Pith. sign in

REVIEW 1 major objections 5 minor 29 references

Recursive Computation of Path Homology for Stratified Digraphs

T0 review · 1 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read The paper proves that full-depth path homology of a stratified digraph can be computed layer by layer through null-space recursions, without enumerating long paths.

desk verdict New recursive method for full-depth path homology of stratified digraphs; sound, well-tested, and a real practical advance, with the known caveat that full-depth depends on the chosen stratification. read the letter →

arxiv 2412.05684 v2 pith:TKYSFJ4B submitted 2024-12-07 cs.CG

classification cs.CG MSC 55N3105C20
keywords pathhomologystratifieddigraphfull-depthcyclespacerecursionBettinumberpersistentfeedforwardnetworksdirectedacyclicgraph
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 addresses the practical difficulty of computing high-dimensional path homology of stratified digraphs—the layer-structured digraphs used to model feedforward neural networks. General path-homology algorithms scale poorly with dimension, roughly $O(n^{6+3p})$ in the general algorithm, so full-depth homology (dimension equal to the number of layers) has been out of reach except in shallow cases. The authors prove that full-depth cycles of a stratified digraph can be updated recursively one layer at a time, using only null-space computations on matrices built from the previous level's cycle space. The result is an algorithm with complexity $O(\sum_{p=1}^{L} |K_{p-1}||K_p|b_{p-1}^3)$, and experiments on random subgraphs of fully connected layered graphs show substantial speedups that grow with depth. The same recursion also yields full-depth persistent homology and, for acyclic digraphs, the maximal path homology.

What carries the argument

The load-bearing object is the cycle space recursion of Theorem 4.3: a linear bijection $\phi_S$ between full-depth cycles on a subgraph layer and tuples $(v_x)_{x\in S^+}$ lying in the null space of $[A_{x_1},\dots,A_{x_{n_p}}]$, where $A_x$ expresses a basis of the predecessor-cycle space in a global basis of the previous level's cycle space. The recursion works by computing, for each vertex in the next layer, the null space of the submatrix that removes the coordinates not connected to that vertex; the new layer's cycle space is then the null space of the concatenated $A_x$'s. This turns a high-dimensional homology computation into repeated linear-algebra null-space computations, with an analytic shortcut when a vertex is connected to all vertices in the previous layer.

What would settle it

Take a small stratified digraph, compute $\beta_L$ with Algorithm 1 and with the general path-homology algorithm of [9], and compare; any mismatch would contradict Theorem 4.3. More sharply, construct a digraph that admits two layerings with different $L$, such as one obtained by adding an empty edge layer, and verify whether the recursion yields different results under the two layerings.

Watch

Extended reading notes

Core claim

The central claim is that over the real numbers, the full-depth cycle space of a stratified digraph $G$ with layers $K_0,\dots,K_L$ has a Markovian structure: every full-depth cycle decomposes uniquely as $\sum_{x\in K_L} \gamma_x x$, where each $\gamma_x$ is a cycle in the subgraph ending at $x$'s predecessors and the sum of the $\gamma_x$ is zero. Because there are no $L$-boundaries, the full-depth path homology $H_L(G)$ is exactly this cycle space. The paper encodes each $\gamma_x$ in a fixed basis of the previous cycle space by a matrix $A_x$, so the whole cycle space becomes the null space of the block matrix $[A_{x_1},\dots,A_{x_{n_p}}]$. Theorem 4.3 proves this cycle space recursion and shows that iterating it from $p=0$ to $p=L$ produces a basis of $H_L(G)$, giving the stated complexity bound and the experimental speedups.

Load-bearing premise

The load-bearing premise is that, over the real numbers and for a fixed stratification, every full-depth cycle has a unique zero-sum decomposition into predecessor cycles; if that uniqueness fails—for example under a different layer assignment or with coefficients admitting torsion—the recursion's output is no longer a well-defined invariant of the digraph.

Editorial extensions

If this is right

  • Full-depth Betti numbers of deep stratified digraphs become computable in time governed by $|K_{p-1}||K_p|b_{p-1}^3$, rather than the general algorithm's $O(n^{6+3L})$.
  • For an increasing filtration of stratified digraphs, full-depth homology classes are never destroyed, so persistent full-depth homology reduces to plotting $\beta_L(G_t)$ against the filtration threshold; the paper demonstrates this on weighted-edge filtrations.
  • For any DAG with longest path length $\ell(G)$, the maximal path homology equals the full-depth homology of the subgraph $G^*$ spanned by longest paths, so the recursion computes it after an $O(|V|+|E|)$ pruning step.
  • Tracking an explicit basis of cycle paths remains possible but slower; without tracking, the recursion can exploit dense layers analytically, making runtime decrease as edge density increases in some regimes.

Reading between the lines

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

  • Going beyond the paper, the same null-space recursion could compute intermediate-dimensional cycle spaces of stratified digraphs by running the recursion up to a fixed $p<L$ and using the tracked basis of $\ker(\partial_p)$.
  • Going beyond the paper, one could test whether a canonical longest-path layering makes $\beta_L$ a digraph invariant, addressing the paper's own remark that layerings are not unique.
  • Going beyond the paper, extending the recursion to integer coefficients would require checking whether the unique-decomposition argument survives torsion; a small counterexample would settle it.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. The paper proposes a recursive algorithm to compute the full-depth reduced path homology of a stratified digraph, i.e., the homology in degree equal to the number of layers L. The main theoretical result (Theorem 4.3) shows that the full-depth cycle space of a stratified digraph can be obtained level by level: at each layer, the cycle space is the null space of a concatenated matrix whose blocks are coordinate-change matrices A_x for predecessor subgraphs. The algorithm is extended to maximal path homology of a DAG via Proposition 3.7, which constructs a stratified subgraph consisting of all longest paths. Experiments on random subgraphs of fully connected layered graphs compare runtime against the general path-homology algorithm [4] and report large speedups for deeper graphs, and the method is also applied to full-depth persistent homology.

Significance. If the technical content is correct, this is a useful contribution: it is, to my knowledge, the first algorithm specifically targeting high-dimensional path homology of stratified digraphs and feedforward networks, a regime where the general algorithm's O(n^{6+3p}) cost is prohibitive. The proofs are largely self-contained, the recursion is derived from the definitions rather than fitted, the code is open source, and the baseline is an external implementation. The main limitation, already acknowledged by the authors, is that 'full-depth' homology depends on the chosen stratification, so the algorithm's output is an invariant of the digraph together with its layer decomposition rather than of the underlying digraph alone. The complexity estimate, however, has a gap that needs repair before the theoretical efficiency claim can be accepted as stated.

major comments (1)
  1. [Section 5.1, complexity estimate] The inequality 'nrow(VKp−1) = P_{x∈K+_{p−1}} ncol(Ax) ≤ |Kp−1|bp−1' is not valid in general. For each x∈K_{p−1}^+, the matrix A_x is constructed from V_{K_{p−2}}, so ncol(A_x) ≤ b_{p−2}, and b_{p−2} can be much larger than b_{p−1}. For example, let L=3, K0={a1,...,am}, K1={u,v}, with a connected bipartite graph having m+2 edges (e.g., pred(u) and pred(v) of sizes about m/2 with a nonempty overlap). Then b0=m−1 and b1=1, while nrow(VK1) = (|pred(u)|−1)+(|pred(v)|−1) = m, so nrow(VK1) = m > |K1|b1 = 2. If K2 contains N vertices each connected only to u, then for each such vertex x, \tilde V_x has nrow about m/2 and ncol=1, so computing Basis(N(\tilde V_x)) costs Ω(m) per vertex; the stated bound O(|K1||K2|b1^3)=O(N) underestimates the true cost Ω(Nm) when N≫m. The complexity formula should be corrected (for example, to a bound involving b_{p−2}b_{p−1}^2 terms, up to base cases) or justified under additional hypotheses.
minor comments (5)
  1. [Section 3, Corollary 3.6] The proof of Corollary 3.6 is omitted for brevity, but the statement is nontrivial and is used to motivate the trimming procedure; please include a proof or relegate it to an appendix.
  2. [Section 5.1, basis tracking] The complexity analysis does not include the cost of basis tracking on line 29, although the experiments show that tracking materially increases runtime; the paper should state explicitly that the given bound applies only to the non-tracking mode.
  3. [Proposition 3.4, proof] In the displayed boundary formula, the term ∂_{s−1}(γ_{xi}x) appears where the path γ_{xi}x has length s; this is presumably a typo for ∂_s(γ_{xi}x) and should be corrected to avoid confusion.
  4. [Algorithm 1, line 29] The notation B(p) ← B(p−1) Σ_i x_i^p A_{x_i^p} V_i^{(p)} mixes path symbols with matrices; a brief explanation of how B(p) represents a basis of paths and how the multiplication is defined would improve readability.
  5. [Definition 1.1, remark] The cyclic-shift example for non-unique layers is correct but dense; a small concrete example, rather than only the general construction, would make the remark easier to follow.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the recursive full-depth path homology computation is derived from proven decompositions with no fitted parameters and no self-citations.

full rationale

The paper's central result, Theorem 4.3, is derived in the text: Proposition 4.2 proves that every full-depth cycle decomposes uniquely as Σ γ_x x with γ_x ∈ ker(∂_{p-1})∩A_{p-1}(G_pred(x)) and Σ γ_x = 0, and Theorem 4.3 then rewrites this coordinate space as V_{K_p} = N([A_{x_1},...,A_{x_{n_p}}]). The proof uses only the definition of the boundary map, Lemma 2.2, and Proposition 3.4, for which a proof is supplied; Proposition 3.4 is described as an extension of Proposition 4 in [8], but it is not imported as an unverified black box. The algorithm computes an exact algebraic quantity, namely the full-depth cycle space of a fixed stratified digraph, rather than predicting a fitted quantity; no parameter is calibrated to data. The only benchmark comparison uses an external implementation [4] of the general algorithm [9], and the bibliography contains no self-citations by the authors. The non-invariance of 'full-depth' under non-canonical layer choices is explicitly acknowledged in the remark after Definition 1.1 and is a scope limitation, not a circular step. Accordingly, no load-bearing derivation step reduces to its inputs by construction.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper introduces no free parameters and no invented entities. It relies on standard path-homology definitions, the defining property of stratified digraphs, and standard longest-path algorithms for DAGs. The axioms listed are background assumptions, not assumptions tailored to force the result.

assumptions (4)
  • standard math Path homology with real coefficients is defined by the chain complex (Λ_p, ∂_p) and the invariant subspace Ω_p; the isomorphism H_p = (ker ∂_p ∩ A_p) / ∂_{p+1}(Ω_{p+1}) holds.
    Invoked throughout Section 2.1, adapted from Grigor'yan et al. [14,15].
  • domain assumption In a stratified digraph with layers K0,...,KL, every allowed p-path has its ith vertex in K_i, so cross sections at different depths are disjoint.
    Direct consequence of Definition 1.1; enables the cycle decomposition in Proposition 3.4 and Proposition 4.2.
  • standard math For a DAG, the longest path length ℓ(G) and the subgraph G* of all longest paths can be computed in O(|V|+|E|) using topological sorting and edge relaxation.
    Used in Section 5.2 and Proposition 3.7; standard algorithms from [22,27] implemented in NetworkX [16].
  • domain assumption The general path homology algorithm of Chowdhury et al. [9] is correct and has complexity O(n^{6+3p}) for p-dimensional homology, serving as a valid baseline.
    Used for the numerical comparison in Section 6.1; accepted from [9] and the implementation [4].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recursive Computation of Path Homology for Stratified Digraphs." pith.science (2026). https://pith.science/paper/TKYSFJ4B

@misc{pith2026241205684,
  author       = {Pith},
  title        = {Pith review of: Recursive Computation of Path Homology for Stratified Digraphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TKYSFJ4B}},
  note         = {Machine review of arXiv:2412.05684}
}
read the original abstract

Stratified digraphs are popular models for feedforward neural networks. However, computation of their path homologies has been limited to low dimensions due to high computational complexity. A recursive algorithm is proposed to compute certain high-dimensional (reduced) path homologies of stratified digraphs. By recursion on matrix representations of homologies of subgraphs, the algorithm efficiently computes the full-depth path homology of a stratified digraph, i.e. homology with dimension equal to the depth of the graph. The algorithm can be used to compute full-depth persistent homologies and for acyclic digraphs, the maximal path homology, i.e., path homology with dimension equal to the maximum path length of a graph. Numerical experiments show that the algorithm has a significant advantage over the general algorithm in computation time as the depth of stratified digraph increases.

Figures

Figures reproduced from arXiv: 2412.05684 by the authors.

Figure 1
Figure 1. Modified persistence diagram for full-depth path homologies. [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages

  1. [4]

    Carranza, D., Doherty, B., Kapulkin, K., Opie, M., Sarazola, M., and Wong, L. Z. (2022). Python script for computing path homology of digraphs. https://github.com/sheaves/path_ homology, version 1.0.0

  2. [1]

    Ballester, R., Casacuberta, C., and Escalera, S. (2023). Topological data analysis for neural network analysis: A comprehensive survey. arXiv preprint arXiv:2312.05840

  3. [2]

    Berry, E., Chen, Y.-C., Cisewski-Kehe, J., and Fasy, B. T. (2020). Functional summaries of persistence diagrams. Journal of Applied and Computational Topology , 4(2):211–262

  4. [3]

    Biscio, C. A. and Møller, J. (2019). The accumulated persistence function, a new useful functional summary statistic for topological data analysis, with a view to brain artery trees and spatial point process applications. Journal of Computational and Graphical Statistics , 28(3):671–681

  5. [5]

    Chazal, F., De Silva, V., and Oudot, S. (2014a). Persistence stability for geometric complexes. Geometriae Dedicata, 173(1):193–214

  6. [6]

    Chazal, F., Glisse, M., Labru` ere, C., and Michel, B. (2014b). Convergence rates for persis- tence diagram estimation in topological data analysis. In International Conference on Machine Learning, pages 163–171

  7. [7]

    Chen, B., Yau, S.-T., and Yeh, Y.-N. (2001). Graph homotopy and Graham homotopy.Discrete Mathematics, 241(1-3):153–170

  8. [8]

    Chowdhury, S., Gebhart, T., Huntsman, S., and Yutin, M. (2019). Path homologies of deep feedforward networks. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pages 1077–1082. 17

Show all 29 references
  1. [9]

    Chowdhury, S., Huntsman, S., and Yutin, M. (2021). Path homology and temporal networks. In Complex Networks & Their Applications IX , pages 639–650

  2. [10]

    and M´ emoli, F

    Chowdhury, S. and M´ emoli, F. (2018). Persistent path homology of directed networks. In Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1152–1169

  3. [11]

    K., Li, T., and Wang, Y

    Dey, T. K., Li, T., and Wang, Y. (2022). An efficient algorithm for 1-dimensional (persistent) path homology. Discrete & Computational Geometry , 68(4):1102–1132

  4. [12]

    Topological persistence and simplification

    Edelsbrunner, Letscher, and Zomorodian (2002). Topological persistence and simplification. Discrete & computational geometry , 28:511–533

  5. [13]

    and Harer, J

    Edelsbrunner, H. and Harer, J. L. (2022). Computational Topology: An Introduction. Amer- ican Mathematical Society. Reprint of the 2010 original

  6. [14]

    Grigor’yan, A., Lin, Y., Muranov, Y., and Yau, S.-T. (2012). Homologies of path complexes and digraphs. arXiv preprint arXiv:1207.2834

  7. [15]

    Grigor’yan, A., Muranov, Y., and Yau, S.-T. (2017). Homologies of digraphs and K¨ unneth formulas. Communications in Analysis and Geometry , 25(5):969–1018

  8. [16]

    A., Schult, D

    Hagberg, A. A., Schult, D. A., and Swart, P. J. (2008). Exploring network structure, dy- namics, and function using NetworkX. In Varoquaux, G., Vaught, T., and Millman, J., editors, Proceedings of the 7th Python in Science Conference , pages 11–15. Python package available at ...

  9. [17]

    Ivashchenko, A. V. (1994). Contractible transformations do not change the homology groups of graphs. Discrete Mathematics, 126(1-3):159–170

  10. [18]

    Li, W., Dasarathy, G., Natesan Ramamurthy, K., and Berisha, V. (2020). Finding the ho- mology of decision boundaries with active learning. Advances in Neural Information Processing Systems, 3:8355–8365

  11. [19]

    Loiseaux, D., Carri` ere, M., and Blumberg, A. (2024). A framework for fast and stable representations of multiparameter persistent homology decompositions. Advances in Neural Information Processing Systems, 36

  12. [20]

    and Villa, A

    Masulli, P. and Villa, A. E. (2016). The topology of the directed clique complex as a network invariant. SpringerPlus, 5:1–12

  13. [21]

    and Lazar, N

    Moon, C. and Lazar, N. A. (2023). Hypothesis testing for shapes using vectorized persistence diagrams. Journal of the Royal Statistical Society Series C: Applied Statistics , 72(3):628–648

  14. [22]

    Pacault, J. F. (1974). Computing the weak components of a directed graph. SIAM Journal on Computing , 3(1):56–61

  15. [23]

    P´ erez-Fern´ andez, D., Guti´ errez-Fandi˜ no, A., Armengol-Estap´ e, J., and Villegas, M. (2021). Characterizing and measuring the similarity of neural networks with persistent homology.arXiv preprint arXiv:2101.07752. 18

  16. [24]

    Rathore, A., Chalapathi, N., Palande, S., and Wang, B. (2021). TopoAct: Visually exploring the shape of activations in deep learning. In Computer Graphics Forum , volume 40, pages 382–397

  17. [25]

    W., Nolte, M., Scolamiero, M., Turner, K., Perin, R., Chindemi, G., D lotko, P., Levi, R., Hess, K., and Markram, H

    Reimann, M. W., Nolte, M., Scolamiero, M., Turner, K., Perin, R., Chindemi, G., D lotko, P., Levi, R., Hess, K., and Markram, H. (2017). Cliques of neurons bound into cavities pro- vide a missing link between structure and function. Frontiers in computational neuroscience , 11:266051

  18. [26]

    Rieck, B., Togninalli, M., Bock, C., Moor, M., Horn, M., Gumbsch, T., and Borgwardt, K. (2018). Neural persistence: A complexity measure for deep neural networks using algebraic topology. arXiv preprint arXiv:1812.09764

  19. [27]

    and Wayne, K

    Sedgewick, R. and Wayne, K. (2011). Algorithms. Addison-Wesley Professional, 4 edition

  20. [28]

    Wasserman, L. (2018). Topological data analysis. Annual Review of Statistics and Its Appli- cation, 5(1):501–532

  21. [29]

    and Yamana, H

    Watanabe, S. and Yamana, H. (2022). Topological measurement of deep neural networks using persistent homology. Annals of Mathematics and Artificial Intelligence , 90(1):75–92. 19

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.