Distributed Stochastic Gradient Method for Non-Convex Problems with Applications in Supervised Learning

He Bai; Jemin George; Prudhvi Gurram; Tao Yang

REVIEW 2 major objections 3 minor 33 references

Distributed Stochastic Gradient Method for Non-Convex Problems with Applications in Supervised Learning

T0 review · 2 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A peer-to-peer SGD algorithm drives networked agents to a critical point of their nonconvex objective, and ten agents training separate MNIST networks nearly match a central trainer.

desk verdict Solid but standard distributed SGD convergence analysis; the proof is fine, but the experiments do not satisfy the paper's own Lipschitz and step-size assumptions. read the letter →

arxiv 1908.06693 v1 pith:36ZY6VI7 submitted 2019-08-19 math.OC cs.SYeess.SY

Jemin George , Tao Yang , He Bai , Prudhvi Gurram This is my paper · ORCID

classification math.OCcs.SYeess.SY MSC 90C2662L2068W15

keywords distributedstochasticgradientdescentnon-convexoptimizationmean-squareconvergenceconsensusmachinelearningneuralnetworksMNIST

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 aims to close a gap: distributed stochastic gradient methods were mostly analyzed for convex problems, and the few non-convex analyses made restrictive assumptions on critical points. It proposes a fully distributed SGD update in which each agent takes a stochastic gradient step toward its own data and a consensus step toward its neighbors' parameter vectors, with no server. The central theoretical claim is that, under smoothness and step-size conditions, this update drives the expected squared norm of the average gradient to zero, i.e., mean-square convergence to a critical point. The paper then shows numerically that ten agents training separate single-hidden-layer networks on a 5000-image MNIST subset reach 7.36% test error, versus 7.12% for centralized training, even when each agent only sees images of one digit.

What carries the argument

The load-bearing object is the descent potential $V(\gamma_k, w_k) = F(w_k) + \frac{1}{2\gamma_k} w_k^\top (L \otimes I_{d_w}) w_k$, with $\gamma_k = \alpha_k/\beta_k$. The first term is the sum of the agents' local objective functions and the second penalizes disagreement across the graph through the Laplacian $L$; its gradient $\nabla V(\gamma_k, w_k) = \nabla F(w_k) + \frac{1}{\gamma_k}(L \otimes I_{d_w}) w_k$ is exactly the expected one-step update direction of the algorithm. The analysis shows a descent inequality on this quantity with a diminishing error term, applies the almost-supermartingale convergence lemma to obtain summability of $\alpha_k \mathbb{E}[\|\nabla V\|^2]$, and finally uses the Hessian Lipschitz assumption to turn that summability into convergence of the expected average gradient norm to zero.

What would settle it

Run update (8) on a two-agent ring with $f_i(w)=\sin(w_i)$ and step sizes satisfying Assumption 3. Theorem 5 predicts the expected squared average gradient tends to zero; a simulation in which this quantity levels off above zero would refute the central claim.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 5: for the distributed SGD update $w(k+1) = (W_k \otimes I_{d_w}) w(k) - \alpha_k g(w(k), \xi(k))$ with $W_k = I_n - \beta_k L$, under assumptions that each local objective is twice continuously differentiable with Lipschitz continuous gradients and Hessians, the expected squared norm of the network-averaged gradient tends to zero, $\lim_{k\to\infty} \mathbb{E}[\|\overline{\nabla F}(w_k)\|_2^2] = 0$. The paper interprets this as mean-square convergence of all agents' parameter estimates to a critical point of the aggregate non-convex objective. The proof first shows the consensus error decays at rate $O((k+1)^{-\delta_2})$, then establishes summability of a weighted expected gradient norm via a descent-potential argument, and then uses the Hessian Lipschitz condition to upgrade a $\liminf$ statement to a full limit. The paper states plainly that the limit point may be a saddle point; escaping saddles is left to future work.

Load-bearing premise

The load-bearing premise is that every local objective has gradients and Hessians that are Lipschitz continuous over the entire unbounded parameter space; the sigmoid neural-network objective used in the experiments does not satisfy this, so the theorem does not directly cover the demonstrated application.

Editorial extensions

If this is right

For any connected undirected graph, step-size sequences $\alpha_k = a/(k+1)^{\delta_2}$ and $\beta_k = b/(k+1)^{\delta_1}$ with $0 < 3\delta_1 < \delta_2 \le 1$, $\delta_1+\delta_2>1$, and $\delta_2>1/2$ are sufficient for the algorithm to reach consensus and converge in mean square to a critical point.
The architecture is fully peer-to-peer: agents exchange only their parameter vectors with neighbors, so raw training data stays local and no server is required.
The MNIST demonstration indicates that performance is essentially independent of how data is split: evenly distributed data and single-class-per-agent data both give 7.36% test error, close to the centralized 7.12%.
The result is a convergence-to-critical-point guarantee, not a global-optimum guarantee; the paper explicitly notes saddle points are included in the limit set.

Reading between the lines

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

Because the theorem requires globally Lipschitz gradients and Hessians but the demonstration uses sigmoid activations and cross-entropy loss, whose gradients and Hessians are not globally Lipschitz over all of $\mathbb{R}^{d_w}$, the experiments should be read as numerical evidence rather than direct verification of Theorem 5; a natural test is to repeat the experiment with a compactly parameteriz
The analysis derives bounded stochastic gradients from the Lipschitz assumptions (Proposition 1) instead of merely assuming bounded variance; this suggests that in practice gradient clipping may be needed to stay inside the theory, and its effect on the observed convergence could be measured directly.
The single-class-per-agent experiment implies a possible privacy-preserving specialization benefit: each agent can learn a distinctive local class distribution while still contributing to a joint model that recognizes all classes; the paper does not quantify communication cost, which would be a natural follow-up measure.
The theorem assumes a fixed connected undirected graph, but the proof's consensus-decay estimate may extend to time-varying or directed graphs; testing the same update on a switching topology would show how much of the guarantee survives outside the stated assumptions.

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.

Desk Editor's Note

read the letter

Colleague,

Quick take on arXiv:1908.06693. The main contribution is a mean-square convergence result for a consensus-based distributed SGD in non-convex problems, extending Zeng and Yin's deterministic analysis to the stochastic setting. That's a real but incremental step. The Lyapunov machinery is conventional, and Theorem 5—which upgrades liminf to full convergence of the average gradient norm using Hessian Lipschitzness—is the most interesting part.

I want to correct one thing in the reader's report: there is no algebra error in Theorem 1. The text correctly uses (1+βλ)(1−βλ) = 1−β²λ². That concern is a misreading.

The soft spots are all at the theory-experiment interface. First, the step sizes in Section V are α_k = 1/(εk+1) and β_k = b/(εk+1)^{1/3}, giving exponents δ2=1 and δ1=1/3, so 3δ1 = δ2 exactly. Assumption 3 requires strict inequality 3δ1 < δ2. The paper says the scaling factor ε doesn't affect the theory, but it never addresses the equality. The proofs might survive with a slightly weaker rate, but as written, the experiment sits outside the stated assumptions. Second, and more importantly, the sigmoid neural network objective is not globally Lipschitz-smooth on R^{d_w}. The stress-test's construction—growing a second-layer weight along a fixed hidden activation—shows the gradient norm grows linearly, so Assumptions 1 and 8 fail. Thus Theorem 5 does not actually cover the reported application. This is a common sin in the optimization-for-ML literature, but it's still a real gap.

The experiments also only report risk and test error, not the quantities the theory bounds, like E[||∇F(w_k)||²] or the consensus error. No error bars or code, either.

Who is this for? Someone in decentralized non-convex optimization who wants a clean template for almost-supermartingale proofs. It won't reshape the field, but the theory is solid and the exposition is readable. The experiment mismatch is fixable; a revision that acknowledges the step-size boundary and the Lipschitz limitation would strengthen it substantially.

I'd send this to peer review. It deserves referee time.

Referee Report

2 major / 3 minor

Summary. The paper develops a distributed stochastic gradient descent algorithm for non-convex optimization over a network of agents. Each agent updates its local parameter by a consensus step on the graph Laplacian and a stochastic gradient step. Under assumptions that the local objective functions have globally Lipschitz continuous gradients and Hessians, and under explicit step-size conditions, the paper proves that the consensus error vanishes and that the expected squared norm of the average gradient tends to zero. The authors then apply the algorithm to distributed training of ten two-layer sigmoid neural networks on MNIST, reporting test errors comparable to centralized training for both random and class-split data distributions.

Significance. If the convergence claims hold, the paper supplies a clean sufficient-condition result for non-convex distributed stochastic optimization, a regime where the literature is sparse. The Lyapunov-based argument is self-contained, the step-size conditions are explicit, and the proof does not rely on circular reasoning or fitted parameters. The experimental setup, including a ring topology and heterogeneous data partitions, addresses a practically relevant federated-learning scenario. The central theorem, Theorem 5, appears sound; I do not find the algebra issue suggested in the reader's report, because the identity (1+βλ)(1−βλ)=1−β^2λ^2 is correct and the displayed bound is a valid relaxation of the inequality obtained from (38). The main weaknesses are in the bridge between the theory and the numerical demonstration, which are addressed in the major comments below.

major comments (2)

[Section V, step-size choices] The experiments in Section V use α_k = 1/(εk+1) and β_k = b/(εk+1)^{1/3}, which correspond to δ2=1 and δ1=1/3. This gives 3δ1 = δ2, violating the strict inequality 3δ1 < δ2 required in Assumption 3. In the proof of Theorem 1, the rate estimate (41) is available only for δ0 < 2δ2−3δ1; with these values 2δ2−3δ1 = 1, so the claimed rate O((k+1)^{−δ2}) with δ2=1 is not covered. The statement that the scale factor ε does not affect the theory does not repair this, because the boundary violation is independent of ε. Please either adjust the experimental step sizes to lie strictly inside Assumption 3 or explicitly state that the simulations are a heuristic demonstration outside the theorem's hypotheses.
[Section V and Assumptions 1, 8] The empirical risk for the two-layer sigmoid network with cross-entropy loss in Section V does not satisfy the global Lipschitz assumptions on gradients and Hessians over R^{d_w}. For example, scaling the second-layer weights by a large factor M while keeping hidden activations fixed introduces a factor linear in M into the derivative with respect to the first-layer weights, so finite constants L and L_H in (12) and (32) cannot exist. Consequently, Theorem 5's conclusion (33) does not apply to the reported application. The paper should either restrict the parameter space (e.g., by considering a compact domain or adding regularization) so that the assumptions hold, or clearly re-label the experiments as numerical illustrations outside the theoretical guarantee.

minor comments (3)

[Section V, experimental validation] The experimental section reports only empirical risk and test error, not the quantities controlled by Theorems 1 and 5, namely E‖\bar{∇F}(w_k)‖² and E‖\tilde{w}_k‖². Including a plot of the empirical average-gradient norm or consensus error would directly connect the numerics to the theory.
[Abstract and Conclusion] The phrase 'asymptotic mean-square convergence to a critical point' is stronger than what Theorem 5 proves; the theorem establishes E‖\bar{∇F}(w_k)‖² → 0, not convergence of the iterates themselves. Please adjust the wording to 'mean-square convergence of the average gradient to zero' or qualify the statement accordingly.
[Appendix B, proof of Theorem 1] The derivation from inequality (38) to (39) skips one relaxation step: the direct substitution gives a coefficient (1+β_kλ_2)(1−β_kλ_2)^2, and the displayed coefficient (1+β_kλ_2)(1−β_kλ_2) is an upper bound. Adding one line to show this monotonicity would make the proof easier to follow.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the convergence proof is a self-contained Lyapunov/supermartingale derivation from stated assumptions, with no fitted parameter or self-citation chain doing load-bearing work.

full rationale

The paper's main convergence claims, Theorems 1-5, are derived from Assumptions 1-8 by a standard Lyapunov and almost-supermartingale argument. Theorem 5's conclusion, lim_k E[||\overline{\nabla F}(w_k)||^2] = 0, is not assumed in any assumption or defined into existence; it follows from the weighted summability of E[||\nabla V(\gamma_k,w_k)||^2] in Theorem 2, the summability of E[||w_{k+1}-w_k||^2] in Theorem 3, and the convergence of E[||\overline{\nabla F}(w_k)||^2] established by applying Lemma 5. The auxiliary lemmas used are cited from external sources: Lemma 1 from [25], Lemma 4 from [32], Lemma 5 from [33], and Lemma 6 from [17]; none of these are self-citations of the present authors, and none simply restate the target conclusion. The numerical MNIST section is an application/demonstration rather than a fitted prediction relabeled as theory: the reported error rates are experimental outcomes, not claimed outputs of the theorem derived from fitted constants. Even if the sigmoid cross-entropy objective does not satisfy the theorem's global Lipschitz assumptions, that is an applicability gap between theory and experiment, not circularity. No equation in the derivation reduces to its own input by construction, and no parameter is fitted to data and then presented as a prediction. The strongest theorem is therefore an independent convergence result, and the paper should not be scored as circular.

Assumptions & free parameters 4 free parameters · 8 assumptions · 0 invented entities

The convergence proof rests on eight assumptions about the objective functions, the graph, and the stochastic gradients. None of these are proved; they are domain assumptions that define the problem class. The experiments further introduce a neural network objective that does not satisfy the global Lipschitz assumptions, and step-size hyperparameters chosen by hand.

free parameters (4)

δ1 = 1/3 (experiments)
Exponent of β_k in (16); chosen as 1/3 in the numerical experiments, which violates the strict condition 3δ1 < δ2.
δ2 = 1 (experiments)
Exponent of α_k in (16); chosen as 1 with ε scaling in the experiments.
a = 1/ε = 1e5 (experiments)
Scaling constant of α_k in (16); set via ε=1e-5 in experiments.
b = 0.2525 (experiments)
Scaling constant of β_k in (16); chosen by hand in experiments.

assumptions (8)

domain assumption Assumption 1: Each local objective f_i and its gradient ∇f_i are Lipschitz continuous with constants L0_i and L_i.
Used throughout the proof, e.g., in Lemma 2 and inequality (12).
domain assumption Assumption 2: The aggregate objective F is lower bounded by F_inf.
Required for the Lyapunov function V to be bounded below in Theorem 2.
domain assumption Assumption 3: Step sizes are α_k=a/(k+1)^{δ2}, β_k=b/(k+1)^{δ1} with 0<3δ1<δ2≤1, δ1+δ2>1, δ2>1/2.
This decay condition is the core sufficient condition for convergence; it is not derived, it is imposed.
domain assumption Assumption 4: The interaction graph is connected and undirected.
Needed for the Laplacian eigenvalue properties in Lemma 3 and the consensus contraction.
domain assumption Assumption 5: b is chosen so that W0=I_n−bL has a single eigenvalue at 1.
Ensures the consensus step is stable at each iteration.
domain assumption Assumption 6: Stochastic gradients are unbiased, Eξ[g(w_k,ξ_k)] = ∇F(w_k).
Standard unbiasedness condition used in (19) and in the proof of Theorem 2.
domain assumption Assumption 7: The second moment of the stochastic gradient is bounded by μv1+μv2||∇F(w)||^2.
Bounded variance assumption, used in Proposition 1 and the proof of Theorem 2.
domain assumption Assumption 8: The Hessians ∇^2 f_i are Lipschitz continuous.
Needed only for Theorem 5 to strengthen the liminf result to a full limit.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Distributed Stochastic Gradient Method for Non-Convex Problems with Applications in Supervised Learning." pith.science (2026). https://pith.science/paper/36ZY6VI7

@misc{pith2026190806693,
  author       = {Pith},
  title        = {Pith review of: Distributed Stochastic Gradient Method for Non-Convex Problems with Applications in Supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/36ZY6VI7}},
  note         = {Machine review of arXiv:1908.06693}
}

read the original abstract

We develop a distributed stochastic gradient descent algorithm for solving non-convex optimization problems under the assumption that the local objective functions are twice continuously differentiable with Lipschitz continuous gradients and Hessians. We provide sufficient conditions on step-sizes that guarantee the asymptotic mean-square convergence of the proposed algorithm. We apply the developed algorithm to a distributed supervised-learning problem, in which a set of networked agents collaboratively train their individual neural nets to recognize handwritten digits in images. Results indicate that all agents report similar performance that is also comparable to the performance of a centrally trained neural net. Numerical results also show that the proposed distributed algorithm allows the individual agents to recognize the digits even though the training data corresponding to all the digits is not locally available to each agent.

Figures

Figures reproduced from arXiv: 1908.06693 by the authors.

**Figure 2.** Empirical risk for all three experiments and a few mis [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 29 canonical work pages

[1]

Scaling distributed ma chine learning with the parameter server,

M. Li, D. G. Andersen, J. W. Park, A. J. Smola, A. Ahmed, V . J osifovski, J. Long, E. J. Shekita, and B.-Y . Su, “Scaling distributed ma chine learning with the parameter server,” in 11th USENIX Symposium on Operating Systems Design and Implementation (OSDI) , 2014, pp. 583 – 598

work page 2014
[2]

A comparison of distributed machine learning platforms,

K. Zhang, S. Alqahtani, and M. Demirbas, “A comparison of distributed machine learning platforms,” in 26th International Conference on Com- puter Communication and Networks (ICCCN) , Jul. 2017, pp. 1–9

work page 2017
[3]

An adaptive synchronous parallel strategy for distributed ma chine learning,

J. Zhang, H. Tu, Y . Ren, J. Wan, L. Zhou, M. Li, and J. Wang, “ An adaptive synchronous parallel strategy for distributed ma chine learning,” IEEE Access , vol. 6, pp. 19 222–19 230, 2018

work page 2018
[4]

Communicati on efﬁcient distributed machine learning with the parameter s erver,

M. Li, D. G. Andersen, A. J. Smola, and K. Y u, “Communicati on efﬁcient distributed machine learning with the parameter s erver,” in Advances in Neural Information Processing Systems , 2014, pp. 19 – 27

work page 2014
[5]

Federated learning: Strategies for improving co mmunication efﬁciency,

J. Kone˘ cn´ u, H. B. McMahan, F. X. Y u, P . Richtarik, A. T. S uresh, and D. Bacon, “Federated learning: Strategies for improving co mmunication efﬁciency,” in NIPS W orkshop on Private Multi-Party Machine Learning, 2016

work page 2016
[6]

Communication-efﬁcient learning of deep networks from de centralized data,

H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y A rcas, “Communication-efﬁcient learning of deep networks from de centralized data,” in Proceedings of the 20th International Conference on Artiﬁc ial Intelligence and Statistics (AISTATS) , 2017

work page 2017
[7]

Optimization meth ods for large- scale machine learning,

L. Bottou, F. Curtis, and J. Nocedal, “Optimization meth ods for large- scale machine learning,” SIAM Review , vol. 60, no. 2, pp. 223–311, 2018

work page 2018
[8]

An approximate dual subgradient a lgorithm for multi-agent non-convex optimization,

M. Zhu and S. Martnez, “An approximate dual subgradient a lgorithm for multi-agent non-convex optimization,” IEEE Transactions on Automatic Control, vol. 58, no. 6, pp. 1534 – 1539, Jun. 2013

work page 2013

Show all 33 references

[9]

Nestt: A non convex primal-dual splitting method for distributed and stochast ic optimization,

D. Hajinezhad, M. Hong, T. Zhao, and Z. Wang, “Nestt: A non convex primal-dual splitting method for distributed and stochast ic optimization,” in Advances in Neural Information Processing Systems , 2016, pp. 3215– 3223

2016
[10]

Prox-PDA: The p roximal primal-dual algorithm for fast distributed nonconvex opti mization and learning over networks,

M. Hong, D. Hajinezhad, and M.-M. Zhao, “Prox-PDA: The p roximal primal-dual algorithm for fast distributed nonconvex opti mization and learning over networks,” in Proc. of the 34th International Conference on Machine Learning , Aug. 2017, pp. 1529 – 1538

2017
[11]

On the converg ence of a distributed augmented lagrangian method for nonconvex opt imization,

N. Chatzipanagiotis and M. M. Zavlanos, “On the converg ence of a distributed augmented lagrangian method for nonconvex opt imization,” IEEE Transactions on Automatic Control , vol. 62, no. 9, pp. 4405–4420, Sep. 2017

2017
[12]

Paralle l and distributed methods for constrained nonconvex optimizationpart i: The ory,

G. Scutari, F. Facchinei, and L. Lampariello, “Paralle l and distributed methods for constrained nonconvex optimizationpart i: The ory,” IEEE Transactions on Signal Processing , vol. 65, no. 8, pp. 1929 – 1944, Apr. 2017

1929
[13]

NEXT: In-network nonconv ex optimiza- tion,

P . D. Lorenzo and G. Scutari, “NEXT: In-network nonconv ex optimiza- tion,” IEEE Transactions on Signal and Information Processing ove r Networks, vol. 2, no. 2, pp. 120–136, Jun. 2016

2016
[14]

A distributed, asynchronous, and incrementa l algorithm for nonconvex optimization: An ADMM approach,

M. Hong, “A distributed, asynchronous, and incrementa l algorithm for nonconvex optimization: An ADMM approach,” IEEE Transactions on Control of Network Systems , vol. 5, no. 3, pp. 935–945, Sep. 2018

2018
[15]

A case for nonconvex dis tributed optimization in large-scale power systems,

J. Guo, G. Hug, and O. K. Tonguz, “A case for nonconvex dis tributed optimization in large-scale power systems,” IEEE Transactions on Power Systems, vol. 32, no. 5, pp. 3842 – 3851, Sep. 2017

2017
[16]

Convergence analys is of alternating direction method of multipliers for a family of nonconvex problems,

M. Hong, Z. Luo, and M. Razaviyayn, “Convergence analys is of alternating direction method of multipliers for a family of nonconvex problems,” SIAM Journal on Optimization , vol. 26, no. 1, pp. 337–364, 2016

2016
[17]

On nonconvex decentralized gradien t descent,

J. Zeng and W. Yin, “On nonconvex decentralized gradien t descent,” IEEE Transactions on Signal Processing , vol. 66, no. 11, pp. 2834– 2848, Jun. 2018

2018
[18]

Zone: Zeroth ord er nonconvex multi-agent optimization over networks,

D. Hajinezhad, M. Hong, and A. Garcia, “Zone: Zeroth ord er nonconvex multi-agent optimization over networks,” IEEE Transactions on Auto- matic Control, vol. Early Access, pp. 1–1, 2019

2019
[19]

Stochastic gradient-push for strongly convex functions on time-varying directed graphs,

A. Nedi´ c and A. Olshevsky, “Stochastic gradient-push for strongly convex functions on time-varying directed graphs,” IEEE Transactions on Automatic Control , vol. 61, no. 12, pp. 3936–3947, Dec. 2016

2016
[20]

Conver gence rates for distributed stochastic optimization over random netwo rks,

D. Jakovetic, D. Bajovic, A. K. Sahu, and S. Kar, “Conver gence rates for distributed stochastic optimization over random netwo rks,” in Proc. IEEE Conference on Decision and Control (CDC) , Dec 2018, pp. 4238– 4245

2018
[21]

SUCAG : Stochastic unbiased curvature-aided gradient method for distributed optimization,

H. Wai, N. M. Freris, A. Nedi´ c, and A. Scaglione, “SUCAG : Stochastic unbiased curvature-aided gradient method for distributed optimization,” in Proc. IEEE Conference on Decision and Control (CDC) , Dec. 2018, pp. 1751–1756

2018
[22]

Distributed stochastic gradient t racking methods,

S. Pu and A. Nedi´ c, “Distributed stochastic gradient t racking methods,” arXiv e-prints , p. arXiv:1805.11454, May 2018

2018 arXiv
[23]

Non-convex distributed opt imization,

T. Tatarenko and B. Touri, “Non-convex distributed opt imization,” IEEE Transactions on Automatic Control, vol. 62, no. 8, pp. 3744 – 3757, Aug. 2017

2017
[24]

Convergence of a multi-a gent projected stochastic gradient algorithm for non-convex optimizatio n,

P . Bianchi and J. Jakubowicz, “Convergence of a multi-a gent projected stochastic gradient algorithm for non-convex optimizatio n,” IEEE Trans- actions on Automatic Control , vol. 58, no. 2, pp. 391–405, Feb. 2013

2013
[25]

Khalil, Nonlinear Systems

H. Khalil, Nonlinear Systems . Upper Saddle River, NJ: Prentice Hall, 2002, ch. 3

2002
[26]

R. A. Horn and C. R. Johnson, Matrix Analysis , 2nd ed. Cambridge University Press, 2012

2012
[27]

Distributed subgradient me thods for multi- agent optimization,

A. Nedi´ c and A. Ozdaglar, “Distributed subgradient me thods for multi- agent optimization,” IEEE Transactions on Automatic Control , vol. 54, no. 1, pp. 48–61, Jan. 2009

2009
[28]

First-order methods almost always avoid saddle p oints,

J. D. Lee, I. Panageas, G. Piliouras, M. Simchowitz, M. I . Jordan, and B. Recht, “First-order methods almost always avoid saddle p oints,” arXiv e-prints, p. arXiv:1710.07406, Oct 2017

2017 arXiv
[29]

Sharp analysis for noncon vex sgd escaping from saddle points,

C. Fang, Z. Lin, and T. Zhang, “Sharp analysis for noncon vex sgd escaping from saddle points,” arXiv e-prints , p. arXiv:1902.00247, Feb 2019

1902 arXiv
[30]

Stochastic gradient descent escapes saddle points efﬁciently,

C. Jin, P . Netrapalli, R. Ge, S. M. Kakade, and M. I. Jorda n, “Stochastic gradient descent escapes saddle points efﬁciently,” arXiv e-prints , p. arXiv:1902.04811, Feb 2019

1902 arXiv
[31]

Bishop, Pattern Recognition and Machine Learning , ser

C. Bishop, Pattern Recognition and Machine Learning , ser. Information Science and Statistics. New Y ork, NY: Springer, 2016

2016
[32]

Distributed linear param eter estimation: Asymptotically efﬁcient adaptive strategies,

S. Kar, J. Moura, and H. Poor, “Distributed linear param eter estimation: Asymptotically efﬁcient adaptive strategies,” SIAM Journal on Control and Optimization , vol. 51, no. 3, pp. 2200–2229, 2013

2013
[33]

A convergence theorem for n on negative almost supermartingales and some applications

H. Robbins and D. Siegmund, “A convergence theorem for n on negative almost supermartingales and some applications.” in Optimizing Methods in Statistics , J. S. Rustagi, Ed. Academic Press, 1971, pp. 233 – 257

1971

Pith tools

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

[1] [1]

Scaling distributed ma chine learning with the parameter server,

M. Li, D. G. Andersen, J. W. Park, A. J. Smola, A. Ahmed, V . J osifovski, J. Long, E. J. Shekita, and B.-Y . Su, “Scaling distributed ma chine learning with the parameter server,” in 11th USENIX Symposium on Operating Systems Design and Implementation (OSDI) , 2014, pp. 583 – 598

work page 2014

[2] [2]

A comparison of distributed machine learning platforms,

K. Zhang, S. Alqahtani, and M. Demirbas, “A comparison of distributed machine learning platforms,” in 26th International Conference on Com- puter Communication and Networks (ICCCN) , Jul. 2017, pp. 1–9

work page 2017

[3] [3]

An adaptive synchronous parallel strategy for distributed ma chine learning,

J. Zhang, H. Tu, Y . Ren, J. Wan, L. Zhou, M. Li, and J. Wang, “ An adaptive synchronous parallel strategy for distributed ma chine learning,” IEEE Access , vol. 6, pp. 19 222–19 230, 2018

work page 2018

[4] [4]

Communicati on efﬁcient distributed machine learning with the parameter s erver,

M. Li, D. G. Andersen, A. J. Smola, and K. Y u, “Communicati on efﬁcient distributed machine learning with the parameter s erver,” in Advances in Neural Information Processing Systems , 2014, pp. 19 – 27

work page 2014

[5] [5]

Federated learning: Strategies for improving co mmunication efﬁciency,

J. Kone˘ cn´ u, H. B. McMahan, F. X. Y u, P . Richtarik, A. T. S uresh, and D. Bacon, “Federated learning: Strategies for improving co mmunication efﬁciency,” in NIPS W orkshop on Private Multi-Party Machine Learning, 2016

work page 2016

[6] [6]

Communication-efﬁcient learning of deep networks from de centralized data,

H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y A rcas, “Communication-efﬁcient learning of deep networks from de centralized data,” in Proceedings of the 20th International Conference on Artiﬁc ial Intelligence and Statistics (AISTATS) , 2017

work page 2017

[7] [7]

Optimization meth ods for large- scale machine learning,

L. Bottou, F. Curtis, and J. Nocedal, “Optimization meth ods for large- scale machine learning,” SIAM Review , vol. 60, no. 2, pp. 223–311, 2018

work page 2018

[8] [8]

An approximate dual subgradient a lgorithm for multi-agent non-convex optimization,

M. Zhu and S. Martnez, “An approximate dual subgradient a lgorithm for multi-agent non-convex optimization,” IEEE Transactions on Automatic Control, vol. 58, no. 6, pp. 1534 – 1539, Jun. 2013

work page 2013