{"id":"4721c787-1824-45ca-a0cb-4675f5cf894f","arxiv_id":"2608.04239","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Spectral submanifolds of equivariant systems are proven equivariant, and a new algorithm exploits this to halve the parameter count and speed up data-driven reduced-order modelling.","lead":"This paper introduces equivariant spectral submanifold reduction, a method for building reduced-order models of high-dimensional dynamical systems that automatically preserves the system's symmetries. The approach cuts fitting time and parameter counts on several benchmarks, which matters for simulating large mechanical and fluid systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The eSSM algorithm freezes the reduced representation R0 from a truncated SVD of orbit-augmented data; if the leading d singular values are not cleanly separated, R0 can belong to the wrong similarity class and all claimed speedups and robustness gains inherit a faulty exact constraint.","rationale":"I agree with the reader's weakest_assumption: Step (i.1) is the load-bearing data-dependent step. The theory (Thm 2.7, Props 3.4-3.5) is mathematically sound for a genuinely equivariant system with a spectral subspace E, but the eSSM algorithm must identify E and its G-representation from data before it can impose equivariance constraints. In Step (i.1) this is done by a truncated SVD of the orbit-augmented Gram matrix, and the resulting R0 is frozen in Step (i.2). Prop 4.4(ii) makes the choice irreversible under continuous optimisation. The failure mode is concrete: when sigma_d is not clearly larger than sigma_{d+1}, the leading d left singular vectors need not span a G-invariant subspace, and even when they do, the subspace may carry a different representation from the true spectral subspace. All subsequent parameter reductions and robustness guarantees are then exact only with respect to the wrong action. This is not an internal inconsistency of the theory; it is a correctness risk in applying the theory to data. The discrete-time CTF extension (Appendix A) is also unproven, but it affects only the third benchmark and is explicitly labelled outside the continuous-time theory. I therefore keep the reader's CONDITIONAL verdict: the central theoretical claim can stand, but the paper should either justify the spectral-gap and data-equivariance conditions under which R0 is reliable, or add a validation step comparing R0 with a model-based representation, and the empirical claims need error bars.","tokens_in":29375,"tokens_out":10745,"duration_ms":101146,"concrete_test":"On the C4 shallow-water experiment of Section 5.2, compute the exact 6-dimensional spectral subspace E_true of the linearised operator A of (31) from the Dedalus discretisation, choose a basis U, and compute R_true(S)=V^T S U with V the dual basis, for a generator S of C4. Run Algorithm 1 with d=6 on the same data, obtain R0(S)=V_{1,0}^T S U_{1,0}, and compare the character vectors chi_R(S)=tr R(S). Repeat after adding 1% relative noise to the snapshots and after setting d=5 and d=7. If chi_R0 differs from chi_R_true, or if chi_R0 changes discontinuously with d or noise, then the frozen representation is unreliable; also check whether the equivariant fit's NMTE on the rotated test trajectory degrades relative to unconstrained SSMLearn in those cases.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The mathematical theorem that a genuinely equivariant system has an equivariant SSM is not in doubt. The load-bearing step is the data-driven bridge, Step (i.1): the algorithm freezes the reduced representation R0 from the d-truncated SVD of the orbit-augmented snapshot matrix (eq. (22), Lemma 4.5) and then, in Step (i.2), constrains every fitted object to commute with this fixed R0 (eq. (24)). Once the similarity class [E] is fixed, Prop. 4.4(ii) guarantees that no continuous optimisation can leave that class. So the advertised exact equivariance, the parameter reductions in Tables 1-2, and the robustness gains all inherit R0. If the leading d singular values are not well separated (sigma_d close to sigma_{d+1}), the truncated SVD can split an eigenspace of the Gram matrix and produce a subspace that is not genuinely G-invariant, or one that is invariant but lies in the wrong similarity class relative to the true spectral subspace E of A. The same failure occurs if the data are not exactly equivariant under the supplied G (noise, symmetry-breaking effects, or incomplete transient removal). Lemma 4.5 only guarantees invariance of whole eigenspaces, not of an arbitrary d-dimensional truncation at a degenerate or near-degenerate threshold. No later step detects or corrects a wrong R0; it is frozen as a hard constraint. This is a data-dependent modeling assumption, distinct from the theory, and it is the least secure link in the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces equivariant spectral submanifold (eSSM) reduction, an extension of SSM-based model reduction that imposes exact linear symmetry constraints on the manifold parametrisation and reduced dynamics. The theoretical sections claim that SSMs of equivariant systems are equivariant submanifolds, that the graph parametrisation and reduced vector field inherit the restricted group action, and that the extended normal form is equivariant. The algorithm freezes a reduced representation R0 from the truncated SVD of an orbit-augmented snapshot matrix, then constrains all fitted objects to commute with R0. Numerical experiments on an oscillator chain, the shallow-water equations on a sphere, and a CTF4Science Kuramoto–Sivashinsky forecasting task report reduced parameter counts, faster fits, and competitive accuracy.","tokens_in":29644,"tokens_out":24241,"duration_ms":212591,"significance":"If the central claim is established, the paper offers a principled way to incorporate symmetries into SSM-based model reduction, with exact parameter-count reductions and a concrete algorithm. The theoretical core is largely self-contained and the parameter-count reductions are exact consequences of the group action rather than fitted quantities. The public availability of the code and the inclusion of a standardized benchmark are strengths. However, the unqualified equivariance theorem is false for arbitrary spectral subspaces, the data-driven step that fixes the reduced representation is not protected against wrong similarity classes, and the reported robustness experiment is in-sample. The significance is therefore conditional on correcting these issues.","major_comments":[{"comment":"Theorem 2.7(i) is false as stated. For example, let A = λI_2 with λ > 0, let G be generated by the 90-degree rotation, and let E = span{(1,0)}. E is a modal eigenspace and hence a spectral subspace under Definition 2.1, and S commutes with A, but S E ≠ E. The proof of the theorem only shows that each full eigenspace of A is invariant under a commuting S, not that an arbitrary one-dimensional modal eigenspace inside a degenerate eigenspace is invariant. The statement, the abstract's claim that SSMs are 'naturally equivariant submanifolds', and the uniqueness argument for W(E) all need to be restricted to spectral subspaces that are direct sums of complete eigenspaces (or isotypic components). This is not pedantic: the global non-resonance condition explicitly allows repeated eigenvalues, so the counterexample is inside the paper's assumptions.","section":"Theorem 2.7"},{"comment":"The data-driven bridge freezes R0 from a truncated SVD of the orbit-augmented snapshot matrix. Lemma 4.5 only guarantees G-invariance of the dominant subspace under a strict singular-value gap; it does not guarantee that this subspace is the spectral subspace E of A from the theory, nor that R0 is the correct representation of G on E. If σ_d is close to σ_{d+1}, or if the data are not exactly equivariant under the supplied G (noise, transients, or symmetry breaking), R0 can belong to the wrong similarity class. By Proposition 4.4(ii) and the hard constraint (24), no subsequent continuous optimisation can leave that class, so the advertised parameter reductions and the 'exactly equivariant by construction' claim inherit the faulty R0. The paper provides no diagnostic or safeguard, such as checking the singular-value gap, testing equivariance of residuals, or validating R0 on held-out data.","section":"Section 4.2, Eq. (22), Lemma 4.5, Eq. (24)"},{"comment":"The 'Rotated' test metric is not an out-of-sample test. The training procedure in Eq. (22) augments the single training trajectory with all four rotations under G = C4, so the 90-degree-rotated trajectory used for evaluation is exactly one block of the orbit-augmented training matrix. The eSSM model is therefore trained on the rotated test trajectory, while the non-equivariant SSMLearn baseline is not. The reported improvement on the 'Rotated' metric is an in-sample consistency check of the imposed symmetry, not evidence of robustness or generalization. A fair evaluation would use a rotation not contained in the supplied group (for example, 45 degrees) or an initial condition outside the training orbit, or would explicitly exclude the rotated trajectory from the augmented training set.","section":"Section 5.2, Eq. (22)"}],"minor_comments":[{"comment":"In the paragraph introducing the extended normal form, 'the change of variables f and n' should read 'the change of variables t and n'.","section":"Section 3.1"},{"comment":"The figures and Table 1 do not clearly indicate which SSMLearn variant is being compared (the Matlab implementation, SSMLearnPy, or the Python implementation with trivial group). Since runtime comparisons are implementation-dependent, this should be stated explicitly in each figure and table caption.","section":"Section 5.1"},{"comment":"The abstract and conclusions present the CTF4Science benchmark as a success without noting the authors' own caveat in Section 5.3 that the chaotic Kuramoto-Sivashinsky data lie outside the theory's assumption of a hyperbolic fixed point. This limitation should be stated in the abstract or at the first mention of the benchmark.","section":"Section 5.3"},{"comment":"The resonant index set uses a strict inequality |Im(m·λ−λ_j)| < δ in Algorithm 1, while Eq. (9) uses ≤ δ. Please make the definition consistent.","section":"Algorithm 1, line 14"},{"comment":"The symbol \\tilde S is used both for the whitened group action in Step (i.1) and for the diagonalised action W^{-1}(S|_E)W in Proposition 3.5. Different symbols would avoid confusion.","section":"Section 4.2 and Prop. 3.5"},{"comment":"The empirical comparisons in Figures 4, 5, 8, and 9 report wall-clock times and NMTE values from single runs without error bars or repeated trials. At least a small number of repeated fits with different initialisations would make the speedup and accuracy claims more robust.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The core idea of equivariant SSMs is viable and the theoretical development is mostly clear, but the current version overstates the equivariance theorem and does not safeguard the data-driven representation-fixing step. The 'Rotated' experiment in Section 5.2 needs to be redesigned before the robustness claim can be accepted. The CTF benchmark inclusion is understandable, but the authors' own scope caveat should be visible in the abstract. The revisions needed are substantial but within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper's main theorem, that SSMs of equivariant systems are equivariant submanifolds with equivariant charts and reduced dynamics, is proven by a short uniqueness argument and is sound. The algorithm built on it, eSSM, constrains the fitted manifold and reduced dynamics to a fixed representation R0 obtained by SVD of orbit-augmented data. That second step is the real weak point: if the leading d singular values are not cleanly separated, the frozen R0 can be in the wrong similarity class, and every advertised speedup and robustness gain inherits that faulty constraint. The theory is not the problem; the bridge from theory to data is.\n\nWhat's new: the paper is the first to put symmetries directly into the SSM parametrization and reduced dynamics, not just restrict ambient dynamics like [34]. Theorem 3.8 gives a clean characterization of equivariant Taylor coefficients, and Corollary 3.11 counts parameters. The parameter-count reductions in Tables 1-2 follow exactly from the group action, not from fitted quantities. The code is public, and the oscillator-chain and shallow-water examples show substantial runtime reductions at matched accuracy. The rotated-trajectory robustness gain is a direct consequence of equivariance and is honestly presented as such.\n\nSoft spots, in order of severity. (1) Step (i.1) freezes R0 once. Lemma 4.5 guarantees whole eigenspaces of the Gram matrix are G-invariant, but a d-truncated SVD at a near-degenerate threshold can split an eigenspace, and no later step detects or corrects a wrong R0. This is a genuine data-dependent modeling assumption, distinct from the theoretical claim, and it is the least secure link. (2) The empirical claims rest on single runs without error bars or repeated trials. (3) The CTF benchmark is explicitly outside the theory (chaotic attractor, not decaying transient), relies on the unproven discrete-time extension in Appendix A, and uses an ad hoc residual-decay correction with a tuned time constant. The paper is candid about this, but the benchmark result should not be used as evidence for the theory without those gaps closed.\n\nThe central argument holds up. I would send this to peer review, with the expectation that the discrete-time equivariance claims get proven or cited, the empirical section gets error bars, and the CTF comparison avoids dependence on the residual correction. It is a serious, useful contribution to symmetry-preserving ROM.","headline":"The equivariance of SSMs is proven cleanly, but the algorithm's exact symmetry constraint hinges on a data-driven subspace selection that can fail without a singular-value gap; still deserving of serious review.","tokens_in":30236,"tokens_out":2200,"would_cite":true,"duration_ms":18972,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37D10","37C81","37M21"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that spectral submanifolds of equivariant systems inherit the full symmetry group, and turns that structure into a data-driven reduced-order algorithm whose fitted manifold and dynamics are exactly symmetric…","keywords":["equivariance","spectral submanifolds","model reduction","nonlinear dynamics","data-driven modelling","extended normal form","delay embedding"],"falsifier":"Take a genuinely $G$-equivariant system, fit an eSSM model, and compare the prediction started at a transformed initial condition $S x_0$ with the group transform of the prediction at $x_0$; agreement within numerical tolerance confirms the equivariance of the fitted model. A sharper test is to repeat the fit when the leading $d$ singular values of the orbit-augmented snapshot matrix are almost degenerate, so that the frozen representation is unstable: the fitted manifold should then visibly break equivariance and the reduced predictions should diverge under symmetry transformations.","tokens_in":29053,"feed_emoji":"🔄","tokens_out":12217,"duration_ms":100110,"temperature":0.7,"pith_summary":"Spectral-submanifold (SSM) reduction builds low-dimensional nonlinear models of high-dimensional systems by fitting the unique smoothest invariant manifold tangent to a chosen spectral subspace of the linearised dynamics. This paper seeks to establish that when the full system has a linear symmetry group, the SSM, its coordinate chart, and the reduced dynamics are automatically equivariant: the symmetry passes through every stage of the reduction. The author converts that fact into an algorithm, eSSM, which restricts every fitted coefficient to a symmetry-adapted space, so the learned model is exactly symmetric by construction. The practical payoff is a large reduction in free parameters and fit time at comparable or better prediction accuracy. This matters because symmetries are common in physical and engineering models, and this gives a principled way to build them into nonlinear reduced-order modelling.","feed_headline":"Symmetries cut model-reduction cost by up to 75 percent","feed_subtitle":"Enforcing exact symmetry in spectral-submanifold fits gives faster and more faithful reduced models.","key_machinery":"The load-bearing object is the spectral submanifold (SSM): the unique smoothest invariant manifold tangent to a chosen spectral subspace $E$ of the linearised dynamics. The argument runs through three equivariance-preserving choices: a $G$-invariant complement $E^c$ (e.g. the orthogonal complement in the $G$-averaged inner product), a restricted representation $R_0$ frozen from an orbit-augmented singular value decomposition of the data, and coefficient parametrisations through the nullspace of the equivariance operator $M_k$ (equivalently, the image of the group-averaging projection $\\mathcal{R}_k$). The same treatment carries into the extended normal form, where the group action preserves the near-resonant support because the diagonalised action commutes with the linear part $\\Lambda$. These linear constraints are what convert a symmetry of the full system into a reduced parameter count.","core_discovery":"The central discovery is that equivariance is not an extra constraint bolted onto spectral-submanifold reduction but a structural property of the manifolds. For a system $\\dot{x}=Ax+f(x)$ equivariant under a compact linear group $G$, the unique smoothest spectral submanifold $W(E)$ of an invariant spectral subspace $E$ is itself $G$-invariant; the graph parametrisation $h$ obeys $h(S|_E\\eta)=S h(\\eta)$, the reduced vector field $r$ obeys $r(S|_E\\eta)=S|_E r(\\eta)$, and the extended normal form is equivariant in diagonalising coordinates. The admissible Taylor coefficients of these maps are exactly the fixed points of the group-averaging projection $\\mathcal{R}_k$, so the symmetry constraints are linear equations on the coefficients. The eSSM algorithm imposes those constraints exactly at every fitted polynomial order, producing a reduced model that is symmetric by construction.","pith_inferences":["The same coefficient-space projection could be applied to other data-driven reduced-order methods whose unknowns are polynomial coefficients, so the equivariance machinery may transfer beyond SSMs; the paper does not explore that transfer.","Because the symmetry group is supplied in advance, a natural extension would be to score candidate groups by the parameter reduction and by the equivariance residual of the fitted model, which would turn symmetry discovery into a model-selection problem.","The theory assumes exact equivariance, so a practical testable extension is to inject a controlled symmetry-breaking perturbation into the data and measure how quickly prediction accuracy degrades; this would quantify how much of the benefit is genuine robustness.","The strong performance on the chaotic benchmark, where the strict assumptions around a fixed point are violated, hints that the equivariance constraint acts as a useful regulariser beyond the rigorous domain; comparing low-data equivariant and unconstrained fits could isolate that effect."],"forward_implications":["The fitted eSSM model is exactly $G$-equivariant by construction, so it cannot drift into symmetry-broken dynamics even when trained on finite or noisy data.","Parameter counts fall sharply with the symmetry group: the parity group prunes all even-degree monomials, and a fourth-order cyclic group cuts free parameters by about 75 percent at every fitted manifold order.","Because a $G$-covariant observable leads to a $G$-equivariant delay embedding, the method applies to partial-state measurements rather than requiring full-state data.","The discrete-time variant uses a multiplicative resonance condition and inherits the same equivariance guarantees, covering iterated maps as well as continuous flows.","On a chaotic benchmark outside the strict theoretical assumptions, the equivariant fit attains accuracy comparable to the unconstrained SSM fit in roughly 40 percent of the runtime."],"supporting_citations":[{"why":"Supplies the existence and uniqueness theorem for the smoothest spectral submanifold that Theorem 2.3 adapts to this setting.","marker":"[27]"},{"why":"Original construction of spectral submanifolds as unique smoothest invariant manifolds tangent to spectral subspaces.","marker":"[28]"},{"why":"Establishes the data-driven SSM-learning pipeline and extended-normal-form fitting that the eSSM algorithm extends.","marker":"[15]"},{"why":"Gives the standard graph parametrisation and reduced-dynamics construction that the equivariant reduction builds on.","marker":"[32]"},{"why":"Introduces the fast data-driven SSM reduction and delay-embedding practices used in the algorithm.","marker":"[3]"},{"why":"Supplies the oblique-projection construction that lets the complement be chosen non-orthogonally yet symmetry-compatible.","marker":"[7]"},{"why":"The delay-embedding theorem that justifies reconstructing the state space from lower-dimensional observations in the equivariant setting.","marker":"[53]"},{"why":"Defines the standardised chaotic-PDE forecasting benchmark and error metrics used for the final evaluation.","marker":"[55]"}],"fun_headline_variants":["Symmetry-aware reduction: 75% cheaper, more faithful models","Equivariant spectral submanifolds: speed and robustness in one","Cut model-reduction cost 75% with symmetry-built manifolds","Symmetry in spectral submanifolds yields faster, better reduced models","Harnessing symmetry for fast, robust nonlinear model reduction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction depends on the data actually possessing the symmetry supplied by the user and on the leading $d$ singular values of the orbit-augmented snapshot matrix being well separated, because the restricted representation is frozen once from that SVD; if either assumption fails, the equivariance constraint imposed on the fit is the wrong one.","fun_headline_variants_meta":{"raw":{"variants":["Symmetry-aware reduction: 75% cheaper, more faithful models","Equivariant spectral submanifolds: speed and robustness in one","Cut model-reduction cost 75% with symmetry-built manifolds","Symmetry in spectral submanifolds yields faster, better reduced models","Harnessing symmetry for fast, robust nonlinear model reduction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001046,"raw_usage":{"total_tokens":4378,"prompt_tokens":906,"completion_tokens":3472,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":3381}},"tokens_in":522,"tokens_out":3472,"duration_ms":21556,"temperature":1.0,"reasoning_tokens":3381,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:08:10.445214+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a genuinely $G$-equivariant system, fit an eSSM model, and compare the prediction started at a transformed initial condition $S x_0$ with the group transform of the prediction at $x_0$; agreement within numerical tolerance confirms the equivariance of the fitted model. A sharper test is to repeat the fit when the leading $d$ singular values of the orbit-augmented snapshot matrix are almost degenerate, so that the frozen representation is unstable: the fitted manifold should then visibly break equivariance and the reduced predictions should diverge under symmetry transformations.","supporting_citations":[{"cited_title":"The Eleventh International Conference on Learning Representations , year=","cited_arxiv_id":null,"evidence_quote":"Original construction of spectral submanifolds as unique smoothest invariant manifolds tangent to spectral subspaces."},{"cited_title":"An Introductory Guide to Koopman Learning","cited_arxiv_id":null,"evidence_quote":"Establishes the data-driven SSM-learning pipeline and extended-normal-form fitting that the eSSM algorithm extends."},{"cited_title":"Nathan , journal=","cited_arxiv_id":null,"evidence_quote":"Gives the standard graph parametrisation and reduced-dynamics construction that the equivariant reduction builds on."},{"cited_title":"Higher Order Dynamic Mode Decomposition , journal =","cited_arxiv_id":null,"evidence_quote":"Introduces the fast data-driven SSM reduction and delay-embedding practices used in the algorithm."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the oblique-projection construction that lets the complement be chosen non-orthogonally yet symmetry-compatible."},{"cited_title":"Nathan and Brunton, Steven L","cited_arxiv_id":null,"evidence_quote":"The delay-embedding theorem that justifies reconstructing the state space from lower-dimensional observations in the equivariant setting."},{"cited_title":"Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=","cited_arxiv_id":null,"evidence_quote":"Defines the standardised chaotic-PDE forecasting benchmark and error metrics used for the final evaluation."}],"review_version":1}