New 'separable graphs' class unifies complex causal and feedback models

A new paper from Christopher Meek and Kayvan Sadeghi (arXiv:2607.01057) tackles a fundamental problem in graphical modeling: how to characterize and identify models whose independence structure is determined by vertex separation in mixed graphs. The authors introduce *separable graphs*, a family that includes all existing graphical model subfamilies (DAGs, chain graphs, ancestral graphs, etc.) but with a stricter property — every missing edge implies the existence of a separating set for its endpoints. This property ensures a clean, consistent mapping between graph structure and conditional independence relations.

The paper goes further to define *essentially separable graphs*, which are separation-equivalent to a separable graph, and provides three key contributions: (1) a purely graphical characterization of separability using ordinary graph properties (extending results for subfamilies), (2) a separational characterization that depends only on the independence relations themselves, and (3) a canonical representation for equivalence classes of essentially separable graphs. The authors also develop an algorithm that, under suitable assumptions, identifies the equivalence class of any essentially separable graph from observed data. At 69 pages with 7 figures, this work offers a unifying framework for handling complex dependencies arising from feedback loops, latent variables, and selection mechanisms — something increasingly important for modern causal inference and machine learning pipelines.