Structural Identifiability in Non-Gaussian Graphical Discrete Lyapunov Models
This paper studies the problem of structural identifiability for graphical discrete Lyapunov models in the non-Gaussian setting from observational data. A graphical discrete Lyapunov model associated with a graph $G$ consists of the set of stationary distributions of stable first-order vector autoregressive models whose transition matrices have sparsity patterns determined by $G$. This model has a causal interpretation as $G$ can be viewed as encoding dynamic causal relationships between random variables of interest. It has been shown that generic parameter identifiability holds for reflexive directed acyclic graphs (DAGs). Structural identifiability for these models remains unstudied beyond certain restrictive cases. In this work, we prove a non-constructive structural identifiability result for reflexive DAGs in non-Gaussian graphical discrete Lyapunov models by leveraging the information contained in the entire collection of stationary cumulants. We also provide a concrete algorithm for learning reflexive directed trees from stationary cumulants of orders $\leq 5$.