Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Hadamard-Riemannian Optimization for Margin-Variance Ensemble

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read An ensemble loss that penalizes margin variance, optimized on the unit sphere via Riemannian gradients, is claimed to exceed the accuracy of 100-tree random forests with only 10 learners.

desk verdict A margin-variance ensemble idea spoiled by a false sphere-simplex equivalence and unconvincing experiments. read the letter →

arxiv 2509.10189 v1 pith:NL27BYJN submitted 2025-09-12 cs.LG

Zexu Jin This is my paper · ORCID
classification cs.LG MSC 68T0590C30
keywords ensemblelearningmargindistributionvarianceregularizationHadamardparameterizationRiemannianoptimizationsimplexconstraintgeneralizationoverfitting
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 proposes an ensemble learning method that, instead of maximizing only the average classification margin, jointly optimizes the average margin and its variance. To avoid costly projections onto the probability simplex, weights are reparameterized as the pointwise square of a vector on the unit sphere, and optimization is carried out with Riemannian gradient descent. On eight benchmark datasets the method reports higher test accuracy than standard margin-based ensembles and, with only 10 base learners, surpasses 100-tree random forests while training faster. The paper also shows the loss is Lipschitz continuous and convex in the margin space, and that the simplex-to-sphere reformulation preserves first- and second-order KKT conditions.

What carries the argument

The load-bearing mechanism is the Hadamard parameterization w = z ⊙ z, which maps the unit sphere ∥z∥=1 onto the probability simplex w≥0, Σw_i=1. Combined with Riemannian gradient descent on the sphere—projecting the Euclidean gradient onto the tangent space and retracting by normalization—this removes the O(n log n) simplex projection step. The margin is smoothed with log-sum-exp, and the loss combines negative expected margin with a variance term; the paper proves the loss is Lipschitz continuous and convex as a function of the margin vector, and states a theorem (Theorem 1) asserting that the simplex and sphere formulations have equivalent first- and second-order KKT conditions.

What would settle it

Take a two-class problem with two base classifiers and define the margin so that the optimal weight is on a simplex boundary (e.g., one classifier dominates). Run Riemannian gradient descent from an initialization that approaches this boundary; if it converges to a point with a zero coordinate that does not satisfy the simplex KKT conditions (e.g., the negative margin gradient still points along the boundary), the claimed equivalence fails. More directly, compare the set of KKT points of min f(w) over the simplex with those of min f(z⊙z) over the sphere for a simple quadratic f; any sphere cri

Watch

Extended reading notes

Core claim

The central claim is that minimizing L(W) = -1/n Σ m_i + λ Var(m) over ensemble weights, where m_i is the log-sum-exp smoothed margin of instance i, produces more accurate and less overfit ensembles than maximizing the expected margin alone. The paper argues this minimization can be performed efficiently by setting w = z ⊙ z (Hadamard parameterization) and running Riemannian gradient descent on the unit sphere ∥z∥=1, eliminating simplex projections. Experiments on eight datasets show the method achieves the highest test accuracy on most of them, with particularly strong results on high-dimensional data, and with only 10 learners exceeds the accuracy of 100-tree random forests.

Load-bearing premise

The proof that optimizing on the sphere is equivalent to optimizing on the simplex assumes that zero-weight boundary points impose no extra constraints in the sphere formulation, but simplex KKT conditions do constrain such points, so spurious sphere critical points could exist.

Editorial extensions

If this is right

  • Small ensembles of 10 learners can match or exceed the accuracy of much larger random forests, lowering inference cost.
  • Including margin variance in the loss reduces overfitting and improves generalization, especially on high-dimensional or noisy data.
  • The Hadamard reparameterization makes weight optimization faster by avoiding simplex projections, with reported speedups up to roughly 59%.
  • The method is data-efficient: competitive accuracy is achieved with limited training samples.
  • The regularization parameter λ controls the accuracy–generalization gap tradeoff, with a moderate value improving both.

Reading between the lines

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

  • The equivalence theorem's boundary case is delicate: when a coordinate z_i = 0, the sphere problem may have critical points that do not correspond to valid simplex KKT points, so the method's success may depend on the optimum being interior or on the retraction avoiding such spurious points.
  • The convexity result holds in margin space, not in weight space after the Hadamard parameterization, so the spherical optimization landscape may still contain local minima; a direct test would be to compare solutions from multiple random initializations.
  • The same Hadamard-plus-Riemannian trick could be applied to other simplex-constrained objectives (e.g., portfolio optimization, mixture weights) whenever the objective's gradient can be cheaply evaluated.
  • The paper uses a fixed train/test split and a single random seed; measuring variance across seeds and folds would test whether the reported accuracy gaps are stable.
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 proposes an ensemble-weight learning method that minimizes a loss combining the negative average margin and the margin variance over training samples. To avoid simplex projections, it reparameterizes the weights as w = z ⊙ z on the unit sphere and applies Riemannian gradient descent. The paper claims three theoretical results (simplex-sphere KKT equivalence, Lipschitz continuity, and convexity in margin space) and reports experiments on eight benchmark datasets in which the proposed ensemble allegedly outperforms random forests, SVM, XGBoost, and LightGBM with shorter runtimes. I find that the central optimization theorem is false as stated, the convexity claim does not apply to the actual optimization landscape, the Lipschitz theorem is internally inconsistent, and the headline experimental comparison selects the regularization parameter using test accuracy. These issues undermine the paper's main claims.

Significance. The margin-variance loss and the Hadamard-sphere reparameterization are reasonable ideas, and the runtime comparison in Table II suggests potential practical efficiency. If the theory were correct and the experiments were properly controlled, the paper could make a useful contribution to margin-based ensemble learning. However, the paper's theoretical foundation fails exactly where it is needed: the claimed equivalence between the simplex and sphere KKT systems is false at boundary points, so the optimization guarantee for Algorithm 1 is unsupported. The empirical claim of state-of-the-art accuracy is also compromised by test-set-based model selection. The paper contains no code, no significance tests, and no reproducible experimental protocol. In its current form the contribution is not established.

major comments (4)
  1. [Section III, Theorem 1 (Eqs. (7)-(9))] The first-order equivalence is false at boundary points. For z_i=0, sphere stationarity 2z_i(∇f_i(w)+γ)=0 is vacuous, whereas simplex KKT requires μ_i=∇f_i(w)+λ≥0 when w_i=0 (Eq. (8)). Concrete counterexample: m=3, f(w)=-(5w_1+2w_2+2w_3), z=(0,1/√2,1/√2), so w=(0,1/2,1/2). With γ=2, the sphere KKT condition holds at this z. But the corresponding simplex KKT point would require λ=2 from the active coordinates, giving μ_1=-3<0, violating Eq. (8). Thus a sphere critical point need not project to a simplex KKT point. Since Algorithm 1 and the optimization rationale in Section III rely on this equivalence, the paper does not establish that Riemannian gradient descent solves the stated margin-variance problem. The proof's assertion that the sphere imposes no additional condition when w_i=0 is exactly the error; the theorem needs a correct treatment of boundary complementarity and second-order
  2. [Theorem 2, Eq. (13) vs. proof] The Lipschitz constant stated in Eq. (13) is M_g(1+√c)(1+4λM_m), but the proof concludes with L=3M_g(1+4λM_m) in Eq. (20). The factor (1+√c) is never derived. Moreover, the bound ∥G_i^T∥_2 ≤ M_g used before Eq. (17) does not follow from the column-wise assumption |g_k(x_i)|_2 ≤ M_g; the operator norm can be as large as √m M_g or √c M_g depending on the matrix dimensions. The theorem is therefore not proved as stated. This is a formal gap in the theoretical claims, even if less central than Theorem 1.
  3. [Section III, Theorem 3] The proof shows that L is convex as a function of the margin vector m, but this does not imply a benign landscape in the weight space W. The margin m_i in Eq. (4) is a concave (not affine) function of W because of the negative log-sum-exp term. Convexity is not preserved under composition with a concave map, so the statement that there is 'only a single global minimum without any local minima' does not follow. Indeed, the counterexample to Theorem 1 shows that the sphere problem can have stationary points that are not KKT points of the simplex problem. This claim should be corrected or removed.
  4. [Section IV-A and Table I] The 'Our-Bestλ' column selects λ per dataset, but no validation split is described; the text only mentions an 80:20 train-test split. If λ is chosen by maximizing test accuracy, the comparison is biased and the claim that the method 'consistently outperforms' baselines is not supported. The λ=0 row is more honest but does not by itself demonstrate the benefit of margin variance. In addition, the number of base learners is not specified in the experimental setup (the conclusion mentions 10 learners without a corresponding table or paragraph), no standard deviations or repeated splits are reported, and the baseline tuning procedure is not described.
minor comments (5)
  1. [Eq. (2)] The max in Eq. (2) lacks a subscript j; as written it is ambiguous which dimension is maximized.
  2. [Section III, complexity claim] The text says simplex projection has O(n log n) complexity, but projection onto the m-dimensional simplex costs O(m log m), where m is the number of learners, not the number of samples n. The algorithmic complexity statements should use consistent notation.
  3. [Theorem 2 proof] Several bounds in the proof are loose or unexplained, e.g., ∥∇s_i m_i∥_2 ≤ 3 in Eq. (16). Constants involving c and m should be made explicit, and the proof should match the statement.
  4. [References] Reference [26] already studies the Hadamard parametrization from simplex to sphere. The paper should position its contribution relative to that work and avoid re-deriving known facts without the correct conditions.
  5. [Figures] Figure 1 shows qualitative decision boundaries but is not connected to a quantitative claim. Figure 2 and Figure 3 would benefit from error bars or repeated runs to support the generalization and sensitivity statements.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the derivation of the loss, the Hadamard reparameterization, and the Riemannian optimization is self-contained. The main flagged issues are a test-set-selected hyperparameter and a flawed boundary argument in Theorem 1, both soundness/correctness concerns rather than circular reductions.

full rationale

The paper's mathematical derivation is not circular. The loss function L(W) in Eq. (5) is defined directly from the smoothed margin m_i, and the variance term is part of the objective rather than a conclusion imported into the optimization. The Hadamard reparameterization w = z⊙z is a genuine change of variables: for ||z||=1, w lies in the probability simplex, so the sphere problem is a reparameterization of the same objective and not an assumption of the desired result. The convergence- and regularity-related theorems (Lipschitz continuity, convexity in margin space) are proven from the explicit definitions rather than assumed. The paper does not rely on load-bearing self-citations; the cited Hadamard-parametrization and KKT-manifold results [26]-[28] are external prior work. The two substantive concerns are non-circular. First, Table I reports 'Our-Bestλ' per dataset, and the setup mentions only an 80:20 train-test split with no validation set; if λ was selected using test accuracy, the comparison between 'Our-Bestλ' and 'Our-λ=0' is a selection-bias artifact rather than independent evidence that margin variance improves generalization. This is an empirical soundness issue, not a definitional circularity. Second, Theorem 1's proof omits the simplex complementarity condition μ_i ≥ 0 on boundary coordinates: the statement 'When w_i=0: The sphere problem imposes no additional condition' ignores that the simplex KKT conditions require ∇f(w)_i + λ ≥ 0 at w_i=0. The sphere problem can therefore admit spurious boundary critical points, so the claimed equivalence is mathematically questionable. That is a correctness/proof-gap issue, not a case where a conclusion is equivalent to its input by construction. Weighing these, the derivation chain itself is self-contained, so the circularity score is low.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

No invented entities: the Hadamard parametrization is a cited prior technique, not a new postulate. The load-bearing assumptions are five hand-chosen or fitted numerical settings (notably per-dataset test-tuned lambda) and four structural premises, three of which are ad hoc to this paper: the false KKT equivalence, the non-sequitur landscape claim, and the unjustified operator-norm bound.

free parameters (5)
  • lambda, variance penalty weight = Per dataset on test accuracy: BASEHOCK 500, Breast 0.1, Chess 40, MNIST 20, Jaffe 0.01, Pathbased 1.0, RELATHE 50, Wine
    The central knob behind the 'variance helps' claim; selected per dataset to maximize test accuracy, so reported results encode test-set information.
  • alpha, log-sum-exp temperature = Not reported
    Required by the smoothed margin in Eqs. 3-4 but never specified in the experiments.
  • learning rate schedule {alpha_t} = Not reported
    Algorithm 1 requires a learning rate sequence for convergence; none is given.
  • number of base learners m = 10 (stated in conclusion only)
    The ensemble size is a hand-chosen design decision; experimental setup never states it.
  • tree depth for baselines and ensemble = 7
    Set 'for fair comparison', but SVM has no depth parameter, so the rule is not well-defined across baselines.
assumptions (7)
  • standard math Log-sum-exp is a valid smoothing of the max function for the margin (Eq. 3)
    Standard smooth approximation; invoked in Eq. 4.
  • standard math Cauchy-Schwarz implies PSD-ness of the variance Hessian (Eq. 23)
    Correct elementary computation backing Theorem 3.
  • domain assumption Classifiers output real-valued score vectors g_k(x) and the ensemble decision is argmax of weighted scores
    Assumed throughout Section II, Eqs. 1-2; never stated as a restriction on base model output type.
  • domain assumption Labels are one-hot encoded and the problem is c-class
    Used in the margin definition, Eq. 2.
  • ad hoc to paper The simplex and sphere KKT systems are equivalent including boundary points z_i = 0
    Theorem 1, Section III Eqs. 6-11: the proof ignores complementarity (mu_i >= 0, mu_i w_i = 0) at zero coordinates, and the equivalence is false as stated due to spurious critical points of the sphere problem.
  • ad hoc to paper Convexity of the loss in margin space implies a benign landscape for gradient descent in W
    Theorem 3, Section III: margins are a non-convex function of W, so convexity in m does not rule out local minima in the actual optimization variable; the paper asserts this inference without proof.
  • ad hoc to paper Column-wise score bounds |g_k(x_i)|_2 <= M_g imply the operator norm bound ||G_i^T||_2 <= M_g
    Theorem 2 proof, Section III: the bound on ||grad_W m_i|| uses ||G_i^T||_2 <= M_g, which does not follow from the stated assumption; the correct bound involves sqrt(m) M_g or row norms.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hadamard-Riemannian Optimization for Margin-Variance Ensemble." pith.science (2026). https://pith.science/paper/NL27BYJN

@misc{pith2026250910189,
  author       = {Pith},
  title        = {Pith review of: Hadamard-Riemannian Optimization for Margin-Variance Ensemble},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NL27BYJN}},
  note         = {Machine review of arXiv:2509.10189}
}
read the original abstract

Ensemble learning has been widely recognized as a pivotal technique for boosting predictive performance by combining multiple base models. Nevertheless, conventional margin-based ensemble methods predominantly focus on maximizing the expected margin while neglecting the critical role of margin variance, which inherently restricts the generalization capability of the model and heightens its vulnerability to overfitting, particularly in noisy or imbalanced datasets. Additionally, the conventional approach of optimizing ensemble weights within the probability simplex often introduces computational inefficiency and scalability challenges, complicating its application to large-scale problems. To tackle these limitations, this paper introduces a novel ensemble learning framework that explicitly incorporates margin variance into the loss function. Our method jointly optimizes the negative expected margin and its variance, leading to enhanced robustness and improved generalization performance. Moreover, by reparameterizing the ensemble weights onto the unit sphere, we substantially simplify the optimization process and improve computational efficiency. Extensive experiments conducted on multiple benchmark datasets demonstrate that the proposed approach consistently outperforms traditional margin-based ensemble techniques, underscoring its effectiveness and practical utility.

Figures

Figures reproduced from arXiv: 2509.10189 by the authors.

Figure 1
Figure 1. Decision boundaries on Pathbased dataset [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Sensitivity analysis of regularization parameter [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 3 linked inside Pith

  1. [1]

    A survey on ensemble learning,

    X. Dong, Z. Yu, W. Cao, Y . Shi, and Q. Ma, “A survey on ensemble learning,”Frontiers of Computer Science, vol. 14, no. 2, pp. 241–258, Apr. 2020

    2020

  2. [2]

    Polikar,Ensemble Learning

    R. Polikar,Ensemble Learning. New York, NY: Springer New York, 2012, pp. 1–34

    2012

  3. [3]

    On diversity and accuracy of homogeneous and heterogeneous ensembles,

    S. Bian and W. Wang, “On diversity and accuracy of homogeneous and heterogeneous ensembles,”International Journal of Hybrid Intelligent Systems, vol. 4, no. 2, pp. 103–128, Jun. 2007

    2007

  4. [4]

    Random Forests,

    L. Breiman, “Random Forests,”Machine Learning, vol. 45, no. 1, pp. 5–32, Oct. 2001

    2001

  5. [5]

    Cutler, D

    A. Cutler, D. R. Cutler, and J. R. Stevens,Random Forests. New York, NY: Springer New York, 2012, pp. 157–175

    2012

  6. [6]

    A comparison of random forest variable selection methods for classification prediction modeling,

    J. L. Speiser, M. E. Miller, J. Tooze, and E. Ip, “A comparison of random forest variable selection methods for classification prediction modeling,” Expert Systems with Applications, vol. 134, pp. 93–101, 2019

    2019

  7. [7]

    A comparative analysis of gradient boosting algorithms,

    C. Bent ´ejac, A. Cs ¨org˝o, and G. Mart ´ınez-Mu˜noz, “A comparative analysis of gradient boosting algorithms,”Artificial Intelligence Review, vol. 54, no. 3, pp. 1937–1967, Mar. 2021

    1937

  8. [8]

    Acceler- ating gradient boosting machines,

    H. Lu, S. P. Karimireddy, N. Ponomareva, and V . Mirrokni, “Acceler- ating gradient boosting machines,” inProceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, ser. Proceedings of Machine Learning Research, S. Chiappa and R. Calandra, Eds., vol. 108. PMLR, 26–28 Aug 2020, pp. 516–526

    2020

Show all 42 references
  1. [9]

    A margin-maximizing fine- grained ensemble method,

    J. Yuan, H. Chen, R. Luo, and F. Nie, “A margin-maximizing fine- grained ensemble method,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  2. [10]

    Variance-reduced methods for machine learning,

    R. M. Gower, M. Schmidt, F. Bach, and P. Richt ´arik, “Variance-reduced methods for machine learning,”Proceedings of the IEEE, vol. 108, no. 11, pp. 1968–1983, 2020

    1968

  3. [11]

    Achieving more with less: A tensor-optimization-powered ensemble method,

    J. Yuan, W. Jiang, Z. Cao, F. Xie, R. Wang, F. Nie, and Y . Yuan, “Achieving more with less: A tensor-optimization-powered ensemble method,”arXiv preprint arXiv:2408.02936, 2024

    2024 arXiv

  4. [12]

    Large dimensional analysis of general margin based classification methods,

    H. Huang and Q. Yang, “Large dimensional analysis of general margin based classification methods,”Journal of Statistical Mechanics: Theory and Experiment, vol. 2021, no. 11, p. 113401, Nov. 2021

    2021

  5. [13]

    Margin Allocation and Tradeoff in Complex Systems Design and Optimization,

    M. D. Guenov, X. Chen, A. Molina-Crist ´obal, A. Riaz, A. S. J. Van Heerden, and M. Padulo, “Margin Allocation and Tradeoff in Complex Systems Design and Optimization,”AIAA Journal, vol. 56, no. 7, pp. 2887–2902, Jul. 2018

    2018

  6. [14]

    Max-margin action prediction machine,

    Y . Kong and Y . Fu, “Max-margin action prediction machine,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 38, no. 9, pp. 1844–1858, 2016

    2016

  7. [15]

    Margin maximization for robust classifi- cation using deep learning,

    A. Matyasko and L.-P. Chau, “Margin maximization for robust classifi- cation using deep learning,” in2017 International Joint Conference on Neural Networks (IJCNN), 2017, pp. 300–307

    2017

  8. [16]

    Maximizing margin quality and quantity,

    Y . Bei and P. Hong, “Maximizing margin quality and quantity,” in2015 IEEE 25th International Workshop on Machine Learning for Signal Processing (MLSP), 2015, pp. 1–6

    2015

  9. [17]

    An Overview of Overfitting and its Solutions,

    X. Ying, “An Overview of Overfitting and its Solutions,”Journal of Physics: Conference Series, vol. 1168, p. 022022, Feb. 2019

    2019

  10. [18]

    Projection onto the probability simplex: An efficient algorithm with a simple proof, and an application,

    W. Wang and M. A. Carreira-Perpi ˜n´an, “Projection onto the probability simplex: An efficient algorithm with a simple proof, and an application,” 2013

    2013

  11. [19]

    l 0 gradient projection,

    S. Ono, “l 0 gradient projection,”IEEE Transactions on Image Process- ing, vol. 26, no. 4, pp. 1554–1564, 2017

    2017

  12. [20]

    A survey on large- scale machine learning,

    M. Wang, W. Fu, X. He, S. Hao, and X. Wu, “A survey on large- scale machine learning,”IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 6, pp. 2574–2594, 2022

    2022

  13. [21]

    Identifying mislabeled data using the area under the margin ranking,

    G. Pleiss, T. Zhang, E. Elenberg, and K. Q. Weinberger, “Identifying mislabeled data using the area under the margin ranking,” inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33. Curran Associates, I...

    2020

  14. [22]

    Log-sum-exp optimization based on continu- ous piecewise linearization techniques,

    X. Xi, J. Xu, and Y . Lou, “Log-sum-exp optimization based on continu- ous piecewise linearization techniques,” in2020 IEEE 16th International Conference on Control & Automation (ICCA), 2020, pp. 600–605

    2020

  15. [23]

    Cooperation of experts: Fusing heterogeneous information with large margin,

    S. Wang, S. Huang, J. Yuan, Z. Shen, and Z. Kang, “Cooperation of experts: Fusing heterogeneous information with large margin,”arXiv preprint arXiv:2505.20853, 2025

    2025 arXiv

  16. [24]

    A mean-variance optimization algorithm,

    I. Erlich, G. K. Venayagamoorthy, and N. Worawat, “A mean-variance optimization algorithm,” inIEEE Congress on Evolutionary Computa- tion, 2010, pp. 1–6

    2010

  17. [25]

    On three concepts in robust design optimization: absolute robustness, relative robustness, and less variance,

    Y . Kanno, “On three concepts in robust design optimization: absolute robustness, relative robustness, and less variance,”Structural and Mul- tidisciplinary Optimization, vol. 62, no. 2, pp. 979–1000, Aug. 2020

    2020

  18. [26]

    From the simplex to the sphere: faster constrained optimization using the Hadamard parametrization,

    Q. Li, D. McKenzie, and W. Yin, “From the simplex to the sphere: faster constrained optimization using the Hadamard parametrization,” Information and Inference: A Journal of the IMA, vol. 12, no. 3, pp. 1898–1937, Apr. 2023

    1937

  19. [27]

    The KKT optimality conditions for optimization problem with interval-valued objective function on Hadamard manifolds,

    S.-l. Chen, “The KKT optimality conditions for optimization problem with interval-valued objective function on Hadamard manifolds,”Opti- mization, vol. 71, no. 3, pp. 613–632, Mar. 2022

    2022

  20. [28]

    Intrinsic Formulation of KKT Conditions and Constraint Qualifications on Smooth Manifolds,

    R. Bergmann and R. Herzog, “Intrinsic Formulation of KKT Conditions and Constraint Qualifications on Smooth Manifolds,”SIAM Journal on Optimization, vol. 29, no. 4, pp. 2423–2444, Jan. 2019

    2019

  21. [29]

    Optimization Techniques on Riemannian Manifolds,

    S. T. Smith, “Optimization Techniques on Riemannian Manifolds,” 2014

    2014

  22. [30]

    Riemannian optimization on relaxed indicator matrix manifold,

    J. Yuan, F. Xie, F. Nie, and X. Li, “Riemannian optimization on relaxed indicator matrix manifold,”arXiv preprint arXiv:2503.20505, 2025

    2025 arXiv

  23. [31]

    Projection-like Retractions on Matrix Mani- folds,

    P.-A. Absil and J. Malick, “Projection-like Retractions on Matrix Mani- folds,”SIAM Journal on Optimization, vol. 22, no. 1, pp. 135–158, Jan. 2012

    2012

  24. [32]

    Global optimization of Lipschitz func- tions,

    C. Malherbe and N. Vayatis, “Global optimization of Lipschitz func- tions,” inProceedings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, D. Precup and Y . W. Teh, Eds., vol. 70. PMLR, 06–11 Aug 2017, pp. 2314– 2323

    2017

  25. [33]

    Constrained, Global Optimization of Unknown Functions with Lipschitz Continuous Gradients,

    A. P. Vinod, A. Israel, and U. Topcu, “Constrained, Global Optimization of Unknown Functions with Lipschitz Continuous Gradients,”SIAM Journal on Optimization, vol. 32, no. 2, pp. 1239–1264, Jun. 2022

    2022

  26. [34]

    Large margin classifiers based on convex class models,

    H. Cevikalp and B. Triggs, “Large margin classifiers based on convex class models,” in2009 IEEE 12th International Conference on Computer Vision Workshops, ICCV Workshops, 2009, pp. 101–108

    2009

  27. [35]

    Path-based clustering for grouping of smooth curves and texture segmentation,

    B. Fischer and J. Buhmann, “Path-based clustering for grouping of smooth curves and texture segmentation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 4, pp. 513–518, 2003

    2003

  28. [36]

    A Generalized Deep Learning Clustering Algorithm Based on Non- Negative Matrix Factorization,

    D. Wang, T. Li, P. Deng, F. Zhang, W. Huang, P. Zhang, and J. Liu, “A Generalized Deep Learning Clustering Algorithm Based on Non- Negative Matrix Factorization,”ACM Transactions on Knowledge Dis- covery from Data, vol. 17, no. 7, pp. 1–20, Aug. 2023

    2023

  29. [37]

    Addressing challenging problems using optimized deep learning classification algorithms on the mnist dataset,

    O. S. Salman and A. S. Salman, “Addressing challenging problems using optimized deep learning classification algorithms on the mnist dataset,” inAdvances in Information and Communication, K. Arai, Ed. Cham: Springer International Publishing, 2022, pp. 247–260

    2022

  30. [38]

    Face recognition under varying expressions and illumination using particle swarm opti- mization,

    S. A. Khan, M. Ishtiaq, M. Nazir, and M. Shaheen, “Face recognition under varying expressions and illumination using particle swarm opti- mization,”Journal of Computational Science, vol. 28, pp. 94–100, 2018

    2018

  31. [39]

    M. J. Wooldridge, J. Dy, S. Natarajan, and Association for the Advance- ment of Artificial Intelligence, Eds.,Thirty-Eighth AAAI Conference on Artificial Intelligence, Thirty-Sixth Conference on Innovative Applica- tions of Artificial Intelligence, Fourteenth Symposium on Educ...

    2024

  32. [40]

    Washington, DC, USA: AAAI Press, 2024

    ——,Thirty-Eighth AAAI Conference on Artificial Intelligence, Thirty- Sixth Conference on Innovative Applications of Artificial Intelligence, Fourteenth Symposium on Educational Advances in Artificial Intelli- gence: February 20-27, 2024, Vancouver, Canada. Washington, DC, USA:...

    2024

  33. [41]

    A comprehensive evaluation of various sensitivity analysis methods: A case study with a hydrological model,

    Y . Gan, Q. Duan, W. Gong, C. Tong, Y . Sun, W. Chu, A. Ye, C. Miao, and Z. Di, “A comprehensive evaluation of various sensitivity analysis methods: A case study with a hydrological model,”Environmental Modelling & Software, vol. 51, pp. 269–285, 2014

    2014

  34. [42]

    Over- fitting, Model Tuning, and Evaluation of Prediction Performance,

    O. A. Montesinos L ´opez, A. Montesinos L ´opez, and J. Crossa, “Over- fitting, Model Tuning, and Evaluation of Prediction Performance,” in Multivariate Statistical Machine Learning Methods for Genomic Pre- diction. Cham: Springer International Publishing, 2022, pp. 109–139

    2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.