{"id":"b951fa85-03df-421c-b96b-2de3584e0248","arxiv_id":"2603.15384","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A signed-diagonal-augmented persistence-sphere map S(μ) is injective, POT1-Lipschitz, and has a continuous inverse on compactly supported measures, yielding a Hilbert-space bi-continuous embedding on controlled-measure classes.","lead":"Persistence spheres map persistence diagrams to functions on a sphere, and this paper refines them so the map is stable and invertible-continuous under partial optimal transport. It gives the first explicit vectorization of this kind with a continuous inverse on compactly supported targets, plus competitive benchmarks across several data modalities.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inverse continuity depends on an unverified shallow-ReLU embedding (Mao et al. 2024) used as a black box in Theorem 3; if that embedding fails, Theorem 4 and Corollary 7 collapse.","rationale":"The reader's weakest_assumption identifies exactly the external shallow-ReLU approximation theorem as the least secure point, and my analysis agrees. The proof of Theorem 3 is otherwise internally consistent: the mollification estimates, the scaling to the unit ball, the use of Kantorovich–Rubinstein duality, and the optimization over r all check out. The stability and injectivity parts are solid, and the compact-support limitation of inverse continuity is stated honestly. The load-bearing vulnerability is genuinely the black-box embedding (29). Without it, the inverse-continuity chain from Theorem 3 to Theorem 4 to Corollary 7 loses its quantitative foundation. I do not find a separate fatal flaw; the issue is verifiability and sensitivity of a key external input. Hence the reader's CONDITIONAL verdict is appropriate, and I recommend no change. A single concrete check—confirming the exact hypotheses of Mao et al. (2024) and whether my W^{5/2,2} embedding is truly valid on B1—would settle the concern. If the check fails, the verdict should be REJECT or at least UNVERDICTED for the inverse-continuity theorems; if it passes, the main mathematical claims stand.","tokens_in":46321,"tokens_out":20309,"duration_ms":173056,"concrete_test":"Consult the proof of Mao et al. (2024, Thm. 1) and verify that it provides an exact integral representation with bounded variation on the unit ball B1 for every g∈W^{5/2,2}(B1), not merely an approximation. If the statement assumes extra conditions (e.g., whole-space domain, different bias range, or a higher Sobolev exponent), re-derive the corresponding bound for the mollified f_r and recompute the exponent in Theorem 3; if the rate worsens, check whether Theorem 4 still follows.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central bi-continuity claim rests on the local Hölder bound in Theorem 3, whose Step 2 imports the external estimate ∥g∥_{K1(P1)} ≤ A0 ∥g∥_{W^{5/2,2}(B1)} (Mao et al. 2024, Thm. 1). This bound is used to represent the mollified test function f_r exactly as an integral of ReLU ridge functions with total variation controlled by a Sobolev norm. If this embedding is false, or if it holds only with a different Sobolev exponent, then inequality (34) fails and the optimization in Step 5 yields no quantitative control. Since Theorem 4 (inverse continuity) and Corollary 7 (L2 bi-continuity) both funnel through Theorem 3, the entire inverse direction of the central claim is conditional on an external, non-peer-reviewed preprint whose constant A0 is not supplied or numerically verified. The paper does not provide an internal proof or a fallback, despite the theorem being load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper refines persistence spheres, mapping an integrable measure μ on the upper half-plane to S(μ) = Λ(μ − (π_Δ)_#μ)|_{S^2}, where Λ is the signed ReLU (lift-zonoid) transform. The main theoretical results are: injectivity (Proposition 7), uniform POT_1-stability with explicit constant 2√2 (Theorem 2), inverse continuity under uniform sphere convergence to compactly supported targets (Theorem 4), a local Hölder-type inverse bound on compact sets with exponent 2/5 in the sphere discrepancy (Theorem 3), and Hilbert-space upgrades giving a bi-continuous embedding of the class M_c(A,B) into L^2(S^2) (Theorem 6, Corollary 7). The paper also contains qualitative comparisons of how various persistence summaries deform the POT_1 geometry, and an experimental section on unsupervised and supervised benchmarks.","tokens_in":46539,"tokens_out":15512,"duration_ms":150921,"significance":"Conditional on the external approximation inequality used in Theorem 3, this is a valuable contribution: an explicit, parameter-free representation with POT_1-Lipschitz stability and inverse continuity on compactly supported targets. The convex-geometric viewpoint and the signed diagonal augmentation are elegant, and the paper goes beyond bounded-cardinality Hilbert embeddings by allowing diverging mass with quantified tail-to-diagonal control. Most of the proof is self-contained and carefully assembled; the stability, injectivity, tail-vanishing, and Hilbert-upgrade arguments are detailed and largely checkable. However, the inverse direction rests on a black-box theorem from a non-peer-reviewed preprint, so the central claim is not yet fully grounded as written.","major_comments":[{"comment":"Theorem 3 imports the embedding ∥g∥_{K_1(P_1)} ≤ A_0 ∥g∥_{W^{5/2,2}(B_1)} from Mao et al. (2024), a preprint. This inequality is the only mechanism producing the quantitative ε^{2/5} inverse estimate: it enters Eq. (32), then Eq. (34), and then the final bound. Theorem 4 and Corollary 7 funnel through this estimate. The paper neither proves the inequality nor supplies the constant A_0, and it is not numerically verified. If the inequality is false, or holds only with a different Sobolev exponent, the inverse direction collapses. Please make this step self-contained (e.g., a proof in an appendix or a fully accepted reference), or at the very least provide a numerical verification of the inequality on a dense set of test functions and state the resulting A_0. This is load-bearing for the paper's main claim.","section":"§6.2.3, Step 2 (Eq. (29))"}],"minor_comments":[{"comment":"The change-of-variables identity for the dilation e_h(x)=h(R_K x) is misstated: the correct scaling is ∥D^m e_h∥_{L^2(B_1)} = R_K^{m-1} ∥D^m h∥_{L^2(B_{R_K})}, not R_K^{m-2}. The qualitative argument is unaffected, but the claimed explicit constants in Eqs. (30)–(31) and in Remark 3 are wrong as written.","section":"§6.2.3, Step 2"},{"comment":"The symbol K is overloaded: first K⊂X is the original compact off-diagonal support, then K := K ∪ π_Δ(K). Step 4 uses Pers(p) ≥ δ_K on K, which is false on the diagonal part π_Δ(K). The proof works if the support assumption refers to the original K; please rewrite with distinct notation, e.g., K_0 and K_0 ∪ π_Δ(K_0).","section":"§6.2.3, Theorem 3"},{"comment":"There is a typo: 'A positive Borel measure μ on E' should be 'on Z'. Also, M is defined by 'If Z = X', but X is not introduced until Section 3.1; define X before introducing M.","section":"§3, Definition 4"},{"comment":"The experimental section reuses all non-PSph baselines from Pegoraro (2026) without rerunning them, and the code is not released. This is acceptable if the protocols are identical, but the paper should state more prominently which numbers are copied, and ideally release code/scripts so the empirical comparisons can be reproduced.","section":"§10–§11 and 'Code' paragraph"},{"comment":"The inverse-continuity theorems are proved only for compactly supported targets, and the unbounded-support regime remains open. This is stated accurately in the abstract, but it should also be listed explicitly as a limitation in the discussion, since practitioners may use the representation for measures with non-compact support.","section":"§12"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a substantial extension of the author's own ICLR paper, and the empirical comparisons partly reuse that paper's baselines. The signed diagonal augmentation is a genuine conceptual improvement. My main concern is the unverified black-box inequality in Theorem 3; if the author can supply a proof or a reliable reference for the needed ReLU-variation-norm embedding, I would be willing to accept. The editor may wish to check novelty disclosure regarding overlap with Pegoraro (2026)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this carefully. The core idea—replacing persistence-dependent reweighting by signed diagonal augmentation—is a genuine improvement, and the paper makes good on the main promise: it gives the first explicit summary map for persistence diagrams with a continuous inverse at every compactly supported target. The stability bound \\|S(μ)-S(ν)\\|_∞ ≤ 2√2 POT_1(μ,ν) is clean, and the injectivity argument via cross-augmentation is elegant. The measure-theoretic formulation and the L² bi-continuity on M_c(A,B) are real extensions of the author's earlier persistence-spheres work, not just tweaks. The local Hölder bound (Theorem 3) is intricate but appears correct, with the caveat below.\n\nThe main soft spot is the dependence of the inverse direction on an external shallow-ReLU approximation result from Mao et al. (2024). Step 2 of Theorem 3 imports the bound \\|g\\|_{K_1(P_1)} ≤ A_0 \\|g\\|_{W^{5/2,2}(B_1)} as a black box; A_0 is never computed or numerically verified. If that theorem is wrong, or holds with a different Sobolev exponent, inequality (34) fails and Theorems 4 and Corollary 7 collapse. I don't think it is wrong—Mao, Siegel, and Xu's result is plausible and known in approximation theory—but the paper should have stated it as an explicit assumption or supplied a proof sketch. As it stands, the main bi-continuity theorem is conditional on an unverified external result. That is a legitimate concern for a referee and should be addressed in revision.\n\nThe empirical section is weaker than the theory. The paper reruns only the PSph and PI pipelines; all other baselines (PL, PSpl, PersLay, SWK) are inherited from the author's ICLR 2026 paper, which is not a consistent comparison. No code or data is released at submission. These issues do not affect the theorems but do weaken the claimed practical superiority. Also, the inverse result is limited to compactly supported targets, which the paper states honestly as a limitation, but it means the result does not cover unbounded-support integrable measures.\n\nThere are a few minor intermediate slips and the paper is quite long (67 pages), but I didn't find any fitting or circular reasoning: the definition is parameter-free, and the main theorems are derived directly. The citation pattern is reasonable; the self-citation to the earlier persistence-spheres paper is justified because this is a direct extension.\n\nBottom line: this is a serious paper with a genuine new result. It deserves a careful referee. The black-box dependency and inherited baselines are fixable, and the theory is strong enough to carry acceptance after revision.","headline":"The signed-augmentation construction is sound and delivers the first explicit vectorization of persistence diagrams with a continuous inverse on compactly supported targets; worth a serious referee, but the inverse bound rests on one imported ReLU-approximation theorem and the empirical comparisons inherit baselines from the author's prior paper.","tokens_in":47011,"tokens_out":6416,"would_cite":true,"duration_ms":53527,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31","62R40","49Q22"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that persistence spheres—functions on the unit sphere built from a signed ReLU integral of a measure minus its diagonal projection—are injective, POT_1-stable, and continuously invertible at every compactly supported t","keywords":["persistence diagrams","partial optimal transport","lift zonoid","topological vectorization","bi-continuous embedding","ReLU approximation","persistence spheres","measure representations"],"falsifier":"Take a compactly supported persistence measure μ and search, for example by gradient descent on point positions and weights, for a sequence μ_n with ∥S(μ_n)−S(μ)∥∞→0 but POT_1(μ_n,μ) bounded away from 0; a single such sequence, or two distinct measures with identical sphere values, would refute Theorem 4 or injectivity directly.","tokens_in":46190,"feed_emoji":"🌐","tokens_out":4999,"duration_ms":47382,"temperature":0.7,"pith_summary":"This paper tries to show that persistence diagrams, and more generally integrable measures on the upper half-plane, can be mapped to functions on the unit sphere in a way that loses no information and preserves the natural partial-transport geometry. The map, called the persistence sphere, is built by subtracting the diagonal projection of a measure and then taking a signed ReLU (lift-zonoid) transform, which encodes the \"delete unmatched mass to the diagonal\" mechanism of partial optimal transport directly into the representation. The central results are a global Lipschitz stability bound (uniform sphere distance is at most 2√2 times POT_1 distance) and a converse: if sphere functions converge uniformly and the target is compactly supported, the underlying measures converge in POT_1. On a growth-controlled class, the same bi-continuity holds in the Hilbert space L^2(S^2), so the representation can be used for regression, clustering, and classification while retaining a provable connection to diagram geometry. A sympathetic reader would care because this is an explicit, parameter-free vectorization with a two-way metric guarantee, something not previously available for this kind of topological summary.","feed_headline":"Persistence spheres gain a continuous inverse at every compact target","feed_subtitle":"Closeness of the sphere functions now implies closeness in partial optimal transport—and the encoding needs no tuning parameters.","key_machinery":"The signed lift-zonoid transform Λ(σ)(v) = ∫ ReLU(⟨v,(1,p)⟩) dσ(p), applied to the augmented signed measure μ_aug = μ − (π_Δ)_#μ. The diagonal augmentation is the load-bearing trick: it builds POT_1's deletion-to-diagonal cost into the representation, and it makes differences of spheres equal to Λ(μ ⊕_Δ ν) − Λ(ν ⊕_Δ μ), so stability follows from Kantorovich–Rubinstein duality and inverse continuity from approximating Lipschitz test functions by ReLU ridges.","core_discovery":"The paper's central claim is that the persistence-sphere map S(μ) = Λ(μ − (π_Δ)_#μ)|_{S^2}, with Λ the signed ReLU (lift-zonoid) transform, gives a linear, parameter-free encoding of measures in which closeness of the sphere functions is equivalent to closeness in POT_1, at least when the target is compactly supported: ∥S(μ)−S(ν)∥∞ ≤ 2√2 POT_1(μ,ν) for all integrable μ,ν, and ∥S(μ_n)−S(μ)∥∞→0 implies POT_1(μ_n,μ)→0 whenever μ is compactly supported. It further proves that on a growth-controlled class the map into L^2(S^2) is bi-continuous onto its image. The mechanism: subtracting the diagonal projection encodes deletions-to-diagonal inside the representation, so S(μ)−S(ν) is an integral aga","pith_inferences":["The diagonal-augmentation trick is transportable: any Lipschitz integral summary whose features approximate Lipschitz functions on compacta should inherit a POT_1-bi-continuity theorem by the same route, suggesting testable upgrades for persistence images and related summaries.","The inverse theorem's compact-support assumption likely has a quantitative form: the proofs suggest a local Hölder estimate with exponent 2/5 that might extend to unbounded targets under the tail conditions of M(A,B), without requiring a bounded-cardinality cap.","If the external ReLU approximation bound can be replaced by a constructive or elementary estimate, the Hölder constant in the local inverse bound would become fully explicit and computable, enabling practical inversion algorithms from sphere values back to diagrams."],"forward_implications":["Any two persistence measures with the same persistence sphere are the same measure (injectivity), so the representation loses no information at the level of measures.","If S(μ_n) converges uniformly to S(μ) for compactly supported μ, then μ_n converges to μ in POT_1: sphere-space convergence can be read back as diagram convergence.","On the growth-controlled class M_c(A,B), the same bi-continuity holds in the Hilbert space L^2(S^2), making the representation compatible with inner-product and gradient-based pipelines while controlling POT_1 error.","The distance from a measure to the empty diagram is exactly √2 times its total persistence, so POT_1-stable comparisons need no persistence-based reweighting or tuning parameters.","Because the map is linear in the augmented measure, smoothed representations such as persistence intensity functions fit in the same framework."],"fun_headline_variants":["Persistence spheres: continuous inverse on compact support","Parameter-free spheres match partial optimal transport","Bi-continuous spheres for lossless measure encoding","Persistence spheres: zero tuning, full invertibility","First explicit bi-continuous map for persistence diagrams"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a literature result bounding ReLU-network approximation in a Sobolev norm holds with a finite constant; on top of that, the inverse result is only proved for compactly supported target measures, so the unbounded-support regime is not covered.","fun_headline_variants_meta":{"raw":{"variants":["Persistence spheres: continuous inverse on compact support","Parameter-free spheres match partial optimal transport","Bi-continuous spheres for lossless measure encoding","Persistence spheres: zero tuning, full invertibility","First explicit bi-continuous map for persistence diagrams"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000294,"raw_usage":{"total_tokens":1653,"prompt_tokens":952,"completion_tokens":701,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":696,"completion_tokens_details":{"reasoning_tokens":631}},"tokens_in":696,"tokens_out":701,"duration_ms":7391,"temperature":1.0,"reasoning_tokens":631,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T18:06:41.437688+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a compactly supported persistence measure μ and search, for example by gradient descent on point positions and weights, for a sequence μ_n with ∥S(μ_n)−S(μ)∥∞→0 but POT_1(μ_n,μ) bounded away from 0; a single such sequence, or two distinct measures with identical sphere values, would refute Theorem 4 or injectivity directly.","supporting_citations":[],"review_version":1}