Pith. sign in

REVIEW 4 major objections 6 minor 53 references

Mobility-Aware Asynchronous Federated Learning with Dynamic Sparsification

T0 review · 4 major / 6 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 The paper claims that device mobility has a dual, speed-dependent effect on asynchronous federated learning: moderate speed aids convergence by reducing staleness, while high speed degrades it by cutting contact time and raising…

desk verdict The convergence analysis is unsupported by the actual proofs; the experiments and the MADS heuristic are fine, but the theory needs major revision before it can be trusted. read the letter →

arxiv 2506.07328 v1 pith:J7DWKBIE submitted 2025-06-08 cs.LG

Jintao Yan , Tan Chen , Yuxuan Sun , Zhaojun Nan , Sheng Zhou , Zhisheng Niu This is my paper
classification cs.LG
keywords asynchronousfederatedlearningdevicemobilitymobileedgeservergradientsparsificationmodelstalenesscontacttimeconvergenceanalysisdynamic
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

This paper aims to establish that device mobility is not uniformly harmful to asynchronous federated learning (AFL): as average speed rises, convergence first improves and then worsens, so there is an optimal mobility regime. The mechanism is a trade-off between two quantities the paper defines and bounds: model staleness, which drops when devices meet the mobile edge server more often, and sparsification error, which grows when contact time is too short to upload all selected gradients. Based on this, the paper derives closed-form rules for how aggressively each device should sparsify gradients and how much power to use, giving the MADS algorithm.

What carries the argument

The machinery is a set of three linked bounds. A virtual model $v^{(r)}$ sums all historical local SGD updates as if they had been applied synchronously, and Lemma 1 bounds the divergence between this virtual model and each device's real local model; that divergence is exactly what staleness injects into the convergence bound. Lemma 2 bounds the second moment of staleness by $\Theta_n$, and Lemma 3 claims that the expected squared top-$k$ sparsification error is at most $(1-\gamma_n)\|x\|^2$, where top-$k$ sparsification keeps only the $k$ largest-magnitude gradient entries. The multiplication of these two effects, staleness and sparsification error, produces the $\Theta_n/\gamma_n^2$ term in Theorem 2, which is the object MADS minimizes online through the Lyapunov drift-plus-penalty reformulation.

What would settle it

A direct numerical check of Lemma 3: fix bit width $u$, model size $s$, rate $A$, and average contact time $c_n$, simulate many top-$k$ uploads, and measure $\mathbb{E}\|x-S(x)\|^2/\|x\|^2$; Lemma 3 predicts this ratio is at most $1-e^{-(u+\log_2 s)/(A c_n)}$ and decays with $c_n$. If the measured ratio stays near 1, the exponential contact-time dependence in Theorem 2 and the U-shaped speed result do not follow from the paper's analysis.

Watch

Extended reading notes

Core claim

The paper's central claim is that in asynchronous federated learning over a mobile network, mobility is not monotonically harmful: the expected squared gradient norm of the global loss after $R$ rounds satisfies $$E\|\nabla F($z^{{(R)}}$)\|^2 \le \frac{8L(F($w^{{(0)}}$)-F(w^*))}{\sqrt{R}}+\frac{2\$\sigma$}{N\sqrt{R}}+\frac{$G^{2}$}{NR}\sum_{n=1}^N \frac{(16-8\gamma_n-11\$gamma_n^{2}$+6\$gamma_n^{3}$)\Theta_n}{\$gamma_n^{2}$}.$$ The third term is the mobility term: $\gamma_n=e^{-(u+\log_2 s)/(A_n c_n)}$ is the fraction of the gradient vector that survives the contact-time-limited top-$k$ upload, and $\Theta_n$ bounds the second moment of model staleness. The coefficient of $\Theta_n$ falls as $\gamma_n$ grows, meaning longer contact time improves the bound twice over, while longer inter-contact time raises $\Theta_n$ and worsens it. Writing $c_n=C/v$ and $lambda_n=\Lambda/v$ for a common speed $v$ gives Corollary 1, a bound that decreases then increases in $v$; the paper reads this as: low mobility helps convergence, high mobility hurts it, and MADS is the control policy that tracks that trade-off.

Load-bearing premise

The main result depends on the claim that longer contact time makes gradient upload loss shrink exponentially; the proof of that claim in the manuscript currently shows only the trivial bound, so if the claim cannot be proven the U-shaped mobility result and the algorithm's motivation lose their foundation.

Editorial extensions

If this is right

  • Under the MADS rule, a device sends $k_n^{(r)}=\tau_n^{(r)}A_n^{(r)}/(u+\log_2 s)$ gradients, so longer contact times and higher rates mean less sparsification, and the optimal transmission power follows a closed-form expression that increases with model staleness.
  • If the convergence bound is right, fixed-sparsification AFL is suboptimal in mobile regimes: at 20 m/s on CIFAR-10, MADS outperforms AFL-Spar by 2.82 percentage points and FedMobile by 8.76 percentage points, and on Argoverse it reduces average displacement error by 9.46 percent.
  • Longer average inter-contact time slows AFL, so any mechanism that increases contact frequency, such as relaying or denser server deployment, should improve the same convergence bound even without changing the sparsification degree.
  • The online energy-constrained controller comes with a formal guarantee: compared with the offline optimum over $R$ rounds, the online solution of MADS has an optimality gap of order $R^2\Phi/V$ while keeping total device energy bounded, so the weight $V$ trades convergence speed against energy consumption.
  • For strongly non-IID data with concentration $\rho=0.1$, MADS reaches 65.66 percent test accuracy, exceeding AFL-Spar by 2.17 points, FedMobile by 7.13 points, AFL by 13.27 points, and SFL-Spar by 52.25 points.

Reading between the lines

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

  • Beyond the paper: because $\gamma_n$ depends on each device's rate $A_n$, the same bound predicts that each device has its own optimal speed; a heterogeneous fleet could schedule which devices upload most aggressively instead of treating speed as a single system parameter.
  • Beyond the paper: since $\gamma_n$ involves $u+\log_2 s$, the optimal mobility regime should shift with model precision and size, with low-precision gradients pushing the sweet spot to higher speeds and large models pulling it lower; this is testable by re-running the same controller with different bit widths.
  • Beyond the paper: the memoryless exponential contact model is a simplification; under heavy-tailed contact times the staleness moment $\Theta_n$ would likely be larger, so the qualitative U-shape may persist but at a lower optimal speed. This is an editorial extrapolation, not a claim of the paper.
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

4 major / 6 minor

Summary. This paper studies asynchronous federated learning (AFL) over a mobile network in which one mobile edge server (MES) is intermittently contacted by N mobile devices. Each device runs local SGD and, upon contact, uploads a top-k sparsified version of its accumulated gradient with error feedback, subject to contact-time and energy limits. The main theoretical contribution is a convergence analysis that closes the round-wise guarantee of Theorem 1 into the parameterized bound of Theorem 2, from which Corollary 1 and Remark 3 conclude that the convergence upper bound is U-shaped in device speed: low mobility helps by reducing model staleness, high mobility hurts by shrinking contact time and inflating sparsification error. On the algorithmic side, the paper reformulates the training-round objective as an online drift-plus-penalty problem (P1), decomposes it per device, and derives closed-form expressions for the sparsification degree and transmission power (Propositions 1 and 2), yielding the MADS algorithm with O(N) per-round complexity. The algorithm is evaluated on CIFAR-10 (ResNet-9, 20 devices) and Argoverse trajectory prediction (LaneGCN), reporting gains of 8.76% in accuracy and 9.46% in ADE over benchmarks, and the experimental accuracy-versus-speed curves in Figs. 5, 9, and 11 show the predicted non-monotone pattern.

Significance. If the central derivations were valid, this paper would provide a valuable quantitative account of mobility's dual role in AFL—reduced staleness versus reduced contact time—and a practically simple policy with closed-form solutions, O(N) per-round complexity, and visible gains over strong baselines on two datasets. The paper deserves credit for a clean system model, for making the falsifiable prediction of a U-shaped accuracy-versus-speed curve and confirming it experimentally (Figs. 5, 9, 11), and for benchmarking against FedMobile and an idealized optimal baseline. The overall assessment is nonetheless negative because the load-bearing theory is not established: Lemma 3's exponential sparsification-error bound is false as stated (Section IV, Eq. (4); the counterexample is detailed in Major Comment 2), Lemma 4 is unproven and deferred to nonexistent reference [58] (Appendix E), and Remark 2's monotonicity statements are sign-reversed. Since Theorem 2, Corollary 1, and Remark 3 all rest on these pieces, the theoretical foundation of the mobility claim cannot be verified.

major comments (4)
  1. [Appendix E, Lemma 4] Lemma 4 (Eq. (21)) is the load-bearing step that converts Theorem 1 into the closed-form bound of Theorem 2: it is the only ingredient that turns the error-feedback memory term E||e_n^(r)||^2 into the (1-gamma_n^2)/gamma_n^2 * Theta_n * eta^2 * G^2 expression that produces the gamma_n- and Theta_n-dependent term in Eq. (5). The entire proof is the sentence 'This is an extension of Lemma 2 in [46]. The proof is provided in [58].' Reference [58] does not appear in the bibliography, which ends at [53], and the cited [46] is a mobility-assisted routing paper that contains no bound on error-feedback memory from gradient sparsification. Without a self-contained proof of (21), the derivation in Appendix E cannot be checked, and Corollary 1 (Eq. (6)) and Remark 3 consequently have no verifiable foundation.
  2. [Section IV, Lemma 3; Appendix C] The exponential sparsification-error bound in Eq. (4) is not established by the proof and is false within the paper's own model. The proof (filed as Appendix C) derives E||x-S(x)||^2 <= (1-e^{-a}) * sum_{K=0}^{s-1} e^{-aK}(s-K)/s * ||x||^2 with a=(u+log2 s)/(A_n c_n), and then applies sum_{K=0}^{s-1} e^{-aK}(s-K)/s <= 1. This inequality is false: for a->0 the sum tends to (s+1)/2, which exceeds 1 for any s>=2. A concrete violation: with s=2, x=(1,2), and a=0.5, the expectation is (1-e^{-0.5})*5 + (e^{-0.5}-e^{-1})*1 ≈ 2.21, while the claimed bound is (1-e^{-0.5})*5 ≈ 1.97. In the paper's own parameter regime (s ≈ 6.6e6, A = 1 MHz, c_n = 10 s) the same series gives an upper bound of roughly 0.97*||x||^2 instead of the claimed 5.5e-6*||x||^2. Because Lemma 3 is the only link that makes sparsification error depend on contact time through gamma_n = e^{-a}, the speed dependence of Theorem 2 and the U-shape analysis in Corollary 1 and Remark 3 rest on a false premise; a corrected bound would carry a factor of order (s+1)/2 and change the closed form of (5).
  3. [Section IV, Remark 2] The monotonicity statements in Remark 2 have the wrong signs. From the definition gamma_n = e^{-(u+log2 s)/(A_n c_n)} it follows that gamma_n increases with c_n, not decreases as the text claims; and from Eq. (3), Theta_n = 1 + lambda_n/(lambda_n+c_n) * f(delta/lambda_n) with lambda_n/(lambda_n+c_n) decreasing in c_n, so Theta_n decreases with c_n, not increases. The remark also states that Theta_n decreases with lambda_n, whereas Theta_n grows without bound as lambda_n -> infinity. The concluding claim that increasing c_n improves convergence is correct in direction, but the reasons given for it (both gamma_n and Theta_n lowering the bound) are back-to-front: with the printed monotonicities, the conclusion would be a degradation, so the remark is internally inconsistent. The sign pattern of these monotonicities matters for the qualitative reading of Corollary 1 and Remark 3 and should be corrected.
  4. [Appendix F, Theorem 3] The proof of Theorem 3 is not complete as written. Equation (24) asserts the equality q_n^(r)*epsilon_n^{*(r)} = (q_n^(r)-q_n^(1))*phi_n, which does not follow from the definitions: epsilon_n^{*(r)} is the instantaneous excess energy of the offline optimum, whereas phi_n = max_r |epsilon_n^(r)| is a constant, and epsilon can be negative, so the bound q_n^(r)*epsilon_n^{*(r)} <= (r-1)*(phi_n)^2 is not established. In addition, the radical in Eq. (11), sqrt(2R^2*Phi - 2*sum_r*sum_n V*U^{*(r)}), can be negative when the penalty term dominates, in which case the claimed energy bound is not a real number. The drift-plus-penalty argument needs to be reworked (steps (23)-(24) and the final row of Appendix F) before Eqs. (10)-(11) can be accepted.
minor comments (6)
  1. [Appendices C and D] Appendix C is titled 'Proof of Lemma 2' but proves the sparsification-error bound stated as Lemma 3 (Eq. (4)), and Appendix D is titled 'Proof of Lemma 3' but proves the staleness bound stated as Lemma 2 (Eq. (3)); the proof pointers in the main text are therefore mismatched.
  2. [Section IV, Eq. (1)] Lemma 1's bound in Eq. (1) is typeset as a fraction with the staleness term over the sparsification term, but Appendix A derives a sum of four terms; as printed the equation is unreadable, and the coefficients in (1) do not match the combination of terms obtained in Appendix A.
  3. [Section VI.C] The text states that as client speed increases 'the ADE first increases and then decreases, consistent with the pattern observed on the CIFAR-10 dataset.' Since a lower ADE is better, the direction is backwards relative to the accuracy pattern in Fig. 5; the sentence should read 'first decreases and then increases.'
  4. [Conclusions and Abstract] In the Conclusions, 'reduces the ADE on the Argoverse trajectory prediction by dataset 9.46%' is garbled; also, the 8.76% accuracy gain and 9.46% ADE gain quoted in the Abstract are not anchored to a specific benchmark in the main text (the body gives 8.76% versus FedMobile at v = 20 m/s).
  5. [Section V.B, P3] For completeness I state that the convexity concern about P3 does not reproduce: the objective (14a) is -C*log2(1+cp) plus terms linear in p with C>0, which is convex in p, so the convexity claim and the KKT derivation in Proposition 2 are sound.
  6. [Section III.B] The relation between the round-based contact indicator zeta_n^(r) in Algorithm 1 and the continuous-time exponential contact/inter-contact process is never formalized (e.g., what happens when a contact spans multiple rounds, or when a contact partially overlaps a round); this ambiguity makes the renewal argument in Appendix D hard to evaluate precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the convergence derivation is not self-fulfilling; missing lemma proofs are support gaps, not circularity.

full rationale

The derivation chain is a standard convergence analysis rather than a construction that fits the target result into its own inputs. Theorem 1 is proved in Appendices A-B from L-smoothness, bounded stochastic gradients, a virtual-model decomposition, and an external error-feedback inequality (Lemma A.1 of [45]); Lemma 2's staleness bound is derived from the assumed exponential contact and inter-contact time model; Lemma 3 attempts to bound sparsification error in terms of contact time; and Theorem 2 then combines these pieces. The MADS algorithm optimizes the Theorem 1 bound, and the experiments report external accuracy and ADE values rather than re-fitting constants, so the main qualitative claims have independent empirical content. The self-citations ([29], [33]) appear only in related-work descriptions and do not carry the proof. No fitted parameter is renamed as a prediction, and no displayed equation is equivalent to an input by construction. What the paper does contain are load-bearing support gaps: the printed Appendix C proof of Lemma 3 does not actually establish the exponential factor (the displayed geometric sum is at least 1, so the final inequality does not follow), and Lemma 4's proof is deferred to a nonexistent reference — the text reads 'This is an extension of Lemma 2 in [46]. The proof is provided in [58].' Reference [58] is absent from the bibliography. These gaps mean Theorem 2 and Corollary 1 are not rigorously established, but they are omissions and mathematical errors, not circularity under the defined patterns.

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

The central claims rest on standard smoothness and bounded-gradient assumptions, on the exponential contact/inter-contact mobility model, and on three externally sourced lemmas. The most fragile is Lemma 4, whose proof is assigned to a nonexistent reference, and Lemma 3's stated bound is also not established by its proof. No new physical entities are introduced.

assumptions (6)
  • standard math Assumption 1: each local loss F_n is L-smooth.
    Used at the start of Section IV and in the proof of Theorem 1 to bound the descent of the virtual model.
  • standard math Assumption 2: stochastic gradient squared norm is bounded by G^2 and variance by sigma^2.
    Used throughout the convergence analysis and in Lemma 1, Theorem 1, and Lemma 4.
  • domain assumption Contact time and inter-contact time are exponentially distributed with means c_n and lambda_n.
    Adopted in Section III-B from the inter-contact model literature [15], [42]-[44]; underpins Lemmas 2 and 3.
  • domain assumption Average contact time and inter-contact time are inversely proportional to device speed, c_n = C/v and lambda_n = Lambda/v.
    Invoked before Corollary 1 with citation to mobility model results [46], [47]; needed to convert speed into contact and staleness effects.
  • standard math Error-feedback memory bound from Lemma A.1 of [45]: ||e||^2 <= (1 - k/s)||x||^2 for top-k sparsification with memory.
    Used in the proof of Lemma 1 and in the derivation of Theorem 1 and Theorem 2; adopted without proof from Stich et al.
  • standard math Lemma 4: E||e_n||^2 <= 4(1-gamma_n^2)/gamma_n^2 Theta_n eta^2 G^2.
    Used to prove Theorem 2, but the proof is deferred to reference [58], which does not appear in the bibliography, leaving this result unsupported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mobility-Aware Asynchronous Federated Learning with Dynamic Sparsification." pith.science (2026). https://pith.science/paper/J7DWKBIE

@misc{pith2026250607328,
  author       = {Pith},
  title        = {Pith review of: Mobility-Aware Asynchronous Federated Learning with Dynamic Sparsification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J7DWKBIE}},
  note         = {Machine review of arXiv:2506.07328}
}
read the original abstract

Asynchronous Federated Learning (AFL) enables distributed model training across multiple mobile devices, allowing each device to independently update its local model without waiting for others. However, device mobility introduces intermittent connectivity, which necessitates gradient sparsification and leads to model staleness, jointly affecting AFL convergence. This paper develops a theoretical model to characterize the interplay among sparsification, model staleness and mobility-induced contact patterns, and their joint impact on AFL convergence. Based on the analysis, we propose a mobility-aware dynamic sparsification (MADS) algorithm that optimizes the sparsification degree based on contact time and model staleness. Closed-form solutions are derived, showing that under low-speed conditions, MADS increases the sparsification degree to enhance convergence, while under high-speed conditions, it reduces the sparsification degree to guarantee reliable uploads within limited contact time. Experimental results validate the theoretical findings. Compared with the state-of-the-art benchmarks, the MADS algorithm increases the image classification accuracy on the CIFAR-10 dataset by 8.76% and reduces the average displacement error in the Argoverse trajectory prediction dataset by 9.46%.

Figures

Figures reproduced from arXiv: 2506.07328 by the authors.

Figure 1
Figure 1. The mobility-aware AFL framework. their AFL achieves a convergence rate comparable to those of synchronous methods. Building on this, Ref. [12] implements AFL in wireless networks and proposes three adaptive trans￾mission scheduling algorithms to enhance learning efficiency. Ref. [13] introduces a data-importance-aware scheduling pol￾icy that accounts for both data distribution and freshness. Ref. [16] leverages the… view at source ↗
Figure 3
Figure 3. Test accuracy w.r.t. training rounds under [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Test accuracy under different device speeds. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figures from the paper (2 more)
Figure 8
Figure 8. Figure 8: Test accuracy of the proposed algorithm compared with the benchmarks under different levels of non-i.i.d. data. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Test accuracy under different device speeds. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 49 canonical work pages

  1. [46]

    Performance analysis of mobility-assisted routing,

    T. Spyropoulos, K. Psounis, and C. S. Raghavendra, “Performance analysis of mobility-assisted routing,” inProc. ACM Int. Symp. Mobile Ad Hoc Netw. Comput. (MobiHoc), Florence, Italy, May 2006, pp. 49– 60

    work page 2006

  2. [53]

    Learning lane graph representations for motion forecasting,

    M. Liang, B. Yang, R. Hu, Y . Chen, R. Liao, S. Feng, and R. Urtasun, “Learning lane graph representations for motion forecasting,” inProc. European Conf. Comput. Vis. (ECCV), Glasgow, UK, Aug. 2020, pp. 541–556

    work page 2020

  3. [1]

    Task- oriented communications for 6G: Vision, principles, and technologies,

    Y . Shi, Y . Zhou, D. Wen, Y . Wu, C. Jiang, and K. B. Letaief, “Task- oriented communications for 6G: Vision, principles, and technologies,” IEEE Wireless Commun., vol. 30, no. 3, pp. 78–85, Sep. 2023

    work page 2023

  4. [2]

    A vision of 6G wireless systems: Applications, trends, technologies, and open research problems,

    W. Saad, M. Bennis, and M. Chen, “A vision of 6G wireless systems: Applications, trends, technologies, and open research problems,”IEEE Network, vol. 34, no. 3, pp. 134–142, May 2020

    2020

  5. [3]

    Edge artificial intelligence for 6G: Vision, enabling technologies, and applications,

    K. B. Letaief, Y . Shi, J. Lu, and J. Lu, “Edge artificial intelligence for 6G: Vision, enabling technologies, and applications,”IEEE J. Sel. Areas Commun., vol. 40, no. 1, pp. 5–36, Jan. 2022

    2022

  6. [4]

    Edge learning with timeliness constraints: Challenges and solutions,

    Y . Sun, W. Shi, X. Huang, S. Zhou, and Z. Niu, “Edge learning with timeliness constraints: Challenges and solutions,”IEEE Commun. Mag., vol. 58, no. 12, pp. 27–33, Dec. 2020

    work page 2020

  7. [5]

    Distributed learning in wireless networks: Recent progress and future challenges,

    M. Chen, D. G ¨und¨uz, K. Huang, W. Saad, M. Bennis, A. V . Feljan, and H. V . Poor, “Distributed learning in wireless networks: Recent progress and future challenges,”IEEE J. Sel. Areas Commun., vol. 39, no. 12, pp. 3579–3605, Oct. 2021

    work page 2021

  8. [6]

    A joint learning and communications framework for federated learning over wireless networks,

    M. Chen, Z. Yang, W. Saad, C. Yin, H. V . Poor, and S. Cui, “A joint learning and communications framework for federated learning over wireless networks,”IEEE Trans. Wireless Commun., vol. 20, no. 1, pp. 269–283, Oct. 2020

    work page 2020

Show all 53 references
  1. [7]

    MEET: Mobility-Enhanced Edge inTelligence for Smart and Green 6G Networks,

    Y . Sun, B. Xie, S. Zhou, and Z. Niu, “MEET: Mobility-Enhanced Edge inTelligence for Smart and Green 6G Networks,”IEEE Commun. Mag., vol. 61, no. 1, pp. 64–70, Oct. 2023

    2023

  2. [8]

    Mobility-aware multi-hop task offloading for autonomous driving in vehicular edge computing and networks,

    L. Liu, M. Zhao, M. Yu, M. A. Jan, D. Lan, and A. Taherkordi, “Mobility-aware multi-hop task offloading for autonomous driving in vehicular edge computing and networks,”IEEE Trans. Intell. Transp. Syst., vol. 24, no. 2, pp. 2169–2182, Jan. 2023

    2023

  3. [9]

    Multi-agent deep reinforcement learning for task offloading in UA V-assisted mobile edge computing,

    N. Zhao, Z. Ye, Y . Pei, Y .-C. Liang, and D. Niyato, “Multi-agent deep reinforcement learning for task offloading in UA V-assisted mobile edge computing,”IEEE Trans. Wireless Commun., vol. 21, no. 9, pp. 6949– 6960, Sept. 2022. 13

    2022

  4. [10]

    A comprehensive survey on mobility-aware D2D communications: Principles, practice and challenges,

    M. Waqas, Y . Niu, Y . Li, M. Ahmed, D. Jin, S. Chen, and Z. Han, “A comprehensive survey on mobility-aware D2D communications: Principles, practice and challenges,”IEEE Commun. Surveys Tuts., vol. 22, no. 3, pp. 1863–1886, 3rd Quart. 2020

    2020

  5. [11]

    Federated learning under arbi- trary communication patterns,

    D. Avdiukhin and S. Kasiviswanathan, “Federated learning under arbi- trary communication patterns,” inProc. Int. Conf. Mach. Learn. (ICML), ser. Proceedings of Machine Learning Research, M. Meila and T. Zhang, Eds., vol. 139, Virtual, Jul. 2021, pp. 425–435

    2021

  6. [12]

    Adaptive transmission scheduling in wireless networks for asynchronous federated learning,

    H.-S. Lee and J.-W. Lee, “Adaptive transmission scheduling in wireless networks for asynchronous federated learning,”IEEE J. Sel. Areas Commun., vol. 39, no. 12, pp. 3673–3687, Dec. 2021

    2021

  7. [13]

    Scheduling and aggregation design for asynchronous federated learning over wireless networks,

    C.-H. Hu, Z. Chen, and E. G. Larsson, “Scheduling and aggregation design for asynchronous federated learning over wireless networks,” IEEE J. Sel. Areas Commun., vol. 41, no. 4, pp. 874–886, Apr. 2023

    2023

  8. [14]

    Qsparse-Local-SGD: Distributed SGD with quantization, sparsification, and local computa- tions,

    D. Basu, D. Data, C. Karakus, and S. N. Diggavi, “Qsparse-Local-SGD: Distributed SGD with quantization, sparsification, and local computa- tions,”IEEE J. Sel. Areas in Inf. Theory, vol. 1, no. 1, pp. 217–226, 2020

    2020

  9. [15]

    Mobility-aware caching in D2D networks,

    R. Wang, J. Zhang, S. H. Song, and K. B. Letaief, “Mobility-aware caching in D2D networks,”IEEE Trans. Wireless Commun., vol. 16, no. 8, pp. 5001–5015, Aug. 2017

    2017

  10. [16]

    Accelerating asynchronous federated learning con- vergence via opportunistic mobile relaying,

    J. Bian and J. Xu, “Accelerating asynchronous federated learning con- vergence via opportunistic mobile relaying,”IEEE Trans. Veh. Technol., vol. 73, no. 7, pp. 10 668–10 680, Jul. 2024

    2024

  11. [17]

    Scheduling policies for federated learning in wireless networks,

    H. H. Yang, Z. Liu, T. Q. Quek, and H. V . Poor, “Scheduling policies for federated learning in wireless networks,”IEEE Trans. Commun., vol. 68, no. 1, pp. 317–333, Sept. 2019

    2019

  12. [18]

    Scheduling for cellular federated edge learning with importance and channel aware- ness,

    J. Ren, Y . He, D. Wen, G. Yu, K. Huang, and D. Guo, “Scheduling for cellular federated edge learning with importance and channel aware- ness,”IEEE Trans. Wireless Commun., vol. 19, no. 11, pp. 7690–7703, Aug. 2020

    2020

  13. [19]

    Broadband analog aggregation for low-latency federated edge learning,

    G. Zhu, Y . Wang, and K. Huang, “Broadband analog aggregation for low-latency federated edge learning,”IEEE Trans. Wireless Commun., vol. 19, no. 1, pp. 491–506, Jan. 2020

    2020

  14. [20]

    Dynamic scheduling for over-the-air federated edge learning with energy constraints,

    Y . Sun, S. Zhou, Z. Niu, and D. G ¨und¨uz, “Dynamic scheduling for over-the-air federated edge learning with energy constraints,”IEEE J. Sel. Areas Commun., vol. 40, no. 1, pp. 227–242, Nov. 2021

    2021

  15. [21]

    Gradient sparsification for communication-efficient distributed optimization,

    J. Wangni, J. Wang, J. Liu, and T. Zhang, “Gradient sparsification for communication-efficient distributed optimization,” inProc. Adv. Neural Inf. Process. Syst. (NIPS), Montr´eal, Canada, Dec. 2018, pp. 1299–1309

    2018

  16. [22]

    High-dimensional stochastic gradient quantization for communication-efficient edge learning,

    Y . Du, S. Yang, and K. Huang, “High-dimensional stochastic gradient quantization for communication-efficient edge learning,”IEEE Trans. Signal Process., vol. 68, pp. 2128–2142, Mar. 2020

    2020

  17. [23]

    Time-correlated sparsification for communication-efficient federated learning,

    E. Ozfatura, K. Ozfatura, and D. G ¨und¨uz, “Time-correlated sparsification for communication-efficient federated learning,” inin Proc. IEEE Int. Symp. Inf. Theory (ISIT), Melbourne, Vic, Australia, Jul. 2021, pp. 461– 466

    2021

  18. [24]

    UVeQFed: universal vector quantization for federated learning,

    N. Shlezinger, M. Chen, Y . C. Eldar, H. V . Poor, and S. Cui, “UVeQFed: universal vector quantization for federated learning,”IEEE Trans. Signal Process., vol. 69, pp. 500–514, Dec. 2021

    2021

  19. [25]

    Lazily aggregated quantized gradient innovation for communication-efficient federated learning,

    J. Sun, T. Chen, G. B. Giannakis, Q. Yang, and Z. Yang, “Lazily aggregated quantized gradient innovation for communication-efficient federated learning,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 4, pp. 2031–2044, Apr. 2022

    2022

  20. [26]

    Dynamic clustering and power control for two-tier wireless federated learning,

    W. Guo, C. Huang, X. Qin, L. Yang, and W. Zhang, “Dynamic clustering and power control for two-tier wireless federated learning,”IEEE Trans. Wireless Commun., vol. 23, no. 2, pp. 1356–1371, Feb. 2024

    2024

  21. [27]

    Communication-efficient and distributed learning over wireless networks: Principles and applications,

    J. Park, S. Samarakoon, A. Elgabli, J. Kim, M. Bennis, S.-L. Kim, and M. Debbah, “Communication-efficient and distributed learning over wireless networks: Principles and applications,”Proc. IEEE, vol. 109, no. 5, pp. 796–819, May 2021

    2021

  22. [28]

    Vehicle selection and resource optimization for federated learning in vehicular edge computing,

    H. Xiao, J. Zhao, Q. Pei, J. Feng, L. Liu, and W. Shi, “Vehicle selection and resource optimization for federated learning in vehicular edge computing,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 8, pp. 11 073–11 087, Aug. 2022

    2022

  23. [29]

    Dynamic scheduling for vehicle-to-vehicle communications enhanced federated learning,

    J. Yan, T. Chen, Y . Sun, Z. Nan, S. Zhou, and Z. Niu, “Dynamic scheduling for vehicle-to-vehicle communications enhanced federated learning,”IEEE Trans. Wireless Commun., early access, Jun. 2025

    2025

  24. [30]

    Ve- hicle selection and resource allocation for federated learning-assisted vehicular network,

    X. Zhang, Z. Chang, T. Hu, W. Chen, X. Zhang, and G. Min, “Ve- hicle selection and resource allocation for federated learning-assisted vehicular network,”IEEE Trans. Mobile Comput., vol. 23, no. 5, pp. 3817–3829, May 2024

    2024

  25. [31]

    Mobility-aware cluster federated learning in hierarchical wireless net- works,

    C. Feng, H. H. Yang, D. Hu, Z. Zhao, T. Q. S. Quek, and G. Min, “Mobility-aware cluster federated learning in hierarchical wireless net- works,”IEEE Trans. Wireless Commun., vol. 21, no. 10, pp. 8441–8458, Oct. 2022

    2022

  26. [32]

    How to tame mobility in federated learning over mobile networks?

    Y . Peng, X. Tang, Y . Zhou, Y . Hou, J. Li, Y . Qi, L. Liu, and H. Lin, “How to tame mobility in federated learning over mobile networks?” IEEE Trans. Wireless Commun., vol. 22, no. 12, pp. 9640–9657, Dec. 2023

    2023

  27. [33]

    Mobility accelerates learning: Convergence analysis on hierarchical federated learning in vehicular networks,

    T. Chen, J. Yan, Y . Sun, S. Zhou, D. G ¨und¨uz, and Z. Niu, “Mobility accelerates learning: Convergence analysis on hierarchical federated learning in vehicular networks,”IEEE Trans. Veh. Technol., vol. 74, no. 1, pp. 1657–1673, Jan. 2025

    2025

  28. [34]

    Exploiting moving intelligence: Delay-optimized computation offloading in vehicular fog networks,

    S. Zhou, Y . Sun, Z. Jiang, and Z. Niu, “Exploiting moving intelligence: Delay-optimized computation offloading in vehicular fog networks,” IEEE Commun. Mag., vol. 57, no. 5, pp. 49–55, May 2019

    2019

  29. [35]

    Adaptive learning-based task offloading for vehicular edge computing systems,

    Y . Sun, X. Guo, J. Song, S. Zhou, Z. Jiang, X. Liu, and Z. Niu, “Adaptive learning-based task offloading for vehicular edge computing systems,” IEEE Trans. Veh. Technol., vol. 68, no. 4, pp. 3061–3074, Jan. 2019

    2019

  30. [36]

    Joint task offloading and resource allocation for vehicular edge computing based on V2I and V2V modes,

    W. Fan, Y . Su, J. Liu, S. Li, W. Huang, F. Wu, and Y . Liu, “Joint task offloading and resource allocation for vehicular edge computing based on V2I and V2V modes,”IEEE Trans. Intell. Transp. Syst., vol. 24, no. 4, pp. 4277–4292, Jan. 2023

    2023

  31. [37]

    A cooperative computation offloading strategy with on-demand deployment of multi-UA Vs in UA V- aided mobile edge computing,

    C. Li, Y . Gan, Y . Zhang, and Y . Luo, “A cooperative computation offloading strategy with on-demand deployment of multi-UA Vs in UA V- aided mobile edge computing,”IEEE Trans. Netw. Service Manage., vol. 21, no. 2, pp. 2095–2110, Apr. 2024

    2024

  32. [38]

    Federated learning via unmanned aerial vehicle,

    M. Fu, Y . Shi, and Y . Zhou, “Federated learning via unmanned aerial vehicle,”IEEE Trans. Wireless Commun., vol. 23, no. 4, pp. 2884–2900, Apr. 2024

    2024

  33. [39]

    UA V-enabled asynchronous federated learning,

    Z. Zhai, X. Yuan, X. Wang, and H. Yang, “UA V-enabled asynchronous federated learning,”IEEE Trans. Wireless Commun., vol. 24, no. 3, pp. 2358–2372, Mar. 2025

    2025

  34. [40]

    Local SGD converges fast and communicates little

    S. U. Stich, “Local SGD converges fast and communicates little.” in Proc. Int. Conf. Learn. Represent. (ICLR), New Orleans, LA, USA, May 2019, pp. 1–17

    2019

  35. [41]

    The error-feedback framework: better rates for SGD with delayed gradients and compressed updates,

    S. U. Stich and S. P. Karimireddy, “The error-feedback framework: better rates for SGD with delayed gradients and compressed updates,”J. Mach. Learning Research, vol. 21, no. 1, Jan. 2020

    2020

  36. [42]

    Fixed point opportunistic routing in delay tolerant networks,

    V . Conan, J. Leguay, and T. Friedman, “Fixed point opportunistic routing in delay tolerant networks,”IEEE J. Sel. Areas Commun., vol. 26, no. 5, pp. 773–782, Jun. 2008

    2008

  37. [43]

    Characterising aggregate inter-contact times in heterogeneous opportunistic networks,

    A. Passarella and M. Conti, “Characterising aggregate inter-contact times in heterogeneous opportunistic networks,” inProc. Int. Conf. Res. Netw.,. Springer, Valencia, Spain, May 2011, pp. 301–313

    2011

  38. [44]

    Analysis of individual pair and aggre- gate intercontact times in heterogeneous opportunistic networks,

    A. Passarella and M. Conti, “Analysis of individual pair and aggre- gate intercontact times in heterogeneous opportunistic networks,”IEEE Trans. Mobile Comput., vol. 12, no. 12, pp. 2483–2495, Dec. 2013

    2013

  39. [45]

    Sparsified SGD with memory,

    S. U. Stich, J.-B. Cordonnier, and M. Jaggi, “Sparsified SGD with memory,” inProc. Adv. Neural Inf. Process. Syst. (NIPS), vol. 31, Montr´eal, Canada, Dec 2018, p. 4452–4463

    2018

  40. [47]

    Contact time in random walk and random waypoint: Dichotomy in tail distribution,

    C. Zhao and M. L. Sichitiu, “Contact time in random walk and random waypoint: Dichotomy in tail distribution,”Ad Hoc Networks, vol. 9, no. 2, pp. 152–163, Mar. 2011

    2011

  41. [48]

    M. J. Neely,Stochastic network optimization with application to com- munication and queueing systems. San Rafael, CA, USA: Morgan & Claypool, 2010

    2010

  42. [49]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, V . Nair, and G. Hinton, “Learning multiple layers of features from tiny images,”Tech. Rep., Toronto, ON, Canada, University of Toronto, Apr. 2009

    2009

  43. [50]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProc. IEEE/CVF Conf. on Comput. Vision and Pattern Recognition (CVPR), Las Vegas, NV , Jun. 2016, pp. 770–778

    2016

  44. [51]

    Study on channel model for frequencies from 0.5 to 100 ghz,

    3rd Generation Partnership Project (3GPP), “Study on channel model for frequencies from 0.5 to 100 ghz,”Tech. Rep. 38.901, v18.0.0, Sophia Antipolis, France, Mar. 2024

    2024

  45. [52]

    Argoverse: 3D tracking and forecasting with rich maps,

    M.-F. Chang, J. Lambert, P. Sangkloy, J. Singh, S. Bak, A. Hartnett, D. Wang, P. Carr, S. Lucey, D. Ramanan, and J. Hays, “Argoverse: 3D tracking and forecasting with rich maps,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Long Beach, CA, USA, Jun. 2019, pp. ...

    2019

Pith tools

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