The distinction between correlation and causation has haunted epistemology since Hume dissolved causal necessity into constant conjunction. For centuries, philosophers debated whether causal relations exist as features of the world or merely as habits of inference. The question seemed intractable because we lacked a formal vocabulary precise enough to state it clearly.

That vocabulary now exists. Since the 1980s, the marriage of probability theory and graph theory—developed by Judea Pearl, Peter Spirtes, and others—has produced a mathematical framework in which causal claims are neither metaphysical mysteries nor reducible to mere correlation. Directed acyclic graphs annotated with conditional probability distributions allow us to represent causal hypotheses as objects that can be compared, tested, and learned from data.

Combined with Bayesian methods, this framework transforms causal inference from a philosophical puzzle into a computational problem. We assign prior probabilities over possible causal structures, update these priors given observed data, and derive posterior distributions over the space of causal models. What emerges is a rigorous epistemology of causation: one that specifies exactly what assumptions we make, what inferences those assumptions license, and where our conclusions become underdetermined by evidence.

A Bayesian network is a directed acyclic graph G = (V, E) whose vertices represent random variables and whose edges encode direct probabilistic dependencies. Attached to each vertex is a conditional probability distribution P(X_i | Pa(X_i)), specifying the behavior of that variable given its parents. Together, graph and distributions define a joint distribution over all variables.

The formalism gains causal significance through the Causal Markov Condition: every variable is probabilistically independent of its non-descendants given its parents. Intuitively, once you know a variable's direct causes, its non-effects tell you nothing more about it. This condition is what elevates a Bayesian network from a mere factorization device into a causal model.

The Markov condition licenses a powerful factorization: P(X₁, ..., X_n) = ∏ P(X_i | Pa(X_i)). This dramatically reduces the parameters needed to specify a joint distribution and, more importantly, encodes testable independence claims. If the graph asserts X ⊥ Y | Z, then this must hold in the data—giving us an empirical grip on causal structure.

The framework also distinguishes observation from intervention through the do-calculus. The distribution P(Y | do(X = x))—representing what happens when we surgically set X to x—generally differs from P(Y | X = x), the conditional distribution given observation. This distinction formalizes what Hume could only gesture at: the difference between mere correlation and causal influence.

Faithfulness serves as the converse assumption: every conditional independence in the true distribution corresponds to a structural feature of the graph. Together, Markov and Faithfulness establish a bidirectional bridge between probabilistic structure and causal structure—the epistemological foundation on which everything else rests.

Takeaway

The Causal Markov Condition is the axiom that transforms statistical dependence into causal knowledge. Without it, graphs are merely convenient representations; with it, they become claims about the world's causal architecture.

Structure learning treats the causal graph itself as an unknown quantity to be inferred. Given data D, we seek the posterior distribution P(G | D) ∝ P(D | G) P(G) over the space of possible directed acyclic graphs. The prior P(G) encodes structural expectations—sparsity, domain knowledge, preferred causal orderings—while the marginal likelihood P(D | G) measures how well each structure explains the data.

The marginal likelihood is computed by integrating over the parameter space: P(D | G) = ∫ P(D | θ, G) P(θ | G) dθ. This integration automatically implements Occam's razor: overly complex graphs incur penalties because their prior probability mass is spread thinly across many parameter configurations. Bayesian model selection thereby balances fit against complexity without requiring ad hoc regularization.

Scoring functions like the Bayesian Information Criterion or the Bayesian Dirichlet equivalent score make these computations tractable for discrete variables. Continuous variables typically require Gaussian assumptions or nonparametric approaches. Each choice smuggles in substantive assumptions about the functional form of causal relationships.

The search space is superexponential—the number of DAGs on n variables grows faster than 2^n²—forcing us to use heuristics. Greedy hill-climbing, MCMC sampling over graph space, and constraint-based algorithms like PC and FCI trade completeness for computational feasibility. Modern approaches like NOTEARS reformulate structure learning as continuous optimization, exploiting differentiable characterizations of acyclicity.

Crucially, observational data alone cannot uniquely identify causal structure. Multiple graphs typically belong to the same Markov equivalence class, sharing all conditional independence relations but differing in edge orientations. Distinguishing among them requires interventional data, temporal information, or additional assumptions about noise distributions—a fundamental epistemic limit encoded directly in the mathematics.

Takeaway

The space of causal hypotheses is only partially identifiable from observation. What data cannot distinguish, no amount of statistical sophistication can recover—this is a feature of reality, not a defect of method.

The graphical framework's mathematical elegance can obscure its substantive philosophical commitments. The Causal Markov Condition, for instance, presupposes that all common causes are represented in the graph—the causal sufficiency assumption. When unmeasured confounders exist, Markov fails, and standard structure learning delivers systematically distorted conclusions.

Faithfulness carries its own metaphysical weight. It assumes that exact cancellations of causal effects—where two paths precisely nullify each other's influence—do not occur. While such cancellations have measure zero under generic parameterizations, they can arise systematically in engineered or evolved systems, where feedback mechanisms specifically produce them.

The framework also assumes a modularity thesis: intervening on one variable leaves other causal mechanisms intact. This is the mathematical content of do-calculus, but philosophically it presupposes that causal relations are locally autonomous rather than holistically constituted. Whether this reflects the world's structure or merely our idealized models remains contested.

These assumptions collectively constitute what we might call the manipulationist metaphysics: causation is what supports interventional counterfactuals, mechanisms are locally decomposable, and confounding is finite and enumerable. This is a substantive worldview, defensible but not self-evident, and incompatible with certain process-based or dispositional accounts of causation.

The virtue of the formal approach is not that it settles these debates but that it makes them precise. Rather than arguing about whether causation is 'really' regularity, counterfactual dependence, or mechanistic transmission, we can specify exactly which mathematical assumptions correspond to which philosophical positions and examine what each licenses empirically.

Takeaway

Formal frameworks do not eliminate metaphysical commitments—they make them auditable. Every equation encodes a philosophical stance that can now be stated, tested, and defended on explicit grounds.

Bayesian causal inference does not dissolve Hume's problem so much as reformulate it with mathematical precision. We cannot derive causal knowledge from correlational data alone—but we can specify exactly which additional assumptions bridge the gap, and how much causal information each assumption purchases.

This is progress of a distinctive kind. The philosophical questions have not vanished; they have been transformed into questions about the justification of specific formal principles: Markov, Faithfulness, causal sufficiency, modularity. Each principle can now be examined, weakened, or defended on both mathematical and empirical grounds.

The deeper lesson concerns epistemology itself. Formal methods succeed not by replacing philosophical reflection but by giving it teeth. When we can compute the posterior over causal structures, we know precisely what our evidence supports—and, equally important, what it does not.