Cherry-Tree Copulas in Multivariate Modeling

• Cherry-tree copulas are multivariate dependence models defined via junction trees that encode conditional independence through k-sized clusters and (k-1)-sized separators.
• They factorize the joint copula density by decomposing it into cluster-based and separator-based copulas, aligning with truncated R-vine constructions and ensuring computational tractability.
• Algorithmic constructions using matrix representations and perfect elimination orderings bridge graphical models, vine copulas, and chordal graphs for efficient multivariate analysis.

Cherry-tree copulas are multivariate dependence models constructed via junction-tree representations where conditional independences are encoded structurally, providing a tractable extension of pair-copula constructions. These copulas generalize the cluster-based graph-theoretic methodology underlying R-vine copulas, allowing factorization of the joint copula density through higher-order cliques ("cherries") and their separators. They play a central role in the study of conditional independence structures in copula theory and establish an explicit connection between graphical models, vine copulas, and matrix representations within the context of multivariate probability distributions (Kovács et al., 2016, Pfeifer et al., 2022).

1. Formal Definition and Graph Structure

For a continuous random vector $X = (X_1,\ldots,X_d)$, a $k$-order cherry-tree is defined as a junction tree over $V = \{1,\ldots,d\}$, where every cluster ("node") $K \subset V$ has cardinality $k$, and every separator (the intersection of adjacent clusters) has cardinality $k-1$. The junction tree must satisfy the running-intersection property: for any pair of clusters, their intersection is contained in each intermediate cluster along the unique path connecting them. In this structure, conditional independence is encoded as follows: if $S$ separates clusters $A$ and $B$, then $X_A \perp X_B \mid X_S$, and vertices not simultaneously present in any cluster are independent conditional on the appropriate separator (Kovács et al., 2016).

2. Copula Density Factorization

A cherry-tree copula factorizes the joint copula density according to the cluster and separator sets of the $k$-order cherry-tree. Let $\mathcal{C}_{ch}$ denote the set of clusters and $\mathcal{S}_{ch}$ the set of separators; $\nu_S$ is the number of clusters containing separator $S$. The copula density is

$$c(\mathbf{u}) = \frac{\prod_{K \in \mathcal{C}_{ch}} c_{X_K}(u_K)}{\prod_{S \in \mathcal{S}_{ch}} [c_{X_S}(u_S)]^{\nu_S - 1}}$$

where $c_{X_T}(u_T)$ is the $|T|$-variate copula density over $T \subseteq V$ evaluated at $u_T = (F_i(x_i))_{i \in T}$, with denominator terms adjusting for over-counting due to cluster overlaps (Kovács et al., 2016).

3. Relationship to Truncated R-vine Copulas

A truncated R-vine copula at level $k$ is an R-vine in which all pair copulas conditioned on sets of cardinality at least $k$ are set to independence; the joint density thus involves only trees $T_1,\ldots,T_{k-1}$ (Kovács et al., 2016). Kovács–Szántai’s theorem provides a necessary and sufficient criterion: a $k$-order cherry-tree copula is a truncated R-vine copula at level $k$ if and only if its set of separators forms a $(k-1)$-order cherry-tree. Equivalently, each $k$-cluster is adjacent to at most two distinct separators. If this criterion is not met, a “lift” procedure can always construct a $(k+1)$-order cherry-tree that is representable as a truncated R-vine at level $k+1$.

4. Cherry-tree Sequences and Chordal Graphs

A regular cherry-tree sequence $\{\Gamma_k\}_{k=1}^{n-1}$ consists of cherry-trees at each order $k$, where nodes of $\Gamma_k$ are the clusters of the corresponding vine tree $T_k$ and separators are intersections of adjacent clusters—the separator sets of $\Gamma_k$ match the clusters of $\Gamma_{k-1}$. This construction is equivalent to both the original R-vine description and the chordal graph (maximal clique) description. Gavril’s theorem asserts that a graph is chordal iff it admits a perfect elimination ordering (PEO), and any cherry-tree (by its construction) admits a PEO, which is central for the matrix representation of vines (Pfeifer et al., 2022).

5. Matrix Representation and Algorithmic Construction

Cherry-tree copula structures admit a unique $n \times n$ lower-triangular matrix representation once a PEO is specified. Two algorithmic approaches—Nápoles’ column-wise method and a row-wise cherry-tree walk—yield the same matrix when initialized with the same PEO. Both fill the main diagonal with the PEO sequence, set the bottom row from $T_1$, and recursively fill subdiagonal entries by matching clusters and separators iteratively through the cherry-tree sequence.

Algorithm	Approach	Complexity
Column-wise	Scan edges in $T_k$ by columns	$O(n^3)$
Cherry-tree	Walk cherry-tree sequence by rows	$O(n^3)$ or $O(n^3\log n)$ (unsorted clusters)

These constructions guarantee that for a fixed PEO, the vine matrix encoding the cherry-tree copula structure is unique, allowing reversible transitions between graphical and matrix representations (Pfeifer et al., 2022).

6. Illustrative Example

Consider $V = \{1,2,3,4,5\}$ and $k = 3$ with clusters $\mathcal{C}_3 = \{\{1,2,3\},\{2,3,4\},\{3,4,5\}\}$, separators $\mathcal{S}_3 = \{\{2,3\},\{3,4\}\}$. The cherry-tree copula density is:

$$c(u_1,\dots,u_5) = \frac{c_{1,2,3}(u_1,u_2,u_3) c_{2,3,4}(u_2,u_3,u_4) c_{3,4,5}(u_3,u_4,u_5)}{c_{2,3}(u_2,u_3) c_{3,4}(u_3,u_4)}$$

Here, $\mathcal{S}_3$ itself forms a 2-order cherry-tree, so the copula is a truncated R-vine at level $3$. The Backward Algorithm reconstructs the vine tree sequence and edge-labeled pair-copulas recursively, allowing expression of the joint density via the vine structure (Kovács et al., 2016).

7. Graphical Equivalence and Applications

Cherry-tree copulas unify multiple graphical descriptions: R-vine sequence, chordal graphs, and junction trees of cliques. The equivalence facilitates factorization, computation, and model selection in high-dimensional settings, leveraging conditional independence and graphical sparsity to overcome computational intractability. When a good-fitting cherry-tree copula is found, and the separator-tree criterion is satisfied, full pair-copula decomposition follows; otherwise, the lift procedure expands the structure for compatibility. These methodologies underpin efficient multivariate modeling in statistics and machine learning, particularly where conditional independence assumptions are interpretable and exploitable (Kovács et al., 2016, Pfeifer et al., 2022).