Essential spectral radius for Markov semigroups. I: Discrete time case (Q1424402)

Let \(P(x,dy)\) be a Markov kernel on some Polish space \(E\). Regarding it as an operator acting on some Banach lattices \(\mathbb{B}\) of measurable functions on \(E\) via \(Pf(x)\equiv \int_E f(y)\,P(x,dy)\) it is interesting to estimate the essential radius \[ r_{\text{ess}}(P|_{\mathbb{B}}):= \sup\{|\lambda|: \lambda\in\sigma_{\text{ess}} (P|_{\mathbb{B}}) \quad (\text{convention:}\sup\varphi:= 0); \] where \(\sigma_{\text{ess}}(P|_{\mathbb{B}})\) denotes the Wolf essential spectrum of \(P|_{\mathbb{B}}\), i.e. the set of \(\lambda\in C\) such that \(\lambda- P\) is not a Fredholm operator in \(\mathbb{B}\) (a bounded linear operator \(A: \mathbb{B}\to \mathbb{B}\) is said Fredholm if its range \(\text{Ran}(A)\) is closed and its kernel \(\text{Ker}(A)\) and the quotient space \(B/\text{Ran}(A)\) are both finite-dimensional). Here \(\mathbb{B}\) may be one of 1) \(b\mathbb{B}\), the space of all real measurable and bounded functions equipped with subnorm \(\| f\|:= \sup_{x\in E}| f(x)|\), 2) \(C_b(E)\), the space of all real continuous and bounded functions equipped with the sup-norm, or 3) \(L^p(\mu)= L^p_{\mathbb{R}}(E,\mathbb{B},\mu)\) equipped with the \(L^p\)-norm \(\| f\|_p\), where \(\mu\) is an invariant probability measure \(\mu\) of \(P\). In probability, the first study was realized on the ``bad'' space \(b\mathbb{B}\) due to the work of Doeblin (1937), (1940) (see also Dobb (1953) or Revuz (1976)). The second section is devoted to the analytic preparation. Section 3 is devoted to the study of Questions 1, 2 and 3 both on \(b\mathbb{B}\) and \(b_u\mathbb{B}\). The main new results are presented in Section 4. The author obtains the equivalence between \(\text{ress}(P| h\mathbb{B})< 1\) and \(\beta_t(P^N)< 1\) for some \(N\) for a Harris recurrent Markov kernel \(P\), based on a deep result of Horowitz (1972), and explains the asymptotic behavior of a Markov kernel \(P\) on \(b_u\mathbb{B}\) if its essential spectral radius is \(< 1\). The counterparts in \(L^p\) of some results in Section 3 are presented in Section 5 where Question 4 is solved (only) in the symmetric case. Sections 6, 7 and 8 are devoted to applications of estimates of \(r_{\text{ess}}(P)\) obtained previously. In Section 6, for a strong Feller and topologically transitive Markov chain the author proves the equivalence between \(r_{\text{ess}}(P)= 0\) of Donsker and Varadhan (1983), and especially the classical sufficient condition of Lyapunov function type found by Donsker and Varadhan for LDP is shown to be necessary. In Section 7, the author shows that the eigenvalues \(\lambda\) of symmetric Markov kernel \(P\) on discrete space \(E\) with \(|\lambda|> r_{\text{ess}}P\) can be estimated by the degrees of geometric recurrence to finite subsets of \(E\), and he establishes that the geometric convergence in the Wasserstein metric implies that in \(b_u\mathbb{B}\) or in \(L^2(\mu)\) in the symmetric case (Chen (1994), (1996)). Applications of general results are studied in Section 8. To keep the continuity of presentation the proofs of several results in Sections 3, 4 and 6 are left to Sections 9 and 10.