By François Pachet and Jean-Daniel Zucker.
Frontiers in Artificial Intelligence, volume 9, article 1840945, 23 October 2026. DOI: 10.3389/frai.2026.1840945.
Read the published paper (PDF)
Open-access journal article
Earlier arXiv preprint · Preprint PDF
About the paper
How can we generate a population of synthetic individuals from statistics about groups? François Pachet and Jean-Daniel Zucker study populations described by categorical attributes, with target frequencies for individual attributes and combinations of attributes. Matching many overlapping combinations becomes computationally difficult as the number of attributes grows.
Abstract summary
The paper proposes a maximum-entropy model that matches the specified frequencies in expectation. Its theoretical formulation is a convex optimization problem over an exponential-family distribution. The practical implementation uses fixed-budget MCMC moment matching to approximate that solution, and the reported results evaluate this finite-budget procedure.
Evaluation
Experiments use NPORS-derived benchmarks with four to 40 attributes and an independently generated benchmark family. Comparisons with generalized raking identify gains for larger problems with higher-order interactions, while raking remains competitive on smaller problems.
Project
Synthetic Populations project — research context, the paper and PDFs, and a link to ImagineAllThePeople.