Multiple integral representation for functionals of Dirichlet processes (Q1002577)

The main purpose of this article is a chaotic representation for square integrable functionals of a Dirichlet process, very analogous to the root usual Wiener or Poisson chaos, except that no regularity of the functional is needed here. The Dirichlet (or Dirichlet-Ferguson) process is a random probability measure \(D\), defined over a measured space \((A,C\ell,\alpha)\), \(\alpha\) a given finite Borel measure, by the requirement that for any finite measurable partition of \(A:(C_1,\dots, C_n)\) the vector \((D(C_1),\dots, D(C_n))\) has the Dirichlet distribution with parameters \((\alpha(C_1),\dots,\alpha(C_n))\). It is known that such Dirichlet process can be got as a.s. limit of infinite urn Pólya scheme \((X_n,n\geq 1)\), \(\alpha\) representing the initial composition of the urn. Then \((X_n)\) is an infinite exchangeable sequence, which is i.i.d. with common law \(D\), conditionally on \(D\). The Fock space associated with \((X_n)\) is known in statistics as Hoeffding space. The chaotic representation for \(L^2(D)\) induces (as for the Wiener chaos) a Hilbert space isomorphism with this Hoeffding space. The author provides finally an application of this chaotic representation, to Bayesian nonparametric \(U\)-statistics.