Efficient Markovian couplings: Examples and counterexamples (Q811755)

This review consists of three parts. The first part introduces the main aim and results of the paper in review. The remainders are some related comments and remarks, which may be helpful for the readers. (A) The intention of the paper is to clarify the time-continuous Markovian couplings which can be used to give a sharp estimate for the spectral gap. The method is to compare the spectral gap \(\lambda_1\) (called \(\mu_2\) in the paper) with some exponential convergence rates. More precisely, the authors introduce a notion ``efficient coupling'' in three different senses as follows. (1) [Definition 2.2]. Let \(\mu =\mu(x_1, x_2)\) \((x_1\neq x_2)\) be the largest constant such that for all \(t\geq 0\), \[ c_1(x_1, x_2) e^{-\mu t} \leq \mathbb P^{x_1, x_2}[\tau >t] \leq c_2(x_1, x_2) e^{-\mu t}, \] where \(\tau \) is the coupling time of the coupled process \((X_t^1, X_t^2)\). Then, the coupling is said to be efficient if \(\min_{x_1\neq x_2} \mu(x_1,x_2)=\lambda_1\). (2) [Definition 3.3]. Let \(g\) be an eigenfunction of \(\lambda_1\). Then a coupling is called efficient if for some \(x_1\) and \(x_2\) with \(g(x_1)\neq g(x_2)\), \[ \mathbb E^{x_1, x_2} |X_t^1-X_t^2|\leq c(x_1, x_2) e^{-\lambda_1 t},\qquad t\geq 0. \] (3) [Definition 3.6]. Everything is the same as in (2) except the last inequality is replaced by \[ \mathbb P^{x_1, x_2}[\tau >t] \leq c(x_1, x_2) e^{-\lambda_1 t},\qquad t\geq 0. \] The case (1) is used for reversible Markov chains with finite state space and the other two cases are used for reflected Brownian motion in a triangle. A general affirmative result obtained in the paper is about birth-death processes with finite state space and reflecting boundary, for which the classical coupling is proved to be efficient in the sense of (1). One may compare this result with the reviewer's variational formulas for \(\lambda_1\) [cf. the reviewer, Sci. China, Ser. A 43, No. 10, 1051-1059 (2000) for details and further references]. The formulas can be easily applied to estimate the lower or upper bounds of \(\lambda_1\). Note that the estimation of \(\lambda_1\) is not treated in the paper in review. Another interesting result of the paper says that the well-known mirror coupling (or coupling by reflection) of reflected Brownian motion can be efficient or not in the sense of (3), according to respectively the triangle has an obtuse angle, or all angles are different each other and acute [Theorem 3.7]. The proof is quite technical. (B) It is worthy to compare the authors' ``efficient coupling'' with the reviewer's ``\(\rho \)-optimal coupling'' since the comments in the last paragraph in page 363 of the paper are somehow not serious. In the study of spectral gap, the reviewer compares the spectral gap with the exponential convergence rate \(\mu\) with respect to a distance \(\rho \). That is, we have \(\lambda_1\geq \mu\) whenever there is a (Markovian) coupling and a distance \(\rho \) such that \(\mathbb E^{x_1, x_2} \rho (X_t^1, X_t^2) \leq \rho (x_1, x_2) e^{-\mu t}\) for all \(x_1\neq x_2\) and \(t\geq 0\). The last condition is indeed equivalent to \(L\rho (x_1, x_2)\leq -\mu \rho (x_1, x_2) \) for all \(x_1\neq x_2\), where \(L\) is a coupling operator of the marginal operators \(L_1\) and \(L_2\) corresponding to the processes \((X_t^1)\) and \((X_t^2)\), respectively, and \(L\) varies over all coupling operators of \(L_1\) and \(L_2\). This naturally leads to the following notion. A coupling operator \(\overline L\) is called \(\rho \)-optimal of \(L_1\) and \(L_2\) if \(\overline L \rho (x_1, x_2)=\inf_{L} L \rho (x_1, x_2)\) for all \(x_1\neq x_2\). This notion gives us a classification of Markovian couplings with respect to a given distance \(\rho \) and is clearly a helpful tool to study the exponential ergodicity (and other topics) of Markov processes. Obviously, in order to get a good estimate of \(\lambda_1\), one needs not only a good coupling but also a good distance since the convergence rate is not a topological concept. Fortunately, it is often true that one coupling is optimal with respect to a large class of distances \(\rho'= \psi \circ \rho \), where \(\psi \) is an arbitrary function on \([0, \infty)\) having properties \(\psi (0)=0\), \(\psi'>0\) and \(\psi''\leq 0\). Thus, we can first fix a coupling and then improve the estimate \(\mu\) by varying \(\psi \). In one-dimensional case for instance, we can always achieve the sharp estimate by using the classical coupling but infinitely many numbers of \(\psi \), depending on the coefficients of the operator \(L_1=L_2\), in contrast to the two fixed types of topology used in (1)--(3). As pointed out in the previous publications, the requirement of ``distance'' can be further relaxed in the study of spectral gap (see also part (C) below). For instance, by replacing the distance \(\rho \) with a nonnegative lower semi-continuous function \(\varphi \), one obtains the notion of \(\varphi \)-optimal couplings. Then, it is proved by \textit{S. Y. Zhang} [Acta Math. Sin., Ser. A 43, No. 5, 773-780 (2000)] that a \(\varphi \)-optimal coupling always exists for (Markov) jump processes with Polish state space. This then leads to solve a long-standed open problem. That is, the equivalence of the stochastic comparability of the semigroups and existence of an order-preserving Markovian coupling for jump processes [cf. \textit{S. Y. Zhang} (loc. cit.) and \textit{Y. Zhang} (see the paper reviewed above)]. However, the equivalence does not hold for diffusions. (C) We now show by an example that the ``inefficiency'' in the senses of (1)--(3) does not really mean the limitation of the coupling method. Let \(E=\{0, 1\}^3= \{(x_1, x_2, x_3): x_k=0\) or \(1\}\). Consider a Markov chain with state space \(E\) and with the following evolution rule. At each point \((x_1, x_2, x_3) \), only one coordinate can be flipped (jumped) between \(0\) and \(1\). Set \(E_k=\{(x_1, x_2, x_3): x_1+x_2+x_3=k\}\), \(k=0, 1, 2, 3\). The rate of the point in \(E_1\) (resp. \(E_3\)) jumps to a point in \(E_1\) (resp. \(E_2\)) is equal to \(e^{- 2 \beta} (\beta >0)\), and the rate to jump back is equal to \(e^{2 \beta }\). The rates of the jumps between the points in \(E_1\) and \(E_2\) are all equal to \(1\). The geometric structure of the chain suggests, and it is actually true, that the eigenfunction \(g\) of \(\lambda_1\) restricted on the set \(E_1\) (resp. \(E_2\)) is constant. This leads us to define a distance on the partition \(\{E_0, E_1, E_2, E_3\}\) of \(E\) rather than on the whole space \(E\) [certainly, one may extend the distance to the whole space by setting the distance between the points in \(E_1\) (resp. \(E_2\)) to be \(\varepsilon\) and then let \(\varepsilon \downarrow 0\)], as follows. \[ \rho (E_1, E_2)=2, \] \[ \rho (E_0, E_1)=\rho (E_2, E_3)= \tfrac{1}{2}\Bigl(-3 e^{-4\beta }-1 +4 e^{-2\beta }+\sqrt{(3 e^{-4\beta }+4 -4 e^{-2\beta })^2 + 16 e^{-2\beta }}\Bigr). \] Then, one can choose a coupling operator \(L\) so that \[ L\rho (x_1, x_2)\leq -\mu \rho (x_1, x_2) \] for all \(x_1\in E_i, x_2\in E_j\) with \(i\neq j\) and \[ \mu =\lambda_1= \tfrac{1}{2} \Bigl(3e^{-2\beta }+2+e^{2\beta }-\sqrt{e^{4\beta }+ 8e^{2\beta }+22-24 e^{-2\beta } +9 e^{-4\beta }}\Bigr) \] (some details are regretted to be omitted here since the limitation of the space). Hence, the coupling does produce a sharp estimate. However, the distance \(\rho \) used here is completely different from the ordinary one used in (2) and is far away to be related to \(\tau \) used in (1) and (3). In conclusion, a good estimate (therefore a good distance) should be related to the eigenfunction, which is a much refined property than couplings.