{"id":"573f6f41-c809-4500-b937-14b827edf133","arxiv_id":"2411.14226","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A regularization and passivity-preserving POD-DEIM model reduction framework for quasilinear magneto-quasistatic field-circuit coupled problems, with an output-perturbation method to restore passivity after DEIM.","lead":"This paper develops a way to shrink the size of computer models for low-frequency electromagnetic devices while preserving a key physical property called passivity, meaning the model cannot create energy on its own. It also shows how to patch a reduced model when an approximation step breaks passivity, using a small correction to the output.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The full column rank of X2 is load-bearing for Theorems 5 and 6 but is only cited to a nonexistent 'Assumption 1 e)(iii)'; it must be stated explicitly.","rationale":"I agree with the reader's identification. This is not merely a typographical issue: the reference to Assumption 1 e)(iii) cannot be satisfied because no item (iii) exists, and the condition is mathematically essential. The proof of Theorem 5 explicitly uses X2 full column rank to establish that \\hat Y_{C2}^T X2 is full rank; the ODE transformation (31) uses (X2^T \\hat Y \\hat Y^T X2)^{-1}. Under the stated assumptions only, X2 can be zero or rank-deficient, so the regularization and passivity-preserving reduction are not guaranteed for all problems the paper claims to cover. The fix is a one-line addition to Assumption 1; the verdict CONDITIONAL is appropriate. Secondary gaps (zero initial state in the passivity enforcement, unproven negativity of µ in the Lipschitz bound) are also present but are less fundamental than the missing rank condition.","tokens_in":24982,"tokens_out":6536,"duration_ms":61740,"concrete_test":"Construct a simply connected 3D (or 2D) domain with a mesh satisfying Assumption 1 a) and two winding functions satisfying e)(i)-(ii) but with both supports in the conducting subdomain (so X2 = 0), or a case where the two columns of X2 are linearly dependent. Assemble the FEM matrices from Section 3.1 and check the rank of X2, then attempt to build the projector \\hat Y_{C2} and verify whether the matrix G1 = Er - Jh Q in Theorem 5 is nonsingular. If G1 becomes singular (or the construction of Z in (31) fails), the missing full-rank hypothesis is confirmed as load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central regularization and the ODE transformation both depend on the matrix X2 ∈ R^{n2×m} having full column rank. Section 3.1 (after eq. (10)) asserts this 'due to Assumption 1 e)(iii)', but Assumption 1 e) terminates at item (ii) (divergence-free and disjoint supports). The rank condition is used in the proof of Theorem 5 to conclude that \\hat Y_{C2}^T X2 has full column rank and hence that the DAE is index-one, and in Theorem 6 and the subsequent construction of the ODE system (31) where (X2^T \\hat Y_{C2} \\hat Y_{C2}^T X2)^{-1} appears. If X2 is rank-deficient, the condensed form (25) and the regularized index-one property can fail. The physical assumptions do not imply the rank condition: windings whose support lies entirely in the conducting subdomain give X2 = 0, and disjoint supports do not prevent linear dependence of the projected winding traces. The condition should either be added as an explicit hypothesis (e.g., in Assumption 1) or proven from the FEM mesh and winding regularity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies the quasilinear magneto-quasistatic FEM system (8) with circuit coupling. It proposes a regularization that projects out singular state components via a condensed form, transforms the regularized DAE into an ODE, proves passivity of the FEM, regularized, and POD-reduced models, and develops a perturbation-based passivity enforcement for POD-DEIM-reduced models driven by a computable state-error bound. Numerical experiments on a 2D transformer illustrate the error bounds and the small output perturbation.","tokens_in":25259,"tokens_out":10759,"duration_ms":101667,"significance":"If the claims hold, the paper delivers a constructive route from a singular 3D FEM MQS DAE to a passive reduced ODE model with a computable a priori bound for DEIM-induced loss of passivity. The storage-function arguments in Theorems 7, 9, and 11 are direct, and the bound in Theorem 15 is explicit in terms of reduced quantities only. These are useful, concrete results for circuit-coupled electromagnetic simulation. The main gaps are missing hypotheses rather than errors in the derivations themselves.","major_comments":[{"comment":"The sentence \"due to Assumption 1 e)(iii) the matrix X2 has full column rank\" refers to an item that does not exist; Assumption 1 e) lists only (i) and (ii). The full-column-rank condition is used in the proof of Theorem 5 to conclude that Yhat_{C2}^T X2 has full column rank, in Theorem 6 to ensure the condensed form, and in Eqs. (26) and (31) where (X2^T Yhat_{C2} Yhat_{C2}^T X2)^{-1} appears. If X2 is rank deficient, e.g., when a winding is supported in the conducting subdomain, the regularized index-one DAE and the subsequent ODE transformation can fail. This condition should be added as an explicit hypothesis in Assumption 1, or an equivalent condition should be proved from the mesh and winding assumptions.","section":"Section 3.1, after Eq. (10)"},{"comment":"The conclusion that the integral of u^T y_POD over [0,t] is nonnegative uses io-passivity of the POD-reduced model, which by Definition 2 and Remark 3 requires zero initial state E x(0)=0. The text does not state x_POD,0=0 before this inequality, and in the general setup of Sections 2 and 3 the initial field A0 is arbitrary. For a nonzero initial state, the free-response contribution to y_POD can make the integral negative, so the perturbation rule (47) need not enforce io-passivity. Add a zero-initial-state hypothesis, or an estimate involving S_POD(x_POD,0), to the passivity enforcement statement.","section":"Section 5, inequality (45)"},{"comment":"The claim that mu = min(mu1, mu2) < 0 is not established. It requires U^T C_d^T M_f C_d U and U_{a1}^T C_1^T M_{f,1} C_1 U_{a1} to be positive definite, but only positive semidefiniteness follows from the given information; in 3D, C_d has a nontrivial kernel, so the reduced matrices can be singular, and A_l is only shown negative semidefinite, so lambda_max(A_l)=0 is possible. If mu=0, the formula in Theorem 15 must be interpreted as a limit and the statement that theta(t) remains bounded needs separate justification. Please state an explicit positive-definiteness condition for the reduced curl Gramians or prove it from the construction of U.","section":"Theorem 13, Eq. (54) and following claim"},{"comment":"The derivation of Delta_DEIM contains two inequalities written with 'lesssim', but the theorem states (49) as a deterministic bound without quantifying the constants. Since the perturbation in (47) is chosen from theta(t), an unquantified constant in Delta_DEIM weakens the guarantee that the perturbed system is io-passive. Please either prove the bound with explicit constants under stated assumptions on the snapshot richness, or state the theorem with the same 'lesssim' convention and discuss the validity of the passivity guarantee.","section":"Theorem 12, proof around Eq. (50)"}],"minor_comments":[{"comment":"The phrase 'Thought for general systems' should read 'Although for general systems'.","section":"Section 4, first paragraph"},{"comment":"The word 'entrees' should be 'entries' in the definition of the matrices M_f and M_{f,1}.","section":"Theorem 13, display (54)"},{"comment":"The displayed error bounds divide by L2[f] and by mu; the paper should state the limiting interpretation when these quantities vanish, since zero is not excluded by the current assumptions.","section":"Theorems 12 and 15"},{"comment":"The text sometimes says 'passivity enforcement', but the proven property is io-passivity with zero initial state; the terminology should be aligned to avoid overstating the result.","section":"Section 5 and Conclusion"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core of this paper is credible and the contributions are real. The authors regularize the quasilinear 3D MQS DAE by projecting out singular state components, prove that POD reduction of the resulting ODE form preserves passivity, and then show that passivity lost through DEIM can be recovered by adding a small output perturbation whose size is controlled by a computable DEIM error bound. The storage-function arguments are clean, and the passivity-preservation proof for the POD-reduced model (Theorem 11) is the best part: it uses the projected basis functions directly rather than forcing the reduced system into a special form.\n\nWhat is new: passivity preservation for POD on quasilinear MQS systems and the DEIM passivity-enforcement-by-output-perturbation. The regularization is an extension of the authors' earlier linear condensed-form work, but the quasilinear extension is nontrivial and the condensed-form derivation is careful.\n\nThe soft spots are real but fixable. First, Section 3.1 asserts that X2 has full column rank \"due to Assumption 1 e)(iii)\", but Assumption 1 e) stops at (ii). That is not just a typo: the physical assumptions do not imply the rank condition. X2 can be zero if the winding support lies inside the conductor, and disjoint supports do not prevent linear dependence of the projected winding traces. This rank condition is used in the proof of Theorem 5 and in the construction of the ODE system (31), so it must be stated explicitly as a hypothesis or proved from the FEM mesh and winding regularity. Second, the io-passivity enforcement argument uses the fact that the POD-reduced system is io-passive, but io-passivity as defined requires zero initial state. The paper never states x_POD(0)=0 or x_DEIM(0)=0. If the initial state is nonzero, the first integral in (45) can be negative, and the proposed δ(t) may not suffice. This is easy to fix by assuming zero initial data or by adding an initial-energy term to δ. Third, the conclusion that µ=min(µ1,µ2)<0 needs the matrices U^T C_d^T M_f C_d U and U_a1^T C_1^T M_{f,1} C_1 U_a1 to be positive definite, not merely positive semidefinite. The paper does not prove that the reduced discrete curl matrices have full column rank; likely true after the regularization, but it should be stated.\n\nThe numerical experiment is 2D only and no code or data are provided, so it does not validate the 3D machinery beyond consistency. That said, the computed error bound θ2 is sharp, and the constants in Table 1 are computed, not fitted to the reported outputs. I see no circularity in the derivations.\n\nBottom line: this paper deserves a serious referee. I would send it out, asking the authors to add the missing rank assumption, clarify the zero-initial-state condition in the passivity enforcement theorem, and justify the positive-definiteness needed for µ<0. These are moderate revisions, not a rewrite.","headline":"A sound, genuinely useful paper on passivity-preserving MOR for quasilinear MQS systems; the main theorems hold, but the X2 rank hypothesis and a zero-initial-state condition need explicit fixing before acceptance.","tokens_in":25743,"tokens_out":4713,"would_cite":false,"duration_ms":49109,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["12H20","15A22","34A09","37L05","78A30","93A15","93C10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The singular components of FEM-discretized quasilinear magneto-quasistatic systems can be projected out, POD model reduction then preserves passivity, and a DEIM-error-sized output perturbation restores passivity when DEIM breaks it.","keywords":["magneto-quasistatic systems","differential-algebraic equations","passivity","regularization","model order reduction","proper orthogonal decomposition","discrete empirical interpolation","passivity enforcement"],"falsifier":"Construct a 3D MQS finite element model with two windings placed in the same non-conducting region so that the coupling matrix $X_2$ has linearly dependent columns, and recompute the condensed form of Theorem 6 and the index-one criterion of Theorem 5; if the pencil remains regular and the criterion holds, the full-rank condition is dispensable, while if either fails, the regularization guarantee collapses. Alternatively, in the paper's own 2D transformer, search for an admissible input where $\\int_0^t u(\\tau)^T y_\\delta(\\tau)\\,d\\tau<0$ for some $t$ despite the bound (56), which would invalidate the passivity-enforcement claim.","tokens_in":24773,"feed_emoji":"🧲","tokens_out":11799,"duration_ms":96402,"temperature":0.7,"pith_summary":"Low-frequency electromagnetic devices coupled to electrical circuits are modelled by quasilinear magneto-quasistatic (MQS) equations; after finite element (FEM) discretization on a 3D domain, the semidiscrete system is a singular differential-algebraic equation (DAE). The paper shows that a constant coordinate transformation can identify and project out the singular state components, turning the DAE into a regular index-one system and then into an ordinary differential equation without changing its input-output behaviour. The paper then proves that proper orthogonal decomposition (POD) model reduction of this system preserves passivity, i.e. the reduced model never generates energy on its own, which is exactly what is needed for safe coupling to circuit simulators. Because the discrete empirical interpolation method (DEIM) used to accelerate the nonlinear evaluation destroys the symmetric structure behind passivity, the paper restores input-output passivity by perturbing the output by an amount controlled by a computable DEIM error bound.","feed_headline":"POD reduction keeps magneto-quasistatic models passive","feed_subtitle":"A DEIM error bound sizes a small output perturbation that restores the energy balance.","key_machinery":"The central object is the condensed form of the matrix pencil $\\lambda E_r - A_r(x_r)$ under a constant coordinate transformation $W$ applied by congruence, which yields blocks $E_{11}$, $I$, $0$ on the energy side and $A_{11}(x_r)$, $0$, $I$ on the operator side; this splitting separates the regular dynamics from the zero- and infinite-eigenvalue parts, and projecting out the infinite part is what regularizes the DAE. Passivity is carried by the magnetic-energy storage function built from the reluctivity integral $\\vartheta(\\xi,s)=\\int_0^{\\sqrt{s}}\\nu(\\xi,\\zeta)\\zeta\\,d\\zeta$, whose dissipation rate is nonnegative because the matrices $M_{11}$ and $R$ are positive definite. For the DEIM correction, the load-bearing estimate is the bound $\\|\\varepsilon(t)\\|\\le \\theta(t)=\\Delta_{\\mathrm{DEIM}}\\mu^{-1}(e^{\\mu t/\\lambda_{\\min}(E)}-1)$, obtained from a logarithmic Lipschitz constant and an exponential comparison inequality for differential inequalities; $\\theta(t)$ sizes the output perturbation.","core_discovery":"On its own terms, the paper establishes a chain of structural guarantees for the quasilinear MQS system discretized by edge elements. The singular components of the FEM DAE can be removed by a constant coordinate transformation: Theorem 5 shows the condensed system has tractability index one and Theorem 6 gives a simultaneous block-diagonalization of the pencil separating regular, zero, and infinite parts. The original FEM model, the regularized DAE, and the equivalent ODE are all passive with the magnetic energy as storage function. POD reduction applied to the structured ODE preserves passivity (Theorem 11). When DEIM is used to approximate the nonlinearity, the symmetric structure is lost, so passivity can fail; the paper proves that adding an output perturbation $\\delta(t) u$ with $\\delta(t)=\\|C\\|\\theta(t)/\\|u(t)\\|$ restores input-output passivity, where $\\theta(t)$ is an explicitly computable DEIM state-error bound (Theorem 15).","pith_inferences":["The same output-perturbation mechanism should work for any approximation of the nonlinearity that comes with a computable state-error bound, not only for DEIM; any such error bound could be plugged into the $\\delta(t)$ formula.","If the winding-coupling matrix $X_2$ is rank-deficient in an industrial device, the paper's chain of guarantees would need a fallback, such as a rank-revealing preprocessing or a generalized-inverse regularization.","The paper proves input-output passivity of the perturbed DEIM system but does not construct a storage function for it, so a state-space passivity certificate would need an additional reachability or storage-function argument.","Because $\\theta(t)$ grows exponentially in $\\mu t/\\lambda_{\\min}(E)$, the perturbation may become large over long time horizons; a tighter, problem-adapted bound or a different enforcement strategy would be needed for long-time simulation."],"forward_implications":["Reduced models of low-frequency electromagnetic devices can be coupled to circuit simulators with a guaranteed nonnegative energy balance, not just with good approximation accuracy.","The projection-based regularization gives a systematic route from a singular FEM DAE to an ODE, so standard ODE model reduction techniques apply without tree-cotree or grad-div gauging.","POD-reduced MQS models are passive by construction and need no post-processing or passivity enforcement.","For POD-DEIM models, passivity enforcement adds only a small, bounded perturbation to the output; in the paper's 2D transformer experiment the perturbed output stays close to the POD-DEIM output.","The DEIM state-error bound $\\theta(t)$ is computed from reduced matrices only, so the passivity guarantee can be evaluated without re-solving the full FEM model."],"supporting_citations":[{"why":"Supplies unique solvability, monotonicity, and Lipschitz properties of the continuous MQS model used throughout the paper.","marker":"[4]"},{"why":"Establishes passivity of the quasilinear MQS system and the magnetic-energy storage function that all later passivity proofs imitate.","marker":"[5]"},{"why":"Provides the POD-DEIM model reduction framework and the transformation of the regularized DAE into an ODE that the paper builds on.","marker":"[17]"},{"why":"Gives the linear version of the condensed-form regularization that the quasilinear Theorem 6 generalizes.","marker":"[18]"},{"why":"Supplies the projector-based tractability-index framework used in Theorem 5 to certify index one.","marker":"[29]"},{"why":"Supplies the DEIM approximation and the standard DEIM error estimate underlying the state-error bound and passivity enforcement.","marker":"[30]"},{"why":"Documents the block structure shared by FEM and FIT discretizations and provides the 2D transformer benchmark used in the numerical experiments.","marker":"[11]"}],"fun_headline_variants":["Passivity preserved in MQS reduction via POD and DEIM","Output perturbation restores passivity in MQS reduced models","Regularized DAEs enable passive model reduction for MQS","A small output fix keeps MQS models passive with DEIM","POD and DEIM yield passive reduced models for quasilinear MQS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole chain of guarantees assumes that the winding currents couple to the non-conducting part of the finite element mesh in linearly independent ways, i.e. the matrix $X_2$ has full column rank; the proofs use this condition, but the stated assumptions never list it.","fun_headline_variants_meta":{"raw":{"variants":["Passivity preserved in MQS reduction via POD and DEIM","Output perturbation restores passivity in MQS reduced models","Regularized DAEs enable passive model reduction for MQS","A small output fix keeps MQS models passive with DEIM","POD and DEIM yield passive reduced models for quasilinear MQS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1579,"prompt_tokens":978,"completion_tokens":601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":515}},"tokens_in":594,"tokens_out":601,"duration_ms":6451,"temperature":1.0,"reasoning_tokens":515,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:25:51.031492+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a 3D MQS finite element model with two windings placed in the same non-conducting region so that the coupling matrix $X_2$ has linearly dependent columns, and recompute the condensed form of Theorem 6 and the index-one criterion of Theorem 5; if the pencil remains regular and the criterion holds, the full-rank condition is dispensable, while if either fails, the regularization guarantee collapses. Alternatively, in the paper's own 2D transformer, search for an admissible input where $\\int_0^t u(\\tau)^T y_\\delta(\\tau)\\,d\\tau<0$ for some $t$ despite the bound (56), which would invalidate the passivity-enforcement claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the POD-DEIM model reduction framework and the transformation of the regularized DAE into an ODE that the paper builds on."},{"cited_title":"In: Beattie, C., Benner, P., Embree, M., Gugercin, S., Lefteriu, S","cited_arxiv_id":null,"evidence_quote":"Gives the linear version of the condensed-form regularization that the quasilinear Theorem 6 generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the block structure shared by FEM and FIT discretizations and provides the 2D transformer benchmark used in the numerical experiments."}],"review_version":1}