Inferring the algorithmic direction of causal pairs

Cristian Grozea – Fraunhofer FIRST

We suggest an approach for obtaining the "correct" direction of a causal pair of variables, based on the AIT (ref Chaitin for entropy H), using only absolute complexity. This is advantageous for practical approximations, as the absolute complexity of an object (as opposed to the conditional one) can be easily upper bounded by using a practical compression algorithm. We assume that an algorithmic process generates output from input: output=f(input,extrainfo) – where “extrainfo” is the “noise”, seen as supplementary information needed to implement a precise mapping (ref Janzig for similar setup). In this framework the generative mapping is implicitly reversible even if the function is not bijective! Specifically, even if the generative process was X->Y, another one always exists that goes Y->X, producing X=g(Y,otherinfo). In fact, the only thing that X can bring to the construction of Y is exactly the same information that Y can bring to the construction of X, that is their mutual information. We break this perfect reversibility by using the principle of least assumption, which amounts to choosing the direction that requires the least amount of extra information to build the target variable. As H(X)=H(X|Y)+I(X,Y), if H(X)>H(Y) then X is elected as “the cause”.

NIPS 2008 workshop on causality