Is a covariance graph a causal graph?

Not by itself. It represents statistical independence assumptions, and causal claims need additional justification.

Why use bidirected edges?

They are a common convention for marginal-independence models and distinguish them from other graphical representations.

Can two variables with no edge become dependent after conditioning?

Yes. The worked Gaussian example demonstrates exactly that possibility.

AI Glossary / Probabilistic machine learning / Intermediate

Covariance graph

Definition

A covariance graph is a graphical representation of marginal independence assumptions, commonly expressed through zeros in a Gaussian covariance matrix. Each node represents a variable, and a missing edge between two nodes indicates that those variables are marginally independent in the model.

What it means

Marginal independence means two variables are independent before conditioning on other measured variables. This differs from conditional independence, which asks whether they become independent after other variables are known.
For a jointly Gaussian vector, zero covariance between two components is equivalent to their marginal independence. Covariance graphs use that relationship to express a structured model. They are often drawn with bidirected edges or dashed lines. The drawing convention needs to be stated clearly because a similar-looking graph can encode a different kind of independence in another modeling framework.

How it works

Choose a set of variables and a graph representing proposed marginal independences. In a Gaussian covariance graph model, constrain the corresponding off-diagonal covariance entries to zero while estimating the remaining parameters and maintaining a valid positive-definite covariance matrix.
Assess the model against the data and consider uncertainty about the structure. An observed sample covariance close to zero is not itself proof of population independence. Conversely, an edge usually allows an association; the model's basic constraints need not force every allowed covariance to be nonzero.
Interpret missing edges using marginal semantics. A Gaussian concentration graph instead places zeros in the inverse covariance, or precision, matrix and describes conditional independence given the other variables. These matrices answer different questions.

Three practical examples

Measurement analysis: represent a hypothesis that two instrument readings are marginally independent while each is associated with another recorded quantity.
Survey modeling: explore structured covariance assumptions among continuous responses, checking whether a Gaussian model is suitable for the measurements.
Model review: compare covariance and precision graphs to clarify whether reported relationships are unadjusted associations or associations conditional on other variables.

Think of it this way

Think of two independent contributors whose work feeds into a common report. Knowing the total report can make their contributions appear related: if one contributed more, the other may have contributed less. The analogy illustrates how conditioning can create dependence. A covariance graph itself does not establish that this causal story generated the data; it only encodes the stated statistical relationships.

Worked example

Consider an original centered Gaussian vector (A, B, C) with covariance matrix:
[1, 0.5, 0]
[0.5, 1, 0.5]
[0, 0.5, 1].
Its leading principal minors are 1, 0.75, and 0.5, all positive, so it is positive definite. A and C have zero covariance and are marginally independent under the joint Gaussian assumption. The covariance graph has edges A ↔ B and B ↔ C, but no A ↔ C edge.
Now condition on B. The conditional covariance of A and C is 0 − (0.5 × 0.5) / 1 = −0.25. Each conditional variance is 1 − 0.5² / 1 = 0.75, giving conditional correlation −0.25 / 0.75 = −1 / 3.
Thus the missing covariance-graph edge does not mean A and C remain independent after B is known. The example also shows why reading this graph with ordinary undirected conditional-independence rules would lead to the wrong conclusion.

Illustrative case study

Fictional case: Rowan Quality reviews three standardized measurements from a production process. An analyst presents a covariance graph with no edge between the first and third measurements. A manager interprets this as meaning one can never help predict the other.
The team clarifies that the claim is marginal and model-dependent. It checks prediction both before and after conditioning on the middle measurement, then records the distinction in the report. Reviewers also inspect the underlying distribution rather than relying on a correlation threshold alone. The case is illustrative, not a report of actual factory data.

Common misconceptions and mistakes

Zero covariance does not imply independence for arbitrary non-Gaussian variables. For example, nonlinear dependence may exist despite zero linear correlation. The Gaussian assumption is therefore material when interpreting covariance zeros as independence.
Bidirected arrows do not mean that two variables cause each other. Causal interpretations require a separate model and assumptions. A path through intermediate nodes also does not automatically override the marginal independence encoded by a missing edge. Finally, covariance selection is not interchangeable with graphical lasso, which ordinarily targets sparse precision structure.

Business use

The value for a business is clearer reasoning about multivariate measurements and the assumptions behind a statistical model. A sparse graph can summarize proposed relationships, but its meaning must match the analysis being presented.
Ask whether edges concern covariance or precision, how the graph was selected, and whether uncertainty or multiple comparisons were considered. Keep variable definitions and conditioning choices visible. If a graph informs a prediction or intervention, validate that specific use rather than treating a visually simple network as a complete explanation of the process.

Check your understanding

A jointly Gaussian model has a zero entry in its precision matrix between A and C. Does that establish a missing edge in its covariance graph?

Answer: No. A precision zero concerns conditional independence given the remaining variables. A covariance-graph missing edge concerns a covariance zero and marginal independence. One does not generally imply the other.

Related terms

Covariance matrix
Conditional independence
Partial correlation
marginal independence
Gaussian graphical model
precision matrix
concentration graph
bidirected graph

Learn next

Learn covariance and conditional covariance, then reproduce the three-variable example. Draw its covariance graph and compare it with the pattern of zeros in its inverse matrix. Next, study how graphical model fitting preserves positive definiteness while imposing selected zero constraints.

Inverse matrix

Sources

Drton and Richardson, “Graphical Methods for Efficient Likelihood Inference in Gaussian Covariance Models,” original paper, introduction and graph semantics: https://www.jmlr.org/papers/volume9/drton08a/drton08a.pdf Drton, official ggm documentation, fitting Gaussian covariance graphs: https://search.r-project.org/CRAN/refmans/ggm/html/fitCovGraph.html Independent source review: Drton–Richardson introduction and graph semantics were read; official ggm fitting documentation was consulted.

Previous: Covariance function Next: Covariance matrix