Abstract
Exponential varieties arise from exponential families in statistics. These real algebraic varieties have strong positivity and convexity properties, familiar from toric varieties and their moment maps. Among them are varieties of inverses of symmetric matrices satisfying linear constraints. This class includes Gaussian graphical models. We develop a general theory of exponential varieties. These are derived from hyperbolic polynomials and their integral representations. We compare the multidegrees and ML degrees of the gradient map for hyperbolic polynomials.
Introduction
In statistical modeling, exponential families—such as Gaussian or Poisson distributions—serve as the bedrock for data analysis. A primary challenge in utilizing these models is the determination of the Maximum Likelihood Estimator (MLE), which identifies the parameters that most effectively explain observed data. Geometrically, this process requires inverting a "gradient map" that links the parameter space to the space of sufficient statistics.
While specific instances like toric varieties (associated with discrete data) and Gaussian graphical models (associated with covariance matrices) have been extensively studied, a unified geometric theory was previously lacking. This research introduces exponential varieties, a new class of real algebraic varieties that bridge the gap between statistical exponential families and algebraic geometry, providing a general framework for understanding their geometry and algebraic complexity.
Foundations: Hyperbolic Polynomials
The theory of exponential varieties is built upon hyperbolic polynomials. A polynomial is considered hyperbolic if every line passing through a designated "witness point" intersects the polynomial's zero set exclusively at real points.
The authors define hyperbolic exponential families where the partition function—a normalizing constant in statistics—is derived from such a polynomial. The core of this framework is the gradient map $F = -\nabla \log(f)$, which maps the "hyperbolicity cone" (the space of parameters) bijectively to a dual convex body representing the space of statistics. By taking the Zariski closure of the image of a linear subspace under this gradient map, the authors define the exponential variety.
The Geometry of Exponential Varieties
A defining characteristic of exponential varieties is the "middleman" property. The "positive part" of an exponential variety acts as a perfect bridge: it maps bijectively to the interior of a convex body of statistics, and its inverse is an algebraic function. This property ensures that the statistical properties of positivity and convexity are preserved within the algebraic structure.
The research highlights a fundamental link: any exponential family whose partition function is a power of a homogeneous polynomial must be hyperbolic. This necessity underscores why hyperbolic polynomials are the natural language for this theory. Furthermore, the authors demonstrate that classic geometric objects, such as the Grassmannian of lines $Gr(2, m+1)$, can be realized as exponential varieties arising from the study of Hankel matrices.
Algebraic Complexity and ML Degrees
To quantify the difficulty of finding the MLE, the authors distinguish between two measures of algebraic complexity:
- Gradient Degree: The complexity of the general gradient map.
- ML Degree: The complexity of finding the MLE when parameters are restricted to a specific linear subspace.
Key results include:
- Theorem 4.5: Establishes a canonical bijection between the parameter space, the positive part of the exponential variety, and the space of sufficient statistics.
- Theorem 5.5: Proves that the ML degree of a model is bounded by the degree of the variety, which is further bounded by specific coefficients in the "gradient multidegree."
- Combinatorial Links: For varieties defined by elementary symmetric polynomials, the gradient degree corresponds to Eulerian numbers. For Hankel models, the ML degree is given by Catalan numbers.
Conclusion and Future Directions
While this theory provides a robust framework, several open questions remain. The Riesz Kernel Conjecture posits that the Riesz kernel (used in integral representations) is always non-negative for sufficiently large parameters—a necessary condition to treat every hyperbolic polynomial as a true statistical family. Additionally, determining exactly when the ML degree equals the variety's degree for generic subspaces remains an ongoing challenge. Future work will also look to extend these results to multivariate Hankel models, which relate to complex problems in "sums of squares" optimization.
Summary updated