1
$\begingroup$

Let $E$ be a separable $\mathbb R$-Banach space, $\rho$ be a complete separable metric on $E$, $\operatorname W_\rho$ denote the Wasserstein metric of order $1$ associated to $\rho$, $\mathcal M_1(E)$ denote the set of probability measures on $(E,\mathcal B(E))$ and $(\kappa_t)_{t\ge0}$ be a Markov semigroup on $(E,\mathcal B(E))$ with $$\operatorname W_\rho(\mu\kappa_t,\nu\kappa_t)\le ce^{-\lambda t}\operatorname W_\rho(\mu,\nu)\;\;\;\text{for all }\mu,\nu\in\mathcal M_1(\mu,\nu)\tag1$$ for some $c\ge0$ and $\lambda>0$.

Are we able to conclude that $(\kappa_t)_{t\ge0}$ has a unique invariant measure $\mu_\ast\in\mathcal M_1(E)$?

By $(1)$, the adjoint semigroup $(\kappa_t^\ast)_{t\ge0}$ is eventually contractive: Let $t_0\ge0$ with $$ce^{-\lambda t}<1\;\;\;\text{for all }t\ge t_0$$ and $t\ge t_0$. Since the Wasserstein space $$\mathcal S^1(E,\rho):=\left\{\mu\in\mathcal M_1(E):(\mu\otimes\delta_0)\rho<\infty\right\}$$ equipped with $\operatorname W_\rho$ is complete and hence we can apply Banach's fixed-point theorem yielding that there is a unique $\mu_\ast\in\mathcal S^1(E,\rho)$ with $$\mu_\ast\kappa_t=\mu_\ast\tag2.$$ Moreover, for any $\mu_0\in\mathcal S^1(E,\rho)$ and $$\mu_n:=\mu_{n-1}\kappa_t\;\;\;\text{for }n\in\mathbb N,$$ it holds $$\operatorname W_\rho(\mu_n,\mu_\ast)\xrightarrow{n\to\infty}0\tag3.$$

So, all what's left to prove is that $\mu_\ast$ does not depend on $t$, i.e. $\mu_\ast$ is invariant with respect to $\kappa_t$ for all $t\ge t_0$.

BTW: Is this all we can hope for or can we even conclude that $\mu_ast$ must be invariant with respect to $\kappa_t$ for all $t\ge\color{red}0$?

$\endgroup$
  • $\begingroup$ That's fine, but it only proves uniqueness in the space of measures with a finite first moment. $\endgroup$ – R W Jun 27 '20 at 17:59
  • $\begingroup$ @RW Yes, this space is $\mathcal S^1(E,\rho)$. I would like to show that there is a $\mu_\ast\in\mathcal S^1(E,\rho)$ with $\mu_\ast\kappa_t=\mu_\ast$ for all $t\ge t_0$. $\endgroup$ – 0xbadf00d Jun 27 '20 at 18:02
3
$\begingroup$

Note that your argument contains an implicit assumption that $\kappa_t \mu \in \mathcal{S}^1$ for every $\mu \in \mathcal{S}^1$ (otherwise the Banach fixed point theorem does not apply). I will also make that assumption. Also, I realized that I have written $\kappa_t \mu$ with $\mu$ on the right; sorry about that.

You have shown that for some fixed $t^* \ge t_0$, that $\kappa_{t^*}$ has an invariant measure $\mu_*$ which is unique in $\mathcal{S}^1$.

Let $t > 0$ be arbitrary. Then we have by the semigroup property that $$\kappa_{t^*} \kappa_t \mu_* = \kappa_{t+ t^*} \mu_* = \kappa_t \kappa_{t^*} \mu_* = \kappa_t \mu_*$$ which proves that $\kappa_t \mu_*$ is invariant for $\kappa_{t^*}$. By uniqueness, $\kappa_t \mu_* = \mu_*$. This proves that $\mu_*$ is invariant for $\kappa_t$.

If $t \ge t_0$, then your argument shows that $\mu_*$ is in fact the unique invariant measure in $\mathcal{S}^1$ for $\kappa_t$. Otherwise, for $t < t_0$, suppose $\mu' \in \mathcal{S}^1$ is another invariant measure for $\kappa_t$. Let $n$ a large enough integer so that $n t \ge t_0$; then $\mu' = \kappa_t^n \mu' = \kappa_{nt} \mu'$. Since $\kappa_{nt}$ has $\mu_*$ as its unique invariant measure, we have $\mu' = \mu_*$.

We have thus shown that for every $t$, $\mu_*$ is invariant for $\kappa_t$, and is the unique such measure in $\mathcal{S}^1$.

$\endgroup$
  • $\begingroup$ Thank you for your answer. (a) For your uniqueness argument in your second argument to work, we need that $\kappa_t^\ast\mu_\ast\in\mathcal S^1$. It's not clear to me why this is the case. Is it a general fact that a Markov kernel is $\mathcal S^1$-preserving? (b) Just to be sure: The result we've obtained is that $(\kappa_t)_{t\ge0}$ has a unique invariant measure in $\mathcal S^1$, but there might be other invariant measures in $\mathcal M_1\setminus\mathcal S^1$, right? $\endgroup$ – 0xbadf00d Jun 28 '20 at 5:09
  • 1
    $\begingroup$ @0xbadf00d: Good point. In fact, consider the following counterexample: let $\nu$ be a probability measure not in $\mathcal{S}^1$ and take $\kappa_t f(x) = \int f\,d\nu$ for all $t$. Then $\kappa_t \mu = \nu$ for every $\mu$ and every $t>0$, so (1) is trivially satisfied (the LHS is always zero). Indeed, this is also a gap in your argument, because $\kappa_t$ has a unique invariant measure and it is not in $\mathcal{S}^1$. I think we need to assume that $\kappa_t$ takes $\mathcal{S}^1$ into $\mathcal{S}^1$ for any of this to work. $\endgroup$ – Nate Eldredge Jun 28 '20 at 12:04
  • $\begingroup$ @0xbadf00d: I agree that even with this assumption, nothing appears to rule out the possibility of another invariant measure which is not in $\mathcal{S}^1$. I don't know of an example where this happens, however. $\endgroup$ – Nate Eldredge Jun 28 '20 at 13:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.