Suppose we have a graphical model defined by a directed acyclic graph $G$. Suppose that a node $a$ divides $G$ into two connected components. By this I mean that:

$G=G_1 \cup G_2 \cup \{a\}$

$G_1 \cap G_2 = \emptyset$

and there are no edges between $G_1$ and $G_2$. ( ok technically by $G$ I mean the set of vertices of $G$, I hope the question is clear anyway... )

Is it true that, given $g_1 \in G_1$ and $g_2 \in G_2$:

$$ p(g_1|g_2,a)=p(g_1|a) ? [1]$$

I conjecture this from how graphical models are build, but is it true ? Of course it is simple to generalize [1] but I wanted to write a simple statement.


Yes it is true.

For example if you follow the d-separation procedure, at step 4, you will delete $a$ which necessarily disconnects $g_1$ and $g_2$, $\forall g_1\in\mathcal{G}_1$, $\forall g_2\in\mathcal{G}_2$, leading to the conditional independence.

  • $\begingroup$ Thanks. Do you have also a reference where such statements, like the one reported in bold in the notes you linked, is discussed ? $\endgroup$
    – Thomas
    Feb 2 at 12:18
  • $\begingroup$ The statement in bold is somehow the way Bayes Net are defined. Maybe you can look here for more details, implications etc. $\endgroup$
    – TheCG
    Feb 2 at 12:36

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.