{"id":"158192fc-6932-456c-92b8-248ba5982e6b","arxiv_id":"2506.18869","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The convex-concave splitting scheme for the Allen-Cahn equation moves interfaces only on a time scale of order epsilon squared, so its guaranteed stability comes from freezing the interface motion.","lead":"This paper shows that a popular unconditionally stable numerical method for the Allen-Cahn equation achieves its stability by nearly freezing the phase boundaries. The effective time step shrinks as the square of the interface width, so choosing a larger step does not speed up the simulation.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Universal ε²-slowness claim is proved only for a special quadratic potential with τ=∞ on R³; bounded domains, finite τ, and the standard potential rest on assertion and numerics.","rationale":"Read in good faith, the paper delivers several genuine contributions: Lemma 3.2/Corollary 3.3 give a clean ε^{1/(p-1)} Hölder-type slowness bound for a broad class of potentials; Appendix B carefully verifies the IPS99 kernel conditions for the screened-Poisson kernel on R³; Appendix D explicitly computes the barrier-potential ball example and shows r_new-r=O(ε²). These should be credited. The concern is not internal inconsistency but scope: the strongest advertised conclusion (universal ε²-slowness) is not what is proved. The rigorous core is the τ=∞, R³, special-potential equivalence. The statements needed for the abstract's universality—bounded domains, finite τ, standard potential—are respectively asserted, not proved, and numerical with fitted prefactors. This matters because the finite-τ update has a natural diffusion scale τ/(1+2τ/ε²), which is τ for small τ; without a proof one cannot claim ε²-slowness independently of τ. The authors are transparent about some of this (intro admits only R³ rigor; Section 4.2 says finite-τ optimality not demonstrated), which is why the appropriate verdict is CONDITIONAL rather than REJECT: the claims should be re-stated as theorems where proved and conjectures otherwise, with code/data for the numerics. My read matches the reader's weakest-assumption analysis.","tokens_in":27878,"tokens_out":18236,"duration_ms":183358,"concrete_test":"Derive the sharp-interface limit of the finite-τ quadratic update on the flat torus T³: localize the periodic Green's function of (1+2τ/ε²-τΔ)^{-1} near an interface point, rescale as in IPS99, and compute the normal velocity v(Σ,n) including periodic-image contributions. If v equals -H with time step h_eff=τ/(1+2τ/ε²) to leading order, the ε²/2 bound is confirmed for τ≫ε² but the universal claim must be restricted to that regime; if v is -H with step ε²/2 independent of τ, the finite-τ claim is proved. If periodic-image terms appear at leading order, the bounded-domain ε² conclusion is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim—convex-concave splitting makes interfaces move on the ε² time scale—is established rigorously only in a narrow setting: W=(|u|-1)², τ=∞, whole space R³, via the equivalence of (4.4) with the IPS99 thresholding scheme (Section 4.2, Appendix B). The universal statement in the abstract/title is not supported at that level. (i) On bounded domains with periodic or Neumann conditions, the Green's function of (1-ε²/2 Δ)^{-1} is not translation-invariant, so the IPS99 hypotheses checked in Appendix B do not directly apply; Section 4.2 only 'asserts that standard proofs apply.' (ii) For finite τ, the quadratic update reads u^{n+1}=(1+2τ/ε²-τΔ)^{-1}(u^n+(2τ/ε²)sign u^n); its natural diffusion scale is τ/(1+2τ/ε²), and no theorem is given showing the interface velocity is ε²/2. The text explicitly says no finite-τ sweet spot has been rigorously excluded. (iii) For the standard W=(u²-1)², Section 6 reports only a numerical fit with prefactor 0.25·ε²; no data/code are provided. These limitations are flagged by the authors, but the abstract presents the ε² conclusion as universal. If the bounded-domain/finite-τ extension fails (e.g. because h_eff behaves like τ for τ≲ε²), the practical conclusion collapses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the convex-concave splitting time discretization of the Allen-Cahn equation, in which the double-well potential is decomposed into a convex and a concave part and the two parts are treated implicitly and explicitly, respectively. The main claimed result is that the scheme's 'effective time step' scales as ε^2, where ε is the interface width, so that unconditional energy stability is achieved only at the price of effectively freezing the interface. The paper contains three strands: (i) a general energy-dissipation estimate (Lemma 3.2 and Corollary 3.3) showing that for potentials with a uniform p-curvature condition, N steps cannot move the solution by more than O((N ε^{1/(p-1)})^{1-1/p}); (ii) for potentials with quadratic convex part, a link between the infinite-step-size iteration (4.4) and the Merriman-Bence-Osher thresholding scheme, with a rigorous identification on the whole space R^3 via the IPS99 convergence theorem, giving effective step ε^2/2; and (iii) evidence for the same ε^2 scaling for barrier potentials (via an explicit radial example) and for the standard quartic potential (via numerical experiments). The authors are explicit that the analysis is only fully rigorous on R^3 and that the standard-potential claim is numerical, but the abstract and title present the ε^2-slowness as a universal conclusion.","tokens_in":28166,"tokens_out":6420,"duration_ms":68188,"significance":"If the ε^2-slowness claim is correct in the generality stated, it is an important and somewhat sobering message for the phase-field community: the popular convex-concave splitting scheme, often advertised as unconditionally stable, achieves that stability by drastically limiting the effective time step, so large nominal time steps do not accelerate interface motion. The rigorous components of the paper are clean and valuable: the energy-dissipation bound in Lemma 3.2/Corollary 3.3 is elementary but sharp in its ε-scaling, Lemma 4.3 gives a variational characterization of the thresholding step, and Appendix B carefully verifies the IPS99 hypotheses for the screened Poisson kernel on R^3. The paper is also commendably honest in flagging its own limitations, explicitly stating that the MBO equivalence is only rigorous on R^3 and that no finite-τ 'sweet spot' has been excluded. However, the headline universal claim is not supported by the proofs as written: the rigorous kernel analysis covers one special potential, τ=∞, and the whole space, while the standard-potential claim is purely numerical and the barrier-potential claim rests on a single explicit example.","major_comments":[{"comment":"The ε^2-thresholding equivalence is rigorously established only for the potential W(u)=(|u|-1)^2, for formally infinite step size τ=∞, and on the whole space R^3. In that setting, (4.4) is rewritten as convolution with the screened Poisson kernel and Appendix B verifies the hypotheses of [IPS99] for that kernel. The sentence immediately after (4.4) asserting that 'standard proofs apply' to bounded domains with periodic or Neumann boundary conditions, and to finite τ, is not a proof: on a bounded domain the Green's function of (1-ε^2/2 Δ)^{-1} is not translation-invariant, and the IPS99 theorem is formulated on R^d. For finite τ the update is u^{n+1}=(1+2τ/ε^2-τΔ)^{-1}(u^n+(2τ/ε^2)sign u^n), whose natural diffusion scale is τ/(1+2τ/ε^2), and the paper itself states that no finite-τ sweet spot has been rigorously excluded. Since the abstract and title assert ε^2-slowness for the full class of potentials and domains, this gap is load-bearing. I request either a proof or a precise theorem for bounded domains and finite τ, or a reformulation of the title, abstract, and conclusions so that the rigorous claim is restricted to the setting actually proved, with the bounded-domain and finite-τ statements explicitly labeled as conjectural.","section":"§4.2 and Appendix B"},{"comment":"The claim that the standard potential W(u)=(u^2-1)^2 is ε^2-slow rests entirely on numerical fits, with an effective step of 0.25 ε^2, at a single spatial resolution n=512 and with Newton-Raphson solves for the convex subproblem. No data, code, error bars, or systematic convergence study in ε are provided. Since the standard potential is the practically most important case and is presented in the abstract as one of the three scenarios, this evidence is not sufficient to support the universal statement. Please either supply reproducible data/code and a convergence study, or explicitly downgrade the standard-potential claim to a numerical observation in the abstract and conclusions.","section":"§6, Figure 7"},{"comment":"For barrier potentials, the ε^2 conclusion is supported by one explicit τ=∞ radial example (Example 5.2), not by a convergence theorem for the iteration. The text says the authors 'demonstrate by example' and Figure 6 shows the displacement r_new-r as O(ε^2), but the abstract counts barrier potentials among the scenarios where the effective time step scales as ε^2. Moreover, the derivation in Appendix D of r_new-r=O(ε^2) uses unquantified 'err' terms and informal 'easy to see' boundedness statements, so even this single example is not fully rigorous as written. Please state precisely what is proved for barrier potentials, what is observed numerically, and what remains conjectural.","section":"§5 and Appendix D"},{"comment":"The central notion of 'effective time step size' is never defined formally. In the rigorous MBO setting it can be identified with the thresholding time h=ε^2/2, but in the general slow-motion bound (Corollary 3.3) and in the numerical sections it is inferred informally from the number of iterations required to move the interface a distance of order one. Without a precise definition, statements such as 'the effective time step size scales as ε^2' are not quantitatively falsifiable and cannot be compared across the three potential classes. I recommend giving a formal definition, for example in terms of the number of iterations needed to achieve a prescribed interface displacement, and then stating each theorem in those terms.","section":"§1 and §4.2"}],"minor_comments":[{"comment":"The phrase 'certain applications certain applications' contains a duplicated word; it should read 'certain applications'.","section":"Page 2, Section 1"},{"comment":"The title contains spacing and hyphenation artifacts: 'CONVEX-CONCA VE' and 'EQUA TION' should be cleaned up.","section":"Title and running header"},{"comment":"In the curvature heuristic, the text says 'If κ>0, there is an incentive to move the transition where u'≫1 to the left; while κ>0 incentivizes transitions further to the right'; the second occurrence should presumably be κ<0, since the two statements are contradictory as written.","section":"§7, heuristic discussion"},{"comment":"The effective-step prefactors 0.29 and 0.5 are stated without explanation of how they were obtained; a sentence describing the fitting procedure would improve reproducibility.","section":"Figure 1 and Figure 3"},{"comment":"The notation in the displayed inner-variation calculation would benefit from a definition of the error term 'err' and from a short justification of why the neglected terms are higher order in ε; as written the O(ε^2) bound is plausible but not fully quantified.","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about its limitations in the body, but the title and abstract substantially overstate the proven scope. The authors may be able to resolve this either by adding proofs for bounded domains and finite τ or by carefully restricting the title and abstract to the rigorous special case and labeling the rest as numerical or conjectural. The numerical section for the standard potential should include code/data or be explicitly acknowledged as a heuristic observation. I would not reject the paper, because the core MBO link and the energy-dissipation estimates are valuable and correct as far as they go."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper gives a genuinely new structural identification: for the quadratic convex part and τ=∞, the convex-concave splitting iteration is exactly an MBO thresholding step of size ε²/2. On R³ they verify the IPS99 conditions in Appendix B, so that link is rigorous and it is the best part of the paper. Second, the headline claim—that the effective time step always scales like ε²—is not proved at that level of generality. Bounded domains, finite τ, and the standard potential are asserted or numerical. The authors flag most of this in the body, but the abstract sells the universal conclusion.\n\nCredit where due: Lemma 3.2 and Corollary 3.3 give a clean, correct energy-dissipation bound: each step dissipates ε/2∥δu∥²_H¹ + (ε/τ)∥δu∥²_L² + (c̄/ε)∥δu∥^p_{L^p}, and the Hölder argument yields the ε^{1/(p-1)} slowness bound independent of τ. That's a solid, citable estimate. The barrier-potential example in Section 5 is also a real computation: for the radial double obstacle, r_new − r = O(ε²), shown by an explicit inner-variation argument. It gives a nontrivial data point in a regime where the curvature conditions fail.\n\nSoft spots, in proportion. The bounded-domain claim in Section 4.2 is an assertion, not a proof: the resolvent kernel is not translation-invariant on a periodic or Neumann domain, and IPS99 is whole-space theory. The finite-τ update has effective step τ/(1+2τ/ε²), so saturation at ε²/2 is plausible, but no theorem shows the interface speed is ε²/2 there; the authors admit they haven't ruled out a sweet spot. For the standard potential, the ε²-scaling is a numerical fit with prefactor 0.25·ε², and no code or data is shipped. I'd call these addressable, not fatal. The practical message—large τ does not buy speed—is supported by Figure 3 and the numerics, but the universal statement needs a rewrite or a proof.\n\nWho is it for: anyone doing phase-field simulations with unconditional stability schemes, and especially people working on MBO-type thresholding. The MBO link is the sort of bridge that will generate follow-up work.\n\nRecommendation: send it to peer review. The rigorous core is new and sound; the framing needs work. I'd push for either proving the bounded-domain/finite-τ statements or labeling them explicitly as conjectural, and for shipping the numerics. My own verdict is 'conditional'—engage with it, but push for honesty in the abstract.","headline":"Solid new MBO link for the τ=∞ quadratic case, but the universal ε²-slowness claim is only partly proved; refereeing it with a toned-down abstract is the right call.","tokens_in":28748,"tokens_out":8262,"would_cite":true,"duration_ms":76537,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M12","35A35","49Q05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Convex-concave splitting for the Allen-Cahn equation moves interfaces on an ε² time scale, not the nominal step size τ.","keywords":["Allen-Cahn equation","convex-concave splitting","MBO scheme","mean curvature flow","effective time step","interface dynamics","energy stability","thresholding"],"falsifier":"Run the convex-concave splitting scheme on a shrinking circle in a periodic square with a large fixed nominal step τ for two values of ε that differ by a factor of two, and measure the number of steps needed to shrink the radius by a fixed fraction. If the effective step scales as ε², the step counts differ by a factor of four; if the interface moves at the nominal τ pace, the counts are nearly equal.","tokens_in":27620,"feed_emoji":"🧊","tokens_out":5608,"duration_ms":56190,"temperature":0.7,"pith_summary":"This paper argues that the convex-concave splitting time-stepping scheme for the Allen-Cahn equation, popular for being unconditionally energy stable and easy to implement, suffers from a severe hidden limitation: for a range of double-well potentials, interfaces move on a time scale proportional to ε², where ε is the width of the transition layer. Choosing a large nominal time step τ does not accelerate the dynamics; stability is bought by effectively freezing the interface in place. For potentials with a quadratic convex part, the authors prove this by showing that the infinite-step iteration is exactly a Merriman–Bence–Osher thresholding step of size ε²/2, linking the Allen-Cahn scheme to mean curvature flow. The same ε² scaling is shown by explicit example for barrier potentials and numerically for the standard potential, and a weaker $ε^{{1/(p-1)}}$ bound is proved in general.","feed_headline":"Allen-Cahn splitting slows interfaces to ε² time scale","feed_subtitle":"A stable, easy scheme that hides a catch: large time steps do not speed up interface motion.","key_machinery":"The central object is the equivalence between the infinite-step-size convex-concave splitting iteration and thresholding dynamics: (1 - τΔ)$u^{{n+1}}$ = sign(u^n) with τ = ε²/2 is the screened-Poisson (Yukawa) kernel iteration K_τ * sign(u^n), which is the MBO scheme for a first-order heat approximation. The second machinery is Lemma 3.2, an energy-dissipation estimate giving a Hölder bound ∥u^N - $u^{0}$∥_{L^p} ≲ (N $ε^{{1/(p-1)}}$)^{1-1/p}, independent of τ, which forces slowness for potentials with convex part |u|^p. The third is the explicit double-obstacle solution for the barrier potential, where r_new = r + O(ε²).","core_discovery":"For the prototypical potential with quadratic convex part, W(u) = (|u|-1)², the convex-concave splitting iteration with formally infinite time step reduces to $u^{{n+1}}$ = (1 - (ε²/2)Δ)^{-1} sign(u^n), which is a first-order implicit Euler approximation to the heat equation applied to the thresholded previous state — i.e. an MBO thresholding step with time step ε²/2. On the whole space ℝ³, the Green's function of (1 - (ε²/2)Δ) is radially symmetric with finite second moment and satisfies the Ishii–Pires–Souganidis conditions, so the iteration is a time discretization of mean curvature flow at speed ε²/2. Hence the effective time step of the scheme is ε²/2 regardless of nominal τ. For barrier potentials, an explicit double-obstacle calculation in a ball shows the zero-level set shifts by O(ε²) per step; numerical experiments with the standard potential confirm the same scaling.","pith_inferences":["The MBO equivalence suggests that quantitative convergence-rate results for thresholding dynamics (for example, in L¹) could transfer to this Allen-Cahn scheme, with an additional error of order ε² per step, once the missing analysis on bounded domains is supplied.","The ε² slowness likely holds for any convex-concave splitting whose convex part has positive second derivative at the wells: evaluating the concave force at the previous iterate produces a fixed force that resists translation of the profile quadratically in the shift, an effect that does not vanish with larger τ.","A testable design consequence is that a semi-implicit scheme with the double-well term treated explicitly and the Laplacian implicitly would move interfaces at the same ε²-limited speed but at a fraction of the per-step cost, making the convex-concave splitting strictly dominated for computing real interface dynamics."],"forward_implications":["For potentials with quadratic convex part, the scheme is literally a first-order approximation to the MBO thresholding scheme, so its interface motion is mean curvature flow on the ε² scale, no matter how large the nominal step τ.","The ε² scaling is independent of τ, so increasing the time step cannot buy faster interface dynamics; the energy-decreasing property is obtained by freezing the interface.","For barrier potentials, the explicit double-obstacle example shows the same ε² slowdown, and the numerical experiment with the standard potential indicates the behavior is universal.","A general energy estimate implies that even when the convex part grows like |u|^p, the effective step is at most ε^{1/(p-1)}, so a wide family of convex-concave splittings are slow compared to the desired mean-curvature time scale.","The scheme does not improve on a simpler semi-implicit treatment (implicit Laplacian, explicit double-well) with a sufficiently small step, and it can be slower in practice because each step may require a new convex minimization."],"supporting_citations":[{"why":"Supplies the general kernel thresholding framework whose convergence to mean curvature flow the screened-Poisson kernel is checked against.","marker":"[Ish95]"},{"why":"Provides the convergence theorem and kernel conditions used to conclude that the (1 - (ε²/2)Δ)^{-1} iteration is a time discretization of mean curvature flow on ℝ³.","marker":"[IPS99]"},{"why":"The Esedoglu-Otto BV-type construction that Lemma 4.3 invokes to obtain a second monotone quantity for the thresholded sequence.","marker":"[EO15]"},{"why":"Gives the unconditional energy-stability theorem for convex-concave splitting that motivates the whole discretization.","marker":"[Bar15]"},{"why":"Establishes Γ-convergence of the Modica-Mortola functional to perimeter, identifying the energy whose decrease the scheme guarantees.","marker":"[MM77]"},{"why":"The OSQP solver used for the numerical experiments with the barrier potential, which provide the evidence for the ε² scaling in that case.","marker":"[SBG+20]"}],"fun_headline_variants":["Allen-Cahn convex-concave splitting hides ε² time-step limit","Stable Allen-Cahn splitting freezes interfaces at ε² speed","Large time steps don't help: Allen-Cahn scheme scales as ε²","Convex-concave splitting reveals MBO thresholding at ε² scale","Effective time step is ε², not τ, in Allen-Cahn splitting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ε²-slow motion for the practically used cases — the standard potential, bounded domains with periodic or Neumann boundary conditions, and finite time steps — rests on the assertion, stated but not proved in Section 4.2, that the whole-space thresholding analysis extends to those settings; for the standard potential, the scaling is only observed in numerics, not derived.","fun_headline_variants_meta":{"raw":{"variants":["Allen-Cahn convex-concave splitting hides ε² time-step limit","Stable Allen-Cahn splitting freezes interfaces at ε² speed","Large time steps don't help: Allen-Cahn scheme scales as ε²","Convex-concave splitting reveals MBO thresholding at ε² scale","Effective time step is ε², not τ, in Allen-Cahn splitting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1231,"prompt_tokens":935,"completion_tokens":296,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":198}},"tokens_in":551,"tokens_out":296,"duration_ms":3309,"temperature":1.0,"reasoning_tokens":198,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:41:29.101443+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the convex-concave splitting scheme on a shrinking circle in a periodic square with a large fixed nominal step τ for two values of ε that differ by a factor of two, and measure the number of steps needed to shrink the radius by a fixed fraction. If the effective step scales as ε², the step counts differ by a factor of four; if the interface moves at the nominal τ pace, the counts are nearly equal.","supporting_citations":[],"review_version":1}