0
$\begingroup$

I'm fairly new to Markov chains. At the moment, I'm trying to understand why a finite state space Markov chain cannot have any null-recurrent states.

Searching on math.SE, I found this answer, which argues as follows:

Consider a communicating class $C$ containing a null, recurrent state. Then $p_{ij}(n)\to 0$ for every $i,j\in C$, but since this class $C$ is closed this gives the contradiction $$1=\lim_n \sum_j p_{ij}(n)=0.$$

My question is: How can we show that $p_{ij}(n) \to 0$ for all $i, j$ in $C$?

I suppose it's sufficient to show that, if $k$ is the null-recurrent state, then $p_{kk}(n) \to 0$. Indeed, for any $i, j \in C$, there exist $m_1$ and $m_2$ such that $p_{ki}(m_1) > 0$ and $p_{jk}(m_2) > 0$, so if $p_{kk}(n) \to 0$, then $$ p_{ij}(n) \leq \frac{p_{kk}(m_1 + n + m_2) }{ p_{ki}(m_1) p_{jk}(m_2) }\to 0.$$

But then, how do I show that $p_{kk}(n) \to 0$? The fact that $k$ is null-recurrent tells us that $\sum_{n = 1}^\infty n f_{kk}(n) = \infty$, where $f_{kk}(n)$ is the probability of the first return from $k$ to $k$ taking place after $n$ steps. I don't see the link between $\sum_{n = 1}^\infty n f_{kk}(n)$ being infinite and $p_{kk}(n)$ tending to zero.

Edit: I thought of using the fact that $$ P_{kk}(s) = 1 + F_{kk}(s) P_{kk}(s), \ \ \ \ -1 < s \leq 1, $$ where $$ P_{kk}(s) := \sum_{n=0}^\infty p_{kk}(n) s^n, \ \ \ \ \ \ \ F_{kk}(s) := \sum_{n=1}^\infty f_{kk}(n)s^n$$ See page 12 of these notes. If you rearrange, and differentiate, you get $$ F_{kk}'(s) = - \frac{P_{kk}'(s)}{(P_{kk}(s))^2}. $$ If $\sum_{n = 1}^\infty n f_{kk}(n) = \infty$, then $F_{kk}'(1)$ must be infinite. But then I'm not sure what to do with the right-hand side.

$\endgroup$
  • $\begingroup$ If $T_k$ is the time of first return to $k$, then $\mathbb{E}[T_k \mid X_0=k] \neq \sum_{n=1}^\infty nf_{kk}(n)$. This is because $f_{kk}(n)$ is $\mathbb{P}(X_n=k \mid X_0=k)$, not $\mathbb{P}(X_n=k\text{ and } X_i \neq k, 1 \leq i <k \mid X_0=k)$. So you cannot say that $\sum_{n=1}^\infty nf_{kk}(n)=\infty$. $\endgroup$ – kccu Feb 10 at 17:52
  • $\begingroup$ @kccu My $f_{kk}(n)$ is defined as $P(X_n = k {\rm \ and \ } X_i \neq k, 1 \leq i < k | X_0 = k)$. Whereas my $p_{kk}(n)$ is defined as $P(X_n = k | X_0 = k)$. $\endgroup$ – Kenny Wong Feb 10 at 18:02
  • $\begingroup$ My apologies, I read too quickly and confused $f_{kk}$ with $p_{kk}$. $\endgroup$ – kccu Feb 10 at 21:23
1
$\begingroup$

I think that claim is not trivial. In the book referenced in your link, that statement is observed as a corollary to the ergodic theorem stating that $p_{ij}(n) \to \frac{1}{\mathbb E_x T_x}$ for any irreducible aperiodic Markov chain. You can find a proof of that result here, for instance.

If all you want to prove is your original claim (that all irreducible finite Markov chains are positive recurrent), I think there's an easier way to do it than by that lemma. Assume aperiodicity for simplicity, but periodic chains just make the proof more annoying (rather than prevent the result from being true). The sketch of the proof is:

  1. For an irreducible, aperiodic finite Markov chain, there exists some $m$ such that $p_{ij}(m) > 0$ for all state pairs $\{i, j\}$.

  2. Moreover, because the chain is finite you can strengthen the above to say that $p_{ij}(m) > \epsilon$ for some universal $\epsilon > 0$.

  3. Hence, you can dominate the excursions of $X_n$ from $x$ to $x$ with a geometric random variable by conditioning on the behavior on every $m^{\text{th}}$ step.

  4. Since the geometric random variable has a finite expectation, so do the excursion lengths.

EDIT: Expanding the answer to address the steps to show step 1 above. I'll work with the following definition of aperiodicity: a state $x$ is aperiodic if $\gcd(k:p_{xx}(k) > 0) = 1$. In all the following, I'll also assume irreducibility of the Markov chain. Without irreducibility, some version of these basic ideas all still hold, but they're just more obnoxious to write down; you'd just have to address an irreducible subclass of states.

Lemma: If state $x$ is aperiodic, then there exists $N$ such that $p_{xx}(n) > 0$ for all $n \geq N$.

Corollary: For the aperiodic state $x$ above, there exists $N'$ such that for all other states $y$ and all $n \geq N'$, $p_{xy}(n) > 0$.

Corollary to the corollary: If a chain has one aperiodic state, all states are aperiodic.

$\endgroup$
  • $\begingroup$ Thanks very much for this! I understand points 2, 3 and 4. Could you please give me a hint as to how to prove point 1? $\endgroup$ – Kenny Wong Feb 12 at 20:48
  • $\begingroup$ I tried to fill in the gaps a bit -- let me know if I should go further. $\endgroup$ – Aaron Montgomery Feb 13 at 3:04
  • $\begingroup$ That's great - thanks! $\endgroup$ – Kenny Wong Feb 13 at 8:07

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.