Summary Sufficiency in Simulation-Based Inference via Mutual Information

Simulation-based inference (SBI) enables posterior estimation when the likelihood is intractable but data can be simulated from the model. For high-dimensional or variable-length data, SBI methods typically rely on low-dimensional summary statistics as conditioning inputs for posterior estimation, making their sufficiency crucial for accurate inference. Mutual information (MI) provides a natural information-theoretic characterization of sufficiency, quantifying how much parameter information a summary retains. However, these MI quantities are generally intractable in likelihood-free settings. Sui et al. (2026) estimate MI from a trained neural posterior using the Barber--Agakov (BA) lower bound, while discussing, but not computing, a Fisher-information upper bound. This makes the estimate ambiguous: a small lower bound may indicate either an insufficient summary or an inaccurate density estimator. We address this limitation with a tractable diagnostic that estimates the BA gap and reveals what the difference between bounds has actually measured. We extend the resulting sufficiency metric to observation-centered restricted data spaces, allowing simulations to focus on a neighborhood of the data while preserving the target posterior and enabling valid sufficiency comparisons between summaries. We illustrate the use of this diagnostic tool on a temporal epidemic-type aftershock sequence model by comparing learned and hand-crafted summary statistics for variable-length earthquake catalogs.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-16
DOI
https://doi.org/10.48550/arxiv.2611.21158
Primary Topic
Markov Chains and Monte Carlo Methods
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Summary Sufficiency in Simulation-Based Inference via Mutual Information

Christian Genest, David A. Stephens, Renjie Peng
arXiv (Cornell University)
Markov Chains and Monte Carlo Methods
preprint

Summary Sufficiency in Simulation-Based Inference via Mutual Information

Christian Genest, David A. Stephens, Renjie Peng
preprint en

Abstract

Simulation-based inference (SBI) enables posterior estimation when the likelihood is intractable but data can be simulated from the model. For high-dimensional or variable-length data, SBI methods typically rely on low-dimensional summary statistics as conditioning inputs for posterior estimation, making their sufficiency crucial for accurate inference. Mutual information (MI) provides a natural information-theoretic characterization of sufficiency, quantifying how much parameter information a summary retains. However, these MI quantities are generally intractable in likelihood-free settings. Sui et al. (2026) estimate MI from a trained neural posterior using the Barber--Agakov (BA) lower bound, while discussing, but not computing, a Fisher-information upper bound. This makes the estimate ambiguous: a small lower bound may indicate either an insufficient summary or an inaccurate density estimator. We address this limitation with a tractable diagnostic that estimates the BA gap and reveals what the difference between bounds has actually measured. We extend the resulting sufficiency metric to observation-centered restricted data spaces, allowing simulations to focus on a neighborhood of the data while preserving the target posterior and enabling valid sufficiency comparisons between summaries. We illustrate the use of this diagnostic tool on a temporal epidemic-type aftershock sequence model by comparing learned and hand-crafted summary statistics for variable-length earthquake catalogs.

arXiv (Cornell University)
Markov Chains and Monte Carlo Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon