Random partitions by semigroup methods (Q1300544): Difference between revisions

Let \({\mathcal P}\) be the set of all partitions of \(\mathbb{N}\). For \(U,V \in {\mathcal P}\), the maximum partition \(U\vee V\in {\mathcal P}\) is defined by ``joining'' \(U\) with \(V\). Then \(({\mathcal P},\vee)\) is an Abelian idempotent semigroup with neutral element \(U_0=\{\{j\}: j\in\mathbb{N}\}\) and absorbing element \(U_\infty= \{\mathbb{N}\}\). Each \(U\in {\mathcal P}\) may be identified with an element of the set \(\{0,1\}^{\mathbb{N} \times\mathbb{N}}\), namely \((j,k)\mapsto 1\) if \(j\) and \(k\) belong to the same class of \(U\), and \((j,k)\mapsto 0\) otherwise. With this identification, \({\mathcal P}\) inherits a topological structure as a subspace of the compact metric space \(\{0,1\}^{\mathbb{N} \times\mathbb{N}}\). Being closed, \({\mathcal P}\) becomes a compact metric space. Let \({\mathfrak B}(P)\) stand for the \(\sigma\)-algebra generated by the open subsets of \({\mathcal P}\). Further, for \(u\subset \mathbb{N}\), let \({\mathcal P}_u\) denote the set of all partitions \(V\in {\mathcal P}\) such that \(u\) is included in some class of \(V\). For \(U\in{\mathcal P}\), set \({\mathcal P}_U= \bigcap_{u\in U}{\mathcal P}_u\). An important result of the paper is as follows: Let \({\mathcal U}=\{U\in {\mathcal P}\): there are only finitely many nontrivial (i.e. having at least 2 elements) classes in \(U\}\). Then a function \(\varphi:(U,\vee)\to\mathbb{R}\) is positively defined and normalized (i.e. \(\varphi (U_0)=1)\) if and only if \(\varphi(U)= \mu({\mathcal P}_U)\), \(U\in {\mathcal U}\), for a (uniquely determined) probability measure \(\mu\) on \({\mathfrak B}(P)\). Now let \(\Sigma_\infty\) denote the group of all permutations of \(\mathbb{N}\) which permute in fact only a finite subset of \(\mathbb{N}\). Every \(\sigma\in \Sigma_\infty\) induces a map \(\overline\sigma: {\mathcal P}\to {\mathcal P}\) by \(\overline\sigma (U)=\{\sigma (u):u\in U\}\), \(U\in {\mathcal P}\). A probability measure \(\mu\) on \({\mathfrak B}({\mathcal P})\) is said to be exchangeable if \(\mu\circ \overline\sigma^{-1} =\mu\) for each \(\sigma\in \Sigma_\infty\). Further, put \(\nabla=\{{\mathbf x}=(x_1,x_2, \dots): x_1\geq x_1\geq 0\), \(\sum_{i\geq 1}x_i\leq 1\}\). For \({\mathbf x}\in \nabla\), consider the probability measure \(\kappa({\mathbf x})= \sum_{i\geq 0} x_i\varepsilon_i\) on \(\mathbb{N}_0\), where \(x_0=1- \sum_{i\geq 1}x_i\), and the product probability measure \(\kappa({\mathbf x})^\infty\) on \(\mathbb{N}_0^\mathbb{N}\). Finally, let \(F:\mathbb{N}^\mathbb{N}_0 \to {\mathcal P}\) be defined by \(f\mapsto\{\{f=c\}: 1\leq c\in f(\mathbb{N})\} \cup\{\{i\}: f(i)=0\}\), and consider the probability measure \(\kappa({\mathbf x})^\infty \circ F^{-1}\). Another main result of the paper is a new proof of the following theorem due to Kingman (Theorem 2): The set of exchangeable probability measures on \({\mathfrak B}({\mathcal P})\) is a Bauer simplex whose extreme points are precisely the probability measures \(\kappa({\mathbf x})^\infty\circ F^{-1}\), \({\mathbf x}\in \nabla\). Reviewer's remarks: Sometimes the authors use awkward notation instead of simply describing by words some mathematical objects. Moreover, there is a gap in the proof of Theorem 2. Namely, to identify the extreme points, the authors appeal to the method of moments applied to a suitable random measure on \([0,1]\). Unfortunately, they are able to handle only the first and the second moment. For the other moments, they claim that ``it is only notationally more cumbersome to prove the general case''.