{"id":"0c243343-6285-4f54-9839-e4e8188f62bd","arxiv_id":"2607.27126","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Voronoi-cell histograms of normalized Expected Persistence Diagrams give a stable, adaptive EPD vectorization that is competitive on topology-sensitive classification and scales better with the number of subsampled diagrams than smooth baselines.","lead":"The paper introduces Voronoi histograms (Vrep) to turn Expected Persistence Diagrams into vectors by counting topological mass in data-dependent cells instead of smoothing each feature with fixed kernels. It offers a practical, scalable alternative for topology-aware point-cloud classification when coarse mass placement matters more than fine local shape.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"Thm 4.5's Lipschitz bound is proved only for population EPDs with a Lipschitz density, but every experiment histograms atomic empirical EPDs subsampled to 50 atoms — a regime where no such deterministic bound holds, since one atom crossing a Voronoi edge shifts Φ by a fixed amount under arbitrarily小","rationale":"The reader identified normalization (mass discard) and the well-resolved-codebook premise as weakest assumptions. Those are real but are explicitly disclosed in the paper (App. D.4 defends normalization; Sec. 4.3 states the bound is \"non-informative\" otherwise), so they function as scoped limitations rather than hidden load-bearing flaws. I partially agree with the reader but locate the least-secure condition elsewhere: the stability theorem is proved for a smoothed population object while the method and all its evidence (including the stability demo itself) live in the atomic finite-n regime, and the proof of Lemma 4.3 has concrete defects (dropped Vol(Ω′) factor; independence assumed for the convolution while the text asserts dependence). This matters because \"Lipschitz-stable (Thm. 4.5)\" is half of the strongest claim and currently does not apply to the object whose empirical performance is reported. It is not disqualifying: an expected-Lipschitz bound for atomic EPDs under Gaussian perturbation is standard and provable, and the paper is otherwise unusually candid — limitations are enumerated, non-dominance over PI/PS/PWGK is stated, and the synthetic Wasserstein-correlation study (F.7) plus the codebook-construction ablation (F.8, including honest reporting that quantization centers match sampled supports) are genuine independent support. Hence the verdict stays CONDITIONAL, with an added explicit condition: either bridge Thm 4.5 to the empirical atomic pipeline (finite-n or expected version, fixing the Lemma 4.3 constants and the independence step) or clearly re-scope the stability claim to the population limit. The reader's own conditions (split variability, public artifacts) remain appropriate.","tokens_in":28648,"tokens_out":7770,"duration_ms":173282,"concrete_test":"On the dynamical-system dataset used for Fig. 5, run the deployed pipeline (n=50 sampled PDs, support subsampled to 50 atoms, the actual codebook set S) and for each EPD compute (a) the fraction of atom mass within distance ∆ of any Voronoi boundary at the noise levels of Fig. 5, and (b) max over single-atom perturbations of size ∆ of ‖Φ̂(µ̄)−Φ̂(µ̄′)‖₁/∆, with and without the 50-atom subsampling. If (b) grows like 1/(min atom weight) or tracks boundary-proximity mass rather than ∆, Thm 4.5 must be restated as a population result and a finite-n (expected-Lipschitz) version proved; if (b) stays bounded and matches the subsampled/unsampled cases, the gap is benign.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 4.3, on which Thm 4.5 rests, assumes µ̄ admits a Lipschitz density p (citing Chazal–Divol [16]) and bounds cell-mass change by M·W1·Vol(V_j), concluding with \"Σ_j M·W1·Vol(V(c_j)) = M·W1\" — which silently requires Vol(Ω′)=1, when in fact Σ_j Vol(V(c_j)) = Vol(Ω′), an arbitrary constant. The convolution step is also internally inconsistent: the text states \"α and ϵ is dependent of each other\" yet writes r = p*q, which is valid only under independence. The deeper problem, though, is the object the theorem covers versus the object deployed. Definition 4.2 and Lemma 4.3 concern µ̄ = lim_{n→∞} (1/n)Σµ_i, the population EPD. The actual representation (Def. 4.1, App. E.1) is computed from the empirical EPD with n=50 sampled PDs, with support further subsampled to |S_µ̄|=50 atoms. For an atomic measure the cell-mass map µ ↦ µ(V(c)) is not Lipschitz in W1: an atom of weight ~1/50 sitting within ε of a Voronoi boundary crosses it under a perturbation of size ε, changing ‖Φ‖₁ by ~2/50 ≈ 0.04 no matter how small ε is. The finite-sample Lipschitz ratio therefore scales with 1/(min atom weight) unless one averages over perturbation realizations. Fig. 5 — the paper's own stability demonstration — is itself run in this uncovered atomic regime and shows only average behavior on one dataset. So the stability half of the central claim (\"Φ is Lipschitz-stable to EPD perturbations\") currently has no theorem connecting it to the pipeline whose accuracies appear in Table 1. This is distinct from the reader's flagged normalization concern: normalization is a disclosed modeling choice (App. D.4/D.5), whereas the population-vs-atomic gap is undisclosed. The gap is likely repairable — under the Gaussian perturbation model of Def. 4.2, expected mass within ∆ of a boundary is O(M_p·L_boundary·∆), giving an expected-Lipschitz finite-n statement — but as written, Thm 4.5 does not cover the method.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper proposes Vrep/Vrepd, a vectorization of empirical Expected Persistence Diagrams (EPDs). An empirical EPD (the average measure of n sampled PDs, ℓ1-normalized to unit mass) is represented by concatenating, over t codebooks of size k sampled from EPD supports, the histogram of EPD mass in each Voronoi cell; Vrepd adds a cell centered at the diagonal. The contributions are: (i) the representation itself; (ii) a stability analysis (Lemmas 4.3–4.4, Thm. 4.5: ‖Φ̂(μ̄)−Φ̂(μ̄′)‖₁ ≤ LΔ) and a conditional Wasserstein lower bound (Thm. 4.6) relating histogram distance to W1 separation minus codebook approximation error; (iii) experiments: Random-Forest classification on Protein, CAD (with two noise levels), and five time-series-derived point-cloud datasets, where Vrep/Vrepd report higher mean accuracy than PI/PS/PL on all datasets; ablations over codebook sampling schemes; scale-up tests showing near-constant cost in n; comparisons to PWGK/SWK (kernels win on several datasets at much higher runtime); an unsupervised t-SNE/CH study; and a synthetic W1-correlation study. The paper is unusually candid about non-dominance and limitations, and the normalization premise is explicitly disclosed.","tokens_in":29204,"tokens_out":8396,"duration_ms":166578,"significance":"If the claims hold, this is a useful, simple, and computationally attractive EPD vectorization: near-constant cost in the number of sampled PDs (Fig. 6b) is a genuine practical advantage over PI/PS/PL, and the data-dependent Voronoi partition is a sensible alternative bias to fixed smooth kernels. The empirical package is stronger than typical: codebook-choice ablations (Table 2), scale-up tests, honest non-dominance results against PWGK/SWK (App. F.2), a controlled synthetic W1-correlation study (App. F.7) that explicitly shows where PI is preferable, codebook-construction alternatives (App. F.8), and promised code. The candid limitation section (D.5) and the conditional reading of Thm. 4.6 are commendable. The weak half is the theory: as written, the stability theorem does not cover the atomic object used in every experiment, and Lemma 4.3's proof has concrete gaps. Fixing or rescoping this would make the paper a solid contribution.","major_comments":[{"comment":"Population/empirical mismatch in the stability results. Def. 4.2, Lemma 4.3 and Thm. 4.5 quantify over the population EPD μ̄ = lim_{n→∞}(1/n)Σμ_i, and Lemma 4.3's proof requires μ̄ to admit a Lipschitz density p (via [16]) perturbed by convolution with Gaussian noise. The deployed pipeline (§4.1, App. E.1) instead histograms an atomic empirical EPD (n=50 sampled PDs, support subsampled to |S_μ̄|=50). For an atomic measure, μ ↦ μ(V(c)) is not Lipschitz in W1: an atom of mass ~1/50 sitting ε from a Voronoi edge crosses it under an ε-perturbation, shifting ‖Φ‖₁ by ~2/50 regardless of how small ε is. Thus no deterministic Lipschitz-in-W1 bound covers the regime of Table 1 and Fig. 5, and Fig. 5 — the only stability evidence — is itself run in that uncovered regime and shows averages on one dataset only. Additionally, Lemma 4.3 is stated for arbitrary μ̄′ (via an optimal matching η), but the ","section":"§4.2, Def. 4.2, Lemma 4.3, Thm. 4.5"},{"comment":"Several specific gaps need repair: (a) the final equality Σ_j M·W1·Vol(V(c_j)) = M·W1 implicitly uses Σ_j Vol(V(c_j)) = 1, but the cells partition Ω′, so the sum is Vol(Ω′); the constant should be M·Vol(Ω′) (directionally harmless, but the stated constant is wrong). (b) The text asserts 'α and ϵ is dependent of each other' and then writes r = p∗q, which is valid only under independence — either assume independence or replace the convolution step. (c) 'W1(μ̄,μ̄′) = ≤ ∫q(z)‖z‖dz': only ≤ holds; W1 is an infimum over couplings. (d) Def. 4.2 uses Δ both as a deterministic radius bound (‖r̂_j − r_j‖ < Δ) and as a Gaussian random variable Δ ~ N(0,Σ); Gaussian noise is unbounded, so the hard bound fails almost surely. Each item is individually fixable, but together they require a careful rewrite of the perturbation model and the proof.","section":"App. C.1 (proof of Lemma 4.3)"},{"comment":"All theory and the main experiments ℓ1-normalize EPDs to unit mass, discarding total feature count / total persistence mass and making Wasserstein (rather than OTp) applicable. The authors disclose this in App. D.4/D.5, which is appropriate, but no experiment quantifies the cost. Since the motivation (§2) explicitly invokes OTp for unequal-mass EPDs, a cheap ablation would substantially clarify when the premise is safe: e.g., augment Φ̂ with log μ̄(Ω) as an extra feature, or an unnormalized variant, on a dataset where class plausibly correlates with feature count. Without it, the 'Wasserstein-scale variation' framing applies to a geometry whose match to the tasks is asserted, not checked.","section":"§4.1 normalization; App. D.4/D.5"},{"comment":"Accuracies are means over 10 random splits on small datasets (BirdChicken n=40, Beef n=60, Protein n=99), with no standard deviations, confidence intervals, or paired tests reported, yet the text claims Vrep/Vrepd 'outperform PI, PS and PL over all the datasets.' Several margins look within plausible split noise (e.g., CAD0.01: 0.912 vs PI 0.911; CAD0.01/CAD0.05 vs PI at 0.900–0.911; Earthquakes vs PL). Please report per-split standard deviations and a paired significance test, mark which differences are significant, and soften the blanket claim where they are not. This is load-bearing for the main empirical claim, not a presentational nicety.","section":"§5, Tables 1–2"}],"minor_comments":[{"comment":"The proof invokes 'Theorem 5.1 and Remark 5.3 in [4]', but [4] is Beer, 'The Hausdorff metric and convergence in measure'; footnote 7's arXiv link (1103.4125) shows Reem [49] is intended. Also 'd_H(V(c_i), V(c_j)) ≤ Δ/γ' should be V(c′_i), and 'bounder area' → 'bounded area'.","section":"App. C.2 (proof of Lemma 4.4)"},{"comment":"d_min(C) is defined with ‖·‖₁ in App. C.5 but with ‖·‖₂ in the App. D.3 restatement of Thm. D.1 — align. In the C.5 proof, '‖Φ(ν̄,C)−Φ(ν̄,C)‖' appears three times where (μ̄,ν̄) is meant.","section":"App. C.5 vs App. D.3 (Thm. D.1)"},{"comment":"The transported mass in the designed plan is exactly ½‖Φ(μ̄,C)−Φ(ν̄,C)‖₁ (the ℓ1 norm double-counts excess and deficit), so the stated bound holds with factor-2 slack — worth a remark, or tighten. Also the phrase 'optimal partial transport metric W1' in the proof is a misnomer since masses are equal here.","section":"App. C.4 (proof of Thm. 4.6)"},{"comment":"L = mt(M_max + kC_0^max) grows linearly with dataset size m; App. D.2 shows mt cancels in the normalized δ, but the raw bound is vacuous at realistic m. A sentence on the practical meaning (and in-principle estimability) of the constants M and C_0, both of which depend on the unknown EPD density, would help readers gauge the bound's content.","section":"Thm. 4.5 constant L"},{"comment":"Main text says C_0 is 'determined by μ̄ and C'; the App. C.2 restatement says 'determined by μ̄' — align.","section":"Lemma 4.4 statement"},{"comment":"Bracketed citation numbers render without brackets throughout ('As pointed by 69', 'motivated by Lemma 2 in 25'). In App. D.3, 'Theorem 0.5' does not exist — presumably Thm. 4.6 is meant. 'V oronoi' appears with a spurious space throughout. Fig. 5 would benefit from error bars across the 450 EPDs.","section":"Citation/formatting"},{"comment":"The example assumes codebook size k very large to get ‖Φ(μ,C)−Φ(ν,C)‖₁ = 2W(μ,ν); please comment on how the conclusion degrades at the k values used in practice (k ≤ 20, App. E.1).","section":"App. D.1 (PI/PS comparison example)"},{"comment":"Since Persistence Bag-of-Words [68] is the closest histogram-style prior art, consider discussing it in §3 rather than only footnote 3, with one sentence on hard Voronoi vs GMM soft assignment. Also state near Def. 4.1 what §5 only says later: for a test EPD the concatenation uses codebooks sampled from training EPDs only.","section":"§3 / Def. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript already carries unusually candid caveats (§4.3 footnote 6, App. D.5 limitation section, App. F.2 non-dominance results, App. F.7 synthetic study showing where PI is better), which suggests the authors have revised in good faith and will be responsive. The remaining issues are technical and concentrated in §4.2/App. C; they look addressable in one revision round. Novelty over ATOL [52] and EPD quantization [25] is incremental but adequately disclosed (App. B, F.8). An LLM-use disclosure is included (App. G). No concerns about fit or citation conduct."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a usable Voronoi-histogram vectorization for empirical EPDs, with clearer scaling in the number of subsampled diagrams than PI/PS/PL, and the authors mostly do not oversell it. The stability half of the central claim is written for the wrong object.\n\nWhat is new is straightforward and real. Instead of averaging smooth transforms of sampled PDs, they treat the empirical EPD as a measure and bin normalized mass in data-dependent Voronoi cells (Vrep/Vrepd), with optional diagonal cell. Relative to Chazal-style EPD averaging and to ATOL/Bag-of-Words/quantization, that is a clean EPD-specific hard-assignment choice. Empirically they do the right checks: RF on topology-heavy sets, codebook ablations, scale-up in n, synthetic W1 correlation, and explicit notes that PWGK/SWK sometimes win. The coarse-mass vs within-cell trade-off is the right story, and Appendix limitations are unusually straight.\n\nSoft spots, in proportion. Normalization to unit mass is a disclosed modeling choice; fine if you care about spatial allocation, not if total mass matters. The free parameters (k, t, subsample size, sampling scheme) are ordinary. The real issue is Theorem 4.5 / Lemma 4.3. The proof assumes a Lipschitz density on a population EPD, has a volume-sum slip (cells sum to Vol(Ω′), not 1), and muddles independence in the convolution step. Experiments histogram atomic empirical EPDs with support cut to 50 atoms. For atomic measures the cell-mass map is not W1-Lipschitz: an atom near a boundary flips a fixed mass under arbitrarily small moves. Figure 5 is average behavior in that uncovered regime. So “Φ is Lipschitz-stable” does not yet connect to Table 1. Likely fixable with an expected-Lipschitz argument under their Gaussian noise model, but as written the theory does not cover the pipeline.\n\nMath elsewhere is standard triangle/transport bookkeeping; citations look appropriate; no circular fitting to labels. This is for people who already vectorize PDs/EPDs and want a fast, adaptive histogram baseline—not a filtration-theory paper.\n\nI would send it to referees. Ask them to repair or correctly scope the stability theorem to finite atomic EPDs, report split variability, and keep the non-dominance caveats. Worth engaging after that fix; not a desk reject.","headline":"Solid EPD tooling with an honest coarse-vs-smooth trade-off, but the Lipschitz stability theorem does not actually cover the atomic histograms used in the experiments.","tokens_in":29335,"tokens_out":619,"would_cite":false,"duration_ms":24114,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31","68T10"],"pacs":[],"model":"grok-4.5","headline":"Expected Persistence Diagrams can be vectorized by counting mass in data-dependent Voronoi cells instead of smoothing each feature with a fixed kernel.","keywords":["Expected Persistence Diagram","Voronoi histogram","topological data analysis","persistence diagram vectorization","Wasserstein stability","point cloud classification","EPD representation"],"falsifier":"Build synthetic EPD pairs whose Wasserstein gap is pure coarse mass shift across well-separated regions with small codebook approximation error: if Vrep’s ℓ₁ distances fail to track W₁ better than PI/PS/PL there—or if tasks that depend on total feature mass lose accuracy under the paper’s unit-mass normalization—the central trade-off claim does not hold.","tokens_in":28943,"feed_emoji":"⬡","tokens_out":895,"duration_ms":38954,"temperature":0.7,"pith_summary":"Persistence diagrams capture the topology of a point cloud, but computing them is expensive. Expected Persistence Diagrams average many diagrams from random subsets, turning topology into a distribution of birth–death features. Most existing vectorizations still smooth each feature with a preset function such as a Gaussian or a landscape and then discretize. This paper proposes Vrep: sample codebooks from the empirical EPD, build Voronoi cells, and record the normalized mass in each cell, optionally with a cell for the diagonal to catch near-diagonal noise. Under normalization and separation conditions the histograms are Lipschitz-stable to small EPD moves and can lower-bound Wasserstein separation when the codebook approximates the measures well. On topology-sensitive classification and dimensionality-reduction tasks the method is competitive with persistence images, silhouettes, and landscapes, and its cost barely grows with the number of sampled diagrams per EPD.","feed_headline":"Voronoi cells vectorize topology without smooth kernels","feed_subtitle":"Adaptive mass counts stay stable and nearly free as you add more sampled diagrams.","key_machinery":"Vrep (Voronoi-based Representation): Φ(μ̄, C) = [μ̄(V(c₁)), …, μ̄(V(c_k))], the histogram of normalized EPD mass over Voronoi cells of a codebook C, concatenated over many sampled codebooks (Vrepd adds a diagonal cell). It carries the argument by replacing smooth functional summaries with adaptive partition-based mass counts, enabling the stability and conditional Wasserstein bounds.","core_discovery":"A Voronoi histogram of a normalized empirical Expected Persistence Diagram—mass counts inside cells of randomly sampled codebooks, concatenated across codebooks—is a stable finite vectorization that needs no explicit smooth point-transformation model, and under stated separation conditions it can preserve Wasserstein-scale differences at the cell level while matching or beating common smooth EPD summaries on topology-rich data.","pith_inferences":["Hard cell assignment will under-resolve tasks driven by fine within-cell shape; hybrid pipelines that keep a smooth local descriptor inside large cells are a natural next test.","Re-attaching total mass or unnormalized counts as extra coordinates would check how much the unit-mass step is costing on density-sensitive problems.","Learned or quantized codebooks under the same hard-histogram readout could tighten the Wasserstein lower bound when random support samples miss modes."],"forward_implications":["When the useful signal is coarse mass allocation on the birth–death plane, adaptive Voronoi counts can replace fixed-kernel EPD vectorizations.","With support subsampling, EPD vectorization cost need not grow with the number of sampled persistence diagrams.","A diagonal Voronoi cell can separate near-diagonal noise from persistent features more stably than interior-only codebooks.","The same unsupervised vectors can feed dimensionality reduction where labeled end-to-end point-cloud models cannot."],"fun_headline_variants":["Voronoi histograms vectorize EPDs without smooth kernels","Adaptive Voronoi cells count EPD mass, skip smooth transforms","Stable Voronoi binning preserves Wasserstein-scale EPD differences","Partition EPD mass with Voronoi codebooks, no Gaussian model","Voronoi histograms match smooth EPD summaries on topology tasks"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Every EPD is normalized to unit total mass before analysis and representation, so any signal in the absolute number or total mass of topological features is thrown away.","fun_headline_variants_meta":{"raw":{"variants":["Voronoi histograms vectorize EPDs without smooth kernels","Adaptive Voronoi cells count EPD mass, skip smooth transforms","Stable Voronoi binning preserves Wasserstein-scale EPD differences","Partition EPD mass with Voronoi codebooks, no Gaussian model","Voronoi histograms match smooth EPD summaries on topology tasks"]},"model":"grok-4.5","effort":"low","cost_usd":0.003177,"raw_usage":{"total_tokens":1062,"prompt_tokens":690,"num_sources_used":0,"completion_tokens":69,"cost_in_usd_ticks":31768000,"prompt_tokens_details":{"text_tokens":690,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":303,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":690,"tokens_out":69,"duration_ms":5045,"temperature":1.0,"reasoning_tokens":303,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T11:43:54.164748+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Build synthetic EPD pairs whose Wasserstein gap is pure coarse mass shift across well-separated regions with small codebook approximation error: if Vrep’s ℓ₁ distances fail to track W₁ better than PI/PS/PL there—or if tasks that depend on total feature mass lose accuracy under the paper’s unit-mass normalization—the central trade-off claim does not hold.","supporting_citations":[],"review_version":1}