Distinguishing between cause and effect via the algorithmic Markov condition

Dominik Janzing and Bernhard Schoelkopf (Max Plank Institute for Biological Cybernetics, Germany)

Inferring the causal structure that links n observables is usually based upon detecting statistical dependences and choosing simple graphs that make the joint measure Markovian. Here we argue why causal inference is also possible when only single observations are present. We develop a theory how to generate causal graphs explaining similarities between single objects. To this end, we replace the notion of conditional stochastic independence in the causal Markov condition with the vanishing of conditional *algorithmic* mutual information. We prove the equivalence of three versions of our algorithmic Markov condition in analogy to a known equivalence theorem for three versions of the statistical Markov condition. We justify the algorithmic causal Markov condition by an algorithmic model of causality, where every effect is influenced by its causes in a computable way. In contrast to causal inference methods that rely on statistical dependences, our theory implies rules for distinguishing between the causal hypotheses X --> Y and Y --> X for two random variables X,Y. This is because the causal graphs relating the individual observations of the statistical ensemble induce different sets of algorithmic independences for the two cases. The impact of our theory is thus threefold. It provides
  1. a graphical-model-based framework for inferring causal relations among individual objects
  2. a probability-free formulation of common independence-based statistical causal inference
  3. a foundation for new statistical causal inference rules by also taking into account non-statistical algorithmic dependences.


NIPS 2008 workshop on causality