{"id":"6f9ec38d-3de9-425d-9750-c70a9017d1c2","arxiv_id":"2506.17626","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A rank-revealing QR based feature filter plus a sparsity-preserving right preconditioner cuts the condition number of domain-decomposed random feature PDE systems by up to 11 orders of magnitude and speeds LSQR convergence 10 to 1000 times.","lead":"This paper adds a local filtering step that removes redundant random neural-network basis functions before a PDE least-squares system is solved, then builds a preconditioner from the remaining factors that makes the solver much better conditioned and faster. If it holds up, it removes a key practical bottleneck for physics-informed random feature PDE solvers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2 is misindexed: the printed A_j is not the adjacent-subdomain overlap product, the stated bound is false as written, and a 2-block scalar example violates it.","rationale":"The reader's verdict is CONDITIONAL, and I am not moving it: the discrepancy is a repair, not a refutation of the whole method. The numerical evidence is extensive, and Proposition 3.1 and the preconditioning construction appear sound. But the central claim as summarized by the reader, that Theorem 3.2 gives a condition-number bound whenever 2alpha<1, is not supported as printed: the alpha used in the theorem is not the actual off-diagonal coupling of Q^T Q for the very block structure the theorem defines. A referee can likely fix this in one line, but until then the deterministic analysis, the discussion of RRQR reducing alpha, and the numerical validation in Figure A.20 that ties kappa to the bound all concern a different quantity. I partially agree with the reader's weakest_assumption: the strip-geometry limitation is real, but the more urgent issue is that even within that strip geometry the theorem's defining quantity is wrong. The other reader issues (no released code, AS baseline without PCA filtering) remain valid but are secondary to the correctness of the main theorem. Verdict stays CONDITIONAL, with the misindexing correction and its re-verification as an explicit acceptance condition.","tokens_in":31275,"tokens_out":19513,"duration_ms":187418,"concrete_test":"Compute the J=2, K=1 example above (b=0.99) from Eq. (8) and compare kappa(Q) with the value predicted by Theorem 3.2 using the printed A_1 = Q21^T Q11. Repeating the same check with the corrected A_1 = Q21^T Q22 should restore the bound. This single arithmetic check settles whether the theorem as written is false and whether the intended alpha is the adjacent overlap block.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical result, Theorem 3.2, is internally inconsistent as written. For the block matrix in Eq. (8), column block j has nonzeros in row blocks j and j+1, namely Q_jj and Q_{j+1,j}, so the (j,j+1) block of Q^T Q is Q_{j+1,j}^T Q_{j+1,j+1}, the product of the two adjacent subdomain blocks sharing row block j+1. The paper instead defines A_j := Q_{j+1,j}^T Q_{j,j} and states E_{j,j+1}=A_j; that is a product of two blocks of the same subdomain, not the overlap coupling between adjacent subdomains. This is substantive, not a notational quibble: take J=2, K=1, with Q11=sqrt(1-b^2), Q21=b, Q22=1. Then Q^T Q = [[1,b],[b,1]], so kappa(Q)=sqrt((1+b)/(1-b)), while the paper's alpha is |b sqrt(1-b^2)|. For b=0.99, alpha ~ 0.0995, so 2alpha<1 and the theorem predicts kappa(Q) <= 1.22, but the actual kappa(Q) ~ 14.1. The intended statement is almost certainly A_j = Q_{j+1,j}^T Q_{j+1,j+1}, and with that correction the proof's ||E||<=2alpha estimate does go through. But as published, the deterministic bound, the interpretation of alpha as overlap strength, and the claim that RRQR filtering reduces alpha by lowering cross-interface correlation are all attached to the wrong matrix. This occurs inside the simplified 1D strip setting the authors do analyze, so it is more basic than the reader's geometry gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a local sigma-RRQR filtering and right-preconditioning strategy for domain-decomposed random-feature least-squares systems arising in ELM-FBPINNs. It defines a filtered global matrix bM and a sparse right preconditioner S^{-1} assembled from local triangular factors, shows that the preconditioned matrix is a sum of local orthogonal factors (Proposition 3.1), and derives a deterministic condition-number bound for overlapping strip-like decompositions (Theorem 3.2), with probabilistic refinements in Appendix A. Numerical experiments on a 1D harmonic oscillator, a 2D multi-scale Laplacian, and a (2+1)D wave equation report reductions in condition number by up to eleven orders of magnitude, LSQR speedups of 10 to 1000 times, and improved accuracy versus unpreconditioned, CG, and additive-Schwarz baselines.","tokens_in":31526,"tokens_out":8639,"duration_ms":83600,"significance":"If the results hold, the paper would remove a key obstacle for domain-decomposed random-feature PDE solvers: overlap- and redundancy-induced ill-conditioning. The experimental methodology is a strength: five seeds with reported ranges, ablations over sigma, activation, and depth, strong and weak scaling studies, and comparisons against exact solutions and a high-fidelity finite-difference solver. The proposed preconditioner is attractive because it is sparse, avoids forming normal equations, and is assembled only from local RRQR factors. However, the central deterministic bound is misstated as written and covers only a simplified 1D strip geometry, so the theoretical support for the headline 2D and (2+1)D claims is currently incomplete.","major_comments":[{"comment":"The off-diagonal block A_j is defined as Q_{j+1,j}^T Q_{j,j}, but the (j,j+1) block of Q^T Q is Q_{j+1,j}^T Q_{j+1,j+1}. With the printed definition, alpha is not the adjacent-subdomain coupling, and the theorem's conclusion is false: for J=2, K=1, Q11=sqrt(1-b^2), Q21=b, Q22=1, and b=0.99, the printed alpha is about 0.0995, so 2alpha<1 holds, but the actual kappa(Q)=sqrt((1+b)/(1-b)) is about 14.1, far above the claimed bound. The proof goes through with A_j:=Q_{j+1,j}^T Q_{j+1,j+1}; please correct the definition, the interpretation of alpha as overlap strength, and Remark 3.4, and re-examine the claim that RRQR filtering reduces this quantity.","section":"Section 3.4, Eq. (8) and Theorem 3.2"},{"comment":"The deterministic bound is derived only for consecutive 1D strips with nearest-neighbor overlap and exactly orthonormal blocks, but the main 2D and (2+1)D experiments use rectangular decompositions with corner overlaps, where each block couples to more than two neighbors and post-filter orthogonality is approximate. The paper's own discussion in Section 4.4.3 concedes that the preconditioner neglects overlapping contributions between subdomains in higher dimensions. Therefore Theorem 3.2 does not justify the condition numbers reported for the 2D and (2+1)D experiments; the abstract and introduction should state this limitation explicitly, and the paper should indicate which of the numerical claims are covered by the theorem.","section":"Section 3.4 vs Section 4"},{"comment":"The probabilistic estimate is derived under a Haar-distributed semi-orthogonal model for the filtered blocks that is asserted but not validated. Figure A.20 only shows that measured condition numbers grow with K and delta; it does not compare measured kappa(Q) against the upper bound (A.3), nor does it test whether the block model is accurate. Consequently, the label 'Numerical validation of bound' overstates the evidence. Either provide a quantitative comparison or present Figure A.20 as a qualitative trend only.","section":"Appendix A.2"},{"comment":"The theorem assumes each subdomain contributes exactly K features, but after RRQR filtering each subdomain retains a different number k_j of columns. The block structure in Eq. (8) and the definition of Q therefore need to be restated with variable block widths, and the condition-number argument must be checked for that case. The proof is likely to survive because it only needs per-block values alpha_j, but the statement as written is not the system actually being solved.","section":"Section 3.4 and Algorithm 1"}],"minor_comments":[{"comment":"The first sentence refers to 'Theorem 3.1', but the result being used is Proposition 3.1; please correct the reference.","section":"Section 3.4"},{"comment":"The text 'the FBPINN and PINN baselines are and nearly an order of magnitude slower' contains a missing word; it should read something like 'are also nearly an order of magnitude slower'.","section":"Section 4.4.1"},{"comment":"The filtered selection matrices \\hat W_j are not defined after column dropping; their dimensions should be stated explicitly (for example, \\hat W_j in R^{k_j x JK}) so that Proposition 3.1 and Algorithm 2 are unambiguous.","section":"Section 3.2 and Algorithm 2"}],"recommendation":"major_revision","confidential_remarks":"The misindexing in Theorem 3.2 appears to be a genuine but local error: the corrected definition A_j = Q_{j+1,j}^T Q_{j+1,j+1} makes the proof go through, and the rest of the paper's algebraic and experimental content is substantially sound. The paper is within the journal's scope, and the experimental study is thorough. I would be willing to review a revised version that fixes the theorem, carefully scopes the theory to the 1D strip case, and either validates or softens the Appendix A claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the algorithmic idea here is real and the experiments are careful, but Theorem 3.2 is wrong as printed. The stress-test note holds up. I checked the block structure in Equation (8), and A_j is indeed misindexed: the (j,j+1) block of Q^T Q is Q_{j+1,j}^T Q_{j+1,j+1}, not Q_{j+1,j}^T Q_{j,j}. I ran the 2-block scalar example from the note; it violates the stated bound. The theorem can probably be fixed by replacing A_j with the correct product, and then the proof's ||E|| ≤ 2α estimate goes through, but as published the deterministic bound is false.\n\nWhat is genuinely new: replacing the SVD-on-Gram-blocks and additive-Schwarz-on-normal-equations of reference [20] with local σ-RRQR filtering plus a sparse right preconditioner assembled from the filtered triangular factors, applied directly to the least-squares system. This is a sensible construction that preserves sparsity and avoids squaring the condition number. The experiments are solid: five seeds, error bars, ablations, strong and weak scaling, and an independent finite-difference ground truth for the wave equation. The reported conditioning improvements and 10–1000× solver speedups are credible.\n\nSoft spots, in proportion. The misindexed theorem is the biggest one—not merely a limitation to strip-like geometry but an error in the simplest case. That said, the numerical evidence stands independent of the theorem, and the intended correction is obvious, so I'd treat it as repairable. The additive Schwarz baseline drops the PCA filtering step from [20], which may flatter the comparison; that should be addressed. No code or data is released, so the tables are not reproducible as shipped. Appendix A's probabilistic refinement assumes a Haar-distributed model for filtered blocks, which is unvalidated. The paper itself concedes degraded convergence in higher dimensions and neglects corner overlaps in the preconditioner—these are honest limitations.\n\nWho this is for: people working on ELM-FBPINNs and random feature PDE solvers. They'll get a clear new tool and a careful empirical study. The theory needs a correction before it can be cited. I'd send it to peer review—a serious referee can sort out the indexing—but I would not cite it myself in its current form. My advice to an editor: send to review, with a strong request that the discussion of Theorem 3.2 be fixed.","headline":"Real algorithmic contribution and careful experiments, but the central condition-number theorem is misindexed and false as written—fixable, but needs fixing.","tokens_in":32192,"tokens_out":6224,"would_cite":false,"duration_ms":59691,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N55","65N35","65F10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Local rank-revealing QR filtering, combined with a sparse preconditioner assembled from the filtered factors, removes most of the ill-conditioning that slows domain-decomposed random feature PDE solvers.","keywords":["random feature methods","physics-informed machine learning","overlapping domain decomposition","rank-revealing QR factorization","preconditioning","least-squares problems","numerical solution of PDEs","ELM-FBPINN"],"falsifier":"Compute the spectral condition number $\\kappa(\\widehat{M}S^{-1})$ for a 2D rectangular decomposition with corner overlaps and compare it with the bound $\\sqrt{(1+2\\alpha)/(1-2\\alpha)}$ using the measured coupling $\\alpha = \\max_j \\|A_j\\|_2$; if the measured condition number exceeds the bound, or if the bound blows up ($2\\alpha \\ge 1$) while LSQR still converges, the strip-geometry assumptions are doing the work. A second check: measure $\\|\\widehat{Q}_j^T \\widehat{Q}_j - I\\|_2$ after filtering; if it is not small, the orthonormality premise of Theorem 3.2 is violated.","tokens_in":30936,"feed_emoji":"🧮","tokens_out":7213,"duration_ms":61464,"temperature":0.7,"pith_summary":"Domain-decomposed random feature methods solve PDEs by fitting many small overlapping neural networks through a single linear least-squares problem, but that problem is often so ill-conditioned that standard iterative solvers barely move. The paper claims that a local rank-revealing QR (RRQR) factorization, applied independently to each subdomain block, can strip out redundant random features and that the remaining factors can be assembled into a sparse right preconditioner that directly attacks the overlap-induced coupling. If right, the condition number of the preconditioned least-squares system is bounded in terms of a single overlap-coupling constant, and in practice the method cuts condition numbers by up to eleven orders of magnitude and accelerates LSQR by factors of 10 to 1000 on 1D, 2D, and (2+1)D test problems. This matters because it removes one of the main obstacles to using randomized neural networks as scalable PDE solvers.","feed_headline":"Local QR filter tames ill-conditioned random feature PDE solvers","feed_subtitle":"A block rank-revealing QR preconditioner cuts least-squares condition numbers by up to 11 orders of magnitude.","key_machinery":"The machinery is the σ-rank-revealing QR (RRQR) factorization applied locally: each subdomain block $M_j$ is permuted and factored as $M_j \\Pi_j = Q_j [R_j; T_j]$ with $R_j$ well-conditioned and the trailing block small, and the columns associated with $R_j$ are retained as the filtered block $\\widehat{M}_j = \\widehat{Q}_j \\widehat{R}_j$. These local factors are then fused into the global sparse right preconditioner $S^{-1} = \\sum_j \\widehat{W}_j^T \\widehat{R}_j^{-1} \\widehat{W}_j$, so that the preconditioned least-squares matrix becomes $Q = \\widehat{M}S^{-1} = \\sum_j V_j^T \\widehat{Q}_j \\widehat{W}_j$. The key identity is that the $\\widehat{R}_j$ factors cancel algebraically, leaving a sum of orthonormal blocks; the conditioning analysis then reduces to bounding the block-tridiagonal overlap matrix $E$ of $Q^T Q = I + E$, whose norm is bounded by $2\\alpha$. This is what converts the ill-conditioning problem into a question of how strongly neighbouring subdomains couple.","core_discovery":"The central claim is that the ill-conditioning of the ELM-FBPINN least-squares system has two separable causes — redundancy of random features within each subdomain and correlation across overlapping subdomains — and that a σ-RRQR factorization of each local block $M_j$ addresses both at once. Filtering keeps only columns whose singular values lie above a relative threshold σ, while the upper-triangular factor $\\widehat{R}_j$ of the filtered block becomes the building block of a right preconditioner $S^{-1} = \\sum_j \\widehat{W}_j^T \\widehat{R}_j^{-1} \\widehat{W}_j$. After preconditioning, the global matrix reduces to $Q = \\sum_j V_j^T \\widehat{Q}_j \\widehat{W}_j$, a sum of orthonormal blocks whose Gram matrix is the identity plus a sparse overlap-coupling matrix. For a simplified strip geometry, Theorem 3.2 bounds the condition number by $\\sqrt{(1+2\\alpha)/(1-2\\alpha)}$ whenever the maximum overlap coupling $\\alpha = \\max_j \\|A_j\\|_2$ satisfies $2\\alpha < 1$, showing that conditioning depends only on nearest-neighbour overlap strength and not on the number of subdomains or total features. The numerical experiments support the practical claim that this strategy consistently improves conditioning and solver speed while matching or exceeding the accuracy of unpreconditioned and additive-Schwarz baselines.","pith_inferences":["If the overlap-coupling constant $\\alpha$ is the true controlling quantity, then an adaptive strategy that increases filtering exactly in the overlap regions—rather than a uniform threshold per subdomain—should improve conditioning further; this is not explored in the paper.","The bound's failure mode for 2D corner overlaps suggests that including diagonal overlap interactions in the preconditioner, as the paper itself flags for future work, could restore the bound in higher dimensions; a testable extension is to add the corner blocks to $S^{-1}$ and measure whether $\\kappa(Q)$ drops back toward the predicted value.","In the random feature setting, the same local RRQR filtering could double as an interpretability tool: the retained columns are actual basis functions rather than linear combinations, so filtering also selects which features carry the solution; the paper notes the interpretability but does not pursue it."],"forward_implications":["Within the strip-decomposition regime covered by Theorem 3.2, the condition number of the preconditioned system stays bounded as the number of subdomains grows, as long as the overlap coupling stays below the 1/2 threshold; this is what makes the method scalable in the 1D strong and weak scaling tests.","Aggressive filtering (larger σ) monotonically improves conditioning but eventually degrades test accuracy by removing basis functions, so a balance between conditioning and representational capacity is needed; the paper shows the trade-off in its ablation tables.","Because the preconditioner is assembled from triangular factors and preserves block sparsity, it can be applied by cheap back-substitution without ever forming the normal equations or a dense Gram matrix, avoiding the squared condition number and memory blow-up of normal-equation preconditioning.","The method matches or beats the additive-Schwarz-preconditioned baseline in accuracy while being roughly an order of magnitude faster and less memory-hungry across the 1D, 2D, and (2+1)D test problems."],"supporting_citations":[{"why":"Defines the random feature method for PDEs that this work preconditions.","marker":"[1]"},{"why":"Introduces the ELM-FBPINN variant whose least-squares system is the target of the filtering and preconditioning.","marker":"[2]"},{"why":"Provides the additive Schwarz preconditioned baseline on the normal equations that the new method is compared against and improves upon.","marker":"[20]"},{"why":"Supplies the rank-revealing QR factorization concept used for local feature filtering.","marker":"[21]"},{"why":"Gives the strong RRQR algorithm used to compute the factorization in practice.","marker":"[23]"},{"why":"LSQR is the iterative least-squares solver whose convergence the preconditioner accelerates.","marker":"[24]"},{"why":"Provides the random matrix bound used in Appendix A to estimate overlap couplings probabilistically.","marker":"[28]"},{"why":"Multilevel domain decomposition cited for the multi-scale test problems and as the proposed extension for larger subdomain counts.","marker":"[15]"}],"fun_headline_variants":["RRQR filter cuts condition numbers by 11 orders","Local feature filtering speeds PDE solvers 1000x","Well-conditioned random features via block QR","Random feature PDE solvers tamed by RRQR preconditioner","Filtering random features yields scalable well-conditioned PDE solvers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The deterministic conditioning bound is proven only for subdomains arranged as consecutive 1D strips with nearest-neighbour overlap and exactly orthonormal filtered columns, whereas the headline 2D and (2+1)D experiments use regular rectangular decompositions with corner overlaps, where each block couples to more than two neighbours and post-filter orthogonality is only approximate.","fun_headline_variants_meta":{"raw":{"variants":["RRQR filter cuts condition numbers by 11 orders","Local feature filtering speeds PDE solvers 1000x","Well-conditioned random features via block QR","Random feature PDE solvers tamed by RRQR preconditioner","Filtering random features yields scalable well-conditioned PDE solvers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000867,"raw_usage":{"total_tokens":3866,"prompt_tokens":1161,"completion_tokens":2705,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":777,"completion_tokens_details":{"reasoning_tokens":2626}},"tokens_in":777,"tokens_out":2705,"duration_ms":20603,"temperature":1.0,"reasoning_tokens":2626,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:07:15.083686+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the spectral condition number $\\kappa(\\widehat{M}S^{-1})$ for a 2D rectangular decomposition with corner overlaps and compare it with the bound $\\sqrt{(1+2\\alpha)/(1-2\\alpha)}$ using the measured coupling $\\alpha = \\max_j \\|A_j\\|_2$; if the measured condition number exceeds the bound, or if the bound blows up ($2\\alpha \\ge 1$) while LSQR still converges, the strip-geometry assumptions are doing the work. A second check: measure $\\|\\widehat{Q}_j^T \\widehat{Q}_j - I\\|_2$ after filtering; if it is not small, the orthonormality premise of Theorem 3.2 is violated.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the random feature method for PDEs that this work preconditions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the rank-revealing QR factorization concept used for local feature filtering."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LSQR is the iterative least-squares solver whose convergence the preconditioner accelerates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the random matrix bound used in Appendix A to estimate overlap couplings probabilistically."},{"cited_title":"Dolean, A","cited_arxiv_id":null,"evidence_quote":"Multilevel domain decomposition cited for the multi-scale test problems and as the proposed extension for larger subdomain counts."}],"review_version":2}