{"id":"93fd9c29-d969-4478-a2de-b8dde492ad01","arxiv_id":"2607.24650","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Density-robust S² coordinates are obtained by rejection-sampling uniform subsamples, computing classical spherical coordinates on each, and aligning them via a spherical Procrustes problem with a proven Euclidean relaxation.","lead":"The paper builds the first density-robust spherical coordinates from degree-two persistent cohomology by subsampling, aligning, and averaging maps on S². It matters because uneven sampling routinely warps topological coordinates, and spheres appear in orientation and surface data where Euclidean charts fail.","discovery_kind":"extension","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The density-robustness mechanism rests on an unproven premise — that Möbius-centered energy minimizers on rejection-sampled subsamples are essentially unbiased — and Theorem 3.4's (1+π)√L* alignment guarantee can be vacuous exactly in the severe-bias regimes where it is invoked.","rationale":"I read the construction in good faith: leaving the Schonsheck–Schonsheck optimizer untouched and adding rejection sampling plus a proved Euclidean-surrogate alignment is a legitimate, well-motivated extension of Blumberg et al. [4] to S², with a genuine scalability benefit and a public reproducibility repo. Theorem 3.4's proof structure checks out: Lemma 3.3 reduces to the circular comparison on a great circle, the lemma's r∈[0,1] parameter correctly handles the un-projected Euclidean centroid Ψ (norm ≤ 1), and the RMS triangle-inequality manipulation is valid. Minor wrinkles: the proof cites \"Theorem 3.3\" where Lemma 3.3 is meant; the spherical problem is defined over SO(3) (Def. 3.2) while the bound is an infimum over O(3), which is only a statement up to reflections (arguably appropriate given the antipodal orientation ambiguity between subsamples, but worth one clarifying sentence). These are presentational, not load-bearing. The substantive soft spot is exactly the one the reader flagged: per-subsample unbiasedness is asserted from the gauge heuristic rather than proved, and the approximation theorem's error scales with √L*, which is uncontrolled in the biased regimes. I agree with the reader's weakest_assumption and with the CONDITIONAL verdict; my stress test does not move it. The proposed test is cheap (the repo and experiment scripts exist) and would either convert the empirical robustness claim into a calibrated, quantitative one or reveal the regime where it fails — either outcome directly informs the conditionality the reader already imposed.","tokens_in":18179,"tokens_out":4631,"duration_ms":173620,"concrete_test":"Using the public repo, on the vMF κ=2 sphere experiment: for each subsample X_i, record (a) the per-subsample coordinate error (mean angular deviation from ground truth after best SO(3) fit) and (b) the residual non-uniformity of the accepted subsample (max/min ratio of the estimated density on X_i); sweep the bandwidth ε, subsample size, and add 1–5% off-sphere outlier contamination. Also compute the achieved spherical Procrustes loss L* in each severe-bias experiment and evaluate the bound (1+π)√L*. If per-subsample error rises sharply with residual non-uniformity or outliers, or if the bound exceeds ~π, the robustness claim is narrower than advertised; if error stays flat and the bound is sub-radian, the empirical claim is well supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a composition of three steps: (i) rejection sampling against a local-count density estimate yields approximately uniform subsamples; (ii) on such subsamples, the Möbius-centered harmonic-energy minimizer of §2.4 is an essentially unbiased S² coordinate; (iii) kernel extension plus spherical Procrustes alignment/averaging preserves that quality globally. Steps (i) and (iii) have formal support (Theorem 3.1 consistency; Theorem 3.4's approximation bound). Step (ii) — the step that actually generates the robustness — has no theorem. The paper's own justification (§4.3) is that on approximately uniform subsamples the zero-center-of-mass gauge is \"essentially satisfied already,\" so per-subsample coordinates are unbiased; §5 concedes the robustness is \"empirical rather than topological,\" in contrast to the Laplacian-convergence grounding available in the circular case. Nothing bounds per-subsample bias as a function of the residual non-uniformity that survives rejection sampling, and Theorem 3.1's conditions (ε_n→0, ε_n n^{1/(2m)}→∞) are asymptotic, giving no finite-sample guidance for the bandwidth actually used; the estimator is also outlier-sensitive, as the authors note. Separately, Theorem 3.4 bounds the Euclidean-to-geodesic gap by (1+π)√L*, where L* is the optimal spherical loss. In the advertised severe-bias regimes, per-subsample coordinates computed on 75–120 points and kernel-extended can disagree substantially; if √L* approaches ~0.76 rad, the bound exceeds the sphere's useful scale and becomes vacuous — precisely where the guarantee is needed. Since consensus averaging cannot remove a shared per-subsample bias, the headline robustness claim inherits the full weight of unquantified step (ii).","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript constructs the first density-robust spherical (degree-two) coordinates from persistent cohomology. The strategy ports the subsampling-and-alignment framework of Blumberg et al. [4] from S^1 to S^2: rejection sampling against a local-count density estimator produces approximately uniform subsamples, the Schonsheck–Schonsheck spherical coordinate is computed on each, kernel-extended to the full dataset, and the maps are aligned and averaged. The new mathematical content is a spherical Procrustes formulation (Def. 3.2) and an approximation theorem (Thm. 3.4, via the great-circle reduction Lem. 3.3) showing the Euclidean O(3) Procrustes solution, projected back to S^2, lies within (1+π)√L* of the geodesic optimum. Experiments show: two implementation accelerations give >100× speedup of the baseline; subsampling removes the O(n^4) memory ceiling and runs at near-constant cost to 10,000 points; under uniform sampling the method matches the baseline, and under von Mises–Fisher and other biased sampling it recovers azimuth/elevation where the baseline fails. Code and reproduction scripts are public.","tokens_in":18545,"tokens_out":4741,"duration_ms":182839,"significance":"If the empirical robustness holds up, this closes a genuine gap: density-robust topological coordinates existed only in degree one, and the authors correctly explain why the weighted-harmonic-cocycle mechanism of Paik–Park does not transfer to the sphere-valued variational problem. The paper has real strengths that should be credited: Lemma 3.3 and Theorem 3.4 are clean, correctly proved, and the great-circle reduction is a neat observation; the spherical Procrustes formulation is a natural contribution in its own right; the scalability results (removal of the quartic memory bottleneck, near-constant runtime to 10^4 points) are practically important for the degree-two pipeline; and the authors ship a public implementation with fixed seeds, five-trial runtime means, and scripts reproducing all figures. The limitations section (§5) is unusually candid about what is not proved. The main caveats concern the evidentiary basis of the headline robustness claim, which is currently qualitative, and the regime in which Theorem 3.4 is informative.","major_comments":[{"comment":"§4.3: the central empirical claim — accurate azimuth/elevation recovery under severe sampling bias — is supported only by scatter-plot figures (Figs. 5–7) at a single configuration each. No numerical error metric is reported: the protocol says coordinates are evaluated 'by correlating' angles against ground truth, but no correlation values or angular errors appear anywhere. The baseline's failure in Fig. 6 is likewise asserted visually. Since robustness is the paper's headline result, I ask for a quantitative table (e.g., circular correlation and/or mean absolute angular error for azimuth and elevation) for baseline vs. subsampling across the uniform, vMF κ=1.3, and the three severe-bias regimes, with variability over seeds and over the subsampling parameters (M, subsample size). This is feasible with the released code and would substantially strengthen the claim.","section":"§4.3, Figs. 5–7"},{"comment":"§4.3 ('Robustness to non-uniform sampling') and §3.1: the mechanism argument — that on each rejection-sampled subsample the Möbius-centering constraint is 'essentially satisfied already,' so per-subsample coordinates are unbiased — is heuristic, and nothing quantifies the residual non-uniformity of the subsamples or relates it to per-subsample coordinate bias. Theorem 3.1 is asymptotic (ε_n→0, ε_n n^{1/(2m)}→∞) and gives no guidance at the n=250–1500 used; the text never states how ε and the acceptance scale C are chosen in the experiments. At minimum, report a subsample-uniformity diagnostic (e.g., the distribution of local counts or estimated density after acceptance) alongside the coordinate errors, and state the estimator parameters used per experiment. If the causal story ('rejection sampling restores uniformity, hence coordinates are unbiased') is to be kept, it needs this quantita","section":"§3.1, §4.3"},{"comment":"Theorem 3.4 is correct as proved, but its informativeness regime should be stated. On S^2 the trivial diameter bound d ≤ π always holds, so (1+π)√L* is non-vacuous only when L* < (π/(1+π))^2 ≈ 0.58 rad², i.e., mean squared geodesic misalignment below about 0.76 rad RMS. Nothing in the paper bounds L* a priori, and in the severe-bias regimes of Fig. 7 — where per-subsample coordinates are computed on 75–120 points and kernel-extended — L* could plausibly approach this threshold. This does not undermine the theorem's purpose (justifying the Euclidean initialization for the subsequent SO(3) refinement), but the paper should (i) state the non-vacuity condition explicitly, and (ii) report the empirical L* (or the bound value) in the experiments so readers can see the guarantee is operative where invoked. Relatedly, the metric d_{(S^2)^n} and d_{(R^3)^n} on configuration space are never define","section":"§3.2, Thm. 3.4 and proof"}],"minor_comments":[{"comment":"Proof of Thm. 3.4: 'by the triangle inequality and Theorem 3.3' should read Lemma 3.3 (there is no Theorem 3.3).","section":"§3.2"},{"comment":"Def. 3.2 restricts rotations to SO(3), but the Euclidean relaxation is solved over O(3) (ten Berge), which includes improper rotations. The theorem legitimately compares against the O(3) optimum, but the initialization passed to the SO(3) refinement may be a reflection; a sentence on how this is handled in practice would help.","section":"§3.2 / §4.1"},{"comment":"Notation inconsistency in §3.1: the kernel-average formula defines Φ_i, but the following sentence refers to 'unit vectors ϕ_i(y)'. Also 'an Euclidean relaxation' (p. 14) and mixed ϵ/ε throughout.","section":"§3.1"},{"comment":"Lem. 3.3 cites 'Lemma 4.3 of [4]', an arXiv preprint whose numbering may shift between versions; give the version number or restate the one-dimensional estimate (it is a line or two) for self-containedness.","section":"§3.2"},{"comment":"Figure 4 appears in the float sequence but is never discussed in the text of §4.2 beyond the caption; either integrate it (which configurations dominate runtime?) or cut it.","section":"§4.2"},{"comment":"§4.3 opening: 'Our main goal is to prove the robustness of the construction' overstates what experiments can do; 'demonstrate' is the right verb, consistent with §5.","section":"§4.3"},{"comment":"The azimuth/elevation evaluation is ill-conditioned near the poles (azimuth is undefined there), and two of the severe-bias distributions concentrate mass at the pole(s). The requested quantitative error table should note how polar points are treated, or use a pole-invariant metric (e.g., geodesic error on S² after a best-fit rotation).","section":"§4.1 / §4.3"},{"comment":"All coordinate-quality experiments are on (noisy) S² itself. Since the motivating applications are data near manifolds with spherical topology rather than literally on S², one non-sphere example (e.g., a genus-0 surface with a prominent H² class) would broaden the evidence; at minimum the scope should be stated.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The mathematics that is new (Lem. 3.3, Thm. 3.4) is correct and the authors are transparent about what is empirical versus proved — §5 explicitly disclaims a topological justification for robustness, which is to their credit rather than against them. My major_revision is driven by the thin quantitative support for the headline robustness claim (no error numbers anywhere in §4.3) and the unremarked vacuity threshold of Thm. 3.4, not by any flaw in the construction. All requested additions are computable with the authors' released code and require no new theory. Fit for a venue in the JACT line is good, though the paper is more methodological than topological; the editor may wish to weigh that. The novelty relative to [4] is real but incremental in conception — the authors represent it accurately."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing: this is a real, usable extension of density-robust circular coordinates to spheres, not a cosmetic rewrite. They leave the nonlinear spherical optimizer alone, run it on rejection-sampled subsamples, and glue the maps with a spherical Procrustes problem whose Euclidean relaxation they actually bound (Lemma 3.3, Theorem 3.4). That alignment piece and the fixed-size-subsample memory fix are the new engineering and math; the rest is honest extension of Blumberg et al. and Schonsheck–Schonsheck.\n\nWhat they do well: background is standard and correctly cited; the great-circle reduction that turns geodesic vs Euclidean comparison into the circular case is short and correct; the approximation guarantee is proved cleanly even if the constant is crude. Implementation work (sparse triangles, numba) and the public repo make the scalability claim checkable—they really do get past the O(n⁴) wall to 10k points. Experiments systematically show baseline distortion under von Mises–Fisher bias and recovery under their method, plus a uniform control that the alignment does not wreck good data. Section 5 is unusually frank about what is still open.\n\nSoft spots, in proportion: the load-bearing step—that Möbius-centered energy minimizers on approximately uniform subsamples are essentially unbiased—has no theorem. They say so themselves; robustness is empirical, unlike the Laplacian story for weighted circles. Theorem 3.4’s (1+π)√L* bound can become loose precisely when subsample maps disagree a lot, i.e., in the severe-bias regime they advertise. Density estimator is local counts and outlier-sensitive; only synthetic spheres; several free parameters (M, subsample size, ε, σ, VR scale). None of that sinks the contribution, but it means the headline “density-robust” is demonstrated, not settled.\n\nWho it is for: people who need sphere-valued latent maps from degree-two cohomology and hit density or memory walls. Worth a serious referee. I would engage—cite the Procrustes formulation and the pipeline if I work near this, and push for real-data tests and a tighter or calibrated alignment guarantee in revision.","headline":"Solid first density-robust S² pipeline via subsample-and-Procrustes; math is clean, robustness still empirical and only synthetic.","tokens_in":19561,"tokens_out":544,"would_cite":true,"duration_ms":16137,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31","62R40","68U05"],"pacs":[],"model":"grok-4.5","headline":"The first density-robust spherical coordinates from persistent cohomology are built by computing classical sphere maps on rejection-sampled uniform subsamples and aligning them with a spherical Procrustes problem.","keywords":["persistent cohomology","spherical coordinates","density robustness","Procrustes alignment","rejection sampling","topological dimensionality reduction","Vietoris–Rips"],"falsifier":"Take a sphere sampled from a strongly biased density (for example von Mises–Fisher with large concentration), run the full subsample-align-average pipeline with the paper’s density estimator, and check whether recovered azimuth and elevation still correlate tightly with ground truth; a clear systematic warp that survives many subsamples and alignments would refute the claim.","tokens_in":19294,"feed_emoji":"🌐","tokens_out":919,"duration_ms":22385,"temperature":0.7,"pith_summary":"Persistent cohomology can turn a degree-two feature into a map from data to the sphere, giving intrinsic nonlinear coordinates when the latent space is spherical rather than Euclidean. Those maps are badly warped by uneven sampling density, and the usual density fixes for circular coordinates do not transfer because spherical coordinates come from a nonlinear energy on sphere-valued maps, not from weighted harmonic cocycles. This paper shows that density robustness can still be recovered without changing that optimization: thin the data by rejection sampling into approximately uniform subsamples, run the classical spherical construction on each, extend the maps, and average them after solving a spherical Procrustes alignment problem. A Euclidean relaxation of that alignment is proved close enough to the geodesic optimum to be usable in practice, and experiments recover accurate azimuth and elevation under severe sampling bias while matching the classical method on uniform data and scaling to ten thousand points by never building the full high-dimensional complex.","feed_headline":"Subsampling yields density-proof spherical coordinates","feed_subtitle":"Align sphere maps on uniform draws to beat sampling bias and scale past the memory wall","key_machinery":"The spherical Procrustes problem (jointly rotating several S²-valued maps and choosing a consensus map to minimize summed squared geodesic discrepancy), together with Theorem 3.4, which bounds the distance between the Euclidean O(3) Procrustes solution (projected back to the sphere) and the true spherical optimum by (1+π) times the square root of the minimal spherical loss.","core_discovery":"Density-robust spherical topological coordinates exist and can be constructed by extending the subsampling-and-alignment paradigm from the circle to the sphere: compute classical spherical coordinates on approximately uniform rejection samples, extend them by kernel-weighted nearest neighbors, and combine them by solving a spherical Procrustes problem whose Euclidean relaxation approximates the intrinsic optimum within a factor controlled by the optimal spherical loss.","pith_inferences":["Because S² is CP¹, the same subsample-and-align pattern is a natural first step toward density-robust projective coordinates from higher cohomology.","If the density estimator fails on outliers, a more stable density model (or explicit outlier rejection) would be the highest-leverage practical upgrade before chasing tighter Procrustes bounds.","The (1+π)√L* guarantee is useful mainly when the spherical loss is already small; experiments that report L* alongside angle error would make the theory–practice link sharper."],"forward_implications":["Sphere-valued latent structure can be recovered from data whose observation density is badly non-uniform without rewriting the underlying energy.","Degree-two persistent cohomology becomes practical on datasets of several thousand to ten thousand points because cohomology is run only on fixed-size subsamples.","The same alignment idea supplies a template for density-robust coordinates targeting other spaces once a Procrustes-type alignment on that space is available.","Under uniform sampling the method matches classical spherical coordinates, so it can replace the baseline without quality loss when density is unknown."],"fun_headline_variants":["Subsample alignment yields density-robust spherical coordinates","Spherical Procrustes merges uniform draws to fix sampling bias","Rejection samples plus alignment build density-robust S^2 coords","Align sphere maps on uniform subsamples to correct density bias","Density-robust spherical coordinates via subsample-and-align"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That local-count rejection sampling really produces subsamples on which the classical sphere-map optimizer is already unbiased, so that extending and averaging those maps yields a faithful global coordinate.","fun_headline_variants_meta":{"raw":{"variants":["Subsample alignment yields density-robust spherical coordinates","Spherical Procrustes merges uniform draws to fix sampling bias","Rejection samples plus alignment build density-robust S^2 coords","Align sphere maps on uniform subsamples to correct density bias","Density-robust spherical coordinates via subsample-and-align"]},"model":"grok-4.5","effort":"low","cost_usd":0.00413,"raw_usage":{"total_tokens":1258,"prompt_tokens":800,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":41304000,"prompt_tokens_details":{"text_tokens":800,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":394,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":800,"tokens_out":64,"duration_ms":7239,"temperature":1.0,"reasoning_tokens":394,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T09:03:16.181643+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Take a sphere sampled from a strongly biased density (for example von Mises–Fisher with large concentration), run the full subsample-align-average pipeline with the paper’s density estimator, and check whether recovered azimuth and elevation still correlate tightly with ground truth; a clear systematic warp that survives many subsamples and alignments would refute the claim.","supporting_citations":[],"review_version":1}