{"id":"603de392-6cd5-4dbb-ae86-daf9e0f1b6bd","arxiv_id":"2505.05134","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Bochner matrices (matrices with Hilbert-space entries) admit cross decompositions and a new adaptive cross-dyadic approximation algorithm that numerically approximates parametric PDE solution maps.","lead":"This paper develops a linear algebra theory for matrices whose entries live in a Hilbert space, and extends cross approximation to these 'Bochner matrices'. It then proposes an adaptive sampling algorithm for approximating parametric PDE solution maps and tests it on linear and nonlinear Stokes problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract promises approximation guarantees and maximum-volume bounds that do not appear in the text; Section 6.3 defers them, and Figure 1 shows random pivots are nearly as good, so the empirical quasioptimality claim for ABCD is unestablished.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the pivot-selection heuristic in Algorithms 3–4 has no theorem showing it identifies near-optimal rows or columns, and the paper explicitly defers approximation guarantees to future work. My stress-test adds two supporting observations. First, the abstract overclaims by stating that approximation guarantees and maximum-volume bounds are derived, when no such results appear anywhere in the text. This is not merely a stylistic mismatch: it directly affects what the paper advertises as its central contribution. Second, the paper's own Figure 1 undercuts the empirical case for the heuristic: random index selection is only marginally worse than ABCD, so the numerical evidence that ABCD selects \"good\" indices is weak. The Stokes benchmark (Table 1) does not include a random baseline, so the quasioptimality there could be coincidental. I also note a likely sign error in the Section 7.1 analytic Fourier formula, which would invalidate that benchmark if confirmed. These issues are serious but addressable: the abstract should be aligned with the actual content, the benchmark should be corrected, and a random-pivot baseline should be reported. The mathematical core—the exact cross decomposition lemmas and the HOSVD bound—appears sound and independently useful. Therefore the appropriate disposition remains CONDITIONAL, matching the reader's verdict; no change is needed.","tokens_in":32216,"tokens_out":6217,"duration_ms":63996,"concrete_test":"Recompute the Stokes experiment (Section 7.3, Table 1) with fully random row and column selection—both indices drawn uniformly, as in the \"random cross\" baseline of Figure 1—and compare the resulting error curves against ABCDX with nrook=1. If the random-pivot error lies within the plotted spread of ABCDX, then the adaptive pivoting is not responsible for the observed quasioptimality, confirming that the missing guarantees are load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The advertised central result—\"approximation guarantees are derived\" and maximum-volume bounds (abstract)—is not present: the only error bounds in the text are for SVD/HOSVD (Theorems 5.12, 5.16), and Section 6.3 explicitly states that approximation guarantees for Bochner cross approximation are future work. Consequently, the claim that ABCD/ABCDX select index sets whose error is within a small constant of HOSVD (Section 7) has no theoretical support. Lemma 6.8 only gives exact interpolation under a rank condition, with no perturbation or quasioptimality statement. Moreover, the numerical evidence is weak: Figure 1 shows a completely random choice of indices is \"only marginally worse\" than ABCD's pivots, and for the Stokes problem no random baseline is reported. The pivot rule (Algorithm 3) maximizes the Hilbert norm of the residual pointwise; without a volume-type or subset-selection bound, this heuristic could be no better than random. The analytic benchmark in Section 7.1 also appears to contain a sign error: for alpha=0, beta=1, k=2, the stated formula for u_k gives 0, but the sine-series coefficient of cos(x) at k=2 is nonzero, so the reported error curves may not correspond to the stated PDE.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops an elementary theory of matrices whose entries lie in a Hilbert space (Bochner matrices), covering norms, ranks, adjoints, QR and SVD factorizations, Mirsky-type low-rank approximation, and HOSVD. It defines a cross component for Bochner matrices and proves exact interpolation on a selected row/column index set under a Tucker-rank condition (Lemmas 6.4 and 6.8). It then proposes adaptive algorithms ABCD and ABCDX, which select pivots by the Hilbert norm of residual entries, and validates them numerically on a parametric linear boundary-value problem and on parametric nonlinear Stokes equations. The abstract claims that approximation guarantees and maximum-volume bounds are derived, but the body explicitly defers those guarantees to future work.","tokens_in":32472,"tokens_out":15636,"duration_ms":150909,"significance":"The algebraic framework is natural, and the core lemmas in Sections 3-6 are mostly correct and self-contained. The exact cross interpolation formula for Bochner matrices and the explicit SVD/HOSVD extensions are useful contributions for parametric PDE surrogate modeling, and the paper contains no fitted parameters in its theoretical results. The authors are transparent about the heuristic status of ABCD in Section 6.3, and the comparison with HOSVD is a sensible evaluation protocol. However, the advertised approximation guarantees and maximum-volume bounds are absent from the manuscript, and the analytic benchmark in Section 7.1 appears inconsistent with the stated PDE. These issues currently prevent acceptance.","major_comments":[{"comment":"The abstract states that \"approximation guarantees are derived\" and that \"the high-dimensional nature of the entries is shown to manifest itself in maximum-volume bounds,\" but neither statement is supported by the text. Section 1 explicitly says \"Neither do we address the question of approximation guarantees for Bochner cross approximation,\" and Section 6.3 contains no theorem about the index sets chosen by ABCD, only an expectation that they are good. The only error bounds in the paper are Theorems 5.12 and 5.16, which concern SVD and HOSVD approximation, not cross approximation, and no maximum-volume definition or bound appears anywhere. This is a load-bearing mismatch between the paper's advertised central result and its actual content; the claims must be removed or the missing results supplied.","section":"Abstract; §1; §6.3"},{"comment":"The displayed solution formula does not solve the stated boundary-value problem (2). For -u'' - α²u = cos(βx), the sine coefficients are proportional to k[1 - (-1)^k cos(βπ)] / [(k² - α²)(k² - β²)], whereas the paper gives the same numerator over (k² + α²)(k² + β²). For example, at α = 0, β = 1, and k = 2, the true coefficient is nonzero and differs from the value produced by the displayed formula. Consequently, the L2 and H1 error curves in Figure 1 are computed with a function that is not the solution of (2), so the validation experiment does not test the stated problem. The PDE, the formula, or both must be corrected.","section":"§7.1, Eq. (2)"},{"comment":"Definition 5.7 is not well-typed as written: it defines a Bochner norm ||·||_• to be unitarily invariant if there is a usual-matrix norm ||·||_∘ with ||QA||_• = ||A||_∘ for every Q with orthonormal columns, but on the right A is a Bochner matrix while ||·||_∘ acts on usual matrices. In the proof of Theorem 5.8, ||UΣV*||_• is identified with ||ΣV*||_∘ and then with ||Σ||_∘ without specifying how a usual matrix is embedded into the Bochner norm. Since Theorem 5.12 and its use in Theorem 5.16 depend on this identification, the definition and proof need to be repaired.","section":"§5.2, Definition 5.7 and Theorem 5.8"},{"comment":"The paper claims that ABCD/ABCDX select index sets whose approximation error is close to HOSVD, but this is not established. No theorem connects the argmax-over-||·||_H pivot rule in Algorithms 3-4 to near-optimal rows or columns; Section 6.3 defers such guarantees. The only direct comparison with random selection, in Figure 1, shows that random cross is \"only marginally worse\" than ABCD, and no random baseline is reported for the Stokes experiments in Figure 3 and Table 1. Please either provide a subset-selection or volume-type bound that supports the pivot rule, or add random baselines and weaken the quasioptimality claim to a heuristic observation.","section":"§6.3, Algorithms 3-4; §7.1, Fig. 1; §7.3, Fig. 3/Table 1"}],"minor_comments":[{"comment":"There is a typo in Definition 5.7: \"unitraily\" should read \"unitarily.\"","section":"§5.2, Definition 5.7"},{"comment":"Section 7 states that a link to the Julia module repository will be added in an updated version; for reproducibility of the numerical claims, the link or a data-availability statement should be present in the submitted version.","section":"§7"},{"comment":"The notation in Figure 1 uses \"H = L2\" and \"H = H1_0\" for the entry space; since the paper consistently uses H for an abstract Hilbert space, a brief note clarifying the two usages would improve readability.","section":"§7.1"},{"comment":"The sentence \"Neither do we address the question of approximation guarantees for Bochner cross approximation\" directly contradicts the abstract; aligning these statements is essential even if the technical results remain unchanged.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"The abstract/body mismatch and the Section 7.1 formula error are the main obstacles. If the promised guarantees and maximum-volume bounds do not exist, the paper should be repositioned as a theory-plus-heuristic numerical study. The Section 7.1 error is independent of the theoretical core and suggests that the numerical section needs a careful recheck before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the mathematical core is a clean, correct extension of cross approximation to Hilbert-space-valued matrices, but the abstract overclaims and the analytic benchmark in Sec. 7.1 has a sign error. It's a conditional accept, not a reject.\n\nWhat's new: Bochner matrices with distinct column/row ranks, the cross component construction with Tucker rank equality (Lemmas 6.4-6.8), and the HOSVD error bound (Thm 5.16). The proofs are elementary and I checked the key steps—they hold up. The paper is honest in the intro that approximation guarantees for Bochner cross approximation are future work, which is why the abstract's 'approximation guarantees are derived' is a real mismatch.\n\nSoft spots. The sign error in (2): for alpha=0, beta=1, k=2, the formula gives u_k=0.2, but the correct coefficient is 1/3; the denominator should be (k^2 - beta^2) not (k^2 + beta^2). So the error curves in Figure 1 are not for the stated PDE. That doesn't touch the theory, but it makes the first numerical experiment unreliable. The pivot-selection heuristic (Algorithms 3-4) has no volume-type bound; the paper defers this to future work. Figure 1 shows random cross is nearly as good as ABCD's chosen indices, so the claim that ABCD selects 'good' index sets is unsupported. The Stokes experiment (Fig 3) is more favorable—rook pivoting clearly helps—but there's no random baseline, so we can't judge quasioptimality there. No code is shipped (repository link is a promise).\n\nOverall: the framework is useful, especially for ROM people who want a non-intrusive way to approximate solution maps. The theory is sound enough for a serious referee, but the abstract, the benchmark, and the numerical claims all need work. I'd send it to review, with a note that the revision must correct or remove the benchmark, align the abstract with the actual results, add a random baseline for Stokes, and release the code. It's not a desk rejection—the core ideas deserve careful vetting, but as is it's not ready.","headline":"A clean but overclaimed extension of cross approximation to Hilbert-space-valued matrices; the core theory is sound, but the abstract promises results the paper does not deliver and the analytic benchmark has a sign error.","tokens_in":32974,"tokens_out":5076,"would_cite":false,"duration_ms":39672,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A23","15B99","65F55"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper extends low-rank cross approximation to matrices whose entries lie in a Hilbert space, proving an interpolatory cross-decomposition identity and validating an adaptive variant on parametric PDE solution maps.","keywords":["Bochner matrices","low-rank approximation","cross approximation","adaptive cross approximation","parametric PDEs","reduced-order modeling","Tucker rank","HOSVD"],"falsifier":"Run ABCD or ABCDX with one rook-pivoting round on a family of Bochner matrices with known singular-value decays and compare each selected-index cross component with the HOSVD error at the same Tucker rank; if the error ratio grows with matrix size, or if randomly selected index sets achieve comparable errors, then the adaptive pivot heuristic is not doing the work attributed to it.","tokens_in":31999,"feed_emoji":"🧮","tokens_out":7786,"duration_ms":74551,"temperature":0.7,"pith_summary":"The paper tries to establish that low-rank cross approximation, a workhorse of numerical linear algebra for ordinary matrices, still works when each entry of the matrix is itself an element of a Hilbert space, such as a function solving a parametric PDE. It proves that a cross component built from selected rows and columns interpolates the Bochner matrix on that index set and carries the same Tucker rank, and it proposes adaptive algorithms, ABCD and ABCDX, that use Hilbert-space norms to pick pivots. Numerically, these adaptive cross components approximate the solution maps of a linear model problem and of parametric nonlinear Stokes equations at errors within a small constant of the HOSVD baseline, while accessing only selected rows and columns of the matrix. The abstract also advertises maximum-volume bounds and approximation guarantees, but those results do not appear in the text; the body explicitly defers approximation guarantees to future work.","feed_headline":"Low-rank cross approximation comes to Hilbert-valued matrices","feed_subtitle":"Adaptive row-and-column selection builds PDE solution-map surrogates at near-HOSVD accuracy.","key_machinery":"The central object is the Bochner matrix, a finite array whose entries lie in a Hilbert space, together with the pseudoinverse acting on it. The workhorse identity is the cross component formula cross(A,I,J)=((G^T)^+ C^T)^T G (G^+ R), where G is the submatrix at the selected index set and C and R are the corresponding columns and rows. The associated concepts are the column and row ranks, the Tucker rank pair, and the SVD-based HOSVD projection that yields quasioptimal low-rank approximations. The formula's job is to express the whole matrix through a small submatrix G and two side factors built from selected columns and rows, so that only a small number of rows and columns must actually be evaluated.","core_discovery":"On its own terms, the paper's central claim is that the identity cross(A,I,J)=((G^T)^+ C^T)^T G (G^+ R), with C, R, G the selected columns, rows, and intersection, is the right Bochner analogue of the classical cross and pseudoskeleton decomposition: it interpolates A on I x J, its Tucker rank equals the Tucker rank of the selected submatrix G, and it reproduces A exactly when rank_T(G)=rank_T(A) (Lemmas 6.4 and 6.8). The paper further claims that the adaptive algorithms ABCD and ABCDX, which choose pivots by the largest Hilbert-space norm in the residual, select index sets whose cross-component error trails the quasioptimal HOSVD error by only a small factor in parametric PDE experiments.","pith_inferences":["If the pivot heuristic were backed by a theorem, the same index-selection strategy would give a non-intrusive reduced-order modeling pipeline whose cost is measured in inner products between PDE solutions, independent of the discretization dimension.","The gap between the dyadic ABCD approximant and its post-processed cross component suggests decoupling selection from approximation: use cheap norm-based pivoting for exploration, then form a one-shot cross component as the final surrogate.","A natural testable extension, not pursued in the paper, would be to fold a higher-order Bochner tensor into a Bochner matrix and run ABCDX mode by mode, connecting to fiber-based tensor completion; the elementary rank and interpolation identities developed here suggest the same strategy would carry over.","If the deferred maximum-volume bounds materialize, they would likely be smaller than in the classical setting because the Hilbert-space dimension of the entries weakens the volume constraint, exactly as the abstract anticipates."],"forward_implications":["A Bochner matrix can be reconstructed exactly from a cross component whenever the selected submatrix has the same Tucker rank as the whole matrix, so a few carefully chosen rows and columns can encode all information for rank-deficient solution maps.","The cross component always interpolates the matrix on the selected index set and has Tucker rank bounded by the selected submatrix, giving a low-rank surrogate whose error is controlled entirely by the index choice.","HOSVD-style projection onto dominant left and right singular subspaces yields a Tucker approximation within a factor of $\\sqrt{2}$ of the best possible error in the $\\ell^2(H)$ norm.","On the parametric nonlinear Stokes equations, ABCDX with rook pivoting reaches errors about twice the HOSVD error at the same rank, demonstrating that the method behaves as a non-intrusive reduced-order model.","The adaptive algorithms can be integrated with existing PDE software, meaning the surrogate is built from solution evaluations and inner products rather than from the discretized system matrices."],"supporting_citations":[{"why":"Supplies the pseudoskeleton cross-approximation theory whose Bochner analogue is built in Lemmas 6.3 and 6.4.","marker":"[28]"},{"why":"Introduces adaptive cross approximation, the method that Algorithms 2 through 5 generalize.","marker":"[6]"},{"why":"Provides the equivalence theorem for CUR decompositions that the Bochner cross-component identity extends.","marker":"[33]"},{"why":"Generalizes column-row matrix decomposition to tensors, the closest existing analogue of the ABCDX algorithm.","marker":"[13]"},{"why":"Defines HOSVD, the quasioptimal baseline against which ABCD and ABCDX approximation errors are measured.","marker":"[19]"},{"why":"Gives the unitarily invariant norm optimality used to prove low-rank approximation bounds for Bochner matrices.","marker":"[55]"},{"why":"Provides volume-sampling guarantees for vector-valued functions, the nearest published support for the pivot-selection guarantees this paper defers.","marker":"[43]"}],"fun_headline_variants":["Bochner matrices get low-rank cross approximation","Cross approximation for Hilbert-valued matrices","Adaptive cross approximation for parametric PDE surrogates","Near-HOSVD accuracy from adaptive cross on Hilbert matrices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that picking entries with the largest Hilbert-space norm during the search identifies near-optimal rows and columns; the paper states this expectation but does not prove it, explicitly deferring approximation guarantees to future work.","fun_headline_variants_meta":{"raw":{"variants":["Bochner matrices get low-rank cross approximation","Cross approximation for Hilbert-valued matrices","Adaptive cross approximation for parametric PDE surrogates","Near-HOSVD accuracy from adaptive cross on Hilbert matrices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000769,"raw_usage":{"total_tokens":3324,"prompt_tokens":782,"completion_tokens":2542,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":398,"completion_tokens_details":{"reasoning_tokens":2484}},"tokens_in":398,"tokens_out":2542,"duration_ms":17574,"temperature":1.0,"reasoning_tokens":2484,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:14:27.583178+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ABCD or ABCDX with one rook-pivoting round on a family of Bochner matrices with known singular-value decays and compare each selected-index cross component with the HOSVD error at the same Tucker rank; if the error ratio grows with matrix size, or if randomly selected index sets achieve comparable errors, then the adaptive pivot heuristic is not doing the work attributed to it.","supporting_citations":[{"cited_title":"Numer Math86,565–89.doi: 10","cited_arxiv_id":null,"evidence_quote":"Introduces adaptive cross approximation, the method that Algorithms 2 through 5 generalize."},{"cited_title":"Generalizing the column–row matrix decomposition to multi-way arrays","cited_arxiv_id":null,"evidence_quote":"Generalizes column-row matrix decomposition to tensors, the closest existing analogue of the ABCDX algorithm."}],"review_version":1}