Pith. sign in

REVIEW 3 major objections 4 minor 25 references

Eigenvalue bounds for preconditioned symmetric multiple saddle-point matrices

T0 review · 3 major / 4 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 This paper proves explicit, N-independent bounds on the eigenvalues of block-diagonally preconditioned multiple saddle-point matrices, and extends them to inexact Schur complements.

desk verdict Useful generalization and perturbation analysis, but the proof of the main eigenvalue containment (Lemma 2/Theorem 1) has a uniform-in-λ gap that looks fatal as written. read the letter →

arxiv 2506.02816 v1 pith:V5YPS6DT submitted 2025-06-03 math.NA cs.NAmath.OC

L. Bergamaschi , A. Martinez , J.W. Pearson , A. Potschka This is my paper
classification math.NAcs.NAmath.OC MSC 65F0865F1065F5049M41
keywords multiplesaddle-pointsystemsblockdiagonalpreconditioningeigenvalueboundsMINRESChebyshevpolynomialsSchurcomplementpreconditionersinexact
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 establishes eigenvalue containment bounds for block-tridiagonal multiple saddle-point systems preconditioned by the block diagonal Schur complement preconditioner. It proves that every eigenvalue of the preconditioned matrix lies in a union of intervals indexed by the number of blocks, and that these intervals are bounded in modulus by 2 regardless of how many saddle-point blocks appear. The results extend earlier bounds for three- and four-block systems to arbitrary block counts, and they also cover the practically important case where the Schur complements are applied only approximately. If the bounds are correct, users of iterative solvers such as MINRES can predict how the spectrum, and hence the iteration count, degrades as the number of blocks grows.

What carries the argument

The argument is carried by the parametric polynomial family $U_k(x,\gamma)$ defined by a three-term recurrence (Definition 1), where the parameters $\gamma_k$ are Rayleigh quotients of the matrices $R_kR_k^T$ built from the scaled off-diagonal blocks. Zeros of $U_{N+1}$ control the eigenvalues of the symmetric similar matrix $Q = P_D^{-1/2}AP_D^{-1/2}$. The proof shows, via an implicit-function and monotonicity argument (Lemmas 3-5), that extremal zeros occur when each $\gamma_k$ is at its boundary value 0 or 1; at those boundaries the polynomials factor into Chebyshev-like polynomials $P_j$ with explicit cosine zeros, yielding the intervals of Theorem 2. Theorem 3 gives an equivalent tridiagonal-matrix viewpoint in which the zeros appear as eigenvalues of decoupled symmetric tridiagonal blocks.

What would settle it

Compute the eigenvalues of $Q = P_D^{-1/2}AP_D^{-1/2}$ for a family of matrices satisfying the assumptions with, say, $N=5$, sampling the $\gamma_k$ parameters over a fine grid in $[0,1]^5$; if any eigenvalue falls outside the union of intervals $\bigcup_{k=1}^{6} I_k$ from Theorem 2, the containment claim is false. Equally, violate one assumption (for example, take $A_0$ indefinite or a rank-deficient $B_1$) and check whether the eigenvalue characterization collapses as the theory predicts.

Watch

Extended reading notes

Core claim

The central claim is Theorem 1 and Theorem 2: for the preconditioned matrix $P_D^{-1}A$, with $P_D = \operatorname{blkdiag}(S_0,\dots,S_N)$ built from the Schur complements $S_k = A_k + B_k S_{k-1}^{-1}B_k^T$, every eigenvalue lies in $\bigcup_{k=1}^{N+1} I_k$, where each $I_k$ is the union of intervals containing the zeros of a recursively defined polynomial family $U_k(x,\gamma)$. Theorem 2 gives the explicit containment $I_{k+1} \subseteq \bigcup_{j=1}^{k+1} (I(P_j) \cup -I(P_{j-1}))$, in which $P_j$ are Chebyshev-like polynomials whose zeros are $2\cos((2s-1)\pi/(2j+1))$. Consequently the extremal eigenvalues of $P_D^{-1}A$ are bounded in modulus by 2 for every $N$, and the eigenvalues closest to zero move toward zero as $N$ grows. A perturbation theorem extends the bounds to approximate Schur complements, and refined field-of-value bounds are derived for double saddle-point systems. This characterization is the statement that the paper argues for, and the numerical experiments are offered as validation.

Load-bearing premise

The bounds require that each Schur complement $S_k$ is symmetric positive definite, which follows from the standing assumptions that $A_0$ is positive definite, the other $A_k$ are positive semidefinite, the coupling blocks $B_k$ have full rank, and $n_k \le n_{k-1}$.

Editorial extensions

If this is right

  • For any number of blocks $N$, every eigenvalue of $P_D^{-1}A$ satisfies $|\lambda| \le 2$; no block count can push the spectrum outside this fixed range.
  • The intervals $I_k$ are explicitly computable from the displayed cosine formulas, so a user can read off worst-case positive and negative eigenvalues, hence a MINRES convergence estimate, before running any iteration.
  • Setting $N=2$ and $N=3$ recovers the previously published bounds for $3\times3$ and $4\times4$ systems, while the null-diagonal case treated earlier for arbitrary $N$ is a special case of Theorem 2.
  • With inexact Schur complements satisfying the perturbation assumptions, the eigenvalues of the approximately preconditioned matrix lie in the Minkowski sum of the exact intervals and a computable perturbation interval $[\sigma_-,\sigma_+]$.
  • Numerical experiments on random and PDE-constrained-optimization problems confirm the predicted interval endpoints and show MINRES iteration counts increasing with $N$ as the theory suggests.

Reading between the lines

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

  • Because the $[-2,2]$ bound is independent of the magnitudes of the individual blocks, the result suggests that block diagonal Schur complement preconditioning is inherently robust to problem scaling, at least within the structural assumptions; this could justify using it on badly scaled multiphysics couplings.
  • The monotonicity lemma (Lemma 5) gives more than containment: it says each zero moves monotonically with the Rayleigh quotients $\gamma_k$, so the theory could be turned into a sensitivity tool for predicting how eigenvalue intervals shrink as an approximate inner solver is improved.
  • A natural stress test is to vary the $\gamma_k$ parameters continuously for $N$ larger than the tested range and check whether the extremal zeros of $U_{N+1}$ always fall inside the Theorem 2 intervals; the implicit-function argument suggests they do, but the paper only validates on random samples.
  • For block counts beyond those tested, the widening gap between the $I_k$ intervals implies that preconditioned MINRES will need more iterations; an extension to time-dependent or multi-field problems with $N>6$ could quantify how quickly this degradation becomes prohibitive.
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

3 major / 4 minor

Summary. The manuscript develops eigenvalue bounds for symmetric block tridiagonal multiple saddle-point systems preconditioned by a block diagonal Schur complement preconditioner. The approach reduces the spectral analysis to the zeros of a recursively defined family of Chebyshev-like polynomials U_k, derives explicit containment intervals I_k with endpoints given by trigonometric formulas, and shows that the extremal eigenvalues are bounded in modulus by 2 independently of the number of blocks. The paper also treats approximately applied Schur complements, giving a general perturbation theorem and a refined analysis for double saddle-point systems, with numerical experiments on random and PDE-constrained problems.

Significance. If the proofs were complete, this would be a valuable extension of the known N=1 and N=2 cases to arbitrary N, with explicit, N-independent outer bounds and a useful framework for inexact Schur complements. The polynomial recurrence, the binary-parameter root formulas of Theorem 3, and the backward perturbation framework of Section 4.1 are elegant and likely to be useful to the preconditioning community. The numerical validation is extensive and generally consistent with the claimed bounds. However, the central theorem currently rests on a false uniform statement in Lemma 2, so the main contribution cannot be considered established until that gap is repaired.

major comments (3)
  1. [Section 2, Lemma 2 and proof of Theorem 1] Lemma 2 is false as stated. The proof produces, via Lemma 1, a vector v whose Rayleigh quotient v^T Z_k(λ)v / v^T v depends on λ, so the parameter vector γ obtained from the induction hypothesis depends on λ, whereas the lemma promises a single γ valid for all λ outside the union of the intervals I_j. A concrete failure occurs already at the k=2 induction step: take R1R1^T = diag(0.2, 0.8), R2 = I, and u = (1,1)^T/√2. Then u^T Z_2(λ)^{-1}u / u^T u = (1/2)[(λ-1)/(λ^2-0.2λ-1) + (λ-1)/(λ^2-0.8λ-1)]. Equating this to (λ-1)/(λ^2-γ1λ-1) forces γ1 ≈ 0.59 at λ=2 and γ1 ≈ 0.5417 at λ=3, so no fixed γ1 exists. Since Theorem 1 invokes Lemma 2 at every induction step, its proof is unsound as written. The theorem may be recoverable from a pointwise version ('for each fixed λ there exists γ'), but that version is neither stated nor proved, and the pointwise induction is not automatic because the γ produced for the auxiliary vector v need not coincide with the γ_k determined by u.
  2. [Section 4.2, Lemma 9] The proof of Lemma 9 is presented as a 'sketch' consisting of a sign table, yet Theorem 6 and Corollary 2 depend on the resulting interval endpoints. The sign table appears correct on inspection, but the derivation of the signs of ∂π/∂γ_* and the treatment of the case γ_E^(0) ≤ μ_b are not written out. Since this is a load-bearing step for the refined double saddle-point bounds, the monotonicity analysis should be given in full rather than as a sketch.
  3. [Section 4.1, Theorem 4] The proof of Theorem 4 bounds only the four extremal eigenvalues using Weyl's inequality, while the statement concerns all eigenvalues of the approximately preconditioned matrix. The missing argument is to apply Weyl's inequality index by index, using Theorem 1 to place every eigenvalue of M1 in I_{N+1}; this is a short addition, but it should be stated explicitly. As printed, the proof does not fully establish the asserted containment for the interior eigenvalues.
minor comments (4)
  1. [Section 2, Definition 1 and Notation] The notation I(Q) is defined for γ-dependent polynomials, but Theorem 2 applies it to the parameter-free polynomials P_j and V_j; please clarify that in this case I(P_j) denotes the union of the intervals between consecutive zeros of P_j.
  2. [Section 3, Lemma 5] The first sentence of Lemma 5 is garbled and contains a duplicated condition: 'Let 0< j≤ k, γ_i ∈ (0,1]^N' should be 'Let 0 < j ≤ k and γ_i ∈ (0,1] for i = j, . . . , k.' Please revise the statement.
  3. [Section 3.2, Table 1] The random test cases are generated without a fixed seed, which hinders reproducibility; please either fix a seed or state explicitly that the random instances are representative. The paper already notes that the bound on the smallest negative eigenvalue is loose, but this should also be mentioned in the caption of Table 1.
  4. [Section 4.2, Lemma 9 sign table] The sign table in the proof of Lemma 9 is difficult to parse; adding column headers and explicitly identifying which interval endpoint is derived from each row would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the eigenvalue bounds are derived from a polynomial recurrence and standard linear algebra, with numerical tests as validation rather than fitted predictions.

full rationale

The central derivation chain is self-contained. The eigenvalues of P_D^{-1}A are characterized through the symmetric similar matrix Q (Eq. 1), the block recurrences (3), and the parameterized polynomials U_k of Definition 1; Theorem 1 then locates eigenvalues in the union of intervals I_k. Theorem 2's interval bounds for the zeros are proved in-paper via Lemma 3 (implicit function theorem), Lemma 4 (interlacing of roots), and Lemma 5 (monotonicity in each gamma_j), with the boundary cases gamma_j in {0,1} resolved by explicit Chebyshev-type formulas. The only citation into the authors' prior work is Lemma 1, taken from [7, Lem. 2.1]; it is a generic Rayleigh-quotient lemma whose hypotheses (symmetric Z with 0 outside the spectral interval) do not include the target eigenvalue bound, so the citation is independent support rather than a load-bearing circular premise. No fitted parameter is renamed as a prediction: the numerical experiments compare computed eigenvalues to analytically derived bounds, and the approximate-Schur-complement results (Theorem 4 and Section 4.2) are perturbation consequences of Theorem 1, not fits. The skeptical objection that Lemma 2's uniform-in-lambda statement may be false is a correctness/quantifier concern about the proof as written, not a circularity: even a pointwise version would not make the conclusion an input. There is no uniqueness claim imported from the authors' own work and no known result merely renamed.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters or invented entities. The bounds are derived from a polynomial recurrence and are worst-case over the admissible Rayleigh quotient range gamma in [0,1]^N. The only external inputs are standard mathematical tools and the structural assumptions on the saddle-point matrices.

assumptions (4)
  • standard math The Implicit Function Theorem (Lemma 3) is used to track zero branches of the parametric polynomials.
    Invoked in Section 3 to justify that extremal zeros occur at parameter endpoints, and in Lemmas 8 and 9 for the refined double saddle-point bounds.
  • standard math Weyl's inequality for eigenvalues of Hermitian matrices.
    Used in the proof of Theorem 4 to bound eigenvalues of the perturbed preconditioned matrix.
  • domain assumption Schur complements S_k are symmetric positive definite under the standing assumptions.
    Section 1; needed for P_D to be positive definite and for the similarity transformation Q = P_D^{-1/2} A P_D^{-1/2} to be meaningful.
  • standard math Sylvester's law of inertia for counting positive and negative eigenvalues of A.
    Used in the proof of Theorem 4 to identify the indices of extremal eigenvalues of the preconditioned matrix.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Eigenvalue bounds for preconditioned symmetric multiple saddle-point matrices." pith.science (2026). https://pith.science/paper/V5YPS6DT

@misc{pith2026250602816,
  author       = {Pith},
  title        = {Pith review of: Eigenvalue bounds for preconditioned symmetric multiple saddle-point matrices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V5YPS6DT}},
  note         = {Machine review of arXiv:2506.02816}
}
abstract

We develop eigenvalue bounds for symmetric, block tridiagonal multiple saddle-point linear systems, preconditioned with block diagonal matrices. We extend known results for $3 \times 3$ block systems [Bradley and Greif, IMA J.\ Numer. Anal. 43 (2023)] and for $4 \times 4$ systems [Pearson and Potschka, IMA J. Numer. Anal. 44 (2024)] to an arbitrary number of blocks. Moreover, our results generalize the bounds in [Sogn and Zulehner, IMA J. Numer. Anal. 39 (2018)], developed for an arbitrary number of blocks with null diagonal blocks. Extension to the bounds when the Schur complements are approximated is also provided, using perturbation arguments. Practical bounds are also obtained for the double saddle-point linear system. Numerical experiments validate our findings.

Figures

Figures reproduced from arXiv: 2506.02816 by the authors.

Figure 1
Figure 1. Bounding intervals Ik+1 = [Bound− l , Bound+ l ] ∪ [Bound− u , Bound+ u ] for Uk+1, k = 1, . . . , 9. This corresponds to eigenvalue bounds for preconditioned multiple saddle￾point systems with k + 1 blocks (i.e., N = k). where the leading entry is cji−1+1 = (−1)ji−1+1. We can view M+ r and −M− r as multiple saddle-point systems of dimension r, and their analogous block diagonal preconditioners are identity matrices… view at source ↗
Figure 2
Figure 2. Extremal eigenvalues of the preconditioned matrix (blue dots) and bounds (red line) after 25 runs with each combination of the parameters from [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

  1. [1]

    Bakrani Balani, L

    F. Bakrani Balani, L. Bergamaschi, A. Mart ´ınez, and M. Hajarian , Some preconditioning techniques for a class of double saddle point problems , Numerical Linear Algebra with Applications, 31 (2024), p. e2551

    work page 2024

  2. [2]

    Bakrani Balani, M

    F. Bakrani Balani, M. Hajarian, and L. Bergamaschi , Two block preconditioners for a class of double saddle point linear systems, Applied Numerical Mathematics, 190 (2023), p. 155 – 167

    work page 2023

  3. [3]

    Beigl, J

    A. Beigl, J. Sogn, and W. Zulehner , Robust preconditioners for multiple saddle point problems and applications to optimal control problems, SIAM Journal on Matrix Analysis and Applications, 41 (2020)

    work page 2020

  4. [4]

    F. P. Beik, C. Greif, and M. Trummer , On the invertibility of matrices with a double saddle-point structure , Linear Algebra and its Applications, 699 (2024), pp. 403–420

    work page 2024

  5. [5]

    F. P. A. Beik and M. Benzi , Iterative methods for double saddle point systems , SIAM J. Matrix Anal. Appl., 39 (2018), pp. 902–921

    work page 2018

  6. [6]

    Math., 150 (2022), pp

    , Preconditioning techniques for the coupled Stokes-Darcy problem: spectral and field-of-values analysis , Numer. Math., 150 (2022), pp. 257–298. EIGENV ALUE BOUNDS FOR PRECONDITIONED SYMMETRIC MULTIPLE SADDLE-POINT MATRICES 19

    work page 2022

  7. [7]

    Bergamaschi, A

    L. Bergamaschi, A. Martinez, J. W. Pearson, and A. Potschka , Spectral analysis of block preconditioners for double saddle-point linear systems with application to PDE-constrained optimization , Computational Optimization with Applica- tions, 91 (2024), pp. 423–455

    work page 2024

  8. [8]

    Bradley and C

    S. Bradley and C. Greif, Eigenvalue bounds for double saddle-point systems , IMA J. Numer. Anal., 43 (2023), pp. 3564– 3592

    work page 2023

Show all 25 references
  1. [9]

    Chidyagwai, S

    P. Chidyagwai, S. Ladenheim, and D. B. Szyld , Constraint preconditioning for the coupled Stokes-Darcy system , SIAM J. Sci. Comput., 38 (2016), pp. A668–A690

    2016

  2. [10]

    Ferronato, A

    M. Ferronato, A. Franceschini, C. Janna, N. Castelletto, and H. Tchelepi , A general preconditioning framework for coupled multi-physics problems , J. Comput. Phys., 398 (2019), p. 108887

    2019

  3. [11]

    G. H. Golub and R. S. V arga, Chebyshev semi-iterative methods, successive over-relaxation iterative methods, and second order Richardson iterative methods. I , Numer. Math., 3 (1961), pp. 147–156

    1961

  4. [12]

    II , Numer

    , Chebyshev semi-iterative methods, successive over-relaxation iterative methods, and second order Richardson iter- ative methods. II , Numer. Math., 3 (1961), pp. 157–168

    1961

  5. [13]

    Han and X

    D. Han and X. Yuan , Local linear convergence of the alternating direction method of multipliers for quadratic programs , SIAM J. Numer. Anal., 51 (2013), pp. 3446–3457

    2013

  6. [14]

    Mardal, B

    K.-A. Mardal, B. F. Nielsen, and M. Nordaas , Robust preconditioners for PDE-constrained optimization with limited observations, BIT, 57 (2017), pp. 405–431

    2017

  7. [15]

    C. C. Paige and M. A. Saunders, Solution of sparse indefinite systems of linear equations , SIAM J. on Numer. Anal., 12 (1975), pp. 617–629

    1975

  8. [16]

    J. W. Pearson and A. Potschka , Double saddle-point preconditioning for Krylov methods in the inexact sequential homotopy method, Numerical Linear Algebra with Applications, 31 (2024), p. e2553

    2024

  9. [17]

    , On symmetric positive definite preconditioners for multiple saddle-point systems , IMA J. Numer. Anal., 44 (2024), pp. 1731–1750

    2024

  10. [18]

    Ramage and E

    A. Ramage and E. C. Gartland , A preconditioned nullspace method for liquid crystal director modeling , SIAM Journal on Scientific Computing, 35 (2013), pp. B226–B247

    2013

  11. [19]

    Rhebergen, G

    S. Rhebergen, G. N. Wells, A. J. W athen, and R. F. Katz , Three-field block preconditioners for models of coupled magma/mantle dynamics , SIAM J. Sci. Comput., 37 (2015), pp. A2270–A2294

    2015

  12. [20]

    Sogn and W

    J. Sogn and W. Zulehner, Schur complement preconditioners for multiple saddle point problems of block tridiagonal form with application to optimization problems , IMA Journal of Numerical Analysis, 39 (2018), pp. 1328–1359

    2018

  13. [21]

    Tao, Topics in random matrix theory , vol

    T. Tao, Topics in random matrix theory , vol. 132 of Graduate Studies in Mathematics, American Mathematical Society, Providence, RI, 2012

    2012

  14. [22]

    A. J. W athen, Realistic eigenvalue bounds for the Galerkin mass matrix , IMA J. Numer. Anal., 7 (1987), pp. 449–457

    1987

  15. [23]

    A. J. W athen and T. Rees, Chebyshev semi-iteration in preconditioning for problems including the mass matrix , Electron. Trans. Numer. Anal., 34 (2008/09), pp. 125–135

    2008

  16. [24]

    J. H. Wilkinson , Handbook Series Linear Algebra. Calculation of the eigenvalues of a symmetric tridiagonal matrix by the method of bisection , Numer. Math., 4 (1962), pp. 362–367

    1962

  17. [25]

    J. Y. Yuan , Numerical methods for generalized least squares problems , in Proceedings of the Sixth International Congress on Computational and Applied Mathematics (Leuven, 1994), vol. 66, 1996, pp. 571–584

    1994

Pith tools

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