# Intuition of Markov structure and how variables relate in the structure

Assume three r.v. $X_1, X_2, X_3$. They are conditionally independent in the following way:

$X_1 \perp X_3 | X_2$

We have that:

$$P(X_3 | X_2) = P(X_3| X_2, X_1)$$

In the notes I am reading it says

Eq. expresses that $X_3$ only depends on $X_1$ through $X_2$.

This statement seems to be to be an intuitive one. However, I am not sure what it means. For me it seems quite the opposite, given $X_2$, information about $X_1$ doesn't change the distribution for $X_3$.

Then the notes proceed to say:

We say such a set of variables form a Markov chain for which we use the notation $X_1 \rightarrow X_2 \rightarrow X_3$

I take that the notation with the arrows says that $X_2$ depends on $X_1$ and that $X_3$ depends on $X_2$, but that $X_3$ depends only on $X_1$ "through" $X_2$ (whatever "through" means). I would like to understand what these ideas mean as precisely as I can in the context of conditional independence and specially, the last diagram and what "through $X_2$ means".

Furthermore, another thing that confuses me is it says "$X_3$ inlet depends on $X_1$ through $X_2$". The word only seems extreme to be because just because we know $X_1 \perp X_3 | X_2$, it does't not mean $P(X_3) =P(X_3 | X_1)$. It isn't the only way, $X_3$ might depend on $X_1$ directly given no knowledge of $X_2$...right?

NOTE: I am aware that the conditional independence condition is symmetric and we achieve the same ideas but the other way round. The mathematical ideas/concepts are confusing, not the algebra. I want the intuition/concepts to be clarified.

The idea is that a Markov chain doesn't have a memory of past events. If I'm playing a game of chess with someone, I might use what I have learned about my opponent's strategy from all his/her previous moves, which would mean my moves would not follow a Markov chain.

However, if before I make each move you wipe my memory of the game so far, my play will follow a Markov chain; it only depends on the position immediately before I move, because that is all the information I have. My move depends on the previous moves in the game only because they got the board to a certain position.

It is true that "given $X_2$, information about $X_1$ doesn't change the distribution for $X_3$," but that means exactly the same thing that the book is saying. $X_3$ does depend on $X_1$, but only because $X_2$ depends on $X_1$; thus the dependence of $X_3$ on $X_1$ is through $X_2$, just as White's second move depends on White's first move only in so far as it defined the position of the board before Black moved.

(Note: in the chess analogy, $X_n$ is a position, not a move; otherwise it would clearly not be a Markov Chain)

• I think I understand what you say but its so counter intuitive for me. Because what you are saying is that since $X_1$ affects $X_2$, then $X_3$ is affected by $X_1$ only through $X_2$, right? But because $X_2$ is already known then we have established that $X_3$ is independent of $X_1$. But established that $X_1$ still affect $X_3$ but through $X_2$. I am not sure if you see the pattern, but It seems circular logic. Is it dependent or is it not? It a little weird. – Pinocchio Jun 27 '14 at 1:54
• Furthermore, another thing that confuses me is it says "$X_3$ inlet depends on $X_1$ through $X_2$". The word only seems extreme to be because just because we know $X_1 \perp X_3 | X_2$, it does't not mean $P(X_3) =P(X_3 | X_1)$. It isn't the only way, $X_3$ might depend on $X_1$ directly given no knowledge of $X_2$...right? – Pinocchio Jun 27 '14 at 2:15
• Or are you saying that, even though $P(X_3|X_2)$ is now conditionally independent of $X_1$, the probability of $X_2$ might have changed because $P(X_1) \neq P(X_1|X_2)$ which in turn affect which value $X_2$ might take (which obviously affect the value $X_2$ might take? – Pinocchio Jun 27 '14 at 2:18