Rational partition models under iterative proportional scaling

TL;DR

This paper studies partition models in log-linear modeling, introducing the Generalized Running Intersection Property (GRIP) that ensures the iterative proportional scaling algorithm computes the exact MLE in one iteration, linking to toric fiber products.

Contribution

It defines the GRIP condition for partition models, enabling exact MLE computation in one IPS cycle and connects this to toric fiber products and existing models.

Findings

GRIP guarantees one-cycle IPS convergence

Connection established between GRIP and toric fiber product

Characterization of balanced, stratified staged trees via GRIP

Abstract

In this work we investigate partition models, the subset of log-linear models for which one can perform the iterative proportional scaling (IPS) algorithm to numerically compute the maximum likelihood estimate (MLE). Partition models include families of models such as hierarchical models and balanced, stratified staged trees. We define a sufficient condition, called the Generalized Running Intersection Property (GRIP), on the matrix representation of a partition model under which IPS algorithm produces the exact MLE in one cycle. Additionally we connect the GRIP to the toric fiber product and to previous results for hierarchical models and balanced, stratified staged trees. This leads to a characterization of balanced, stratified staged trees in terms of the GRIP.