Characteristic Imset Framework

• The characteristic imset framework is an algebraic and polyhedral approach that encodes DAG properties as 0–1 vectors, uniquely representing Markov equivalence classes.
• It enables efficient linear optimization through block factorization, toric ideals, and explicit polyhedral characterizations to support causal discovery and Bayesian network structure learning.
• Its applications extend from diagnosis models and ordered-node networks to interventional settings, offering practical insights in combinatorial and algebraic analysis of graphical models.

The characteristic imset (CIM) framework is an algebraic and polyhedral approach to the combinatorial and statistical analysis of graphical models, particularly directed acyclic graphs (DAGs), Markov equivalence classes, and Bayesian network structure learning. It encodes combinatorial and statistical properties of DAGs via 0–1 vectors called characteristic imsets, whose convex hulls—the CIM polytopes—form the feasible regions for linear optimization in score-based causal discovery and model selection. The framework provides unique representatives for Markov equivalence classes, admits explicit polyhedral characterizations in certain structured cases, and supports algebraic and combinatorial study of causal inference through its tight connection to integer programming, toric ideals, and supermodularity.

1. Algebraic Foundations: Characteristic Imsets and Standard Imsets

Given a finite set $N$ of nodes and a DAG $G$ on $N$, the characteristic imset $c_G$ is a binary vector indexed by all subsets $S \subseteq N$ with $|S| \geq 2$, defined as follows: $$c_G(S) = \begin{cases} 1, & \exists\,i\in S:\;S\setminus\{i\} \subseteq \mathrm{pa}_G(i) \ 0, & \text{otherwise} \end{cases}$$ Here, $\mathrm{pa}_G(i)$ is the parent set of node $i$ in $G$. Thus, $c_G(S) = 1$ if $S$ forms a star with sink $i$ covered by its parents, i.e., $S$ is "complete toward" $i$ in the induced subgraph $G[S]$ (Xi et al., 2012, Restadh, 2023).

A closely related object is the standard imset $u_G$, an integer vector (possibly non-binary) that provides another unique representation of the Markov equivalence class. The characteristic imset is obtained from the standard imset by Möbius inversion, yielding a 0–1 vector while preserving equivalence: $c_G(S) = 1 - \sum_{T \supseteq S} u_G(T)$ Standard imsets have less immediate graph-theoretic interpretations, while characteristic imsets directly reflect underlying parent sets and v-structures.

2. The Characteristic-Imset Polytope: Definition and Structure

For a class $\mathcal{G}$ of DAGs on $N$, the characteristic-imset polytope (CIM polytope) is the convex hull of all characteristic imsets from $\mathcal{G}$: $$\mathbf{P}_{\mathcal{G}} = \operatorname{conv}\{c_G : G \in \mathcal{G}\} \subseteq \mathbb{R}^{2^{|N|} - |N| - 1}$$ For the full family of DAGs, this is the full CIM polytope; faces corresponding to restricted skeletons (e.g., fixed undirected graph $G$) are subpolytopes simplicially defined within $\mathbf{P}_{\mathcal{G}}$ (Xi et al., 2012, Restadh, 2023, Linusson et al., 2022).

Block Structure for Structured Families

• Diagnosis models (bipartite DAGs): The CIM polytope factorizes as a direct product of simplices:

$$\mathbf{P}_{m,n} = \underbrace{\Delta_{2^m-1} \times \cdots \times \Delta_{2^m-1}}_{n\ \text{factors}}$$

Each factor corresponds to all possible parent sets for a single 'symptom' node (Xi et al., 2012).

• Ordered-node networks: For DAGs with a fixed node ordering, the polytope is also a product of simplices, the $i$th factor reflecting the possible parent sets for node $i$ among nodes $< i$ (Xi et al., 2012):

$$\mathbf{P}_{[n]} = \Delta_{2^1-1} \times \Delta_{2^2-1} \times \cdots \times \Delta_{2^{n-1}-1}$$

This block factorization enables modular, efficient optimization procedures on the polytope for these structured classes.

3. Polyhedral Geometry: Faces, Edges, and Facets

Vertices, Edges, and Facets

Every vertex of the CIM polytope corresponds to a characteristic imset associated to a unique Markov equivalence class (Restadh, 2023, Cussens et al., 2015). The combinatorial and geometric properties are as follows:

• Edges: In the diagnosis model, two vertices (imsets) are connected by an edge if and only if they differ in the parent set of exactly one symptom node (Xi et al., 2012).
• Facets: Since structured polytopes like the diagnosis model are products of simplices, their facets arise from those of the factors. For a simplex $\Delta_{2^m-1}$ (corresponding to a single variable's parent set), the facets have explicit combinatorial form:

$$\sum_{t \supseteq s} (-1)^{|t| - |s|} x_t \geq 0 \quad \text{for all } s \subseteq A$$

with $x_t = c_G(\{t\} \cup \{b\})$ (Xi et al., 2012).

General Families: Connectivity and Edge Structure

For general undirected skeleton $G$, the CIM polytope's face structure is rich. The diameter of $\mathrm{CIM}_n$, defined as the maximal edge-walk distance between any two vertices, obeys the bound $\mathrm{diam}(\mathrm{CIM}_n) \leq 2n - 2$, and for face polytopes corresponding to fixed skeleton $G$, the diameter is $\leq |E|$ (Restadh, 2023). In the tree case, explicit combinatorial characterizations (essential flips, v-structure reversals) permit even sharper bounds and enable full edge descriptions (Linusson et al., 2022).

4. Algebraic and Toric Aspects: Ideals and Fiber Products

Characteristic imsets are crucial in algebraic statistics via their role in toric ideals:

• Toric ideals: The kernel of the monomial map sending equivalence classes to their monomials in $t_S$ encodes the relations (moves) in the polytope and is called the characteristic imset ideal (Hollering et al., 2022). For tree-skeletons, these ideals admit a quadratic Gröbner basis, with each move corresponding to minimal relations between DAGs connected by single essential flips.
• Quasi-independence gluing (QIG): This is a generalization of toric fiber products, permitting polytope (and ideal) assembly along shared substructures (e.g., gluing tree polytopes over edge-overlaps). Such iterative constructions yield explicit Gröbner bases and are essential in decomposing the algebraic and polyhedral structure of CIM polytopes for chordal and tree-like skeletons (Hollering et al., 2022).

5. Applications to Causal Discovery and Bayesian Network Learning

The CIM framework underpins several key advances in computational learning of Bayesian networks and causal graphical models:

• Score-equivalence and linear optimization: For many decomposable, score-equivalent criteria (BIC, MDL, AIC), the problem of selecting an optimal network reduces to linear integer programming over the CIM polytope (Restadh, 2023, Cussens et al., 2015, Studeny et al., 2011). For block-structured polytopes (e.g., diagnosis models, ordered DAGs), this enables efficient, parallelizable optimization.
• Causal discovery algorithms: Greedy edge-walk algorithms (such as GES and specialized tree-learning algorithms exploiting the essential-flip structure) operate on the CIM polytope and traverse Markov equivalence classes efficiently. The edge structure directly informs the design of move-sets for efficient search (Linusson et al., 2022, Restadh, 2023).
• Polyhedral relaxations and LP bounds: Explicit and implicit LP-relaxations of the CIM and standard imset polytopes (using cluster, facet, and supermodular inequalities) facilitate tractable lower bounds and tighten feasible regions in mixed-integer programming solvers (Studeny et al., 2011, Cussens et al., 2015).

6. Generalizations, Interventions, and Extensions

Extensions of the CIM framework include:

• Interventional CIM polytopes: When causal discovery leverages both observational and interventional data, the corresponding CIM polytopes (faces associated to DAGs with fixed interventions) admit explicit polyhedral descriptions, particularly in tree settings, using Möbius-type facet constructions and toric fiber product decompositions (Hollering et al., 2024).
• Cyclic graphs and general graphs: The CIM definition and associated toric ideals generalize to directed graphs with cycles, where the imset is an integer vector with coordinates counting covering patterns within cycles; such imsets parametrize covariance equivalence classes for linear structural equation models (Johnson et al., 16 Jun 2025).
• Conditional independence structures: Underlying the CIM approach is the semi-elementary imset framework, which encodes the semi-graphoid axioms of CI as linear identities among imsets, supporting decomposition, enumeration, and Markov basis analysis in algebraic statistics (Kashimura et al., 2011).

7. Open Problems and Future Directions

Several key problems remain open:

• Facet description: For general graphs, the complete facet characterization of the CIM polytope is unknown except for small $n$ or strong structural restrictions (tree, star, order) (Restadh, 2023, Cussens et al., 2015).
• Diameter and edge structure: For arbitrary skeletons and especially for the full polytope, determining all minimal generating moves remains challenging (Restadh, 2023).
• Computational trade-offs: While explicit relaxations enable efficient optimization, tighter implicit relaxations (via supermodular inequalities) pose computational challenges due to their large or intractable description (Studeny et al., 2011).
• Extensions to mixed/chain graphs: Generalizing CIM and semi-elementary imset approaches to BDAGs, ADMGs, or chain graphs remains an important active area (Andrews et al., 2022).

The characteristic imset framework continues to provide powerful tools for high-dimensional structure learning, causal discovery, and algebraic-statistical analysis of graphical models, with ongoing advances in polyhedral, combinatorial, and algorithmic complexity (Xi et al., 2012, Restadh, 2023, Hollering et al., 2022, Linusson et al., 2022, Hollering et al., 2024, Johnson et al., 16 Jun 2025).