A relationship between randomized manipulation and parameter independence

We argue that a Bayesian will usually need to specify a joint prior density of the conditional probabilities of Causal Bayesian network (CBN). We show that in order to do this, under very mild conditions it will be necessary to demand that this joint prior density exhibits the properties of local and global independence. To make the connection between prior independence and causality, it is first necessary to strengthen slightly the assumptions of factorization invariance under manipulation which induces randomized intervention. We introduce the hypercausal BN (HCBN) that asserts a set of factorizations of densities which are invariant to a class of "do" operations larger than those considered by Pearl. We show that if a BN is assumed to be hypercausal, then the prior distributions on the probabilities of the idle system must exhibit local and global independence