Limit theorems for a class of identically distributed random variables. (Q1879813)

Let \(X= (X_1,X_2,\dots)\) be a sequence of \(E\)-valued random variables adapted to a filtration \({\mathcal G} =({\mathcal G}_0,{\mathcal G}_1,{\mathcal G}_2,\dots)\). Then \(X\) is called conditionally identically-distributed with respect to \({\mathcal G}\), or \({\mathcal G}\)-c.i.d. when \(E[f(X_k) \mid {\mathcal G}_n] = E[f(X_{n+1})\mid {\mathcal G}_n]\), a.s., for all \(k > n\geq 0\) and all bounded measurable \(f: E\to R\). Equivalently for \({\mathcal G}={\mathcal G}(X)\): When \([X_1,\dots,X_n,X_{n+2}]\sim [X_1,\dots,X_n,X_{n+1}]\), \(n\geq 0\). Examples are given showing that c.i.d. is weaker than exchangeability. Conditions for equivalence are derived, e.g. \(X\) is c.i.d. under \(P (.\mid H)\) for any shift-invariant \(H\), or \(X\) is c.i.d. under all finite permutations of the \(X_i\). A strong law for c.i.d. \(X\) is proved: \(E[f(X_{n+1})\mid {\mathcal G}_n]\to V(f)\), a.s. and in \(L^1\), when \(E| f(X_1)|<\infty\). Also \(\sum_{i<n}f(X_i)/n\to V(f)\), a.s., for bounded \(f\). Three central limit theorems (CLT) are proved, with different random centerings, by application of a martingale CLT. The convergence is stable, i.e. there is convergence in law conditionally given any \(H\) with \(P(H) > 0\). In the final section the CLT is shown to hold as convergence in distribution of processes to a limiting process, using the uniform distance topology, e.g. for the processes \[ W(n,t)=n^{-1/2}(f(X_1) +\cdots +f(X_n) - nV (f)), \quad t\geq 0, \] with \(f= I((-\infty,t])\).