nucleation

CLAIM-135: Recovering latent structure from relational data is three achievements, reconstruction, identification up to stated ambiguities, and constitution of a communicative kind, and identifiability theory reaches the second at most

6 min read

Properties13
StatusActive
Version3
Rolethesis
Warrantinterpretive
Tagsphilosophy-of-science, mathematical-statistics, individuation
Linepragmatic-transport
Bears onQUESTION-001 — To what extent can an utterance's pragmatic identity be reconstructed from a globally consistent collection of context-relative judgements, and what fails when it cannot?
GroundsCASE-042 — Collaborative filtering and topic models: latent relational structure recovered from interaction matrices, up to gauge, or uniquely under anchor conditions, THEORY-032 — The Yoneda lemma determines an object only up to isomorphism, from its whole hom-functor, among objects the category already has; as a formal version of "an object is determined by its relations" it presupposes the relata
SupportsCLAIM-139 — A communicative kind's identity is grounded in its relational organization rather than in any expression or encoding; statistical methods recover aspects of that organization from restricted observations, translation transports it, and fidelity is which independently identifiable, decision-relevant relational distinctions persist, CLAIM-206 — If a communicative kind can outrun every probe, human same-kind judgements, themselves one probe family, can neither confirm nor defeat the October thesis; if it cannot, the kind does in the tests only what an observational class does, which the record's term for a kind says it is not, CLAIM-204 — A communicative kind's identity is grounded in its relational organization, for which competent participants' uptake is the standard; statistical methods recover aspects of it from restricted observations; translation transports it; and what persists is what the receivers can still recover, of which the analyst can certify the independently identifiable part, CLAIM-454 — A latent feature or mode recovered from an organization's members' activity is at most a constituent of an organizational perspective, not the perspective: a feature allocation says which members take part in which modes, and whether those modes are organized into one perspective depends on the relations of access, uptake and authority among their carriers, which the same allocation leaves open
ComplementsCLAIM-150 — An ontology of intrinsically individuated objects with relational properties can reproduce the categorical and operational identifications, and the predictions, of a relational one, so identification underdetermines constitution, CLAIM-209 — Probes and statistical models recover only the distinctions identifiable from the relational evidence they are given, and a signature sufficient for a stated family of communicative decisions outside the observed probes can be identified when the full latent factorization is not
UsesTERM-014 — communicative object, as an observational equivalence class, TERM-040 — communicative kind, as a repeatable relational organization that numerically distinct events realize, distinct from a token and from a structural position
Illustrated byCASE-040 — Topic and act crossed: a factorial design that decorrelates what a text is about from what it does
Cited as “rests on” byCLAIM-543 — What separates a language model's distribution over personas from a distribution over organizational perspectives is not whether the latent model is hierarchical, since a mixture or an LDA-style admixture already is, but whether some latent components condition which others are available or consequential: a model whose components are exchangeable given their weights is organizationally flat however many levels its prior has, and dependencies among components count as organization (organization as relational arrangement) only where interventions, not fit, support them

Graph View

The claim

A214 §1 sets out three questions: whether a low-dimensional structure explains the observed relationships, whether it can be recovered uniquely “apart from transformations that do not affect its identity”, and whether the recovered structure does “remain meaningful when we change the observational system, representation, or causal conditions”. Then: “Matrix factorization addresses the first. Identifiable factorization and topic-model theory address the second under specified assumptions. Our manuscript is principally concerned with the third.”

A214 names three layers: “X: observed relational evidence, (W,H): latent explanatory representation, 𝒦: structural kinds posited by the explanation”. And two warrants: “The transition from good reconstruction to identified structure requires assumptions and evidence. The transition from identified structure to real relational kind requires an additional scientific and philosophical argument.”

The first two steps are A207’s. A207 §5 had distinguished “Observed relational profile / Model-imputed relational profile / Structurally identifiable relational profile”, and said “The third requires additional assumptions and evidence.” A214 added the third step, from identified structure to kind. A218 §5 keeps the distinction and splits it finer: “Observed relational evidence. Latent explanatory representations. Statistically identifiable structure. Task-sufficient structural information. Philosophically proposed constitutive relations.”

The record holds the three as distinct achievements, each with its own warrant.

  1. Reconstruction is warranted by fit, on held-out entries.
  2. Identification up to stated ambiguities is warranted by a theorem and its conditions: a stated gauge (CLAIM-148), and conditions such as separability or anchor words for topic models (CASE-042). Identifiability theory gives sufficient conditions for this step and reaches no further.
  3. Constitution of a communicative kind (TERM-040) needs evidence that the identified structure is what the act consists in, not only what one family of observations reveals of it. That evidence must come from probes chosen for their bearing on the act, and from an argument about what the act is (CLAIM-150).

The second step is relative to the probes in the way TERM-014 says an observational class is: two items are indistinguishable relative to a probe set and a model class. The third step is not relative in that way. A kind can be real and still unseparated by an insufficient family of probes. A218 Ch7: “A structural kind may be real while remaining indistinguishable under an insufficient family of empirical probes.”

THEORY-032 supplies the formal limit on the third step. Even a complete relational profile, the whole hom-functor, determines an object only up to isomorphism and presupposes the relata. A data matrix is less than that, since it has no composition.

QUESTION-001 asks how far pragmatic identity can be reconstructed from context-relative judgements. This claim answers part of it. Even a reconstruction that is consistent and identified falls short of the kind until the third warrant is given.

Completion is not identification

The first step can succeed while the second fails, and A207 §5 gives the mechanism. “A sparse matrix can often admit multiple completions consistent with every observed entry. Those completions may disagree on the similarity of two items”. So “a latent representation may look stable because the model’s regularization selected one completion, not because the observed relational evidence uniquely determined it.”

The theory here is standard. A rank-k m×n matrix has k(m+n−k) degrees of freedom, so fewer observed entries cannot identify it. With more, the sampling pattern and the matrix’s incoherence decide whether the completion is unique. Candès and Recht’s guarantee for exact completion (doi:10.1007/s10208-009-9045-5; not held in either record) assumes entries sampled uniformly at random. Missing-not-at-random exposure violates exactly that, and A207 §5 names it: it “can make learned factors reflect selection patterns rather than preferences or communicative distinctions.” Randomized assignment of realizations to interpreters restores the standard conditions. Active querying improves the design but guarantees nothing.

One pointer bears on A207’s own proposal and is not worked out here. A207’s judgement data form a four-way tensor Y_uiac. A CP decomposition of a tensor of order three or more is essentially unique, up to scaling and permutation, under Kruskal’s rank condition. So unlike the matrix case, its components can be identifiable without nonnegativity. A Tucker decomposition keeps a gauge on each mode. A207 calls tensor factorization “not a novel contribution”, which is true, but its bearing on the second step goes unsaid.

What it does not say

It does not say identified factors are never kinds. It says only that identifiability does not make them kinds. It does not claim the second step is open: for topic models and latent-structure models there are sufficient conditions (Donoho and Stodden; Arora, Ge and Moitra; Allman, Matias and Rhodes, arXiv:0809.5032). None of these is held in either record. It does not say the third step is out of reach. It says the third step needs evidence of a different kind from the second, and that neither A214 nor A218 supplies it.

Note of 2026-10-26: three latent-structure models, and the second achievement

The exchange “Experimental Infrastructure Review” proposed hLDA, the author–topic model and the mixed membership stochastic blockmodel for inferring an organization’s perspectives from its members’ activity (QUESTION-276). All three were read on 2026-10-26, and none reaches this claim’s second achievement with a result.

  • hLDA (NOTE-1176) has no identifiability result. Its one check is recovery of trees from data simulated under its own prior (THEORY-1010, eight of ten), and its own caveat is that recovery depends on “how identifiable the topics are”.
  • The author–topic model (NOTE-1180) names only the permutations of topics. Its likelihood leaves a continuum of changes of topic basis free (THEORY-972).
  • The blockmodel (NOTE-1179) says in its §5 that the difference between the block matrix and the memberships “practically eliminates” the identifiability issue, without argument, and the likelihood given the memberships contradicts it (THEORY-993). Whether the Dirichlet prior on memberships separates them in the marginal likelihood is left open. Allman, Matias and Rhodes, named under “What it does not say” and still not held, is the kind of result that question needs.

So each stops at the first achievement, and the third is not attempted. The author–topic paper’s reading of Enron email as showing “the roles and interests of different individuals” is the step from structure to kind taken by inspection (CLAIM-454). The claim is unchanged.

Backlinks