Abstract
Many statistical hypotheses can be formulated in terms of polynomial equalities and inequalities in the unknown parameters and thus correspond to semi-algebraic subsets of the parameter space. We consider large sample asymptotics for the likelihood ratio test of such hypotheses in models that satisfy standard probabilistic regularity conditions. We show that the assumptions of Chernoff’s theorem hold for semi-algebraic sets such that the asymptotics are determined by the tangent cone at the true parameter point. At boundary points or singularities, the tangent cone need not be a linear space and limiting distributions other than chi-square distributions may arise. While boundary points often lead to mixtures of chi-square distributions, singularities give rise to nonstandard limits. We demonstrate that minima of chi-square random variables are important for locally identifiable models, and in a study of the factor analysis model with one factor, we reveal connections to eigenvalues of Wishart matrices.
Introduction
In classical statistical inference, Wilks’ Theorem serves as a cornerstone, asserting that the likelihood ratio test (LRT) statistic asymptotically follows a chi-square ($\chi^2$) distribution. However, this result relies on the assumption that the null hypothesis is a smooth manifold. In practice, many common models—including factor analysis, hidden variable models, and those with inequality constraints—are defined by polynomial equations and inequalities. These define semi-algebraic sets, which frequently possess singularities or boundary points where the smooth manifold assumption fails.
At these "rough" points, the standard $\chi^2$ approximation breaks down, potentially leading to incorrect $p$-values and flawed scientific conclusions. This paper bridges the gap between statistics and algebraic geometry to rigorously define the asymptotic behavior of the LRT at such problematic points.
Geometric Foundations of Hypotheses
To address the breakdown of Wilks' Theorem, we must characterize the local geometry of the parameter space. Instead of a flat tangent space (like a tangent plane to a sphere), we utilize the tangent cone. The tangent cone captures all possible directions one can move away from a point while remaining within the hypothesis set.
As shown in Figure 1, a singularity such as a self-intersection results in a tangent cone that is not a linear space. The paper proves that semi-algebraic sets are "everywhere Chernoff-regular," a technical property ensuring that the LRT statistic converges to the squared Mahalanobis distance between a multivariate normal draw and this tangent cone.
Singularities and Nonstandard Limits
While boundary points often lead to "chi-bar-square" distributions (mixtures of $\chi^2$), singularities can lead to entirely different limits. For instance, in models that are only "locally identifiable," the LRT limit is often the minimum of several independent $\chi^2$ variables. This reflects the model's ability to "see" multiple possible parameter directions simultaneously at the singularity.
Figure 2 illustrates a cusp, another type of singularity where the tangent cone is restricted (e.g., to a half-line). In such cases, the limiting distribution deviates significantly from the standard $\chi^2$ form. For complex models where the true tangent cone is difficult to compute, the paper introduces the "algebraic tangent cone" (derived via Gröbner bases) to provide provable lower bounds for asymptotic $p$-values.
Case Study: One-Factor Analysis
The theoretical framework is applied to the one-factor analysis model. A major singularity exists in this model when the off-diagonal entries of the covariance matrix are zero. While standard software often assumes a $\chi^2$ distribution, this research proves that near these singularities, the true distribution is related to the eigenvalues of $2 \times 2$ Wishart matrices. Specifically, the LRT statistic converges to a distribution involving the smaller of two eigenvalues. This insight allows for the calculation of conservative $p$-values using the larger eigenvalue, offering a practical path for testing in geometrically complex spaces.
Conclusion and Future Work
This work establishes that for semi-algebraic hypotheses, the LRT statistic converges to the squared distance to the tangent cone (Theorems 2.6 & 4.3). While the one-factor case is resolved, the geometry of multi-factor models remains significantly more complex. Future research may focus on the validity of bootstrap procedures, as the limiting distributions at singularities often depend on unknown nuisance parameters, making them non-pivotal. Furthermore, the computational cost of finding exact tangent cones in high-dimensional models remains a challenge for symbolic algebra.
Summary updated