Abstract
We argue that Algorithmic Information Theory (AIT) admits a principled way toquantify outliers in terms of so-called randomness deficiency. For theprobability distribution generated by a causal Bayesian network, we show thatthe randomness deficiency of the joint state decomposes into randomnessdeficiencies of each causal mechanism, subject to the Independence ofMechanisms Principle. Accordingly, anomalous joint observations can bequantitatively attributed to their root causes, i.e., the mechanisms thatbehaved anomalously. As an extension of Levin's law of randomness conservation,we show that weak outliers cannot cause strong ones when Independence ofMechanisms holds. We show how these information theoretic laws provide a betterunderstanding of the behaviour of outliers defined with respect to existingscores.