Fusions of a probability distribution (Q1184099)

The authors study a new concept: The fusions of a probability. To make the results more accessible, we shall present (unlike the authors) only the finite-dimensional case. So, \(X\) is \(R^n\) (they consider any Banach separable space) and \({\mathcal P}\) is the convex set of all the probabilistic measures on \({\mathcal B}(X)\). If \(A\) is a Borel set and \(P\) is from \({\mathcal P}\), the barycenter of \(A\) with respect to \(P\) is the point \(b=b(A,P)=(\int_A x\,dP(x))/P(A)\) provided that \(P(A)\neq 0\) and the integral exists. Fix a \(P\) from \({\mathcal P}\). Let \(A\in{\mathcal B}(X)\) and \(0\leq t\leq 1\). Define \(Q_{A,t}(P)\) as \(P\) if \(tP(A)=0\) and as \(P\mid_{A^c}+tP(A)\delta(b(A,P))+(1-t)P\mid_A\) otherwise, provided that \(A\) has a barycenter. Say that \(Q_{A,t}(P)\) is the elementary fusion of \(P\) given \(A\) and \(t\). Let \({\mathcal E}(P)\) be the set of all elementary fusions of \(P\), let \({\mathcal S}(P)\) be the set of all the probabilities which can be obtained from \(P\) by using a finite number of elementary fusions \(({\mathcal S}(P)={\mathcal E}(P)\cup{\mathcal E}({\mathcal E}(P))\cup{\mathcal E}({\mathcal E}({\mathcal E}(P)))\cup\dots\;)\). Finally, let \({\mathcal F}(P)\) be the weak closure of \({\mathcal S}(P)\). A \(Q\) from \({\mathcal F}(P)\) is called a fusion of \(P\). Now the set \({\mathcal F}(P)\) has remarkable properties. It is convex; a fusion of a fusion of \(P\) is again a fusion of \(P\); if \(P\) has a barycenter and \(Q\in{\mathcal F}(P)\), then \(Q\) has the same barycenter; if \(P\) has no barycenter, strange things may happen (if \(X=R\) and \(P=\) the Cauchy distribution, then any distribution is a fusion of \(P\)!); if \(P\) has a first moment (i.e. \(\int \| x\| \,dP(x)<\infty)\), then \({\mathcal F}(P)\) is tight and conversely; if \(Q\in{\mathcal F}(P)\), then there must exist a dilation of \(X\), \(\mu\), such that \(P=\mu Q\) (it makes beautiful connections with the Choquet theory); the converse does not necessarily hold; if \(P\) is a fusion of \(Q\) and \(Q\) is a fusion of \(P\) and either has first moments, then \(P=Q\). If \(P\) has a first moment, the results are even nicer. So, \(Q\) is a fusion of \(P\) iff there exists a dilation of \(X\) such that \(P=\mu Q\). This is a strong continuation of the Cartier-Fell-Meyer theorem [\textit{P. Cartier}, \textit{J. M. G. Fell} and \textit{P.-A. Meyer}, Bull. Soc. Math. France 92, 435--445 (1965; Zbl 0154.39402)]. Some applications in cost-reward problems are given.