Pith. sign in

REVIEW 2 major objections 5 minor 41 references

On Convergence of Distributed Approximate Newton Methods: Globalization, Sharper Bounds and Beyond

T0 review · 2 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read Line search and a master-only local solve let DANE converge globally and, in the quadratic case, reduce communication complexity from O(κ²/n) to O(κ/√n) rounds.

desk verdict Deterministic theory is solid and worth a serious referee, but the advertised stochastic non-quadratic rates rest on an unproved uniform Hessian closeness condition. read the letter →

arxiv 1908.02246 v1 pith:DYQXAXNX submitted 2019-08-06 stat.ML cs.LGstat.CO

classification stat.MLcs.LGstat.CO MSC 68W1590C2590C53
keywords distributedoptimizationDANEapproximateNewtonmethodcommunicationcomplexityheavy-ballaccelerationbacktrackinglinesearchstronglyconvexempiricalriskminimization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is about DANE, a communication-efficient distributed optimization method in which each machine solves a local subproblem and the results are averaged. The authors propose two variants: DANE-LS, which adds backtracking line search and assigns the local solve to a single master machine, and DANE-HB, which adds heavy-ball momentum. They claim that for quadratic losses DANE-LS reaches ε-accuracy in O(κ/√n log(p/δ) log(1/ε)) communication rounds, improving the previously established O(κ²/n) bound for DANE. They further claim that for strongly convex non-quadratic losses DANE-LS converges globally with local rate O(γ/μ log(1/ε)), and that DANE-HB attains O(√(γ/μ) log(1/ε)) locally, with a global version for linear prediction models. If these guarantees hold, distributed training could communicate far less whenever each machine's local data are statistically representative of the whole.

What carries the argument

The proof rests on a two-part mechanism. First, with only the master machine solving, DANE-LS's update obeys an error recursion driven by the preconditioned correction I - (H1 + γI)⁻¹H, and the spectral norm bound ||I - $H^{{1/2}}$(H1 + γI)⁻$¹H^{{1/2}}$|| ≤ 2γ/(μ+2γ) (Lemma 23) delivers a contraction factor 1 - μ/(2(μ+2γ)). Second, a backtracking Armijo line search—using either the global objective or a locally computable sufficient-descent condition—guarantees descent for arbitrary γ > 0, and the unit step is eventually accepted, turning global asymptotic convergence into a local non-asymptotic rate. For DANE-HB, the same contraction is embedded in the 2-block companion matrix of heavy-ball momentum, whose spectral radius is bounded by 1 - √(μ/(μ+2γ)) (Lemma 24). The uniform Hessian-closeness condition sup_w ||∇²F1(w) - ∇²F(w)|| ≤ γ is the thread connecting local and global Hessian information in Lemmas 27, 13, and 18.

What would settle it

Construct a strongly convex twice-differentiable pair F, F1 for which ||∇²F1(w*) - ∇²F(w*)|| ≤ γ but sup_w ||∇²F1(w) - ∇²F(w)|| ≫ γ, run DANE-LS with unit step in a neighborhood of w*, and check whether the contraction rate is the predicted O(γ/μ). If the observed rate degrades to the larger uniform value, the uniform-closeness condition is genuinely load-bearing; if not, the theorem's assumption is stronger than needed.

Watch

Extended reading notes

Core claim

The paper proves that a modest architectural change to DANE—letting one master machine solve a regularized local problem and using backtracking line search to force descent—yields both global convergence and markedly sharper communication bounds. For a quadratic objective with local Hessian H1 and global Hessian H satisfying ||H1 - H|| ≤ γ, DANE-LS contracts the distance to the optimum at a rate 1 - μ/(2(μ+2γ)) per round, giving O(γ/μ log(1/ε)) communication. In the stochastic setting with γ = L√(32 log(p/δ)/n), this becomes O(κ/√n log(p/δ) log(1/ε)), improving on prior DANE-type bounds of O(κ²/n). For strongly convex twice-differentiable objectives, DANE-LS converges globally from any starting point, and once the iterate enters a neighborhood of the minimizer the same O(γ/μ log(1/ε)) rate applies. Adding heavy-ball momentum gives a local rate O(√(γ/μ) log(1/ε)), and for linear prediction models DANE-HB-LM makes that rate global.

Load-bearing premise

The non-quadratic results assume that the local Hessian and the global Hessian differ by at most γ everywhere in the domain, not just near the minimizer; in the stochastic setting only pointwise closeness is proved, so the global/local non-quadratic rates depend on an unproved uniform version of that bound.

Editorial extensions

If this is right

  • For quadratic ERM, DANE-LS improves the communication complexity from O(κ²/n) to O(κ/√n); when κ = O(√(mn)) this is O(√m) rounds, closing much of the gap to known lower bounds.
  • DANE-LS's global convergence means practitioners can safely use small regularization γ (e.g., O(1/√n)) without risking divergence, and can terminate by monitoring the objective decrease or the norm of the iterate difference.
  • DANE-HB matches the communication complexity of AIDE and MP-DANE for quadratic objectives using an elementary heavy-ball analysis, giving a simpler route to nearly tight second-order distributed optimization.
  • For linear prediction models, DANE-HB-LM provides a global O(√(γ/μ) log²(1/ε)) communication bound, which the paper presents as the first nearly optimal non-asymptotic DANE-type result for non-quadratic convex objectives.
  • The local non-quadratic rates show that the benefit of a good local Hessian approximation—measured by γ/μ or √(γ/μ)—is not an artifact of quadratic structure, so the method's advantage should persist for smooth strongly convex losses near their minimizers.

Reading between the lines

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

  • The paper's non-quadratic stochastic conclusions appear conditional: the uniform Hessian-closeness bound sup_w ||∇²F1(w) - ∇²F(w)|| ≤ γ is assumed, while Lemma 28 only proves pointwise closeness; without a uniform convergence theorem for smooth losses, the local rates in Theorems 13 and 18 do not yet transfer to the stochastic setting.
  • The master-only design suggests an asynchronous or energy-saving implementation in which workers only compute gradients at broadcast iterates; if the convergence bounds hold under asynchrony, this could further cut wall-clock time in federated settings.
  • The companion-matrix heavy-ball technique used for DANE-HB is a template that could accelerate other approximate Newton methods whose local Hessian error is bounded by γ, potentially yielding analogous √(γ/μ) rates for inexact quasi-Newton-distributed algorithms.
  • A sharp testable prediction follows from Corollary 4 and Theorem 19: for logistic regression with n large relative to p, communication rounds should grow roughly like m^{1/4} for DANE-HB-LM and √m for DANE-LS; these scalings could serve as benchmarks for comparing distributed optimizers.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. This paper studies communication-efficient distributed optimization of strongly convex empirical-risk problems. It proposes DANE-LS, a DANE variant with master-only local solves and backtracking line search, and DANE-HB, a heavy-ball accelerated variant. For quadratic objectives the authors prove stochastic communication complexity O(kappa/sqrt(n) log(1/epsilon)) for DANE-LS and O(sqrt(kappa)/n^{1/4} log(1/epsilon)) for DANE-HB, improving on previous DANE bounds. For non-quadratic strongly convex functions they prove global asymptotic convergence for DANE-LS and local linear rates O(gamma/mu log(1/epsilon)) and O(sqrt(gamma/mu) log(1/epsilon)) under a uniform Hessian closeness condition sup_w ||nabla^2 F1(w) - nabla^2 F(w)|| <= gamma; for linear prediction models a globally convergent variant DANE-HB-LM is analyzed. Numerical experiments on synthetic and real logistic-regression tasks illustrate the stability and communication advantages of the proposed methods.

Significance. If the results hold, the paper makes a solid contribution. The quadratic bounds are the cleanest part: they improve the known DANE bound from O(kappa^2/n) to O(kappa/sqrt(n)) for DANE-LS and match AIDE/MP-DANE for DANE-HB using fairly elementary spectral-radius lemmas. The heavy-ball lemmas and the convergence recurrences are transparent and give parameter-free contraction factors. The line-search globalization also addresses a practically observed instability of plain DANE. The deterministic non-quadratic local rates are natural and potentially useful. However, the stochastic non-quadratic claims currently outrun the supplied proof machinery, and the global-convergence proof has an internal gap. The quadratic results and the deterministic conditional non-quadratic rates are defensible and worth publishing after revision.

major comments (2)
  1. [Section 2.3, Lemma 28, Remark 14; Theorems 13 and 18] The non-quadratic results (Lemma 27, Lemma 11, Theorem 13, Theorem 18, and the local part of Theorem 8(b)) all assume the uniform bound sup_w ||nabla^2 F1(w) - nabla^2 F(w)|| <= gamma. In the stochastic sampling setting the only concentration result supplied, Lemma 28, is pointwise in w: for each fixed w, with probability at least 1 - delta, ||H1(w) - H(w)|| <= L sqrt(32 log(p/delta)/n). Since the iterates w(t) are random and depend on the same data that defines F1, this pointwise statement does not imply the uniform bound needed to apply the contraction arguments. Remark 14 cites Mei et al. (2018) for gamma = O(sqrt(p/n)), but no uniform-convergence theorem is stated and its hypotheses are not checked for the specific losses and regularizers used here. Thus the advertised non-quadratic communication improvements in Table 1 are not derived as stochastic statements. For quadratic objectives this gap is harmless because the Hessians are constant, but for logistic-type losses it is load-bearing. Please either prove a uniform concentration bound under explicit conditions or explicitly restrict the non-quadratic theorems to the deterministic setting where sup_w ||nabla^2 F1(w) - nabla^2 F(w)|| <= gamma is an assumption.
  2. [Appendix B.2, proof of Theorem 8(a)] The paragraph 'From Lemma 27 we know...' invokes Lemma 27, whose hypothesis sup_w ||nabla^2 F1(w) - nabla^2 F(w)|| <= gamma is not assumed in part (a), so the cited lemma cannot be used there. In addition, the displayed descent inequality eta_t rho(gamma + mu) ||w~^t - w^{t-1}||^2 <= 2(F(w^{t-1}) - F(w^t)) does not by itself imply ||w~^t - w^{t-1}|| -> 0 unless one has a lower bound on the accepted step sizes eta_t, and no such lower bound is proved. The conclusion that the iterates converge to the minimizer also requires showing that any limit point of the bounded sequence is stationary, e.g., by passing to the limit in the gradient of P^{t-1}; this step is absent. The result is likely repairable with a standard Armijo backtracking lower-bound argument, but as written the global-convergence theorem is not fully proved.
minor comments (5)
  1. [Corollary 4] The displayed bound t >= (1 + 2 kappa sqrt(32 log(p/delta)/n)) log(2 sqrt(kappa) ||w(0) - w*|| / epsilon) is not algebraically equivalent to Theorem 3's bound t >= 2(mu + 2 gamma)/mu log(sqrt(kappa) ||w(0) - w*|| / epsilon) with gamma = L sqrt(32 log(p/delta)/n); a factor of 2 appears to be missing. Please check the constants.
  2. [Appendix C] The opening sentence of Appendix C says 'proofs of results in Section 2'; it should refer to Section 3.
  3. [Proof of Theorem 18] In the paragraph defining the matrices at the fixed point, both Hessians are denoted H^*: the line 'Let H* = nabla^2 F(w*) and H* = nabla^2 F1(w*)' assigns the same symbol to two different matrices; the second should be H1^*.
  4. [Proof of Lemma 6(b)] The notation Delta F-tilde(w(t-1), w~^t) in the proof of Lemma 6(b) is undefined; it should presumably be Delta F1(w~^t, w^{t-1}).
  5. [Proof of Theorem 8(a)] The word 'uncles' in the proof of Theorem 8(a) is a typo for 'unless'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all rates are conditional on explicit Hessian-closeness assumptions and derived from spectral-radius lemmas; no fitted parameter is renamed as a prediction.

full rationale

The paper's derivation chain is not circular. The quadratic-rate theorems (Theorems 3 and 15) take ||H1 - H|| <= gamma as an explicit premise and prove contraction through deterministic spectral-radius lemmas (Lemma 23 and Lemma 24); gamma is the algorithm's regularization input, not a constant fitted to reproduce the claimed convergence. Corollaries 4 and 16 instantiate gamma via Lemma 28, a matrix concentration bound, and in the quadratic case the Hessians are constant so the pointwise concentration statement is sufficient. The non-quadratic local rates (Theorems 13 and 18) explicitly assume sup_w ||∇²F1(w) - ∇²F(w)|| <= gamma, and Lemma 27 is derived from exactly that same assumption; the rate O(gamma/mu) is a conditional guarantee, not a self-definitional equivalence. Remark 14's appeal to Mei et al. for gamma = O(sqrt(p/n)) is an external uniform-convergence citation, not a self-citation, and the absence of a proof of that uniform version is a correctness/completeness gap rather than circularity. No step defines a target quantity in terms of itself, no prediction is statistically forced by fitted inputs, and no load-bearing premise rests on an author self-citation chain. The written proof of Theorem 8(a) does invoke Lemma 27 without that part's assumptions, an internal inconsistency that a careful reader should flag, but it does not make the central derivation circular. Overall, the central claims have independent mathematical content and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central results rest on standard convex optimization assumptions plus a uniform Hessian-closeness condition that is proved only pointwise in the stochastic setting; no new physical entities are introduced. Hyperparameters gamma, beta, rho, and epsilon_t are algorithmic tuning choices, not fitted to the convergence claims.

free parameters (4)
  • gamma (local subproblem regularization) = L sqrt(32 log(p/delta)/n) in Corollaries 4, 16, 21; 40/sqrt(n) in experiments; arbitrary in deterministic theorems
    Controls the trust region of the local DANE subproblem. The theorems require it to satisfy ||H1 - H|| <= gamma, so its value is chosen from a Hessian concentration bound, not fitted to the convergence outcome.
  • beta (heavy-ball momentum) = (1 - sqrt(mu/(mu + 2*gamma)))^2
    Optimal heavy-ball coefficient from spectral radius Lemma 24; no data fitting.
  • epsilon_t (local subproblem tolerance) = Theorem-dependent decreasing bounds
    Each theorem sets an upper bound on the local gradient norm of the inexact solver; it is a stopping rule, not a fitted parameter.
  • rho (line search parameter) = Any value in (0,1/3) for Theorem 13
    Line search acceptance constant; no data dependence.
assumptions (5)
  • domain assumption F and F1 are mu-strongly convex and L-smooth, with nu-Lipschitz Hessian
    Assumed in Definitions 1 and 2 and throughout Theorems 8, 13, and 18; standard for second-order analyses.
  • domain assumption Data is evenly and randomly split across m machines; each local objective is an average of n i.i.d. samples
    Equation (2) and Section 1; needed so local Hessian H1 approximates global H.
  • domain assumption Uniform Hessian closeness sup_w ||nabla^2 F1(w) - nabla^2 F(w)|| <= gamma
    Used in Lemma 27, Theorems 8(b), 13, 18, and 19. In stochastic settings only a pointwise version is proved in Lemma 28, making this the load-bearing assumption.
  • standard math Spectral radius lemmas (Lemmas 24 and 25) and matrix concentration (Tropp 2012, Lemma 28)
    Used without proof as background; standard results.
  • domain assumption Line search iterates stay bounded so D in Lemma 6(b) exists
    Option-II feasibility requires ||w_tilde^t - w^(t-1)|| <= D; the paper assumes a bounded domain in Remark 7.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On Convergence of Distributed Approximate Newton Methods: Globalization, Sharper Bounds and Beyond." pith.science (2026). https://pith.science/paper/DYQXAXNX

@misc{pith2026190802246,
  author       = {Pith},
  title        = {Pith review of: On Convergence of Distributed Approximate Newton Methods: Globalization, Sharper Bounds and Beyond},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DYQXAXNX}},
  note         = {Machine review of arXiv:1908.02246}
}
read the original abstract

The DANE algorithm is an approximate Newton method popularly used for communication-efficient distributed machine learning. Reasons for the interest in DANE include scalability and versatility. Convergence of DANE, however, can be tricky; its appealing convergence rate is only rigorous for quadratic objective, and for more general convex functions the known results are no stronger than those of the classic first-order methods. To remedy these drawbacks, we propose in this paper some new alternatives of DANE which are more suitable for analysis. We first introduce a simple variant of DANE equipped with backtracking line search, for which global asymptotic convergence and sharper local non-asymptotic convergence rate guarantees can be proved for both quadratic and non-quadratic strongly convex functions. Then we propose a heavy-ball method to accelerate the convergence of DANE, showing that nearly tight local rate of convergence can be established for strongly convex functions, and with proper modification of algorithm the same result applies globally to linear prediction models. Numerical evidence is provided to confirm the theoretical and practical advantages of our methods.

Figures

Figures reproduced from arXiv: 1908.02246 by the authors.

Figure 1
Figure 1. (a) The number of communication rounds (y-axis) ve [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Theory verification: the number of communication r [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Algorithm evaluation with comparison to DANE-typ [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Algorithm evaluation with comparison to other met [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 31 canonical work pages

  1. [1]

    Communication complexity of distributed convex learning and optimization

    Yossi Arjevani and Ohad Shamir. Communication complexity of distributed convex learning and optimization. In Advances in Neural Information Processing Systems (NIPS), pages 1756--1764, 2015

  2. [2]

    Distributed optimization and statistical learning via the alternating direction method of multipliers

    Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, and Jonathan Eckstein. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine Learning , 3 0 (1): 0 1--122, 2011

  3. [3]

    Mapreduce: Simplified data processing on large clusters

    Jeffrey Dean and Sanjay Ghemawat. Mapreduce: Simplified data processing on large clusters. Commun. ACM, 51 0 (1): 0 107--113, January 2008. ISSN 0001-0782. doi:10.1145/1327452.1327492. URL http://doi.acm.org/10.1145/1327452.1327492

  4. [4]

    Global convergence of the heavy-ball method for convex optimization

    Euhanna Ghadimi, Hamid Reza Feyzmahdavian, and Mikael Johansson. Global convergence of the heavy-ball method for convex optimization. In 2015 European Control Conference (ECC), pages 310--315. IEEE, 2015

  5. [5]

    Result analysis of the nips 2003 feature selection challenge

    Isabelle Guyon, Steve Gunn, Asa Ben-Hur, and Gideon Dror. Result analysis of the nips 2003 feature selection challenge. In Advances in Neural Information Processing Systems (NIPS), pages 545--552, 2005

  6. [6]

    Communication-efficient distributed dual coordinate ascent

    Martin Jaggi, Virginia Smith, Martin Tak \'a c, Jonathan Terhorst, Sanjay Krishnan, Thomas Hofmann, and Michael I Jordan. Communication-efficient distributed dual coordinate ascent. In Advances in Neural Information Processing Systems (NIPS), 2014

  7. [7]

    Accelerating stochastic gradient descent using predictive variance reduction

    Rie Johnson and Tong Zhang. Accelerating stochastic gradient descent using predictive variance reduction. In Advances in Neural Information Processing Systems (NIPS), pages 315--323, 2013

  8. [8]

    Communication-efficient distributed statistical inference

    Michael I Jordan, Jason D Lee, and Yun Yang. Communication-efficient distributed statistical inference. Journal of the American Statistical Association, pages 1--14, 2018

Show all 41 references
  1. [9]

    Sgdlibrary: A matlab library for stochastic optimization algorithms

    Hiroyuki Kasai. Sgdlibrary: A matlab library for stochastic optimization algorithms. Journal of Machine Learning Research, 18: 0 215--1, 2017

  2. [10]

    Federated optimization: Distributed machine learning for on-device intelligence

    Jakub Kone c n \`y , H Brendan McMahan, Daniel Ramage, and Peter Richt \'a rik. Federated optimization: Distributed machine learning for on-device intelligence. arXiv preprint arXiv:1610.02527, 2016

  3. [11]

    Distributed stochastic variance reduced gradient methods by sampling extra data with replacement

    Jason D Lee, Qihang Lin, Tengyu Ma, and Tianbao Yang. Distributed stochastic variance reduced gradient methods by sampling extra data with replacement. Journal of Machine Learning Research, 18 0 (1): 0 4404--4446, 2017

  4. [12]

    Rcv1: A new benchmark collection for text categorization research

    David D Lewis, Yiming Yang, Tony G Rose, and Fan Li. Rcv1: A new benchmark collection for text categorization research. Journal of Machine Learning Research, 5 0 (Apr): 0 361--397, 2004

  5. [13]

    Communication efficient distributed machine learning with the parameter server

    Mu Li, David G Andersen, Alex J Smola, and Kai Yu. Communication efficient distributed machine learning with the parameter server. In Advances in Neural Information Processing Systems (NIPS), 2014

  6. [14]

    A universal catalyst for first-order optimization

    Hongzhou Lin, Julien Mairal, and Zaid Harchaoui. A universal catalyst for first-order optimization. In Advances in Neural Information Processing Systems (NIPS), pages 3384--3392, 2015

  7. [15]

    Distributed inexact newton-type pursuit for non-convex sparse learning

    Bo Liu, Xiao-Tong Yuan, Lezi Wang, Qingshan Liu, Junzhou Huang, and Dimitris Metaxas. Distributed inexact newton-type pursuit for non-convex sparse learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), 2019

  8. [16]

    Linearly convergent stochastic heavy ball method for minimizing generalization error

    Nicolas Loizou and Peter Richt \'a rik. Linearly convergent stochastic heavy ball method for minimizing generalization error. arXiv preprint arXiv:1710.10737, 2017

  9. [17]

    Adding vs

    Chenxin Ma, Virginia Smith, Martin Jaggi, Michael Jordan, Peter Richtarik, and Martin Takac. Adding vs. averaging in distributed primal-dual optimization. In International Conference on Machine Learning(ICML), pages 1973--1982, 2015

  10. [18]

    Communication-efficient learning of deep networks from decentralized data

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 1273--1282, 2017

  11. [19]

    The landscape of empirical risk for nonconvex losses

    Song Mei, Yu Bai, Andrea Montanari, et al. The landscape of empirical risk for nonconvex losses. The Annals of Statistics, 46 0 (6A): 0 2747--2774, 2018

  12. [20]

    B. Polyak. Some methods of speeding up the convergence of iteration methods. USSR Computational Mathematics and Mathematical Physics , 4 0 (5): 0 1--17, 1964

  13. [21]

    N. Qian. On the momentum term in gradient descent learning algorithms. Neural networks, 12 0 (1): 0 145--151, 1999

  14. [22]

    AIDE : Fast and communication efficient distributed optimization

    Sashank J Reddi, Jakub Kone c n \`y , Peter Richt \'a rik, Barnab \'a s P \'o cz \'o s, and Alex Smola. AIDE : Fast and communication efficient distributed optimization. arXiv preprint arXiv:1608.06879, 2016

  15. [23]

    Distributed coordinate descent method for learning with big data

    Peter Richt \'a rik and Martin Tak \'a c . Distributed coordinate descent method for learning with big data. Journal of Machine Learning Research, 17 0 (1): 0 2657--2681, 2016

  16. [24]

    Stochastic convex optimization

    Shai Shalev-Shwartz, Ohad Shamir, Nathan Srebro, and Karthik Sridharan. Stochastic convex optimization. In Annual Conference on Learning Theory (COLT), 2009

  17. [25]

    Without-replacement sampling for stochastic gradient methods

    Ohad Shamir. Without-replacement sampling for stochastic gradient methods. In Advances in Neural Information Processing Systems (NIPS), pages 46--54, 2016

  18. [26]

    Communication-efficient distributed optimization using an approximate newton-type method

    Ohad Shamir, Nati Srebro, and Tong Zhang. Communication-efficient distributed optimization using an approximate newton-type method. In International Conference on Machine Learning (ICML), pages 1000--1008, 2014

  19. [27]

    Cocoa: A general framework for communication-efficient distributed optimization

    Virginia Smith, Simone Forte, Ma Chenxin, Martin Tak \'a c , Michael I Jordan, and Martin Jaggi. Cocoa: A general framework for communication-efficient distributed optimization. Journal of Machine Learning Research, 18: 0 230, 2018

  20. [28]

    User-friendly tail bounds for sums of random matrices

    Joel A Tropp. User-friendly tail bounds for sums of random matrices. Foundations of Computational Mathematics, 12 0 (4): 0 389--434, 2012

  21. [29]

    Efficient distributed learning with sparsity

    Jialei Wang, Mladen Kolar, Nathan Srebro, and Tong Zhang. Efficient distributed learning with sparsity. In International Conference on Machine Learning (ICML), pages 3636--3645, 2017 a

  22. [30]

    Memory and communication efficient distributed stochastic optimization with minibatch prox

    Jialei Wang, Weiran Wang, and Nathan Srebro. Memory and communication efficient distributed stochastic optimization with minibatch prox. In Annual Conference on Learning Theory (COLT), pages 1882--1919, 2017 b

  23. [31]

    Giant: Globally improved approximate newton method for distributed optimization

    Shusen Wang, Farbod Roosta-Khorasani, Peng Xu, and Michael W Mahoney. Giant: Globally improved approximate newton method for distributed optimization. In Advances in Neural Information Processing Systems (NeurIPS), pages 2338--2348, 2018

  24. [32]

    A lyapunov analysis of momentum methods in optimization

    Ashia C Wilson, Benjamin Recht, and Michael I Jordan. A lyapunov analysis of momentum methods in optimization. arXiv preprint arXiv:1611.02635, 2016

  25. [33]

    Dscovr: Randomized primal-dual block coordinate algorithms for asynchronous distributed optimization

    Lin Xiao, Adams Wei Yu, Qihang Lin, and Weizhu Chen. Dscovr: Randomized primal-dual block coordinate algorithms for asynchronous distributed optimization. Journal of Machine Learning Research, 20 0 (43): 0 1--58, 2019

  26. [34]

    Petuum: A new platform for distributed machine learning on big data

    Eric P Xing, Qirong Ho, Wei Dai, Jin Kyu Kim, Jinliang Wei, Seunghak Lee, Xun Zheng, Pengtao Xie, Abhimanu Kumar, and Yaoliang Yu. Petuum: A new platform for distributed machine learning on big data. IEEE Transactions on Big Data, 1 0 (2): 0 49--67, 2015

  27. [35]

    Xin, Patrick Wendell, Tathagata Das, Michael Armbrust, Ankur Dave, Xiangrui Meng, Josh Rosen, Shivaram Venkataraman, Michael J

    Matei Zaharia, Reynold S. Xin, Patrick Wendell, Tathagata Das, Michael Armbrust, Ankur Dave, Xiangrui Meng, Josh Rosen, Shivaram Venkataraman, Michael J. Franklin, Ali Ghodsi, Joseph Gonzalez, Scott Shenker, and Ion Stoica. Apache spark: A unified engine for big data processin...

  28. [36]

    Di SCO : Distributed optimization for self-concordant empirical loss

    Yuchen Zhang and Lin Xiao. Di SCO : Distributed optimization for self-concordant empirical loss. In International Conference on Machine Learning (ICML), pages 362--370, 2015

  29. [37]

    Stochastic primal-dual coordinate method for regularized empirical risk minimization

    Yuchen Zhang and Lin Xiao. Stochastic primal-dual coordinate method for regularized empirical risk minimization. Journal of Machine Learning Research, 18 0 (1): 0 2939--2980, 2017

  30. [38]

    Efficient stochastic gradient hard thresholding

    Pan Zhou, Xiaotong Yuan, and Jiashi Feng. Efficient stochastic gradient hard thresholding. In Advances in Neural Information Processing Systems (NeurIPS), pages 1984--1993, 2018

  31. [39]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  32. [40]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  33. [41]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.