pith. sign in

arxiv: 2604.18505 · v1 · submitted 2026-04-20 · 💻 cs.IT · math.IT· stat.ML

Bayesian experimental design: grouped geometric pooled posterior via ensemble Kalman methods

Huchen Yang , Xinghao Dong , Jinlong Wu This is my paper

Pith reviewed 2026-05-10 03:19 UTC · model grok-4.3

classification 💻 cs.IT math.ITstat.ML
keywords Bayesian experimental designexpected information gainensemble Kalman inversionpooled posteriorgrouped samplingimportance samplingamortized inference
0
0 comments X p. Extension

The pith

Grouping outer samples into group-specific pooled posteriors via tailored ensemble Kalman inversion yields more accurate and stable expected information gain estimators at amortized cost.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Bayesian experimental design requires estimating expected information gain through nested inference, but each outer sample produces a distinct posterior, forcing a choice between expensive per-sample inference or less accurate fully amortized inference across all samples. The paper proposes partitioning the outer samples into groups, constructing a geometric pooled proposal distribution for each group, and applying a tailored ensemble Kalman inversion formulation that reuses existing computations to generate the group-specific samples without additional forward model evaluations. A conservative diagnostic assesses the importance-sampling quality within each group to decide on the grouping. This alignment improvement produces more accurate and stable estimators while keeping total cost at the amortized level, with demonstrations on both linear Gaussian and high-dimensional network calibration problems.

Core claim

We propose a grouped geometric pooled posterior framework that partitions outer samples into groups and constructs a pooled proposal for each group. While such grouping strategy would normally require generating separate proposal samples for different groups, our tailored ensemble Kalman inversion (EKI) formulation generates these samples without extra forward-model evaluation cost. We also introduce a conservative diagnostic to assess importance-sampling quality to guide grouping. This grouping strategy improves within-group proposal-target alignment, yielding more accurate and stable estimators while keeping the cost comparable to amortized approaches.

What carries the argument

The grouped geometric pooled posterior framework, realized through a tailored ensemble Kalman inversion formulation that produces group-specific proposal samples by reusing forward-model evaluations already performed for the outer loop.

If this is right

  • Expected information gain estimators become more accurate and stable than those obtained from fully amortized inference across all outer samples.
  • Total computational cost, measured in forward-model evaluations, remains comparable to standard amortized approaches.
  • Within-group proposal-target alignment improves, reducing the degradation that occurs when a single proposal is forced to cover highly heterogeneous posteriors.
  • The framework applies directly to both simple Gaussian-linear settings and complex high-dimensional model discrepancy calibration tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the diagnostic and zero-cost generation hold across problems, the method could make Bayesian experimental design feasible for a wider range of physical systems where full nested sampling has been computationally prohibitive.
  • The reuse pattern in the ensemble Kalman formulation suggests similar grouping tactics could improve efficiency in other ensemble-based inverse problems that face heterogeneous targets.
  • Adaptive or data-driven grouping rules derived from the same diagnostic might further tighten the accuracy-cost tradeoff beyond the fixed partitioning explored here.

Load-bearing premise

The tailored ensemble Kalman inversion must generate the necessary group-specific proposal samples at zero extra forward-model cost, and the conservative diagnostic must correctly identify groupings where the improved alignment actually raises estimator quality above the fully amortized baseline.

What would settle it

A controlled run on the high-dimensional network-based calibration problem in which the grouped method produces EIG estimator variance no lower than the amortized baseline, or requires additional forward-model calls to maintain the same accuracy, would falsify the central efficiency-accuracy claim.

Figures

Figures reproduced from arXiv: 2604.18505 by Huchen Yang, Jinlong Wu, Xinghao Dong.

Figure 1
Figure 1. Figure 1: Posteriors of the physical parameter. From top to bottom, the rows present [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Updating trajectory of the error parameter. [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Scatter plots comparing, for each outer sample, the distance from its individual posterior to the global geometric pooled posterior versus the distance to the prior. The PDE costs of the different methods, reported in [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Posteriors of the physical parameter. From top to bottom, the rows present [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Relative error between the solution field and the true field. From top to [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: illustrates the grouping of outer samples y induced by the proposed diagnostic criterion. We use the conservative diagnostic to predict the importance sampling performance and select groups directly in the observation space. The resulting grouping closely approximates the grouping that would be obtained by actually running importance sampling and then clustering based on the realized weights, but without t… view at source ↗
Figure 8
Figure 8. Figure 8: Scatter plots comparing, for each outer sample, the distance from its individual [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Corner plot of selected coordinates of the parameter vector, comparing individ [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
read the original abstract

Bayesian experimental design (BED) for complex physical systems is often limited by the nested inference required to estimate the expected information gain (EIG) or its gradients. Each outer sample induces a different posterior, creating a large and heterogeneous set of inference targets. Existing methods have to sacrifice either accuracy or efficiency: they either perform per-outer-sample posterior inference, which yields higher fidelity but at prohibitive computational cost, or amortize the inner inference across all outer samples for computational reuse, at the risk of degraded accuracy under posterior heterogeneity. To improve accuracy and maintain cost at the amortized level, we propose a grouped geometric pooled posterior framework that partitions outer samples into groups and constructs a pooled proposal for each group. While such grouping strategy would normally require generating separate proposal samples for different groups, our tailored ensemble Kalman inversion (EKI) formulation generates these samples without extra forward-model evaluation cost. We also introduce a conservative diagnostic to assess importance-sampling quality to guide grouping. This grouping strategy improves within-group proposal-target alignment, yielding more accurate and stable estimators while keeping the cost comparable to amortized approaches. We evaluate the performance of our method on both Gaussian-linear and high-dimensional network-based model discrepancy calibration problems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes a grouped geometric pooled posterior framework for Bayesian experimental design to estimate expected information gain. Outer samples are partitioned into groups, each with a pooled proposal constructed via a tailored ensemble Kalman inversion (EKI) formulation that is claimed to incur no extra forward-model evaluations compared to fully amortized single-proposal baselines. A conservative diagnostic is introduced to assess importance-sampling quality and guide grouping. The approach is asserted to yield more accurate and stable estimators under posterior heterogeneity while preserving amortized-level cost, with supporting evaluations on Gaussian-linear problems and high-dimensional network-based model discrepancy calibration.

Significance. If the no-extra-cost property of the tailored EKI holds without hidden reductions in ensemble size or iterations, and the reported accuracy gains prove robust, the work would meaningfully advance practical BED for complex physical systems by addressing the accuracy-efficiency tradeoff in nested inference. The dual evaluation on linear-Gaussian and high-dimensional network settings provides a reasonable initial validation of the grouping strategy.

major comments (1)
  1. Abstract and Methods: The pivotal claim that the tailored EKI formulation 'generates these samples without extra forward-model evaluation cost' must be substantiated with the explicit update rule or shared-ensemble construction. Standard EKI costs scale with ensemble size times iterations; the manuscript needs to show precisely how group-specific proposals reuse the identical forward evaluations as the single amortized baseline, without implicit trade-offs such as smaller per-group ensembles or fewer iterations, since this directly determines whether accuracy gains can be attributed to improved within-group alignment.
minor comments (2)
  1. Notation section: The term 'geometric pooled posterior' should be defined with an explicit equation contrasting it to standard importance-weighted pooling to avoid ambiguity in the grouping construction.
  2. Results: The high-dimensional network experiments would benefit from reporting the number of groups chosen by the diagnostic and the corresponding within-group alignment metrics to allow readers to assess the practical impact of the conservative diagnostic.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. We address the single major comment below and will revise the manuscript to provide the requested substantiation.

read point-by-point responses

  1. Referee: Abstract and Methods: The pivotal claim that the tailored EKI formulation 'generates these samples without extra forward-model evaluation cost' must be substantiated with the explicit update rule or shared-ensemble construction. Standard EKI costs scale with ensemble size times iterations; the manuscript needs to show precisely how group-specific proposals reuse the identical forward evaluations as the single amortized baseline, without implicit trade-offs such as smaller per-group ensembles or fewer iterations, since this directly determines whether accuracy gains can be attributed to improved within-group alignment.

    Authors: We agree that explicit details on the no-extra-cost property are necessary to support our claims. In the revised manuscript we will add a new subsection in Methods that presents the tailored EKI update rule, including the precise form of the Kalman gain and the geometric pooling operator. The construction reuses a single shared ensemble of forward-model evaluations (identical size and iteration count to the amortized baseline) that is computed once; group-specific proposals are then obtained by post-processing this ensemble with the pooling step, which requires no additional model runs. We will also include a complexity table and pseudocode confirming that forward evaluations, ensemble size, and iteration count are unchanged relative to the single-proposal baseline. This revision will make clear that accuracy improvements arise from improved within-group alignment rather than any hidden computational trade-offs. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper introduces a grouped geometric pooled posterior framework and a tailored ensemble Kalman inversion (EKI) formulation as algorithmic constructions for Bayesian experimental design. The claims center on partitioning outer samples into groups, constructing pooled proposals, and generating samples without extra forward-model evaluations via the tailored EKI. These are presented as novel methodological contributions with a conservative diagnostic for importance-sampling quality, evaluated empirically on Gaussian-linear and network-based problems. No derivation steps reduce by construction to inputs, fitted parameters renamed as predictions, or load-bearing self-citations. The framework is self-contained against external benchmarks and does not rely on self-referential definitions or uniqueness theorems from prior author work.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Based on abstract only: the approach rests on standard Bayesian inference assumptions and the properties of ensemble Kalman inversion; no free parameters, invented entities, or ad-hoc axioms are explicitly introduced in the provided text.

axioms (2)
  • domain assumption Standard assumptions of Bayesian experimental design and importance sampling hold for the outer and inner loops.
    Invoked implicitly when discussing EIG estimation and proposal quality.
  • domain assumption Ensemble Kalman inversion can be formulated to produce group-specific samples without additional forward evaluations.
    Central to the cost claim; appears in the description of the tailored EKI.

pith-pipeline@v0.9.0 · 5510 in / 1390 out tokens · 59320 ms · 2026-05-10T03:19:36.772500+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 15 canonical work pages

  1. [1]

    Alexanderian,Optimal experimental design for infinite-dimensional bayesian inverse problems governed by pdes: A review, Inverse Problems, 37 (2021), p

    A. Alexanderian,Optimal experimental design for infinite-dimensional bayesian inverse problems governed by pdes: A review, Inverse Problems, 37 (2021), p. 043001

    2021

  2. [2]

    Ao and J

    Z. Ao and J. Li,On estimating the gradient of the expected information gain in bayesian experimental design, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 20311–20319

    2024

  3. [3]

    Y. Z. Barlas, S. J. Sloman, and S. Kaski,Robust experimental design via generalised bayesian inference, arXiv preprint arXiv:2511.07671, (2025)

    work page arXiv 2025

  4. [4]

    J. Beck, B. M. Dia, L. F. Espath, Q. Long, and R. Tempone,Fast bayesian experimental design: Laplace-based importance sampling for the expected information gain, Computer Methods in Applied Mechanics and Engineering, 334 (2018), pp. 523–553

    2018

  5. [5]

    Callahan, A

    J. Callahan, A. Chin, J. Pacheco, and T. Catanach,Reverse-annealed sequential monte carlo for efficient bayesian optimal experiment design, in The Thirty-ninth Annual Conference on Neural Information Processing Systems

  6. [6]

    Chaloner, I

    K. Chaloner and I. Verdinelli,Bayesian experimental design: A review, Statistical Science, 10 (1995), pp. 273–304, https://doi.org/10.1214/ss/1177009939, https://projecteuclid.org/ journals/statistical-science/volume-10/issue-3/Bayesian-Experimental-Design-A-Review/ 10.1214/ss/1177009939.full (accessed 2024-06-05). Publisher: Institute of Mathematical Statistics

    work page doi:10.1214/ss/1177009939 1995

  7. [7]

    Ding and Q

    Z. Ding and Q. Li,Ensemble kalman inversion: mean-field limit and convergence analysis, Statistics and Computing, 31 (2021), pp. 1–21

    2021

  8. [8]

    J. Dong, C. Jacobsen, M. Khalloufi, M. Akram, W. Liu, K. Duraisamy, and X. Huan, Variational bayesian optimal experimental design with normalizing flows, Computer Methods in Applied Mechanics and Engineering, 433 (2025), p. 117457. 28H. YANG, X. DONG, JL. WU

    2025

  9. [9]

    X. Dong, H. Yang, and J.-L. Wu,Stochastic and non-local closure modeling for nonlinear dynamical systems via latent score-based generative models, arXiv preprint arXiv:2506.20771, (2025)

    work page arXiv 2025

  10. [10]

    C. C. Drovandi, J. M. McGree, and A. N. Pettitt,A sequential monte carlo algorithm to incorporate model uncertainty in bayesian sequential design, Journal of Computational and Graphical Statistics, 23 (2014), pp. 3–24

    2014

  11. [11]

    M. R. Ebers, K. M. Steele, and J. N. Kutz,Discrepancy modeling framework: Learning missing physics, modeling systematic residuals, and disambiguating between deterministic and random effects, SIAM Journal on Applied Dynamical Systems, 23 (2024), pp. 440–469

    2024

  12. [12]

    Englezou, T

    Y. Englezou, T. W. Waite, and D. C. Woods,Approximate laplace importance sampling for the estimation of expected shannon information gain in high-dimensional bayesian design for nonlinear models, Statistics and Computing, 32 (2022), p. 82

    2022

  13. [13]

    C. Feng,Optimal Bayesian experimental design in the presence of model error, thesis, Massachu- setts Institute of Technology, 2015, https://dspace.mit.edu/handle/1721.1/97790 (accessed 2024-07-09). Accepted: 2015-07-17T19:46:48Z

    2015

  14. [14]

    A layered multiple importance sampling scheme for focused optimal Bayesian experimental design

    C. Feng and Y. M. Marzouk,A layered multiple importance sampling scheme for focused optimal bayesian experimental design, arXiv preprint arXiv:1903.11187, (2019)

    work page Pith review arXiv 1903

  15. [15]

    Foster, D

    A. Foster, D. R. Ivanova, I. Malik, and T. Rainforth,Deep adaptive design: Amortizing sequential Bayesian experimental design, June 2021, https://doi.org/10.48550/arXiv.2103. 02438, http://arxiv.org/abs/2103.02438 (accessed 2024-06-05). arXiv:2103.02438 [cs, stat]

    work page doi:10.48550/arxiv.2103 2021

  16. [16]

    Foster, M

    A. Foster, M. Jankowiak, M. O’Meara, Y. W. Teh, and T. Rainforth,A unified stochastic gradient approach to designing bayesian-optimal experiments, in International Conference on Artificial Intelligence and Statistics, PMLR, 2020, pp. 2959–2969

    2020

  17. [17]

    Go and P

    J. Go and P. Chen,Sequential infinite-dimensional Bayesian optimal experimental design with derivative-informed latent attention neural operator, Journal of Computational Physics, (2025), p. 113976

    2025

  18. [18]

    T. Goda, T. Hironaka, W. Kitade, and A. Foster,Unbiased mlmc stochastic gradient-based optimization of bayesian experimental designs, SIAM Journal on Scientific Computing, 44 (2022), pp. A286–A311

    2022

  19. [19]

    R. B. Grosse, S. Ancha, and D. M. Roy,Measuring the reliability of mcmc inference with bidirectional monte carlo, Advances in Neural Information Processing Systems, 29 (2016)

    2016

  20. [20]

    X. Huan, J. Jagalur, and Y. Marzouk,Optimal experimental design: Formulations and computations, Acta Numerica, 33 (2024), pp. 715–840

    2024

  21. [21]

    Simulation-based optimal Bayesian experimental design for nonlinear systems,

    X. Huan and Y. M. Marzouk,Simulation-based optimal Bayesian experimental design for nonlinear systems, Journal of Computational Physics, 232 (2013), pp. 288–317, https://doi.org/10.1016/j.jcp.2012.08.013, https://www.sciencedirect.com/science/article/ pii/S0021999112004597 (accessed 2024-07-30)

    work page doi:10.1016/j.jcp.2012.08.013 2013

  22. [22]

    Huan and Y

    X. Huan and Y. M. Marzouk,Gradient-based stochastic optimization methods in Bayesian experimental design, International Journal for Uncertainty Quantification, 4 (2014)

    2014

  23. [23]

    Huang, Y

    D. Huang, Y. Guo, L. Acerbi, and S. Kaski,Amortized bayesian experimental design for decision-making, Advances in Neural Information Processing Systems, 37 (2024), pp. 109460– 109486

    2024

  24. [24]

    Iollo, C

    J. Iollo, C. Heinkel´e, P. Alliez, and F. Forbes,Bayesian experimental design via contrastive diffusions, arXiv preprint arXiv:2410.11826, (2024)

    work page arXiv 2024

  25. [25]

    Jones, M

    M. Jones, M. Goldstein, P. Jonathan, and D. Randell,Bayes linear analysis for Bayesian optimal experimental design, Journal of Statistical Planning and Inference, 171 (2016), pp. 115–129, https://doi.org/10.1016/j.jspi.2015.10.011, https://www.sciencedirect.com/ science/article/pii/S0378375815001950 (accessed 2024-07-23)

    work page doi:10.1016/j.jspi.2015.10.011 2016

  26. [26]

    M. C. Kennedy and A. O’Hagan,Bayesian calibration of computer models, Journal of the Royal Statistical Society Series B: Statistical Methodology, 63 (2001), pp. 425–464, https://doi. org/10.1111/1467-9868.00294, https://academic.oup.com/jrsssb/article/63/3/425/7083367 (accessed 2024-07-09)

    work page doi:10.1111/1467-9868.00294 2001

  27. [27]

    N. B. Kovachki and A. M. Stuart,Ensemble Kalman inversion: a derivative-free technique for machine learning tasks, Inverse Problems, 35 (2019), p. 095005, https://doi.org/10.1088/ 1361-6420/ab1c3a, https://dx.doi.org/10.1088/1361-6420/ab1c3a (accessed 2024-10-18). Publisher: IOP Publishing

    work page doi:10.1088/1361-6420/ab1c3a 2019

  28. [28]

    Levine and A

    M. Levine and A. Stuart,A framework for machine learning of model error in dynamical systems, Communications of the American Mathematical Society, 2 (2022), pp. 283–344

    2022

  29. [29]

    D. V. Lindley,On a measure of the information provided by an experiment, The Annals of Math- ematical Statistics, 27 (1956), pp. 986–1005, https://doi.org/10.1214/aoms/1177728069, https://projecteuclid.org/journals/annals-of-mathematical-statistics/volume-27/ issue-4/On-a-Measure-of-the-Information-Provided-by-an-Experiment/10.1214/aoms/ GROUPED POOLED PO...

    work page doi:10.1214/aoms/1177728069 1956

  30. [30]

    Q. Long, M. Scavino, R. Tempone, and S. Wang,Fast estimation of expected information gains for bayesian experimental designs based on laplace approximations, Computer Methods in Applied Mechanics and Engineering, 259 (2013), pp. 24–39

    2013

  31. [31]

    R. M. Neal,Annealed importance sampling, Statistics and computing, 11 (2001), pp. 125–139

    2001

  32. [32]

    J. N. Neuberger, A. Alexanderian, and B. van Bloemen Waanders,Goal oriented optimal design of infinite-dimensional bayesian inverse problems using quadratic approximations, Journal of Scientific Computing, 105 (2025), p. 55

    2025

  33. [33]

    M. L. Parks, E. De Sturler, G. Mackey, D. D. Johnson, and S. Maiti,Recycling krylov subspaces for sequences of linear systems, SIAM Journal on Scientific Computing, 28 (2006), pp. 1651–1674

    2006

  34. [34]

    Rainforth, A

    T. Rainforth, A. Foster, D. R. Ivanova, and F. Bickford Smith,Modern Bayesian experimental design, Statistical Science, 39 (2024), pp. 100–114

    2024

  35. [35]

    E. G. Ryan, C. C. Drovandi, J. M. McGree, and A. N. Pettitt,A review of modern computational algorithms for Bayesian optimal design, International Statistical Review, 84 (2016), pp. 128–154, https://doi.org/10.1111/insr.12107, https://onlinelibrary.wiley.com/ doi/10.1111/insr.12107 (accessed 2024-04-22)

    work page doi:10.1111/insr.12107 2016

  36. [36]

    K. J. Ryan,Estimating expected information gains for experimental designs with application to the random fatigue-limit model, Journal of Computational and Graphical Statistics, 12 (2003), pp. 585–603. [38]D. Sanz-Alonso and Z. Wang,Bayesian update with importance sampling: Required sample size, Entropy, 23 (2020), p. 22

    2003

  37. [37]

    Shen and X

    W. Shen and X. Huan,Bayesian sequential optimal experimental design for nonlinear models using Policy gradient reinforcement learning, Computer Methods in Applied Mechanics and Engineering, 416 (2023), p. 116304, https://doi.org/10.1016/j.cma.2023.116304, http: //arxiv.org/abs/2110.15335 (accessed 2024-04-22). arXiv:2110.15335 [cs, stat]

    work page doi:10.1016/j.cma.2023.116304 2023

  38. [38]

    J.-L. Wu, M. E. Levine, T. Schneider, and A. Stuart,Learning about structural errors in models of complex dynamical systems, Journal of Computational Physics, (2024), p. 113157

    2024

  39. [39]

    K. Wu, P. Chen, and O. Ghattas,A fast and scalable computational framework for large-scale high-dimensional Bayesian optimal experimental design, SIAM/ASA Journal on Uncertainty Quantification, 11 (2023), pp. 235–261

    2023

  40. [40]

    H. Yang, C. Chen, and J.-L. Wu,Active learning of model discrepancy with bayesian experimental design, Computer Methods in Applied Mechanics and Engineering, 446 (2025), p. 118198, https://doi.org/https://doi.org/10.1016/j.cma.2025.118198, https:// www.sciencedirect.com/science/article/pii/S0045782525004700

    work page doi:10.1016/j.cma.2025.118198 2025

  41. [41]

    H. Yang, X. Dong, and J.-L. Wu,Bayesian experimental design for model discrepancy calibra- tion: An auto-differentiable ensemble kalman inversion approach, Journal of Computational Physics, 545 (2026), p. 114469, https://doi.org/https://doi.org/10.1016/j.jcp.2025.114469, https://www.sciencedirect.com/science/article/pii/S002199912500751X

    work page doi:10.1016/j.jcp.2025.114469 2026