{"id":"3d51307f-3f7b-4052-9806-b34c347196a7","arxiv_id":"2607.27795","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"DG–SAV applies the discrete-gradient chain rule to the SAV-quadratized extended system, yielding three exact-modified-energy integrators, fully explicit separable-Hamiltonian schemes, and a Casimir-preserving Poisson variant.","lead":"This paper combines two standard numerical techniques — discrete gradients and scalar auxiliary variables — into one family of integrators that conserve or dissipate a modified energy exactly, with user-tunable cost. It matters because it gives simulation practitioners a practical dial between cheap energy-stable SAV schemes and accurate, nonlinear-solve discrete-gradient methods, plus a Casimir-preserving Poisson extension.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uniform positivity of E_SAV + C0 is the load-bearing premise: near its zero set the SAV coefficients and the order estimates in Theorem 4.3 degenerate, and the paper's own Appendix C flags this dependence.","rationale":"The reader's weakest_assumption — uniform positivity of E_SAV+C0 — is indeed the load-bearing premise. I verified that every theorem in Sections 4 and C uses this assumption essentially: Theorem 4.3 needs \\bar r/Σ* = 1+O(h^2) uniformly, Proposition 4.4 needs Σ*^2 ≥ δ for the contraction constant, and even Theorem 3.1's statement implicitly requires Σ* ≠ 0. The algebra of the energy identity itself is a two-line telescoping argument and is correct; the concern is not about the identity but about whether the method is defined and whether the stated orders hold in the claimed generality. The paper is honest about this dependence (Appendix C.1) and explicitly excludes unbounded energies, so the conditional verdict is right. The numerical experiments all use C0 safely away from zero, so the concern does not undermine the empirical claims as reported; it limits the scope of the framework. I do not see an additional load-bearing flaw in the Casimir extension beyond the already-stated linear-independence assumption, which is similarly explicit. Hence no change to the reader's verdict.","tokens_in":26846,"tokens_out":14197,"duration_ms":125544,"concrete_test":"Run the Predictive DG–SAV variant (Definition 4.1(iii)) on the double-well oscillator of Section 7.3 with C0 = 0 (so V(q)+C0 vanishes at q=±1) and initial condition (q0,p0)=(0,1.5), which crosses q=1. If the first crossing produces NaN/Inf or the modified energy \\hat H is no longer conserved to machine precision, the failure confirms that δ>0 is necessary. Then repeat with C0 = 1e-6 and the same trajectory: measure the observed order of the original-energy error at fixed T against a high-accuracy reference; if the error constant grows as C0→0 and the h^2 asymptotic rate is lost before the scheme becomes singular, the δ-dependence of Theorem 4.3 is not merely technical.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—Theorem 3.1's exact modified-energy identity and the second-order accuracy of the Midpoint and Predictive variants—rests on the scheme being well-defined and on the local error analysis being uniform. Both require E_SAV(x)+C0 ≥ δ > 0 on a neighborhood of the trajectory. The scheme itself divides by Σ* = sqrt(E_SAV(x*)+C0) in the x-update and in the r-update; if Σ* → 0 the step is undefined or the coefficients blow up. More subtly, the order proofs in Appendix C.1 use the expansion rbar/Σ* = 1 + O(h^2). On the constraint manifold with exact data, rbar = Σ(x_m)+O(h^2) and Σ* = Σ(x_m)+O(h^2), so the relative error is O(h^2/Σ(x_m)). The O(h^2) statement is only uniform when Σ(x_m) ≥ δ > 0. If the trajectory approaches the zero set, the residual becomes O(h^3/δ) and the global error bound carries a 1/δ factor; the claimed second order is no longer uniform. Proposition 4.4 has the same issue: its contraction constant contains ||G*||^2/(4δ), so h0 → 0 as δ → 0. The paper acknowledges this in Appendix C ('the positivity assumption ... is used to make Σ^{-1} and all SAV coefficient ratios smooth and locally bounded') and defers unbounded energies to Kemmochi–Sato [21]. This is therefore a real boundary of the framework, not an error in the algebra; the conditional verdict is appropriate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a discretization template (DG-SAV) for gradient flows, Hamiltonian systems, and Poisson systems, based on splitting the energy as E = E_DG + E_SAV, introducing an auxiliary variable r with r^2 = E_SAV + C0, and forming the modified energy E-hat = E_DG + r^2. The central identity (Theorem 3.1) shows that any scheme of the template satisfies E-hat_{n+1} - E-hat_n = h <eta, S_* eta>, so skew-symmetric S_* gives exact modified-energy conservation and negative-semidefinite S_* gives dissipation for every solvable step. Three variants are proposed: Forward Euler (first order), self-adjoint Midpoint (second order), and Predictive (second order, reduced implicitness). A projected tensor construction (Theorem 5.1) enforces selected discrete Casimir conditions for Poisson systems. Numerical experiments on the Allen-Cahn equation, a nonlocal Ohta-Kawasaki-type gradient flow, a double-well Hamiltonian oscillator, and a cubic-Casimir Poisson system support the theoretical claims.","tokens_in":27124,"tokens_out":11741,"duration_ms":99868,"significance":"The exact modified-energy identity is a simple but useful structural observation that unifies DG and SAV approaches and gives the user a tunable cost/structure trade-off. The order and well-posedness theorems are carefully stated under a uniform positivity assumption, and Appendix C supplies detailed residual calculations. The numerical benchmarks use external references and report work-precision comparisons, which is a strength. If the results hold, the Predictive variant is a valuable second-order low-cost structure-preserving scheme. The main limitation is that only the modified energy is guaranteed; original-energy behavior is controlled only through the SAV gap, and the accuracy analysis is not uniform as E_SAV + C0 approaches zero.","major_comments":[{"comment":"The order and well-posedness results are correct under the stated hypothesis E_SAV + C0 >= delta > 0, but the stated O(h^2) and O(h) global errors should be read as having constants that may blow up as delta -> 0. In particular, the expansion bar_r/Sigma_* = 1 + O(h^2) is only meaningful when Sigma_* is bounded below, and the contraction constant in Proposition 4.4 contains ||G_*||^2/(4 delta). The manuscript already acknowledges this in Appendix C, but the main text and conclusions should state more prominently that trajectories approaching the zero set of E_SAV + C0 are outside the scope of the accuracy and well-posedness theorems, and that the constants depend on 1/delta.","section":"Theorem 4.3; Appendix C.1"},{"comment":"The well-posedness of the Midpoint variant is asserted via the implicit function theorem but not demonstrated. Since this is a central second-order scheme, the authors should state the smoothness and invertibility conditions on the Jacobian of the midpoint map at (x_n, r_n, h) = (x_n, r_n, 0) that guarantee a unique local branch. Without this, the claim that Midpoint is locally well-posed is a proof sketch rather than a theorem.","section":"Section 4.3, Remark 4.5"}],"minor_comments":[{"comment":"The statement that source code and data 'are being prepared for public release' makes the numerical results not currently reproducible. Please either provide the code and data with this submission or explicitly state the reason for the delay.","section":"Section 7, Code availability"},{"comment":"The reduction of the Predictive predictor to 'q_pred = q_n + h/2 M^{-1} p_n' is correct only because the SAV coefficients depend on q alone; this should be stated explicitly to avoid the impression that the p-component of the predictor is being ignored.","section":"Section 6.2 and Corollary 6.2"},{"comment":"The convergence studies use a numerical Full DG reference (e.g., Delta t = 5e-4) or DOP853 with rtol/atol. Please state explicitly how the reference accuracy compares with the finest tested step sizes, so that the reported error floors are not misread as asymptotic orders.","section":"Section 7.1 and 7.2"},{"comment":"The bullet claiming SAV gaps 'six to seven orders of magnitude smaller' lacks the caveat given in Section 7.2 that the compared schemes use different energy splits and that the absolute gaps are not scale-invariant. Please add the caveat to the contribution bullet or soften the claim.","section":"Section 1.2, contributions bullet"}],"recommendation":"minor_revision","confidential_remarks":"The paper is mathematically sound and the central identity is verified. The main limitation is the uniform positivity of E_SAV + C0, which is stated in the theorems but should be more clearly reflected in the abstract-level claims. The numerical section is informative but currently lacks reproducible code. I recommend minor revision rather than major revision because the load-bearing algebra is correct and the issues are local presentation and scope-clarity items."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the DG–SAV framework is the real thing. The central identity, Theorem 3.1, follows directly from the discrete chain rule plus the r-update, and it holds for any frozen reference state. I re-derived the key algebra, including the rbar/Sigma expansions in Appendix C.1, and it checks out. The three variants give a clean cost–accuracy interpolation, and the split between E_DG and E_SAV is a legitimate design handle. The paper is also honest about its boundaries: unbounded energies are deferred to Kemmochi–Sato, and the polynomial double-well case has no cost advantage. That transparency earns credit.\n\nThe main soft spot is the uniform positivity assumption E_SAV + C0 ≥ δ > 0. It is load-bearing: near the zero set, Sigma^{-1} and the coefficient ratios in the order proofs blow up, and the error bounds carry a 1/delta factor. The paper flags this in Appendix C, so it is a stated boundary rather than an algebraic error, but it means the second-order claims are only uniform on regions where the quadratized term stays bounded away from zero. For energies without a usable lower bound, the framework needs separate analysis.\n\nNumerics: the benchmarks are useful, but the paper says code and data are 'being prepared for public release.' That is a real gap for a methods paper. Timings also lack variance, and the abstract's headline 'six to seven orders of magnitude smaller' SAV gap omits the split-dependence caveat that the body states. All three are fixable.\n\nThe Poisson extension is neat, and the manufactured example shows exact Casimir preservation. I would like to see where it sits relative to Cohen–Hairer's linear energy-preserving integrators, which the paper mentions but does not delineate. Minor.\n\nBottom line: solid mathematical core, honest statements of scope, and a useful framework. The empirical efficiency claims need reproduction, but the framework itself deserves referee time. I would accept for peer review and ask for code/data release in the revised version.","headline":"A genuinely useful unification of DG and SAV whose central algebra checks out; the main caveat is a well-flagged positivity assumption, and the numerical claims need the promised code/data.","tokens_in":27742,"tokens_out":1641,"would_cite":true,"duration_ms":16344,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65P10","65L05","37M15","70H05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that discrete-gradient and scalar-auxiliary-variable methods are two endpoints of a single template whose modified-energy law is determined by one quadratic form, yielding three practical integrators and an extension that p","keywords":["discrete gradient","scalar auxiliary variable","structure-preserving integration","modified energy","energy dissipation","Poisson systems","Casimir preservation","Hamiltonian systems"],"falsifier":"Integrate a scalar gradient flow ẋ = −(E_DG'(x)+E_SAV'(x)) with E_DG(x)=x²/2 and E_SAV(x)=x³, C₀=0, starting at x₀>0 so that E_SAV+C₀ crosses zero at x=0, using the Forward Euler DG–SAV variant. If the scheme remains bounded and converges at the advertised order across the crossing, the positivity assumption is not necessary; a blow-up or an order drop settles that it is.","tokens_in":26622,"feed_emoji":"⚙️","tokens_out":5413,"duration_ms":45891,"temperature":0.7,"pith_summary":"The paper tries to establish that discrete-gradient (DG) and scalar-auxiliary-variable (SAV) methods are not separate paradigms but two ends of one template: SAV quadratization lifts the dynamics to an extended state (x,r) with r²=E_SAV+C₀, and applying a discrete-gradient chain rule to the modified energy Ê=E_DG+r² yields an exact energy law whose sign is set solely by the structure matrix. This gives three concrete integrators—first-order Forward Euler, second-order self-adjoint Midpoint, and second-order Predictive with reduced implicitness—all conserving or dissipating the modified energy without any time-step restriction. A Poisson extension with a projected tensor preserves selected Casimirs exactly alongside the modified Hamiltonian. If the framework is right, structure preservation becomes a user-tunable design choice rather than a fixed property of a given integrator.","feed_headline":"One framework turns energy-preserving integration into a tunable choice","feed_subtitle":"Three new schemes conserve or dissipate the modified energy exactly while letting users trade cost against accuracy.","key_machinery":"The DG–SAV template itself: x_{n+1}−x_n = h S_* (∇̃E_DG(x_{n+1},x_n)+(r̄/Σ_*)G_*), r_{n+1}−r_n = ⟨G_*, x_{n+1}−x_n⟩/(2Σ_*), with r̄=(r_{n+1}+r_n)/2 and Σ_*=\\sqrt{E_SAV(x_*)+C₀}. The second equation is chosen so that the r² increment telescopes with the discrete chain rule into one quadratic form. The split E=E_DG+E_SAV is the central design dial: pure DG at one end, pure SAV-type quadratization at the other. The projected tensor of Theorem 5.1 replaces the Poisson tensor B by PBP to enforce selected discrete Casimir conditions.","core_discovery":"Theorem 3.1: for any instance of the DG–SAV template, Ê(x_{n+1},r_{n+1}) − Ê(x_n,r_n) = h⟨η, S_* η⟩, where η is the generalized gradient combining the discrete gradient of E_DG with the SAV terms. Consequently, if S_* is skew-symmetric the modified energy is conserved exactly, and if S_* = −M_* with M_*⪰0 it dissipates exactly—no matter how the SAV coefficients are frozen. The three variants are then shown to be first-order (Forward Euler), second-order and self-adjoint (Midpoint), and second-order (Predictive), each with local well-posedness. In the Poisson setting, the projected tensor restores the discrete Casimir condition that midpoint evaluation loses, giving exact modified-Hamiltonian","pith_inferences":["The split could be selected adaptively during a simulation: assign more of the energy to E_DG in regions where the SAV gap matters, and more to E_SAV where cost dominates; the paper's identity suggests exact conservation would hold for any per-step split.","The framework invites higher-order variants: since the energy law does not depend on the order of the scheme, symmetric composition or deferred correction of the three schemes should preserve the exact modified-energy law while raising the order—an extension the paper flags but does not test.","The positivity assumption limits the method to energies where E_SAV+C₀ stays away from zero; a robust implementation would need to detect incipient zeros and shift C₀ or redraw the split mid-simulation.","The Casimir projection requires linear independence of the selected discrete Casimir gradients; near points where two Casimir gradients become nearly dependent, the projection may become ill-conditioned, suggesting a regularized variant worth testing."],"forward_implications":["If the core identity holds, any future DG–SAV scheme with skew-symmetric S_* conserves the modified energy exactly without any step-size restriction, regardless of how the SAV coefficients are frozen.","The three variants give users a concrete menu: first-order cheap (Forward Euler), second-order self-adjoint (Midpoint), and second-order with reduced implicitness (Predictive), each with proven local well-posedness.","The projected-tensor construction means selected Casimirs can be preserved exactly even when the Casimir is nonlinear and the midpoint tensor fails the discrete condition, at the price of solving a low-dimensional projection.","In the tested dissipative benchmarks, the second-order DG–SAV variants reach full discrete-gradient accuracy while keeping the SAV gap orders of magnitude smaller than classical SAV schemes, so the modified energy tracks the physical energy more closely at similar cost."],"fun_headline_variants":["Exact energy identity drives three structure-preserving integrators","Three schemes, one identity: exact energy control","DG+SAV: a tunable trinity of energy-exact integrators","Midpoint scheme restores Casimir loss in Poisson systems","Trade cost for accuracy with exact energy in one framework"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is uniform positivity: E_SAV(x)+C₀ ≥ δ > 0 must hold along the trajectory and in the local neighborhood sampled by the scheme; if a trajectory approaches the zero set of E_SAV+C₀, the coefficient 1/Σ blows up and the order and well-posedness theorems collapse.","fun_headline_variants_meta":{"raw":{"variants":["Exact energy identity drives three structure-preserving integrators","Three schemes, one identity: exact energy control","DG+SAV: a tunable trinity of energy-exact integrators","Midpoint scheme restores Casimir loss in Poisson systems","Trade cost for accuracy with exact energy in one framework"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3266,"prompt_tokens":715,"completion_tokens":2551,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":2478}},"tokens_in":459,"tokens_out":2551,"duration_ms":16155,"temperature":1.0,"reasoning_tokens":2478,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:17:10.833245+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Integrate a scalar gradient flow ẋ = −(E_DG'(x)+E_SAV'(x)) with E_DG(x)=x²/2 and E_SAV(x)=x³, C₀=0, starting at x₀>0 so that E_SAV+C₀ crosses zero at x=0, using the Forward Euler DG–SAV variant. If the scheme remains bounded and converges at the advertised order across the crossing, the positivity assumption is not necessary; a blow-up or an order drop settles that it is.","supporting_citations":[],"review_version":1}