Markov chains: accessibility and state classes

Dependencies:

  1. Markov chain
  2. Chapman-Kolmogorov equation
  3. /sets-and-relations/equivalence-relation
  4. /sets-and-relations/poset

Let $X = [X_0, X_1, \ldots]$ be a markov chain with transition function $P$ and state space $D$.

State $j$ is said to be accessible from state $i$ iff $P^{(t)}(i, j) > 0$ for some $t \ge 0$. We can prove that $j$ is accessible from $i$ iff there exist states $k_0, k_2, \ldots, k_r$ such that $P(k_t, k_{t+1}) > 0, \forall 0 \le t < r$ and $k_0 = i$ and $k_r = j$. We can also prove that accessibility is a reflexive and transitive relation.

States $i$ and $j$ are said to communicate iff $i$ is accessible from $j$ and $j$ is accessible from $i$. We can prove that communication is an equivalence relation. Hence, communication partitions the state space into equivalence classes, called state classes. If there is just one state class (i.e., all states communicate), then the markov chain is called irreducible.

State class $J$ is said to be accessible from state class $I$ iff $\exists j \in J$ and $\exists i \in I$ such that $j$ is accessible from $i$. Then we can show that $\forall i \in I$ and $\forall j \in J$, $j$ is accessible from $i$. Furthermore, we can prove that accessibility induces a partial ordering on the state classes.

Intermediate transitions of accessibility

$j$ is accessible from $i$ iff $\exists m \ge 0$ such that $P^{(m)}(i, j) > 0$. Let $k_0 = i$ and $k_m = j$. \begin{align} & P^{(m)}(i, j) \\ &= \Pr(X_m = j \mid X_0 = i) \\ &= \sum_{k_1 \in D} \sum_{k_2 \in D} \ldots \sum_{k_{m-1} \in D} \Pr\left(\bigwedge_{t=1}^m X_t = k_t \mid X_0 = k_0\right) \\ &= \sum_{k_1 \in D} \sum_{k_2 \in D} \ldots \sum_{k_{m-1} \in D} \prod_{t=1}^m \Pr\left(X_t = k_t \mid \bigwedge_{s=0}^{t-1} X_s = k_s\right) \\ &= \sum_{k_1 \in D} \sum_{k_2 \in D} \ldots \sum_{k_{m-1} \in D} \prod_{t=1}^m \Pr\left(X_t = k_t \mid X_{t-1} = k_{t-1}\right) \tag{by markov property} \\ &= \sum_{k_1 \in D} \sum_{k_2 \in D} \ldots \sum_{k_{m-1} \in D} \prod_{t=1}^m P(k_{t-1}, k_t). \end{align} All these terms are non-negative. Hence, $P^{(m)}(i, j) > 0 \iff \exists k_0, k_2, \ldots, k_m$ such that $P(k_{t-1}, k_t) > 0, \forall 1 \le t \le m$ and $k_0 = i$ and $k_m = j$.

Proof that accessibility of states is a reflexive and transitive relation

$i$ is accessible from $i$, since $P^{(0)}(i, i) = 1 > 0$. Hence, accessibility is reflexive.

Suppose $j$ is accessible from $k$ and $k$ is accessible from $i$. Then $\exists m \ge 0$ and $\exists n \ge 0$, such that $P^{(m)}(i, k) > 0$ and $P^{(n)}(k, j) > 0$. Then by the Chapman-Kolmogorov equation, \[ P^{(m+n)}(i, j) = \sum_{r} P^{(m)}(i, r)P^{(n)}(r, j) \ge P^{(m)}(i, k)P^{(n)}(k, j) > 0. \] Hence, $j$ is accessible from $i$. Hence, accessibility is transitive.

Proof that communication is an equivalence relation

$i$ communicates with $i$, since $P^{(0)}(i, i) = 1 > 0$. Hence, communication is reflexive.

It is easy to see that communication is symmetric.

Suppose $i$ and $k$ communicate and $k$ and $j$ communicate. By the transitivity of accessibility, we get that $i$ and $j$ communicate. Hence, communication is transitive.

Accessibility of state classes

Suppose state class $J$ is accessible from state class $I$. Then $\exists i' \in I$ and $\exists j' \in J$ such that $j'$ is accessible from $i'$. Now let $i \in I$ and $j \in J$. Since $i$ and $i'$ are in the same class $I$, $i'$ is accessible from $i$. Since $j$ and $j'$ are in the same class $J$, $j$ is accessible from $j'$. Hence, by transitivity of accessibility of states, $i$ is accessible from $j$.

Proof that accessibility induces a partial ordering on state classes

A state class $I$ is accessible from itself, since for any $i \in I$, $i$ is accessible from itself.

Suppose state class $K$ is accessible from state class $I$ and state class $J$ is accessible from state class $K$. This means that $\exists k_1 \in K$ and $\exists i \in I$ such that $k_1$ is accessible from $i$, and $\exists j \in J$ and $\exists k_2 \in K$ such that $j$ is accessible from $k_2$. Then $k_2$ and $k_1$ are accessible from each other, since they belong to the same state class $K$. Hence, by the transitivity of accessibility of states, we get that accessibility of state classes is also transitive.

Suppose two distinct state classes $I$ and $J$ are accessible from each other. Then $\exists i_1 \in I$ and $\exists j_1 \in J$ such that $i_1$ is accessible from $j_1$, and $\exists i_2 \in I$ and $\exists j_2 \in J$ such that $j_2$ is accessible from $i_2$. $i_1$ and $i_2$ are accessible from each other since they are from the same state class $I$. $j_1$ and $j_2$ are accessible from each other since they are from the same state class $J$. By transitivity of accessibility, this means that $j_1$ is accessible from $i_1$. Hence, $j_1$ and $i_1$ communicate. But this is a contradiction, since $i_1$ and $j_1$ are in different state classes. Hence, no two state classes are accessible from each other. Hence, accessibility of state classes is anti-symmetric. Hence, accessibility of state classes is a partial ordering.

Dependency for:

  1. Markov chains: finite sink is positive recurrent
  2. Markov chains: positive recurrence is a class property
  3. Markov chains: long run proportion is inverse of time to reenter (incomplete)
  4. Markov chains: recurrence is a class property
  5. Markov chains: finite sink is recurrent (incomplete)
  6. Markov chains: recurrent class is sink

Info:

Transitive dependencies:

  1. /analysis/topological-space
  2. /sets-and-relations/countable-set
  3. /sets-and-relations/poset
  4. /sets-and-relations/equivalence-relation
  5. /sets-and-relations/de-morgan-laws
  6. σ-algebra
  7. Generated σ-algebra
  8. Borel algebra
  9. Measurable function
  10. Generators of the real Borel algebra (incomplete)
  11. Measure
  12. σ-algebra is closed under countable intersections
  13. Group
  14. Ring
  15. Semiring
  16. Matrix
  17. Identity matrix
  18. Probability
  19. Conditional probability (incomplete)
  20. Random variable
  21. Markov chain
  22. Chapman-Kolmogorov equation