On asymptotically optimal tests (Q5906255)

Let \({\mathcal P}\) be a collection of probability measures P defined on a probability space (\(\Omega\),\({\mathcal S})\) and let \({\mathcal S}_ n\) be an increasing sequence of sub-\(\sigma\)-fields of \({\mathcal S}\). Consider the problem of testing the hypothesis \(P\in {\mathcal P}_ 1\) against the alternative \(P\in {\mathcal P}_ 0\), where \({\mathcal P}_ 1\) is a proper subset of \({\mathcal P}\) and \({\mathcal P}_ 0={\mathcal P}-{\mathcal P}_ 1\). For a sequence of test functions \(\phi_ n\), let \[ \alpha_ n(P)=\int \phi_ ndP_ n\quad and\quad \beta_ n(P)=1-\alpha_ n(P) \] be the probabilities of type-I and type-II error, respectively, where \(P_ n\) is the restriction of P to \({\mathcal S}_ n\). Let \(\alpha_ n=\sup_{{\mathcal P}_ 0}\) \(\alpha_ n(P)\) and suppose that \(\alpha_ n\) is subject to certain upper bound conditions. Then, for Q in \({\mathcal P}_ 1\), the rate at which \(\beta_ n(Q)\) tends to 0 may be a measure of asymptotic performance of the sequence of test functions \(\phi_ n\). More precisely, \(\phi_ n\) with \(\limsup_{n}\alpha_ n<1\) is of rate \(A\geq 0\), if \[ \limsup_{n}n^{-1}\log \alpha_ n\leq -A. \] Next, let \(\Phi_ A\) be the collection of all sequences of tests of rate A and set \[ -B(A,Q,{\mathcal P}_ 0)=\inf_{\Phi_ a}\lim_{n}\inf n^{-1}\log \beta_ n(Q). \] Then \(\phi_ n\) is said to be exponential rate optimal at Q, if \[ \lim_{n}n^{-1}\log \beta_ n(Q)=-B(A,Q,{\mathcal P}_ 0). \] Another line of optimality is established as follows: Let \({\mathcal Y}\) be the collection of all sequences of test statistics \(\eta_ n\), where large values of \(\eta_ n\) are deemed significant. Let \(G_ n(t)=\sup_{{\mathcal P}_ 0} P_ n(\eta_ n\geq t).\) Then the rate at which \(G_ n(\eta_ n)\) tends to 0 in Q-probability, \(Q\in {\mathcal P}_ 1\), may serve as a measure of the asymptotic performance of the sequence of statistics \(\eta_ n\). More precisely, let \[ C(Q,{\mathcal P}_ 0)=\inf \{u: \lim_{n}Q_ n\{n^{-1}\log G_ n(\eta_ n)\geq -u\}=1,\quad for\quad all\quad \eta_ n\in {\mathcal Y}\}. \] Then a sequence of statistics \(\eta_ n\) in \({\mathcal Y}\) is said to be exact slope optimal at Q, if \[ \lim_{n}n^{-1}\log G_ n(\eta_ n)=-C(Q,{\mathcal P}_ 0)\quad in\quad Q-probability. \] In the remaining part of the paper, the author obtains upper bounds for B(A,Q,\({\mathcal P}_ 0)\) and C(Q,\({\mathcal P}_ 0)\) which are expressed in terms of the underlying statistical model (\(\Omega\),\(\{\) \({\mathcal S}_ n\},{\mathcal P})\). More precisely, let \(\rho_ n=\rho_ n(P,Q)=dP_ n/dQ_ n\), let \(A_ n(u)=\{\rho_ n\leq \exp (nu)\}\), let \(K(P,Q)=\inf \{u: \lim_{n} P_ n(A_ n(u))=1\}\), and let \[ K({\mathcal R},Q)=\inf_{{\mathcal R}} K(R,Q),\quad K(P,{\mathcal R})=\inf_{{\mathcal R}} K(P,R), \] where \({\mathcal R}\) is an arbitrary subset of \({\mathcal P}\). Then, under suitable regularity conditions, it is shown that \[ B(A,Q,{\mathcal P}_ 0)\leq K({\mathcal P}_ A,Q),\quad where\quad {\mathcal P}_ A=\{R: K(R,{\mathcal P}_ 0)<A\},\quad A>0; \] also, C(Q,\({\mathcal P}_ 0)\leq K(Q,{\mathcal P}_ 0)\). The methodology used applies equally well to discrete \(n\in \{0,1,...\}\) as well as for \(n\in [0,\infty)\). Furthermore, in a Markov processes framework, sequences of tests are given which are exponential rate optimal, and exact slope optimal for a large class of alternatives. The paper is concluded with a discussion on refinements of the results presented and on possible extensions.