Ancestral graph¶

A mixed graph using directed, bidirected and undirected edges to encode conditional independences left by latent-variable marginalization and selection conditioning in a DAG.

Version: v2 · 2026-09-19 · History
Domain-specific #: 3287
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomain: Graphical Models → Experimental Design & Statistics

Core Idea¶

Arrowheads cannot point to ancestors, undirected-edge endpoints cannot carry arrowheads and maximal ancestral graphs add a separate condition ensuring every missing edge represents an independence. Hidden common causes become bidirected edges, selection effects can become undirected edges and graph separation rules preserve observable conditional-independence structure without explicitly retaining latent nodes. The abstraction is therefore identified by a declared carrier, a transformation or constraint over that carrier, and an invariant that tells an analyst whether the named structure is genuinely present.

How would you explain it like I'm…

The Hidden-Cause Arrow Map

An Ancestral graph is a map of dots and lines showing which things might make other things happen. Sometimes a hidden thing we cannot see is pushing on two things at once. Instead of drawing the hidden thing, the map draws a line with arrowheads on both ends between the two things it pushes.

Hidden-Cause Map

Scientists draw graphs with dots for things they measure and arrows for 'this can affect that'. Often, some causes are hidden because nobody measured them. An Ancestral graph handles this without drawing the hidden causes: if something unseen affects two measured things, it links them with a two-headed arrow; if the way the data was picked links two things, it uses a plain line with no arrowheads. The graph has rules so that arrows never loop back in a way that makes something its own cause. Reading the graph tells you which measured things should be unrelated once you know certain others.

Latent-Aware Independence Graph

An Ancestral graph is a graph over only the observed variables that still captures the effects of hidden variables. It uses three edge types: directed arrows for causal influence, bidirected edges (arrowheads at both ends) for a hidden common cause, and undirected edges for selection effects, meaning the way the sample was chosen. Two rules keep it consistent: an arrowhead can never point into a variable that is an ancestor of the variable at the other end of that edge, and an endpoint of an undirected edge cannot carry an arrowhead. A graph-reading rule called m-separation tells you which observed variables are conditionally independent. A maximal ancestral graph adds one more condition: every missing edge must correspond to some independence. This lets researchers reason about the observable independence pattern without listing the hidden variables.

Ancestral graphs are mixed graphs over an observed vertex set, with directed (→), bidirected (↔), and undirected (—) edges. They let you represent the conditional-independence structure implied by a causal DAG after marginalizing latent variables and conditioning on selection variables, without explicitly retaining those nodes: hidden common causes become bidirected edges, and selection effects can become undirected edges. Constraints: the graph has no directed cycles, arrowheads cannot point into ancestors (so if a is an ancestor of b, the a–b edge can't be ↔ or ←), and endpoints of undirected edges cannot receive arrowheads. Independence is read off via m-separation, generalizing d-separation. A maximal ancestral graph (MAG) additionally requires that every non-adjacent pair be m-separated by some set, so each missing edge represents an independence (no inducing paths between non-adjacent vertices). MAGs relate to DAGs (which they summarize) and PAGs (equivalence classes of MAGs used in causal discovery).

Scope of Application¶

Ancestral graph belongs to graphical models and is useful where the analyst can specify the typed graphical models carrier, including objects, relations, parameters, conventions, evidence, boundaries, and comparison targets, then evaluate the observed vertex set, directed bidirected and undirected edge semantics, ancestor relation and acyclicity, arrowhead and undirected-endpoint restrictions, latent and selection variables represented, m-separation criterion, maximality, inducing paths and relation to DAG MAG and PAG models are explicit. The scope is broad within that domain but bounded by the need for the observed vertex set, directed bidirected and undirected edge semantics, ancestor relation and acyclicity, arrowhead and undirected-endpoint restrictions, latent and selection variables represented, m-separation criterion, maximality, inducing paths and relation to DAG MAG and PAG models are explicit.

Clarity¶

The abstraction clarifies a crowded vocabulary by making the observed vertex set, directed bidirected and undirected edge semantics, ancestor relation and acyclicity, arrowhead and undirected-endpoint restrictions, latent and selection variables represented, m-separation criterion, maximality, inducing paths and relation to DAG MAG and PAG models are explicit the center of the account. A claim should name the carrier, the governing operation or relation, the applicable assumptions, and the recognition test.

Manages Complexity¶

Without the abstraction, an analyst must reason directly over many local details: the carrier roles, admissibility assumptions, competing conventions, derived invariants, boundary cases, and proof or validation obligations specific to Ancestral graph. Ancestral graph compresses them into the roles in the structural signature. That compression permits comparison across instances without erasing the variables that determine validity. It also exposes which details may be varied safely and which are constitutive.

Abstract Reasoning¶

Identify the carrier. State what the elements, states, objects, or observations are: the typed graphical models carrier, including objects, relations, parameters, conventions, evidence, boundaries, and comparison targets. Reject examples whose alleged carrier belongs to a different problem.
Lock the constitutive rule. Express the observed vertex set, directed bidirected and undirected edge semantics, ancestor relation and acyclicity, arrowhead and undirected-endpoint restrictions, latent and selection variables represented, m-separation criterion, maximality, inducing paths and relation to DAG MAG and PAG models are explicit independently of one notation or implementation.

Knowledge Transfer¶

Knowledge transfers strongly among subfields of graphical models because they reuse the typed graphical models carrier, including objects, relations, parameters, conventions, evidence, boundaries, and comparison targets, Hidden common causes become bidirected edges, selection effects can become undirected edges and graph separation rules preserve observable conditional-independence structure without explicitly retaining latent nodes., and type the carrier, state every parameter and convention in the definition, test that the observed vertex set, directed bidirected and undirected edge semantics, ancestor relation and acyclicity, arrowhead and undirected-endpoint restrictions, latent and selection variables represented, m-separation criterion, maximality, inducing paths and relation to DAG MAG and PAG models are explicit, compare the nearest accepted identity, and report counterexamples, uncertainty, and limiting cases.

Relationships to Other Abstractions¶

Current abstraction Ancestral graph Domain-specific

Parents (1) — outgoing typed relations

Ancestral graph is a kind of Representation Prime

It remains its own entry because its identity is fixed by the observed vertex set, directed bidirected and undirected edge semantics, ancestor relation and acyclicity, arrowhead and undirected-endpoint restrictions, latent and selection variables represented, m-separation criterion, maximality, inducing paths and relation to DAG MAG and PAG models.

Hierarchy path (1) — routes to 1 parentless root

Ancestral graph → is a kind of Representation → presupposes Abstraction

Neighborhood in Abstraction Space¶

Ancestral graph sits in a crowded region of the domain-specific corpus (14^th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Graphical & Network Models of Dependence (16 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-25