{"id":"cf091fe5-19c7-4459-bee4-75ffb8ddbe75","arxiv_id":"2603.03420","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Polynomial-structured nonlinear systems can be projected onto a reduced basis with all Newton-solver operators precomputed offline, so online cost is independent of full-model dimension and hyper-reduction is unnecessary.","lead":"This paper shows that when a nonlinear simulation can be written with polynomial (quadratic) terms, a reduced-order model can precompute every operator once and then run in a space whose size does not depend on the full simulation—so no hyper-reduction approximation is needed. The authors demonstrate the idea on Burgers' and cubic-reaction heat equations, where the Galerkin version gives 10–100× speedups.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's 'no dependence on full-order dimension' fails when input dimension N_u grows with N; online cost scales with N_u·n^2.","rationale":"The reader's weakest_assumption concerned the need for accessible, polynomial/liftable operators—a scope condition that the paper itself acknowledges. My concern is different and more directly challenges the abstract's headline claim: even within the stated operator-accessible setting, the claim of N-independence requires an unstated bound on the input dimension. If N_u is O(N), the offline-online decomposition collapses because online evaluations involve matrices with N_u rows/columns. This is not merely a theoretical edge case; spatially distributed actuation/control is common in PDE applications, and nothing in the paper's formulation (Section 2.1) restricts N_u. However, the numerical experiments use N_u=1 and N_u=2, and the method works as intended for small N_u. The paper can be fixed by scoping the abstract and the operator-complexity statements to N_u=O(1), or by explicitly extending the approach to handle distributed inputs through additional approximations. Since the core derivation is correct for the demonstrated setting and the verdict was already CONDITIONAL, my recommendation remains CONDITIONAL; I therefore mark the verdict as unchanged. The proposed concrete test would decisively show whether the N-dependence appears when N_u grows with N.","tokens_in":25742,"tokens_out":12303,"duration_ms":114905,"concrete_test":"Analytic and computational check: In (15), the bilinear contribution requires multiplying an n × (N_u n) matrix by an (N_u n)-vector, i.e., O(N_u n²) operations. Implement a minimal HRF-G solver for (1) with N_u=N, for example by setting B(μ)=I_N, N(μ)=0, and a sparse quadratic F. Fix n=8 (or another small value) and vary N from 10² to 10⁴. Measure the per-time-step online wall-clock time for evaluating (15). If the runtime scales approximately as N (through N_u), the claim that evaluation is independent of N is falsified for distributed inputs. Alternatively, simply perform a FLOP count: multiplying the precomputed n×N operator ΦᵀB by the N-vector u costs O(Nn), which already contradicts the abstract when N_u=N.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract; Eqs. 14–15; Appendix A) is that online evaluation of the projected residual and Jacobian depends only on the reduced dimension n, not on the full-order dimension N. This is true only if the number of inputs N_u is treated as a small constant. In (15), the term ΦᵀN(μ)(I_{N_u}⊗Φ)(u_{m-j}⊗x̂) multiplies an n × (N_u n) precomputed operator by an (N_u n)-vector; the Jacobian term in (14) multiplies the same operator by (u_m⊗I_n), an (N_u n) × n matrix. Both operations cost O(N_u n²) per evaluation. The abstract's unconditional 'does not depend on the dimension of the full-order model' is unsupported for distributed inputs, where N_u may equal N (e.g., a spatially resolved control signal). In that case the online cost scales linearly with N and no offline-online speedup is obtained. The paper's two experiments use N_u=1 and N_u=2, so the tests never exercise this dependency. The condition N_u=O(1) is not stated in Section 2.1 or anywhere in the claims. Thus the strongest claim is overbroad as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an intrusive projection-based model-order reduction framework for semidiscrete dynamical systems of the form (1), where the nonlinearity is quadratic in the state and bilinear in state/input, with affine parameter dependence. The key idea is that all operators appearing in the projected residual and Jacobian of a reduced Newton solver can be precomputed offline using Kronecker identities; the online evaluation then costs only in the reduced dimension and reproduces the Galerkin or LSPG projection of the full-order residual exactly, avoiding the additional approximation layer of hyper-reduction. The HRF-G and HRF-LSPG schemes are tested on a 1D Burgers equation and on a heat equation with a cubic reaction term handled by the lifting w=q^2, with comparisons to ECSW-based Galerkin and LSPG ROMs. Numerical results show that the HRF schemes match the unreduced projected ROMs in accuracy and achieve speedups relative to the FOM, although HRF-LSPG is markedly slower than HRF-G.","tokens_in":26076,"tokens_out":7084,"duration_ms":68624,"significance":"The central algebraic observation is correct and clearly presented: for polynomial systems of the stated form, the projected residual and Jacobian terms factor into precomputable reduced operators. The exact-reproduction property is convincingly demonstrated by ROM evaluation errors that are orders of magnitude lower than those of ECSW. The lifting example extends the method to a cubic nonlinearity and shows that the lifting error is small. The availability of public code is a plus for reproducibility. The framework is incremental relative to existing quadratic-bilinear MOR and operator-inference literature, but it is a clean and potentially useful treatment of the Newton-solver setting, with honest experimental comparisons against ECSW. The main weaknesses are overbroad claims in the abstract and in Section 3 about independence from the full-order dimension and about the speedups obtained by HRF-LSPG.","major_comments":[{"comment":"The paper's central claim that online evaluation \"does not depend on the dimension of the full-order model\" is not fully supported. In Eq. (15), the bilinear-input term is [Phi^T N(I_Nu (x) Phi)](u_{m-j} (x) xhat), where the underbraced operator has dimension n x (N_u n); each evaluation costs O(N_u n^2). The Jacobian term in Eq. (14) has the same cost. If N_u is a fixed constant, the cost is independent of N, and the claim holds. However, the manuscript nowhere states the condition N_u=O(1). For distributed or spatially resolved inputs, N_u can scale with N, in which case the online cost scales linearly with N and no offline-online speedup is obtained. The two numerical experiments use N_u=1 and N_u=2, so they do not exercise this dependence. The condition should be stated explicitly in Section 2.1 and used to qualify the abstract and Section 5.","section":"Section 3, Eqs. (14)-(15); Appendix A"},{"comment":"The abstract states that HRF-G and HRF-LSPG achieve two and one order of magnitude speedup, respectively, with respect to the FOM. This is not uniformly supported. Section 4.1 explicitly reports that HRF-LSPG \"becomes roughly as expensive as the FOM\" at eps_POD=10^-4, and Figure 5 shows that HRF-LSPG generally attains the smallest speedup among the methods considered. The one-order speedup claim is therefore parameter-dependent and potentially misleading. The authors should report the range of observed speedups and state the conditions (modal energy, test parameter set) under which the one-order figure is achieved, or revise the abstract to avoid overclaiming.","section":"Abstract vs. Section 4.1, Figure 3 and Figure 5"}],"minor_comments":[{"comment":"The caption reads \"ECSW-G, and ECSW-G\" twice in the same phrase. One of these should presumably be ECSW-LSPG.","section":"Section 4.2, Figure 10 caption"},{"comment":"For the lifted HRF schemes, it is unclear whether the reference Galerkin-ROM/LSPG-ROM used in Eq. (19) is the lifted or the non-lifted ROM. The text says all other methods use the non-lifted formulation, but the definition of ROM evaluation error in Eq. (19) requires the same projection scheme. If the reference is the non-lifted ROM, the metric conflates lifting error with ROM evaluation error. Please clarify.","section":"Section 4.2, Figure 9 and ROM evaluation error definition"},{"comment":"The sentence \"the superscript (k) from x_m^(k) has been dropped for notational convenience\" is slightly confusing because the same notation is used for the previous time-step values. A brief notational clarification would help.","section":"Eq. (15)"},{"comment":"The text states that ECSW results are presented for tolerances 10^-5 and 10^-9. In some panels the speedup values are difficult to read, especially where curves overlap. Consider adding a table with numerical speedup values for selected eps_POD values.","section":"Section 4.1, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically sound and the central derivation is correct. The main issue is that the abstract and Section 3 overstate the dimensionality independence by omitting the role of N_u, and the headline HRF-LSPG speedup claim is contradicted by the paper's own results. Both are fixable with explicit qualifications; I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core contribution is concrete. For polynomial (quadratic plus bilinear) FOMs with affine parameter dependence, the authors write out all projected residual and Jacobian operators in reduced coordinates, precompute them offline, and solve the exact Galerkin/LSPG Newton systems with no hyper-reduction layer. Appendix A is the genuinely new part—explicit LSPG Newton operators, with the nested loops made visible. The Galerkin piece overlaps with QLMOR and operator inference, but applying the idea to Newton solvers for both projections is a solid extension. The derivation is clean, and the ROM evaluation error plots confirm that HRF schemes reproduce the conventional Galerkin/LSPG solutions essentially exactly. The lifting example for the cubic reaction term is a good demonstration of scope. Code and data are public, which helps.\n\nThe main problem is the abstract. The claim that evaluation 'scales only with the dimension of the reduced space, and does not depend on the dimension of the full-order model' is only true if the input dimension N_u is fixed, O(1). Equations (14) and (15) contain terms of size N_u n and N_u n^2; if N_u grows with N, the online cost scales with N and the offline-online speedup disappears. That condition is not stated anywhere. The two experiments use N_u=1 and N_u=2, so the tests never exercise this dependency. The stress-test note lands. The abstract also says HRF-LSPG achieves one order of magnitude speedup, but Section 4.1 reports that at eps_POD=1e-4 HRF-LSPG becomes roughly as expensive as the FOM, and throughout the paper HRF-LSPG is slower than ECSW-LSPG. The conclusion is more honest about this. Both issues are presentation flaws rather than fatal to the method, but they should be fixed before publication.\n\nOther soft spots are minor and mostly acknowledged: the numerical evidence is limited to 1D problems with modest N, and the method is intrusive, so legacy or black-box codes are out of scope. The authors say that in the conclusion.\n\nThis paper deserves a serious referee. It is a useful extension within projection-based MOR, not a paradigm shift, and the referee should push for the abstract to be scoped precisely and for an experiment with larger N_u. I would cite Appendix A in my own work.","headline":"Useful but narrowly scoped: new explicit HRF reduced Newton operators for Galerkin/LSPG, with an abstract that overreaches on N-dependence and HRF-LSPG speed.","tokens_in":680,"tokens_out":830,"would_cite":true,"duration_ms":32389,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M22"],"pacs":[],"model":"deepseek-v4-flash","headline":"For polynomial dynamical systems, reduced Newton solvers can be computed exactly without hyper-reduction, because all projected residual and Jacobian operators are precomputable offline and online cost scales only with the reduced dimension","keywords":["projection-based model-order reduction","hyper-reduction-free","polynomial nonlinear systems","Galerkin projection","least-squares Petrov-Galerkin","reduced Newton solver","lifting transformation","offline-online decomposition"],"falsifier":"For a fixed polynomial full-order model and reduced basis, compute the direct projected residual Φᵀr(Φx̂) and projected Jacobian Φᵀ(∂r/∂x)(Φx̂)Φ at several random reduced states and parameter values, and compare them numerically with the HRF formulas; any difference beyond machine precision would falsify the claimed exactness. Separately, fix the reduced dimension n and double N while keeping the operators fixed in structure; if per-Newton wall-clock time grows measurably, the claimed independence from N is false.","tokens_in":1472,"feed_emoji":"🧮","tokens_out":1860,"duration_ms":58092,"temperature":0.7,"pith_summary":"The paper tries to establish that a large class of nonlinear reduced-order models, those whose full-order equations are polynomial with affine parameter dependence, can be solved with a Newton method in the reduced space without the usual extra approximation layer called hyper-reduction. The authors show that every operator appearing in the projected residual and its Jacobian can be computed once offline, so the online cost depends only on the reduced dimension, not on the size of the full-order model. This means Galerkin and least-squares Petrov-Galerkin projections are solved exactly, with no sparse-sampling error and no hyper-reduction tolerance to tune. In two test problems, the Galerkin variant matches full-projection accuracy while outperforming sparse-sampling approaches in speed; the lifted cubic heat equation demonstrates how non-polynomial nonlinearities can be brought into the required form.","feed_headline":"Polynomial nonlinearities make reduced Newton solves hyper-reduction-free","feed_subtitle":"Exact Galerkin and least-squares Petrov-Galerkin solves now scale only with reduced dimension, not full-model size.","key_machinery":"The central device is a pair of Kronecker-splitting identities: writing x̂⊗Iₙ as Σₗ Gₗ x̂ₗ and Iₙ⊗x̂ as Σₗ Hₗ x̂ₗ, where Gₗ and Hₗ are fixed Boolean block matrices of size n²×n. Substituting these into the projected Jacobian turns every term into a product of an offline-precomputed reduced operator and a scalar component of the reduced state; the same expansion handles the residual. This reduces the Newton solve to sums over reduced-coordinate indices only, and for LSPG it introduces one additional nested sum relative to Galerkin.","core_discovery":"For semidiscrete systems of the form dx/dt = C + A x + F(x⊗x) + B u + N(u⊗x), with operators that depend affinely on parameters, the paper derives explicit expressions for the projected residual and Newton Jacobian of a reduced-order model. Using Kronecker-product identities, each quadratic and bilinear term is rewritten as a sum over the reduced state components. Substituting these identities into the Galerkin and LSPG Newton systems yields operators, such as ΦᵀF(Φ⊗Φ)Gₗ and ΦᵀN(I⊗Φ), that can be precomputed offline and depend only on the reduced dimension n and the number of inputs, not on the full-order dimension N. The online Newton iteration therefore evaluates the exact projected residu","pith_inferences":["If full-order operators are only partially accessible, the same precomputation idea could in principle be coupled with data-inferred full-space operators; the paper flags this as future work, and it is the natural next test of the framework.","The storage footprint of precomputed reduced operators grows with powers of n, so for very large reduced dimension the memory requirement, rather than arithmetic, may become the limiting factor, a trade-off the paper's experiments do not stress.","The lifting route trades a larger latent space for polynomial structure; a direct comparison at equal polynomial degree (cubic HRF versus lifted quadratic HRF) on a more complex benchmark would separate lifting-induced error from reduction error.","For legacy or black-box codes whose nonlinearities cannot be lifted exactly, the method does not apply; a data-driven operator-inference extension would be required to make the approach non-intrusive."],"forward_implications":["The projected Galerkin and LSPG Newton systems are solved exactly: no sparse-sampling approximation enters the reduced residual or Jacobian, so the ROM evaluation error relative to the projection is effectively zero up to floating-point arithmetic.","For the targeted polynomial class, online cost per Newton iteration is independent of the full-order dimension N, making very large full-order models practical in many-query settings.","Cubic and higher-degree polynomial terms can be handled by summing over Kronecker products of the reduced state, and lifting transformations extend the method to non-polynomial nonlinearities such as cubic reaction terms.","Hyper-reduction tolerance tuning disappears for these systems, and in the paper's experiments the Galerkin variant avoids the convergence failures that loose sparse-sampling tolerances produce.","The LSPG variant is exact but carries extra nested sums over reduced coordinates, so the paper's data favor Galerkin for computational speed while retaining LSPG as a stability-oriented alternative."],"fun_headline_variants":["Skip hyper-reduction: polynomial systems get exact reduced Newton","Reduced Newton solvers exact for polynomial nonlinearities","Polynomial ROM Newton: no hyper-reduction, scale with n not N","Two orders of magnitude speedup without hyper-reduction","Exact reduced Newton solves: offline precomputed, online fast"],"cache_read_input_tokens":27904,"weakest_assumption_plain":"The full-order model must be known and expressible, exactly or through an exact lifting transformation, as a polynomial system with affine parameter dependence; if the operators are inaccessible or the nonlinearity cannot be lifted, the offline precomputation breaks down.","fun_headline_variants_meta":{"raw":{"variants":["Skip hyper-reduction: polynomial systems get exact reduced Newton","Reduced Newton solvers exact for polynomial nonlinearities","Polynomial ROM Newton: no hyper-reduction, scale with n not N","Two orders of magnitude speedup without hyper-reduction","Exact reduced Newton solves: offline precomputed, online fast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000476,"raw_usage":{"total_tokens":2264,"prompt_tokens":879,"completion_tokens":1385,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":1300}},"tokens_in":623,"tokens_out":1385,"duration_ms":11583,"temperature":1.0,"reasoning_tokens":1300,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T19:06:58.973147+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed polynomial full-order model and reduced basis, compute the direct projected residual Φᵀr(Φx̂) and projected Jacobian Φᵀ(∂r/∂x)(Φx̂)Φ at several random reduced states and parameter values, and compare them numerically with the HRF formulas; any difference beyond machine precision would falsify the claimed exactness. Separately, fix the reduced dimension n and double N while keeping the operators fixed in structure; if per-Newton wall-clock time grows measurably, the claimed independence from N is false.","supporting_citations":[],"review_version":1}