Spreading and predictable sampling in exchangeable sequences and processes (Q1105909): Difference between revisions

Some results on exchangeable sequences are given in terms of martingales and stopping times. Extensions to exchangeable processes in continuous time are also considered. A sequence \(X=(X_ 1,X_ 2,...)\) of random variables is said to be spreadable if \[ (X_{k_ 1},X_{k_ 2},...)=d(X_ 1,X_ 2,...),\quad k_ 1<k_ 2<.... \] Define the measure-valued processes \((\pi_ n)\) and \((\lambda_ n)\) by \[ \pi_ n=P[T_ n\circ X\in \cdot | {\mathcal F}_ n],\quad \lambda_ n=P[X_{n+1}\in \cdot | {\mathcal F}_ n],\quad n\geq 0, \] where \(T_ 0,T_ 1,..\). denote the shift operators on \({\mathbb{R}}^{\infty}\), \({\mathcal F}_ n=\sigma (X_ 1,...,X_ n)\), \(n\geq 1\). It is proved that the following conditions are equivalent: (i) X is spreadable; (ii) \(T_{\tau}\circ X=d X\) for every finite stopping time \(\tau\); (iii) \((\pi_ nf)\) is a martingale for every bounded \(f:R^{\infty}\to R.\) The following are also \(equivalent:\) (i') \((X_ 1,...,X_ n\), \(X_{n+2})=d(X_ 1,...,X_ n\) \(X_{n+1})\), \(n\geq 1;\) (ii') \(X_{\tau +1}=d X_ 1\) for every stopping \(time;\) (iii') \((\lambda_ nf)\) is a martingale for every bounded \(f: R\to R.\) Let \(\{\nu_ i,i\geq 1\}\) be a sequence of i.i.d. random variables and independent of X with \[ P[\nu_ i=1]=1-P[\nu_ i=0]=p,\quad i\in {\mathbb{N}}, \] for some \(p\in (0,1]\) and set \[ \tau_ j=\inf \{k\in {\mathbb{N}}:\sum^{k}_{i=1}\nu_ i=j\},\quad j\in {\mathbb{N}}. \] The sequence \(Y=(X_{\tau_ 1},X_{\tau_ 2},...)\) is called a p-thinning of X. The following statement characterizes exchangeability. Fix \(p\in (0,1]\) and let X be a stationary sequence of random variables with p-thinning Y. Then X is exchangeable iff \(X=d Y.\) Fix a filtration \({\mathcal F}=({\mathcal F}_ 0\), \({\mathcal F}_ 1,...)\). The following theorem shows that the distribution of an exchangeable sequence is invariant under predictable sampling. Let X be a finite or infinite \({\mathcal F}\)-exchangeable sequence with index set I and let \(\tau_ 1,...,\tau_ k\) be a.s. distinct I-valued predictable stopping times. Then \[ (X_{\tau_ 1},...,X_{\tau_ k})=d(X_ 1,...,X_ k). \] It is difficult to quote here the results on martingale characterizations in continuous time.