Pith. sign in

REVIEW 3 major objections 4 minor 55 references

Theory of Decentralized Robust Kernel-Based Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Decentralized robust kernel learning reaches the centralized minimax rate at every node.

desk verdict Solid, technically demanding extension of kernel learning theory to decentralized robust regression; the central rates look right, though a couple of flagged assumptions need tightening before publication. read the letter →

arxiv 2506.05215 v2 pith:24DBYEP7 submitted 2025-06-05 cs.LG

classification cs.LG MSC 68T0562G0846E22
keywords decentralizedlearningrobustregressionreproducingkernelHilbertspacegradientdescentminimaxoptimalratesconsensusspectralgaploss
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 a decentralized version of kernel-based gradient descent for robust regression, where each agent updates a local estimator with a robust loss and then averages with neighbors through a doubly stochastic consensus matrix. The aim is to show that no central server is needed: each local estimator, not just a global average, can approximate the regression function. Under balanced local sample sizes and suitable choices of step size, communication rounds, and robustness scaling, the paper proves that every node's estimator achieves the same optimal rate (up to logarithmic factors) as the centralized least-squares kernel gradient descent algorithm, in both the mean-square and RKHS norms. The significance is that this removes the single point of failure and synchronization bottleneck of divide-and-conquer distributed learning, while retaining full statistical optimality.

What carries the argument

The engine is a three-term error decomposition: $f_{t+1,D_u}-\hat f_{t+1,D}=T_{1,t}+T_{2,t}+T_{3,t}$, where $\hat f_{t,D}$ is the centralized kernel gradient descent sequence. $T_{1,t}$ measures the deviation from consensus, controlled by the mixing inequality $\sum_v \left|[M^k]_{uv}-1/m\right|\le 2(\sqrt{m}\,\gamma_M^k\wedge 1)$ for the doubly stochastic communication matrix; $T_{2,t}$ captures the difference between local empirical integral operators and the true operator; $T_{3,t}$ accounts for the bias introduced by the robust windowing function $W$ and the scaling parameter $\sigma$. The auxiliary data-free sequence $\tilde f_t$ and refined estimates for the operators $(I-\alpha W'_+(0)L_{K,D_v})$ carry the argument, and the effective dimension condition $\mathcal{N}(\lambda)\le C_0\lambda^{-s}$ together with the regularity assumption $f_\rho=L_K^r g_\rho$ fix the achievable rates.

What would settle it

Run the algorithm on a two-node network where one node holds 90% of the data and the other 10%, with the same step size and $\sigma$ as in Theorem 2, and check whether both nodes still converge at the predicted $|D|^{-r/(2r+s)}$ rate. If they do, the balanced-size assumption is unnecessary; if the smaller node's rate degrades, the assumption is load-bearing.

Watch

Extended reading notes

Core claim

The central discovery is that decentralization costs nothing in statistical rate. The main theorem (Theorem 2) states that if each node holds $|D_u|=|D|/m=n$ samples, the number of iterations is $t=|D|^{1/(2r+s)}$, the consensus averaging window is $\bar{t}$ proportional to $\log|D|/(1-\gamma_M)$, and the robustness parameter $\sigma$ satisfies (2.8), then with high probability each local estimator obeys $\|f_{t,D_u}-f_\rho\|_{L^2_{\rho_X}} \lesssim_\delta (\log 512/\delta)^{4\vee(2p+2)} |D|^{-r/(2r+s)}$. The RKHS analogue (Theorem 5) gives the corresponding rate $|D|^{-(r-1/2)/(2r+s)}$ in the RKHS norm. These are exactly the minimax optimal rates of centralized kernel regression up to logarithmic factors. The proof reaches this by bounding the gap between each local robust estimator and the classical centralized gradient descent sequence, and showing that the network-mixing, sampling, and robustness error terms can all be driven below the target rate under the stated selection rules.

Load-bearing premise

Every theorem assumes the local datasets all have exactly the same size, $|D_v|=|D|/m$, and the clean cancellation at the heart of the proof uses that equality; unbalanced local data is not covered.

Editorial extensions

If this is right

  • Each node's local estimator is a final usable approximation, so no central aggregation step is needed.
  • The selection rules (2.7) and (2.11) tell a practitioner how much local data is required given the network's spectral gap.
  • The robustness parameter $\sigma$ must grow with the total sample size and the unknown regularity indices $r$ and $s$ to reach the optimal rate.
  • The windowing function $W$ covers losses such as Cauchy, Welsch, and Fair losses, so the rates apply to a family of robust estimators at once.

Reading between the lines

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

  • Inference: the balanced-dataset requirement is a genuine restriction; real deployments with heterogeneous data volumes would likely need a weighted consensus step or an additional bias-correction term.
  • Inference: because the $\sigma$ selection rule depends on the unknown regularity parameters $r$ and $s$, a practical implementation would need a data-driven or adaptive procedure to satisfy (2.8).
  • Inference: the same techniques should transfer to decentralized minimum error entropy and maximum correntropy algorithms, since these fit the windowing function framework.
  • Inference: the predicted dependence on the spectral gap suggests ring or grid networks need more local data or more communication rounds to keep the same global rate, which the experiments qualitatively support.
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

3 major / 4 minor

Summary. The paper proposes a decentralized robust kernel-based gradient descent algorithm in an RKHS setting, where each agent runs a robust local update based on a windowing loss L_σ and averages only with neighbors through a doubly stochastic communication matrix M. The authors derive an error decomposition (Proposition 1) separating the decentralized estimator from a centralized least-squares gradient descent sequence, and then provide high-probability bounds for the L2, RKHS, and generalization-error gaps. The main advertised results (Theorems 2, 3, and 5) claim that, under the proposed local sample-size rules (2.7) and (2.11) and suitable choices of σ, every local estimator achieves the centralized minimax optimal rates up to logarithmic factors. The paper also contains numerical experiments illustrating convergence and robustness with Laplace noise.

Significance. If the main theorems were established as stated, this would be a substantial contribution: it would be the first decentralized (non-divide-and-conquer) robust kernel regression method with per-node optimal minimax rates, and the proof strategy via the decomposition (3.14) and the T1/T2/T3 estimates is a serious technical framework. The paper is also commendable for covering both L2 and RKHS norms, for making the dependence on the spectral gap explicit, and for including numerical support for robustness against non-Gaussian noise. However, the central rate claims rest on sample-size conditions that do not, as written, imply the required bounds; this is a load-bearing gap that must be repaired before the main results can be accepted.

major comments (3)
  1. [§7, proof of Theorem 2; condition (2.7)] The proof of Theorem 2 lists the first required rate condition as (1/(1-γ_M))(√m/√n) ≤ |D|^{-r/(2r+s)}. Under the balanced-data assumption |D|=mn, this is equivalent to n ≥ (1/(1-γ_M)) |D|^{1/2+r/(2r+s)}. Since \bar t = (2(2r+s+1)/((2r+s)(1-γ_M))) log|D|, the first term of (2.7) is only \bar t |D|^{1/2}, which is strictly weaker by a factor |D|^{r/(2r+s)}. For example, with r=s=1 and a bounded-degree expander, (2.7) permits n ≈ \bar t |D|^{1/2}, giving m ≈ |D|^{1/2}/\bar t and making the offending term ≈ 1/\bar t, which is not O(|D|^{-1/3}) up to logarithmic factors. Thus the stated sample-size rule (2.7) does not imply the first rate condition, and the optimal L2-rate claim in Theorem 2 is not established as stated; either (2.7) must be strengthened to include n ≥ \bar t |D|^{1/2+r/(2r+s)} (or an equivalent m-scaling condition), or an additional assumption such as m=O(1) must be imposed.
  2. [Theorem 5; condition (2.11)] The analogous gap appears in the RKHS-norm result. The first term in the Theorem 4 bound is \bar t(√m/√n), so the required condition for the optimal RKHS rate is \bar t√m/√n ≤ |D|^{-(r-1/2)/(2r+s)}, which translates to n ≥ \bar t |D|^{1/2+(r-1/2)/(2r+s)} after substituting m=|D|/n. The first term of (2.11), however, is only \bar t |D|^{1/2-1/(2(2r+s))}, which is strictly weaker by a factor |D|^{r/(2r+s)}. Consequently, the proof of Theorem 5 does not establish the claimed optimal RKHS rate under the stated condition (2.11); the same repair as in Theorem 2 is needed.
  3. [§7, paragraph before Lemma 6] The paper asserts that the moment condition (2.2) is equivalent to condition (7.1) because f_ρ∈H_K and ||f_ρ||_∞≤κ||f_ρ||_K. This equivalence is not proved and is not immediate: (2.2) is a polynomial moment condition on Y, while (7.1) is an exponential moment condition on the residual Y-f_ρ(x). The optimal-rate conclusions of Theorems 2 and 5 depend on Lemma 6, which is imported from [22] under condition (7.1). The authors should either supply a proof of the equivalence (e.g., showing that the Bernstein condition implies the required subexponential tail with constants depending on ||f_ρ||_∞) or state (7.1) as an explicit assumption.
minor comments (4)
  1. [Throughout Theorems 1–5] All main theorems assume balanced local datasets, |D_u|=|D|/m, and the proof uses this in the transition from (3.7) to (3.8). Since decentralized deployments often have heterogeneous data volumes, a remark or extension covering unbalanced partitions would improve the paper's practical relevance.
  2. [Proof of Proposition 6] In the proof, the line 'When |D_1|=|D_2|=...=|D_m|=1/n' appears to be a typo; it should read '= |D|/m = n'.
  3. [Section 8, numerical experiments] The experiments average over 10 runs without reporting standard deviations or error bars; adding them would make the convergence comparisons in Figures 1–4 more informative.
  4. [Theorems 2 and 5] The choice t=|D|^{1/(2r+s)} is not in general an integer; the paper should state how the iteration count is rounded or that all rates are unaffected by rounding.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the decentralized bounds are proven against an external centralized baseline, and the noted gaps are correctness risks rather than circular reductions.

full rationale

The proof chain is not circular. The paper decomposes the decentralized estimator's error relative to the centralized gradient descent sequence fhat_t,D via terms T1,T2,T3, then applies the triangle inequality together with a centralized rate for fhat_t,D - f_rho. The centralized rate lemma (Lemma 6) is imported from [22], which shares an author with the present paper, but it is a published, checkable, parameter-free result about a centralized least-squares algorithm, not about the present decentralized robust algorithm; it serves as an external benchmark rather than a restatement of the paper's assumptions or fitted values. No parameter is fitted to a subset and then renamed a prediction; the sample-size conditions (2.7) and (2.11) are presented as sufficient inequalities derived from the error bounds, and the sigma selection rules are explicit. There are genuine correctness concerns that belong in a different assessment: the assertion in Section 7 that the Bernstein condition (2.2) is equivalent to the exponential condition (7.1) is unproved and is at least very doubtful, and the hidden m-dependence from |D_u|=|D|/m means condition (2.7) does not seem to imply the required inequality (1/(1-gamma_M)) sqrt(m/n) <= |D|^{-r/(2r+s)} when m grows with |D|. These are proof gaps, not reductions of the claimed result to its own inputs, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 8 assumptions · 0 invented entities

The paper's central claim rests on a stack of standard learning-theoretic assumptions (moment, capacity, source) and the unproved equivalence between its moment condition and the noise condition used by the cited centralized-rate theorem. No invented entities are introduced.

free parameters (3)
  • step size alpha = alpha ~ 1 (constant order)
    Chosen by user; theorems require 0<alpha<=kappa^{-2} min{1/W'+(0),1/C_W} and, for optimal rates, alpha constant order. Not fitted to data.
  • robustness scale sigma = sigma >= |D|^{(p+r+1)/(2p(2r+s))} or |D|^{(p+r+2)/(2p(2r+s))} n^{-1/(4p)} (lower bound, (2.8))
    The optimal-rate results require sigma large enough that the robust perturbation term T3 is negligible; this depends on unknowns r, s, p, n. In experiments sigma=10 is fixed.
  • local sample size n = n satisfying (2.7) or (2.11)
    The rate results hold only when the per-agent sample size meets the stated lower bounds; this is a design choice, not estimated from data.
assumptions (8)
  • domain assumption Bernstein moment condition on outputs (2.2)
    Assumed for the concentration results; standard for unbounded outputs.
  • domain assumption Effective dimension bound N(lambda) <= C0 lambda^{-s}, 0<s<=1 (2.4)
    Capacity condition on the RKHS, standard in the literature.
  • domain assumption Source condition f_rho = L_K^r g_rho, r>1/2 (2.5)
    Regularity of the target function, standard.
  • domain assumption Doubly stochastic communication matrix M with 0<gamma_M<1 (2.1)
    Network model needed for consensus convergence; standard in distributed optimization.
  • ad hoc to paper Equivalence of (2.2) and (7.1) in the current setting
    Asserted in Section 7 before Lemma 6 without proof; used to import centralized optimal rates from [22].
  • standard math Centralized least-squares gradient descent optimal rates (Lemma 6, from [22])
    Published result used as baseline; independently checkable.
  • standard math Concentration inequalities for Hilbert-valued random variables (Lemmas 2-3, from [30], [40])
    Used to control all random error terms.
  • standard math Operator norm estimate for polynomials of L_K (Lemma 5, from [49])
    Used for spectral bounds on the iteration operators.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Theory of Decentralized Robust Kernel-Based Learning." pith.science (2026). https://pith.science/paper/24DBYEP7

@misc{pith2026250605215,
  author       = {Pith},
  title        = {Pith review of: Theory of Decentralized Robust Kernel-Based Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/24DBYEP7}},
  note         = {Machine review of arXiv:2506.05215}
}
abstract

We propose a new decentralized robust kernel-based learning algorithm within the framework of reproducing kernel Hilbert spaces (RKHSs) by utilizing a networked system that can be represented as a connected graph. The robust loss function $\huaL_\sigma$ induced by a windowing function $W$ and a robustness scaling parameter $\sigma>0$ can encompass a broad spectrum of robust losses. Consequently, the proposed algorithm effectively provides a unified decentralized learning framework for robust regression, which fundamentally differs from the existing distributed robust kernel-based learning schemes, all of which are divide-and-conquer based. We rigorously establish a learning theory and offer comprehensive convergence analysis for the algorithm. We show each local robust estimator generated from the decentralized algorithm can be utilized to approximate the regression function. Based on kernel-based integral operator techniques, we derive general high confidence convergence bounds for the local approximating sequence in terms of the mean square distance, RKHS norm, and generalization error, respectively. Moreover, we provide rigorous selection rules for local sample size and show that, under properly selected step size and scaling parameter $\sigma$, the decentralized robust algorithm can achieve optimal learning rates (up to logarithmic factors) in both norms. The parameter $\sigma$ is shown to be essential for enhancing robustness and ensuring favorable convergence behavior. The intrinsic connection among decentralization, sample selection, robustness of the algorithm, and its convergence is clearly reflected.

Figures

Figures reproduced from arXiv: 2506.05215 by the authors.

Figure 1
Figure 1. Maximum, minimum, and median of the excess generalization errors across all agents versus the data sample size |D| of the DKBRGD algorithm [PITH_FULL_IMAGE:figures/full_fig_p040_1.png] view at source ↗
Figure 2
Figure 2. Maximum of the excess generalization errors across all agents versus the data sample size |D| of the DKBRGD algorithm for three different choices of the number of agents m. 40 [PITH_FULL_IMAGE:figures/full_fig_p040_2.png] view at source ↗
Figure 3
Figure 3. Maximum of the excess generalization errors across all agents versus the data sample size |D| of the DKBRGD algorithm for two different choices of the noise distribution. number of data sample size |D|. The findings shown in [PITH_FULL_IMAGE:figures/full_fig_p041_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Maximum of the excess generalization errors across all agents versus the data sample size |D| of the DKBGD algorithm for two different choices of the noise distribution. [4] Andreas Christmann, and Ingo Steinwart. Consistency and robustness of kernel-based regres￾sion …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 55 canonical work pages

  1. [22]

    Distributed kernel-based gradient descent algorithms

    Shao-Bo Lin, and Ding-Xuan Zhou. Distributed kernel-based gradient descent algorithms. Constructive Approximation, 47(2): 249-276, 2018

  2. [1]

    On regularization algorithms in learning theory, Journal of complexity, 23.1: 52-72, 2007

    Frank Bauer, Sergei Pereverzev, and Lorenzo Rosasco. On regularization algorithms in learning theory, Journal of complexity, 23.1: 52-72, 2007

  3. [2]

    Hendrickx, Alex Olshevsky, and John N

    Vincent Blondel, Julien M. Hendrickx, Alex Olshevsky, and John N. Tsitsiklis. Convergence in multiagent coordination, consensus, and flocking, Proceedings of the 44th IEEE Conference on Decision and Control (pp. 2996-3000), 2005

  4. [3]

    Optimal rates for the regularized least-squares algorithm

    Andrea Caponnetto, and Ernesto De Vito. Optimal rates for the regularized least-squares algorithm. Foundations of Computational Mathematics 7 (2007): 331-368. 41 Figure 4: Maximum of the excess generalization errors across all agents versus the data sample size|D| of the DKBGD algorithm for two different choices of the noise distribution

  5. [4]

    Consistency and robustness of kernel-based regres- sion in convex risk minimization

    Andreas Christmann, and Ingo Steinwart. Consistency and robustness of kernel-based regres- sion in convex risk minimization. Bernoulli (2007): 799-819

  6. [5]

    Learning theory: An Approximation Theory Viewpoint

    Felipe Cucker, and Ding-Xuan Zhou. Learning theory: An Approximation Theory Viewpoint. Vol. 24. Cambridge University Press, 2007

  7. [6]

    Duchi, Alekh Agarwal, Martin J

    John C. Duchi, Alekh Agarwal, Martin J. Wainwright. Dual averaging for distributed opti- mization: Convergence analysis and network scaling. IEEE Transactions on Automatic control 57.3 (2011): 592-606

  8. [7]

    Yunlong Feng, Xiaolin Huang, Lei Shi, Yuning Yang, Johan A. K. Suykens. Learning with the maximum correntropy criterion induced losses for regression. Journal of Machine Learning Research, 16 (2015), 993–1034

Show all 55 references
  1. [8]

    A framework of learning through empirical gain maximization

    Yunlong Feng, Qiang Wu. A framework of learning through empirical gain maximization. Neural Computation 33.6 (2021): 1656-1697

  2. [9]

    Mapreduce: Distributed computing for machine learning

    Dan Gillick, Arlo Faria, and John DeNero. Mapreduce: Distributed computing for machine learning. Berkley, Dec 18 (2006)

  3. [10]

    Distributed minimum error entropy algorithms

    Xin Guo, Ting Hu, and Qiang Wu. Distributed minimum error entropy algorithms. Journal of Machine Learning Research 21.126: 1-31, 2020

  4. [11]

    Optimality of robust online learning

    Zheng-Chu Guo, Andreas Christmann, and Lei Shi. Optimality of robust online learning. Foundations of Computational Mathematics 24.5 (2024): 1455-1483

  5. [12]

    Gradient descent for robust kernel-based regression

    Zheng-Chu Guo, Ting Hu, Lei Shi. Gradient descent for robust kernel-based regression. Inverse Problems 34.6 (2018): 065009

  6. [13]

    Learning theory of distributed spectral algorithms, Inverse Problems, 33.7 (2017): 074009

    Zheng-Chu Guo, Shao-Bo Lin, and Ding-Xuan Zhou. Learning theory of distributed spectral algorithms, Inverse Problems, 33.7 (2017): 074009

  7. [14]

    Robust guided image filtering using nonconvex potentials

    Bumsub Ham, Minsu Cho, and Jean Ponce. Robust guided image filtering using nonconvex potentials. IEEE transactions on pattern analysis and machine intelligence 40.1 (2017): 192-207. 42

  8. [15]

    Distributed robust regression with correntropy losses and regularization kernel networks

    Ting Hu, Renjie Guo. Distributed robust regression with correntropy losses and regularization kernel networks. Analysis and Applications (2024): 1-36

  9. [16]

    Distributed kernel gradient descent algorithm for minimum error entropy principle, Applied and Computational Harmonic Analysis, 49(1): 229-256, 2020

    Ting Hu, Qiang Wu, and Ding-Xuan Zhou. Distributed kernel gradient descent algorithm for minimum error entropy principle, Applied and Computational Harmonic Analysis, 49(1): 229-256, 2020

  10. [17]

    Kernel gradient descent algorithm for information theoretic learning

    Ting Hu, Qiang Wu, Ding-Xuan Zhou. Kernel gradient descent algorithm for information theoretic learning. Journal of Approximation Theory 263 (2021): 105518

  11. [18]

    Huber, and Elvezio M

    Peter J. Huber, and Elvezio M. Ronchetti. Robust Statistics. John Wiley & Sons, 2011

  12. [19]

    Decentralized online learning with kernels, IEEE Transactions on Signal Processing, 66(12): 3240-3255, 2018

    Alec Koppel, Santiago Paternain, Cedric Richard, and Alejandro Ribeiro. Decentralized online learning with kernels, IEEE Transactions on Signal Processing, 66(12): 3240-3255, 2018

  13. [20]

    Decentralized learning over a network with Nyström approximation using SGD

    Heng Lian, and Jiamin Liu. Decentralized learning over a network with Nyström approximation using SGD. Applied and Computational Harmonic Analysis, 66 (2023): 373-387

  14. [21]

    Distributed learning with regularized least squares, The Journal of Machine Learning Research, 18.1: 3202-3232, 2017

    Shao-Bo Lin, Xin Guo, and Ding-Xuan Zhou. Distributed learning with regularized least squares, The Journal of Machine Learning Research, 18.1: 3202-3232, 2017

  15. [23]

    Distributed filtered hyperinterpolation for noisy data on the sphere, SIAM Journal on Numerical Analysis, 59(2): 634-659, 2021

    Shao-Bo Lin, Yu Guang Wang, and Ding-Xuan Zhou. Distributed filtered hyperinterpolation for noisy data on the sphere, SIAM Journal on Numerical Analysis, 59(2): 634-659, 2021

  16. [24]

    Statistical optimality of divide and conquer kernel-based functional linear regression, Journal of Machine Learning Research, 25.155 (2024): 1-56

    Jiading Liu, and Lei Shi. Statistical optimality of divide and conquer kernel-based functional linear regression, Journal of Machine Learning Research, 25.155 (2024): 1-56

  17. [25]

    Pokharel, and Jose C

    Weifeng Liu, Puskal P. Pokharel, and Jose C. Principe. Correntropy: Properties and applica- tions in non-Gaussian signal processing. IEEE Transactions on signal processing 55.11 (2007): 5286-5298

  18. [26]

    Approximating functions with multi-features by deep convolutional neural networks, Analysis and Applications 21, 93–125, 2023

    Tong Mao, Zhongjie Shi, and Ding-Xuan Zhou. Approximating functions with multi-features by deep convolutional neural networks, Analysis and Applications 21, 93–125, 2023

  19. [27]

    Distributed optimization over time-varying directed graphs, IEEE Transactions on Automatic Control 60.3 (2014): 601-615

    Angelia Nedic, and Alex Olshevsky. Distributed optimization over time-varying directed graphs, IEEE Transactions on Automatic Control 60.3 (2014): 601-615

  20. [28]

    Distributed subgradient methods for multi-agent optimization, IEEE Transactions on Automatic Control, 54(1): 48-61, 2009

    Angelia Nedic, and Asuman Ozdaglar. Distributed subgradient methods for multi-agent optimization, IEEE Transactions on Automatic Control, 54(1): 48-61, 2009

  21. [29]

    Distributed SGD in overparametrized linear regression

    Mike Nguyen, Charly Kirst, and Nicole Mücke. Distributed SGD in overparametrized linear regression. Analysis and Applications 22.03 (2024): 425-466

  22. [30]

    Optimum bounds for the distributions of martingales in Banach spaces, Annals of Probability, 1679–1706, 1994

    Iosif Pinelis. Optimum bounds for the distributions of martingales in Banach spaces, Annals of Probability, 1679–1706, 1994

  23. [31]

    Sundhar Ram, Angelia Nedić, and Venugopal V

    S. Sundhar Ram, Angelia Nedić, and Venugopal V. Veeravalli. Distributed stochastic sub- gradient projection algorithms for convex optimization, Journal of Optimization Theory and Applications, 147 (2010): 516-545

  24. [32]

    Decentralised learning with dis- tributed gradient descent and random features, Proceedings of 37th International Conference on Machine Learning, PMLR, 119, 2020

    Dominic Richards, Patrick Rebeschini, and Lorenzo Rosasco. Decentralised learning with dis- tributed gradient descent and random features, Proceedings of 37th International Conference on Machine Learning, PMLR, 119, 2020. 43

  25. [33]

    Learning theory estimates via integral of operators and their approximations, Constructive Approximation 26(2): 153-172, 2007

    Steve Smale, Ding-Xuan Zhou. Learning theory estimates via integral of operators and their approximations, Constructive Approximation 26(2): 153-172, 2007

  26. [34]

    Support Vector Machines

    Ingo Steinwart, Andreas Christmann. Support Vector Machines. Springer, 2008

  27. [35]

    Estimating conditional quantiles with the help of the pinball loss

    Ingo Steinwart, and Andreas Christmann. Estimating conditional quantiles with the help of the pinball loss. Bernoulli, 17 (1) 211 - 225, February 2011

  28. [36]

    Optimal rates of distributed regression with imperfect kernels, Journal of Machine Learning Research, 22: 7732-7765, 2021

    Hongwei Sun, and Qiang Wu. Optimal rates of distributed regression with imperfect kernels, Journal of Machine Learning Research, 22: 7732-7765, 2021

  29. [37]

    Distributed learning with dependent samples, IEEE Transactions on Information Theory, 68(9): 6003-6020, 2022

    Zirui Sun, and Shao-Bo Lin. Distributed learning with dependent samples, IEEE Transactions on Information Theory, 68(9): 6003-6020, 2022

  30. [38]

    Distributed least squares prediction for functional linear regression, Inverse Problems, 38(2): 025002, 2021

    Hongzhi Tong. Distributed least squares prediction for functional linear regression, Inverse Problems, 38(2): 025002, 2021

  31. [39]

    On the convergence of gradient descent for robust functional linear regression, Journal of Complexity, 84: 101858, 2024

    Cheng Wang, and Jun Fan. On the convergence of gradient descent for robust functional linear regression, Journal of Complexity, 84: 101858, 2024

  32. [40]

    Online minimum error entropy algorithm with unbounded sampling, Analysis and Applications, 17(2): 293-322, 2019

    Cheng Wang, and Ting Hu. Online minimum error entropy algorithm with unbounded sampling, Analysis and Applications, 17(2): 293-322, 2019

  33. [41]

    Weak Convergence and Empirical Processes: with Applications to Statistics

    Jon Wellner. Weak Convergence and Empirical Processes: with Applications to Statistics. Springer Science & Business Media, 2013

  34. [42]

    Learning rates of least-square regularized regression

    Qiang Wu, Yiming Ying, and Ding-Xuan Zhou. Learning rates of least-square regularized regression. Foundations of Computational Mathematics 6 (2006): 171-192

  35. [43]

    Data mining with big data

    Xindong Wu, Xingquan Zhu, Gong-Qing Wu, Wei Ding. Data mining with big data. IEEE Transactions on Knowledge and Data Engineering 26.1 (2013): 97-107

  36. [44]

    Fast linear iterations for distributed averaging, Systems Control Letters, 53.1 (2004): 65-78

    Lin Xiao, and Stephen Boyd. Fast linear iterations for distributed averaging, Systems Control Letters, 53.1 (2004): 65-78

  37. [45]

    Distributed average consensus with least- mean-square deviation

    Lin Xiao, Stephen Boyd, and Seung-Jean Kim. Distributed average consensus with least- mean-square deviation. Journal of Parallel and Distributed Computing 67.1: 33-46, 2007

  38. [46]

    COKE: Communication-censored decentral- ized kernel learning, Journal of Machine Learning Research, 22(1): 8813-8847, 2021

    Ping Xu, Yue Wang, Xiang Chen, and Zhi Tian. COKE: Communication-censored decentral- ized kernel learning, Journal of Machine Learning Research, 22(1): 8813-8847, 2021

  39. [47]

    On early stopping in gradient descent learning

    Yuan Yao, Lorenzo Rosasco, and Andrea Caponnetto. On early stopping in gradient descent learning. Constructive Approximation 26.2 (2007): 289-315

  40. [48]

    Online gradient descent learning algorithms

    Yiming Ying, and Massimiliano Pontil. Online gradient descent learning algorithms. Founda- tions of Computational Mathematics 8 (2008): 561-596

  41. [49]

    Distributed gradient descent for functional learning, IEEE Transactions on Information Theory, 70(9), 6547 - 6571, 2024, 2024

    Zhan Yu, Jun Fan, Zhongjie Shi, and Ding-Xuan Zhou. Distributed gradient descent for functional learning, IEEE Transactions on Information Theory, 70(9), 6547 - 6571, 2024, 2024

  42. [50]

    Robust kernel-based distribution regression, Inverse Problems, 37(10): 105014, 2021

    Zhan Yu, Daniel Ho, Zhongjie Shi, and Ding-Xuan Zhou. Robust kernel-based distribution regression, Inverse Problems, 37(10): 105014, 2021

  43. [51]

    Distributed randomized gradient-free mirror descent algorithm for constrained optimization, IEEE Transactions on Automatic Control, 67(2): 957-964, 2022

    Zhan Yu, Daniel Ho, and Deming Yuan. Distributed randomized gradient-free mirror descent algorithm for constrained optimization, IEEE Transactions on Automatic Control, 67(2): 957-964, 2022. 44

  44. [52]

    Zhan Yu, Block coordinate type methods for optimization and learning, Analysis and Appli- cations 21, 777–817, 2023

  45. [53]

    Generalization analysis of deep CNNs under maximum correntropy criterion

    Yingqiao Zhang, Zhiying Fang, and Jun Fan. Generalization analysis of deep CNNs under maximum correntropy criterion. Neural Networks 174 (2024): 106226

  46. [54]

    Divide and conquer kernel ridge re- gression: A distributed algorithm with minimax optimal rates, Journal of Machine Learning Research, 16(1): 3299–3340, 2015

    Yuchen Zhang, John Duchi, and Martin Wainwright. Divide and conquer kernel ridge re- gression: A distributed algorithm with minimax optimal rates, Journal of Machine Learning Research, 16(1): 3299–3340, 2015

  47. [55]

    Capacity of reproducing kernel spaces in learning theory, IEEE Transactions on Information Theory, 49(7): 1743-1752, 2003

    Ding-Xuan Zhou. Capacity of reproducing kernel spaces in learning theory, IEEE Transactions on Information Theory, 49(7): 1743-1752, 2003. 45

Pith tools

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