Abstract
Stein’s method provides a way of bounding the distance of a probability distribution to a target distribution $\mu$. Here we develop Stein’s method for the class of discrete Gibbs measures with a density $e^V$, where $V$ is the energy function. Using size bias couplings, we treat an example of Gibbs convergence for strongly correlated random variables due to Chayes and Klein [Helv. Phys. Acta 67 (1994) 30–42]. We obtain estimates of the approximation to a grand-canonical Gibbs ensemble. As side results, we slightly improve on the Barbour, Holst and Janson [Poisson Approximation (1992)] bounds for Poisson approximation to the sum of independent indicators, and in the case of the geometric distribution we derive better nonuniform Stein bounds than Brown and Xia [Ann. Probab. 29 (2001) 1373–1403].
Introduction
In probability theory, quantifying the distance between a complex distribution and a simpler target distribution is a fundamental challenge. Stein’s method has long been the gold standard for this task, particularly for Normal and Poisson approximations. However, its application to more complex, interacting systems—specifically those described by Gibbs measures—has historically been less developed.
Gibbs measures are central to statistical mechanics, describing systems where particle interactions are governed by an energy function $V(k)$. This paper extends Stein’s method to the broad class of discrete Gibbs measures, providing a rigorous framework to analyze interacting particle systems and deriving sharper error bounds for several classical distributions.
Methodology: The Generator Approach
The authors utilize the generator approach to Stein’s method, which views the target Gibbs distribution $\mu$ as the stationary distribution of a Markov process. Specifically, they employ a birth-death process.
The Stein Operator
By defining an operator $\mathcal{A}$ such that for a random variable $X$, $E[\mathcal{A}g(X)] = 0$ if and only if $X$ follows the distribution $\mu$, the authors create a functional tool to measure "distance." If $E[\mathcal{A}g(W)]$ is close to zero for a broad class of functions $g$, then the distribution of $W$ is close to $\mu$.
Size-Bias Coupling
To manage the dependencies inherent in Gibbs measures, the paper integrates size-bias coupling. This technique transforms a random variable $X$ into $X^*$ such that $E[Xf(X)] = E[X]E[f(X^*)]$. The authors demonstrate that size-biasing is not merely a specialized trick for Poisson variables but a fundamental property of any distribution where the "death rate" of the system is proportional to the number of individuals. This coupling fits naturally with birth-death processes where the death rate is linear (unit per capita).
Comparison of Generators
A key innovation in this work is the ability to bound the distance between two distributions by directly comparing their birth and death rates. This allows for the comparison of discrete lattice models to continuous limits without the need to calculate often-intractable normalizing constants, known as partition functions.
Key Insights and Theoretical Refinements
By treating the Poisson and Geometric distributions as specific instances of Gibbs measures, the authors derived bounds that are sharper than those found in literature dedicated solely to those distributions:
- Poisson Approximation: The authors refined the "magic factor" for Poisson approximation to $\frac{1}{\lambda}(1 - e^{-\lambda})$, a meaningful improvement over the standard $1 \wedge 1/\lambda$ established by Barbour, Holst, and Janson (1992).
- Geometric Distribution: The paper derives new non-uniform Stein bounds for the geometric distribution, demonstrating that the solution to the Stein equation is bounded by $1/p$, improving upon the results of Brown and Xia (2001).
- Lattice-to-Continuum Bridge: The framework provides a quantitative rate for when a discrete gas on a grid begins to behave like a continuous gas. This offers a rigorous mathematical foundation for assumptions frequently made in statistical physics.
Applications in Statistical Physics
The utility of this framework is demonstrated through the Chayes-Klein model, which involves strongly correlated random variables. The authors provide the first explicit rates of convergence for this model to the grand-canonical Gibbs ensemble.
Furthermore, the paper introduces a general theorem (Theorem 3.5) that allows researchers to bound the Total Variation (TV) distance between two Gibbs measures simply by examining the ratio of their birth rates. This simplifies the process of comparing complex interacting systems significantly.
Limitations and Future Directions
While the framework is powerful, it faces certain limitations. The general Gibbs bound for the Binomial distribution was found to be $O(e^{n \ln n})$, which grows too rapidly for many practical applications; in this specific case, specialized Binomial methods remain superior. Additionally, comparing distributions with different supports (e.g., finite vs. infinite) remains a technical hurdle.
Future work aims to extend these discrete measures on non-negative integers ($\mathbb{N}_0$) to higher-dimensional spaces, including spatial Gibbs measures and Gibbs point processes, which would further bridge the gap between abstract probability and complex physical simulations.
Summary updated