{"id":"bf1c7412-34d3-4d64-bb07-028e20ab6a2d","arxiv_id":"2607.05437","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"A Matlab/Fortran toolbox implements unblocked, recursive and blocked GTH algorithms that deliver componentwise-accurate linear algebra on M-matrices supplied via left or right triplets.","lead":"The paper ships a Matlab toolbox (with Lapack-style Fortran kernels) that solves linear systems, factorizations, inverses, square roots, singular values and Riccati equations for M-matrices given by triplet representations, using subtraction-free GTH algorithms. Practitioners in Markov chains, networks and PDE discretizations can obtain componentwise-accurate results even when the matrix is severely ill-conditioned.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The Reader’s strongest claim matches the paper’s stated contribution and is backed by both theory and experiments. The weakest assumption (availability of a triplet) is accurately diagnosed and is already flagged by the authors; it is a usage precondition rather than a flaw in the algorithms. Because no additional load-bearing technical concern surfaces under scrutiny, the ACCEPT verdict with high confidence stands without adjustment.","tokens_in":23636,"tokens_out":386,"duration_ms":4392,"concrete_test":"Clone the public repository, rebuild the Mex files, and re-run Example 3 (n=512, σ=10^{-12}) and Example 6 (ill-conditioned u) with the shipped mmatrix objects; verify that the reported componentwise errors remain O(u) while Matlab’s lu/svd degrade, confirming that the published accuracy figures are reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the Lapack-style GTH kernels (unblocked/recursive/blocked, left and right triplets) and the derived solvers deliver componentwise accuracy for the listed problems when a usable triplet is supplied. That claim is supported by the classical subtraction-free analysis of GTH, the explicit perturbation theory collected in Theorem 3 / Appendix, the safety-factor argument for the non-subtraction-free steps (square-root, SDA), and the numerical experiments of §5 that keep componentwise errors near machine precision while standard Lapack/Matlab routines lose digits. The only practical caveat—the need for an a-priori triplet—is already stated by the authors (Figure 1, §2.2) and is correctly identified by the Reader as the weakest assumption; it does not undermine the algorithmic correctness of the toolbox itself. No internal inconsistency, hidden cancellation, or unsupported accuracy claim appears in the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","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.","tokens_in":23863,"tokens_out":841,"duration_ms":6870,"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.","major_comments":[],"minor_comments":[{"comment":"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.","section":null},{"comment":"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.","section":null},{"comment":"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.","section":null},{"comment":"Minor typographical inconsistencies appear (e.g., “Themmatrixtoolbox” / “mmatrix-toolbox” / “mmatrix toolbox” in the title and abstract; “themmatrixclass” in §4). Uniform naming would improve readability.","section":null},{"comment":"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.","section":null}],"recommendation":"accept","confidential_remarks":"The manuscript is a solid software/algorithms paper for a numerical linear algebra or mathematical software venue. The central accuracy claims rest on well-established GTH theory plus careful implementation; I see no load-bearing gaps. The only soft point is the a-priori triplet requirement, which the authors already flag. Fit for a cs.MS / numerical software journal is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean software paper that does what it says. The authors take the classical GTH algorithm (and the higher-level routines already in the literature for square roots, MAREs, and high-relative-accuracy SVD) and ship production-quality Fortran kernels—unblocked, recursive cache-oblivious, and blocked Level-3—together with a usable Matlab mmatrix class that overloads the usual operators. That packaging is the actual novelty; the underlying theory is not new, and they do not pretend otherwise.\n\nWhat they do well is the engineering and the evidence. The Lapack-style interfaces, the careful left/right triplet variants, the work-buffer accounting, and the safety-factor argument for the non-subtraction-free steps (sqrtm, SDA) are all spelled out with clear invariants. The numerical section is honest: componentwise errors stay near machine precision on the FEM and random-triplet tests while Matlab’s lu/inv/svd/eig lose digits exactly as the condition number predicts. The appendix mildly extends the perturbation bounds to allow u to be perturbed as well; that is useful and cleanly proved via the all-minors matrix-tree representation.\n\nThe soft spot is the one they themselves flag in Figure 1 and §2.2: you need a usable triplet up front. Recovering one from the matrix entries is generally ill-conditioned, so the toolbox is only as good as the analytic triplet the application already supplies. That is a real practical limitation for some users, but it is not a flaw in the algorithms or the claims. Free parameters (γ = 1.5, block size) are minor and documented.\n\nWho this is for: people who already work with Markov chains, networks, or discretizations that naturally produce triplets and who care about componentwise accuracy on small entries. The math, the citations, and the experiments look solid; no circularity, no hidden cancellation. I would send it to referees without hesitation and would cite the toolbox myself if I needed these kernels. Engage with it.","headline":"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.","tokens_in":24439,"tokens_out":515,"would_cite":true,"duration_ms":5201,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F30","65Y15","65F05","15B48"],"pacs":[],"model":"grok-4.5","headline":"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.","keywords":["M-matrix","mmatrix-toolbox","componentwise error","GTH algorithm","triplet representation","Lapack-style factorization","accurate SVD","matrix square root"],"falsifier":"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.","tokens_in":24581,"feed_emoji":"📐","tokens_out":651,"duration_ms":5181,"temperature":0.7,"pith_summary":"The paper introduces the mmatrix toolbox: a Matlab package whose Fortran core implements unblocked, recursive and blocked variants of the GTH algorithm for M-matrices that are given by a left or right triplet. A triplet stores the off-diagonal entries together with two nonnegative vectors that reconstruct the diagonal without cancellation; the same representation lets every subsequent elimination step stay subtraction-free. Consequently the toolbox returns componentwise-accurate LU factors, linear-system solutions (nonnegative right-hand sides), inverses, Schur complements, singular values, square roots and solutions of associated Riccati equations, even for matrices whose ordinary condition numbers are huge. The Matlab layer wraps these kernels in an object that overloads ordinary linear-algebra operations, so users can keep working with familiar syntax while the library silently preserves the triplet. Numerical experiments confirm that the new routines stay near machine precision in every entry while standard Lapack-style solvers lose all correct digits in the small ones.","feed_headline":"Toolbox keeps every entry of ill-conditioned M-matrices accurate","feed_subtitle":"Subtraction-free GTH kernels on triplet data beat standard solvers on tiny entries","key_machinery":"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.","core_discovery":"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.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["GTH on triplets keeps M-matrix entries accurate without cancellation","Blocked recursive GTH yields precise LU and inverses for ill-conditioned M-matrices","mmatrix toolbox computes componentwise-accurate Schur complements via GTH","Triplet GTH algorithms solve M-matrix systems and Riccati equations accurately","Subtraction-free Lapack-style GTH factors guarantee entrywise M-matrix precision"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"A usable left or right triplet must already be known analytically; recovering one from the matrix entries alone is generally an ill-conditioned operation.","fun_headline_variants_meta":{"raw":{"variants":["GTH on triplets keeps M-matrix entries accurate without cancellation","Blocked recursive GTH yields precise LU and inverses for ill-conditioned M-matrices","mmatrix toolbox computes componentwise-accurate Schur complements via GTH","Triplet GTH algorithms solve M-matrix systems and Riccati equations accurately","Subtraction-free Lapack-style GTH factors guarantee entrywise M-matrix precision"]},"model":"grok-4.5","effort":"low","cost_usd":0.005072,"raw_usage":{"total_tokens":1378,"prompt_tokens":703,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":50720000,"prompt_tokens_details":{"text_tokens":703,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":591,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":703,"tokens_out":84,"duration_ms":4544,"temperature":1.0,"reasoning_tokens":591,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T02:44:53.675192+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"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.","supporting_citations":[],"review_version":1}