Pith. sign in

REVIEW 2 major objections 6 minor 41 references

Sketch low-rank dynamics: orthogonal vs. oblique projections

T0 review · 2 major / 6 minor · reviewed 2026-07-12 · grok-4.5

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

Pith's one-line read Sketching the Galerkin condition in dynamical low-rank approximation produces a different ODE that fails when the perpendicular residual is large; evolving sketch-orthogonal bases under the classical orthogonal projector keeps the original

desk verdict Clean negative theorem plus a structure-preserving sketch DLRA that actually works; the math is elementary but the distinction is useful and the experiments match the theory. read the letter →

arxiv 2607.03402 v1 pith:GJ42WUOZ submitted 2026-07-03 math.NA cs.NA

Benjamin Carrel , Laura Grigori This is my paper · ORCID
classification math.NAcs.NA MSC 65F5565L0565M9915A23
keywords dynamicallow-rankapproximationsketchingobliqueprojectionorthogonalrandomizedGram–Schmidtprojector-splittingBUGintegratorVlasov–Poisson
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

Dynamical low-rank approximation compresses large matrix ODEs by projecting the vector field onto the tangent space of fixed-rank matrices. This paper asks how randomized sketching should enter that construction. Sketching the Galerkin condition itself yields an oblique projection whose dynamics coincide with classical DLRA only when the sketched perpendicular residual vanishes against every tangent direction. On problems where that residual is large, the oblique scheme drifts from the classical trajectory and loses invariants such as energy and mass. The paper therefore proposes an orthogonal alternative: keep the classical orthogonal projector for the dynamics and use sketching only to orthogonalize the bases via randomized Gram–Schmidt (or randomized Cholesky QR when the bases are well conditioned). The resulting integrators inherit classical accuracy and conservation, while the basis update needs fewer global synchronizations on row-distributed data and can run as BLAS-3 kernels on accelerators. Sketch versions of the projector-splitting and BUG integrators are derived and tested on Allen–Cahn, Fokker–Planck, and Vlasov–Poisson.

What carries the argument

The bijective map φ that converts between sketch-orthogonal factorizations (P,S,W) and classical orthogonal factorizations (U,Σ,V), together with the equivalence condition of Proposition 4.4 that isolates when the oblique residual vanishes.

What would settle it

On a problem with large perpendicular residual such as the Vlasov–Poisson two-stream instability, measure whether the orthogonal sketch integrators track the classical electric-energy trajectory and conserve mass and total energy, while the oblique integrators drift and lose those invariants; any orthogonal run that drifts would falsify the claim that the orthogonal scheme preserves classical dynamics.

Watch

Extended reading notes

Core claim

The oblique sketch DLRA obtained by sketching the Galerkin condition is equivalent to classical DLRA if and only if the sketched perpendicular residual is orthogonal to every tangent direction; otherwise it solves a genuinely perturbed ODE. Evolving sketch-orthogonal bases under the standard orthogonal projector coincides with classical DLRA for every vector field, inherits its error bound, and preserves its conservation properties.

Load-bearing premise

The sketches must stay injective on the column and row spaces of every rank-r matrix that appears during the integration; if they become rank-deficient on those evolving spaces, both the conversion map and the oblique formulas break.

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

2 major / 6 minor

Summary. The paper studies two ways of incorporating randomized sketching into dynamical low-rank approximation (DLRA) of matrix ODEs. Sketching the Galerkin condition yields an oblique tangent-space projection; the authors prove (Proposition 4.4, residual identity (15)) that this coincides with classical DLRA if and only if the sketched perpendicular residual vanishes against every tangent direction, and that it otherwise solves a perturbed ODE. As an alternative they introduce an orthogonal sketch DLRA that evolves sketch-orthogonal bases (via randomized Gram–Schmidt / Cholesky QR) while retaining the classical orthogonal projector, so the dynamics and error bounds of Koch–Lubich are preserved. Sketch versions of the projector-splitting and BUG integrators are derived, communication costs are analyzed, and the theory is illustrated on Allen–Cahn, Fokker–Planck, and Vlasov–Poisson, with the orthogonal scheme matching classical DLRA and the oblique scheme failing once the perpendicular residual is large.

Significance. The orthogonal-versus-oblique distinction is a clear, load-bearing contribution for the DLRA community: it shows that the most natural sketching of the Galerkin condition is not structure-preserving in general, and supplies a simple alternative that is. The continuous analysis (bijective map φ, Proposition 3.1, Lemmas 4.1–4.2, Proposition 4.4) is self-contained and elementary, the residual diagnosis on Vlasov–Poisson is falsifiable and confirmed numerically, and the implementation is public. The communication analysis of RGS versus Householder QR is a useful secondary contribution for large-scale settings even though wall-clock speedups are not measured here. If the results hold as stated, the paper gives a practical design rule—prefer orthogonal sketch DLRA unless the problem is known to be approximately low-rank compatible—and a clean geometric explanation of when sketching the variational principle is safe.

major comments (2)
  1. Abstract and §5.2 emphasize fewer global synchronizations and a BLAS-2→BLAS-3 shift as a computational advantage of the sketch basis update, but §6 explicitly validates only accuracy on n=64 single-CPU runs and disclaims wall-clock measurement. The mathematical comparison of orthogonal vs. oblique projections does not depend on those claims, yet they are foregrounded as part of the contribution. Either add at least one large-scale timing/communication experiment (or a clear multi-GPU/distributed sketch of the sBUG pipeline), or rephrase the abstract and introduction so that the hardware advantage is presented strictly as an asymptotic/communication analysis whose empirical verification is left to future work.
  2. Remark 4.6 correctly notes that the approximate-(η,r,k)-LRC residual is not controlled by an OSE alone and that no trajectory error bound is proved. The abstract and Table 1 nevertheless state that the oblique approach “approximately reproduces” classical DLRA on LRC problems. On Allen–Cahn and Fokker–Planck the numerics support this, but the theoretical claim should be tightened to match what is proved (exact equivalence iff (14); residual of order η only heuristically) so that the abstract does not overstate the LRC regime.
minor comments (6)
  1. §5 and Remark 4.5 discuss drawing a fresh (Θ,Ω) at each macro step and a union bound over N steps, while §6 holds a single pair fixed over the whole trajectory. A short sentence reconciling the two settings (and stating that the fixed-sketch experiments still probe the pointwise theory of §4) would avoid confusion.
  2. Assumption 1 and the domain of φ require injectivity of Θ,Ω on the column/row spaces of every rank-r iterate. A brief practical note on what the integrator should do if RGS/rCholQR detects rank deficiency (re-draw, increase ℓ, or fall back to Householder) would make the discrete method more complete.
  3. Figure 1 and Figure 4 both show electric energy; cross-referencing them and aligning axis scales/labels would help the reader connect the motivating example to the full Vlasov–Poisson study.
  4. In §2.3 the cost of RGS is stated as mr² flops with ℓ=O(r); a one-line comparison table (Householder / MGS / RGS / rCholQR: flops, synchronizations, BLAS level) would make §5.2 easier to scan.
  5. Typographical: “July 2026” on the title page; “Sonnendr¨ ucker” spacing; occasional missing spaces before citations. Also clarify once that (ΘP)+ under sketch-orthogonality equals (ΘP)T (already in Remark 2.6) wherever the oblique formulas first appear in Table 3.
  6. Related work: a short pointer to how the present basis-update sketching differs from the fully randomized range-finding integrators of Carrel (2025) and the randomized Runge–Kutta methods of Lam et al. (2025)—already mentioned briefly—would help position the contribution for readers outside the DLRA community.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: central equivalence (Prop. 4.4) and orthogonal construction are self-contained manifold/OSE arguments, not forced by fit or self-cited identity.

full rationale

The paper’s load-bearing claims are definitional constructions and elementary derivations, not predictions that collapse to their inputs. Orthogonal sketch DLRA is defined to use the classical orthogonal projector (7)/(8) on sketch-orthogonal factors related by the bijective map φ; coincidence with classical DLRA is therefore by construction of the method, not a circular “prediction.” Proposition 4.4 derives the iff condition for the oblique scheme from the sketched Galerkin condition (10), the tangent-space decomposition, and the two-sided OSE embedding (Corollary 2.3)—standard linear-algebra steps that do not assume the conclusion. Self-citations (Carrel 2025; Balabanov–Grigori RGS; Carrel et al. integrators) supply background tools and complementary methods; none is invoked as a uniqueness theorem or as the sole justification of the equivalence/non-equivalence statements. There is no parameter fit re-labeled as prediction, no ansatz smuggled via citation, and no renaming of a known empirical law. Numerical sections illustrate rather than close the argument. Score 1 only for ordinary non-load-bearing self-citation of coauthors’ prior RNLA/DLRA work.

Assumptions & free parameters 1 free parameters · 4 assumptions · 2 invented entities

The work rests on standard manifold geometry of fixed-rank matrices, the classical DLRA Galerkin condition, and the definition of oblivious subspace embeddings. The only paper-specific modeling choices are the sketch size ℓ=O(r), the non-degeneracy Assumption 1, and the approximate low-rank-compatibility notion (Definition 2.4). No free parameters are fitted to data; the numerical experiments use fixed, stated values.

free parameters (1)
  • sketch size ℓ = ℓ=2r in all reported experiments
    Chosen by hand as ℓ=2r (or [2r,4r] for large-scale) following Balabanov–Grigori heuristics; not fitted to the PDE data but still a free algorithmic parameter that controls both cost and embedding quality.
assumptions (4)
  • domain assumption Classical DLRA is defined by the Dirac–Frenkel / Galerkin condition on the tangent space of the fixed-rank manifold (Koch–Lubich 2007).
    Invoked throughout Sections 2–4 as the reference dynamics that the sketch variants are compared against.
  • standard math An oblivious subspace embedding OSE(k,ε,δ) approximately preserves inner products on every k-dimensional subspace with high probability (Sarlós, Halko–Martinsson–Tropp).
    Used to control the two-sided sketching error in Corollary 2.3 and Proposition 4.4.
  • ad hoc to paper Assumption 1: the sketches Θ,Ω remain injective on the column and row spaces of every rank-r iterate that arises.
    Stated explicitly so that sketch-orthogonal factorizations and left pseudo-inverses are well-defined; conditioned on the favorable OSE event.
  • ad hoc to paper Definition 2.4 of (η,r,k)-low-rank compatibility of the vector field.
    Introduced to classify when the oblique residual is expected to be small; used in Section 4.2 and the numerical discussion.
invented entities (2)
  • sketch Stiefel manifold St_Θ(m,r) and the associated manifold M_Θ,Ω_r of sketch-orthogonal factorizations
    purpose: Provide a geometric home for bases that are orthonormal only after sketching, so that RGS can be used while the classical orthogonal projector is retained.
    Defined in Section 2.3 / 3.1; the bijective map φ relates them to the ordinary Stiefel manifold. No independent physical existence claimed.
  • oblique tangent-space projector P_Θ,Ω_Y
    purpose: Realize the sketched Galerkin condition as a full-space projection so that its deviation from the classical projector can be analyzed.
    Derived in Lemma 4.2 from the sketched velocities; the paper’s main negative result is that this projector differs from P_Y when the perpendicular residual is large.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sketch low-rank dynamics: orthogonal vs. oblique projections." pith.science (2026). https://pith.science/paper/GJ42WUOZ

@misc{pith2026260703402,
  author       = {Pith},
  title        = {Pith review of: Sketch low-rank dynamics: orthogonal vs. oblique projections},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GJ42WUOZ}},
  note         = {Machine review of arXiv:2607.03402}
}
read the original abstract

We study how sketching techniques from randomized numerical linear algebra can be incorporated into the dynamical low-rank approximation (DLRA) of large-scale matrix differential equations. A natural approach is to sketch the Galerkin condition that defines the DLRA, which leads to an oblique tangent space projection. We show that this oblique projection approximately reproduces the standard DLRA only under restrictive conditions on the vector field, and that it fails on problems with a large perpendicular residual. As an alternative, we propose an orthogonal sketch DLRA that evolves sketch-orthogonal bases while using standard orthogonal projections for the dynamics. This approach preserves the geometric structure of the classical DLRA and is numerically stable. The computational advantage of randomized Gram--Schmidt over Householder QR lies in fewer global synchronizations on a row-distributed basis, at a comparable flop count; when the basis is well conditioned, randomized Gram--Schmidt can be replaced by randomized Cholesky QR, which additionally shifts the basis update from BLAS-2 to BLAS-3 kernels, making it well-suited to modern accelerators. We derive sketch versions of the projector-splitting and BUG integrators, and demonstrate the approach on the Allen--Cahn, Fokker--Planck, and Vlasov--Poisson equations.

Figures

Figures reproduced from arXiv: 2607.03402 by the authors.

Figure 1
Figure 1. Motivating example: electric energy on the two-stream instability (Vlasov–Poisson equation ( [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Allen–Cahn equation (17) with n = 64, r = 8, ε = 0.01, D = [0, 2π] 2 , sketch size ℓ = 2r = 16. Each sketch curve is the median over 20 Gaussian sketches; the shaded band is the interquartile range. The solid gray curve is the unsketched BUG baseline and the black dotted curve is the best rank-r SVD truncation of the reference. Left: orthogonal sketch variants. Right: oblique sketch variants. 6.2 Fokker–Planck equat… view at source ↗
Figure 3
Figure 3. Fokker–Planck equation (19) with n = 64, r = 8, σ = 2, D = [0, 2π] 2 , sketch size ℓ = 2r = 16. Each sketch curve is the median over 20 Gaussian sketches; the shaded band is the interquartile range. The solid gray curve is the unsketched BUG baseline and the black dotted curve is the best rank-r SVD truncation of the reference. Left: orthogonal sketch variants. Right: oblique sketch variants. 2Expanding µ1(x1, x2) =… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Vlasov–Poisson equation (21) (two-stream instability) with nx = nv = 64, r = 8, sketch size ℓ = 2r = 16, and t ∈ [0, 60]. Each sketch curve is the median over 5 Gaussian sketches; the shaded band is the interquartile range. Left: relative Frobenius error. Right: electr…
Figure 5
Figure 5. Figure 5: Conservation laws on the Vlasov–Poisson two-stream instability, same parameters as Figure [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 3 linked inside Pith

  1. [1]

    Optimization algorithms on matrix manifolds

    P-A Absil, Robert Mahony, and Rodolphe Sepulchre. Optimization algorithms on matrix manifolds. Princeton University Press, 2008

    2008

  2. [2]

    Approximate nearest neighbors and the fast Johnson--Lindenstrauss transform

    Nir Ailon and Bernard Chazelle. Approximate nearest neighbors and the fast Johnson--Lindenstrauss transform. In Proceedings of the thirty-eighth annual ACM symposium on Theory of computing, pages 557--563, 2006

    2006

  3. [3]

    A microscopic theory for antiphase boundary motion and its application to antiphase domain coarsening

    Samuel M Allen and John W Cahn. A microscopic theory for antiphase boundary motion and its application to antiphase domain coarsening. Acta metallurgica, 27 0 (6): 0 1085--1095, 1979

    1979

  4. [4]

    Randomized Cholesky QR factorizations

    Oleg Balabanov. Randomized Cholesky QR factorizations. arXiv preprint arXiv:2210.09953, 2022

    arXiv 2022

  5. [5]

    Randomized Gram--Schmidt process with application to GMRES

    Oleg Balabanov and Laura Grigori. Randomized Gram--Schmidt process with application to GMRES . SIAM Journal on Scientific Computing, 44 0 (3): 0 A1450--A1474, 2022

    2022

  6. [6]

    Randomized block Gram--Schmidt process for the solution of linear systems and eigenvalue problems

    Oleg Balabanov and Laura Grigori. Randomized block Gram--Schmidt process for the solution of linear systems and eigenvalue problems. SIAM Journal on Scientific Computing, 47 0 (1): 0 A553--A585, 2025

    2025

  7. [7]

    Randomized linear algebra for model reduction

    Oleg Balabanov and Anthony Nouy. Randomized linear algebra for model reduction. Part I : Galerkin methods and error estimation. Advances in Computational Mathematics, 45 0 (5): 0 2969--3019, 2019

    2019

  8. [8]

    Randomized linear algebra for model reduction

    Oleg Balabanov and Anthony Nouy. Randomized linear algebra for model reduction. Part II : minimal residual methods and dictionary-based approximation. Advances in Computational Mathematics, 47 0 (2): 0 26, 2021

    2021

Show all 41 references
  1. [9]

    Randomized methods for dynamical low-rank approximation

    Benjamin Carrel. Randomized methods for dynamical low-rank approximation. Journal of Computational Physics, page 114421, 2025

    2025

  2. [10]

    Projected exponential methods for stiff dynamical low-rank approximation problems

    Benjamin Carrel and Bart Vandereycken. Projected exponential methods for stiff dynamical low-rank approximation problems. arXiv preprint arXiv:2312.00172, 2023

    2023 arXiv

  3. [11]

    Low-rank Parareal : a low-rank parallel-in-time integrator

    Benjamin Carrel, Martin J Gander, and Bart Vandereycken. Low-rank Parareal : a low-rank parallel-in-time integrator. BIT Numerical Mathematics, 63 0 (1): 0 13, 2023

    2023

  4. [12]

    Interpolatory dynamical low-rank approximation: Theoretical foundations and algorithms

    Benjamin Carrel, Daniel Kressner, Hei Yin Lam, and Bart Vandereycken. Interpolatory dynamical low-rank approximation: Theoretical foundations and algorithms. arXiv preprint arXiv:2510.19518, 2025

    2025

  5. [13]

    An unconventional robust integrator for dynamical low-rank approximation

    Gianluca Ceruti and Christian Lubich. An unconventional robust integrator for dynamical low-rank approximation. BIT Numerical Mathematics, 62 0 (1): 0 23--44, 2022

    2022

  6. [14]

    A rank-adaptive robust integrator for dynamical low-rank approximation

    Gianluca Ceruti, Jonas Kusch, and Christian Lubich. A rank-adaptive robust integrator for dynamical low-rank approximation. BIT Numerical Mathematics, 62 0 (4): 0 1149--1174, 2022

    2022

  7. [15]

    A robust second-order low-rank BUG integrator based on the midpoint rule

    Gianluca Ceruti, Lukas Einkemmer, Jonas Kusch, and Christian Lubich. A robust second-order low-rank BUG integrator based on the midpoint rule. BIT Numerical Mathematics, 64 0 (3): 0 30, 2024

    2024

  8. [16]

    The integration of the Vlasov equation in configuration space

    Chio-Zong Cheng and Georg Knorr. The integration of the Vlasov equation in configuration space. Journal of Computational Physics, 22 0 (3): 0 330--351, 1976

    1976

  9. [17]

    Randomized implicitly restarted Arnoldi method for the non-symmetric eigenvalue problem

    Jean-Guillaume De Damas and Laura Grigori. Randomized implicitly restarted Arnoldi method for the non-symmetric eigenvalue problem. SIAM Journal on Matrix Analysis and Applications, 46 0 (4): 0 2395--2422, 2025

    2025

  10. [18]

    Randomized orthogonalization and Krylov subspace methods: principles and algorithms

    Jean-Guillaume de Damas, Laura Grigori, Igor Simunec, and Edouard Timsit. Randomized orthogonalization and Krylov subspace methods: principles and algorithms. arXiv preprint arXiv:2512.15455, 2025

    2025

  11. [19]

    Rank-adaptive tensor methods for high-dimensional nonlinear PDEs

    Alec Dektor, Abram Rodgers, and Daniele Venturi. Rank-adaptive tensor methods for high-dimensional nonlinear PDEs . Journal of Scientific Computing, 88 0 (2): 0 36, 2021

    2021

  12. [20]

    Communication-optimal parallel and sequential QR and LU factorizations

    James Demmel, Laura Grigori, Mark Hoemmen, and Julien Langou. Communication-optimal parallel and sequential QR and LU factorizations. SIAM Journal on Scientific Computing, 34 0 (1): 0 A206--A239, 2012

    2012

  13. [21]

    A low-rank algorithm for weakly compressible flow

    Lukas Einkemmer. A low-rank algorithm for weakly compressible flow. SIAM Journal on Scientific Computing, 41 0 (5): 0 A2795--A2814, 2019

    2019

  14. [22]

    A mass, momentum, and energy conservative dynamical low-rank scheme for the Vlasov equation

    Lukas Einkemmer and Ilon Joseph. A mass, momentum, and energy conservative dynamical low-rank scheme for the Vlasov equation. Journal of Computational Physics, 443: 0 110495, 2021

    2021

  15. [23]

    A low-rank projector-splitting integrator for the Vlasov--Poisson equation

    Lukas Einkemmer and Christian Lubich. A low-rank projector-splitting integrator for the Vlasov--Poisson equation. SIAM Journal on Scientific Computing, 40 0 (5): 0 B1330--B1360, 2018

    2018

  16. [24]

    A quasi-conservative dynamical low-rank algorithm for the Vlasov equation

    Lukas Einkemmer and Christian Lubich. A quasi-conservative dynamical low-rank algorithm for the Vlasov equation. SIAM Journal on Scientific Computing, 41 0 (5): 0 B1061--B1081, 2019

    2019

  17. [25]

    Comparison of Eulerian Vlasov solvers

    Francis Filbet and Eric Sonnendr \"u cker. Comparison of Eulerian Vlasov solvers. Computer Physics Communications, 150 0 (3): 0 247--266, 2003

    2003

  18. [26]

    Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions

    Nathan Halko, Per-Gunnar Martinsson, and Joel A Tropp. Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions. SIAM Review, 53 0 (2): 0 217--288, 2011

    2011

  19. [27]

    Analysis of randomized Householder--Cholesky QR factorization with multisketching

    Andrew J Higgins, Daniel B Szyld, Erik G Boman, and Ichitaro Yamazaki. Analysis of randomized Householder--Cholesky QR factorization with multisketching. Numerische Mathematik, 157 0 (5): 0 1695--1737, 2025

    2025

  20. [28]

    Projection methods for dynamical low-rank approximation of high-dimensional problems

    Emil Kieri and Bart Vandereycken. Projection methods for dynamical low-rank approximation of high-dimensional problems. Computational Methods in Applied Mathematics, 19 0 (1): 0 73--92, 2019

    2019

  21. [29]

    Discretized dynamical low-rank approximation in the presence of small singular values

    Emil Kieri, Christian Lubich, and Hanna Walach. Discretized dynamical low-rank approximation in the presence of small singular values. SIAM Journal on Numerical Analysis, 54 0 (2): 0 1020--1038, 2016

    2016

  22. [30]

    Dynamical low-rank approximation

    Othmar Koch and Christian Lubich. Dynamical low-rank approximation. SIAM Journal on Matrix Analysis and Applications, 29 0 (2): 0 434--454, 2007

    2007

  23. [31]

    Randomized low-rank Runge--Kutta methods

    Hei Yin Lam, Gianluca Ceruti, and Daniel Kressner. Randomized low-rank Runge--Kutta methods. SIAM Journal on Matrix Analysis and Applications, 46 0 (2): 0 1587--1615, 2025

    2025

  24. [32]

    A projector-splitting integrator for dynamical low-rank approximation

    Christian Lubich and Ivan V Oseledets. A projector-splitting integrator for dynamical low-rank approximation. BIT Numerical Mathematics, 54 0 (1): 0 171--188, 2014

    2014

  25. [33]

    Randomized numerical linear algebra: Foundations and algorithms

    Per-Gunnar Martinsson and Joel A Tropp. Randomized numerical linear algebra: Foundations and algorithms. Acta Numerica, 29: 0 403--572, 2020

    2020

  26. [34]

    Randomized numerical linear algebra: A perspective on the field with an eye to software

    Riley Murray, James Demmel, Michael W Mahoney, N Benjamin Erichson, Maksim Melnichenko, Osman Asif Malik, Laura Grigori, Piotr Luszczek, Micha Derezi \'n ski, Miles E Lopes, et al. Randomized numerical linear algebra: A perspective on the field with an eye to software. arXiv p...

    2023 arXiv

  27. [35]

    Error analysis of the dynamically orthogonal approximation of time dependent random PDEs

    Eleonora Musharbash, Fabio Nobile, and Tao Zhou. Error analysis of the dynamically orthogonal approximation of time dependent random PDEs . SIAM Journal on Scientific Computing, 37 0 (2): 0 A776--A810, 2015

    2015

  28. [36]

    Fast and accurate randomized algorithms for linear systems and eigenvalue problems

    Yuji Nakatsukasa and Joel A Tropp. Fast and accurate randomized algorithms for linear systems and eigenvalue problems. SIAM Journal on Matrix Analysis and Applications, 45 0 (2): 0 1183--1214, 2024

    2024

  29. [37]

    Dynamically orthogonal field equations for continuous stochastic dynamical systems

    Themistoklis P Sapsis and Pierre FJ Lermusiaux. Dynamically orthogonal field equations for continuous stochastic dynamical systems. Physica D: Nonlinear Phenomena, 238 0 (23-24): 0 2347--2360, 2009

    2009

  30. [38]

    Improved approximation algorithms for large matrices via random projections

    Tamas Sarlos. Improved approximation algorithms for large matrices via random projections. In 2006 47th annual IEEE symposium on foundations of computer science (FOCS'06), pages 143--152. IEEE, 2006

    2006

  31. [39]

    Practical sketching algorithms for low-rank matrix approximation

    Joel A Tropp, Alp Yurtsever, Madeleine Udell, and Volkan Cevher. Practical sketching algorithms for low-rank matrix approximation. SIAM Journal on Matrix Analysis and Applications, 38 0 (4): 0 1454--1485, 2017

    2017

  32. [40]

    Low-rank matrix completion by Riemannian optimization

    Bart Vandereycken. Low-rank matrix completion by Riemannian optimization. SIAM Journal on Optimization, 23 0 (2): 0 1214--1236, 2013

    2013

  33. [41]

    Sketching as a tool for numerical linear algebra

    David P Woodruff. Sketching as a tool for numerical linear algebra. Foundations and Trends in Theoretical Computer Science, 10 0 (1--2): 0 1--157, 2014

    2014

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.