Hoeffding decompositions and urn sequences (Q2519683)

Let \({\mathbf X}= X_1,X_2,\dots\) be an infinite exchangeable sequence of \(D\)- valued random variables, with \(D= \{d_1,\dots, d_m\}\). Let \(SU_0(X_1,\dots, X_n)= R\) and \(SU_k\), \(k= 1,\dots, n\), be the space of all random variables \[ F(X_1,\dots, X_n)= \sum\varphi(X_{j1},\dots, X_{jk})\tag{1} \] with sum over all \(1\leq j_1<\cdots< j_k\leq n\) and \(\varphi= \varphi_{nk}\) a unique symmetric function \(D^k\to R\). Define orthogonality by \(EYZ= 0\) and then inductively the spaces \(SH_k(X_1,\dots, X_n)\), \(k= 0,\dots, n\), by \(SH_0= SU_0= R\), \(SH_k= SU_k\cap (SU_{k-1})^{\perp}\), \(k= 1,\dots, n\), so that the set \(SU_n\) of all symmetric functions \(T(X_1,\dots, X_n)\) has the orthogonal decomposition into spaces \(SH_k\), \(k= 0,\dots, n\). The random sequence \({\mathbf X}\) is Hoeffding decomposable if for every \(n> 1\) and \(k= 1,\dots, n\) we have \(F(X_1,\dots, X_n)\). \(SH_k(X_1,\dots, X_n)\) if and only if the kernel \(\varphi\) in (1) is degenerate in the sense that \(E[\varphi(X_1,\dots, X_k)\mid X_2,\dots, X_k]=0\), a.s. It was shown by \textit{G. Peccati} [Ann. Probab. 32, No. 3A, 1796--1829 (2004; Zbl 1055.62060)], that \({\mathbf X}\) is Hoeffding decomposable iff it is, in a certain technical sense, weakly independent. The paper studies the characterization of Hoeffding decomposable \({\mathbf X}\) in terms of its de Finetti representation. Now let \(m= 2\). Then \(P(X_1= x_1,\dots, X_n= x_n)= \int_{[0,1]}\theta^j(1- \theta)^{n-j}\gamma(d\theta)\), where \(\sum x_i\) and \(\gamma\) the de Finetti measure for \(m= 2\). It is shown then that nondeterministic \({\mathbf X}\) is Hoeffding decomposable iff either \({\mathbf X}\) is i.i.d. or \(\gamma\) is a beta distribution so that \({\mathbf X}\) is a two- colour Pólya sequence. Nondeterministic here means that \(\gamma\) is not restricted to the two-point set \(\{0, 1\}\). The if part is in \textit{G. Peccati} [loc. cit.]. The proof of the difficult only if part starts with a characterization of Hoeffding decomposability in terms of conditional probabilities for zeroes in \(X_1,\dots, X_n\). From this a third-order recurrence for the moments \(\mu_i\) of \(\gamma\) is derived. Then it is shown that they are the moments of a beta distribution, for \(i= 0,1,2\) directly and for \(i> 2\) by the recurrence. The assertion follows since a distribution on \([0,1]\) is determined by its moments. When \(m\geq 2\) the situation is more complicated. For Hoeffding decomposable \({\mathbf X}\) the authors prove a recurrence for certain sets. It follows then that the law of \({\mathbf X}\) is determined by the mean vector and the covariance matrix of the de Finetti measure \(\gamma\). Corollary: When \(\gamma\) is the distribution of \(\xi_i/\sum\xi_j\), \(i= 1,\dots, m\) with independent and infinitely divisible \(\xi_1,\dots, \xi_m\) the distribution \(\gamma\) must be \(m\)-dimensional beta.