scientific article; zbMATH DE number 1302143
Although recent work has shown that dual geometry can be used outside of exponential families, we only consider exponential families and particular subsets called curved exponential families. A random variable \(Y\) having density from an exponential family is usually written \(p(y\mid\eta) =\text{exp} \{y^T\eta-\psi(\eta)\}h(y)\), with \(h(y)\) usually being Lebesgue or counting measure. Two important sets associated with an exponential family are the natural parameter space \[ N =\{\eta\in\mathbb{R}^k:\int \text{exp}(y^T\eta)h(y)d\nu(y)<\infty\}, \] and the expectation space \[ M=\{\mu=\int yp(y\mid\eta)d\nu(y):\eta\in N\}. \] For a given \(\psi\) and \(h\), we consider subsets of the full exponential family \(S=\{p(y\mid\eta):\eta\in N\}\). The subset \(S_0\) is a one-dimensional curved exponential if there is an open subset \(\Theta\subset\mathbb{R}\) and a function \(\eta(\cdot) :\Theta\to N\) such that (1) \(\eta(\cdot)\) is infinitely differentiable, one-to-one, and rank 1, and (2) a sequence of points \(\eta_n\in N_0=\eta(\Theta)\) converging to \(\eta_0\in N_0\) implies the pre-image of this sequence converges to the pre-image of its limit in \(\Theta\). Higher dimensional curved exponential families can be defined in a similar manner, but we restrict ourselves to the one-dimensional case for simplicity. Under these conditions, according to standard geometrical terminology, \(\theta\to\eta(\theta)\) is said to be an imbedding and \(N_0\) is an imbedded submanifold in \(N\), and thus one also may call the family \(S_0\) an imbedded subfamily of \(\mathcal S\). For the entire collection see [Zbl 0914.00073].
This page was built for publication:
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4247408)