Pith. sign in

REVIEW 3 major objections 5 minor 49 references

FedCritic-MIMO: Communication-Efficient Serverless Federated Critic Learning for Massive-MIMO Resource Control in Open and Disaggregated 6G RANs

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Serverless exchange of compressed critic parameters lets interference-coupled 6G controllers coordinate without a central trainer, cutting critic traffic by roughly 76% while matching centralized QoS.

desk verdict A solid engineering contribution with an honest but conditional theory; the central coordination claim is plausible and worth refereeing, but the headline rate depends on an unproved tracking-error condition and Lemma 1's statement outruns its proof. read the letter →

arxiv 2608.03852 v1 pith:FNB7KN6N submitted 2026-08-04 cs.LG cs.MAcs.NI

Amin Farajzadeh , Melike Erol-Kantarci This is my paper · ORCID
classification cs.LGcs.MAcs.NI
keywords AI-native6GRANOpenfederatedreinforcementlearningmulti-agentserverlessmassiveMIMOresourceallocationcommunication-efficient
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 argues that coordinating a dense, interference-limited 6G radio access network does not require a central trainer, a parameter server, or shared policies: neighboring base-station controllers can coordinate by exchanging only the shared part of their critic (value-estimation) networks, sent only when the update matters and compressed per layer. It proposes FedCritic-MIMO, which combines an event trigger driven by critic change, queue pressure, and interference intensity; layer-wise top-$k$ sparsification with error feedback; and balanced fusion that weights strongly coupled neighbors while preserving the network average of the shared critic. Under a fixed-policy, frozen-target critic-regression model, the paper derives conditional finite-time stationarity and consensus bounds for the shared-critic recursion, with a stationarity rate of $O(T^{-1/2})+O(\log T/T)$ when the accumulated tracking error decays as $O(\log T/T)$. In reuse-1 massive-MIMO OFDMA simulations the method achieves the best performance-communication tradeoff among the considered baselines, raising held-out throughput and mean SINR while cutting critic-communication overhead by about 76% relative to uncompressed distributed critic exchange. A sympathetic reader would take the central claim to be that critic-only, sparsely exchanged value information is sufficient to coordinate otherwise independent controllers.

What carries the argument

The load-bearing object is the balanced, compressed peer-to-peer critic recursion $\Psi^{t+1} = (\Psi^t - \eta_c G^t)(W^t)^T + R^t$, where $\Psi^t$ stacks the shared critic parameters across base stations, $G^t$ stacks local critic gradients, $W^t$ is a symmetric doubly stochastic fusion matrix built from two-sided interference-relevance scores, and $R^t$ is the perturbation caused by event triggering and top-$k$ compression. The key identity is Lemma 2: because $W^t$ is doubly stochastic and symmetric, the network average satisfies $\bar{\psi}^{t+1} = \bar{\psi}^t - \eta_c \bar{g}^t$, exactly the update of a centralized gradient method, while the perturbation obeys $\frac{1}{N}\mathbb{E}\|R^t\|_F^2 \le 4\varepsilon_t^2$, with $\varepsilon_t^2$ the mean squared distance between local shared-critic parameters and the public reconstructions neighbors hold. The disagreement recursion $Z^{t+1} = (Z^t - \eta_c G^t\Pi)(A^t)^T + R^t$ with $\|A^t\|_2 \le \lambda_W < 1$ then yields the consensus bound, so both the theory and the practical benefit hinge on keeping the tracking error small relative to the spectral gap $1-\lambda_W$.

What would settle it

Measure the mean squared tracking error $\varepsilon_t^2 = \frac{1}{N}\sum_{n=1}^N \mathbb{E}\|e_t^n\|_2^2$ between each controller's local shared-critic parameters and the public reconstruction its neighbors hold, in the same 7-cell reuse-1 setup; if the running average $\frac{1}{T}\sum_{t=0}^{T-1}\varepsilon_t^2$ does not decay at the $O(\log T/T)$ rate, Corollary 1's advertised stationarity rate does not follow from the theory and the method's practical value rests on the simulation operating point alone. A second, cheaper check is to run a periodic-compressed baseline at the same sparsification budget $\rho_c = 0.20$; the paper's own discussion predicts the event trigger contributes little at this operating point, because the trigger fires in most training rounds.

Watch

Extended reading notes

Core claim

FedCritic-MIMO's central claim is that compatible shared critic parameters exchanged peer-to-peer over an interference-aware graph can coordinate RAN controllers without centralized trajectory collection, parameter-server aggregation, or actor homogenization. Each base station keeps its own actor, local experience, and personalized critic head; only a designated shared critic subnetwork $\psi^n_t$ is transmitted, and only when a utility score $\Gamma^n_t$ combining critic innovation, queue urgency, and interference intensity crosses a threshold. Messages are layer-wise top-$k$ compressed with error feedback, and fusion uses symmetric interference-relevance weights $W_t$ that are doubly stochastic, so the network average of the shared critic parameters evolves exactly like a centralized gradient recursion while a perturbation controlled by the tracking error $\varepsilon_t^2$ enters the stationarity and consensus bounds (Theorem 1). Under the decay condition $\frac{1}{T}\sum_{t=0}^{T-1}\varepsilon_t^2 = O(\log T/T)$, the randomized-iterate stationarity rate is $O(T^{-1/2})+O(\log T/T)$ (Corollary 1); Remark 1 states explicitly that this decay condition is not proved. In simulations over seven reuse-1 massive-MIMO cells, the framework attains the highest held-out reward, a mean SINR about 1.7–2.5 dB above the strongest competing learning baselines, QoS comparable to the centralized-training baseline, the lowest interference cost per delivered bit, and about 76% lower critic-communication overhead than uncompressed distributed critic exchange.

Load-bearing premise

The finite-time stationarity rate is conditioned on a decay property of the accumulated tracking error — the combined error from skipped communications and compressed critic updates — that the paper explicitly does not prove; if that error does not shrink at least as fast as $O(\log T/T)$, the guarantee collapses to the simulation results alone.

Editorial extensions

If this is right

  • Critic-only, sparsely exchanged value information is sufficient for coordination: controllers improve throughput, user-rate distribution, mean SINR, QoS satisfaction, and interference efficiency while keeping actors and personalized critic heads private.
  • The serverless architecture matches the centralized-training baseline (CTDE-MAPPO) on QoS and beats it on interference cost, so removing the central trainer need not sacrifice coordination quality.
  • The advertised $O(T^{-1/2})+O(\log T/T)$ stationarity rate holds only for the fixed-policy, frozen-target critic-regression phase and only if the accumulated tracking error decays as $O(\log T/T)$; it does not certify the full actor-critic process or global optimality.
  • At the reported operating point, the roughly 76% communication saving comes mainly from layer-wise sparse exchange with error feedback rather than from fewer communication rounds, since the event trigger fires in most training rounds.

Reading between the lines

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

  • A direct empirical check of whether $\frac{1}{T}\sum_{t=0}^{T-1}\varepsilon_t^2$ decays as $O(\log T/T)$ in the same 7-cell setup would reveal whether the advertised rate is realized in practice or whether the method's value rests on simulations alone; the paper does not perform this check.
  • The same trigger-compress-fuse machinery should transfer to asynchronous controllers, time-varying interference graphs, and imperfect CSI, which the paper lists as future work; the doubly stochastic fusion property and the average-preservation identity would need revisiting when exchange is asymmetric or graph edges change.
  • The paper's own note that the QoS bar and the user-rate CDF must come from the same samples if both measure the same quantity invites a re-derivation of the reported satisfaction ratio directly from the held-out rate samples as a consistency check.
  • Sweeping the compression budget $\rho_c$ and the trigger floor $\tau_{th,\min}$ should trace a communication-performance frontier shaped by $\bar{\varepsilon}^2/(1-\lambda_W)^2$; a reader could test whether the tradeoff curve bends as the theory predicts.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes FedCritic-MIMO, a decentralized multi-agent reinforcement learning framework for downlink resource control in reuse-1 multi-cell massive-MIMO OFDMA networks. Each cell-level controller keeps its own actor and personalized critic head, and coordination is achieved by exchanging only a shared critic subnetwork over an interference-aware graph, using utility-aware event-triggered communication, adaptive layer-wise top-k compression with error feedback, and balanced interference-aware fusion. The authors formulate the problem as a Dec-POMDP, provide a conditional convergence analysis of the shared-critic recursion under fixed-policy and frozen-target assumptions, and evaluate the method against heuristic, independent-learning, centralized-training, and communication-ablation baselines. The simulations report improved held-out throughput, SINR, interference efficiency, and an approximately 76% reduction in critic communication relative to uncompressed distributed critic exchange, with QoS performance comparable to a centralized-training baseline.

Significance. If the claims hold, the paper addresses a genuinely important architectural problem: coordinating independently deployable RAN controllers without a central trainer or parameter server. The empirical study is carefully set up with common warm starts, multiple seeds, held-out evaluation, and communication-aware baselines, and the paper is unusually candid in its limitations (Remark 1 on the tracking-error condition and Remark 2 on the scope of the theory). The main significance risk is that the advertised finite-time rate depends on an unproven condition, and the stated one-step tracking-error lemma is internally inconsistent. These issues do not invalidate the empirical contribution, but they must be repaired before the theoretical claims can be taken as established.

major comments (3)
  1. [Section IV-G, Corollary 1, Eq. (77), Remark 1] The advertised O(T^{-1/2}) + O(log T / T) stationarity rate is conditional on the accumulated tracking-error decay 1/T Σ_{t=0}^{T-1} ε_t^2 = O(log T / T), and Remark 1 explicitly states that this condition is not proved. Since no explicit trigger-threshold schedule, compression budget, or auxiliary argument is given that would imply (77), the finite-time guarantee is not established for the proposed mechanism under its stated assumptions. The abstract and the contribution list present this rate as a central result; the paper should either prove (77) under an explicit schedule or clearly downgrade the claim to a conditional statement whose hypothesis remains unverified.
  2. [Lemma 1 and Appendix A] The statement of Lemma 1 disagrees with its proof. In the no-trigger case, (45) gives Γ_n^t = ||~ψ^{t+1}_n - hatψ^t_n||^2 / (||hatψ^t_n||^2 + ε_tr) · (1 + α_Q Q̄^t_n + α_I Ī^t_n) < τ_th, so ||e^{t+1}_n||^2 < τ_th(||hatψ^t_n||^2 + ε_tr). The lemma instead claims the much tighter bound τ_th^2(||hatψ^t_n||^2 + ε_tr)^2, which does not follow from the trigger condition as written. Because Lemma 1 is the paper's only concrete tool for controlling the tracking error that enters Theorem 1 and Corollary 1, this inconsistency is load-bearing and must be corrected, along with any subsequent uses of the stated bound.
  3. [Section V-F and Section I-B (Contributions)] The contribution list claims improved QoS satisfaction, but the held-out results show overlapping confidence intervals with CTDE-MAPPO, and the text in Section V-F explicitly states that the figure supports comparable QoS rather than a statistically significant advantage. The abstract and contribution wording should be qualified so that the QoS claim matches the evidence; the stronger and better-supported improvements are throughput, SINR, and interference efficiency.
minor comments (5)
  1. [Section V-G] The QoS-satisfaction metric should be defined precisely in one place, and the relationship between the QoS bar in Fig. 4(a) and the rate CDF in Fig. 5 should be made explicit. The paper itself flags this ambiguity; it would be better to resolve it in the text.
  2. [Section V-E] The comparison between Periodic-Full and Event-Uncompressed shows that the trigger remains active in most rounds, so the measured communication savings are attributed mainly to compression. Adding a periodic-compressed baseline would isolate the contribution of event triggering; the paper notes this limitation, but the missing baseline weakens the ablation story.
  3. [Appendix A] The proof of Lemma 1 uses a strict inequality in the no-trigger case while the lemma statement uses a non-strict inequality; trivial, but the statement and proof should be consistent after the bound is corrected.
  4. [Section IV-H] Equation (79) defines the directed logical payload with the factor |B_n|, but the reported overhead in Fig. 3 should state explicitly whether this factor is included or whether physical broadcast is assumed. The text explains both options, but the figure caption does not say which one was used.
  5. [Section II-B] The claim that repeatedly solving (17) steers the system toward satisfying the long-term average-rate constraints would benefit from a brief justification or a pointer to the standard drift-plus-penalty argument, since this link is used to motivate the surrogate objective.

Circularity Check

0 steps flagged · score 2.0 of 10

No meaningful circularity: the convergence claim is explicitly conditional and self-contained, with only a minor non-load-bearing self-citation.

full rationale

The paper's derivation chain is self-contained rather than circular. Theorem 1 is proved from Assumption 1 and Lemma 2 using a standard decentralized SGD recursion in which the event-triggering and top-k tracking error appears as an explicit perturbation term ε_t^2. Corollary 1 does not silently assume away the mechanism: it states condition (77) on the accumulated tracking error as an additional, unproved decay assumption, and Remark 1 openly says that the paper does not establish this condition. Thus the advertised O(T^{-1/2}) + O(log T/T) rate is a conditional implication, not a tautology, and not a fitted parameter renamed as a prediction. The stated Lemma 1 proof gives ||e^{t+1}_n||^2 < τ_th(||hat-ψ^t_n||^2 + ε_tr) while the lemma statement contains τ_th^2(||hat-ψ^t_n||^2 + ε_tr)^2; this is a proof-strength mismatch and a correctness risk, but not circularity, and Theorem 1's proof does not use Lemma 1 to obtain its main bound. The empirical communication reduction is also not forced by construction: it is measured against Periodic-Full and Event-Uncompressed baselines, and the paper explicitly acknowledges that isolating the event-trigger gain would require a periodic-compressed baseline, so no missing control arm is hidden. The sole self-referential element is citation [34], the authors' own earlier FedCritic framework, used only as the architectural starting point and positioning statement; no load-bearing theorem, uniqueness claim, or fitted value from [34] is imported. Therefore no circular step can be exhibited, and the appropriate finding is no significant circularity, with a minor novelty-reducing self-citation only.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The empirical results depend on several hand-chosen communication and fusion hyperparameters (trigger thresholds, compression budget, weighting coefficients). The convergence theorem rests on standard decentralized-optimization regularity assumptions plus an unproved tracking-error decay condition. No new physical entities are introduced; all components are algorithmic.

free parameters (5)
  • Trigger threshold schedule τ_th,0 and τ_th,min = 0.020 initial, 0.001 floor
    Sets how often critic messages are sent; the 76% communication saving depends on this operating point.
  • Compression budget ρ_c and per-layer ratio bounds = ρ_c=0.20, bounds [0.10,0.25]
    Determines top-k sparsity per layer and therefore the measured communication reduction.
  • Utility trigger coefficients α_Q, α_I = α_Q=1.0, α_I=1.5
    Balance queue urgency and interference intensity in the event trigger.
  • Local fusion weight ω_self = 0.55
    Controls how much each controller trusts its own shared critic relative to neighbors in balanced fusion.
  • Queue normalization q0 and reward scaling = q0=10, scaling=0.01
    Shape virtual-queue urgency and reward magnitudes; chosen by hand, not justified from first principles.
assumptions (6)
  • domain assumption Each local critic objective F_n is L-smooth and lower bounded, with unbiased stochastic gradients and bounded variance and heterogeneity (Assumptions A2-A4).
    Invoked in Theorem 1; not verified for the PPO actor-critic training used in the simulations.
  • domain assumption The fusion matrices W^t are symmetric, doubly stochastic, and satisfy a uniform mixing condition with λ_W<1 (Assumption A1).
    Needed for the consensus part of Theorem 1; depends on the interference graph and the defined weights.
  • ad hoc to paper Fixed joint policy, frozen personalized critic heads, and frozen regression targets during the analyzed optimization window (Section IV-G).
    The actual algorithm updates local actors and critic heads concurrently; the theorem covers a simplified fixed-policy, frozen-target recursion (Remark 2).
  • ad hoc to paper Tracking-error decay 1/T Σ ε_t^2 = O(logT/T) (Eq. 77).
    The central rate result depends on this condition, which the paper explicitly does not prove (Remark 1).
  • domain assumption Common shared-critic initialization ψ_0 and feasibility of the minimum-rate vector.
    Used to start the recursion and to keep virtual queues finite; standard for queue-based QoS formulations.
  • domain assumption Reliable synchronous peer-to-peer communication over a fixed undirected coordination graph.
    The algorithm and analysis assume no message loss or asynchrony during training; future work lists asynchrony as open.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FedCritic-MIMO: Communication-Efficient Serverless Federated Critic Learning for Massive-MIMO Resource Control in Open and Disaggregated 6G RANs." pith.science (2026). https://pith.science/paper/FNB7KN6N

@misc{pith2026260803852,
  author       = {Pith},
  title        = {Pith review of: FedCritic-MIMO: Communication-Efficient Serverless Federated Critic Learning for Massive-MIMO Resource Control in Open and Disaggregated 6G RANs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FNB7KN6N}},
  note         = {Machine review of arXiv:2608.03852}
}
abstract

This paper proposes FedCritic-MIMO, a communication-efficient serverless federated multi-agent reinforcement learning framework for AI-native resource control across independently deployable cell-level controllers in open and disaggregated 6G RANs. Controllers share no trainer, retain local actors and personalized critic components, and exchange only compatible shared critic parameters. FedCritic-MIMO targets reuse-$1$ multi-cell massive-MIMO OFDMA deployments, where RAN controllers jointly manage user scheduling, per-stream power allocation, beamforming, interference, and long-term QoS with limited inter-controller signaling. Each base station locally executes its actor without centralized training or actor federation, while critic knowledge is exchanged peer-to-peer over an interference-aware graph. It enables this collaboration through wireless-aware event triggering, adaptive layer-wise top-$k$ sparse critic exchange with error feedback, and balanced interference-aware fusion. We establish conditional finite-time stationarity and consensus guarantees for the balanced, compressed peer-to-peer critic recursion under a fixed-policy, frozen-target critic-regression model. In strongly interference-coupled reuse-$1$ simulations, FedCritic-MIMO achieves the best performance-communication tradeoff among heuristic, independent-learning, centralized-training, and communication-ablation baselines. It achieves the highest held-out throughput, improves user-rate distribution and mean SINR, increases QoS satisfaction, and attains the lowest interference cost per delivered bit among learning baselines. It reduces critic-communication overhead by $76\%$ relative to uncompressed distributed critic exchange. These results demonstrate that serverless exchange of compatible shared critic parameters can coordinate RAN controllers without centralized trajectory collection or parameter-server aggregation.

Figures

Figures reproduced from arXiv: 2608.03852 by the authors.

Figure 2
Figure 2. Held-out episodic reward distribution. 0 0 00  0 00  0 " %!$ 00 0 0 0 [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Cumulative training-side communication overhead. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 1
Figure 1. Validation learning dynamics over training updates [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Held-out QoS and SINR performance of the selected [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Empirical CDF of per-UE time-average held-out rates [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: QoS–interference operating points of the evaluated [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 43 canonical work pages

  1. [1]

    AI-Native Control in Open and Disaggregated RANs: Open and disaggregated RAN architectures expose programmable control functions and interfaces for intelli gent radio-resource management [4]. Recent work has studied learning-enabled control for O-RAN resource management, including joint scheduling, O-RU association, and power allocation [12], as well as f...

    work page

  2. [2]

    These methods enable wireless controllers to adapt to time-varying channels and traffic without repeatedly solving nonconvex optimization proble ms online

    Learning-Based Wireless Resource Allocation and Beamforming: Deep reinforcement learning (DRL) and MARL have been widely studied for dynamic spectrum access, user scheduling, power control, and interference management [14], [15]. These methods enable wireless controllers to adapt to time-varying channels and traffic without repeatedly solving nonconvex opt...

    work page

  3. [3]

    However, collecting channel, queue, interference, schedu ling, and beamforming information from multiple BSs creates substantial signaling and scalability requirements [21]

    Centralized and Decentralized MARL: Centralized and CTDE-based MARL methods use centralized critics, joint observations, value decomposition, or global training sig nals to improve coordination among interacting agents [19], [20 ]. However, collecting channel, queue, interference, schedu ling, and beamforming information from multiple BSs creates substant...

    work page

  4. [4]

    Federated and Communication-Efficient MARL: Federated learning supports collaborative model training without exchanging raw local data, while FedAvg and FedProx provide foundational aggregation mechanisms under statis tical and systems heterogeneity [26], [27]. Federated MARL has been applied to wireless and edge control [28], including channel assignment ...

    work page

  5. [5]

    Our ea rlier FedCritic framework introduced serverless critic collabo ration for scheduling and power control in multi-cell OFDMA networks [34]

    Positioning of This W ork: General decentralized actor–critic and compressed-consensus methods provide th e algorithmic foundations for peer-to-peer learning. Our ea rlier FedCritic framework introduced serverless critic collabo ration for scheduling and power control in multi-cell OFDMA networks [34]. FedCritic-MIMO addresses the additional architectural...

    work page

  6. [6]

    critic-only federation : only critic-side information is exchanged across BSs, while actors remain local

    work page

  7. [7]

    utility-aware event-triggered communication : a BS communicates only when its critic update is sufficiently important, as determined jointly by critic innovation, local queue urgency, and interference intensity

    work page

  8. [8]

    compressed balanced interference-aware fusion : when communication is triggered, only a compressed shared- critic increment is exchanged and fused over the interference graph using symmetric interference- relevance weights. B. Local Actor–Critic Parameterization and Objectives For BS n ∈ N , let θt n denote the local actor parameters. To allow value-funct...

    work page

Show all 49 references
  1. [9]

    (65) Proof: See Appendix B

    (62) Then Rt1 = 0, (63) ¯ψt+1 = ¯ψt −ηc¯gt, ¯gt ≜ 1 N N∑ n=1 gt n, (64) 1 N E‖Rt‖2 F ≤ 4ε2 t. (65) Proof: See Appendix B. G. Conditional Convergence of the Shared Critic Recursion Convergence analysis considers a fixed joint policy π whose induced trajectory process admits a st...

  2. [10]

    (A1) The matrices Wt are symmetric and doubly stochastic, and satisfy the following uniform mixing condition: there exists λW ∈ [0,1) such that‖ ‖ ‖ ‖Wt − 1 N 11T ‖ ‖ ‖ ‖ 2 ≤λW, ∀t

    (67) Assumption 1. (A1) The matrices Wt are symmetric and doubly stochastic, and satisfy the following uniform mixing condition: there exists λW ∈ [0,1) such that‖ ‖ ‖ ‖Wt − 1 N 11T ‖ ‖ ‖ ‖ 2 ≤λW, ∀t. (68) (A2) Each Fn is lower bounded and L-smooth, i.e., ‖∇Fn(ψ)−∇Fn(ψ′)‖2 ≤L‖...

  3. [11]

    APPENDIX B PROOF OF LEMMA 2 Proof: Because Wt is doubly stochastic, (IN − (Wt)T )1 = 0

    These two cases prove (54). APPENDIX B PROOF OF LEMMA 2 Proof: Because Wt is doubly stochastic, (IN − (Wt)T )1 = 0. Equation (61) therefore gives Rt1 = 0. Right-multiplying (60) by 1/N then gives (64). Finally, ‖IN −(Wt)T ‖2 ≤ 2 for a symmetric stochastic matrix; hence, 1 N E‖...

  4. [12]

    For fixed ηc, taking the limit superior in (74) and using limsupT →∞T −1∑ t<Tε2 t ≤ ¯ε2 gives the stated stationarity neighborhood.REFERENCES

    Substituting ηc =cη/ √ T and (77) into (74) gives, respectively, O(T −1/2), O(T −1/2), O(T −1), and O(log T/T ) for its four right- hand-side terms, proving (78). For fixed ηc, taking the limit superior in (74) and using limsupT →∞T −1∑ t<Tε2 t ≤ ¯ε2 gives the stated stationari...

  5. [13]

    6G cellular networks: Mapping the landscape for the IMT-2030 framework,

    E. Hossain and A. I. V era-Rivera, “6G cellular networks: Mapping the landscape for the IMT-2030 framework,” IEEE Trans. Technol. Soc. , vol. 6, no. 4, pp. 377–392, Dec. 2025

    2025

  6. [14]

    Operator’s perspective on 6G: 6G services, vision, and spectrum,

    M. Na et al. , “Operator’s perspective on 6G: 6G services, vision, and spectrum,” IEEE Commun. Mag., vol. 62, no. 8, pp. 178–184, Aug. 2024

    2024

  7. [15]

    Open RAN for 6G networks: Architecture, use cases and open issues,

    B. Agarwal et al. , “Open RAN for 6G networks: Architecture, use cases and open issues,” IEEE Commun. Surveys Tuts. , vol. 28, pp. 2881–2924, Secondquarter 2026

    2026

  8. [16]

    Empowering the 6G cellular architecture with open RAN,

    M. Polese et al. , “Empowering the 6G cellular architecture with open RAN,” IEEE J. Sel. Areas Commun. , vol. 42, no. 2, pp. 245–262, Feb. 2024

    2024

  9. [17]

    Time synchronous OFDMA for dense wireless access in open-RAN,

    F. Mazzenga and A. Vizzarri, “Time synchronous OFDMA for dense wireless access in open-RAN,” IEEE Commun. Lett. , vol. 30, pp. 66–70, 2026

    2026

  10. [18]

    Toward standardization of 6G and NextG: Key technologies to enable fundamental enhancements,

    W. Chen et al. , “Toward standardization of 6G and NextG: Key technologies to enable fundamental enhancements,” IEEE J. Sel. Areas Commun., vol. 44, pp. 4333–4365, Mar. 2026

    2026

  11. [19]

    Massive MIMO over correlated fading channels: Multi-cell MMSE processing, pilot assignment an d power control,

    M. Elyasi and A. V osoughi, “Massive MIMO over correlated fading channels: Multi-cell MMSE processing, pilot assignment an d power control,” IEEE Trans. Wirel. Commun. , vol. 25, pp. 30–46, Jun. 2026

    2026

  12. [20]

    Interference burden in wireless communica- tions: A comprehensive survey from PHY layer perspective,

    A. Tusha and H. Arslan, “Interference burden in wireless communica- tions: A comprehensive survey from PHY layer perspective,” IEEE Com- mun. Surveys Tuts. , vol. 27, no. 4, pp. 2204–2246, Fourthquarter 2025

    2025

  13. [21]

    Applications of multi-agent reinforcement learning in future internet: A comprehensive survey,

    T. Li et al. , “Applications of multi-agent reinforcement learning in future internet: A comprehensive survey,” IEEE Commun. Surveys Tuts., vol. 24, no. 2, pp. 1240–1279, Secondquarter 2022

    2022

  14. [22]

    Emergent communication in multi-agent reinforcement learning for future wireless networks,

    M. Chafii et al., “Emergent communication in multi-agent reinforcement learning for future wireless networks,” IEEE Internet Things Mag. , vol. 6, no. 4, pp. 18–24, Dec. 2023

    2023

  15. [23]

    A collaborative m ulti-agent deep reinforcement learning-based wireless power allocation w ith centralized training and decentralized execution,

    A. Kopic, E. Perenda, and H. Gacanin, “A collaborative m ulti-agent deep reinforcement learning-based wireless power allocation w ith centralized training and decentralized execution,” IEEE Trans. Commun. , vol. 72, no. 11, pp. 7006–7016, Nov. 2024

    2024

  16. [24]

    Joint scheduling, O-RU association, and power allocation in O-RAN via model-based optimization and PPO-based reinforcement learning,

    A. E. Matemu, M. Kim, and K. Lee, “Joint scheduling, O-RU association, and power allocation in O-RAN via model-based optimization and PPO-based reinforcement learning,” IEEE Trans. Wirel. Commun., vol. 25, pp. 20102–20117, Jul. 2026

    2026

  17. [25]

    Federated dee p reinforcement learning for resource allocation in O-RAN sl icing,

    H. Zhang, H. Zhou, and M. Erol-Kantarci, “Federated dee p reinforcement learning for resource allocation in O-RAN sl icing,” in Proc. IEEE Glob. Commun. Conf. (GLOBECOM) , Rio de Janeiro, Brazil, 2022, pp. 958–963

    2022

  18. [26]

    Multi-agent deep reinforcement learning for dynamic power allocation in wireless networks,

    Y . S. Nasir and D. Guo, “Multi-agent deep reinforcement learning for dynamic power allocation in wireless networks,” IEEE J. Sel. Areas Commun., vol. 37, no. 10, pp. 2239–2250, Oct. 2019

    2019

  19. [27]

    Resource management in wireless networks via multi-agent deep reinforcement learning,

    N. Naderializadeh et al. , “Resource management in wireless networks via multi-agent deep reinforcement learning,” IEEE Trans. Wirel. Commun., vol. 20, no. 6, pp. 3507–3523, Jun. 2021

    2021

  20. [28]

    Decentralized and joint resource allocation, beamforming, and beamcombining for 5G networks with hetero geneous MARL,

    A. Al-Habashna et al. , “Decentralized and joint resource allocation, beamforming, and beamcombining for 5G networks with hetero geneous MARL,” IEEE Access , vol. 13, pp. 101491–101506, 2025

    2025

  21. [29]

    Proportional fair resource scheduling for dynamic beyond 5G networks: A distributed hierarchical DRL approac h,

    Y . Chen et al. , “Proportional fair resource scheduling for dynamic beyond 5G networks: A distributed hierarchical DRL approac h,” IEEE Trans. Mobile Comput. , vol. 25, no. 7, pp. 10893–10909, Jul. 2026

    2026

  22. [30]

    Joint beamforming and resource allocation for delay optimization in RIS-assisted OFDM systems: A DRL approach,

    Y . Ma et al. , “Joint beamforming and resource allocation for delay optimization in RIS-assisted OFDM systems: A DRL approach, ” arXiv preprint arXiv:2506.03586, 2025

    2025 arXiv

  23. [31]

    An introduction to centralized training for decentralized execution in cooperative multi-agent reinforcement learn ing,

    C. Amato, “An introduction to centralized training for decentralized execution in cooperative multi-agent reinforcement learn ing,” arXiv preprint arXiv:2409.03052, 2024

    2024 arXiv

  24. [32]

    Multi-agent reinforcement learning in wire less distributed networks for 6G,

    J. Zhang, “Multi-agent reinforcement learning in wire less distributed networks for 6G,” arXiv preprint arXiv:2502.05812, 2025

    2025 arXiv

  25. [33]

    An offline multi-agent reinforc ement learning framework for radio resource management,

    E. Eldeeb and H. Alves, “An offline multi-agent reinforc ement learning framework for radio resource management,” IEEE Trans. Mobile Comput., vol. 25, no. 1, pp. 1137–1150, Jan. 2026

    2026

  26. [34]

    Fully decentralized multi-agent reinforcement learnin g with networked agents,

    K. Zhang et al., “Fully decentralized multi-agent reinforcement learnin g with networked agents,” in Proc. 35th Int. Conf. Mach. Learn. (ICML) , Stockholm, Sweden, 2018, pp. 5872–5881

    2018

  27. [35]

    Sample and communication-efficient decentralized actor–critic algorithms with finite-time analysis,

    Z. Chen et al. , “Sample and communication-efficient decentralized actor–critic algorithms with finite-time analysis,” in Proc. 39th Int. Conf. Mach. Learn. (ICML) , Baltimore, MD, USA, 2022, pp. 3794–3834

    2022

  28. [36]

    Fully decentralized coopera tive multi-agent reinforcement learning: a survey,

    J. Jiang, K. Su, and Z. Lu, “Fully decentralized coopera tive multi-agent reinforcement learning: a survey,” arXiv preprint arXiv:2 401.04934, 2024

    2024

  29. [37]

    Learn-to-share: A decentralized multi-agent spectrum sharing framework for heterogeneous networks in the 6G era,

    Y . Tao, J.-C. He, Z.-J. Liu, and S. Y ang, “Learn-to-share: A decentralized multi-agent spectrum sharing framework for heterogeneous networks in the 6G era,” IEEE J. Sel. Areas Commun. , vol. 44, pp. 3490–3506, Jan. 2026

    2026

  30. [38]

    Communication-efficient learning of deep networks from decentralized data,

    H. B. McMahan et al. , “Communication-efficient learning of deep networks from decentralized data,” in Proc. 20th Int. Conf. Artif. Intell. Statist. (AISTATS), Fort Lauderdale, FL, USA, 2017, pp. 1273–1282

    2017

  31. [39]

    Federated optimization in heterogeneous networks,

    T. Li et al., “Federated optimization in heterogeneous networks,” in Proc. Mach. Learn. Sys. (MLSys), vol. 2, Austin, TX, USA, 2020, pp. 429–450

    2020

  32. [40]

    Federated reinforcement learning for energy- efficient D2D-IoT networks with AoI awareness,

    P . Parhizgar et al. , “Federated reinforcement learning for energy- efficient D2D-IoT networks with AoI awareness,” IEEE Open J. V eh. Technol., vol. 6, pp. 2828–2841, Sep. 2025

    2025

  33. [41]

    Meta federated reinforcement learning for distributed resource allocation,

    Z. Ji, Z. Qin, and X. Tao, “Meta federated reinforcement learning for distributed resource allocation,” IEEE Trans. Wirel. Commun. , vol. 23, no. 7, pp. 7865–7876, Jul. 2024

    2024

  34. [42]

    Communication-efficient soft actor–critic policy collaboration via regulated segment mixture,

    X. Y u et al. , “Communication-efficient soft actor–critic policy collaboration via regulated segment mixture,” IEEE Internet Things J. , vol. 12, no. 4, pp. 3929–3947, Feb. 2025

    2025

  35. [43]

    Pareto actor–critic for communication and computation co-optimization in non-cooperative federated learning se rvices,

    R. Tan et al. , “Pareto actor–critic for communication and computation co-optimization in non-cooperative federated learning se rvices,” IEEE Trans. Mobile Comput. , vol. 25, no. 2, pp. 1628–1643, Feb. 2026

    2026

  36. [44]

    Decentralized stochastic optimization and gossip algorithms with compressed commun ication,

    A. Koloskova, S. U. Stich, and M. Jaggi, “Decentralized stochastic optimization and gossip algorithms with compressed commun ication,” in Proc. 36th Int. Conf. Mach. Learn. (ICML) , Long Beach, CA, USA, 2019, pp. 3478–3487

    2019

  37. [45]

    Error feedback fixes SignSGD and other gradient compression schemes,

    S. P . Karimireddy et al. , “Error feedback fixes SignSGD and other gradient compression schemes,” in Proc. 36th Int. Conf. Mach. Learn. (ICML), Long Beach, CA, USA, 2019, pp. 3252–3261

    2019

  38. [46]

    FedCritic: Serve rless federated critic learning-based resource allocation for multi-cell OFDMA in 6G,

    A. Farajzadeh and M. Erol-Kantarci, “FedCritic: Serve rless federated critic learning-based resource allocation for multi-cell OFDMA in 6G,” arXiv preprint arXiv:2605.21418, 2026

    2026 arXiv

  39. [47]

    Partitioned edge learning over fast fading channels,

    Z. Jiang et al. , “Partitioned edge learning over fast fading channels,” IEEE Trans. V eh. Technol., vol. 74, no. 6, pp. 8561–8576, Jun. 2025

    2025

  40. [48]

    Asymptotic analysis of federated learning under event- triggered communication,

    X. He et al. , “Asymptotic analysis of federated learning under event- triggered communication,” IEEE Trans. Signal Process. , vol. 71, pp. 2654–2667, 2023

    2023

  41. [49]

    Robust federated learning for unreliable and resource- limited wireless networks,

    Z. Chen et al. , “Robust federated learning for unreliable and resource- limited wireless networks,” IEEE Trans. Wireless Commun. , vol. 23, no. 8, pp. 9793–9809, Aug. 2024

    2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.