{"id":"d87479b3-cea2-4140-ab6e-d88c596ab49b","arxiv_id":"2501.00389","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The accelerated Allen-Cahn equation formally converges to the hyperbolic interface law ∂_t v = (1-v^2)(h-αv), and a large-step FISTA discretization empirically accelerates Ginzburg-Landau minimization.","lead":"This paper studies a momentum-based ('accelerated') version of the Allen-Cahn equation for minimizing diffuse perimeter energies, and derives a formal geometric law for its interfaces. It also reports that a FISTA-type discretization with large steps can converge faster than gradient descent on planar curves, triply periodic minimal surfaces, and semi-supervised graph classification tasks.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma A.2 is internally inconsistent: the variational proof yields the opposite sign of the corrector equation (10), which is the sign needed for the O(1/ε) cancellation in §3.2.2. If uncorrected, equation (3) does not follow even formally.","rationale":"The paper's central claim is the formal singular limit (3), and the reader's CONDITIONAL verdict is appropriate. My check focused on whether the formal derivation is internally consistent. The refined ansatz in §3.2.2 is algebraically coherent if the corrector satisfies Eq. (10) with the plus sign. However, Lemma A.2's proof as written establishes the opposite sign: the displayed variational functional F has a minus sign in front of φ'+2xφ'', so its Euler-Lagrange equation is −ψ''+W''ψ=φ'+2xφ'', which is not Eq. (10). Step 3 of the proof even states that Euler-Lagrange equation, and the numerical construction solves a homogeneous ODE that also matches the opposite sign. This is a concrete internal inconsistency, not merely a missing rigorous convergence proof. If the wrong-sign corrector is used, the O(1/ε) cancellation in §3.2.2 fails, so Eq. (3) is not derived even at the formal level. The circle experiment in §3.2.3 cannot rescue the issue because for a circle the normal velocity is spatially constant, making ω constant and all tangential-gradient and corrector-sensitive terms trivial; it only validates the ODE (6). This concern is specific and checkable, and it does not affect the paper's sound elementary results (energy decrease, finite speed, conditional critical-point convergence). Because the flaw is likely a fixable sign error in the proof rather than a refutation of the underlying idea, the reader's CONDITIONAL verdict remains appropriate; no verdict change is needed.","tokens_in":39998,"tokens_out":19079,"duration_ms":185924,"concrete_test":"Re-derive the O(1/ε) balance of the refined ansatz in §3.2.2 using the corrector equation actually obtained from the Euler-Lagrange equation of the displayed functional F in Lemma A.2 (opposite sign from (10)). If the cancellation identity fails, correct Lemma A.2 (or the sign of F) and recompute the resulting interface law. Separately, solve the inhomogeneous ODE (10) numerically for the same double-well potential used in Figure 14, verify ψ''−W''ψ=φ'+2xφ'', and check whether the proof's construction (homogeneous ODE with ψ'(0)=0) reproduces it; if it does not, the printed existence proof needs revision before Eq. (3) can be cited.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The formal singular limit (3) hinges on the refined ansatz in §3.2.2. There, the O(1/ε) terms reduce to φ'(r/ε)B/ε + (v∂tω)(−ψ''+W''ψ+2xφ'')/ε with x=r/ε, and cancellation requires the corrector equation ψ''−W''ψ=φ'+2xφ'' — exactly Eq. (10). Lemma A.2, however, states (10) but proves a different statement. The functional F displayed there is 1/2∫(|ψ'|²+W''(ϕ)ψ²)dx − ∫ψ(φ'+2xφ'')dx, whose Euler-Lagrange equation is −ψ''+W''ψ=φ'+2xφ'', i.e. ψ''−W''ψ=−(φ'+2xφ''). Step 3 of the proof explicitly writes this EL equation, and the numerical construction solves a homogeneous ODE ψc''=W''ψc with ψc'(0)=0, which likewise corresponds to the opposite sign. If the corrector satisfies the EL sign instead of (10), the term 2xφ'' is not cancelled and the coefficient of φ' in the interface equation acquires extra contributions; Eq. (3) is then not the formal limit of the stated ansatz. The circle validation cannot reveal this, because for a circle v is spatially constant, ω is constant, all tangential gradients ∇ω and ∂tπ·∇ω vanish, and the corrector-sensitive cancellation is trivial; the experiment only confirms the ODE (6) for the radius.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the momentum-based (damped hyperbolic) Allen-Cahn equation, ∂_tt u + α∂_t u = Δu − W′(u)/ε², as a tool for minimizing the Ginzburg-Landau functional. The authors prove three elementary properties: monotone decay of a total energy (Theorem 3.1), conditional subsequential convergence to critical points (Theorem 3.2), and finite speed of propagation (Theorem 3.3). The main analytic contribution is a formal singular-limit derivation, Eq. (3), asserting that interfaces evolve by ∂_t v = (1−v²)(h − αv) with v the normal velocity and h the mean curvature. The paper also proposes two time discretizations, CINEMA and a FISTA-type convex-concave splitting scheme, proves stability of CINEMA, and provides numerical experiments for curves, surfaces, and graph-based semi-supervised learning.","tokens_in":40433,"tokens_out":23517,"duration_ms":205658,"significance":"The proposed singular limit, if valid, is a novel hyperbolic geometric flow: it has finite propagation speed, a velocity-dependent interface width with Lorentz-like factor (1−v²)^{−1/2}, permits corner formation in analogy to hyperbolic conservation laws, and is not the gradient flow of the limiting perimeter functional. The rigorous results Theorems 3.1–3.3 and 4.1–4.2 are correct and useful, and the numerical experiments (circle validation, curve evolution, Schwarz P and Gyroid surfaces, MNIST) are extensive and clearly presented. The authors are also explicit that the singular limit is formal and that large-time-step FISTA behavior is not captured by the PDE analysis. However, the central formal derivation is currently undermined by a sign inconsistency in the corrector lemma (Lemma A.2), so the paper's main new analytic claim needs repair before the results can be relied upon.","major_comments":[{"comment":"Lemma A.2 states the corrector equation as ψ′′ − W′′(ϕ)ψ = ϕ′ + 2xϕ′′ (Eq. (10)), but the variational proof in Step 3 derives the Euler-Lagrange equation −ψ′′ + W′′(ϕ)ψ = ϕ′ + 2xϕ′′, which is the opposite sign. Step 4's numerical construction, which solves the homogeneous problem ψc′′ = W′′(ϕ)ψc and combines with ϕ′, likewise corresponds to the Euler-Lagrange sign. Section 3.2.2 uses exactly Eq. (10) to cancel the O(1/ε) terms in the refined ansatz, so the stated lemma does not support the cancellation. If the corrector satisfies the sign derived in the proof, the ϕ′ and 2xϕ′′ terms in the interface equation would add rather than cancel, and Eq. (3) would not follow even formally. This is a load-bearing error: the singular limit is the paper's central analytic claim, and the proof of the corrector lemma must be reconciled with the statement (or the ansatz and derivation adjusted accordingly) and the formal calculation re-verified.","section":"Appendix A, Lemma A.2; Section 3.2.2"},{"comment":"The formal singular-limit derivation discards all terms of order sdist ('we discard the terms in the equation which retain sdist') and supplements this with the assumption of 'well-prepared initial data', which is never defined. No error estimates or compactness arguments are supplied, so Eq. (3) is not a proven asymptotic statement about solutions of the accelerated Allen-Cahn equation. The numerical confirmation in Section 3.2.3 only treats circles, for which v and ω are spatially constant, all tangential gradients vanish, and the corrector-sensitive cancellation is trivial; it therefore cannot validate the refined ansatz in the general case. The authors are transparent about the formal nature of the derivation, but as the central claim of the paper, this limitation should be stated more prominently and, ideally, the well-prepared condition should be made precise.","section":"Section 3.2.2"}],"minor_comments":[{"comment":"The manuscript contains several typos, e.g., 'minization' in the header, 'Eudlidean' in Section 1, and 'are not be observed' in the abstract.","section":"Throughout"},{"comment":"The displayed algebraic lines following the refined ansatz are difficult to parse because of ambiguous parentheses in the typeset expressions; please clarify the grouping of the terms involving −(v∂tω)ψ′′ and W′′(ϕ)ψ/ε.","section":"Section 3.2.2"},{"comment":"In the computation of the cross term in F(ξ), the identity ∫(ψ̂′ϕ′′ + W′′ϕ′ψ̂) = 2∫ϕ′′ψ̂′ is not correct; the left-hand side equals 0 by integration by parts (since ∫W′′ϕ′ψ̂ = ∫ϕ′′′ψ̂ = −∫ϕ′′ψ̂′). The conclusion that the cross term vanishes is unaffected, but the displayed identity should be corrected.","section":"Lemma A.2, Step 3"},{"comment":"The sentence 'We can find the perimeter of (a phase field approximation to) an evolving disk in two ways' switches between 'we' and passive constructions; please streamline for readability.","section":"Section 3.2.3"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Lemma A.2 is a technical but load-bearing issue: the statement and proof must be made consistent, and the sign convention in the ansatz in Section 3.2.2 adjusted accordingly. Given the paper's otherwise careful presentation and the fact that Theorems 3.1–3.3 and the CINEMA stability analysis appear sound, I see this as a major revision rather than a rejection. I also note that the paper's use of the companion work [DOSW25] is contextual, not circular, and that the numerical results are reproducible in principle given the detailed algorithm descriptions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nThe thing to know: the paper's headline formal limit (3) is not actually derived. Lemma A.2 states the corrector equation ψ'' − W''ψ = φ' + 2xφ'', but the variational proof solves the opposite sign: the functional's Euler–Lagrange equation is −ψ'' + W''ψ = φ' + 2xφ''. The §3.2.2 cancellation uses the stated (plus) sign, so with the lemma as proved the O(1/ε) terms do not cancel and equation (3) does not follow, even formally. This is a load-bearing flaw, not a typo.\n\nWhat's genuinely good: Theorems 3.1–3.3 are clean and correct — energy dissipation, conditional convergence to critical points, and finite speed of propagation. The CINEMA scheme with its unconditional energy stability (Theorem 4.2) is a nice contribution, and the convex-concave splitting comparison with FISTA is useful. The graph experiments are interesting, and the paper is honest about the limits of the PDE analysis for large time steps.\n\nThe soft spots are proportionate. The sign error is the big one. The circle validation in Figure 2 cannot reveal it because for a circle v is constant, all tangential gradients and ∂tπ·∇ω vanish, so the corrector-sensitive terms drop out; the experiment only tests the ODE for the radius, not the full geometric law. The formal derivation also drops O(sdist) terms with no error estimate and never defines \"well-prepared\" initial data. The 5x speedup on the Schwarz P surface is a single run, with no code or error bars.\n\nThere is no circularity problem: the derivation is parameter-free and the self-citation is used only to explain ε²-slow movement, not as input to the new claims.\n\nBottom line: this is worth a serious referee. The idea is good, the rigorous parts are solid, and the numerical observations are suggestive. But as written the central formal result is unsupported. The sign in Lemma A.2 is fixable — change the sign in the functional — and then the derivation likely goes through, but the current version shouldn't be published without that correction and a re-check of the limit. I'd send it to review with the expectation of major revision.","headline":"The paper's headline formal singular limit is not derived as written: Lemma A.2 proves the opposite sign of the corrector equation, so the O(1/ε) cancellation in §3.2.2 fails; the rigorous parts and numerical comparisons are decent but the central claim needs a fix.","tokens_in":40950,"tokens_out":6838,"would_cite":false,"duration_ms":62494,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q05","53E10","35R02","53Z50"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper's central claim is that the momentum-based accelerated Allen-Cahn equation has a sharp-interface limit given by the hyperbolic law $\\partial_t v = (1-v^2)(h-\\alpha v)$, so interfaces move at bounded speed, form corners instead…","keywords":["accelerated Allen-Cahn equation","Ginzburg-Landau functional","mean curvature flow","singular limit","momentum methods","convex-concave splitting","semi-supervised learning","graph Laplacian"],"falsifier":"Simulate the accelerated Allen-Cahn equation at several small values of $\\varepsilon$ for a non-circular well-prepared initial curve, extract the zero-level-set normal velocity $v$ and curvature $h$, and compare them with an independent numerical solution of $\\partial_t v=(1-v^2)(h-\\alpha v)$; if the deviation does not shrink at the expected $O(\\varepsilon)$ rate, or if the width of a moving interface does not scale like $\\varepsilon(1-v^2)^{-1/2}$, the formal limit is wrong.","tokens_in":39816,"feed_emoji":"⚡","tokens_out":11037,"duration_ms":103692,"temperature":0.7,"pith_summary":"The paper studies momentum-based minimization of the Ginzburg-Landau diffuse perimeter functional, replacing the parabolic Allen-Cahn gradient flow by the damped hyperbolic equation $(\\partial_{tt}+\\alpha\\partial_t)u = \\Delta u - W'(u)/\\varepsilon^2$. Its central claim is that in the sharp-interface limit $\\varepsilon\\to 0$, well-prepared interfaces do not follow mean curvature flow but instead obey $\\partial_t v = (1-v^2)(h-\\alpha v)$, where $v$ is the normal velocity and $h$ the mean curvature. If this is right, momentum changes the geometry of perimeter minimization: information travels at speed at most one, corners can persist, and a moving interface is thinner than a stationary one. The paper also argues, with numerical evidence on planar curves, three-dimensional minimal surfaces, and graph-based classification, that a particular large-time-step FISTA-type convex-concave splitting can substantially accelerate convergence to minimizers even though the continuous hyperbolic PDE is less regular than Allen-Cahn.","feed_headline":"Momentum gives phase-field interfaces a speed limit and corners","feed_subtitle":"In the sharp-interface limit, velocity and curvature obey a Lorentz-type law instead of mean curvature flow.","key_machinery":"The central object is the refined traveling-wave/corrector ansatz for the accelerated Allen-Cahn equation: $u \\approx \\varphi(r/\\varepsilon) + \\varepsilon\\, v\\,\\partial_t\\omega\\, \\psi(r/\\varepsilon)$ with $r=(1-v^2)^{-1/2}\\operatorname{sdist}$. The optimal profile $\\varphi$ satisfies $\\varphi'' = W'(\\varphi)$, the corrector $\\psi$ satisfies $\\psi'' - W''(\\varphi)\\psi = \\varphi' + 2x\\varphi''$, and $\\omega=(1-v^2)^{-1/2}$ is the Lorentz factor. Inserting this ansatz and discarding off-interface terms turns the PDE into an averaged interface equation whose solvability condition is exactly $\\partial_t v=(1-v^2)(h-\\alpha v)$; the corrector sets the exchange ratio between the two leading-order terms. On the discrete side, the key machinery is a convex-concave splitting whose convex part is quadratic, making each implicit step a linear solve; this yields CINEMA, which is unconditionally energy-stable but does not accelerate, and FISTA, which accelerates for intermediate-to-large step sizes at negligible extra cost.","core_discovery":"The authors derive that the accelerated Allen-Cahn equation, the momentum version of the $L^2$ gradient flow of the Ginzburg-Landau energy, has a sharp-interface limit described by $\\partial_t v = (1-v^2)(h-\\alpha v)$. The derivation uses the refined traveling-wave ansatz $u = \\varphi(r/\\varepsilon) + \\varepsilon\\, v\\,\\partial_t\\omega\\, \\psi(r/\\varepsilon)$ with $r = (1-v^2)^{-1/2}\\operatorname{sdist}$, where $\\varphi$ is the optimal interface profile, $\\psi$ solves the corrector equation $\\psi'' - W''(\\varphi)\\psi = \\varphi' + 2x\\varphi''$, and $(1-v^2)^{-1/2}$ acts as a Lorentz factor for the signed distance. The corrector is needed because a moving interface changes both its width and its shape, and it fixes the balance between the $\\varphi'$ and $x\\varphi''$ terms that would otherwise not match. The resulting geometric law is hyperbolic: finite speed of propagation, no parabolic smoothing of corners, and an interface width contracted by $(1-v^2)^{-1/2}$. Numerical experiments on shrinking circles match the law, while large-time-step FISTA discretizations on planar curves, triply periodic minimal surfaces, and graph label propagation show that momentum can accelerate minimization outside the regime where the PDE analysis applies.","pith_inferences":["The appearance of the Lorentz factor suggests that other phase-field gradient flows, when inertial terms are added, may acquire similar velocity-dependent width corrections; a direct check would be to measure the width of a fast planar interface and compare it with $\\varepsilon(1-v^2)^{-1/2}$.","Since no tangential derivative of $v$ enters the limiting law, the flow is purely local along the interface; this raises the possibility of analysing corner formation and defining continuation past singularities by characteristic methods, as in hyperbolic conservation laws.","The FISTA acceleration occurs precisely in the large-step regime where the continuous PDE limit does not apply, so the mathematically honest explanation of the acceleration may be an $\\varepsilon$-fixed, large-time-step discrete limit rather than the singular limit (3).","On graphs, finite speed of propagation would mean label information cannot cross more than about one graph edge per unit of hyperbolic time, and momentum's inertia may be what lets FISTA escape poor local minima that stall plain gradient descent."],"forward_implications":["If the singular limit is correct, the sharp-interface limit of momentum-based diffuse perimeter minimization is a hyperbolic geometric flow with speed bound $|v|\\le 1$, so interfaces can form corners and a shrinking circle can collapse and then re-expand.","Because a moving interface is thinner by the factor $(1-v^2)^{-1/2}$, fast-moving interfaces require a finer spatial resolution than the Allen-Cahn equation would need at the same $\\varepsilon$.","In the large-time-step FISTA regime, momentum can reduce the number of iterations several-fold compared with convex-concave splitting gradient descent, for example reaching a triply periodic minimal surface in about 93 steps versus 471 for gradient descent, and it can speed up label propagation in graph-based semi-supervised learning.","CINEMA decreases the total energy in every step for any time-step size and is therefore a stable momentum scheme, but it does not provide acceleration; where the concave gradient is evaluated decides whether momentum helps.","The conditional convergence result implies that any long-time limit of the accelerated Allen-Cahn flow is a critical point of the Ginzburg-Landau energy, so the method is a legitimate minimization tool despite its non-monotone energy landscape."],"supporting_citations":[{"why":"Rigorous convergence of the Allen-Cahn equation to Brakke mean curvature flow; it sets the parabolic limit that the paper's hyperbolic law replaces.","marker":"[Ilm93]"},{"why":"Convergence of perturbed Allen-Cahn equations to forced mean curvature flow, the standard result the accelerated equation is contrasted with.","marker":"[MR11]"},{"why":"Short relative-entropy proof of Allen-Cahn to mean curvature flow convergence rates, supplying the quantitative parabolic baseline for the formal expansion.","marker":"[FLS20]"},{"why":"Source of the corrector-equation technique for Allen-Cahn-type profiles, which the refined ansatz for the accelerated equation relies on.","marker":"[GM10]"},{"why":"Second-order flows for non-convex energies with convex-splitting schemes, providing existence and stable discretizations for the accelerated Allen-Cahn equation.","marker":"[CDI+24]"},{"why":"Shows convex-concave splitting slows Allen-Cahn interface motion by $\\varepsilon^2$, which motivates the splitting and explains why large FISTA steps can beat gradient descent.","marker":"[DOSW25]"},{"why":"The FISTA algorithm whose convex-concave splitting version the paper adapts for momentum-based Ginzburg-Landau minimization.","marker":"[BT09]"},{"why":"Continuous-time ODE model of Nesterov acceleration that connects momentum schemes to damped hyperbolic dynamics and guides the step-size scaling.","marker":"[SBC14]"}],"fun_headline_variants":["Momentum gives phase-field interfaces a Lorentz-style speed limit","Momentum method makes interfaces obey a relativistic speed limit","Hyperbolic sharp-interface limit: momentum gives finite speed","Momentum replaces mean curvature flow with a Lorentz-type law","Accelerated Allen-Cahn: interfaces move with a finite speed bound"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation of the limiting law assumes that the solution remains in the refined traveling-wave form $u=\\varphi(r/\\varepsilon)+\\varepsilon\\, v\\,\\partial_t\\omega\\, \\psi(r/\\varepsilon)$ with $r=(1-v^2)^{-1/2}\\operatorname{sdist}$, that the initial data are well prepared, and that all off-interface terms of order $\\operatorname{sdist}$ can be discarded; no rigorous error estimate or stability proof for this ansatz is given.","fun_headline_variants_meta":{"raw":{"variants":["Momentum gives phase-field interfaces a Lorentz-style speed limit","Momentum method makes interfaces obey a relativistic speed limit","Hyperbolic sharp-interface limit: momentum gives finite speed","Momentum replaces mean curvature flow with a Lorentz-type law","Accelerated Allen-Cahn: interfaces move with a finite speed bound"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000753,"raw_usage":{"total_tokens":3378,"prompt_tokens":1003,"completion_tokens":2375,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":2299}},"tokens_in":619,"tokens_out":2375,"duration_ms":15997,"temperature":1.0,"reasoning_tokens":2299,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:52:12.287559+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the accelerated Allen-Cahn equation at several small values of $\\varepsilon$ for a non-circular well-prepared initial curve, extract the zero-level-set normal velocity $v$ and curvature $h$, and compare them with an independent numerical solution of $\\partial_t v=(1-v^2)(h-\\alpha v)$; if the deviation does not shrink at the expected $O(\\varepsilon)$ rate, or if the width of a moving interface does not scale like $\\varepsilon(1-v^2)^{-1/2}$, the formal limit is wrong.","supporting_citations":[],"review_version":1}