General criteria for the study of quasi-stationarity (Q6164925)

The main theme of this article consists in studying the general criteria for quasi-stationarity. In the paper, the authors provide general criteria ensuring the existence and the exponential non-uniform convergence in weighted total variation norm to a quasi-stationary distribution for Markov processes with absorption. And besides they also characterize a subset of its domain of attraction by an integrability condition, and prove the existence of a right eigenvector for the semigroup of the process and the existence and exponential ergodicity of the \(Q\)-process. \par More precisely, let \(( X_t, t \in I )\) be a Markov process in \(E \cup \{ \partial \}\), where \(E\) is a measurable space and \(\partial \notin E\), with set of time indices \(I\) which might be \(\mathbb{Z}_+\), \(\mathbb{R}_+\) or \(\frac{1}{2} \mathbb{Z}_+\) for some \(k \in \mathbb{N}\). \[ \tau_{\partial} = \inf \{ t \in I, X_t = \partial \}\tag{1} \] indicates the absorption time. \(( P_n ), n \in \mathbb{Z}_+\), defined as \[ P_n f(x) = \mathbb{ E}_x [ f( X_n) \mathbf{1}_{ \{ n < \tau_{\partial} \} } ], \qquad \text{ for all } n \in \mathbb{Z}_+,\tag{2} \] gives the sub-Markovian transition semigroup of \(X\) considered at integer times, for all bounded or nonnegative measurable functions \(f\) on \(E\) and all \(x \in E\). The usual left-action of \(P_n\) on measures is defined as \[ \mu P_n f = \int_E P_n f(x) \mu(dx),\tag{3} \] for all probability measures \(\mu\) on \(E\). We need the following assumption. Assumption (E). There exist a positive integer \(n_1\), positive real constants \(\theta_1, \theta_2, c_1, c_2, c_3\), two functions \({\varphi}_1, {\varphi}_2 : E \to \mathbb{R}_+\) and a probability measure \(\nu\) on a measurable subset \(K \subset E\) such that (E 1) (Local Dobrushin coefficient) For \(\text{ for all } x \in K\), \[ \mathbb{P}_x( X_{n_1} \in ( \cdot) ) \geq c_1 \cdot \nu ( ( \cdot ) \cap K ).\tag{4} \] (E 2) (Global Lyapunov criterion) We have \(\theta_1 < \theta_2\) and \[ \inf_{x \in E} \varphi_1(x) \geq 1, \qquad \sup_{x \in K } \varphi_1 (x) < \infty, \qquad \inf_{ x \in K} \varphi_2(x) > 0, \qquad \sup_{ x \in E} \varphi_2(x) \leqslant 1,\tag{5} \] \[ P_1 \varphi(x) \leqslant \theta_1 \varphi(x) + c_2 \mathbf{1}_K (x), \qquad \text{ for all } x \in E, \qquad \qquad P_1 \varphi_1(x) \leqslant \theta_2 \varphi_2(x), \qquad \text{ for all } x \in E.\tag{6} \] (E 3) (Local Harnack inequality) We have \[ \sup_{ n \in \mathbb{Z}_+ } \frac{ \sup_{ y \in K} \mathbb{P}_y ( n < \tau_{\partial} )}{ \inf_{y \in K} \mathbb{P}_y ( n < \tau_{\partial} ) } \leqslant c_3.\tag{7} \] (E 4) (Aperiodicity) For all \(x \in K\), there exists \(n_4(x)\) such that, for all \(n \geq n_4(\partial)\), \[ \mathbb{P}_x ( X_n \in K ) > 0.\tag{8} \] Note that it follows from (E 2) that \(\theta_2 \leqslant 1\) and thus \(\theta_1 < 1\). Assumption (E) is an extension of the ergodicity criteria developed in [\textit{S. P. Meyn} and \textit{R. L. Tweedie}, Adv. Appl. Probab. 25, No. 3, 518--548 (1993; Zbl 0781.60053)]. Indeed, if the condition that \(\tau_{\partial} = \infty\), \(\mathbb{P}_x\)-almost surely for all \(x \in E\) is assumed, then Condition (E 3) becomes void and one can take \(\varphi_2 \equiv 1\) in (E 2), so that \(\theta_2 = 1\). We may recognize in (E 1) the standard small set assumption, in (E 2) for \(\varphi_1\) a standard Foster-Lyapunov criterion, and in (E 4) an aperiodicity condition. In the general case, the authors provide conditions ensuring the existence of Lyapunov functions satisfying (E 2) in terms of exponential moment of hitting times for \(\varphi_1\) and exponential decay of the probability to be in \(K\) for \(\varphi_2\), and conditions ensuring (E 1) and (E 3) based on comparisons between transition probabilities. \par The next is the first main result of this paper. Theorem 1. There exist a constant \(C > 0\), a constant \(\alpha \in (0, 1)\), a probability measure \(\nu_{QSD}\) on \(E\) such that \(\nu_{QSD} (K) > 0\) and such that \[ \left\Vert \frac{ \mu P_n}{ \mu P_n \mathbf{1}_E } - \nu_{QSD} \right\Vert_{ TV( \varphi_1)} \leqslant C \cdot \alpha^n \frac{ \mu( \varphi_1)}{ \mu( \varphi_2) }, \qquad \qquad \text{ for all } n \geq 0,\tag{9} \] for all probability measures \(\mu\) on \(E\) such that \(\mu( \varphi_1) < \infty\) and \(\mu( \varphi_2) > 0\). In addition, \(\nu_{QSD}\) is the unique quasi-stationary distribution satisfying \(\nu_{QSD} ( \varphi_2) > 0\) and \(\nu_{QSD} ( \varphi_1) < \infty\). Here \(\mu(\varphi) := \int_E \varphi(x) \mu( d x)\) for all probability measures \(\mu\) on \(E\), and for all probability measures \(\mu_1\) and \(\mu_2\), the total variation distance \(\Vert \cdot \Vert_{ TV( \varphi )}\) is given by \[ \Vert \mu_1 -\mu_2 \Vert_{TV( \varphi_1)} := \sup_f \vert \mu_1 (f) - \mu_2(f) \vert,\tag{10} \] where the supremum ``sup'' is taken over the set of \(f : E \to \mathbb{R}\) being measurable such that \(\vert f \vert \leqslant \varphi_1\). \par In order to state the assertion of the next main theorem, we need to define the space \(L^{\infty}(\psi)\) as the set measurable real functions \(f\) on \(E\) such that \[ \Vert f \Vert_{ L^{\infty}(\psi)} := \sup_{ x \in E} \frac{ f(x)}{\psi(x)} < \infty, \tag{11} \] for all positive functions \(\psi\) on \(E\). Note that \(( L^{\infty}(\psi), \Vert \cdot \Vert_{ L^{\infty}(\psi)} )\) is a Banach space. Theorem 2. There exists a function \(\eta : E \to \mathbb{R}_+\) such that \[ \eta(x) = \lim_{ n \to + \infty} \frac{ \mathbb{P}_x ( n < \tau_{\partial} )}{ \mathbb{P}_{ \nu_{QSD} } ( n < \tau_{\partial} ) } = \lim_{n \to + \infty} \theta_0^{-n} \cdot \mathbb{P}_x ( n < \tau_{\partial} ), \quad \text{ for all } x \in E,\tag{12} \] where the convergence is geometric in \(L^{\infty}( \varphi_1)\). \par It is interesting to note that \(\inf_{y \in K} \eta(y) > 0\), \(\nu_{QSD}(\eta)=1\), \(\eta \in L^{\infty}( \varphi_1^{F(*)} )\), \(P_1 \eta =\theta_0 \eta\) and \(\theta_0 \geq \theta_2 > \theta_1\), with \(F(*) = \log ( 1 / \theta_0) / \log ( 1 / \theta_1)\). Moreover, the authors consider the \(Q\)-process and its ergodicity properties under Condition (E). The next result states that \(\Omega = E^{ \mathbb{Z}_+ }\) is the canonical state space of Markov chains on \(E\) and \(( \mathcal{F}_n )\), \(n \in \mathbb{Z}_+\), is the associated canonical filtration. We define \(E'\) \(:=\) \(\{ x \in E\), \(\eta(x) > 0 \}\). Theorem 3. (a) (Existence of the \(Q\)-process) There exists a family \(( Q_x)\), \(x \in E'\), of probability measures on \(\Omega\), defined by \[ \lim_{ n \to + \infty} \mathbb{P}_x ( A \vert n < \tau_{\partial} ) = \mathbb{Q}_x (A), \qquad \text{ for all } x \in E',\tag{13} \] for all \(\mathcal{F}_m\)-measurable sets \(A\) and for all \(m \geq 0\). The process \(( \Omega\), \(( \mathcal{F}_n )_{ n \in \mathbb{Z}_+ }\), \(( X_n )_{n \in \mathbb{Z}_+}\), \(( \mathbb{Q}_x )_{ x \in E'} )\) is an \(E'\)-valued homogeneous Markov chain. (b) (Semigroup) The semigroup of the Markov process \(X\) under \(( \mathbb{Q}_x )_{ x \in E'}\) is given for all bounded measurable functions \(\varphi\) on \(E'\) and \(n \geq 0\) by \[ \tilde{P}_n \varphi (x) = \frac{ \theta_0^{-n} }{ \eta (x)} P_n ( \eta \varphi)(x). \tag{14} \] (c) (Exponential ergodicity) The probability measure \(\beta\) on \(E'\) defined by \[ \beta(dx) = \eta(x) \nu_{QSD} ( dx)\tag{15} \] is the unique invariant distribution of the Markov process \(X\) under \(( \mathbb{Q}_x )_{ x \in E'}\). Moreover, there exist constants \(C > 0\) and \(\alpha \in (0,1)\) such that, for all initial distributions \(\mu\) on \(E'\) such that \(\mu( \varphi_1 / \eta) < \infty\) and \[ \Vert \mu \tilde{P}_n - \beta(h) \Vert_{TV( \varphi_1 / \eta)} \leqslant C \alpha^n \mu( \varphi_1 / \eta), \qquad \text{ for all } n \geq 0.\tag{16} \] In addition, for all initial distributions \(\mu\) on \(E'\), \[ \Vert \mu \tilde{P}_n - \beta \Vert_{TV} \longrightarrow 0 \qquad ( \text{as} \quad n \to \infty).\tag{17} \] For other related works, see, e.g., [\textit{M. Benaim} et al., Ann. Appl. Probab. 28, No. 4, 2370--2416 (2018; Zbl 06974754)] for stochastic approximation of quasi-stationary distributions on compact spaces, [\textit{J. R. Chazottes} et al., Ann. Inst. Henri Poincaré, Probab. Stat. 55, No. 4, 2249--2294 (2019; Zbl 1434.60237)] for quasi-stationary distributions for multitype birth-and-death processes, [\textit{W. Oçafrain}, Electron. J. Probab. 26, Paper No. 83, 30 p. (2021; Zbl 1470.39058)] for convergence to quasi-stationary through Poincaré inequalities and Bakry-Émery criteria, and [\textit{A. Velleret}, Stochastic Processes Appl. 148, 98--138 (2022; Zbl 1491.60128)] for unique quasi-stationary distribution with a possibly stabilizing extinction.