Causal Faithfulness

The primary risk of assuming faithfulness is its vulnerability to exact cancellation or parameter tuning. This occurs when the causal effects along different paths in the graph perfectly balance out, creating a conditional independence that does not correspond to d-separation.

• Example: In a graph X → Y ← Z and X → Z, the path X → Z → Y and the direct path X → Y could have effects that exactly cancel, making X and Y independent despite being d-connected. This is a violation of faithfulness. In practice, such exact cancellation is considered 'unstable' or 'measure-zero', but approximate violations can occur with finite data.

Faithfulness is the converse of the Causal Markov Condition. Together, they form a complete link between causality and probability.

• Causal Markov Condition: Every d-separation in the graph implies a conditional independence in the data. (Graph → Data).
• Faithfulness Assumption: Every conditional independence in the data implies a d-separation in the graph. (Data → Graph).

This bidirectional relationship is what allows us to use statistical tests on data to make inferences about the unobserved causal graph. Violating faithfulness breaks the Data → Graph link, making discovery from independencies unreliable.

While faithfulness is often assumed, its plausibility can be assessed. Robust causal discovery involves:

• Using multiple hypothesis tests with different significance levels to check for stable independence judgments.
• Employing score-based methods (like GES) that are less sensitive to individual independence test failures, though they typically assume a version of faithfulness.
• Considering background knowledge to rule out graph structures where parameter cancellation is likely.
• Utilizing interventional data where possible, as faithfulness violations are less likely under interventions that disrupt the system's natural equilibrium.

Faithfulness is closely related to the minimality assumption, which states that the true causal graph is the simplest one that can explain the observed independencies (i.e., it has no superfluous edges). A minimal graph is always faithful to the distribution it generates, but a distribution faithful to a graph does not guarantee that graph is minimal (a supergraph may also be faithful). In practice, causal discovery algorithms often seek the minimal faithful graph—the sparsest graph that satisfies the faithfulness condition with the data—as the preferred explanation, adhering to a causal version of Occam's Razor.

d-Separation (directional separation) is a graphical criterion used to read conditional independence relationships directly from a causal DAG. A set of variables Z d-separates X from Y if it blocks all paths between X and Y. Paths can be blocked by conditioning on colliders (common effects) or non-colliders.

• Core Mechanism: Faithfulness assumes that all conditional independencies in the data are a consequence of d-separation in the true causal graph. If X and Y are d-separated given Z in the graph, they must be independent in the data.
• Violation Detection: If statistical tests show an independence that is not implied by d-separation (or vice versa), it signals a violation of faithfulness.

Causal Discovery is the algorithmic process of inferring a causal graph from data. Faithfulness is a critical, often necessary, assumption for these algorithms to work correctly. Methods like the PC and FCI algorithms rely on testing conditional independencies to deduce graph structure.

• Faithfulness Dependency: These algorithms assume that independencies found in the data correspond exactly to d-separation in the true graph. An unfaithful independence (e.g., from cancellation) can lead the algorithm to incorrectly remove a true causal edge.
• Robustness: Modern causal discovery research focuses on making algorithms more robust to approximate violations of faithfulness.

Parameter Cancellation is the primary cause of faithfulness violations. It occurs when the specific numerical parameters of a causal model (e.g., path coefficients in a linear system) align in such a way that they cancel each other out, creating a conditional independence that is not implied by the graph structure alone.

• Classic Example: In a chain X → Y → Z, X and Z are generally dependent. However, if the effect of X on Y and the effect of Y on Z multiply to zero, X and Z become independent, violating faithfulness.
• Implication: Such cancellations are considered 'fine-tuned' or 'measure-zero' events in parameter space, meaning they are theoretically possible but unlikely in practice without a specific mechanism.

In philosophy of science and causal inference, Stability (or Robustness) is a concept closely related to faithfulness. A causal model is stable if the conditional independence relations it entails remain invariant under small perturbations to the causal parameters. Faithfulness violations are inherently unstable.

• Theoretical Foundation: Judea Pearl and others argue that only stable models are scientifically useful, as they represent robust mechanisms rather than coincidental parameter alignments.
• Justification for Faithfulness: The assumption is justified by the belief that true causal mechanisms in nature are stable; therefore, we should prefer causal graphs that imply stable, faithful independence patterns.