REVIEW 3 major objections 3 minor 5 references
Data and homotopy types
T0 review · 3 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read Under explicit density and scale-gap assumptions, an inclusion of data sets induces homotopy equivalences of Vietoris-Rips and Lesnick complexes, and a controlled homotopy equivalence of branch-point hierarchies.
desk verdict The Vietoris-Rips and Lesnick interleaving theorems are solid and checkable; the advertised controlled homotopy equivalence of branch point posets is only sketched and needs formalization. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The argument runs on two mechanisms. The first is an interleaving homotopy diagram built from a choice function θ that sends each point of Y to an r-close point of X, making V_s(Y) deform into V_{s+2r}(X) through a homotopy that factors through face inclusions; the same construction works on Lesnick complexes once the density assumption is upgraded to configurations of k+1 distinct points. The second is the poset of branch points Br_k(X) inside the hierarchy tree Γ_k(X): least upper bounds of branch points are again branch points, and the inclusion Br_k(X) ⊂ Γ_k(X) has a homotopy inverse given by sending each vertex to its unique maximal branch point below it. This calculus of least upper bounds is what turns the pointwise interleavings into controlled homotopy equivalences of branch-point posets, with the shift map s_* recording the bounded scale distortion.
What would settle it
Place a=(0,0), b=(0.1,0) in X and add c=(0.1,1), d=(0,1) to get Y, with r=0.2 and k=1. The configuration $X^{2}$_dis is not r-dense in $Y^{2}$_dis: the pair (c,a) is farther than r from both (a,b) and (b,a). At scale s=1 the phase gap is 0.9 > 2r, yet L_{1,1}(X) is an edge while L_{1,1}(Y) is a 4-cycle, so the inclusion is not a homotopy equivalence; this shows the configuration-density assumption is doing real work.
Extended reading notes
Core claim
The central claim is that the pair of assumptions - r-density plus the scale gap 2r < s_{i+1} - s_i - turns an inclusion X ⊂ Y into homotopy equivalences of simplicial complexes. Corollary 2 states that if X is r-dense in Y and 2r < s_{i+1} - s_i, then i: V_{s_i}(X) → V_{s_i}(Y) is a weak homotopy equivalence; Corollary 4 states the same for L_{s_i,k}(X) → L_{s_i,k}(Y) when $X^{{k+1}}$_{dis} is r-dense in $Y^{{k+1}}$_{dis}. Theorems 1 and 3 establish the underlying interleaving homotopy commutative diagrams using a choice function θ that sends each point (or each (k+1)-tuple) of Y to an r-close point (or tuple) of X. On the hierarchy side, the paper defines branch points of the tree Γ_k(X) = Γ(π_0 L_{*,k}(X)), proves that Br_k(X) ⊂ Γ_k(X) is a homotopy equivalence via the maximal-branch-point map, and shows that the induced map i_*: Br_k(X) → Br_k(Y) satisfies θ_* i_* ≤ s_* and i_* θ_* ≤ s_*, where s_* shifts scale by 2r; hence it is a controlled homotopy equivalence, becoming an actual equivalence when 2r is below the next phase-change gap.
Load-bearing premise
The result depends on the larger data set being well approximated not just point by point but at the level of configurations: every collection of k+1 distinct points of the larger set must have a corresponding collection of k+1 distinct points of the smaller set within distance r; without that, the Lesnick and branch-point statements do not follow.
Editorial extensions
If this is right
- For point clouds with a dense subsample and no phase-change gap below 2r, persistent homology of the Vietoris-Rips filtration is unchanged by the subsample at the sampled scale parameters.
- The same holds for Lesnick complexes, so density-filtered clustering outputs are unchanged at those scales.
- The branch-point posets Br_k(X) and Br_k(Y) have homotopies θ_* i_* ≤ s_* and i_* θ_* ≤ s_*, with s_* a shift by 2r; when 2r is smaller than the next phase gap these homotopies collapse to an actual homotopy equivalence.
- The least upper bound in Γ_k(X) induces an ultrametric on π_0 L_{s,k}(X), giving a quantitative merge-time distance for density clusters.
- If the configuration-space density condition holds for all k up to a fixed bound, the entire hierarchy of branch-point posets is stable under the inclusion up to the same 2r scale shift.
Reading between the lines
- One could read the configuration-space density condition as a requirement that the subsample realizes the local k-th order correlation structure of the larger set; pointwise density alone is insufficient exactly when coarse k-tuples are sampled too sparsely.
- The controlled homotopy equivalence suggests a practical stability estimate for hierarchical density clustering: subsampling moves branch points by at most 2r in the scale coordinate, so the dendrogram timetable is stable up to a known shift.
- Because Br_k(X) is a tree that is homotopy equivalent to the full hierarchy Γ_k(X), it may serve as a sparse replacement for the full cluster hierarchy in persistence computations, potentially reducing the number of vertices needed to represent the same connectivity information.
- A numerical experiment could measure the least upper bounds of branch points before and after subsampling and check whether the inequality (s,x) ≤ (t+2r, θ(y)) ≤ (s+2r,x) holds with t within 2r; this would test the controlled equivalence directly.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies inclusions X⊂Y⊂R^n of finite data sets and gives explicit density conditions under which the induced maps of Vietoris-Rips complexes V_s(X)→V_s(Y) and Lesnick complexes L_{s,k}(X)→L_{s,k}(Y) are weak homotopy equivalences at scale parameters separated from the next phase change. The proofs use elementary interleaving maps based on choosing r-close points. The second half defines hierarchy posets Γ_k(X) and their branch point subposets Br_k(X), develops a least-upper-bound calculus, and claims that the induced branch point map is a controlled homotopy equivalence with shift 2r.
Significance. The interleaving results in Section 1 are clean, self-contained, and checkable; the density assumption on distinct tuples for Lesnick complexes is explicit, and no fitted parameters appear. If the branch-point stability claim can be formalized, it would give a concrete stability statement for hierarchical clusterings relevant to HDBSCAN and DBSCAN. However, as it stands, the branch-point part of the paper is a sketch rather than a theorem, so the main advertised consequence is not yet established.
major comments (3)
- [Section 2, after Lemma 12] The claim in the abstract and introduction that the branch point map is a 'controlled homotopy equivalence' is never stated as a precise theorem. The inequalities θ_* i_* ≤ s_* and i_* θ_* ≤ s_* are described informally, but the paper does not define 'controlled homotopy equivalence' for posets, does not state the hypotheses under which these inequalities imply an equivalence (e.g., whether the scale gap 2r < s_{i+1}-s_i is required), and does not prove that s_*(s,[x]) = (s,[x]) under that condition. Since this is advertised as a main consequence, the central claim is currently not checkable.
- [Section 2, paragraph beginning 'This map takes a branch point...'] There is an unresolved contradiction: the text first says 'The map i∗ preserves least upper bounds by Lemma 7' and later says 'the map i∗ only preserves least upper bounds up to homotopy.' Lemma 7 concerns the inclusion Br_k(X)⊂Γ_k(X), not i_*; the second statement is the correct one, and the first must be deleted or replaced. As written, a reader cannot tell which statement is intended.
- [Section 2, final paragraph] 'If the bound 2r is sufficiently small, then (s,[x]) is the largest branch point below (s+2r,[x]) and s∗(s,[x])=(s,[x]) in that case.' This condition is used implicitly to convert the interleaving inequalities into an equivalence, but it is only asserted in prose and not stated as a lemma with proof. If the scale-gap condition 2r < s_{i+1}-s_i is the intended hypothesis, it should be stated and proved in the same way as Corollaries 2 and 4.
minor comments (3)
- [Theorem 3 proof, first sentence] The proof begins 'Suppose that y∈ L_{s,k}(X)_0−L_{s,k}(X)_0', which is an empty set; the intended statement is likely about vertices of L_{s,k}(Y).
- [Theorems 1 and 3] The homotopy commutative diagrams are established via natural transformations on posets of non-degenerate simplices, i.e., on subdivisions; the paper should say explicitly that this gives homotopy commutativity after subdivision and hence for the original complexes, since the reader may otherwise be confused about the category in which the diagrams commute.
- [Section 2, shift map notation] The notation s∗ is used for a poset morphism on Br_k(X) but its domain is often left implicit; specify whether s∗ is defined on X or on Y in each displayed inequality.
Circularity Check
No circular derivation: Corollaries 2 and 4 are direct consequences of explicit density and scale-gap hypotheses; the branch-point section is informal but does not presuppose its conclusion.
full rationale
The central derivation chain is self-contained. Theorem 1 constructs the map θ directly from the r-density assumption and verifies the interleaving inequalities d(θ(y_i),θ(y_j)) ≤ s+2r; Corollary 2 then uses the phase-gap assumption 2r < s_{i+1}-s_i to conclude that the horizontal maps are identities, so the weak equivalence is forced by the stated hypotheses rather than by a fitted or pre-supposed quantity. Theorem 3 and Corollary 4 follow the same pattern, with θ built from the explicit distinct-tuple density assumption X^{k+1}_dis r-dense in Y^{k+1}_dis; no parameter in the conclusion is fitted to the data. The branch-point inequalities θ_* i_* ≤ s_*, i_* θ_* ≤ s_* and id ≤ s_* are derived in Section 2 from the definitions of maximal branch points and from the interleaving maps, with s_* the 2r-shift; even though Section 2 does not assemble these into a formally stated 'controlled homotopy equivalence' theorem, the argument does not assume the equivalence. The only self-citation, to the author's preprint [3] for stable components, is motivational: branch points are defined independently and the later inequalities are proved from those definitions, so the citation is not load-bearing. The text's conflicting Section 2 statements about whether i_* preserves least upper bounds pointwise or only up to homotopy are a correctness/formalization defect, not a circular reduction. Score 1 reflects the minor self-citation and informal branch-point presentation, not a circular derivation.
Assumptions & free parameters
assumptions (5)
- domain assumption X and Y are finite subsets of R^n with the Euclidean metric.
- domain assumption X is r-dense in Y for Theorems 1-2, or X^{k+1}_dis is r-dense in Y^{k+1}_dis in the configuration space for Theorems 3-4.
- domain assumption The scale gap 2r < s_i+1 - s_i holds for the phase-change numbers.
- domain assumption The poset Gamma_k(X) of path components is a tree, meaning it is a contractible poset, and has least upper bounds.
- standard math Standard homotopy theory of simplicial complexes and posets, including weak equivalences, mapping cylinders, and natural transformations as homotopies.
invented entities (2)
-
Branch point poset Br_k(X)
-
Least upper bound distance d on pi0 L_{s,k}(X)
Cite this review
Pith. "Pith review of Data and homotopy types." pith.science (2026). https://pith.science/paper/ELMIY4YC
@misc{pith2026190806323,
author = {Pith},
title = {Pith review of: Data and homotopy types},
year = {2026},
howpublished = {\url{https://pith.science/paper/ELMIY4YC}},
note = {Machine review of arXiv:1908.06323}
}
read the original abstract
This paper presents explicit assumptions for the existence of interleaving homotopy equivalences of both Vietoris-Rips and Lesnick complexes associated to an inclusion of data sets. Consequences of these assumptions are investigated on the space level, and for corresponding hierarchies of clusters and their sub-posets of branch points. Hierarchy posets and branch point posets admit a calculus of least upper bounds, which is used to show that the map of branch points associated to the inclusion of data sets is a controlled homotopy equivalence.
Reference graph
Works this paper leans on
-
[1]
Andrew J. Blumberg and Michael Lesnick. Universality of the homo topy interleaving distance. CoRR, abs/1705.01690, 2017
arXiv 2017
-
[2]
Characterization, stabilit y and con- vergence of hierarchical clustering methods
Gunnar Carlsson and Facundo M´ emoli. Characterization, stabilit y and con- vergence of hierarchical clustering methods. J. Mach. Learn. Res. , 11:1425– 1470, 2010
work page 2010
-
[3]
J.F. Jardine. Stable components and layers. Preprint, 2019
work page 2019
-
[4]
M. Lesnick and M. Wright. RIVET: visualization and analysis of two- dimensional persistent homology. http://rivet.online, 2019
work page 2019
-
[5]
Accelerated hierarchical densit y based clustering
Leland McInnes and John Healy. Accelerated hierarchical densit y based clustering. In 2017 IEEE International Conference on Data Mining Work- shops, ICDM Workshops 2017, New Orleans, LA, USA, November 1 8-21, 2017, pages 33–42, 2017. 12
work page 2017
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.