{"id":"1dfb75ef-9fdf-445e-a501-641f361bedbc","arxiv_id":"2412.17695","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Quadratic-manifold Neural Galerkin reduced models give locally unique, residual-minimizing trajectories and, for linear full models, online cost independent of the full dimension.","lead":"This paper introduces a nonlinear model-reduction method that combines quadratic manifolds with Neural Galerkin dynamics, proving uniqueness of the reduced trajectory and online costs independent of the full model size for linear problems. It reports large speedups on wave and particle-transport tests, and shows that fewer collocation points can act as built-in hyper-reduction.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central online-efficiency identity (4.12) is unverified and Eq. (4.10) is printed inconsistently; if the H tensor is wrong, the O(n^4) independence-from-N claim and the corresponding speedup experiments fail.","rationale":"I agree with Reader's general shape of the paper: the residual-minimizing formulation and the local-uniqueness result are clean ideas, the experiments appear plausible, and the main concerns are about verifiability. However, I think the most load-bearing weak point is not the collocation/grid separation caveat in the uniqueness lemma but the fact that the entire online-efficiency story is a set of algebraic identities that (a) are not proven, (b) contain as-printed an index error in Eq. (4.10), and (c) underlie the speedup plots that support the major claim of online efficiency. If the tensor precomputation is wrong, then even the uniqueness property V^T W=0 cannot save the prediction cost claim. The manuscript itself flags a related restriction for interpolated manifolds (V_Ξ^T W_Ξ=0 can fail), but that is a limitation of scope, not an internal error; the internal algebra in 4.3 is more fragile. Given the reader already marked CONDITIONAL due to the typo and missing code, my recommendation stays CONDITIONAL but sharpens the condition: the preprint should provide a verified derivation of (4.10)-(4.12) or release the code that implements the tensors, so the O(n^4) claim is reproducible. This is not a rejection because the remaining theoretical contributions (residual minimization, local uniqueness for greedy-trained manifolds, hyper-reduction-style collocation separation) do not depend on the disputed tensor calculation and are supported by the manuscript's own argumentation.","tokens_in":22232,"tokens_out":2101,"duration_ms":17212,"concrete_test":"Re-derive (4.12) from scratch by substituting J(theta) = V + K·theta and f(g(theta)) = A(s0 + V theta + W h(theta)) into normal equations J^T J theta_dot = J^T f, imposing V^T W=0 and V^T V=I. Check whether the printed (4.10) recovers the resulting polynomial terms: in particular the term (theta·H)(theta⊗theta) must contain A acting on W h(theta), so H must be defined using W, not V. If the rederivation shows that H as printed equals A, then every linear experiment in Section 5 (Figures 2 and 6) must be rerun with the corrected tensor, and the runtime claim O(n^4) must be re-verified.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central computational claim that QMNG reduced models for linear full models inherit a precomputable online cost of O(n^4) independent of N rests entirely on the tensor precomputation in Section 4.3, specifically the identities (4.8)-(4.12). The loaded assumption is that the reported expressions J, S, A, H, s0-hat, A-hat, H-hat are correctly derived and correctly implemented. Two concrete red flags are visible in the manuscript text itself: (i) The definition of H in Eq. (4.10) is written identically to the definition of A: both read 'Akjm = Kijk AilVlm, Hkjm = Kijk AilVlm'. Since H should contain the action of A on W (the quadratic part), the right-hand side for H must involve W, not V; as printed, H=A and the subsequent term (theta·H)(theta⊗theta) is dimensionally and structurally wrong. (ii) The text states 'the solution to the least-squares problem (4.5)' is obtained by solving the normal equations with left-hand side I + theta·J·theta, using the identity J^T J = I + K^T K and V^T W = 0. The identity itself is plausible under V^T W=0, but the tensor formulas are never derived in the paper; no code or data are provided, and no independent check is offered. Since the numerical experiments on acoustic waves and charged particles rely on Algorithm 3 to assemble these tensors, an error in (4.10)-(4.12) would invalidate the online-efficiency experiment and the claim of O(n^4) online cost. The most load-bearing fix is therefore to independently derive eq. (4.12) from J(theta) = V + K(theta) and f(q)=A q, and to check whether (4.10)'s H and A definitions differ at all; as printed they do not.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper combines Neural Galerkin schemes with quadratic-manifold parametrizations for nonlinear model reduction. For the vector-valued decoder with collocation points equal to the full-model grid, the authors prove local uniqueness of the reduced dynamics and residual-norm minimization, and for linear full models they derive precomputable tensor quantities that purportedly make online predictions cost O(n^4), independent of the full state dimension N. A second variant uses an interpolated parametrization that allows collocation points to differ from the grid points, which acts as a form of hyper-reduction. The method is tested on advecting acoustic waves, charged-particle densities, and Burgers' equation, with honest reporting of where speedups are and are not obtained.","tokens_in":22746,"tokens_out":8924,"duration_ms":83376,"significance":"If the main technical claims hold, the paper makes a useful contribution: residual-norm minimization distinguishes QMNG reduced models from constant-test-space approaches and appears to stabilize predictions on tightly fitted quadratic manifolds, while the O(n^4) online complexity for linear problems is a structurally interesting result. The paper is also careful in stating assumptions and limitations, including the absence of speedup for Burgers' equation and the dependence of the uniqueness result on the greedy-manifold property V^T W = 0. However, the central tensor precomputation in Section 4.3 contains a visible indexing/definition error and is never derived, so the online-efficiency claim and the linear-problem experiments are not yet fully supported.","major_comments":[{"comment":"The definition of H_{kjm} is printed identically to A_{kjm}, namely H_{kjm} = K_{ijk} A_{il} V_{lm}, so the last index is m (of size n) even though H is used in (4.12) to multiply θ ⊗ θ, which has length n^2. Moreover the right-hand side must involve A W, not A V, to represent the term K(θ)^T A W (θ ⊗ θ). As printed, the definition of H cannot be correct, and since Algorithm 3 assembles this tensor for the linear-problem experiments, the O(n^4)-independent-of-N claim is not currently verified.","section":"Section 4.3, Eq. (4.10)"},{"comment":"Please provide a term-by-term derivation of (4.12) from the normal equations of (4.5). The displayed formula does not follow as written: with f(q)=Aq and q=s0+Vθ+W(θ⊗θ), the first projection term is V^T A s0, not the V^T s0 defined in (4.9), and the term θ·S with S_{jk}=K_{ikj}[s0]_i cannot by itself account for A s0. Either the tensor definitions or the formula for fhat need to be corrected, and the derivation should be included so that the precomputation can be checked.","section":"Section 4.3, Eqs. (4.11)-(4.12)"},{"comment":"The tensor products in (3.11)-(3.12) are defined for T ∈ R^{n×m×p}, but (3.13) declares KΞ ∈ R^{m×n×n}; under the stated convention the expression KΞ·θ is not well defined. Likewise the remark below (4.8) that θ·J·θ = (θ·K)^T(K·θ) mixes the two tensor-vector products in a way that is dimensionally inconsistent. Please make all tensor index conventions explicit and consistent throughout Section 4.3, since the complexity table and Algorithms 3-4 rely on these operations.","section":"Sections 3.4 and 4.3, tensor-index conventions"},{"comment":"The abstract and conclusions state that solutions of the proposed QMNG reduced models are locally unique without the qualifications required by Lemma 4.1 and Proposition 4.2. Section 4.2 itself notes that the condition V_Ξ^T W_Ξ = 0 can be violated when interpolation separates collocation points from grid points, so the uniqueness guarantee does not cover the interpolated QMNG models used for the Burgers experiment. Please qualify the claims to state that local uniqueness holds for the grid-collocated vector-valued version under the greedy-manifold construction with V^T W = 0.","section":"Abstract, Section 4.2, Section 6"}],"minor_comments":[{"comment":"The text says the QMNG approximations are plotted for t ∈ {0,2,4,6.8}, while Figure 1 uses t ∈ {0,2,4,6,8}; the former appears to be a typo.","section":"Section 5.1.2"},{"comment":"The procedure name is printed as OMNGOnline; this should be QMNGOnline.","section":"Algorithm 2"},{"comment":"The Jacobian matrix is said to be in R^{N×n}, but the displayed entries use ∂/∂θ_p; p is undefined and should presumably be n.","section":"Eq. (4.3)"},{"comment":"The paper states that solving (4.5) via normal equations uses J^T J = I + K^T K, which is plausible under V^T W = 0, but the corresponding four-tensor J in (4.8) is never explicitly constructed. A short construction of J from K would improve verifiability.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be a solid contribution once the tensor identities in Section 4.3 are corrected and fully derived. The central issue is localized but load-bearing: the online-efficiency results and the linear-problem experiments depend on these formulas. The local-uniqueness scope should also be qualified in the abstract. I do not see grounds for rejection, provided the authors can fix the tensor section."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful paper, and the core theoretical claim—locally unique solutions and residual-norm minimization for Neural Galerkin on quadratic manifolds—looks sound. The soft spot is Section 4.3: there's an apparent typo in Eq (4.10) where A and H are defined identically, and since the online-efficiency claim for linear systems rests on those tensor formulas, that section needs a careful pass before the O(n^4) independence-from-N statement is bulletproof.\n\nWhat's new: the combination of Neural Galerkin dynamics (time-varying test space) with quadratic manifolds is new, and it pays off. The local-uniqueness result (Lemma 4.1, Prop 4.2) is a real step beyond the constant-in-time test spaces in [9] and the Hamiltonian-specific construction in [33]. The precomputed-tensor route in 4.3 is clever, and the separate-collocation-points idea in Section 3 is a nice baked-in hyper-reduction that they demonstrate on Burgers. The experiments are honest: they report orders-of-magnitude speedups on acoustic waves and charged particles, and they explicitly say there's no speedup for Burgers in 1D. The comparison to constant test spaces makes a fair case that residual minimization is what stabilizes the well-fit manifolds.\n\nSoft spots, in proportion. First, the (4.10) typo: as printed, Hkjm = Kijk AilVlm is identical to A, and the right-hand side of (4.12) has terms with H that structurally must involve W. This is almost certainly a typo, and the surrounding derivation in Section 4.3 is plausible, but the manuscript never derives (4.8)-(4.12), and there's no code or data to check against. A referee should ask for the step-by-step derivation and ideally a reproducible implementation. Second, the uniqueness guarantee depends on V^T W = 0 from the greedy training in [36]; the paper explicitly notes this fails when interpolation separates collocation from grid points, so the guarantee is narrower than the abstract might suggest. That's a stated limitation, not a hidden one. Third, the online cost scales as O(n^4), which they honestly show leads to slower-than-full-model runtimes for n >= 40 in the wave example; that's an inherent tradeoff, not a flaw.\n\nWho this is for: anyone working on nonlinear model reduction for transport-dominated problems, especially people who want a principled alternative to autoencoder-based manifolds with hyper-reduction. The paper deserves a serious referee; the fix list is short (correct (4.10), add derivations or code) but the load-bearing part of the online-efficiency claim should be verified before publication. I'd engage with it.","headline":"A solid, honest nonlinear-reduction paper whose core theorem is plausible; the online-efficiency claim rests on a tensor section with an apparent typo that needs fixing and verification.","tokens_in":23190,"tokens_out":3216,"would_cite":true,"duration_ms":26920,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M60","65M99"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes reduced models that run Neural Galerkin dynamics on quadratic manifolds, giving locally unique, residual-minimizing trajectories and, for linear problems, online costs independent of the full-model state dimension.","keywords":["nonlinear model reduction","Neural Galerkin schemes","quadratic manifolds","Kolmogorov barrier","online efficiency","hyper-reduction","residual minimization","transport-dominated problems"],"falsifier":"Train a quadratic manifold by ordinary least-squares fitting of $W$ without enforcing $V^\\top W = 0$, or use collocation points distinct from grid points, then run QMNG on a linear wave problem and compute the smallest singular value of $J(\\theta)$ along the trajectory; a zero singular value and a nonunique least-squares minimizer would refute the local-uniqueness claim outside the greedy setting.","tokens_in":22034,"feed_emoji":"🧮","tokens_out":7633,"duration_ms":67724,"temperature":0.7,"pith_summary":"The paper proposes reduced models that run Neural Galerkin dynamics on quadratic manifolds, a nonlinear parametrization built from a linear basis plus a quadratic feature map. The authors claim that these reduced solutions are locally unique and that, by construction, they minimize the residual norm at every time step, which keeps predictions stable even when the manifold is tightly fitted to training data. For linear full models, all quantities that depend on the full state dimension are moved into an offline phase, so each online step costs O($n^{4}$) with n the reduced dimension, independent of the full-model size N. For nonlinear problems, the residual can be evaluated at collocation points that are separate from the full-model grid, giving a built-in form of hyper-reduction without an extra empirical-interpolation layer. Numerical experiments with acoustic waves and charged-particle densities report orders-of-magnitude speedups over the full models.","feed_headline":"Quadratic manifolds make reduced models unique and fast","feed_subtitle":"The dynamics minimize residual norm at each step, and linear problems precompute all N-dependent costs offline.","key_machinery":"The machinery is the quadratic decoder $g_{V,W}(\\theta) = s_0 + V\\theta + W h(\\theta)$ with $h(\\theta) = \\theta \\otimes \\theta$, whose Jacobian is affine in $\\theta$: $J(\\theta) = V + K\\cdot\\theta$. Neural Galerkin schemes choose the time derivative $\\dot{\\theta}$ by the least-squares projection of the residual onto the columns of this Jacobian, which is equivalent to minimizing the residual norm. The orthogonality $V^\\top W = 0$ inherited from the greedy training algorithm makes $J(\\theta)$ full column rank for all $\\theta$ and lets the normal-equation matrix simplify to $I + K(\\theta)^\\top K(\\theta)$, the key step in the local-uniqueness proof. For linear problems, tensor contractions precompute all $N$-dependent products offline, leaving only reduced-dimension operations in the online phase.","core_discovery":"The central claim is that quadratic manifolds are not just expressive enough to approximate transport-dominated solutions; they also carry enough structure for Neural Galerkin schemes to define a well-posed reduced dynamics. The paper shows that when the manifold is trained with the greedy algorithm, the linear and quadratic parts satisfy $V^\\top W = 0$, which forces the $N \\times n$ Jacobian $J(\\theta)$ to have full column rank for every $\\theta$. Consequently, the least-squares problem that selects $\\dot{\\theta}$ has a unique solution, and the right-hand side of the reduced ODE is locally Lipschitz, so trajectories are locally unique. Because Neural Galerkin dynamics solve that least-squares problem, the reduced solution minimizes the residual norm over time; experiments on the acoustic wave and Vlasov problems show this residual-minimizing choice remains stable with tightly fitted manifolds where a constant-in-time test space fails. For linear full models, the same affine-Jacobian structure lets every $N$-dependent term be contracted in advance, so prediction costs scale as $O(n^4)$, independent of $N$; for nonlinear problems, the paper demonstrates that using roughly ten times fewer collocation points than grid points still tracks the reconstruction error.","pith_inferences":["The local-uniqueness guarantee is tied to the greedy training property $V^\\top W = 0$; the paper notes this property can fail once interpolation separates collocation points from grid points, so users of the interpolated variant should verify the rank of the batch Jacobian before relying on uniqueness.","The $O(n^4)$ online cost means the practical speedup window is limited in the reduced dimension: the acoustic-wave experiment itself shows speedup disappears around $n = 40$, so the method will be most useful for small $n$ with large $N$.","For nonlinear problems, the paper reports no speedup on the Burgers example because the full model is cheap and automatic differentiation dominates; extending the precomputation idea to nonlinear right-hand sides, or approximating derivatives without automatic differentiation, would be the natural next step.","The stability comparison suggests a direct test of the time-continuous versus time-discrete residual formulation on the same quadratic manifolds; if QMNG remains stable where discrete-residual fits fail, the paper's variational choice is validated as the cause."],"forward_implications":["Quadratic-manifold Neural Galerkin reduced models have locally unique trajectories, so they do not suffer the tangent-space collapse that can occur with other nonlinear parametrizations.","For linear full models, online predictions of QMNG reduced models cost $O(n^4)$ per step independently of the full-model dimension $N$, which is what the paper means by online efficiency.","Because collocation points can differ from grid points, nonlinear QMNG reduced models carry hyper-reduction within the dynamics and need no separate empirical-interpolation step.","Residual-norm minimization is the ingredient that stabilizes QMNG predictions on tightly fitted manifolds; replacing the Jacobian by a constant-in-time test space destabilizes the reduced dynamics, as the acoustic-wave and charged-particle experiments show."],"supporting_citations":[{"why":"Supplies the Neural Galerkin least-squares dynamics and the residual-minimization condition used throughout the paper.","marker":"[5]"},{"why":"Supplies the greedy quadratic-manifold construction and the $V^\\top W = 0$ property that carries the full-rank and uniqueness proofs.","marker":"[36]"},{"why":"Defines the constant-in-time test-space reduced dynamics on quadratic manifolds that QMNG is compared against and that can destabilize.","marker":"[9]"},{"why":"Introduces quadratic approximation manifolds for nonlinear model reduction and motivates the Kolmogorov-barrier setting.","marker":"[10]"},{"why":"Documents the tangent-space collapse of nonlinear parametrizations that the QMNG full-rank result is designed to avoid.","marker":"[7]"},{"why":"Establishes the prior use of time-varying test spaces for quadratic manifolds in Hamiltonian systems, the main comparative approach with time-dependent test spaces.","marker":"[33]"}],"fun_headline_variants":["Quadratic manifolds beat Kolmogorov barrier in model reduction","Neural Galerkin on quadratic manifolds: stable and fast","Quadratic-manifold reduction: unique dynamics, fast","Transport problems tamed by quadratic manifold reduction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The uniqueness guarantee rests on training the quadratic manifold with the paper's greedy algorithm so that the linear and quadratic parts are exactly orthogonal; if that orthogonality is lost, the Jacobian can lose rank and the reduced trajectory may not be unique.","fun_headline_variants_meta":{"raw":{"variants":["Quadratic manifolds beat Kolmogorov barrier in model reduction","Neural Galerkin on quadratic manifolds: stable and fast","Quadratic-manifold reduction: unique dynamics, fast","Transport problems tamed by quadratic manifold reduction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000315,"raw_usage":{"total_tokens":1799,"prompt_tokens":975,"completion_tokens":824,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":769}},"tokens_in":591,"tokens_out":824,"duration_ms":6795,"temperature":1.0,"reasoning_tokens":769,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:15:08.036898+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a quadratic manifold by ordinary least-squares fitting of $W$ without enforcing $V^\\top W = 0$, or use collocation points distinct from grid points, then run QMNG on a linear wave problem and compute the smallest singular value of $J(\\theta)$ along the trajectory; a zero singular value and a nonunique least-squares minimizer would refute the local-uniqueness claim outside the greedy setting.","supporting_citations":[{"cited_title":"Bruna, B","cited_arxiv_id":null,"evidence_quote":"Supplies the Neural Galerkin least-squares dynamics and the residual-minimization condition used throughout the paper."},{"cited_title":"Geelen, S","cited_arxiv_id":null,"evidence_quote":"Defines the constant-in-time test-space reduced dynamics on quadratic manifolds that QMNG is compared against and that can destabilize."},{"cited_title":"Barnett, C","cited_arxiv_id":null,"evidence_quote":"Introduces quadratic approximation manifolds for nonlinear model reduction and motivates the Kolmogorov-barrier setting."}],"review_version":1}