Comparison of sequential experiments (Q1922417): Difference between revisions

The decision theoretic comparison of statistical experiments can be extended from observation vectors \((X_1,\dots, X_n)\) to observation sequences \((X_n)_{n\geq 1}\). More precisely, let \(J:=\mathbb{N}\) or \(J=\mathbb{N}_n: =\{1,\dots,n\}\). A sequential experiment is defined as a triplet \((E,{\mathcal E},{\mathcal P})\) with \(E= \prod_{j\in J}E_j\), \({\mathcal E}:= \sigma(\{X_j: j\in J\})\) and \({\mathcal P}:= \{\{P^{ (j)}_\vartheta: \vartheta\in \Theta\}: j\in J\}\). Applying sequential decision procedures in terms of sequences \(\{\delta_j: j\in J\}\) of Markov kernels \(\delta_j\) on corresponding action spaces \((K_j, {\mathfrak K}_j)\) one introduces the notion of sequential sufficiency (sequential \(o\)-deficiency) between two sequential experiments \((E,{\mathcal E},{\mathcal P})\) and \((F,{\mathcal F},Q)\) by comparing their associated risk functions. The author of the present paper continues his joint work with the late E. N. Torgersen by proving a ``sequential Markov kernel criterion'' for the comparison \((E,{\mathcal E},{\mathcal P}) \geq_\mathbb{N}(F, {\mathcal F},Q)\). He also shows that this comparison is generally not implied by \((E,{\mathcal E},{\mathcal F})\geq_{\mathbb{N}_n}(F,{\mathcal F},Q)\) for all \(n\geq 1\). An interesting additional reference to the subject is the paper of \textit{H. Heckendorff}, Wiss. Z. Tech. Hochsch. Karl-Marx-Stadt 23, 383-386 (1981; Zbl 0489.62008).