View source for Causal Markov condition

The '''Causal Markov condition''' (CMC) is the foundational principle that connects causal structure to probabilistic structure: it states that every variable in a causal system is probabilistically independent of its non-effects, conditional on its direct causes. In the language of '''[[Probabilistic graphical model|graphical models]]''', this means that once you condition on a variable's parents in the causal graph, the variable is independent of all other variables except its descendants. The CMC is not an empirical claim about the world; it is a semantic stipulation about what it means for a graph to represent causation. If the graph is causally interpreted, the Markov condition must hold by definition.

Formally, given a directed acyclic graph G and a probability distribution P over its nodes, the CMC asserts that P factorizes according to G: each variable is conditionally independent of its non-descendants given its parents. This factorization is what makes '''[[Causal discovery]]''' possible at all. Without it, observing that X and Y are independent given Z would tell us nothing about whether Z mediates a causal relationship between them. The '''[[PC algorithm]]''', '''[[GES]]''', and every other constraint-based causal discovery algorithm implicitly assumes the CMC every time it reads a conditional independence test as evidence about graph structure.

== d-Separation and the Semantics of Causation ==

The operational content of the CMC is captured by '''[[d-separation]]''', a graph-theoretic criterion for reading conditional independencies from a directed acyclic graph. Two sets of nodes X and Y are d-separated by a third set Z if every undirected path between X and Y is blocked by Z according to specific rules involving colliders (nodes where two arrows meet head-to-head). When X and Y are d-separated by Z, the CMC guarantees that X is independent of Y given Z in the probability distribution. Conversely, if the distribution is faithful — that is, if it contains no ''accidental'' independencies — then every conditional independence corresponds to a d-separation in the graph.

The pairing of CMC with the '''[[Faithfulness assumption]]''' creates a powerful inferential engine. The CMC says: structure implies probability. Faithfulness says: probability implies structure. Together, they license the inference from patterns of conditional independence to equivalence classes of causal graphs. But this inferential license is conditional on both assumptions holding, and neither holds universally. The CMC fails when there are unmeasured common causes (violating '''[[Causal sufficiency]]'''), when causation operates through feedback loops (violating acyclicity), or when the variables themselves are not well-defined (violating the very premise of graph-based causation).

== The Limits of Markovian Causation ==

The CMC encodes a particular metaphysics: the world is composed of discrete variables connected by directed, acyclic causal arrows, and the absence of an arrow means the absence of a direct causal influence. This is a powerful and tractable framework, but it is not the only framework. '''[[Dynamical systems]]''' theorists argue that causation is better understood as the evolution of trajectories in a state space, where the notion of direct