Accuracy Controlled Structure-Preserving ${\cal H}^2$-Matrix-Matrix Product in Linear Complexity with Change of Cluster Bases

Dan Jiao; Miaomiao Ma

REVIEW 3 major objections 3 minor 13 references

Accuracy Controlled Structure-Preserving ${\cal H}^2$-Matrix-Matrix Product in Linear Complexity with Change of Cluster Bases

T0 review · 3 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read An accuracy-controlled H2 matrix-matrix product can run in linear complexity.

desk verdict Real novelty, real gap: the new cluster-basis update is clever, but the paper asserts accuracy control without proving it, and the complexity claims outrun the data. read the letter →

arxiv 1908.05218 v1 pith:65LDJMT7 submitted 2019-08-14 math.NA cs.NA

Miaomiao Ma , Dan Jiao This is my paper · ORCID

classification math.NAcs.NA MSC 65F3065N38

keywords H2-matrixmatrix-matrixproductlinearcomplexityaccuracycontrolclusterbaseselectromagneticanalysisvolumeintegralequationhierarchicalmatrices

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 claims that the product of two H2-matrices—data-sparse matrices whose off-diagonal blocks are low-rank and share nested cluster bases—can be computed with explicitly controlled accuracy while keeping the same asymptotic complexity as the existing formatted product. The usual linear-complexity H2 matrix-matrix product fixes the cluster bases of the product in advance and can therefore be inaccurate when the true product content is not representable in those bases. Here the cluster bases are recomputed on the fly, level by level from leaf to root, from Gram-matrix sums of all relevant block products followed by singular-value truncation at a prescribed tolerance. If the claim is right, H2 multiplication becomes a dependable arithmetic primitive for fast solvers: linear time and memory for constant-rank matrices, and O(N log N) time with O(N) memory for electrically large 3-D volume integral equations whose ranks grow linearly with electrical size.

What carries the argument

The load-bearing mechanism is the instantaneous renewal of the cluster bases. A cluster basis is a low-rank factor V_t shared by all admissible blocks in a row or column cluster and nested through transfer matrices T. The algorithm computes, for each cluster, a Gram-matrix sum of the actual matrix content that would be produced by the different multiplication cases, adds the original-basis contribution, and then takes an SVD, keeping singular vectors above epsilon_trunc. This yields new bases that represent the product's own content rather than the factors' content, and the same construction at non-leaf levels yields new transfer matrices. Once the bases exist, each admissible product block is just a coupling matrix of rank size, so the traversal stays within O($k_l^{3}$) per block.

What would settle it

Compute an H2-matrix product on a family where the exact product has off-diagonal ranks growing faster than the input rank model, for example by multiplying two H2 matrices engineered so the product's admissible blocks have quadratic-in-level rank, and measure raw wall time and memory without normalizing by $C_sp^{2}$ or C_sp. If the observed scaling departs from O(N) or O(N log N) on that family, the complexity claim would fail. A cheaper check is to rerun the paper's 3-D cube-array suite and plot raw time and memory versus N without dividing out C_sp, since Table III already reports C_sp growing from 16 to 126.

Watch

Extended reading notes

Core claim

The central claim is that an H2 matrix-matrix product can be both accuracy-controlled and structure-preserving at linear complexity if, during a bottom-up traversal, the product's cluster bases and transfer matrices are regenerated rather than reused. For each row and column cluster, the algorithm accumulates Gram matrices of the four block-multiplication cases (full times full, full times low-rank, low-rank times full, low-rank times low-rank), normalizes the contributions, and takes an SVD with truncation tolerance epsilon_trunc to produce the new cluster basis. The same procedure at non-leaf levels produces new transfer matrices. With the new bases in hand, all four cases reduce to coupling-matrix computations of rank size, so each level costs O($C_sp^{2}$ 2^l $k_l^{3}$) in time and O(C_sp 2^l $k_l^{2}$) in memory. The paper demonstrates the algorithm on capacitance extraction matrices and on 2-D and 3-D scattering problems with over a million unknowns, reporting that the relative error tracks epsilon_trunc.

Load-bearing premise

The complexity bounds in equations (55) through (58) require that the truncation ranks k_l of the newly generated product bases grow no faster than the ranks of the input H2-matrices (constant for static capacitance extraction, linear in electrical size for 3-D volume integral equations), and that the per-cluster block count C_sp remains a bounded constant; the paper's own 3-D cube-array test shows C_sp rising from 16 to 126 as N grows, so the raw scaling curves must be read accordingly.

Editorial extensions

If this is right

H2 multiplication becomes a reliable building block for fast direct solvers: matrix inversion, LU factorization, and block solves can use it without the risk that formatted products silently lose accuracy.
For constant-rank H2 matrices, two matrices can be multiplied in O(N) time and memory, putting matrix-matrix products on the same asymptotic footing as H2 matrix-vector products.
For 3-D electrodynamic volume-integral equations, where rank grows linearly with electrical size, the product costs O(N log N) time and O(N) storage, allowing million-unknown products on a single core.
The product retains H2 form with rank-minimized cluster bases, so downstream operations on C = A times B do not need to convert formats.
Every multiplication and addition in the traversal is either exact or truncated with tolerance epsilon_trunc, giving a direct accuracy knob tied to a user-specified error.

Reading between the lines

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

A natural extension is to use the same Gram-matrix/SVD basis-renewal idea in other hierarchical formats such as H-matrix, HSS, or HODLR; the per-level cost would change, but the accuracy-control mechanism should transfer.
Because the new bases are built from the product content itself, the algorithm implicitly computes a near-optimal H2 approximation of the exact product; a direct dense-product comparison on small problems would quantify how close the SVD truncation comes to optimal.
The complexity guarantee is conditional on product-rank growth; if multiplying certain operator classes produces intrinsically higher-rank off-diagonal blocks, raw complexity would degrade, so an a priori or adaptive rank estimator for the product would strengthen the method.
The backward-split post-processing mirrors existing H2 backward transformations, so the new bases could be recycled into an accuracy-controlled H2 inverse or LU factorization by chaining this product with hierarchical inversion stages.

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.

Desk Editor's Note

read the letter

Two things to know about this paper. First, the core idea is genuinely new: instead of using fixed cluster bases during H2 matrix-matrix multiplication, it computes new nested bases from normalized Gram sums of the products and truncates via SVD to a user-set epsilon_trunc. That is a real advance over Borm's formatted multiplication and over the superlinear posteriori route. Second, the paper does not prove the central claim that this controls accuracy. Section V-A is a paragraph of assertion. No theorem links the truncation of Gram sums to the error of the assembled product. This is not a minor omission; it is the load-bearing claim of the paper. The stress-test note is right on this point. Truncating a Gram sum controls a weighted sum of projections, not worst-case error per block, and the recursive construction propagates lower-level truncations with no accounting. The numerical tables show errors that are consistently several times epsilon_trunc and that grow with N at fixed epsilon (Table IV), so the data don't remedy the gap.

What the paper does well: the algorithm is described at real equation-level detail, the nested-basis machinery is coherent, and the experiments are substantial—millions of unknowns on one core, with clear complexity trends for constant-rank problems. The comparison to existing formatted MMP is a genuine plus.

Soft spots, in proportion: - The complexity analysis treats C_sp and the rank growth model as given. Table III shows C_sp growing from 16 to 126 across their own cube-array suite, so the plotted "true scaling" is time/C_sp^2, not raw scaling. If C_sp grows with N, the claimed O(N log N) for 3D VIE isn't demonstrated. - The normalization of the Gram sums (before summation) is not specified; the outcome of SVD truncation depends on it. That's a reproducibility issue. - The accuracy control is the headline, and it is asserted. The error metric (59) uses an MVP reference that is itself exact only if the H2 MVP is exact; that's fine, but the absence of an error bound remains.

Who is this for? Researchers building fast direct solvers for integral equations, particularly in computational electromagnetics. They will want this method if the gap is closed. The paper deserves a serious referee; I'd send it to review, but the revision should be major: prove a global error bound or state a precise conjecture with supporting analysis, and show raw time/memory scaling for cases where C_sp and ranks vary.

My verdict: conditional, leaning positive on novelty, negative on the correctness of the accuracy-control claim as currently supported.

Referee Report

3 major / 3 minor

Summary. The manuscript proposes an H2-matrix-matrix product (MMP) algorithm that computes the cluster bases of the product matrix on the fly, by SVD-truncating sums of normalized Gram matrices formed from the four leaf-level and non-leaf-level multiplication cases. The claimed advantages are explicit accuracy control through a user-prescribed truncation parameter epsilon_trunc, preservation of the H2 structure, and linear (or O(N log N) for variable-rank 3-D problems) time and memory complexity. The algorithm is developed in detail in Sections III and IV, and numerical experiments for capacitance extraction and 2-D/3-D scattering up to about 1.5 million unknowns are reported in Section VI.

Significance. If fully established, the algorithm would be a useful contribution: an accuracy-controlled, structure-preserving H2-MMP with complexity matching the input rank model would remove a known limitation of existing formatted H2 multiplications, and the numerical results show credible scaling and monotone error reduction across large problems on a single core. The paper's strengths are the equation-level detail of the algorithm, the explicit treatment of all four multiplication cases at both leaf and non-leaf levels, and the large-scale experiments comparing against the existing formatted MMP. However, the central accuracy-control claim is not supported by a theorem or error bound, and the complexity analysis relies on unproved rank and sparsity-constant assumptions for the product matrix. These gaps are load-bearing and need to be addressed before the core claims can be accepted.

major comments (3)

[V-A, Eqs. (9)-(16), (40)-(47), Table IV] The central claim that epsilon_trunc controls the accuracy of the computed product is asserted in Section V-A but is not proved. The SVD truncation of a sum of normalized Gram matrices controls a weighted sum of squared projection errors over the sampled blocks, not the worst-case error of any individual admissible block or the global assembled error ||C_H2 - A_H2 B_H2||. Row and column bases are truncated independently, and the non-leaf bases in (40)-(47) are built from collected blocks that already contain lower-level truncation errors, so any error propagation is recursive and unaccounted for. No lemma or theorem links the truncation of these Gram matrices to a bound on the final product error. The numerical data are consistent with monotone improvement as epsilon_trunc shrinks, but they do not supply the missing bound: in Table IV the relative error at epsilon_trunc=1e-4 grows from 1.48e-3 to 8.09e-3 as N grows, and at epsilon_trunc=1e-2 it reaches 6.58e-2, which is larger than the nominal tolerance.
[V-B, Eqs. (53)-(58), Table III, Fig. 7] The complexity claims for variable-rank 3-D problems assume that the post-truncation ranks k_l of the product follow the same growth model as the input ranks from the authors' earlier work [5], and that C_sp is a bounded constant. Neither is proved for the product matrix C. In fact, Table III reports C_sp growing from 16 to 126 as N grows to about 1.5 million, so Fig. 7's normalization by C_sp^2 and C_sp removes a growth factor rather than demonstrating raw O(N log N) time and O(N) memory scaling. The raw scaling data should be shown and compared with the theoretical rates without dividing out the growth of C_sp, or an explicit bound on C_sp for the product matrices should be proved.
[III-C, Eqs. (12), (16), (43), (47)] The normalization applied to the Gram matrices before summation is not specified. The text says only that the matrices are normalized because their magnitudes may differ greatly, but it does not state which norm is used (Frobenius, spectral, trace-normalized, etc.). Different normalization choices produce different truncated subspaces and different actual errors for the same epsilon_trunc, and no analysis is provided for the effect of this weighting. This is a free parameter in the algorithm and should be fixed and justified if the accuracy-control claim is to be reproducible and meaningful.

minor comments (3)

[V-B] The sentence 'the cost of changing the cluster bases at the leaf level ... as can be seen from (13), (14), and (15)' appears to cite the column Gram matrix equations; the row cluster basis cost is described by (9)-(11), so the equation references should be corrected.
[Eq. (17), case-4] The expression 'BjSBj,k' is typeset without a separating space and could be misread as a single symbol; it should be written as 'B_j S^B_{j,k}' for clarity.
[Eqs. (9)-(16)] The notation O(C_sp) above the summation signs is nonstandard and potentially confusing; it should be stated explicitly that the number of terms is bounded by a constant multiple of C_sp or C_sp^2, respectively.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the cluster bases are computed adaptively from the product data via Gram-matrix SVDs, and the accuracy tolerance is a user input rather than a fitted parameter.

full rationale

The central derivation is self-contained in the sense required for circularity analysis. The new row and column cluster bases are obtained by truncated SVDs of Gram matrices built from the actual leaf- and non-leaf-level product blocks, e.g., Eqs. (9)-(16) and (40)-(47). This is adaptive low-rank compression of the product data itself, not a parameter fitted to a target quantity and then renamed as a prediction. The tolerance epsilon_trunc is a user-specified input, and the coupling-matrix formulas (17), (21), (48), and (52) exactly express each admissible block in the chosen bases once those bases are fixed. No equation in the paper defines the final product error in terms of epsilon_trunc, and the paper does not prove a bound linking the normalized Gram SVD truncation to the assembled product error; Section V-A merely asserts that 'the accuracy of the product is directly controlled by epsilon_trunc.' That is a missing error bound and a correctness/validation gap, not a circular reduction. The variable-rank complexity claims (57)-(58) import the rank-growth model from the authors' earlier work [5]. This is self-citation, but [5] is an external, falsifiable study of integral-operator ranks, and the present paper's target result, the accuracy-controlled H2 matrix-matrix product, is not an input to [5]. The paper does not prove that the post-truncation ranks of the product follow that model, and Table III shows C_sp growing with N; again, these are unverified assumptions or validation weaknesses rather than cases where a prediction reduces to its inputs by construction. Overall, no load-bearing circular step is identifiable in the derivation chain.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The ledger shows the paper introduces no new fitted constants or physical entities; the only free choice is the ad hoc equal-weight normalization of Gram matrices. The four axioms are standard H2 assumptions plus two unproven load-bearing assumptions: the rank behavior of the product bases and the accuracy of the SVD truncation heuristic.

free parameters (1)

Normalization weighting for Gram matrix sums
In (12) and (16), the component Gram matrices are normalized then summed with equal weight. The specific norm used for normalization is not stated, and the equal-weight choice is ad hoc; it affects the resulting cluster bases and hence the product accuracy.

assumptions (4)

domain assumption H2-matrix nested cluster bases and strong admissibility with bounded C_sp (from [1])
The algorithm presupposes the standard H2 framework: nested bases (3), block tree with bounded number of admissible neighbors C_sp per cluster. Table III shows C_sp growing in the 3D cube-array example, so boundedness is not automatic for the tested regime.
domain assumption Rank of the product's new cluster bases follows the input matrices' rank model (constant for electrostatic, linear in electrical size for 3D VIE from [5])
The complexity formulas (53)-(58) use k_l as the post-truncation rank of the product at level l, but no bound on these ranks is proven; the 3D VIE claim imports the rank-growth model from the authors' prior work [5].
domain assumption The exact product A*B is representable as an H2 matrix with the same block cluster tree as A and B
The algorithm assigns each product to an admissible or inadmissible block of the prescribed C structure. It is assumed that the product's off-diagonal content is captured by this partition plus the newly generated low-rank bases; no fill-in analysis is given.
ad hoc to paper SVD truncation of the normalized Gram matrix sum yields bases that represent all admissible product blocks to accuracy epsilon_trunc
This is the core accuracy-control mechanism in Section III-C and IV-C. It is asserted rather than proven: no error bound links the truncated bases to the global product error, and the normalization weighting is unspecified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Accuracy Controlled Structure-Preserving ${\cal H}^2$-Matrix-Matrix Product in Linear Complexity with Change of Cluster Bases." pith.science (2026). https://pith.science/paper/65LDJMT7

@misc{pith2026190805218,
  author       = {Pith},
  title        = {Pith review of: Accuracy Controlled Structure-Preserving $\cal H^2$-Matrix-Matrix Product in Linear Complexity with Change of Cluster Bases},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/65LDJMT7}},
  note         = {Machine review of arXiv:1908.05218}
}

abstract

${\cal H}^2$-matrix constitutes a general mathematical framework for efficient computation of both partial-differential-equation and integral-equation-based operators. Existing linear-complexity ${\cal H}^2$ matrix-matrix product (MMP) algorithm lacks explicit accuracy control, while controlling accuracy without compromising linear complexity is challenging. In this paper, we develop an accuracy controlled ${\cal H}^2$ matrix-matrix product algorithm by instantaneously changing the cluster bases during the matrix product computation based on prescribed accuracy. Meanwhile, we retain the computational complexity of the overall algorithm to be linear. Different from the existing ${\cal H}^2$ matrix-matrix product algorithm where formatted multiplications are performed using the original cluster bases, in the proposed algorithm, all additions and multiplications are either exact or computed based on prescribed accuracy. Furthermore, the original ${\cal H}^2$-matrix structure is preserved in the matrix product. While achieving optimal complexity for constant-rank matrices, the computational complexity of the proposed algorithm is also minimized for variable-rank ${\cal H}^2$-matrices. The proposed work serves as a fundamental arithmetic in the development of fast solvers for large-scale electromagnetic analysis. Applications to both large-scale capacitance extraction and electromagnetic scattering problems involving millions of unknowns on a single core have demonstrated the accuracy and efficiency of the proposed algorithm.

Figures

Figures reproduced from arXiv: 1908.05218 by the authors.

**Figure 2.** An H2 -matrix structure. (a) AH2 . (b) BH2 . (c) CH2 . where Tt1 and Tt2 are called transfer matrices. Because of such a nested relationship, the cluster bases only need to be stored for leaf clusters. For non-leaf clusters, only transfer matrices need to be stored. The H2 -matrix is stored in a tree structure, with the size of leaf-level clusters denoted by leafsize. The number of blocks formed by a single cluster … view at source ↗

**Figure 1.** Illustration of a block cluster tree and resulting [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 3.** H2 -matrix at leaf level. (a) A L H2 . (b) B L H2 . (c) C L H2 [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: H2 -matrix block at non-leaf level (L − 1). (a) A L−1 H2 . (b) B L−1 H2 . (c) C L−1 H2 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: MMP performance for AH2 × BH2 of large-scale capacitance extraction matrices. (a) Time scaling v.s. N. (b) Memory scaling v.s. N. The proposed solution is generated by first computing an MMP of AH2 BH2 to obtain CH2 , and then compute CH2 x. From Table I, we can see th…

**Figure 6.** Figure 6: MMP performance for AH2 × AH2 of 2-D slab from 4λ to 28λ. (a) Time scaling v.s. N. (b) Memory scaling v.s. N. 0 5 10 15 Number of unknowns(N) 105 0 1 2 3 4 5 MMP time(s) / Csp 2 1E-4 1E-3 1E-2 (a) 0 5 10 15 Number of unknowns(N) 105 0 200 400 600 800 MMP memory(MB) / C…

**Figure 7.** Figure 7: MMP performance for AH2 × AH2 of 3-D cube array. (a) Time scaling v.s. N. (b) Memory scaling v.s. N. Memory/Csp and Multiplication time/C2 sp respectively to examine the true scaling rate. In [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

[5]

Theoretical study on the rank of integral operators for broadband electromagnetic modeling from static to electrodynamic frequencies,

W. Chai and D. Jiao, “Theoretical study on the rank of integral operators for broadband electromagnetic modeling from static to electrodynamic frequencies,” IEEE Trans. Compon., Packag., and Manuf. Technol. , vol. 3, no. 12, pp. 2113-2126, Dec. 2013

work page 2013
[1]

H2-matrix in linear complexity,

S. Borm, “ H2-matrix in linear complexity,”Computing, vol. 77, pp. 1-28, Feb. 2006

work page 2006
[2]

Efﬁcient Numerical Methods for Non-local Operators: H2- matrix compression, Algorithms and Analysis,

S. Borm, “Efﬁcient Numerical Methods for Non-local Operators: H2- matrix compression, Algorithms and Analysis,” European Mathematical Society Tracts in Mathematics 14, 2006

work page 2006
[3]

Existence of H-matrix Representations of the Inverse Finite-Element Matrix of Electrodynamic Problems and H-Based Fast Direct Finite-Element Solvers,

H. Liu and D. Jiao, “Existence of H-matrix Representations of the Inverse Finite-Element Matrix of Electrodynamic Problems and H-Based Fast Direct Finite-Element Solvers,” IEEE Trans. MTT , vol. 58, no. 12, pp. 3697-3709, 2010

work page 2010
[4]

Direct Finite Element Solver of Linear Complexity for Large-Scale 3-D Electromagnetic Analysis and Circuit Extraction,

B. Zhou and D. Jiao, “Direct Finite Element Solver of Linear Complexity for Large-Scale 3-D Electromagnetic Analysis and Circuit Extraction,” IEEE Trans. MTT , vol. 63, no. 10, pp. 3066-3080, Oct. 2015

work page 2015
[6]

W. C. Chew, J. M. Jin, E. Michielssen, and J. M. Song, Fast and efﬁcient algorithms in computational electromagnetics . Norwood, MA: Artech House, 2001

work page 2001
[7]

Dense matrix inversion of linear complexity for integral-equation based large-scale 3-D capacitance extraction,

W. Chai and D. Jiao, “Dense matrix inversion of linear complexity for integral-equation based large-scale 3-D capacitance extraction,” IEEE Trans. MTT, vol. 59, no. 10, pp. 2404-2421, Oct. 2011

work page 2011
[8]

Accuracy-Controlled and Rank-Minimized H2- Matrix-Matrix Product with Change of Cluster Bases in Linear Com- plexity,

M. Ma and D. Jiao, “Accuracy-Controlled and Rank-Minimized H2- Matrix-Matrix Product with Change of Cluster Bases in Linear Com- plexity,” IEEE MTT-S International Conference on Numerical Electro- magnetic and Multiphysics Modeling and Optimization (NEMO), May 2019

work page 2019

Show all 13 references

[9]

A tetrahedral modeling method for electromagnetic scattering by arbitrarily shaped inhomogeneous dielectric bodies,

D. H. Schaubert, D. R. Wilton and A. W. Glisson, “A tetrahedral modeling method for electromagnetic scattering by arbitrarily shaped inhomogeneous dielectric bodies,” IEEE Trans. Antennas Propag. , vol. AP-32, no. 1, pp. 77-85, 1984

1984
[10]

An O(N ) iterative and O(N logN) direct volume integral equation solvers for large-scale electrodynamic analysis,

S. Omar and D. Jiao, “An O(N ) iterative and O(N logN) direct volume integral equation solvers for large-scale electrodynamic analysis,” the 2014 International Conference on Electromagnetics in Advanced Applications (ICEAA), Aug. 2014

2014
[11]

Minimal-rank H2-matrix based iterative and direct volume integral equation solvers for large-scale scattering analysis,

D. Jiao and S.Omar, “Minimal-rank H2-matrix based iterative and direct volume integral equation solvers for large-scale scattering analysis,”Proc. IEEE Int. Symp. Antennas Propag. , Jul. 2015

2015
[12]

Accuracy Controlled H2-Matrix-Matrix Product in Linear Complexity and Its Applications,

M. Ma and D. Jiao, “Accuracy Controlled H2-Matrix-Matrix Product in Linear Complexity and Its Applications,” 2018 IEEE International Symposium on Antennas and Propagation & USNCURSI National Radio Science Meeting, Jul. 2018

2018
[13]

Accuracy-Controlled and Structure-Preserved H2-Matrix-Matrix Product in Linear Complexity,

M. Ma and D. Jiao, “ Accuracy-Controlled and Structure-Preserved H2-Matrix-Matrix Product in Linear Complexity,” 2018 International Conference on Electromagnetics in Advanced Applications (ICEAA), Sep. 2018

2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [5]

Theoretical study on the rank of integral operators for broadband electromagnetic modeling from static to electrodynamic frequencies,

W. Chai and D. Jiao, “Theoretical study on the rank of integral operators for broadband electromagnetic modeling from static to electrodynamic frequencies,” IEEE Trans. Compon., Packag., and Manuf. Technol. , vol. 3, no. 12, pp. 2113-2126, Dec. 2013

work page 2013

[2] [1]

H2-matrix in linear complexity,

S. Borm, “ H2-matrix in linear complexity,”Computing, vol. 77, pp. 1-28, Feb. 2006

work page 2006

[3] [2]

Efﬁcient Numerical Methods for Non-local Operators: H2- matrix compression, Algorithms and Analysis,

S. Borm, “Efﬁcient Numerical Methods for Non-local Operators: H2- matrix compression, Algorithms and Analysis,” European Mathematical Society Tracts in Mathematics 14, 2006

work page 2006

[4] [3]

Existence of H-matrix Representations of the Inverse Finite-Element Matrix of Electrodynamic Problems and H-Based Fast Direct Finite-Element Solvers,

H. Liu and D. Jiao, “Existence of H-matrix Representations of the Inverse Finite-Element Matrix of Electrodynamic Problems and H-Based Fast Direct Finite-Element Solvers,” IEEE Trans. MTT , vol. 58, no. 12, pp. 3697-3709, 2010

work page 2010

[5] [4]

Direct Finite Element Solver of Linear Complexity for Large-Scale 3-D Electromagnetic Analysis and Circuit Extraction,

B. Zhou and D. Jiao, “Direct Finite Element Solver of Linear Complexity for Large-Scale 3-D Electromagnetic Analysis and Circuit Extraction,” IEEE Trans. MTT , vol. 63, no. 10, pp. 3066-3080, Oct. 2015

work page 2015

[6] [6]

W. C. Chew, J. M. Jin, E. Michielssen, and J. M. Song, Fast and efﬁcient algorithms in computational electromagnetics . Norwood, MA: Artech House, 2001

work page 2001

[7] [7]

Dense matrix inversion of linear complexity for integral-equation based large-scale 3-D capacitance extraction,

W. Chai and D. Jiao, “Dense matrix inversion of linear complexity for integral-equation based large-scale 3-D capacitance extraction,” IEEE Trans. MTT, vol. 59, no. 10, pp. 2404-2421, Oct. 2011

work page 2011

[8] [8]

Accuracy-Controlled and Rank-Minimized H2- Matrix-Matrix Product with Change of Cluster Bases in Linear Com- plexity,

M. Ma and D. Jiao, “Accuracy-Controlled and Rank-Minimized H2- Matrix-Matrix Product with Change of Cluster Bases in Linear Com- plexity,” IEEE MTT-S International Conference on Numerical Electro- magnetic and Multiphysics Modeling and Optimization (NEMO), May 2019

work page 2019