1
$\begingroup$

Let's say that we have a graph A->B->C, and its coditional independence representation is like follows:

| a0   | a1      |      
|:-----|--------:|   
| 0.6  | 0.4     |  

| A    | P(b0|A) | P(b1|A) |  
|:-----|--------:|--------:|
| a0   | 0.30    | 0.70    |
| a1   | 0.20    | 0.80    |

| B    | P(c0|B) | P(c1|B) |  
|:-----|--------:|--------:|
| b0   | 0.20    | 0.80    |
| b1   | 0.90    | 0.10    |

This graph is valid since every probability is non-negative and the sum of the graph distribution is 1: $P(A, B, C)=P(A)P(B|A)P(C|B)=0.6*0.3*0.2+0.6*0.3*0.8+\cdots + 0.4*0.8*0.9+0.4*0.7*0.1=1$

But what if I change the C to A? The first two probability tables don't change:

| a0   | a1      |      
|:-----|--------:|   
| 0.6  | 0.4     |  

| A    | P(b0|A) | P(b1|A) |  
|:-----|--------:|--------:|
| a0   | 0.30    | 0.70    |
| a1   | 0.20    | 0.80    |

But how to draw the third one?

| B    | P(a0|B) | P(a1|B) |  
|:-----|--------:|--------:|
| b0   | 0.60    | 0.40    |
| b1   | 0.60    | 0.40    | 

I am wonder if it is an infinit loop and the probability tables are not possible for such graphs. Any suggestions are hight appreciated. Thanks.

$\endgroup$
3
  • $\begingroup$ What do you mean by changing $C$ to $A$? Are you asking $P(A|B)$ when the network is still in the same form, i.e. $A\rightarrow B\rightarrow C$ $\endgroup$
    – gunes
    Commented Jan 24, 2019 at 12:44
  • $\begingroup$ @gunes I want to see what would happen. I want to make sure the network will be invalid if I substitute C by A(indicating the network has changed). If P(A|B) is just P(A) the sum would be larger than 1, I thought. $\endgroup$ Commented Jan 24, 2019 at 13:14
  • $\begingroup$ I think it would lead to an endless loop and the joint distribution would be summed larger than 1. $\endgroup$ Commented Jan 25, 2019 at 2:18

1 Answer 1

0
$\begingroup$

When you construct a network in the form $A\rightarrow B\rightarrow C$, since every node depends on only its parents, by definition the joint probability is written as $P(A,B,C)=P(A)P(B|A)P(C|B)$. When we set $C=A$, i.e. form a cyclic relationship, this formula turns into $$P(A,B,A)=P(A,B)=P(A)P(B|A)P(A|B)$$ Here, $P(A)P(B|A)=P(A,B)$, then to satisfy the equation it should be $P(B|A)=1$, assuming non-zero probabilities. Conversely thinking, we'll have $P(A|B)=1$. If they're both $1$, then $P(A,B)=P(A)$, which means $B=A$ or $A$ and $B$ are deterministic, i.e. with probability $1$, e.g. The world is spinning.

$\endgroup$
1
  • $\begingroup$ So is my head. LOL. $\endgroup$ Commented Jan 25, 2019 at 22:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.