{"id":"8f6489fb-1c3c-47fd-9ce4-b682a66ae6cb","arxiv_id":"2509.02720","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A spacetime wavelet method for linear PDEs produces a Sylvester matrix equation, solved efficiently with Global GMRES and a wavelet-based recursive initial guess.","lead":"A new numerical solver discretizes linear partial differential equations in both space and time using wavelets, turning the problem into a Sylvester matrix equation solved with Global GMRES. The method shows high-order accuracy and beat a standard Kronecker-product formulation in tests on diffusion and convection-diffusion problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Own scaling data in §4.3 complicate the claimed Gl-GMRES advantage: the fitted exponents (Gl-GMRES ~ O(N^2.5), restarted GMRES ~ below O(N^2)) imply a possible crossover at larger N, so the 'consistently faster' claim is not yet supported.","rationale":"I read the paper as making two separable claims: (1) the spacetime wavelet discretization yields high-order accurate solutions to linear IBVPs, and (2) solving the resulting Sylvester equation with Gl-GMRES, augmented by the recursive wavelet initial guess, is computationally preferable to the Kronecker-product/restarted-GMRES route. The accuracy claim is supported by manufactured-solution convergence studies and a priori wavelet error estimates, and I have no substantive objection there. The efficiency claim is the load-bearing part of the paper's novelty, and it is the least secure. The paper itself reports fitted complexities with Gl-GMRES at O(N^2.5) and restarted GMRES below O(N^2); if true, the demonstrated speed advantage at j = 1..5 cannot persist indefinitely. The test I propose runs the same comparison at two larger resolution levels, which directly checks whether the claimed advantage is asymptotic or merely a small-system effect. The reader's weakest assumption was broader Gl-GMRES convergence robustness; I partially agree, but I sharpen it to the concrete scaling crossover implied by the paper's own data. This concern does not overturn the numerical verification, but it does mean the headline performance claim should be stated conditionally until the scaling behavior is checked at larger N.","tokens_in":15819,"tokens_out":7310,"duration_ms":72072,"concrete_test":"Extend the Section 4.3 comparison to j = 6 and j = 7 for the diffusion problem (N approximately 3.9e5 and 1.5e6) with px = 6, pt = 4, m = 30(j+1), nu = 0.1, zero initial guess, and identical hardware. Record wall time and total flops/iterations for both Gl-GMRES and restarted GMRES. If Gl-GMRES time or operation count grows faster and crosses over by j = 7, the 'consistently faster' claim should be limited to small systems; if the gap persists or widens, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central efficiency claim rests on Section 4.3, where Gl-GMRES on the Sylvester form is compared with restarted GMRES on the Kronecker system for j = 1 through 5. The paper's own complexity fit in Fig. 8b gives Gl-GMRES approximately O(N^2.5) and restarted GMRES below O(N^2). If these fits describe the asymptotic regime, the advantage shown in Fig. 8a must reverse at some sufficiently large N; at minimum, the comparison has not established that solving in Sylvester form is the better strategy for larger spacetime grids. The demonstration also relies on a single heuristic restart rule m = 30(j+1), a single viscosity value, and no preconditioning, so the convergence behavior of Gl-GMRES is not shown to generalize beyond the two model problems. This does not invalidate the high-order convergence results, but it does mean the headline performance claim is extrapolated from a small, possibly non-representative range.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spacetime wavelet discretization for linear initial-boundary value problems, based on Deslauriers-Dubuc wavelets and dense wavelet derivative operators. The discretization produces a Sylvester matrix equation, solved here with global GMRES (Gl-GMRES) using a modified global Arnoldi process; boundary and initial conditions are enforced with permutation matrices, and a recursive wavelet-synthesis algorithm supplies initial guesses for successive resolution levels. Verification studies on one-dimensional linear diffusion and convection-diffusion equations with manufactured solutions report high-order convergence rates for both solution and derivative approximations, consistent with the a priori estimates cited from wavelet theory. The paper also claims that the Gl-GMRES Sylvester-form solver outperforms restarted GMRES applied to the Kronecker-product form, and that the recursive initial-guess technique reduces iteration counts and wall time.","tokens_in":16031,"tokens_out":6195,"duration_ms":58688,"significance":"If fully supported, the paper would be a useful practical contribution: it demonstrates an alternative to Kronecker-form spacetime discretizations, where the Sylvester form substantially reduces the number of stored nonzeros, and it provides a simple recursive initialization that accelerates Gl-GMRES. The convergence portion is credible and well executed: manufactured solutions are used, the observed rates match the predicted p-alpha orders, derivative superconvergence is documented, and the algorithms are given in sufficient detail to be reproduced. The main weakness is that the headline performance claim rests on a narrow, confounded comparison in Section 4.3; the authors' own complexity fits suggest a possible crossover at larger problem sizes, so the claim that the Sylvester formulation is 'consistently faster' is not yet established. The convergence results are sound, but the efficiency claim needs additional work before the paper can be accepted as a demonstration of the Sylvester form's superiority.","major_comments":[{"comment":"The claim that Gl-GMRES on the Sylvester form is 'consistently faster' is not supported beyond the tested range. The text reports fitted scalings of approximately O(N^2.5) for Gl-GMRES and slightly below O(N^2) for restarted GMRES on the Kronecker system. If these fits describe the asymptotic behavior, the advantage shown in Fig. 8a must eventually reverse at sufficiently large N. Please provide results at larger N, or present the fits with confidence intervals and an explicit crossover estimate; otherwise the claim should be restricted to the tested problem sizes.","section":"§4.3, Fig. 8"},{"comment":"The performance comparison changes two variables simultaneously: the algebraic formulation (Sylvester versus Kronecker) and the iterative solver (Gl-GMRES versus restarted GMRES). This does not isolate the benefit of the Sylvester form. To substantiate the abstract and conclusion claims that the Sylvester formulation outperforms the Kronecker-product approach, the comparison should either hold the solver fixed across formulations, include a solver appropriate to the vectorized Kronecker system (for example a preconditioned GMRES variant), or be explicitly framed as a comparison of the authors' combined solver choice, not of the algebraic form alone.","section":"§4.3"},{"comment":"The efficiency and recursive-initialization results depend on the heuristically chosen restart parameter m = 30(j+1) and on the tolerances tol_H = 10^{-8} and residual tolerance 10^{-8}. The paper states that similar results are obtained for other values of m, but gives no sensitivity data and no convergence analysis for the restarted Gl-GMRES iteration. Since the central performance claim relies on this restart rule, please include a sensitivity study (e.g., iteration counts and times for several m values at each level) or a theoretical justification for the rule; otherwise the generality of the conclusions is unclear.","section":"§3.3 and §4.3"}],"minor_comments":[{"comment":"The text states that 'Fig. 9 shows that the number of iterations is drastically attenuated,' but Fig. 9 shows relative time; the iteration counts appear in Fig. 10. Please correct the citation.","section":"§4.4, Fig. 9 and Fig. 10"},{"comment":"The caption for Fig. 7 says the convection-diffusion problem is at j = 2 with px = 6, pt = 4, while the surrounding text states that px = pt = 8 is used for this example. The caption and text should be made consistent.","section":"Fig. 7 caption"},{"comment":"If the modified global Arnoldi loop breaks early because H(z+1,z) < tol_H, the least-squares problem in Algorithm 1 is written for y in R^m and H of size (m+1) x m, while the actual H has smaller dimensions. Please clarify how the reduced H and V are handled during a restart cycle.","section":"Algorithms 1 and 2"},{"comment":"The level ranges in the captions are inconsistent: Table 1 and Fig. 3 use j = 3, 4, 5, while Table 2 and Fig. 6 use different ranges (j = 2, 3, 4 in the text). Please harmonize the level ranges in the captions and bodies.","section":"Tables 1 and 2 and Figs. 3 and 6"},{"comment":"The abstract promises 'user-prescribed error estimates,' but the paper presents a priori asymptotic rates of the form O(delta q^{p-alpha}), not user-specified tolerance-controlled error estimates. Please rephrase to avoid overstating what is demonstrated.","section":"Abstract and §2.3"},{"comment":"The vertical axis of Fig. 8b is labeled 'number of operations,' but the text does not define how this quantity is computed. Please specify whether it counts matrix-vector products, flops, or some other measure, and how the fits were obtained.","section":"Fig. 8b"},{"comment":"The title contains a spacing artifact, 'Proble ms'; this should be corrected in the final version.","section":"Title, page 1"}],"recommendation":"major_revision","confidential_remarks":"The convergence results are solid and the paper is generally well organized, but the performance claim in the abstract and conclusions is not yet backed by the evidence in Section 4.3. The fitted scalings reported by the authors themselves make the crossover concern concrete, and the comparison conflates formulation with solver choice. I believe this can be fixed within the scope of the manuscript by extending the experiments, adding a cleaner comparison, or tempering the claims. The overlap with the authors' earlier nonlinear spacetime wavelet paper is acknowledged, but the Sylvester-form and recursive-initialization aspects are sufficiently distinct for this venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the convergence/verification side of this paper is honest and solid enough to referee; the performance claim in the abstract, as stated, is not supported by the paper's own data. The genuinely new bit is the recursive wavelet-synthesis initial guess for Global GMRES on Sylvester-form spacetime wavelet systems. That is a small but real contribution, and the paper shows it reduces iteration counts on the two test problems.\n\nWhat it does well: the MMS verification is careful. Convergence rates match wavelet theory, derivative superconvergence is demonstrated, and the Sylvester form obviously saves memory versus the Kronecker system. The error estimates are a priori, not fitted. This part is reproducible in principle, though no code or data are provided.\n\nSoft spots: Section 4.3's comparison is the load-bearing efficiency claim, and it doesn't hold as stated. The comparison confounds formulation (Sylvester vs Kronecker) with solver (Gl-GMRES vs restarted GMRES). More importantly, their own fitted complexities in Figure 8b give Gl-GMRES about O(N^2.5) while restarted GMRES is below O(N^2). If these fits mean anything, the Sylvester route loses at larger N. The 'consistently faster' claim in the abstract and conclusions is therefore only true for the tested range, j=1 to 5. The stress-test note is right to flag this. The recursive initialization is plausible but lacks theoretical backing; it is demonstrated on two small problems with a single restart heuristic m=30(j+1). There is no spectrum-based argument that Gl-GMRES will behave well for broader PDE classes. These are addressable, but they should be fixed before publication. Also, no preconditioning is used, which limits the practical value claim—and including preconditioned competitors would be fair.\n\nWho this is for: people working on wavelet-based spacetime solvers or matrix-equation formulations of discretized PDEs. For that reader, the recursive initialization and careful verification are useful. It is not a field-changing paper, but it is a legitimate incremental contribution.\n\nRecommendation: it deserves peer review. A good referee should push for a corrected performance section, a discussion of the complexity crossover, and ideally code release. I would accept it conditionally, not reject.","headline":"The accuracy results are solid and the recursive initialization is a real small contribution, but the paper's own scaling fits undercut the headline performance claim, so the comparison needs rework before acceptance.","tokens_in":16542,"tokens_out":2630,"would_cite":true,"duration_ms":23486,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65T60","65F10","65M70","65M12"],"pacs":[],"model":"deepseek-v4-flash","headline":"A spacetime wavelet discretization turns linear initial-boundary value problems into Sylvester matrix equations solved by Global GMRES with recursive initial guesses, achieving predicted high-order convergence at lower memory cost than…","keywords":["spacetime wavelet method","Sylvester matrix equation","Global GMRES","linear initial-boundary value problems","Deslauriers-Dubuc wavelets","wavelet derivative connection coefficients","recursive initial-guess algorithm","high-order convergence"],"falsifier":"Run the method on an advection-dominated linear problem, such as convection-diffusion with $c \\gg \\nu$, or on a hyperbolic or oscillatory linear PDE, at successive resolution levels while tracking Global GMRES iterations and cumulative wall time against the zero-initial-guess baseline. If iteration counts grow superlinearly with level $j$, or the recursive solver becomes slower than solving directly at $j_{\\max}$, the efficiency claim is refuted.","tokens_in":1579,"feed_emoji":"🧮","tokens_out":1845,"duration_ms":45006,"temperature":0.7,"pith_summary":"The paper proposes solving linear initial-boundary value problems by discretizing both space and time simultaneously with Deslauriers-Dubuc wavelets, so the PDE becomes a Sylvester matrix equation AX + XB = C rather than a single large Kronecker-product system. To solve it, the authors use Global GMRES and feed it recursively generated wavelet-synthesized initial guesses. They report high-order convergence rates, matching wavelet theory, for both solutions and derivatives on diffusion and convection-diffusion test problems, and report that the Sylvester/Global GMRES route is consistently faster than restarted GMRES on the Kronecker formulation. A sympathetic reader would care because spacetime methods avoid timestep stability limits and offer a priori error control, while the Sylvester form avoids the memory blow-up of vectorized systems.","feed_headline":"Wavelet spacetime solver beats Kronecker GMRES on linear PDEs","feed_subtitle":"The Sylvester form skips the Kronecker blow-up and recursive wavelet guesses cut iteration counts.","key_machinery":"The central objects are the Deslauriers-Dubuc wavelet basis and its derivative connection coefficients $\\Gamma$, which discretize derivatives directly in space and time. Applying these operators to a linear PDE puts the discrete system in Sylvester form $AX+XB=C$; semi-orthogonal permutation matrices $P_x$, $P_t$ enforce initial and boundary conditions while keeping the reduced system in the same Sylvester structure. Global GMRES with a Modified Global Arnoldi process, which generates orthogonal matrices rather than vectors, solves the equation, and the paper's recursive algorithm synthesizes coarser-grid solutions into informed initial guesses for finer grids.","core_discovery":"On the paper's own terms, the central claim is that a fully wavelet-based spacetime discretization of linear PDEs produces Sylvester equations that can be solved accurately and efficiently: numerical solutions and derivative approximations converge at the rates predicted by wavelet theory, with user-prescribed error estimates, and the Sylvester form can be solved with Global GMRES rather than the Kronecker-product vectorized system. The recursive wavelet algorithm synthesizes the solution at one resolution level into the initial guess for the next level, reducing the number of Global GMRES iterations and cumulative time to solution compared with starting from zero at the target level. The paper further claims that this Sylvester route outperforms the standard Kronecker-product formulation, which requires storing orders of magnitude more nonzero entries.","pith_inferences":["If the convergence behavior is generic, the method could be applied to stiff regimes such as small-diffusion, strong-advection linear problems where explicit timestepping fails, since the spacetime formulation has no visible CFL restriction.","A natural next step, absent from the paper, would be a convergence theory for Global GMRES on this class of Sylvester systems linking the spectra of $A$ and $B$ to iteration bounds; the paper only provides empirical spectra.","The recursive initial-guess idea should transfer to nonlinear problems, where prior-level solutions could initialize Newton-like iterations for both the state and the Jacobian.","The K-matrix spectra for the convection-diffusion case show a vertical band absent in the diffusion case, suggesting that advection-dominated limits may hit a performance cliff requiring different restart values or preconditioning; this is a testable boundary of the efficiency claim."],"forward_implications":["The discretization provides a built-in a priori error estimate: truncation error is $O(\\Delta q^{p-\\alpha})$, dictated by the weakest derivative order in the PDE, so users can predict convergence order before solving.","The full spacetime formulation avoids timestep stability restrictions, sidestepping the explicit/implicit timestep trade-off of conventional semi-discretizations.","For the diffusion test at level $j=5$, the Sylvester matrices together store roughly 4,449 nonzeros while the equivalent Kronecker product matrix stores 1,267,589 nonzeros, giving a concrete memory advantage.","The recursive wavelet-synthesized initial-guess procedure reduces Global GMRES iteration counts and cumulative wall time relative to a zero initial guess at every tested resolution level.","Boundary and initial condition enforcement via permutation matrices preserves well-posedness, with uniqueness guaranteed when the spectra of $A$ and $-B$ are disjoint."],"supporting_citations":[{"why":"Supplies the spacetime wavelet discretization framework for nonlinear PDEs that this paper adapts to the linear Sylvester setting.","marker":"[20]"},{"why":"Provides the semi-orthogonal permutation-matrix technique used to enforce initial and boundary conditions while preserving Sylvester structure.","marker":"[44]"},{"why":"Defines the Global GMRES and Modified Global Arnoldi algorithms used as the core iterative solver.","marker":"[53]"},{"why":"Provides the wavelet derivative connection coefficients and their eigenvector construction used to build the discrete operators.","marker":"[38]"},{"why":"Supplies the nested multiresolution spaces and a priori error estimates the paper's convergence claims rely on.","marker":"[45]"},{"why":"Documents the computational inefficiency of the Kronecker/vectorization formulation that the Sylvester route avoids.","marker":"[79]"}],"fun_headline_variants":["Wavelet spacetime solver turns linear PDEs into Sylvester equations","Sylvester form spares wavelet PDE solver from Kronecker blow-up","Recursive wavelet initial guesses accelerate GMRES for PDEs","High-order spacetime wavelets meet GMRES for linear PDEs","Wavelet spacetime method solves Sylvester equations from PDEs"],"cache_read_input_tokens":18688,"weakest_assumption_plain":"The load-bearing premise is that Global GMRES converges quickly enough on the Sylvester systems produced by the spacetime wavelet discretization, without preconditioning and with the fixed restart rule $m=30(j+1)$, across the class of linear PDEs the method claims to target.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet spacetime solver turns linear PDEs into Sylvester equations","Sylvester form spares wavelet PDE solver from Kronecker blow-up","Recursive wavelet initial guesses accelerate GMRES for PDEs","High-order spacetime wavelets meet GMRES for linear PDEs","Wavelet spacetime method solves Sylvester equations from PDEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000625,"raw_usage":{"total_tokens":2841,"prompt_tokens":842,"completion_tokens":1999,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":1910}},"tokens_in":458,"tokens_out":1999,"duration_ms":13747,"temperature":1.0,"reasoning_tokens":1910,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:35:10.286535+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on an advection-dominated linear problem, such as convection-diffusion with $c \\gg \\nu$, or on a hyperbolic or oscillatory linear PDE, at successive resolution levels while tracking Global GMRES iterations and cumulative wall time against the zero-initial-guess baseline. If iteration counts grow superlinearly with level $j$, or the recursive solver becomes slower than solving directly at $j_{\\max}$, the efficiency claim is refuted.","supporting_citations":[{"cited_title":"Cochran and Karel Matouˇ s","cited_arxiv_id":null,"evidence_quote":"Supplies the spacetime wavelet discretization framework for nonlinear PDEs that this paper adapts to the linear Sylvester setting."},{"cited_title":"Sylvester equations a nd the numerical solution of partial fractional differential equations","cited_arxiv_id":null,"evidence_quote":"Provides the semi-orthogonal permutation-matrix technique used to enforce initial and boundary conditions while preserving Sylvester structure."},{"cited_title":"Global FOM and GMRES algorithms for matrix equations","cited_arxiv_id":null,"evidence_quote":"Defines the Global GMRES and Modified Global Arnoldi algorithms used as the core iterative solver."},{"cited_title":"Van Loan","cited_arxiv_id":null,"evidence_quote":"Documents the computational inefficiency of the Kronecker/vectorization formulation that the Sylvester route avoids."}],"review_version":2}