pith. sign in

arxiv: 2605.23571 · v1 · pith:N5JBHXKZnew · submitted 2026-05-22 · 🧮 math.NA · cs.NA

Accelerating an ensemble of variational data assimilations with randomized preconditioning

I. Dau\v{z}ickait\.e , S. G\"urol , M. Destouches , L. Berre , A. T. Weaver This is my paper

Pith reviewed 2026-05-25 03:33 UTC · model grok-4.3

classification 🧮 math.NA cs.NA
keywords variational data assimilationensemble methodsrandomized algorithmspreconditioningsketching matrixHessian eigenpairsiterative linear solvers
0
0 comments X

The pith

A sketching matrix built from right-hand side differences in EDA yields a preconditioner that accelerates solves for all ensemble members.

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

The paper aims to show that the special structure of ensembles of variational data assimilations can be used to create an effective random sketching matrix for low-rank Hessian approximation. This matrix is made from the differences between the right-hand sides of the linear systems solved for each ensemble member. A sympathetic reader would care because EDA involves many similar large linear systems whose solution dominates the cost, so a single good preconditioner that works for the whole ensemble could cut computation time substantially. The approach is tested in idealised experiments with the Lorenz-96 model.

Core claim

The structure of the EDA can be exploited to construct a suitable sketching matrix using the differences of the right-hand sides of the linear systems of equations, resulting in a preconditioner able to accelerate the EDA solution process for all ensemble members, even if constructed from the control member only.

What carries the argument

The randomized sketching matrix constructed from differences of the right-hand sides, which is used to obtain a low-rank representation of the Hessian for computing its leading eigenpairs and building the limited memory preconditioner.

If this is right

  • The preconditioner accelerates convergence of the iterative solver for every member of the ensemble.
  • It remains effective even when the sketching matrix is built using only the control member.
  • This reduces the overall cost of running the full EDA by avoiding the need to build separate preconditioners for each member.

Where Pith is reading between the lines

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

  • If the RHS differences capture the main variability, similar sketching could be applied in other ensemble Kalman filter variants.
  • Real-world tests with full-scale weather models would be needed to confirm the speedups seen in Lorenz-96.
  • The method might combine with other randomized techniques for even lower-rank approximations.

Load-bearing premise

The differences between the right-hand sides of the ensemble systems are rich enough to produce a sketching matrix whose range captures the important directions of the Hessian.

What would settle it

Numerical tests on the Lorenz-96 model or similar where the iteration count for ensemble members does not decrease when using this preconditioner compared to no preconditioner or random sketching without structure.

read the original abstract

Ensembles of variational data assimilations (EDA) require solving systems of linear equations with iterative methods. The solution process can be accelerated using a limited memory preconditioner constructed with approximations of the leading eigenpairs of the Hessian matrix. Randomized methods for low-rank matrix approximations provide a feasible approach for computing these eigenpairs. These methods use a random sketching matrix to obtain a low-rank representation of the Hessian matrix, which is then used for computing the eigendecomposition. The sketching matrix highly influences the quality of the approximation. In this paper, we show how the structure of the EDA can be exploited to construct a suitable sketching matrix, i.e., using the differences of the right-hand sides of the linear systems of equations. Idealised numerical experiments with the Lorenz-96 model show that the resulting preconditioner is able to accelerate the EDA solution process for all ensemble members, even if constructed from the control member only.

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

2 major / 2 minor

Summary. The paper proposes exploiting the structure of ensemble variational data assimilation (EDA) to construct a sketching matrix for randomized low-rank Hessian approximation from differences in the right-hand sides of the linear systems. This yields a limited-memory preconditioner that accelerates iterative solvers for all ensemble members, including when built solely from the control member. The approach is demonstrated through idealized numerical experiments with the Lorenz-96 model.

Significance. If the empirical acceleration holds under more rigorous statistical scrutiny, the method offers a low-overhead way to improve EDA efficiency by reusing existing ensemble information rather than requiring new computations or larger ensembles. The construction avoids introducing free parameters and directly leverages EDA structure, which is a strength for practical adoption in data assimilation workflows.

major comments (2)
  1. [§5] §5 (numerical experiments): The reported positive results on acceleration lack specification of ensemble size, number of independent trials, error bars on iteration counts or wall-clock times, and any statistical significance tests. This prevents assessment of whether the observed speedups are robust or could be due to variability in the Lorenz-96 setup.
  2. [§4] The central claim that the preconditioner works for all members even when constructed from the control only rests on the quality of the RHS-difference sketching matrix; however, no quantitative comparison (e.g., approximation error norms or eigenvalue capture rates) is provided between this sketching choice and standard random sketching to isolate the contribution of the EDA-specific construction.
minor comments (2)
  1. [Abstract] Abstract and §1: The description of the sketching matrix construction would benefit from an explicit equation or pseudocode early in the paper to clarify how RHS differences are assembled into the matrix.
  2. [§5] Figure captions in the experimental section should include the precise values of ensemble size, observation density, and iteration tolerance used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive recommendation of minor revision and the constructive comments. We address each major comment below and will revise the manuscript accordingly to strengthen the presentation of the numerical results.

read point-by-point responses

  1. Referee: [§5] §5 (numerical experiments): The reported positive results on acceleration lack specification of ensemble size, number of independent trials, error bars on iteration counts or wall-clock times, and any statistical significance tests. This prevents assessment of whether the observed speedups are robust or could be due to variability in the Lorenz-96 setup.

    Authors: We agree that the current manuscript lacks sufficient specification of the experimental protocol. In the revised version we will explicitly report the ensemble size, the number of independent trials, error bars on iteration counts and wall-clock times, and the results of statistical significance tests (e.g., paired t-tests) to allow readers to assess robustness. These additions will appear in Section 5. revision: yes

  2. Referee: [§4] The central claim that the preconditioner works for all members even when constructed from the control only rests on the quality of the RHS-difference sketching matrix; however, no quantitative comparison (e.g., approximation error norms or eigenvalue capture rates) is provided between this sketching choice and standard random sketching to isolate the contribution of the EDA-specific construction.

    Authors: We acknowledge that a side-by-side quantitative comparison would better isolate the benefit of the EDA-specific RHS-difference sketching matrix. In the revision we will add such a comparison, reporting relative approximation error norms and eigenvalue capture rates for the proposed sketching matrix versus standard Gaussian random matrices of identical dimensions. The new material will be placed in Section 4. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical acceleration claim is independent of inputs

full rationale

The paper describes an algorithmic construction that exploits EDA structure to build a sketching matrix from RHS differences, then reports numerical results on Lorenz-96 showing acceleration even when using only the control member. No derivation chain, fitted parameter renamed as prediction, or self-citation load-bearing step is present; the central claim is an empirical observation from experiments rather than a tautological reduction of outputs to inputs by construction. The method is presented as a direct exploitation of problem structure without circular redefinition.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated assumption that RHS differences yield a sufficiently accurate low-rank Hessian approximation for preconditioning purposes.

pith-pipeline@v0.9.0 · 5708 in / 1037 out tokens · 89842 ms · 2026-05-25T03:33:58.587771+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

62 extracted references · 62 canonical work pages

  1. [1]

    Enhanced parallelization of the incremental 4

    Bousserez, Nicolas and Guerrette, Jonathan J and Henze, Daven K , journal=. Enhanced parallelization of the incremental 4. 2020 , publisher=

    work page 2020

  2. [2]

    Quarterly Journal of the Royal Meteorological Society , volume=

    Optimal and scalable methods to approximate the solutions of large-scale Bayesian problems: Theory and application to atmospheric inversion and data assimilation , author=. Quarterly Journal of the Royal Meteorological Society , volume=. 2018 , publisher=

    work page 2018

  3. [3]

    Randomised preconditioning for the forcing formulation of weak-constraint

    Dau. Randomised preconditioning for the forcing formulation of weak-constraint. Quarterly Journal of the Royal Meteorological Society , volume=. 2021 , publisher=

    work page 2021

  4. [4]

    Randomized numerical linear algebra methods with application to data assimilation

    Alexandre Scotto Di Perrotolo. Randomized numerical linear algebra methods with application to data assimilation. 2022

    work page 2022

  5. [5]

    arXiv preprint arXiv:2405.04811 , year=

    A general error analysis for randomized low-rank approximation with application to data assimilation , author=. arXiv preprint arXiv:2405.04811 , year=

    work page arXiv

  6. [6]

    Randomized preconditioned solvers for strong constraint

    Subrahmanya, Amit N and Rao, Vishwas and Saibaba, Arvind K , journal=. Randomized preconditioned solvers for strong constraint. 2025 , publisher=

    work page 2025

  7. [7]

    arXiv preprint arXiv:2603.28969 , year=

    A Spectral Preconditioner for the Conjugate Gradient Method with Iteration Budget , author=. arXiv preprint arXiv:2603.28969 , year=

    work page arXiv

  8. [8]

    A strategy for operational implementation of

    Courtier, PHILIPPE and Th. A strategy for operational implementation of. Quarterly Journal of the Royal Meteorological Society , volume=. 1994 , publisher=

    work page 1994

  9. [9]

    Finding structure with randomness:

    Halko, Nathan and Martinsson, Per-Gunnar and Tropp, Joel A , journal=. Finding structure with randomness:. 2011 , publisher=

    work page 2011

  10. [10]

    Randomized numerical linear algebra:

    Martinsson, Per-Gunnar and Tropp, Joel A , journal=. Randomized numerical linear algebra:. 2020 , publisher=

    work page 2020

  11. [11]

    Drineas, Petros and Mahoney, Michael W , journal=. On the

    work page

  12. [12]

    Using the

    Williams, Christopher and Seeger, Matthias , journal=. Using the

    work page

  13. [13]

    Advances in Neural Information Processing Systems , volume=

    Fixed-rank approximation of a positive-semidefinite matrix from streaming data , author=. Advances in Neural Information Processing Systems , volume=

    work page

  14. [14]

    Matrix Algorithms:

    Stewart, Gilbert W , year=. Matrix Algorithms:

    work page

  15. [15]

    SIAM Journal on Matrix Analysis and Applications , volume=

    Randomized low-rank approximation of monotone matrix functions , author=. SIAM Journal on Matrix Analysis and Applications , volume=. 2023 , publisher=

    work page 2023

  16. [16]

    ACM Transactions on Mathematical Software (TOMS) , volume=

    Algorithm 971: An implementation of a randomized algorithm for principal component analysis , author=. ACM Transactions on Mathematical Software (TOMS) , volume=. 2017 , publisher=

    work page 2017

  17. [17]

    The 2020 global operational

    Bouyssel, F and Berre, L and B. The 2020 global operational. Data Assimilation for Atmospheric, Oceanic and Hydrologic Applications (Vol. IV) , pages=. 2022 , publisher=

    work page 2020

  18. [18]

    Seminar on recent development in data assimilation for atmosphere and ocean , pages=

    Background error covariance modelling , author=. Seminar on recent development in data assimilation for atmosphere and ocean , pages=. 2003 , organization=

    work page 2003

  19. [19]

    A variational assimilation ensemble and the spatial filtering of its error covariances: increase of sample size by local spatial averaging , booktitle =

    Berre, Lo. A variational assimilation ensemble and the spatial filtering of its error covariances: increase of sample size by local spatial averaging , booktitle =. 2007 , pages =

    work page 2007

  20. [20]

    The use of an ensemble approach to study the background error covariances in a global

    Belo-Pereira, Margarida and Berre, Lo. The use of an ensemble approach to study the background error covariances in a global. Monthly Weather Review , volume=

    work page

  21. [21]

    On the use of

    Bonavita, Massimo and Isaksen, Lars and H. On the use of. Quarterly Journal of the Royal Meteorological Society , volume=. 2012 , publisher=

    work page 2012

  22. [22]

    Quarterly Journal of the Royal Meteorological Society , volume=

    Accelerating and parallelizing minimizations in ensemble and deterministic variational assimilations , author=. Quarterly Journal of the Royal Meteorological Society , volume=. 2012 , publisher=

    work page 2012

  23. [23]

    Minimization algorithms for variational data assimilation , author=. Proc. of ECMWF Seminar on Recent Development in Numerical Methods for Atmospheric Modelling (7-11 September 1998, Reading, UK) , year=

    work page 1998

  24. [24]

    Mercier, Fran. Block. Quarterly Journal of the Royal Meteorological Society , volume=. 2018 , publisher=

    work page 2018

  25. [25]

    Speeding up the ensemble data assimilation system of the limited-area model of

    Mercier, Fran. Speeding up the ensemble data assimilation system of the limited-area model of. Quarterly Journal of the Royal Meteorological Society , volume=. 2019 , publisher=

    work page 2019

  26. [26]

    Quarterly Journal of the Royal Meteorological Society , volume=

    Limited-memory preconditioners, with application to incremental four-dimensional variational data assimilation , author=. Quarterly Journal of the Royal Meteorological Society , volume=. 2008 , publisher=

    work page 2008

  27. [27]

    Accounting for model error in the

    Raynaud, Laure and Berre, Lo. Accounting for model error in the. Quarterly Journal of the Royal Meteorological Society , volume=. 2012 , publisher=

    work page 2012

  28. [28]

    Towards understanding

    Carson, Erin and Liesen, J. Towards understanding. Linear Algebra and its Applications , volume=. 2024 , publisher=

    work page 2024

  29. [29]

    Journal of the Meteorological Society of Japan

    Development of an operational variational assimilation scheme , author=. Journal of the Meteorological Society of Japan. Ser. II , volume=. 1997 , publisher=

    work page 1997

  30. [30]

    SIAM Journal on Optimization , volume=

    On a class of limited memory preconditioners for large scale linear systems with multiple right-hand sides , author=. SIAM Journal on Optimization , volume=. 2011 , publisher=

    work page 2011

  31. [31]

    Randomized algorithms for low-rank matrix approximation:

    Tropp, Joel A and Webber, Robert J , journal=. Randomized algorithms for low-rank matrix approximation:

    work page

  32. [32]

    Predictability:

    Lorenz, Edward N , booktitle=. Predictability:. 1996 , organization=

    work page 1996

  33. [33]

    A two-layer quasi-geostrophic model of summer trough formation in the

    Fandry, CB and Leslie, LM , journal=. A two-layer quasi-geostrophic model of summer trough formation in the

    work page

  34. [34]

    Numerical Linear Algebra with Applications , volume=

    Impact of correlated observation errors on the conditioning of variational data assimilation problems , author=. Numerical Linear Algebra with Applications , volume=. 2024 , publisher=

    work page 2024

  35. [35]

    and Snyder, Chris , year =

    Hamill, Thomas M. and Snyder, Chris , year =. A hybrid ensemble. Monthly Weather Review , volume =

    work page

  36. [36]

    , year =

    Lorenc, Andrew C. , year =. The potential of the ensemble. Quarterly Journal of the Royal Meteorological Society , volume =

    work page

  37. [37]

    Ensemble-derived stationary and flow-dependent background-error covariances:

    Buehner, Mark , year =. Ensemble-derived stationary and flow-dependent background-error covariances:. Quarterly Journal of the Royal Meteorological Society , volume =. doi:10.1256/qj.04.15 , keywords =

    work page doi:10.1256/qj.04.15

  38. [38]

    1999 , month = jan, journal =

    Variational quality control , author =. 1999 , month = jan, journal =

    work page 1999

  39. [39]

    2011 , month = oct, journal =

    Observation errors in all-sky data assimilation , author =. 2011 , month = oct, journal =. doi:10.1002/qj.830 , urldate =

    work page doi:10.1002/qj.830 2011

  40. [40]

    M. An. 2015 , month = oct, journal =

    work page 2015

  41. [41]

    and Lawless, A

    Gratton, S. and Lawless, A. S. and Nichols, N. K. , title =. SIAM Journal on Optimization , volume =. 2007 , doi =

    work page 2007

  42. [42]

    2007 , publisher=

    Applied multivariate statistical analysis , author=. 2007 , publisher=

    work page 2007

  43. [43]

    A comparative study of

    Fertig, Elana J and Harlim, John and Hunt, Brian R , journal=. A comparative study of. 2007 , publisher=

    work page 2007

  44. [44]

    Coupling ensemble

    Zhang, Fuqing and Zhang, Meng and Hansen, James A , journal=. Coupling ensemble. 2009 , publisher=

    work page 2009

  45. [45]

    Quarterly Journal of the Royal Meteorological Society , volume=

    Parallelization in the time dimension of four-dimensional variational data assimilation , author=. Quarterly Journal of the Royal Meteorological Society , volume=. 2017 , publisher=

    work page 2017

  46. [46]

    Quarterly Journal of the Royal Meteorological Society , volume=

    Exploring the potential and limitations of weak-constraint 4D-Var , author=. Quarterly Journal of the Royal Meteorological Society , volume=. 2020 , publisher=

    work page 2020

  47. [47]

    Journal of Advances in Modeling Earth Systems , volume=

    Online model error correction with neural networks in the incremental 4D-Var framework , author=. Journal of Advances in Modeling Earth Systems , volume=. 2023 , publisher=

    work page 2023

  48. [48]

    SIAM Journal on Scientific Computing , volume=

    A deflated version of the conjugate gradient algorithm , author=. SIAM Journal on Scientific Computing , volume=. 2000 , publisher=

    work page 2000

  49. [49]

    and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and

    Virtanen, Pauli and Gommers, Ralf and Oliphant, Travis E. and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and. Nature Methods , year =

    work page

  50. [50]

    Revisiting the

    Gittens, Alex and Mahoney, Michael W , journal=. Revisiting the. 2016 , publisher=

    work page 2016

  51. [51]

    Randomized

    Persson, David and Boull. Randomized. SIAM Journal on Mathematics of Data Science , volume=. 2025 , publisher=

    work page 2025

  52. [52]

    2006 , publisher=

    Numerical optimization , author=. 2006 , publisher=

    work page 2006

  53. [53]

    Acta Numerica , volume=

    Preconditioning , author=. Acta Numerica , volume=. 2015 , publisher=

    work page 2015

  54. [54]

    2013 , school=

    Solving regularized nonlinear least-squares problem in dual space with application to variational data assimilation , author=. 2013 , school=

    work page 2013

  55. [55]

    Quarterly Journal of the Royal Meteorological Society , volume=

    An observation-space formulation of variational assimilation using a restricted preconditioned conjugate gradient algorithm , author=. Quarterly Journal of the Royal Meteorological Society , volume=. 2009 , publisher=

    work page 2009

  56. [56]

    2013 , publisher=

    Krylov subspace methods: principles and analysis , author=. 2013 , publisher=

    work page 2013

  57. [57]

    2003 , publisher=

    Iterative methods for sparse linear systems , author=. 2003 , publisher=

    work page 2003

  58. [58]

    Data assimilation in weather forecasting: a case study in

    Fisher, Mike and Nocedal, Jorge and Tr. Data assimilation in weather forecasting: a case study in. Optimization and Engineering , volume=. 2009 , publisher=

    work page 2009

  59. [59]

    International Conference on Learning Representations , year =

    A generalization of the randomized singular value decomposition , author=. International Conference on Learning Representations , year =

    work page

  60. [60]

    A Global Oceanic Data Assimilation System

    John Derber and Anthony Rosati. A Global Oceanic Data Assimilation System. Journal of Physical Oceanography. 1989. doi:10.1175/1520-0485(1989)019<1333:AGODAS>2.0.CO;2

    work page doi:10.1175/1520-0485(1989)019 1989

  61. [61]

    1999 , publisher=

    Frommer, Andreas and Maass, Peter , journal=. 1999 , publisher=

    work page 1999

  62. [62]

    Journal of research of the National Bureau of Standards , volume=

    Methods of conjugate gradients for solving linear systems , author=. Journal of research of the National Bureau of Standards , volume=

    work page