{"id":"4a60a64b-6da9-48fe-aef6-10f664ddbb48","arxiv_id":"2608.12010","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The symmetry increase that degrades equivariant networks on symmetric inputs is governed by a unique, computable 'symmetry infimum' determined by the feature space.","lead":"This paper shows that when symmetric inputs pass through equivariant neural networks, the output often gains extra symmetry, which destroys information, and this gain has a computable lower bound set by the network's feature space. The authors derive that bound, give algorithms to compute it, and verify it on synthetic and molecular data.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.2 proves only C∞-density of almost-isovariant maps, not that most parameters of a TFN are almost isovariant; the paper's \"for most equivariant maps\" overstates what is shown.","rationale":"The reader's weakest assumption is the manifold hypothesis for data. That is a legitimate limitation, but the theorem states its assumptions explicitly and the empirical section is framed as validation rather than as a proof of the manifold assumption. I found a more direct gap between the theorem and the paper's language: the genericity transfer from equivariant function space to TFN is stated as C∞-density, but the abstract and §5.2 repeatedly say \"most equivariant maps\" and \"generic property.\" In a finite-dimensional parameter space, density does not imply probability-one typicality, and the proof does not establish openness in parameter space. The empirical random-initialization results in §6.2 are suggestive and may in practice support the stronger claim, but no theorem in the paper quantifies the measure of good parameters. I therefore recommend a conditional acceptance: the mathematical existence theorem is sound as far as I can verify, but the practical controllability claim should be either restricted to the existence/dense formulation or supplemented with an explicit typicality statement and numerical evidence that random/trained parameters are almost isovariant with high probability.","tokens_in":43784,"tokens_out":26337,"duration_ms":289481,"concrete_test":"Fix the TFN architecture used in §6.2 (e.g., one layer, l0=2, k=4-fold inputs) and draw N=200 independent random initializations. For each run, compute the embedding-difference norm between the original k-fold and rotated copies, as in §F.2.1, and compare the resulting degeneration pattern against Table 1. If a non-negligible fraction of initializations deviate from the predicted pattern (e.g., an embedding norm below 10^{-6} where Table 1 predicts distinguishability, or vice versa), then the dense-subset theorem is not sufficient for the paper's \"most maps\" claim. If the failure fraction is negligible or concentrates on a measure-zero-looking set, the concern does not land.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central practical claim is that, under regularity assumptions, most equivariant maps or sufficiently expressive ENNs achieve the symmetry infimum almost everywhere (Abstract, §5.2, Theorem 5.2). What Theorem 5.2 and Proposition D.20 actually establish is a C∞-dense subset G of the parameterized family F on which the dimension-theoretic generic behavior holds. Density does not imply measure-theoretic typicality: an open dense subset of a finite-dimensional parameter space can have arbitrarily small Lebesgue measure, and Section D.4 explicitly remarks (after Prop. D.19) that the relevant transversality property is not generally open. Thus the theorem guarantees existence of arbitrarily good approximating maps that are almost isovariant on the data manifold, but it does not show that a random initialization or a gradient-descent solution will lie in this good subset. The paper's phrase \"for most equivariant maps\" and its use of \"generic property\" for the TFN family are therefore stronger than the proof. This matters because the proposed design guidelines are meant to make symmetry increase predictable in actual ENNs, not merely achievable by some carefully chosen parameter setting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the phenomenon of symmetry increase in equivariant neural networks: for a G-equivariant map f:X→Y, the isotropy subgroup of x is contained in that of f(x). The authors prove (Thm. 3.1) that for every closed H⊆G the fixed-point subspace X^H has a unique minimal orbit type, defining a symmetry infimum, and give necessary conditions for the existence of (relative) isovariant maps (Thms. 3.2–3.3). For SO(3)/O(3) feature spaces they propose a high-multiplicity sufficiency condition for Michel's criterion (Prop. 4.2), algorithms to test orbit types and compute the infimum, and detailed tables (Appendix E) predicting full/axial/half degenerations for k-fold inputs. Under the manifold hypothesis, they prove C∞-density of TFNs in smooth equivariant maps (Thm. 5.1) and a genericity result (Thm. 5.2) asserting that, for any target map, there exist arbitrarily close approximating equivariant maps that are almost isovariant on the data manifold, with full isovariance at sufficiently high multiplicity. Experiments on synthetic k-fold point clouds and QM9 polarizability are offered as validation.","tokens_in":44043,"tokens_out":26929,"duration_ms":286449,"significance":"If the results hold, the paper provides a unifying and computable framework for an important failure mode of ENNs. The symmetry-infimum concept is natural and the derivation is largely self-contained; the proofs in Appendix B–D are detailed, the SO(3)/O(3) tables are consistent with the Michel/Ihrig-Golubitsky/Linehan-Stedman criteria, the C∞-density theorem extends Dym & Maron in a useful way, and the experiments—especially the sharp 2D/3D k-fold classification in Table 21—align with the tables. The paper also ships a reproducibility commitment with code. My reservations concern the gap between the proved density statement and the advertised 'for most equivariant maps' claim, and a missing chain argument in the general version of Prop. 4.2; both are fixable and do not undermine the core infimum calculus.","major_comments":[{"comment":"Theorem 5.2 and Prop. D.20 establish only a C∞-dense (Baire-category) statement: there is a dense subset G⊂F on which the relevant transversality and dimension bounds hold, so for every target f and ϵ one can find a good approximating map g. This does not imply that 'most' parameters of a TFN, in the measure-theoretic sense relevant to random initialization or gradient descent, are almost isovariant; an open dense subset of a finite-dimensional parameter space can have arbitrarily small Lebesgue measure, and §D.4 (Remark after Prop. D.19) explicitly notes that the openness of the transversality property does not generally hold. The abstract's 'for most equivariant maps' and §5.2's 'a significant portion of maps within a dense parameterization' are therefore stronger than what is proved. Since contribution (iii) is the basis for the practical claim that symmetry increase is predictable in actual ENNs, the paper should either prove a measure-theoretic typicality statement or replace 'most'/'significant portion' by 'topologically generic/existence' throughout.","section":"§5.2, Theorem 5.2, Prop. D.20"},{"comment":"The proof of Prop. 4.2 passes from the Michel condition on adjacent supergroups to the assertion 'for any closed subgroup H′ of G containing H, dim V^{H′}<dim V^H' without justification. This step is nontrivial for compact Lie groups whose subgroup lattice is infinite (e.g., tori): one must show that any proper supergroup H′ contains an adjacent supergroup K of H with dim V^K = dim V^{H′} (or otherwise derive the inequality by a chain argument). As written, the sufficiency of the adjacent-subgroup check—and hence the correctness of Algo. 1 for arbitrary compact G—is not fully established. The concrete SO(3)/O(3) computations in §E are explicit and checked, so the paper's applications survive, but the claimed algorithm's generality should be restricted to those groups, or the missing chain argument should be supplied.","section":"§4.1, Prop. 4.2, Algorithms 1–2"}],"minor_comments":[{"comment":"The threshold-based interpretation of Fig. 5—values above 10^-3 indicate distinguishable and below 10^-6 indistinguishable—is a numerical heuristic; it should be stated as such and ideally supported by a sensitivity analysis or explicit numerical-error bound.","section":"§6.2"},{"comment":"Several point groups in Fig. 6 and Table 22 have very small sample sizes (e.g., Td with 5, C3h with 4, D3h with 15); the text should explicitly acknowledge that the QM9 evidence for these groups is anecdotal rather than a statistical validation.","section":"§6.3"},{"comment":"The notation D^k is used in Eq. (10) but the surrounding text sometimes writes 'Dkf'; please define D^k before first use and make the notation uniform.","section":"Theorem 5.1, Eq. (10)"},{"comment":"Equation (7) contains a garbled symbol 'L∞ l0=0' and should be typeset as a direct sum over l0, e.g., ⊕_{l0=0}^{∞}.","section":"§4.1, Eq. (7)"},{"comment":"The return value 'min(O)' relies on the partial order of orbit types; please state explicitly at that point that the minimum is with respect to the orbit-type ordering and is unique by Theorem 3.1.","section":"Algorithm 2"}],"recommendation":"major_revision","confidential_remarks":"This is a strong theory submission with a sound core. The main issue is claim-Proof alignment: the advertised 'for most equivariant maps' is only a C∞-density statement, and the general compact-Lie-group version of Algorithm 1 needs a missing chain argument or a scoped statement. Both points are fixable without changing the central infimum calculus; I would support publication after the authors either weaken the relevant claims or supply the missing arguments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper gives a clean, computable characterization of symmetry increase for equivariant networks on symmetric inputs, and the main infimum machinery checks out. Theorem 3.1 is known in equivariant topology (Azzi et al. Prop. 2.10), but the application to ENN feature design, the high-multiplicity orbit-type algorithm, and the complete SO(3)/O(3) tables are genuinely new and useful. The three degeneration types for k-fold structures unify earlier observations, and the k-fold experiments align with Table 1 remarkably well. The QM9 case studies are small but the full-degeneration predictions check out on the tested point groups.\n\nThe soft spot is the genericity claim. Theorem 5.2 proves C∞-density of almost-isovariant maps, not measure-theoretic typicality. Section D.4 itself remarks that the relevant transversality property is not generally open, so the dense set could have arbitrarily small measure in parameter space. The abstract's \"for most equivariant maps\" and the use of \"generic property\" for the TFN family overstate what is actually shown. This matters because the design guidelines are meant to predict behavior of trained networks, not merely assert that some approximating map exists. The empirical visualizations with random initializations partially mitigate this, since the predicted degenerations do occur in practice, but the theoretical statement should be softened to \"for a dense set\" or supplemented with an openness argument.\n\nThe manifold hypothesis is also strong for discrete molecular data, and the QM9 validation is qualitative rather than a direct test of the topological assumptions. That said, the necessity results and the infimum tables stand on their own; the sufficiency gap doesn't undermine the core framework.\n\nI'd take this paper seriously. The right fix is to rewrite the genericity claims to match the proof and add a short discussion of why density is the best one can currently show. The tables and the feature-design guidelines are worth publishing, and I'd cite them in my own work. Send it to reviewers; with a careful revision, it should be accepted.","headline":"A genuinely useful and mostly rigorous framework for computing symmetry increase in equivariant nets, held back by a density-vs-measure gap in the 'most equivariant maps' claim.","tokens_in":44577,"tokens_out":2715,"would_cite":true,"duration_ms":30165,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For equivariant networks on symmetric inputs, the minimum possible symmetry gain is fixed by the feature space, and generic networks achieve it almost everywhere.","keywords":["equivariant neural networks","symmetry increase","orbit types","isotropy subgroups","symmetry infimum","isovariant maps","manifold hypothesis","SO(3) and O(3) representations"],"falsifier":"Train a $C^\\infty$-expressive equivariant network on a compact $G$-invariant submanifold whose orbit types satisfy $(p_Y(H))\\in O_G(Y)$, and compute the Hausdorff measure of the set of inputs where $\\rho_Y(G_x)\\neq \\rho_Y(G_{f(x)})$; if this set has positive measure for any approximant with enough multiplicity, the generic almost-isovariance claim of Theorem 5.2 is false.","tokens_in":43589,"feed_emoji":"🔄","tokens_out":10574,"duration_ms":101405,"temperature":0.7,"pith_summary":"This paper argues that the loss of expressivity known to afflict equivariant neural networks on symmetric inputs is not arbitrary: for any chosen feature space and any input symmetry, the smallest unavoidable increase in symmetry exists and is fixed by the representation structure of the feature space. It proves this lower bound, the symmetry infimum, is unique, gives a computable orbit-type algorithm for it, and shows that for generic equivariant maps, including tensor-field networks, which form a $C^\\infty$-dense family, the output symmetry equals this infimum on all but measure-zero inputs when the data lies on a finite union of compact smooth $G$-invariant submanifolds. A reader should care because this converts a known failure mode into a design tool: before training, one can predict which orientations of a symmetric object will be erased by a given feature space, and choose feature components that avoid the most damaging collapses.","feed_headline":"Feature space sets the floor for symmetry gain in equivariant nets","feed_subtitle":"A new infimum formula predicts when symmetric inputs collapse, and generic networks hit that limit almost everywhere.","key_machinery":"The central object is the symmetry infimum $I_G(Y,H)$, the unique smallest orbit type appearing in the fixed-point subspace $Y^H$; it is what the feature space algebraically forces an equivariant map to reach. The argument is carried by three devices: the operator $p_Y$ that enlarges a subgroup by the kernel of the feature representation and thereby separates designed, unavoidable increases from accidental ones; an orbit-type test based on Michel's Criterion, which compares fixed-point-space dimensions at adjacent supergroups and becomes sufficient for high-multiplicity representations where every isotypic component has multiplicity exceeding $\\dim G$, with a trace formula computing those dimensions; and a genericity argument on Whitney regular orbit-type stratifications showing that, in a $C^\\infty$-dense equivariant family, transversality makes the dimension of the set of too-symmetric outputs strictly smaller than the data stratum, so almost-isovariance holds almost everywhere. A counterexample in the appendix shows that orbit-type inclusion alone is not enough for full isovariance, which is why the multiplicity condition $r>\\max_j\\dim M_j$ carries real weight.","core_discovery":"The paper's central claim is that the degradation has a universal lower bound. For a representation $Y$ of a compact Lie group $G$ and a closed subgroup $H$ of $G$, the fixed-point subspace $Y^H$ contains points of several orbit types, and Theorem 3.1 asserts that among them there is a unique minimal one, the symmetry infimum $I_G(Y,H)$. Equivariance forces the isotropy subgroup of the output to contain that of the input, and generically it should be exactly the image of $I_G(Y,H)$ under the kernel-projection operator $p_Y$ rather than something larger. Theorem 5.2 makes this precise: if the data is supported on a finite union of compact smooth $G$-invariant submanifolds and the map family has $C^\\infty$ approximation capability, then dense approximants are almost isovariant relative to $Y$, with $\\rho_Y(G_x)=\\rho_Y(G_{f(x)})$ holding on every stratum except a subset of measure zero; if the feature space contains more than $\\max_j \\dim M_j$ copies of each needed component, full relative isovariance can be achieved. Thus the paper claims that the expressive collapse of equivariant networks on symmetric data is predictable and, in principle, avoidable by design.","pith_inferences":["This suggests a compositional design principle that goes beyond the paper's stated guidelines: because the infimum of a direct sum can only be smaller than or equal to the infimum of each summand, adding features induces a partial order on symmetry loss, and bottleneck-type subgroups determine when new components can introduce new orbit types.","The same orbit-type stratification should predict degeneracies in any smooth equivariant map, not just neural networks, so learned equivariant simulators or generative models on rotationally symmetric data should display the same collapse near symmetric states.","A testable extension is to perturb data off the manifold: if almost-isovariance is the mechanism, adding noise should break the almost-everywhere guarantee at finite sample size, with the observed degeneracies controlled by the local dimension of the data distribution.","The closed-form infimum tables for $SO(3)$ and $O(3)$ could seed an architecture search that scores candidate irrep selections by whether $I_G(Y,H)$ equals $p_Y(H)$ for the molecular point groups present in a dataset, and uses that score as a feature-selection regularizer."],"forward_implications":["A fixed feature space has a fixed minimal symmetry gain: every equivariant network built on that feature space must lose at least the information erased by $I_G(Y,H)$, regardless of architecture or training.","The algorithm gives a pre-training check: compute $I_G(Y,H)$ for the input symmetries present in a dataset, and the degeneracies, such as full, axial, or half collapse for $k$-fold objects, are known before any experiment.","Because almost isovariance is generic for expressive maps, sufficiently capable equivariant models should exhibit exactly the predicted infimum behavior on symmetric inputs, not a worse collapse; the paper's TFN, HEGNN, and QM9 experiments are consistent with this.","Increasing the multiplicity of feature components above the dimension of the data manifold removes the remaining exceptional inputs and yields full relative isovariance, providing a constructive remedy.","Feature-selection guidelines follow: orientation-dependent tasks should include components containing the projected orbit type $p_Y(H)$, and general tasks should avoid components whose infimum is the full symmetry group, since those components are annihilated."],"supporting_citations":[{"why":"States Curie's principle for equivariant maps, the inclusion $G_x\\subseteq G_{f(x)}$ from which the whole analysis of forced symmetry increase starts.","marker":"Kaba & Ravanbakhsh (2023)"},{"why":"Supplies the fixed-point-dimension criterion used by the orbit-type test to decide whether a subgroup is an isotropy subgroup.","marker":"Michel (1980)"},{"why":"Provides the sufficiency criterion that makes Michel's condition into an exact orbit-type test for high-multiplicity representations.","marker":"Ihrig & Golubitsky (1984)"},{"why":"Gives the trace formula used to compute fixed-point-space dimensions in the infimum algorithm.","marker":"Golubitsky et al. (1988)"},{"why":"Establishes TFN universality, which the paper strengthens to $C^\\infty$ density in Theorem 5.1.","marker":"Dym & Maron (2021)"},{"why":"Supplies the higher-order MLP approximation theorem used to convert polynomial equivariant approximation into $C^\\infty$ density.","marker":"Pinkus (1999)"},{"why":"Provides the Whitney regular orbit-type stratifications and $G$-manifold structure theory used in the genericity proof of Theorem 5.2.","marker":"Field (2007)"},{"why":"Documents the empirical symmetry-dependent degradation that the paper's theory is built to explain and reproduce.","marker":"Joshi et al. (2023)"},{"why":"Gives the earlier collapse-to-zero theory and the HEGNN architecture used as the model backbone in the QM9 validation.","marker":"Cen et al. (2024)"}],"fun_headline_variants":["Equivariant nets: symmetry gain has a floor","Infimum theorem explains symmetric input collapse","Design features to dodge symmetry increase","Generic equivariant maps hit the infimum","A computable floor for symmetry increase"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proofs that generic equivariant maps realize the infimum assume the data lies exactly on a finite union of compact, smooth, $G$-invariant submanifolds of the input space; real point clouds and molecules are discrete, finite, and noisy, so the almost-everywhere statement applies to that idealization rather than to raw data.","fun_headline_variants_meta":{"raw":{"variants":["Equivariant nets: symmetry gain has a floor","Infimum theorem explains symmetric input collapse","Design features to dodge symmetry increase","Generic equivariant maps hit the infimum","A computable floor for symmetry increase"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000302,"raw_usage":{"total_tokens":1784,"prompt_tokens":1031,"completion_tokens":753,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":688}},"tokens_in":647,"tokens_out":753,"duration_ms":8325,"temperature":1.0,"reasoning_tokens":688,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:19:33.554194+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a $C^\\infty$-expressive equivariant network on a compact $G$-invariant submanifold whose orbit types satisfy $(p_Y(H))\\in O_G(Y)$, and compute the Hausdorff measure of the set of inputs where $\\rho_Y(G_x)\\neq \\rho_Y(G_{f(x)})$; if this set has positive measure for any approximant with enough multiplicity, the generic almost-isovariance claim of Theorem 5.2 is false.","supporting_citations":[{"cited_title":"Pattern selection with","cited_arxiv_id":null,"evidence_quote":"Provides the sufficiency criterion that makes Michel's condition into an exact orbit-type test for high-multiplicity representations."}],"review_version":1}