Théorie spectrale d'un opérateur de transition sur un espace métrique compact. (Spectral decompositions induced by equicontinuous Markov operators on a compact metric space) (Q1190311): Difference between revisions

Methods of martingale theory and ergodic theory are combined to study the spectral properties of a Markov operator \(P\) on a compact metric space \(E\). The hypotheses imposed on \(P\) is at least \[ \text{(HO):}\quad P\text{ operates on the space }{\mathcal C}(E)\text{ of continuous functions on }E. \] Let \({\mathcal H}_ c\) be the space of continuous \(P\)-harmonic functions and let \(G(\cdot,\cdot)\) be the potential kernel of \(P\). Let \(\sim\) be the equivalence relation defined by \({\mathcal H}_ c\), i.e., \(x\sim u\) iff for all \(f\) in \({\mathcal H}_ c\), \(f(x)=f(u)\). Let \(F\) be the set of all \(u\) such that for all \(x\) in \(\text{supp}(G(u,\cdot))\), \(u\sim x\). \(F\) and its ``ergodic'' classes are compact and \(P\)-absorbing. The space \(B=F/\sim\) equipped with the Hausdorff metric is compact metric and is potential-theoretically a boundary of \(E\): (2.3) For the canonical Markov chain \((X_ n)\) on \((\Omega,(\mathbb{P}_ x))\), there is a shift-invariant mapping \(\sigma: \Omega\to B\), such that \(\lim d(X_ n(\omega),\sigma(\omega))=0\), for \(\omega\) in \(\Omega_ 0\), a shift-invariant \(P\)-sure set. (2.4) With \(\mu_ x=\mathbb{P}_ x\cdot\sigma^{-1}\), for any \(h\) in \({\mathcal H}_ c\), there is \(\hat h\) in \({\mathcal C}(B)\), such that (Poisson formula) \(h(x)=\int\hat h(\gamma)\mu_ x(d\gamma)\). The Banach algebra \({\mathcal H}_ c\) is isomorphic to \({\mathcal C}(B)\). If \(\nu\) is an extremal \(P\)-invariant probability measure, \(\nu\) is necessarily supported by a single class. Assume \[ \text{(H1): for } f\in{\mathcal C}(E), \text{ the Cesaro sequence of the iterates of }P\text{ on } f, \] \[ \{n^{-1}\sum^{n-1}_ 0P^ mf: n\in\mathbb{N}\} \text{ is equi-continuous}. \] (3.2) Any class \(\gamma\) supports a unique \(P\)- invariant probability measure \(\nu_ \gamma\). And any \(p\)-invariant probability measure is an integral of such \(\nu_ \gamma\). The Cesaro limit (uniform in \(x\)) \[ {1\over n}\sum^{n-1}_ 0P^ mf(x)\to\int \nu_ \gamma(f)\mu_ x(d\gamma). \] (3.3) And \(\mathbb{P}\)-surely, the Cesaro-limit \[ {1\over n}\sum^{n-1}_ 0 f(X_ m)\to \nu_{\sigma(\;)}(f). \] Assume \[ \text{(H2): for }f\in{\mathcal C}(E),\text{ the iterates }\{P^ nf: n\in\mathbb{N}\}\text{ are equi-continuous.} \] (4.6) For a unimodular complex number \(\lambda\) let \({\mathcal H}_{\lambda c}\) be the eigenspace of \(P\), with eigenvalue \(\lambda\), of continuous functions. Let \(\approx\) be the equivalence relation defined by \(\cup{\mathcal H}_{\lambda c}\), namely: \(x\approx u\) iff for all \(f\) in \(\cup{\mathcal H}_{\lambda c}\), \(f(x)=f(u)\). Let \(\mathbb{C}\) be the quotient space of these ``cyclic'' classes. \(\mathbb{C}\) is a compact metric space, and \(P\) induces an isometry \(\tau\) on \(\mathbb{C}\). Indeed \(\tau\) acts transitively on each \(\gamma/\approx\) for \(\gamma\in B\). Then the space \({\mathcal C}(E)\) is the direct sum \({\mathcal C}(E)={\mathcal E}_ 1\oplus{\mathcal E}_ 2\), where \({\mathcal E}_ 1\) is generated by \(\cup{\mathcal H}_{\lambda c}\) and \({\mathcal E}_ 2=\{f:\lim_ n P^ nf=0\}\).