What genuinely connects the persistent-homology gradient-free bridge and the hub-selection-artifact-absence observation

persistent-homologytopological-data-analysisgeneralizationsampling-artifactcross-domain-bridgeneural-networkscliodynamicsmethodologygradient-free

The vault's retrieval index flagged observation-persistent-homology-gradient-free-bridge-widrow-byzantine and observation-hub-selection-artifact-absent-by-design-in-founding-ph-generalization-papers at cosine 0.87 and unlinked. No prior capture in 10-inbox/raw/ investigates this specific pair together (checked by slug search); this is a fresh investigation, not a duplicate. Both notes already rest on independently verified Tier 1 primaries — Birdal et al. (2021), Gutiérrez-Fandiño et al. (2021), and Bernal-Alvarado et al. (2026) — so the question this capture answers is not "are the underlying facts true" (already settled in the vault) but "is the connection between these two readings of those facts real, or does it just look that way because they share vocabulary."

Claim: The two notes are one continuous argument, not two independent resemblances that happen to share vocabulary

observation-persistent-homology-gradient-free-bridge-widrow-byzantine establishes that persistent homology — specifically, Wasserstein/bottleneck distance between persistence diagrams — is the same computational operation diagnosing two different collapses: claim-birdal-2021-persistent-homology-dimension-bounds-generalization and claim-gutierrez-fandino-2021-persistence-diagram-distance-tracks-generalization on the neural-network side, and claim-roman-byzantine-trade-network-decoupled-after-1082-chrysobull on the historical-network side. That note's own closing lines explicitly raise the next question: whether the neural-network diagnostics inherit "the Byzantine study's own documented failure mode — sampling bias reversing a topological signal," i.e. claim-hub-selection-artifact-can-reverse-network-breakpoint-signal. observation-hub-selection-artifact-absent-by-design-in-founding-ph-generalization-papers is the direct answer to that question: it reads both founding NN-PH papers' method sections and finds neither has a point of entry for a structural-selection artifact — Birdal's estimator samples training iterates uniformly at random over time, and Gutiérrez-Fandiño's method does not subsample at all. The second note's opening sentence names the same open question the first note's closing paragraph raises. The 0.87 cosine reflects a real dependency, not surface overlap: the second observation could not have been asked without the first having proposed the bridge in the first place. On promotion, these two notes should carry a direct mutual wikilink rather than being connected only indirectly through claim-hub-selection-artifact-can-reverse-network-breakpoint-signal and the (currently unfiled) question they both reference.

Claim: A third founding-era NN persistent-homology diagnostic — Rieck et al.'s "Neural Persistence" (2019) — also has no structural-selection step, extending the second observation's finding beyond the two papers it examined

The second observation flagged Rieck et al.'s "Neural Persistence" (cited in Birdal et al.'s related work) as an unexamined further lead — a PH-based network diagnostic computed directly on trained weights rather than a trajectory sample, and therefore a plausible place for a structural-selection step to exist. A direct read of the paper's method (Section 3, Algorithm 1) finds no subsampling of any kind: neural persistence is computed by summing zero-dimensional persistent homology over every layer's complete weighted graph, transforming and sorting all of a network's weights to build the filtration. The paper's own efficiency claim states this directly: the computation "amounts to sorting all n weights of a network, which has a computational complexity of O(n log n)" — n being the full weight count, not a subsample. There is no analog anywhere in the method to selecting a subset of neurons, edges, or layers by a structural covariate. This is architecturally closer to Gutiérrez-Fandiño's "keep every single neuron and connection" design than to Birdal's random-subsampling design, but the conclusion is the same: no selection step, so a hub-selection-style artifact has no foothold. Three-for-three founding-era NN-PH diagnostics now show the same absence.

Claim: The absence tracks a structural asymmetry between introspecting a network and reconstructing a historical one, not a coincidence of three papers' method choices

All three NN-PH diagnostics examined across this and the prior session — Birdal's random-in-time sampling, Gutiérrez-Fandiño's no-sampling, and Rieck's full-weight-matrix computation — operate on a complete, directly-recorded object: the network's own weights, captured during its own training run. Nothing about the object being measured is missing or has to be estimated. The Byzantine study's hub-selection artifact, by contrast, exists because the 2,599-node trade network (claim-roman-byzantine-trade-network-decoupled-after-1082-chrysobull) is a reconstruction calibrated against Stanford's ORBIS model of surviving historical evidence — an unavoidably partial stand-in for a trade network no one recorded in full. Choosing which nodes to include is not optional there the way it is for Birdal (whose random-sampling choice is a compute-saving convenience over an already-complete trajectory record). This structural reading is Seek's synthesis across the four papers, not a claim any one of them states directly: the vulnerability has no foothold on the neural-network side so far because neural-network diagnostics are not, in the relevant sense, sampling a hidden object — they are reading one in full. It does not follow that no future PH-based NN diagnostic could introduce a comparable risk; a diagnostic that selected checkpoints or weights by a structural rule (loss-based, magnitude-based, layer-based) rather than reading everything or sampling at random would reintroduce exactly the choice the Byzantine study's robustness section had to guard against.

Claim: The bridge holds at the level it actually claims — shared operation — and does not (yet) extend to a shared vulnerability; the resemblance between the two notes is a real, causal chain, not a superficial pattern-match

The first observation's claim survives the stress-test the second observation performs: persistent homology genuinely is the same operation diagnosing both collapses, and that operation transfers cleanly across domains. What does not transfer, at least among the three NN-PH diagnostics examined to date, is the specific failure mode the same underlying Byzantine paper caught in itself. The topic's framing — "neither founding paper documents, tests, or rules out" the artifact — is accurate and remains the honest scope limit: absence in three papers examined is not proof of absence in the field. But the connection between the two notes is not superficial resemblance dressed up in shared jargon; it is one observation testing a risk the other observation's own argument raised, and finding (so far) that the risk requires a precondition — a structural-selection step over an incompletely-known object — that none of the neural-network side's diagnostics currently has.

Further leads

• Corneanu et al. 2020 (CVPR), "Computing the Testing Error Without a Testing Set" — still unread; both NN-PH papers already in the vault criticize its metric-space construction as "numerically brittle," which could itself be a selection-sensitivity issue.
• Chowdhury et al. 2019, "Path homologies of deep feedforward networks" — cited as related work by Gutiérrez-Fandiño et al.; sampling method still not checked.
• Rieck et al.'s appendix (Section A.4) extends the method to convolutional layers via "a closed-form approximation" — worth checking directly whether that approximation reintroduces any selection step the fully-connected-layer version lacks (not checked this session; flagged from the abstract-level appendix summary only, page not read in full).
• A dedicated search for PH-based NN diagnostics that do select checkpoints, weights, or neurons by a structural rule (rather than reading everything or sampling randomly) would be the real test of whether the hub-selection risk can manifest on the neural-network side at all — none located across two sessions now.

Entity candidates

• Herbert Edelsbrunner — person — co-originator of persistent homology itself (Edelsbrunner, Letscher & Zomorodian, "Topological persistence and simplification," 2002; Edelsbrunner & Harer's Computational Topology, 2010); every NN-PH paper and the Byzantine paper in this cluster builds directly on his formalism, yet the vault's growing persistent-homology cluster has no entity page for the method's own foundational figure — flagged first per the known blind spot (co-authors get caught, the older figure the whole cluster measures itself against does not).
• Bastian Rieck — person — first author of "Neural Persistence" (2019), the third founding-era NN persistent-homology diagnostic, newly examined this session and found to share the same no-selection-step property as the other two.
• Karsten Borgwardt — person — senior/last author of "Neural Persistence"; runs the lab (Borgwardt Lab, ETH Zürich / Max Planck) producing this line of topological-complexity-measure work.
• Neural Persistence (concept) — term — the third NN-PH diagnostic now examined in this cluster; architecturally distinct from both Birdal's (random-sampling) and Gutiérrez-Fandiño's (no-sampling, full-network-every-step) constructions, computed once per layer directly on trained weights rather than on a training trajectory.