Pith. sign in

REVIEW 1 major objections 5 minor 2 references

Persistence Spheres: a Bi-continuous Linear Representation of Measures for Partial Optimal Transport

T0 review · 1 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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

desk verdict 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. read the letter →

arxiv 2603.15384 v2 pith:NB4Y7YAN submitted 2026-03-16 stat.ML cs.LGmath.ATmath.STstat.TH

Matteo Pegoraro This is my paper · ORCID
classification stat.MLcs.LGmath.ATmath.STstat.TH MSC 55N3162R4049Q22
keywords persistencediagramspartialoptimaltransportliftzonoidtopologicalvectorizationbi-continuousembeddingReLUapproximationspheresmeasurerepresentations
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

1 major / 5 minor

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.

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 (1)
  1. [§6.2.3, Step 2 (Eq. (29))] 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.
minor comments (5)
  1. [§6.2.3, Step 2] 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.
  2. [§6.2.3, Theorem 3] 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).
  3. [§3, Definition 4] 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.
  4. [§10–§11 and 'Code' paragraph] 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.
  5. [§12] 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.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation-level circularity: the core proofs are self-contained or rest on an external ReLU-approximation theorem; only minor non-load-bearing self-citations to Pegoraro (2026) appear.

full rationale

I walked the main derivation chain: Definition 13 (persistence sphere) -> Proposition 7 (injectivity) -> Theorem 2 (POT_1 stability) -> Lemmas 7-9 (tail vanishing) -> Theorem 3 (local Hölder bound on compact cores) -> Theorem 4 (inverse continuity) -> Theorem 6/Corollary 7 (L^2 bi-continuity on M_c(A,B)). No step renames a fitted parameter as a prediction and no step defines the output in terms of the target quantity. The construction is parameter-free: S(µ) = Λ(µ - (π_Δ)#µ)|_{S^2}, with Λ a fixed ReLU integral. Stability follows from Kantorovich-Rubinstein duality applied to cross-augmented positive measures; injectivity follows from the external lift-zonoid injectivity theorem of Koshevoy-Mosler/Hendrych-Nagy; inverse continuity funnels through Theorem 3. The only load-bearing external input is the shallow-ReLU Sobolev-to-variation estimate (29), imported from Mao et al. (2024) at Theorem 3, Step 2. That estimate is not equivalent to the theorem being proved: it is an independent approximation statement about representing Sobolev test functions by ReLU ridge integrals, and it does not assume POT_1 convergence, sphere convergence, or the inverse-continuity conclusion. The fact that A_0 is not supplied or numerically verified is a correctness/completeness risk, not circularity. The self-citations to Pegoraro (2026) appear as lineage, as the weighted comparison construction PSph*, and as the source of empirical baselines; none of the mathematical proofs of Theorem 4 or Corollary 7 invokes a theorem from Pegoraro (2026). I therefore find no circular step, and the score reflects only the presence of minor, non-load-bearing self-citation.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The mathematical construction has no fitted constants: the persistence sphere is parameter-free up to numerical discretization. The main unproved inputs are standard convex-analysis and optimal-transport facts, a recent ReLU approximation theorem used as a black box, and the compact-support/growth-class assumptions that delimit the inverse-continuity and Hilbert-space results. Experimental hyperparameters (grid sizes, kernel bandwidths, spline grids) are not part of the central theorem and are not counted as free parameters of the mathematical claim.

assumptions (5)
  • standard math Lift-zonoid transform is injective on integrable positive measures and Hausdorff convergence of lift zonoids is equivalent to weak convergence plus uniform integrability.
    Used in Proposition 4 and Proposition 7; cited from Koshevoy and Mosler (1998) and Hendrych and Nagy (2022), not reproved in the paper.
  • standard math Kantorovich-Rubinstein duality for OT1 holds for finite integrable measures on X with the ∞-norm.
    Used throughout the stability and inverse estimates, e.g., Theorem 2 Step 2 and Theorem 3 Step 0.
  • standard math Shallow ReLU ridge functions approximate functions in W^{5/2,2}(B_1) with variation-norm control: ∥g∥_{K_1(P_1)} ≤ A_0 ∥g∥_{W^{5/2,2}(B_1)}.
    This is the external black-box bound from Mao et al. (2024) that Theorem 3 relies on for the local Hölder inverse estimate; the constant A_0 is not reproduced or certified in the paper.
  • domain assumption The target measure μ in the inverse-continuity theorem is compactly supported in the upper half-plane.
    Theorem 4 and Corollary 2 explicitly assume compact support; continuity of the inverse is proved only at every compactly supported target, not for general integrable measures.
  • domain assumption The growth-controlled class M_c(A,B) satisfies the compatibility condition A(B_R) B(R^2 \ B_{R-1}) / R → 0.
    This assumption is introduced in Definition 14 to make the L2-to-uniform upgrade in Theorem 6 work; it is a new modeling condition, not a standard theorem.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Persistence Spheres: a Bi-continuous Linear Representation of Measures for Partial Optimal Transport." pith.science (2026). https://pith.science/paper/NB4Y7YAN

@misc{pith2026260315384,
  author       = {Pith},
  title        = {Pith review of: Persistence Spheres: a Bi-continuous Linear Representation of Measures for Partial Optimal Transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NB4Y7YAN}},
  note         = {Machine review of arXiv:2603.15384}
}
abstract

We improve and extend persistence spheres, introduced in~\cite{pegoraro2025persistence}. Persistence spheres map an integrable measure $\mu$ on the upper half-plane, including persistence diagrams (PDs) as counting measures, to a function $S(\mu)\in C(\mathbb{S}^2)$, and the map is stable with respect to 1-Wasserstein partial transport distance $\mathrm{POT}_1$. Moreover, to the best of our knowledge, persistence spheres are the first explicit representation used in topological machine learning for which continuity of the inverse on the image is established at every compactly supported target. Recent bounded-cardinality bi-Lipschitz embedding results in partial transport spaces, despite being powerful, are not given by the kind of explicit summary map considered here. Our construction is rooted in convex geometry: for positive measures, the defining ReLU integral is the support function of the lift zonoid. Building on~\cite{pegoraro2025persistence}, we refine the definition to better match the $\mathrm{POT}_1$ deletion mechanism, encoding partial transport via a signed diagonal augmentation. In particular, for integrable $\mu$, the uniform norm between $S(0)$ and $S(\mu)$ depends only on the persistence of $\mu$, without any need of ad-hoc re-weightings, reflecting optimal transport to the diagonal at persistence cost. This yields a parameter-free representation at the level of measures (up to numerical discretization), while accommodating future extensions where $\mu$ is a smoothed measure derived from PDs (e.g., persistence intensity functions~\citep{wu2024estimation}). Across clustering, regression, and classification tasks involving functional data, time series, graphs, meshes, and point clouds, the updated persistence spheres are competitive and often improve upon persistence images, persistence landscapes, persistence splines, and sliced Wasserstein kernel baselines.

Figures

Figures reproduced from arXiv: 2603.15384 by the authors.

Figure 1
Figure 1. Example of the lift zonoid construction for a discrete measure [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. One-point diagrams drifting along (1, 1). Left: attenuation of a pure diagonal offset between Dk and D′ k . Right: behavior of the deletion-to-diagonal cost for Dk. The updated definition matches the POT1 geometry by construction, whereas in Pegoraro (2026) both effects are mediated by the decay of the reweighting ωK along diagonal lines. 45 [PITH_FULL_IMAGE:figures/full_fig_p045_2.png] view at source ↗
Figure 3
Figure 3. Two elementary mechanisms by which standard topological summaries deform [PITH_FULL_IMAGE:figures/full_fig_p048_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: FDA simulations. Left: realizations from the baseline generative model used in scenarios (i)–(iii), where the discriminative signal is more likely to be carried by large oscillations while noise induces abundant low-persistence clutter. Right: realizations from scenari…
Figure 5
Figure 5. Figure 5: PP simulation. Top row: one simulated point pattern for each point process. Bottom row: corresponding persistence diagrams in degrees H0 (blue) and H1 (red). Or￾dering: Poisson, Thomas, Mat´ern, jittered lattice. “Eyeglasses” regression. This synthetic regression task …
Figure 6
Figure 6. Figure 6: Examples of data, persistence diagrams, and PSph representations appearing in [PITH_FULL_IMAGE:figures/full_fig_p059_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 1 linked inside Pith

  1. [209]

    Theodore Papamarkou, Tolga Birdal, Michael M Bronstein, Gunnar E Carlsson, Justin Curry, Yue Gao, Mustafa Hajij, Roland Kwitt, Pietro Lio, Paolo Di Lorenzo, et al

    American Mathematical Society Providence, 2015. Theodore Papamarkou, Tolga Birdal, Michael M Bronstein, Gunnar E Carlsson, Justin Curry, Yue Gao, Mustafa Hajij, Roland Kwitt, Pietro Lio, Paolo Di Lorenzo, et al. Position: Topological deep learning is the new frontier for relational learning. In Inter- national Conference on Machine Learning , pages 39529–...

    arXiv 2015

  2. [2008]

    Sara Kaliˇ snik

    doi: 10.1002/9780470725160. Sara Kaliˇ snik. Tropical coordinates on the space of persistence barcodes. Foundations of Computational Mathematics , 19(1):101–129, 2019. Olav Kallenberg. Foundations of modern probability. Springer, 1997. Gleb Koshevoy and Karl Mosler. Lift zonoids, random convex hulls and the variability of random vectors. Bernoulli, 4(3):3...

    arXiv 2019

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.