De Finetti-type theorems for random selection processes. Necessary and sufficient conditions (Q847766)

This paper looks for non-exchangeable families of random variables where the conditional independence of De Finetti's theorem still holds. It continues the search by the authors [e.g. J. Math. Anal. Appl. 347, No. 2, 407--415 (2008; Zbl 1209.60032); Electron. J. Probab. 14, 72--86, electronic only (2009; Zbl 1190.60027)]. Here they consider a random selection process (RSP) \(N\) on a set \(E\), i.e. a family \(N(x)\), \(x\in E\), of \(0-1\) random variables satisfying for every finite \(B\subset E\) the conditions (C1) The families \(N(x)\), \(x\in B\), and \(N(x)\), \(x\in B^c\), are conditionally independent given \(N(B)=\sum_{x\in B}N(x)\) and (C2) \(P(N(x)=1\mid N(B)=1= m(x)/m(B))\), \(x\in B\), where \(m\) is a weight function on \(E\) and \(m(B)=\sum_{x\in B}m(x)\). Interpretation: elements of \(E\) are selected and \(N(x)= 1\) or \(0\) means that \(x\) is or is not selected. It was shown by the authors [De Finetti's-type theorems for nonexchangeable 0-1 random variables, submitted ] that the \(N(x)\), \(x\in E\), are exchangeable iff (C1) and (C2) hold with constant \(m(x)\). In a RSP the distribution of \(N(x)\), \(x\in B\subset E\) is determined by the marginal distribution of \(N(B)\). These marginal distributions should be compatible for different \(B\), i.e. they should satisfy the ``total probability rule'' (1.2). This holds if they are weighted means, with a fixed distribution \(F\), of certain generalized binomials as in (1.3) and then the conditional independence holds. In [11] it is shown that under one of three conditions (A1), (A2), (A3) these generalized binomials are the extreme elements of the convex set \(D\) of all \(P(N(B)= s)\), \(B\) finite, \(B\subset E\), satisfying (1.2). Then (1.3) and the conditional independence follow. In order to prove that these conditions are necessary the present paper determines the extreme elements of \(D\) when no \(A_i\) holds. These elements are shown to be ``generalized hypergeometrics'' or mixtures of them and then the De Finetti independence does not hold. In section 4 an example is given of a RSP with non-constant function \(m\).