Abstract:This paper integrates two strands of the literature on stability of general state Markov chains: conventional, total variation based results and more recent order-theoretic results. First we introduce a complete metric over Borel probability measures based on partial stochastic dominance. We then show that many conventional results framed in the setting of total variation distance have natural generalizations to the partially ordered setting when this metric is adopted.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to unify the stability results in the classical Markov chain theory with those in the Markov chain theory based on monotonicity. Specifically, the paper aims to construct a unified theoretical framework that can cover both the traditional stability theory based on total variation distance and the monotonicity - based method simultaneously, thereby providing a stability analysis tool for a wider range of Markov chains.
### Background and Motivation
1. **Classical Theory**: Traditionally, the stability analysis of Markov chains mainly relies on the Total Variation Distance. This method is very effective in many models, but for some models, the total variation convergence condition is too strict, resulting in inapplicability.
2. **Monotonicity Method**: In recent years, researchers have developed a new method based on monotonicity. This method does not require total variation convergence but utilizes the monotonic properties of the state - transition probabilities. This method is applicable to some models that cannot be handled by traditional methods, but it may not be powerful enough in other cases.
### Main Contributions of the Paper
- **Introduction of a New Metric**: The paper introduces a new metric \(\gamma\), which is a complete metric on the set of Borel probability measures on a partially ordered Polish space. This metric includes the total variation distance, the Kolmogorov consistent distance, and the Bhattacharya distance as special cases.
- **Unified Framework**: By adopting this new metric \(\gamma\), the paper shows how to unify the results of the classical theory and the monotonicity method under one framework. This framework not only covers all uniformly ergodic models but also extends to other partially monotonic models.
### Specific Examples
- **Example 1**: Consider a simple Markov chain \(\{X_t\}\) defined as:
\[
X_{t + 1}=\frac{X_t+W_{t + 1}}{2}
\]
where \(\{W_t\}_{t\geq1}\) is an independent and identically distributed Bernoulli(1/2) random sequence. For the state space \(S = [0,1]\), if \(X_t\) is rational, then \(X_{t + 1}\) is also rational; if \(X_t\) is irrational, then \(X_{t + 1}\) is also irrational. Therefore, for a rational number \(x\) and an irrational number \(y\), the distributions \(P_t(x,\cdot)\) and \(P_t(y,\cdot)\) are concentrated on disjoint sets, resulting in a total variation distance of always 2, that is:
\[
\|P_t(x,\cdot)-P_t(y,\cdot)\| = 2
\]
However, due to the monotonicity of \(X_{t+1}\) with respect to the current state \(X_t\), the monotonicity method can be used to prove the existence of a unique stationary distribution, and the distribution of \(X_t\) converges to this stationary distribution under this new metric \(\gamma\).
- **Example 2**: Consider an inventory model:
\[
X_{t + 1}=\begin{cases}
(X_t - W_{t + 1})^+&\text{if }X_t>0\\
(K - W_{t + 1})^+&\text{if }X_t = 0
\end{cases}
\]
where \(x^+=\max\{x,0\}\), \(\{W_t\}\) is independent and identically distributed, and assume that \(\ln W_t\) follows a standard normal distribution. The state space is taken as \(S = [0,K]\). For this model, the monotonicity method is not applicable, but the classical theory can easily handle it, proving the existence of a unique stationary distribution, and the distribution of \(X_t\) converges to this stationary distribution in total variation distance.
### Conclusion
The paper successfully unifies the classical theory and the monotonicity method by introducing the new metric \(\gamma\) and the concept of ordered affinity, providing a new tool for stability analysis of a wider range of Markov chain models. This framework not only covers the existing results but also extends to other partially monotonic models.