{"id":"513547ed-d967-4d10-b4b8-45f9c1903a5b","arxiv_id":"2607.13382","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Residual-space Christoffel sampling with whitening makes random-feature PDE collocation systems well-conditioned while keeping sample complexity linear in the retained residual dimension.","lead":"For linear PDEs solved with random feature collocation, this paper builds the collocation sampling density and a coefficient whitening map from the operator-applied features themselves, improving system conditioning. It offers a principled fix for a known instability in strong-form random feature methods, with theorems and experiments in 1D-3D.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Rank truncation via gap rule can discard approximation directions; the method may be well-conditioned yet inaccurate regardless of collocation size.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the retention rule (5) and the whitening map T_r restrict the coefficient solve to a subspace that may not contain directions essential for accuracy, and the method has no safeguard against this. This is the most consequential issue because it threatens the method's usefulness as a PDE solver, not merely its conditioning. The paper's central theoretical claim—Gram concentration on the retained space—is not invalidated, but the practical claim of 'stable strong-form random feature collocation' is weakened if stability is achieved at the cost of accuracy. The proposed concrete test directly measures the approximation power of the restricted trial space, isolating the truncation effect from sampling and conditioning. If the restricted best-approximation error is large, then no sampling strategy can close the accuracy gap; if it is small, the concern is empirically refuted for that class of problems. Since this concern is already reflected in the reader's CONDITIONAL verdict, no change in verdict is recommended.","tokens_in":23964,"tokens_out":9880,"duration_ms":95044,"concrete_test":"For a representative benchmark (e.g., high-frequency Poisson or Helmholtz k=16), compute the L2 best-approximation error of the exact solution in the restricted trial space {Σ_m (T_r d)_m φ_m : d ∈ R^r} and in the full feature span span{φ_m}, using a dense quadrature to evaluate the projections. If the restricted best-approximation error is orders of magnitude larger than the full-space error, the rank truncation is a fundamental accuracy barrier for the method; if the errors are comparable, the concern is mitigated for that benchmark.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The whitening map P_coef = T_r (Eq. 6) restricts the coefficient solve to range(T_r), an r-dimensional subspace selected by the heuristic gap rule (5), not by any accuracy criterion. Theorem 1 only guarantees that the whitened interior Gram concentrates about I_r on this retained subspace; it does not establish that range(T_r) contains the directions needed to approximate the PDE solution. Since the solve is confined to range(T_r), any component of the best approximation in the full feature span lying outside that subspace is unrecoverable, no matter how many collocation points are added. The paper's own ablations corroborate this risk: Table 4 shows pruning feature columns to k = r degrades the relative error from 4.1e-5 to 1.6e-2, and the L-shaped enrichment control demonstrates that missing solution directions cannot be repaired by point selection alone. Thus the method can produce a numerically full-rank but inaccurate system, undermining its practical value as a PDE solver even though the Gram-concentration claim remains valid on the retained space.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes operator-aware row selection and coefficient whitening for random feature collocation of linear PDEs. For a fixed realized trial space, it defines the operator-residual feature vector ψ=(Lφ_1,...,Lφ_M)^T, the reference interior Gram G=E[ψψ^T], a retained rank r, and a whitening map P_coef=T_r. The randomized method samples collocation points from the residual-Christoffel density ∥T_r^Tψ∥²/r relative to a reference measure and weights the rows inversely to that density, so that the empirical whitened interior Gram is unbiased for I_r. A deterministic greedy method selects rows by the largest regularized leverage score and is shown, locally, to maximize the log-determinant increment. Theorem 1 gives a Chernoff-type concentration bound with sample complexity N≥(3/ε²)r log(2r/δ); Corollary 1 bounds the resulting condition number. Theorems 2–4 give an operator limit, stretched-exponential spectral decay for uniformly analytic residual kernels, and a ridge-leverage concentration result. Numerical experiments span Poisson, Helmholtz, convection–diffusion, variable-coefficient, and elasticity problems; they report substantially reduced condition numbers, faster LSQR convergence, and good small-sample accuracy for the greedy rule. The paper is careful to distinguish conditioning claims from accuracy claims and to report ablations (boundary weighting, whitening, feature pruning, enrichment) that expose the limitations of rank truncation.","tokens_in":24258,"tokens_out":8747,"duration_ms":103674,"significance":"The residual-Christoffel construction is a principled and novel extension of Christoffel sampling to differential residuals, and the population-level concentration theory is clean and appears correct. The proofs use standard matrix concentration and are transparent. The explicit sample-complexity bound and the spectral analysis of the residual kernel are valuable contributions, and the numerical comparisons are extensive. The main theoretical guarantee, however, is established for the population construction (or for a finite candidate set that is also used for whitening), while the implemented Algorithm 1 uses a separate reference quadrature and a separately drawn candidate set; the resulting deterministic mismatch ε_cand is isolated in Remark 2 but never controlled or reported. The rank-truncation step is heuristic and the method can be well-conditioned yet inaccurate, as the paper's own Table 4 and enrichment control demonstrate. These issues are fixable within the manuscript's scope.","major_comments":[{"comment":"Theorem 1 guarantees concentration about I_r only for the population sampler or when the finite candidate set is also the whitening measure. In Algorithm 1, the reference Gram G_Q is formed on Q while candidates are drawn independently from μ, so the conditional target is H_P, not I_r, and the theorem does not control ∥Ĝ⊥−I_r∥ unless ε_cand is controlled. No bound on ε_cand is given and the experiments do not report it. The abstract/conclusion phrase 'sample complexity linear in retained dimension' therefore applies to a construction that differs from the numerical algorithm. Please either prove a finite-candidate concentration bound with explicit assumptions on P and Q, or restrict the claims and add a numerical report of ε_cand.","section":"Section 3.1, Remark 2, Theorem 1"},{"comment":"The coefficient map P_coef=T_r and the gap-based rank rule (5) restrict the coefficient solve to range(T_r). Theorems 1 and 4 say nothing about whether this retained subspace contains the directions needed for solution accuracy. The paper separates conditioning from accuracy, but the failure mode is real: Table 4 shows RRQR pruning to k=r degrades the relative L2 error from 4.1e-5 to 1.6e-2, and the L-shaped enrichment control shows that missing solution directions cannot be repaired by point selection alone. This limitation should be stated explicitly near the statement of Corollary 1 and in the conclusion, and the paper should discuss what accuracy guarantees, if any, can be made for the gap-rule choice of r.","section":"Section 5.5, Table 4, Section 2.2"},{"comment":"The deterministic leverage-greedy construction is a headline contribution and is reported as the most accurate method at small sample sizes, but no convergence or conditioning theorem is given for it. Proposition 1 only shows that each greedy step maximizes the one-step log-determinant increment; it does not imply a bound on the final assembled Gram or on the condition number after N steps. Given the prominence of this algorithm in the experiments, the paper should either develop a stability guarantee or explicitly state that greedy selection is heuristic and supported only by numerical evidence.","section":"Section 3.2, Algorithm 2"}],"minor_comments":[{"comment":"The sizes and construction details of the reference quadrature Q and the candidate pool P are not reported in the general setup. Some figure captions give P for one experiment, but a reproducibility section should list Q, P, and how candidates are sampled for every benchmark.","section":"Section 5.1"},{"comment":"The connection between ϑ and β is stated conditionally ('If the augmented reference Gram is chosen to match... then ϑ=β'). This is confusing because β is the boundary weight in (9), while G_aug^ϑ is a reference object. Please clarify whether ϑ=β is used in the experiments or only in the remark.","section":"Section 2.3, Eq. (13)"},{"comment":"Panel (d) is labeled only 'raw whitened'; the figure caption should state which raw system (uniform, Christoffel, or both) is compared, and the conditioning values should be readable from the panel.","section":"Figure 8"},{"comment":"The analytic-kernel hypothesis is verified for sine features and for tanh under a pole-separation condition, but the experiments do not check whether the chosen a_rf and parameter support actually satisfy the condition. A short note on this would avoid overclaiming the applicability of the stretched-exponential decay result.","section":"Theorem 3"},{"comment":"There are minor typographical issues, e.g., 'Schrödinger' with unusual spacing in the introduction, and the notation M sys/s is introduced only informally. These do not affect the substance.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The population-level theory is sound and the experiments are valuable, but the central sample-complexity claim is not matched to the implemented finite-candidate algorithm. The authors are transparent about this in Remark 2, yet the abstract and conclusion overstate the reach of the theorem. I recommend major revision: either adapt the algorithm so the theorem applies (e.g., use the same quadrature as candidate set) or extend the analysis to bound ε_cand, and also make the rank-truncation accuracy limitation explicit. With these changes the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the takeaway: this is a solid, useful paper for the RFM collocation community. The residual-Christoffel idea is genuinely new — using the operator-applied features to define both the sampling density and the whitening map — and the main concentration theorem is correct for the population construction. The experiments convincingly show that the resulting systems are better conditioned and need far fewer LSQR iterations. If you work on randomized collocation, it's worth a careful read.\n\nThe coupling of the residual Gram to the sampling measure and the whitening map is not in CS4ML or the earlier leverage-sampling literature. The paper is also honest: Remark 2 separates the finite-candidate mismatch from the concentration bound, and the ablations carefully isolate point selection from whitening and from feature pruning. The deterministic greedy variant looks like a nice practical addition, especially for small scalar sample sizes.\n\nThe biggest soft spot is exactly what the stress-test flags. The whitening map restricts the coefficient solve to range(T_r), and r is chosen by the gap heuristic, not by any accuracy criterion. Theorem 1 only guarantees concentration on that retained subspace. If the true solution has significant components outside it, extra collocation points won't help. The paper's own Table 4 shows RRQR pruning to k=r wrecks accuracy (4.1e-5 to 1.6e-2), and the L-shaped enrichment control shows a missing singularity direction cannot be rescued by point selection. That doesn't kill the central conditioning claim — it just means the method stabilizes the represented directions, not the ones the trial space failed to represent. The finite-candidate issue is real but minor; Figure 3 shows the 3D case doesn't reach unit relative deviation in the tested range, so the practical constants are dimension-dependent. Data availability is still incomplete, though an initial code release exists.\n\nWho's this for? People working on random feature methods, Christoffel sampling, or preconditioned iterative solvers for collocation. It's not a general-purpose PDE breakthrough, but it's a solid within-subfield contribution with a clear argument.\n\nIt deserves a serious referee. The main revisions I'd ask for: bound the approximation error of the retained subspace or offer a practical accuracy-aware criterion for r, and deposit the complete experiment scripts. I'd engage with it.","headline":"A solid, useful conditioning fix for random-feature collocation, with a real but contained limitation around rank truncation.","tokens_in":24685,"tokens_out":1738,"would_cite":true,"duration_ms":18081,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N35","65F35","62K05","65D15","68W20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Sampling collocation points from the residual Christoffel density of the operator-applied features provably stabilizes random feature PDE solves.","keywords":["random feature method","operator-residual sampling","Christoffel sampling","PDE collocation","residual Gram","whitening","effective dimension","least-squares conditioning"],"falsifier":"Take a manufactured solution on a domain with a known singularity or a feature map deliberately band-limited so that the top-r residual subspace misses the solution's main frequency, apply the method with the heuristic rank rule, and check the relative L2 error: if the error stays large while the condition number of the whitened system stays small, the retained-subspace premise is violated. The paper's RRQR pruning to k = r gives a concrete instance where accuracy collapses despite good conditioning.","tokens_in":23833,"feed_emoji":"🎯","tokens_out":5575,"duration_ms":51444,"temperature":0.7,"pith_summary":"The paper claims that the instability of random feature collocation for linear PDEs can be traced to a mismatch between the sampling measure and the geometry of the operator-applied trial space, and that this mismatch is curable by construction. It builds the interior residual Gram from the operator-applied features, retains its leading eigenspace, and uses the induced Christoffel function as a sampling density with inverse-density weights, together with a whitening map that makes the population Gram the identity on the retained subspace. Conditional on the realized trial space, the sampled whitened Gram concentrates around the identity with O(r log(r/δ)/ε²) points, giving a condition number bound independent of feature count and population anisotropy. For analytic residual kernels, the operator spectrum decays stretched-exponentially, giving a polylogarithmic ridge effective dimension. The consequence would be a principled design principle for stable strong-form collocation that transfers across geometries and operator classes.","feed_headline":"Residual-space sampling stabilizes random feature PDE solves","feed_subtitle":"Operator-applied features set both the collocation density and a whitening map, making the sampled Gram concentrate near the identity.","key_machinery":"The central objects are the residual-Christoffel function K_M^L(x) = ψ(x)^T G_r^† ψ(x) and the whitening map T_r = [v_1/√λ_1, …, v_r/√λ_r], where G is the population Gram of operator-applied features and G_r^† is its rank-r truncated inverse. K measures how a point's residual row is represented in the retained eigenspace; its expectation is r, so K/r is a probability density. Sampling from K/r with weights r/(N K) makes each weighted whitened row a PSD summand of trace r/N, which is the key to the matrix-Chernoff concentration bound (N ≥ 3r ln(2r/δ)/ε²) and the resulting condition-number bound κ ≤ √((1+ε)/(1−ε)). A deterministic variant maximizes successive regularized log-determinant increm","core_discovery":"The central discovery is that the operator-residual Gram G = E[ψψ^T] defines both the correct sampling measure and the correct solver coordinates. The residual-Christoffel function K(x) = ψ(x)^T G_r^† ψ(x) quantifies the leverage of a point in the retained residual space; sampling points proportional to K/r and weighting them by 1/K makes every sampled row an unbiased, bounded-rank-one estimator of the identity on the retained subspace, so the empirical whitened Gram concentrates about I_r. The whitening map T_r simultaneously restricts the coefficient solve to the retained residual coordinates. Thus stability is achieved not by covering the domain but by resolving the operator-induced geome","pith_inferences":["A direct corollary the paper leaves implicit: the same residual Gram and Christoffel weights can be reused for multiple right-hand sides sharing the operator, domain, and trial space, so the leverage-construction cost is amortized across a parameter study.","The finite-candidate mismatch ε_cand = ||H_P − I_r|| is the practical bottleneck between the population theorem and the discrete sampler; one testable extension is to build the candidate Gram H_P itself with the same Christoffel-weighted quadrature, which would close the gap between the concentration bound and the implemented algorithm.","The theory is conditional on a fixed trial space; combining residual-Christoffel sampling with trial-space enrichment (e.g., appending singularity functions) addresses the demonstrated failure mode where the retained subspace lacks required directions—the paper's own L-shape experiment shows enrichment, not point selection, fixes the error.","The sketched range-finder construction for large M trades retained residual subspace size against construction cost; a natural extension is a sketch size adaptive to a user-specified accuracy target, blending the dense-eigendecomposition and sketch regimes."],"forward_implications":["With O(r log(r/δ)/ε²) collocation points, the whitened interior Gram lies within ε of the identity, so the interior residual block has condition number at most √((1+ε)/(1−ε)), independent of feature count M and of the population anisotropy λ₁/λ_r.","For uniformly analytic residual kernels, the random-feature operator eigenvalues decay as τ_j ≤ C e^{−c j^{1/d}}, so the ridge effective dimension grows only polylogarithmically with 1/ℓ; the observed retained dimension r stays small relative to M (r/M falls from 0.44 to 0.05 as M grows from 400 to 6400).","Residual-space sampling plus whitening removes numerical rank deficiency and reduces iterative-solver counts from caps to tens or hundreds across Poisson, Helmholtz, variable-coefficient, convection–diffusion, and elasticity benchmarks, on square, L-shaped, and annular domains in one to three spatial dimensions.","The deterministic leverage-greedy rule attains the lowest median errors at the smallest sample-to-rank ratios in scalar tests, improving high-frequency Poisson error by more than an order of magnitude relative to uniform and residual-adaptive sampling.","Boundary-aware whitening (including a boundary Gram block) reduces the full interior-plus-boundary condition number by nearly three orders of magnitude without changing error, while the boundary weight β trades accuracy against conditioning as expected."],"fun_headline_variants":["Sampling by residual Christoffel stabilizes random feature PDE solves","Operator-aware sampling and whitening tame PDE collocation systems","Residual-space geometry fixes unstable random feature collocation","Whitening makes operator Gram concentrate near identity","PDE collocation: sample in residual space, whiten, solve stably"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that the retained rank-r subspace of the operator-residual Gram contains the trial-space directions the true solution actually needs; if the solution's residual has significant components outside that subspace (or the rank heuristic truncates them), the method produces a well-conditioned but inaccurate system, as the paper's own feature-pruning and L-shape ablation show.","fun_headline_variants_meta":{"raw":{"variants":["Sampling by residual Christoffel stabilizes random feature PDE solves","Operator-aware sampling and whitening tame PDE collocation systems","Residual-space geometry fixes unstable random feature collocation","Whitening makes operator Gram concentrate near identity","PDE collocation: sample in residual space, whiten, solve stably"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1138,"prompt_tokens":747,"completion_tokens":391,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":491,"tokens_out":391,"duration_ms":14757,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T05:19:51.703579+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a manufactured solution on a domain with a known singularity or a feature map deliberately band-limited so that the top-r residual subspace misses the solution's main frequency, apply the method with the heuristic rank rule, and check the relative L2 error: if the error stays large while the condition number of the whitened system stays small, the retained-subspace premise is violated. The paper's RRQR pruning to k = r gives a concrete instance where accuracy collapses despite good conditioning.","supporting_citations":[],"review_version":1}