Pith. sign in

REVIEW 4 major objections 5 minor 43 references

ALBATROSS: Cheap Filtration Based Geometry via Stochastic Sub-Sampling

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper's central claim is that stochastic sub-sampling makes topological data analysis (TDA) cheap enough to map the human cortex's functional geometry, and the resulting map is mostly hyperbolic.

desk verdict Real memory win; circular geometry testing. read the letter →

arxiv 2509.03681 v1 pith:SCTKAQDK submitted 2025-09-03 q-bio.QM

Andrew J. Stier , Naichen Shi , Raed Al Kontar , Chad Giusti , Marc G. Berman This is my paper · ORCID
classification q-bio.QM
keywords topologicaldataanalysisBetticurvesstochasticsub-samplinghyperbolicgeometrypersistenthomologyresting-statefMRIfunctionalconnectivitycorticalparcellation
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

Topological data analysis can reveal whether a dataset lives on a curved space, but computing filtered simplicial complexes on large matrices once required hundreds of gigabytes of memory. The paper introduces ALBATROSS, a protocol that computes Betti curves on many small random sub-samples of an adjacency matrix and averages them, so memory use depends on the sub-sample size rather than the full dataset. It validates the protocol by reproducing earlier findings on strawberry odor chemistry and rat hippocampal place-cell firing. It then runs TDA where it was previously infeasible: voxel-wise resting-state fMRI correlation matrices for 392 cortical parcels in ten people. The headline finding is that all ten global 392-by-392 matrices, and 353 of 392 local parcels, are statistically consistent with hyperbolic geometry.

What carries the argument

The object that carries the argument is the Betti curve of a filtered simplicial complex, which counts n-dimensional holes in the clique complex of a weighted adjacency matrix as the threshold varies. ALBATROSS replaces the full complex with an average over many random sub-samples of size n, a stochastic sub-sampling estimate; it then searches candidate geometries (Euclidean, hyperbolic shell or full space, or a shuffled null) by minimizing a normalized l1-plus-integrated-Betti-value objective, and converts the fit into p-values through a central-limit-theorem z-statistic with Bonferroni and Fisher corrections.

What would settle it

Split the sub-samples into two halves: use one half to select each model's parameters (minimizing normalized l1 plus integrated-Betti-value distance) and the other half to compute the reported p-values. If ALBATROSS's geometry claims are genuine, 'consistent with hyperbolic' should survive the split for most parcels and for the global matrices; if the p-values collapse, the geometry finding is an artifact of using the same data to fit and test.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that ALBATROSS turns filtered-simplicial-complex TDA from a supercomputer-scale computation into a laptop-scale one, and that doing so exposes a coherent geometric organization in the human cortex: the resting-state functional connectivity of every individual studied is consistent with a hyperbolic space at the global level, and most parcels show local hyperbolic geometry. The authors are careful to frame this as statistical consistency with candidate model spaces rather than direct proof that the cortex 'is' hyperbolic. What would follow is that cortical functional organization has a negatively curved, hierarchical structure, with local topology va

Load-bearing premise

The protocol assumes that p-values obtained by comparing the empirical Betti curves to the model that was chosen to best fit those same curves are valid evidence of geometric consistency; if that post-model-selection comparison is not controlled, the reported 'consistent with hyperbolic' results are in-sample fit statistics rather than confirmatory tests.

Editorial extensions

If this is right

  • TDA on matrices with hundreds of rows and columns moves from a 475 GB compute node to a 24 GB laptop, so researchers without institutional clusters can run geometry checks.
  • The same protocol applies to any weighted adjacency matrix, not just neuroimaging, and the two validation datasets show it generalizes across data types.
  • Because classic network-science measures are a subset of filtered-complex information, ALBATROSS also provides a route to large-scale versions of measures such as rich-club structure.
  • If the cortical finding is correct, most of the brain's functional parcels sit in a negatively curved space, linking local connectivity topology to the brain's global hierarchical organization.
  • Memory use stays roughly constant for a chosen sub-sample size, so the approach scales to datasets that would otherwise be untouchable.

Reading between the lines

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

  • The candidate-geometry family is limited to Euclidean, hyperbolic, and shuffled null spaces; including spherical or product geometries could reclassify the 39 parcels currently inconsistent with all three, so the 'hyperbolic' label is relative to that family.
  • Because the same data choose and then test the model, the reported p-values are likely optimistic; a held-out split of sub-samples would turn ALBATROSS into a general model-selection tool whose geometry calls are confirmatory.
  • The correlation between local hyperbolic radius and global radial coordinate predicts a testable signature: interventions that shift a region's global hierarchical position should change its local Betti curves even when local signal quality is unchanged.
  • Replacing the radial sampling density in the hyperbolic model with the exact hyperbolic volume element could change the best-fit radii and noise levels, though probably not the coarse geometry family.
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

4 major / 5 minor

Summary. This paper introduces ALBATROSS, a protocol for computing Betti curves of filtered simplicial complexes of large adjacency matrices by stochastic sub-sampling. The protocol computes average empirical Betti curves from random subsamples, searches over Euclidean, hyperbolic, and shuffled candidate models using integrated Betti value (ibv) and L1-distance statistics, and then uses a CLT-based z-test combined by Fisher's method to produce p-values. The authors validate the approach on strawberry odor and hippocampal place-cell datasets and apply it to resting-state fMRI functional connectivity matrices from 392 cortical parcels in 10 HCP subjects, reporting that all ten global matrices and 353/392 local parcels are consistent with hyperbolic geometry, and that local hyperbolic radius correlates with the global radial coordinate (Spearman rs = 0.52).

Significance. If the statistical inference were valid, this paper would make a useful practical contribution: the memory benchmarks (e.g., roughly 71 GB for a direct n=300 computation versus 1.5 GB for ALBATROSS with n=30 on a laptop) and the public code would substantially lower the barrier to TDA on large biological datasets. The HCP application is also novel in scale. However, the statistical engine of the paper is not valid as implemented: the hypothesis tests use the same statistics that were minimized during model selection, and the CLT variance is mis-scaled. The headline geometric conclusions are therefore unsupported, and the central claim that ALBATROSS provides valid statistical geometry inference is not established.

major comments (4)
  1. [Online Methods, Eqs. 10–12 and Eqs. 4–5] Circular model selection and testing. The best-fit model is chosen by minimizing the objective Σ_{k=1}^3 (l1^k + ibv^k) in Eq. 12, where l1^k and ibv^k are exactly the normalized L1-distance and integrated Betti value statistics defined in Eqs. 10–11. The subsequent hypothesis test in Eqs. 4–5 evaluates the same empirical Betti curves against that selected model using the same L1 and ibv statistics. No split-sample, hold-out, or post-selection adjustment is provided. Consequently, the observed statistics are minimizers of the same functional used to build the null distribution, and the p-values in Supplementary Table 5 and the 'consistent with hyperbolic' statements in Figure 3 are in-sample fit statistics rather than valid hypothesis tests. The same issue affects the strawberry and hippocampal validation results.
  2. [Online Methods, Eq. 3] Incorrect CLT variance scaling. For a mean of r independent resamplings, the variance of the mean is Var(β_r)/r, not Var(β_r)·r. As written, Eq. 3 inflates the variance by a factor of r^2 relative to the correct expression and deflates every Z statistic in Eq. 4 by a factor of r. For r=100, this is a two-order-of-magnitude change in the test statistic, so all reported p-values would change after correction.
  3. [Online Methods, Eq. 7] Hyperbolic radial density and unspecified dimension. The radial density p(r) ∼ sinh((d−1)r) is not the radial marginal of the uniform measure in d-dimensional hyperbolic space; the correct density is proportional to sinh^{d−1}(r). For d>2, the simulated 'hyperbolic' model is a different, non-uniform ensemble, so consistency with that model does not directly imply consistency with hyperbolic geometry. The manuscript applies 3-D hyperbolic MDS but does not report the d used in the Betti-curve fits; d is absent from the Bayesian-optimization parameter list and from Supplementary Tables 3–5. The sampler should be corrected or the model explicitly re-described as a non-uniform radial model.
  4. [Online Methods, Statistical Inference] The 'shuffled' candidate space is mentioned repeatedly in the main text, Figure 3, and Supplementary Table 5, but the manuscript never specifies how the shuffled adjacency matrices are generated. Without this definition, the claim that zero parcels are consistent with the a-geometric shuffled model is not reproducible or interpretable.
minor comments (5)
  1. [Online Methods, Eq. 8] The denominator in the cosine formula reads ||xi||||xi||; it should be ||xi|| ||xj||.
  2. [Figure 3 and main text] Figure 3a states that all parcels had topology consistent with a hyperbolic shell, but the main text says 353 of 392 regions are consistent and 39 are inconsistent with any candidate space. This contradiction needs to be resolved.
  3. [Supplementary Table 3] The listed n values contain duplicates and omissions: n ∈ [10,15,20,25,40,45,50,45,50,55] should be cleaned. Also, the table reports only rmax and rmin; the best-fit d and noise level ε are described as varying but are not tabulated, which hampers reproducibility.
  4. [Main text and Online Methods] There are repeated typos: 'stoachstic' (main text), 'itegrated' (Online Methods), and 'the first n Betti numbers' should be 'the first three Betti numbers' in context.
  5. [Supplementary Table 5] p-values are printed as 0.0. After Bonferroni correction and Fisher's method, p-values cannot be exactly zero; they should be reported with inequalities such as p < 1e-10.

Circularity Check

1 steps flagged · score 7.0 of 10

P-values supporting hyperbolic-geometry claims are computed after model selection on the same l1/ibv statistics, making them in-sample fit statistics rather than valid tests.

  1. fitted input called prediction [Online Methods, 'Statistical Inference' (Eqs 3-5) vs. 'Grid search for best fit euclidean geometry' and 'Bayesian optimization for best fit hyperbolic geometry' (Eqs 10-12)]
    "Finally, the grid point with the smallest combined metric,P3 k=1 l1k + ibvk, is chosen as the best fit model. ... Our objective function is defined as: argmax(ε,rmin,rmax)obj = − 3X k=1 lk 1 + ibvk (12) ... These Z statistics are computed for the L1-distance between empirical and best fit geometric Betti curves and for the area under the Betti curves (11)."

    The normalized l1 and ibv discrepancies in Eqs 10-11 are the terms of the model-selection objective (Eq 12: minimized by grid search for Euclidean models; the negative sum is maximized by Bayesian optimization for hyperbolic models). The same 'L1-distance between empirical and best fit geometric Betti curves and ... area under the Betti curves' define the Z statistics in Eq 4, whose null distribution is obtained by re-sampling from the selected best-fit model (Eq 3). Thus the empirical test statistic is the minimum of the objective used to pick the model: it is a selection-minimized in-sample quantity, not a draw from the null distribution. Consequently the p-values (Bonferroni-corrected, Fisher-combined in Eq 5) and the resulting 'consistent with Hyperbolic geometry' classifications (Figu

full rationale

The central circularity is selective inference: model parameters are chosen by minimizing the exact l1 and ibv discrepancies that are then used as test statistics against a null derived from the chosen model, with no split-sample or post-selection adjustment. The p-values in Figure 3 and Supplementary Table 5 are therefore in-sample fit statistics, undermining the load-bearing geometric conclusions for the cortex. However, the paper is not wholly circular: the memory-reduction engineering claim is benchmarked independently, and ALBATROSS reproduces earlier full-data conclusions for strawberry odors and hippocampal place cells, providing external support for the pipeline's ability to recover known geometry. These independent elements keep the score below 8. Separately, Eq 7's radial density p(r) ~ sinh((d-1)*r) is not the standard hyperbolic volume element; that is a modeling/correctness concern distinct from circularity and is not counted in the score.

Assumptions & free parameters 8 free parameters · 6 assumptions · 1 invented entities

The central inference rests on imported and hand-set assumptions: subsampling representativeness from prior literature, a CLT applied to Betti measures, Fisher's independence, an unspecified shuffled null, and a possibly incorrect hyperbolic radius distribution. The candidate-geometry parameters (d, epsilon, rmax, rmin) are fitted to the same curves later tested, so they are free parameters, not predictions.

free parameters (8)
  • subsample_size_n = n=35 for HCP; n in [10,55] in validation
    Number of rows/columns randomly sampled per iteration in ALBATROSS; controls memory and statistical behavior, chosen based on the authors' claim that inference is reliable for n>30.
  • iterations_I = I=150 for HCP; 100 to 1000 in validation
    Number of random subsamples averaged to form each empirical/model Betti curve; larger I reduces variance and is a user-chosen parameter.
  • euclidean_dimension_d = Grid searched over d in [1,100]; e.g., 12-64 for place cells
    Dimension of the Euclidean candidate model; fitted by grid search to minimize distance to empirical Betti curves.
  • euclidean_noise_epsilon = Grid searched over [0.0,0.15]; low values for place cells
    Multiplicative Gaussian noise added to Euclidean model distances; fitted to the empirical curves.
  • hyperbolic_rmax = Selected by BO over [1e-6,1]; e.g., rmax=5 for strawberry
    Maximum radius of the hyperbolic shell model; fitted via Bayesian optimization to the empirical Betti curves.
  • hyperbolic_rmin = Selected by BO over [1e-6,0.99]; e.g., 0.9 for strawberry, 0.0 for place cells
    Minimum radius as a fraction of rmax in the hyperbolic shell; fitted during optimization.
  • hyperbolic_noise_epsilon = Selected by BO over [1e-6,1]; varies considerably with n and I
    Noise level in the hyperbolic distance model; fitted rather than predicted.
  • hyperbolic_dimension_d = Fixed; 3 for the cortex analysis, not searched
    Dimension of the hyperbolic model; chosen by hand rather than fitted, and it affects all Betti-curve fits.
assumptions (6)
  • domain assumption Stochastic subsampling of an adjacency matrix yields Betti curves representative of the full matrix
    Relied on throughout; justified by citing refs 23 and 24 rather than proved here. The main text explicitly notes this had not previously been studied for clique complexes of weighted networks.
  • standard math The central limit theorem applies to Betti-curve measures across independent subsample iterations
    Invoked in Eq 3 to construct normal null distributions for the test statistics; requires independence and finite variance, which is not established for the dependent, threshold-coupled Betti curve measures.
  • standard math Fisher's method is valid for the six combined p-values
    Eq 5 combines six p-values arising from Betti curves that are not independent; Fisher's method requires independent p-values, so the resulting chi-squared statistic does not have the stated null distribution.
  • domain assumption The shuffled adjacency matrix is a valid a-geometric null model
    Used as the third candidate model in the cortex analysis, but no definition, construction, or validation of the shuffled model is given in the methods.
  • ad hoc to paper Uniform sampling of the hyperbolic shell uses radial density p(r) ~ sinh((d-1)r)
    Eq 7 in Online Methods. The standard uniform volume element in d-dimensional hyperbolic space is sinh^{d-1}(r), not sinh((d-1)r); this appears to bias all hyperbolic model Betti curves.
  • ad hoc to paper Best-fit model parameters need not be held out before goodness-of-fit testing
    The pipeline fits (epsilon, rmax, rmin, d) to the empirical curves and then computes p-values against the same fitted model (Eqs 4-5 vs Eqs 10-12). No post-selection correction or demonstration of validity is provided.
invented entities (1)
  • No new physical or mathematical entities introduced
    purpose: The protocol introduces no new particle, force, dimension, conserved quantity, or ledger entry
    The only named objects are the ALBATROSS protocol and candidate null models, which are procedural constructs rather than invented entities with independent falsifiable handles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ALBATROSS: Cheap Filtration Based Geometry via Stochastic Sub-Sampling." pith.science (2026). https://pith.science/paper/SCTKAQDK

@misc{pith2026250903681,
  author       = {Pith},
  title        = {Pith review of: ALBATROSS: Cheap Filtration Based Geometry via Stochastic Sub-Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SCTKAQDK}},
  note         = {Machine review of arXiv:2509.03681}
}
read the original abstract

Topological data analysis (TDA) detects geometric structure in biological data. However, many TDA algorithms are memory intensive and impractical for massive datasets. Here, we introduce a statistical protocol that reduces TDA's memory requirements and gives access to scientists with modest computing resources. We validate this protocol against two empirical datasets, showing that it replicates previous findings with much lower memory requirements. Finally, we demonstrate the power of the protocol by mapping the topology of functional correlations for the human cortex at high spatial resolution, something that was previously infeasible without this novel approach.

Figures

Figures reproduced from arXiv: 2509.03681 by the authors.

Figure 1
Figure 1. Euclidean multi dimensional scaling fails to capture the perceptual axes of strawberry [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. a) Computation of Betti curves requires exponentially more memory with increasing [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Topology of functional correlations across human cortex. a) ALBATROSS was run [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 39 canonical work pages

  1. [1]

    A. E. Sizemore, J. E. Phillips-Cremins, R. Ghrist, D. S. Bassett, Network Neuroscience 3, 656 (2019)

    work page 2019

  2. [2]

    Wasserman, Annual review of statistics and its application 5, 501 (2018)

    L. Wasserman, Annual review of statistics and its application 5, 501 (2018)

    work page 2018

  3. [3]

    J. M. Shine, et al., Nature neuroscience 22, 289 (2019)

    work page 2019

  4. [4]

    Nair, et al., Cell 186, 178 (2023)

    A. Nair, et al., Cell 186, 178 (2023)

    work page 2023

  5. [5]

    A. K. Suresh, et al., elife 9, e58848 (2020)

    work page 2020

  6. [6]

    Hinton, S

    G. Hinton, S. T. Roweis, NIPS (Citeseer, 2002), vol. 15, pp. 833–840

    work page 2002

  7. [7]

    Mead, Journal of the Royal Statistical Society: Series D (The Statistician) 41, 27 (1992)

    A. Mead, Journal of the Royal Statistical Society: Series D (The Statistician) 41, 27 (1992)

    work page 1992

  8. [8]

    Dahmen, S

    D. Dahmen, S. Gr ¨un, M. Diesmann, M. Helias, Proceedings of the National Academy of Sciences 116, 13051 (2019)

    work page 2019

Show all 43 references
  1. [9]

    Giusti, Algebraic topology and neuroscience: A bibliography (2018)

    C. Giusti, Algebraic topology and neuroscience: A bibliography (2018). 11

    2018

  2. [10]

    Giusti, E

    C. Giusti, E. Pastalkova, C. Curto, V . Itskov, Proceedings of the National Academy of Sci- ences 112, 13455 (2015)

    2015

  3. [11]

    Y . Zhou, B. H. Smith, T. O. Sharpee,Science advances 4, eaaq1458 (2018)

    2018

  4. [12]

    Y . Zhou, T. O. Sharpee,Iscience 24, 102225 (2021)

    2021

  5. [13]

    Singh, et al., Journal of Vision 8, 11 (2008)

    G. Singh, et al., Journal of Vision 8, 11 (2008)

    2008

  6. [14]

    R. K. Luneburg (1947)

    1947

  7. [15]

    Indow, The global structure of visual space , vol

    T. Indow, The global structure of visual space , vol. 1 (World Scientific, 2004)

    2004

  8. [16]

    Bauer, Journal of Applied and Computational Topology 5, 391 (2021)

    U. Bauer, Journal of Applied and Computational Topology 5, 391 (2021)

    2021

  9. [18]

    Curto, Bulletin of the American Mathematical Society 54, 63 (2017)

    C. Curto, Bulletin of the American Mathematical Society 54, 63 (2017)

    2017

  10. [19]

    Giusti, R

    C. Giusti, R. Ghrist, D. S. Bassett, Journal of Computational Neuroscience 41, 1 (2016)

    2016

  11. [20]

    D. S. Bassett, O. Sporns, Nature neuroscience 20, 353 (2017)

    2017

  12. [21]

    Sporns, Networks of the Brain (MIT press, 2010)

    O. Sporns, Networks of the Brain (MIT press, 2010)

    2010

  13. [22]

    Bendich, J

    P. Bendich, J. S. Marron, E. Miller, A. Pieloch, S. Skwerer,The Annals of Applied Statistics 10 (2016)

    2016

  14. [23]

    N. O. Malott, A. M. Sens, P. A. Wilsey, 2020 IEEE International Conference on Big Data (Big Data) (IEEE, 2020), pp. 2681–2690

    2020

  15. [24]

    Solomon, A

    E. Solomon, A. Wagner, P. Bendich, arXiv preprint arXiv:2101.12288 (2021)

    2021 arXiv

  16. [25]

    M. A. Koyama, V . Robins, K. Turner,arXiv preprint arXiv:2412.07805 (2024). 12

    2024 arXiv

  17. [26]

    M. A. Koyama, F. Memoli, V . Robins, K. Turner,arXiv preprint arXiv:2307.16333 (2023)

    2023 arXiv

  18. [27]

    Zhang, P

    H. Zhang, P. D. Rich, A. K. Lee, T. O. Sharpee, Nature Neuroscience 26, 131 (2023)

    2023

  19. [28]

    Bubenik, M

    P. Bubenik, M. Hull, D. Patel, B. Whittle, Inverse Problems 36, 025008 (2020)

    2020

  20. [29]

    R. C. Craddock, G. A. James, P. E. Holtzheimer III, X. P. Hu, H. S. Mayberg,Human brain mapping 33, 1914 (2012)

    1914

  21. [30]

    Jiang, T

    X. Jiang, T. Zhang, S. Zhang, K. M. Kendrick, T. Liu, Psychoradiology 1, 23 (2021)

    2021

  22. [31]

    Henselman, R

    G. Henselman, R. Ghrist, arXiv preprint arXiv:1606.00199

    arXiv

  23. [32]

    R. W. Ghrist, Elementary applied topology, vol. 1 (Createspace Seattle, 2014)

    2014

  24. [33]

    Ghrist, Bulletin of the American Mathematical Society 45, 61 (2008)

    R. Ghrist, Bulletin of the American Mathematical Society 45, 61 (2008)

    2008

  25. [34]

    C. R. Harris, et al., Nature 585, 357 (2020)

    2020

  26. [35]

    Sedgwick, Bmj 344 (2012)

    P. Sedgwick, Bmj 344 (2012)

    2012

  27. [36]

    R. A. Fisher, Breakthroughs in statistics (Springer, 1992), pp. 66–70

    1992

  28. [37]

    P. I. Frazier, arXiv preprint arXiv:1807.02811 (2018)

    2018 arXiv

  29. [38]

    C. K. Williams, C. E. Rasmussen, Gaussian processes for machine learning , vol. 2 (MIT press Cambridge, MA, 2006)

    2006

  30. [39]

    Balandat, et al., Advances in neural information processing systems 33, 21524 (2020)

    M. Balandat, et al., Advances in neural information processing systems 33, 21524 (2020)

    2020

  31. [40]

    Praturu, T

    A. Praturu, T. O. Sharpee, Iscience 27 (2024)

    2024

  32. [41]

    M. L. Schwieterman, et al., PloS one 9, e88446 (2014). 13

    2014

  33. [42]

    Y . Wang, S. Romani, B. Lustig, A. Leonardo, E. Pastalkova, Nature neuroscience 18, 282 (2015)

    2015

  34. [43]

    Pastalkova, V

    E. Pastalkova, V . Itskov, A. Amarasingham, G. Buzs´aki, Science 321, 1322 (2008)

    2008

  35. [44]

    M. F. Glasser, et al., Neuroimage 80, 105 (2013). Acknowledgments The authors thank Tatyana Sharpee for sharing code and helping to replicate the sampling of points from candidate hyperbolic spaces and Salar Fattahi for helpful discussions around Bayesian optimization. Concept...

    2013

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.