Pith. sign in

REVIEW 3 major objections 5 minor 37 references

LR-WaveHoltz: A Low-Rank Helmholtz Solver

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Low-rank WaveHoltz solves Helmholtz with bounded ranks

desk verdict A credible proof-of-concept for low-rank WaveHoltz; the numerics support the claims, but the theory is thin and one fixable inconsistency needs correcting. read the letter →

arxiv 2510.09352 v2 pith:LJGQPP5U submitted 2025-10-10 math.NA cs.NA

classification math.NAcs.NA MSC 65M0665M2265F10
keywords low-rankmethodstensortrainsHelmholtzequationWaveHoltzAndersonaccelerationsteptruncationSBPfinitedifferencesunderwateracoustics
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 introduces LR-WaveHoltz, a low-rank version of the WaveHoltz iterative Helmholtz solver. WaveHoltz solves the Helmholtz equation by time-stepping a related wave equation over one period and filtering, avoiding the indefinite linear systems that make direct frequency-domain solvers hard. The new method stores each time step's solution in compressed form, using SVD in two dimensions and tensor trains in three, and truncates after every update to control rank growth. The central empirical claim is that the ranks of these compressed iterates stay bounded by the rank of the correspondingly truncated full-rank solution, so the compression does not drive the iteration away from the true solution. If correct, this makes large point-source underwater acoustics problems tractable, especially in 3D where full-rank grids become prohibitive.

What carries the argument

The load-bearing object is the WaveHoltz operator, a time-domain filter that maps initial data to the filtered wave-field integral and has the Helmholtz solution as its fixed point. The low-rank machinery consists of the step-truncation operator applied after every summation and time step, the SVD format in 2D, the tensor-train format in 3D, and a low-rank Anderson acceleration that solves for the combination coefficients using only low-rank inner products. At corner blocks the implicit boundary update is handled through a factored alternating direction iteration solve of a Sylvester equation, which also returns the solution directly in low-rank factors.

What would settle it

Run LR-WaveHoltz on a point-source free-space problem with the paper's tolerance scheduling and compare the converged low-rank iterate with a full-rank WaveHoltz solve on the same grid; if the difference is much larger than the truncation tolerance while the reported residual is below it, the truncated fixed point has shifted. A simpler check: monitor blockwise ranks and look for any block whose rank grows beyond the rank of the full-rank solution truncated at the same tolerance.

Watch

Extended reading notes

Core claim

The paper claims that the WaveHoltz fixed-point iteration can be run entirely in low-rank form without losing the Helmholtz solution. In each iteration, the wave equation is advanced with a step-truncated leapfrog scheme, then the WaveHoltz filter integral is accumulated and truncated; in 2D the solution is stored as an SVD and in 3D as a tensor train. Low-rank Anderson acceleration is added in 2D to speed up the fixed-point iteration. Extensive experiments on free-space, half-space, and layered problems with point sources show convergence to the expected solutions, and in every block the numerical rank remains bounded by the rank of the full-rank solution truncated at the same tolerance. Th

Load-bearing premise

The load-bearing premise is that compressing the solution after every time step and every WaveHoltz update does not move the iteration away from the true fixed point of the WaveHoltz operator; the paper supports this only through numerical observation, not with an error bound.

Editorial extensions

If this is right

  • In 3D, the tensor-train representation allows solves on grids with 301^3 points per block, which the paper notes would be prohibitive for a full-rank solver.
  • Throughout the iteration, ranks stay bounded by the rank of the truncated full-rank solution, so memory requirements can be predicted from the solution's own separability rather than from the mesh size.
  • For open free-space problems, WaveHoltz already converges quickly, so low-rank Anderson acceleration gives little iteration-count improvement; acceleration matters mainly for problems with trapped waves.
  • At convergence, the discrete error is about the same size as the stopping residual, so the scheduled truncation tolerance directly controls the achieved accuracy.

Reading between the lines

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

  • A hybrid solver that keeps full-rank blocks near the point source and low-rank blocks in the far field would likely extend the method to problems where the source region cannot be compressed, a direction the authors mention as future work.
  • If the bounded-rank observation holds at higher frequencies and larger domains, the method may offer a practical way to compress point-source Helmholtz fields beyond the setting of directional fast methods.
  • Because the paper validates the bounded-rank premise only empirically, a rigorous perturbation analysis connecting the truncation tolerance to the distance between the low-rank fixed point and the true Helmholtz solution would be the natural next step.
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 / 5 minor

Summary. The paper proposes LR-WaveHoltz, a low-rank implementation of the WaveHoltz fixed-point iteration for the Helmholtz equation with point sources. The wave equation is discretized with multiblock SBP-SAT finite differences, and the solution is compressed with an SVD in 2D and tensor trains in 3D. Rank growth is controlled by step truncation, and in 2D the WaveHoltz iteration is accelerated with low-rank Anderson acceleration. Numerical experiments cover free-space, half-space, and layered problems in 2D and 3D, and report residual convergence, rank behavior, runtimes, and compression gains. The paper is explicitly presented as a proof of concept, and it honestly notes that 2D gains are limited and that Anderson acceleration gives little improvement for open problems.

Significance. If the claimed behavior is robust, the paper makes a useful algorithmic contribution: it demonstrates that a WaveHoltz solver can be run entirely in low-rank SVD/TT format for point-source Helmholtz problems, with large memory and runtime gains in 3D. The numerical study is extensive and well organized, and the paper is commendably honest about the limitations of the 2D method and of Anderson acceleration. The main weakness is that the convergence claim rests on empirical observations — especially rank-boundedness by T_ε_k(W) — rather than on an analysis of the perturbed fixed-point map induced by step truncation. The 2D error comparisons against the full-rank WaveHoltz solution are reassuring, but the corresponding validation is missing in 3D, where the claimed gains are largest. This is a promising proof of concept that needs additional evidence or analysis for the fixed-point consistency before it can be accepted as a fully supported solver.

major comments (3)
  1. [§5, Algorithm 5, lines 23–26; §6.1.3, Fig. 8] The stopping residual ρ_{k+1}^{ij} = ||W^{k+1}_{ij} − W^{k}_{ij}|| is computed from truncated iterates only. Since each WaveHoltz application is followed by the truncation operator T_ε, the iteration actually solves a perturbed fixed-point problem. A small consecutive-iterate residual does not by itself imply that the untruncated WaveHoltz residual ||W* − Π(W*)|| is small, and rank-boundedness by T_ε_k(W) does not control this gap. The 2D experiments partially mitigate this by reporting the error to the full-rank WaveHoltz solution, but the paper should either add a perturbation/contractivity argument showing that the truncation tolerance ε_k controls the fixed-point shift, or add an explicit numerical comparison between the untruncated WaveHoltz residual and the stopping residual at the final LRWH iterate for representative 2D and 3D cases. Without this, the claim that LRWH converges to
  2. [§4.1, Algorithm 3; §5, lines 25–26] The step-truncation tolerance is scheduled as ε_{k+1}^{ij} = max{K, θhρ_{k+1}^{ij}} with a fixed positive floor K. Since K never vanishes, the truncated WaveHoltz operator has an O(K) perturbation that remains even as ρ_k → 0. The experiments choose K below ε* (e.g., K = 1e−5, ε* = 1e−3), so the practical effect is probably small, but no analysis or parameter study is given. Please justify the choice of K, state how K is selected relative to ε* in each experiment, and show that the converged solution is insensitive to K. This is load-bearing because it concerns whether the iteration actually converges to the true fixed point of the untruncated WaveHoltz operator.
  3. [§6.2, Figs. 16–19] The 3D TTWH experiments are the main motivation for the paper, but the validation there consists only of residual decay and rank stability under mesh refinement. No reference solution is computed, so it is possible that the iteration converges to a low-rank artifact of the truncated TT operator rather than to the Helmholtz solution. I request at least one small-scale 3D problem — e.g., a single block or a 2×1×1 partition at moderate resolution — where the TTWH solution can be compared with a full-rank WaveHoltz solution or an analytic Green's function. The 2D error-vs-residual comparisons in Figures 6, 9, 11, and 13 provide a good model; the same check should be supplied in 3D.
minor comments (5)
  1. [§4.2 vs Appendix A] The text in §4.2 states that fADI converges in exactly two iterations, while Lemma A.1 and the derivation in Appendix A use three iterations. Please reconcile this discrepancy.
  2. [Fig. 15 caption] The caption uses ǫ = √2θ/||G|| while the text defines δ = √2ε/||G|| for the TT rounding tolerance. Make the notation consistent.
  3. [References] Reference [35] and reference [29] appear to be the same arXiv preprint (Rotem, Runborg, Appelö, arXiv:2407.06929). Please merge or remove the duplicate.
  4. [Throughout] There are several typos: 'paritioned' in Figure 9 caption, 'correpsonding' in §6.2.2, 'the the overall magnitude' in §6.1.1, and 'Acelleration' in the keywords. A light proofreading pass is needed.
  5. [§6.1.3] The description of the full-rank reference solution (“performing an explicit inverse”) is vague. Provide the precise way the reference WaveHoltz solution is computed so the error comparison is reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LR-WaveHoltz is a new combination with independent full-rank benchmarks, and the self-citations are to external convergence theorems.

full rationale

Walked the claimed derivation chain. The WaveHoltz fixed-point identity Π[u,iωu]^T = [u,iωu]^T is true by construction of the kernel (cos ωt − 1/4) in Eq. (5), but the paper uses this as an equivalence, not as evidence of convergence: the actual convergence claim is explicitly delegated to [8,29], which are parameter-free convergence theorems with proofs in separate publications and are not fitted to the present experiments. The low-rank step-truncation (Algorithm 3) and TT rounding (Eq. 15) are standard rank-adaptive approximations; no target quantity is defined in terms of a fitted parameter. The Anderson acceleration is inherited from [25], but its effectiveness here is demonstrated numerically against LRWH without acceleration (Figs. 5–7, 9–14), so it is not an unverified self-citation chain. The empirical checks compare against full-rank SBP references and explicit inverse solutions, so the error claims are not self-referential. The residual schedule ε_k = max{K, θhρ_k} does couple the truncation tolerance to the residual, but it does not define the claimed Helmholtz accuracy; the same experiments also report direct discrete error. The open issue that step-truncation may shift the WaveHoltz fixed point is a correctness/approximation risk, not circularity, since the paper does not claim a theorem ruling out such a shift. No step reduces to its input by construction, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard numerical discretization ingredients, plus two assumptions that are not proven: that step-truncated WaveHoltz retains the full-rank fixed point, and that iterates stay low-rank. The free parameters are algorithmic tuning choices (tolerances, scheduling, stability parameters) and test-specific source/damping definitions; none are fitted to the target Helmholtz solution.

free parameters (7)
  • Truncation tolerance ε (wave-solver tolerance ε̃ = ε/(2N_t)) = ε = 1e-2 to 1e-5 h-scaled in 2D; δ = √2 ε / ||G|| in 3D
    Controls SVD/TT rank and accuracy at every time step and WaveHoltz update; scaling with h is chosen empirically so errors are grid independent.
  • Scheduling parameter θ = θ = 1 in 2D, θ = 0.5 in 3D
    Sets how aggressively truncation tolerances follow the current residual; chosen by hand to balance rank growth and convergence.
  • Floor constant K = K = 1e-5 in 2D, K = 1e-4 in 3D
    Prevents artificial rank inflation when the residual in a block is very small before the signal arrives; hand-tuned.
  • SBP penalty parameter τ = τ = 15
    Stability parameter in the SAT boundary treatment; chosen sufficiently large for the 4th-order SBP operators.
  • Time-step ratio Δt/h = 0.15
    Leap-frog time step relative to grid size; chosen for stability and accuracy of the wave solver.
  • Gaussian source width δ = δ = 1/(2ω)
    Width of the Gaussian approximating the point source in (17); test-specific and tied to frequency.
  • Initial and stopping tolerances ε_0, ε* = ε* = 1e-3; ε_0 not stated for 2D
    Stopping tolerance is taken from an external rule of thumb [36]; the initial blockwise truncation tolerance is an input to Algorithm 5 but its value for 2D examples is not reported.
assumptions (6)
  • domain assumption The SBP-SAT semidiscretization (11)–(12) is stable and satisfies an energy estimate when τ is sufficiently large.
    Invoked in §3.2, following [14], to guarantee the underlying full-rank wave solver is stable. The low-rank version inherits this only informally.
  • domain assumption The WaveHoltz filter operator Π has the stated frequency-selection property and the fixed-point iteration converges under natural assumptions.
    Used in §2.2 to justify that the full-rank WaveHoltz iteration converges to the Helmholtz solution; taken from [8,29].
  • ad hoc to paper Step-truncation with tolerance ε does not accumulate errors large enough to break convergence of the WaveHoltz iteration.
    No error analysis is given for Algorithms 3–5; the paper relies on residual plots to show that truncation does not prevent convergence.
  • ad hoc to paper The WaveHoltz iterates have low SVD/TT rank at all times, with rank bounded by the rank of the truncated full-rank solution.
    Central to the method; validated only empirically in the numerical sections (Figs. 8, 10, 12, 14). No proof relates solution regularity or source separation to iterate rank.
  • standard math TT rounding in the TT-Toolbox satisfies the relative tolerance bound (15).
    Standard property of TT rounding from [19]; used in §4.3 for all 3D compression operations.
  • domain assumption The Gaussian forcing (17) is an adequate model of a point source for the Helmholtz solutions studied.
    Used in all experiments instead of a Dirac delta; the width δ is tied to frequency but no convergence with δ→0 is shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LR-WaveHoltz: A Low-Rank Helmholtz Solver." pith.science (2026). https://pith.science/paper/LJGQPP5U

@misc{pith2026251009352,
  author       = {Pith},
  title        = {Pith review of: LR-WaveHoltz: A Low-Rank Helmholtz Solver},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LJGQPP5U}},
  note         = {Machine review of arXiv:2510.09352}
}
read the original abstract

We propose a low-rank method for solving the Helmholtz equation. Our approach is based on the WaveHoltz method, which computes Helmholtz solutions by applying a time-domain filter to the solution of a related wave equation. The wave equation is discretized by high-order multiblock summation-by-parts finite differences. In two dimensions we seek to compress the solution in matrix form, and in three dimensions using tensor trains. To control rank growth we use step-truncation during time stepping and a low-rank Anderson acceleration for the WaveHoltz fixed point iteration. We have carried out extensive numerical experiments demonstrating the convergence and efficacy of the iterative scheme for free- and half-space problems in two and three dimensions with constant and piecewise constant wave speeds.

Figures

Figures reproduced from arXiv: 2510.09352 by the authors.

Figure 1
Figure 1. (a) An illustration of a multiblock partitioning o [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. The free-space Green’s function G(x, y) of the Helmholtz equation centered at (−0.1, 0.5) throughout the domain [0, 5] × [0, 1]. Degrees of freedom PPW 512 20 1012 40 2012 80 4012 160 8012 320 16012 640 [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Ranks of the truncated Greens function Tǫ(G(X, Y )) in each of the 5×1 blocks using the the truncation tolerances ǫ = 10−3h −1 , ǫ = 10−4h −1 and ǫ = 10−5h −1 for varying number of points per wavelength. Points per wavelength Runtime [sec] 40 80 160 320 10-2 100 40 80 160 320 40 80 160 320 40 80 160 320 40 80 160 320 Points per wavelength Error 40 80 160 320 10-7 10-5 10-4 10-3 10-2 40 80 160 320 40 80 160 320 40 80… view at source ↗
Figures from the paper (18 more)
Figure 4
Figure 4. Figure 4: Time to simulate hundred time steps in each block us [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Residuals obtained using the LRWH and LRAA(4) meth [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 3
Figure 3. Figure 3: The errors stagnate at higher level for higher tole [PITH_FULL_IMAGE:figures/full_fig_p016_3.png]
Figure 6
Figure 6. Figure 6: Free-space problem. The resulting Helmholtz solu [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Free-space problem. The total AA(M) residuals [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Free-space problem. Distribution of ranks for the [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Half-space problem. The resulting Helmholtz solu [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Half-space problem. Distribution of ranks for th [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Layered free-space problem. The resulting Helmh [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Layered free-space problem. Distribution of ran [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: Layered half-space problem. The resulting Helmh [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Layered half-space problem. Distribution of ran [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]
Figure 15
Figure 15. Figure 15: Runtimes for taking a time step in each block (top) [PITH_FULL_IMAGE:figures/full_fig_p024_15.png]
Figure 16
Figure 16. Figure 16: Free-space problem with unit wave speed solved us [PITH_FULL_IMAGE:figures/full_fig_p024_16.png]
Figure 17
Figure 17. Figure 17: Free-space problem with unit wave speed solved us [PITH_FULL_IMAGE:figures/full_fig_p025_17.png]
Figure 18
Figure 18. Figure 18: Mixed reflecting and outflow boundary conditions w [PITH_FULL_IMAGE:figures/full_fig_p025_18.png]
Figure 19
Figure 19. Figure 19: Half-space problem with unit wave speed solved us [PITH_FULL_IMAGE:figures/full_fig_p026_19.png]
Figure 20
Figure 20. Figure 20: Moving point source. Average TT-ranks of the nume [PITH_FULL_IMAGE:figures/full_fig_p026_20.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 1 linked inside Pith

  1. [1]

    J. M. Hovem, Ray trace modeling of underwater sound propa gation. Documentation and use of the PlaneRay model, IntechOpen, 2011

  2. [2]

    C. L. Pekeris, Theory of propagation of explosive sound i n shallow water (1948)

  3. [3]

    F. D. Tappert, The parabolic approximation method, Wave propagation and underwater acous- tics (2005) 224–287

  4. [4]

    O. G. Ernst, M. J. Gander, Why it is difficult to solve Helmho ltz problems with classical iterative methods, Numerical analysis of multiscale probl ems (2011) 325–363

  5. [5]

    Y. A. Erlangga, Advances in iterative methods and precon ditioners for the Helmholtz equation, Archives of Computational Methods in Engineering 15 (2008) 37–66

  6. [6]

    Kreiss, J

    H.-O. Kreiss, J. Oliger, Comparison of accurate methods for the integration of hyperbolic equations, Tellus 24 (3) (1972) 199–215

  7. [7]

    Hagstrom, G

    T. Hagstrom, G. Hagstrom, Grid stabilization of high-or der one-sided differencing ii: Second- order wave equations, Journal of Computational Physics 231 (23) (2012) 7907–7931

  8. [8]

    Appel¨ o, F

    D. Appel¨ o, F. Garcia, O. Runborg, Waveholtz: Iterative solution of the Helmholtz equation via the wave equation, SIAM Journal on Scientific Computing 4 2 (4) (2020) A1950–A1983

Show all 37 references
  1. [9]

    Appel¨ o, F

    D. Appel¨ o, F. Garcia, A. A. Loya, O. Runborg, El-WaveHol tz: A time-domain iterative solver for time-harmonic elastic waves, Computer Methods in Appli ed Mechanics and Engineering 401 (2022) 115603

  2. [10]

    Z. Peng, D. Appel¨ o, EM-WaveHoltz: A flexible frequency -domain method built from time- domain solvers, IEEE Transactions on Antennas and Propagat ion 70 (7) (2022) 5659–5671

  3. [11]

    Kreiss, G

    H.-O. Kreiss, G. Scherer, Finite element and finite differ ence methods for hyperbolic partial differential equations, in: Mathematical aspects of finite el ements in partial differential equa- tions, Elsevier, 1974, pp. 195–212

  4. [12]

    M. H. Carpenter, D. Gottlieb, S. Abarbanel, Time-stabl e boundary conditions for finite- difference schemes solving hyperbolic systems: methodology and application to high-order compact schemes, Journal of Computational Physics 111 (2) ( 1994) 220–236

  5. [13]

    Mattsson, F

    K. Mattsson, F. Ham, G. Iaccarino, Stable boundary trea tment for the wave equation on second-order form, Journal of Scientific Computing 41 (2009 ) 366–383

  6. [14]

    Mattsson, F

    K. Mattsson, F. Ham, G. Iaccarino, Stable and accurate w ave-propagation in discontinuous media, Journal of Computational Physics 227 (19) (2008) 875 3–8767

  7. [15]

    Bachmayr, Low-rank tensor methods for partial differe ntial equations, Acta Numerica 32 (2023) 1–121

    M. Bachmayr, Low-rank tensor methods for partial differe ntial equations, Acta Numerica 32 (2023) 1–121

  8. [16]

    O. Koch, C. Lubich, Dynamical low-rank approximation, SIAM Journal on Matrix Analysis and Applications 29 (2) (2007) 434–454

  9. [17]

    Dektor, A

    A. Dektor, A. Rodgers, D. Venturi, Rank-adaptive tenso r methods for high-dimensional non- linear pdes, Journal of Scientific Computing 88 (2) (2021) 36 . 28

  10. [18]

    Appel¨ o, Y

    D. Appel¨ o, Y. Cheng, Robust implicit adaptive low rank time-stepping methods for matrix differential equations, Journal of Scientific Computing 102 ( 3) (2025) 81

  11. [19]

    I. V. Oseledets, Tensor-train decomposition, SIAM Jou rnal on Scientific Computing 33 (5) (2011) 2295–2317

  12. [20]

    S. V. Dolgov, B. N. Khoromskij, I. V. Oseledets, Fast sol ution of parabolic problems in the ten- sor train/quantized tensor train format with initial appli cation to the Fokker–Planck equation, SIAM Journal on Scientific Computing 34 (6) (2012) A3016–A30 38

  13. [21]

    S. V. Dolgov, A. P. Smirnov, E. Tyrtyshnikov, Low-rank a pproximation in the numerical modeling of the Farley–Buneman instability in ionospheric plasma, Journal of Computational Physics 263 (2014) 268–282

  14. [22]

    Bebendorf, W

    M. Bebendorf, W. Hackbusch, Existence of H-matrix approximants to the inverse FE-matrix of elliptic operators with l∞-coefficients, Numerische Mathematik 95 (1) (2003) 1–28

  15. [23]

    Engquist, H

    B. Engquist, H. Zhao, Approximate separability of the G reen’s function of the Helmholtz equation in the high frequency limit, Communications on Pur e and Applied Mathematics 71 (11) (2018) 2220–2274. doi:10.1002/cpa.21755

  16. [24]

    Engquist, L

    B. Engquist, L. Ying, Fast directional multilevel algo rithms for oscillatory kernels, SIAM Journal on Scientific Computing 29 (4) (2007) 1710–1737

  17. [25]

    Appel¨ o, Y

    D. Appel¨ o, Y. Cheng, lrAA: Low-Rank Anderson Accelera tion, arXiv preprint arXiv:2503.03909 (2025)

  18. [26]

    F. B. Jensen, W. A. Kuperman, M. B. Porter, H. Schmidt, A. Tolstoy, Computational ocean acoustics, Vol. 2011, Springer, 2011

  19. [27]

    Hagstrom, Radiation boundary conditions for the num erical simulation of waves, Acta numerica 8 (1999) 47–106

    T. Hagstrom, Radiation boundary conditions for the num erical simulation of waves, Acta numerica 8 (1999) 47–106

  20. [28]

    Appel¨ o, S

    D. Appel¨ o, S. Wang, An energy-based discontinuous gal erkin method for coupled elasto- acoustic wave equations in second-order form, Internation al Journal for Numerical Methods in Engineering 119 (7) (2019) 618–638

  21. [30]

    Y. Yang, A. Townsend, D. Appel¨ o, Anderson acceleratio n based on the h- s sobolev norm for contractive and noncontractive fixed-point operators, Jou rnal of Computational and Applied Mathematics 403 (2022) 113844

  22. [31]

    Mattsson, J

    K. Mattsson, J. Nordstr¨ om, Summation by parts operato rs for finite difference approximations of second derivatives, Journal of Computational Physics 19 9 (2) (2004) 503–540

  23. [32]

    J. E. Hicken, D. W. Zingg, Summation-by-parts operator s and high-order quadrature, Journal of Computational and Applied Mathematics 237 (1) (2013) 111 –125

  24. [33]

    Almquist, S

    M. Almquist, S. Wang, J. Werpers, Order-preserving int erpolation for summation-by-parts operators at nonconforming grid interfaces, SIAM Journal o n Scientific Computing 41 (2) (2019) A1201–A1227. 29

  25. [34]

    Benner, R.-C

    P. Benner, R.-C. Li, N. Truhar, On the adi method for sylv ester equations, Journal of Com- putational and Applied Mathematics 233 (4) (2009) 1035–104 5

  26. [35]

    Rotem, O

    A. Rotem, O. Runborg, D. Appel¨ o, Convergence of the sem i-discrete waveholtz iteration, arXiv:2407.06929 (2024)

  27. [36]

    Appel¨ o, J

    D. Appel¨ o, J. W. Banks, W. D. Henshaw, D. W. Schwendeman , A rule of thumb for choosing points-per-wavelength for finite difference approximations of Helmholtz problems, submitted (2025)

  28. [37]

    Sabino, Solution of large-scale Lyapunov equations via the block modified Smith method, Ph.D

    J. Sabino, Solution of large-scale Lyapunov equations via the block modified Smith method, Ph.D. thesis, Rice University, Houston (2006)

  29. [38]

    K. B. Petersen, M. S. Pedersen, et al., The matrix cookbo ok, Technical University of Denmark 7 (15) (2008) 510. A Enforcing time dependent boundary conditions In this section, we provide a short introduction to the metho d used to impose the outflow conditions in matrix form. ...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.