Standard statistical theory relies on a comforting assumption: that our models are "smooth." We typically imagine the space of possible parameters as a flat sheet or a gentle curve—a manifold. In this smooth world, if you zoom in closely on any point, the model looks like a linear subspace. This linearity is the engine behind the classical Likelihood Ratio Test (LRT) and the ubiquitous $\chi^2$ distribution used to calculate p-values.
However, modern models—including neural networks, mixture models, and latent variable models—are rarely smooth. They are singular. They contain "pinched" points, self-intersections, and sharp corners where the standard rules of calculus and statistics break down.
The Failure of the Smoothness Assumption
In a regular (smooth) model, the Fisher information matrix is invertible, and the local geometry is a flat tangent space. But as Kenji Fukumizu explains in Statistical Estimation in Singular Models, many models used in machine learning possess singularities where the Fisher information matrix becomes non-invertible.
Consider a Gaussian mixture model. If you try to determine if a dataset comes from one component or two, you are testing a hypothesis at a singularity. When the weights of the second component go to zero, or when the means of two components merge, the model’s geometry doesn't just flatten out; it intersects itself or hits a boundary. At these specific coordinates, the model is no longer a smooth manifold, and the standard $\chi^2$ approximation for the LRT statistic fails, as noted in Algebraic Techniques for Gaussian Models.
From Tangent Spaces to Tangent Cones
To fix this, researchers have turned to algebraic geometry. If a model is defined by polynomial constraints—what Seth Sullivant calls Algebraic statistical models—we can describe its shape using the language of algebraic varieties.
The key conceptual shift is replacing the tangent space with a tangent cone.
- Tangent Space: A linear approximation (like a flat plane) that assumes you can move in any direction symmetrically.
- Tangent Cone: A more general set of all "feasible" directions at a point.
At a singularity, the tangent cone might look like a "V" shape or a cross. As Likelihood ratio tests and singularities demonstrates, the asymptotic distribution of the LRT is actually determined by the shape of this cone. Instead of a simple $\chi^2$ distribution, the test statistic behaves like the distance from a random variable to this cone. This explains why standard p-values are often "wrong" for complex models: they are measuring distance to a flat plane when they should be measuring distance to a complex, non-convex cone.
Identifying the Singular Locus
Identifying where these problems occur is a major focus of recent work. Han Xiao and colleagues have investigated the Smoothness of Gaussian conditional independence models, showing that even simple constraints on correlations can create geometric singularities like self-intersections.
In latent variable modeling, Yunxiao Chen and Irini Moustaki point out in A Note on Likelihood Ratio Tests for Models with Latent Variables that these singularities are not just theoretical curiosities—they are common in factor analysis and random effects models. When parameters sit on the boundary of the allowed space (e.g., a variance being zero), the model is singular, and the "standard" p-values reported by software are fundamentally flawed.
Robust Inference in a Singular World
If the standard tools fail, how do we perform valid inference? Several paths have emerged:
- Geometric Correction: By characterizing the specific geometry of the tangent cone, researchers can derive non-standard asymptotic distributions that replace the $\chi^2$ test.
- Algebraic Complexity: Rosa Winter and Yuhan Jiang show in Linear Spaces of Symmetric Matrices with Non-Maximal Maximum Likelihood Degree that the "complexity" of solving for a maximum likelihood estimator is tied to the model's algebraic structure, providing a map of where estimation becomes unstable.
- Asymptotic Bayesian Frameworks: To bypass the difficulties of the likelihood surface, Ryohei Fujimaki and Satoshi Morinaga developed Factorized Asymptotic Bayesian (FAB) inference. This method remains valid even when singularities are present, allowing for consistent model selection in mixture models by accounting for the non-regularity of the parameter space.
By treating statistical models as geometric objects rather than just sets of equations, this research direction provides the tools necessary to navigate the "non-smooth" reality of modern data science.