Abstract:In this paper we introduce some recent progresses on the convergence rate in Wasserstein distance for empirical measures of Markov processes. For diffusion processes on compact manifolds possibly with reflecting or killing boundary conditions, the sharp convergence rate as well as renormalization limits are presented in terms of the dimension of the manifold and the spectrum of the generator. For general ergodic Markov processes, explicit estimates are presented for the convergence rate by using a nice reference diffusion process, which are illustrated by some typical examples. Finally, some techniques are introduced to estimate the Wasserstein distance of empirical measures.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is about the asymptotic behavior of the empirical measure of Markov processes under the Wasserstein distance. Specifically, the author has studied the following aspects:
1. **Convergence rate of the empirical measure of diffusion processes**:
- For diffusion processes on compact manifolds (possibly with reflective or absorptive boundary conditions), the author gives the optimal convergence rate of the empirical measure under the Wasserstein distance and obtains the renormalized limit through the dimension of the manifold and the spectrum of the generator operator.
2. **Convergence estimates for general ergodic Markov processes**:
- For general ergodic Markov processes, by introducing a suitable reference diffusion process, the author gives an explicit estimate of the convergence rate and illustrates it with some typical examples.
3. **Behavior of the empirical measure in the quasi - ergodic case**:
- For the Markov process \(X_t\) with a finite lifetime (for example, a diffusion process with an absorptive boundary), the author studies the long - time behavior of the empirical measure under the condition \(\{t < \tau\}\), where \(\tau\) is the time of the first hitting of the boundary.
- The author defines the quasi - invariant probability measure \(\mu_0\) and discusses the long - time behaviors of the conditional expectations \(E_\nu(W_p(\mu_t, \mu_0)\mid\tau > t)\) and \(W_p(E_\nu(\mu_t\mid\tau > t), \mu_0)\).
4. **Specific convergence rates of the Wasserstein distance in different dimensions**:
- For different manifold dimensions \(d\), the author analyzes in detail the specific convergence rate of \(E_\nu[W_2(\mu_t, \mu_0)^2\mid\tau > t]\), and in particular, gives the exact convergence speeds in the cases of \(d\leq3\) and \(d\geq4\).
### Formula summary
- **p - Wasserstein distance**:
\[
W_p(\mu_1, \mu_2):=\left(\inf_{\pi\in C(\mu_1, \mu_2)}\int_{M\times M}\rho^p d\pi\right)^{\frac{1}{p}}, \quad \mu_1, \mu_2\in P,
\]
where \(C(\mu_1, \mu_2)\) is the set of all couplings of \(\mu_1\) and \(\mu_2\).
- **Empirical measure**:
\[
\mu_t:=\frac{1}{t}\int_0^t\delta_{X_s}ds, \quad t > 0,
\]
where \(\delta_{X_s}\) is the Dirac measure at \(X_s\).
- **Ergodicity condition**:
\[
\lim_{t\rightarrow\infty}\|P_t^*\nu - \mu\|_{L^2(\mu)} = 0,
\]
where \(\mu\) is the invariant probability measure.
- **Quasi - ergodicity condition**:
\[
\lim_{t\rightarrow\infty}E_\nu(f(X_t)\mid\tau > t)=\mu(f), \quad \nu\in P_0,
\]
where \(\nu\in P_0\) means that \(\nu\) satisfies \(E_\nu[1_{\{\tau > t\}}]>0\) for all \(t > 0\).
Through these formulas and theoretical derivations, the author reveals the asymptotic behavior and convergence speed of the empirical measure of different types of Markov processes under the Wasserstein distance.