Imperial College London Turns Causal Discovery Into a Local Information-Theory Read
A new theoretical framework links Partial Information Decomposition atoms to causal graph structure, letting you infer edges locally without searching the full graph space.
ProTakeaways
- First theoretical link between Partial Information Decomposition atoms and causal graph structure in Bayesian networks and hypergraphs
- Unique information identifies direct causal neighbors; synergy identifies colliders and co-parents
- Enables local causal discovery: characterize a variable's neighborhood without global graph search
- Extended to Bayesian hypergraphs, distinguishing parents, children, co-heads, and co-tails
- Uses bivariate PID with supervariables to avoid the Dedekind-number explosion of multivariate PID
- No code released; cannot break Markov equivalence and depends on scalable PID estimators
Figuring out which variables cause which in a complex system usually means running conditional independence tests across every possible subset of variables, or scoring candidate graphs with algorithms like hill-climbing. Both approaches treat causality as a global optimization problem. A new paper from Imperial College London proposes a different route: read the causal structure directly off the information-theoretic footprint of each variable, one neighborhood at a time.
The work establishes the first formal correspondence between Partial Information Decomposition (PID) components and causal roles in both Bayesian networks and the more expressive Bayesian hypergraphs. If the theory holds up in practice, it turns causal discovery into a local computation rather than a global search.
Why pairwise graphs keep tripping over higher-order structure
Classical causal discovery relies on directed acyclic graphs where every edge connects exactly two variables. These edge-based models can in principle capture higher-order interactions, but inferring them requires aggregating information across the global graph structure, since graphs intrinsically encode pairwise relations. Such higher-order interactions are crucial in neural, biological and social systems, where a single event may depend on the joint state of many variables at once.
PID, introduced by Williams and Beer, is a framework from information theory that splits the mutual information a set of sources carries about a target into three kinds of contributions: redundant (information any source alone could provide), unique (information only one specific source has), and