Pith. sign in

REVIEW 3 major objections 4 minor 28 references

Distributed gradient methods under heavy-tailed communication noise

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

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

Pith's one-line read A clipped-consensus distributed gradient method converges in MSE to an arbitrarily small neighborhood of the global minimizer under heavy-tailed, possibly infinite-variance communication noise.

desk verdict A genuinely new setting and a sensible algorithm, but two load-bearing proof gaps—the disagreement contraction in Lemma 7 and the a-scaling of the uniform bound—mean Theorem 1 is not established as written. read the letter →

arxiv 2505.24464 v1 pith:PQJBWQBC submitted 2025-05-30 math.OC cs.LG

Manojlo Vukovic , Dusan Jakovetic , Dragana Bajovic , Soummya Kar This is my paper
classification math.OCcs.LG MSC 90C2590C1568W15
keywords distributedoptimizationheavy-tailednoiseinfinitevarianceclippedconsensustwotime-scalestepsizesmeansquarederrorconvergencestronglyconvexwirelesssensornetworks
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 addresses distributed optimization over a network when messages between nodes are corrupted by heavy-tailed noise whose variance may be infinite, a setting that standard finite-variance analyses exclude. It proposes a gradient scheme in which the consensus part is passed through a bounded nonlinearity (clipping) and the consensus and gradient step sizes decay on two different time scales. Under strongly convex, heterogeneous local costs, it claims the mean squared error converges to a neighborhood of the global minimizer with rate $O(a^2 + 1/t^\gamma)$, where $a$ is the gradient step-size constant; by choosing $a$ small, the limiting error can be made arbitrarily small at the price of slower convergence. This matters because heavy-tailed interference is common in dense wireless and IoT networks, and existing distributed methods designed for finite-variance noise are shown by simulation to fail there.

What carries the argument

The load-bearing object is the noise-averaged nonlinearity $\phi(a) = \mathbb{E}[\Psi(a+\xi)]$, the expected value of the clipping function under the noise density. Because $\Psi$ is bounded, $\phi$ is bounded and odd and, by the Polyak lemma, $\phi'(0)>0$; the argument linearizes the consensus update through $\phi$ and splits the error into a disagreement component $\tilde{x}^t = (I-J)x^t$ and an average component. Lemma 7 shows the disagreement contracts by factor $1 - \beta_t \phi'(0)G/(2M_t)$, using the almost-sure bound $\|x^t\| \le M_t \sim t^{1-\delta}$ to control the mean-value points; combining that contraction with the average-dynamics recursion yields the $O(a^2 + t^{-\gamma})$ bound of Theorem 1.

What would settle it

Run algorithm (2) with $\Psi(a) = \arctan(a)$ and the noise density (22) with $\beta = 2.05$, as in the paper's Example 1, and record $D_t = \max_{i,j} \|x_i^t - x_j^t\|$. At each time $t$, numerically compute $\phi'(c) = \mathbb{E}[1/(1+(c+\xi)^2)]$ over $c \in [0, D_t]$ and compare with $\phi'(0)G/(2M_t)$. If for any observed $D_t$ the inequality fails while the recursion (17) is still being applied, that step of the proof is unjustified at that iterate; conversely, if the simulated MSE still tracks $O(a^2)$, the theorem's conclusion may survive but needs a different argument for large disagreements.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the update $x_i^{t+1} = x_i^t - \beta_t \sum_{j \in \Omega_i} \Psi(x_i^t - x_j^t + \xi_{ij}^t) - \alpha_t \nabla f_i(x_i^t)$, with $\alpha_t = a/(t+1)$, $\beta_t = b/(t+1)^\delta$, $\delta \in (1/2,1)$, and $\Psi$ bounded, odd, monotone, and differentiable, drives every node's estimate to a common point whose MSE distance from $x^\star$ is eventually at most $256 a^2 c_1'^2 (MN-1)(L \tilde{c} + \|\nabla F(\theta^\star)\|^2)/(((1-\delta)\phi'(0)G)^2) \cdot (1+4L/\mu)$, with finite-time bound $O(a^2) + O(t^{-\gamma})$. Because the bound scales as $a^2$, the asymptotic accuracy is user-tunable, while the rate $\gamma = \min(\phi'(0)G(1-\delta)/(4c_1'), a\mu)$ exhibits the accuracy-speed tradeoff. The same claim includes the negative results that the equal-time-scale clipping variant and the finite-variance method of [24] fail under infinite-variance noise.

Load-bearing premise

The contraction step that drives the whole proof assumes the derivative $\phi'(c)$ of the noise-averaged clipping function stays bounded below by $\phi'(0)G/(2M_t)$ not just near zero but for every disagreement size $c$ up to $M_t \sim t^{1-\delta}$; the paper establishes that lower bound only on a fixed neighborhood $|c| < g$.

Editorial extensions

If this is right

  • The asymptotic MSE can be driven below any desired threshold by picking the gradient step-size constant $a$ small enough, so the method offers a tunable accuracy level rather than only boundedness.
  • The convergence rate is governed by $\gamma = \min(\phi'(0)G(1-\delta)/(4c_1'), a\mu)$: smaller $a$ buys accuracy but linearly slows the transient error decay.
  • Finite-variance methods such as [24] cannot be safely deployed when link noise may have infinite variance; the clipping and the two time scales are both needed.
  • For realistic IoT and wireless deployments where interference has been modeled as heavy-tailed, the algorithm gives a first principled convergence guarantee rather than an empirical fix.
  • The equal-time-scale clipped variant (23) does not inherit the guarantee, so simply porting the estimation method of [11] to optimization is insufficient.

Reading between the lines

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

  • Editorial inference: the same two-time-scale clipped design could be combined with an adaptive or slowly decaying $a$ to seek exact convergence ($a_t \to 0$), a regime the paper does not analyze but which its tradeoff suggests is the natural next step.
  • Editorial inference: because the proof only needs $\phi'(c)$ bounded below on the actual range of disagreements, a case-split argument that treats large disagreements separately (where clipping saturates and the consensus term behaves like a sign function) could either complete the theorem or reveal a genuine counterexample.
  • Editorial inference: the comparison with finite-variance methods suggests a practical detector of infinite-variance noise could switch a network from standard consensus to the clipped two-time-scale update; the paper does not propose such a switching rule.
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 considers distributed strongly convex optimization over a fixed connected graph, with additive communication noise that may have infinite variance. It proposes a consensus-plus-gradient algorithm in which the consensus innovation is passed through a bounded odd nonlinearity and the consensus and gradient step sizes decay at different rates (beta_t = b/(t+1)^delta with delta in (1/2,1), alpha_t = a/(t+1)). The main result, Theorem 1, claims MSE convergence to a neighborhood of the global minimizer whose size can be made arbitrarily small by decreasing a, with a rate O(t^{-gamma}). The proof proceeds by bounding the disagreement from the global average (Lemma 7) and then the convergence of the global average, using a uniform MSE bound (Lemma 6) and a deterministic comparison lemma (Lemma 4). Numerical experiments on synthetic quadratic problems and two real-data logistic regression problems compare the method with an equal-time-scale variant and with the finite-variance method of [24].

Significance. The problem is timely: heavy-tailed and infinite-variance communication noise is well documented in wireless sensor and IoT settings, while existing distributed optimization analyses assume finite noise variance. The algorithmic idea of combining a bounded nonlinearity with mixed time scales is natural, and the paper gives explicit algorithm constants and a meaningful comparison with existing methods. If Theorem 1 were rigorously established, the paper would be a useful contribution to distributed optimization under heavy-tailed communication noise. The manuscript does not provide code or machine-checked proofs, and the numerical evidence is conventional Monte Carlo; the central value of the paper is the claimed theoretical guarantee, which is exactly where the current proof has serious gaps.

major comments (3)
  1. [§4.3, Lemma 7, Eq. (16)] The key contraction inequality ||z_t|| <= (1 - beta_t phi'(0)G/(2M_t))||xtilde_t|| + alpha_t||zeta_t|| is obtained by invoking Lemma 2 and Lemma 5, but Lemma 2 only gives phi'(c) >= phi'(0)G/(2g) for mean-value points c with |c| < g. Lemma 5 only provides ||x_t|| <= M_t with M_t = O(t^{1-delta}), so pairwise disagreements can be far outside (-g,g). For the paper's own example Psi = arctan, phi'(c) decays like c^{-2} for large c, so the required lower bound over the whole range |c| <= M_t is false. No saturation or case-split argument is supplied to handle large disagreements. Since this inequality is the source of the contraction in Eqs. (16)-(18), Lemma 7 and consequently Theorem 1 are not established as written.
  2. [§4.3, Theorem 1 and following paragraph] The displayed limsup bound in Theorem 1 is proportional to a^2 * tilde_c, where tilde_c is the uniform MSE bound from Lemma 6. In Lemma 6, the proof drops the beta_t^2 c_1^2 term in Eq. (14), which forces t_mse of order a^{-1/(2delta-1)}. Lemma 5 then gives tilde_c >= (t_mse)^{1-delta} = O(a^{-(1-delta)/(2delta-1)}). Hence the proof's limsup bound scales at best as a^{2-(1-delta)/(2delta-1)}, which diverges as a -> 0 whenever delta < 3/5; for the experimental value delta = 0.51 the exponent is approximately -22.5. Therefore the statement after Theorem 1 that 'the bound can be made arbitrarily small by setting parameter a small enough' does not follow from the proof. The proof would need a uniform-in-a bound for tilde_c, or a different argument controlling the limsup constant independently of tilde_c.
  3. [§4.2, Lemma 6] The choice of t_mse in Lemma 6 is the source of the bad a-dependence described above. A comparison recurrence that retains the t^{-2delta} forcing term instead of dropping it would lead to a finite limit independent of a, because the t^{-2delta} term is summable for delta > 1/2. The manuscript, however, applies Lemma 4 after discarding this term and thereby introduces the a-dependent threshold. This is not merely a cosmetic issue: Lemma 6, as stated and proved, is insufficient to support the small-a claim made in Theorem 1.
minor comments (4)
  1. [§4.2, Lemma 6 statement] The first entry of the max defining tilde_c, namely ||x0 - theta*|| + 2c_1' b (t_mse)^{1-delta}/(1-delta), is not squared, although tilde_c is used as a bound on E||x_t - theta*||^2. This should either be squared or the statement should be adjusted.
  2. [§4.3, Theorem 1 statement] The quantity tilde_c appears in the right-hand side of Theorem 1 but is not defined in the theorem statement; it should be explicitly tied to Lemma 6.
  3. [Abstract and Introduction] The informal claim that the MSE is 'of the order O(a^2 + 1/t^gamma)' is looser than the theorem, where the two contributions are not simply additive and the rate exponent is gamma = min(phi'(0)G(1-delta)/(4c_1'), a mu). Please restate the precise bound.
  4. [Throughout] There are numerous typos and stylistic issues, including 'communication frojjtoi' in Section 2, the capitalized 'We prove' in the introduction, and 'semipositive' for positive semidefinite. These should be corrected in a revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reduction: Theorem 1 does not reduce to a fitted or self-defined input; reliance on authors' earlier lemmas is technical, not premise-smuggling.

full rationale

The paper's derivation does not, by its own equations, reduce a predicted quantity to an input that was fitted or defined in terms of that quantity. The step sizes α_t = a/(t+1), β_t = b/(t+1)^δ and the nonlinearity Ψ are explicit design parameters, and no constant in the MSE bound is estimated from the numerical experiments; a, b, δ are set by the user, so the 'prediction' is not a refitted input. The proof imports Lemma 1 from Polyak [17] and the first part of Lemma 2 and Lemma 3 from the authors' prior work [10], and the φ-transformation framework from [11,21]; this is self-citation, but it is not circular in the sense prohibited here. The cited lemmas are parameter-free statements about φ (defined by an independent averaging identity) and about a deterministic scalar recursion; their assumptions (Assumptions 3-4 on p and Ψ) do not include the target MSE convergence of the distributed algorithm, and they are peer-reviewed results rather than assumptions invented for Theorem 1. Lemma 2's second part is proved in the text. The other two issues raised in the manuscript are correctness gaps, not circularity: (i) Lemma 7 applies the local lower bound of Lemma 2, valid for |a| < g, to pairwise disagreements as large as M_t ~ t^{1-δ}, which grows, so the contraction inequality is not established; (ii) the claimed a→0 accuracy is not supported by the proof because \tilde c in Lemma 6 is at least M_{t_mse} ≍ a^{-(1-δ)/(2δ-1)} while t_mse ≍ a^{-1/(2δ-1)}, making a^2\tilde c diverge as a→0 for δ near 1/2. Neither is a reduction of the conclusion to the hypotheses; both are mathematical weaknesses in the derivation chain. Therefore no circular step is identified and the circularity score is 0.

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

No new physical entities are postulated. The free parameters are explicit algorithm constants (a, b, delta) and the clipping nonlinearity, none fitted to data. The derivation leans on a chain of lemmas from [17] and from the authors' own [10, 11]; the most fragile is Lemma 2 (first part deferred to [10]) and its use at time-varying scale in Lemma 7, which is asserted without proof and is questionable for the paper's own examples.

free parameters (4)
  • a (gradient step-size constant) = 1.0, 0.5, 0.1, 0.05, 0.01 (experiments)
    alpha_t = a/(t+1). Bias of Theorem 1 is O(a^2) and rate is O(1/t^{a mu}); a is an explicit user-chosen design constant, not a hidden fit.
  • b (consensus step-size constant) = 1.0 (all experiments)
    beta_t = b/(t+1)^delta. Enters the analysis through M_t and c_tilde; user-chosen design constant.
  • delta (consensus step-size decay rate) = 0.51, 0.75, 0.99 (experiments), restricted to (0.5, 1)
    Decay rate of beta_t; the bias bound scales as 1/(1-delta)^2 and the disagreement rate as t^{-phi'(0)G(1-delta)/(4c1')}. The claim that delta outside (0.5, 1) causes failure is asserted, not proven.
  • Nonlinearity psi (and its derived constants c1, c1', G) = psi(u) = arctan(u) is the named example
    The bias and rate depend on phi'(0), G, and c1', which are derived from psi and the noise density, but neither psi nor the constants are fit to data; psi is a design choice satisfying Assumption 4.
assumptions (8)
  • domain assumption Graph is connected, simple, and static (Assumption 1)
    Standard network model; guarantees lambda_2(L) > 0, used in Lemma 7 for the disagreement contraction.
  • domain assumption Local costs f_i are twice differentiable, mu-strongly convex and L-smooth (Assumption 2)
    Gives the mean-value-theorem Hessian bounds, the contraction coefficient (1 - mu alpha_t), and the bounds on grad F(theta*); heterogeneous minimizers are allowed by construction.
  • domain assumption Communication noise is i.i.d. across time and arcs, with symmetric pdf, positive near 0, and finite first absolute moment (Assumption 3)
    Defines the heavy-tail setting: variance may be infinite. Symmetry plus finite mean gives zero mean; p > 0 near 0 gives phi'(0) > 0.
  • domain assumption Nonlinearity psi is odd, sign-preserving, nondecreasing, bounded, differentiable with bounded derivative (Assumption 4)
    Design requirements for clipping; differentiability (new vs [11, 21]) is needed for the mean-value linearization in Lemmas 6 and 7.
  • standard math Polyak lemma: phi(a) = integral psi(a+w)p(w)dw is odd, bounded, nondecreasing, differentiable with phi'(0) > 0 (Lemma 1)
    Imported from [17]; the basis for replacing the noisy nonlinearity by the smooth phi and a zero-mean residual eta_t.
  • standard math Lower bound |phi(a)| >= phi'(0)G|a|/(2g) and phi'(c) >= phi'(0)G/(2g) for c in (0, a), |a| < g (Lemma 2)
    The first part is deferred to the authors' own [10]; the sketch proves only the derivative part. Lemma 7 applies this with time-varying g = M_t, which the statement does not support.
  • standard math Deterministic comparison lemmas: v_{t+1} <= (1 - p/(t+1)^delta) v_t + q/(t+1)^delta has limit q/p with O(t^{-p}) rate (Lemmas 3-4)
    Lemma 3 is from the authors' [10] (Lemma 8.1 there); Lemma 4 is proven in the text. Used to turn one-step recursions into limsup and O(1/t^p) statements.
  • ad hoc to paper Weighted-Laplacian spectral gap: lambda_min(L_c) >= phi'(0)G/(2M_t) on the whole range of possible disagreements (Lemma 7)
    The load-bearing asserted step. Requires phi'(c) >= phi'(0)G/(2M_t) for mean-value points as large as M_t ~ t^{1-delta}; not proven and false for psi = arctan, whose derivative decays as c^{-2}.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Distributed gradient methods under heavy-tailed communication noise." pith.science (2026). https://pith.science/paper/PQJBWQBC

@misc{pith2026250524464,
  author       = {Pith},
  title        = {Pith review of: Distributed gradient methods under heavy-tailed communication noise},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PQJBWQBC}},
  note         = {Machine review of arXiv:2505.24464}
}
read the original abstract

We consider a standard distributed optimization problem in which networked nodes collaboratively minimize the sum of their locally known convex costs. For this setting, we address for the first time the fundamental problem of design and analysis of distributed methods to solve the above problem when inter-node communication is subject to \emph{heavy-tailed} noise. Heavy-tailed noise is highly relevant and frequently arises in densely deployed wireless sensor and Internet of Things (IoT) networks. Specifically, we design a distributed gradient-type method that features a carefully balanced mixed time-scale time-varying consensus and gradient contribution step sizes and a bounded nonlinear operator on the consensus update to limit the effect of heavy-tailed noise. Assuming heterogeneous strongly convex local costs with mutually different minimizers that are arbitrarily far apart, we show that the proposed method converges to a neighborhood of the network-wide problem solution in the mean squared error (MSE) sense, and we also characterize the corresponding convergence rate. We further show that the asymptotic MSE can be made arbitrarily small through consensus step-size tuning, possibly at the cost of slowing down the transient error decay. Numerical experiments corroborate our findings and demonstrate the resilience of the proposed method to heavy-tailed (and infinite variance) communication noise. They also show that existing distributed methods, designed for finite-communication-noise-variance settings, fail in the presence of infinite variance noise.

Figures

Figures reproduced from arXiv: 2505.24464 by the authors.

Figure 1
Figure 1. (a) Monte Carlo-estimated MSE error of proposed algorithm for the different choices of [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Monte Carlo-estimated MSE error for proposed algorithm and for algorithm (23) for (a) [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. (a) Monte Carlo-estimated CAP curve with proposed algorithm and algorithm form [24] (b) Monte [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: (a) Monte Carlo-estimated CAP curve with proposed algorithm and algorithm form [24] (b) Monte [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 27 canonical work pages

  1. [24]

    Wang and A

    Y. Wang and A. Nedic , Tailoring gradient methods for differentially private distributed optimization , IEEE Transactions on Automatic Control, 69 (2024), pp. 872--887

    work page 2024

  2. [1]

    Agarwal, M

    A. Agarwal, M. J. Wainwright, and J. C. Duchi , Distributed dual averaging in networks , Advances in Neural Information Processing Systems, 23 (2010)

    work page 2010

  3. [2]

    S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein, et al. , Distributed optimization and statistical learning via the alternating direction method of multipliers , Foundations and Trends in Machine learning, 3 (2011), pp. 1--122

    2011

  4. [3]

    Chatzipanagiotis, D

    N. Chatzipanagiotis, D. Dentcheva, and M. M. Zavlanos , An augmented lagrangian method for distributed optimization , Mathematical Programming, 152 (2015), pp. 405--434

    work page 2015

  5. [4]

    Clavier, T

    L. Clavier, T. Pedersen, I. Larrad, M. Lauridsen, and M. Egan , Experimental evidence for heavy tailed interference in the IoT , IEEE Communications Letters, 25 (2021), pp. 692--695

    work page 2021

  6. [5]

    J. C. Duchi, A. Agarwal, and M. J. Wainwright , Dual averaging for distributed optimization: Convergence analysis and network scaling , IEEE Transactions on Automatic Control, 57 (2012), pp. 592--606

    work page 2012

  7. [6]

    Haenggi and R

    M. Haenggi and R. Ganti , Interference in large wireless networks , Foundations and Trends in Networking, 3 (2009), pp. 127--248

    work page 2009

  8. [7]

    Hughes , Alpha-stable models of multiuser interference , in 2000 IEEE International Symposium on Information Theory (Cat

    B. Hughes , Alpha-stable models of multiuser interference , in 2000 IEEE International Symposium on Information Theory (Cat. No.00CH37060), 2000, pp. 383--

    work page 2000

Show all 28 references
  1. [8]

    Ilow and D

    J. Ilow and D. Hatzinakos , Analytic alpha-stable noise modeling in a poisson field of interferers or scatterers , Signal Processing, IEEE Transactions on, 46 (1998), pp. 1601 -- 1611

    1998

  2. [9]

    Jakovetic, D

    D. Jakovetic, D. Bajovic, A. K. Sahu, and S. Kar , Convergence rates for distributed stochastic optimization over random networks , in 2018 IEEE Conference on Decision and Control (CDC), 2018, pp. 4238--4245

    2018

  3. [10]

    Jakoveti c , D

    D. Jakoveti c , D. Bajovi c , A. K. Sahu, S. Kar, N. Milosevic, and D. Stamenkovi c , Nonlinear gradient mappings and stochastic optimization: A general framework with applications to heavy-tail noise , SIAM Journal on Optimization, 33 (2023), pp. 394--423

    2023

  4. [11]

    Jakovetic, M

    D. Jakovetic, M. Vukovic, D. Bajovic, A. K. Sahu, and S. Kar , Distributed recursive estimation under heavy-tail communication noise , SIAM Journal on Control and Optimization, 61 (2023), pp. 1582--1609

    2023

  5. [12]

    Jakovetić, J

    D. Jakovetić, J. M. F. Moura, and J. Xavier , Linear convergence rate of a class of distributed augmented lagrangian algorithms , IEEE Transactions on Automatic Control, 60 (2015), pp. 922--936

    2015

  6. [13]

    C. Liu, L. Chen, L. Luo, and J. C. Lui , Communication efficient distributed newton method with fast convergence rates , in Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023, pp. 1406--1416

    2023

  7. [14]

    Lobel and A

    I. Lobel and A. Ozdaglar , Distributed subgradient methods for convex optimization over random networks , IEEE Transactions on Automatic Control, 56 (2011), pp. 1291--1306

    2011

  8. [15]

    Nedic , Distributed gradient methods for convex machine learning problems in networks: Distributed optimization , IEEE Signal Processing Magazine, 37 (2020), pp

    A. Nedic , Distributed gradient methods for convex machine learning problems in networks: Distributed optimization , IEEE Signal Processing Magazine, 37 (2020), pp. 92--101

    2020

  9. [16]

    Nedic and A

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

    2009

  10. [17]

    Polyak and Y

    B. Polyak and Y. Tsypkin , Adaptive estimation algorithms: Convergence, optimality, stability , Automation and Remote Control, 1979 (1979)

    1979

  11. [18]

    Pu and A

    S. Pu and A. Nedic , A distributed stochastic gradient tracking method , in 2018 IEEE Conference on Decision and Control (CDC), 2018, pp. 963--968

    2018

  12. [19]

    Selim, M

    B. Selim, M. S. Alam, V. Carvalho, G. Kaddoum, and B. L. Agba , Noma-based iot networks: Impulsive noise effects and mitigation , IEEE Communications Magazine, 58 (2020), pp. 69--75

    2020

  13. [20]

    Shorinwa and M

    O. Shorinwa and M. Schwager , Distributed quasi-newton method for multi-agent optimization , IEEE Transactions on Signal Processing, 72 (2024), pp. 3535--3546

    2024

  14. [21]

    Vukovic, D

    M. Vukovic, D. Jakovetic, D. Bajovic, and S. Kar , Nonlinear consensus+innovations under correlated heavy-tailed noises: Mean square convergence rate and asymptotics , SIAM Journal on Control and Optimization, 62 (2024), pp. 376--399

    2024

  15. [22]

    Wang , Distributed optimization with noisy information sharing , in 2023 42nd Chinese Control Conference (CCC), 2023, pp

    Y. Wang , Distributed optimization with noisy information sharing , in 2023 42nd Chinese Control Conference (CCC), 2023, pp. 1743--1748

    2023

  16. [23]

    Wang and T

    Y. Wang and T. Başar , Gradient-tracking-based distributed optimization with guaranteed optimality under noisy information sharing , IEEE Transactions on Automatic Control, 68 (2023), pp. 4796--4811

    2023

  17. [25]

    Wei and A

    E. Wei and A. Ozdaglar , Distributed alternating direction method of multipliers , in 2012 IEEE 51st IEEE Conference on Decision and Control (CDC), 2012, pp. 5445--5450

    2012

  18. [26]

    M. Win, P. Pinto, and L. Shepp , A mathematical theory of network interference and its applications , Proceedings of the IEEE, 97 (2009), pp. 205 -- 230

    2009

  19. [27]

    Yang and A

    X. Yang and A. Petropulu , Co-channel interference modeling and analysis in a poisson field of interferers in wireless communications , IEEE Transactions on Signal Processing, 51 (2003), pp. 64--76

    2003

  20. [28]

    S. Yu, D. Jakovetic, and S. Kar , Smoothed gradient clipping and error feedback for decentralized optimization under symmetric heavy-tailed noise , 2024

    2024

Pith tools

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