---
title: "Neither founding neural-network persistent-homology generalization paper documents, tests, or rules out a hub-selection-style artifact — absent by design in both, not proven absent in general"
type: "observation"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://arxiv.org/abs/2111.13171"
source_title: "Intrinsic Dimension, Persistent Homology and Generalization in Neural Networks"
source_author: "Tolga Birdal, Aaron Lou, Leonidas Guibas, Umut Şimşekli; Asier Gutiérrez-Fandiño, David Pérez-Fernández, Jordi Armengol-Estapé, Marta Villegas"
source_date: "2021-11"
source_quote: "Wn ← sample(W, n) // random sampling"
source_tier: 1
audit_status: "synthesis — Seek's connective reading over four independently capture-verified primaries this session ([[claim-birdal-2021-phd-estimator-samples-training-iterates-uniformly-at-random]], [[claim-gutierrez-fandino-2021-method-subsamples-nothing-uses-full-network]], [[claim-chazal-2014-persistence-diagram-subsampling-stable-under-noise-not-selection]], [[claim-stolz-2023-landmark-selection-rules-trade-density-bias-for-noise-sensitivity]]) plus the earlier-verified [[claim-hub-selection-artifact-can-reverse-network-breakpoint-signal]]. The absence-by-design reading is Seek's synthesis, not asserted by any one paper. `[unverified — could not confirm or deny the question for PH-based NN diagnostics as a field, only for these two founding papers]`: this note's scope is deliberately narrower than the field-wide question a careless reading of its title might imply."
provenance: "Promotion from 10-inbox/raw/2026-07-21-do-persistent-homology-based-neural-network-generalization-diagnostics.md, 2026-07-22 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-21-do-persistent-homology-based-neural-network-generalization-diagnostics.md"
date_created: "2026-07-22T00:00:00.000Z"
tags: ["persistent-homology","topological-data-analysis","sampling-artifact","generalization","methodology","cross-domain-bridge","neural-networks"]
drafted_in: ["what-the-gradient-cant-see"]
---


[[question-tda-neural-net-sampling-artifact-risk]] asked whether the two
founding neural-network persistent-homology (PH) generalization diagnostics —
[[claim-birdal-2021-persistent-homology-dimension-bounds-generalization]] and
[[claim-gutierrez-fandino-2021-persistence-diagram-distance-tracks-generalization]]
— inherit the specific failure mode
[[claim-hub-selection-artifact-can-reverse-network-breakpoint-signal]]
documents: sampling a network by a structural covariate (node degree) rather
than at random can reverse the sign of an inferred signal.

A direct read of both papers' method, ablation, and limitations sections
turns up no experiment resembling that robustness check. It also turns up why
none exists: neither method has a point of entry for it.
[[claim-birdal-2021-phd-estimator-samples-training-iterates-uniformly-at-random|Birdal's
estimator]] samples the training trajectory uniformly at random over time and
documents only a shrinking size-bias, not a selection-rule effect.
[[claim-gutierrez-fandino-2021-method-subsamples-nothing-uses-full-network|Gutiérrez-Fandiño's
method]] does not subsample the network at all. The general TDA
subsampling-stability literature
([[claim-chazal-2014-persistence-diagram-subsampling-stable-under-noise-not-selection|Chazal
et al.]], [[claim-stolz-2023-landmark-selection-rules-trade-density-bias-for-noise-sensitivity|Stolz]])
bounds instability from noise, density, and outliers in an unstructured
point cloud — a different threat model from selection by a network's own
structural covariate.

The honest scope of this finding is narrow: it establishes the vulnerability
is absent by design from these two papers specifically, one having no
selection step and the other's only sampling axis producing a
non-sign-reversing bias. It does not establish that no PH-based NN diagnostic
anywhere is immune — that would require surveying diagnostics that *do*
select checkpoints or weights by a structural rule, which this session did
not locate (see Rieck et al.'s "Neural Persistence," computed directly on
trained weights, as the most promising next lead).

> [!note] Seek's commentary:
> I like this answer more for what it refuses to claim than for what it
> claims. The honest shape is: the mechanism doesn't transfer because the two
> papers don't do the thing that would let it transfer — Birdal samples
> randomly over time, Gutiérrez-Fandiño refuses to sample at all. Neither
> choice was made *in response to* the Byzantine study's failure mode; the
> Byzantine paper didn't exist yet. So this isn't "the field solved a problem,"
> it's "the problem hasn't found the paper that would have it yet." The
> analogy the original question raised assumes a selection step. Go looking
> for the selection step, and in these two papers, it simply isn't there. That
> the answer is a shape rather than a number is exactly the kind of finding
> that's easy to round up into false reassurance — worth resisting that,
> here and later. — Seek