tokens_in":16799,"feed_emoji":"⚡","tokens_out":8001,"duration_ms":73835,"temperature":0.7,"pith_summary":"The paper proposes an algorithm for simulating generalized Wasserstein gradient flows—partial differential equations with concentration-dependent mobility—by solving the fully discrete dynamic JKO scheme exactly in a structure-preserving sense. The algorithm, called VPTPD, extends the transformed primal-dual (TPD) method to nonsmooth objective functionals by combining a proximal gradient step for the nonsmooth transport cost with an explicit gradient step for the smooth energy, and uses variable-dependent diagonal preconditioners built from the Hessian of a regularized objective. The authors prove that the component-wise proximal subproblem with these preconditioners has a unique bounded solution computable by a convergent Newton iteration, and they introduce an adaptive step-size strategy to improve reliability. Numerical tests in 1D, 2D, and 3D show that the method reduces the number of iterations by up to 99.8% and CPU time by up to a factor of 20 compared to the PrePDJKO method, while preserving energy dissipation, mass conservation, and bounds. The method targets large-scale simulations of degenerate, nonlinear, nonlocal gradient flows such as Cahn-Hilliard, Keller-Segel, and droplet wetting models.","feed_headline":"New solver runs Wasserstein gradient flows up to 20x faster","feed_subtitle":"Variable preconditioners and proximal splitting cut iteration counts 20x while preserving energy, mass, and bounds.","key_machinery":"The central mechanism is the variable-preconditioned transformed primal-dual flow, with the specific preconditioners I_u = diag(∇²F̂_h(u^k)) (the diagonal of the Hessian of a regularized objective at the previous JKO solution) and I_p = B I_u^{-1} B^T (the Schur complement). This transformation makes the saddle-point system nearly upper triangular and diagonally dominant, almost decoupling primal and dual updates; the nonsmooth transport cost is handled through a separable proximal operator, and each component-wise subproblem is solved by a bound-preserving Newton iteration.","core_discovery":"Each step of the dynamic JKO scheme for generalized Wasserstein gradient flows is a nonsmooth, linearly constrained minimization. The paper shows it can be solved efficiently by a transformed primal-dual (TPD) method extended with proximal splitting. The method uses variable-dependent diagonal preconditioners—I_u from the Hessian of a regularized objective, I_p its Schur complement—that nearly decouple the primal and dual updates. The paper proves the component-wise proximal subproblem has a unique bounded solution for concave mobilities and a convergent, bound-preserving Newton solver. Numerical tests in 1D–3D report up to 99.8% fewer iterations and a 20x CPU speedup, while preserving energ","pith_inferences":["We infer that the diagonal-Hessian preconditioning idea could transfer to other saddle-point problems where the objective is a sum of a smooth term and a separable nonsmooth term, such as optimal transport-based inverse problems or imaging.","A rigorous convergence analysis of the outer iteration would likely require a sharp bound on λ and σ in terms of the Lipschitz constant of ∇J_h under the I_u metric and the condition number of B I_u^{-1}B^T; such a bound is absent from the paper.","A testable extension is to replace the heuristic adaptive rule with a backtracking line search on a Lyapunov functional; if convergence can be certified, the method would gain a verifiable guarantee without losing its speed.","The success of the component-wise proximal Newton solver suggests that the same bound-preserving strategy could accelerate other primal-dual schemes that currently rely on expensive exact proximal evaluations."],"forward_implications":["The reported speedups imply that for the tested problems, the computational bottleneck shifts from iteration count to per-iteration matrix inversion, where the paper notes that parallel and high-performance computing techniques could yield further gains.","The proved existence, uniqueness, and bound-preserving Newton solver for the proximal subproblem makes the method applicable to a broad class of concave mobilities, including degenerate cases M(ρ)=ρ and M(ρ)=1−ρ².","The semi-implicit-explicit splitting provides a template for extending transformed primal-dual methods to other nonsmooth optimization problems that combine a smooth energy with a separable convex nonsmooth term under linear constraints.","The adaptive step-size rule demonstrates a practical way to maintain acceleration when the smooth part's Lipschitz constant is not explicitly known, and the paper's numerical results show it further improves efficiency beyond the constant-step version."],"fun_headline_variants":["Wasserstein flows: 20x faster with variable preconditioners","Solver cuts Wasserstein flow iterations up to 99.8%","Variable preconditioners accelerate Wasserstein flows 20x","Wasserstein flows solved 20x faster by preconditioned solver","New primal-dual method speeds Wasserstein flows 20x"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The outer iteration's convergence and the reported speedups depend on the heuristic adaptive step-size rule staying within a stability bound for λ and σ that the paper does not estimate rigorously; if for some energy or mobility that hidden bound is violated, the iteration counts and CPU improvements do not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Wasserstein flows: 20x faster with variable preconditioners","Solver cuts Wasserstein flow iterations up to 99.8%","Variable preconditioners accelerate Wasserstein flows 20x","Wasserstein flows solved 20x faster by preconditioned solver","New primal-dual method speeds Wasserstein flows 20x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001193,"raw_usage":{"total_tokens":4768,"prompt_tokens":761,"completion_tokens":4007,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":3919}},"tokens_in":505,"tokens_out":4007,"duration_ms":30338,"temperature":1.0,"reasoning_tokens":3919,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:12:32.546165+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 4.1 on the 1D saturation test with λ0 set 100 times above the paper's value (λ0=5 instead of 0.05). If iteration counts still drop by 99% and energy remains monotone, the hidden step-size bound is not load-bearing; if the iteration count or energy error blows up, the adaptive rule is genuinely fixing an otherwise unstable regime.","supporting_citations":[],"review_version":1}