Graphical Modeling from Marginal Information via Optimal Transport
stat.ME, math.ST, stat.TH
Submitted: 2026-10-29
Updated: 2026-10-29
License: http://creativecommons.org/licenses/by/4.0/
The gist: A graphical model represents multivariate interactions through conditional independence, but requires jointly observed samples to be fitted to data.
Terminology
Abstract
A graphical model represents multivariate interactions through conditional independence, but requires jointly observed samples to be fitted to data. When individuals cannot be tracked and only marginal observations are available, as with destructive measurements leading to cross-sectional data, the problem of graphical modeling is confounded with that of assembling a joint law. We show how simultaneous graphical modeling and joint law assembly can be realized via a multi-marginal optimal transport problem with graphical cost. Focusing on Gaussian marginals, we generate a joint law satisfying the Markov properties prescribed by a given graph. Within the class of joint distributions respecting the graph and marginals constraints, our construction selects a member by making the interactions set by the graph as strong as possible under a dependence budget measured by relative entropy. In the empirical setting, where only estimates of the marginals are available, we examine the stability of the problem under perturbations of both the marginals and of the graph weights. The resulting error splits into an estimation term, vanishing with the sample size, and a specification term, which quantifies the cost of a misspecified graph. Finally, we illustrate these findings by numerical experiments.
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes