Colored Gaussian Graphical Models

• Colored Gaussian Graphical Models are multivariate normal models that impose sparsity and equality constraints via vertex and edge coloring, reducing parameter complexity.
• They leverage algebraic geometry and combinatorial symmetries to yield tractable vanishing ideals and facilitate closed-form solutions in model estimation.
• The approach enhances Bayesian inference, efficient model search, and maximum likelihood estimation, making it valuable for high-dimensional network analysis.

Colored Gaussian Graphical Models (CGGMs) are a class of multivariate normal models in which both conditional independence (sparsity) and equality (symmetry) constraints are imposed on the precision (concentration) matrix through vertex-and-edge coloring of an underlying graph structure. This framework generalizes classical Gaussian graphical models by reducing the number of free parameters via combinatorial symmetries and has substantial impact on model selection, algebraic geometry, and Bayesian inference in high-dimensional statistics (Coons et al., 2021, Chojecki et al., 23 Jan 2026, Biaggi et al., 8 Jul 2025).

1. Model Definition and Combinatorial Structure

A CGGM is defined by an undirected graph $G = (V, E)$ together with disjoint partitions of vertices and edges:

• Vertex colors $\mathcal V = \{ V_1, ..., V_r \}$, each imposing $K_{ii} = K_{jj}$ for $i, j \in V_a$.
• Edge colors $\mathcal E = \{ E_1, ..., E_R \}$, each imposing $K_{ij} = K_{kl}$ for $\{i, j\}, \{k, l\} \in E_{b}$.

The concentration matrix $K$ is restricted to be symmetric positive definite and satisfy both the sparsity (conditional independence) and symmetry constraints: $$K_{ij} = 0 \quad \text{if } \{i, j\} \notin E; \quad K_{ii} = K_{jj} \text{ if } i, j \text{ in same vertex color}; \quad K_{ij} = K_{kl} \text{ if edges share color}.$$ The parameter space is a linear subspace of $\mathrm{Sym}^+(p)$ of dimension $r + R$. Special subclasses include the RCOP models where coloring arises as orbits under a permutation symmetry group $\Gamma \subset \mathrm{Aut}(G)$, leading to full symmetry within color classes (Coons et al., 2021, 2207.13330).

2. Algebraic Geometry and Vanishing Ideals

CGGMs have rich algebraic structure: the Zariski closure of all covariance matrices $\Sigma = K^{-1}$ produces a vanishing ideal $I_G$ in the polynomial ring $\mathbb{R}[\sigma_{ij}]$. For RCOP-colorings on block graphs (chordal graphs with clique cycles), the variety is toric in the covariance coordinates, and the ideal is binomial—generated by linear and quadratic binomials:

• Linear binomials: $\sigma_{ij} - \sigma_{kl}$ whenever the color-multisets along shortest paths coincide.
• Quadratic binomials: $\sigma_{ij}\sigma_{kl} - \sigma_{ik}\sigma_{jl}$ under certain edge-multiset conditions (Coons et al., 2021, Biaggi et al., 8 Jul 2025).

A key combinatorial theorem asserts that binomiality of the vanishing ideal is equivalent to the graph being a triangle-regular block graph—a property encoded by a (symmetric) Jordan scheme on the adjacency matrices of the coloring, generalizing association schemes but requiring closure under the Jordan product $K \circ K' = \frac{1}{2}(KK' + K'K)$ (Biaggi et al., 8 Jul 2025).

Notably, this binomiality does not require RCOP coloring in general—there exist triangle-regular colorings not arising from group orbits, refuting prior conjectures (Biaggi et al., 8 Jul 2025).

3. Bayesian Inference and Explicit Normalizing Constants

A pivotal challenge in Bayesian learning of CGGMs is the computation of Diaconis–Ylvisaker normalizing constants for the colored $G$-Wishart prior: $$p(K) \propto |\det K|^{(\delta-2)/2} e^{-\frac{1}{2}\text{tr}(D K)} 1_{K \in \mathcal{P}_G^{\mathcal{C}}}$$ The normalizing constant

$$I_{G, \mathcal{C}}(\delta, D) = \int_{\mathcal{P}_G^{\mathcal{C}}} |\det K|^{(\delta-2)/2} e^{-\frac{1}{2}\text{tr}(D K)} dK$$

is intractable except for decomposable graphs and certain RCOP models (2207.13330).

Recent advances identify broad combinatorially defined subclasses—Block-Cholesky spaces (BC-spaces) and Diagonally Commutative Block-Cholesky spaces (DCBC-spaces)—over which these integrals admit closed-form solutions. These arise if the colored graph satisfies a color perfect elimination ordering (CPEO) and a 2-path regularity condition ("Color Elimination-Regular (CER) graphs"). All RCOP models on decomposable graphs are symmetric CER graphs, but the class is strictly larger (Chojecki et al., 23 Jan 2026). Computation in the commutative case is tractable via Jordan frame decompositions and Gram–Cholesky factorizations.

4. Symmetry, Model Search, and Statistical Regularity

CGGMs motivate a lattice-theoretic organization of model classes:

• Edge-regular, vertex-regular, regular (both), and permutation-generated subclasses form complete, though non-distributive, lattices of models under refinement.
• The Edwards–Havrânek procedure exploits this lattice to efficiently search for optimal colored models, drastically reducing search complexity compared to enumeration. In edge-regular models one need only explore polynomially many rejection duals at each stage (Gehrmann, 2011).

Model selection and estimation leverage penalized composite likelihood or Bayesian MCMC approaches, explicitly incorporating symmetry constraints and sparsity via grouping and selection penalties or carefully constructed reversible-jump Markov chains (Li et al., 2020, Li et al., 2020).

5. Maximum Likelihood Estimation and Sample Size Thresholds

Existence and uniqueness of the MLE in CGGMs depends crucially on the interplay between graph structure, coloring, and data sample size:

• An algebraic elimination criterion determines the minimal sample size for generic existence of the MLE; colored models often yield lower thresholds than their uncolored counterparts due to parameter space reduction (Uhler, 2010, Makam et al., 2021).
• For colored cycles and other symmetric graphs, explicit formulas for thresholds and ML degrees are available. For instance, colored 4-cycles require only two samples, below the treewidth bound (Uhler, 2010).
• Closed-form MLEs exist for certain block- or symmetry-regular models, and mean estimation reduces to least squares if and only if the mean partition is finer than the vertex coloring and equitable across all edge colors (Gehrmann et al., 2011).

6. Extensions to Directed, Ancestral, and DAG Models

Recent work generalizes the colored symmetry paradigm to directed graphical models (RDAGs) and DAG-based CGGMs. In such models, coloring controls equivalence classes for error variances and structural coefficients, with identifiability and Markov characterization tightly coupled to the color-induced block structure (Boege et al., 2024, Makam et al., 2021).

• BPEC-DAGs, a subclass of colored DAGs with properly blocked edge colors, permit structurally identifiable causal modeling and efficient learning via greedy edge-colored search algorithms.
• The algebraic saturation framework for vanishing ideals extends from undirected to directed and ancestral graphs, providing tools for identifying Markov properties in rationally parameterized colored models.

7. Practical and Theoretical Implications

CGGMs underpin parsimony in high-dimensional network inference—genomics, neuroscience, finance—where block-symmetry and conditional independence co-occur. The toric and binomial structure yields tractable Markov bases for algebraic statistics and closed-form ML degree calculations (Coons et al., 2021, Biaggi et al., 8 Jul 2025). The ability to analytically compute normalizing constants for large classes of CGGMs has transformed Bayesian model selection and causal inference in symmetric graphical settings (Chojecki et al., 23 Jan 2026, 2207.13330).

The algebraic characterization via Jordan schemes exposes deep connections to combinatorics and invariant theory, reshaping understanding of which symmetries matter in structured probabilistic models.

In summary, Colored Gaussian Graphical Models unify symmetry and sparsity in networked statistical models, offering tractable algebraic structure, powerful model selection methodologies, and principled Bayesian inference for high-dimensional graphical data.