To uncover the causal architecture of the world from purely observational data, scientists rely on a fundamental bridge connecting probability distributions to directed graphs. This bridge is built on two pillars: the Causal Markov Condition, which guarantees that graphical separation implies statistical independence, and the Causal Faithfulness Assumption, which guarantees the reverse. Faithfulness asserts that every conditional independence observed in the data is a deliberate consequence of the graph's structure, rather than an accidental byproduct of perfectly canceling parameters.
If faithfulness holds, causal discovery is a clean, combinatorial puzzle. If variable $A$ and variable $B$ become independent when conditioning on $C$, we can confidently erase the edge between $A$ and $B$.
But what happens when nature violates this assumption? In real-world systems, faithfulness frequently breaks down. When it does, standard causal discovery algorithms fail silently, generating incorrect graphs that can lead to disastrous real-world decisions. A growing body of research focuses on understanding why faithfulness fails, characterizing the severity of these violations, and engineering robust algorithms that can see through unfaithful data.
The Mechanics of Unfaithfulness
To understand why faithfulness fails, consider a classic biological example: a feedback loop or a redundant pathway. Suppose a drug ($A$) directly increases a patient's survival rate ($C$), but also triggers a side effect ($B$) that decreases survival ($C$) by the exact same margin.
[Drug A] ------> [Side Effect B]
| |
| (Positive) | (Negative)
v v
[Survival C] <---------+
When we measure the total statistical correlation between the Drug ($A$) and Survival ($C$), the positive and negative pathways perfectly cancel out. Statistically, $A$ and $C$ appear completely independent. A traditional constraint-based algorithm, such as the PC algorithm, will observe this independence and mistakenly conclude that the drug has no causal relationship with survival.
This phenomenon is not a rare edge case. As Garvesh Raskutti et al. (2013) demonstrated using algebraic geometry, the "strong faithfulness" assumption—which requires partial correlations to be bounded safely away from zero—is mathematically highly restrictive. The authors showed that even in sparse, low-dimensional graphs, the volume of the parameter space that violates strong faithfulness grows rapidly with the size of the network. In finite-sample settings, weak causal effects become statistically indistinguishable from zero, creating a fundamental failure mode for structure learning.
Furthermore, this problem scales poorly. Mateusz Olko et al. (2025) proved that the probability of a data distribution satisfying the strong faithfulness requirement decreases exponentially as the number of nodes and the density of the causal graph increase. In large-scale real-world systems, unfaithfulness is virtually guaranteed.
Beyond path cancellations, deterministic relationships also violate faithfulness. When one variable perfectly predicts another, it introduces accidental conditional independencies that do not map to missing edges, causing theoretical error bounds to collapse, as noted in constraint-based Bayesian network learning frameworks like those evaluated by Ioannis Tsamardinos et al. (2006).
The Danger of Silent Failure
The core issue plaguing modern causal inference is that algorithms typically lack internal mechanisms to detect when their mathematical assumptions are violated. As Xiu-Chuan Li et al. (2025) point out, this leads to the silent generation of incorrect causal graphs. If an algorithm encounters a path cancellation, it simply omits the edges without warning, leading to false conclusions in high-stakes applications like root cause analysis or clinical trials.
Relaxing the Assumption
To build more reliable systems, researchers have proposed decomposing and weakening the faithfulness assumption rather than abandoning it entirely.
Adjacency-Faithfulness
One prominent strategy is to separate structure identification (finding where the edges are) from edge orientation (determining the direction of the arrows). Joseph Ramsey et al. (2012) introduced Conservative PC (CPC), which relaxes global faithfulness to a weaker condition called Adjacency-Faithfulness. Adjacency-faithfulness only assumes that if two variables are directly connected by an edge, they cannot be rendered independent by conditioning on any subset of other variables. By being significantly more cautious during the orientation phase and treating ambiguous independence tests with skepticism, the authors showed that CPC drastically reduces false causal claims while maintaining computational efficiency.
Interaction-Based Faithfulness
Other frameworks look beyond pairwise conditional independence to capture hidden links. Alexander Marx et al. (2021) proposed 2-adjacency faithfulness, a relaxation that leverages "triple interactions"—dependencies involving three variables simultaneously. This allows algorithms to detect causal links that are invisible to traditional conditional independence tests, such as XOR-type relationships and perfectly canceling paths. Similarly, Shuyan Wang et al. (2022) utilized a $k$-Triangle-Faithfulness condition to construct uniformly consistent estimators of causal effects in non-Gaussian, non-parametric settings, bypassing the physically implausible constraints of strong faithfulness.
Bypassing Faithfulness Entirely
Can we discover causal structures without relying on faithfulness at all? Recent algorithmic breakthroughs suggest that by changing the statistical tools we use, we can bypass the requirement entirely.
Instead of testing for conditional independence, we can exploit algebraic and structural properties of the data distribution:
- Precision Matrices: For linear structural equation models (SEMs), Asish Ghoshal et al. (2018) developed an algorithm that learns graph structures in polynomial time by leveraging the properties of the precision (inverse covariance) matrix. Their method operates under a broad set of conditions without requiring the faithfulness assumption, ensuring reliability even when partial correlations vanish.
- Composition Properties: In the context of chain graphs, José M. Peña et al. (2014) proved that structure learning is possible under the relatively weak assumption of the composition property (which asserts that if $A$ is independent of $B$ given $C$, and $A$ is independent of $D$ given $C$, then $A$ is independent of $B \cup D$ given $C$). This allowed the creation of algorithms that outperform traditional faithfulness-dependent methods on real-world datasets.
- Rank Constraints: When dealing with hidden (latent) variables and selection bias, Haoyue Dai et al. (2025) demonstrated that looking at the algebraic rank of covariance submatrices, rather than standard conditional independence, allows researchers to recover latent causal structures even when parameters perfectly cancel out.
- Sparsity Optimization: For cyclic systems containing feedback loops, Soheun Yi et al. (2024) introduced the Filter, Rank, and Prune (FRP) framework. Instead of assuming strong faithfulness, FRP searches for the sparsest graph capable of explaining the observed statistical distribution, making it highly robust to weak or noisy causal signatures.
By shifting the foundation of causal discovery from fragile, global independence mappings to robust algebraic constraints and localized relaxations, the field is moving toward a more realistic paradigm—one where causal inference remains trustworthy even when nature hides its tracks.