REVIEW 5 minor 42 references
The mmatrix toolbox: componentwise accurate algorithms for M-matrices with triplet representation
T0 review · 0 major / 5 minor · reviewed 2026-07-12 · grok-4.5
Pith's one-line read A Matlab toolbox uses subtraction-free GTH algorithms on triplet M-matrices to keep every entry accurate even when the matrix is severely ill-conditioned.
desk verdict Solid Lapack-style software packaging of known GTH accuracy results; the blocked/recursive kernels and OO Matlab class are the real additions, and the experiments back the claims. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The GTH algorithm: a subtraction-free Gaussian elimination that, at every pivot step, recovers the diagonal entry from the current triplet by a formula that only adds and multiplies nonnegative quantities, then updates the triplet of the Schur complement so that the invariant is preserved.
What would settle it
Run the toolbox and ordinary Lapack solvers on a family of triplet M-matrices whose condition number grows without bound (for example the finite-element matrix of Example 3 with successively smaller sigma) and check whether the componentwise relative error of every entry of the inverse or of the singular-value vector stays near machine epsilon while the classical solvers lose all digits.
Extended reading notes
Core claim
When an M-matrix is supplied through a left or right triplet, the GTH algorithm (and its blocked and recursive Lapack-style variants) can compute LU factors, inverses, Schur complements and several matrix functions without any subtractive cancellation, thereby guaranteeing high componentwise accuracy even for severely ill-conditioned instances.
Load-bearing premise
A usable left or right triplet must already be known analytically; recovering one from the matrix entries alone is generally an ill-conditioned operation.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents the mmatrix toolbox, a Matlab package with a Lapack-style Fortran core for componentwise-accurate computations on M-matrices given by left or right triplet representations. It implements unblocked, recursive (cache-oblivious), and blocked GTH algorithms for LU factorization, linear systems with nonnegative right-hand sides, inversion, Schur complements, singular values (via accurate LDU + Jacobi SVD), matrix square roots (Cyclic Reduction with a safety factor), MARE solutions (SDA/ADDA with auxiliary triplets), and the smallest eigenvalue (shifted inverse iteration). Componentwise perturbation theory is collected and mildly extended (Theorem 3 and Appendix Theorem 5), and an object-oriented mmatrix class overloads standard Matlab operations while preserving triplets when possible. Extensive numerical experiments (Examples 3–8) show near-machine componentwise accuracy on ill-conditioned problems where standard Matlab/Lapack routines lose digits, with timings competitive for the blocked/recursive kernels.
Significance. If the claims hold, the work supplies a practical, reusable software artifact that makes classical subtraction-free GTH technology and its extensions usable for a range of M-matrix problems that arise in Markov chains, networks, and discretizations. Strengths include the Lapack-style Fortran kernels (left and right triplets, multiple blocking strategies), the systematic collection of componentwise perturbation results with a mild generalization allowing u-perturbations, the safety-factor argument that keeps non-subtraction-free steps (square root, SDA) under control, and reproducible numerical evidence that componentwise errors stay near machine precision while standard solvers do not. The open Matlab interface lowers the barrier for users who already possess analytic triplets. The main practical limitation—the need for a usable a-priori triplet—is already stated by the authors and does not undermine the algorithmic contribution.
minor comments (5)
- In §2.6.4 the SVD path uses diagonal pivoting on the recovered diagonal entries rather than complete pivoting when u is not the all-ones vector; the text correctly notes that κ(L) and κ(Û) may then be larger. A short remark or numerical check quantifying the degradation for the ill-conditioned-u case of Example 6 would help users decide when the method remains reliable.
- Table 1 and §3.7 list work-buffer sizes; it would be useful to state explicitly whether the recursive right-triplet kernel (dmmtrf2) re-uses the same n-length buffer across recursion levels or allocates additional temporary space, for readers who wish to call the Fortran routines directly.
- The default safety factor γ = 1.5 is used for square-root and SDA; a one-sentence sensitivity note (or a pointer to the earlier literature) on how larger γ trades accuracy for slower quadratic convergence would make the parameter choice more transparent.
- Minor typographical inconsistencies appear (e.g., “Themmatrixtoolbox” / “mmatrix-toolbox” / “mmatrix toolbox” in the title and abstract; “themmatrixclass” in §4). Uniform naming would improve readability.
- In Algorithm 1 the update v ← u ⊙ (p − min(p)) is subtraction-free only after the min is taken; a brief parenthetical that the resulting v remains nonnegative would reassure readers scanning the pseudocode.
Circularity Check
No significant circularity; GTH accuracy claims rest on classical subtraction-free analysis, independent perturbation theory, and external numerical benchmarks.
full rationale
The paper's central claims concern Lapack-style implementations (unblocked/recursive/blocked GTH kernels for left/right triplets, plus derived solvers for systems, inverses, Schur complements, SVD, square roots, and MAREs) that deliver componentwise accuracy when a usable triplet is supplied. These rest on the well-known subtraction-free property of the GTH algorithm (cited to Grassmann–Taksar–Heyman and Alfa–Xue–Ye) together with the componentwise perturbation bounds collected in Theorem 3 (proved in the appendix from the all-minors matrix-tree lemma of Chaiken and related literature). No parameters are fitted to data and then re-presented as predictions; the numerical experiments of §5 generate independent test matrices (finite-element, random triplets) and compare against high-precision (vpa) or exact references. Self-citations (e.g., [5] for the square-root iteration, [14] for multilinear PageRank, [16] for fluid queues) supply algorithmic building blocks whose correctness is independently established; they are not load-bearing uniqueness theorems that close a circular loop. The only acknowledged limitation—the need for an a-priori triplet—is stated explicitly (Figure 1, §2.2) and does not render the algorithmic claims circular. The derivation chain is therefore self-contained against external theory and benchmarks.
Assumptions & free parameters
free parameters (2)
- safety factor γ (default 1.5) =
1.5
- block size n_b
assumptions (3)
- standard math An M-matrix possesses a right (resp. left) triplet if and only if, in its Frobenius normal form, every singular diagonal block has zero off-block entries in its row (resp. column).
- standard math Componentwise relative perturbations of a triplet of size ε produce componentwise relative perturbations of the inverse of size O(n)ε (first-order).
- domain assumption The GTH update never performs a subtraction of two quantities of the same sign when a triplet is maintained.
Cite this review
Pith. "Pith review of The mmatrix toolbox: componentwise accurate algorithms for M-matrices with triplet representation." pith.science (2026). https://pith.science/paper/YKJUBLMH
@misc{pith2026260705437,
author = {Pith},
title = {Pith review of: The mmatrix toolbox: componentwise accurate algorithms for M-matrices with triplet representation},
year = {2026},
howpublished = {\url{https://pith.science/paper/YKJUBLMH}},
note = {Machine review of arXiv:2607.05437}
}
read the original abstract
We introduce the mmatrix toolbox, a Matlab software package for componentwise accurate computations with M-matrices described through left or right triplet representations. The core of the toolbox is a Fortran implementation, in the Lapack style, of the unblocked, recursive, and blocked versions of the GTH algorithm and its applications for the accurate computation of the solution of linear systems with M-matrix coefficient and nonnegative right-hand side, the LU factorization of an M-matrix and its inverse. These algorithms avoid subtractive cancellation; this property ensures high componentwise accuracy even for ill-conditioned problems. The toolbox contains also accurate algorithms for related problems, such as computing the Schur complement, the singular values, the square root of an M-matrix, and the solution of nonsymmetric algebraic Riccati equations associated with M-matrices. The Matlab interface is based on an object-oriented implementation allowing one to use standard Matlab operations on M-matrices with triplet representation.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
-
[1]
, TITLE =
Berman, Abraham and Plemmons, Robert J. , TITLE =. 1994 , PAGES =
1994
-
[2]
Linear Algebra Appl
Computing the singular value decomposition with high relative accuracy , author=. Linear Algebra Appl. , volume=. 1999 , DOI =
1999
-
[3]
Demmel, James and Koev, Plamen , fjournal=. Numer. Math. , volume=. 2004 , DOI =
2004
-
[4]
Linear Algebra Appl
Kuo, I Wen , TITLE =. Linear Algebra Appl. , FJOURNAL =. 1977 , NUMBER =
1977
-
[5]
and Iannazzo, Bruno and Meini, Beatrice and Meng, Jie , year=
Bini, Dario A. and Iannazzo, Bruno and Meini, Beatrice and Meng, Jie , year=. Component-wise accurate computation of the square root of an. 2605.21679 , archivePrefix=
-
[6]
Regenerative analysis and steady state distributions for
Grassmann, Winfried K and Taksar, Michael I and Heyman, Daniel P , fjournal=. Regenerative analysis and steady state distributions for. Oper. Res. , volume=. 1985 , publisher=
1985
-
[7]
Accurate solutions of
Xue, Jungong and Xu, Shufang and Li, Ren-Cang , FJournal =. Accurate solutions of. Numer. Math. , volume=. 2012 , publisher=
2012
-
[8]
Royal Society Open Science , volume=
Comparison of social structures within cities of very different sizes , author=. Royal Society Open Science , volume=. 2016 , publisher=
2016
Show all 42 references
-
[9]
2006 , URL =
Pajek datasets , author=. 2006 , URL =
2006
-
[10]
Numerical computing with
Moler, Cleve B , year=. Numerical computing with
-
[11]
Multilinear
Gleich, David F and Lim, Lek-Heng and Yu, Yongyang , FJournal=. Multilinear. SIAM J. Matrix Anal. Appl. , volume=. 2015 , publisher=
2015
-
[12]
Proceedings of the 2015 SIAM International Conference on Data Mining , pages=
Tensor spectral clustering for partitioning higher-order network structures , author=. Proceedings of the 2015 SIAM International Conference on Data Mining , pages=. 2015 , organization=
2015
-
[13]
and Gleich, David F
Benson, Austin R. and Gleich, David F. and Lim, Lek-Heng , title =. SIAM Rev. , issn =. 2017 , language =. doi:10.1137/16M1074023 , keywords =
2017 doi
-
[14]
The University of
Davis, Timothy A and Hu, Yifan , Fjournal=. The University of. ACM Trans. Math. Softw. , volume=. 2011 , publisher=
2011
-
[15]
Statistical Science , volume=
The mixture transition distribution model for high-order Markov chains and non-Gaussian time series , author=. Statistical Science , volume=. 2002 , publisher=
2002
-
[16]
Estimation and modelling repeated patterns in high order
Raftery, Adrian and Tavar. Estimation and modelling repeated patterns in high order. Journal of the Royal Statistical Society Series C: Applied Statistics , volume=. 1994 , publisher=
1994
-
[17]
Linear and Multilinear Algebra , volume=
On the limiting probability distribution of a transition probability tensor , author=. Linear and Multilinear Algebra , volume=. 2014 , publisher=
2014
-
[18]
Iterative
Saad, Yousef , year=. Iterative
-
[19]
Varga, Richard S. , year=. Matrix
-
[20]
Introduction to
Stoer, Josef and Bulirsch, Roland and Bartels, R and Gautschi, Walter and Witzgall, Christoph , volume=. Introduction to. 1980 , publisher=
1980
-
[21]
Gleich, David F and Yu, Yongyang , title=
-
[22]
Entrywise perturbation theory for diagonally dominant
Alfa, Attahiru Sule and Xue, Jungong and Ye, Qiang , FJournal =. Entrywise perturbation theory for diagonally dominant. Numer. Math. , volume=. 2002 , publisher=
2002
-
[23]
Accurate computation of the smallest eigenvalue of a diagonally dominant
Alfa, Attahiru Sule and Xue, Jungong and Ye, Qiang , fjournal=. Accurate computation of the smallest eigenvalue of a diagonally dominant. Math. Comp. , volume=. doi:10.1090/S0025-5718-01-01325-4 , year=
-
[24]
SIAM Journal on Mathematics of Data Science , volume=
Ergodicity coefficients for higher-order stochastic processes , author=. SIAM Journal on Mathematics of Data Science , volume=. 2020 , publisher=
2020
-
[25]
Perron-based algorithms for the multilinear
Meini, Beatrice and Poloni, Federico , FJournal=. Perron-based algorithms for the multilinear. Numer. Linear Algebra Appl. , volume=. 2018 , publisher=
2018
-
[26]
Linear Algebra Appl
Poloni, Federico , Title =. Linear Algebra Appl. , ISSN =. 2013 , Language =. doi:10.1016/j.laa.2011.05.036 , Keywords =
2013 doi
-
[27]
Xue, Jungong and Xu, Shufang and Li, Ren-Cang , Title =. Numer. Math. , ISSN =. 2012 , DOI =
2012
-
[28]
Extrapolation methods for fixed-point multilinear
Cipolla, Stefano and Redivo-Zaglia, Michela and Tudisco, Francesco , FJournal=. Extrapolation methods for fixed-point multilinear. Numer. Linear Algebra Appl. , volume=. 2020 , publisher=
2020
-
[29]
Anderson accelerated fixed-point iteration for multilinear
Lai, Fuqi and Li, Wen and Peng, Xiaofei and Chen, Yannan , FJournal=. Anderson accelerated fixed-point iteration for multilinear. Numer. Linear Algebra Appl. , volume=. 2023 , publisher=
2023
-
[30]
Relative-error bounds for the
O'Cinneide, Colm Art , FJournal =. Relative-error bounds for the. Numer. Math. , volume=. 1996 , publisher=
1996
-
[31]
Abdesselam, Abdelmalek , FJournal=. The. Adv. Appl. Math. , volume=. 2004 , publisher=
2004
-
[32]
A combinatorial proof of the all minors matrix tree theorem , author=. SIAM J. Alg. Disc. Meth. , volume=. 1982 , publisher=
1982
-
[33]
Newton's method in floating point arithmetic and iterative refinement of generalized eigenvalue problems , fjournal =
Tisseur, Fran. Newton's method in floating point arithmetic and iterative refinement of generalized eigenvalue problems , fjournal =. SIAM J. Matrix Anal. Appl. , issn =. 2001 , language =. doi:10.1137/S0895479899359837 , keywords =
2001 doi
-
[34]
, title =
Higham, Nicholas J. , title =. 2002 , publisher =. doi:10.1137/1.9780898718027 , keywords =
2002 doi
-
[35]
and Poloni, Federico , title =
Nguyen, Giang T. and Poloni, Federico , title =. Numer. Math. , issn =. 2015 , language =. doi:10.1007/s00211-014-0675-4 , keywords =
2015 doi
-
[36]
and Sun, Ji-guang , title =
Stewart, Gilbert W. and Sun, Ji-guang , title =. 1990 , publisher =
1990
-
[37]
Accuracy and componentwise accuracy in multilinear
Mehdi Najafi Kalyani and Federico Poloni , year=. Accuracy and componentwise accuracy in multilinear. 2506.18356 , archivePrefix=
-
[38]
Linear Algebra Appl
Guo, Chun-Hua , title =. Linear Algebra Appl. , issn =. 2013 , language =. doi:10.1016/j.laa.2013.08.018 , keywords =
2013 doi
-
[39]
Linear Algebra Appl
Guo, Chun-Hua and Lu, Di , title =. Linear Algebra Appl. , issn =. 2016 , language =. doi:10.1016/j.laa.2015.11.024 , keywords =
2016 doi
-
[40]
and Iannazzo, Bruno and Meini, Beatrice , TITLE =
Bini, Dario A. and Iannazzo, Bruno and Meini, Beatrice , TITLE =. 2012 , PAGES =
2012
-
[41]
2018 , publisher =
Huang, Tsung-Ming and Li, Ren-Cang and Lin, Wen-Wei , title =. 2018 , publisher =. doi:10.1137/1.9781611975369 , keywords =
2018 doi
-
[42]
and Prokop, Harald and Ramachandran, Sridhar , title =
Frigo, Matteo and Leiserson, Charles E. and Prokop, Harald and Ramachandran, Sridhar , title =. Proceedings of the 40th Annual Symposium on Foundations of Computer Science , series =. 1999 , pages =
1999
Reviewed July 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.