{"id":"daa7cd79-527e-4e8a-b22b-145166cf4c38","arxiv_id":"2506.01574","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Grassmann interpolation via maximum-volume local coordinates is matrix-decomposition-free and preserves Euclidean asymptotic error orders.","lead":"This paper introduces a faster way to interpolate data on the Grassmann manifold, the space of low-dimensional subspaces, using coordinate charts that need no matrix decompositions. The authors add maximum-volume coordinates to keep these charts well-conditioned and prove that the interpolation error behaves the same as in ordinary Euclidean space.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2's row-selection lacks a multi-sample guarantee; the chosen MV permutation can leave another data point's upper p×p block singular, so the practical method's order-of-error claim is not established for arbitrary data.","rationale":"The paper's strongest contribution is the clean error-transfer theorem, and the reader correctly identifies the unproven multi-sample MV selection as the weakest assumption. I agree with the CONDITIONAL verdict; the risk is that in applications the single permutation chosen by Algorithm 2 may not keep all upper blocks well-conditioned, voiding the chart and the error bound. The reader's description of Algorithm 2 as selecting a permutation from a single sample point is slightly imprecise—the pseudocode evaluates all per-sample maxvol candidates and picks the best worst-case—but the conclusion (no multi-sample guarantee) is correct. A decisive counterexample is the orthogonal pair above: no single row-selection can keep both upper blocks nonsingular, so the method can fail outright. The finite-difference step and single-run numerics are secondary reproducibility concerns; they do not affect the theoretical core. Overall, the central theorem is sound, and the CONDITIONAL verdict is appropriate.","tokens_in":23794,"tokens_out":28431,"duration_ms":284181,"concrete_test":"Set p=2 and take U1=[I2;0] and U2=[0;I2] as 4×2 Stiefel matrices. Run Algorithm 2 with the maxvol routine from [53] on the pair {U1,U2}, and inspect the inverse Frobenius norms of the upper 2×2 blocks of both permuted matrices for each candidate permutation. If either candidate yields a singular block (infinite norm), the absence of a multi-sample guarantee is confirmed; if an implementation returns P with finite norms for both matrices, the counterexample needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 2 and the error-preservation claim are internally sound: the arcsin bound follows from Lemma 1, Lemma 5, and the operator-norm bound of Lemma 4. The load-bearing weakness is upstream: the entire method requires a fixed chart Ψ for all sample points, and maximum-volume row selection is what guarantees the chart is defined and well-conditioned. Algorithm 2 actually applies maxvol to each sample separately and then picks the permutation minimizing the worst inverse-block norm across samples, but no a priori bound ties the selected permutation to the other matrices. A simple adversarial example shows the guarantee cannot be automatic: with p=2, take U1=[I2;0] and U2=[0;I2] in R^{4×2}. The maxvol permutation for U1 is the first two rows; for U2 it is the last two rows. Whichever candidate is selected, the upper 2×2 block of the other matrix is zero, so the MV chart is undefined. Hence the practical method can fail outright on valid input, and the order-of-magnitude result is not established for the algorithm on arbitrary data sets. The authors acknowledge the multi-sample generalization as open in Section 5, but the conditional nature of the central practical claim should be stated prominently.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a matrix-decomposition-free local coordinate chart on the Grassmann manifold Gr(n,p), introduces maximum-volume (MV) coordinates obtained by row-permuting Stiefel representatives, and analyzes interpolation in these coordinates. The main theoretical result (Theorem 2, Section 3.6) shows that, under a single-chart assumption and a smallness condition, the Grassmann distance between the true and interpolated subspaces is bounded by arcsin(M||B - Btilde||_F), with M <= sqrt(5/2)+1, so the asymptotic interpolation order is preserved: O(h^2) for two-point Lagrange and O(h^4) for two-point Hermite interpolation. The paper also proves that the proposed parameterizations form a retraction (Theorem 1), bounds the conditioning of the chart and its inverse (Lemmas 3 and 4), and bounds the curvature of unit-speed Grassmann geodesics (Lemma 5). Two numerical experiments, one academic and one on FitzHugh-Nagumo POD subspaces, compare MV-coordinate interpolation with Riemannian normal coordinate interpolation.","tokens_in":23979,"tokens_out":7887,"duration_ms":89008,"significance":"If the claims are accepted, the paper provides a useful alternative to Riemannian normal coordinates for Grassmann interpolation, with the practical advantage that preprocessing does not require large-scale SVDs or matrix exponentials. The error-propagation theorem is cleanly derived with explicit constants, and the conditioning analysis of the coordinate chart and parameterization is a valuable contribution. The authors make their source code available, which strengthens reproducibility. The principal weakness is that the row-selection algorithm for MV coordinates has no multi-sample guarantee; this is acknowledged as an open question in Section 5, but it means the headline error-order claim is established only conditionally on the existence of a single well-conditioned chart covering all data. The paper is honest about this gap, but the scope of the central claim should be stated more prominently.","major_comments":[{"comment":"Algorithm 2 selects a single row-permutation matrix P by applying maxvol to each sample individually and then choosing the candidate that minimizes the worst inverse-block norm, but it provides no a priori guarantee that the selected P is admissible for all samples. A concrete failure mode is obtained with p=2, U1 = [I2; 0] and U2 = [0; I2] in R^{4x2}: the maxvol permutation for U1 must select rows {1,2}, while for U2 it must select rows {3,4}, so whichever candidate permutation is chosen, the upper 2x2 block of the other matrix is identically zero and the MV chart is undefined. Since the asymptotic error estimate in Theorem 2 presupposes a single well-conditioned chart covering all data, the practical Algorithm 2 pipeline does not inherit the estimate for arbitrary data sets.","section":"Section 3.4, Algorithm 2"},{"comment":"The headline statement that the asymptotic interpolation error on Gr(n,p) has the same order as in Euclidean space is proven only under the explicit assumption that all sample points and interpolated images lie in a single chart with a bounded inverse-block norm. This condition is not verified by Algorithm 2 and is left as an open question in Section 5. The authors should state this condition explicitly in Theorem 2 and its surrounding discussion, and should qualify the abstract and Remark 2 so that readers do not attribute an unconditional guarantee to Algorithm 2.","section":"Abstract / Remark 2 / Section 5"}],"minor_comments":[{"comment":"The proof that Q(P(t)) can be chosen smoothly in t is only sketched via a Gram-Schmidt extension; since Gram-Schmidt on a fixed complement can fail when the complement ceases to be transverse, a short argument using a local parallel frame or a standard reference on smooth frame completion would make the proof complete.","section":"Theorem 1"},{"comment":"The finite-difference approximation used for d/dI_a u(t) and d/dI_a v(t) is not specified; please state the step size and the stencil used to compute the derivative data.","section":"Section 4.2"},{"comment":"The selection criterion in Algorithm 2 is written as ||(f_i[1:p,1:p])^{-1}||_F, but the upper block must be understood as the block after applying the candidate permutation; please make this explicit in the pseudocode.","section":"Algorithm 2"},{"comment":"The statement that the coordinates 'do not require any matrix decompositions' is too strong given that Section 3.2 evaluates the parameterization via a matrix square root or a Cholesky decomposition; the authors should rephrase the claim, for example to 'no SVD or matrix exponential of n x n matrices'.","section":"Abstract and Section 3.2"},{"comment":"The text says 'when permuting the columns of the point data' before computing MV coordinates; since the permutation acts on the rows of the Stiefel representation, this should read 'rows'.","section":"Section 4.1"},{"comment":"There are several typographical errors, including 'discrete empericial' (Section 3.4), 'quanitfied' and 'forward to' (Section 5), 'conditons' and 'Boundary conditons' (Section 4.2), and 'goedesicallly' (Appendix A); a careful proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"I concur with the reader's high-confidence assessment that the main theoretical derivations are sound. The requested revision should not require solving the multi-sample maxvol problem, but it must make the conditional nature of the practical error-order claim explicit and ideally add the simple adversarial example or a sufficient condition for Algorithm 2 to succeed. The paper is otherwise mathematically solid and within the scope of math.NA."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is worth reading and worth refereeing. The core theoretical result, Theorem 2, is clean: under a single-chart assumption, the asymptotic order of the Euclidean interpolation error is preserved on Gr(n,p), with an explicit constant M <= sqrt(5/2)+1. The lemmas leading to it (Lipschitz bound for the parameterization, curvature bound for Grassmann geodesics) check out algebraically. That alone is a useful contribution, and the paper is honest about where the limits are.\n\nWhat's genuinely new: Helmke-Moore style local coordinates (U2 U1^{-1}) have been around for decades but were not analyzed as interpolation charts. The conditioning analysis and the maxvol row-selection strategy for choosing a chart are new, and the numerical experiments show MV coordinates can reduce preprocessing errors by orders of magnitude and match Riemannian normal coordinates in accuracy. The paper ships code, which is good practice.\n\nSoft spots, in order of importance. First, the MV chart selection in Algorithm 2 has no multi-sample guarantee. The algorithm picks the best permutation from per-sample maxvol candidates, but nothing a priori ties that permutation to the other data points. The stress-test example is correct: with U1=[I2;0], U2=[0;I2], whichever permutation you pick, the other matrix has a singular upper 2x2 block, so the chart is undefined. The authors flag this generalization as open in Section 5, but the abstract's phrasing could leave a reader thinking the guarantee covers the algorithm as stated. That should be tightened.\n\nSecond, the numerical evidence is suggestive, not systematic: single runs, and the finite-difference step in the neuron experiment is not specified, so the derivative data quality is hard to assess. Third, minor: the smoothness proof of Theorem 1 is sketched rather than detailed.\n\nOverall, the paper's central claim holds up, and its main weakness is a clearly-labeled open question rather than a hidden flaw. The right outcome is a serious peer-review process, ideally with an added section on when MV row selection provably works (or a counterexample-based caveat) and a small multi-run numerical study.\n\nRecommendation: send to peer review. The paper deserves referee time and is likely to be citable after modest revision.","headline":"A clean, useful paper on decomposition-free Grassmann coordinates, with an honest gap between the MV heuristic and the error theorem.","tokens_in":24521,"tokens_out":2112,"would_cite":true,"duration_ms":22796,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15B10","15B57","65F99","53C30","53C80"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that Grassmann manifold interpolation can be done in matrix-decomposition-free maximum-volume local coordinates while preserving the same asymptotic error orders as Euclidean interpolation.","keywords":["Grassmann manifold","maximum-volume coordinates","interpolation","Hermite interpolation","local coordinates","retraction","parametric model reduction","error bounds"],"falsifier":"Run the FitzHugh-Nagumo POD interpolation experiment with maxvol terminated after a few iterations, as in the paper's Figure 6, at a parameter interval where the left and right POD bases have nearly disjoint dominant row patterns, and record the relative interpolation error as the step size $h$ shrinks; if the error stops following $O(h^2)$ or $O(h^4)$ scaling despite the MV-chart being used, the practical claim that MV coordinates keep interpolation numerically reliable would be refuted.","tokens_in":23550,"feed_emoji":"📐","tokens_out":5201,"duration_ms":50766,"temperature":0.7,"pith_summary":"This paper introduces maximum-volume (MV) local coordinates for interpolating functions whose values lie on the Grassmann manifold Gr(n,p). Unlike Riemannian normal coordinates, these coordinates require no matrix decompositions: data are mapped through the chart $B = U_2 U_1^{-1}$ and mapped back through $\\varphi(B)$, which only needs solving with $I+B^T B$. The paper proves an explicit error-transmission bound, $\\mathrm{dist}(P,\\tilde P) \\le \\arcsin(M\\|B-\\tilde B\\|_F)$ with $M \\le \\sqrt{5/2}+1$, which implies the same asymptotic error order as in Euclidean space: $O(h^2)$ for two-point Lagrange and $O(h^4)$ for Hermite interpolation. Because ill-conditioned row blocks can corrupt the chart, the paper proposes selecting the $p$ rows of each Stiefel representative with maximum absolute determinant (maxvol), and shows numerically that this makes interpolation accuracy comparable to Riemannian normal coordinates while avoiding decomposition costs. A reader should care because this offers a cheaper route to Grassmann interpolation in parametric model reduction and other high-dimensional applications.","feed_headline":"Grassmann interpolation keeps Euclidean error rates without matrix decompositions","feed_subtitle":"Max-volume coordinates give Lagrange O(h^2) and Hermite O(h^4) errors on the manifold, matching flat-space interpolation.","key_machinery":"The load-bearing objects are the chart-parameterization pair $(\\Psi, \\varphi)$ and the maxvol row-selection rule. $\\Psi$ sends a projector to the matrix $B = U_2 U_1^{-1}$ of a linear fractional map between complementary coordinate patches; $\\varphi(B)$ is the orthogonal projector onto $\\mathrm{span}([I_p; B])$ and is a retraction on $\\mathrm{Gr}(n,p)$. The maxvol rule selects, among all $p \\times p$ subblocks of a Stiefel representative, one with maximal absolute determinant, which controls $\\|U_1^{-1}\\|_F$ and hence the chart's conditioning. Theorem 2 combines Lemma 1's sandwich bound with Lemma 5's curvature bound ($\\kappa \\le 2$) and Lemma 4's bound on $d\\varphi$ to turn coordinate-domain interpolation errors into manifold-distance errors.","core_discovery":"The central discovery is that Grassmann interpolation can be formulated in decomposition-free local coordinates without losing asymptotic accuracy. The coordinate chart $\\Psi$ maps a projector represented by $U = [U_1; U_2]$ to $B = U_2 U_1^{-1}$; its inverse $\\varphi(B) = [I; B](I + B^T B)^{-1}[I, B^T]$ is globally defined and forms a retraction. The paper quantifies the conditioning of $\\varphi$ (worst-case amplification $M \\le \\sqrt{5/2}+1$) and of the chart, showing the chart can be arbitrarily ill-conditioned unless the row block $U_1$ is chosen well. Choosing $U_1$ as a maxvol subblock (maximizing $|\\det U_1|$) bounds $\\|U_1^{-1}\\|_F$ by $\\sqrt{p(n-p)+p}$ and makes the coordinates usable. Theorem 2 then shows $\\mathrm{dist}(P,\\tilde P) \\le \\arcsin(M\\|B-\\tilde B\\|_F)$, so Euclidean interpolation errors of order $h^k$ survive transfer to the manifold; in particular, two-point Lagrange interpolation is $O(h^2)$ and Hermite interpolation is $O(h^4)$.","pith_inferences":["The single-permutation assumption could be tested directly: if a data set contains two samples whose maxvol-optimal row blocks are nearly disjoint, a chart chosen from one sample will be ill-conditioned at the other; a multi-sample maxvol that optimizes the worst case over all samples would close the open question posed in the paper.","The bound $M \\le \\sqrt{5/2}+1$ is global and not sharp, as the paper's own Figure 3 shows slack; a tighter curvature- or data-dependent constant would sharpen the error estimates and might be derivable from the principal-angle parametrization.","Because the maxvol row-selection problem is the same one faced by DEIM and CUR decompositions, algorithms and guarantees from that literature could be imported to build provably good multi-sample charts.","The three-step error accounting (preprocessing, interpolation, postprocessing) extends naturally to other retractions on the Grassmann manifold, offering a way to rank retractions by their preprocessing conditioning."],"forward_implications":["Because the coordinates avoid SVDs and matrix exponentials, Hermite interpolation no longer requires differentiating matrix decompositions, lowering preprocessing cost.","Theorem 2 guarantees that any Euclidean interpolation scheme with asymptotic error $O(h^k)$ in the coordinate domain yields $O(h^k)$ interpolation error on the Grassmann manifold.","For two-point data, Lagrange interpolation scales as $O(h^2)$ and Hermite interpolation as $O(h^4)$, matching the Euclidean rates.","MV coordinates are not just an accuracy improvement but a numerical necessity: without them, the FitzHugh-Nagumo POD experiment shows interpolation errors that are orders of magnitude larger and spike-driven.","The same coordinate framework extends to more than two data points, provided all samples lie in a single chart, making higher-order interpolants straightforward."],"supporting_citations":[{"why":"Supplies the decomposition-free local parameterization of the Grassmann manifold that the paper builds on.","marker":"[22]"},{"why":"Provides the Grassmann geometry background, tangent-space characterization, and geodesic formulas used throughout.","marker":"[38]"},{"why":"Gives the volume lemma bounding $\\|U_1^{-1}\\|_F$ for a maxvol subblock, justifying the MV chart.","marker":"[47]"},{"why":"Provides the maxvol algorithm used in Algorithm 2 to compute the permutation matrix.","marker":"[53]"},{"why":"The Riemannian-normal-coordinate Hermite interpolation and error-propagation baseline that the new method is compared against.","marker":"[23]"},{"why":"The sandwich bounds between Euclidean and manifold distances used as Lemma 1.","marker":"[34]"},{"why":"The classical Lagrange and Hermite interpolation error estimates that yield $O(h^2)$ and $O(h^4)$.","marker":"[46]"}],"fun_headline_variants":["Maxvol Grassmann coords match flat-space interpolation errors","Decomposition-free Grassmann interpolation reaches Euclidean rates","Grassmann maxvol charts: Euclidean accuracy without factorizations","No matrix decompositions: Grassmann interpolation error stays Euclidean"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The practical success of the method rests on the assumption, which the paper leaves as an open question, that a single row-permutation matrix picked by maxvol at one sample point keeps the upper $p \\times p$ blocks of all other sample matrices well-conditioned; the paper provides no a priori bound for multiple data matrices.","fun_headline_variants_meta":{"raw":{"variants":["Maxvol Grassmann coords match flat-space interpolation errors","Decomposition-free Grassmann interpolation reaches Euclidean rates","Grassmann maxvol charts: Euclidean accuracy without factorizations","No matrix decompositions: Grassmann interpolation error stays Euclidean"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000416,"raw_usage":{"total_tokens":2204,"prompt_tokens":1059,"completion_tokens":1145,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":1079}},"tokens_in":675,"tokens_out":1145,"duration_ms":9199,"temperature":1.0,"reasoning_tokens":1079,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:39:26.632580+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the FitzHugh-Nagumo POD interpolation experiment with maxvol terminated after a few iterations, as in the paper's Figure 6, at a parameter interval where the left and right POD bases have nearly disjoint dominant row patterns, and record the relative interpolation error as the step size $h$ shrinks; if the error stops following $O(h^2)$ or $O(h^4)$ scaling despite the MV-chart being used, the practical claim that MV coordinates keep interpolation numerically reliable would be refuted.","supporting_citations":[{"cited_title":"Communications & Control Engineering","cited_arxiv_id":null,"evidence_quote":"Supplies the decomposition-free local parameterization of the Grassmann manifold that the paper builds on."},{"cited_title":"In: Matrix Methods: Theory, Algo- rithms And Applications: Dedicated to the Memory of Gene Golub, pp","cited_arxiv_id":null,"evidence_quote":"Provides the maxvol algorithm used in Algorithm 2 to compute the permutation matrix."},{"cited_title":"In: Proceedings of the 2007 ACM Symposium on Solid and Physical Modeling","cited_arxiv_id":null,"evidence_quote":"The sandwich bounds between Euclidean and manifold distances used as Lemma 1."},{"cited_title":"Springer, ??? (2007)","cited_arxiv_id":null,"evidence_quote":"The classical Lagrange and Hermite interpolation error estimates that yield $O(h^2)$ and $O(h^4)$."}],"review_version":1}