Theoretical and methodological challenges in hierarchical Bayesian inference for model-form uncertainty
Portone, Teresa, White, Rebekah Dale · 2025
Read Theoretical and methodological challenges in hierarchical Bayesian inference for model-form uncertainty by Portone, Teresa, White, Rebekah Dale (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/doe-osti-2587508-theoretical-and-methodological-challenge.txt.
Introduction
Teresa Portone and Rebekah White wrote this report, SAND2025-10780, printed in September 2025 at Sandia National Laboratories. It examines the theory and practice of hierarchical Bayesian inference. The goal is to capture model-form uncertainty: the error that comes from the model itself being wrong, not just from noisy data. Such methods stack layers of probability distributions, with hyperparameters governing the spread of the discrepancy terms.
The authors work through where the machinery strains. One thread asks when data can actually inform a scale parameter that sets the size of model error. Their numerical exploration suggests it may take near-degenerate cases for the data to speak. Two demonstration problems carry the argument: a spatially varying linear model, and subsurface contaminant transport. Calibration plots compare the method's predictive quantiles against held-out data. The report is aimed at analysts who build uncertainty estimates for high-stakes simulations.
Catalog record
| CREATOR | Portone, Teresa, White, Rebekah Dale |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| LICENSE | Public Domain |
| RIGHTS | PUBLIC DOMAIN |
| SOURCE | DOE OSTI |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
SAND2025-10780 SANDIA REPORT SAND2025-10780 Printed September 2025 Theoretical and methodological challenges in hierarchical Bayesian inference for model-form uncertainty Teresa Portone & Rebekah D. White Prepared by Sandia National Laboratories Albuquerque, New Mexico 87185 Livermore, California 94550 DE-NA0003525. 5 6 CONTENTS Acknowledgement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2. Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3. Theoretical and methodological challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 4. Numerical demonstrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 4.1. Spatially varying linear model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 4.2. Subsurface contaminant transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 5. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 7 8 LIST OF FIGURES Figure 2-1. An illustration of different predictive uncertainty scenarios where data is non- linear but the model is linear. On the left, MFU is not addressed, resulting in an overly-concentrated predictive distribution that assigns low probability to many of the data points. In the middle, the data is probable, but the uncertainty bounds are so great that they couldn’t meaningfully inform a decision. On the right, the data is probable according to the predictive distribution, but the bounds may be tight enough to inform a decision. Note that an “adequately disperse” predictive distribution is necessary but not sufficient to determine whether an uncertainty characterization is meaningful for extrapolative prediction. . . . . . . . . . . . . . . . . . . 15 Figure 4-1. Left: Marginal prior and posterior densities for 𝜇 and 𝜎. Right: 95% high- probability region for prior and posterior predictive distributions based on 0.025 and 0.975 quantiles, compared to calibration data (black). . . . . . . . . . . . . . . . . . . . 20 Figure 4-2. Left: Marginal prior and posterior densities for 𝜇 and 𝜎. Right: 95% high- probability region for prior and posterior predictive distributions based on 0.025 and 0.975 quantiles, compared to calibration data (black). . . . . . . . . . . . . . . . . . . . 20 Figure 4-3. Top left: Marginal prior densities for 𝜃 under different hyperprior specifications, which are differentiated by shading and solid vs. dashed lines. Top right and bot- tom: 95% high-probability region for prior and posterior predictive distributions under different hyperprior specifications. The prior predictive distributions are nearly identical, making them difficult to distinguish. . . . . . . . . . . . . . . . . . . . . . . . 21 Figure 4-4. A comparison of prior and posterior marginal densities for 𝜇 and 𝜎 (left, center) and predictive densities (right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 Figure 4-5. Marginal prior and posterior densities for (top) hyperparameters and (bottom left) outflow concentration QoI. Bottom right: 95% high-probability region for prior and posterior predictive distributions based on 0.025 and 0.975 quantiles compared to calibration data (black). Data was produced using the fractional derivative to represent dispersion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 Figure 4-6. Calibration data predictive distributions (top) and outflow concentration QoI predictive distributions (bottom). Limited data (left) is gathered to time 0.3 with 10 data points. Copious data (right) is gathered to time 0.9 with 60 data points (a doubling in time resolution). Increased data did not substantively modify the predictive distributions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 9 Figure 4-7. Marginal prior and posterior densities for (top) hyperparameters and (bottom left) outflow concentration QoI. Bottom right: 95% high-probability region for prior and posterior predictive distributions based on 0.025 and 0.975 quantiles compared to calibration data (black). Data was produced using the general linear operator to represent dispersion. . . . . . . . . . . . . . . . . . . . . . .