Pith. sign in

REVIEW 5 major objections 6 minor 25 references

Structured Variational $D$-Decomposition for Accurate and Stable Low-Rank Approximation

T0 review · 5 major / 6 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 The paper claims that a regularized, optimally-fit three-factor product $A \approx P D Q$ achieves lower reconstruction error than truncated SVD, CUR, and NMF on sparse and noisy data, at a per-iteration cost of $O(n^2 k)$.

desk verdict The main empirical claim—that D-decomposition beats truncated SVD on dense full matrices—is impossible under Eckart-Young, and the paper's own Table 1 exposes an unstated or incorrect evaluation protocol. read the letter →

arxiv 2506.08535 v1 pith:QIXCYOYA submitted 2025-06-10 math.NA cs.LGcs.NA

Ronald Katende This is my paper
classification math.NAcs.LGcs.NA MSC 15A2365F3090C26
keywords D-decompositionlow-rankapproximationvariationalmatrixfactorizationalternatingminimizationregularizedoptimizationperturbationstabilityconditionnumbercontrolstructureddecomposition
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

The paper introduces the D-decomposition, a non-orthogonal matrix factorization $A \approx P D Q$ computed by minimizing a regularized Frobenius loss with alternating minimization. It aims to show that this variational factorization is more accurate than classical algebraic decompositions such as truncated SVD, CUR, and nonnegative matrix factorization, especially when the data matrix is sparse, noisy, or ill-conditioned. The authors also claim existence, perturbation stability, uniqueness up to permutation, and a per-iteration complexity of $O(n^2 k)$. If the claims hold, the method offers a practical alternative to SVD in regimes where structural constraints and conditioning matter as much as reconstruction accuracy.

What carries the argument

The central object is the regularized three-factor product $A \approx P D Q$, where $D$ is a non-diagonal core matrix allowed to encode conditioning and spectral structure. The workhorse is alternating minimization: each block update solves a regularized normal equation (for $P$, for $D$, and for $Q$), with the regularizers $\alpha_i\|\cdot\|_F^2$ and $\beta \log \kappa(D)$ keeping the subproblems positive definite and bounding the condition number of $D$. This machinery is what links reconstruction accuracy, numerical stability, and spectral energy alignment in a single variational objective.

What would settle it

Run truncated SVD at the same target rank $k$ on the same full dense matrices used in Table 4 and compare relative Frobenius error. By the Eckart-Young theorem, no rank-$k$ factor can beat truncated SVD on the full matrix, so any reported lower error for the D-decomposition would prove that the benchmark used a different preprocessing or evaluation protocol; checking whether the input was mean-imputed and whether error is computed on observed entries settles the claim.

Watch

Extended reading notes

Core claim

The paper proposes that the D-decomposition is a structured, optimization-based alternative to classical low-rank approximations. In its formulation, $P \in \mathbb{R}^{n \times k}$, $D \in \mathbb{R}^{k \times k}$, and $Q \in \mathbb{R}^{k \times n}$ are found by minimizing $\|A - P D Q\|_F^2 + \lambda R(P, D, Q)$, where $R$ enforces sparsity, bounded condition numbers, or other structural properties. The paper reports that this regularized three-factor decomposition achieves lower relative Frobenius reconstruction error than truncated SVD, CUR, and sparse PCA on MovieLens, gene expression, and MNIST covariance matrices, and that the log-condition penalty controls the conditioning of $D$ without degrading accuracy. It also proves that the optimization problem has a global minimizer under a coercive regularizer, that the core $D$ changes continuously under additive perturbations, that identifiability holds up to permutation, and that the alternating minimization updates are well-posed and converge to a stationary point.

Load-bearing premise

The comparison assumes that all methods are evaluated under the same rank-$k$ Frobenius reconstruction protocol on the same matrices; if the D-decomposition's lower error is instead measured on observed entries or mean-imputed data, the comparison with truncated SVD must be rerun under that exact protocol.

Editorial extensions

If this is right

  • If the D-decomposition is correct, regularized non-orthogonal three-factor minimization is a viable reconstruction method in noisy, sparse, and ill-conditioned regimes where algebraic factorizations are brittle.
  • The conditioning penalty $\beta \log \kappa(D)$ can stabilize the core matrix without sacrificing reconstruction accuracy, as shown by the ablation study where $\kappa(D)$ drops from $1.8\times 10^6$ to $34$ with error nearly unchanged.
  • The per-iteration cost of $O(n^2 k)$ with quadratic empirical scaling makes the method practical for moderate-scale problems where full SVD would be too expensive.
  • The variational framework allows direct incorporation of structural priors (sparsity, conditioning, energy localization) into the factorization itself, which classical algebraic decompositions do not support.

Reading between the lines

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

  • The reported improvements over truncated SVD on full-dense matrices conflict with the Eckart-Young optimality of truncated SVD for rank-$k$ Frobenius approximation, so the comparisons in Tables 1 and 4 must rely on an unstated alternative evaluation protocol, such as observed-entry masking or mean-imputed input; the abstract's claim is only supported under that protocol.
  • A testable extension is to hold out a random subset of entries, fit all methods on the observed part, and compare only the held-out error; this would separate the regularizer's imputation benefit from any pure approximation gain.
  • The core $D$ could be reused as a well-conditioned preconditioner or as a low-dimensional latent coordinate system for downstream tasks, since the theory guarantees $\kappa(D)$ is bounded when the log-condition penalty is active.
  • The claim that spectral energy concentrates in the dominant eigenspace (Example 3.3) is directly checkable on any dataset by computing $\sum_{i=1}^k \|u_i^\top P\|_2^2$ and comparing with the same quantity for the SVD basis.
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

5 major / 6 minor

Summary. This paper introduces the D-decomposition, a regularized variational low-rank factorization A ≈ P D Q with P ∈ R^{n×k}, D ∈ R^{k×k}, Q ∈ R^{k×n}, obtained by alternating minimization of a Frobenius loss plus a regularizer. The authors claim existence, perturbation stability, and identifiability results (Theorems 2.1–2.3), an O(n^2 k) per-iteration complexity, and empirical improvements over truncated SVD, CUR, and NMF on synthetic and real datasets. The central claim is that the method achieves lower reconstruction error than truncated SVD under a fixed-rank Frobenius protocol.

Significance. The idea of a learnable, regularized non-orthogonal factorization is potentially useful for applications such as imputation and ill-conditioned problems, and the paper provides a fairly detailed framework with several worked examples. However, the paper does not provide machine-checked proofs or executable code, and its central empirical claim contradicts the Eckart-Young theorem under the stated Frobenius-norm protocol. The theoretical results have serious gaps. If the method were correctly evaluated under a different protocol (e.g., missing-entry imputation), the contribution could be of interest, but as written the claims are not supported.

major comments (5)
  1. [3.1, Table 1] The reported relative Frobenius errors in Table 1 are impossible under the stated protocol. For the low-rank matrix A = U V^T with U,V ∈ R^{500×50} orthonormal, rank(A)=50, so the rank-50 truncated SVD has zero error to machine precision, yet the table reports 0.0021 for Truncated SVD and 0.0020 for D-Decomposition. Since any output P D Q has rank at most 50, by the Eckart-Young theorem no rank-50 approximation can have lower Frobenius error than truncated SVD. The positive SVD error indicates either a different evaluation protocol (e.g., missing entries), a different rank, or non-faithful reporting; the central empirical claim is unsupported as written.
  2. [3.2, Table 4] The same Eckart-Young contradiction applies to Table 4, where D-decomposition is reported to achieve lower relative Frobenius error than truncated SVD for rank k=50 on MovieLens, leukemia expression, and MNIST covariance matrices. For any matrix, the rank-k truncated SVD is the exact minimizer of the Frobenius error over all rank-k approximations. The reported improvements are therefore only possible under an unstated alternative protocol (such as evaluating on observed entries with missing data), which is not described in Section 3.1 or the table captions. The abstract's claim of lower reconstruction error is not justified.
  3. [2.1, Theorem 2.1(b)] The proof of Theorem 2.1(b) is not valid. It shows that the objective gap between the perturbed and unperturbed problems is O(ε), but it does not establish that the distance between specific minimizers, ||D* - D0||_F, is O(ε). Without strong convexity or some form of uniqueness, the set of minimizers may be large and the selected global minimizer can jump discontinuously under perturbation. The appeal to 'smooth sensitivity results in constrained Euclidean minimization' is made without verifying the required constraint qualifications, second-order sufficient conditions, or Lipschitz properties of the solution mapping. The claimed stability bound is therefore unproven.
  4. [2.2, Theorem 2.2] Theorem 2.2 is not established. The proof assumes that two global minimizers yield the same product B = P*D*Q* = \tilde P \tilde D \tilde Q, but this equality is not shown; the regularized objectives could have different products. Even if the products coincide, the invariance group of the factorization P D Q is all invertible T ∈ R^{k×k}, not just diagonal T. Condition (ii) only rules out non-identity diagonal T, leaving the possibility of non-diagonal transformations that change the regularization; the conclusion that the only ambiguity is permutation is therefore unsupported. The proof also incorrectly asserts that T must make R invariant, but condition (ii) does not imply that for non-diagonal T.
  5. [2.3, Theorem 2.3] Theorem 2.3(a) is circular: the bound on κ(D*) contains the constant C0 = α1||P*||_F^2 + α2||D*||_F^2 + α3||Q*||_F^2, which depends on the very minimizer being bounded. The proof of part (c) asserts an inequality δ ≤ c λ^{-1} R(P,D,Q)^{-1} without derivation, and the constant c is left unspecified. Additionally, Algorithm 1 in Appendix A implements only the quadratic regularizer with α_i and does not include the β log κ(D) term used in Theorem 2.3 and the ablation study of Section 3.4, so it is unclear what objective is actually solved in those experiments.
minor comments (6)
  1. [2, Algorithm 1] The update equations in Section 2 and Algorithm 1 omit the factor λ in the regularized normal equations (e.g., the P update should have P (DQQ^T D^T + λ α1 I_k) = A Q^T D^T), whereas Example 3.2 correctly includes λ. This inconsistency should be resolved.
  2. [3.2, Table 4] The abstract and conclusion mention comparisons with nonnegative matrix factorization, but NMF is not included in Table 1 or Table 4; the comparison is incomplete or the text should be amended.
  3. [2, Problem Formulation] The problem formulation assumes square matrices A ∈ R^{n×n}, but Table 4 uses rectangular matrices (943×1682, 72×5000) without explaining the adaptation of the method or the theoretical results.
  4. [3.1, Figure 2] There are typographical issues: 'FIgure 2' (capital I) and 'F robenius' in the Table 2 caption; these should be corrected.
  5. [1, Footnote 0] The code availability footnote states that the implementation 'will be made publicly available upon publication'; for reproducibility, the code should be provided at review time.
  6. [2.1, Theorem 2.1(c)] The claim of uniqueness 'up to scaling and permutation' is imprecise because the invariance group of the factorization includes all invertible T ∈ R^{k×k}; the normalization constraints that would remove this freedom are not fully specified.

Circularity Check

2 steps flagged · score 6.0 of 10

Two theoretical 'confirmations' are tautological: Theorem 2.3(a)'s bounding constant contains the very minimizer it bounds, and Theorem 2.3(c)'s energy-localization estimate has a free constant c, so the reported 98% alignment is consistent by construction.

  1. self definitional [Section 2.3, proof of Theorem 2.3(a)]
    "Letting C0 = α1∥P ∗∥2F + α2∥D∗∥2F + α3∥Q∗∥2F , this yields κ(D∗) ≤ exp(1/β(1/λ∥A∥2F + C0))."

    The claimed bound on κ(D*) has a constant C0 that contains α2∥D*∥²_F, i.e., the norm of the very matrix D* being bounded. The inequality is therefore an implicit identity in D*: once D* is computed, C0 can be set to the observed value, so the theorem imposes no a priori restriction on the condition number. Example 3.3's report that κ(D*) ≤ 20 is not a test of a prediction; the bound is satisfied by construction rather than by the log-condition regularizer.

  2. other [Section 2.3, proof of Theorem 2.3(c); Example 3.3]
    "∥(I − UkU⊤k )P ∥2F ≤ c/λR(P, D, Q), for some constant c, using coercivity of R and the fact that energy outside the top eigenspace contributes to the residual. ... We find P10 i=1 ηi ≥ 0.98, indicating that 98% of the energy is aligned with the top-10 eigenspace."

    The energy-localization inequality contains an unspecified constant c that is never derived, bounded, or related to A, k, λ, or the regularizer. Because c is free, any observed alignment level can be made to satisfy the inequality after the fact. Calling the reported 0.98 alignment a confirmation of Theorem 2.3(c) is therefore circular: the free constant is effectively fitted to the data, and the empirical result carries no independent predictive content.

full rationale

The paper defines the D-decomposition variationally and then validates its own theorems with examples built from the same definitions. The two strongest circular steps are in Theorem 2.3. Part (a) purports to bound κ(D*), but the constant C0 in the bound is defined using ∥D*∥²_F, so the inequality is a tautology that cannot be falsified and does not predict the observed conditioning. Part (c) asserts an energy-localization bound with an unspecified constant c; the subsequent Example 3.3 reports 98% alignment and calls it a confirmation, but because c is free, any alignment value would be consistent. These are not parameter-free predictions; they reduce to the existence of constants chosen after the fact. The benchmark claim that D-decomposition beats truncated SVD on the same full Frobenius protocol is also unsupported: in Table 1 the exact rank-50 matrix A = UV^T with k=50 should give truncated SVD error near zero, yet the table reports 0.0021, and the D-decomposition is reported as 0.0020, which contradicts Eckart-Young under the stated protocol. This is a correctness/protocol inconsistency rather than a circular derivation, so I have not counted it as a separate circular step, but it reinforces that the central reconstruction-accuracy claim is not established. There are no load-bearing self-citations; the cited stability and convergence results are external. Overall, the conditioning and localization 'confirmations' are circular by construction, giving a partial-circularity score of 6.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claims rest on fitted hyperparameters lambda, alpha, beta, and k; on a coercive-regularizer assumption; on an unstated benchmark protocol; and on a nonstandard symmetry assumption. None is independently supported by shipped code or data.

free parameters (4)
  • target rank k = 50 in most benchmarks; 10 in Example 3.3
    Chosen by hand for each experiment; all comparative claims depend on it.
  • regularization weight lambda = 1e-4 to 1e-1 in Table 9
    Tuned per experiment; the reported stability properties change with lambda.
  • factor weights alpha1, alpha2, alpha3 = 1e-3 in Tables 1 and 4
    Set by hand; Table 9 shows the condition number of D varies strongly with alpha.
  • condition penalty weight beta = 1e-2 in Example 3.3; 0 to 1e-1 in ablation
    Controls kappa(D); the paper's own ablation shows kappa(D) changes by over five orders of magnitude with beta.
assumptions (4)
  • ad hoc to paper The regularizer R is continuous, coercive, and enforces bounded condition numbers on D.
    Used in Theorem 2.1; the quadratic regularizer alone does not enforce bounded kappa(D), and the log-condition term is not part of the alternating updates in Algorithm 1.
  • domain assumption The benchmark results are evaluated as full-matrix rank-k Frobenius reconstruction on identical matrices.
    This is the natural reading of Tables 1 and 4; under this premise the reported SVD-beating numbers are impossible, so the paper silently depends on an unstated protocol.
  • domain assumption A smooth sensitivity result from Bonnans and Shapiro applies to nonunique minimizers of the non-convex objective.
    Invoked in the proof of Theorem 2.1(b) to pass from an O(epsilon) objective gap to an O(epsilon) bound on D; the cited work is not in the reference list and the required regularity is not established.
  • ad hoc to paper The full gauge freedom of the factorization is captured by P maps to P T, D maps to T^{-1} D T^{-1}, Q maps to T Q with a common invertible T.
    Theorem 2.2 uses this specific symmetry, but the actual freedom of a rank-k factorization is more general, so uniqueness up to permutation is not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Structured Variational $D$-Decomposition for Accurate and Stable Low-Rank Approximation." pith.science (2026). https://pith.science/paper/QIXCYOYA

@misc{pith2026250608535,
  author       = {Pith},
  title        = {Pith review of: Structured Variational $D$-Decomposition for Accurate and Stable Low-Rank Approximation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QIXCYOYA}},
  note         = {Machine review of arXiv:2506.08535}
}
abstract

We introduce the $D$-decomposition, a non-orthogonal matrix factorization of the form $A \approx P D Q$, where $P \in \mathbb{R}^{n \times k}$, $D \in \mathbb{R}^{k \times k}$, and $Q \in \mathbb{R}^{k \times n}$. The decomposition is defined variationally by minimizing a regularized Frobenius loss, allowing control over rank, sparsity, and conditioning. Unlike algebraic factorizations such as LU or SVD, it is computed by alternating minimization. We establish existence and perturbation stability of the solution and show that each update has complexity $\mathcal{O}(n^2k)$. Benchmarks against truncated SVD, CUR, and nonnegative matrix factorization show improved reconstruction accuracy on MovieLens, MNIST, Olivetti Faces, and gene expression matrices, particularly under sparsity and noise.

Figures

Figures reproduced from arXiv: 2506.08535 by the authors.

Figure 1
Figure 1. Structure of the D-decomposition: A ≈ P DQ, with regularization applied to P, D, and Q individually. 2.1 Existence, Stability, and Computational Guarantees To ensure that the proposed D-decomposition is both theoretically sound and computationally practical, we establish four foundational properties, that is, the existence of a minimizer, stability under additive perturbations, uniqueness under convex regularization… view at source ↗
Figure 2
Figure 2. Face reconstruction results on the Olivetti Faces dataset, comparing original images, mean [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

  1. [1]

    A maximally split and adaptive relaxed al- ternating direction method of multipliers for regularized extreme learning machines,

    Z. Wang, S. Huo, X. Xiong, K. Wang, and B. Liu, “A maximally split and adaptive relaxed al- ternating direction method of multipliers for regularized extreme learning machines,” Mathematics, vol. 11, no. 14, p. 3198, 2023

    work page 2023

  2. [2]

    Sparse pattern selection strategies for robust frobenius-norm minimization preconditioners in electromagnetism,

    B. Carpentieri, I. Duff, and L. Giraud, “Sparse pattern selection strategies for robust frobenius-norm minimization preconditioners in electromagnetism,” Numerical Linear Algebra with Applications , vol. 7, no. 7–8, pp. 667–685, 2000

    work page 2000

  3. [3]

    Generalized matrix spectral factorization with symmetry and construction of quasi-tight framelets over algebraic number fields,

    R. Lu, “Generalized matrix spectral factorization with symmetry and construction of quasi-tight framelets over algebraic number fields,” Mathematics, vol. 12, no. 6, p. 919, 2024. 14

    work page 2024

  4. [4]

    A penalized matrix decomposition, with applications to sparse principal components and canonical correlation analysis,

    D. M. Witten, R. Tibshirani, and T. Hastie, “A penalized matrix decomposition, with applications to sparse principal components and canonical correlation analysis,” Biostatistics, vol. 10, no. 3, pp. 515–534, 2009

    work page 2009

  5. [5]

    G. H. Golub and C. F. Van Loan, Matrix Computations , 4th ed. Baltimore, MD: Johns Hopkins University Press, 2013

    work page 2013

  6. [6]

    Nonconvex factorization and manifold formulations are almost equivalent in low-rank matrix optimization,

    Y. Luo, X. Li, and A. R. Zhang, “Nonconvex factorization and manifold formulations are almost equivalent in low-rank matrix optimization,” SIAM Journal on Optimization , vol. 33, no. 1, p. 123–149, 2023

    work page 2023

  7. [7]

    A Validation Approach to Over-parameterized Matrix and Image Recovery

    L. Ding, Z. Qin, L. Jiang, J. Zhou, and Z. Zhu, “A validation approach to over-parameterized matrix and image recovery,” 2024. [Online]. Available: https://arxiv.org/abs/2209.10675

    work page Pith review arXiv 2024

  8. [8]

    Why are big data matrices approximately low rank?

    M. Udell and A. Townsend, “Why are big data matrices approximately low rank?” SIAM Journal on Mathematics of Data Science , vol. 1, no. 1, pp. 144–160, 2019

    work page 2019

Show all 25 references
  1. [9]

    Rank-revealing qr factorizations and the singular value decomposition,

    T. F. Chan, “Rank-revealing qr factorizations and the singular value decomposition,” SIAM Journal on Scientific and Statistical Computing , vol. 9, no. 4, pp. 727–741, 1988

    1988

  2. [10]

    Guaranteed minimum-rank solutions of linear matrix equa- tions via nuclear norm minimization,

    B. Recht, M. Fazel, and P. A. Parrilo, “Guaranteed minimum-rank solutions of linear matrix equa- tions via nuclear norm minimization,” SIAM Review, vol. 52, no. 3, pp. 471–501, 2010

    2010

  3. [11]

    A deterministic cur matrix decomposition algorithm for improved large-scale data analysis,

    C. Boutsidis, M. W. Mahoney, and P. Drineas, “A deterministic cur matrix decomposition algorithm for improved large-scale data analysis,” SIAM Journal on Matrix Analysis and Applications , vol. 30, no. 3, pp. 1139–1178, 2008

    2008

  4. [12]

    Sparse non-negative matrix factorizations via alternating non-negativity- constrained least squares for microarray data analysis,

    H. Kim and H. Park, “Sparse non-negative matrix factorizations via alternating non-negativity- constrained least squares for microarray data analysis,” Bioinformatics, vol. 23, no. 13, pp. 1495– 1502, 2007

    2007

  5. [13]

    Cur matrix decompositions for improved data analysis,

    M. W. Mahoney and P. Drineas, “Cur matrix decompositions for improved data analysis,” Proceed- ings of the National Academy of Sciences , vol. 106, no. 3, pp. 697–702, 2009

    2009

  6. [14]

    A regularized matrix factorization approach to induce structured sparse–low–rank solutions in the eeg inverse problem,

    J. Montoya-Mart ´ ınez, A. Art´ es-Rodr ´ ıguez, M. Pontil, and L. K. Hansen, “A regularized matrix factorization approach to induce structured sparse–low–rank solutions in the eeg inverse problem,” EURASIP Journal on Advances in Signal Processing , vol. 2014, no. 1, p. 97, 2014

    2014

  7. [15]

    Matrix factorization techniques in machine learning, signal processing, and statistics,

    A. Cichocki, R. Zdunek, A. H. Phan, and S.-i. Amari, “Matrix factorization techniques in machine learning, signal processing, and statistics,” Mathematical Foundations of Computational Engineer- ing, vol. 11, no. 12, p. 2674, 2023

    2023

  8. [16]

    Sparse principal component analysis,

    H. Zou, T. Hastie, and R. Tibshirani, “Sparse principal component analysis,” Journal of Computa- tional and Graphical Statistics , vol. 15, no. 2, pp. 265–286, 2006

    2006

  9. [17]

    General tensor decomposition, moment matrices and applications,

    V. Kuleshov, S. Ermon, M. Pathak, and S. Ermon, “General tensor decomposition, moment matrices and applications,” Journal of Machine Learning Research , vol. 18, pp. 1–35, 2017

    2017

  10. [18]

    Low-rank matrix recovery via iteratively reweighted least squares min- imization,

    K. Mohan and M. Fazel, “Low-rank matrix recovery via iteratively reweighted least squares min- imization,” Advances in Neural Information Processing Systems (NeurIPS) , vol. 23, pp. 845–853, 2010

    2010

  11. [19]

    A framework of regularized low-rank matrix models for regression and classification,

    H.-H. Huang, F. Yu, X. Fan et al., “A framework of regularized low-rank matrix models for regression and classification,” Statistics and Computing , vol. 34, no. 10, p. 10, 2024

    2024

  12. [20]

    Accelerating nuclear-norm regularized low-rank matrix op- timization through burer–monteiro decomposition,

    C. Lee, L. Liang, T. Tang, and K. Toh, “Accelerating nuclear-norm regularized low-rank matrix op- timization through burer–monteiro decomposition,” Journal of Machine Learning Research, vol. 25, pp. 379:1–52, 2024, bM-Global algorithm escaping spurious minima, with automatic r...

    2024

  13. [21]

    Deep matrix factorization with noise-adaptive regularization,

    Y. Zhang, Y. Wu, and L. Yang, “Deep matrix factorization with noise-adaptive regularization,” Journal of Machine Learning Research , vol. 24, pp. 1–35, 2023

    2023

  14. [22]

    Variational bayesian matrix factorization with spectral priors,

    R. Chen, K. Gupta, and A. Lee, “Variational bayesian matrix factorization with spectral priors,” Neural Computation, vol. 36, no. 1, pp. 55–88, 2024. 15

    2024

  15. [23]

    Neural parameterized low-rank models for structured matrix approx- imation,

    H. Lee, Q. Tran, and V. Rao, “Neural parameterized low-rank models for structured matrix approx- imation,” in International Conference on Machine Learning , 2022, pp. 12 401–12 414

    2022

  16. [24]

    Bayesian nonparametric low-rank matrix estimation with hierar- chical priors,

    X. Wang, W. Liu, and Z. Feng, “Bayesian nonparametric low-rank matrix estimation with hierar- chical priors,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2023, in press

    2023

  17. [25]

    Efficient algorithms for regularized nonnegative scale-invariant low- rank approximation models,

    J. E. Cohen and V. Leplat, “Efficient algorithms for regularized nonnegative scale-invariant low- rank approximation models,” arXiv preprint arXiv:2403.18517 , 2024, majorization–minimization algorithm with convergence guarantees for ℓ(p) p -regularized low-rank approximations...

    2024 arXiv

Pith tools

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