A startup data team ran a causal discovery algorithm on rich observational data—months of click streams, session histories, and purchase records. They wanted to know what causes high-value customer churn. The algorithm produced something unexpected: not a single causal diagram, but a graph with undirected edges scattered throughout. Some arrows had direction. Others were just lines. The team's first reaction was that the algorithm was broken. The second was that they needed more data. Both reactions were wrong. What they were looking at was a mathematical confession: the data alone cannot determine which direction these arrows should point. This is not a failure of the algorithm. This is a theorem about what observational data can and cannot determine. The output they received is called a CPDAG—a Completed Partially Directed Acyclic Graph. And it is the honest answer to the question: what has the data learned?

Chapter 5 established the limit of observation: Rung 2 causal questions cannot be answered from observational data alone. Chapter 6 introduced the structural causal model as the mathematical apparatus to overcome that limit. This chapter proves exactly where the limit lies. The equivalence problem is the boundary where data stops and expert judgment must begin. It is not a practical limitation of sample size or algorithmic sophistication. It is a mathematical boundary. When multiple causal structures produce identical conditional independence patterns, no amount of additional observational data will ever distinguish them. This forces a three-part choice: accept the uncertainty, seek interventional data through experiments, or elicit expert domain knowledge to resolve ambiguity. Understanding the equivalence problem transforms how you approach causal discovery. It tells you when to trust data and when to demand expertise. It separates problems data can solve from problems requiring human judgment.

Markov equivalence is the formal name for a simple but profound fact: different causal diagrams can produce identical statistical signatures in data. Consider three variables X, Y, and Z. If X causes Y and Y causes Z—a chain—then in observational data, X and Z are marginally correlated but conditionally independent given Y. Now suppose Y causes both X and Z—a fork. In the data, you observe the identical pattern: marginal correlation between X and Z, conditional independence given Y. Or suppose Z causes Y and Y causes X—a reverse chain. Again, identical statistical pattern. All three causal stories are Markov equivalent: they predict the same conditional independences and the same joint distribution. A statistician examining any amount of observational data could never tell them apart. This is not because the causal structures are truly the same—they are incompatible stories about reality. It is because they make identical predictions about what you will observe in the world. The data simply cannot encode the difference.

The three structures demonstrate the core problem. In the chain, information flows from X through Y to Z in a straight line. In the fork, Y sits at the center and influences both X and Z independently. The connection between X and Z is mediated entirely through their shared cause Y. In the reverse chain, causality points backward: Z influences Y which influences X. The causal paths are opposite. The roles of source and sink are reversed. Yet when you examine observational data from any of these systems, you find the same pattern. X and Z are correlated marginally. Once you condition on Y—once you stratify the data by Y's value—that correlation vanishes. X and Z become independent. This conditional independence pattern is a signal about Y's structural role. It tells you Y blocks all paths between X and Z. But it does not tell you which of the three structures generated that blocking. Direction is hidden. Only data-generating structure, not data itself, knows direction.

The conditional independence X ⊥ Z | Y is what a causal discovery algorithm will extract from observational data. It captures a fact: Y has a special structural role. Either as a mediator, a confounder, or reverse-direction mediator. But conditional independence is mute on direction. It says Y blocks information flow between X and Z. It does not say which direction that information naturally flows. Which variable is the source? Which is the sink? For causal discovery algorithms like PC or GES, conditional independence is the primary signal. These algorithms infer edge presence from marginal correlation and infer direction from conditional independence patterns. When the same conditional independence pattern is compatible with multiple directions, the algorithm cannot choose. It must produce an undirected edge. This is mathematically sound. It is not weakness. It is an honest accounting of what the data knows and what it does not. The undirected edge is a flag: direction cannot be determined from observation alone.

A CPDAG does not present a single causal diagram. It represents an equivalence class—all the diagrams compatible with the conditional independences found in the data. A directed edge X→Y appears with that orientation in every DAG in the equivalence class. The data has determined direction. An undirected edge X—Y appears pointing both ways X→Y and X←Y in different DAGs in the equivalence class. The data permits both orientations. Understanding a CPDAG requires three skills. First, distinguish directed from undirected edges. Second, understand what each type means for inference. Third, assess the consequences if undirected edges are oriented incorrectly. A directed edge is data-robust. An undirected edge is data-ambiguous and requires external knowledge. The startup's result was not failure. It was the algorithm correctly identifying which causal relationships it could determine and which it could not. Some arrows pointed. Some remained open. That mixed result is the honest answer observational data provides.

Reading a CPDAG requires three skills. First, distinguish directed from undirected edges visually. A directed edge X→Y is an arrow. An undirected edge X—Y is a line. Second, understand what these types mean. Directed edges are data-determined. They appear with the same orientation in every DAG that could have generated the observed independence patterns. Undirected edges are ambiguous. They appear with opposite orientations in different equivalent DAGs. Knowing the difference tells you where the data speaks authoritatively and where it remains silent. Third, assess consequences. For directed edges, inference is robust to how other edges are oriented. For undirected edges, causal estimates will depend critically on eventual orientation. If you assume X→Y when the truth is X←Y, your causal effect estimate will be systematically wrong. The CPDAG forces these distinctions explicit. You cannot hide behind a single point estimate. You must recognize which parts of your causal story rest on data and which parts rest on assumption.

The Verma-Pearl condition is the mathematical foundation of Markov equivalence. It answers the question: which DAGs are equivalent? The answer has two parts. First, equivalent DAGs have the same skeleton—the same set of edges, ignoring direction. The chain X→Y→Z and the fork X←Y→Z share the same skeleton: edges X-Y and Y-Z. Second, equivalent DAGs have the same set of v-structures. A v-structure is a pair of edges converging into a single node with no edge between their sources, X→Y←Z with no edge between X and Z. The chain has no v-structures. The fork has no v-structures. The reverse chain has no v-structures. All three have the same skeleton and the same v-structures, so they are equivalent. This condition is complete and precise. It fully describes when two DAGs make identical statistical predictions. It provides the theoretical foundation for why CPDAGs represent the maximal information observational data can extract. Every undirected edge in a CPDAG is an edge that can point either direction without violating the skeleton-v-structure condition. Orientation does not matter for the independence structure.

The startup's instinct was understandable but wrong. More data will not resolve undirected edges. This is the core lesson. If a CPDAG contains an undirected edge, the equivalence class contains at least two DAGs with opposite directions on that edge. These two DAGs make identical predictions about all conditional independences. They predict the same marginal correlations, the same partial correlations, the same patterns of dependence and independence across all possible conditioning sets. As you collect more observational data, estimates of these quantities become more precise. But precision does not break equivalence. With infinite observational data and perfect estimates of all conditional independences, equivalent DAGs would still be indistinguishable. The conditional independence patterns would be identical. The marginal distributions would be identical. The only way to distinguish them is intervention—deliberately change one variable and observe how others respond. Observational data, no matter how rich or precise, cannot create causal variation. It can only estimate relationships already present in the data. Direction must come from somewhere else.

The equivalence problem forces a choice, and it is explicit in the CPDAG. You cannot ignore undirected edges and pretend they are determined. The first route—accepting uncertainty—is viable when undirected edges do not affect your target inference. If you estimate a causal effect and the ambiguous edge is not on a path you traverse, you can report estimates robust across all possible orientations. Sensitivity analysis quantifies how estimates change when you vary undirected orientations. The second route—seeking interventional data—is the gold standard where resources permit. Randomized controlled trials, quasi-experiments, natural experiments, or instrumental variables can orient edges. The startup's next step might be to run experiments on their key ambiguities. This is expensive but provides data-driven, definitive answers. The third route—expert knowledge—is what happens when the first two are infeasible. An expert might know from domain experience that customer behavior causes churn, not reverse. But this knowledge must be explicit, transparent, and stated as an assumption that could be wrong. All three routes require owning the choice.

The equivalence problem is the mathematical form of the limit posed at the beginning. Chapter 5 asked: can observational data answer causal Rung 2 questions? The answer was no, not alone. Chapter 6 responded: here is the structural apparatus—the causal diagram and do-operator—that formalizes what knowledge you need. This chapter proves exactly what you need. Where Markov equivalence occurs, you have reached the boundary of what data provides. You can choose to live with uncertainty, seek experiments, or ask experts. But you cannot algorithm your way out. This is not a deficiency of current science or insufficient cleverness. It is a theorem. Understanding it transforms practice. It tells you when to trust data and when to demand expertise. When sensitivity analysis is sufficient and when you need interventional evidence. It separates problems data solves from problems requiring human judgment. The startup's undirected edges were not a failure. They were honesty—mathematics telling them precisely where data ends and knowledge must begin.