Pith. sign in

REVIEW 2 major objections 4 minor 41 references

Selecting Interpretable Circular Coordinates from Data

T0 review · 2 major / 4 minor · reviewed 2026-07-10 · grok-4.5

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

Pith's one-line read Detected loops in data can be explained by the lowest-energy circle-valued dictionary candidates whose classes span the selected cohomology.

desk verdict Clean dictionary-selection method for persistent H1 with a real consistency proof and usable diagnostics; soft spots are the usual clean-manifold/single-scale ones, not a broken core. read the letter →

arxiv 2607.08230 v1 pith:3FUE7K4F submitted 2026-07-09 math.AT stat.ML

Vincent P. Grande , Marina Meila This is my paper
classification math.ATstat.ML MSC 55N3155U1057R1968T09
keywords circularcoordinatespersistentcohomologydictionaryselectioncochaininnerproductvectormatroidDirichletenergymoleculartorsionshead-directioncells
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

Persistent cohomology can find circular structure in data, but the resulting coordinates are abstract: they do not say which measured angle, phase, torsion, or decoder is responsible. This paper formulates that interpretation task as selecting a small set of user-supplied circle-valued candidates so that the cohomology classes of their pulled-back angular forms span the relevant H1 subspace at minimal Dirichlet energy. Continuously the problem is a minimum-weight basis in a vector matroid and is solvable by the greedy algorithm. For finite point clouds the method CIRCOL builds density-corrected cochain weights, projects dictionary edge cochains onto discrete harmonic representatives of the selected persistent classes, and runs the same greedy selection. The authors prove that the cochain inner product consistently estimates the continuous L2 product of fixed smooth 1-forms under non-uniform sampling, and that the resulting projection matrix also diagnoses topologically trivial candidates and unexplained classes. Synthetic circles and tori, molecular torsion dictionaries, and head-direction cell orderings confirm that the selected coordinates recover the expected generators.

What carries the argument

The density-corrected cochain inner product on oriented edges (kernel weights divided by empirical kernel masses, normalized to match continuum L2) that converges in probability to the Riemannian L2 product of fixed smooth 1-forms; its Gram matrix produces the projection coefficients that drive both selection and diagnostics.

What would settle it

On a non-uniformly sampled noisy torus whose dictionary contains the true angles, higher-winding multiples, and mixed combinations, check whether CIRCOL still returns the two ground-truth generators as the unique lowest-energy full-rank basis once sample size and bandwidth enter the theorem’s asymptotic regime; systematic selection of mixed or multi-winding candidates would refute the selection claim.

Watch

Extended reading notes

Core claim

CIRCOL recovers a minimum-energy subset of dictionary circular coordinates whose pulled-back classes form a basis for the selected persistent H1, using a density-corrected cochain inner product that is a consistent estimator of the continuous L2 product of smooth 1-forms and a greedy vector-matroid basis algorithm.

Load-bearing premise

The data must be independent samples from a clean smooth compact manifold without boundary, with all important loops appearing together as integer-liftable classes at one filtration scale whose 1-skeleton contains every short kernel pair.

Editorial extensions

If this is right

  • A domain scientist can replace ad-hoc colorings of embeddings by an automatic ranking of which supplied angles, torsions, or decoders explain each persistent loop.
  • Rows of the projection matrix that stay near zero flag either topologically trivial dictionary entries or persistent classes that no candidate explains, giving an immediate diagnostic.
  • Homology classes that do not come from the true manifold are orthogonal to every valid dictionary form and can be dropped before the matroid step.
  • The same pipeline identifies physical torsions in molecular trajectories and the correct cyclic order among head-direction cells without using ground-truth labels during selection.

Reading between the lines

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

  • Because the continuous problem is already a matroid, the same greedy energy argument could be tried for other integral-period coordinates (for example spherical or toroidal coordinates) once an analogous discrete energy is defined.
  • Consistency of the bilinear form on fixed smooth forms does not automatically give convergence of the discrete harmonic representatives; closing that gap would strengthen finite-sample recovery guarantees.
  • Relaxing exact cocycle conditions to nearly harmonic cochains, as the discussion already flags, would let the method audit noisy or outlier-obstructed cycles common in single-cell and sensor data.
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

2 major / 4 minor

Summary. The paper proposes CIRCOL for selecting scientifically meaningful circle-valued dictionary functions that explain persistent H1 classes detected in data. In the continuous setting, each candidate is represented by the cohomology class of its pulled-back angular form, and minimum-Dirichlet-energy selection of a spanning set is cast as a minimum-weight basis problem in a vector matroid. For point clouds, the authors construct a density-corrected cochain inner product, prove its consistency for fixed smooth 1-forms under non-uniform sampling (Theorem 3.1 / Appendix A), form a projection matrix onto discrete harmonic representatives of integer-lifted persistent classes, and greedily select a low-energy spanning dictionary subset (Algorithm 1). The projection matrix is also used diagnostically for trivial candidates and unexplained classes. Experiments on synthetic circle/torus data, molecular torsion dictionaries, and head-direction neural recordings support the method.

Significance. If the claims hold, this is a useful and well-scoped contribution at the interface of TDA and interpretable manifold learning: it turns abstract circular coordinates into audited dictionary explanations rather than another coordinate construction. The continuous matroid reduction is clean, and the consistency theorem for the density-corrected cochain inner product is carefully proved with standard nonparametric rates. The diagnostic use of the projection matrix and the successful recovery of torsions and cyclic neural orderings are concrete strengths. The work is complementary to ManifoldLasso/TSLasso and to existing circular-coordinate constructions, and the limitations (clean manifold, single scale, cost of PH) are stated honestly.

major comments (2)
  1. The consistency result (Theorem 3.1 / A.8) is carefully scoped to fixed smooth 1-forms, but the algorithm optimizes over discrete harmonic representatives obtained by constrained minimization on C1(Sε;R). The manuscript itself notes (end of Appendix A) that cochain-inner-product consistency does not yet imply convergence of those harmonic representatives or of the selected bases. For the central discrete claim, a short statement of what is and is not proved—or a sketch of the additional conditions needed—would make the load-bearing gap explicit rather than only implicit.
  2. Section 3 and the discrete problem definition assume a single filtration scale ε at which all selected integer-lifted classes are simultaneously alive and approximate the true manifold cohomology, together with a 1-skeleton containing all kernel-supported pairs. Experiments appear to work under this regime, but the paper would be stronger with a brief sensitivity check (or explicit protocol) for how ε, h, and the choice of k affect the projection matrix and selected basis on at least one real example (e.g., ethanol or head-direction). Without that, the free parameters remain load-bearing for practical use even though they do not invalidate the proved estimator.
minor comments (4)
  1. Figure 1 panels D–G and Figure 4 are informative, but axis labels and the meaning of “relative weighted Dirichlet energy” could be stated more explicitly in the captions so that the integer winding recovery is immediately readable without the main text.
  2. Notation for the cochain weight matrix M, the Gram matrix Q, and the projection matrix P is introduced cleanly in Section 3; a short summary table of continuous vs discrete objects would still help readers moving between Sections 2 and 3.
  3. Related work on spectral exterior calculus and diffusion geometry is appropriately cited; a one-sentence contrast with Maggs et al. (already mentioned) on gene-set cyclic views versus external dictionary auditing would further clarify positioning.
  4. Minor typographical inconsistencies appear (e.g., “circol” vs “CIRCOL”, occasional spacing around citations). A light copy-edit pass would suffice.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: dictionary and persistent classes are independent inputs; the cochain inner-product estimator is derived from continuum L2 geometry with a self-contained proof.

full rationale

The central claims do not reduce to their inputs by construction. Continuous selection is a standard min-weight basis problem in a vector matroid on projections of pulled-back angular forms (Section 2); dictionary elements and the target H1 subspace are independent. CIRCOL (Algorithm 1) discretizes this via persistent cohomology, integer lifts, harmonic representatives, and a density-corrected cochain inner product; Theorem 3.1/A.8 proves consistency of that inner product for fixed smooth 1-forms under i.i.d. non-uniform sampling by a four-step continuum-to-discrete argument (sphere averages, geodesic integrals, density correction, edge discretization) fully proved in Appendix A, not fitted to recover a pre-chosen answer. Costs cj = omega_j^T M omega_j and projections pj = Q^{-1} A^T M omega_j are computed from independent cochains; greedy selection is ordinary matroid optimization. Theorem 3.2 uses the dictionary only as an external filter for noise classes (if i^*h = 0 then pairings vanish), not as a self-definition of the classes. Experiments hold out ground-truth angles/orders from complex construction and selection. Related-work citations (circular coordinates [30], ManifoldLasso/TSLasso [18,19], authors' prior Hodge work [13,14]) provide context or contrast and are not load-bearing premises that force the estimator or selection result. No fitted-input-as-prediction, uniqueness-from-self-citation, or ansatz-smuggling chain closes the loop.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claims rest on standard de Rham / persistent-cohomology background, a clean manifold sampling model, and a few algorithmic choices (bandwidth, filtration scale, prime for coefficients). No new physical entities are postulated; CIRCOL is an algorithmic construction whose independent evidence is the consistency theorem and the recovery experiments.

free parameters (3)
  • bandwidth h (and relation to filtration scale ε)
    Kernel bandwidth for the cochain weights; must satisfy h→0 and N h^d / log N → ∞ for consistency, but is chosen by the user in practice and affects both bias and sparsity.
  • filtration scale ε and number k of selected persistent classes
    User/heuristic choice of which bars and which single scale are treated as the true H1; the algorithm assumes all selected classes are simultaneously alive.
  • odd prime p for Z/pZ persistent cohomology
    Standard TDA choice for coefficient field before integer lift; affects which classes lift cleanly via the Bockstein obstruction.
assumptions (5)
  • domain assumption Data are i.i.d. samples from a smooth compact Riemannian manifold without boundary with strictly positive smooth density π.
    Required for the consistency theorem (Theorem 3.1 / A.8) and for the continuous de Rham picture; authors note real data may be noisy or stratified.
  • standard math de Rham theorem and existence of harmonic representatives with integer periods for integral classes.
    Used throughout Section 2 to identify circular coordinates with closed 1-forms of integer periods.
  • domain assumption A single filtration scale exists at which all selected integer-lifted classes approximate the manifold cohomology.
    Stated in Section 3; multi-scale selection is mentioned as possible but not developed.
  • domain assumption Dictionary candidates admit consistent short angular unwrapping so that triangle defects vanish and cochains are cocycles.
    Required for ω(gj) to represent a cohomology class; non-zero defects are diagnosed but not automatically corrected.
  • standard math Vector matroid greedy algorithm solves the minimum-weight basis problem optimally.
    Cited via Edmonds; used to justify Algorithm 1 selection step.
invented entities (1)
  • CIRCOL algorithm and density-corrected cochain weight matrix M independent evidence
    purpose: Discrete estimator of L2 inner product on 1-forms and the selection/diagnostic pipeline that uses it.
    New algorithmic object; independent evidence is the consistency theorem for fixed smooth forms and empirical recovery on synthetic and real data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Selecting Interpretable Circular Coordinates from Data." pith.science (2026). https://pith.science/paper/3FUE7K4F

@misc{pith2026260708230,
  author       = {Pith},
  title        = {Pith review of: Selecting Interpretable Circular Coordinates from Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3FUE7K4F}},
  note         = {Machine review of arXiv:2607.08230}
}
abstract

Circular coordinates obtained from persistent cohomology reveal loop structure in data, but they usually remain abstract: A detected circle does not tell us which measured angle, phase, torsion, or decoder explains it. We propose a method for selecting interpretable circle-valued coordinates from a user-supplied dictionary of scientifically meaningful candidates explaining the detected cohomology. In the continuous setting, each candidate is represented by the cohomology class of its pulled-back angular form, and selecting a minimum-energy set of candidates spanning the relevant $H^1$ subspace becomes a minimum-weight basis problem in a vector matroid. We then introduce CIRCOL, a method for discrete point clouds sampled from the manifold. We prove that the introduced cochain inner product is a consistent estimator of the $L^2$ inner product of fixed smooth 1-forms under non-uniform sampling. The resulting projection matrix both helps selecting a basis of low-energy dictionary coordinates and diagnoses topologically trivial candidates or unexplained persistent classes. Finally, we verify the effectiveness of our method on synthetic examples, on molecular simulations, and neural recordings of head-direction cells.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages

  1. [1]

    Baird and J

    P . Baird and J. C. Wood. Harmonic morphisms between Riemannian manifolds. 29. Oxford University Press, 2003

    work page 2003

  2. [2]

    Belkin and P

    M. Belkin and P . Niyogi. ‘Laplacian eigenmaps for dimensionality reduction and data representation’. In:Neural computation 15.6 (2003), pp. 1373–1396

    work page 2003

  3. [3]

    Berry and D

    T. Berry and D. Giannakis. ‘Spectral exterior calculus’. In:Communications on Pure and Applied Mathematics 73.4 (2020), pp. 689–770

    work page 2020

  4. [4]

    C heng, J

    M. C heng, J. J ansen, K. R eimer, V . P . Grande, J. S. N agai, Z. L i, P . Kießling, M. Grasshoff, C. Kuppe, M. T. Schaub, R. Kramann and I. G. Costa. ‘PHLOWER lever- ages single-cell multimodal data to infer complex, multi-branching cell differentiation trajectories’. In:Nature Methods (2025)

    work page 2025

  5. [5]

    Chmiela, H

    S. Chmiela, H. E. Sauceda, K.-R. Müller and A. Tkatchenko. ‘Towards exact molecular dynamics simulations with machine-learned force fields’. In: Nature Communications 9.1 (2018), p. 3887

    work page 2018

  6. [6]

    R. R. C oifman and S. Lafon. ‘Diffusion maps’. In:Applied and computational harmonic analysis 21.1 (2006), pp. 5–30

    work page 2006

  7. [7]

    Duszkiewicz

    A. Duszkiewicz. Local origin of excitatory-inhibitory tuning equivalence in a cortical network (Duszkiewicz et al. 2024). Data set and code. 2025. url: https://doi.org/10.6084/m9. figshare.24921252

    work page doi:10.6084/m9 2024

  8. [8]

    Duszkiewicz, S

    A. Duszkiewicz, S. Skromne Carrasco and A. Peyrache. Large-scale recordings of head direction cells in mouse postsubiculum. Version 0.250207.0025. Data set. 2025. url: https: //dandiarchive.org/dandiset/000939/0.250207.0025. 22

    work page arXiv 2025

Show all 41 references
  1. [9]

    A. J. D uszkiewicz, P . Orhan, S. S kromne Carrasco, E. H. B rown, E. O wczarek, G. R. Vite, E. R. Wood and A. Peyrache. ‘Local origin of excitatory-inhibitory tuning equivalence in a cortical network’. In: Nature Neuroscience 27 (2024), pp. 782–792

    2024

  2. [10]

    B. Eckmann. ‘Harmonische Funktionen und Randwertaufgaben in einem Komplex.’ In: Commentarii mathematici Helvetici 17 (1944–1945), pp. 240–255

    1944

  3. [11]

    J. Edmonds. ‘Matroids and the greedy algorithm’. In:Mathematical programming 1.1 (1971), pp. 127–136

    1971

  4. [12]

    R. J. Gardner, E. Hermansen, M. Pachitariu, Y. Burak, N. A. Baas, B. A. Dunn, M.-B. Moser and E. I. Moser. ‘Toroidal topology of population activity in grid cells’. In:Nature 602.7895 (2022), pp. 123–128

    2022

  5. [13]

    V . P . Grande and M. T. Schaub. ‘Disentangling the Spectral Properties of the Hodge Laplacian: Not All Small Eigenvalues Are Equal’. In: arXiv preprint arXiv: 2311.14427 (2024). arXiv: 2311.14427

    2024 arXiv

  6. [14]

    V . P . Grande and M. T. Schaub. ‘Point-Level Topological Representation Learning on Point Clouds’. In: Forty-second International Conference on Machine Learning. 2025

    2025

  7. [15]

    Hélein and J

    F. Hélein and J. C. Wood. ‘Harmonic maps’. In:Handbook of global analysis 1213 (2008), pp. 417–491

    2008

  8. [16]

    I. J ones. ‘Diffusion Geometry’. In:arXiv preprint arXiv:2405.10858 (2024)

    2024 arXiv

  9. [17]

    I. Jones. ‘Manifold Diffusion Geometry: Curvature, Tangent Spaces, and Dimension’. In: arXiv preprint arXiv:2411.04100 (2024)

    2024

  10. [18]

    S. J. Koelle, H. Zhang, M. Meila and Y.-C. Chen. ‘Manifold coordinates with physical meaning’. In:Journal of Machine Learning Research 23.133 (2022), pp. 1–57

    2022

  11. [19]

    S. J. Koelle, H. Zhang, O.-V . Murad and M. Meila. ‘Consistency of dictionary-based manifold learning’. In: International Conference on Artificial Intelligence and Statistics. PMLR. 2024, pp. 4348–4356

    2024

  12. [20]

    Kovacev-Nikolic, P

    V . Kovacev-Nikolic, P . Bubenik, D. Nikoli ´c and G. Heo. ‘Using persistent homology and dynamical distances to analyze protein binding’. In: Statistical Applications in Genetics and Molecular Biology 15.1 (Jan. 2016)

    2016

  13. [21]

    Maehara and Y

    K. Maehara and Y. Ohkawa. ‘Modeling latent flows on single-cell data using the Hodge decomposition’. In:bioRxiv (2019), p. 592089

    2019

  14. [22]

    Maggs, M

    K. Maggs, M. K. Youssef, C. Pulver, J. Isma, T. J. Nguyên, M. Arzt, W. Karthaus, H. A. Harrington, K. Hess and G. P . Dotto. ‘Topology identifies concurrent cyclic processes in single-cell transcriptomics and androgen receptor function’. In: bioRxiv (8th Dec. 2025). Preprint, ...

    2025 doi

  15. [23]

    Paik and J

    T. Paik and J. Park. ‘Circular coordinates for density-robust analysis’. In:arXiv preprint arXiv:2301.12742 (2023)

    2023 arXiv

  16. [24]

    J. A. P erea. ‘Multiscale Projective Coordinates via Persistent Cohomology of Sparse Filtrations’. In:Discrete & Computational Geometry 59.1 (2018), pp. 175–225

    2018

  17. [25]

    J. A. Perea. ‘Sparse Circular Coordinates via Principal Z-Bundles’. In:Topological Data Analysis. Ed. by N. A. B aas, G. E. C arlsson, G. Q uick, M. S zymik and M. Thaule. Vol. 15. Abel Symposia. Cham: Springer International Publishing, 2020, pp. 435–458. url: https://doi.org/...

    2020 doi

  18. [26]

    J. A. Perea and J. Harer. ‘Sliding Windows and Persistence: An Application of Topo- logical Methods to Signal Analysis’. In: Foundations of Computational Mathematics 15.3 (2015), pp. 799–838

    2015

  19. [27]

    Rybakken, N

    E. Rybakken, N. A. Baas and B. A. Dunn. ‘Decoding of Neural Data Using Cohomological Feature Extraction’. In: Neural Computation 31.1 (2019), pp. 68–93

    2019

  20. [28]

    N. C. S chonsheck and S. C. S chonsheck. ‘Spherical coordinates from persistent co- homology’. In:Journal of Applied and Computational Topology 8.1 (2024), pp. 149–173

    2024

  21. [29]

    Scoccola, H

    L. Scoccola, H. Gakhar, J. Bush, N. Schonsheck, T. Rask, L. Zhou and J. A. Perea. ‘Toroidal Coordinates: Decorrelating Circular Coordinates with Lattice Reduction’. In:39th International Symposium on Computational Geometry (SoCG 2023). Ed. by E. W. Chambers and J. Gudmundsson....

    2023 doi

  22. [30]

    M orozov and M

    V .de Silva, D. M orozov and M. Vejdemo-Johansson. ‘Persistent Cohomology and Circular Coordinates’. In: Discrete & Computational Geometry 45.4 (2011), pp. 737–759

    2011

  23. [31]

    Z. Su, Y. Tong and G.-W. Wei. ‘Hodge decomposition of single-cell RNA velocity’. In: Journal of chemical information and modeling 64.8 (2024), pp. 3558–3568

    2024

  24. [32]

    J. S. Taube, R. U. Muller and J. Ranck James B. ‘Head-direction cells recorded from the postsubiculum in freely moving rats. I. Description and quantitative analysis’. In: The Journal of Neuroscience 10.2 (1990), pp. 420–435

    1990

  25. [33]

    Vandereyken, A

    K. Vandereyken, A. Sifrim, B. Thienpont and T. Voet. ‘Methods and applications for single-cell and spatial multi-omics’. In: Nature Reviews Genetics 24.8 (2023), pp. 494–515. A. P roof of Convergence of the 1-cochain inner product In this section, we will give a proof of the c...

    2023

  26. [38]

    31 Let M = MN,hN be the diagonal matrix defined in Step 4 from this sample and bandwidth hN

    the 1-skeleton of the chosen simplicial complex on X contains every unordered pair {xi, xj} for which KhN (xi, xj) ̸= 0. 31 Let M = MN,hN be the diagonal matrix defined in Step 4 from this sample and bandwidth hN. Then, for every fixed smooth α, β ∈ Ω1(M), the sampled cochains...

  27. [39]

    (M, g) is a smooth compact d-dimensional Riemannian manifold without boundary

  28. [40]

    ,xN} is drawn i.i.d

    the sample X = {x1, . . . ,xN} is drawn i.i.d. from a measure ν = π µ, where π ∈ C2(M) is strictly positive

  29. [41]

    the kernel profile κ is bounded, supported in [0, 1), and satisfies m2 > 0

  30. [42]

    h = hN → 0, Nh d N/ log N → ∞, and h N < inj(M) for all sufficiently large N

  31. [43]

    Let M = MN,hN be the diagonal matrix defined in Step 4 from this sample and bandwidth hN

    the 1-skeleton of the chosen simplicial complex on X contains every unordered pair {xi, xj} for which KhN (xi, xj) ̸= 0. Let M = MN,hN be the diagonal matrix defined in Step 4 from this sample and bandwidth hN. Then, for every fixed smooth α, β ∈ Ω1(M), the sampled cochains fr...

  32. [44]

    the term h2 N is the geometric bias from replacing an infinitesimal inner product by an average over a ball of radius hN

  33. [45]

    Corollary A.9 (Balanced bandwidth)

    the term q log N/(Nhd N) is the finite-sample error from estimating that local average using only finitely many nearby pairs. Corollary A.9 (Balanced bandwidth). Under the assumptions of theorem A. 8, if hN ≍ log N N 1/(d+4) , (A. 84) then, for every fixed smooth α, β ∈ Ω1(M),...

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.