Causal discovery is the ambitious attempt to learn the "gears" of a system—the directed arrows of cause and effect—solely by observing its outputs. To bridge the gap between raw statistical correlations and structural causal links, researchers have long relied on a cornerstone assumption: Faithfulness.
In an ideal world, the Faithfulness assumption posits that every statistical independence we observe in data is the result of a missing link in the underlying causal graph. If two variables are independent, they aren't directly connected; if they are dependent, they are. However, as Philippe Brouillard and colleagues argue in The Landscape of Causal Discovery Data, real-world systems are rarely so cooperative. From biological feedback loops to perfectly balanced economic forces, the "Faithfulness" of our data is under constant threat, leading to "silent failures" where algorithms hallucinate or overlook critical causal relationships.
The Intuition of Faithfulness
To understand why this assumption is so vital—and so fragile—imagine a light switch ($X$) connected to a bulb ($Y$). If flipping the switch changes the light, they are dependent. Faithfulness assumes that if we observe no statistical relationship between $X$ and $Y$, it must be because there is no wire between them.
A violation occurs when there is a wire, but the effect is hidden. Imagine a second, hidden wire ($Z$) that acts as an inverter. If $X$ turns the light on through one path but simultaneously turns it off through another, the net effect is zero. To an observer, $X$ and $Y$ appear independent. A standard algorithm, relying on faithfulness, would incorrectly conclude that $X$ does not cause $Y$. As noted by Peter Spirtes in Causal discovery and inference, these "perfectly balanced opposing forces" render the underlying structure invisible to traditional constraint-based methods.
Why Faithfulness Fails in the Wild
The research indicates several distinct ways this assumption breaks down in practice:
- Path Cancellation: As described by Saber Salehkaleybar and AmirEmad Ghassami in their work on Linear Non-Gaussian Causal Models, multiple causal paths can cancel each other out. This is common in homeostatic biological systems where a change in one variable triggers a compensatory response in another.
- Deterministic Relationships: When variables are perfectly predictable from one another (e.g., in physics or dataset identifiers), standard independence tests fail. Wiebke Günther and Jakob Runge demonstrate in J-PCMCI+ that these deterministic "dummy variables" can mislead algorithms into pruning valid edges.
- Aggregation and Scale: Causal properties don't always scale. Jonas Wahl and Urmi Ninad show in Foundations of Causal Discovery on Groups of Variables that even if faithfulness holds for individual neurons or atoms, it frequently breaks down when we aggregate them into brain regions or climate zones.
Detecting and Bypassing the Trap
If we cannot trust faithfulness, how do we proceed? The research direction has shifted toward two strategies: falsification and identifiability.
Falsification involves checking if the algorithm's assumptions are consistent with the data. Philipp M. Faller and Francesco Locatello introduced Self-Compatibility in Evaluating Causal Discovery without Ground Truth. By checking if causal structures inferred from overlapping subsets of variables contradict each other, researchers can detect when faithfulness (or other assumptions) has been violated without needing to know the "true" graph.
Identifiability focuses on finding new mathematical "handles" that don't rely on independence tests alone.
- Functional Models: Jonas Peters and Bernhard Schölkopf proved in Identifiability of Causal Graphs using Functional Models that by assuming specific noise structures (like additive noise), we can orient arrows even when faithfulness is shaky.
- Algebraic Properties: Asish Ghoshal and Jean Honorio developed algorithms for learning linear SEMs that leverage the precision matrix, allowing for structure recovery in polynomial time without requiring the restrictive faithfulness condition.
- Dormant Independence: Ilya Shpitser and Judea Pearl introduced the concept of Dormant Independence—independencies that only appear after an intervention. This provides a way to verify structural constraints that are "hidden" in purely observational data.
Toward Real-World Robustness
The ultimate goal of this research is to move causal discovery out of the "synthetic trap." As benchmarks like CausalTime by Yuxiao Cheng show, algorithms that perform perfectly on idealized data often fail on realistic time-series. By relaxing the faithfulness assumption and developing diagnostic tools like CAFÉ for verifying causal influence, the field is building a more resilient framework for understanding the complex, often "unfaithful" mechanisms of the real world.