Pith. sign in

REVIEW 4 major objections 6 minor 47 references

Arnoldi Singular Vector perturbations for machine learning weather prediction

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper argues that an adjoint-free Arnoldi singular-vector method turns random noise into flow-dependent initial-condition perturbations for a 24-hour machine-learning weather model, and that these perturbations, unlike random noise…

desk verdict A useful proof-of-concept that adjoint-free Arnoldi perturbations grow in Pangu Weather, but the method is not actually computing singular vectors, and the paper needs more rigor before the SV claim can stand. read the letter →

arxiv 2506.22450 v1 pith:KBOSSELU submitted 2025-06-13 physics.ao-ph cs.LG

Jens Winkler , Michael Denhard This is my paper · ORCID
classification physics.ao-phcs.LG
keywords ArnoldiSingularVectorsMachineLearningPanguWeatherGenCastForecastUncertaintyPredictionEnsemblesadjoint-freeerrorgrowth
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a particular adjoint-free variant of singular-vector analysis—the directions in which small initial errors grow fastest over a chosen window—called Arnoldi Singular Vectors (A-SV), can find the fastest-growing initial-condition errors of a machine-learning weather model using only forward runs of the nonlinear forecast model. Applied to the 24-hour Pangu Weather model, the method starts from Gaussian noise and produces perturbation patterns concentrated in regions the atmosphere actually amplifies: near-surface temperature, the upper-troposphere jet streams, and the polar front. The paper reports that these A-SV perturbations grow from the very first forecast step, while random perturbations of the same amplitude are damped by the model and need about four days to recover their initial size. If true, the method offers a practical route to initializing ensemble forecasts for machine-learning weather prediction without writing tangent-linear or adjoint code.

What carries the argument

The load-bearing object is the evolved increment matrix (EIM), $A_{\tau,x_0,h}$, whose $i$-th column is the nonlinear evolved increment $M_\tau(x_0+h e_i)-M_\tau(x_0)$ of the forecast map around a reference state $x_0$. Equation (7) approximates the evolved increment of an arbitrary perturbation vector $v$ as the matrix-vector product $A v$, so each Arnoldi step needs only forward model integrations, not a tangent-linear or adjoint model. The Arnoldi iteration keeps an orthonormal basis $Q$ and a small projected operator $H$; the SVD of $H$ yields approximate singular vectors, which are mapped back into model space by multiplying with $Q$. In the experiments the forecast map is the 24-hour Pangu inference with $\tau=24$h, the perturbation amplitude is $h=500$, distances use a massless total-energy measurement on temperature and wind variables, and the block Arnoldi variant uses blocksize 8 over 12 loops, giving a 96-dimensional Krylov space.

What would settle it

Repeat the A-SV construction on the same reference state with amplitudes $h=50$, $h=200$, $h=500$, and $h=2000$; if the leading singular vectors or their 24-hour growth rates change discontinuously, the secant-linear approximation is not stable at the chosen scale. Alternatively, sample random directions $v$ and compare $\|M_{24h}(x_0+h v)-M_{24h}(x_0)-A_{\tau,x_0,h}v\|$ with $\|M_{24h}(x_0+h v)-M_{24h}(x_0)\|$; large relative mismatches would show that equation (7) does not represent the model's local error growth.

Watch

Extended reading notes

Core claim

On its own terms, the central claim is that the leading right singular vectors of an evolved increment matrix are genuine local unstable modes of the 24-hour Pangu Weather model. The A-SV algorithm constructs that matrix column by column from nonlinear forecasts, builds a 96-dimensional Krylov subspace, and projects the singular value decomposition of a small operator back into the full model state. In daily experiments from December 2024 to February 2025, the leading singular vectors show flow-dependent, weather-system-scale structure: temperature increments concentrated near the surface and wind increments in the upper-troposphere jet streams, consistent with a displacement of the polar front. The same runs show their amplitudes growing from forecast hour zero, in contrast to random noise, which Pangu dampens during the first steps and which needs roughly four days to regain its initial amplitude. The paper reads this as evidence that A-SV transforms unstructured noise into perturbations conditioned on the specific reference state, a process it compares with the denoising step of diffusion-based forecasters such as GenCast.

Load-bearing premise

The paper's method stands on the secant-linearity assumption in equation (7): the response of the 24-hour forecast model to a combined perturbation is treated as the same weighted sum of its responses to each basis perturbation, so the evolved increment matrix only describes true local error growth if the model is smooth enough at the chosen amplitude $h=500$.

Editorial extensions

If this is right

  • Any forecast model, numerical or machine-learned, can in principle be probed for its fastest-growing initial-condition perturbations without building tangent-linear or adjoint versions of that model.
  • For machine-learning weather prediction, the leading A-SV perturbations are a concrete candidate basis for initializing ensemble forecasts, with paired plus/minus or cluster-based schemes available to keep the ensemble mean unchanged.
  • Starting from random noise and letting the model's own dynamics shape the patterns means the algorithm simultaneously diagnoses where a model is locally unstable (singular values above one) and where it damps perturbations.
  • Relative to diffusion-based ensemble generation such as GenCast, the A-SV approach defines its ensemble members from one run at initial time rather than from many stochastic sampling trajectories.
  • Because ML inference is cheap and the block Arnoldi loop parallelizes, much larger Krylov subspaces are affordable and could expose unstable modes that a 96-dimensional subspace misses.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper leaves implicit is to use the A-SV singular spectrum as a compact fingerprint of a machine-learning model's error-growth behavior, letting different models be compared by how many quickly growing modes they have and where those modes sit.
  • Since the experiments fix the amplitude at $h=500$ and the optimization window at the model's 24-hour step, a natural test is to make the amplitude adaptive to local flow curvature or to chain shorter A-SV windows; the results could change the measured growth rates even if the leading structures stay similar.
  • The paper's comparison with diffusion models suggests a hybrid: use A-SV patterns as the starting seeds for a reverse-diffusion sampler such as GenCast, so the perturbation geography is set by the model dynamics rather than by the learned score function.
  • A forecast-verification experiment comparing ensemble spread and a probabilistic skill score such as CRPS for A-SV-initialized ensembles against randomly perturbed ensembles on the same Pangu model would tell whether these growing directions actually improve probabilistic skill, which the paper does not yet test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces Arnoldi Singular Vectors (A-SV), an adjoint-free method to compute singular-vector-like perturbations for machine learning weather prediction, and applies it to the 24h Pangu Weather model. The method iteratively applies the nonlinear forecast model to perturbed states, builds a Krylov subspace from the evolved increments, and computes the SVD of the resulting Hessenberg matrix. The authors report that the leading A-SV perturbations grow faster than random noise from the start of the forecast, display flow-dependent structures located near jets and the polar front, and could serve as ensemble perturbations. They also discuss conceptual similarities to and differences from diffusion-based ensembles such as GenCast.

Significance. If the A-SV perturbations indeed approximate the leading singular vectors of the local propagator, the method is significant: it offers a practical, matrix-free and adjoint-free route to initial-condition perturbation construction for both NWP and MLWP, and the demonstration on Pangu Weather opens a low-cost path to ensemble generation for a class of models that currently lack traditional singular vector machinery. The paper also contributes a clean description of the evolved increment matrix and explicitly acknowledges the norm- and amplitude-dependence of the construction. However, the central mathematical claim linking the Arnoldi process to true singular vectors of a non-normal propagator is not substantiated in the manuscript, and the empirical evidence is drawn from averaged three-month diagnostics without uncertainty quantification. The significance is therefore conditional on resolving the algorithmic-convergence question and on providing statistical support for the growth claims.

major comments (4)
  1. [Section 3.4, eq. (3)] The central argument that the SVD of the low-dimensional Hessenberg matrix H approximates the SVD of the full operator A is not generally valid for non-normal A. The Arnoldi process constructs a Krylov subspace K_m(A,v) = span{v, Av, ..., A^{m-1}v}, while the right singular vectors of A are eigenvectors of A^T A, whose Krylov subspace is generated by A^T A, not by A. For a strongly non-normal operator such as the 24h atmospheric propagator, the Ritz vectors extracted from the SVD of H need not converge to the dominant right singular vectors of A; the paper itself notes that L-SV and A-SV set up different Krylov subspaces (Section 3.5) but does not quantify the resulting gap. The paper refers to Winkler et al. (2020) for mathematical plausibility and error bounds, but provides no convergence criterion, no demonstration of convergence with increasing Krylov dimension, and no comparison with true singular vectors obtained from, e.g., a finite-difference Jacobian or a Lanczos-SV implementation on the same model. Without such validation, the interpretation of the leading vectors as 'Singular Vectors' that describe the fastest-growing local modes is unsupported.
  2. [Section 3.3, eq. (7)] The secant-linearity assumption in eq. (7) is load-bearing: it equates the nonlinear evolved increment of an arbitrary perturbation vector to a weighted sum of nonlinear evolved increments of the standard basis vectors. The paper sets h=500 (Section 5.1) and states that this choice is 'based on observed variances of real states', but provides no quantitative justification, no sensitivity analysis with respect to h, blocksize, or number of loops, and no test of how well the linear superposition holds for the Pangu model in the h=500 neighborhood. The potential failure is not hypothetical: MLWP models are trained on smooth fields but may have highly non-smooth local response in the latent space; moreover, the paper's own remark in Section 3.5 that the amplitude choice is 'nontrivial' and the suggested indicator 'if a set of computed A-SV perturbations does not grow right from the beginning of the forecast, this is a relevant indicator that the amplitude was chosen too large' creates a circularity in which the amplitude is tuned precisely to make the leading perturbations grow, and that same growth is then reported as evidence for the method's skill.
  3. [Section 5.4, Figures 6 and 7] The central empirical claim that A-SVs grow faster than random noise is based on three-month averages without any measure of spread, confidence intervals, or significance tests. The figures show that the A-SV curves lie above the random-noise curve, but the sample variability across the daily experiments is not reported, and the differences could in principle be driven by a few outlier days or by the inherent selection mechanism (the perturbations are explicitly constructed to maximize growth under the same model, so faster growth is partially built into the definition). The paper also does not quantify the potential contamination from the fact that the optimization window is 24h while the growth is evaluated over 168h; a rigorous comparison should include a verification against random perturbations that are rescaled to the same initial amplitude in the same energy norm, and should report the distribution of growth rates over the three-month sample.
  4. [Section 5.3 and 5.5] The claim that the A-SV perturbations are located in dynamically sensitive regions (jet streams, polar front) is supported only by visual inspection of selected case studies (15 January 2025 and 14 January 2025). No statistical evaluation is provided, such as a comparison with forecast error covariance, analysis error covariance, or a null test using random perturbations projected onto the same norm. The statement in Section 5.3 that 'the wind pattern ... coincides with the increased temperatures in polar regions, what indicates a southward shift of the polar front' is a qualitative interpretation; the paper should either provide a quantitative diagnostic (e.g., correlation of the perturbation energy with local baroclinicity or Eady growth rate) or soften the claim to a hypothesis for future work.
minor comments (6)
  1. [Section 6] The heading 'SUMMERY' should be corrected to 'SUMMARY'.
  2. [Figure 7 caption] The cross-reference in the Figure 7 caption appears as 'see section [ref]'; a proper section number should be inserted.
  3. [Section 7.1] The text states that eqs. (8) and (9) do not define a valid norm and proposes the term 'total energy measurement' (TEM), yet later in the same section and in Section 3.5 the term 'norm' is still used (e.g., 'the used norm' and 'Since a common atmospheric data state ... in the range 100-10000'). The terminology should be made consistent throughout.
  4. [Section 3.4, Algorithm 1] In the pseudocode, the matrix dimension of Q is denoted n x l m, but earlier 'm' is the number of loops; for the non-block algorithm m equals the number of basis vectors, while for the block algorithm the basis dimension is l*m. The notation is ambiguous; please define clearly whether m is the number of loops or the target subspace dimension.
  5. [Abstract and Section 4] The word 'therefor' appears in the abstract and should be 'therefore'; the sentence 'therefor we briefly discuss similarities and differences.' is grammatically awkward and should be rewritten.
  6. [Section 5.1] The choice of the northern-hemisphere domain and the restriction to latitudes above 35 degrees is stated, but the exact number of degrees of freedom (13 million variables) is given later in Section 5.4 without a derivation; please state how this number is computed from the Pangu grid and the 13 levels, so that readers can assess the subspace-to-domain ratio.

Circularity Check

1 steps flagged · score 5.0 of 10

The A-SV growth result is partly built into the SVD objective, but the physical-structure evidence and random-noise damping are independent; no deeper circularity found.

  1. self definitional [Section 3.3 (Eq. 6-7) and Section 5.3 / Section 6]
    "The leading SVs of an EIM will point in directions, where an integration with the full non-linear model generates the largest increments Iτ,x0,h after time τ (large singular values; unstable manifold) ... The A-SV method is designed to set perturbations, where it gets the larger effects in terms of perturbation energy and this is indeed observable. ... In contrast to the random perturbations the A-SVs start growing right from the beginning of the forecast, which is the intended goal of the algorithm."

    By Eq. (7), the evolved increment matrix A is built from secant forecasts, and the A-SV perturbations are the right singular vectors of H projected by Q; they are, by definition, the perturbations with maximal evolved-increment norm inside the Krylov subspace. Reporting that these perturbations grow faster than random noise over the first 24 h (Figures 6-7) is therefore largely a restatement of the optimization objective, not an independent empirical prediction. The paper's own wording—'designed to set perturbations where it gets the larger effects' and 'intended goal'—confirms this. The non-circular content is the flow-dependent placement near jets/polar front, vertical structure, and the independent observation that Pangu dampens random noise.

full rationale

The central growth claim is partly tautological: the method constructs perturbations by maximizing the 24 h evolved-increment norm, so showing that the leading A-SVs grow in that same 24 h window substantially restates the SVD objective. The paper itself labels this the 'intended goal of the algorithm.' This is a genuine but limited circularity, and the score reflects it. The paper's other evidence is not circular: the localization of A-SV patterns near jet streams and the polar front, the vertical structure of the perturbations, and the fact that random noise is damped by Pangu Weather are all independent of the optimization objective and would not be forced by the definition of the SVs. I do not treat the reference to Winkler et al. (2020) as load-bearing circularity, because it is a peer-reviewed methods paper and the present work applies the method as a tool; the mathematical concern about Arnoldi approximating singular vectors of a non-normal operator without an adjoint is a correctness/validation issue rather than a circularity. Weighing these, the paper deserves a moderate score for partial self-definition, not a high score for wholesale circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central computational claim rests on a series of modeling and numerical choices, most importantly the secant linearity of evolved increments (eq. 7), the small size of the Krylov subspace relative to state dimension, and the hand-chosen amplitude and norm. These are not independently justified beyond plausibility arguments.

free parameters (6)
  • perturbation amplitude h = 500
    Chosen 'on the basis of observed variances of real states' (Section 5.1), but no systematic sensitivity study is provided; growth rates and perturbation structure may depend on this value.
  • blocksize = 8
    Algorithmic choice for the block Arnoldi variant; the spectrum tail shows steps at this size (Section 5.2), indicating it affects the Krylov subspace structure.
  • number of loops = 12
    Together with blocksize gives a 96-dimensional Krylov subspace; the authors acknowledge the Krylov dimension is small relative to the 13 million variables (Section 5.4), so results depend on this truncation.
  • optimization window tau = 24h
    Set equal to the Pangu 24h forecast step; the paper notes longer windows become dominated by nonlinearity, so this choice is a model-specific compromise.
  • domain and variables = Northern Hemisphere >35N; T,U,V on 13 levels plus T2m,U10m,V10m; exclude surface pressure
    Reduced total energy norm (Section 7.1) and domain selection are modeling choices that affect the norm and thus the SV definition.
  • number of returned SVs k = first two leading SVs plus SV17, with discussion of tails
    The number of SVs used for analysis and potential ensembles is not systematically justified; the paper recommends only the leading part of the spectrum.
assumptions (5)
  • domain assumption The dynamical system M_tau is sufficiently smooth so that nonlinear evolved increments of an arbitrary vector v can be approximated by a linear combination of the nonlinear evolved unitary vectors.
    Section 3.3, eq. (7): this secant linearity is the foundation of the EIM, but its validity for Pangu Weather is not tested.
  • domain assumption The projected leading singular vectors of the 96-dimensional H approximate the leading singular vectors of the full n-dimensional EIM.
    Section 3.2, eq. (3): 'The assumption is that the projection of the SVs of H into n-dimensional model space by multiplication with Q sufficiently covers relevant modes of A.' Standard Krylov assumption, not verified for this problem.
  • domain assumption The total energy measurement, with surface pressure neglected, defines a meaningful metric for perturbation growth in this system.
    Section 5.1 and 7.1: the norm is central to SV computation; neglecting surface pressure and varying the reference temperature could alter the SVs.
  • domain assumption Pangu Weather 24h model is treated as a deterministic perfect model whose sensitivity to ICON-analysis initial states reflects real atmospheric predictability limits.
    Section 2: the paper uses the pretrained Pangu model as a black box, ignoring possible model error and the discrepancy between the ERA5-trained model and ICON analyses.
  • domain assumption The reference state (ICON 00UTC analysis) is an appropriate point for local SV analysis.
    Section 5.1: the SV analysis is local to a given reference state; the choice of ICON analysis rather than Pangu's native ERA5-like state could bias the computed sensitivities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Arnoldi Singular Vector perturbations for machine learning weather prediction." pith.science (2026). https://pith.science/paper/KBOSSELU

@misc{pith2026250622450,
  author       = {Pith},
  title        = {Pith review of: Arnoldi Singular Vector perturbations for machine learning weather prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KBOSSELU}},
  note         = {Machine review of arXiv:2506.22450}
}
read the original abstract

Since weather forecasts are fundamentally uncertain, reliable decision making requires information on the likelihoods of future weather scenarios. We explore the sensitivity of machine learning weather prediction (MLWP) using the 24h Pangu Weather ML model of Huawei to errors in the initial conditions with a specific kind of Singular Vector (SV) perturbations. Our Arnoldi-SV (A-SV) method does not need linear nor adjoint model versions and is applicable to numerical weather prediction (NWP) as well as MLWP. It observes error growth within a given optimization time window by iteratively applying a forecast model to perturbed model states. This creates a Krylov subspace, implicitly based on a matrix operator, which approximates the local error growth. Each iteration adds new dimensions to the Krylov space and its leading right SVs are expected to turn into directions of growing errors. We show that A-SV indeed finds dynamically meaningful perturbation patterns for the 24h Pangu Weather model, which grow right from the beginning of the forecast rollout. These perturbations describe local unstable modes and could be a basis to initialize MLWP ensembles. Since we start A-SV from random noise perturbations, the algorithm transforms noise into perturbations conditioned on a given reference state - a process that is akin to the denoising process of the generic diffusion based ML model of GenCast, therefor we briefly discuss similarities and differences.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 35 canonical work pages

  1. [2]

    Arnoldi, W. E. (1951) The Principle of Minimized Iterations in the Solution of the Matrix Eigenvalue Problem . Quarterly of Applied Mathematics, 9, 17--29

    work page 1951

  2. [3]

    Ben-Bouallegue, Z., Clare, M. C. A., Magnusson, L., Gascon, E., Maier-Gerber, M., Janousek, M., Rodwell, M., Pinault, F., Dramsch, J. S., Lang, S. T. K., Raoult, B., Rabier, F., Chevallier, M., Sandu, I., Dueben, P., Chantry, M. and Pappenberger, F. (2023) The rise of data-driven weather forecasting. ://arxiv.org/abs/2307.10128

    work page Pith review arXiv 2023

  3. [4]

    and Tian, Q

    Bi, K., Xie, L., Zhang, H., Chen, X., Gu, X. and Tian, Q. (2023) Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619, 533--538

    work page 2023

  4. [5]

    Chen, K., Han, T., Gong, J., Bai, L., Ling, F., Luo, J.-J., Chen, X., Ma, L., Zhang, T., Su, R. et al. (2023) Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv preprint arXiv:2304.02948

    arXiv 2023

  5. [6]

    Croitoru, F.-A., Hondru, V., Ionescu, R. T. and Shah, M. (2023) Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45, 10850--10869

    work page 2023

  6. [7]

    Diaconescu, E. P. and Laprise, R. (2012) Singular vectors in atmospheric sciences: A review . Earth-Science Reviews 113, 113, 161--175

    work page 2012

  7. [8]

    Durran, D. R. and Gingrich, M. (2014) Atmospheric predictability: Why butterflies are not of practical importance. Journal of the Atmospheric Sciences, 71, 2476 -- 2488. ://journals.ametsoc.org/view/journals/atsc/71/7/jas-d-14-0007.1.xml

    work page 2014

  8. [9]

    Kalnay, M

    E. Kalnay, M. Crazza, M. C. (2002) A re bred vectors the same as L yapunov vectors? Symp. on Observations, Data Assimilation an Probablistic Prediction, Amer. Meteor. Soc., 173--177

    work page 2002

Show all 47 references
  1. [10]

    Part V: Ensemble Prediction System

    ECMWF (2019) IFS Documentation CY46R1, chap. Part V: Ensemble Prediction System. ECMWF. ://www.ecmwf.int/node/19309

    2019

  2. [11]

    and Errico, R

    Ehrendorfer, M. and Errico, R. M. (1995) Mesoscale predictability and the spectrum of optimal perturbations. J. Atmos. Sci., 52, 3475--3500

    1995

  3. [12]

    Golub, G. H. and Loan, C. F. V. (1996) Matrix Computations (3rd ed.) . Baltimore, MD, USA: Johns Hopkins Univ. Press

    1996

  4. [13]

    Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D. et al. (2020) The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146, 1999–2049

    2020

  5. [14]

    and Abbeel, P

    Ho, J., Jain, A. and Abbeel, P. (2020) Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20. Red Hook, NY, USA: Curran Associates Inc

    2020

  6. [15]

    (2003) Atmospheric modeling, data assimilation and predictability

    Kalnay, E. (2003) Atmospheric modeling, data assimilation and predictability. Cambridge university press

    2003

  7. [16]

    and Aila, T

    Karras, T., Aittala, M., Laine, S. and Aila, T. (2022) Elucidating the design space of diffusion-based generative models. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22. Red Hook, NY, USA: Curran Associates Inc

    2022

  8. [17]

    (2022) Forecasting global weather with graph neural networks

    Keisler, R. (2022) Forecasting global weather with graph neural networks. arXiv preprint arXiv:2202.07575

    2022 arXiv

  9. [18]

    and Anandkumar, A

    Kurth, T., Subramanian, S., Harrington, P., Pathak, J., Mardani, M., Hall, D., Miele, A., Kashinath, K. and Anandkumar, A. (2022) FourCastNet : Accelerating global high-resolution weather forecasting using adaptive fourier neural operators. arXiv preprint arXiv:2208.05419

    2022 arXiv

  10. [19]

    Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W. et al. (2023) Learning skillful medium-range global weather forecasting. Science, 382, 1416--1421

    2023

  11. [20]

    (1950) An Iteration Method for the Solution of the Eigenvalue Problem of Linear Differential and Integral Operators

    Lanczos, C. (1950) An Iteration Method for the Solution of the Eigenvalue Problem of Linear Differential and Integral Operators . Journal of research of the National Bureau of Standards, 45 (4), 255--282

    1950

  12. [21]

    and Palmer, T

    Leutbecher, M. and Palmer, T. N. (2008) Ensemble forecasting . Journal of computational physics, 227, 3515--3539

    2008

  13. [22]

    and Qi, Y

    Li, H., Chen, L., Zhong, X., Zhang, F., Cheng, Y., Xu, Y. and Qi, Y. (2023) FuXi : A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6

    2023

  14. [23]

    Lorenz, E. N. and Haman, K. (1996) The essence of chaos. Pure and Applied Geophysics, 147, 598–599

    1996

  15. [24]

    (2009) Sampling Uncertainities in Ensemble Weather Forecasting

    Magnusson, L. (2009) Sampling Uncertainities in Ensemble Weather Forecasting. Ph.D. thesis, Department of Meteorology, Stockholm University, Stockholm, Sweden. Universitetsservice US-AB

    2009

  16. [25]

    and Tebbens, J

    Meurant, G. and Tebbens, J. D. (2020) Krylov Methods for Nonsymmetric Linear Systems , vol. 57 of Springer Series in Computational Mathematics . Springer Nature Switzerland AG. ://doi.org/10.1007/978-3-030-55251-0

    2020 doi

  17. [26]

    and Petroliagis, T

    Molteni, F., Buizza, R., Palmer, T. and Petroliagis, T. (1996) The new ECMWF ensemble prediction system: Methodology and validation . Quarterly Journal of the Royal Meteorological Society, 122, 73--119

    1996

  18. [27]

    and Palmer, T

    Mureau, R., Molteni, F. and Palmer, T. (1993) Ensemble prediction using dynamically conditional perturbations . Quarterly Journal of the Royal Meteorological Society, 119, 299--323

    1993

  19. [28]

    and Grover, A

    Nguyen, T., Shah, R., Bansal, H., Arcomano, T., Madireddy, S., Maulik, R., Kotamarthi, V., Foster, I. and Grover, A. (2023) Scaling transformer neural networks for skillful and reliable medium-range weather forecasting. arXiv preprint arXiv:2312.03876

    2023 arXiv

  20. [29]

    (2019) The ECMWF ensemble prediction system: Looking back (more than) 25 years and projecting forward 25 years

    Palmer, T. (2019) The ECMWF ensemble prediction system: Looking back (more than) 25 years and projecting forward 25 years. Quarterly Journal of the Royal Meteorological Society, 145, 12–24

    2019

  21. [30]

    and Hagedorn, R

    Palmer, T. and Hagedorn, R. (2006) Predictability of weather and climate. Cambridge University Press

    2006

  22. [31]

    Palmer, T. N. (1993) Extended-range atmospheric prediction and the lorenz model. Bulletin of the American Meteorological Society, 74, 49--66. ://journals.ametsoc.org/view/journals/bams/74/1/1520-0477_1993_074_0049_erapat_2_0_co_2.xml

    1993

  23. [32]

    R., El-Kadi, A., Masters, D., Ewalds, T., Stott, J., Mohamed, S., Battaglia, P., Lam, R

    Price, I., Sanchez-Gonzalez, A., Alet, F., Andersson, T. R., El-Kadi, A., Masters, D., Ewalds, T., Stott, J., Mohamed, S., Battaglia, P., Lam, R. and Willson, M. (2024) Gencast: Diffusion-based ensemble forecasting for medium-range weather. ://arxiv.org/abs/2312.15796

    2024 arXiv

  24. [33]

    Nature, 637, 84–90

    --- (2025) Probabilistic weather forecasting with machine learning. Nature, 637, 84–90

    2025

  25. [34]

    Roberts, N., Ayliffe, B., Evans, G., Moseley, S., Rust, F., Sandford, C., Trzeciak, T., Abernethy, P., Beard, L., Crosswaite, N. et al. (2023) Improver: The new probabilistic postprocessing system at the Met office. Bulletin of the American Meteorological Society, 104, E680–E697

    2023

  26. [35]

    (1979) Implementation aspects of band lanczos algorithms for computation of eigenvalues of large sparse symmetric matrices

    Ruhe, A. (1979) Implementation aspects of band lanczos algorithms for computation of eigenvalues of large sparse symmetric matrices. MATHEMATICS OF COMPUTATION, 33, 680--687

    1979

  27. [36]

    (2003) Iterative Methods for Sparse Linear Systems

    Saad, Y. (2003) Iterative Methods for Sparse Linear Systems. Society for Industrial and Applied Mathematics, second edn. ://epubs.siam.org/doi/abs/10.1137/1.9780898718003

    2003 doi

  28. [37]

    and Schultz, M

    Saad, Y. and Schultz, M. H. (1986 a ) GMRES: A generalized minimal residual algorithm for solving nonsymmetric linear systems . SIAM J. Sci. Stat. Comput., 7, 856--869

    1986

  29. [38]

    SIAM Journal on Scientific and Statistical Computing, 7, 856--869

    --- (1986 b ) Gmres: A generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM Journal on Scientific and Statistical Computing, 7, 856--869. ://doi.org/10.1137/0907058

    1986 doi

  30. [39]

    and Craig, G

    Selz, T. and Craig, G. C. (2023) Can artificial intelligence-based weather prediction models simulate the butterfly effect? Geophysical Research Letters, 50, e2023GL105747. ://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2023GL105747. E2023GL105747 2023GL105747

    2023 doi

  31. [40]

    and Craig, G

    Selz, T., Riemer, M. and Craig, G. C. (2022) The transition from practical to intrinsic predictability of midlatitude weather. Journal of the Atmospheric Sciences, 79, 2013 -- 2030. ://journals.ametsoc.org/view/journals/atsc/79/8/JAS-D-21-0271.1.xml

    2022

  32. [41]

    P., Kumar, A., Ermon, S

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S. and Poole, B. (2021) Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations

    2021

  33. [42]

    and Schmitt, B

    Winkler, J., Denhard, M. and Schmitt, B. A. (2020) Krylov methods for adjoint-free singular vector based perturbations in dynamical systems. Quarterly Journal of the Royal Meteorological Society, 146, 225–239. ://rmets.onlinelibrary.wiley.com/doi/abs/10.1002/qj.3668

    2020 doi

  34. [43]

    and Yoshida, T

    Yamaguchi, H., Hotta, D., Kanehama, T., Ochi, K., Ota, Y., Sekiguchi, R., Shimpo, A. and Yoshida, T. (2018) Introduction to JMA ’s new global ensemble prediction system. CAS/JSC WGNE, Research Activities in Atmospheric and Oceanic Modelling, 42, 6–13

    2018

  35. [44]

    and Cui, B

    Zhu, Y., Toth, Z., Wobus, R., Wei, M. and Cui, B. (2012b) May 2006 upgrade of the GEFS and first implementation of NAEFS systems

    2012

  36. [45]

    and Baldauf, M

    Zängl, G., Reinert, D., Rípodas, P. and Baldauf, M. (2015) The icon (icosahedral non-hydrostatic) modelling framework of dwd and mpi-m: Description of the non-hydrostatic dynamical core. Quarterly Journal of the Royal Meteorological Society, 141, 563--579. ://rmets.onlinelibra...

    2015 doi

  37. [46]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.senten...

  38. [47]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  39. [48]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.