EXTRA: An Exact First-Order Algorithm for Decentralized Consensus Optimization
Open accessThe DOI matches its Crossref record; 47 of the 47 references it cites were found in a registry or an index; 7 of the 19 sources its claims cite were checked for support.
Abstract
Recently, there has been growing interest in solving consensus optimization problems in a multiagent network. In this paper, we develop a decentralized algorithm for the consensus optimization problem $\mathrm{minimize}_{x\in\mathbb{R}^p}~\bar{f}(x)=\frac{1}{n}\sum_{i=1}^n f_i(x),$ which is defined over a connected network of $n$ agents, where each function $f_i$ is held privately by agent $i$ and encodes the agent's data and objective. All the agents shall collaboratively find the minimizer while each agent can only communicate with its neighbors. Such a computation scheme avoids a data fusion center or long-distance communication and offers better load balance to the network. This paper proposes a novel decentralized exact first-order algorithm (abbreviated as EXTRA) to solve the consensus optimization problem. “Exact” means that it can converge to the exact solution. EXTRA uses a fixed, large step size, which can be determined independently of the network size or topology. The local variable of every agent $i$ converges uniformly and consensually to an exact minimizer of $\bar{f}$. In contrast, the well-known decentralized gradient descent (DGD) method must use diminishing step sizes in order to converge to an exact minimizer. EXTRA and DGD have the same choice of mixing matrices and similar per-iteration complexity. EXTRA, however, uses the gradients of the last two iterates, unlike DGD which uses just that of the last iterate. EXTRA has the best known convergence rates among the existing synchronized first-order decentralized algorithms for minimizing convex Lipschitz--differentiable functions. Specifically, if the $f_i$'s are convex and have Lipschitz continuous gradients, EXTRA has an ergodic convergence rate $O(\frac{1}{k})$ in terms of the first-order optimality residual. In addition, as long as $\bar{f}$ is (restricted) strongly convex (not all individual $f_i$'s need to be so), EXTRA converges to an optimal solution at a linear rate $O(C^{-k})$ for some constant $C>1$.
Overview
CoChat Summary
TL;DR
EXTRA is a synchronous decentralized exact first-order algorithm for consensus optimization that can use a fixed step size independent of network size and converges to an exact solution. It achieves an ergodic O(1/k) rate for convex Lipschitz-differentiable objectives and a linear rate under restricted strong convexity.
Key findings
- EXTRA uses a fixed large step size and converges exactly, unlike DGD with a fixed step size, which stalls in an O(α)-neighborhood and requires diminishing step sizes for exact convergence.
- Under Assumptions 1–3 and step size \(0<\alpha<2\lambda_{\min}(\tilde W)/L_f\), the sequence \(z_k\) converges to an optimal \(z^*\), with contraction factor \(\zeta=1-\alpha L_f/(2\lambda_{\min}(\tilde W))\); running-average progress is \(O(1/k)\), running-best progress is \(o(1/k)\), and running-average optimality residuals are \(O(1/k)\).
- If \(g(x)=f(x)+\frac{1}{4\alpha}\|x\|^2_{\tilde W-W}\) is restricted strongly convex with constant \(\mu_g>0\), then with proper step size \(\alpha<2\mu_g\lambda_{\min}(\tilde W)/L_f^2\), \(\|z_k-z^*\|_G^2\) converges Q-linearly at \(O((1+\delta)^{-k})\), and \(\|x_k-x^*\|_{\tilde W}^2\) converges R-linearly at the same rate.
- Numerical experiments: in decentralized least squares with \(n=10\), \(m_i=1\), \(p=5\), connectivity ratio \(r=0.5\), \(L_f=1\), and \(\|x^*-x_0\|=300\), DGD with fixed step size and EXTRA show similar linear convergence for the first 200 iterations, after which DGD slows and stalls while EXTRA continues. In decentralized logistic regression with \(n=200\), \(r=0.2\), \(m_i=10\), \(p=20\), and 2000 samples, EXTRA outperforms DGD and shows linear exact convergence.
Methods
- Design: a synchronous decentralized first-order algorithm where each agent holds a local copy \(x^{(i)}\) and communicates only with neighbors; EXTRA is derived by taking the difference of two DGD updates and uses gradients of the last two iterates.
- Data/sample: numerical tests on decentralized least squares, decentralized robust least squares with Huber loss, and decentralized logistic regression over randomly generated connected networks.
- Analysis: convergence is established under mixing-matrix assumptions, convex Lipschitz-differentiable objectives, and solution existence; rates are derived via a contraction recursion, ergodic averaging, and restricted strong convexity.
Limitations
- The paper states EXTRA is “far from perfect” and that more work is needed to adapt it to asynchronous and dynamic network settings; these are left as open questions.
- The paper notes it is open whether \(\|z_k-z_{k+1}\|_G^2\) is monotonic; if monotonicity could be shown, the rates would hold for the last point in the running sequence. It also states that proving linear convergence under the numerically observed larger step size \(O(1/L_f)\) remains open.
- The paper states that determining \(\mu_g\) or a lower bound is tricky except when each \(f_i\) is restricted strongly convex, and that deriving an optimal step size and explicit linear rate is “quite tricky and messy.” The numerical results are described as preliminary.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures

![Fig. 4.2 Plot of residuals ‖ x k -x ∗ ‖ F ‖ x 0 -x ∗ ‖ F . Constant α = 0 . 5276 is the theoretical critical step size given for DGD in [36]. For DGD with diminishing step sizes O (1 /k 1 / 3 ) and O (1 /k 1 / 2 ) , we have hand-optimized their initial step sizes as 10 α and 20 α , respectively. The initial large step sizes have helped them (the red and purple curves) realize faster convergence initially.](https://figures.cochat.ai/paper-figures/4f2c85fb730e4c6eaeaa955cda38aeac/extra-an-exact-first-order-algorithm-for-decentralized-consensus-optimization-figure-2.png)
![Fig. 4.3 Plot of residuals ‖ x k -x ∗ ‖ F ‖ x 0 -x ∗ ‖ F . Constant α = 0 . 0059 is the theoretical critical step size given for DGD in [36]. For DGD with diminishing step sizes O (1 /k 1 / 3 ) and O (1 /k 1 / 2 ) , we have hand-optimized their initial step sizes as 10 α and 20 α , respectively.](https://figures.cochat.ai/paper-figures/d5f55fab59f84fe6834d6cb2909bb298/extra-an-exact-first-order-algorithm-for-decentralized-consensus-optimization-figure-3.png)
Verification
Claims the paper makes
6 Valid · 1 Not valid · 9 Not checked · 58 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“On the other hand, properly reducing α k enables exact convergence , namely, that each x k ( i ) converges to the same exact solution.”
What the sources say
- Misattributedcites [36] [36] K. Yuan, Q. Ling, and W. Yin , On the Convergence of Decentralized Gradient Descent , arXiv preprint arXiv:1310.7063, (2013). 1.1, 2.1, 2.4, 4.1, 4.1, 4.2, 4.3 - sourceExperimental — accuracy not yet measured
Our copy disclaims the diminishing-stepsize result for itself: §1.3 (Contribution and notation) states 'Since our analysis uses a fixed stepsize, the local solutions will not be asymptotically consensual. To adapt our analysis to diminishing stepsizes, significant changes will be needed,' and every rate in Theorems 2-3 and Remarks 1-3 lands in an O(alpha/(1-beta))-neighborhood rather than the exact solution. The only place the paper states the claim's point is §1.2 (Related methods): 'Convergence to x* can be achieved by using diminishing stepsizes in [8,11,29] at the price of slower rates of convergence' — i.e., it credits exact convergence under diminishing stepsizes to other works it cites, without reporting or relying on it. Reference [8] in our copy is Duchi, Agarwal and Wainwright, 'Dual averaging for distributed optimization: convergence analysis and network scaling,' IEEE TAC 57 (2012) (with [11] Jakovetic, Xavier, Moura and [29] also credited). (LLM, conf=0.82) It matches “Dual averaging for distributed optimization: convergence analysis and network scaling” (2012) instead.
Checked by a reasoning model — may be wrong.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Misattributed
Every source the paper cites for this claim was checked and supports it.
In the paper
“In contrast, the well-known decentralized gradient descent (DGD) method must use diminishing step sizes in order to converge to an exact minimizer.”
- Also stated in: § EXTRA: AN EXACT FIRST-ORDER ALGORITHM FOR DECENTRALIZED CONSENSUS OPTIMIZATION ∗ · p. 4 · plain
“properly reducing α k enables exact convergence , namely, that each x k ( i ) converges to the same exact solution.”
What the sources say
- Supportedcites [36] [36] K. Yuan, Q. Ling, and W. Yin , On the Convergence of Decentralized Gradient Descent , arXiv preprint arXiv:1310.7063, (2013). 1.1, 2.1, 2.4, 4.1, 4.1, 4.2, 4.3 - source
Section 1.2 (Related methods) states that with a fixed stepsize the listed algorithms 'and iteration (4) in particular' converge to a neighborhood of x* rather than x* itself, and that convergence to x* can be achieved by using diminishing stepsizes (at the price of slower rates) — exactly the necessity the claim attributes. Section 1.3 reinforces this: 'Since our analysis uses a fixed stepsize, the local solutions will not be asymptotically consensual,' and the Abstract/Theorems 2-3 give only O(α/(1-β))-neighborhood landing levels for fixed α. The only mild caveat is that the diminishing-stepsize exact-convergence results are credited to references [8,11,29] rather than proved here for iteration (4), but the paper's own statement of the fixed-stepsize limitation for DGD covers the claim. (LLM, conf=0.85)
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § EXTRA: AN EXACT FIRST-ORDER ALGORITHM FOR DECENTRALIZED CONSENSUS OPTIMIZATION ∗ · p. 4 · plain
Every source the paper cites for this claim was checked and supports it.
In the paper
“(iii) min t ≤ k { a t } = o ( 1 k ) .”
What the sources say
- Supportedcites [6] [6] D. Davis and W. Yin , Convergence Rates of Splitting Algorithms for Optimization . arXiv preprint arXiv:1406.4834, 2014. 1 - source
Section 2 ('Summable sequence convergence lemma', Lemma 3, p. 8) states Part 4 'No monotonicity', defining the best-so-far sequence and asserting that '(a_j^best) is monotonically nonincreasing and the above bounds continue to hold when a_k is replaced with a_k^best', while the Part 1a proof on the same page gives a_k = o(1/(Λ_k − Λ_⌈k/2⌉)) = o(1/(k+1)) when (λ_j) is bounded away from 0 (i.e. unweighted summability). Together these yield exactly the claim: a nonnegative sequence with finite sum has running minimum o(1/k) (the source writes o(1/(k+1)), the same rate up to index shift). The remark on p. 9 confirms the setting — Part 1 generalizes the known result that a nonnegative, summable, monotonic sequence converges at o(1/(k+1)) — and p. 12 applies Lemma 3 to a running minimum, min_{i=0,…,k}{‖e_i‖}. (LLM, conf=0.85)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“An EXTRA update is, therefore, a DGD update with a cumulative correction term.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In this subsection we prove that EXTRA with a proper step size reaches linear convergence if the original objective ¯ f is restricted strongly convex.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
31 more claims
Create a free account to see all 74 claims and to verify the 7 references still open.
- Probably not a substantial claim. 21
- It may state two separate claims as one. 26
- The quote only partly states this claim. 1
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“EXTRA converges at an o (1 /k ) rate when the objective function in (1) is convex, and it has a Q-linear rate when the objective function is strongly convex.”
From Achieving Geometric Convergence for Distributed Optimization Over Time-Varying Graphs (2017) · cites this paper as [62]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Proposition 1 ( [62]). Assume null { I -W } = span { 1 } where W ∈ R n × n . If x ∗ satisfies conditions: (i) x ∗ = Wx ∗ (consensus), (ii) 1 ⊤ ∇ f ( x ∗ ) = 0 (optimality), then the rows of x ∗ are the same as each other and the transpose of each row is an optimal solution of the problem (1) .”
From Achieving Geometric Convergence for Distributed Optimization Over Time-Varying Graphs (2017) · cites this paper as [62]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Algorithm 1: EXTRA · p. 9
“null { U } = (span { 1 } ) ⊥”
§ Algorithm 1: EXTRA · p. 7
“null { W -˜ W } = span { 1 } , null { I -˜ W } ⊇ span { 1 } .”
Citations
Cites
Distributed Subgradient Methods for Multi-Agent Optimization
Nedic, A., & Ozdaglar, A. (2009) · IEEE Transactions on Automatic Control · doi:10.1109/tac.2008.2009515 · Cited by 3,719 · cited 1×
“1) include the (sub)gradient method [21, 25, 36], the (sub)gradient-push method [23, 24], the fast (sub)gradient method [5, 14], and the dual averaging method [8].”
MethodFast linear iterations for distributed averaging
Xiao, L., & Boyd, S. (2004) · Systems & Control Letters · doi:10.1016/j.sysconle.2004.02.022 · Cited by 2,706
Distributed Optimization Over Time-Varying Directed Graphs
Nedic, A., & Olshevsky, A. (2015) · IEEE Transactions on Automatic Control · doi:10.1109/tac.2014.2364096 · Cited by 1,187
Distributed average consensus with least-mean-square deviation
Xiao, L., Boyd, S., & Kim, S.-J. (2007) · Journal of Parallel and Distributed Computing · doi:10.1016/j.jpdc.2006.08.010 · Cited by 1,107
Optimal decentralized protocol for electric vehicle charging
Gan, L., Topcu, U., & Low, S. H. (2013) · IEEE Transactions on Power Systems · doi:10.1109/tpwrs.2012.2210288 · Cited by 997
Dual Averaging for Distributed Optimization: Convergence Analysis and Network Scaling
Duchi, J. C., Agarwal, A., & Wainwright, M. J. (2012) · IEEE Transactions on Automatic Control · doi:10.1109/tac.2011.2161027 · Cited by 947
On the Linear Convergence of the ADMM in Decentralized Consensus Optimization
Shi, W., Ling, Q., Yuan, K., Wu, G., & Yin, W. (2014) · IEEE Transactions on Signal Processing · doi:10.1109/tsp.2014.2304432 · Cited by 820
Self-citationDistributed Stochastic Subgradient Projection Algorithms for Convex Optimization
Sundhar Ram, S., Nedić, A., & Veeravalli, V. V. (2010) · Journal of Optimization Theory and Applications · doi:10.1007/s10957-010-9737-7 · Cited by 814 · cited 1×
“1) has additional constraints that force x into a bounded set, which also leads to bounded (sub)gradients and a Hessian, projected first-order algorithms are applicable [27, 37].”
BackgroundOn Distributed Convex Optimization Under Inequality and Equality Constraints
Zhu, M., & Martinez, S. (2012) · IEEE Transactions on Automatic Control · doi:10.1109/tac.2011.2167817 · Cited by 759
Fastest Mixing Markov Chain on a Graph
Boyd, S., Diaconis, P., & Xiao, L. (2004) · SIAM Review · doi:10.1137/s0036144503423264 · Cited by 744
Research works printed but not linked: 3
- [5] I. Chen , Fast Distributed First-Order Methods , master's thesis, Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, 2012. 1.1, 2.1, 2.4, 4.1 -
- [15] B. Johansson , On Distributed Optimization in Networked Systems , PhD thesis, KTH, 2008. 1 -
- [31] J. Tsitsiklis , Problems in Decentralized Decision Making and Computation , PhD thesis, Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, 1984. 2.4 -
Cited by
Items 1,301–1,315 of 1,315
Decentralized Primal-Dual Learning Over Directed Graphs
Zhang, X., Han, H., Chen, H., & Sayed, A. H. (2026) · IEEE Transactions on Signal Processing · doi:10.1109/tsp.2026.3714876
“The symmetry of such networks allows for the use of doubly-stochastic combination matrices, where both rows and columns sum to one [5], [6], [7], [8], [9], [10], [11], [12], [13], [14].”
Das, S., Schenato, L., & Dey, S. (2026) · European Journal of Control · doi:10.1016/j.ejcon.2026.101610
Decentralized Elastic-Net Broad Learning System over Directed Graphs
Liang, J., & Zhou, M. (2026) · Journal of Computer and Communications · doi:10.4236/jcc.2026.147008
Proximal dogleg opportunistic majorization for nonconvex and nonsmooth optimization
Zhou, Y., & Dai, W. (2026) · Signal Processing · doi:10.1016/j.sigpro.2026.110849 · Cited by 2
A Weighted Bundle Method of Multipliers with Improved Computational Efficiency
Zheng, Z., Li, C., Xu, X., & Wu, X. (2026) · International Conference on Control and Automation · doi:10.1109/icca69928.2026.11618153
Hou, Q., & Hao, L. (2026) · Ocean Engineering · doi:10.1016/j.oceaneng.2026.127430
Variance reduced stochastic diffusion algorithm for large-scale distributed optimisation
Zhao, J., & Gao, W. (2026) · International Journal of Control · doi:10.1080/00207179.2026.2717761
Decentralized low tubal rank tensor estimation via factorized gradient tracking
Song, Z., Peng, L., Zhao, W., & Lian, H. (2026) · Neurocomputing · doi:10.1016/j.neucom.2026.134884
Distributed pull-primal-dual stochastic gradient algorithm over directed networks with noisy links
Wu, J., Zhang, M., Zhang, S., & Zhou, Z. (2026) · Signal Processing · doi:10.1016/j.sigpro.2026.110885
Similarity-Promoting Resilient Adaptive Aggregation in Peer-to-Peer Machine Learning
Bhowmick, C., & Koutsoukos, X. (2026) · Machine Learning · doi:10.1007/s10994-026-07162-3
AdGT: Decentralized Gradient Tracking With Adaptive Per-Agent Stepsizes
Ghaderyan, D., & Werner, S. (2026) · IEEE Transactions on Signal Processing · doi:10.1109/tsp.2026.3733814 · Cited by 5
“We compare AdGT with both accelerated and non-accelerated decentralized algorithms, including OPTRA [44], APM-C [62], Acc-DNGD-NSC [38], EXTRA [8], NEXT/DIGing [20], [21], and DPSGD [1].”
Influence of Link Noise on Exact Diffusion Stochastic Optimization Over Distributed Networks
Han, H., Zhang, S., Chen, H., et al. (2026) · IEEE Transactions on Signal Processing · doi:10.1109/tsp.2026.3733659
“…methods have been proposed using gradient-tracking (GT), primal-dual, and alternating direction multiplier techniques to solve the problem (1); see, e.g., the presentation in [3] as well as DIGing [11], Aug-DGM [12], EXTRA [13], DLM [14], NIDS [15], and exact diffusion (ED) [16], [17] methods.”
Edge-Based Event-Triggered Online Distributed Inexact Gradient Descent Algorithm
Luo, D., Chen, S., Wu, X., et al. (2026) · Algorithms · doi:10.3390/a19100818
Distributed Optimization Under Uncertainty: A Cross-Paradigm Review and Future Directions
Mugenga, I. R., Geletu, A., Mirau, S., & Li, P. (2026) · Mathematics · doi:10.3390/math14193489
A Two‐Average‐Tracking Scheme in Distributed Optimization With a Small‐Gain Convergence Analysis
Xiao, Y., & Ni, W. (2026) · Asian Journal of Control · doi:10.1002/asjc.70243
Metadata
- Authors
- Shi, W., Ling, Q., Wu, G., & Yin, W.
- Venue
- SIAM Journal on Optimization
- Published
- 2015-01
- Volume
- 25
- Issue
- 2
- Pages
- 944-966
- ISSN
- 1052-6234 (ISSN-L)
- 1095-7189
- OpenAlex
- W1571416372
- Cited by
- 1,261
- References
- 25
- Citations by year
- 2026: 75 · 2025: 116 · 2024: 131 · 2023: 131 · 2022: 145 · 2021: 138 · 2020: 158 · 2019: 140 · 2018: 97 · 2017: 62 · 2016: 43 · 2015: 21
- FWCI
- 85.01
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- SIAM Journal on Optimization (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- National Science Foundation (0748839, 1317602)
- National Natural Science Foundation of China (61004137)
- China Scholarship Council
- Ministry of Industry and Information Technology of the People's Republic of China
- Topics
- Physical Sciences › Engineering › Control and Systems Engineering › Distributed Control Multi-Agent Systems
- Physical Sciences › Computer Science › Artificial Intelligence › Stochastic Gradient Optimization Techniques
- Physical Sciences › Computer Science › Computer Networks and Communications › Distributed Sensor Networks and Detection Algorithms
- Keywords
- multi-agent systems, large step size, linear convergence, ergodic convergence rate, strongly convex functions, distributed optimization
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.