Equations defining probability tree models (Q2284967): Difference between revisions

This paper studies staged trees using algebraic geometry. Staged trees are particular graph structures encoding sequences or chains of events that are conditionally independent; the edges on the graph are probability weights. The graphical structure of staged trees together with the probabilistic characteristic on the edges appeals to an algebraic geometric approach to studying these objects. In algebraic geometry, one is interested in studying the geometry of zero sets of systems of polynomials, known as algebraic varieties. This paper studies the systems of equations characterizing classes of staged trees using algebraic geometry to understand certain probabilistic properties and features of algebraic geometric objects associated with staged trees. Staged trees are similar in spirit to statistical graphical models; both are graphical structures that encode relationships between events in terms of conditional independence, but staged trees are strictly sequential and events (nodes) arise based on the ``stages'' of these depicted events, whereas events (nodes) in graphical models are random variables. Also, the conditional independence in staged trees is between the stages, while in graphical models, it arises in the edges between the random events. Graphical models have been studied using algebraic geometry; in particular, it has been shown that decomposable graphical models can be characterized as toric varieties [\textit{G. Pistone} et al., Algebraic statistics: Computational commutative algebra in statistics. Boca Raton, FL: Chapman \& Hall (2001); Zbl 0960.62003; \textit{D. Geiger} et al., Ann. Stat. 34, No. 3, 1463--1492 (2006; Zbl 1104.60007)]. Toric varieties are important objects in algebraic geometry: they are algebraic varieties that contain algebraic tori, where algebraic actions on the torus extend to the entire variety. In other words, they encode a local-to-global property, locally from the torus, to the entire variety, globally. Such local-to-global characterizations are of great interest in statistical inference frameworks. A local-to-global property is desirable for staged trees, because it would provide insight on the full sequence of events on the entire staged tree based on studying local observations of events (stages). Staged trees, in general, do not correspond to toric varieties, so an algebraic characterization of such a local-to-global property is missing for staged trees. This paper fills in this gap and gives conditions under which staged trees can be described by toric varieties. This paper also provides other algebraic geometric properties of this toric variety of staged trees. In particular, it studies the toric ideal of staged trees using its combinatorial structure and gives its generators.