{"id":"77f80e9d-c8b5-468b-b56f-d6d25b20e881","arxiv_id":"2504.16890","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A min-max gradient flow with a dynamically adapted KL penalty is proposed as a particle method for approximating optimal transport couplings, with claimed convergence to the optimal plan.","lead":"This paper recasts the optimal transport problem as a two-player min-max game, with a penalty weight that grows over time, and proposes a particle algorithm that approximates the optimal transport plan. The authors claim convergence guarantees for the dynamics and report faster runtimes than an earlier particle method, while acknowledging the method is not faster than the standard Sinkhorn algorithm.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3.4's key identity (3.3) asserts b1(x)b2(y) = σ[Λ](x,y), which forces e^{-c/Λ}μν to factor; for any non-separable cost such as quadratic it is impossible, so all downstream Danskin and convergence results rest on a false premise.","rationale":"The reader's weakest-assumption pinpoints Lemma 3.4, and my reading agrees: this lemma is the load-bearing step. My analysis sharpens the objection: (3.3) is not just a missing proof or an unverified regularity condition; it algebraically forces the Gibbs kernel to be a rank-one product, which fails for any non-separable cost (e.g. quadratic cost). The paper's subsequent Lemmas 3.5, 3.7, 3.8, 3.11 and Theorem 3.9 all use the marginal formulas (3.5)-(3.6) or the measure σ introduced in (3.3), so they inherit the defect. Corollary 3.14, the statement that b[Λ(t)] converges in W to an OT plan and Ed(Λ)→c*, is therefore unproved. I did not need to adjudicate the separate displacement-convexity results (Theorems 3.2-3.3) or the numerics; the central advertised theoretical contribution fails as written. The numerical implementation and code are real, and some weaker estimates (notably Ed≤c* and the O(√t) growth bound) survive because they follow from plugging in an OT plan rather than from (3.3). However, those weaker estimates alone do not establish convergence of the best response to an OT plan, and the proof supplied for that convergence is invalid. A repaired argument would need to work directly with the true minimizer of E(·,Λ), which may be singular, and would have to prove continuity and Danskin-type properties without the product ansatz; no such argument is present. Hence the appropriate verdict remains REJECT rather than ACCEPT or CONDITIONAL: the main theorem is not established, even though the underlying method may be salvageable.","tokens_in":12700,"tokens_out":13390,"duration_ms":133325,"concrete_test":"Compute the mixed partial derivative of log σ[Λ] for the quadratic cost: with μ = ν = Uniform([0,1]) and c(x,y) = (x-y)^2, (3.3) would imply log(b1(x)b2(y)) = -(x-y)^2/Λ + log μ(x) + log ν(y) - log Z. The left side has zero mixed partial derivative, while the right side has ∂²/∂x∂y = -2/Λ ≠ 0. This contradiction settles that (3.3) cannot hold for a non-separable cost. A discrete analog: for μ=ν uniform on {0,1} with c(0,0)=c(1,1)=0, c(0,1)=c(1,0)=1 and Λ=1, the matrix σ is [[0.269,0.099],[0.099,0.269]], which has rank 2 and therefore cannot be written as b1(i)b2(j); this refutes the claimed product formula without any regularity assumptions.","verdict_should_be":"REJECT","load_bearing_attack":"The central convergence proof rests on Lemma 3.4. Even taking the Euler-Lagrange derivation at face value, equation (3.3) is internally inconsistent for the intended setting. It states b1(x)b2(y) = (1/Z)e^{-c(x,y)/Λ} μ(x)ν(y), where b1,b2 are marginals of the best response b[Λ]. Since the left side is a product of a function of x and a function of y, the right side must have rank one, i.e. e^{-c/Λ}μν must factor. That forces c(x,y) = a(x)+b(y) (up to constants) on the support of b[Λ]. For c(x,y) = ||x-y||^2, the log-density of the proposed σ has mixed partial derivative ∂²/∂x∂y = -2/Λ ≠ 0, while any factored RHS has zero mixed partial derivative; hence no absolutely continuous coupling can satisfy (3.3). The actual minimizer of E(·,Λ) in this equal-marginal case is the diagonal coupling, which is singular and has product of marginals μ⊗μ, not the Gibbs kernel σ. Consequently the explicit marginal formulas (3.5)-(3.6), the closed form of V in (3.7), Lemma 3.11's covariance formula for dV/dΛ, and the continuity/Danskin arguments (Lemmas 3.7-3.8, Theorem 3.9) that invoke them are unsupported. Corollary 3.14's conclusion b[Λ(t)] → ρ* uses exactly these ingredients, so the advertised convergence theorem is not established by the paper. The estimate Ed(Λ) ≤ c* does survive without (3.3), and the numerical method may work, but the proof of the main claim collapses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reformulates the Kantorovich optimal transport problem as a min-max problem, where the marginal constraints are relaxed by KL divergences whose weight Λ is dynamically adjusted, and studies a two-timescale Wasserstein gradient flow. The main theoretical claims are (i) displacement convexity of the relaxed energy (Theorems 3.2–3.3), (ii) a Danskin-type theorem for the best-response map (Theorem 3.9), and (iii) convergence of the best-response dynamics to an optimal transport plan (Corollary 3.14). The paper also proposes a particle method with a binning density estimator and presents numerical experiments, including cases with non-log-concave marginals and reverse-KL regularizers.","tokens_in":13069,"tokens_out":12562,"duration_ms":115410,"significance":"If the theoretical claims were correct, the paper would supply a new dynamic/particle approach to optimal transport plans with rigorous convergence guarantees, and it would resolve an open question from [16] on displacement convexity of the relaxed energy. The numerical experiments and the released code are useful and suggest that the particle algorithm may work well in practice. However, the central theoretical structure rests on a false identity (Lemma 3.4) and on a proof gap in the claimed displacement convexity (Lemma 3.1); the advertised convergence theorem is therefore not established. The numerical contribution is real, but the manuscript's main selling point is the theory, and that theory collapses as written.","major_comments":[{"comment":"Equation (3.3) is not a valid consequence of the Euler–Lagrange condition for the minimization of E(·, Λ) over couplings. The objective depends on ρ only through its marginals and the linear cost term, so the correct first-order system couples the Kantorovich potentials of the inner optimal-transport problem between b1[Λ] and b2[Λ] with the KL terms; the raw cost c does not appear pointwise over the whole product space. For c(x,y)=||x−y||² the minimizer b[Λ] is supported on the graph of an optimal map, whereas σ[Λ] is absolutely continuous, so (3.3) cannot hold as an identity of densities. Indeed, when μ=ν with quadratic cost, the unique minimizer is the diagonal coupling, whose marginals give b1 b2 = μ⊗μ, not σ[Λ]; the asserted identity would require e^{−||x−y||²/Λ} to be constant on the support. The formulas (3.5)–(3.6), the closed form (3.7) for V, and the covariance formula in Lemma 3.11 all derive from (3.3); hence Lemma 3.5, Lemma 3.7, Lemma 3.12, Proposition 3.13, and Corollary 3.14 are unsupported.","section":"§3.2, Lemma 3.4, Eq. (3.3)"},{"comment":"The proof of zero-displacement convexity of H1(ρ)=∫ρ1 log ρ1 does not justify the reduction to the marginal space. The sufficient condition (2.1) must hold for every optimal coupling γ between ρ and ρ̃ in the joint space, but the x-projection of γ is generally not an optimal coupling between ρ1 and ρ̃1; it is only some coupling. The known displacement convexity of the entropy on P2(R^d) gives the required monotonicity only along the Wasserstein geodesic between the marginals, not along arbitrary projected couplings. Without an additional argument showing the inequality along the projected coupling, the claimed displacement convexity of E in Theorem 3.2 and the exponential contraction in Theorem 3.3, which are advertised as resolving the open question from [16], are not proved.","section":"§3.1, Lemma 3.1"},{"comment":"The claimed equivalence between convergence of the second moment of b[Λn] and convergence of the second moment of σ[Λn] is false. The second moment of a joint distribution includes the cross term ∫ x·y db[Λn], which is determined by the optimal-transport coupling between the marginals b1[Λn] and b2[Λn]. The density σ[Λn] is a product (after normalization) and carries no information about this cross-term; its second moment is the sum of the two marginal second moments of the product measure. Convergence of ∫∥z∥² dσ[Λn] therefore does not imply convergence of ∫∥z∥² db[Λn]. This invalidates Lemma 3.5 and hence the tightness argument in Lemma 3.7, on which the continuity of the best-response map and the subsequent Danskin argument depend.","section":"§3.2, Lemma 3.5"}],"minor_comments":[{"comment":"The displayed contraction inequality appears to contain a typo: the right-hand side reads W(ρ(t), ρ(0)), but the natural estimate is W(ρ(t), ρ̄) ≤ e^{−(λ̃+Λλ)t} W(ρ(0), ρ̄).","section":"§3.1, Theorem 3.3"},{"comment":"In the proof, the case Λ∞ = +∞ is not handled carefully: the expression V(Λ∞) is undefined if Λ∞ is infinite, yet the proof writes V∞ := V(Λ∞). The authors need an argument that V(Λ(t)) has a limit as t→∞ before assigning V∞.","section":"§3.3, Proposition 3.13"},{"comment":"The relationship between the continuous-time two-timescale system (2.6) and the discrete particle update is not fully spelled out, particularly the role of the additive Gaussian noise with standard deviation 0.02√dt; a brief consistency argument or reference would help.","section":"§4, Eq. (4.1)"},{"comment":"The column headers 'KL reverse KL' and 'total KL' are not defined in the text; please clarify which of the four KL terms are combined in each column.","section":"Table 1"}],"recommendation":"reject","confidential_remarks":"The core theoretical lemma (Lemma 3.4) is demonstrably false, and the proof of Lemma 3.1 has a substantive gap; the convergence theorem is therefore not established. The numerical method may be independently useful, but the manuscript as submitted promises rigorous convergence and does not deliver it. A rejection is appropriate, possibly with an invitation to resubmit if the theory is reworked around the correct first-order conditions for the marginal-constrained problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nThe key thing to know: the advertised convergence theorem is not established. Lemma 3.4's identity (3.3) is false in the paper's own intended setting, and everything downstream of it—the explicit marginals, the expression for V, the covariance formula, the Danskin argument, and Corollary 3.14—is unsupported. The paper is an honest attempt, not a fake, but the central proof does not work.\n\nWhat is genuinely new and good: replacing the fixed regularizer of Liu–Sun–Zha with a dynamically tuned Lambda is a natural and useful idea. The particle scheme with binning density estimation is clearly faster than the earlier method, code is provided, and the empirical comparison of forward vs reverse KL is interesting. The paper correctly identifies an open question from [16] and tries to answer it.\n\nThe soft spots, in order of severity. First, (3.3) asserts b1(x)b2(y) = (1/Z)e^{-c(x,y)/Lambda} mu(x)nu(y). For c(x,y)=|x-y|^2, the right side has nonzero mixed partial derivative in x and y, while any product of a function of x and a function of y has zero mixed derivative. No absolutely continuous coupling can satisfy the equation. Worse, when mu=nu is Gaussian, the minimizer of E(.,Lambda) is the diagonal coupling, which is singular; the formula (3.3) is not the Euler-Lagrange condition at that point. This is not a minor gap; it is a false premise. Lemma 3.1 also has a hole: displacement convexity of int rho1 log rho1 as a functional of the marginal does not transfer to the joint without an argument, because the Wasserstein geodesic between two joint measures does not project to the Wasserstein geodesic between their marginals. So Theorem 3.2 is also unsupported as written.\n\nThe numerical experiments may still work, and the code is a useful starting point, but the paper's main advertised contribution—convergence with guarantees—is not proven. The paper honestly notes it will not beat Sinkhorn in speed, so the remaining value is the expressivity claim, which is plausible but not formally backed. The citation pattern is fine; leaning on [8] and [16] is standard.\n\nWho this is for: readers working on particle methods for OT may find the numerics worth a look. The theory section should not be taken at face value.\n\nRecommendation: send to peer review—the ideas deserve scrutiny—but expect reject or major revision. If the authors can fix or replace the best-response analysis, there may be a viable paper here; as it stands, the central theorem is unsupported.\n\nBest.","headline":"The paper's advertised convergence theorem is not established: Lemma 3.4's identity (3.3) is impossible for non-separable costs, and the particle method is left without the claimed guarantees.","tokens_in":13610,"tokens_out":7077,"would_cite":false,"duration_ms":67377,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","49J35","90C47"],"pacs":[],"model":"deepseek-v4-flash","headline":"Best-response dynamics with a self-tuning KL penalty converge to an optimal transport plan.","keywords":["optimal transport","Kantorovich problem","min-max","Wasserstein gradient flow","best-response dynamics","Danskin theorem","KL divergence","particle method"],"falsifier":"Take $\\mu=\\nu=\\mathcal{N}(0,1)$ and $c(x,y)=\\frac{1}{2}(x-y)^2$. The true optimal plan is the diagonal coupling $(x,x)$, so the best response satisfies $b_1[\\Lambda](x)b_2[\\Lambda](y)=\\mu(x)\\mu(y)$; substituting into (3.3) would require $e^{-(x-y)^2/(2\\Lambda)}/Z[\\Lambda]=1$ almost everywhere, which is false for any finite $\\Lambda$.","tokens_in":12486,"feed_emoji":"🎯","tokens_out":7323,"duration_ms":68222,"temperature":0.7,"pith_summary":"This paper treats the Kantorovich optimal transport problem as a zero-sum game: the coupling minimizes the transport cost plus KL penalties on its marginals, while the penalty weight is maximized. In the best-response limit, where the coupling always sits at the current minimizer, the paper proves that the dynamics for the penalty weight, driven by the marginal mismatch, converge so that the mismatch tends to zero and the best-response coupling converges to a true optimal transport plan. The result matters because it turns a conceptually clean min-max formulation into a convergence-certified particle method for approximating optimal couplings, without the factorization bias of Sinkhorn-type entropy regularization.","feed_headline":"Best-response flow provably reaches the optimal transport plan","feed_subtitle":"A two-player game with self-tuning KL weights converges to the optimal coupling.","key_machinery":"The load-bearing mechanism is the best response $b[\\Lambda]$ to the energy $E(\\rho,\\Lambda)=\\int\\int c\\,d\\rho+\\Lambda(\\mathrm{KL}(\\rho_1|\\mu)+\\mathrm{KL}(\\rho_2|\\nu))$, together with the Danskin identity $\\frac{d}{d\\Lambda}E_d(\\Lambda)=V(\\Lambda)=\\mathrm{KL}(b_1[\\Lambda]|\\mu)+\\mathrm{KL}(b_2[\\Lambda]|\\nu)$. Formula (3.3) asserts $b_1[\\Lambda](x)b_2[\\Lambda](y)=\\frac{1}{Z[\\Lambda]}e^{-c(x,y)/\\Lambda}\\mu(x)\\nu(y)$, giving explicit marginal formulas (3.5)-(3.6); those formulas turn $V$ into an explicit function of $\\Lambda$, make $V$ Lipschitz, and yield the $O(\\sqrt t)$ growth bound. Displacement convexity supplies the contraction of the fast Wasserstein flow, and the Danskin theorem closes the loop by justifying that $\\dot{\\Lambda}=V(\\Lambda)$ is the correct slow dynamics.","core_discovery":"The paper's central claim is Corollary 3.14: under the best-response dynamics, $b[\\Lambda(t)]\\to \\rho_*$ in $W_2$, where $\\rho_*$ minimizes the Kantorovich problem, and the regularized cost $E_d(\\Lambda)$ converges to the optimal cost $c_*$. The proof route is: the energy is displacement convex, the best-response map is continuous, Danskin's theorem identifies the derivative of $E_d$ with $V(\\Lambda)$, the penalty weight grows only as $O(\\sqrt t)$, the marginal mismatch $V$ must tend to zero, and continuity of the best response then pushes the coupling to an optimal plan.","pith_inferences":["If the product formula in Lemma 3.4 fails for finite $\\Lambda$, the convergence argument likely needs an additional term involving Kantorovich potentials: the Gibbs form would be replaced by $e^{-(c-\\phi-\\psi)/\\Lambda}$, and that change would alter the whole analysis.","The property $V(\\Lambda)\\to 0$ suggests a natural stopping criterion for the particle algorithm in practice: stop when the empirical KL mismatch between the coupling's marginals and the input marginals falls below a tolerance, avoiding the need to pre-specify a regularization schedule.","The observed asymmetry between the two KL directions hints at a broader design principle: the choice of KL direction interacts with which marginal is being flowed toward which target, and a directional displacement-convexity condition might predict when mixing directions outperforms a uniform choice."],"forward_implications":["The min-max flow offers a convergence-certified particle method for optimal transport plans: the particle scheme approximates the continuum flow, and the theory guarantees the continuum flow reaches the optimal coupling.","The regularization weight is not a user-chosen hyperparameter: the dynamics select it adaptively, and the square-root growth bound means the weight stays manageable over long runs.","Because the coupling is represented by particles rather than reweighted samples, approximants can have support different from the input marginal samples, unlike a Sinkhorn-style approximation.","The same dynamics can use reverse KL penalties, and the numerics show that mixing the two KL directions can reduce the empirical marginal error in the examples studied."],"supporting_citations":[{"why":"Supplies the particle-evolving method this work adapts and the cited existence result for the best-response minimizer for fixed $\\Lambda$.","marker":"[16]"},{"why":"Provides the general min-max Wasserstein gradient flow framework and the envelope-theorem argument that Theorem 3.9 adapts.","marker":"[8]"},{"why":"Furnishes the displacement-convexity criterion (Proposition 16.4) used in Theorem 3.2.","marker":"[22]"},{"why":"Gives the Wasserstein gradient flow, tightness, and existence facts used in the continuity and optimal-transport arguments.","marker":"[2]"},{"why":"Provides the Gamma-convergence background used in Proposition 3.6 to prove continuity of the best response.","marker":"[4]"},{"why":"Justifies the timescale-separated and best-response limit used for the dynamics (2.7).","marker":"[3]"}],"fun_headline_variants":["Two-player game provably finds optimal transport","Self-tuning KL yields provable optimal transport","Gradient flow provably reaches optimal coupling","Min-max game converges to optimal transport plan","Particle method with adaptive KL solves optimal transport"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof hinges on Lemma 3.4's identity $b_1[\\Lambda](x)b_2[\\Lambda](y)=\\frac{1}{Z[\\Lambda]}e^{-c(x,y)/\\Lambda}\\mu(x)\\nu(y)$ for every finite $\\Lambda$; if that identity fails, the explicit $V$, the growth bound, and the convergence theorem lose their footing.","fun_headline_variants_meta":{"raw":{"variants":["Two-player game provably finds optimal transport","Self-tuning KL yields provable optimal transport","Gradient flow provably reaches optimal coupling","Min-max game converges to optimal transport plan","Particle method with adaptive KL solves optimal transport"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1238,"prompt_tokens":774,"completion_tokens":464,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":390,"completion_tokens_details":{"reasoning_tokens":397}},"tokens_in":390,"tokens_out":464,"duration_ms":4686,"temperature":1.0,"reasoning_tokens":397,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:54:51.226192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $\\mu=\\nu=\\mathcal{N}(0,1)$ and $c(x,y)=\\frac{1}{2}(x-y)^2$. The true optimal plan is the diagonal coupling $(x,x)$, so the best response satisfies $b_1[\\Lambda](x)b_2[\\Lambda](y)=\\mu(x)\\mu(y)$; substituting into (3.3) would require $e^{-(x-y)^2/(2\\Lambda)}/Z[\\Lambda]=1$ almost everywhere, which is false for any finite $\\Lambda$.","supporting_citations":[{"cited_title":"Approximating the Optimal Transport Plan via Particle-Evolving Method","cited_arxiv_id":"2105.06088","evidence_quote":"Supplies the particle-evolving method this work adapts and the cited existence result for the best-response minimizer for fixed $\\Lambda$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Furnishes the displacement-convexity criterion (Proposition 16.4) used in Theorem 3.2."},{"cited_title":"Ambrosio, N","cited_arxiv_id":null,"evidence_quote":"Gives the Wasserstein gradient flow, tightness, and existence facts used in the continuity and optimal-transport arguments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Gamma-convergence background used in Proposition 3.6 to prove continuity of the best response."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Justifies the timescale-separated and best-response limit used for the dynamics (2.7)."}],"review_version":1}