Paperlane

no. 17 · Wednesday 18 November 2026

Identifiability of latent variable models with flexible latent distributions

Chengyu Cui and Gongjun Xu

math.ST

Abstract

Latent variable models are popularly used to characterize dependence among many observed variables through underlying latent variables. Recent advances in statistics and machine learning methods have encouraged growing interest in studying latent variable models with flexible latent distributions that capture complex latent structures. This flexibility, however, raises a fundamental question of whether model parameters can be identified when the latent distribution is unknown and must itself be learned. In this paper, we show that exact identifiability fails almost everywhere, in the sense that distinct but arbitrarily close pairs of model parameters and latent distributions can yield exactly the same distribution for the observed variables. Nonetheless, the resulting identification ambiguity is small. In particular, we prove that the identification ambiguity, measured by the diameter of the set of parameters inducing the same marginal distribution, vanishes at a superpolynomial rate of $\exp\{-c(\log p)^2\}$ as the number of observed variables $p$ grows. Building on this result, we show that the failure of exact identifiability need not preclude consistent estimation or valid inference of the model parameters and latent distribution. These results suggest that, when many variables are observed, the identifiability issue is far less of an obstacle to estimation and inference in latent variable models than is commonly expected.

fields
Statistics Theory
pdf ↗ page arXiv:2611.23052 ↗