Tackling Heterogeneity in Federated Learning via Variance-Reduced Boltzmann Sampling within Homogeneous Social Coalitions

Alessandro Licciardi; Anton Proskurnikov; Lamberto Rondoni; Lorenzo Zino; Roberta Raineri

REVIEW 3 major objections 4 minor 20 references

Tackling Heterogeneity in Federated Learning via Variance-Reduced Boltzmann Sampling within Homogeneous Social Coalitions

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 FedCVR-Bolt claims that federated learning heterogeneity can be mitigated at the client-selection stage by forming model-similarity coalitions and sampling one variance-reducing representative per coalition, with provable convergence to a…

desk verdict A clearly-specified client-selection heuristic with a promising empirical setup, but the convergence proof assumes the very property it claims to establish. read the letter →

arxiv 2506.02897 v3 pith:5QDOTJJR submitted 2025-06-03 cs.LG

Alessandro Licciardi , Roberta Raineri , Anton Proskurnikov , Lamberto Rondoni , Lorenzo Zino This is my paper

classification cs.LG

keywords federatedlearningclientselectiondataheterogeneityvariancereductionspectralclusteringBoltzmannexplorationopiniondynamicscoalitionformation

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 claims that a large part of the damage done by non-IID data in federated learning can be repaired at the client-selection stage, before any gradient is computed. It proposes FedCVR-Bolt, which partitions clients into coalitions of similar local models using spectral clustering on a homophily-inspired similarity matrix, then samples one representative from each coalition with probability proportional to the estimated reduction in variance of the global model update. The authors argue this two-stage scheme provably converges to a neighborhood of a stationary point of the global loss under smoothness, bounded-variance, and a gradient-alignment assumption, and that on MNIST, CIFAR-10, and CIFAR-100 it outperforms FedAvg, Power-of-Choice, ActiveFL, FedProx, IFCA, and FeSEM. If right, it means heterogeneity can be tamed by smarter participant sampling rather than by changing the loss or the aggregation rule.

What carries the argument

The load-bearing object is the variance-reduction score v_k = ∑_d (C^d α)_k² / C^d_kk, computed from the online covariance estimate C^d, updated by a Robbins-Monro rule; this score measures how much observing client k's model reduces the conditional variance of each component of the global model. The score feeds a Boltzmann distribution π_p(k;t) ∝ exp(β v_k) that selects one representative per spectral cluster, with β=1 in the experiments, and it drives both the coalition-formation step (spectral clustering on the homophily matrix) and the within-coalition selection step. The convergence argument rests on a Descent-Lemma bound combined with Assumption A3, an alignment condition ⟨G(θ),∇L(θ)⟩ ≥ c‖∇L(θ)‖², and on the assertion that after a thermalization time the clustering is stationary, so the sampling distribution is fixed.

What would settle it

Run FedCVR-Bolt on a heterogeneous benchmark and compute the empirical alignment ratio ⟨G(θ_t),∇L(θ_t)⟩/‖∇L(θ_t)‖² at each round; if it is not bounded below by a positive constant, or frequently becomes negative, during training, Assumption A3 is violated and the ε-bound of Proposition A3 cannot hold. Alternatively, construct a two-cluster dataset where spectral clustering provably oscillates across rounds and check whether the average squared gradient norm still falls below ε.

Watch

Extended reading notes

Core claim

FedCVR-Bolt is a client-sampling algorithm for federated learning that treats each client's current local model as an opinion in a social network. At every round the server computes an RBF-style homophily matrix W_kj ∝ exp(-γ‖θ_k - θ_j‖²), applies spectral clustering to partition the K clients into P coalitions, and then draws one client j_p from each coalition with Boltzmann probability π_p(k;t) = exp(β v_k)/∑_{j∈C_p} exp(β v_j), where v_k = ∑_d (C^d α)_k² / C^d_kk is the total variance reduction of the global model estimate obtained by observing client k's update, computed from the online-estimated covariance C^d of the model components. The variance-reduction formula is derived from the best linear predictor of the global model given a subset of observed clients, generalizing results from opinion-dynamics node selection. The paper's central claim is that selecting one variance-maximizing representative per model-similarity coalition lowers the variance of the aggregated gradient, and that under standard smoothness and bounded-variance assumptions plus an alignment condition the procedure converges in expectation to a neighborhood of a stationary point of the global loss, with experiments showing higher accuracy than six heterogeneity-aware baselines on three benchmarks.

Load-bearing premise

The proof assumes that the expected model update produced by the client-selection policy stays aligned with the true loss gradient by a fixed positive margin c, and that after some time the coalitions stop changing; if either fails, the stated convergence bound does not follow.

Editorial extensions

If this is right

The average expected squared gradient norm of the global model is guaranteed to fall below ε = LηM/(2c), so the iterates concentrate near a stationary point of the global loss rather than diverging, for a sufficiently small learning rate.
Empirically, FedCVR-Bolt achieves higher test accuracy than all six baselines on MNIST, CIFAR-10, and CIFAR-100 under Dirichlet(0.1) partitioning, with the largest gains over FedAvg and Power-of-Choice on CIFAR-10.
The default implementation spends the first 30 rounds on uniform random sampling to build a stable covariance estimate before switching to Boltzmann variance-reduction sampling, so the method is compatible with an exploration warm-up phase.
Computation is server-side and O(K²D) per round, dominated by similarity-matrix construction, so selection overhead does not grow with model dimension beyond the pairwise-similarity cost.

Reading between the lines

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

The authors leave implicit that the variance-reduction score depends on an online estimate of the model covariance; under concept drift, that estimate may lag, and the Boltzmann temperature β would need to be raised to maintain exploration—a tunable knob the paper fixes at β=1 without analysis.
The same selection machinery could extend to other aggregation rules beyond weighted averaging, such as SGD with momentum or variance-reduced aggregators, where the variance-reduction objective would generalize naturally.
Because the coalition structure is recomputed each round, persistent spectral clusters could serve as a cheap, interpretable diagnosis of which clients share data distributions, potentially informing personalized federated learning without training per-client models.
A testable extension would be to estimate the gradient-alignment constant c empirically during training and adapt the learning rate or β when alignment degrades, making the convergence guarantee adaptive rather than assumed.

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.

Desk Editor's Note

read the letter

The paper describes FedCVR-Bolt, a client-selection method for heterogeneous FL that groups clients by spectral clustering on their model parameters and then picks one representative per group using a Boltzmann distribution over a variance-reduction score. That specific combination is new, and the algorithm is clearly specified. The experiments span a synthetic regression dataset, MNIST, CIFAR-10, and CIFAR-100, and the results are mostly positive, if modest, against FedAvg, Power-of-Choice, ActiveFL, FedProx, IFCA, and FeSEM. If the method works as advertised, it is a useful, server-side way to think about client selection.

The soft spots are real. The central theoretical claim — Proposition A3 in Appendix B.4 — does not deliver what the abstract promises. The proof relies on Assumption A3, which states that the expected update of the selection policy has a positive inner product with the true gradient. That is essentially the convergence property you would want the theorem to establish. Calling it "very weak" and saying the experiments can verify it is not a proof; no verification of the pointwise inequality is reported. The proof also inserts, "without losing generality," the assumption that after some time the spectral clustering has converged to stationary clusters. That is not without loss of generality by any argument in the paper; the clustering is recomputed every round from fresh client models, and nothing shows it stabilizes. Finally, the proof treats a single SGD step with a known covariance matrix, while the algorithm runs S local steps and updates the covariance online. So the theorem, as written, is about a simplified process, and the simplification is doing the work.

Empirically, the gains are a few percentage points on classification and a few percent on synthetic regression, with overlapping error bars and no significance tests. The main text and appendix disagree on the number of rounds for MNIST (100 vs 200). These are not fatal to the heuristic, but they undercut the confident tone of the conclusions.

Who gets value from this? Researchers who work on client-selection policies might find the variance-reduction-with-coalitions idea worth trying as a baseline. The paper is not there yet as a theoretical contribution; the theory should either be fixed or removed, and the experimental claims need significance testing and more careful reporting. That said, I think a serious editor should send it to review. The algorithm is concrete, the direction is relevant, and a good referee report could push the authors to either prove something honest about FedCVR-Bolt or reframe the paper as a purely empirical study. I would bring it to a reading group, but I would not cite it in my own work until the theory is sorted out.

Referee Report

3 major / 4 minor

Summary. The paper proposes FedCVR-Bolt, a federated-learning client-selection algorithm that first partitions clients into coalitions by spectral clustering of normalized client models and then selects one representative per coalition according to a Boltzmann distribution over an estimated variance-reduction score. The authors claim convergence of the global model to a neighborhood of a stationary point of the global loss under mild standard assumptions, and they report accuracy improvements over FedAvg, Power-of-Choice, ActiveFL, FedProx, IFCA, and FeSEM on synthetic regression, MNIST, CIFAR-10, and CIFAR-100 benchmarks.

Significance. If the advertised convergence guarantee were valid, the paper would make a useful contribution by connecting social-consensus ideas to client selection and by giving a principled variance-reduction objective for selecting one representative per cluster. The variance-reduction identities in Section 3.1 and Appendix B.2 are correctly derived in the linear best-linear-predictor setting, and the paper includes a reproducibility statement with a code link and detailed implementation notes. However, the central theoretical claim is not established: the proof of Proposition A3 relies on an alignment assumption that essentially states the desired conclusion, and it further assumes without proof that the spectral clustering has converged. Because the main advertised contribution—a convergence guarantee under mild, standard assumptions—is not supported, the significance of the paper in its present form is limited to its empirical and heuristic contribution.

major comments (3)

[Appendix B.4, Assumption A3 and Proposition A3] Assumption A3 postulates that the expected client-selection update satisfies <G(theta), grad L(theta)> >= c ||grad L(theta)||^2. This inequality is exactly the property the proof of Proposition A3 requires to convert the Descent Lemma into a negative drift term; it is not derived from the coalition-formation or Boltzmann-exploration mechanisms. The statement that the assumption 'can be verified by our experimental results' is not a mathematical verification, and a finite set of accuracy curves cannot establish a pointwise inequality over the whole parameter space. The convergence guarantee is therefore circular: it assumes the alignment property that the theorem claims to establish.
[Appendix B.4, proof of Proposition A3] The proof inserts 'Without losing generality, we can assume that after T0 we are in a thermalized regime, where the spectral clustering has converged, and the clusters are stationary.' This is not a WLOG reduction: Algorithm 1 recomputes the spectral clustering at every round from the normalized client models, and no argument is given that the partition stabilizes in finite time. If the clusters change, the Boltzmann measure pi(t) and the expectation defining G(theta) change with t, so the telescoping sum from T0 to T has no stationary object to telescope over. The stationarity of the clusters after T0 must be stated as an explicit assumption, and the dependence of the final bound on T0 must be handled; the current proof simply discards T0 in the limit.
[Section 3 (Algorithm 1) versus Appendix B.4 (Proposition A3)] Proposition A3 analyzes a single SGD step of the form theta_gl(t+1) = theta_gl(t) - eta g(t) and assumes a known model covariance C, whereas Algorithm 1 estimates C_d(t) online via the Robbins-Monro update in Eq. (11), performs S local training steps, and updates the means of unobserved clients through Eq. (9). The proof does not bound the error between the estimated covariance C_d(t) and the true C_d, nor does it account for multiple local steps or the mean-update rule, so the theorem does not apply to the algorithm as actually implemented.

minor comments (4)

[Section 4, first paragraph] The informal convergence comment states that the global loss is non-increasing because it is a convex combination of non-increasing local losses; this is false in general, since each local loss is evaluated at a different local model and the averaged global model can increase the global loss under heterogeneity. This paragraph should be removed or qualified.
[Throughout] There are numerous typos and duplicated phrases that should be corrected, including 'identifying the most the most informative' in the abstract, 'results form estimation theory' and 'coherently' in Section 1, 'Warrren B Powell' in the references, 'termalization time' in Appendix B.4, and 'equation equation 2' in Section 2.
[Section 5.2 and Appendix C.1] The main text says S=10 local epochs while the setup in Appendix C.1 describes S=10 local steps/epochs with a batch size of 100; the terminology 'epochs' versus 'steps' should be made consistent.
[Appendix B.4, Assumption labels] The proof says 'Under Assumptions 1 and 2' but the assumptions are labeled Assumption A2 and Assumption A3; the cross-references should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the variance-reduction derivation is reproduced in the appendix, and the convergence theorem is a conditional result, not a tautology.

full rationale

The paper's main derivation chain is not circular. The variance-reduction formula (Corollary 1 and Proposition A1) is proved in Appendix B.2 from standard linear least-squares projection, so the credit to Raineri et al. (2025) — a prior work by one of the present authors — is not load-bearing: the argument is self-contained. The convergence claim (Proposition A3) is a genuine conditional theorem: given smoothness (A2) and the gradient-alignment assumption (A3), the Descent Lemma yields the stated ε-neighborhood bound. A3 is a strong and unverified sufficient condition, and the 'thermalized regime' WLOG in the proof is an unsupported assertion that narrows the theorem's applicability to the actual Algorithm 1. These are soundness/completeness gaps, not cases where a conclusion is definitionally identical to an input or where a fitted parameter is renamed as a prediction. The empirical benchmarks (MNIST, CIFAR-10/100, synthetic regression) are external and provide independent content. Hence no significant circularity.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claims rest on several assumptions that are either strong or unproven: the gradient alignment assumption essentially encodes the desired convergence behavior; the thermalized regime is asserted without proof; and the covariance matrices are treated as known in the theory while estimated online in practice. The empirical setup also introduces free hyperparameters (gamma, beta, warm-up rounds, weight subsampling) that are not accounted for in the theory.

free parameters (6)

gamma (homophily kernel bandwidth)
Parameter in the RBF similarity kernel W_kj = exp(-gamma ||theta_k - theta_j||^2) / normalization; no value or sensitivity analysis is reported.
beta (Boltzmann inverse temperature) = 1
Set to 1 in all experiments (Section 3, Appendix C.1); controls exploration-exploitation trade-off.
gamma_t (Robbins-Monro step size) = 1/t
Online covariance update schedule chosen as 1/t to ensure estimator convergence (Appendix C.1).
Warm-up rounds = 30
Initial uniform random sampling phase before FedCVR-Bolt policy is active; not in Algorithm 1 or theory.
Number of coalitions P = 10
Set equal to the participation rate P=10, assumed to be known a priori, though the true number of clusters is not known.
Sampled weight dimension for covariance = 300
For CIFAR models, covariance is computed on a random subset of D=300 final-layer weights (Appendix C.2), deviating from full-model theory.

assumptions (6)

domain assumption Covariance matrices C_d are invertible (Assumption A1)
Needed for the variance-reduction formula (C_AA)^{-1}; paper argues exact duplicate models are unlikely.
standard math Global loss L is L-smooth and federated gradient has bounded variance (Assumption A2)
Standard in FL convergence analysis.
ad hoc to paper Gradient alignment: expected update satisfies <G(theta), grad L(theta)> >= c ||grad L(theta)||^2 (Assumption A3)
Assumes the selection policy is already biased in a descent direction; this is effectively the desired convergence property.
ad hoc to paper Thermalized regime: after time T0, spectral clustering converges and clusters are stationary
Used in the proof of Proposition A3 to fix the Boltzmann sampling distribution; stated as 'without loss of generality' but not proven.
domain assumption Client models are random variables with known joint covariance, and the global model is a convex combination (Eq. 2)
Statistical learning setup used for variance reduction; in practice C_d is estimated online, not known.
domain assumption Homophily-based RBF similarity (Eq. 4) is a valid proxy for client data similarity
Spectral clustering relies on this kernel to define coalitions; alternative kernels are tested in an ablation but the choice remains heuristic.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Tackling Heterogeneity in Federated Learning via Variance-Reduced Boltzmann Sampling within Homogeneous Social Coalitions." pith.science (2026). https://pith.science/paper/5QDOTJJR

@misc{pith2026250602897,
  author       = {Pith},
  title        = {Pith review of: Tackling Heterogeneity in Federated Learning via Variance-Reduced Boltzmann Sampling within Homogeneous Social Coalitions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5QDOTJJR}},
  note         = {Machine review of arXiv:2506.02897}
}

read the original abstract

Federated Learning (FL) enables privacy-preserving collaborative model training, but its effectiveness is often limited by client data heterogeneity. We introduce a client-selection algorithm that (i) dynamically forms nonoverlapping coalitions of clients based on asymptotic agreement and (ii) selects one representative from each coalition to minimize the variance of model updates. Our approach is inspired by social-network modeling, leveraging homophily-based proximity matrices for spectral clustering and techniques for identifying the most informative individuals to estimate a group's aggregate opinion. We provide theoretical convergence guarantees for the algorithm under mild, standard FL assumptions. Finally, we validate our approach by benchmarking it against three strong heterogeneity-aware baselines; the results show higher accuracy and faster convergence, indicating that the framework is both theoretically grounded and effective in practice.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 11 canonical work pages

[1]

Due to Assumption 1, if the loss is L-smooth, we can apply the Descent Lemma, from Chapter 2, Theorem 2.1.5 in Nesterov (2014)

=θ gl(t)−ηg(t) , where g(t) is the federated gradient produced by FedCVR-Bolt. Due to Assumption 1, if the loss is L-smooth, we can apply the Descent Lemma, from Chapter 2, Theorem 2.1.5 in Nesterov (2014). Thus, we bound as follows: L(θgl(t+ 1))≤ L(θgl(t))−η⟨∇L(θ gl(t)), g(t)⟩+ Lη2 2∥g(t)∥2 . By taking the expectation up to thet-th round, denoted byE t, ...

work page 2014
[6]

Mime: Mimicking centralized stochastic algorithms in federated learning.arXiv preprint arXiv:2008.03606, 2020a

Sai Praneeth Karimireddy, Martin Jaggi, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. Mime: Mimicking centralized stochastic algorithms in federated learning.arXiv preprint arXiv:2008.03606, 2020a. Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh....

arXiv 2008
[8]

Quality-aware client selection and resource optimization for federated learning in computing networks

Yanyan Liao, Jie Feng, Zongjie Zhou, Bodong Shang, Lei Liu, and Qingqi Pei. Quality-aware client selection and resource optimization for federated learning in computing networks. InICC 2024-IEEE International Conference on Communications, pp. 2628–2633. IEEE,

work page 2024
[10]

Social network community detection using agglomerative spectral clustering.Complexity, 2017:1–10, 11

Ulzii-Utas Narantsatsralt and Sanggil Kang. Social network community detection using agglomerative spectral clustering.Complexity, 2017:1–10, 11

work page 2017
[14]

A game-theoretic federated learning framework for data quality improvement.IEEE Transactions on Knowledge and Data Engineering, 35(11):10952–10966, 2022a

Lefeng Zhang, Tianqing Zhu, Ping Xiong, Wanlei Zhou, and Philip S Yu. A game-theoretic federated learning framework for data quality improvement.IEEE Transactions on Knowledge and Data Engineering, 35(11):10952–10966, 2022a. Sai Qian Zhang, Jieyu Lin, and Qi Zhang. A multi-agent reinforcement learning approach for efficient client selection in federated l...

arXiv
[15]

A NOTATION For readability, we summarize below the notation conventions adopted throughout the paper

13 Preprint, Under Review Symbol Description Knumber of clients Kset of clients A ⊆ Ksubset of clients W(t)influence matrix among clients at roundt Tnumber of communication rounds Dnumber of model parameters Snumber of training iterations Pnumber of participating clients ηlearning rate θgl ∈R D global model θgl(t)∈R D global model at roundt θk(t)∈R D loca...

work page 2010
[16]

Given now that C d is positive definite since it is an invertible covariance matrix, then from Sylvester Criterion also its submatrixC d AA is invertible and the thesis follows

Thus, it holds C d AA(αA −ˆαA) +C A−Aα−A = 0, whereC d A−A ={C d ij}i∈A,j /∈Aandα −A ={α i}i /∈A. Given now that C d is positive definite since it is an invertible covariance matrix, then from Sylvester Criterion also its submatrixC d AA is invertible and the thesis follows. Building on this Lemma, we can now prove the main result used in the definition o...

work page 2025
[18]

+ LηM 2c For T→ ∞, T0 is negligible with respect to T . Hence, by finally adopting a change of variable and taking the limit with respect toT, we obtain the claim as lim T→∞ 1 T TX t=0 E[∥∇L(θgl(t))∥2]≤ LηM 2c =:ε 19 Preprint, Under Review C IMPLEMENTATIONDETAILS ANDFURTHEREXPERIMENTS In this appendix, we focus on the experimental setup, hyperparameters o...

work page 2018

Show all 20 references

[19]

While the distinctive server-side operations of FedCVR-Bolt , such as clustering based on individual models {θk(t)} and deriving ¯θk(t+

), it is possible to employ secure aggregation protocols (Bonawitz et al., 2016). While the distinctive server-side operations of FedCVR-Bolt , such as clustering based on individual models {θk(t)} and deriving ¯θk(t+

2016
[20]

for advanced tasks like personalization or clustering (Smith et al., 2017). Thus, FedCVR-Bolt ensures foundational data privacy by localizing data processing, and the principles of secure aggregation offer a complementary mechanism for enhancing model confidentiality during ag...

2017
[1998]

Fedmd: Heterogenous federated learning via model distillation.arXiv preprint arXiv:1910.03581,

Daliang Li and Junpu Wang. Fedmd: Heterogenous federated learning via model distillation.arXiv preprint arXiv:1910.03581,

1910 arXiv
[2001]

Distributed coalition formation in energy-aware cloud federations: A game-theoretic approach (extended version)

Marco Guazzone, Cosimo Anglano, Roberto Aringhieri, Matteo Sereno, et al. Distributed coalition formation in energy-aware cloud federations: A game-theoretic approach (extended version). CoRR, vol. abs/1309.2444,

arXiv
[2007]

Optimizing federated learning on non-iid data with reinforcement learning

Hao Wang, Zakhary Kaplan, Di Niu, and Baochun Li. Optimizing federated learning on non-iid data with reinforcement learning. InIEEE INFOCOM 2020-IEEE Conference on Computer Communications, pp. 1698–1707. IEEE, 2020a. Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papaili...

2020 arXiv
[2014]

Masato Ota, Yuko Sakurai, and Satoshi Oyama

ISBN 1461346916. Masato Ota, Yuko Sakurai, and Satoshi Oyama. Coalitional game theoretic federated learning. In 2022 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), pp. 48–55. IEEE,

2022
[2018]

Heterofl: Computation and communication efficient federated learning for heterogeneous clients.arXiv preprint arXiv:2010.01264,

Enmao Diao, Jie Ding, and Vahid Tarokh. Heterofl: Computation and communication efficient federated learning for heterogeneous clients.arXiv preprint arXiv:2010.01264,

2010 arXiv
[2019]

Leaf: A benchmark for federated settings.arXiv preprint arXiv:1812.01097,

Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Koneˇcn`y, H Brendan McMa- han, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated settings.arXiv preprint arXiv:1812.01097,

arXiv
[2020]

Active federated learning.arXiv preprint arXiv:1909.12641,

Jack Goetz, Kshitiz Malik, Duc Bui, Seungwhan Moon, Honglei Liu, and Anuj Kumar. Active federated learning.arXiv preprint arXiv:1909.12641,

1909 arXiv
[2021]

Practical secure aggregation for federated learning on user-held data.arXiv preprint arXiv:1611.04482,

Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for federated learning on user-held data.arXiv preprint arXiv:1611.04482,

arXiv
[2023]

22nd IFAC World Congress

ISSN 2405-8963. 22nd IFAC World Congress. Roberta Raineri, Giacomo Como, and Fabio Fagnani. Optimal selection of the most informative nodes for a noisy degroot model with stubborn agents.arXiv preprint arXiv:2504.08622,
[2024]

Noah Mark

doi: 10.1109/TMC.2024.3368473. Noah Mark. Culture and competition: Homophily and distancing explanations for cultural niches. American Sociological Review, 68:319–345, 06

2024

Pith tools

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

[1] [1]

Due to Assumption 1, if the loss is L-smooth, we can apply the Descent Lemma, from Chapter 2, Theorem 2.1.5 in Nesterov (2014)

=θ gl(t)−ηg(t) , where g(t) is the federated gradient produced by FedCVR-Bolt. Due to Assumption 1, if the loss is L-smooth, we can apply the Descent Lemma, from Chapter 2, Theorem 2.1.5 in Nesterov (2014). Thus, we bound as follows: L(θgl(t+ 1))≤ L(θgl(t))−η⟨∇L(θ gl(t)), g(t)⟩+ Lη2 2∥g(t)∥2 . By taking the expectation up to thet-th round, denoted byE t, ...

work page 2014

[2] [6]

Mime: Mimicking centralized stochastic algorithms in federated learning.arXiv preprint arXiv:2008.03606, 2020a

Sai Praneeth Karimireddy, Martin Jaggi, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. Mime: Mimicking centralized stochastic algorithms in federated learning.arXiv preprint arXiv:2008.03606, 2020a. Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh....

arXiv 2008

[3] [8]

Quality-aware client selection and resource optimization for federated learning in computing networks

Yanyan Liao, Jie Feng, Zongjie Zhou, Bodong Shang, Lei Liu, and Qingqi Pei. Quality-aware client selection and resource optimization for federated learning in computing networks. InICC 2024-IEEE International Conference on Communications, pp. 2628–2633. IEEE,

work page 2024

[4] [10]

Social network community detection using agglomerative spectral clustering.Complexity, 2017:1–10, 11

Ulzii-Utas Narantsatsralt and Sanggil Kang. Social network community detection using agglomerative spectral clustering.Complexity, 2017:1–10, 11

work page 2017

[5] [14]

A game-theoretic federated learning framework for data quality improvement.IEEE Transactions on Knowledge and Data Engineering, 35(11):10952–10966, 2022a

Lefeng Zhang, Tianqing Zhu, Ping Xiong, Wanlei Zhou, and Philip S Yu. A game-theoretic federated learning framework for data quality improvement.IEEE Transactions on Knowledge and Data Engineering, 35(11):10952–10966, 2022a. Sai Qian Zhang, Jieyu Lin, and Qi Zhang. A multi-agent reinforcement learning approach for efficient client selection in federated l...

arXiv

[6] [15]

A NOTATION For readability, we summarize below the notation conventions adopted throughout the paper

13 Preprint, Under Review Symbol Description Knumber of clients Kset of clients A ⊆ Ksubset of clients W(t)influence matrix among clients at roundt Tnumber of communication rounds Dnumber of model parameters Snumber of training iterations Pnumber of participating clients ηlearning rate θgl ∈R D global model θgl(t)∈R D global model at roundt θk(t)∈R D loca...

work page 2010

[7] [16]

Given now that C d is positive definite since it is an invertible covariance matrix, then from Sylvester Criterion also its submatrixC d AA is invertible and the thesis follows

Thus, it holds C d AA(αA −ˆαA) +C A−Aα−A = 0, whereC d A−A ={C d ij}i∈A,j /∈Aandα −A ={α i}i /∈A. Given now that C d is positive definite since it is an invertible covariance matrix, then from Sylvester Criterion also its submatrixC d AA is invertible and the thesis follows. Building on this Lemma, we can now prove the main result used in the definition o...

work page 2025

[8] [18]

+ LηM 2c For T→ ∞, T0 is negligible with respect to T . Hence, by finally adopting a change of variable and taking the limit with respect toT, we obtain the claim as lim T→∞ 1 T TX t=0 E[∥∇L(θgl(t))∥2]≤ LηM 2c =:ε 19 Preprint, Under Review C IMPLEMENTATIONDETAILS ANDFURTHEREXPERIMENTS In this appendix, we focus on the experimental setup, hyperparameters o...

work page 2018