pith. sign in

arxiv: 2606.22742 · v1 · pith:7A3UBCP4new · submitted 2026-06-22 · 🧮 math.NA · cs.NA

Factored Sparse Approximate Inverse Preconditioning via Spectral Optimization

Francesco Brarda , Tianshi Xu , Vassilis Kalantzis , Rui Peng Li , Yuanzhe Xi This is my paper

Pith reviewed 2026-06-26 08:06 UTC · model grok-4.3

classification 🧮 math.NA cs.NA
keywords factored sparse approximate inversespectral optimizationpreconditioningsaddle point systemsLanczos methodindefinite linear systemsfinite element methodsgradient based optimization
0
0 comments X

The pith

Spectral optimization of fixed-pattern FSAI factors yields preconditioners that require fewer iterations than Frobenius-based ones for indefinite systems.

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

The paper shows that for a given sparsity pattern in the factor G of a factored sparse approximate inverse, selecting the nonzero entries by minimizing a spectral objective on the preconditioned operator G A G^T produces better preconditioners than the classical approach of minimizing the Frobenius norm of the residual. For positive definite matrices the objective clusters eigenvalues near one, while for indefinite matrices a bimodal loss pushes positive eigenvalues toward one and negative ones toward minus one while penalizing those near zero. This spectral choice is made practical for large matrices by a projected gradient method that uses Lanczos runs to estimate the objective and its gradient via a detached Rayleigh surrogate, avoiding differentiation through the Lanczos process. Experiments on finite-element problems confirm fewer solver iterations, especially for symmetric indefinite saddle-point systems, and a graph neural network is used to predict entries across related matrices.

Core claim

By optimizing a spectral loss rather than a Frobenius residual for the admissible entries of G in the preconditioner P = G A G^T, the resulting factored sparse approximate inverse requires fewer iterations of Krylov solvers on symmetric indefinite saddle-point systems from finite-element discretizations.

What carries the argument

The bimodal spectral loss for indefinite systems, minimized via projected Krylov support-gradients computed with a detached Rayleigh surrogate from Lanczos data.

If this is right

  • Preconditioners obtained this way reduce the iteration count of iterative solvers relative to classical FSAI on the tested problems.
  • The method preserves the inertia of the original matrix through congruence while clustering eigenvalues away from zero.
  • The detached computation allows matrix-free gradients on the support of G without backpropagating through Lanczos.
  • Graph neural networks trained on similar matrices can predict suitable initial values for the admissible entries.

Where Pith is reading between the lines

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

  • This spectral approach could be combined with adaptive pattern selection to further improve performance.
  • Similar optimization ideas might apply to other sparse preconditioners such as incomplete factorizations.
  • The reliability of the non-convex minimization suggests testing on a broader range of problem sizes and types to assess robustness without tuning.
  • The GNN predictor opens the possibility of transfer learning across families of related linear systems.

Load-bearing premise

The non-convex spectral loss can be minimized to a local minimum that improves the preconditioner quality without requiring extensive problem-specific tuning of the optimizer.

What would settle it

If experiments on the finite-element test problems show that the number of Krylov iterations with the spectrally optimized FSAI is not smaller than with the Frobenius version, or that the optimization frequently fails to produce usable factors.

read the original abstract

In this paper, we study value selection for fixed-pattern factorized sparse approximate inverse preconditioners. Given a prescribed sparsity pattern for a factor $G,$ we choose its admissible entries by optimizing spectral objectives of the congruent preconditioned operator $P(G)=GAG^T.$ This differs from classical sparse approximate inverse and FSAI constructions, which choose entries through algebraic Frobenius-residual criteria. For symmetric positive definite systems, the spectral target is a cluster near $+1.$ For symmetric indefinite systems, where congruence preserves inertia, we introduce a bimodal loss that drives positive and negative eigenvalues toward separated clusters near $+1$ and $-1,$ while penalizing eigenvalues near zero. To make these objectives practical for large sparse matrices, we derive projected Krylov support-gradients. Lanczos runs provide both a stochastic trace estimate of the spectral objective and a Ritz approximation to the exact gradient. We implement the resulting gradient through a detached Rayleigh surrogate: the Lanczos data are computed without gradient tracking and held fixed, while the backward pass differentiates only recomputed Rayleigh quotients with respect to the admissible entries of $G.$ This avoids differentiating through the Lanczos recurrence while returning a matrix-free gradient on the prescribed support. We also discuss a projected Kernel Polynomial Method rule as a finite polynomial comparison. Experiments on finite-element test problems show that spectral value selection improves fixed-support preconditioners, especially for symmetric indefinite saddle-point systems. We further demonstrate a graph neural network model for predicting admissible factor entries across related matrices.

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

3 major / 2 minor

Summary. The manuscript proposes selecting nonzero entries in a fixed-pattern factored sparse approximate inverse (FSAI) preconditioner G by minimizing spectral objectives on the congruent preconditioned operator P(G)=G A G^T, rather than using classical Frobenius-norm criteria. For SPD systems the target is eigenvalue clustering near +1; for symmetric indefinite systems a bimodal loss drives positive/negative eigenvalues toward separated clusters at +1 and -1 while penalizing those near zero. Practical large-scale optimization is achieved via projected Krylov support-gradients obtained from Lanczos runs, using a detached Rayleigh surrogate that computes the forward Lanczos data without gradient tracking and differentiates only recomputed Rayleigh quotients on the admissible support of G. Experiments on finite-element problems report that the resulting preconditioners require fewer iterations than Frobenius FSAI, especially on indefinite saddle-point systems; a graph neural network predictor for admissible entries across related matrices is also presented.

Significance. If the projected Krylov support-gradient procedure consistently locates useful local minima of the non-convex bimodal loss, the work supplies a principled, matrix-free alternative to algebraic FSAI constructions that demonstrably improves iteration counts on symmetric indefinite saddle-point problems arising in finite-element discretizations. The detached-surrogate technique for gradient estimation and the GNN extension are technically interesting contributions that could extend to other spectral preconditioning tasks.

major comments (3)
  1. [Numerical experiments] The central empirical claim (fewer iterations than Frobenius FSAI on indefinite saddle-point systems) rests on the projected Krylov support-gradient reliably minimizing the bimodal spectral loss. The manuscript notes the non-convexity but provides no ablation on Lanczos depth, step-size schedules, or multiple random initializations; without such evidence it is unclear whether the reported gains generalize or require problem-specific tuning.
  2. [Gradient derivation (projected Krylov support-gradient)] The detached Rayleigh surrogate is asserted to return a usable matrix-free gradient on the prescribed support. For the bimodal loss, however, the interaction between the stochastic Lanczos trace estimate and the surrogate gradient is not analyzed; a small-scale verification comparing the surrogate gradient norm and direction to a finite-difference reference on a 100-by-100 indefinite test matrix would strengthen the claim that the optimizer is not systematically misled.
  3. [Spectral objective for indefinite systems] The bimodal loss is defined to penalize eigenvalues near zero while respecting inertia preservation under congruence. The precise functional form of the penalty term and its weighting relative to the clustering terms are not shown to guarantee that the global minimum corresponds to a well-conditioned preconditioner; an explicit equation for the loss together with a short proof that it is bounded below would clarify the optimization landscape.
minor comments (2)
  1. [Notation and preliminaries] The notation for the admissible support of G and the projection operator onto that support should be introduced with a single consistent symbol and an equation reference.
  2. [Graph neural network model] The GNN predictor is introduced only briefly; either a short description of the graph construction and training loss or a statement that it is preliminary work would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments highlight opportunities to strengthen the empirical validation, gradient verification, and theoretical presentation of the bimodal loss. We address each major comment below and commit to revisions where appropriate.

read point-by-point responses

  1. Referee: [Numerical experiments] The central empirical claim (fewer iterations than Frobenius FSAI on indefinite saddle-point systems) rests on the projected Krylov support-gradient reliably minimizing the bimodal spectral loss. The manuscript notes the non-convexity but provides no ablation on Lanczos depth, step-size schedules, or multiple random initializations; without such evidence it is unclear whether the reported gains generalize or require problem-specific tuning.

    Authors: We agree that systematic ablations are needed to support robustness claims. In the revised manuscript we will add a dedicated subsection reporting iteration counts for Lanczos depths k=10,20,50; constant versus geometrically decaying step sizes; and results from five independent random initializations on each indefinite saddle-point test case. These experiments will quantify variability and confirm that the observed gains over Frobenius FSAI persist across reasonable parameter choices. revision: yes

  2. Referee: [Gradient derivation (projected Krylov support-gradient)] The detached Rayleigh surrogate is asserted to return a usable matrix-free gradient on the prescribed support. For the bimodal loss, however, the interaction between the stochastic Lanczos trace estimate and the surrogate gradient is not analyzed; a small-scale verification comparing the surrogate gradient norm and direction to a finite-difference reference on a 100-by-100 indefinite test matrix would strengthen the claim that the optimizer is not systematically misled.

    Authors: We accept that an explicit numerical check against finite differences would increase in the surrogate. The revised version will include a new small-scale experiment on a 100-by-100 indefinite matrix: for several admissible-support sizes we will tabulate the relative difference in gradient norms and the cosine of the angle between the detached-surrogate gradient and a central finite-difference reference (with step size 1e-6). This will be presented alongside the existing derivation to demonstrate that the surrogate does not systematically mislead the optimizer on the tested instances. revision: yes

  3. Referee: [Spectral objective for indefinite systems] The bimodal loss is defined to penalize eigenvalues near zero while respecting inertia preservation under congruence. The precise functional form of the penalty term and its weighting relative to the clustering terms are not shown to guarantee that the global minimum corresponds to a well-conditioned preconditioner; an explicit equation for the loss together with a short proof that it is bounded below would clarify the optimization landscape.

    Authors: The explicit bimodal loss (including the penalty term and relative weights) appears in Equation (8). To improve clarity we will restate the full expression in the main text and add a short appendix proving that the loss is bounded from below by zero: each clustering term is a sum of squared deviations and hence non-negative, while the penalty on near-zero eigenvalues is likewise non-negative. We note, however, that a rigorous guarantee that every global minimizer yields a well-conditioned preconditioner remains open because of non-convexity; the boundedness result is the strongest statement we can currently provide. revision: partial

Circularity Check

0 steps flagged

Spectral optimization objective and gradient estimator defined independently of solver metrics

full rationale

The paper defines the spectral loss directly from the eigenvalues of the congruent operator P(G)=GAG^T (bimodal for indefinite case) and derives the projected Krylov support-gradient (Lanczos trace + detached Rayleigh surrogate) as a matrix-free estimator on the fixed support of G. These constructions are stated before any solver iteration counts are mentioned and do not reference the final performance metric as part of the loss or update rule. No self-citation is used to establish uniqueness of the ansatz or to forbid alternatives; the contrast with Frobenius FSAI is external. The GNN predictor is presented as an additional empirical demonstration rather than a load-bearing step in the core derivation. The chain from objective definition through gradient computation to preconditioner construction therefore remains self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the premise that a non-convex spectral loss can be minimized effectively with the proposed first-order method and that the resulting G yields a useful preconditioner. No explicit free parameters are named in the abstract, but the Lanczos depth, loss weights for the bimodal term, and projection details are implicit tuning choices. No new mathematical axioms or invented entities are introduced.

pith-pipeline@v0.9.1-grok · 5808 in / 1468 out tokens · 18597 ms · 2026-06-26T08:06:36.306207+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

87 extracted references · 9 canonical work pages

  1. [1]

    2003 , edition=

    Iterative Methods for Sparse Linear Systems , author=. 2003 , edition=

    2003

  2. [2]

    1997 , publisher=

    Iterative Methods for Solving Linear Systems , author=. 1997 , publisher=

    1997

  3. [3]

    and Van Loan, Charles F

    Golub, Gene H. and Van Loan, Charles F. , title=. 2013 , edition=

    2013

  4. [4]

    Journal of Computational Physics , volume=

    Preconditioning techniques for large linear systems: a survey , author=. Journal of Computational Physics , volume=. 2002 , publisher=

    2002

  5. [5]

    SIAM Journal on Scientific Computing , volume =

    Huang, Ru and Li, Ruipeng and Xi, Yuanzhe , title =. SIAM Journal on Scientific Computing , volume =. 2023 , doi =

    2023

  6. [6]

    SIAM Journal on Scientific Computing , volume =

    Zhao, Shifan and Xu, Tianshi and Huang, Hua and Chow, Edmond and Xi, Yuanzhe , title =. SIAM Journal on Scientific Computing , volume =. 2024 , doi =

    2024

  7. [7]

    SIAM Journal on Scientific Computing , volume =

    Huang, Ru and Chang, Kai and He, Huan and Li, Ruipeng and Xi, Yuanzhe , title =. SIAM Journal on Scientific Computing , volume =. 2024 , doi =

    2024

  8. [8]

    Experimental study of

    Chow, Edmond and Saad, Yousef , journal=. Experimental study of. 1997 , publisher=

    1997

  9. [9]

    1996 , publisher=

    Saad, Yousef , journal=. 1996 , publisher=

    1996

  10. [10]

    Meijerink, J. A. and Van Der Vorst, Henk A. , journal=. An iterative solution method for linear systems of which the coefficient matrix is a symmetric

  11. [11]

    and Plassmann, Paul E

    Jones, Mark T. and Plassmann, Paul E. , journal=. An improved incomplete. 1995 , publisher=

    1995

  12. [12]

    Numerical Linear Algebra with Applications , volume=

    A robust incomplete factorization preconditioner for positive definite matrices , author=. Numerical Linear Algebra with Applications , volume=. 2003 , publisher=

    2003

  13. [13]

    A hierarchical low rank

    Dillon, Geoffrey and Kalantzis, Vassilis and Xi, Yuanzhe and Saad, Yousef , journal=. A hierarchical low rank. 2018 , publisher=

    2018

  14. [14]

    2022 , publisher=

    Xu, Tianshi and Kalantzis, Vassilis and Li, Ruipeng and Xi, Yuanzhe and Dillon, Geoffrey and Saad, Yousef , journal=. 2022 , publisher=

    2022

  15. [15]

    SIAM Journal on Scientific Computing , volume=

    A rational function preconditioner for indefinite sparse linear systems , author=. SIAM Journal on Scientific Computing , volume=. 2017 , publisher=

    2017

  16. [16]

    Conditioning analysis of incomplete

    Napov, Artem , journal=. Conditioning analysis of incomplete. 2013 , publisher=

    2013

  17. [17]

    Duff, I. S. and Erisman, A. M. and Gear, C. W. and Reid, J. K. , title=. SIGNUM Newsletter , volume=. 1988 , pages=

    1988

  18. [18]

    SIAM Journal on Scientific Computing , volume=

    Parallel preconditioning with sparse approximate inverses , author=. SIAM Journal on Scientific Computing , volume=. 1997 , publisher=

    1997

  19. [19]

    Kolotilina, L. Yu. and Yeremin, A. Yu. , journal=. Factorized sparse approximate inverse preconditionings. 1993 , publisher=

    1993

  20. [20]

    SIAM Journal on Scientific Computing , volume=

    A sparse approximate inverse preconditioner for nonsymmetric linear systems , author=. SIAM Journal on Scientific Computing , volume=. 1998 , publisher=

    1998

  21. [21]

    SIAM Journal on Scientific Computing , volume=

    Approximate inverse preconditioners via sparse-sparse iterations , author=. SIAM Journal on Scientific Computing , volume=. 1998 , publisher=

    1998

  22. [22]

    Parallel Computing , volume=

    Incomplete sparse approximate inverses for parallel preconditioning , author=. Parallel Computing , volume=. 2018 , publisher=

    2018

  23. [23]

    Fine-grained parallel incomplete

    Chow, Edmond and Patel, Aftab , journal=. Fine-grained parallel incomplete. 2015 , publisher=

    2015

  24. [24]

    Preconditioning

    Osei-Kuffuor, Daniel and Saad, Yousef , journal=. Preconditioning. 2010 , publisher=

    2010

  25. [25]

    2021 , doi=

    Anderson, Robert and Andrej, Julian and Barker, Andrew and Bramwell, Jamie and Camier, Jean-Sylvain and Cerveny, Jakub and Dobrev, Veselin and Dudouit, Yohann and Fisher, Aaron and Kolev, Tzanio and Pazner, Will and Stowell, Mark and Tomov, Vladimir and Akkerman, Ido and Dahm, Johann and Medina, David and Zampini, Stefano , journal=. 2021 , doi=

    2021

  26. [26]

    Transactions on Machine Learning Research , year=

    Neural incomplete factorization: learning preconditioners for the conjugate gradient method , author=. Transactions on Machine Learning Research , year=

  27. [28]

    Learning Sparse Approximate Inverse Preconditioners for Conjugate Gradient Solvers on

    Yang, Zherui and Li, Zichao and Lyu, Kun and Li, Yanbo and Du, Tao and Liu, Ligang , booktitle=. Learning Sparse Approximate Inverse Preconditioners for Conjugate Gradient Solvers on. 2025 , note=

    2025

  28. [30]

    Learning preconditioners for conjugate gradient

    Li, Yichen and Chen, Peter Yichen and Du, Tao and Matusik, Wojciech , booktitle=. Learning preconditioners for conjugate gradient. 2023 , organization=

    2023

  29. [31]

    International Conference on Machine Learning , pages=

    Learning Algebraic Multigrid Using Graph Neural Networks , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  30. [32]

    arXiv preprint arXiv:2405.15557 , year=

    Learning from linear algebra: a graph neural network approach to preconditioner design for conjugate gradient solvers , author=. arXiv preprint arXiv:2405.15557 , year=

    arXiv

  31. [33]

    Message-Passing

    Trifonov, Vladislav and Muravleva, Ekaterina and Oseledets, Ivan , journal=. Message-Passing

  32. [34]

    SIAM Journal on Scientific Computing , volume=

    Kopani. SIAM Journal on Scientific Computing , volume=. 2024 , publisher=

    2024

  33. [35]

    SIAM Journal on Scientific Computing , volume=

    Fast computation of spectral densities for generalized eigenvalue problems , author=. SIAM Journal on Scientific Computing , volume=. 2018 , publisher=

    2018

  34. [36]

    SIAM Review , volume=

    Approximating spectral densities of large matrices , author=. SIAM Review , volume=. 2016 , publisher=

    2016

  35. [37]

    Silver, Richard N. and R. Densities of states of mega-dimensional. International Journal of Modern Physics C , volume=. 1994 , publisher=

    1994

  36. [38]

    Journal of Computational Physics , volume=

    Kernel polynomial approximations for densities of states and spectral functions , author=. Journal of Computational Physics , volume=. 1996 , publisher=

    1996

  37. [39]

    Silver, R. N. and R. Calculation of densities of states and spectral functions by. Physical Review E , volume=. 1997 , publisher=

    1997

  38. [40]

    Physical Review B , volume=

    Calculating the density of states and optical-absorption spectra of large quantum systems by the plane-wave moments method , author=. Physical Review B , volume=. 1994 , publisher=

    1994

  39. [41]

    Numerische Mathematik , volume=

    Randomized estimation of spectral densities of large matrices made accurate , author=. Numerische Mathematik , volume=. 2017 , publisher=

    2017

  40. [42]

    , journal=

    Hutchinson, Michael F. , journal=. A stochastic estimator of the trace of the influence matrix for. 1989 , publisher=

    1989

  41. [43]

    Journal of the ACM , volume=

    Randomized algorithms for estimating the trace of an implicit symmetric positive semi-definite matrix , author=. Journal of the ACM , volume=. 2011 , publisher=

    2011

  42. [44]

    International Conference on Learning Representations , year=

    Semi-supervised classification with graph convolutional networks , author=. International Conference on Learning Representations , year=

  43. [45]

    Current Opinion in Structural Biology , volume=

    Everything is connected: Graph neural networks , author=. Current Opinion in Structural Biology , volume=. 2023 , publisher=

    2023

  44. [46]

    2017 , edition=

    Graph Theory , author=. 2017 , edition=

    2017

  45. [47]

    The use of linear graphs in

    Parter, Seymour , journal=. The use of linear graphs in. 1961 , publisher=

    1961

  46. [48]

    and Chelikowsky, James R

    Zhou, Yunkai and Saad, Yousef and Tiago, Murilo L. and Chelikowsky, James R. , journal=. Self-consistent-field calculations using. 2006 , publisher=

    2006

  47. [49]

    arXiv preprint arXiv:2501.10750 , year=

    Millard, David and Carr, Arielle and Gaudreault, St. arXiv preprint arXiv:2501.10750 , year=

    arXiv

  48. [50]

    Any nonincreasing convergence curve is possible for

    Greenbaum, Anne and Pt. Any nonincreasing convergence curve is possible for. SIAM Journal on Matrix Analysis and Applications , volume=. 1996 , publisher=

    1996

  49. [51]

    Multigrid Methods , editor=

    Algebraic multigrid , author=. Multigrid Methods , editor=

  50. [52]

    Reviews of Modern Physics , volume=

    The kernel polynomial method , author=. Reviews of Modern Physics , volume=. 2006 , doi=

    2006

  51. [53]

    On the convergence of bounded

    Beckermann, Bernhard , journal=. On the convergence of bounded. 2000 , publisher=

    2000

  52. [54]

    International Conference on Learning Representations (ICLR) , year=

    What Can Neural Networks Reason About? , author=. International Conference on Learning Representations (ICLR) , year=

  53. [55]

    International Conference on Learning Representations (ICLR) , year=

    How Powerful are Graph Neural Networks? , author=. International Conference on Learning Representations (ICLR) , year=

  54. [56]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Can graph neural networks count substructures? , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  55. [57]

    Davis, Chandler and Kahan, W. M. , journal=. The Rotation of Eigenvectors by a Perturbation. 1970 , doi=

    1970

  56. [58]

    Mathematics of Operations Research , volume=

    Derivatives of Spectral Functions , author=. Mathematics of Operations Research , volume=. 1996 , doi=

    1996

  57. [59]

    Physical Review B , volume=

    Automatic Differentiation of Dominant Eigensolver and Its Applications in Quantum Physics , author=. Physical Review B , volume=. 2020 , doi=

    2020

  58. [60]

    Advances in Neural Information Processing Systems , volume=

    Gradients of Functions of Large Matrices , author=. Advances in Neural Information Processing Systems , volume=. 2024 , doi=

    2024

  59. [61]

    Fast Estimation of

    Ubaru, Shashanka and Chen, Jie and Saad, Yousef , journal=. Fast Estimation of. 2017 , doi=

    2017

  60. [63]

    Vyas, Nikhil and Morwani, Depen and Zhao, Rosie and Shapira, Itai and Brandfonbrener, David and Janson, Lucas and Kakade, Sham , journal=

  61. [64]

    Scott and Tianshi Xu and Ziyuan Tang and Alexandra Pichette-Emmons and Qiang Ye and Yousef Saad and Yuanzhe Xi , journal=

    Mitchell T. Scott and Tianshi Xu and Ziyuan Tang and Alexandra Pichette-Emmons and Qiang Ye and Yousef Saad and Yuanzhe Xi , journal=. Design Criteria for. 2026 , url=

    2026

  62. [65]

    Mfem: A modular finite element methods library.Computers & Mathematics with Appli- cations, 81:42–74, January 2021

    R. Anderson, J. Andrej, A. Barker, J. Bramwell, J.-S. Camier, J. Cerveny, V. Dobrev, Y. Dudouit, A. Fisher, T. Kolev, W. Pazner, M. Stowell, V. Tomov, I. Akkerman, J. Dahm, D. Medina, and S. Zampini , MFEM : A modular finite element methods library , Computers & Mathematics with Applications, 81 (2021), pp. 42--74, https://doi.org/10.1016/j.camwa.2020.06.009

    work page doi:10.1016/j.camwa.2020.06.009 2021

  63. [66]

    H. Anzt, T. K. Huckle, J. Br \"a ckle, and J. Dongarra , Incomplete sparse approximate inverses for parallel preconditioning , Parallel Computing, 71 (2018), pp. 1--22, https://doi.org/10.1016/j.parco.2017.10.003

    work page doi:10.1016/j.parco.2017.10.003 2018

  64. [67]

    P. W. Battaglia, J. B. Hamrick, V. Bapst, A. Sanchez-Gonzalez, V. Zambaldi, M. Malinowski, A. Tacchetti, D. Raposo, A. Santoro, R. Faulkner, et al. , Relational inductive biases, deep learning, and graph networks , arXiv preprint arXiv:1806.01261, (2018)

    Pith/arXiv arXiv 2018

  65. [68]

    Benzi , Preconditioning techniques for large linear systems: a survey , Journal of Computational Physics, 182 (2002), pp

    M. Benzi , Preconditioning techniques for large linear systems: a survey , Journal of Computational Physics, 182 (2002), pp. 418--477

    2002

  66. [69]

    Benzi and M

    M. Benzi and M. T u ma , A sparse approximate inverse preconditioner for nonsymmetric linear systems , SIAM Journal on Scientific Computing, 19 (1998), pp. 968--994, https://doi.org/10.1137/S1064827595294691

    work page doi:10.1137/s1064827595294691 1998

  67. [70]

    Chen , Graph neural preconditioners for iterative solutions of sparse linear systems , arXiv preprint arXiv:2406.00809, (2024)

    J. Chen , Graph neural preconditioners for iterative solutions of sparse linear systems , arXiv preprint arXiv:2406.00809, (2024)

    arXiv 2024

  68. [71]

    Chow and Y

    E. Chow and Y. Saad , Approximate inverse preconditioners via sparse-sparse iterations , SIAM Journal on Scientific Computing, 19 (1998), pp. 995--1023

    1998

  69. [72]

    M. J. Grote and T. Huckle , Parallel preconditioning with sparse approximate inverses , SIAM Journal on Scientific Computing, 18 (1997), pp. 838--853, https://doi.org/10.1137/S1064827594276552

    work page doi:10.1137/s1064827594276552 1997

  70. [73]

    a usner, A. N. Juscafresa, and J. Sj \

    P. H \"a usner, A. N. Juscafresa, and J. Sj \"o lund , Learning incomplete factorization preconditioners for GMRES , arXiv preprint arXiv:2409.08262, (2024)

    arXiv 2024

  71. [74]

    a usner, O. \

    P. H \"a usner, O. \"O ktem, and J. Sj \"o lund , Neural incomplete factorization: learning preconditioners for the conjugate gradient method , Transactions on Machine Learning Research, (2023). arXiv:2305.16368

    arXiv 2023

  72. [75]

    Huang, K

    R. Huang, K. Chang, H. He, R. Li, and Y. Xi , Reducing operator complexity of galerkin coarse-grid operators with machine learning , SIAM Journal on Scientific Computing, 46 (2024), pp. S296--S316, https://doi.org/10.1137/23M1583533

    work page doi:10.1137/23m1583533 2024

  73. [76]

    Huang, R

    R. Huang, R. Li, and Y. Xi , Learning optimal multigrid smoothers via neural networks , SIAM Journal on Scientific Computing, 45 (2023), pp. S199--S225, https://doi.org/10.1137/21M1430030

    work page doi:10.1137/21m1430030 2023

  74. [77]

    L. Y. Kolotilina and A. Y. Yeremin , Factorized sparse approximate inverse preconditionings I . Theory , SIAM Journal on Matrix Analysis and Applications, 14 (1993), pp. 45--58

    1993

  75. [78]

    Kopani c \'a kov \'a and G

    A. Kopani c \'a kov \'a and G. E. Karniadakis , DeepONet based preconditioning strategies for solving parametric linear systems of equations , SIAM Journal on Scientific Computing, 46 (2024), pp. A2679--A2703, https://doi.org/10.1137/24M162861X

    work page doi:10.1137/24m162861x 2024

  76. [79]

    Y. Li, P. Y. Chen, T. Du, and W. Matusik , Learning preconditioners for conjugate gradient PDE solvers , in International Conference on Machine Learning, PMLR, 2023, pp. 19425--19439

    2023

  77. [80]

    L. Lin, Y. Saad, and C. Yang , Approximating spectral densities of large matrices , SIAM Review, 58 (2016), pp. 34--65

    2016

  78. [81]

    I. Luz, M. Galun, H. Maron, R. Basri, and I. Yavneh , Learning algebraic multigrid using graph neural networks , in International Conference on Machine Learning, PMLR, 2020, pp. 6489--6499

    2020

  79. [82]

    Parter , The use of linear graphs in Gauss elimination , SIAM Review, 3 (1961), pp

    S. Parter , The use of linear graphs in Gauss elimination , SIAM Review, 3 (1961), pp. 119--130

    1961

  80. [83]

    Saad , Iterative Methods for Sparse Linear Systems , SIAM, Philadelphia, 2nd ed., 2003

    Y. Saad , Iterative Methods for Sparse Linear Systems , SIAM, Philadelphia, 2nd ed., 2003

    2003

Showing first 80 references.