conditional independence

Two events can be dependent on their own yet become independent once you know some third fact. This is conditional independence: A and B are conditionally independent given C when, inside the world where C is known, learning A tells you nothing more about B. The third fact C explains away the apparent connection between A and B.

The definition mirrors ordinary independence but everything is conditioned on C: A and B are independent given C when P(A and B given C) = P(A given C) times P(B given C). A homely example: among all people, having yellow-stained fingers and having lung cancer are positively associated — see one and the other becomes more likely. But that link runs through smoking. Condition on 'is a smoker', and within smokers, stained fingers and lung cancer become (roughly) independent: once you know someone smokes, their stained fingers no longer carry extra information about their cancer risk. The common cause C created the association, and conditioning on it removes it.

Conditional independence is the backbone of modern probabilistic modeling. It is what lets a naive Bayes spam filter treat words as independent given the class, and what makes Bayesian networks and hidden Markov models tractable — they encode exactly which variables become independent once others are observed. The crucial honesty point cuts both ways: conditional independence given C neither implies nor is implied by plain (unconditional) independence. Conditioning can create dependence just as easily as it can remove it.