I cannot understand how there can be positive recurrent and null recurrent Markov Chains. Markov Chains can be split up into transient and recurrent states, where recurrent means that it will be able to go back to that state sooner or later, as compared to a transient state whereby it may escape without ever being able to come back to the state.
Since by definition, a recurrent state means that the Markov chain will be able to return to the state in finite time, why is there a need to define another subset of recurrent Markov chain (null recurrent), whose definition (I feel, even though I know it's not true) violates the whole point of a recurrent Markov Chain in the first place?
Could someone please help with the intuition behind this?
Best Answer
Mars Plastic puts it rather nicely. Here are additional elements.
In order to better understand this concept of positive recurrent and null recurrent Markov chains, first it is good to set ourselves in a context where it becomes important.
One of the fundamental questions for Markov chains is whether there exists a stationary distribution. If you restrict yourself to finite state chains, then there is always one (Brouwer's fixed point theorem), and the notion of a null-recurrent state simply does not exist.
In the infinite case, you can start asking new questions. Even fully connected chains can fail to have a stationary distribution. It can be proven that if the chain is positive recurrent then it must exist, and $\pi(i) = 1/E[\tau_i]$.
If it's null recurrent, that means $\pi$ does not exist, but you still have a guarantee of returning to every state.
In other words, even if the concept of a mixing time does not make sense, you still have finite hitting times.