Conditional Independence Graphs

• Conditional Independence Graphs are graphical structures that represent CI relationships by mapping the absence of edges to independence among random variables.
• They employ various forms such as undirected, directed, and mixed graphs to succinctly capture complex dependency and conditional relationships.
• Graphical operations like arc addition, node deletion, and graph combination ensure that CI statements can be completely derived using the graphoid axioms.

Conditional independence graphs (also called Markov networks or CI graphs) are graphical structures that encode conditional independence (CI) relationships among collections of random variables. In these graphs, the absence of an edge, as specified by separation criteria appropriate to the graph type, corresponds to a conditional independence statement. CI graphs provide a central formalism bridging algebraic, combinatorial, and semantic approaches to probabilistic reasoning, particularly for inferring, verifying, and exploiting independence relations without reference to explicit probability distributions.

1. Formal Preliminaries: Graphoid Axioms and Graphical Encodings

A dependency model is specified as a ternary predicate $I(X,Z,Y)$, where $X$, $Y$, $Z$ are disjoint subsets of a finite set $U$ (of random variables), interpreted as "$X$ is conditionally independent of $Y$ given $Z$." The probabilistic criterion is $P(X\,|\,Z,Y) = P(X\,|\,Z)$ whenever $P(Z)>0$. The manipulation of such CI statements is governed by the graphoid axioms [Dawid 1979]:

• Symmetry: $I(X,Z,Y)\iff I(Y,Z,X)$
• Decomposition: $I(X,Z, Y\cup W)\implies I(X,Z,Y)$
• Weak Union: $I(X,Z, Y\cup W)\implies I(X, Z\cup Y, W)$
• Contraction: $I(X,Z\cup Y, W)$ and $I(X,Z,Y)\implies I(X,Z, Y\cup W)$

These axioms are both sound and, modulo positivity restrictions (e.g., for Intersection), often complete for the CI structures encountered in probabilistic and graphical models (Shachter, 2013).

CI graphs come in several primary categories:

• Undirected graphs (Markov networks): An edge between nodes $i$ and $j$ is present if and only if $X_i$ and $X_j$ are dependent given all other variables; absence implies $I(X_i, V\setminus\{i,j\}, X_j)$.
• Directed acyclic graphs (Bayesian networks): CI is encoded via $d$-separation, which specifies conditional independence based on the presence/absence of active paths in the DAG.
• Chain graphs, summary graphs, and mDAGs: These allow more general edge types (e.g., directed, bidirected, or hyperedges) to represent marginal, conditional, and context-specific CI in the presence of hidden variables or after marginalization (Evans, 2014).

2. Multiple Undirected Graphs and Purely Graphical Inference

A fundamental insight due to Shachter is that manipulating CI statements graphically using sets of undirected graphs—multiple undirected graphs (MUGs)—and three elementary graphical operations (arc addition, node deletion, graph combination) is equivalent to derivation in the graphoid axiomatic system (Shachter, 2013):

• Arc Addition: Adding edges can only destroy separations, never create new independencies.
• Node Deletion (Marginalization): Deleting a node after fully connecting its neighbors preserves all previously valid separations.
• Graph Combination (Contraction): Duplicating a graph, augmenting with new nodes, and fully connecting selected nodes implements the contraction property.

The main theorem is that, for any initial set of CI statements encoded in a MUG, a statement is derivable via the graphoid axioms if and only if it can be obtained via a finite sequence of these graphical operations, making graphical CI reasoning a complete alternative to algebraic proof for these axioms (Shachter, 2013). This equivalence forms the basis for the graphical analysis of commutation properties, as well as for practical algorithms (e.g., junction-tree inference, moralization, d-separation reasoning).

3. Markov Properties and Maximal Independent Sets

In the context of undirected CI graphs, three Markov properties (pairwise, local, global) are central:

• Pairwise Markov: Nonadjacent nodes are conditionally independent given the rest.
• Local Markov: Each node is independent of all non-neighbors given its boundary.
• Global Markov: Separation in the graph corresponds to conditional independence.

The mutual conditional independence property (MCIP) states that, in a maximal independent set $I$ of the graph, $\{X_{i_1},...,X_{i_k}\}$ are mutually independent given $V\setminus I$. This property is both necessary and sufficient—subject to positivity—for the equivalence of the three Markov properties. Factorization theorems (e.g., Hammersley-Clifford) specify that, for strictly positive distributions, independence structures induced by a graph coincide precisely with its separation properties (Gauraha, 2016).

In decomposable (chordal) graphs, Lauritzen’s factorization allows for efficient inference by mapping CI structure to a junction-tree over maximal cliques. Inference methods exploit MCIP to split high-dimensional computations into lower-dimensional ones whenever possible.

4. Variants: Directed, Mixed, and Generalized CI Graphs

Directed and Mixed Graphs

Bayesian networks, encoded as DAGs, employ $d$-separation for CI queries. When introducing latent variables, marginalization can lead to mixed graphs with bidirected (and possibly higher-order) edges, encoded graphically as acyclic directed mixed graphs (ADMGs) or mixed DAGs (mDAGs) (Evans, 2014). Marginal conditional-independence structure may necessitate the use of hyperedges to preserve the full set of marginal dependencies, which cannot be accurately represented by pairwise bidirected edges alone.

Summary Graphs

Summary graphs are used to capture CI after sequences of marginalization and conditioning, with path-based separation rules tailored for these operations. Such graphs can alert analysts to direct and indirect confounding, and correctly represent identifiability status of path coefficients and the preservation of certain independence or dependence relations that would be misrepresented by standard marginal graphs (Wermuth, 2010).

Chain Event Graphs

For statistical problems with logical zeros or asymmetric unfolding (e.g., event sequences), chain event graphs (CEGs) use positions (sets of indistinguishable event-tree vertices) and stages to encode context-specific CI, with a separation theorem based on the existence of cut-vertices separating position-variables (Thwaites et al., 2015).

5. Extensions and Specialized CI Graphs

Stationary Diffusions and Stochastic Processes

In continuous-time and time-series settings, conditional-independence structure in stationary diffusions is characterized by the sparsity pattern of the drift vector field. The bidirected trek graph, constructed via treks in the directed drift graph, encodes the CI structure in the stationary distribution: separation in this trek graph corresponds exactly to conditional independence in the law of the process (Boege et al., 2024). In time series, differential conditional-independence graphs can be constructed by estimating changes in the inverse power spectral density (IPSD), correspondingly inferring differences in CI structure between time periods (Tugnait, 7 Dec 2025).

Nonlinear or Nonstandard Algebraic Models

Max-linear Bayesian networks, motivated by extreme-value modeling, exhibit CI structures that go beyond what is encoded by $d$-separation. Context-specific, context-free, and support-induced CI relations arise, and are precisely represented only via adapted "source DAGs" and $*$-separation, often involving tropical algebraic criteria (Améndola et al., 2020).

Categorical and Lattice Perspectives

Universal algebraic frameworks recast CI structures as objects in categoroids, which simultaneously encode preordered (binary) and ternary (CI) relationships, with bridge morphisms linking the two (Mahadevan, 2022). Lattice CI models and Hibi ideals provide an algebraic (toric, ideal-theoretic) representation, connecting to distributive lattices and graph-theoretic transitive closures (Caines et al., 2021).

6. Redundancy, Axiomatization, and Verification

The existence of multiple redundant CI tests in structure-learning procedures is both a challenge and an opportunity for empirical and theoretical work (Faller et al., 12 Feb 2025). Various kinds of redundancy are delineated:

• Graphoid redundancy: Follows from graphoid axioms given prior CI tests.
• Graphical redundancy: Imposed uniquely by the assumed graphical structure.
• Purely graphical redundancy: A statement that is graphically but not axiomatically redundant, detecting errors arising from unfaithfulness or sample noise.

The Markov and faithfulness properties are essential for ensuring that graphical CI structures correspond to those of probability distributions. Compositional graphoids—semigraphoids closed under Intersection and Composition—exactly characterize the CI models that are representable by graphs, with the new information-theoretic criteria supplying sufficient discrete-variable conditions for these properties (Boege, 16 Apr 2025).

In structure-learning, constraint-based algorithms such as PC, SGS, and SP exploit these properties for efficient recovery and verification. Redundant or purely graphical CI statements can be leveraged to detect and correct errors due to unfaithfulness or statistical uncertainty.

7. Faithfulness, Interventions, and Context-Specific Structures

Faithfulness—the requirement that all and only those CI relations induced by the graph are present in the distribution—is critical for the soundness of inference and learning (Mogensen, 2023). When faithfulness fails (e.g., when the set of CIs in the data is strictly larger than those implied by the graph), appropriate algorithms (e.g., trimming, minimum-distance selection) can be used to construct faithful supergraphs or to correct for discrepancies.

In settings with interventions, such as causal inference, latent projection (as in mDAGs) commutes with interventions by properly adjusting both directed and bidirected (hyper)edges, preserving the causal interpretation under node manipulation (Evans, 2014).

Specialized frameworks such as local independence graphs for stochastic processes extend CI notions to allow for Granger causal modeling and generalizations ($\delta$- and $\mu$-separation), with nuanced characterizations of faithfulness and explicit algorithms for graph recovery given independence data (Mogensen, 2023).

---

References:

• Shachter, R. D. "A Graph-Based Inference Method for Conditional Independence," (Shachter, 2013)
• Uhler, C., et al. "Mutual Conditional Independence and its Applications to Inference in Markov Networks," (Gauraha, 2016)
• Neal, R. M. "On Deducing Conditional Independence from d-Separation in Causal Graphs with Feedback," (Neal, 2011)
• Boege, T., Drton, M., Hollering, A., and Weiss, S. "Conditional Independence in Stationary Diffusions," (Boege et al., 2024)
• Evans, R. J. "Graphs for Margins of Bayesian Networks," (Evans, 2014)
• Boege, T. "On the Intersection and Composition properties of conditional independence," (Boege, 16 Apr 2025)
• Amendola, C., Klüppelberg, C., Lauritzen, S., Tran, N. M. "Conditional Independence in Max-linear Bayesian Networks," (Améndola et al., 2020)
• Maier, M., Marazopoulou, K., and Jensen, D. "Reasoning about Independence in Probabilistic Models of Relational Data," (Maier et al., 2013)
• Wermuth, N. "Probability distributions with summary graph structure," (Wermuth, 2010)
• Mogensen, P. "Faithful graphical representations of local independence," (Mogensen, 2023)
• Lam, W.Y., Andrews, J.L., Ramsey, J. "On Different Notions of Redundancy in Conditional-Independence-Based Discovery of Graphical Models," (Faller et al., 12 Feb 2025)
• Gad N., Balakrishnan A. "Knowledge Propagation over Conditional Independence Graphs," (Chajewska et al., 2023)
• Sullivant, S., Lin, S., and Yoshida, R. "Lattice Conditional Independence Models and Hibi Ideals," (Caines et al., 2021)
• Kolmogorov, V. "Categoroids: Universal Conditional Independence," (Mahadevan, 2022)
• Tugnait, J. et al. "Learning Conditional Independence Differential Graphs From Time-Dependent Data," (Tugnait, 7 Dec 2025)
• Geiger, D., Pearl, J. "Logical and algorithmic properties of conditional independence and graphical models," Ann. Statist. 1993.