{"id":"5e311717-5f4f-4f24-a6b3-8c2978f69ccd","arxiv_id":"2607.22905","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"AlgMortar constructs a fully algebraic, provably well-posed multiscale mortar preconditioner from the SPD matrix alone, and on high-contrast Darcy-flow and SPE11 corner-point systems it is competitive with BoomerAMG.","lead":"AlgMortar is a new preconditioner built straight from the matrix of a simulation, with no grid geometry needed, that speeds up pressure-flow solves in porous media. On high-contrast reservoir test cases with tens of millions of unknowns it matches or beats the standard algebraic-multigrid preconditioner on wall-clock time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SPE11C corner-point experiments are never checked against the -W>0 hypothesis; TPFA on non-K-orthogonal grids can yield positive interface couplings that break Theorem 4.4 and the Cholesky step.","rationale":"The reader identified the same weakest assumption: the unverified M-matrix property on the SPE11 corner-point grids. I agree this is the most load-bearing concern because the entire theoretical edifice—Theorem 4.4, Corollaries 4.5/5.1, and the spectral bound—rests on -W being positive definite, which is exactly the M-matrix/nonpositive-off-diagonal hypothesis. The SPE11C experiments are the paper's most challenging and practically relevant evidence for the 'competitive with AMG' claim, yet they are precisely where TPFA can leave the hypothesis. A positive off-diagonal coupling on an interface would invalidate the proof of Q>0 and could make A_i indefinite, causing Cholesky failure; the paper provides no check or fallback. I do not escalate to REJECT because the central theorem is correctly proven for the named matrix class (SPD M-matrices), and the Cartesian-grid experiments (Sections 7.1-7.3) fall within that class. The gap is addressable: verify the hypothesis or explicitly qualify the corner-point results as heuristic. Thus the reader's CONDITIONAL verdict is appropriate, and my read does not change it. The unproved assertion A_i > 0 in Section 3.3 is a related but secondary gap; for M-matrices it is true and easily proved, so the core concern remains the unverified hypothesis on the actual test matrices.","tokens_in":22922,"tokens_out":8186,"duration_ms":71736,"concrete_test":"Extract the assembled 50M and 100M SPE11C matrices (or reconstruct them with the OPM TPFA pipeline in Section 7.4) and (1) scan all off-diagonal entries for a_kl > 0; (2) re-run the AlgMortar graph partition and record the set of interface edges E; (3) check whether any a_kl > 0 falls on E. If none, the M-matrix hypothesis is satisfied on the relevant interfaces and the concern does not land. If some do, run Algorithm 5.1: if Cholesky of any A_i fails or PCG diverges, the corner-point claims are outside Theorem 4.4; if it still converges, report that the method is heuristic there, and state the observed behavior.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central well-posedness result, Theorem 4.4, proves Q>0 only under the joint hypotheses that A is SPD and -W is SPD, where W is the diagonal matrix of interface couplings a_kl (Section 3.2.3, 'We assume that the diagonal entries of W ... are negative, as is the case for M-matrices'). This is satisfied whenever the fine-grid matrix is an M-matrix with strictly negative graph-edge entries. However, the flagship irregular-grid experiments in Section 7.4 use TPFA on SPE11B/C corner-point grids. TPFA is monotone (M-matrix) only on K-orthogonal grids; on general corner-point grids the transmissibilities can be non-positive-off-diagonal, i.e., some a_kl > 0. The paper never verifies that the assembled SPE11C matrices satisfy the M-matrix condition or that the specific graph-partition interfaces have negative couplings. If any interface edge has a_kl > 0, then -W is not positive definite, the quadratic form in (4.4) is no longer bounded below by (1/4) p^T A p, the conclusion Q>0 fails, and the local operators A_i = A_ii - sum_j D_ij (Eq. 3.32) can lose SPD so the Cholesky factorization in Algorithm 5.1 can break. The assertion in Section 3.3 that each A_i is SPD is stated without proof; for M-matrices it follows from A_ii being an M-matrix and -D_ij being nonnegative diagonal, but this argument is absent and does not extend beyond M-matrices. Thus the numerical 'effectiveness on corner-point grids' in the Conclusion is an empirical claim lying outside the proven theory; the abstract's 'SPD with nonpositive off-diagonal entries' hypothesis is not established for those runs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AlgMortar, an algebraic multiscale mortar preconditioner for SPD systems. It constructs a domain decomposition from the matrix graph, defines local modified Dirichlet operators, and enforces a weak flux-continuity condition through an algebraic interface system. The authors prove well-posedness under the assumptions that the fine matrix A is SPD and the interface-weight matrix W (whose entries are the off-diagonal couplings on graph interfaces) is negative definite. They also derive a spectral bound kappa(M^{-1}A) <= 1 + 2 K_Lambda for the preconditioned system. Numerical experiments on large Darcy-flow systems, including SPE10 and SPE11 corner-point grids, compare AlgMortar with BoomerAMG and report competitive or faster times in several large cases.","tokens_in":23157,"tokens_out":7031,"duration_ms":68180,"significance":"If the theoretical claims hold, this is a useful contribution: a genuinely algebraic multiscale mortar preconditioner with self-contained proofs and a non-fitted spectral indicator. I re-derived the central identities (Prop. 3.1, Lemmas 4.2-4.3, identity (4.5), Theorem 4.4) and they are correct. The preconditioner exactly recovering MMMFEM on uniform Cartesian grids with K=I is a nice structural result. The numerical study is substantial, with weak and strong scaling on systems up to about 10^8 unknowns. However, the paper's main theoretical guarantee is conditional on an M-matrix-type sign condition, and the flagship irregular-grid experiments do not verify that condition; this gap is load-bearing for the conclusion that AlgMortar is effective on corner-point grids.","major_comments":[{"comment":"The SPE11 corner-point experiments are not covered by the proved theory. Theorem 4.4 requires A and -W to be SPD, which for a TPFA discretization means nonpositive off-diagonal entries (M-matrix) and strictly negative interface couplings. On non-K-orthogonal corner-point grids, TPFA can produce positive off-diagonal couplings. The paper never verifies that the assembled SPE11B/C matrices satisfy these conditions, yet the Conclusion states that AlgMortar 'remained effective on corner-point grids.' If an interface edge has a_kl > 0, then -W is not SPD, the quadratic form (4.4) loses its positivity argument, and, more critically, the local operators A_i = A_ii - sum_j D_ij (Eq. 3.32) can become indefinite, so the Cholesky factorization in Algorithm 5.1 can fail. The empirical claim therefore lies outside the theorem's hypotheses. Please either verify the sign conditions for the actual SPE11","section":"§7.4 and Theorem 4.4"},{"comment":"The sentence 'If A is symmetric positive definite and the diagonal entries of W are negative, then each A_i is also symmetric positive definite' is asserted without proof at the point where it first appears. This property is needed for the Cholesky factorization used in Algorithm 5.1. It follows later from Lemma 4.3 and identity (4.5), but only under the joint hypotheses A SPD and -W SPD. The paper should state the argument explicitly in Section 3.3 and, more importantly, acknowledge that outside this class A_i may lose definiteness. In its current form, the setup phase has no fallback or check for this failure mode.","section":"§3.3 and Algorithm 5.1"},{"comment":"The paper sometimes describes AlgMortar as operating on 'an SPD matrix' or 'using only an SPD matrix,' which overstates the domain of the theoretical results. The well-posedness theorem and the spectral bound require the additional sign condition that -W be positive definite (equivalently, all graph-interface couplings be strictly negative). The abstract does state 'nonpositive off-diagonal entries,' but the introduction and several remarks would benefit from the same qualification to avoid implying that the proof covers all SPD matrices.","section":"Introduction / Abstract"}],"minor_comments":[{"comment":"The header 'AlgMortar AMGPCG-block Jacobi interface solver' appears corrupted; it should probably be split into 'AlgMortar' (with two interface-solver columns) and 'AMG'.","section":"Table 1 header"},{"comment":"Typo: 'Darcy flow proble' should be 'Darcy flow problem.'","section":"Section 7"},{"comment":"The phrase 'We assume that the diagonal entries of W ... are negative, as is the case for M-matrices' is clear but should explicitly state that equality to zero is excluded on edges of E, since an edge is only present when a_kl != 0; strict negativity is what makes -W positive definite.","section":"Section 3.2.3"}],"recommendation":"major_revision","confidential_remarks":"The algebraic core is solid and the numerical study is extensive, so I do not see a fatal flaw in the method itself. The main issue is the mismatch between the assumptions of Theorem 4.4 and the SPE11 corner-point experiments; the authors should either demonstrate that those matrices satisfy the sign conditions or clearly restrict the claims. If they can address that, the paper is likely acceptable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The punchline: AlgMortar's algebraic core is real and mostly rigorous. The construction — graph-partition subdomains, diagonal modification to mimic Dirichlet problems, W-weighted weak flux continuity — is genuinely new as a fully algebraic multiscale mortar preconditioner, even if each ingredient traces to prior work. I re-derived the main algebra by hand and it holds: Prop. 3.1 (B^T = 4WM), Lemmas 4.2–4.3, the quadratic-form identity, Theorem 4.4, and Corollaries 4.5/5.1. The spectral bound κ ≤ 1 + 2K_Λ is a legitimate a priori estimate with K_Λ defined, not fitted. That is a solid contribution.\n\nThe numerics are honestly reported. Tables 1, 3–5 include cases where AMG is faster, and the conclusion states 'matched or outperformed' rather than a sweeping claim. For large Darcy-flow systems the method looks competitive, and the weak- and strong-scaling behavior is plausible.\n\nNow the soft spots, in proportion. The load-bearing one is the M-matrix hypothesis. The paper's theory assumes A SPD and -W SPD. The SPE11B/C corner-point experiments are the flagship claims, yet the paper never checks that the assembled TPFA matrices satisfy the nonpositive-off-diagonal condition. TPFA on non-K-orthogonal corner-point grids can produce positive off-diagonal couplings, and if any interface edge has a_kl > 0, -W is not SPD and the proof of Theorem 4.4 collapses. The local operators A_i can also lose SPD, which would break the Cholesky step in Algorithm 5.1. The assertion in Section 3.3 that each A_i is SPD is stated without proof; it follows for M-matrices, but not beyond. So the corner-point results are empirically interesting but sit outside the proven theory. This needs to be addressed before the paper is accepted: verify the hypothesis on those matrices, or extend/relax the theory.\n\nMinor but real: no code or data shipped, the SPE11 pipeline is partly unreleased, and all timings are single runs with no error bars. The hyperparameters (400 subdomains/process, 15+15 smoothing, 80 blocks, interface rtol 1e-2) and the non-default BoomerAMG settings from [22] are reasonable but have clear tuning cost. These are addressable in revision.\n\nWho benefits: people working on algebraic multiscale preconditioners, reservoir simulation, or PDE preconditioning. I'd cite it once the hypothesis gap is closed or at least stated honestly. The paper deserves a serious referee; desk rejection would be wrong.","headline":"Algebraically sound, honestly reported, and worth refereeing — but the corner-point experiments are not covered by the paper's own theory.","tokens_in":23939,"tokens_out":1894,"would_cite":true,"duration_ms":18530,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F08","65N55","65Y05","65F10"],"pacs":[],"model":"deepseek-v4-flash","headline":"AlgMortar rebuilds the multiscale mortar method from the matrix alone: for SPD M-matrices it is provably well posed, and in large Darcy-flow tests it matches or beats algebraic multigrid.","keywords":["algebraic multiscale method","mortar preconditioner","domain decomposition","M-matrix","conjugate gradient","Darcy flow","corner-point grid","multiscale basis functions"],"falsifier":"Take an SPD system from a two-point flux finite-volume discretization on a deliberately distorted corner-point grid, or add small positive off-diagonal couplings to an M-matrix, run AlgMortar, and check whether every local $A_i$ stays positive definite and whether PCG still converges in the predicted iteration range; a single local Cholesky failure or a preconditioned matrix with eigenvalues below zero would show the M-matrix hypothesis is load-bearing.","tokens_in":22558,"feed_emoji":"💧","tokens_out":7750,"duration_ms":68135,"temperature":0.7,"texified_at":"2026-08-05T21:41:34.474562+00:00","pith_summary":"AlgMortar proposes to precondition large sparse symmetric positive definite linear systems using only the matrix itself. The paper's central claim is that the multiscale mortar mixed finite element method—normally defined through grids, coefficients, and geometry—can be rebuilt algebraically: graph partition the matrix, solve local Dirichlet-like problems, and couple them with an interface condition that enforces weak flux continuity. The authors prove that for SPD matrices with nonpositive off-diagonal entries (M-matrices) the resulting preconditioner is well posed and positive definite, with a preconditioned condition number bounded by $1 + 2 K_\\Lambda$, and they show numerically that on 50–110 million unknown Darcy-flow systems it matches or beats algebraic multigrid in solution time, especially when a relaxed tolerance is acceptable.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":5098,"prompt_tokens":801,"completion_tokens":4297,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":801,"completion_tokens_details":{"reasoning_tokens":3566}},"feed_headline":"Matrix-only mortar preconditioner matches or beats multigrid at scale","feed_subtitle":"AlgMortar builds its coarse space from the system matrix alone; on the largest Darcy-flow tests it matches or beats algebraic multigrid.","key_machinery":"The load-bearing identity is $B^T = 4 W M$, where $W$ is the diagonal matrix of inter-subdomain entries $a_{kl}$, $M$ averages endpoint pressures on each graph edge, and $B$ assembles weighted interface contributions back into subdomains. Around it, the method uses graph partitioning to define subdomains, modified local matrices $A_i = A_{ii} - \\sum D_{ij}$ that mimic Dirichlet problems, multiscale basis functions computed by local Cholesky solves, and a piecewise-constant mortar space on each interface. This system yields the interface operator $S_\\lambda = -(W + 4 W M \\hat{A}^{-1} M^T W)$, whose SPD-ness on the mortar space is what makes conjugate gradients applicable; the quadratic-form identity in Theorem 4.4","core_discovery":"On its own terms, AlgMortar is the claim that MMMFEM's geometric ingredients—subdomains, local Dirichlet problems, and the weak continuity of normal flux—can be replaced by algebraic counterparts read off an SPD matrix $A$. Subdomains are graph partitions; local operators $A_i$ are $A_{ii}$ minus sums of diagonal blocks formed from inter-subdomain columns; the interface coupling is assembled from the matrix entries $a_{kl}$ on graph edges, with incidence and averaging operators satisfying $B^T = 4 W M$. The paper proves the coupled system matrix $Q$ is positive definite when $A$ and $-W$ are SPD, so the interface system and the reduced pressure system are SPD and PCG applies; the exact-recovery statement is tha","pith_inferences":["The paper's largest corner-point results sit outside the proved hypothesis: the well-posedness theorem requires strictly negative inter-subdomain entries, yet the SPE11 matrices are never checked for positive off-diagonal couplings. Reporting the sign pattern of those matrices would show whether the flagship runs are covered by the proof.","Because the graph partition and local factorizations are reusable when only the right-hand side changes, AlgMortar may be especially attractive for repeated solves in time-dependent or multiphase simulation; the paper lists this as future work, not a present claim.","The spectral bound suggests a quantitative predictor: PCG iterations should track the best approximation of edge averages by the mortar space, so adaptive enrichment—adding basis functions where the algebraic projection error is large—could be guided by the same quantity K_Lambda."],"forward_implications":["For any SPD matrix with nonpositive off-diagonal entries, the AlgMortar interface and reduced pressure systems are SPD, so PCG with AlgMortar is guaranteed to converge (Theorem 4.4 and Corollaries 4.5, 5.1).","The preconditioned condition number is bounded by 1 + 2 K_Lambda, so convergence is controlled by how well the piecewise-constant mortar space captures average pressures on graph interfaces (Theorem 5.6).","On uniform Cartesian grids with K = I, AlgMortar exactly reproduces MMMFEM, meaning the algebraic interface condition is the same weak flux-continuity condition (Remarks 3.2–3.4).","After one preconditioner application, residuals are nonzero only on partition interfaces, so inexpensive local smoothing can be applied without changing the parallel structure (Section 5.4).","On 10^7–10^8 unknown Darcy-flow systems, AlgMortar scales well and matches or beats algebraic multigrid in solution time on the largest cases, with the edge widening when the tolerance is relaxed (Tables 1, 3–5)."],"fun_headline_variants":["Matrix-only mortar preconditioner scales and matches AMG","Algebraic multiscale mortar: matrix in, multigrid out","Fully algebraic mortar preconditioner rivals AMG on hard flows","No geometry: AlgMortar builds coarse space from matrix alone"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof requires the fine-grid matrix to be an M-matrix—symmetric positive definite with nonpositive off-diagonal entries—with strictly negative entries on every graph-interface edge, so the diagonal weight matrix $-W$ and the local modified operators $A_i$ are all positive definite; the corner-point-grid experiments in the paper never verify this sign condition, and TPFA on non-K-orthogonal corner-point grids can produce positive off-diagonal couplings.","fun_headline_variants_meta":{"raw":{"variants":["Matrix-only mortar preconditioner scales and matches AMG","Algebraic multiscale mortar: matrix in, multigrid out","Fully algebraic mortar preconditioner rivals AMG on hard flows","No geometry: AlgMortar builds coarse space from matrix alone"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000868,"raw_usage":{"total_tokens":3628,"prompt_tokens":808,"completion_tokens":2820,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2749}},"tokens_in":552,"tokens_out":2820,"duration_ms":21764,"temperature":1.0,"reasoning_tokens":2749,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:13:15.836404+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an SPD system from a two-point flux finite-volume discretization on a deliberately distorted corner-point grid, or add small positive off-diagonal couplings to an M-matrix, run AlgMortar, and check whether every local $A_i$ stays positive definite and whether PCG still converges in the predicted iteration range; a single local Cholesky failure or a preconditioned matrix with eigenvalues below zero would show the M-matrix hypothesis is load-bearing.","supporting_citations":[],"review_version":1}