{"id":"32a3d65c-405f-4218-b2a5-68d6979e0507","arxiv_id":"2603.23645","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For planar rotations, real phase retrieval, and finite reflection groups, linear transforms of max filter banks achieve distortion arbitrarily close to the Euclidean distortion of the orbit space.","lead":"Near-optimal Euclidean embeddings of orbit spaces for rotations, phase retrieval, and reflection groups can be obtained by composing a max filter bank with a linear map. This gives a practical route to low-distortion group-invariant features for signal processing and equivariant machine learning.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The reader's weakest-assumption note correctly isolates the non-vanishing of Fourier/Gegenbauer coefficients as the analytic hinge for cases (a)–(b). Those coefficients are computed explicitly and shown nonzero by elementary (if tedious) calculations; the remainder of the argument (differentiation under the integral, dominated convergence for the Lip-norm approximation, finite-dimensional spanning for reflections) uses only standard tools and is free of hidden gaps. Numerics are supportive but not required for the proved theorems. Consequently the ACCEPT verdict stands; the open generalization questions (Problem 6) are already flagged by the authors and do not undermine the claims that are proved.","tokens_in":24854,"tokens_out":478,"duration_ms":5373,"concrete_test":"Independently recompute the Fourier coefficients of the 2π/r-periodic max filter f(\theta)=cos \theta on [-\\pi/r,\theta/r] (last display of the proof of Theorem 14) and the even Gegenbauer integrals c_{2m} via the Rodrigues formula (end of the proof of Theorem 16); confirm they remain nonzero for all r>1 and all even degrees 2m with d\nge3. If any vanishes, the integral-transfer step fails for that polynomial.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 3) rests on the coordinate functions of the known optimal embeddings lying in the Lipschitz closure of the span of max filters (via Lemma 4 + Theorem 5). For cases (a) and (b) this is obtained by writing every G-invariant polynomial as an integral combination of max filters against a Lipschitz density q (Theorems 14 and 16) and then approximating the integral by Riemann sums in the Lip norm (Theorem 13). The only potential soft spot is non-vanishing of the relevant Fourier/Gegenbauer coefficients; the paper supplies explicit formulae (Fourier series of the periodic max filter for SO(2); Rodrigues formula + integration by parts for even Gegenbauer polynomials when d>2) that establish the coefficients are nonzero. Case (c) is elementary linear algebra (span M equals the finite-dimensional space of coordinate functions of the Weyl-chamber projection). No internal inconsistency or missing hypothesis appears in these arguments.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper studies bilipschitz embeddings of orbit spaces R^d/G for finite G ≤ O(d). Its main theorem (Theorem 3) asserts that for three classical families—nontrivial finite subgroups of SO(2), the phase-retrieval group {±I}, and finite reflection groups—the Euclidean distortion c_2(R^d/G) is nearly achieved by a linear post-composition of a max-filter bank. The argument proceeds by showing that the coordinate functions of the known optimal embeddings lie in the Lipschitz closure of the span of max filters (Theorem 5), then invoking a continuity-of-distortion lemma (Lemma 4). For the first two families this is realized by writing every G-invariant polynomial on the sphere as an integral combination of max filters against a Lipschitz density (Theorems 14 and 16) and approximating the integral in the Lipschitz norm by Riemann sums controlled via coarea estimates on Voronoi boundaries (Theorem 13). For reflection groups the claim reduces to elementary linear algebra: the span of max filters coincides with the finite-dimensional space of Weyl-chamber projections. Numerical experiments on additional groups and two shape datasets support the broader utility of the linear-max-filter architecture.","tokens_in":25016,"tokens_out":764,"duration_ms":7691,"significance":"The result closes a concrete gap between the known Euclidean distortions of three fundamental orbit spaces and the distortions previously obtained from max-filter banks alone. The architecture—integral transfer of invariant polynomials followed by Lipschitz-norm approximation—is new and supplies an explicit, constructive route from harmonic analysis to bilipschitz embeddings. The proofs are fully written, the non-vanishing of the relevant Fourier and Gegenbauer coefficients is established by direct computation (Rodrigues formula and integration by parts), and the numerical section demonstrates that the same linear-max-filter construction continues to approach optimal distortion on groups outside the three theoretical cases. These features make the paper a solid contribution to bilipschitz invariant theory and to the design of group-invariant feature maps.","major_comments":[],"minor_comments":[{"comment":"The abstract and title supplied in the submission metadata describe an entirely different paper on synchronized singular forms and coarea reduction. The body is the bilipschitz-invariant-theory manuscript. The metadata should be corrected before publication.","section":null},{"comment":"In the proof of Theorem 13 the constant hidden in the O(ε) bound depends on |G|, ∥q∥_Lip and surface measures ω_{d-2}; an explicit dependence would make the quantitative approximation rate clearer.","section":null},{"comment":"Section 5 reports empirical distortions for several groups not covered by Theorem 3 (e.g., C_2/S^1, (R^2)^2/O(2)). A short remark clarifying that these are numerical evidence only, not theorems, would prevent misreading.","section":null},{"comment":"The notation spanM versus its Lipschitz closure is introduced late; a single sentence in §1.3 defining the bar notation would improve readability.","section":null}],"recommendation":"accept","confidential_remarks":"The mismatch between the arXiv abstract/title (coarea reduction for synchronized kernels) and the actual manuscript (bilipschitz max-filter approximation) is almost certainly a packaging error on the authors’ side; the body itself is coherent and self-contained. Once the metadata are fixed the paper is ready for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The paper that actually sits in the manuscript is Cahill–Iverson–Mixon–Willey on bilipschitz invariants (arXiv:2603.23643), not the coarea/sparse-transfer abstract in the header. What is new is clean: for planar rotations, real phase retrieval, and finite reflection groups, the coordinate functions of the known optimal embeddings lie in the Lipschitz closure of the span of max filters (Theorem 5). Lemma 4 then immediately upgrades that to near-optimal distortion for a linear layer on a max-filter bank (Theorem 3). For reflections the approximation is exact and finite-dimensional; for the other two cases they write every invariant polynomial as an integral combination of max filters (Theorems 14 and 16) and approximate the integral by Riemann sums in the Lip norm (Theorem 13).\n\nThe analytic steps look solid. Non-vanishing of the Fourier coefficients for SO(2) and of the even Gegenbauer coefficients for phase retrieval is checked by hand (explicit series plus Rodrigues + integration by parts). Differentiability and piecewise-linear arguments for the proper containments in Theorem 5 are standard and carefully written. Numerics on a few extra groups and two real shape datasets are supportive, though unreproducible without code.\n\nSoft spots are minor and already flagged by the authors. Everything is case-by-case; they openly leave a unified argument and the general-G question as open problems. The numerics are illustrative rather than systematic. Citation pattern is normal for a group that has been developing the subject.\n\nThis is for people already working on bilipschitz invariant theory or equivariant feature maps. It does not invent a new general method, but it closes a concrete gap that the community has been staring at. I would send it to peer review without hesitation; the proofs are there and the result is useful. Worth reading if you care about max filters or orbit-space embeddings.","headline":"Solid case-by-case proof that linear post-composition of max filters recovers near-optimal Euclidean distortion for the three classical families of finite orthogonal groups.","tokens_in":25627,"tokens_out":474,"would_cite":true,"duration_ms":5701,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["46E35","43A85"],"pacs":[],"model":"grok-4.5","headline":"For planar rotations, real phase retrieval and finite reflection groups, a linear map applied to a max filter bank nearly achieves the optimal Euclidean distortion of the orbit space.","keywords":["bilipschitz invariants","max filtering","orbit spaces","Euclidean distortion","phase retrieval","reflection groups","Lipschitz functions","positively homogeneous maps"],"falsifier":"Exhibit a single G-invariant polynomial on the sphere whose associated Fourier or Gegenbauer coefficients all vanish; the integral-transfer step then fails and the Lipschitz-closure claim collapses for that polynomial.","tokens_in":25687,"feed_emoji":"📐","tokens_out":908,"duration_ms":20729,"temperature":0.7,"pith_summary":"Bilipschitz invariant theory seeks low-distortion Euclidean embeddings of orbit spaces R^d/G for finite subgroups G of the orthogonal group. Optimal distortions are known exactly for three classical families (planar rotations, sign flips for phase retrieval, and reflection groups), but the generic max-filter-bank construction only reaches strictly larger distortion. This paper proves that post-composing a sufficiently rich max filter bank with a single linear layer recovers the optimal (or arbitrarily near-optimal) distortion in all three cases. The argument reduces the geometric claim to a pure function-space inclusion: the coordinate functions of the known optimal embeddings lie in the Lipschitz closure of the linear span of max filters. Once that inclusion is established, a general continuity lemma for distortion upgrades approximation in Lipschitz norm into approximation of distortion constants.","feed_headline":"Max filters plus a linear map nearly hit optimal orbit embeddings","feed_subtitle":"Closes the distortion gap for rotations, phase retrieval and reflection groups","key_machinery":"The Lipschitz-closure inclusion (1) for the optimal coordinate functions. It is proved by expressing those functions as integral combinations of max filters against Lipschitz densities (via Fourier series on the circle or spherical harmonics), then showing that Riemann-sum approximations converge in Lipschitz norm by a dominated-convergence argument on gradients.","core_discovery":"Theorem 3 asserts that, for each of the three families, and for every ε>0, there exist a max filter bank Φ and a linear map L such that the composition L∘Φ has distortion at most the Euclidean distortion of the orbit space plus ε (and exactly equal, with no ε, when G is a reflection group). The same statement is equivalent, via a continuity-of-distortion lemma, to the claim that the optimal coordinate functions lie in the Lipschitz closure of the span of max filters.","pith_inferences":["The same linear-post-processing idea is likely to work for any closed subgroup once a positively homogeneous optimal embedding is known, even if a unified analytic proof remains out of reach.","The failure of max filters alone to reach the optimal distortion is not a defect of the templates but a defect of the geometry of the span; the linear layer supplies the missing second-order corrections.","The numerical success on shape datasets suggests that LMF feature maps can serve as drop-in, trainably near-isometric layers for group-invariant machine-learning pipelines."],"forward_implications":["The gap between universal max-filter banks and known optimal distortions can be closed by a linear layer for the three classical families.","For reflection groups the optimal embedding is exactly a linear image of a d-dimensional max filter bank, so no approximation or extra dimension is required.","Any future optimal embedding that is positively homogeneous will automatically lie in the same Lipschitz closure once its coordinate functions are known.","Numerical training of linear-plus-max-filter maps recovers near-optimal empirical distortion on several additional orbit spaces beyond the three proved cases."],"fun_headline_variants":["Exact coarea reduction for synchronized singular forms","Sparse transfer of one-dimensional Dini domination","Geometric recomposition after pushforward density control","Uniform versus critical regimes via phase pullbacks","Abstract operator criterion from Lebesgue-layer densities"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Every group-invariant polynomial on the sphere can be written as an integral against max filters with a Lipschitz density; this requires that certain Fourier or Gegenbauer coefficients never vanish.","fun_headline_variants_meta":{"raw":{"variants":["Exact coarea reduction for synchronized singular forms","Sparse transfer of one-dimensional Dini domination","Geometric recomposition after pushforward density control","Uniform versus critical regimes via phase pullbacks","Abstract operator criterion from Lebesgue-layer densities"]},"model":"grok-4.5","effort":"low","cost_usd":0.003212,"raw_usage":{"total_tokens":1062,"prompt_tokens":790,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":32120000,"prompt_tokens_details":{"text_tokens":790,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":201,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":790,"tokens_out":71,"duration_ms":2782,"temperature":1.0,"reasoning_tokens":201,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T19:26:38.943327+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Exhibit a single G-invariant polynomial on the sphere whose associated Fourier or Gegenbauer coefficients all vanish; the integral-transfer step then fails and the Lipschitz-closure claim collapses for that polynomial.","supporting_citations":[],"review_version":1}