{"id":"fd82b6a6-bd0f-425a-81d3-8063c3cc9ea1","arxiv_id":"2411.15814","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A nonlocal Ermentrout-Cowan-type equation, anisotropically rescaled, formally converges to horizontal mean curvature flow in Heisenberg and SE(2) geometries, with a coefficient theta that is fitted in the numerics.","lead":"This paper derives horizontal mean curvature flow in the Heisenberg group as the formal scaling limit of a nonlocal mean-field equation, extending a classical Euclidean result to a sub-Riemannian geometry used in visual cortex models. The result offers a new multiscale bridge from cell-level activity to curvature-based image processing, with numerical evidence on an exact solution that has characteristic points.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The explicit coefficient θ in (21) is missing a factor 1/2 from the quadratic Taylor term in (19)-(20); since §4.2 calibrates θ numerically rather than evaluating the formula, the stated interface speed can be wrong by a factor of two.","rationale":"The paper's central contribution is the formal derivation of the coefficient θ in the horizontal mean curvature flow. The missing 1/2 in (21) is the weakest point because it is an internal inconsistency in the main calculation: the factor is forced by the Taylor expansion and the solvability condition, not by any external regularity assumption. The numerical section cannot detect it because θ is fitted, not computed from the formula; the exact-solution comparison then uses the fitted value, so it tests the algorithm's self-consistency rather than formula (21). This is more load-bearing than the acknowledged comparison-principle assumption in Definition 4.2, since the latter is explicitly labeled an assumption and the paper's formal claim does not depend on it. The exact-solution comparison and the SE(2) extension give real independent support to the overall framework, and the error is localized and fixable. I therefore agree with the reader's CONDITIONAL verdict: the paper needs a corrected (21), a numerical evaluation of the formula rather than a fit, and a clear statement that Theorem 4.1 uses the corrected coefficient.","tokens_in":17263,"tokens_out":16363,"duration_ms":134685,"concrete_test":"Recompute θ from (19)-(20) for the kernel and β=1.2 used in §4.2, evaluating both (21) and the halved formula β/(2N) ∫∫ Ĵ(r1²+s²) m′(r)m′(r+r1) s² dr dr1 ds, using the instanton m for that kernel. Compare with the fitted θ=0.56561. If the halved expression matches the fit and the printed expression is twice the fit, correct (21) and state that §4.2 confirms the corrected formula only after replacing calibration by direct evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Most load-bearing: the printed formula (21) for θ is inconsistent with the expansion preceding it. In (18)-(20), the second-order term is ε/2 ∫ J(y) m′(r+ν·y) y^T H y dy, so after diagonalizing the Hessian (eigenvalues 0, k0) the coefficient R(r) is defined in §4.1 as R(r)=1/2 ∫ Ĵ(ŷ) m′(r+ŷ1) ŷ2^2 dŷ. The solvability condition for L(m1), obtained by multiplying V m′ = β(1−m²) k0 R + L(m1) by m′/(1−m²) and integrating, gives V N = β k0 ∫ R(r) m′(r) dr, hence θ = V/k0 = β/(2N) ∫ m′(r)m′(r+r1) Ĵ(r1²+s²) s² dr dr1 ds. Formula (21) as printed omits the 1/2. The numerical validation in §4.2 determines θ=0.56561 by linear regression on a shrinking cylinder; it never evaluates (21), so the factor error is invisible to the numerics. This is not a matter of missing regularity: if (21) is used as stated, the limiting interface speed is twice the coefficient derived from the paper's own expansion. The reader's concern about the unproved comparison hypothesis in Definition 4.2 is real but secondary to this internal inconsistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives a formal asymptotic expansion for the nonlocal mean-field equation (15) under the anisotropic Heisenberg rescaling and concludes that, away from characteristic points, the zero level set evolves by horizontal mean curvature flow with normal velocity V0 = θ k0, where θ is given by (21). The same formal derivation is transferred to SE(2) by freezing the Lie algebra. Numerical experiments with the Heisenberg heat kernel are reported, including a test against an exact axisymmetric solution with characteristic points from [22]. The paper explicitly presents the result as a regularization of Heisenberg mean curvature flow and does not claim a full rigorous convergence proof.","tokens_in":17568,"tokens_out":3330,"duration_ms":31737,"significance":"If the derivation and coefficient were correct, the paper would give a new multiscale link from a particle/mean-field model to sub-Riemannian curvature flow, with potential relevance to cortical models of V1. The paper also proposes a numerical scheme interpolating between an Ermentrout-Cowan type equation and a Bence-Merriman-Osher type algorithm, and it reproduces known exact solutions. These are valuable contributions. However, two load-bearing issues affect the central claim: the printed formula for θ appears to be off by a factor of 1/2 relative to the expansion in §4.1, and Theorem 4.1 asserts convergence although only a formal expansion is provided and the Euclidean trapping argument is not transferred to the Heisenberg setting.","major_comments":[{"comment":"Formula (21) is inconsistent with the expansion preceding it. In (20) the second-order term carries a factor ε/2, and immediately afterwards R(r) is defined as R(r) = (1/2) ∫ Ĵ(ŷ) m′(r+ŷ1) ŷ2² dŷ. Applying the solvability condition from Lemma 2.3, namely ⟨(1−m²)k0 R + L(m1), m′/(1−m²)⟩ = 0, gives V0 = k0 θ with θ = (β/(2N)) ∫ m′(r)m′(r+r1) Ĵ(r1²+s²) s² dr dr1 ds. The factor 1/2 is missing in the printed expression (21). The numerical validation in §4.2 determines θ by linear regression on a shrinking cylinder and never evaluates (21), so the numerics cannot detect the factor error. This is a load-bearing issue because Theorem 4.1 refers to 'θ as in (21)', and the stated interface speed would be twice the speed derived from the paper's own expansion.","section":"§4.1, Eq. (21)"},{"comment":"Theorem 4.1 states a locally uniform convergence result, but the proof supplied in §4.1 is only a formal asymptotic expansion with no error estimates. The Euclidean proof summarized in Section 2 relies on comparison principles and a trapping argument for biased solutions; in the Heisenberg setting the paper itself notes (Section 3 and after Definition 4.2) that comparison principles for horizontal mean curvature flow are only partially known. Definition 4.2 assumes the existence of smooth sets M_t^δ satisfying V0 = θκ0 + δ and a comparison principle, but this assumption does not by itself imply convergence of the solutions mε of (15) to the indicator of M_t. No argument is given that the zero level sets of mε are trapped between M_t^δ and M_t^{-δ}. Thus the theorem, as stated, is not proven by the material in §4.1; it should either be downgraded to a formal/conjectural statement or supplied with the missing convergence argument.","section":"Theorem 4.1 and Definition 4.2"}],"minor_comments":[{"comment":"The convergence statement says 'locally uniformly in RN \\ ∂Mt', but the setting is R3; the formula should read R3 (or H1), and the two limit values should be stated consistently with the conventions for Mt and its complement.","section":"Theorem 4.1"},{"comment":"The phrase 'M_0^δ has Hausdorff distance δ from M0' is imprecise: Hausdorff distance is between sets, and it is not clear whether the intended meaning is dist_H(M0^δ, M0) = δ or dist_H(∂M0^δ, ∂M0) = δ. Please clarify.","section":"Definition 4.2"},{"comment":"There are typographical errors: 'Ementrout-Cowan' in the abstract and 'mena curvature flow' in §4.2 should be 'Ermentrout-Cowan' and 'mean curvature flow', respectively.","section":"Abstract and §4.2"},{"comment":"Reference [25] is cited as 'Katzoulakis' in the bibliography; the correct spelling is 'Katsoulakis'.","section":"References"},{"comment":"The notation alternates between R(r) and 'hat R' in the splitting step before equation (21); please use a single symbol consistently, especially since the missing factor 1/2 is tied to the definition of R.","section":"§4.1 notation"}],"recommendation":"major_revision","confidential_remarks":"The factor 1/2 error in (21) and the unsupported convergence statement in Theorem 4.1 are both fixable in revision, but they are central to the paper's main claim. The formal expansion is a useful contribution, and the numerical study is suggestive, but the stated theorem and the printed coefficient should be corrected or reformulated before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before reading. First, the main deliverable is real: a clean formal asymptotic expansion showing that the zero level set of the nonlocal mean-field equation, under the Heisenberg anisotropic rescaling, moves by horizontal mean curvature flow away from characteristic points, with an explicit coefficient θ. The SE(2) extension via the freezing procedure is a natural and interesting addition. Second, the printed formula (21) for θ is inconsistent with the expansion preceding it. The second-order term in (18)-(20) carries an explicit 1/2, and the function R(r) is defined with that 1/2. The solvability condition then gives θ = (1/(2N)) ∫ m'(r)m'(r+r1) Ĵ(r1²+s²)s² dr dr1 ds. Formula (21) as stated omits the 1/2. This is an internal algebraic error, not a regularity gap. The numerics do not catch it because the code determines θ by linear regression on a shrinking cylinder and never evaluates (21). If (21) is used as printed, the interface speed is twice what the paper's own expansion implies. Adding the 1/2 restores consistency.\n\nWhat the paper does well: it carries the De Masi-Orlandi-Presutti-Triolo / Katsoulakis-Souganidis expansion into a genuinely different geometric setting, and the reduction of the Hessian term to the curvature k0 via the eikonal equation is clean. The numerical scheme interpolating between Ermentrout-Cowan and Bence-Merriman-Osher is a nice observation, and the test against the Ferrari-Liu-Manfredi exact solution with characteristic points is worthwhile even though it uses the fitted θ.\n\nSoft spots, in proportion. Theorem 4.1 states convergence, but no proof is given; the paper explicitly says comparison principles in H1 are only partially known, and Definition 4.2 makes the existence of approximating surfaces and a comparison principle an assumption. So the theorem is conditional on hypotheses the authors themselves flag as unproved. That is honest, but the statement goes beyond what the formal expansion establishes. The numerical validation is also somewhat circular: fitting θ on a cylinder and reproducing the sphere does not test the formula for θ. These are correctable weaknesses, not a broken central idea.\n\nThis paper is for anyone working on sub-Riemannian mean curvature flow, nonlocal reaction-diffusion equations, or geometric models of vision. It deserves serious refereeing, but it needs revision: fix the factor in (21), either prove Theorem 4.1 under stated assumptions or reframe it as a formal result, and ideally add a numerical check that evaluates (21) instead of fitting it.","headline":"A genuinely new formal derivation of horizontal MCF from a nonlocal mean-field equation in H1 and SE(2), but the printed θ formula is off by a factor 1/2 and Theorem 4.1 is only a conditional formal statement.","tokens_in":18122,"tokens_out":6449,"would_cite":true,"duration_ms":54723,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35K93","35R03","53C17","35B25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under the anisotropic rescaling of the Heisenberg group, a nonlocal mean-field equation has zero level sets that evolve by horizontal mean curvature flow $V_0=\\theta\\kappa_0$.","keywords":["mean curvature flow","Heisenberg group","nonlocal mean-field equation","asymptotic expansion","sub-Riemannian geometry","Carnot groups","roto-translation group","visual cortex model"],"falsifier":"Take a compact, non-axisymmetric initial surface in $\\mathbb{H}^1$ with a known non-characteristic point, solve (15) numerically for decreasing $\\varepsilon$, and measure the normal velocity of the zero level set at that point; the claim predicts $V_0=\\theta\\kappa_0$ with $\\theta$ from (21). A persistent deviation beyond discretization error as $\\varepsilon\\to0$ would falsify the scaling limit. A complementary check is whether any family $M_t^\\delta$ satisfying Definition 4.2 can exist up to the claimed time: if a surface develops a characteristic singularity before time $T$, the theorem's main hypothesis fails, and the formal expansion supplies no error estimate to replace it.","tokens_in":5,"feed_emoji":"🌀","tokens_out":14329,"duration_ms":184595,"temperature":0.7,"pith_summary":"This paper claims that, under the anisotropic dilations of the Heisenberg group, solutions of the nonlocal mean-field equation $\\varepsilon^{2}\\partial_t m_\\varepsilon=-m_\\varepsilon+\\tanh(\\beta\\int J^\\varepsilon(y^{-1}\\circ x)\\,m_\\varepsilon(y)\\,dy)$ have zero level sets that move, away from characteristic points, by horizontal mean curvature flow $V_0=\\theta\\kappa_0$, with $\\theta$ the explicit kernel-dependent constant in (21). The derivation is a formal asymptotic expansion: a one-dimensional instanton profile is placed along the signed distance to the moving surface, and the leading-order part of the expanded convolution produces a curvature term through the horizontal Hessian of the distance. The paper states the convergence as Theorem 4.1, conditional on a regularity and comparison assumption for $\\delta$-perturbed surfaces (Definition 4.2), and stresses that this assumption is needed because characteristic points can be singular even for smooth surfaces. The same calculation extends locally to the roto-translation group $\\mathrm{SE}(2)$, and numerical experiments reproduce an exact axisymmetric shrinking solution, including motion at characteristic points.","feed_headline":"Mean-field equation scales to Heisenberg curvature flow","feed_subtitle":"Zero level sets follow horizontal mean curvature as ε shrinks, linking cell models to visual-cortex geometry.","key_machinery":"Three ingredients carry the argument. First, the anisotropic dilation $\\delta_\\lambda(x_1,x_2,x_3)=(\\lambda x_1,\\lambda x_2,\\lambda^2 x_3)$ and the rescaled kernel $J^\\varepsilon(x)=\\varepsilon^{-4}J(\\delta_{1/\\varepsilon}x)$, with homogeneous dimension $Q=4$, make the nonlocal convolution compatible with the sub-Riemannian geometry. Second, the one-dimensional instanton $m$, the odd monotone solution of $-m+\\tanh(\\beta J*m)=0$ on the line, whose derivative $m'$ is the zero eigenfunction of the linearized operator $Lf=-f+(1-m^2)\\int J(s)f(r+s)\\,ds$; solvability of the first corrector is controlled by the $L^2(\\mu)$-orthogonality condition of Lemma 2.3. Third, the homogeneous Taylor expansion of the signed distance, whose symmetrized horizontal Hessian has eigenvalue $0$ along the horizontal normal and eigenvalue $\\kappa_0$ along the orthogonal horizontal direction, so the surviving quadratic term in the expanded convolution is exactly the curvature term. Projecting the curvature remainder onto $m'$ fixes the speed coefficient $\\theta$ in (21); the corresponding argument on the Lie algebra with local dilations gives the $\\mathrm{SE}(2)$ extension.","core_discovery":"The central claim is that the anisotropic scaling limit of a nonlocal mean-field equation provides a new regularisation and approximation of Heisenberg mean curvature flow. Concretely, if $m_\\varepsilon$ solves (15) with kernel $J^\\varepsilon(x)=\\varepsilon^{-4}J(\\delta_{1/\\varepsilon}x)$, then as $\\varepsilon\\to0$ the zero level set of $m_\\varepsilon$ tends, locally uniformly outside $\\partial M_t$, to the surface $\\partial M_t$ evolving by $V_0=\\theta\\kappa_0$, where $\\kappa_0=\\operatorname{div}_{\\mathbb{H}^1} n_0$ is the horizontal mean curvature and $\\theta$ is the $L^2(\\mu)$-projection constant in (21). The theorem is conditional: it assumes the existence of smooth surfaces $M_t^\\delta$ without characteristic points that satisfy $V_0=\\theta\\kappa_0+\\delta$ classically and obey a comparison principle, with Hausdorff distance to $M_t$ vanishing as $\\delta\\to0$. The paper presents the expansion as a formal argument at non-characteristic points and offers numerical evidence that the approximation persists near characteristic points, where the flow itself is not classically defined. By freezing the Lie algebra, the same expansion is transferred to the roto-translation group $\\mathrm{SE}(2)$.","pith_inferences":["A direct consequence the authors leave implicit is that the numerical scheme can be tuned through the kernel and inverse temperature to produce a prescribed speed coefficient $\\theta$; measuring $\\theta$ from (21) for a compactly supported kernel would be a cheap experimental check.","The argument suggests a route to a fully rigorous convergence theorem if a comparison principle for Heisenberg mean curvature flow were available in the needed generality; the paper identifies exactly this gap when it labels Definition 4.2 an assumption.","The same formal machinery should apply to other step-two Carnot or sub-Riemannian structures by replacing the Heisenberg dilation with the appropriate homogeneous dilation, with characteristic points again the main obstruction."],"forward_implications":["For small but finite $\\varepsilon$, the nonlocal equation is globally defined and gives a regularisation of Heisenberg mean curvature flow rooted in a multiscale derivation.","Away from characteristic points, the effective speed is $V_0=\\theta\\kappa_0$, so the coefficient $\\theta$ from (21) determines the time scale of the limiting flow for a given kernel and inverse temperature.","The same derivation applies to the roto-translation group $\\mathrm{SE}(2)$, so the nonlocal equation approximates the sub-Riemannian curvature flow used in a cortical model of image completion and denoising.","The numerical scheme interpolates between a nonlocal mean-field evolution and a diffusion-concentration thresholding algorithm; in the appropriate parameter limits it reduces to those two classical procedures.","The exact axisymmetric shrinking solution with characteristic points is reproduced numerically, indicating that the approximation is meaningful even where the curvature flow is classically undefined."],"supporting_citations":[{"why":"Supplies the Euclidean nonlocal scaling-limit method: the instanton profile, the asymptotic expansion, and the spectral structure of the linearized operator.","marker":"[15]"},{"why":"Provides the comparison-principle and viscosity framework for nonlocal mean-field convergence to mean curvature that the paper adapts to the Heisenberg setting.","marker":"[25]"},{"why":"Gives the Hessian of the signed distance in $\\mathbb{H}^1$ and the identity $\\Delta_{\\mathbb{H}^1} d=\\kappa_0$ at non-characteristic points, from which the curvature term in the expansion is read.","marker":"[2]"},{"why":"Supplies the Taylor expansion of functions on homogeneous groups used to expand the convolution term under Heisenberg dilations.","marker":"[6]"},{"why":"Used for the eikonal equation satisfied by the Carnot–Carathéodory distance, giving the zero eigenvalue of the horizontal Hessian along the normal.","marker":"[29]"},{"why":"Provides the comparison principle and the exact axisymmetric shrinking solution used as a numerical benchmark, including behaviour at characteristic points.","marker":"[22]"},{"why":"Provides the Lie-algebra freezing construction that transfers the Heisenberg computation locally to $\\mathrm{SE}(2)$.","marker":"[33]"},{"why":"Explains the tangent-cone relation between the Heisenberg group and $\\mathrm{SE}(2)$, which justifies transferring the local computation from one group to the other.","marker":"[27]"},{"why":"Defines the sub-Riemannian cortical model on the roto-translation group whose curvature flow is the target application of the extension.","marker":"[14]"}],"fun_headline_variants":["Mean-field limit yields Heisenberg curvature flow","Anisotropic scaling gives Heisenberg mean curvature flow","From cell equations to Heisenberg curvature flow","New route to Heisenberg mean curvature via mean field","Heisenberg curvature flow from nonlocal mean-field scaling"],"cache_read_input_tokens":20224,"weakest_assumption_plain":"The load-bearing premise is that, for small $\\delta$, the evolving surface can be surrounded by smooth surfaces without characteristic points that move with speed $\\theta\\kappa_0+\\delta$, obey a comparison principle, and converge to the true surface as $\\delta\\to0$; the paper explicitly labels this an assumption because characteristic points can be singular even for smooth surfaces.","fun_headline_variants_meta":{"raw":{"variants":["Mean-field limit yields Heisenberg curvature flow","Anisotropic scaling gives Heisenberg mean curvature flow","From cell equations to Heisenberg curvature flow","New route to Heisenberg mean curvature via mean field","Heisenberg curvature flow from nonlocal mean-field scaling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1365,"prompt_tokens":1058,"completion_tokens":307,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":674,"completion_tokens_details":{"reasoning_tokens":249}},"tokens_in":674,"tokens_out":307,"duration_ms":3420,"temperature":1.0,"reasoning_tokens":249,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:53:54.769627+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a compact, non-axisymmetric initial surface in $\\mathbb{H}^1$ with a known non-characteristic point, solve (15) numerically for decreasing $\\varepsilon$, and measure the normal velocity of the zero level set at that point; the claim predicts $V_0=\\theta\\kappa_0$ with $\\theta$ from (21). A persistent deviation beyond discretization error as $\\varepsilon\\to0$ would falsify the scaling limit. A complementary check is whether any family $M_t^\\delta$ satisfying Definition 4.2 can exist up to the claimed time: if a surface develops a characteristic singularity before time $T$, the theorem's main hypothesis fails, and the formal expansion supplies no error estimate to replace it.","supporting_citations":[{"cited_title":"De Masi, E","cited_arxiv_id":null,"evidence_quote":"Supplies the Euclidean nonlocal scaling-limit method: the instanton profile, the asymptotic expansion, and the spectral structure of the linearized operator."},{"cited_title":"Communications in Mathematical Physics, vol 169, vol 1, pp","cited_arxiv_id":null,"evidence_quote":"Provides the comparison-principle and viscosity framework for nonlocal mean-field convergence to mean curvature that the paper adapts to the Heisenberg setting."},{"cited_title":"Arcozzi, F.Ferrari","cited_arxiv_id":null,"evidence_quote":"Gives the Hessian of the signed distance in $\\mathbb{H}^1$ and the identity $\\Delta_{\\mathbb{H}^1} d=\\kappa_0$ at non-characteristic points, from which the curvature term in the expansion is read."},{"cited_title":"Bonfiglioli","cited_arxiv_id":null,"evidence_quote":"Supplies the Taylor expansion of functions on homogeneous groups used to expand the convolution term under Heisenberg dilations."},{"cited_title":"Serra Cassano Surface measures in Carnot-Carath` eodory spaces Calc","cited_arxiv_id":null,"evidence_quote":"Used for the eikonal equation satisfied by the Carnot–Carathéodory distance, giving the zero eigenvalue of the horizontal Hessian along the normal."},{"cited_title":"Ferrari, Q","cited_arxiv_id":null,"evidence_quote":"Provides the comparison principle and the exact axisymmetric shrinking solution used as a numerical benchmark, including behaviour at characteristic points."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Lie-algebra freezing construction that transfers the Heisenberg computation locally to $\\mathrm{SE}(2)$."},{"cited_title":"Le Donne, Lecture notes on subRiemannian geometry - Carnot- Carath´ eodory spaces from the Lie Group viewpoint","cited_arxiv_id":null,"evidence_quote":"Explains the tangent-cone relation between the Heisenberg group and $\\mathrm{SE}(2)$, which justifies transferring the local computation from one group to the other."},{"cited_title":"Citti and A","cited_arxiv_id":null,"evidence_quote":"Defines the sub-Riemannian cortical model on the roto-translation group whose curvature flow is the target application of the extension."}],"review_version":1}