Natural exponential families and Lie algebras (Q2563406)

Let us recall first what is called a natural exponential family of probabilities. If \(\mu\) is a positive measure on \(\mathbb{R}^d\), and \(L_\mu\) its Laplace transform, \(L_\mu(\theta)=\int_{\mathbb{R}^d}e^{\langle\theta,x\rangle}\mu(dx)\), the family \(F\) of probabilities \[ P(\theta,\mu)(dx)={1\over L_\mu(\theta)}e^{\langle\theta,x\rangle}\mu(dx) \] is called a natural exponential family. The mean \(m\) of \(P(\theta,\mu)\) is given by \(m=\int_{\mathbb{R}^d} xP(\theta,\mu)(dx)\), and the map \(\theta\mapsto m\) is one to one. One writes \(P(\theta,\mu)(dx)=f_\mu(x,m)\mu(dx)\) and one considers the variance \(V_F\) as a function of \(m\), \[ V_F(m)=\int_{\mathbb{R}^d}(x-m)^t(x-m)f_\mu(x,m)\mu(dx). \] The natural exponential family \(F\) is said to be quadratic if \(V_F\) is a second order polynomial in \(m\), and simple quadratic if the top term of \(V_F(m)\) is a scalar multiple of \(\| m\|^2\). To such a family \(F\) the author associates a family of polynomials in \(d\) variables constructed from the Taylor expansion of the function \(m\mapsto f_\mu(x,m)\) at a point \(m=0\). The property for these polynomials to be orthogonal with respect to \(\mu\) is related to the property for the family \(F\) to be quadratic. These polynomials satisfy recurrence relations, which lead to the introduction of a Lie algebra of operators, whose properties reflect properties of the family \(F\). Then the author studies the inverse problem: to construct, from a Lie algebra \(\mathcal A\), a natural exponential family \(F\), the Lie algebra \(\mathcal A\) belonging to a a class of Lie algebras of operators acting on a space of polynomials in \(d\) variables. By using the Campbell-Hausdorff formula one constructs a positive definite function on \(\mathbb{R}^d\), and a measure \(\mu\) by the Bochner theorem.