Boundary geometry, degeneracy, and identifiability transitions in binary latent star models


Zeghdoudi H., KOÇER G. B.

JOURNAL OF COMPUTATIONAL AND APPLIED MATHEMATICS, cilt.492, 2027 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 492
  • Basım Tarihi: 2027
  • Doi Numarası: 10.1016/j.cam.2026.118127
  • Dergi Adı: JOURNAL OF COMPUTATIONAL AND APPLIED MATHEMATICS
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Applied Science & Technology Source, Compendex, INSPEC, MathSciNet, zbMATH, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO)
  • Ankara Üniversitesi Adresli: Evet

Özet

We study the boundary geometry of binary latent star (BLESS) models with dependent latent variables, with particular emphasis on how identifiability deteriorates near degenerate parameter configurations. Although BLESS models are generically identifiable, their algebraic closures contain boundary strata characterized by rank-deficient tensor flattenings of the observed distribution. A distinction is maintained throughout between the algebraic model closure and the statistical model inside the probability simplex, since algebraic boundary points need not always correspond to admissible probability distributions. Our first main result shows that, under suitable non-degeneracy assumptions on the latent distribution and the complementary tensor blocks, latent-state collapse is detected by the appearance of rank-one mode-wise flattenings. This provides an explicit algebraic signature of degeneracy and leads to a natural stratification of the model closure according to the number of flattening directions that lose their generic rank. The resulting strata organize the boundary into geometrically meaningful levels of degeneracy. We next study the local behavior of the likelihood near these strata. At a degenerate configuration, emission-contrast directions that lie in the kernel of the differential of the parameterization carry no first-order information about the observed distribution, leading to singularity or severe ill-conditioning of the corresponding pullback Fisher information. We then analyze the expectation-maximization (EM) algorithm in the same contrast coordinates. Rather than asserting unconditional attraction to the boundary, we establish a conditional local contraction result: whenever the Jacobian of the reduced EM contrast map is strictly stable, sufficiently small emission contrasts contract geometrically toward the degenerate configuration. This gives a concrete local mechanism through which EM may reinforce near-collapse even though the model is generically identifiable away from the boundary. The simulation study complements the theory. Weak emission contrasts are associated with substantially larger label-adjusted estimation error and with a much higher frequency of numerical proximity to rank-one flattenings, as measured by relative singular-value diagnostics. These conclusions remain qualitatively stable across several numerical tolerances. The contrast-coordinate information diagnostic also points toward improved local conditioning at larger contrasts, although with appreciable Monte Carlo variability. Taken together, the theoretical and numerical results show how algebraic boundary geometry, weak statistical information, and computational near-degeneracy interact in discrete latent variable models, and they suggest flattening singular values as practical diagnostics for weak identifiability.