{"id":"ef6b5061-bc86-4117-98b2-9a8fa1c1782c","arxiv_id":"2411.17403","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A weighted SAV method for gradient flows adaptively blends nonlinear energy-based and Lagrange multiplier SAV schemes to keep solutions at large time steps while tracking the original energy more closely.","lead":"The authors propose a weighted scalar auxiliary variable (SAV) method for simulating gradient flows, combining two existing SAV approaches with a tunable weight. The method stays solvable for large time steps while tracking the original physical energy more closely, which could help long-time phase-field simulations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's 'existence for all λ∈(λ,1]' is false even for bounded E_N: a scalar counterexample with E_N=cos φ has no solution for λ=0.99 although λ=1 is solvable, so the central solvability claim collapses.","rationale":"The reader's weakest assumption (boundedness of E_N) is real but not the deepest problem. Theorem 3 is the load-bearing result for the headline claim 'solutions for any time step with sufficiently large λ'. Its proof uses (3.12) with a fixed fmin, but fmin depends on λ; in degenerate cases f1(1)=0, fmin ∼ -(1-λ)^2 while the g-term is O(1-λ), so the sufficient condition fails near λ=1. The explicit scalar data above gives a bounded E_N where Eq. (3.7) has no root for λ=0.99, refuting the theorem. The energy-dissipation part (Theorem 4) is a correct variational identity and the numerical evidence may still be useful, but the central existence guarantee and the associated bisection rationale are not supported; hence the paper should be rejected unless the theorem is corrected to give the actual solvability set.","tokens_in":24456,"tokens_out":31831,"duration_ms":298252,"concrete_test":"Recompute the scalar Eq. (3.7) for L=G=1, τ=1, φ^n=4, E_N=cos φ, H=-sin φ, C=1.410, p=2, q=-0.4351, r^n=0.8697. For λ=0.99, evaluate h(x)=2.3586x^2+0.017x+0.01(cos(2-0.4351x)-cos4) and confirm its global minimum is positive (≈2.3e-3), meaning no real root exists. Also solve at λ=0.5 to confirm a root does exist, demonstrating that the solvability set is not an interval [λ,1] as Theorem 3 claims.","verdict_should_be":"REJECT","load_bearing_attack":"The central existence guarantee (Theorem 3) is not merely restricted by boundedness of E_N; it can fail even for bounded E_N. Consider the scalar (1D) case L=G=1, τ=1, φ^n=4, E_N(φ)=cos φ, C=1.410, so r^n=sqrt(cos4+C)=0.8697 and H(φ^n)=-sin4=0.7568. Then p=(1+τ)^{-1}φ^n=2 and q=-τH/[(1+τ)r^n]=-0.4351. One checks f1(1)=2r^n+(H(φ^n),p-φ^n)/r^n=0, so at λ=1 the only root of (3.7) is the trivial x=0. For λ=0.99, Eq. (3.7) becomes h(x)=2.3586x^2+0.017x+0.01(cos(2-0.4351x)-cos4), and h(x)>0 for every real x (global minimum ≈2.3e-3), so there is no solution. By continuity the same holds for all λ<1 sufficiently close to 1. Hence there is no λ∈(0,1) such that solutions exist for every λ∈(λ,1]; the theorem's conclusion is false. The proof fails because fmin in (3.12) is not independent of λ and can vanish quadratically in (1-λ) while the g-term is only first-order.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a weighted scalar auxiliary variable (SAV) method for gradient flows, unifying the nonlinear energy-based SAV approach (λ=1) and the Lagrange multiplier SAV approach (λ=0) through a weight coefficient λ∈[0,1]. Backward Euler and Crank–Nicolson time discretizations are formulated, leading to nonlinear scalar equations (3.7) and (3.32). The main theoretical claims are: (i) for any time step τ, solutions exist for all λ in some interval (λ_bar,1] (Theorem 3 and, without proof, Theorem 6); (ii) the modified energy (3.14) dissipates unconditionally, and the original energy dissipates under an additional condition (Theorems 4 and 7); and (iii) an adaptive bisection algorithm selects the smallest feasible λ. Numerical experiments on Allen–Cahn and Cahn–Hilliard type equations with a truncated quartic potential demonstrate first- and second-order accuracy, mass conservation, and smaller energy deviation for smaller λ.","tokens_in":24763,"tokens_out":16185,"duration_ms":157893,"significance":"The unification of the two SAV formulations is conceptually attractive, and the discrete energy-dissipation identities in Theorems 4 and 7 are derived cleanly. The numerical study is extensive and supports the practical observation that smaller λ reduces the deviation between the modified and original energies, and that the Lagrange multiplier limit can fail for large τ. However, the central existence theorem for arbitrary τ is false as stated; a concrete bounded nonlinearity gives no solution for λ close to 1 even though λ=1 is solvable. Since the claimed solvability guarantee is the main theoretical contribution and is also used to justify the adaptive algorithm, the paper in its current form does not establish its headline claim. With a corrected existence statement and a revised algorithm justification, the underlying weighting idea could still be a useful contribution.","major_comments":[{"comment":"Theorem 3 is false as stated; the proof's sufficient condition (3.12) is invalid because f_min depends on λ. A scalar counterexample: take L=G=1, τ=1, φ^n=4, E_N(φ)=cos φ, and choose C = -sin 4 - cos 4 ≈ 1.4104 so that r^n = sqrt(cos 4 + C) and f_1(1)=2r^n + (H(φ^n), p^{n+1}-φ^n)/r^n = 0. With p^{n+1}=2 and q^{n+1} = -H(φ^n)/(2r^n) ≈ -0.4351, the equation (3.7) at λ=1 has only the root x=0, while at λ=0.99 the left-hand side h(x)=2.3586x^2+0.017x+0.01(cos(2-0.4351x)-cos 4) is strictly positive for all real x (global minimum ≈2.3×10^-3). Hence (3.7) has no solution for λ=0.99, and by continuity none for all λ<1 sufficiently close to 1, so no λ_bar<1 can have the property stated. The flaw is that f_min in (3.12) can vanish quadratically as λ→1 when f_1(1)=0, while the bound 2M(1-λ) is only linear. Theorem 6, whose proof is omitted, inherits this problem.","section":"Section 3.1, Theorem 3 and Eq. (3.12)"},{"comment":"The statement of Theorem 3 does not include the boundedness assumption |E_N|≤M that is introduced in its proof, and the paper's main numerical model (4.1) has E_N(φ)=∫(1/4)(φ^2-1-γ)^2 dx, which is unbounded on the natural function space. The truncated potential (4.2) makes E_N bounded, but the paper does not establish convergence of the truncated model to the original one as δ→∞ (nor as γ varies). Thus as written the theoretical solvability guarantee applies only to the truncated model, while the abstract presents it as a property of the weighted SAV method for gradient flows with bounded nonlinear energy. This gap should be stated explicitly in the theorem, and either the analysis extended or the numerical examples reinterpreted as solving the truncated problem.","section":"Section 3.1, Theorem 3 and Section 4, Eq. (4.2)"},{"comment":"The bisection search for λ_min assumes that solvability of (3.7) is monotone in λ in the sense that solvable values form an interval (λ_bar,1]. This is exactly the statement that Theorem 3 was supposed to provide, and it is false by the counterexample above. The assertion after Theorem 5 that 'there is no λ<λ_a that guarantees a solution' is not a consequence of Theorem 5(ii): that result only says that if a solution at λ* satisfies 2r*(r*-r^n)<E_N[φ*]-E_N[φ^n], then every λ∈[λ*,1] is solvable; it does not exclude isolated solvable values below λ*. Consequently, when the algorithm encounters an unsolvable midpoint and moves the bracket upward, it may discard all solvable λ, and the returned λ is not certified to be the minimal one. Algorithm 2 and the analogous claim for the CN scheme require the same correction.","section":"Section 3.1, Algorithm 1 and the paragraph after Theorem 5"}],"minor_comments":[{"comment":"There are several typographical errors, including 'utiliezd' in the Introduction, 'oganized' in the Introduction, and 'form' instead of 'from' in the step descriptions of Algorithms 1 and 2; these should be corrected.","section":"Throughout"},{"comment":"In Algorithm 2, the solvability checks inside the bisection loop refer to 'Eq. (3.7)', but they should refer to the CN scalar equation (3.32).","section":"Section 3.2, Algorithm 2"},{"comment":"Theorem 6 is stated without a proof; given that the corresponding first-order theorem is false as stated, the sentence 'We will omit the detailed proof' is not acceptable unless the corrected first-order result is supplied and the analogous proof is shown to carry over.","section":"Section 3.2, Theorem 6"},{"comment":"The statement that smaller λ gives a modified energy closer to the original is essentially a restatement of the definition (3.14) combined with the discrete relation between r and E_N; the numerical confirmation in Table 3 is therefore expected and should be presented as verification of the formula rather than as a new discovery.","section":"Section 4.3 and Eq. (3.14)"}],"recommendation":"major_revision","confidential_remarks":"The stress-test counterexample is decisive for Theorem 3 as stated. I would not accept the paper until the authors either prove a corrected existence theorem (for example, under a nondegeneracy condition such as f_1(1)≠0, or by a different argument showing existence of some λ for each τ) and revise the algorithm's monotonicity claim, or explicitly restrict all claims to settings where they are true. If no correct existence theorem is supplied, the central contribution of the paper disappears and rejection would be appropriate. The energy-stability identities and the numerical experiments are valuable; the main issue is concentrated in the existence and algorithm sections."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The punchline: Theorem 3, the central existence guarantee, is false as stated. I checked the stress-test counterexample and it holds. Take L=G=1, τ=1, φ^n=4, E_N(φ)=cos φ, C=1.410. Then the scalar equation (3.7) is solvable at λ=1, but for λ=0.99 it has no real root, and by continuity no λ<1 sufficiently close to 1 works. E_N is bounded and smooth, so boundedness doesn't rescue the claim. The proof's inequality (3.12) treats f_min as a constant independent of λ, but f_min(λ) vanishes quadratically as λ→1 while the g-term is only O(1−λ), so the bound cannot hold. This is a load-bearing failure, not a corner case: the abstract and introduction specifically advertise solvability for any time step with a sufficiently large weight.\n\nCredit where it's due. The weighting idea itself is a natural and useful interpolation between the two SAV families, and the discrete energy E=1/2(φ,Lφ)+λ(r^2−C)+(1−λ)E_N[φ] is well chosen. Theorem 4's dissipation identity is clean, no fitted constants. The second-order CN extension is straightforward. The numerics are thorough: temporal convergence tables, several 2D interface evolutions including the fractional Cahn–Hilliard case, and a 3D torus. The λ_min plots give concrete evidence that the plain Lagrange multiplier scheme loses solvability at practical time steps.\n\nSoft spots. The adaptive bisection for λ_min is heuristic; Theorem 5 gives only conditional monotonicity, and the algorithm probes for solvability without a rigorous guarantee. The original-energy stability in Theorem 4 depends on an a posteriori inequality that the experiments never check. Boundedness of E_N is a genuine restriction; the tests enforce it by truncating the double-well at δ=5, but the counterexample shows boundedness alone is not enough for Theorem 3. Minor: Algorithm 2 references Eq. (3.7) instead of (3.32), and no code or data is posted.\n\nWho this is for: numerical analysts working on SAV schemes and phase-field computations. Read it for the weighted-energy idea and the numerical comparisons; don't rely on Theorem 3. The paper deserves a serious referee because the flaw is subtle and the method may be repairable, but the current version should not be accepted without a corrected or weakened existence statement.","headline":"The weighted-SAV idea is solid and the numerics are convincing, but Theorem 3's existence guarantee is false as stated—a bounded-E_N counterexample with cos nonlinearity shows the proof's f_min bound fails.","tokens_in":25308,"tokens_out":8141,"would_cite":false,"duration_ms":66613,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M22","65M70","35K35","35K55"],"pacs":[],"model":"deepseek-v4-flash","headline":"The two main scalar auxiliary variable (SAV) approaches for gradient flows—one that always has solutions but controls only a modified energy, the other that preserves the original energy but can fail at large time steps—are unified by a…","keywords":["weighted scalar auxiliary variable","gradient flow","energy stability","Lagrange multiplier","phase-field model","Cahn–Hilliard equation","Allen–Cahn equation","solution existence"],"falsifier":"Run the first-order scheme (3.3) on the Cahn–Hilliard equation with an untruncated quartic $E_N$ and a large time step, following Algorithm 1; if at some step no $\\lambda\\in[0,1]$ makes the scalar equation (3.7) solvable, then Theorem 3 fails for the untruncated energy. A more direct check is to exhibit a configuration where the quadratic coefficient in (3.8) is positive but $f(x;\\lambda)+g(x;\\lambda)$ stays positive for all real $x$ at every $\\lambda\\in[0,1]$, contradicting the claimed solution existence.","tokens_in":24228,"feed_emoji":"⚖️","tokens_out":6474,"duration_ms":53409,"temperature":0.7,"pith_summary":"The paper introduces a weighted scalar auxiliary variable (SAV) method for gradient flows, blending two existing SAV strategies: the nonlinear energy-based approach, which always has solutions but only guarantees stability for a modified energy, and the Lagrange multiplier approach, which preserves the original energy but can fail for large time steps. The central claim is that a weight coefficient $\\lambda \\in [0,1]$ interpolating between the two produces schemes that are solvable for any time step size whenever $\\lambda$ is large enough, while choosing $\\lambda$ as small as possible keeps the discrete energy close to the original free energy. First- and second-order discretizations are analyzed, and numerical experiments on Allen–Cahn, Cahn–Hilliard, and space-fractional Cahn–Hilliard equations demonstrate the expected accuracy, energy stability, and improved faithfulness of the energy. A reader should care because the method addresses a known trade-off in structure-preserving simulation between solution existence at large time steps and accurate energy dissipation.","feed_headline":"One weight coefficient rescues SAV solvers at large time steps","feed_subtitle":"A weighted blend of two SAV schemes keeps energy close to the original while staying solvable for any step size.","key_machinery":"The central object is the weighted scalar auxiliary variable reconstruction (3.1), where $r(t)=\\sqrt{E_N[\\varphi]+C}$, $L$ is the linear, symmetric, non-negative operator in the quadratic energy, and $\\lambda\\in[0,1]$ is the weight; $\\lambda=1$ reproduces the nonlinear energy-based SAV scheme and $\\lambda=0$ reproduces the Lagrange multiplier formulation. The load-bearing mechanism is the decoupling $\\varphi^{n+1}=p^{n+1}+r^{n+1}q^{n+1}$, which reduces each time step to two linear elliptic solves with constant coefficients plus one nonlinear scalar equation (3.7) for $r^{n+1}$. The positivity identity (3.11) makes the quadratic coefficient of that scalar equation positive for any $\\tau>0$, and the assumed uniform boundedness of $E_N$ supplies the extremal argument in Theorem 3. The bisection algorithm then chooses the minimal $\\lambda$ for which the scalar equation has a solution, and the weighted energy (3.14) is what is guaranteed to dissipate; original-energy dissipation follows when the inequality connecting $2r^{n+1}(r^{n+1}-r^n)$ and $E_N[\\varphi^{n+1}]-E_N[\\varphi^n]$ holds.","core_discovery":"On its own terms, the paper's central claim is that the nonlinear energy-based SAV method and the Lagrange multiplier SAV method are two ends of a single spectrum, and that the spectrum removes each method's defect. With auxiliary variable $r(t)=\\sqrt{E_N[\\varphi]+C}$, the weighted scheme (3.3) couples the update of $r^{n+1}$ to the actual change in the nonlinear energy $E_N[\\varphi^{n+1}]-E_N[\\varphi^n]$ through the weight $\\lambda$. Theorem 3 states that for any time step $\\tau>0$ there exists $\\bar\\lambda\\in(0,1)$ such that for all $\\lambda\\in(\\bar\\lambda,1]$ the resulting nonlinear scalar equation (3.7) has a solution, so the scheme is solvable for arbitrary time steps. Theorem 4 proves unconditional dissipation of the weighted energy $E[\\varphi,r] = \\tfrac12(\\varphi,L\\varphi)+\\lambda(r^2-C)+(1-\\lambda)E_N[\\varphi]$, and shows that when $2r^{n+1}(r^{n+1}-r^n)\\ge E_N[\\varphi^{n+1}]-E_N[\\varphi^n]$ the original energy dissipates as well. By bisection on $\\lambda$, the method selects the smallest weight for which the scalar equation is solvable at each step, numerically yielding an energy close to the original while avoiding the nonexistence failure of the pure Lagrange multiplier approach.","pith_inferences":["Because the existence proof relies on the uniform bound $|E_N|\\le M$, the theoretical guarantee applies to the truncated double-well potential used in the numerics; extending Theorem 3 to the untruncated quartic phase-field energy would require an argument that does not use uniform boundedness.","The observed power-law decay of the energy discrepancy with respect to $\\lambda$ in Table 3 suggests a provable rate of the form $\\|E_\\lambda-E_0\\|\\sim\\lambda^p$, which could guide the choice of the initial bracket in the bisection algorithm.","The constrained-optimization reformulation in Remark 2 could be solved directly instead of by bisection, potentially yielding even smaller weights and therefore energies closer to the original.","The weighting idea should transfer to exponential or logarithmic auxiliary variables and to other linear multi-step or Runge–Kutta discretizations, since the stability proofs are variational and the scalar equation remains one-dimensional."],"forward_implications":["First-order backward Euler and second-order Crank–Nicolson weighted SAV schemes retain their temporal convergence orders while permitting arbitrarily large time steps once the weight is chosen sufficiently large.","The adaptive minimal-weight strategy makes the computed discrete energy approximate the original free energy more closely than the pure nonlinear energy-based SAV scheme, as measured by energy deviation in the numerical experiments.","When $2r^{n+1}(r^{n+1}-r^n)\\ge E_N[\\varphi^{n+1}]-E_N[\\varphi^n]$, the scheme dissipates the original energy, not merely a modified energy.","The computational cost stays consistent with classical SAV methods: two constant-coefficient linear elliptic solves per step plus one scalar root-finding problem.","The method covers $L^2$, $H^{-1}$, and fractional $H^{-\\nu}$ gradient flows, including Allen–Cahn, Cahn–Hilliard, and space-fractional Cahn–Hilliard equations, and preserves mass conservation for the Cahn–Hilliard family."],"supporting_citations":[{"why":"Supplies the nonlinear energy-based SAV approach and its modified-energy stability result that the weighted method generalizes.","marker":"[38]"},{"why":"Introduces the Lagrange multiplier SAV approach whose original-energy stability is recovered in the $\\lambda=0$ limit.","marker":"[10]"},{"why":"Proves solution existence for the Lagrange multiplier SAV scheme for sufficiently small time steps, the limitation that Theorem 3 overcomes for large steps.","marker":"[35]"},{"why":"Demonstrates numerically that the Lagrange multiplier scalar equation may have no solution for large time steps, the failure mode the weight is designed to avoid.","marker":"[3]"},{"why":"Introduces relaxation into SAV methods to improve energy consistency, the gap the weighted approach addresses by construction.","marker":"[32]"},{"why":"Provides the linear multi-step SAV framework and numerical stability baselines used for the second-order schemes.","marker":"[25]"}],"fun_headline_variants":["Weighted SAV: one parameter fixes step size and energy","SAV with a dial: tune weight for solvable steps and true energy","Blending SAV schemes: any step size, energy near original","A single weight bridges two SAV solvers, fixing both flaws","Weighted SAV method: solvable for all steps, energy stays close"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The solvability theorem assumes the nonlinear part of the free energy, $E_N$, is uniformly bounded; the paper enforces this by truncating the double-well potential in all numerical tests, so the guarantee does not directly cover the standard untruncated quartic phase-field energy.","fun_headline_variants_meta":{"raw":{"variants":["Weighted SAV: one parameter fixes step size and energy","SAV with a dial: tune weight for solvable steps and true energy","Blending SAV schemes: any step size, energy near original","A single weight bridges two SAV solvers, fixing both flaws","Weighted SAV method: solvable for all steps, energy stays close"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1615,"prompt_tokens":1010,"completion_tokens":605,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":511}},"tokens_in":626,"tokens_out":605,"duration_ms":5364,"temperature":1.0,"reasoning_tokens":511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:09:39.956643+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the first-order scheme (3.3) on the Cahn–Hilliard equation with an untruncated quartic $E_N$ and a large time step, following Algorithm 1; if at some step no $\\lambda\\in[0,1]$ makes the scalar equation (3.7) solvable, then Theorem 3 fails for the untruncated energy. A more direct check is to exhibit a configuration where the quadratic coefficient in (3.8) is positive but $f(x;\\lambda)+g(x;\\lambda)$ stays positive for all real $x$ at every $\\lambda\\in[0,1]$, contradicting the claimed solution existence.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the nonlinear energy-based SAV approach and its modified-energy stability result that the weighted method generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Lagrange multiplier SAV approach whose original-energy stability is recovered in the $\\lambda=0$ limit."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves solution existence for the Lagrange multiplier SAV scheme for sufficiently small time steps, the limitation that Theorem 3 overcomes for large steps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates numerically that the Lagrange multiplier scalar equation may have no solution for large time steps, the failure mode the weight is designed to avoid."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces relaxation into SAV methods to improve energy consistency, the gap the weighted approach addresses by construction."},{"cited_title":"Z., Zhang, G.: Linear mu lti-step methods and their numerical stability for solving gradient flow equations","cited_arxiv_id":null,"evidence_quote":"Provides the linear multi-step SAV framework and numerical stability baselines used for the second-order schemes."}],"review_version":1}