{"id":"f5809482-77be-42f8-bcfa-0ec612e92d2f","arxiv_id":"1908.07984","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper introduces minimal residual multistep (MRMS) methods that adapt explicit multistep coefficients to minimize a BDF residual, reducing per-step linear algebra to a thin least-squares solve while preserving the order and zero-stability of the underlying BDF formula.","lead":"A new class of multistep ODE solvers, MRMS methods, chooses its coefficients step by step so that the residual of an implicit BDF formula is as small as possible, avoiding the factorization of large matrices. For large stiff linear systems the method can be several times faster than standard BDF while giving similar accuracy, though convergence remains unproved.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The only large-scale benchmark has a separable exact solution, so the MRMS search subspace V contains the exact solution; general accuracy depends on an unproved subspace approximation.","rationale":"The reader's weakest assumption identifies the same load-bearing point: practical accuracy depends on the unproved closeness of y_BDF_k to V. I agree, and I sharpen it by noting that the only large-scale numerical experiment is constructed so that the exact solution lies exactly in V, so the experiment does not test this assumption. The other potential concerns are real but secondary: the mean-value step in Theorem 1 needs an invertibility/uniform-bounded-inverse condition on the BDF residual Jacobian, but this is fixable under standard dissipativity assumptions and is an asymptotic statement; the zero-stability lemma is sound for f=0. The paper honestly lists the missing convergence proof, but the consequence is that the central practical claim is not established beyond the favorable benchmark. This does not contradict the reader's conditional verdict; it reinforces it. Therefore the verdict remains unchanged: CONDITIONAL, pending either a convergence proof or evidence on non-separable stiff linear problems.","tokens_in":13284,"tokens_out":21843,"duration_ms":223576,"concrete_test":"Run the same 2D heat equation benchmark with non-separable, non-low-rank data, e.g., random smooth initial data with zero forcing, so V has dimension up to 2k rather than 1. For N=400 and the same step-size sequence M_s, compute the maximum-norm error of MRMS(k,k) and BDF-k against a fine spectral reference solution, and record ||y_MRMS_k - y_BDF_k|| / ||y_BDF_k|| per step. If the MRMS errors no longer track the BDF errors, or the empirical order drops below p, the reported agreement came from the separability of the constructed solution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 6.1 constructs the exact solution as w_ij(t) = p(t) q_ij and defines b(t) so that the ODE right-hand side is p'(t) q_ij. Consequently every vector y_j and f_j appearing in the MRMS subspace V (Eq. 8) is a scalar multiple of the single vector q; V is one-dimensional regardless of k and contains the exact solution at all times. The Section 6.4 observation that y_BDF_k is well approximated by V is therefore almost forced by the test problem and gives no evidence for general stiff non-autonomous linear systems. The order theorem is local, and the paper explicitly concedes in the Conclusion that a rigorous convergence proof is missing; because MRMS coefficients depend on the computed data, Dahlquist's equivalence theorem does not automatically apply. Thus the central practical claim, similar accuracy to BDF for large N, rests on an untested assumption that dist(y_BDF_k, V) stays small. If this distance is not small for a problem whose solution is not separable or low-rank, MRMS will deviate from BDF and the speed advantage becomes irrelevant.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Minimal Residual Multistep (MRMS) methods for large stiff linear non-autonomous ODEs. On each step, an explicit k-step linear multistep form is used, and its 2k coefficients are chosen to minimize the norm of the residual of an implicit BDF formula. This reduces the per-step linear algebra from solving an n×n system to solving an n×2k least-squares problem. The paper claims (Theorem 1) order min{2k−1,p} and (Lemma 1) zero-stability inherited from the underlying BDF method, gives a partial linear stability analysis for the one-step case, and reports numerical experiments on a 2D heat equation showing similar accuracy to BDF with growing speed advantage as the spatial dimension increases. The Conclusion explicitly states that a rigorous convergence proof is still missing and that general linear stability analysis is open.","tokens_in":13452,"tokens_out":7123,"duration_ms":76378,"significance":"If substantiated, the MRMS approach would be a new matrix-free alternative to factorization-based BDF for large stiff linear non-autonomous problems, reducing the dominant per-step cost from O(n^3) factorization or O(n^2) backsubstitution to O(nk^2) least-squares work. The paper is honest about its limitations: the missing convergence proof and the open stability questions are stated explicitly, and the experimental code is provided. The partial stability analysis for the one-step method is carefully derived. However, the main practical claim—that MRMS methods match BDF accuracy for general large stiff systems—currently rests on an untested subspace approximation property, and the numerical experiment is constructed so that the approximation subspace exactly contains the solution, so it provides no evidence for that property.","major_comments":[{"comment":"The proof of Theorem 1 uses the step: 'Using mean-value theorem for vector-valued functions it can be shown that ... ||yk−y(tk)|| ≤ C||r(yk)−r(y(tk))||.' This requires a uniform lower bound on the singular values of Dr over a neighborhood of y(tk); for the linear stiff case r'(x)=τA(tk)−ckI, such a bound is exactly what fails when the BDF system is ill-conditioned, and for nonlinear r the map need not be injective. Thus the order result is not established as a global convergence statement. The paper itself concedes in the Conclusion that a rigorous convergence proof is missing; the theorem should be restated as a local consistency result, or the missing hypotheses should be supplied.","section":"Section 2, Theorem 1 and Conclusion, item 1"},{"comment":"The numerical experiment is constructed from the separable exact solution w_ij(t)=p(t)q_ij with b(t) defined by (40)–(41). Because f(t_j,w(t_j))=p'(t_j)q, every column of the matrix V in (8) is a scalar multiple of the single vector q, so V is one-dimensional for every k. The BDF solution y_BDF_k also lies in span(q) because the right-hand side of (43) is a multiple of q. Therefore the observation in Section 6.4 that 'the exact solutions y_BDF_k are well approximated by the elements of subspaces V' is an identity forced by the test problem, not evidence for the general stiff non-autonomous case. The paper should include experiments with non-separable or multi-mode solutions, random initial data, or a direct report of dist(y_BDF_k,V), before claiming that MRMS accuracy matches BDF for large systems.","section":"Section 6.1 and Section 6.4"},{"comment":"The stability analysis for the one-step MRE method covers the equal-magnitude case (Corollary 2) and shows that instability in individual components can be harmless when the corresponding η_i is small. Example 3, however, demonstrates that R(0) can be much smaller than 1, heavily damping slow modes, and that |R(z3)|>1 occurs for a whole neighborhood of η=0. The statement that the method is 'applicable in stiff case' is therefore only a partial conclusion; the paper should clearly state that a full stability guarantee for general initial data is not obtained, beyond what is admitted in Section 4.3 and the Conclusion.","section":"Section 4.2, Example 3, and Section 4 conclusions"}],"minor_comments":[{"comment":"There are several typos: 'experimantal', 'absoltely', 'Kramer' (for Cramer), 'withouth', 'whithout', and 'consequense'.","section":"Throughout"},{"comment":"The text says the graphs of −R(z3) are shown in 'the first row of Figure 3', but the illustration caption is 'Figure 2: Illustration of example 3.' The figure references should be corrected.","section":"Section 4.2, Example 3"},{"comment":"The sign convention for α in the general method (α_j with a minus sign in Eq. (2)) and in the one-step method (α used positively in Eq. (11)) is confusing; the paper notes this, but a consistent notation would improve readability.","section":"Section 1, Eq. (2) and Section 4.1, Eq. (11)"},{"comment":"The phrase 'The computing of matrix W' and similar constructions are awkward; these sentences should be rephrased for clarity.","section":"Section 5.1"},{"comment":"The text refers to 'Figure 6.3' in the discussion; this should be 'Figure 6' to match the figure caption.","section":"Section 6.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is an honest preprint with a genuinely novel idea, but the two main pillars of the practical claim—the order/convergence result and the numerical demonstration—are both weaker than the abstract suggests. The order proof has a missing coercivity/invertibility argument, and the numerical benchmark is algebraically degenerate for the subspace approximation. I would not reject the paper, because the central idea is worth pursuing and the limitations are openly acknowledged, but the current manuscript needs substantial additional analysis or new experiments before it can support the claimed advantages."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on matrix-free stiff integrators. The core idea is real: replace each implicit BDF solve by a projection of the residual onto the low-dimensional subspace spanned by recent states and right-hand sides, then solve a thin least-squares problem. That is a different trade-off from Newton-Krylov or Rosenbrock, and I do not know an earlier paper doing exactly this. The paper is also honest about what it does not prove.\n\nThe order and zero-stability theorems look right to me. The order argument uses a vector-valued mean-value step that deserves a careful referee, but the claim itself is plausible because the optimization always beats the explicit 2k-1 order method inside the subspace. The zero-stability inheritance argument is clean. The MRE stability analysis is partial, as the author says, and the numerical stability studies are suggestive rather than conclusive.\n\nThe soft spots are real but not fatal to the idea. Most important: the strong numerical claim rests on Section 6.1, where the exact solution is w_ij(t)=p(t)q_ij and b(t) is chosen so every f_j is a scalar multiple of q. That makes the MRMS subspace one-dimensional and it contains the exact solution at every time, so the observed 'similar accuracy to BDF' is much less informative than the narrative implies. The author's own sentence that y_BDF_k is well approximated by V is nearly forced by that construction and gives no evidence for general stiff non-autonomous problems. The conclusion also explicitly says the convergence proof is missing and the Dahlquist conditions may not be sufficient; with coefficients depending on computed data, equivalence does not automatically follow. Those two concessions, together, mean the method is not yet established as reliable for general stiff linear systems.\n\nThe experiments at N=1000 do show the cost structure is favorable: one thin least-squares solve per step can beat repeated sparse backsubstitution. That comparison is fair in the fixed-step, single-factorization setting tested, even though the accuracy comparison is stacked. The code is public, which helps.\n\nWho gets value from this: researchers working on matrix-free time integration, especially for large semi-discrete PDEs, and anyone interested in residual-minimization time steppers. It deserves serious refereeing; the idea is new enough and the open questions are stated precisely enough that a careful referee could help turn a conditional paper into a solid one. I would send it to review.\n\nMy own verdict: conditional, not reject. The central concept is sound; the missing convergence theory and the structurally favorable benchmark are the two things a revision must address.","headline":"A genuinely new matrix-free idea for stiff linear ODEs, with honest limitations: the order and zero-stability results are plausible, but the main benchmark is structurally favorable and the missing convergence proof is load-bearing for general stiff problems.","tokens_in":13985,"tokens_out":650,"would_cite":true,"duration_ms":9410,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65L04","65L05","65L06"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new family of multistep methods, MRMS(k,p), is claimed to have the order and zero-stability of the underlying BDF formula while reducing the per-step unknown count from n to O(1) and replacing matrix factorization with a thin…","keywords":["ordinary differential equations","stiff systems","linear multistep methods","BDF methods","least squares","minimal residual multistep methods","non-autonomous linear systems","matrix factorization-free integration"],"falsifier":"Run MRMS(k,p) on a fixed linear system whose BDF solution is known and measure the distance from the BDF step to the subspace V at each step; if the MRMS error exceeds the bound from the order theorem whenever that distance fails to shrink like O($τ^{{p+1}}$), the central premise is refuted.","tokens_in":13059,"feed_emoji":"⚡","tokens_out":10341,"duration_ms":85572,"temperature":0.7,"pith_summary":"This paper introduces minimal residual multistep (MRMS) methods, a family of solvers for large stiff linear ODE systems that avoid factorizing an n-by-n matrix at every step. Instead of fixing the coefficients of an explicit k-step linear multistep method in advance, MRMS chooses them on each step to minimize the norm of the residual of an implicit BDF formula, so the new approximation is forced into a low-dimensional subspace spanned by the most recent solution vectors and right-hand sides. The paper proves that MRMS(k,p) has order min{2k−1,p} and inherits zero-stability from the underlying BDF method when p≤k. For linear non-autonomous problems each step costs one thin least-squares solve rather than a full linear system solve, and experiments on a 2D heat equation show accuracy comparable to BDF with an increasing speed advantage as the system grows.","feed_headline":"Stiff ODE step: from n-by-n solve to one least-squares fit","feed_subtitle":"MRMS methods match BDF accuracy and stability while swapping each n-by-n solve for a thin least-squares fit.","key_machinery":"The central object is the residual-minimization step: at each integration step the method selects coefficients (α,β) of an explicit k-step linear multistep formula y_k=∑(τβ_j f_j−α_j y_j) to minimize ‖r(y_k)‖, where r(x)=τf(t_k,x)−(c_k x+c_{k−1} y_{k−1}+...+c_{k−p} y_{k−p}) is the residual of a p-step BDF formula. Equivalently, the new value is the least-squares projection of the BDF step onto the subspace V=span{y_0,...,y_{k−1},τf_0,...,τf_{k−1}}. In the linear non-autonomous case the residual is affine, so each step reduces to the thin least-squares problem min_γ ‖Wγ−g‖ with W=(τA(t_k)−c_k I)V, solvable by QR or SVD in O(nk²) flops.","core_discovery":"The central claim is that a multistep method need not have precomputed coefficients to be useful: choosing the coefficients adaptively by residual minimization against a BDF formula gives a method whose order and zero-stability match the BDF formula. Concretely, MRMS(k,p) is defined by y_k = argmin_{x∈V} ‖r(x)‖, where V is the span of the last k solution vectors and their scaled right-hand sides, and r is the residual of the p-step BDF formula. Theorem 1 states the order is min{2k−1,p}; Lemma 1 states that for p≤k the generating polynomial of the zero-stability analysis is proportional to $z^{{k−p}}$ times the BDF generating polynomial, so zero-stability is inherited. For the one-step case (minimal residual Euler), the paper shows well-posedness conditions and derives sufficient conditions for absolute stability, including stability for all systems of dimension n≤415 with equal-magnitude initial components. The numerical experiment reports that MRMS and BDF give nearly identical errors on a 2D heat equation, while MRMS becomes faster as the spatial grid N grows from 20 to 1000.","pith_inferences":["A natural next step, not taken in the paper, is to enrich V with a few additional vectors, such as powers of (τA−c_k I) applied to the last residual, to make the subspace assumption hold for a wider class of stiff problems; the order and stability theorems suggest the method would tolerate this without changing its structure.","The observed heavy damping of slow modes in the minimal residual Euler method implies the method acts as a filter; comparing MRMS with exponential integrators could reveal whether this filtering is a bug or a feature for problems with many near-zero eigenvalues.","The paper's complexity accounting suggests a sharper implementation: for constant A and τ, the matrix W changes by only two columns per step, so a QR update would reduce the per-step least-squares cost from O(nk²) to O(nk); this is noted but not implemented."],"forward_implications":["For large linear non-autonomous stiff systems, each step costs O(nk²) for the thin least-squares solve plus matrix-vector products, versus O(n³) for a direct factorization; the paper's heat-equation experiments show the crossover near N=400 and a several-fold speedup at N=1000.","MRMS(k,p) inherits the order and zero-stability of the underlying BDF formula, so BDF users can switch to MRMS for large problems without changing the expected convergence order in the fixed-step regime.","The one-step minimal residual Euler method is shown to be well-posed under explicit conditions and absolutely stable for all systems with n≤415 (and beyond for sufficiently negative eigenvalues), supporting its use on stiff problems.","Because the method is nonlinear even for linear ODEs, classical absolute-stability regions do not apply; stability depends on the joint eigenvalue distribution, which the paper identifies as needing further research."],"supporting_citations":[{"why":"Defines the BDF formula family whose residual MRMS minimizes; the underlying method that supplies order and zero-stability.","marker":"[6]"},{"why":"Introduces the idea of restricting stiff-system solution to a low-dimensional subspace, which MRMS adapts.","marker":"[4]"},{"why":"Provides the mean-value theorem for vector-valued functions used in the proof of the order theorem.","marker":"[7]"},{"why":"Supplies the Chebyshev minimax polynomial used to derive sufficient stability bounds for the minimal residual Euler method.","marker":"[8]"},{"why":"Sets out the classical linear multistep stability framework that the paper adapts and contrasts with MRMS.","marker":"[9]"},{"why":"Gives the O(nk²) complexity of least-squares solvers used in the per-step cost comparison.","marker":"[10]"},{"why":"States the Dahlquist convergence conditions for linear multistep methods, which the paper identifies as an open question for MRMS convergence.","marker":"[13]"}],"fun_headline_variants":["Skip the big solve: adapt coefficients on the fly","Adaptive multistep methods dodge matrix factorizations","Stiff ODEs: no factorization needed, just least squares","MRMS: multistep accuracy, one small fit per step","Cut n-by-n solves to O(1) per step for stiff ODEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the exact BDF solution at each step stays close to the low-dimensional subspace V spanned by the most recent solution vectors and right-hand sides; if that approximation is poor, the MRMS solution will deviate from the BDF result and the favorable accuracy comparison collapses.","fun_headline_variants_meta":{"raw":{"variants":["Skip the big solve: adapt coefficients on the fly","Adaptive multistep methods dodge matrix factorizations","Stiff ODEs: no factorization needed, just least squares","MRMS: multistep accuracy, one small fit per step","Cut n-by-n solves to O(1) per step for stiff ODEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000294,"raw_usage":{"total_tokens":1798,"prompt_tokens":1118,"completion_tokens":680,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":734,"completion_tokens_details":{"reasoning_tokens":590}},"tokens_in":734,"tokens_out":680,"duration_ms":106280,"temperature":1.0,"reasoning_tokens":590,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:52:40.224243+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MRMS(k,p) on a fixed linear system whose BDF solution is known and measure the distance from the BDF step to the subspace V at each step; if the MRMS error exceeds the bound from the order theorem whenever that distance fails to shrink like O($τ^{{p+1}}$), the central premise is refuted.","supporting_citations":[{"cited_title":"Curtiss, J","cited_arxiv_id":null,"evidence_quote":"Defines the BDF formula family whose residual MRMS minimizes; the underlying method that supplies order and zero-stability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the mean-value theorem for vector-valued functions used in the proof of the order theorem."},{"cited_title":"Saad, Iterative Methods for Sparse Linear Systems, 2nd Edition, Society for Industrial and Applied Mathematics, Philadelphia, PA, USA, 2003","cited_arxiv_id":null,"evidence_quote":"Supplies the Chebyshev minimax polynomial used to derive sufficient stability bounds for the minimal residual Euler method."},{"cited_title":"Hairer, G","cited_arxiv_id":null,"evidence_quote":"Sets out the classical linear multistep stability framework that the paper adapts and contrasts with MRMS."},{"cited_title":"Dahlquist, Convergence and stability in the numerical integration of ordinary differential equations, Mathematica Scandinavica 4 (1) (1956) 33– 53","cited_arxiv_id":null,"evidence_quote":"States the Dahlquist convergence conditions for linear multistep methods, which the paper identifies as an open question for MRMS convergence."}],"review_version":1}