{"id":"625fa16a-3c08-4c40-a7f8-67c48e0840d5","arxiv_id":"2504.14930","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Gaussian process regression trained on small linear systems predicts near-optimal AMG strong-threshold parameters that match grid-search iteration counts at reduced cost for several PDEs.","lead":"This paper trains a Gaussian process regression model on small problems to predict the best algebraic multigrid tuning parameter for larger problems, and tests it on Poisson, diffusion, and Helmholtz equations. A reader might care because it is a cheap autotuning alternative to expensive parameter sweeps for a widely used linear solver.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GP extrapolation from n≤600 to n=1024 is untested, and the reported iteration-count matches likely reflect AMG's flat theta-performance plateau rather than accurate theta prediction.","rationale":"The reader's weakest_assumption identifies the same load-bearing point: theta*(n) must be smooth enough for a GP to extrapolate from n up to 600 to n=1024. I agree with that assessment, and add that the paper's own tables make the non-smoothness visible. The central demonstration is a single test point per PDE, with no error bars; the diffusion experiment is even less controlled because the coefficient matrix depends on random T and seed, so theta* is not a function of n alone. The kernel-combination claim is also undercut by Table 4, where the Gaussian+Laplacian combination has higher MAE, BIC, and LOO-SPE than the single Gaussian kernel, so the stated superiority is not consistently supported. The abstract's promise of 'theoretical support' is not fulfilled by any theorem or analysis in the paper. These issues are all fixable by additional experiments, so a conditional verdict is appropriate. The use of standard packages (IFEM, hypre) and a concrete grid-search protocol is a positive feature; the concern is not about fabrication but about whether the evaluation design can support the claimed generalization. My concern does not move the reader's verdict but specifies the decisive test that would settle it.","tokens_in":14915,"tokens_out":6820,"duration_ms":61604,"concrete_test":"Run the AMG theta sweep at n=700, 800, 900, and 1024 for the constant-coefficient Poisson problem with the same settings as Section 3.1, and record both the grid-search-optimal theta and the set of theta values whose iteration count is within one iteration of the minimum. The concern is confirmed if the GP trained exactly as in Tables 2-3 has large theta error at the intermediate n values, or if the width of the near-optimal plateau at n=1024 is wider than 0.05 and contains both 0.276 and 0.368; either outcome would show that the reported iteration-count match is a plateau artifact rather than evidence of accurate extrapolation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 models theta as a zero-mean GP with stationary kernels and noise sigma=1e-4; training and retraining inputs lie in [64,600] (Tables 2 and 8), but the only reported test point is n=1024 (Tables 3 and 9). For a stationary GP, the posterior mean at an input far beyond the training range reverts to the prior mean and the posterior variance grows, so the n=1024 prediction is dominated by prior assumptions unless theta*(n) is known to be smooth and extrapolable. The paper's own Table 1 contradicts smoothness: theta* jumps from 0.269 at n=80 to 0.387 at n=96 and back to 0.324 at n=112; Table 7 is similarly non-monotonic. The evaluation metric is the AMG iteration count, not theta accuracy. Figure 1 shows that iteration count is flat over wide theta intervals, and the predicted theta differs from the grid-search optimum by 0.092 for Poisson and 0.021 for Helmholtz while still giving the same iteration count. This is the signature of a performance plateau, not of accurate prediction. Without intermediate extrapolation points, repeated trials, or baseline comparisons (default theta=0.25, nearest neighbor, constant predictor), the central claim that small-scale training predicts near-optimal parameters is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes to select the strong threshold parameter theta in algebraic multigrid (AMG) by Gaussian process regression (GPR) trained on small-scale systems. The training data are grid-search-optimal theta values for n=64,...,400 (with additional retraining points up to n=600), and the method is evaluated by predicting theta at n=1024 for the constant-coefficient Poisson equation, a variable-coefficient diffusion equation, and the Helmholtz equation. A kernel library with linear combinations is compared using MSE, RMSE, MAE, R2, BIC, correlation, MdAPE, LOO-SPE, and PICP. The paper reports that GPR-predicted theta gives iteration counts close to grid-search values at n=1024 while reducing the cost of parameter sweeping.","tokens_in":15169,"tokens_out":4981,"duration_ms":40515,"significance":"The use of GPR to avoid parameter sweeps in AMG is a practical and timely idea, and the paper makes an effort to compare several kernel combinations with multiple evaluation metrics. If the extrapolation from small n to large n were demonstrated robustly, the method would be useful. However, the current evidence is thin: only one extrapolation point per equation, no repeated trials, no error bars, and the reported iteration counts are consistent with the flat performance plateau of AMG with respect to theta rather than with accurate theta prediction. The contribution is also incremental with respect to the authors' prior kernel-learning frameworks [10,11]. The conclusions are therefore plausible but not yet substantiated.","major_comments":[{"comment":"The central claim is tested at a single extrapolation point n=1024, with no intermediate extrapolation points or repeated trials. Table 1 shows that the optimal theta*(n) is highly non-smooth (0.269 at n=80, 0.387 at n=96, 0.324 at n=112), so the GP's stationarity/smoothness assumption used in §2.2 is not satisfied; the fact that theta_pred=0.368 and theta_opt=0.276 give the same iteration count is likely a consequence of the flat iteration plateau visible in Figure 1, not of accurate parameter prediction.","section":"§3.1, Tables 1 and 3"},{"comment":"For the diffusion equation, the GPR prediction yields 489.34 iterations versus 432.67 for the grid-search optimum and 78.89 s versus 51.17 s; this is roughly 13% more iterations and 54% more time, so the statement that GPR results 'closely approximate the optimal parameters' efficiency' is not supported by the numbers in the table. The table also omits the predicted and optimal theta values, the number of test matrices, and any standard deviation, so the comparison cannot be assessed.","section":"§3.2, Table 6"},{"comment":"The table lists a test set with n in [200,600] and 10 randomly selected points, but Table 9 reports results at n=1024; if the 10 points were used only for model selection or retraining, the distinction between test and retraining sets must be clarified, and if they were used as a test set, the extrapolation claim at n=1024 is not the one being tested. Either way, the internal inconsistency prevents the reader from verifying the experimental design.","section":"§3.3, Table 8"},{"comment":"The evaluation metrics are computed on an implied test set of only 7 points (PICP fractions such as 4/7, 3/7), and the reported R2 values are low or negative (e.g., R2=0.3965 for Gaussian+Laplacian in Table 4, and R2=0.0957 for Rational Quadratic+Laplacian in Table 10). These values do not support the abstract's claim that a 'comprehensive analysis ... confirms GPR's efficiency and reliability'; the number of test points and the exact n values used for the metrics need to be stated.","section":"§3.1, Tables 4–5 and §3.3, Tables 10–11"}],"minor_comments":[{"comment":"Section 3.2 states that the training set consists of 23 matrices, but n=64,80,...,400 with increment 16 gives 22 values; please reconcile the count.","section":"§3.2"},{"comment":"The text refers to 'Table reft2dks' and 'Table 3.3' instead of stable table numbers; the cross-references need to be corrected.","section":"§3.2 and §3.3"},{"comment":"Figure 2 is referenced in Section 3.2 for the block partition, but the caption of Figure 2 in the text actually describes regression curves in Section 3.1; the figure numbering and cross-references need correction.","section":"§3.2"},{"comment":"The manuscript does not state the GPR hyperparameters, the full kernel-library elements, the random seeds for the diffusion experiments, or the number of Monte Carlo runs; without these, the numerical results are not reproducible.","section":"§3"},{"comment":"The abstract and Section 5 claim 'theoretical support for AMG's practical application,' but the paper provides no theoretical analysis; please rephrase to match the actual contribution.","section":"Abstract and §5"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline paper. The idea is reasonable and the authors have a related prior framework, but the experimental evidence for the headline extrapolation claim is thin and partly inconsistent with the tables. I would like to see additional experiments (multiple test sizes, repeated random instances, baselines such as default theta=0.25 and nearest-neighbor prediction) before accepting. I also suggest the editors check the relationship to the authors' previous work [10,11] for novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the take: the paper applies the authors' existing GPR-plus-kernel-learning machinery to AMG strong-threshold prediction for three PDEs, and the n=1024 iteration counts are about what you'd expect if the method works. That's the true core. The diffusion experiment is the most interesting bit: GPR-predicted theta gives 489 iterations vs 935 for the default theta=0.25, so there is real practical signal. The constant-Poisson and Helmholtz cases show smaller gaps, consistent with a fairly flat performance plateau.\n\nWhat's actually new is the application-level evaluation: GPR-based theta prediction for these specific equations, plus the kernel-combination comparison. The experiments use IFEM and hypre, so they're reproducible in principle. That's good.\n\nThe soft spots are in proportion. The evaluation is thin: one test size per equation, no error bars or repeated trials, no baselines like nearest neighbor or a constant predictor. The kernel-combination superiority claim is not supported by Table 4: the plain Gaussian kernel has a lower BIC and much lower LOO-SPE than Gaussian+Laplacian, so the text's 'outperforms' is selective reporting. On the GP side, the stress-test concern lands. Table 1 shows theta*(n) jumping from 0.269 to 0.387 and back at consecutive n values—that is not a smooth function. With stationary kernels and a zero mean, a GP extrapolating from n≤600 to n=1024 is mostly guessing, and the fact that the predicted theta differs from the swept optimum by 0.09 yet still yields the same iteration count is a plateau signature, not evidence of accurate prediction. The 'theoretical support' sentence in the abstract is also unsupported; there is no theory in the paper. And the abstract promises a variable-coefficient Poisson experiment that I cannot find—Section 3.2 is a diffusion equation with random coefficients, not labeled as Poisson.\n\nOne minor thing: Definition 2.1 forgets the theta factor in the strong-dependence set; as written, the set is always the max column, which can't be right.\n\nVerdict: the paper deserves a serious referee, because the question is relevant and the experiments are reproducible enough to check. But it needs major revision: interpolation tests within the training range, baselines, error bars, and a direct comparison of theta accuracy against iteration-count plateau. I'd bring it to reading group as a cautionary example of GP extrapolation overreach, but I wouldn't cite it in its current form.","headline":"A plausible application of the authors' GPR pipeline to AMG strong-threshold tuning, but the evidence is too thin and the GP extrapolation to n=1024 likely rides on AMG's flat performance plateau rather than accurate theta prediction.","tokens_in":15732,"tokens_out":4022,"would_cite":false,"duration_ms":34847,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["52B10","65D18","68U05","68U07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Gaussian process models trained on small grids predict near-optimal AMG parameters for larger grids.","keywords":["algebraic multigrid","Gaussian process regression","strong threshold parameter","kernel learning","parameter prediction","partial differential equations","Helmholtz equation","Poisson equation"],"falsifier":"For a held-out family of diffusion or Helmholtz matrices at $n=1024$, sweep $\\theta$ finely near the GPR prediction, record the true optimal iteration count, and compare: if the GPR-chosen $\\theta$ is beaten by more than a few percent on many instances, or if its posterior interval frequently misses the swept optimum, then the extrapolation claim is falsified.","tokens_in":14659,"feed_emoji":"🧮","tokens_out":7533,"duration_ms":63056,"temperature":0.7,"pith_summary":"This paper tries to establish that Gaussian process regression (GPR), trained only on small algebraic multigrid (AMG) solves, can replace the expensive grid search over AMG's strong threshold parameter. The target is the threshold $\\theta$ that controls which variables are treated as strongly connected and therefore how the coarse grid is built. Across Poisson, diffusion, and Helmholtz problems, the paper reports that GPR-predicted $\\theta$ at $n=1024$ reproduces grid-search iteration counts while using far less computation, and that a default $\\theta$ often performs worse or fails to converge. If true, this gives AMG users a cheap, data-driven way to set a parameter that is normally chosen by trial and error.","feed_headline":"GP picks near-optimal AMG parameters after small-grid training","feed_subtitle":"Grid-search iteration counts are matched at n=1024 while parameter sweeps are skipped.","key_machinery":"The central object is a Gaussian process $f(n)$ with a kernel formed as a linear combination of a library of basic kernels $k(x,x') = \\sum_{\\xi=1}^N c_\\xi k_\\xi(x,x')$, with coefficients fit by maximizing the log marginal likelihood. The GP posterior mean $\\mu_* = K(n_*,n)(K(n,n)+\\sigma^2 I)^{-1}(\\theta - \\mu)$ converts observed optimal thresholds at small sizes into a predicted threshold at a new size. The target parameter is the AMG strong threshold $\\theta$, defined by the condition $|a_{ij}| \\geq \\theta \\max_{k \\neq i} |a_{ik}|$, which controls C/F splitting and hence the whole multigrid hierarchy. The machinery's job is to make $\\theta^*(n)$ extrapolable without running AMG at the target size.","core_discovery":"The central claim is that the mapping from problem size $n$ to grid-search-optimal threshold $\\theta^*(n)$ can be learned by a Gaussian process from training data on $n \\in [64,400]$ (plus targeted retraining points in $[200,600]$) and extrapolated to $n=1024$. At $n=1024$, the predicted $\\theta$ yields the same iteration count as the grid-search optimum for the constant-coefficient Poisson equation (17 iterations at $\\theta=0.368$ versus 17 at $\\theta=0.276$), close counts for the Helmholtz equation (27 at $\\theta=0.344$ versus 25 at $\\theta=0.323$, while the default $\\theta=0.25$ fails to converge), and a large improvement over the default for diffusion problems (489 average iterations versus 936, at 79 seconds versus 179). The paper interprets these matches as evidence that GPR-based parameter prediction is a practical substitute for parameter sweeping and that combined kernel functions improve reliability.","pith_inferences":["A sharper test of the method would report the width of the iteration plateau around the predicted $\\theta$; if AMG's iteration count barely changes over a wide $\\theta$ range, the reported $n=1024$ matches may mostly reflect robustness rather than prediction skill.","The same small-to-large GP transfer could be tried on other AMG controls (smoother sweeps, coarsening aggressiveness), where the smoothness of the optimal parameter as a function of size may differ.","Because the paper retrains separately for each PDE family, the method as presented does not yet show cross-equation transfer; testing one GP on an unseen equation type would be a natural extension.","Predicting on $n=1024$ from training on $n \\leq 600$ relies on extrapolation; a Bayesian treatment that also reports the posterior variance at the target size could flag cases where the prediction should not be trusted."],"forward_implications":["AMG setup for a new, larger problem can start from a GP prediction instead of a sweep over the threshold parameter, saving the dominant tuning cost.","Combined kernel functions, especially Gaussian+Laplacian and Rational Quadratic+Laplacian, can reach prediction-interval coverage at or above the single Gaussian kernel, making the reliability claim stronger than the accuracy claim alone.","The multi-metric evaluation template (R2, correlation, PICP, MSE/RMSE/MAE, MdAPE, BIC, LOO-SPE) gives a reusable way to compare parameter predictors for iterative solvers.","For repeated solves at many sizes, a trained GP can be reused at each new matrix size without recomputing the sweep, since the model maps size to parameter directly."],"supporting_citations":[{"why":"Supplies the finite-element discretization used to build the test linear systems.","marker":"[3]"},{"why":"Provides the Gaussian process posterior formulas and kernel-selection background underlying the predictor.","marker":"[4]"},{"why":"Provides the AMG solver platform on which all iteration counts and timings are measured.","marker":"[5]"},{"why":"Supplies the GPR parameter-prediction strategy that this paper adapts to AMG threshold selection.","marker":"[10]"},{"why":"Supplies the kernel-library and linear-combination construction used to choose the GPR kernel.","marker":"[11]"},{"why":"Defines the classical AMG setup phase and the strong-threshold splitting rule that theta controls.","marker":"[13]"},{"why":"Supplies the leave-one-out squared prediction error criterion used to evaluate the GPR models.","marker":"[19]"},{"why":"Cited for the GPR methodology used to train the predictor.","marker":"[24]"},{"why":"Supplies the diffusion equation test problem and the strong-threshold auto-tuning context.","marker":"[34]"}],"fun_headline_variants":["GPR predicts AMG parameters matching grid search at n=1024","Kernel learning speeds AMG parameter tuning without sweeps","Small-dataset GP sets AMG thresholds to grid-search parity","Gaussian process finds AMG parameters, cuts sweep time","Learned kernels pick AMG thresholds as well as grid search"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The grid-search-optimal threshold, viewed as a function of problem size, changes smoothly enough that a model trained on sizes up to 400 can extrapolate it to size 1024.","fun_headline_variants_meta":{"raw":{"variants":["GPR predicts AMG parameters matching grid search at n=1024","Kernel learning speeds AMG parameter tuning without sweeps","Small-dataset GP sets AMG thresholds to grid-search parity","Gaussian process finds AMG parameters, cuts sweep time","Learned kernels pick AMG thresholds as well as grid search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1342,"prompt_tokens":1013,"completion_tokens":329,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":243}},"tokens_in":629,"tokens_out":329,"duration_ms":3062,"temperature":1.0,"reasoning_tokens":243,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:37:11.890113+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a held-out family of diffusion or Helmholtz matrices at $n=1024$, sweep $\\theta$ finely near the GPR prediction, record the true optimal iteration count, and compare: if the GPR-chosen $\\theta$ is beaten by more than a few percent on many instances, or if its posterior interval frequently misses the swept optimum, then the extrapolation claim is falsified.","supporting_citations":[{"cited_title":"A tutorial on Gaussian process regression: Modelling, exploring, and exploiting functions","cited_arxiv_id":null,"evidence_quote":"Cited for the GPR methodology used to train the predictor."},{"cited_title":"iFEM: an innovative finite element methods package in MATLAB","cited_arxiv_id":null,"evidence_quote":"Supplies the finite-element discretization used to build the test linear systems."},{"cited_title":"Gaussian processes for machine learning","cited_arxiv_id":null,"evidence_quote":"Provides the Gaussian process posterior formulas and kernel-selection background underlying the predictor."},{"cited_title":"hypre: A library of high performance preconditioners","cited_arxiv_id":null,"evidence_quote":"Provides the AMG solver platform on which all iteration counts and timings are measured."},{"cited_title":"A general alternating-direction implicit framework with Gaussian process regression parameter prediction for large sparse linear systems","cited_arxiv_id":null,"evidence_quote":"Supplies the GPR parameter-prediction strategy that this paper adapts to AMG threshold selection."},{"cited_title":"Multitask kernel-learning parameter prediction method for solv- ing time-dependent linear systems","cited_arxiv_id":null,"evidence_quote":"Supplies the kernel-library and linear-combination construction used to choose the GPR kernel."},{"cited_title":"Algebraic multigrid","cited_arxiv_id":null,"evidence_quote":"Defines the classical AMG setup phase and the strong-threshold splitting rule that theta controls."},{"cited_title":"Parameter selection in Gaussian process interpolation: an empirical study of selection criteria","cited_arxiv_id":null,"evidence_quote":"Supplies the leave-one-out squared prediction error criterion used to evaluate the GPR models."},{"cited_title":"AutoAMG (θ): An Auto-tuned AMG Method Based on Deep Learning for Strong Threshold","cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion equation test problem and the strong-threshold auto-tuning context."}],"review_version":1}