Pith. sign in

REVIEW 3 major objections 6 minor 15 references

Nodal AMG Coarsening and Interpolation for PDE Systems

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read An algebraic AMG coarsening for PDE systems preserves near-kernel components by averaging along length-two paths of a nodal dual problem, and matches geometric re-discretization when the paths align with refinement.

desk verdict A useful new coarsening idea for AMG on H(curl)/H(div)/Stokes systems, with honest accounting of the parts that are still heuristic; worth a serious referee. read the letter →

arxiv 2501.16248 v1 pith:MQRHKPPS submitted 2025-01-27 math.NA cs.NA

classification math.NAcs.NA MSC 65N5565F1065N30
keywords algebraicmultigridcompatiblerelaxationidealinterpolationH(curl)H(div)Stokessystemnear-kernelapproximationnodalcoarsening
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

This paper tries to make algebraic multigrid work for systems of PDEs whose matrices have large near-nullspaces—the curl-curl, grad-div, and Stokes problems—where ordinary pointwise smoothers and injection coarsening fail. Its proposal is to coarsen a nodal 'dual' problem built from an automatically computed near-kernel matrix and then to form each coarse variable as an orientation-consistent average of two fine unknowns along a short path, which preserves the near-kernel components on the coarse grid without using any geometric mesh information. If this works, AMG for these systems becomes a practical two-grid method with interpolation that is sparser than the ideal harmonic extension, and on structured meshes it exactly reproduces the coarse-grid re-discretization one would write down geometrically.

What carries the argument

The load-bearing mechanism is Algorithm 4.1, which works as follows: the near-kernel finder Algorithm 3.1 (local eigen-decompositions on diameter-$m$ sets, filtered by an eigenvalue threshold) supplies a matrix $N$ of near-kernel vectors; the algorithm forms the nodal dual matrix $A_N = N^T A N$ and coarsens it with a classical C/F splitting; then for every pair of coarse nodes that are distance-two neighbors in the graph of $A_N$, it picks one mutual distance-one neighbor and selects two fine DoFs forming a length-two path, adding a row to $R^T$ that has $\pm 1$ entries at those DoFs with signs fixed by $N$. The remaining DoFs become the fine basis $S = [I, 0]^T$, and the interpolation is the local harmonic extension $P_\star = (I - S(S^T A S)^{-1} S^T A) R^T$. The property that matters is that this path averaging keeps the orientation information of the underlying near-kernel in the coarse space, something injection coarsening cannot do when near-kernels overlap.

What would settle it

Run the two-grid solver on a uniform triangular mesh with Dirichlet boundary conditions for the curl-curl problem at $\beta = 0.01$ and watch the convergence rate as the mesh is refined: the paper's own result shows the rate settling at 0.94, so if the rate does not improve once boundary-adjacent coarse variables are added, the claim that the algebraic coarse space always contains the full near-kernel is false in that setting.

Watch

Extended reading notes

Core claim

The central claim is that the near-kernels of H(curl), H(div), and Stokes systems can be preserved algebraically in an AMG coarse space by running a classical C/F splitting on the nodal dual matrix $N^T A N$, where $N$ is a matrix of local near-kernel vectors, and then defining each coarse variable as the sum of two fine DoFs that lie on a length-two path between coarse nodes, with signs taken from the corresponding rows of $N$. This construction decouples the fine and coarse variables completely, so the resulting ideal interpolation $P_\star = (I - S(S^T A S)^{-1} S^T A) R^T$ is sparse and local. When the path selection coincides with geometric mesh refinement, the algebraic interpolation reproduces the geometric interpolation (and hence re-discretization) for structured and unstructured meshes alike, and the same machinery applies to the Stokes saddle-point system by identifying it with a curl-curl problem on a dual mesh.

Load-bearing premise

The whole construction assumes the local near-kernel finder (Algorithm 3.1) catches every smooth mode the smoother cannot remove, and that coarsening $N^T A N$ with length-two paths preserves every one of those modes on the coarse grid; the Dirichlet-triangle test (rate 0.94) shows this can fail near boundaries.

Editorial extensions

If this is right

  • AMG setup for H(curl), H(div), and grad-div systems reduces to a classical AMG coarsening of the nodal dual matrix $N^T A N$, so no geometric mesh information is required.
  • The coarse and fine variables are completely decoupled, giving an interpolation that is sparser than the global ideal harmonic extension while still preserving near-kernels.
  • Wherever the algebraic length-two paths agree with the geometric refinement pattern, the method reproduces the re-discretized coarse operator, so convergence matches the geometric multigrid baseline.
  • The Stokes system can be coarsened with the same rules by treating pressure as nodal and velocity as edge degrees of freedom on a dual mesh; the coarse velocities then preserve the discrete divergence constraint.
  • Compatible relaxation on the selected coarse variables predicts the two-grid convergence rate, making CR a practical monitor for coarsening quality.

Reading between the lines

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

  • The Dirichlet-triangle result (rate 0.94) suggests a cheap remedy the paper leaves implicit: run CR globally, locate regions where the coarse variables miss the near-kernel (near boundaries), and add extra coarse variables there before building interpolation.
  • The path selection in Algorithm 4.1 picks one arbitrary length-two path when several exist; one could replace this with a matching optimization minimizing the energy of the resulting interpolation, which may improve robustness on very unstructured meshes.
  • Because the method identifies curl-curl, grad-div, and Stokes coarsening, the same nodal-dual construction may transfer to other de Rham complex discretizations, such as H(div)-conforming magnetostatics or nearly incompressible elasticity, where the near-kernel is a gradient or curl of a smaller space.
  • The equivalence between path-based coarse variables and geometric refinement could be turned into an algebraic test: the degree to which the algebraic coarsening matches re-discretization measures how 'geometric' an unstructured mesh actually is.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes an algebraic multigrid (AMG) coarsening and interpolation scheme for symmetric positive definite systems with large near-kernels, targeting H(curl), H(div), and Stokes-type PDE systems. The setup has three stages: (i) an automatic smoother construction that detects local near-kernels N via local eigen-decompositions (Algorithm 3.1, based on prior work); (ii) a coarsening stage that performs a classical C/F splitting on the nodal dual matrix N^T A N and then builds coarse variables R and fine variables S by forming oriented length-two paths between coarse nodes, with signs taken from the rows of N (Algorithm 4.1); and (iii) an interpolation stage that computes a sparse approximation of the ideal interpolation operator P* = (I - S(S^T A S)^{-1} S^T A) R^T and uses the Galerkin coarse operator. The authors claim that, when the coarse variables align with the geometric refinement, the method reproduces re-discretization on unstructured meshes. Numerical experiments on uniform quadrilateral and triangular meshes for curl-curl problems (periodic and Dirichlet boundary conditions) and on a periodic MAC Stokes problem show convergence rates that are competitive with geometric interpolation and far better than classical pointwise ideal interpolation, although the Dirichlet triangle case has a slow asymptotic rate of 0.94.

Significance. If the proposed construction works as broadly as claimed, it would be a valuable contribution: it provides a genuinely algebraic way to build coarse variables that preserve the near-kernel structure of H(curl), H(div), and Stokes systems, without explicit geometric mesh information, and it packages this with an automatic smoother and a complete two-grid method. The idea of coarsening the nodal dual problem N^T A N and then defining coarse variables as oriented averages along paths is concrete and novel, and the numerical evidence on periodic quad and Stokes problems is encouraging. The paper also correctly uses compatible relaxation as a diagnostic: the reported CR rates match the observed two-grid rates, including the degraded Dirichlet case, which lends credibility to the measurement methodology. However, the strength of the central claim in the abstract is not matched by the evidence: no unstructured mesh is tested, the one boundary-sensitive test degrades substantially, and two load-bearing steps rest on numerical observation rather than analysis.

major comments (3)
  1. [Algorithm 4.1 and Section 4.3] The arbitrary length-two path selection is a load-bearing step that is not justified. The paragraph in Section 4.3 states that 'If multiple paths exist between them, we only pick one of them based on the labeling order for simplicity.' The coarse variable constructed in line 9 of Algorithm 4.1 preserves a near-kernel component only if the chosen path carries the correct orientation information; no argument is given that the labeling-order path does so, and no characterization is given of the configurations in which such a path exists. The paper's own Dirichlet triangle experiment (Section 6.1, Figure 6.1b) demonstrates that the construction can miss near-kernel components near boundaries, producing an asymptotic two-grid rate of 0.94, and the text attributes this precisely to missing paths to boundary coarse nodes. This is direct evidence that the path-selection heuristic can fail in realistic settings. Consequently, the abstract's claim that the method 'reproduces re-discretization on unstructured meshes' is not established by the presented construction or experiments.
  2. [Section 4.2, Eq. (4.4)-(4.5)] Dropping S_E from the coarse variables is justified only by a numerically computed post-scaling transformation, not by analysis. The text says: 'we find that we can completely drop the SE term from the definition of R and that the range of this approximation and that of using RE and SE defined in (4.2) are the same. Though we have not found an appropriate analytical form, we are able to compute a post-scaling transformation numerically that establishes the equivalence ... for a set of example problems and coarse grids.' This is a central algorithmic step because the final choices R := REG and S := SI do not form the full GAMG splitting assumed in Section 1 (Eq. (1.4) with RS=0, S^T S=I, R R^T=I); the paper itself notes that the splitting is only partial, with dim(span(R^T)) + dim(span(S)) <= n. The convergence theory in Section 1, including Eq. (1.3) and the ideal interpolation formula (1.5), relies on the full-splitting orthogonality conditions. Without an analysis of the range equivalence or of the error introduced by the partial splitting, the theoretical support for the final interpolation operator is incomplete, and the claim that the method follows the GAMG ideal-interpolation framework is only heuristic.
  3. [Abstract and Section 6] The numerical experiments do not support the unstructured-mesh claim in the abstract. All curl-curl experiments in Section 6 are on uniform quadrilateral meshes with periodic boundary conditions or on a uniform triangular mesh with a straight Dirichlet boundary; the Stokes experiment is also on a uniform periodic quadrilateral MAC mesh. No genuinely unstructured mesh (variable topology, irregular node distribution, curved or reentrant boundaries) is tested. Moreover, the only non-periodic test, the Dirichlet triangle in Figure 6.1b, shows a notably degraded rate (0.94) that the authors attribute to missing near-kernel components near the boundary. The abstract states that 'when the coarse variables align with the geometric refinement, our method reproduces re-discretization on unstructured meshes,' but the evidence only supports a much weaker statement: on structured meshes where the algorithm happens to recover geometric refinement, the two-grid rate is close to the geometric interpolation rate. The claim should either be substantiated with unstructured-mesh experiments or qualified accordingly.
minor comments (6)
  1. [Section 1, Eq. (1.3)] The matrix R is used in the expression K(P R) before it is defined; the definition of R as a restriction matrix with RP = I_c appears only after Eq. (1.3). Please introduce R before its first use.
  2. [Algorithm 4.1, line 9] The notation for constructing the column of R^T is difficult to parse: the expression 'RT = [RT · · · 0 [±1]DoF1 · · · 0 · · · [±1]DoF2 0 · · ·]^T' mixes a recursive appearance of R^T with the new column. A clearer notation, e.g., appending a column vector with two nonzero entries, would aid reproducibility.
  3. [Section 4.2, paragraph after Eq. (4.4)] The claim that a post-scaling transformation establishes range equivalence 'for a set of example problems and coarse grids' is not accompanied by any details, table, or figure documenting those examples. Since this is a key step, at least a concise summary of the verification should be provided.
  4. [Section 5] The sentence 'we focus only the primary global Stokes system' appears to be missing the word 'on' and should read 'we focus only on the primary global Stokes system.' The identity 'Ae = −∆ + βI = ∇ × ∇ × −∇∇ · +βI' also appears typographically garbled; the second term should presumably be '-∇∇·'.
  5. [Section 6.1, Figure 6.1] The text states that the asymptotic rate in Figure 6.1b is 0.94, but this value is not directly annotated in the plot; please add a label or a text reference so the reader can verify the number against the curve.
  6. [Section 7, Conclusions] The concluding sentence describing future work ('use CR to identify inadequately coarsened regions and add coarse variables iteratively') essentially acknowledges that the current algorithm, without adaptive enrichment, is not fully robust near boundaries. This is a useful limitation statement, but it should be moved to or echoed in the numerical results section so that the boundary degradation is presented as a known limitation of the current method, not only as future work.

Circularity Check

1 steps flagged · score 2.0 of 10

No material circularity: near-kernel preservation is a design tautology because Algorithm 4.1 builds R from N, but the convergence-rate and CR measurements are independent; the arbitrary path selection is an acknowledged unproven assumption, not a circular step.

  1. self definitional [Section 4.2, after Eq. (4.3); Algorithm 4.1, lines 5-9]
    "Here, the + or − must be consistent with the underlying orientations in the fine grid near kernels. ... The coarse variables now preserve the near kernels (grad, curl and/or div) exactly."

    Algorithm 4.1 constructs each coarse row of R by selecting two fine DoFs and placing ±1 entries whose signs are taken from the near-kernel matrix N ('the orientations of the near kernel for the fine DoFs can be determined from the corresponding rows of N', Section 4.3). Hence the statement that the coarse variables 'preserve the near kernels exactly' is a direct consequence of how R was defined, not an independently derived prediction or a first-principles result. The paper does not use this tautology as its evidence: the two-grid convergence rates, CR diagnostics, and comparisons to geometric interpolation and P♯ are measured independently. The circularity is therefore minor and does not affect the central numerical claims.

full rationale

The central derivation chain is otherwise self-contained. Algorithm 4.1 builds coarse variables by averaging fine DoFs along length-two paths, with signs read from the near-kernel matrix N, so any claim that the coarse space contains N is true by construction. I flag this as a minor self-definitional design tautology, not as a discovered prediction. The paper's main evidence is independent: two-grid convergence rates in Section 6 are compared against geometric interpolation, classical P⋆, and the optimal P♯; the CR rates are a theoretical diagnostic on the same coarse variables, not a fit to the measured rates. The abstract's re-discretization claim is empirical and explicitly conditional on 'when the coarse variables align with the geometric refinement', and the paper itself documents the failure mode on the Dirichlet triangular mesh (rate 0.94) due to missing near-kernel components near the boundary, which is an acknowledged correctness limitation rather than a circular step. The GAMG theory in references [6,7,1] is cited as a framework and is not used to force the new numerical results; no uniqueness theorem is imported from the authors' prior work, and no load-bearing conclusion reduces to a self-citation chain. The arbitrary choice of one path in Algorithm 4.1 ('we only pick one of them based on the labeling order for simplicity') is an unproven assumption and a correctness risk, but it is not circular. Overall the paper is externally benchmarked and its central content is independent; score 2 reflects the minor design tautology only.

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

The paper introduces no new physical entities. It relies on several algorithmic parameters (m, epsilon, damping, a numerically fitted transformation) and on domain assumptions about the completeness of the near-kernel matrix, the validity of nodal dual coarsening, and the Stokes reduction to curl-curl. The central claim depends on these assumptions, none of which are proven in the paper.

free parameters (4)
  • coarsening factor m = not stated
    Used in Algorithm 3.1 to define diameter-m neighborhoods for local near-kernel detection; the value is not reported in the numerical experiments.
  • near-kernel filter threshold epsilon = not stated
    Algorithm 3.1 uses epsilon to select near-kernel modes; its value is not specified.
  • L1-Jacobi damping factor = 0.5
    Section 6.1: smoothing uses L1-Jacobi with damping 0.5; this is a hand-set parameter, not tuned per problem.
  • post-scaling transformation = not provided
    Section 4.2: a post-scaling transformation is computed numerically to demonstrate range equivalence between interpolation with and without S_E; this transformation is not derived or reported.
assumptions (4)
  • standard math GAMG two-grid convergence theory applies to the constructed method
    The method is built on this theory (Section 1, Eqs. 1.3-1.6), but the paper notes the splitting is only partial, so the theory's orthogonality conditions are not strictly met.
  • domain assumption The near-kernel matrix N from Algorithm 3.1 captures the relevant near-nullspace of the system
    Section 3 and the two-grid setup: N is computed from local eigen-decompositions; if it misses modes (e.g., near boundaries), interpolation degrades (Section 6.1).
  • domain assumption Coarsening the nodal dual problem N^T A N with classical AMG yields the correct coarse nodes
    Section 4.3 Step 2: the algorithm first coarsens N^T A N; the quality of this coarsening controls the coarse grid quality (Section 7).
  • domain assumption For Stokes, the block structure and the relation A_N = G^T A_e G justify reducing coarsening to a curl-curl problem
    Section 5: the equivalence relies on assumptions on boundary conditions (periodic) and the discrete gradient G; the paper notes Hiptmair and Scott results for no-slip and Freudenthal meshes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nodal AMG Coarsening and Interpolation for PDE Systems." pith.science (2026). https://pith.science/paper/MQRHKPPS

@misc{pith2026250116248,
  author       = {Pith},
  title        = {Pith review of: Nodal AMG Coarsening and Interpolation for PDE Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MQRHKPPS}},
  note         = {Machine review of arXiv:2501.16248}
}
read the original abstract

We present an approach to constructing a practical coarsening algorithm and interpolation operator for the algebraic multigrid (AMG) method, tailored towards systems of partial differential equations (PDEs) with large near-kernels, such as H(curl) and H(div). Our method builds on compatible relaxation (CR) and the ideal interpolation model within the generalized AMG (GAMG) framework but introduces several modifications to define an AMG method for PDE systems. We construct an interpolation operator through a coarsening process that first coarsens a nodal dual problem and then builds the coarse and fine variables using a matching algorithm. Our interpolation follows the ideal formulation; however, we enhance the sparsity of ideal interpolation by decoupling the fine and coarse variables completely. When the coarse variables align with the geometric refinement, our method reproduces re-discretization on unstructured meshes. Together with an automatic smoother construction scheme that identifies the local near kernels, our approach forms a complete two-grid method. Finally, we also show numerical results that demonstrate the effectiveness of this interpolation scheme by applying it to targeted problems and the Stokes system.

Figures

Figures reproduced from arXiv: 2501.16248 by the authors.

Figure 4.1
Figure 4.1. Curl-curl on uniform quadrilateral mesh. [PITH_FULL_IMAGE:figures/full_fig_p005_4_1.png] view at source ↗
Figure 4.2
Figure 4.2. Uniform triangular mesh with Dirichlet boundary conditions [PITH_FULL_IMAGE:figures/full_fig_p007_4_2.png] view at source ↗
Figure 6.1
Figure 6.1. curl-curl and Stokes tests simplicity. The convergence rate using the classical P⋆ also deteriorates with increas￾ing problem size. While our interpolation by Algorithm 4.1 still achieves optimal convergence, the asymptotic rate is slow at 0.94. This is related to our discussion of [PITH_FULL_IMAGE:figures/full_fig_p009_6_1.png] view at source ↗
Figures from the paper (1 more)
Figure 6.2
Figure 6.2. Figure 6.2: Stokes: sparse approximation P Numerical tests show that in this case, gradients are preserved on the coarse grid, and the coarse grid error correc￾tion remains divergence-free. The cases where RT ’s are smoothed with Richard￾son are also provided, which results in a…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 14 canonical work pages

  1. [1]

    A. Ali, J. Brannick, K. Kahl, O. A. Krzysik, J. B. Schroder, and B. S. Southworth, Generalized optimal AMG convergence theory for nonsymmetric and indefinite problems , 2024, https://arxiv.org/abs/2401.11146, https://arxiv.org/abs/2401.11146

  2. [2]

    W. M. Boon, R. Hiptmair, W. Tonnon, and E. Zampa, H(curl)-based approximation of the stokes problem with slip boundary conditions , 2024, https://arxiv.org/abs/2407.13353

  3. [3]

    Brannick, F

    J. Brannick, F. Cao, K. Kahl, R. D. F algout, and X. Hu, Optimal interpolation and compatible relaxation in classical algebraic multigrid , SISC, 40 (2018), pp. A1473–A1493

  4. [4]

    Brezina and P

    M. Brezina and P. S. V assilevski, Smoothed aggregation spectral element agglomeration amg: SA-ρAMGe, in Proceedings of the 8th International Conference on Large-Scale Scientific Computing, LSSC’11, 2011, pp. 3–15

  5. [5]

    Claus, Multigrid smoothers for saddle point systems , PhD thesis, Universit¨ atsbibliothek Wuppertal, 2019

    L. Claus, Multigrid smoothers for saddle point systems , PhD thesis, Universit¨ atsbibliothek Wuppertal, 2019

  6. [6]

    R. D. F algout and P. S. V assilevski, On generalizing the AMG framework , SINUM, 42 (2004), pp. 1669–1693. UCRL-JC-150807

  7. [7]

    R. D. F algout, P. S. V assilevski, and L. T. Zikatanov, On two-grid convergence estimates, 10 Numer. Linear Algebra Appl., 12 (2005), pp. 471–494. UCRL-JRNL-203843

  8. [8]

    P. E. F arrell, L. Mitchell, and L. R. Scott, Two conjectures on the stokes complex in three dimensions on freudenthal meshes , SISC, 46 (2024), pp. A629–A644

Show all 15 references
  1. [9]

    Hiptmair, Multigrid method for maxwell’s equations , SINUM, 36 (1998), pp

    R. Hiptmair, Multigrid method for maxwell’s equations , SINUM, 36 (1998), pp. 204–225

  2. [10]

    Hiptmair, Finite elements in computational electromagnetism , Acta Numerica, 11 (2002), pp

    R. Hiptmair, Finite elements in computational electromagnetism , Acta Numerica, 11 (2002), pp. 237–339

  3. [11]

    Lee and C

    B. Lee and C. Tong, A novel algebraic multigrid-based approach for maxwell’s equations , LLNL Technical Report - UCRL-JC-218750, (2009)

  4. [12]

    J. C. Nedelec, Mixed finite elements in r3 , Numerische Mathematik, 35 (1980), pp. 315–341

  5. [13]

    Reitzinger and J

    S. Reitzinger and J. Sch¨oberl, An algebraic multigrid method for finite element discretiza- tions with edge elements , Numerical Linear Algebra with Applications, 9 (2002), pp. 223– 238

  6. [14]

    J. W. Ruge and K. St¨uben, Algebraic multigrid (AMG) , in Multigrid Methods, S. F. Mc- Cormick, ed., Frontiers Appl. Math., SIAM, Philadelphia, 1987, pp. 73–130

  7. [15]

    Voronin, S

    A. Voronin, S. MacLachlan, L. N. Olson, and R. Tuminaro, Monolithic algebraic multigrid preconditioners for the stokes equations , 2024, https://arxiv.org/abs/2306.06795. 11

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.