Suppose I have a conditional discrete distribution $P_{M|A}(m,a)$. If I take sum over $m$, i.e. $\sum_{m}P_{M|A}(m,a)$, do I get $f_A(a)$, where $f_A(a)$ is the pmf of the random variable A?

Thanks for clarifying. I am trying to read about graphical models from http://ai.stanford.edu/~paskin/gm-short-course/lec2.pdf. Here are the equations I am confused about:
\begin{align} &∑_e∑_a∑_{mp}E(e)p_B(b)p_{A|EB}(a,e,b)p_{J|A}(true,a)p_{M|A}(m,a) \\ &\qquad = ∑_e∑_{ap}E(e)p_B(b)p_{A|E}B(a,e,b)p_{J|A}(true,a)∑_mp_{M|A}(m,a) \\ &\qquad = ∑_e∑_{ap}E(e)p_B(b)p_{A|EB}(a,e,b)p_{J|A}(true,a)ψ_A(a) \end{align}
I am assuming $ψ_A(a)$ is pmf of $A$, and there comes the question, that how to marginalize a conditional density function. Ideally, $\sum_mp_{M|A}(m,a)$ should be equal to 1. What am I missing?

(I have asked this as a new question here: http://stats.stackexchange.com/questions/45622/marginalizing-a-conditional-distribution-for-bayesian-networks.)


For each $a$, $P_{M|A}(m|A=a)$ is a pmf, so if you sum over all $m$ in its domain, you should get

$$ \sum_m P_{M|A}(m|A=a) = 1 $$

Otherwise, the function is not a valid pmf. The pmf of $A$ is the sum of joint pmf $P_{A,M}(a,m)$ over all $m$,

$$ f_{A}(a)= \sum_{m} P_{A,M}(a,m) $$

which cannot be recovered from $P_{M|A}(m|A=a)$ without already having the marginal pmf.

| cite | improve this answer | |

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.