{"id":"988b1f2f-0ee1-4a9b-a1a4-039f5e611b7c","arxiv_id":"2504.17258","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A subgroup-based downsampling layer with an equivariant anti-aliasing projection for finite groups, shown to cut parameters and modestly improve equivariance and rotated-input accuracy in small G-CNN experiments.","lead":"This paper introduces a downsampling layer for group-equivariant neural networks that reduces the symmetry group to a subgroup and filters out aliasing before subsampling. It includes an algorithm to pick the subgroup, a sampling theorem for finite-group signals, and experiments on rotated MNIST, CIFAR-10, and STL-10 showing moderate accuracy and equivariance gains at lower parameter counts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'equivariant anti-aliasing' operator is only approximately equivariant (Eq.","rationale":"The Reader's verdict identified Eq. (14)'s relaxed equivariance constraint and the absence of an error bound as the weakest assumption. My stress-test concurs and sharpens the concern: the issue is not merely a missing numerical bound but a possible structural incompatibility. An exactly equivariant real projection onto a bandlimited subspace must select whole real irreducible subrepresentations of the regular representation, whereas the classical ideal low-pass subspace (the natural 'bandlimited' set for real signals) generally cuts through those irreps. Therefore the trade-off encoded in Eq. (14) may be forced, and the reported Lequi improvements cannot be cleanly separated from the non-equivariance of the anti-aliasing layer itself. That said, Claim 2 as a conditional reconstruction theorem is correct, and the empirical results do show consistent Lequi reductions, so the paper's core engineering contribution remains plausible. The appropriate remedy is not rejection but a required qualification: the paper should state that the anti-aliasing layer is approximately equivariant, provide the achieved equivariance residual, and temper the abstract's 'recovers standard ideal low-pass downsampling' claim. Since the Reader already reached CONDITIONAL for overlapping reasons, my independent assessment does not move the verdict; I would keep it CONDITIONAL and ask for the hard-constraint feasibility check and residual reporting before final acceptance.","tokens_in":24770,"tokens_out":15837,"duration_ms":164319,"concrete_test":"Solve Eq. (14) for C16→C8 with the hard equivariance constraint vec(\\hat P_M) = \\bar T vec(\\hat P_M) imposed as an equality (e.g., by projected gradient or SLSQP with λ→∞), and compare the resulting M with the ideal low-pass M from Example 3 / Eq. (12). Also report the residual ||vec(\\hat P_M*) - \\bar T vec(\\hat P_M*)||_2 for the published λ=5 and compare it with the Lequi improvements in Table 2 (e.g., R=2 CIFAR O(2): 0.398→0.336; R=3: 0.549→0.379). If no real solution coincides with the low-pass subspace, or if the residual is on the same scale as those Lequi gaps, then the equivariance gain cannot be attributed to the anti-aliasing operator's equivariance and the 'equivariant anti-aliasing' claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Claim 2 is internally sound but close to definitional: it guarantees perfect reconstruction only for signals already in range(P_M), the column space of the interpolation matrix, and Table 1 validates that identity numerically rather than testing a substantive hypothesis. The genuinely load-bearing part of the paper is the assertion that the anti-aliasing projection is G-equivariant and behaves like an ideal low-pass filter. That part is not proven. Eq. (14) replaces the hard equivariance constraint vec(\\hat P_M) = \\bar T vec(\\hat P_M) with a penalty term weighted by λ=5, so the reported P_M* is not exactly equivariant. No theorem bounds ||vec(\\hat P_M*) - \\bar T vec(\\hat P_M*)||_2, and no propagation analysis connects this residual to the measured Lequi after several equivariant layers and pooling. Furthermore, for real-valued feature maps the two objectives are in structural tension: an exactly equivariant real projection must have a range that is a direct sum of real irreducible subrepresentations of the regular representation, while the classical low-pass subspace is generally not of this form. For C8→C4, the real invariant subspace of dimension 4 necessarily includes the Nyquist component and omits part of the k=2 pair, so it is not the low-pass subspace. Thus the abstract's claim that the method recovers 'the standard downsampling of an ideal low-pass filter' cannot hold simultaneously with exact equivariance in general, and the penalty in Eq. (14) may be concealing an infeasibility rather than a small numerical relaxation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for downsampling feature maps on finite groups inside group equivariant CNNs. Given a finite group G and a subsampling rate R, it provides an algorithm (Algorithm 1) that returns a subgroup G_down by modifying the Cayley graph of G, together with a heuristic for choosing which generator to subsample along. It then defines bandlimited signals on G relative to a matrix M satisfying F^{-1}_{G_down} = S F^{-1}_G M (Eq. 9), states a Subgroup Sampling Theorem (Claim 2) guaranteeing perfect reconstruction from subgroup samples for such bandlimited signals, and proposes to construct an anti-aliasing operator P_{M*} by solving the constrained optimization problem in Eq. (14), which trades off an equivariance penalty against a smoothness penalty. Experiments on rotated MNIST, CIFAR-10, and STL-10 report that the proposed layer reduces parameter count, improves classification accuracy, and lowers measured equivariance error Lequi compared to naive subgroup subsampling.","tokens_in":25016,"tokens_out":9101,"duration_ms":88007,"significance":"The paper addresses a real gap in group equivariant architectures: how to downsample a feature map on a general finite group at a specified rate. The proposed subgroup selection algorithm and the interpolation framework are novel, and the authors ship code and provide numerical validation of the reconstruction claim (Table 1). The experimental results consistently show that the anti-aliasing operator improves accuracy and equivariance over naive subsampling in small-scale settings, which is a useful practical contribution. However, the theoretical core is thin: Claim 2 is essentially a direct consequence of the defining constraint in Eq. (9), and the headline claim that the anti-aliasing operator is 'equivariant' is not actually established because Eq. (14) relaxes exact equivariance to a penalty. If the equivariance residual and its propagation are left unquantified, the central claim of the paper is not supported as stated.","major_comments":[{"comment":"The anti-aliasing operator is not proven to be G-equivariant. The exact equivariance condition vec(P_hat_M) = T vec(P_hat_M) is replaced by a penalty term with λ = 5, so the reported P_{M*} is not guaranteed to be exactly equivariant. The paper provides no bound on the residual ||vec(P_hat_{M*}) - T vec(P_hat_{M*})||_2, nor any analysis connecting this residual to the measured Lequi in Table 2 or to the layer-wise equivariance errors in Figures A1-A8. Since 'equivariant anti-aliasing' is a central contribution and the empirical equivariance improvements are attributed to it, this missing quantification is load-bearing. I request that the authors either enforce exact equivariance (e.g., by projecting the solution onto the Reynolds fixed-point space) or provide a residual bound and a propagation analysis, and that they qualify the term 'equivariant' if only approximate equivariance is achieved.","section":"§4.2, Eq. (14)"},{"comment":"The Subgroup Sampling Theorem is definitional rather than a substantive sampling-theoretic result. The bandlimited subspace is defined as range(B) with B = F^{-1}_G M, and the interpolation operator is defined as I = B F_{G_down}; Eq. (9), namely F^{-1}_{G_down} = S B, is imposed as a constraint. Under these definitions, I S is the identity on range(B) by construction, so Claim 2 is a verification that the constraint achieves what it was designed to achieve. Table 1 numerically validates this identity but does not test an independent hypothesis. The presentation as a 'sampling theorem' generalizing classical sampling theory overstates the novelty; the paper should clearly acknowledge that the reconstruction guarantee is built into the definition of the bandlimited subspace and the interpolation operator.","section":"§4.2, Claim 2"},{"comment":"The claimed recovery of classical low-pass downsampling on cyclic groups is problematic for real-valued signals. The matrix M in Eq. (12) selects the first N/2 complex Fourier coefficients; for a real signal, this coefficient set is not conjugate-symmetric, so the inverse Fourier transform is not real-valued in general. Moreover, for the concrete case C8 → C4, the classical low-pass subspace (DC, the k=1 cosine/sine pair, and the k=2 cosine) splits the 2D real irrep at k=2, so no exactly G-equivariant real projection can have that subspace as its range. Thus the method can only approximately recover the ideal low-pass operation, and the approximation error is not characterized. The abstract and Section 4.2 should be qualified, or a precise condition should be given under which equality (rather than approximation) holds.","section":"Example 3 and Abstract"},{"comment":"There is a direct contradiction between the heuristic described in the text and the provided pseudocode. Section 4.1 states that the algorithm selects the generator with the maximum order satisfying the constraints of Claim 1, but Algorithm 3, line 11, selects the generator with the minimum order via the condition 'if (index = NULL OR O[j] < O[index])'. This makes the subgroup selection procedure ambiguous and prevents reproducibility of the reported experiments from the paper alone. The text or the pseudocode must be corrected, and the choice used in the experiments should be stated explicitly.","section":"§4.1 and Algorithm 3"}],"minor_comments":[{"comment":"The paper sets λ = 5 in Eq. (14) but provides no sensitivity analysis and no report of the actual equivariance residual of P_{M*} for the groups used in the experiments. Reporting the residual ||vec(P_hat_{M*}) - T vec(P_hat_{M*})||_2 for C24 and D24 would directly address the main concern about approximate equivariance.","section":"§A7"},{"comment":"The displayed optimization objective is typeset in a garbled way in the provided text (the norm, Diag, and absolute-value expressions are not rendered cleanly). The equation should be typeset carefully so that the equivariance objective and smooth selection objective are legible.","section":"§4.2, Eq. (14)"},{"comment":"The example mixes the complex Fourier basis (used for the matrix M in Eq. (12)) with the real cosine/sine basis discussed in the preceding sentences. The relationship between these bases and the notion of 'ascending frequency' for real signals should be clarified, especially regarding conjugate symmetry.","section":"Example 3"},{"comment":"The 'Recon. Err.' column uses an 'a/b' format to present errors with and without anti-aliasing, but the caption does not define this notation. Please clarify the format in the caption or table note.","section":"Table 1"},{"comment":"The proof of Claim 1 contains several typographical errors and garbled expressions, including 'od mod r' (should be o_d mod R) and unreadable products such as 'sRnm d' and 'ski_d'. The proof should be carefully proofread and reformatted.","section":"§A3.3"},{"comment":"The heading 'RESUTL' is a typo; it should be 'RESULTS'.","section":"§A2.1"},{"comment":"The pseudocode for Check-Compliance lists 'Generators' in the input but then uses the variable s_d without defining it; the notation should be made consistent with Algorithm 1.","section":"Algorithm 2"}],"recommendation":"major_revision","confidential_remarks":"The structural tension between exact equivariance and the classical low-pass subspace for real signals is a correctness-risk concern that the authors do not address; it is the kind of issue that might be resolved by reframing the claims as approximate, but it needs to be addressed explicitly. The paper has already appeared at ICLR 2025; if this arXiv version is a prior submission, the referee report should focus on the scientific content rather than venue fit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new thing here is the subgroup-selection algorithm (Alg. 1/3) that picks a subgroup from a rate R, plus the optimization in Eq. (14) that learns a projection approximating equivariant anti-aliasing. The paper places this well in the G-CNN context, and the experiments show the anti-aliasing layer consistently lowers equivariance error and often improves rotated test accuracy. Credit where due: the subgroup selection is useful, and I would consider using it.\n\nThe soft spots are real. Claim 2 is close to a tautology: bandlimited is defined as being in the range of the interpolation operator, so perfect reconstruction follows from the definition. Table 1 verifies that identity numerically; it does not test a substantive hypothesis. More importantly, the \"equivariant\" anti-aliasing is not exactly equivariant. Eq. (14) enforces vec(P_M)=Tbar vec(P_M) only as a penalty with λ=5, and the paper gives no bound on the residual or how it propagates through stacked layers. The stress-test note is right that for real feature maps there is a structural tension: an exactly equivariant real projection has a range that decomposes into real irreps, while a classical low-pass subspace is not of that form for, say, C8→C4. So the abstract's claim that the method recovers ideal low-pass filtering while maintaining equivariance cannot hold exactly in general. The empirical results suggest the approximation is acceptable in practice, but the claim should be softened.\n\nThere are also small errors: Algorithm 3's text contradicts the pseudocode (the reader flags max order), and the notation in Eq. (14) is under-specified. These are minor.\n\nWho is this for? Researchers building scale-efficient equivariant CNNs who need a concrete downsampling layer with a principled anti-aliasing step. The theory is elementary but the packaging is useful. It deserves a serious referee, but a conditional accept with demands to fix the pseudocode, temper the abstract, and add a bound or at least a numerical study of the equivariance residual.","headline":"A useful subgroup-sampling routine for G-CNNs whose 'equivariant anti-aliasing' is an unproven approximation rather than an exact operator.","tokens_in":25617,"tokens_out":2862,"would_cite":true,"duration_ms":27094,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sampling theory for finite groups makes subgroup downsampling perfectly reversible","keywords":["group equivariant networks","subgroup downsampling","anti-aliasing","sampling theorem","Fourier transform on finite groups","equivariance","bandlimited signals","Cayley graph"],"falsifier":"Measure the residual equivariance error of the learned anti-aliasing operator directly by computing the norm of $\\mathrm{vec}(P_{M^*})$ minus the Reynolds-projected version of $\\mathrm{vec}(P_{M^*})$, and check whether it grows with group order or with network depth; alternatively, propagate a fixed input through a deep G-CNN with and without anti-aliasing and test whether the layerwise equivariance error diverges at higher layers. If the residual error grows sharply on larger groups or deeper networks, the empirical equivariance claim would fail.","tokens_in":1624,"feed_emoji":"🔄","tokens_out":2129,"duration_ms":63672,"temperature":0.7,"pith_summary":"The paper aims to show that the familiar CNN operation of downsampling—low-pass filtering followed by keeping every R-th sample—can be generalized from translation grids to feature maps indexed by any finite group. It supplies an algorithm that, given a group and a downsampling rate, returns a suitable subgroup to downsample to, and it defines bandlimited-ness relative to that subgroup. Its central theoretical result, the Subgroup Sampling Theorem, says that a signal whose Fourier coefficients lie in a specific eigenspace can be reconstructed exactly from samples on the subgroup. The paper then constructs an equivariant anti-aliasing projection that approximately enforces that bandlimited condition. If the claim is right, group-equivariant networks can shrink in parameter count and memory without sacrificing accuracy or symmetry guarantees.","feed_headline":"Sampling theorem makes group downsampling perfectly reversible","feed_subtitle":"Equivariant anti-aliasing lets G-CNNs shed parameters without losing accuracy or symmetry.","key_machinery":"The load-bearing object is the matrix M, a map from Fourier coefficients on the subgroup to Fourier coefficients on the whole group, together with the sampling matrix S and the interpolation operator $I = B F_{G_{\\downarrow}}$ where $B = F_G^{-1} M$. The perfect reconstruction condition is that $\\hat{x}$ equals $M(M^*M)^{-1}M^*\\hat{x}$, and the anti-aliasing operator is the projection $P_M = B(B^*B)^{-1}B^*$ onto the column space of B. The paper constructs M by optimizing Eq. (14), whose equivariance term penalizes deviation from the Reynolds-operator fixed space and whose smoothness term penalizes high graph-Laplacian energy. Subgroup selection is driven by a Cayley-graph algorithm that replaces a generator $s_d$ with its R-th power $s_d^R$, giving a concrete generalization of 'keep every R-th element' to arbitrary finite groups.","core_discovery":"The central claim is the Subgroup Sampling Theorem: for any signal x on a finite group G, if the Fourier coefficient vector $\\hat{x}$ lies in the 1-eigenspace of the projection $M(M^*M)^{-1}M^*$, where M connects the Fourier basis of the subgroup $G_{\\downarrow}$ to that of G, then x can be reconstructed perfectly from the subsampled signal on $G_{\\downarrow}$. The paper finds such an M by solving a constrained optimization that balances an equivariance objective (membership in the fixed space of the Reynolds operator) against a smoothness objective based on the Cayley-graph Laplacian, with exact perfect reconstruction as a constraint. For cyclic groups, the construction reduces to ideal low-pass filtering followed by ordinary subsampling, which is the paper's argument that it generalizes classical sampling theory. Empirically, inserting this downsampling layer into group-equivariant CNNs reduces parameter count, improves or matches accuracy, and lowers equivariance error relative to subsampling without anti-aliasing.","pith_inferences":["Because the reconstruction condition is linear, one could in principle compose multiple sequential downsampling stages using the same M machinery; the paper does not derive such a composition rule.","The equivariance penalty in Eq. (14) uses a fixed balancing weight, so a natural test is whether the residual equivariance error grows with group order or with the depth of the equivariant network.","The subgroup-choice heuristic maximizes the number of generators kept; task-adaptive or learned subgroup selection is an untested alternative that could outperform the fixed heuristic on data with anisotropic symmetry content.","The interpolation operator gives a concrete way to reconstruct latent group features, which could be used as a regularizer or as a bridge for skip connections in equivariant architectures with multiple resolutions."],"forward_implications":["Group-equivariant CNNs can downsample by an integer rate R along chosen generators, with a concrete subgroup produced by Algorithm 1 instead of requiring hand-picked subgroups.","For any such subgroup, anti-aliasing via $P_M$ makes downsampled feature maps losslessly reconstructible from the subgroup samples, provided the bandlimited condition holds.","When the group is cyclic, the method reduces to classical downsampling: an ideal low-pass filter followed by subsampling.","In image classification experiments, the layer reduces model size while improving or matching accuracy and lowering equivariance error compared with subsampling without anti-aliasing.","The anti-aliasing filter needs to be computed only once per architecture before training, because the optimization in Eq. (14) does not depend on the network weights."],"supporting_citations":[{"why":"Introduces group-equivariant CNNs and the subsampling conventions that this paper generalizes; also supplies the lifting and group-convolution machinery used in the experiments.","marker":"Cohen & Welling, 2016"},{"why":"Provides the classical sampling-theory setup of bandlimited signals, sampling matrices, and interpolation that the paper extends to finite groups.","marker":"Vetterli et al., 2014"},{"why":"Defines the ideal low-pass anti-aliasing filter that the cyclic-group case recovers as a special case.","marker":"Shannon, 1949"},{"why":"Gives the orthonormal Fourier basis for functions on groups, which underlies the group Fourier transform and the construction of M.","marker":"Peter & Weyl, 1927"},{"why":"Supplies the Reynolds-operator characterization of equivariant linear maps used in the anti-aliasing objective.","marker":"Mouli & Ribeiro, 2021"},{"why":"Presents previous group-equivariant subsampling that requires a given subgroup and lacks anti-aliasing; serves as a comparison point for the proposed method.","marker":"Xu et al., 2021"},{"why":"Provides the real-irrep Fourier basis construction and the steerable CNN framework used in the experiments.","marker":"Cesa et al., 2021"},{"why":"Supplies BlurPool-style anti-aliasing for spatial subsampling, used alongside the proposed group-dimension anti-aliasing in the architecture.","marker":"Zhang, 2019"}],"fun_headline_variants":["Equivariant anti-aliasing makes group downsampling lossless","Subgroup sampling theorem: reversible downsampling for group CNNs","Perfect reconstruction downsampling on any finite group","Anti-aliased downsampling for G-CNNs: sharper symmetry"],"cache_read_input_tokens":27648,"weakest_assumption_plain":"The anti-aliasing projection is only approximately equivariant: equation (14) relaxes exact equivariance into a penalty term with a fixed balancing weight, and the paper gives no bound on how residual equivariance error accumulates through stacked convolution layers.","fun_headline_variants_meta":{"raw":{"variants":["Equivariant anti-aliasing makes group downsampling lossless","Subgroup sampling theorem: reversible downsampling for group CNNs","Perfect reconstruction downsampling on any finite group","Anti-aliased downsampling for G-CNNs: sharper symmetry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000849,"raw_usage":{"total_tokens":3700,"prompt_tokens":957,"completion_tokens":2743,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2669}},"tokens_in":573,"tokens_out":2743,"duration_ms":18078,"temperature":1.0,"reasoning_tokens":2669,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:47:16.319345+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the residual equivariance error of the learned anti-aliasing operator directly by computing the norm of $\\mathrm{vec}(P_{M^*})$ minus the Reynolds-projected version of $\\mathrm{vec}(P_{M^*})$, and check whether it grows with group order or with network depth; alternatively, propagate a fixed input through a deep G-CNN with and without anti-aliasing and test whether the layerwise equivariance error diverges at higher layers. If the residual error grows sharply on larger groups or deeper networks, the empirical equivariance claim would fail.","supporting_citations":[{"cited_title":"Group equivariant convolutional networks","cited_arxiv_id":null,"evidence_quote":"Introduces group-equivariant CNNs and the subsampling conventions that this paper generalizes; also supplies the lifting and group-convolution machinery used in the experiments."},{"cited_title":"Communication in the presence of noise","cited_arxiv_id":null,"evidence_quote":"Defines the ideal low-pass anti-aliasing filter that the cyclic-group case recovers as a special case."},{"cited_title":"The completeness of the primitive representations of a closed continuous group","cited_arxiv_id":null,"evidence_quote":"Gives the orthonormal Fourier basis for functions on groups, which underlies the group Fourier transform and the construction of M."},{"cited_title":"A program to build E(N) -equivariant steerable CNNs","cited_arxiv_id":null,"evidence_quote":"Provides the real-irrep Fourier basis construction and the steerable CNN framework used in the experiments."},{"cited_title":"Making convolutional networks shift-invariant again","cited_arxiv_id":null,"evidence_quote":"Supplies BlurPool-style anti-aliasing for spatial subsampling, used alongside the proposed group-dimension anti-aliasing in the architecture."}],"review_version":1}