{"id":"0c0a1444-6838-4e80-a76e-46a45d122e03","arxiv_id":"2501.03459","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For general p>1, discrete gradient flows of a particle-based energy on the L^p-Wasserstein space converge to continuum gradient flows, covering doubly nonlinear diffusion in one dimension.","lead":"This paper proves that a particle method for approximating gradient flows on the L^p-Wasserstein space converges to the continuum gradient flow, which in one dimension is a doubly nonlinear diffusion equation. It extends the p=2 result of Carrillo, Patacchini, Sternberg, and Wolansky to all p>1 on bounded intervals.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4 is invoked with squared metric derivatives, but Proposition 2 verifies only a p-th power condition; for p != 2 the chain of implications is incomplete.","rationale":"The paper makes a serious attempt to extend Carrillo-Patacchini-Sternberg-Wolansky from p=2 to general p, and the reader's conditional verdict is reasonable: the C3/Hypothesis 2 issue is real, the proofs of Lemmas 3-5 are omitted, and the interpolation argument in Proposition 4 is delicate. My concern is different and more elementary: the proof of Theorem 3 depends on a Gamma-convergence-of-gradient-flows theorem whose stated hypotheses do not match the p-setting in which it is used. Proposition 2 verifies liminf of p-th powers of metric derivatives, while Theorem 4 as printed asks for liminf of squares; for p != 2 neither condition follows from the other. The paper also concludes with L^p convergences in (8), which is inconsistent with a theorem whose hypotheses are squared. This is not a question of the central claim being false; the result may well be true and provable by a p,q version of Serfaty's argument. But as written, the logical bridge from the discrete estimates to the continuum gradient flow is missing for the main range p > 1. Because this is an unproved invocation rather than a demonstrated falsehood, the appropriate verdict remains CONDITIONAL rather than REJECT or UNVERDICTED.","tokens_in":942,"tokens_out":1391,"duration_ms":217200,"concrete_test":"Compare the statement of Theorem 4 with the original theorem in Serfaty [17]. If Serfaty's theorem is stated only with squared metric derivatives and squared slopes, then the present Theorem 4 is not a theorem for p != 2. To settle the concern, supply a self-contained proof of the p,q analogue, replacing (C1) by liminf integral |mu'_N|^p >= integral |rho'|^p and adding liminf |partial E_N|^q >= |partial E|^q in (C3), and verify that Proposition 2 and Proposition 4 prove this analogue. As a quick concrete check, take p = 3 and note that the printed Proposition 2 does not verify the squared condition used in Theorem 4.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Serfaty's theorem is the mechanism that converts the discrete estimates (C1)-(C3) into the continuum gradient flow and the convergences (8). As printed, Theorem 4 requires (C1) liminf_N integral_0^t |mu'_N|^2 ds >= integral_0^t |rho'|^2 ds and (C3) liminf_N |partial E_N|(mu_N(t)) >= |partial E|(rho(t)), with no exponent in (C3). The flows in this paper are p-curves of maximal slope (Definition 5), and Proposition 2 establishes only the p-th power statement liminf_N integral_0^t |mu'_N|^p ds >= integral_0^t |rho'|^p ds. On a bounded time interval, a liminf lower bound on L^p norms does not control the liminf of L^2 norms when p != 2, and the conclusion (8) is itself in L^p. Therefore, unless a p,q version of Serfaty's theorem is stated and proved, with (C1) in p-th powers and (C3) as liminf |partial E_N|^q >= |partial E|^q, the proof of Theorem 3 is not logically closed for the paper's advertised range p > 1. This gap is independent of the C3/Hypothesis 2 interpolation issue identified by the reader; even a fully correct C3 does not repair the exponent mismatch in the invoked theorem.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper adapts the particle method of Carrillo, Patacchini, Sternberg, and Wolansky to gradient flows on the L^p-Wasserstein space over R, for p>1. The continuum energy is discretized by replacing each empirical measure with a piecewise-constant density on nonoverlapping balls centered at the particles, and the discrete gradient flow is defined as a p-curve of maximal slope in a weighted l^p norm. The main theorem (Theorem 3) claims that, for well-prepared initial data in a class G(Omega), the discrete gradient flows converge, up to a subsequence, narrowly to an absolutely continuous curve, and that on a bounded interval, under an additional scaling condition on B'', the limit is a continuum gradient flow in the sense of a p-curve of maximal slope, with convergence of metric derivatives, energies, and slopes. For p=2 or for power-law energies, convergence of the whole sequence to the doubly nonlinear diffusion equation is claimed. The proof follows Serfaty's Gamma-convergence-of-gradient-flows framework and attempts to verify conditions (C1)-(C3).","tokens_in":26414,"tokens_out":11214,"duration_ms":106175,"significance":"If the result is established, it would be a useful extension of the one-dimensional particle method to p-curves of maximal slope and to doubly nonlinear diffusion equations. The discrete energy is a direct, parameter-free discretization of the continuum energy, and the claimed limit statements are concrete and falsifiable. However, as the manuscript stands, the central proof is not closed: the invoked Serfaty theorem is stated with the wrong exponents, and several load-bearing lemmas and the recovery-sequence construction are delegated to [3] without proof. The significance is therefore conditional on a substantial revision.","major_comments":[{"comment":"Theorem 4, as stated, invokes Serfaty's theorem with condition (C1) involving squared metric derivatives and condition (C3) with no exponent, while the discrete flows of Definition 9 are p-curves of maximal slope. Proposition 2 proves only the p-th-power inequality liminf_N ∫_0^t |μ'_N|^p ds ≥ ∫_0^t |ρ'|^p ds. For p≠2, a lower bound on L^p norms over a bounded interval does not imply the corresponding L^2 lower bound, and the slopes in (C3) are not stated in the correct conjugate exponent. Moreover, the convergence of slopes in (8) is stated in L^p, although for a p-curve of maximal slope the natural integrability of the slope is L^q with 1/p+1/q=1. Since Theorem 4 is the mechanism that converts (C1)-(C3) into the continuum gradient flow, the proof of Theorem 3 is not logically closed for the advertised range p>1. The manuscript should either state and prove a p,q version of Serfaty's theorem with (C1) in p-th powers and (C3) in q-th powers, or restrict the main theorem to p=2.","section":"Section 3, Theorem 4 and Proposition 2"},{"comment":"Lemmas 3, 4, and 5 are load-bearing: Lemma 3 gives the lower bound |η_i| ≥ |σ_i - σ_{i+1}| used in Lemma 6, Lemma 4 fixes the behavior of the boundary particles, and Lemma 5 provides the uniform spacing bounds E_1/N ≤ Δx_i ≤ E_2/N used throughout the proof of Proposition 4 and in the interpolation. The paper states that these are obtained by the same strategy as in [3] and omits all details. Because the discrete gradient flow here is taken with respect to a weighted l^p norm, the subdifferential is different from the p=2 case, and the adaptation is not automatic. These proofs must be supplied.","section":"Section 3.2, Lemmas 3-5"},{"comment":"The proof of (C3) is not self-contained. It asserts without proof that the interpolation ρ̃_N(t) converges narrowly to ρ(t), that the normalization constant A_N converges to 1, and that the limit passage (14)-(15) is valid after using Hypothesis 2. The step relating liminf_N I_p(ρ̂_N) and liminf_N I_p(ρ̃_N) via the scaling inequality B''(αx) ≥ φ(α)B''(x) is only sketched, and the finiteness assumption in Lemma 6 is not justified for every t∈[0,T]. Since (C3) is the last condition needed to apply Theorem 4, this is a major gap.","section":"Section 3.2, proof of Proposition 4"},{"comment":"Theorem 5 states the Gamma-convergence of the discrete energies, but its proof consists entirely of 'we follow the same strategy as in [3]' and 'we omit details here.' If this theorem is part of the paper's claims, it needs a proof; if it is only meant as a reference to known results, it should be stated as such. Moreover, the construction of recovery sequences is needed to make the well-preparedness assumption in Definition 11 non-vacuous, so this omission affects the interpretation of the main theorem.","section":"Section 4, Theorem 5"}],"minor_comments":[{"comment":"After extracting a subsequence for the weak convergence of |μ'_N| in L^p, the notation reverts to the full sequence in equation (10); the subsequence should be relabeled consistently throughout the proof.","section":"Section 3.1, proof of Proposition 2"},{"comment":"The displayed formula for the Fisher information of ρ̂_N appears to contain an exponent inconsistency: with the definition in Lemma 1 the denominator should involve ρ^{-1}, not ρ^{1-3p}; this should be checked.","section":"Section 3.2, proof of Proposition 4"},{"comment":"The condition 'minsupp ρ ρ > 0' should be written as min over the support of ρ, for example min_{ρ>0} on supp ρ, to avoid ambiguity.","section":"Section 3, Definition 10"},{"comment":"The uniqueness references for the p-heat equation and the Leibenson equation should be accompanied by precise hypotheses matching the class G(Ω); as written it is unclear whether the cited uniqueness theorems apply to the weak solutions obtained in Theorem 3.","section":"Section 3, Remark 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a direct extension of [3] and currently reads as a draft: the central theorem depends on a misstated external theorem and on several lemmas whose proofs are omitted. I would not reject outright because the strategy is plausible and the gaps are localizable, but the author should be asked to provide full proofs or precise references for the p,q version of Serfaty's theorem and for the omitted lemmas. I would also ask the editor to verify the relationship to [17]: the theorem as quoted in Section 3 does not appear to match the p=2 statement in Serfaty's paper, and this mismatch is not a presentation issue but a logical gap in the proof of the main theorem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"New here is a genuine extension: the particle method of Carrillo–Patacchini–Sternberg–Wolansky, previously for p=2, is pushed to general p>1 and applied to doubly nonlinear diffusion in 1D. The energy discretization, the slope computations, and especially Lemma 6 and the renormalized interpolation in Proposition 4 are real new estimates. The paper also credits [3] honestly and has no fitted parameters or circularity.\n\nThe soft spot is not the originality; it is the logic of the convergence proof. Theorem 4, as stated, is Serfaty's theorem with (C1) in squared metric derivatives. The flows here are p-curves of maximal slope (Definition 5), and Proposition 2 verifies only the p-th power version of (C1): liminf ∫|μ'_N|^p ≥ ∫|ρ'|^p. On a fixed time interval, that liminf in L^p does not imply the L^2 liminf when p≠2, and the conclusion (8) is written in L^p. So unless there is a p,q version of Serfaty's theorem stated and proved (with (C1) in p-th powers and (C3) in q-th powers), Theorem 3 is not logically closed for the advertised range p>1. This is independent of the C3/Hypothesis 2 interpolation issue; it is the load-bearing step.\n\nThere are also several important lemmas not proved in the text: Lemmas 3–5 and the details of Theorem 5 are dismissed with 'same strategy as in [3]'. That may be acceptable for a follow-up paper, but here p≠2 changes the inequalities, so the reader cannot check them without redoing [3] by hand. Lemma 6 is new and treated; Proposition 4 is sketched at a level that is plausible but has a confusing normalization (ν_N = A_N \\tilde ρ_N after \\tilde ρ_N was already normalized).\n\nBottom line: the result is probably true and the paper is worth engaging with, but in its current form the main theorem is not fully proved for p≠2 because of the Serfaty exponent mismatch. The author should be asked to either prove a p,q Gamma-convergence theorem or show how the p-th power liminf reduces to the squared one. Send it to referees, with a request for major revision; readers in numerical analysis and optimal transport will get value once the gap is fixed.","headline":"A genuine p>1 extension of the CPSW particle method, with new slope estimates, but the proof as written invokes Serfaty's theorem with squared metric derivatives while verifying only p-th powers, so the main theorem is not logically closed for p≠2.","tokens_in":26945,"tokens_out":7423,"would_cite":false,"duration_ms":73074,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","35K55","65M75"],"pacs":[],"model":"deepseek-v4-flash","headline":"Particle method is proved convergent for L^p-Wasserstein gradient flows.","keywords":["L^p-Wasserstein space","gradient flows","particle method","doubly nonlinear diffusion equation","curve of maximal slope","Gamma-convergence","porous medium equation"],"falsifier":"Choose a one-dimensional energy density $B$ that satisfies Hypothesis 1 but violates the dilation inequality of Hypothesis 2, prepare particles with spacing bounds as in Definition 11, and compute the discrete slopes $|\\partial\\mathcal{E}_N|(\\mu_N)$ together with the continuum slope of the interpolation $\\tilde{\\rho}_N$. If the ratio of these quantities does not tend to a limit at least 1 as $N\\to\\infty$, or if the normalization constant in $\\tilde{\\rho}_N$ diverges, condition (C3) fails and the conclusion of Theorem 3 would not hold for that $B$.","tokens_in":25913,"feed_emoji":"📐","tokens_out":17445,"duration_ms":137178,"temperature":0.7,"pith_summary":"This paper proves that a particle approximation to gradient flows on the L^p-Wasserstein space over the real line converges to the continuum gradient flow. The continuum limit is the doubly nonlinear diffusion equation, of which the porous medium equation and the $p$-Laplacian heat equation are special cases. The proof preserves the steepest-descent structure at the particle level by discretizing the energy through nonoverlapping intervals centered on the particles, and then passes the discrete flows to the continuum limit. On bounded intervals, with an extra growth condition on the energy density $B$, the metric derivatives, the energies, and the slopes of the discrete flows converge to their continuum counterparts. If $p=2$ or the energy is power-law, the whole sequence of particle flows converges rather than only a subsequence.","feed_headline":"Particle method proved convergent for L^p-Wasserstein gradient flows","feed_subtitle":"A particle cloud's discrete steepest descent reproduces the continuum doubly nonlinear diffusion equation in 1D.","key_machinery":"The load-bearing object is the discrete energy $\\mathcal{E}_N$ built from nonoverlapping balls centered at the $N$ particles: in one dimension the balls are intervals, so $\\mathcal{E}_N(\\mu_N)=\\sum_i |B_i| B(1/(N|B_i|))=(1/N)\\sum_i h(N\\Delta x_i)$ with $h(x)=xB(1/x)$. This is exactly the continuum energy of the piecewise-constant interval density, and it gives the particle system a gradient-flow structure at the discrete level. To pass to the limit, the proof builds an interpolation $\\tilde{\\rho}_N$ from the affine interpolation of the dual variables $\\sigma_i=-N h'(N\\Delta x_i)$ by inverting through $h'$, and shows that the discrete slope controls the continuum Fisher-type information of this interpolation. Hypothesis 2 is used at the last step, to control the effect of the normalization constant in $\\tilde{\\rho}_N$ when comparing discrete and continuum slopes.","core_discovery":"The central claim, Theorem 3, is that if the energy density $B$ satisfies the convexity and doubling assumptions in Hypothesis 1 and the initial particle configurations are well-prepared for a continuum density $\\rho$, then the discrete gradient flows $\\mu_N$ converge narrowly, up to a subsequence, to an absolutely continuous curve $\\rho$ in $P_p(\\Omega)$. When $\\Omega=[-\\ell,\\ell]$ and Hypothesis 2 holds, that is $B''>0$ together with the dilation comparison $B''(\\alpha x)\\ge \\varphi(\\alpha)B''(x)$, the limit is a continuum gradient flow of the energy, and the discrete metric derivatives, energies, and local slopes converge to the continuum ones in the sense of (8). For $p=2$ or for power-law energies, uniqueness of the continuum solution upgrades the convergence to the full sequence, and the limit solves the doubly nonlinear diffusion equation in one dimension.","pith_inferences":["Because the proof forces neighboring particle spacings to equalize as $N$ grows, the maximum ratio $|\\Delta x_{i+1}/\\Delta x_i-1|$ is a natural numerical diagnostic: for well-prepared initial data it should decay to zero, and its decay rate may track the practical accuracy of the method.","Hypothesis 2 is a mild self-similarity condition on $B$; if it is close to being violated, the slope comparison suggests that non-power-law energy densities will converge more slowly, a statement the paper does not make explicitly.","An analogous result in more than one dimension would need a different cell construction, because nonoverlapping balls leave gaps and the one-dimensional ordering of particles is used essentially in the interpolation argument."],"forward_implications":["The discrete gradient-flow equations for the particle positions are a faithful approximation of the doubly nonlinear diffusion equation in one dimension.","When the theorem applies, discrete energies, metric derivatives, and local slopes pass to the continuum limit, so the approximation preserves the energy-dissipation structure of the flow.","The method extends the variational particle approximation from the $L^2$-Wasserstein setting to every $p>1$, covering $p$-Laplacian heat and porous-medium type flows.","On bounded intervals with $p=2$ or with power-law energies, uniqueness of the continuum solution turns subsequential convergence into convergence of the entire sequence of particle flows."],"supporting_citations":[{"why":"It introduces the particle method and the ball discretization of the energy that this paper extends from $p=2$ to general $p>1$.","marker":"[3]"},{"why":"It establishes the variational formulation of diffusion equations as Wasserstein gradient flows, the conceptual starting point for the paper.","marker":"[10]"},{"why":"It supplies the Gamma-convergence criterion for metric gradient flows used as Theorem 4 to transfer convergence from the discrete to the continuum flow.","marker":"[17]"},{"why":"It provides the theory of curves of maximal slope, local slopes, and Wasserstein gradient flows used throughout the proof.","marker":"[1]"},{"why":"It provides the displacement convexity of the energy that underlies Hypothesis 1 and makes local slopes act as upper gradients.","marker":"[13]"},{"why":"It identifies the Leibenson equation as the target doubly nonlinear diffusion equation.","marker":"[12]"},{"why":"It establishes uniqueness for the doubly nonlinear equation, which is used to promote subsequential convergence to whole-sequence convergence.","marker":"[9]"},{"why":"It establishes uniqueness for the $p$-heat flow, another special case where full-sequence convergence is obtained.","marker":"[11]"}],"fun_headline_variants":["Particle method converges on L^p-Wasserstein flows","1D particle method yields doubly nonlinear diffusion","Discrete gradient flows converge in L^p-Wasserstein","Particle scheme proven for L^p-Wasserstein gradients","Convergence of particle method for L^p-Wasserstein flows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of the slope lower-semicontinuity condition (C3) requires the energy density $B$ to satisfy $B''>0$ with a dilation scaling inequality $B''(\\alpha x)\\ge \\varphi(\\alpha)B''(x)$, and it requires the initial particle spacings to be bounded between two constants divided by $N$; if either condition fails, the key comparison between discrete and continuum slopes has no justification.","fun_headline_variants_meta":{"raw":{"variants":["Particle method converges on L^p-Wasserstein flows","1D particle method yields doubly nonlinear diffusion","Discrete gradient flows converge in L^p-Wasserstein","Particle scheme proven for L^p-Wasserstein gradients","Convergence of particle method for L^p-Wasserstein flows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000483,"raw_usage":{"total_tokens":2320,"prompt_tokens":811,"completion_tokens":1509,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":1425}},"tokens_in":427,"tokens_out":1509,"duration_ms":49174,"temperature":1.0,"reasoning_tokens":1425,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:51:34.594833+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a one-dimensional energy density $B$ that satisfies Hypothesis 1 but violates the dilation inequality of Hypothesis 2, prepare particles with spacing bounds as in Definition 11, and compute the discrete slopes $|\\partial\\mathcal{E}_N|(\\mu_N)$ together with the continuum slope of the interpolation $\\tilde{\\rho}_N$. If the ratio of these quantities does not tend to a limit at least 1 as $N\\to\\infty$, or if the normalization constant in $\\tilde{\\rho}_N$ diverges, condition (C3) fails and the conclusion of Theorem 3 would not hold for that $B$.","supporting_citations":[{"cited_title":"A., Patacchini, F","cited_arxiv_id":null,"evidence_quote":"It introduces the particle method and the ball discretization of the energy that this paper extends from $p=2$ to general $p>1$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It establishes the variational formulation of diffusion equations as Wasserstein gradient flows, the conceptual starting point for the paper."},{"cited_title":"Discrete Contin","cited_arxiv_id":null,"evidence_quote":"It supplies the Gamma-convergence criterion for metric gradient flows used as Theorem 4 to transfer convergence from the discrete to the continuum flow."},{"cited_title":"Birkh ¨auser Basel, 2005","cited_arxiv_id":null,"evidence_quote":"It provides the theory of curves of maximal slope, local slopes, and Wasserstein gradient flows used throughout the proof."},{"cited_title":"J.: A convexity principle for interacting gases, Adv","cited_arxiv_id":null,"evidence_quote":"It provides the displacement convexity of the energy that underlies Hypothesis 1 and makes local slopes act as upper gradients."},{"cited_title":"izv akad","cited_arxiv_id":null,"evidence_quote":"It identifies the Leibenson equation as the target doubly nonlinear diffusion equation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It establishes uniqueness for the doubly nonlinear equation, which is used to promote subsequential convergence to whole-sequence convergence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It establishes uniqueness for the $p$-heat flow, another special case where full-sequence convergence is obtained."}],"review_version":1}