In the classical approach to causal discovery, researchers rely on two pillars: the Causal Markov Condition (which says the graph's separations imply independencies in the data) and the Causal Faithfulness Condition. Faithfulness is the "no accidents" rule: it assumes that every conditional independence we observe in the data must be explained by the structure of the causal graph, rather than by a coincidental cancellation of path coefficients.
However, in real-world datasets—especially those with dense connections, deterministic relationships, or grouped variables—strict faithfulness is frequently violated. Recent research has shifted toward relaxed faithfulness assumptions, developing algorithms that can recover causal structures even when the data doesn't perfectly mirror the graph.
The Problem with Strict Faithfulness
Strict faithfulness is a fragile assumption. If two causal paths between variables $X$ and $Y$ happen to have coefficients that sum to zero, $X$ and $Y$ will appear independent even though a direct causal link exists. Standard algorithms like PC or GES would see this independence and incorrectly remove the edge.
As Sanghack Lee and Vasant G Honavar demonstrate in Lifted Representation of Relational Causal Models Revisited, this problem is exacerbated in relational data and macro-variable groupings. When we aggregate variables, the abstract representation often fails to capture all independencies, leading to a fundamental breakdown of strict faithfulness.
A Hierarchy of Weaker Assumptions
To make causal discovery more robust, researchers have proposed a hierarchy of weaker constraints that prioritize different aspects of the graph structure:
- Causal Minimality: The weakest alternative, requiring only that the true causal graph is the simplest one (the one with the fewest edges) that satisfies the Markov condition. Asish Ghoshal and Jean Honorio show in Learning linear structural equation models in polynomial time and sample complexity that minimality is often sufficient to recover the structure of linear SEMs without needing full faithfulness.
- Adjacency and Orientation Faithfulness: Instead of requiring all independencies to be faithful, these assumptions only require that if two variables are adjacent in the graph, they are dependent in the data (adjacency), and that certain "collider" structures are detectable (orientation). Jonas Wahl and colleagues found that while strict faithfulness is lost when grouping variables, adjacency faithfulness is often preserved, allowing for conservative discovery.
- Triangle and Shielded-Collider Faithfulness: These are localized versions of the assumption. Jiji Zhang and Wolfgang Mayer utilize SAT-based solvers to encode these local constraints, finding that they significantly reduce logical contradictions in noisy data.
Robust Algorithms for Noisy Environments
The shift toward relaxed assumptions has led to a new generation of algorithms. For instance, the GRaSP algorithm, developed by Wai-Yin Lam and Joseph D. Ramsey, uses permutation-based search to navigate variable orderings. By weakening the faithfulness requirement, GRaSP can accurately recover dense graphs where path cancellations are likely.
Similarly, in dynamical systems, Søren Wengel Mogensen introduces Ancestral Faithfulness in Causal screening in dynamical systems. This assumes that if a directed path exists, its influence won't be perfectly neutralized. This allows for the identification of causal ancestors in complex stochastic processes using fewer statistical tests.
Why It Works: Statistical Regularization
Perhaps the most counterintuitive finding in this area is that complexity can be a virtue. Arash Mehrjou and colleagues argue in Theoretical Guarantees for Causal Discovery on Large Random Graphs that as the number of variables grows, the error rate in orienting edges actually vanishes. In high-dimensional systems, the sheer scale of the network acts as a form of statistical regularization, making the interventional signals of causal parents more detectable against the background noise, provided we use a relaxed faithfulness threshold.
By moving away from the "all-or-nothing" requirement of strict faithfulness, these researchers have created a more flexible toolkit for causal inference that is better suited to the messy, interconnected realities of modern data science.