I am learning about Bayesian networks (and PGMs in general) and I am stuck into this. Basically I'm trying to find a mistake in my reasoning.

Consider this example (copied from https://youtu.be/Nis3O4CVJAU?t=632):

enter image description here

I can obtain the given correct result by using the chain rule if I traverse the nodes from top to bottom (i.e. from $N$ to $1$):

\begin{align} p(x_1, \dots, x_N) &= \prod_{k=1}^N p(x_k | x_{k-1}, \dots, x_1) \\ &= p(x_N | x_{N-1} \dots, x_1) \cdots p(x_2 | x_1) p(x_1) \end{align}

In this case $N=7$ and, from the graph: \begin{align} p(x_7 | x_6 \dots, x_1) &= p(x_7 | x_4, x_5) \\ p(x_6 | x_5 \dots, x_1) &= p(x_6 | x_4) \\ &\cdots \\ p(x_3 | x_2, x_1) &= p(x_3) \\ p(x_2 | x_1) &= p(x_2) \\ p(x_1) &= p(x_1) \\ \end{align}

The problem appears when I try to traverse the nodes from $1$ to $N$:

\begin{align} p(x_1, \dots, x_N) &= p(x_1 | x_{2}, \dots, x_7) p(x_2 | x_{3}, \dots, x_7) \cdots p(x_7) \\ \end{align}

For the root nodes, this seems reasonable to me: \begin{align} p(x_1 | x_{2}, \dots, x_7) &= p(x_1)\\ p(x_2 | x_{3}, \dots, x_7) &= p(x_2)\\ p(x_3 | x_{4}, \dots, x_7) &= p(x_3)\\ \end{align}

But then I have:

\begin{align} p(x_4 | x_{5}, \dots, x_7) &= p(x_4) \\ p(x_5 | x_{6}, x_7) &= p(x_5) \\ p(x_6 | x_7) &= p(x_6) \\ p(x_7) &= p(x_7) \\ \end{align}

Which, on the one hand, has to be wrong, because: $$ p(x_1, \dots, x_N) =^{??} p(x_1)\cdots p(x_7) $$

And therefore the variables would be independent (!?).

But, on the other hand, we have to apply the property defined in the graph (is this right??): $$p(x_3 | x_1, x_2, x_4, x_5, x_6, x_7) = p(x_3)$$

So, I'm guessing the mistake is that it doesn't follow from that property that $$p(x_3 | x_6, x_7) = p(x_3)$$

But if that's the case, why is it the following correct? $$ p(x_3 | x_1, x_2, x_4, \dots, x_7) = p(x_3) \implies p(x_3 | x_1, x_2) = p(x_3) $$

I suspect I may have the definition of the property wrong, but I'm not sure.

  • 1
    $\begingroup$ You have to be careful, $x_1, x_4$ and $x_5$ are dependent, in particular $p(x_1 | x_2, ..., x_7) = p(x_1 | x_4, x_5)$. $\endgroup$
    – mbartczak
    Sep 13, 2019 at 22:44

1 Answer 1


Your independence assumptions are in fact not reasonable at all.

I'm not sure what gave you the idea that, for example,

$$p(x_1 | x_2,\ldots,x_7)=p(x_1).$$

No variable in a Bayesian network can be assumed to be independent of its immediate neighbors.


The direction of the arrows inform the dependence structure for variables which are not immediate neighbors. Their development was an attempt to formalize human intuition about causality.

There should be some discussion in your text about d-separation, Markov blankets, colliders, etc. but essentially for A-junctions, as seen in the relationship between $x_1,x_4,x_5$. The variables $x_4$ and $x_5$ are independent precisely when one conditions on $x_1$, while for V-junctions, such as seen in the relationship between $x_1,x_2,x_4$, the variables $x_1$ and $x_2$ are independent precisely when one doesn't condition on $x_4$.

  • $\begingroup$ Oh, I assumed (don't know why) that as it was a DAG, the dependency was only in the direction of the arrow, and now I understand it is complete nonsense (e.g. in a two-node graph, dependency has to be reciprocal). But then, what does the "directed" part add in comparison to an acyclic Markov network? $\endgroup$
    – jp48
    Sep 13, 2019 at 22:59
  • 1
    $\begingroup$ @jp48 I've added some more material to the answer on the purpose of the arrows, however for a proper treatment you'll need to read the relevant parts of your text. $\endgroup$
    – Set
    Sep 14, 2019 at 2:09
  • $\begingroup$ Thanks a lot. I'm trying to find a suitable text and it's surprisingly difficult. I have some background on estimation theory and I'm looking for a middle ground between entry-level and texts that assume too much knowledge; e.g. measure theory, σ-algebras, etc. may possibly be "overkill" (I plan to learn about that in the future, anyways). The best I could find is "Artificial Intelligence - A Modern Approach" by Russel & Norvig. Do you have any suggestions? $\endgroup$
    – jp48
    Sep 14, 2019 at 15:24
  • 1
    $\begingroup$ @jp48 you don't need measure theory for probabilistic graphical models, just a solid understanding of calculus based probability theory. I would recommend either Daphne Koller's Probabilistic Graphical Models: Principles and Techniques, or the relevant chapters in Kevin Murphy's Machine Learning: A Probabilistic Perspective. $\endgroup$
    – Set
    Sep 14, 2019 at 15:28
  • $\begingroup$ Thanks again! Those look certainly what I wanted. $\endgroup$
    – jp48
    Sep 14, 2019 at 15:37

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .