{"id":"b083d76d-99e9-4b17-b0e4-4fcd67744a10","arxiv_id":"2412.08076","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Neural networks predict the weights of momentum-accelerated long-step Richardson iterations, reducing iteration counts on anisotropic diffusion and Helmholtz problems versus Chebyshev-based iterations in numerical experiments.","lead":"A new family of neural solvers learns the step sizes of a long-step Richardson iterative method from PDE parameters, and embeds them in multilevel networks to solve anisotropic diffusion and Helmholtz equations faster than standard Chebyshev-accelerated iterations. The experiments show large reductions in iteration counts, but the claims rely on in-distribution fitting and lack runtime measurements.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The multilevel robustness claim is contradicted by the paper's own scaling tables: FNS iterations jump from 12 to 83 between N=256 and N=512, and WANS counts grow from 8.8 to 336.6, so 'parameter-independent convergence' is not supported.","rationale":"The reader's weakest assumption was that the meta-network trained at N=64 might not generalize to larger grid sizes. My concern is related but stronger: the paper's own reported tables show a clear, and in some cases superlinear, growth of iteration counts with grid size, so the multilevel 'parameter-independence' claim is not merely unproven—it is contradicted by the evidence presented. I agree with the reader's conditional verdict because the core algorithmic idea and the reported improvements over FNS/WANS are still plausible and worth publishing in revised form, but the abstract and introduction should be toned down to claim 'better robustness than the original FNS/WANS on tested cases' rather than 'effective overcoming' of single-level limitations or 'parameter-independent convergence.' The proposed concrete test is an analytical reanalysis of Tables 8 and 9, so it can be run immediately by the authors or referees. This does not require rejecting the paper, but it makes the conditional acceptance more pointed: the stated grid-independence claims must be removed or supported by additional large-scale experiments.","tokens_in":13046,"tokens_out":8117,"duration_ms":89595,"concrete_test":"Compute the empirical scaling exponents from Tables 8 and 9 by fitting log10(iterations) versus log10(N) over the last three grid sizes in each table. For Richardson(10)-FNS, the N=256-to-512 segment gives an exponent of roughly log2(83/12) ≈ 2.8; for NAG-Richardson(5)-WANS, the N=2048-to-4096 segment gives roughly log2(336.6/134.9) ≈ 1.3. If these exponents are not close to zero, the 'mild grid dependence' and 'parameter-independent convergence rate' claims fail on the paper's own data. Optionally run Richardson(10)-FNS at N=1024 with the same trained networks to confirm that the N=512 jump is not an outlier.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central multilevel claim is that the proposed solvers 'effectively overcome the drawback of single-level methods' and, for WANS, have a 'parameter-independent convergence rate.' The paper's own tables contradict this. For Richardson(10)-FNS, Table 8 shows iteration counts of 5, 6, 9, 12, and 83 for N = 32, 64, 128, 256, and 512; the jump from 12 to 83 is more than a sixfold increase over a single grid doubling. For NAG-Richardson(5)-WANS, Table 9 shows iteration counts rising from 8.8 at N=128 to 336.6 at N=4096, which is not a parameter-independent rate. This is not merely an untested extrapolation: the reported in-range data already fail the stated property. Moreover, the Meta-omega network is trained at N=64 with input (lg epsilon, theta) and no dependence on grid size, so there is no mechanism for the learned weights to adapt to the grid-size-dependent spectrum; the single-level Table 4 exhibits the expected failure, with iterations growing from 27 at N=64 to 977 at N=512. The evidence supports the weaker claim that the new solvers outperform the original FNS and WANS on the tested cases, but not the abstract's stronger robustness and parameter-independence claims. A separate issue is that 'lower computational complexity' is asserted from iteration counts without runtime or operation-count measurements, but the scale-dependence contradiction is the more direct defect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a family of learned long-step Richardson iterative solvers, termed Richardson(m)-NS, together with momentum-accelerated variants and SSOR preconditioning, and integrates them as smoothers into the multilevel neural solvers FNS and WANS. The central claims are that these solvers achieve faster convergence and lower computational complexity than Chebyshev iteration and Chebyshev semi-iteration, and that the multilevel versions Richardson(m)-FNS and NAG-Richardson(m)-WANS overcome the grid- and parameter-dependence of single-level solvers, with the WANS version having a parameter-independent convergence rate. These claims are supported by tables of iteration counts from numerical experiments on anisotropic diffusion and high-wavenumber Helmholtz problems.","tokens_in":13356,"tokens_out":6748,"duration_ms":67126,"significance":"The algorithmic contribution is interesting and the experiments are broad in their coverage of PDE parameters. The Richardson(m)-NS framework with momentum and preconditioning is clearly described, and the alternating optimization training for the multilevel FNS variant is a practically useful technique. If the performance were reproducible and the claims were restricted to the settings actually tested, the paper would be a useful addition to the neural-solver literature. However, the headline claims of parameter independence and lower computational complexity are not supported by the data presented, and the multilevel scaling results in the paper's own tables contradict the stated robustness. The current form therefore overstates the significance of the experiments.","major_comments":[{"comment":"The abstract and §3.2 state that NAG-Richardson(m)-WANS has a 'parameter-independent convergence rate,' but Table 9 shows FGMRES iteration counts growing from 8.8 at N=128 to 336.6 at N=4096, i.e., roughly a 38-fold increase over a 32-fold grid-size increase and an approximately linear growth at the largest sizes. This is not parameter-independent, and the improvement over the original WANS at N=4096 is marginal (336.6 vs. 344.3 iterations). The claim should be removed or replaced with a precise statement about the range and behavior actually observed.","section":"Abstract and §3.2, Table 9"},{"comment":"The Introduction asserts that Richardson(m)-FNS is 'mildly dependent on grid size,' but Table 8 shows iteration counts rising from 12 at N=256 to 83 at N=512, a six-fold increase over a single grid doubling. The single-level solver in Table 4 grows from 27 iterations at N=64 to 977 at N=512. The data therefore do not support the assertion that the multilevel versions 'effectively overcome the drawback of single-level methods' with respect to grid-size dependence. The authors should either report experiments at larger N that demonstrate a bounded or slowly growing iteration count, or substantially weaken this claim.","section":"Introduction, §3.1, Table 8"},{"comment":"The claim of 'lower computational complexity' relative to Chebyshev iteration and Chebyshev semi-iteration is not established by the evidence. Only iteration counts are reported; no runtimes, flop counts, or matrix-vector product counts are given. The Chebyshev iteration in Table 2 is based on Eq. (6), which requires knowledge of λmax and λmin, but the computational cost of obtaining these (for example, via the power method or a dense eigensolver) is not accounted for in the comparison. Per-iteration costs also differ among the proposed variants because of momentum terms, the SSOR preconditioner, and the Meta-omega network evaluation. The paper should either provide a complexity analysis that includes these costs or restrict the claims to iteration counts.","section":"Abstract and §2.3"},{"comment":"The training loss L2 (Eq. (15)) is the same relative-residual metric used as the stopping criterion in all reported tables. This means the reported iteration counts measure the objective for which Meta-omega was trained, and the test parameters in the single-level experiments are within the training distribution (ε ∈ [10^-6,1], θ ∈ [0,π]) at the training grid size N=64. The degradation observed at N=512 in Tables 4 and 8 is an out-of-distribution extrapolation, and the Meta-omega network has no input that encodes grid size or spectral information beyond (lg ε, θ). To support claims of robustness or parameter independence, the authors should test on parameters outside the training distribution and report statistics over many random samples rather than single deterministic points.","section":"§2.2, Eq. (15), §2.3, §3.1"},{"comment":"The comparison to 'Chebyshev iteration (m=3 in (6))' is not apples-to-apples because Eq. (6) requires the exact extreme eigenvalues λmax and λmin, and the text itself notes that computing λmin is costly. The paper does not state how these reference results were obtained. If the eigenvalues were computed exactly, the reference method is an oracle that does not correspond to an implementable solver with the same per-iteration cost; otherwise, the reader cannot judge whether the reported iteration savings translate into real computational savings. The authors should describe the reference computation and account for its cost.","section":"§2.3, Table 2"}],"minor_comments":[{"comment":"The phrase 'This paper first introduce' should be 'This paper first introduces'; there are several similar grammar errors throughout.","section":"Abstract"},{"comment":"The sentence 'we obtain a Richardson(m)-FNS that is independent of PDE parameters and mildly dependent on grid size' is imprecise; Table 8 does not support 'mildly dependent.'","section":"Introduction"},{"comment":"There is a typo in 'x− and y− directionss' and the spacing in 'W ANS' is inconsistent across the text and tables.","section":"§3.2"},{"comment":"The phrase 'preprocessing techniques' in the Conclusions should read 'preconditioning techniques.'","section":"Conclusions"},{"comment":"The description of m as 'continuous m (with m ≥ 1)' is odd, since m is an integer number of inner iterations.","section":"§2.1"},{"comment":"The claim that NAGex-Richardson(3)-NS with SSOR achieves 'nearly a 20 times speedup' is not consistent with Table 2, where the iteration ratio relative to Richardson(3)-NS is about 77 at ε=10^-6 and about 28 at ε=1; please re-check the calculation and state the baseline clearly.","section":"Introduction and §2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's useful core is the demonstration that meta-learned Richardson weights with momentum and preconditioning can accelerate existing neural solvers on specific test cases. The biggest risk is that the authors' framing overstates the results: the abstract promises parameter-independent convergence and lower complexity, but the paper's own scaling tables contradict both. A revised version that focuses on the empirical improvements, provides a proper complexity accounting, and includes out-of-distribution scaling tests would be much more defensible. Also note that the improvement over the original WANS at the largest grid is marginal (336.6 vs. 344.3), so the 'improved performance' claim for NAG-Richardson(5)-WANS is weaker than the abstract suggests."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The single-level results are the real contribution: the learned Richardson weights land close to Chebyshev-optimal without eigenvalue estimates, and the momentum variants show consistent speedups. But the paper overreaches in the abstract, claiming parameter-independent multilevel convergence and lower computational complexity when its own tables and the absence of runtime measurements don't support either.\n\nWhat's new: the meta-network weight prediction for long-step Richardson, the NAG-extension tweak, and the integration of these as smoothers in FNS and WANS. The experiments are thoughtful: they compare against Chebyshev variants, report iteration counts across anisotropy, direction, and grid size, and the qualitative effects (momentum helps, multilevel helps) come through. Credit where due: the Richardson(m)-NS at m=15 hits 29 iterations at epsilon=1 versus 3999 for Richardson(1)-NS, and the SSOR-preconditioned NAGex-Richardson(3)-NS reaches 20-43 iterations. That is a real practical gain.\n\nThe soft spots are three. First, 'lower computational complexity' is asserted from iteration counts alone; no wall-clock time or operation counts. Chebyshev semi-iteration also needs only an eigenvalue estimate, and the meta-network inference has a cost. So that claim is unsupported. Second, the multilevel robustness claims are contradicted by the paper's own tables: Richardson(10)-FNS goes from 12 iterations at N=256 to 83 at N=512, and NAG-Richardson(5)-WANS goes from 8.8 at N=128 to 336.6 at N=4096. That is not 'parameter-independent convergence' by any reading; it is better than the baselines, which is the defensible claim. Third, the meta-network is trained at N=64 only, with no mechanism to adapt to discretization size, and the authors themselves note (Remark 3) that test performance worsened as training loss decreased. That is in-distribution fitting, not generalization. No code or hyperparameters are released, which makes independent verification harder.\n\nWho this is for: people working on learned iterative solvers and parametric PDEs will want to see the single-level weight-learning approach. The multilevel claims should be taken with skepticism until runtime benchmarks and out-of-distribution tests appear. A serious referee should engage with it; I'd send it to review with a request for runtime data, code, and a revision that scales the robustness claims back to what the tables actually show.","headline":"The single-level learned Richardson weights are a genuine, well-tested contribution, but the abstract overclaims multilevel parameter-independence and computational complexity that the paper's own tables contradict.","tokens_in":13934,"tokens_out":2585,"would_cite":false,"duration_ms":24217,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F10","65N55","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A meta-network learns the weights of long-step Richardson iteration, and momentum, SSOR preconditioning, and multilevel embedding make the resulting solvers converge faster than optimal-weight Chebyshev iteration without eigenvalue…","keywords":["Richardson(m) iteration","neural solver","momentum acceleration","SSOR preconditioning","multilevel neural solver","Fourier neural solver","anisotropic diffusion","Helmholtz equation"],"falsifier":"A direct test: train Richardson(3)-NS on epsilon drawn uniformly from 1e-6 to 1 and theta from 0 to pi at N=64, as in the paper, then freeze the network and measure the relative residual after a fixed number of iterations on N=512 anisotropic systems with the same parameter ranges. If the iteration count to reach 1e-6 grows at the same rate as the paper's own Table 4 (27 to 977 iterations), the claimed transfer across grid sizes fails, and the multilevel results cannot be credited to the learned weights.","tokens_in":12761,"feed_emoji":"🧮","tokens_out":8128,"duration_ms":75492,"temperature":0.7,"pith_summary":"The paper's aim is to replace hand-tuned or eigenvalue-derived weights in the long-step Richardson iteration with weights predicted by a small neural network, and to show that the resulting solvers beat classical Chebyshev acceleration. It introduces Richardson(m)-NS, where one outer iteration bundles m weighted inner iterations and a meta-network maps PDE coefficients to the weights. Adding momentum (especially a Nesterov-style correction), SSOR preconditioning, and then embedding the solver inside two multilevel neural solvers yields, in the paper's experiments, faster convergence and lower computational cost than optimal-weight Chebyshev iteration and Chebyshev semi-iteration. The multilevel versions, Richardson(m)-FNS for anisotropic diffusion and NAG-Richardson(m)-WANS for Helmholtz, are the paper's answer to the single-level solvers' sensitivity to PDE parameters and grid size. If true, the practical consequence is that expensive eigenvalue estimation can be replaced by a cheap learned map from PDE parameters to iteration weights.","feed_headline":"Neural weights make Richardson iteration up to 20x faster","feed_subtitle":"A small network picks Richardson(m) weights, cutting eigenvalue costs and stabilizing multilevel convergence across grids.","key_machinery":"The central object is the Richardson(m) iteration, which treats m consecutive weighted updates $u_k = (I-\\omega_i A)u_{k-1}+\\omega_i f$ as one outer iteration with polynomial operator $T_m(A,\\omega)=\\prod_{i=1}^m (I-\\omega_i A)$. The load-bearing mechanism is a meta-network $\\mathrm{Meta}(\\theta;\\mu)$ that maps PDE parameters $\\mu$ (anisotropy strength and angle, transformed to $\\lg\\varepsilon$) directly to the weight vector $\\omega$, replacing eigenvalue-based weight formulas. The paper then layers on momentum variables ($\\alpha_i$), a Nesterov-style lookahead (with an extension $\\tilde{\\alpha}_i$), and an SSOR preconditioner for the single-level variants, and an alternating-optimization training algorithm when the solver is embedded as a smoother in the FNS and WANS multilevel architectures.","core_discovery":"The paper's central claim is that the weights of the m-step Richardson iteration do not need to come from the spectrum of A: a meta-network trained on random PDE parameters can predict weights that make Richardson(m) converge nearly as fast as the Chebyshev iteration with optimal weights, and faster than Chebyshev semi-iteration, while avoiding eigenvalue estimates. The claim is made in two stages. At the single level, momentum and preconditioning variants, MOM-Richardson(m), NAG-Richardson(m), and NAGex-Richardson(m), are proposed, with the SSOR-preconditioned NAGex version reported to be nearly 20 times faster than plain Richardson(3)-NS on the test problem. At the multilevel level, replacing the fixed smoother in FNS with Richardson(m)-NS, and the Chebyshev semi-iterative smoother in WANS with NAG-Richardson(m), is claimed to overcome the strong dependence on PDE parameters and grid size: Richardson(10)-FNS converges in 4 to 10 iterations for anisotropy strengths from 1 down to 1e-8 and all tested angles, and NAG-Richardson(5)-WANS preconditioning FGMRES gives lower iteration counts than WANS with Chebyshev semi-iteration up to N=4096. The paper presents these as numerical findings, with the theoretical backbone being the known faster convergence of long-step Richardson and the Chebyshev optimal weights as benchmark.","pith_inferences":["The paper's tables show that the single-level learned weights degrade sharply as N grows (e.g., 27 iterations at N=64 versus 977 at N=512 for SSOR-preconditioned NAGex); a natural inference is that the multilevel wrapping, not the learned weights alone, is what earns grid robustness, and that testing the meta-network on unstructured meshes would likely require retraining.","Because the FNS and WANS gains are demonstrated on structured-grid model problems, the cleanest extension would be to check the same alternating-trained smoothers on non-uniform grids and on advection-diffusion-reaction equations, where the paper itself says future work lies.","The WANS comparison at large N shows a modest iteration reduction (336.6 vs 344.3 at N=4096); an editor's inference is that the practical advantage there may be the removed eigenvalue/power-method cost rather than dramatically faster convergence."],"forward_implications":["If the learned weights generalize as reported, solving a parametric family of linear systems no longer requires per-matrix eigenvalue estimates: the same small network supplies the Richardson weights for new parameter values.","The reported single-level speedups imply that momentum and SSOR preconditioning can be composed with learned long-step weights: in the m=3 anisotropic test, NAGex-Richardson(3)-NS with SSOR needs about 20 iterations where Richardson(3)-NS needs hundreds.","Embedding learned Richardson smoothers in FNS gives a concrete recipe for anisotropic problems: Richardson(10)-FNS converges in 4 to 10 iterations across the tested anisotropy strengths and directions, whereas the original FNS can diverge or need many more.","For Helmholtz systems, using NAG-Richardson(m) as a WANS smoother removes the power-method eigenvalue estimation and keeps FGMRES iteration counts competitive or better than Chebyshev semi-iteration at all tested grid sizes up to N=4096."],"supporting_citations":[{"why":"Supplies the theoretical result that long-step (m-step) Richardson iteration converges faster, motivating the Richardson(m) formulation.","marker":"[38]"},{"why":"Gives the Chebyshev optimal-weight construction for Richardson(m), the baseline that Richardson(m)-NS is compared against.","marker":"[39, 40]"},{"why":"Supplies the empirical alpha=1/30 choice for the Chebyshev semi-iterative method used as the single-level baseline.","marker":"[42]"},{"why":"Introduces the Fourier neural solver (FNS) whose smoothing and neural-iteration structure is extended by replacing the smoother with Richardson(m)-NS.","marker":"[30]"},{"why":"Provides the CNN architecture for Meta-T and the convergence framework used in the FNS construction.","marker":"[31]"},{"why":"Supplies the Wave-ADR neural solver (WANS), its Helmholtz training/test data, and the smoother baseline replaced by NAG-Richardson(m).","marker":"[12]"},{"why":"Provides the Nesterov accelerated gradient method that is adapted into the NAG-Richardson(m) variants.","marker":"[3, 4]"}],"fun_headline_variants":["Neural-predicted Richardson weights outpace Chebyshev iteration","Meta-network sets Richardson(m) weights, beating Chebyshev","Momentum and preconditioning boost neural Richardson solvers","Multilevel neural Richardson cracks anisotropic and Helmholtz"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that carries the argument is that the meta-network, trained only on 64x64 grids with anisotropy parameters drawn uniformly, predicts Richardson(m) weights that remain effective on finer grids (up to N=512 or N=4096 in the multilevel tests) and at Helmholtz wavenumbers; if the learned weights do not transfer across discretization scale, the claimed multilevel robustness and parameter-independent convergence rate collapse.","fun_headline_variants_meta":{"raw":{"variants":["Neural-predicted Richardson weights outpace Chebyshev iteration","Meta-network sets Richardson(m) weights, beating Chebyshev","Momentum and preconditioning boost neural Richardson solvers","Multilevel neural Richardson cracks anisotropic and Helmholtz"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000705,"raw_usage":{"total_tokens":3216,"prompt_tokens":1022,"completion_tokens":2194,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":2126}},"tokens_in":638,"tokens_out":2194,"duration_ms":17144,"temperature":1.0,"reasoning_tokens":2126,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:15:03.037570+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: train Richardson(3)-NS on epsilon drawn uniformly from 1e-6 to 1 and theta from 0 to pi at N=64, as in the paper, then freeze the network and measure the relative residual after a fixed number of iterations on N=512 anisotropic systems with the same parameter ranges. If the iteration count to reach 1e-6 grows at the same rate as the paper's own Table 4 (27 to 977 iterations), the claimed transfer across grid sizes fails, and the multilevel results cannot be credited to the learned weights.","supporting_citations":[{"cited_title":"Grimmer, Provably faster gradient descent via long steps, SIAM Journal on Optimization 34 (3) (2024) 2588–2608","cited_arxiv_id":null,"evidence_quote":"Supplies the theoretical result that long-step (m-step) Richardson iteration converges faster, motivating the Richardson(m) formulation."},{"cited_title":"Adams, M","cited_arxiv_id":null,"evidence_quote":"Supplies the empirical alpha=1/30 choice for the Chebyshev semi-iterative method used as the single-level baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Fourier neural solver (FNS) whose smoothing and neural-iteration structure is extended by replacing the smoother with Richardson(m)-NS."},{"cited_title":"A Hybrid Iterative Neural Solver Based on Spectral Analysis for Parametric PDEs","cited_arxiv_id":"2408.08540","evidence_quote":"Provides the CNN architecture for Meta-T and the convergence framework used in the FNS construction."}],"review_version":1}