{"id":"ffdfeb9d-5487-4e76-9011-86b5ffdc88fd","arxiv_id":"2506.15259","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"New splitting-randomized dynamical low-rank methods solve stiff matrix differential equations at low rank, showing numerical first-order Lie-Trotter and second-order Strang convergence.","lead":"Splitting-based randomized low-rank algorithms for stiff matrix differential equations are proposed, treating the stiff linear part exactly with matrix exponentials and the nonlinear part with randomized low-rank solvers. Numerical tests on Allen-Cahn and differential Riccati equations show first-order (Lie-Trotter) and second-order (Strang) convergence.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Range-closure in Eq. (2.5) is unproven: the B-step replaces N(t) by its projection onto Range(Omega), so the reported first/second-order rates may not hold when row spaces rotate.","rationale":"The reader identified precisely the same load-bearing premise: Eq. (2.5) assumes the row space of N(t) is captured by the random sketch Omega, and the paper gives no bound on the resulting error. This is not an external debate about consensus; it is an internal gap between the derivation of the B-step and its use in every numerical experiment. If the range condition fails, the low-rank trajectory solved by the method is not the low-rank representation of the true nonlinear flow, and the clean first- and second-order tables could be artifacts of the specific test problems, parameter choices, or the unreported p and q values. The concern is serious enough to keep the paper conditional: the numerics are plausible and the experiments are extensive, but the central convergence claim needs either a local error analysis for the randomized substep or a direct numerical check of the range-inclusion assumption. Since the reader already recommended conditional acceptance with convergence analysis as a condition, no verdict change is needed. The stress-test pass therefore confirms the reader's assessment rather than overturning it, while sharpening the specific identity whose failure would break the argument.","tokens_in":25475,"tokens_out":5962,"duration_ms":66195,"concrete_test":"Construct a rank-1 2x2 test problem whose exact solution rotates its row space, e.g., N(t)=u sigma v(t)^T with v(t)=R(t)v0. Run Algorithm A.4 (DRSVD) with p=q=0 and one fixed Gaussian Omega over a single substep of size tau in {2^-2, ..., 2^-8}, and compare the returned rank-1 factor against the exact flow Phi^F_tau(N0). If the relative error does not decay like tau (or tau^2) and instead saturates at the level of ||N(t)-N(t)Omega(Omega^T Omega)^-1 Omega^T||_F, the missing range condition in Eq. (2.5) is a real failure rather than a harmless idealization. Repeating with increased p and q would also reveal how strongly the reported tables depend on the currently unreported randomization parameters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central convergence claim depends on the randomized dynamical range finder solving (2.5)-(2.7) faithfully. In Section 2.3.1, B(t)=N(t)Omega has derivative F(t,N(t))Omega, but Eq. (2.5) closes by using B(t)(Omega^T Omega)^-1 Omega^T in place of N(t). This is exact only when Range(N(t)^T) is contained in Range(Omega) for all t in the substep, an identity the paper neither states nor bounds. The same condition enters Algorithm A.3 line 7 and Algorithm A.4 step 3, where exactness is quoted from [8, Lemma 3.1] under Range(N(t)) subset Range(Q) without checking that condition for the Allen-Cahn or Riccati flows. Since every convergence table uses this substep, the observed first- and second-order rates could reflect row spaces staying nearly aligned with the random sketch over short intervals, or could be masked by unreported oversampling and power-iteration parameters; neither is demonstrated. The absence of an error analysis is therefore not merely a stylistic gap: the closed B-step is a core premise of the method, and Section 5 explicitly defers its justification to future work.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes splitting-based randomized dynamical low-rank (RDLR) approximations for stiff semilinear matrix differential equations of the form X'(t) = AX(t) + X(t)A^T + F(t,X(t)). The stiff linear part is integrated exactly with matrix exponentials, while the nonstiff nonlinear part is approximated by the randomized dynamical low-rank solvers DRSVD and DGN due to Carrel. Lie-Trotter and Strang compositions are formulated, and rank-adaptive variants are introduced. Numerical experiments on semi-discretized Allen-Cahn equations and differential Riccati equations report first-order (Lie-Trotter) and second-order (Strang) temporal convergence when the target rank is sufficiently large, together with spatial convergence studies and long-time interface simulations.","tokens_in":25661,"tokens_out":5137,"duration_ms":50535,"significance":"If the convergence claims are correct, the paper offers a potentially efficient alternative to projector-splitting low-rank integrators for stiff matrix-valued ODEs, combining operator splitting with randomized range finding. The manuscript is usefully explicit: Algorithms 1 and 2 and A.3-A.6 give complete pseudocode, and the tables report errors over many ranks, step sizes, and grid sizes, which allows a reader to see the rank dependence directly. However, the central claim is empirical: no error bounds are proved, a key range-closure assumption in the B-step is unstated, and the observed convergence order degrades at low ranks. The contribution is therefore best read as a numerical demonstration whose robustness conditions still need to be pinned down.","major_comments":[{"comment":"The B-step closes the nonlinearity by replacing N(t) with B(t)(Ω^T Ω)^{-1} Ω^T, which is exact only if Range(N(t)^T) is contained in Range(Ω) for all t in the substep; this condition is never stated and no bound is given for the projection error. The argument preceding Eq. (2.5) establishes Range(B(t0+τ)) = Range(N(t0+τ)) only at the endpoint and only under the rank assumption rank(N(t0+τ)) ≤ r+p, not for intermediate times. The same unverified condition enters Algorithm A.3 (lines 2 and 7), Algorithm A.4 (step 3), and the exactness claims imported from [8, Lemmas 3.1 and 3.2]. Because every convergence table in Section 4 is produced through this substep, the observed first- and second-order rates cannot be separated from the assumption that the random sketch tracks the row and column spaces of the flow; Section 5 explicitly defers an error analysis.","section":"§2.3.1, Eq. (2.5)"},{"comment":"The observed convergence orders degrade or become negative at low ranks, so the claimed 'desired convergence orders' are not a uniform property of the methods. In Table 4.2, DRSVD-ST at r=12 gives rates -0.0942 and -0.1169 for M=64 and 128, respectively; in Table 4.8, DRSVD-ST at r=4 gives rates -0.0924 and -0.1083 for M=256 and 512. Even the Lie-Trotter variants saturate at low ranks, e.g., Table 4.1 with r=12 shows a rate of only 0.4351 between M=128 and 256, and Table 4.7 with r=4 shows 0.3670 between M=256 and 512. The paper should either quantify the rank threshold above which the nominal orders hold or explicitly qualify the convergence statements in the abstract and introduction.","section":"§4.1.1, Tables 4.1-4.2; §4.1.2, Tables 4.7-4.8"},{"comment":"The numerical experiments never report the oversampling parameters p (and ℓ for DGN) or the number of power iterations q used in the randomized range finders, although these are free inputs to every method. Without these values the convergence tables cannot be reproduced, and the possibility that the observed high-order rates depend on generous oversampling or power iterations cannot be assessed. Please report the parameter values used in Tables 4.1-4.9 and, ideally, include a sensitivity study with respect to p and q.","section":"§4.1, Algorithms A.3-A.5"},{"comment":"The manuscript contains no error analysis for either the randomized range finder or the composed splitting scheme; Section 5 item (ii) acknowledges this gap. Because the nonlinear substep is approximated by a stochastic low-rank solver, the standard Lie-Trotter and Strang splitting order results do not automatically transfer to the composed method. At minimum, the authors should state the conditions (e.g., invariant subspaces, Lipschitz constants, and rank requirements) under which first- and second-order convergence is expected, and should clearly distinguish measured numerical rates from proven rates.","section":"§5, item (ii)"}],"minor_comments":[{"comment":"The equation is described as 'autonomous' although F depends explicitly on t; this wording should be corrected to avoid confusion.","section":"§2.3.1, after Eq. (2.5)"},{"comment":"The formula F(t, B(t)Ω^T Ω)^{-1} Ω^T) Ω has mismatched parentheses; it should read F(t, B(t)(Ω^T Ω)^{-1} Ω^T) Ω.","section":"Algorithm A.6, line 3"},{"comment":"Several numerical entries contain typographical artifacts such as '1.66 04E-5', '1.9 162E-6', and '1.6 604E-5'; these should be cleaned before publication.","section":"Tables 4.1-4.8"},{"comment":"The caption states both 'different time step size τ' and 'the number of time steps M = 1024'; these are inconsistent and the caption should clarify whether the abscissa is τ or M.","section":"Fig. 4.1 caption"},{"comment":"The control system writes x ∈ R^m while A ∈ R^{d×d}; the notation for the state dimension should be made consistent.","section":"§4.1.2"},{"comment":"The initial condition B(t0) = A0 Q2 should presumably be B(t0) = N0 Q2, since A0 is not defined in the algorithm.","section":"Algorithm A.5, line 3"}],"recommendation":"major_revision","confidential_remarks":"The paper leans heavily on Carrel [8], an arXiv preprint, for the exactness lemmas underlying DRSVD and DGN; if that preprint changes or is not published, the present conclusions are affected. The codes are promised but not yet released, which would help verify the missing parameter settings. The paper may fit the journal's scope, but the central claims need to be qualified or supported before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a reasonable engineering contribution—splitting exponential integrators with Carrel's randomized DLRA for stiff matrix ODEs—and the experiments show the advertised orders when the rank is big enough. But there is a real gap at the core of the nonlinear substep, and the paper needs a referee who pushes on it.\n\nWhat's new: the combination of Lie-Trotter/Strang splitting with DRSVD/DGN for the nonstiff part, plus rank-adaptive versions, is not in the prior literature. The linear part is rank-preserving and handled exactly via matrix exponentials, which is clean. The tests on Allen-Cahn and differential Riccati are standard but well chosen, and the accuracy advantage over the projector-splitting baseline is visible. Tables 4.1–4.9 report convergence orders that match first/second order at sufficient rank, and the degradation at low rank (e.g. Table 4.2, r=12) is honestly visible rather than hidden.\n\nWhere it gets shaky: Eq. (2.5) closes the B-step by replacing N(t) with B(t)(Omega^T Omega)^-1 Omega^T. That is exact only if Range(N(t)^T) is contained in Range(Omega), which for a random Gaussian sketch is not true for a generic flow—it can hold approximately when the row space is slow and the sketch is wide, but the paper never states the condition or bounds the error. The same issue carries into Algorithms A.3–A.5 via the exactness lemmas from Carrel [8]. Since every convergence table depends on this substep, the observed rates could in principle be an artifact of the sketch staying aligned with the true row space over short intervals. That is a load-bearing soft spot, not a stylistic one. The authors themselves defer error analysis to future work (Section 5, item ii), which is honest but leaves the central claim unproven.\n\nOther issues are more ordinary: no variance reporting across random sketches, no seeds, no oversampling p or power iteration q values in the tables, no code release, and the efficiency claim against [8] has no runtime comparison. These are fixable with a revision.\n\nWho it is for: people working on low-rank integrators for large stiff matrix ODEs, especially control/Riccati and phase-field problems. It's a useful variant that deserves a serious referee, but the right outcome is major revision, not acceptance as is.","headline":"A sensible combination of exponential splitting and randomized DLRA with honest experiments, but the B-step closure in Eq. (2.5) is unanalyzed and should gate acceptance.","tokens_in":26275,"tokens_out":2866,"would_cite":true,"duration_ms":31579,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F55","68W20","65L04"],"pacs":[],"model":"deepseek-v4-flash","headline":"Randomized low-rank integrators reach order one and two on stiff matrix ODEs","keywords":["low-rank splitting","dynamical low-rank approximation","rank-adaptive","matrix differential equation","differential Riccati equation","randomized numerical linear algebra","Lie-Trotter splitting","Strang splitting"],"falsifier":"Take a small semilinear matrix ODE with a known exact solution, run one step of the nonlinear subproblem, and compare the true $N(t_0+\\tau)$ with the sketched reconstruction $B(t_0+\\tau)(\\Omega^\\top\\Omega)^{-1}\\Omega^\\top$ for a few Gaussian sketches of increasing oversampling $p$; if this reconstruction error fails to decay with $p$ or $\\tau$, or if the overall method's error is dominated by this term, the range-inclusion premise fails and the claimed convergence orders do not hold in general.","tokens_in":25196,"feed_emoji":"🧮","tokens_out":16472,"duration_ms":133769,"temperature":0.7,"pith_summary":"This paper proposes and tests a family of low-rank integrators for stiff semilinear matrix differential equations of the form $X'(t) = AX(t) + X(t)A^\\top + F(t,X(t))$. The core idea is to split the equation into a stiff linear subproblem, solved exactly with matrix exponentials, and a nonstiff nonlinear subproblem, advanced by randomized dynamical low-rank methods that sketch the evolving row and column spaces. The paper claims that the Lie-Trotter composition gives first-order accuracy, the Strang composition gives second-order accuracy, and that both extend to rank-adaptive schemes, with numerical evidence on Allen-Cahn and differential Riccati problems. A sympathetic reader would care because existing projector-splitting low-rank integrators for stiff problems are sensitive to small singular values and expensive, while this approach is designed to remove step-size restrictions and avoid forming full-rank matrices.","feed_headline":"Low-rank randomized integrators hit orders one and two on stiff ODEs","feed_subtitle":"The method reaches first order (Lie-Trotter) and second order (Strang) on Allen-Cahn and Riccati equations.","key_machinery":"The central object is the pair of randomized dynamical low-rank solvers, DRSVD and DGN, built on a dynamical randomized range finder. The range finder advances the sketched ODE $\\dot B(t) = F(t, B(t)(\\Omega^\\top\\Omega)^{-1}\\Omega^\\top)\\Omega$ (and its transpose analogue) to estimate the dominant column space of the nonlinear flow, then post-processes by solving reduced ODEs on the augmented bases $Q = \\mathrm{orth}([U_0, Q_\\tau])$; DRSVD projects the nonlinear ODE onto that basis, while DGN additionally samples the row space and couples three small ODEs. The splitting with the exactly integrated linear flow is what keeps the solution on the rank-$r$ manifold, and the randomized steps are what make the nonlinear subproblem cheap to advance.","core_discovery":"The paper's central claim is that operator splitting and randomized dynamical low-rank approximation can be combined into practical integrators for (1.1). The stiff linear flow $\\Phi^A_\\tau(M_0) = e^{\\tau A}M_0 e^{\\tau A^\\top}$ preserves rank exactly, so the approximation stays on the rank-$r$ manifold; the nonlinear flow $\\Phi^F_\\tau$ is approximated by a dynamic randomized range finder followed by either a dynamic randomized SVD (DRSVD) or a dynamic generalized Nyström (DGN) post-processing step that solves only small sketched ODEs. Composing these as $\\widetilde L_\\tau = \\Phi^A_\\tau \\circ \\widetilde\\Phi^F_\\tau$ yields a first-order Lie-Trotter method and $\\widetilde S_\\tau = \\Phi^A_{\\tau/2} \\circ \\widetilde\\Phi^F_\\tau \\circ \\Phi^A_{\\tau/2}$ yields a second-order Strang method, with rank-adaptive variants based on tolerance-driven truncation. On the Allen-Cahn equation discretized on $1024\\times1024$ grids and on a differential Riccati equation, the methods show the expected temporal orders and lower relative errors than the low-rank projector-splitting reference.","pith_inferences":["A natural extension the paper leaves implicit is to monitor the sketched residual $\\|B(t)-QQ^\\top B(t)\\|_F$ during the nonlinear update and restart or enrich the sketch when it grows, which would guard against the unstated range-inclusion error behind Eq. (2.5).","Because the composition is symmetric, higher-order methods could likely be built by composing the same Strang step with weighted coefficients, although the paper does not attempt this.","The authors note in the concluding remarks that truncated SVD may become unstable for large initial perturbations; a concrete extension would be to swap the SVD truncation for a regularized or QLP-based truncation, or to combine the adaptive rank selection with a restart strategy.","The DGN variant's double sketching of both row and column spaces should make it preferable for nonsymmetric or rectangular problems, an advantage not directly tested by the symmetric Allen-Cahn and Riccati examples."],"forward_implications":["The Lie-Trotter variant converges at order one and the Strang variant at order two in time on stiff semilinear matrix ODEs of the form (1.1).","Because the stiff linear part is integrated exactly by matrix exponentials, the methods avoid explicit CFL-type step-size restrictions and are suited to long-time simulations.","Both DRSVD and DGN avoid forming full-rank matrices in the nonlinear step, working with sketched matrices of size $m \\times (r+p)$ and related small cores, so the cost scales with the chosen rank rather than with the full state dimension.","The rank-adaptive extensions replace fixed-rank truncation by tolerance-based truncation, so users do not need to preselect the target rank."],"supporting_citations":[{"why":"Supplies the randomized dynamical low-rank methods (range finder, DRSVD, DGN) that the paper embeds into the splitting framework for the nonlinear subproblem.","marker":"[8]"},{"why":"Supports the exact exponential integration of the stiff linear subproblem, which removes step-size restrictions.","marker":"[14]"},{"why":"Provides the randomized range finder and randomized SVD theory that underpin the dynamic sketching steps of the nonlinear solver.","marker":"[19]"},{"why":"The BUG-type integrators used as a benchmark for single-step efficiency, which the randomized methods are claimed to match.","marker":"[27]"},{"why":"The projector-splitting integrator for dynamical low-rank approximation that the paper's randomized post-processing is designed to avoid.","marker":"[32]"},{"why":"The low-rank Lie-Trotter splitting for stiff matrix differential equations used as the numerical baseline for convergence and accuracy comparisons.","marker":"[39]"},{"why":"Provides the generalized Nyström method and sketching framework from which the DGN variant is adapted.","marker":"[49]"}],"fun_headline_variants":["Randomized splitting tames stiff matrix ODEs at orders 1 and 2","First- and second-order randomized low-rank integrators for stiff ODEs","Split-and-randomize low-rank solver hits order 1 and 2 on stiff problems","Randomized low-rank splitting achieves first and second order on stiff ODEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole convergence story depends on the randomized range finder tracking the true dominant row and column spaces of the nonlinear flow closely enough that replacing $F(t,N(t))$ with $F(t, B(t)(\\Omega^\\top\\Omega)^{-1}\\Omega^\\top)$ (and the analogous post-processing projection) introduces errors smaller than the splitting error; the paper does not state or bound the error from the unsketched complement.","fun_headline_variants_meta":{"raw":{"variants":["Randomized splitting tames stiff matrix ODEs at orders 1 and 2","First- and second-order randomized low-rank integrators for stiff ODEs","Split-and-randomize low-rank solver hits order 1 and 2 on stiff problems","Randomized low-rank splitting achieves first and second order on stiff ODEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000593,"raw_usage":{"total_tokens":2786,"prompt_tokens":958,"completion_tokens":1828,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":1753}},"tokens_in":574,"tokens_out":1828,"duration_ms":13189,"temperature":1.0,"reasoning_tokens":1753,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:38:54.469289+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small semilinear matrix ODE with a known exact solution, run one step of the nonlinear subproblem, and compare the true $N(t_0+\\tau)$ with the sketched reconstruction $B(t_0+\\tau)(\\Omega^\\top\\Omega)^{-1}\\Omega^\\top$ for a few Gaussian sketches of increasing oversampling $p$; if this reconstruction error fails to decay with $p$ or $\\tau$, or if the overall method's error is dominated by this term, the range-inclusion premise fails and the claimed convergence orders do not hold in general.","supporting_citations":[{"cited_title":"part 1: Dirichlet boundary conditions","cited_arxiv_id":null,"evidence_quote":"Supports the exact exponential integration of the stiff linear subproblem, which removes step-size restrictions."},{"cited_title":"SIAM Rev","cited_arxiv_id":null,"evidence_quote":"Provides the randomized range finder and randomized SVD theory that underpin the dynamic sketching steps of the nonlinear solver."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The BUG-type integrators used as a benchmark for single-step efficiency, which the randomized methods are claimed to match."},{"cited_title":"BIT 54, 171–188 (2014)","cited_arxiv_id":null,"evidence_quote":"The projector-splitting integrator for dynamical low-rank approximation that the paper's randomized post-processing is designed to avoid."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The low-rank Lie-Trotter splitting for stiff matrix differential equations used as the numerical baseline for convergence and accuracy comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the generalized Nyström method and sketching framework from which the DGN variant is adapted."}],"review_version":2}