pith. sign in

arxiv: 2606.00274 · v2 · pith:MHFQYQ5Wnew · submitted 2026-05-29 · 🧮 math.NA · cs.NA

Error bounds for approximate posteriors from likelihood-informed reduced-order models

Han Cheng Lie , Jakob Scheffels , Elisabeth Ullmann This is my paper

Pith reviewed 2026-06-28 21:05 UTC · model grok-4.3

classification 🧮 math.NA cs.NA
keywords Bayesian inverse problemsreduced-order modelslikelihood-informed subspaceserror boundsposterior approximationPetrov-Galerkin projectionlinear Gaussian models
0
0 comments X

The pith

Reduced-order models recover exact posteriors when rank matches the intrinsic dimension of the inverse problem

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

The paper derives error bounds for approximate posteriors obtained from reduced-order forward models that project parameters onto likelihood-informed subspaces. These subspaces are chosen to capture the directions where the prior-to-posterior update occurs in Bayesian inverse problems. For linear Gaussian settings the analysis shows that the resulting approximate posterior covariance and mean become exact once the reduced-order rank equals the rank of the prior-preconditioned Hessian. The work supplies explicit bounds on the error in the root prior-preconditioned Hessian of the data misfit and demonstrates the bounds on two structural-engineering examples.

Core claim

In linear Gaussian inverse problems with possibly singular prior covariance, a reduced-order model that employs a Petrov-Galerkin projection onto likelihood-informed subspaces bounds the error incurred in approximating the root prior-preconditioned Hessian of the data misfit; the same bounds control the errors in the approximate posterior covariance and mean. The model recovers the exact posterior whenever its rank equals the intrinsic dimension, defined as the rank of the prior-preconditioned Hessian.

What carries the argument

Petrov-Galerkin projection onto likelihood-informed subspaces obtained from optimal low-rank approximations of the posterior covariance matrix

If this is right

  • Error bounds hold for the approximation of the root prior-preconditioned Hessian of the data misfit.
  • Error bounds hold for the approximate posterior covariance and mean.
  • Exact posterior recovery occurs precisely when the reduced-order rank equals the rank of the prior-preconditioned Hessian.
  • The bounds are illustrated numerically on structural-engineering inverse problems.

Where Pith is reading between the lines

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

  • The recovery result suggests a practical stopping criterion for choosing the rank in expensive sampling or optimization tasks.
  • The same subspace construction could be tested as a dimension-reduction device in related uncertainty-quantification settings that are not strictly linear.
  • The error bounds may be combined with existing sampling algorithms to certify the accuracy of low-rank posterior approximations.

Load-bearing premise

The analysis assumes the inverse problem is linear and Gaussian, possibly with a singular prior covariance matrix.

What would settle it

A linear Gaussian inverse problem in which the reduced-order posterior fails to match the exact posterior once the reduced-order rank equals the rank of the prior-preconditioned Hessian would falsify the recovery result.

Figures

Figures reproduced from arXiv: 2606.00274 by Elisabeth Ullmann, Han Cheng Lie, Jakob Scheffels.

Figure 1
Figure 1. Figure 1: Comparison of S −1 obs(G − Gb (r))Spr, for Gb = Gb OLR from (2.9) and Gb = Gb ROM from (2.14), for the cantilever bar problem in Section 4.1. (a): Singular values (δi)i of S −1 obsGSpr from (2.4). (b): Errors ∥S −1 obs(G − Gb (r))Spr∥∞ for Gb = Gb OLR and Gb = Gb ROM, the OLR bound t(∞, r) from (2.11), and the ROM bound b(∞, r) from (3.5). (c): Errors ∥µpos(y) − µbpos(y)∥2 resulting from using Gb OLR and G… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of S −1 obs(G − Gb (r))Spr, for Gb = Gb OLR from (2.9) and Gb = Gb ROM from (2.14), for the plate problem in Section 4.2. (a): Singular values (δi)i of S −1 obsGSpr from (2.4). (b): Errors ∥S −1 obs(G − Gb (r))Spr∥∞ for Gb = Gb OLR and Gb = Gb ROM, the OLR bound t(∞, r) from (2.11), and the ROM bound b(∞, r) from (3.5). (c): Errors ∥µpos(y) − µbpos(y)∥2 resulting from using Gb OLR and Gb ROM in … view at source ↗
read the original abstract

In the design of computational methods for Bayesian inverse problems, costly forward model evaluations make it difficult to sample from or compute the posterior. This motivates the need for approximate forward models that are cheaper to evaluate. We consider reduced-order forward models which exploit the lower-dimensional structure in the Bayesian inverse problem by projecting to the "likelihood-informed subspace" of the parameter space where the prior-to-posterior update is significant. However, the theoretical properties of these reduced-order forward models and their impact on the solution of the Baysian inverse problem are not always well-understood. In this work we consider linear Gaussian inverse problems with a possibly singular prior covariance matrix. We analyse a recently proposed reduced-order model which uses a Petrov-Galerkin projection to likelihood-informed subspaces that arise in optimal low-rank approximations of the posterior covariance matrix. We bound the error in the resulting approximation of the root prior-preconditioned Hessian of the data misfit. Based on this we also bound the errors of the approximate posterior covariance and mean. Our analysis shows that this reduced-order model recovers the exact posterior when the rank of the reduced-order model is equal to the "intrinsic dimension" of the inverse problem, i.e. the rank of the prior-preconditioned Hessian. Two numerical experiments from structural engineering illustrate the performance of our bounds.

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

0 major / 2 minor

Summary. The paper analyzes error bounds for a Petrov-Galerkin reduced-order model (ROM) applied to linear Gaussian Bayesian inverse problems, including cases with singular prior covariances. The ROM projects onto likelihood-informed subspaces derived from optimal low-rank approximations of the posterior covariance. The authors bound the error in the root prior-preconditioned Hessian of the data misfit, and from this derive bounds on the approximate posterior covariance and mean. They prove that the ROM recovers the exact posterior precisely when its rank equals the intrinsic dimension (rank of the prior-preconditioned Hessian). Two numerical experiments from structural engineering are used to illustrate the bounds.

Significance. If the derivations hold, the work supplies explicit, vanishing error bounds for likelihood-informed ROMs in the linear-Gaussian setting, together with a clean characterization of exact recovery at the intrinsic dimension. This is useful for guiding reduced-rank selection in computational inverse problems. The explicit treatment of singular priors and the direct link between the subspace rank and the prior-preconditioned Hessian rank are strengths of the analysis.

minor comments (2)
  1. §2.2: the notation for the prior-preconditioned Hessian operator H and its low-rank factors could be introduced with an explicit reference to the eigenvalue decomposition used later in the bounds.
  2. The numerical experiments section would benefit from a short table summarizing the observed versus predicted error decay rates for the covariance and mean as a function of reduced rank.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of the manuscript, accurate summary of its contributions, and recommendation to accept. We are pleased that the explicit error bounds, treatment of singular priors, and exact recovery result at the intrinsic dimension were viewed as strengths.

Circularity Check

0 steps flagged

Derivation is self-contained mathematical analysis with no circularity

full rationale

The paper derives error bounds on the root prior-preconditioned Hessian, posterior covariance, and mean for a Petrov-Galerkin reduced-order model in linear Gaussian inverse problems (including singular priors). The central observation that the approximate posterior recovers the exact posterior when the reduced rank equals the rank of the prior-preconditioned Hessian follows directly from the definition of the likelihood-informed subspace as the range of that operator; once the subspace is full-rank, the projection is the identity on the relevant space and the bounds vanish by standard linear algebra. No fitted parameters are renamed as predictions, no self-citations are invoked as load-bearing uniqueness theorems, and the analysis remains independent of any external fitted quantities or prior self-referential results. This is the expected outcome for a rigorous a-priori error analysis in numerical analysis.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, preventing identification of specific free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5762 in / 1100 out tokens · 25519 ms · 2026-06-28T21:05:13.404828+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

49 extracted references · 35 canonical work pages · 2 internal anchors

  1. [1]

    Aziznejad, Shayan and Unser, Michael , title =. Numer. Funct. Anal. Optim. , volume =. 2021 , doi =

    2021

  2. [2]

    Case of equality in spectral norm matrix triangle inequality , NOTE =

  3. [3]

    Likelihood-informed Model Reduction for Bayesian Inference of Static Structural Loads

    Scheffels, Jakob and Qian, Elizabeth and Papaioannou, Iason and Ullmann, Elisabeth , year =. doi:10.48550/arXiv.2510.07950 , note =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.07950

  4. [4]

    and Lie, H

    Carere, G. and Lie, H. C. , year =. Optimal low-rank posterior covariance approximation in linear. doi:10.48550/arXiv.2411.01112 , note =

    work page doi:10.48550/arxiv.2411.01112

  5. [5]

    Inverse Probl

    Giuseppe Carere and Han Cheng Lie , fjournal =. Inverse Probl. Imag. , pages =. 2026 , doi =

    2026

  6. [6]

    Flath, H. P. and Wilcox, L. C. and Akcelik, V. and Hill, J. and Van Bloemen Waanders, B. and Ghattas, O. , title =. SIAM J. Sci. Comput. , volume =. 2011 , doi =

    2011

  7. [7]

    Spantini, Alessio and Solonen, Antti and Cui, Tiangang and Martin, James and Tenorio, Luis and Marzouk, Youssef , title =. SIAM J. Sci. Comput. , volume =. 2015 , doi =

    2015

  8. [8]

    Henderson, H. V. and Searle, S. R. , title =. SIAM Review , volume =. 1981 , doi =

    1981

  9. [9]

    , year =

    Eiermann, Michael and Ernst, Oliver G. , year =. Geometric aspects of the theory of. doi:10.1017/S0962492901000046 , journal =

    work page doi:10.1017/s0962492901000046

  10. [10]

    2025 , doi =

    Posterior error bounds for prior-driven balancing , author =. 2025 , doi =

    2025

  11. [11]

    Choosing Observation Operators to Mitigate Model Error in

    Cvetkovi\'. Choosing Observation Operators to Mitigate Model Error in. SIAM/ASA J. Uncertain. Quantif. , volume =. 2024 , doi =

    2024

  12. [12]

    PEM , volume=

    The stochastic finite element method in structural reliability , author=. PEM , volume=. doi:10.1016/0266-8920(88)90019-7 , year=

    work page doi:10.1016/0266-8920(88)90019-7

  13. [13]

    Hsein and Zhang, Dongming and Wang, Lei , month = mar, year =

    Huang, Hongwei and Gong, Wenping and Khoshnevisan, Sara and Juang, C. Hsein and Zhang, Dongming and Wang, Lei , month = mar, year =. Simplified procedure for finite element analysis of the longitudinal performance of shield tunnels considering spatial soil variability in longitudinal direction , volume =. doi:10.1016/j.compgeo.2014.11.010 , journal =

    work page doi:10.1016/j.compgeo.2014.11.010 2014

  14. [14]

    Vittorio Erba, Emanuele Troiani, Lenka Zdeborová, and Florent Krzakala

    The. Psychometrika , author =. 1936 , pages =. doi:10.1007/BF02288367 , number =

    work page doi:10.1007/bf02288367 1936

  15. [15]

    Dimension and model reduction approaches for linear Bayesian inverse problems with rank-deficient prior covariances

    König, Josie and Qian, Elizabeth and Freitag, Melina A. , year =. Dimension and model reduction approaches for linear. doi:10.48550/arXiv.2506.23892 , publisher =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.23892

  16. [16]

    twiddle” operation and bounds on derivatives of pro- jectors A.1 The “twiddle

    Bhatia, Rajendra , TITLE =. 1997 , PAGES =. doi:10.1007/978-1-4612-0653-8 , URL =

    work page doi:10.1007/978-1-4612-0653-8 1997

  17. [17]

    Stuart, A. M. , year =. Inverse problems:. doi:10.1017/S0962492910000061 , journal =

    work page doi:10.1017/s0962492910000061

  18. [18]

    On the local

    Sprungk, Björn , year =. On the local. doi:10.1088/1361-6420/ab6f43 , journal =

    work page doi:10.1088/1361-6420/ab6f43

  19. [19]

    Adaptive multi-fidelity polynomial chaos approach to

    Yan, Liang and Zhou, Tao , year =. Adaptive multi-fidelity polynomial chaos approach to. doi:10.1016/j.jcp.2018.12.025 , journal =

    work page doi:10.1016/j.jcp.2018.12.025 2018

  20. [20]

    , year =

    Novák, Lukáš and Sharma, Himanshu and Shields, Michael D. , year =. Physics-informed polynomial chaos expansions , volume =. doi:10.1016/j.jcp.2024.112926 , journal =

    work page doi:10.1016/j.jcp.2024.112926 2024

  21. [21]

    , year =

    Lu, Fei and Morzfeld, Matthias and Tu, Xuemin and Chorin, Alexandre J. , year =. Limitations of polynomial chaos expansions in the. doi:10.1016/j.jcp.2014.11.010 , journal =

    work page doi:10.1016/j.jcp.2014.11.010 2014

  22. [22]

    Dinkel, Maximilian and Geitner, Carolin M and Rei, Gil Robalo and Nitzler, Jonas and Wall, Wolfgang A , journal=. Solving. 2024 , publisher=

    2024

  23. [23]

    Adaptive

    Villani, Paolo and Unger, Jörg and Weiser, Martin , year =. Adaptive. doi:10.48550/arXiv.2404.19459 , publisher =

    work page doi:10.48550/arxiv.2404.19459

  24. [24]

    Deep surrogate accelerated delayed-acceptance

    Deveney, Teo and Mueller, Eike and Shardlow, Tony , year =. Deep surrogate accelerated delayed-acceptance. doi:https://doi.org/10.1137/22M1513113 , journal =

    work page doi:10.1137/22m1513113

  25. [25]

    An adaptive surrogate modeling based on deep neural networks for large-scale

    Yan, Liang , year =. An adaptive surrogate modeling based on deep neural networks for large-scale. doi:10.4208/cicp.OA-2020-0186 , journal =

    work page doi:10.4208/cicp.oa-2020-0186 2020

  26. [26]

    and Graham-Brady, Lori and Shields, Michael D

    Pasparakis, George D. and Graham-Brady, Lori and Shields, Michael D. , year =. Bayesian neural networks for predicting uncertainty in full-field material response , volume =. doi:10.1016/j.cma.2024.117486 , journal =

    work page doi:10.1016/j.cma.2024.117486 2024

  27. [27]

    Physics-informed

    Pf. Physics-informed. arXiv preprint arXiv:2212.12474 , year=. doi:10.48550/arXiv.2212.12474 , note =

    work page doi:10.48550/arxiv.2212.12474

  28. [28]

    Antoulas , title =

    Athanasios C. Antoulas , title =. 2005 , publisher =

    2005

  29. [29]

    SIAM review , volume=

    A survey of projection-based model reduction methods for parametric dynamical systems , author=. SIAM review , volume=. 2015 , doi=

    2015

  30. [30]

    , booktitle=

    Bui-Thanh, Tan and Burstedde, Carsten and Ghattas, Omar and Martin, James and Stadler, Georg and Wilcox, Lucas C. , booktitle=. Extreme-scale. 2012 , volume=

    2012

  31. [31]

    A. J. Sci. Comput. , author =. 2013 , pages =. doi:10.1137/12089586X , number =

    work page doi:10.1137/12089586x 2013

  32. [32]

    Inverse Problems , author =

    Likelihood-informed dimension reduction for nonlinear inverse problems , volume =. Inverse Problems , author =. 2014 , pages =. doi:10.1088/0266-5611/30/11/114015 , number =

    work page doi:10.1088/0266-5611/30/11/114015 2014

  33. [33]

    Certified dimension reduction in nonlinear

    Zahm, Olivier and Cui, Tiangang and Law, Kody and Spantini, Alessio and Marzouk, Youssef , journal=. Certified dimension reduction in nonlinear. doi:10.1090/mcom/3737 , year=

    work page doi:10.1090/mcom/3737

  34. [34]

    Model. J. Sci. Comput. , author =. 2022 , pages =. doi:10.1007/s10915-022-01798-8 , number =

    work page doi:10.1007/s10915-022-01798-8 2022

  35. [35]

    Time-. J. Sci. Comput. , author =. 2023 , pages =. doi:10.1007/s10915-023-02358-4 , number =

    work page doi:10.1007/s10915-023-02358-4 2023

  36. [36]

    An ensemble

    Stavrinides, Pavlos and Qian, Elizabeth , year =. An ensemble. doi:10.48550/arXiv.2507.03207 , publisher =

    work page doi:10.48550/arxiv.2507.03207

  37. [37]

    PAMM , author =

    Inference-. PAMM , author =. 2024 , pages =. doi:10.1002/pamm.202400051 , number =

    work page doi:10.1002/pamm.202400051 2024

  38. [38]

    Lumley, J. L. , month = jan, year =. Coherent. Transition and. doi:10.1016/B978-0-12-493240-1.50017-X , pages =

    work page doi:10.1016/b978-0-12-493240-1.50017-x

  39. [39]

    Turbulence and the dynamics of coherent structures

    Sirovich, Lawrence , journal=. Turbulence and the dynamics of coherent structures. doi:10.1090/qam/910462 , year=

    work page doi:10.1090/qam/910462

  40. [40]

    Learning physics-based models from data: perspectives from inverse problems and model reduction , volume =

    Ghattas, Omar and Willcox, Karen , year =. Learning physics-based models from data: perspectives from inverse problems and model reduction , volume =. doi:10.1017/S0962492921000064 , journal =

    work page doi:10.1017/s0962492921000064

  41. [41]

    Asia Pac

    Model order reduction for. Asia Pac. J. Comput. Eng. , author =. 2014 , pages =. doi:10.1186/2196-1166-1-2 , number =

    work page doi:10.1186/2196-1166-1-2 2014

  42. [42]

    and Willcox, Karen , title =

    Cui, Tiangang and Marzouk, Youssef M. and Willcox, Karen , title =. Int. J. Numer. Meth. Engng , volume =. 2015 , doi =

    2015

  43. [43]

    Electron

    Accelerating the. Electron. Res. Arch. , author =. 2021 , pages =. doi:10.3934/era.2021044 , number =

    work page doi:10.3934/era.2021044 2021

  44. [44]

    , year =

    Rao, Preetham P. , year =. Inverse parameter estimation using compressed sensing and. doi:10.1016/j.cma.2024.116820 , journal =

    work page doi:10.1016/j.cma.2024.116820 2024

  45. [45]

    Acta Numerica , volume=

    Reduced basis methods for time-dependent problems , author=. Acta Numerica , volume=. 2022 , doi=

    2022

  46. [46]

    SISSA springer series, Springer Cham , year=

    Real time reduced order computational mechanics , author=. SISSA springer series, Springer Cham , year=. doi:10.1007/978-3-031-49892-3 , publisher=

    work page doi:10.1007/978-3-031-49892-3

  47. [47]

    Stein. J. Sci. Comput. , author =. 2021 , pages =. doi:10.1137/20M1321589 , number =

    work page doi:10.1137/20m1321589 2021

  48. [48]

    A reduced basis ensemble. GEM Int. J. Geomath. , author =. 2023 , pages =. doi:10.1007/s13137-023-00235-8 , number =

    work page doi:10.1007/s13137-023-00235-8 2023

  49. [49]

    Cotter, S. L. and Dashti, M. and Stuart, A. M. , year =. Approximation of. doi:10.1137/090770734 , journal =

    work page doi:10.1137/090770734