Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Structured Column Subset Selection for Bayesian Optimal Experimental Design

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that structured Bayesian optimal experimental designs can be found near-optimally by reshaping the design matrix into a tensor and selecting columns per mode, with speedups up to 50× over greedy methods.

desk verdict Useful extension of CSSP to multiway OED, but the near-optimal claim rests on an unproved heuristic transfer that the authors themselves acknowledge. read the letter →

arxiv 2506.00336 v1 pith:QLF5UMV2 submitted 2025-05-31 math.NA cs.NA

Hugo Díaz , Arvind K. Saibaba , Srinivas Eswar , Vishwas Rao , Zichao Wendy Di This is my paper · ORCID
classification math.NAcs.NA MSC 58F1558F1753C35
keywords OptimalexperimentaldesignColumnsubsetselectionBayesianinverseproblemsTensordecompositionsRandomizedmethodsExpectedinformationgainGolub–Klema–StewartSensorplacement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper extends column-subset-selection-based optimal experimental design (OED) from problems with a single design variable to problems where design choices come in a multiway structure—for instance, choosing both sensor locations and measurement times, or sources and receivers. The key move is to reshape the weighted forward operator into a tensor and apply a standard column-selection routine to each mode of that tensor, so the selected design is automatically a product of per-mode choices ($S = S_d \otimes \cdots \otimes S_1$). Three algorithmic templates are proposed—independent, sequential, and iterative mode selection—alongside randomized Sketch-First variants that compress the matrix before selection. In numerical experiments on time-dependent inverse problems, seismic tomography, X-ray tomography, and flow reconstruction, the methods deliver expected information gain close to the best achievable designs and run up to 50 times faster than greedy approaches. If the mode-by-mode transfer is valid, the paper provides a scalable, practical way to do structured Bayesian experimental design.

What carries the argument

The central object is the tensor $\mathcal{X}$ obtained by reshaping the matrix $A = \Gamma_{pr}^{1/2} F^T R^{-1/2}$, whose columns index candidate experimental designs. The GKS method—a truncated SVD of a mode unfolding followed by column-pivoted QR on the top right singular vectors—serves as the column subset selection engine applied to each mode, yielding selection matrices $S_j$ whose Kronecker product $S = S_d \otimes \cdots \otimes S_1$ guarantees the multiway constraint. The three templates differ in how they use the other modes: IndSelect applies GKS independently to each mode unfolding; SeqSelect conditions each mode on previously selected indices, paralleling sequentially truncated HOSVD; and IterSelect sweeps through modes repeatedly until the expected information gain stops improving, paralleling higher-order orthogonal iteration. The Sketch-First variant replaces $A$ with a random row sketch $\Omega A$, so the algorithm never needs to form $A$ explicitly and never needs the adjoint of the forward operator.

What would settle it

Run exhaustive search on a small structured problem—say $d=3$ modes, $m_i=9$ choices per mode, and $k_i=3$ selections per mode—and compare the best expected information gain found by IndSelect, SeqSelect, and IterSelect (with random restarts) against the global optimum under the product constraint. If any of the proposed methods falls meaningfully short of the optimum while the one-mode Lemma 1 bound would predict near-optimality, the mode-by-mode transfer is falsified.

Watch

Extended reading notes

Core claim

The central claim is that structured OED—where $d$ categorical design variables each contribute $k_i$ selections from $m_i$ choices—can be solved by reshaping the matrix $A$ (the prior-weighted, noise-whitened forward operator) into a $(d+1)$-way tensor $\mathcal{X}$ and applying the Golub–Klema–Stewart column subset selection method, mode by mode. The selection operator produced by the algorithm has the Kronecker product form $S = S_d \otimes \cdots \otimes S_1$, which enforces the structured constraint by construction. For $d=1$, Lemma 1 bounds the expected information gain of the GKS selection against the optimal design; the paper's contribution is to carry this machinery to $d>1$ and to demonstrate through four inverse problems that the resulting designs are near-optimal in EIG, with computational savings up to $50\times$ over greedy baselines when the Sketch-First randomized variants are used. The authors present no multiway analog of Lemma 1, leaving the mode-by-mode transfer as an empirical conclusion.

Load-bearing premise

The load-bearing premise is that the GKS column-selection heuristic, proved near-optimal only for a single mode ($d=1$), continues to produce near-optimal expected information gain when applied mode by mode to tensor unfoldings under the product constraint; the paper states that it could not derive the multiway analog of Lemma 1.

Editorial extensions

If this is right

  • Existing CSSP subroutines used for $d=1$ can be reused as black boxes for each tensor mode, so structured OED inherits the maturity of standard column selection software.
  • The Sketch-First variants require only forward operator applications, not the adjoint, making the method usable with legacy or implicit forward solvers.
  • IterSelect converged in at most four iterations on all tested problems, and its cost depends on the number of selected designs $K$ rather than the full candidate count $M$, making it the fastest template.
  • Structured designs that select roughly 7% of the candidate source–receiver pairs increase reconstruction error only from 9.4% to 11.5% relative to using all data, so the near-optimal designs retain most of the information.
  • The greedy variants achieve slightly higher EIG on some problems but at one-to-two-orders-of-magnitude higher runtime, so the GKS-based templates are the practical default.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to run IterSelect from many random initializations and check whether the selected EIG distribution is tightly concentrated; if not, the mode-by-mode landscape has multiple local optima and the deterministic templates may be unreliable.
  • The same tensor-unfolding template could be applied to other optimality criteria such as A-, C-, or E-optimality, or to nonlinear OED via local linearization of the forward map, since the machinery only requires a candidate matrix of column sensitivities.
  • If the mode-by-mode transfer is validated theoretically, the connection to HOSVD and sequentially truncated HOSVD suggests that adaptively choosing the mode processing order could yield further speedups or accuracy gains.
  • The Sketch-First substitution of $\log\det(I + (YS)(YS)^T)$ for the true EIG has only a one-mode justification; a rigorous multiway bound linking the sketched EIG to the true EIG under the product constraint would turn an empirical observation into a guarantee.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper extends the column subset selection approach to Bayesian D-optimal experimental design from the unstructured setting (d=1) to structured selections, where the design variables are arranged in d categories and one must select k_j indices per category. The authors reshape the weighted forward operator A into a tensor, apply CSSP-style selection (primarily GKS) to mode unfoldings either independently, sequentially, or iteratively, and introduce a Sketch-First randomized variant that avoids forming A explicitly and does not require the adjoint. The paper provides computational cost analyses, greedy variants for comparison, and numerical experiments on a time-dependent heat equation, seismic tomography, X-ray tomography, and flow reconstruction. The reported experiments show that the proposed methods outperform random designs and are much faster than the greedy baselines.

Significance. If the central effectiveness claim is accepted, the paper would fill a genuine gap: prior CSSP-based OED handles only independent column selection, whereas many applications require group or multiway selection. The algorithmic framework is clearly specified, the cost analysis is useful, and the adjoint-free sketching variant is a practical contribution for problems where A^T is unavailable. The four application studies demonstrate broad applicability, and the greedy counterparts provide a reasonable baseline. However, the theoretical support for multiway near-optimality is absent: the paper explicitly states that no analog of Lemma 1 was derived for the structured case, and the only exhaustive validation collapses to d=1. The numerical evidence against random designs is suggestive but does not by itself establish near-optimality. The contribution is therefore best viewed as a heuristic algorithmic framework whose practical utility is plausible but whose main claim needs either a proof, a counterexample analysis, or a substantively weakened statement.

major comments (4)
  1. [Section 5] The central claim that the proposed methods produce near-optimal EIG designs is not supported by a theoretical guarantee for d>1. Lemma 1, the only result linking GKS column selection to EIG optimality, is proved only for d=1. Section 5 explicitly states 'we were not able to derive analogs of Lemma 1 for the structured column selection.' Applying GKS mode-by-mode to unfoldings optimizes a different objective, and no bound connects the resulting Kronecker-structured selection S_d⊗...⊗S_1 to logdet(I + AS(AS)^T). The paper should either provide a structured analog of Lemma 1 (even under restrictive conditions), give a concrete counterexample showing the limitation, or explicitly reframe the multiway results as heuristics that are validated only empirically.
  2. [Section 4.1] The exhaustive-search validation does not actually test the multiway case. In the time-dependent problem, the selection is of the form S = I_T ⊗ S_1 with k_2 = T, so only the sensor mode is nontrivial; the tensor is 28×10×401 but the effective problem reduces to d=1 column selection. Thus the only experiment that establishes near-optimality by comparing against all designs provides no evidence for the interaction between modes, which is the paper's new contribution. A small d=2 exhaustive experiment (for example, small candidate sets in the seismic or X-ray setting with all source-receiver pairs enumerated) would directly test whether the mode-by-mode heuristics preserve EIG near-optimality.
  3. [Sections 4.2-4.4] In the seismic, X-ray, and flow experiments, the proposed methods are compared only against random designs. Since random designs can be arbitrarily far from the optimum, 'better than every one of 5,000 random draws' does not establish near-optimality. The paper should report where the selected design falls in the full distribution of random designs (for example, percentile relative to a much larger sample), and should compare against a stronger baseline such as a greedy or local-search method over the full structured search space. Without such evidence, the phrase 'near-optimal EIG designs' in Section 5 overstates what the numerics demonstrate.
  4. [Section 3.4] The computational cost analysis is useful but does not compensate for the missing approximation guarantee. In particular, the IterSelect cost in Table 3 contains a duplicated term '(niter+1)·C_EIG(K,N_rows) + (niter+1)·C_EIG(K,N_rows)'; even after correcting this typo, the cost model depends on the number of iterations n_iter, which is claimed to be at most 3 in all experiments but is not analyzed. Since IterSelect has no monotonicity guarantee, the stopping criterion may terminate at a local optimum or cycle, and the paper should discuss this behavior or present convergence evidence beyond the four experiments.
minor comments (5)
  1. [Section 3.3] The text says 'In Algorithm 2 we show how to combine the sketching with IndSelect,' but the pseudocode for Sketch-First-IndSelect is Algorithm 5; the reference should be corrected.
  2. [Table 3] The IterSelect row contains a duplicated cost term; the line should list only one '(niter+1)·C_EIG(K,N_rows)' term.
  3. [Section 4.1, Eq. (24)] The block-diagonal noise covariance is written as blockdiag(R_0,...,R_N) ∈ R^{n_s(J+1)×n_s(J+1)}; the last block should be R_J, not R_N, since the time index runs from 0 to J.
  4. [Article metadata] The MSC Classification (58F15, 58F17, 53C35) appears to be unrelated to numerical analysis or experimental design; the authors should replace it with appropriate MSC classes such as 65Fxx or 62Kxx.
  5. [Section 4.2, Table 4] The runtime comparison should be interpreted with care, as the text notes that the greedy implementations were not optimized; this caveat should also be reflected in the abstract's 'up to 50× speedup' claim, which is based on this comparison.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the EIG objective and benchmarks are external, and the multiway heuristic gap is explicitly admitted rather than disguised as a theorem.

full rationale

The paper's central claim—tensor-based structured CSSP methods produce effective (and in the d=1 case near-optimal by exhaustive comparison) EIG designs—is not a circular derivation. The EIG objective, phi_EIG(S)=logdet(I+AS(AS)^T) (Eq. 13), is defined independently of the algorithms, and no fitted parameter is used to construct it. IndSelect, SeqSelect, IterSelect, and the Sketch-First variants select indices via CSSP on mode unfoldings, and their outputs are evaluated against externally generated benchmarks: exhaustive enumeration in Section 4.1 and random designs in Sections 4.2 through 4.4. No equation in the paper equates the target EIG with the algorithms' internal selection criterion by construction, and no fitted 'prediction' is relabeled as a result. The authors do cite their own prior work [9] for Lemma 1 and for the spectral-preservation property of the sketch, but these are cited mathematical results used as motivation for GKS, not as inputs that define the structured selection output. The paper explicitly acknowledges the missing multiway analog: 'we were not able to derive analogs of Lemma 1 for the structured column selection.' That acknowledged gap is an unproved heuristic transfer, which is a correctness or support concern, not a circularity. The self-citations are real but not load-bearing in the sense of making the derivation equivalent to its inputs.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central algorithmic claims rest on standard linear algebra facts, on the unproved transfer of GKS-type column selection to multiway unfoldings, and on the randomized sketching assumption that Omega A preserves leading singular values. No new physical or mathematical entities are introduced.

free parameters (1)
  • Oversampling parameter p = 10
    Chosen by hand for all Sketch-First experiments in Section 4. Larger p improves sketch accuracy but increases cost; the paper does not vary p or report sensitivity to it.
assumptions (5)
  • standard math Sylvester's determinant lemma equates logdet(I+AA^T) and logdet(I+A^TA).
    Used in Section 2.2, Eq (8), to express EIG as logdet(I+AA^T).
  • domain assumption A is assumed to be generated with a diagonal noise covariance R=sigma_R^2 I, so each column of A corresponds to an individual design variable.
    Stated in Section 2.2 following [9]; non-diagonal R would mix design variables and break the column-selection interpretation.
  • ad hoc to paper The GKS column-selection heuristic, with its EIG bound for d=1 (Lemma 1), remains an effective selection rule when applied to mode unfoldings of the tensor.
    This is the key unproved transfer; Section 5 explicitly says no analogs of Lemma 1 were derived for the structured case.
  • domain assumption A Gaussian sketch Omega with r=K+p rows approximately preserves the leading singular values of A, so selection computed on Omega A transfers to A.
    Invoked in Section 3.3 and justified by reference to randomized linear algebra [37] and to [9, Section 4] for the d=1 case; no structured-case verification is provided.
  • domain assumption The product-structured selection operator S=S_d (tensor) ... (tensor) S_1 captures the intended experimental constraints.
    Assumed in Section 3.1; this restricts designs to Cartesian products of per-mode choices, which is the paper's definition of structured selection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Structured Column Subset Selection for Bayesian Optimal Experimental Design." pith.science (2026). https://pith.science/paper/QLF5UMV2

@misc{pith2026250600336,
  author       = {Pith},
  title        = {Pith review of: Structured Column Subset Selection for Bayesian Optimal Experimental Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QLF5UMV2}},
  note         = {Machine review of arXiv:2506.00336}
}
abstract

We consider optimal experimental design (OED) for Bayesian inverse problems, where the experimental design variables have a certain multiway structure. Given $d$ different experimental variables with $m_i$ choices per design variable $1 \le i\le d$, the goal is to select $k_i \le m_i$ experiments per design variable. Previous work has related OED to the column subset selection problem by mapping the design variables to the columns of a matrix $\mathbf{A}$. However, this approach is applicable only to the case $d=1$ in which the columns can be selected independently. We develop an extension to the case where the design variables have a multi-way structure. Our approach is to map the matrix $\mathbf{A}$ to a tensor and perform column subset selection on mode unfoldings of the tensor. We develop an algorithmic framework with three different algorithmic templates, and randomized variants of these algorithms. We analyze the computational cost of all the proposed algorithms and also develop greedy versions to facilitate comparisons. Numerical experiments on four different applications -- time-dependent inverse problems, seismic tomography, X-ray tomography, and flow reconstruction -- demonstrate the effectiveness and scalability of our methods for structured experimental design in Bayesian inverse problems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 27 canonical work pages

  1. [1]

    The European Physical Journal A57(2021) https://doi.org/10.1140/ epja/s10050-021-00382-2

    Melendez, J., Furnstahl, R., Griesshammer, H., McGovern, J., Phillips, D., Pra- tola, M.: Designing optimal experiments: an application to proton Compton scattering. The European Physical Journal A57(2021) https://doi.org/10.1140/ epja/s10050-021-00382-2

    work page 2021

  2. [2]

    Human Brain Mapping8(2-3), 109–114 (1999) https://doi.org/10.1002/(SICI)1097-0193(1999) 8:2/3x109::AID-HBM7y3.0.CO;2-W

    Dale, A.M.: Optimal Experimental Design for Event-Related fMRI. Human Brain Mapping8(2-3), 109–114 (1999) https://doi.org/10.1002/(SICI)1097-0193(1999) 8:2/3x109::AID-HBM7y3.0.CO;2-W

    work page doi:10.1002/(sici)1097-0193(1999 1999

  3. [3]

    Magnetic Resonance in Medicine88(1), 239–253 (2022) https://doi.org/10.1002/ mrm.29212 https://onlinelibrary.wiley.com/doi/pdf/10.1002/mrm.29212

    Scope Crafts, E., Lu, H., Ye, H., Wald, L.L., Zhao, B.: An efficient approach to optimal experimental design for magnetic resonance fingerprinting with B-splines. Magnetic Resonance in Medicine88(1), 239–253 (2022) https://doi.org/10.1002/ mrm.29212 https://onlinelibrary.wiley.com/doi/pdf/10.1002/mrm.29212

    work page doi:10.1002/mrm.29212 2022

  4. [4]

    Computational Geosciences20, 375–383 (2016)

    Seidler, R., Padalkina, K., B¨ ucker, H.M., Ebigbo, A., Herty, M., Marquart, G., Niederau, J.: Optimal experimental design for reservoir property estimates in geothermal exploration. Computational Geosciences20, 375–383 (2016)

    work page 2016

  5. [5]

    Geophysical Jour- nal International236(3), 1309–1331 (2023) https://doi.org/10.1093/gji/ggad492 https://academic.oup.com/gji/article-pdf/236/3/1309/55270901/ggad492.pdf

    Strutz, D., Curtis, A.: Variational Bayesian experimental design for geophys- ical applications: seismic source location, amplitude versus offset inversion, and estimating CO2 saturations in a subsurface reservoir. Geophysical Jour- nal International236(3), 1309–1331 (2023) https://doi.org/10.1093/gji/ggad492 https://academic.oup.com/gji/article-pdf/236/3...

    doi:10.1093/gji/ggad492 2023

  6. [6]

    Acta Numerica33, 715–840 (2024) https://doi.org/10.1017/ s0962492924000023

    Huan, X., Jagalur, J., Marzouk, Y.: Optimal experimental design: Formulations and computations. Acta Numerica33, 715–840 (2024) https://doi.org/10.1017/ s0962492924000023

    work page 2024

  7. [7]

    (eds.) The Bayesian Approach to Inverse Problems, pp

    Dashti, M., Stuart, A.M.: In: Ghanem, R., Higdon, D., Owhadi, H. (eds.) The Bayesian Approach to Inverse Problems, pp. 311–428. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-12385-1 7

    doi:10.1007/978-3-319-12385-1 2017

  8. [8]

    Inverse Problems37(4), 043001 (2021)

    Alexanderian, A.: Optimal experimental design for infinite-dimensional Bayesian inverse problems governed by PDEs: A review. Inverse Problems37(4), 043001 (2021)

    work page 2021

Show all 43 references
  1. [9]

    arXiv preprint arXiv:2402.16000 (2024)

    Eswar, S., Rao, V., Saibaba, A.K.: Bayesian D-optimal experimental designs via column subset selection. arXiv preprint arXiv:2402.16000 (2024)

    2024 arXiv

  2. [10]

    Statistical Science10(3), 273–304 (1995) https://doi.org/10.1214/ss/1177009939 28

    Chaloner, K., Verdinelli, I.: Bayesian Experimental Design: A Review. Statistical Science10(3), 273–304 (1995) https://doi.org/10.1214/ss/1177009939 28

    1995

  3. [11]

    Acta Numerica29, 403–572 (2020)

    Martinsson, P.-G., Tropp, J.A.: Randomized numerical linear algebra: Founda- tions and algorithms. Acta Numerica29, 403–572 (2020)

    2020

  4. [12]

    Theoretical Computer Science410(47-49), 4801– 4811 (2009)

    Civril, A., Magdon-Ismail, M.: On selecting a maximum volume sub-matrix of a matrix and related problems. Theoretical Computer Science410(47-49), 4801– 4811 (2009)

    2009

  5. [13]

    SIAM Journal on Scientific Computing17(4), 848–869 (1996) https://doi.org/10.1137/0917055

    Gu, M., Eisenstat, S.C.: Efficient algorithms for computing a strong rank- revealing QR factorization. SIAM Journal on Scientific Computing17(4), 848–869 (1996) https://doi.org/10.1137/0917055

    1996 doi

  6. [14]

    SIAM Journal on Matrix Analysis and Applications 37(3), 1223–1249 (2016)

    Saibaba, A.K.: HOID: Higher order interpolatory decomposition for tensors based on Tucker representation. SIAM Journal on Matrix Analysis and Applications 37(3), 1223–1249 (2016)

    2016

  7. [15]

    SIAM Journal on Mathemat- ics of Data Science2(1), 189–215 (2020) https://doi.org/10.1137/19M1261043 https://doi.org/10.1137/19M1261043

    Minster, R., Saibaba, A.K., Kilmer, M.E.: Randomized Algorithms for Low-Rank Tensor Decompositions in the Tucker Format. SIAM Journal on Mathemat- ics of Data Science2(1), 189–215 (2020) https://doi.org/10.1137/19M1261043 https://doi.org/10.1137/19M1261043

    2020 doi

  8. [16]

    Journal of Fluid Mechanics962(2023) https:// doi.org/10.1017/jfm.2023.269

    Farazmand, M., Saibaba, A.: Tensor-based flow reconstruction from optimally located sensor measurements. Journal of Fluid Mechanics962(2023) https:// doi.org/10.1017/jfm.2023.269

    2023 doi

  9. [17]

    The Annals of Statistics 38(4), 1978–2004 (2010) https://doi.org/10.1214/09-AOS778

    Huang, J., Zhang, T.: The benefit of group sparsity. The Annals of Statistics 38(4), 1978–2004 (2010) https://doi.org/10.1214/09-AOS778

    2010 doi

  10. [18]

    In: Proceedings of the 30th International Conference on Neural Information Processing Systems

    Wu, T., Benson, A.R., Gleich, D.F.: General tensor spectral co-clustering for higher-order data. In: Proceedings of the 30th International Conference on Neural Information Processing Systems. NIPS’16, pp. 2567–2575. Curran Associates Inc., Red Hook, NY, USA (2016)

    2016

  11. [19]

    Magnetic Resonance in Medicine58(6), 1182–1195 (2007) https://doi.org/10.1002/mrm.21391 https://onlinelibrary.wiley.com/doi/pdf/10.1002/mrm.21391

    Lustig, M., Donoho, D., Pauly, J.M.: Sparse MRI: The application of compressed sensing for rapid MR imaging. Magnetic Resonance in Medicine58(6), 1182–1195 (2007) https://doi.org/10.1002/mrm.21391 https://onlinelibrary.wiley.com/doi/pdf/10.1002/mrm.21391

    2007 doi

  12. [20]

    Investigative Radiology51(6), 349–364 (2016) https://doi.org/10.1097/RLI.0000000000000274

    Yang, A.C., Kretzler, M., Sudarski, S., Gulani, V., Seiberlich, N.: Sparse Recon- struction Techniques in Magnetic Resonance Imaging: Methods, Applications, and Challenges to Clinical Adoption. Investigative Radiology51(6), 349–364 (2016) https://doi.org/10.1097/RLI.0000000000000274

    2016 doi

  13. [21]

    SIAM review 51(3), 455–500 (2009)

    Kolda, T.G., Bader, B.W.: Tensor decompositions and applications. SIAM review 51(3), 455–500 (2009)

    2009

  14. [22]

    SIAM Journal on Matrix Analysis and Applications21(4), 1253– 1278 (2000)

    De Lathauwer, L., De Moor, B., Vandewalle, J.: A Multilinear Singular Value 29 Decomposition. SIAM Journal on Matrix Analysis and Applications21(4), 1253– 1278 (2000)

    2000

  15. [23]

    SIAM Journal on Sci- entific Computing34(2), 1027–1052 (2012) https://doi.org/10.1137/110836067

    Vannieuwenhoven, N., Vandebril, R., Meerbergen, K.: A New Truncation Strat- egy for the Higher-Order Singular Value Decomposition. SIAM Journal on Sci- entific Computing34(2), 1027–1052 (2012) https://doi.org/10.1137/110836067 . Accessed 2025-04-01

    2012 doi

  16. [24]

    ArXivabs/2010.16262(2020)

    Bakker, T., Hoof, H., Welling, M.: Experimental design for MRI by greedy policy search. ArXivabs/2010.16262(2020)

    2020 arXiv

  17. [25]

    In: 2020 59th IEEE Conference on Decision and Control (CDC), pp

    Rodrigues, D., Makrygiorgos, G., Mesbah, A.: Tractable Global Solutions to Bayesian Optimal Experiment Design. In: 2020 59th IEEE Conference on Decision and Control (CDC), pp. 1614–1619 (2020). https://doi.org/10.1109/ CDC42340.2020.9304226

    2020

  18. [26]

    SIAM Journal on Scientific Computing45(1), 57–77 (2023) https://doi.org/10.1137/21M1466542

    Wu, K., Chen, P., Ghattas, O.: An Offline-Online Decomposition Method for Efficient Linear Bayesian Goal-Oriented Optimal Experimental Design: Applica- tion to Optimal Sensor Placement. SIAM Journal on Scientific Computing45(1), 57–77 (2023) https://doi.org/10.1137/21M1466542

    2023 doi

  19. [27]

    Proceedings of the National Academy of Sciences106(3), 697–702 (2009)

    Mahoney, M.W., Drineas, P.: CUR matrix decompositions for improved data analysis. Proceedings of the National Academy of Sciences106(3), 697–702 (2009)

    2009

  20. [29]

    SIAM Jour- nal on Scientific Computing38(3), 1454–1482 (2016) https://doi.org/10.1137/ 140978430 https://doi.org/10.1137/140978430

    Sorensen, D.C., Embree, M.: A DEIM Induced CUR Factorization. SIAM Jour- nal on Scientific Computing38(3), 1454–1482 (2016) https://doi.org/10.1137/ 140978430 https://doi.org/10.1137/140978430

    2016 doi

  21. [30]

    Foundations and Trends®in Machine Learning3(2), 123–224 (2011) https://doi.org/10.1561/ 2200000035

    Mahoney, M.W.: Randomized Algorithms for Matrices and Data. Foundations and Trends®in Machine Learning3(2), 123–224 (2011) https://doi.org/10.1561/ 2200000035

    2011

  22. [31]

    In: Proceedings of the 33rd International Conference on Interna- tional Conference on Machine Learning - Volume 48

    Altschuler, J., Bhaskara, A., Fu, G., Mirrokni, V., Rostamizadeh, A., Zadi- moghaddam, M.: Greedy column subset selection: new bounds and distributed algorithms. In: Proceedings of the 33rd International Conference on Interna- tional Conference on Machine Learning - Volume 48....

  23. [32]

    SIAM Jour- nal on Scientific Computing39(4), 263–291 (2017) https://doi.org/10.1137/ 15M1044680 https://doi.org/10.1137/15M1044680 30

    Duersch, J.A., Gu, M.: Randomized QR with Column Pivoting. SIAM Jour- nal on Scientific Computing39(4), 263–291 (2017) https://doi.org/10.1137/ 15M1044680 https://doi.org/10.1137/15M1044680 30

    2017 doi

  24. [33]

    arXiv preprint arXiv:2410.03969 (2024)

    Epperly, E.N., Tropp, J.A., Webber, R.J.: Embrace rejection: Kernel matrix approximation by accelerated randomly pivoted Cholesky. arXiv preprint arXiv:2410.03969 (2024)

    2024 arXiv

  25. [34]

    SIAM Journal on Matrix Analysis and Applications21(4), 1324–1342 (2000) https://doi.org/10

    De Lathauwer, L., De Moor, B., Vandewalle, J.: On the Best Rank-1 and Rank- (R1,R 2,...,R N) Approximation of Higher-Order Tensors. SIAM Journal on Matrix Analysis and Applications21(4), 1324–1342 (2000) https://doi.org/10. 1137/S0895479898346995 https://doi.org/10.1137/S08954...

    2000 doi

  26. [35]

    In: Proceedings of the 2007 SIAM International Conference on Data Mining, pp

    Sheehan, B.N., Saad, Y.: Higher order orthogonal iteration of tensors (HOOI) and its relation to PCA and GLRAM. In: Proceedings of the 2007 SIAM International Conference on Data Mining, pp. 355–365 (2007). SIAM

    2007

  27. [36]

    SIAM journal on Optimization23(4), 2037–2060 (2013)

    Beck, A., Tetruashvili, L.: On the convergence of block coordinate descent type methods. SIAM journal on Optimization23(4), 2037–2060 (2013)

    2013

  28. [37]

    SIAM Review53(2), 217–288 (2011) https://doi.org/10.1137/090771806 https://doi.org/10.1137/090771806

    Halko, N., Martinsson, P.G., Tropp, J.A.: Finding Structure with Randomness: Probabilistic Algorithms for Constructing Approximate Matrix Decomposi- tions. SIAM Review53(2), 217–288 (2011) https://doi.org/10.1137/090771806 https://doi.org/10.1137/090771806

    2011 doi

  29. [38]

    Demmel, J.W.: Applied Numerical Linear Algebra, p. 419. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA. https://doi.org/10.1137/1. 9781611971446

    doi

  30. [39]

    GAMM- Mitteilungen43(3), 202000014 (2020) https://doi.org/10.1002/gamm.202000014

    Freitag, M.A.: Numerical linear algebra in data assimilation. GAMM- Mitteilungen43(3), 202000014 (2020) https://doi.org/10.1002/gamm.202000014

    2020 doi

  31. [40]

    arXiv preprint arXiv:2502.00150 (2025)

    Alexanderian, A., D´ ıaz, H., Rao, V., Saibaba, A.K.: Optimal sensor place- ment under model uncertainty in the weak-constraint 4D-Var framework. arXiv preprint arXiv:2502.00150 (2025)

    2025

  32. [41]

    Numerical Algorithms 81(3), 773–811 (2019)

    Gazzola, S., Hansen, P.C., Nagy, J.G.: IR Tools: a MATLAB package of itera- tive regularization methods and large-scale test problems. Numerical Algorithms 81(3), 773–811 (2019)

    2019

  33. [42]

    IEEE Transactions on Medical Imaging5, 170–176 (1986)

    Radon, J.: On the determination of functions from their integral values along certain manifolds. IEEE Transactions on Medical Imaging5, 170–176 (1986)

    1986

  34. [43]

    arXiv:2505.04004 (2025) arXiv:2505.04004 [math.NA]

    Kakasenko, L., Alexanderian, A., Farazmand, M., Saibaba, A.K.: Bridging the Gap Between Deterministic and Probabilistic Approaches to State Estimation. arXiv:2505.04004 (2025) arXiv:2505.04004 [math.NA]. arXiv preprint

    2025 arXiv

  35. [44]

    Journal of Atmo- spheric and Oceanic Technology21(10), 1575–1589 (2004) https://doi.org/10

    Popinet, S., Smith, M., Stevens, C.: Experimental and numerical study of the turbulence characteristics of airflow around a research vessel. Journal of Atmo- spheric and Oceanic Technology21(10), 1575–1589 (2004) https://doi.org/10. 1175/1520-0426(2004)021x1575:EANSOTy2.0.CO;2...

    2004

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.