Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

A structural bound for cluster robustness of randomized small-block Lanczos

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A structural bound ties cluster robustness of small-block Lanczos to the block-level spectral gap.

desk verdict This paper has a real structural theorem for small-block Lanczos, but the advertised cluster-robustness bound only follows from an unproved conjecture that the author explicitly identifies as a hard open problem. read the letter →

arxiv 2507.10144 v2 pith:RXPJ6G5M submitted 2025-07-14 math.NA cs.NA

Nian Shao This is my paper · ORCID
classification math.NAcs.NA MSC 65F1515A1860B20
keywords blockLanczosclusterrobustnessmatrixpolynomialsKrylovsubspacemethodsrandomizedeigenvaluealgorithmslow-rankapproximationspectralgapVandermondematrices
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 sets out to explain why Randomized Small-Block Lanczos (RSBL) — a block Lanczos variant whose block size is larger than one but far smaller than the target eigenvalue cluster — stays robust when eigenvalues cluster tightly. Its main theorem bounds the angle between the target cluster subspace and the $d$-step block Krylov subspace by a structurally explicit expression in which the controlling quantity is the $b$th-order relative gap $\mathrm{relgap}_b$ between $b\times b$ eigenvalue blocks, not the individual eigenvalue gaps that slow single-vector Lanczos. If valid, this gives theoretical support for extracting $bd$ eigenvalues from a cluster at matvec cost governed by the block-level gap. The proof works through a new interpolation formula for matrix polynomials that generalizes Lagrange interpolation and reduces the gap exponent from $bd-1$ to $d-1$. A conjectured high-probability bound on two non-normality factors would turn the structural bound into a clean robustness estimate; the paper supports the conjecture with experiments.

What carries the argument

The load-bearing mechanism is block Vandermonde interpolation with matrix polynomials. The paper constructs fundamental matrix polynomials $F_k$ satisfying $F_k(B_j)=\delta_{kj}I$ for the $b\times b$ blocks $B_i=\Omega_i^{-1}\Lambda_i\Omega_i$, using a chain-of-solvents factorization (Lemma 3 and Theorem 1) that generalizes Lagrange interpolation to non-commuting block variables. The factorization expresses $F_1(\lambda I)$ as a product of linear factors $\lambda I-\hat B_i$ generated by a backward recurrence of matrices $S_{i,j}$, and its growth away from the cluster interval is controlled by two quantities, $\chi_{\mathrm{mono}}$ and $\chi_{\mathrm{coef}}$, that measure the departure of the similarity-transformed blocks from normality. Inequality (13) shows that the degree-$(d-1)$ growth of each fundamental polynomial costs only one inverse power of the block-level gap $\mathrm{relgap}_b$, which is what converts classical scalar interpolation into the cluster-robust RSBL statement.

What would settle it

Run the paper's own $d=2$ experiment for moderate $b$ (say $b=2$ or $3$) with the inter-block gap $\alpha$ fixed and let the intra-block spacing $\beta$ shrink far below the values in Figure 4. Since $\chi_{\mathrm{coef}}$ then equals $\|(B_1-B_2)^{-1}\|\cdot\min|\lambda_1-\lambda_2|$, a growing $\chi_{\mathrm{coef}}$ as $\beta\to0$ — equivalently, a smallest singular value of the block Vandermonde matrix that shrinks with the intra-block gap — would falsify Conjecture 1 and the corollary's independence of intra-block gaps.

Watch

Extended reading notes

Core claim

For a symmetric matrix $A$ with desired eigenvalues partitioned into $d$ blocks $\Lambda_1,\dots,\Lambda_d$ of size $b$, and a Gaussian starting block $\Omega$, the paper proves, with probability one, that $\tan\angle(\mathrm{range}(Q),K_d(A,\Omega))\le c_\Omega G_d$, where $G_d$ is the maximum over the unwanted spectrum of the norms of the fundamental matrix polynomials in the interpolation formula. Combined with the growth estimate in Theorem 2, this gives $\tan\angle(\mathrm{range}(Q),K_d(A,\Omega))\le c_\Omega(\chi_{\mathrm{mono}}\chi_{\mathrm{coef}}/\mathrm{relgap}_b)^{d-1}$. The author's claim is that $\mathrm{relgap}_b$, the gap between $b\times b$ diagonal blocks rather than between individual eigenvalues, is the correct resolution scale for RSBL: clusters may be much tighter than the block size without destroying convergence, provided no eigenvalue multiplicity exceeds $b$. Conjecture 1 asserts that the two non-normality factors are benign, and the experiments show the angle tracking $\mathrm{relgap}_b^{-(d-1)}$ while staying independent of the intra-block spacing.

Load-bearing premise

The proof delivers its clean robustness bound only if Conjecture 1 is true: the two non-normality factors $\chi_{\mathrm{mono}}$ and $\chi_{\mathrm{coef}}$ stay bounded, with high probability, by a constant depending only on $b$ and $d$ and not on the eigenvalue gaps inside or between the blocks; if that fails, the bound can blow up as clusters tighten.

Editorial extensions

If this is right

  • If the structural bound holds, RSBL can recover $bd$ eigenvalues of a cluster whose diameter far exceeds the block size, with the convergence rate depending only on the gap between $b\times b$ blocks, so adding more eigenvalues to a cluster of fixed block size does not by itself slow the method.
  • Eigenvalue computation inherits a Chebyshev-accelerated bound (23): after $d$ steps, further matvecs give exponential convergence with rate controlled by $\mathrm{relgap}_b$, recovering and extending the single-vector result.
  • For low-rank approximation via $A=A^T A$ of a tall matrix, the bound yields sample complexity whose $\log(1/\mathrm{relgap}_b)$ term is linear in the target rank, matching the 'linear complexity in target rank for any block size' conjecture of [20].
  • The multiplicity condition $\mathrm{relgap}_b>0$ is shown necessary: if an eigenvalue has multiplicity larger than $b$, no block Krylov space of any length can contain its full eigenspace, which explains null-space failures.
  • Because the bound is independent of the gap between the cluster interval and the rest of the spectrum, the result applies equally to interior and exterior eigenvalue clusters.

Reading between the lines

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

  • If Conjecture 1 is established, the same structural framework should transfer to other block Krylov solvers such as block GMRES or block rational-Krylov methods, where the decisive quantity would again be the block-level gap rather than individual gaps.
  • The $d=2$ analysis suggests a concrete random-matrix target: bound the smallest singular value of $B_1-B_2$ for two independent Gaussian similarities of diagonal matrices; progress there would likely resolve the general conjecture through the $S_{i,j}$ recurrence.
  • A practical design principle follows: choose the block size to match the maximal expected eigenvalue multiplicity rather than the cluster size; beyond that threshold, the theory predicts little additional robustness gain from larger blocks, only higher cost.
  • The bound's independence of the distance to the unwanted spectrum suggests that restarting or filtering strategies for RSBL can be tuned using only cluster-internal information, potentially reducing communication in parallel settings.
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 / 4 minor

Summary. The paper studies Randomized Small-Block Lanczos (RSBL) for symmetric eigenvalue problems, aiming to explain its observed robustness when the target eigenvalues form clusters of size larger than the block size b. It develops a structural bound for tan ∠(range(Q), K_d(A, Ω)), where Q is the invariant subspace associated with a cluster and Ω is a Gaussian b-column initial block. The main theorem (Theorem 3) bounds this tangent by c_Ω G_d, and Theorem 2's inequality (13) converts the factor G_d into (χ_mono χ_coef / relgap_b)^{d-1}. Because χ_mono and χ_coef depend on random similarity transformations of the cluster blocks, the paper proposes Conjecture 1, stating that their product is bounded with high probability independently of the spectra Λ_i; under this conjecture the clean bound tan ∠ ≤ (c_{b,d,n}/relgap_b)^{d-1} follows. Numerical experiments support the conjecture, and the paper discusses consequences for eigenvalue computation and low-rank approximation.

Significance. The structural part is a genuine contribution: Theorem 3 and inequality (13) are derived through matrix-polynomial interpolation, with no fitted constants, and the paper honestly separates what is proved from what is conjectured. The public reproduction scripts are a strength. If Conjecture 1 were proved, the clean relgap_b^{-(d-1)} estimate would be an important result, improving the block-size dependence over prior work and providing the first rigorous explanation of small-block cluster robustness. However, the advertised cluster-robustness conclusion is conditional on a conjecture that the paper itself shows is connected to an open anti-concentration problem already in the d=2 case. The significance is therefore real but currently conditional.

major comments (2)
  1. [§3.2, Conjecture 1 and Corollary of Conjecture 1] Conjecture 1 is the load-bearing bridge from the deterministic structural bound to the advertised robustness estimate. Theorem 3 only gives tan ∠ ≤ c_Ω (χ_mono χ_coef / relgap_b)^{d-1}; the clean bound tan ∠ ≤ (c_{b,d,n}/relgap_b)^{d-1} is not a theorem but a corollary of the unproved conjecture. This is not a minor technical gap: Section 3.2.2 shows that already for d=2 one must control χ_coef = ||(B_1−B_2)^{-1}|| · min|λ_1−λ_2|, i.e. the smallest singular value of a difference of two independent non-normal random similarities, which the paper identifies as an intrinsic open difficulty. The experiments in Figure 4 are suggestive but do not establish uniformity in Λ_i. The manuscript should either prove the conjecture, at least in a nontrivial special case, or explicitly state in the abstract, introduction, and applications that the relgap_b-based bound and its consequences are conditional on Conjecture 1.
  2. [§3.1, Theorem 3, high-probability bound on c_Ω^2] The displayed high-probability bound on c_Ω^2 appears under-derived. From the definition c_Ω = sqrt(d(n−bd)) · max_i ||Ω_i^{-1}|| · max_j ||Ω_j|| · max_j ||Ω_j|| ||Ω_j^{-1}||, a direct application of the quoted Gaussian tail bounds to the three independent max-factors gives powers of (2√b + sqrt(2 log(2(m−d)/δ))) that are not reflected in the printed statement: both max_j ||Ω_j|| and max_j ||Ω_j|| ||Ω_j^{-1}|| contribute, so the factor (2√b + ...) should appear at least to the fourth power in c_Ω^2, with an additional b^2/δ^4 block-size factor, and the prefactor should be d(n−bd), not (dn−bd^2)^3. In addition, the statement 'with probability at least 1−4δ' does not account for the maxima over d and m−d independent Gaussian blocks unless δ is rescaled by m. This is a technical error in a stated theorem and should be corrected; it does not affect the deterministic structural bound (16), but it does affect the 'with high probability' formulation used later.
minor comments (4)
  1. [§1, Remark 1] The bound tan ∠(range(Q), K_ℓ(A,Ω)) ≤ c γ^{ℓ−d} tan ∠(range(Q), K_d(A,Ω)) is asserted without proof or a precise reference; Section 3.2.3 gives a related estimate only under Conjecture 1, so please state the exact lemma from [15] that justifies the deterministic decay.
  2. [§2.4, proof of Proposition 3] In the sentence 'where we use Ω_i = I_d in the last equality', the identity should be I_b, since Ω_i ∈ R^{b×b}; this is a typo but may confuse readers.
  3. [§3.2, Conjecture 1] The phrase 'with high probability' should be made quantitative: specify whether the failure probability is allowed to depend on b and d only, and state explicitly that the constant in the conjecture is independent of the eigenvalues in Λ_i.
  4. [§3.2.1, Figure 4] The experiments in Figure 4 report medians with quartiles, whereas Conjecture 1 is a high-probability statement; reporting tail probabilities or maxima over trials would provide more directly relevant evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the structural bound is derived from external matrix-polynomial and random-matrix results, and the advertised clean gap-only estimate is explicitly conditional on a stated open conjecture rather than being assumed into the derivation.

full rationale

The paper's central result, Theorem 3 together with inequality (13), is derived from external ingredients: matrix-polynomial interpolation from Dennis-Traub-Weber, the angle identity from Zhu-Knyazev, and Gaussian random-matrix tail bounds from Rudelson-Vershynin. No fitted constants enter the derivation, and the quantity c_Omega is controlled by standard random-matrix bounds rather than calibrated to the conclusion. The subsequent clean estimate (c_{b,d,n}/relgap_b)^{d-1} is explicitly marked as contingent on Conjecture 1, which the paper states as an open probabilistic statement about chi_mono chi_coef and supports only with numerical experiments; this is conditionality, not circularity, because the conjecture is not secretly assumed as an input to Theorem 3. Author self-citations, notably [15] and [20], are used in the applications section to translate the angle bound into convergence statements for eigenvalues and low-rank approximation, and in the introduction to frame the problem; these citations do not supply the load-bearing content of the structural bound. The d=2 discussion in Section 3.2.2 honestly identifies the estimation of k(B_1-B_2)^{-1}k as an intrinsic open difficulty, which further confirms that the paper is not disguising a fitted or self-referential argument as a proof. A secondary technical concern about the stated high-probability bound on c_Omega^2 possibly dropping a factor in the union-bound computation affects correctness of a tail estimate, not circularity of the derivation.

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

No parameters are fitted in the derivation; the theory relies on standard external theorems plus the explicitly conjectural bound on chi_mono and chi_coef.

assumptions (5)
  • standard math Generalized Bezout theorem for matrix polynomials (Gantmacher, Chap. 4.3, Thm. 1) guarantees a factorization (lambda I - B) * Phi_B(lambda I) whenever B is a solvent.
    Used in the induction proof of Lemma 3 to strip off chain-of-solvent factors.
  • standard math Gaussian random matrix tail bounds for extreme singular values from Rudelson-Vershynin (Thm. 2.6 and Thm. 3.1).
    Used in Theorem 3 to control c_Omega; the paper's application of these bounds has a union-bound gap.
  • standard math Identity tan angle(range(Q), K_d(A,Omega)) = ||K_perp K^{-1}|| from Zhu and Knyazev (Thm. 3.1).
    The starting point of the Theorem 3 proof, valid when K is full column rank.
  • domain assumption m = n/b is an integer (Theorem 3).
    Simplifies the partition of the Gaussian initial matrix into b-by-b blocks; the author notes non-multiple cases can be handled by augmentation.
  • ad hoc to paper Conjecture 1: chi_mono * chi_coef is bounded with high probability by a constant depending only on b and d, independent of eigenvalue gaps.
    This is unproved and is exactly what converts Theorem 3 into the clean cluster-robustness corollary; numerical experiments support but do not prove it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A structural bound for cluster robustness of randomized small-block Lanczos." pith.science (2026). https://pith.science/paper/RXPJ6G5M

@misc{pith2026250710144,
  author       = {Pith},
  title        = {Pith review of: A structural bound for cluster robustness of randomized small-block Lanczos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RXPJ6G5M}},
  note         = {Machine review of arXiv:2507.10144}
}
read the original abstract

The Lanczos method is a fast and memory-efficient algorithm for solving large-scale symmetric eigenvalue problems. However, its rapid convergence can deteriorate significantly when computing clustered eigenvalues due to a lack of cluster robustness. A promising strategy to enhance cluster robustness -- without substantially compromising convergence speed or memory efficiency -- is to use a random small-block initial, where the block size is greater than one but still much smaller than the cluster size. This leads to the Randomized Small-Block Lanczos (RSBL) method. Despite its empirical effectiveness, RSBL lacks the comprehensive theoretical understanding already available for single-vector and large-block variants. In this paper, we develop a structural bound that supports the cluster robustness of RSBL by leveraging tools from matrix polynomials. We identify an intrinsic theoretical challenge stemming from the non-commuting nature of matrix multiplication. To provide further insight, we propose a conjectured probabilistic bound for cluster robustness and validate it through empirical experiments. Finally, we discuss how insights into cluster robustness can enhance our understanding of RSBL for both eigenvalue computation and low-rank approximation.

Figures

Figures reproduced from arXiv: 2507.10144 by the authors.

Figure 1
Figure 1. Total number of matvecs and computational time by block Lanczos required for the conver￾gence of 32 eigenvalues in a cluster with β being the cluster diameter. While the number of matvecs required by the single-vector Lanczos method is consistently lower than that of block variants, the relative overhead incurred by RSBL quickly becomes negligible as the target eigenvalues form a tighter cluster. When communication … view at source ↗
Figure 2
Figure 2. Diagram for defining Bbi and Ωbi . Proof of Lemma 3. The proof is based on mathematical induction. Denote Φd−1 = F1 ∈ Pd−1(R b×b ). For the base case, since Φd−1(Bd) = 0, the generalized Bézout’s theorem yields the existence of a matrix polynomial Φd−2 ∈ Pd−2(R b×b ), such that Φd−1(λI) = (λI − Bd)Φd−2(λI) = (λI − Bbd)Φd−2(λI) ∀ λ ∈ R. We then consider the general case. Suppose that we have Φd−1(λI) =  Y d j=i+1 (λ… view at source ↗
Figure 3
Figure 3. Cluster robustness for exterior eigenvalues (top) and interior eigenvalues (down) with different eigenvalue gaps. The left panels contain matrices with varying eigenvalue gaps between the clusters, whereas the right panels contain matrices with different eigenvalue gaps within each cluster. The median of empirical tan ∠(range(Q), Kd(A, Ω)) over 1000 independent trials is plotted, with the 25th and 75th quartiles sha… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: χmono (top) and χcoef (bottom) with different eigenvalue gaps. The left panels contain matrices with varying eigenvalue gaps between the clusters, whereas the right panels contain matrices with different eigenvalue gaps within each cluster. The median of the empirical …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Convergence analysis of a nonlinear eigensolver based on rational approximation of the resolvent

    math.NA 2026-07 accept novelty 6.0 of 10

    Block probing and zooming-in make resolvent-polefinding for nonlinear eigenproblems accurate and multiplicity-aware, with proved rates and stable barycentric rootfinding.

Reference graph

Works this paper leans on

36 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ballard, E

    G. Ballard, E. Carson, J. Demmel, M. Hoemmen, N. Knight, and O. Schwartz. Communication lower bounds and optimal algorithms for numerical linear algebra. Acta Numer. , 23:1–155, 2014. doi:10.1017/S0962492914000038

    doi:10.1017/s0962492914000038 2014

  2. [2]

    Bathe and E

    K.-J. Bathe and E. L. Wilson. Solution methods for eigenvalue problems in structural mechanics. Internat. J. Numer. Methods Engrg. , 6(2):213–226, 1973. doi:10.1002/nme.1620060207

    work page doi:10.1002/nme.1620060207 1973

  3. [3]

    Brailovskaya and R

    T. Brailovskaya and R. van Handel. Universality and sharp matrix concentration inequalities. Geom. Funct. Anal. , 34(6):1734–1838, 2024. doi:10.1007/s00039-024-00692-9

    doi:10.1007/s00039-024-00692-9 2024

  4. [4]

    T. Chen, E. N. Epperly, R. A. Meyer, C. Musco, and A. Rao. Does block size matter in randomized block Krylov low-rank approximation? In Proceedings of the 2026 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages 1026–1046. SIAM, Philadelphia, PA, 2026. doi:10.1137/ 1.9781611978971.42

    work page 2026

  5. [5]

    N. Cook. Lower bounds for the smallest singular value of structured random matrices. Ann. Probab., 46(6):3442–3500, 2018. doi:10.1214/17-AOP1251

    doi:10.1214/17-aop1251 2018

  6. [6]

    J. E. Dennis, Jr., J. F. Traub, and R. P. Weber. The algebraic theory of matrix polynomials. SIAM J. Numer. Anal. , 13(6):831–845, 1976. doi:10.1137/0713065

    doi:10.1137/0713065 1976

  7. [7]

    Drineas and I

    P. Drineas and I. C. F. Ipsen. Low-rank matrix approximations do not need a singular value gap. SIAM J. Matrix Anal. Appl. , 40(1):299–319, 2019. doi:10.1137/18M1163658

    doi:10.1137/18m1163658 2019

  8. [8]

    Drineas, I

    P. Drineas, I. C. F. Ipsen, E.-M. Kontopoulou, and M. Magdon-Ismail. Structural convergence results for approximation of dominant subspaces from block Krylov spaces. SIAM J. Matrix Anal. Appl., 39(2):567–586, 2018. doi:10.1137/16M1091745

    work page doi:10.1137/16m1091745 2018

Show all 36 references
  1. [9]

    Frommer, K

    A. Frommer, K. Lund, and D. B. Szyld. Block Krylov subspace methods for functions of matrices II: Modified block FOM. SIAM J. Matrix Anal. Appl. , 41(2):804–837, 2020. doi:10.1137/ 19M1255847

    2020

  2. [10]

    F. R. Gantmacher. The theory of matrices. Vol. 1 . AMS Chelsea Publishing, Providence, RI, 1998

    1998

  3. [11]

    M. Hoemmen. Communication-avoiding Krylov subspace methods, 2010. Thesis (Ph.D.)– University of California, Berkeley. URL: http://gateway.proquest.com/openurl?url_ ver=Z39.88-2004&rft_val_fmt=info:ofi/fmt:kev:mtx:dissertation&res_dat=xri: pqdiss&rft_dat=xri:pqdiss:3413388. 18

    2010

  4. [12]

    Ikegami and T

    T. Ikegami and T. Sakurai. Contour integral eigensolver for non-Hermitian systems: a Rayleigh-Ritz-type approach. Taiwanese J. Math. , 14(3A):825–837, 2010. doi:10.11650/twjm/ 1500405869

    2010 doi

  5. [13]

    Ikegami, T

    T. Ikegami, T. Sakurai, and U. Nagashima. A filter diagonalization for generalized eigenvalue problems based on the Sakurai-Sugiura projection method. J. Comput. Appl. Math. , 233(8):1927– 1936, 2010. doi:10.1016/j.cam.2009.09.029

    1927 doi

  6. [14]

    S. Kaniel. Estimates for some computational techniques in linear algebra. Math. Comp. , 20:369– 378, 1966. doi:10.2307/2003590

    1966 doi

  7. [15]

    Kressner and N

    D. Kressner and N. Shao. On a randomized small-block Lanczos method for large-scale null space computations. SIAM J. Sci. Comput. , pages S462–S485, 2026. doi:10.1137/24M1674881

    2026 doi

  8. [16]

    Kubinova and K

    M. Kubinova and K. M. Soodhalter. Admissible and attainable convergence behavior of block Arnoldi and GMRES. SIAM J. Matrix Anal. Appl. , 41(2):464–486, 2020. doi:10.1137/ 19M1272469

    2020

  9. [17]

    Lancaster

    P. Lancaster. Lambda-matrices and vibrating systems . Dover Publications, Inc., Mineola, NY, 2002

    2002

  10. [18]

    R. B. Lehoucq, D. C. Sorensen, and C. Yang. ARPACK users’ guide , volume 6 of Software, Environments, and Tools. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1998. doi:10.1137/1.9780898719628

    1998 doi

  11. [19]

    Li and L.-H

    R.-C. Li and L.-H. Zhang. Convergence of the block Lanczos method for eigenvalue clusters. Numer. Math. , 131(1):83–113, 2015. doi:10.1007/s00211-014-0681-6

    2015 doi

  12. [20]

    Meyer, C

    R. Meyer, C. Musco, and C. Musco. On the unreasonable effectiveness of single vector Krylov methods for low-rank approximation. In Proceedings of the 2024 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages 811–845. SIAM, Philadelphia, PA, 2024. doi:10.1137/ 1.9781...

    2024

  13. [21]

    Musco and C

    C. Musco and C. Musco. Randomized block Krylov methods for stronger and faster ap- proximate singular value decomposition. Advances in Neural Information Processing Sys- tems, 28, 2015. URL: https://proceedings.neurips.cc/paper_files/paper/2015/file/ 1efa39bcaec6f3900149160693...

    2015

  14. [22]

    Z. Nie. Matrix anti-concentration inequalities with applications. In STOC ’22—Proceedings of the 54th Annual ACM SIGACT Symposium on Theory of Computing , pages 568–581. ACM, New York, [2022] ©2022

    2022

  15. [23]

    B. N. Parlett. The symmetric eigenvalue problem , volume 20 of Classics in Applied Mathematics . Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1998. doi:10.1137/ 1.9781611971163

    1998

  16. [24]

    Persson, T

    D. Persson, T. Chen, and C. Musco. Randomized block-krylov subspace methods for low-rank approximation of matrix functions. arXiv preprint: 2502.01888 , 2025. doi:10.48550/arXiv. 2502.01888

    2025 doi

  17. [25]

    Rudelson and R

    M. Rudelson and R. Vershynin. Non-asymptotic theory of random matrices: extreme singular values. In Proceedings of the International Congress of Mathematicians. Volume III , pages 1576–

  18. [26]

    Rudelson and O

    M. Rudelson and O. Zeitouni. Singular values of Gaussian matrices and permanent estimators. Random Structures Algorithms , 48(1):183–212, 2016. doi:10.1002/rsa.20564

    2016 doi

  19. [27]

    Y. Saad. On the rates of convergence of the Lanczos and the block-Lanczos methods. SIAM J. Numer. Anal. , 17(5):687–706, 1980. doi:10.1137/0717059

    1980 doi

  20. [28]

    A. P. Seyranian, E. Lund, and N. Olhoff. Multiple eigenvalues in structural optimization problems. Structural optimization , 8:207–227, 1994. doi:10.1007/BF01742705. 19

    1994 doi

  21. [29]

    Simoncini and E

    V. Simoncini and E. Gallopoulos. Convergence properties of block GMRES and matrix polyno- mials. Linear Algebra Appl. , 247:97–119, 1996. doi:10.1016/0024-3795(95)00093-3

    1996 doi

  22. [30]

    E. D. Sontag. Mathematical control theory , volume 6 of Texts in Applied Mathematics . Springer- Verlag, New York, second edition, 1998. doi:10.1007/978-1-4612-0577-7

    1998 doi

  23. [31]

    J. A. Tropp. Randomized block Krylov methods for approximating extreme eigenvalues. Numer. Math., 150(1):217–255, 2022. doi:10.1007/s00211-021-01250-3

    2022 doi

  24. [32]

    J. A. Tropp. Comparison theorems for the minimum eigenvalue of a random positive-semidefinite matrix. arXiv preprint arXiv:2501.16578 , 2025. doi:10.48550/arXiv.2501.16578

    2025 doi arXiv

  25. [33]

    J. A. Tropp and R. J. Webber. Randomized algorithms for low-rank matrix approximation: Design, analysis, and applications. arXiv preprint: 2306.12418 , 2023. doi:10.48550/arXiv. 2306.12418

    2023 doi arXiv

  26. [34]

    Wu and H

    K. Wu and H. Simon. Thick-restart Lanczos method for large symmetric eigenvalue problems. SIAM J. Matrix Anal. Appl. , 22(2):602–616, 2000. doi:10.1137/S0895479898334605

    doi

  27. [35]

    Zhu and A

    P. Zhu and A. V. Knyazev. Angles between subspaces and their tangents. J. Numer. Math. , 21(4):325–340, 2013. doi:10.1515/jnum-2013-0013 . 20

    2013 doi

  28. [1602]

    Hindustan Book Agency, New Delhi, 2010

    2010

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.