Decentralized Stochastic Optimization and Machine Learning: A Unified Variance-Reduction Framework for Robust Performance and Fast Convergence
The DOI matches its Crossref record; 51 of the 53 references it cites were found in a registry or an index.
Abstract
Decentralized methods to solve finite-sum minimization problems are important in many signal processing and machine learning tasks where the data samples are distributed across a network of nodes, and raw data sharing is not permitted due to privacy and/or resource constraints. In this article, we review decentralized stochastic first-order methods and provide a unified algorithmic framework that combines variance reduction with gradient tracking to achieve robust performance and fast convergence. We provide explicit theoretical guarantees of the corresponding methods when the objective functions are smooth and strongly convex and show their applicability to nonconvex problems via numerical experiments. Throughout the article, we provide intuitive illustrations of the main technical ideas by casting appropriate tradeoffs and comparisons among the methods of interest and by highlighting applications to decentralized training of machine learning models.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
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?
In the citing paper
“Specifically, the variance reduction with the gradient tracking method was investigated in [146] to achieve a fast convergence rate.”
From Edge artificial intelligence for 6g: vision, enabling technologies, and applications (2022) · cites this paper as [146]
Citations
Cites
Book Review: Pattern Recognition and Machine Learning
Nasrabadi, N. M. (2007) · Journal of Electronic Imaging · doi:10.1117/1.2819119 · Cited by 22,066 · cited 3×
“…where the objective functions are smooth and strongly convex [2]; text classification with support vector machines, where the objectives are convex but not necessarily smooth [1]; and perception tasks with deep neural networks, where the cost functions are nonconvex, in general [1], [2].”
BackgroundMethodInfluentialCommunication-Efficient Learning of Deep Networks from Decentralized Data
McMahan, H. B., Moore, E., Ramage, D., Hampson, S., & Arcas, B. A. Y. (2016) · arXiv · doi:10.48550/arxiv.1602.05629 · Cited by 14,215 · cited 1×
“Various programming models and several variants of master–worker configurations that are tailored for specific computing needs and environments have been proposed, such as MapReduce, AllReduce, and federated learning [50].”
BackgroundConsensus and Cooperation in Networked Multi-Agent Systems
Olfati-Saber, R., Fax, J. A., & Murray, R. M. (2007) · Proceedings of the IEEE · doi:10.1109/jproc.2006.887293 · Cited by 10,522 · cited 3×
“…, ; i 6 and 2) optimality—the agreement should be on the minimizer of the global objective F , i.e., . cons i i = ) Average-consensus algorithms [10] are information fusion protocols that enable each node to appropriately combine the vectors received from its neighbors and agree on the average…”
BackgroundMethodResultInfluentialHorn, R. A., & Johnson, C. R. (2012) · doi:10.1017/cbo9781139020411 · Cited by 4,642
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×
“A well-known solution to this problem is DGD [13], [14], described as follows.”
MethodOptimization Methods for Large-Scale Machine Learning
Bottou, L., Curtis, F. E., & Nocedal, J. (2018) · SIAM Review · doi:10.1137/16m1080173 · Cited by 3,128
Accelerating Stochastic Gradient Descent using Predictive Variance Reduction
Johnson, R., & Zhang, T. (2013) · Cited by 1,889
Gossip-based computation of aggregate information
Kempe, D., Dobra, A., & Gehrke, J. (2004) · 44th Annual IEEE Symposium on Foundations of Computer Science, 2003. Proceedings. · doi:10.1109/sfcs.2003.1238221 · Cited by 1,480 · cited 1×
“A well-studied solution to this issue is based on the push-sum (type) algorithms [39] that enable consensus with non-DS weights through the help of eigenvector estimation.”
MethodPattern Recognition and Machine Learning
Maindonald, J. (2007) · Journal of Statistical Software · doi:10.18637/jss.v017.b05 · Cited by 1,349
EXTRA: An Exact First-Order Algorithm for Decentralized Consensus Optimization
Shi, W., Ling, Q., Wu, G., & Yin, W. (2015) · SIAM Journal on Optimization · doi:10.1137/14096668x · Cited by 1,261
Research works printed but not linked: 1
- lecun, 1998, The MNIST Database of Handwritten Digits
Cited by
Items 101–143 of 143
Personalized and Resilient Distributed Learning Through Opinion Dynamics
Ballotta, L., Bastianello, N., Ferrari, R., & Johansson, K. H. (2025) · IEEE Transactions on Control of Network Systems · doi:10.1109/tcns.2025.3649108 · Cited by 2
“These algorithms are leveraged to solve distributed learning problems [9], [10].”
Distributed Optimization in Multilayer Networks via Tensor-Laplacian Dynamics
Rodríguez-Camargo, C. D., Urquijo-Rodríguez, A. F., & Mojica‐Nava, E. (2025) · IEEE Systems Journal · doi:10.1109/jsyst.2025.3640067 · Cited by 2
On Parameter Selection for First-Order Methods: A Matrix Analysis Approach
Baron-Prada, E., Alsubaihi, S., Alshehri, K. A., & Albalawi, F. (2022) · International Conference on Control, Decision and Information Technologies (CoDIT) · doi:10.1109/codit58514.2023.10284153 · Cited by 2
“Significant contributions to such algorithms have been made from both a theoretical point of view and an application point of view (e.g., in machine learning [3] and control theory [4]).”
BackgroundDecentralized Learning-based Scenario Identification Method for Intelligent Vehicular Communications
Zhou, Y., Wangt, Y., Liu, P., Yang, J., Ohtsuki, T. O., & Sari, H. (2021) · IEEE Vehicular Technology Conference · doi:10.1109/vtc2021-fall52928.2021.9625563 · Cited by 2
“Recently,decentralized learning has been considered as one of effective acceleration tools for DL-based classification problems in communication systems [15], [16].”
Accelerated Distributed Stochastic Non-Convex Optimization over Time-Varying Directed Networks
Chen, Y., Hashemi, A., & Vikalo, H. (2023) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp49357.2023.10094584 · Cited by 2
“Such studies include the decentralized stochastic gradient descent (DSGD) and its variants, [3, 4], which combine the decentralized average consensus with local gradient updates.”
A stochastic primal–dual algorithm for composite constrained optimization
Su, E., Hu, Z., Xie, W. C., Li, L., & Zhang, W. (2024) · Neurocomputing · doi:10.1016/j.neucom.2024.128285 · Cited by 2
Asynchronous Distributed Learning with Quantized Finite-Time Coordination
Bastianello, N., Rikos, A. I., & Johansson, K. H. (2024) · IEEE Conference on Decision and Control (CDC) · doi:10.1109/cdc56724.2024.10885917 · Cited by 2
“Alterna-tively, the average consensus can be replaced with dynamic average consensus, which gives rise to gradient tracking algorithms [12].”
InfluentialZhang, J., Ling, Q. H., & So, A. M. (2020) · IEEE Conference on Decision and Control (CDC) · doi:10.1109/cdc42340.2020.9303887 · Cited by 2
Peng, J., Li, W., & Ling, Q. H. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2308.05292 · Cited by 2
“For reducing the negative e ff ect of stochastic gradient noise, variance reduction methods have been employed in decentralized stochastic optimization [35, 36, 37].”
MethodAccelerated methods for distributed optimization
Hendrikx, H. (2021) · theses.fr (ABES) · doi:10.70675/285acaadze92cz460az885azb3887e474582 · Cited by 2
Decentralized Machine Learning over the Internet
Zhu, D., & Peng, J. (2022) · Chinese Control Conference (CCC) · doi:10.23919/ccc55666.2022.9901831 · Cited by 1
“Most of the existing decentralized machine learning algorithms operate within data centers that are equipped with multiple CPUs and/or GPUs [16, 17].”
BackgroundA Complex-Based Rethinking of Communications
Scatà, M. (2025) · Studies in systems, decision and control · doi:10.1007/978-3-032-04243-9_3 · Cited by 1
Distributed Dual Operator Splitting Algorithm for Nonsmooth Composite Optimization
Li, H., Lü, Q., Xia, D., Wang, X., Wang, Z., Zheng, L., Li, J. Q., & Ran, L. (2025) · Computational intelligence methods and applications · doi:10.1007/978-981-95-0911-9_5 · Cited by 1
Resource- and Convergence-Aware Neural Algorithm Selection for 5G/B5G RAN
AlZailaa, A., Chi, H. R., Radwan, A., & Aguiar, R. L. (2026) · IEEE Transactions on Consumer Electronics · doi:10.1109/tce.2026.3653112 · Cited by 1
“The convergence rate α is modeled as [32], [33]: where { α 0 , d s , N , B } are model-dependent constants, selected feature subset size, dataset size, and batch size, respectively.”
A novel distributed gradient algorithm for composite constrained optimization over directed network
Ou, M., Zhang, H., Yan, Z., Yang, Z., & Ran, H. (2026) · Scientific Reports · doi:10.1038/s41598-026-36058-4 · Cited by 1
Federated LSTM-LRP Networks with SMOTE and LIME for Interpretable Gait Classification
Pushpalatha, O., & Premkumar, R. (2026) · Engineering Technology & Applied Science Research · doi:10.48084/etasr.14598 · Cited by 1
Robust-to-Noise Algorithms for Distributed Resource Allocation and Scheduling
Doostmohammadian, M., & Aghasi, A. (2023) · RSI International Conference on Robotics and Mechatronics (ICRoM) · doi:10.1109/icrom60803.2023.10412486 · Cited by 1
Ran, L., Hu, J., Liu, H., & Li, H. (2021) · Chinese Control Conference (CCC) · doi:10.23919/ccc52363.2021.9550251
“sumed to be a general case of existing consensus problems [1, 4, 11, 12, 13, 20, 21, 24, 25, 26].”
BackgroundMethodInfluentialD4.1: Initial results on adaptive directional LOS and NLOS reliable connectivity
Renzo, M. D., Aravanis, A. I., Qian, X., Hrasnica, H., Ntontin, K., Katzouris, N., Manganaris, K., Lazarakis, F., Alexiou, A., Boulogeorgos, A. A., et al. (2021) · Zenodo (CERN European Organization for Nuclear Research) · doi:10.5281/zenodo.5707432
StochaLM: a Stochastic alternate Linearization Method for distributed optimizationPreprint
Almeida, I., & Xavier, J. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2004.11432
“Other methods which don’t use a token include, but are certainly not limited to, EXTRA [32], DIGing [33], ADD-OPT [34], Nesterov methods [35, 36], augmented Lagrangean methods [37, 38, 39], variance reduction and average gradient methods [40, 41], and others [42, 43, 44, 45, 46].”
MethodQuantized Adaptive Subgradient Algorithms and Their ApplicationsPreprint
Xu, K., Wangni, J., Zhang, Y., Ye, D., Wu, J., & Zhao, P. (2022) · arXiv (Cornell University) · doi:10.48550/arxiv.2208.05631
“alleviate the negative effect of slow machines and decentralized algorithms [42, 43] get rid of the dependency on high communication costs on the central node.”
BackgroundDistributed Composite Optimization for Multi-agent Systems with Asynchrony
Li, H., Hu, J., Ran, L., Wang, Z., Lü, Q., Du, Z., & Huang, T. (2022) · Studies in systems, decision and control · doi:10.1007/978-3-030-99776-2_7
Projection Algorithms for Distributed Stochastic Optimization
Lü, Q., Liao, X., Li, H., Deng, S., & Gao, S. (2022) · Wireless networks · doi:10.1007/978-981-19-8559-1_2
Event-Triggered Acceleration Algorithms for Distributed Stochastic Optimization
Lü, Q., Liao, X., Li, H., Deng, S., & Gao, S. (2022) · Wireless networks · doi:10.1007/978-981-19-8559-1_5
Proximal Algorithms for Distributed Coupled Optimization
Lü, Q., Liao, X., Li, H., Deng, S., & Gao, S. (2022) · Wireless networks · doi:10.1007/978-981-19-8559-1_3
Constrained distributed online convex optimization with bandit feedback for unbalanced digraphs
Tada, K., Hayashi, N., & Takai, S. (2023) · IET Control Theory and Applications · doi:10.1049/cth2.12548
Moreau Envelope ADMM for Decentralized Weakly Convex Optimization
Mirzaeifard, R., Venkategowda, N. K. D., Jung, A., & Werner, S. (2023) · Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · doi:10.1109/apsipaasc58517.2023.10317303
“For convex objective functions, subgradient methods and ADMM are guaranteed to converge to a global optimum under suitable parameter choices [6], [10], [11].”
MethodFinite-Time Convergence Rates of Decentralized Local Markovian Stochastic Approximation
Wang, P., & Zheng, N. (2024) · International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2024/562
“The weighted connectivity matrix M = [ M ij ] ∈ R n × n satisfies the following: (1) Doubly Stochastic: Assumption 2 is fairly standard in the analysis of decentralized optimization; see e.g., [Xin et al. , 2020].”
BackgroundFully Distributed and Quantized Algorithm for MPC-based Autonomous Vehicle Platooning OptimizationPreprint
Doostmohammadian, M., Aghasi, A., & Rabiee, H. R. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2501.18889
“Emerging notions of cloud-computing and multi-agent processing motivate distributed and parallelized algorithms in large-scale applications from filtering and estimation over sensor networks [1]–[3] to distributed machine learning [4]– [6] and computing resource allocation [7], [8].”
Accelerated online averaging over networks
Liu, W., Liu, X., Mao, X., & Wu, Z. (2025) · Scientia Sinica Mathematica · doi:10.1360/ssm-2024-0223
Li, C., Yang, M., Cao, J., Wu, X., He, Y., & Wang, X. (2024) · IEEE International Conference on Frontiers Technology of Information and Computer (ICFTIC) · doi:10.1109/icftic64248.2024.10913272
“…the averaging gradient and uses it to correct the gradient of each sample, improving the stability and convergence speed of the algorithm. mixSVRG [10] and CheapSVRG [11] combine the advantages of different algorithms by using a batch of random gradients or sampling instances to estimate the…”
Kalwar, D., Baranwal, M., & Khadilkar, H. (2025) · ACM International Conference on Information and Knowledge Management (CIKM) · doi:10.1145/3746252.3761288
Distributed Dual Proximal Gradient Algorithms for Nonsmooth Constrained Composite Optimization
Li, H., Lü, Q., Xia, D., Wang, X., Wang, Z., Zheng, L., Li, J. Q., & Ran, L. (2025) · Computational intelligence methods and applications · doi:10.1007/978-981-95-0911-9_9
Parameter Mapping of Distribution Substitution for Inter-Point Distances in Random Networks
Dang, S., Ye, J., Guo, S., Shubair, R. M., & Chafii, M. (2025) · IEEE Transactions on Signal and Information Processing over Networks · doi:10.1109/tsipn.2025.3642229
“Because of the ergodicity of random initialization, either (cid:12) → ∞ or ψ → 0 will necessarily result in the global optimum of (6) [50].”
Dahl, M. (2023) · DiVA (Digitala Vetenskapliga Arkivet)
Decentralized Constrained Optimization: a Novel Convergence Analysis
Shahriari-Mehr, F. (2023) · Chalmers Research (Chalmers University of Technology)
Kalwar, D., Baranwal, M., & Khadilkar, H. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2509.22174
Tang, R., Li, H., Xia, D., Ran, L., Li, J., & Zhu, W. (2026) · Swarm and Evolutionary Computation · doi:10.1016/j.swevo.2026.102282
On Gossip Algorithms for Machine Learning with Pairwise ObjectivesPreprint
Colin, I., Bellet, A., Clémençon, S., & Salmon, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.24128
“Gossip algorithms have been designed to solve various statistical tasks in this context, such as data aggregation or decentralized multi-agent optimization, see e.g., Loizou & Richtárik (2016), Zantedeschi et al. (2020), Xin et al. (2020a) or Hendrikx et al. (2020).”
Chen, Z., & Wang, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.27962
“Existing distributed optimization/learning algorithms (see, e.g., [9]–[19]) universally assume that participating agents are honest and behave truthfully.”
Shahriari-Mehr, F. (2025) · Chalmers Research (Chalmers University of Technology) · doi:10.63959/chalmers.dt/5789
DeFL-VP: Decentralized Federated Learning with Variance Reduction and Differential Privacy
Liu, R., Xiu, R., Zhang, W., Liao, X., & Sakurai, K. (2026) · Journal of Machine Learning and Information Security · doi:10.53941/jmlis.2026.100012
Chaotic-AES Integration for Privacy-Preserving Distributed Personalized Optimization
Tang, Q., Zhou, X., Yuan, Y., Lü, Q., & Li, H. (2026) · IEEE Signal Processing Letters · doi:10.1109/lsp.2026.3711542
“…first-order solvers have been developed to address distributed optimization, including subgradient and gradient descent methods [10], [11], [12], [13], as well as more advanced schemes such as EXTRA [14], [15], [16], DIGing [17], [18], [19], push–pull methods [20], [21], [22],…”
Metadata
- Authors
- Xin, R., Kar, S., & Khan, U. A.
- Affiliations
- Carnegie Mellon University (US)
- Tufts University (US)
- Venue
- IEEE Signal Processing Magazine
- Published
- 2020-05
- Volume
- 37
- Issue
- 3
- Pages
- 102-113
- ISSN
- 1053-5888 (ISSN-L)
- 1558-0792
- OpenAlex
- W3023944502
- Cited by
- 114
- References
- 53
- Citations by year
- 2026: 5 · 2025: 15 · 2024: 13 · 2023: 16 · 2022: 28 · 2021: 22 · 2020: 13
- FWCI
- 7.76
- Type
- journal article
- Language
- en
- Open access
- closed
- Licence
- Locations
- IEEE Signal Processing Magazine (publishedVersion)
- Funders
- National Science Foundation (CCF-1350264; CCF-1513936; CMMI-1903972; CBET-1935555)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Stochastic Gradient Optimization Techniques
- Physical Sciences › Engineering › Control and Systems Engineering › Distributed Control Multi-Agent Systems
- Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
- Keywords
- distributed optimization, stochastic optimization, variance reduction, gradient tracking, finite-sum minimization, decentralized machine learning, decentralized learning, non-convex optimization, strong convexity, convergence guarantees, privacy-preserving machine learning
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.
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.