Minimal Trails in Restricted DAGs
Alexis Derumigny, Niels Horsman, Dorota Kurowicka
TL;DR
This work addresses the problem of characterizing minimal trails activated by a subset $Z$ in directed acyclic graphs that exclude active cycles, with applications to copula-based Bayesian networks. The authors define TRAILS$(X,Y|Z)$, the converging-connection set ConvCon$(T)$, closest descendants, subtrails, and a partial order $<_{TRAIL}$ to compare trails, proving structural properties of minimal trails under the no-active-cycle constraint. A key contribution is showing how converging nodes interact with the activation set $Z$, especially when $Y 1d Z$ has local relationships, leading to precise subgraph configurations and d-separation implications. These results yield a framework for efficient conditional independence reasoning in constrained DAGs and PCBNs, potentially aiding computations in copula-based BN models.
Abstract
In this paper, the properties of minimal trails in a directed acyclic graph that is restricted not to contain an active cycle are studied. We are motivated by an application of the results in the copula-based Bayesian Network model developed recently. We propose a partial order on the set of trails activated by a certain subset of nodes, and show that every minimal trail, according to such an order, has a simple structure.
