{"id":"129d8d59-15ee-400b-b9fb-4ec6d5f8d4f0","arxiv_id":"2507.14521","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A convex variational formulation for vector potential magnetostatics with energy-based hysteresis is derived, with existence, uniqueness, and global convergence of two finite element solvers.","lead":"This paper shows that a magnetostatic field problem with an energy-based hysteresis model can be rewritten as one convex minimization problem, and it proves that two finite element solvers for this problem always converge. For engineers simulating transformers and motors with ferromagnetic cores, this offers a rigorous way to include hysteresis, the main cause of core losses.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The global convergence proof of the regularized Newton method (Assertion 9) is delegated to unpublished self-cited preprints [28],[29] without verifying their assumptions for the saturating energy U, leaving the abstract's second main contribution unsupported within the manuscript.","rationale":"The reader identified the same load-bearing concern: the global convergence of the regularized Newton method is asserted on the basis of external preprints rather than demonstrated in the manuscript. I agree that this is the weakest point of the central claim. The variational equivalence itself (Assertion 6) is well argued: the first-order optimality conditions derived from (9) coincide with the strong form (10)–(11) and with the hysteresis minimization (3), so the transfer from the minimizer of (9) to a solution of (1)–(3) is credible. The block-coordinate descent result (Assertion 11) cites a published paper [30], so it is less problematic, although the multi-pinning generalization used in the benchmarks is not proven. The decisive gap is that Assertion 9's proof is effectively a pointer to [28] and [29], which the paper does not verify against the specific structure of the energy U(J) and the pinning term. Since the abstract explicitly promises 'two globally convergent iterative methods,' and Newton is the faster of the two, the missing verification is load-bearing. The intended fix is addressable: either fold the convergence proof into the paper or state the precise assumptions under which the external results apply. Hence the existing CONDITIONAL verdict is appropriate; no change is needed.","tokens_in":10291,"tokens_out":9578,"duration_ms":123970,"concrete_test":"Formally verify the transfer step: write the optimality system as a monotone operator equation T(A,J) = 0 with T = (curl(ν0(curlA−J)), −ν0 curlA + ν0J + U′(J) + χ(J−Jp)/|J−Jp|ε) and check whether T satisfies the strong-monotonicity and Lipschitz conditions assumed in Heid [28] and Egger–Engertsberger–Radu [29] on the sublevel set {f ≤ f(A0,J0)} of the objective (9). Compute the monotonicity modulus as |J| → Js: if it tends to zero or the Lipschitz constant of U′(J) blows up, the infinite-dimensional global linear convergence proof does not transfer and Assertion 9 must be weakened to local or finite-dimensional-only convergence. An independent check is to rerun the Newton iteration of Section III-B from a starting field with |J| = 0.99 Js on the single-cell model; if the Armijo step count grows without bound, the global claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central advertised contributions are the variational characterization (Assertions 3–6) and two globally convergent solvers (Assertions 9 and 11). Assertion 6 seems sound: the first-order conditions of the strongly convex problem (9) reproduce (10)–(12), and these are exactly the field equations plus the hysteresis minimization (3). The load-bearing gap is Assertion 9. Its proof in the manuscript is a single sentence: the discrete case is 'fully covered by [27]' and the infinite-dimensional case is delegated to preprints [28],[29]. The objective in (9) contains U(J) = −2AsJs/π log(cos(π/2 |J|/Js)), whose gradient behaves like tan(π/2 |J|/Js) and blows up as |J| → Js. Thus the Hessian is not globally Lipschitz on Q; it is Lipschitz only on sublevel sets where |J| stays bounded away from saturation. Standard global Newton–Armijo theorems require a Lipschitz Hessian and bounded sublevel sets, so the finite-dimensional argument from [27] needs a quantification of how the monotonicity and smoothness constants behave as iterates approach saturation. Preprints [28] and [29] address nonlinear magnetostatics, but the pinning term χ(J−Jp)/|J−Jp|ε enters the derivative and the energy has a singular boundary; it is not automatic that their assumptions hold here. Since Assertion 9 is the support for the claimed mesh-independent linear convergence of the faster solver, this is a genuine unsupported step. It does not invalidate the variational equivalence, but it does mean the manuscript delivers only the formulation, not the promised convergence proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a variational reformulation of magnetoquasistatic field equations coupled to the energy-based vector hysteresis model of Henrotte et al. For a single load step, the coupled system (1)-(3) is shown to be equivalent to the minimization of a strongly convex functional (9) over the vector potential A and the magnetic polarization J. The authors claim well-posedness of the continuous and discrete problems, present a regularized Newton method and a block-coordinate descent method, and illustrate the approach with numerical experiments on the T-joint transformer benchmark.","tokens_in":10659,"tokens_out":14849,"duration_ms":187132,"significance":"The central variational equivalence is a genuinely useful contribution: if it holds, it provides a clean formulation that avoids the numerical inversion of the hysteresis operator needed in earlier scalar-potential or vector-potential approaches. The first-order optimality argument in Assertion 6 is self-contained and non-circular, and the discrete well-posedness in Assertion 7 is immediate. The two proposed solvers, if fully justified, would give practitioners two globally convergent options. However, the manuscript's support for the Newton method is only a reference to the authors' preprints, and the same is largely true for the block-coordinate descent method. Thus the practical significance of the paper depends on completing or making explicit those convergence proofs.","major_comments":[{"comment":"The proof of Assertion 9 is a two-sentence reference to [27] and the preprints [28], [29], yet this assertion carries one of the two advertised global-convergence claims. The objective in (9) contains U(J) = -2 As Js / pi log(cos(pi/2 |J|/Js)), whose gradient is unbounded as |J| tends to Js; the energy is finite only for |J| < Js almost everywhere. Consequently the Newton operator (19) contains U''(J), which is not a bounded multiplication operator on L2 unless the iterates are pointwise bounded away from saturation. Standard damped-Newton global-convergence theorems require assumptions such as Lipschitz continuity of the Hessian on the relevant sublevel set and control of the Newton decrement; the manuscript does not verify these for (9), and it is not automatic that the magnetostatics results in [28] or [29] cover the singular boundary and the pinning term chi |J - Jp|_eps. Please provide a self-contained proof of the claimed global linear convergence, or state the theorem from [28]/[29] and verify its hypotheses step by step.","section":"Section III-B, Assertion 9"},{"comment":"The proof of Assertion 11 likewise consists of a reference to [30] with the sentence 'The results follow immediately'. The abstract's second global-convergence claim therefore also rests on an external result whose hypotheses are not stated. In particular, [30] concerns a non-smooth convex minimization problem in plasticity, not the present coupled field problem; the linear-rate conclusion for the infinite-dimensional Hilbert space setting of (23) is not immediate. Please give the precise convergence theorem from [30] and check its assumptions, or prove the assertion directly.","section":"Section III-C, Assertion 11"}],"minor_comments":[{"comment":"The displayed formula for U(J) omits the absolute value; as written it defines a scalar function of a vector argument. It should read U(J) = -2 As Js / pi log(cos(pi/2 |J| / Js)).","section":"Assumption 2"},{"comment":"The denominator in the pinning-term derivative is missing from the displayed formula: it should read chi / |J - Jp|_eps (I - v otimes v), not chi |J - Jp|_eps (I - v otimes v).","section":"Remark 10 and equation (19)"},{"comment":"The phrase 'mesh-independent with constant average iteration counts' is presented as an expected and confirmed property, but Assertions 9 and 11 only state linear convergence and no mesh-independent rate is proved. Please either state a rate theorem or qualify this conclusion as an empirical observation.","section":"Section IV-D"},{"comment":"The name 'Berqvist' should be 'Bergqvist'.","section":"Introduction, second paragraph"},{"comment":"The numerical section uses a five-cell model with multiple pinning forces, whereas the analysis in Sections II and III is stated for a single internal variable. Please state explicitly that the uniqueness and convergence assertions extend to the multi-cell case, for example by noting that the functional (26) is again strongly convex and satisfies the same structural assumptions.","section":"Section IV"},{"comment":"Reference [29] is cited only as an arXiv preprint. Since the proof of Assertion 9 relies on it, please provide a full statement of the cited result or update the reference to a published version if one exists.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's main theoretical novelty is the variational equivalence, which is clean and likely correct. The global-convergence claims for the two solvers, especially Assertion 9, are supported only by references to the authors' own preprints, which makes external validation difficult. I would advise the editor to require a self-contained proof or a detailed verification of the cited hypotheses before publication. The citation pattern is not improper, but it does place a burden on the reader."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core idea is a simultaneous minimization over A and J for the Henrotte–François-Lavet hysteresis model, and the equivalence proof (Assertion 6) is clean and genuinely useful. The paper is worth taking seriously. But the advertised global convergence result for the regularized Newton method is not proved here; it is outsourced to two self-cited preprints, and the stress-test note is right that the singular energy makes the transfer non-automatic. That is the main soft spot, and it is addressable.\n\nWhat is actually new: the variational characterization (9) with the saturating energy and the regularized pinning term. The first-order conditions reproduce the field equations and the hysteresis minimization, so existence, uniqueness, and a finite element discretization follow by convexity. That is a real contribution. The block-coordinate descent interpretation (Assertion 11) is also sensible, and the numerical results show mesh-independent iteration counts, which is what you expect from a strongly convex problem.\n\nThe soft spots, in order of size. First and biggest: Assertion 9. The proof is one sentence delegating to a textbook (for the discrete case) and to preprints [28],[29]. The stress-test note identifies the real issue: the Hessian contains terms like tan(pi/2 |J|/Js), which blows up as |J| approaches Js, so the Hessian is not globally Lipschitz. Standard global Newton–Armijo theorems require bounded sublevel sets and Lipschitz Hessians, and the preprints are about nonlinear magnetostatics without this singular boundary. It is plausible the argument can be adapted, but that has not been shown here. The claim of two globally convergent methods is therefore overstated for the Newton method.\n\nSecond: the 2D benchmark uses the 5-pinning-force extension, but the theory section only covers a single internal variable. The paper says everything generalizes immediately, and it probably does, but there is no proof. Minor.\n\nThird: the numerical section has no comparison with existing formulations, e.g., the inverse-operator approach of Jacques et al. or scalar-potential methods. The tables show iteration counts and losses, not accuracy against a reference. Acceptable for an illustration, but it limits what we can conclude about efficiency.\n\nThe citation pattern is fine: self-cites are used where prior work is needed, and the core derivation is self-contained. The central equivalence proof does not depend on the preprints, so the main formulation stands regardless.\n\nBottom line: this deserves a serious referee. The referee should push the authors to either prove the Newton convergence for this specific energy or downgrade the claim to local convergence or empirical observation. With that fixed, it is a solid methods paper for computational electromagnetics.","headline":"A clean variational formulation for vector-potential hysteresis that deserves referee time, but the Newton convergence claim is delegated to unpublished preprints and needs either proof or softening.","tokens_in":11156,"tokens_out":2512,"would_cite":true,"duration_ms":29580,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65K10","35Q61","78M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that a magneto-quasistatic field problem coupled to an energy-based magnetic hysteresis model is equivalent to one strongly convex minimization problem, and supplies two globally convergent iterative solvers for it.","keywords":["ferromagnetic materials","vector potential formulation","finite element analysis","minimization algorithms","magnetic hysteresis","convex variational problem","Newton method","block-coordinate descent"],"falsifier":"Run the block-coordinate descent and the Newton method on the two-dimensional T-joint benchmark with a deliberately poor initial guess (e.g. $A_0=0$, $J_0$ far outside the polarization range), and check whether the iterates' residuals in the optimality system (10)-(11) decrease to zero at the claimed linear rate; if either method stalls or converges to a point that does not satisfy the first-order condition, the global convergence assertions would be contradicted.","tokens_in":10096,"feed_emoji":"🧲","tokens_out":5691,"duration_ms":54282,"temperature":0.7,"pith_summary":"The paper establishes that a magneto-quasistatic field problem coupled to an energy-based magnetic hysteresis model can be reformulated as a single convex minimization problem over the magnetic vector potential and the magnetic polarization. This makes it possible to prove existence and uniqueness of solutions by standard convex analysis, and to justify finite element discretization. The authors then propose two iterative solvers, a regularized Newton method and a block-coordinate descent method, and claim both converge globally at a linear rate. A numerical simulation of a transformer T-joint confirms mesh-independent iteration counts and shows the Newton method is substantially faster. A careful reader would care because most prior vector-potential hysteresis solvers lacked a rigorous convergence guarantee.","feed_headline":"One convex problem tames magnetic hysteresis in field solvers","feed_subtitle":"A single variational formulation guarantees existence, uniqueness, and two provably convergent solvers for ferromagnetic devices.","key_machinery":"The load-bearing object is the convex functional (9): $\\int_\\Omega \\frac{\\nu_0}{2}|\\mathrm{curl}\\,A-J|^2 - \\langle H_s,\\mathrm{curl}\\,A\\rangle + U(J) + \\chi |J-J_p|_\\varepsilon\\,dx$, minimized over $A\\in V\\subset H_0(\\mathrm{curl})$ and $J\\in L^2(\\Omega)^3$. Strong convexity in $(A,J)$ gives a unique minimizer and makes the first-order optimality conditions necessary and sufficient; those conditions are exactly the field equations (10)-(11) (or the inclusion (12) when $\\varepsilon=0$). The regularization $|x|_\\varepsilon$ of the pinning term makes the problem twice differentiable for $\\varepsilon>0$, allowing the Newton method, while the block method minimizes alternately in $A$ and $J$. The splitting $H_0(\\mathrm{curl})=\\nabla H^1_0 \\oplus V$ encodes gauging and boundary conditions.","core_discovery":"On its own terms, the central discovery is Assertion 6: the unique minimizer $(A,J)$ of the strongly convex functional (9), with $J$ a polarization and $B=\\mathrm{curl}\\,A$, yields fields $B$ and $H=\\nu_0(B-J)$ that satisfy the field equations (1)-(2) and reproduce the hysteresis minimization (3) for $J$. The equivalence is obtained from the first-order optimality conditions, and it extends to the non-smooth case $\\varepsilon=0$ through subdifferential inclusions. Consequently the whole coupled problem is a variational problem with a unique solution, and its discretization by Nédélec and piecewise-constant finite elements inherits that well-posedness. The paper further asserts (Assertions 9 and 11) that both a damped Newton method and a block-coordinate descent converge globally, with linear rate, to that minimizer from any starting point.","pith_inferences":["The same variational characterization might be used to couple the energy-based hysteresis model to other dissipative effects, e.g. eddy currents or mechanical deformation, by adding further convex terms to the functional; the paper does not explore this.","The block-coordinate descent method here is a slight variant of the original iterative scheme (6)-(7); if the convexity arguments are robust, they may be adaptable to prove convergence of the original scheme under an additional convexity or coercivity condition, which the paper leaves open.","The equivalence suggests a natural multigrid or domain-decomposition solver based on the functional, since the block updates are local in $J$ and elliptic in $A$.","The paper's regularization error $\\sqrt{\\varepsilon}$ (Remark 1) provides a ready-made tolerance: one can choose $\\varepsilon$ so that the error of the smooth model is below the discretization error, though the paper does not give such an adaptive strategy."],"forward_implications":["Finite element discretization of (9) inherits well-posedness; any conforming discretization yields a unique discrete solution.","The Newton method's per-iteration cost is essentially that of a linear magnetostatic problem, since the polarization increment can be eliminated locally (Remark 10).","Since the functional already contains all information of the field equations, any a posteriori error estimate for the functional directly controls the error in the fields.","The equivalence gives a constructive route to hysteresis loss computation: the pinning term $\\chi |J-J_p|_\\varepsilon$ integrated over the load step yields the loss density.","The convergence theory likely extends to multiple pinning forces and to the higher-order and curved-element discretizations mentioned in the paper, because the variational structure is preserved."],"supporting_citations":[{"why":"Supplies the energy-based variational hysteresis model in the incremental form that is incorporated into the vector-potential formulation.","marker":"[9]"},{"why":"Original energy-based vector hysteresis model with pinning, the basis of [9].","marker":"[8]"},{"why":"Standard convex analysis reference used to prove existence and uniqueness of the minimizer of (9).","marker":"[20]"},{"why":"Provides the Newton method with line search whose finite-dimensional convergence is used in Assertion 9.","marker":"[27]"},{"why":"Provides infinite-dimensional convergence theory for damped Newton methods, referenced in the proof of Assertion 9.","marker":"[28]"},{"why":"Provides global convergence of iterative solvers for nonlinear magnetostatics, referenced in the proof of Assertion 9.","marker":"[29]"},{"why":"Provides the convergence theory for block-coordinate descent / nonsmooth convex minimization used in Assertion 11.","marker":"[30]"}],"fun_headline_variants":["Convex formulation yields unique hysteresis solution","Hysteresis as convex minimization with unique fields","Two globally convergent methods for vector hysteresis","Unique solution from convex hysteresis problem","Convex optimization solves magnetic hysteresis uniquely"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that the regularized Newton method converges globally from any starting point relies on two external preprints by the same authors; if those results do not apply to the infinite-dimensional function spaces and nonsmooth pinning terms used here, the paper itself does not yet contain a complete proof of that convergence claim.","fun_headline_variants_meta":{"raw":{"variants":["Convex formulation yields unique hysteresis solution","Hysteresis as convex minimization with unique fields","Two globally convergent methods for vector hysteresis","Unique solution from convex hysteresis problem","Convex optimization solves magnetic hysteresis uniquely"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000236,"raw_usage":{"total_tokens":1456,"prompt_tokens":850,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":544}},"tokens_in":466,"tokens_out":606,"duration_ms":7143,"temperature":1.0,"reasoning_tokens":544,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:54:30.476443+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the block-coordinate descent and the Newton method on the two-dimensional T-joint benchmark with a deliberately poor initial guess (e.g. $A_0=0$, $J_0$ far outside the polarization range), and check whether the iterates' residuals in the optimality system (10)-(11) decrease to zero at the claimed linear rate; if either method stalls or converges to a point that does not satisfy the first-order condition, the global convergence assertions would be contradicted.","supporting_citations":[{"cited_title":"Francois-Lavet, F","cited_arxiv_id":null,"evidence_quote":"Supplies the energy-based variational hysteresis model in the incremental form that is incorporated into the vector-potential formulation."},{"cited_title":"Ekeland and R","cited_arxiv_id":null,"evidence_quote":"Standard convex analysis reference used to prove existence and uniqueness of the minimizer of (9)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides infinite-dimensional convergence theory for damped Newton methods, referenced in the proof of Assertion 9."},{"cited_title":"Global convergence of iterative solvers for problems of nonlinear magnetostatics","cited_arxiv_id":"2403.18520","evidence_quote":"Provides global convergence of iterative solvers for nonlinear magnetostatics, referenced in the proof of Assertion 9."},{"cited_title":"Carstensen","cited_arxiv_id":null,"evidence_quote":"Provides the convergence theory for block-coordinate descent / nonsmooth convex minimization used in Assertion 11."}],"review_version":1}