Directed Acyclic Graph

A directed acyclic graph (DAG) is a directed graph that contains no directed cycles — there is no way to start at a node and follow a sequence of directed edges that eventually loops back to the starting node. This seemingly simple topological constraint has profound consequences: it guarantees the existence of a topological ordering, makes certain computational problems tractable, and provides the mathematical foundation for causal reasoning in Judea Pearl's framework.

In causal modeling, DAGs serve as the syntax for expressing causal assumptions. Nodes represent variables; directed edges represent direct causal effects; the acyclicity constraint encodes the assumption that causes precede effects in time. A DAG encodes not merely associations but a claim about the data-generating process: it specifies which variables are causally prior to which others and which paths represent spurious correlations that should be blocked by conditioning. The do-calculus — the set of rules for determining when causal effects are identifiable from observational data — is derived entirely from the graphical structure of the DAG. Without the acyclicity constraint, these rules would not hold, and causal identification would be far more difficult.

DAGs also appear in Bayesian networks, where the acyclicity constraint ensures that the joint probability distribution factors into a product of local conditional probabilities. The same structure that makes causal inference tractable makes probabilistic inference tractable — a convergence that suggests the DAG is not merely a convenient representation but a deep feature of how structured systems generate data.

The directed acyclic graph is the simplest mathematical object that can distinguish correlation from causation. Its acyclicity is not a limitation but a representational choice: it encodes the temporal asymmetry of causation in graphical form. To abandon DAGs in favor of more general graph structures — cyclic graphs, hypergraphs, undirected graphs — is not necessarily wrong, but it is to give up the one representation that makes causal identification both transparent and computationally tractable. The DAG is not the only way to think about causation, but it may be the only way that scales.

The Acyclicity Assumption in Dynamical Systems

The power of DAGs is purchased at a price: the acyclicity constraint forbids feedback. In any system where variables influence each other through closed causal loops — where A affects B and B affects A, or where the effect of an intervention feeds back to alter its own conditions — the DAG framework either breaks down or must be applied through a contortion.

Consider the central examples of system dynamics. Prices affect production, production affects inventory, inventory affects prices. The immune system attacks pathogens, pathogens evolve, the immune system adapts. These are not directed acyclic graphs. They are directed cyclic graphs, or more accurately, dynamical systems on graphs where the causal structure is time-indexed rather than static. A DAG can represent one time-slice of such a system — what causes what at a single moment — but it cannot represent the feedback that makes the system behave as a system rather than as a collection of independent causal chains.

The standard response within the DAG tradition is to unroll the cycle: to create a time-indexed DAG in which each variable at time t is a separate node, and edges run only forward in time. This works formally. A cyclic graph becomes acyclic when each node is duplicated across time steps. But this unrolling is not merely a representational convenience. It is an ontological commitment: it treats the system as a sequence of independent snapshots rather than as an integrated dynamical whole. The unrolled DAG can tell you that X at time t causes Y at time t+1. It cannot tell you that the system has attractors, that perturbations may be amplified or damped by feedback, or that the same intervention at different points in the cycle may produce opposite effects.

The alternative is to abandon the acyclicity constraint and work with cyclic causal models. Cyclic causality — the study of causal relations in feedback systems — asks different questions than DAG-based causal inference. It does not ask "what is the effect of do(X) on Y?" It asks "what are the stable regimes of this dynamics, and which perturbations can shift the system from one regime to another?" This is the language of feedback topology and convergent cross mapping, not of do-calculus.

The synthesizer's judgment is that the DAG framework and the dynamical-systems framework are not competitors but territorially distinct. DAGs are the right tool when the system decomposes into modules with unidirectional influence. Dynamical models are the right tool when the system is integrated by feedback. The error is not in using DAGs where they fit, but in exporting the DAG framework to domains — market dynamics, ecology, climate science, neuroscience — where feedback is not a complication to be controlled but the central phenomenon to be understood. The acyclicity assumption is not a neutral modeling choice. It is a claim that the system has no feedback, and in systems where feedback dominates, that claim is false.

Cyclic causality and the dynamical causal model represent two emerging frameworks for causal reasoning without the acyclicity constraint. Whether they can achieve the tractability and transparency of the DAG framework remains an open question — but it is a question that cannot even be asked within the DAG framework itself.