acceptodds
Under review as a conference paper at ICLR 2027

Decentralized Personalized Federated Learning with Compressed Messages via Finite-Time Quantized Coordination

Abstract

Decentralized Federated Learning (DFL) enables privacy-preserving collaborative training without a central server. However, it faces two major challenges: statistical data heterogeneity, which degrades global model performance on local tasks, and limited communication bandwidth, which hinders collaboration and degrades model accuracy. To address these issues, we propose Q-pFedMe, a decentralized personalized FL algorithm with quantized communication. Q-pFedMe decouples local personalized model optimization from global reference learning using Moreau envelopes. Furthermore, it embeds fixed-point quantization into finite-time averaging, making all nodes reach exact agreement on a common quantized aggregate. We theoretically prove that Q-pFedMe converges linearly to a quantization-dependent neighborhood of the optimal solution for strongly convex objectives, and has an optimization term with an explicit error floor for smooth non-convex objectives. We also derive a finite-time upper bound on the number of communication steps required by the quantized average consensus subroutine at each global round. Experiments on Synthetic, MNIST, and CIFAR-10 show that Q-pFedMe achieves accuracy comparable to centralized unquantized pFedMe. Additional CIFAR-100 experiments on two directed topologies demonstrate the highest personalized accuracy and lowest per-transmission payload among the evaluated decentralized baselines. Our code is available at https://anonymous.4open.science/r/N8HgpH8UoWHorEt/.

Primary area:
optimization
Keywords:
Decentralized Federated Learning, Personalized Federated Learning, Quantized Communication, Finite-Time Consensus, Directed Networks

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_sLi3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a new method for personalized decentralized federated learning under quantization. The authors propose a new method that combines existing methods on quantization and consensus and obtain a set of convergence analysis.

Critical strengths
  1. The paper targets a valid problem (i.e., quantized decentralized FL).
  2. Overall the paper is easy to read, although there is room for improvement (I will explain in my comments below).
Critical weaknesses

The paper suffers from a verity of issues that deserve careful considerations:

  1. The paper claims tries to sell the convergence analysis more ambitious as what it really is. For example, in the abstract, it is mentioned that "We theoretically prove that Q-pFedMe converges linearly to a quantization-dependent neighborhood of the optimal solution for strongly convex objectives, and has an O(1/R) optimization term with an explicit error floor for smooth non-convex objectives." This claims, especially, the part for strongly convex case, is highly standard in decentralized FL literature. In particular, any decentralized FL method with inexact local solver or consensus (caused by quantization noise in this paper) will exhibit the same convergence behavior. Thus, it is not clear what is the main advantage of the proposed method, or what is the uniqueness of the analysis (in terms of the tightness of the bounds, etc.).

  2. Line 77: It is not clear what does FP32 mean.

  3. Line 111: The set notation (i.e., { }) is missed in both definitions of N^_ and N^+.

  4. Eq. (7), as L approaches the value of \lambda, the value of the bound reaches infinity. This is questionable as the smoothness value (especially its smaller values) should not cause such extreme conditions.

  5. The same issue arises in line 278: The denominator \lambda^2-8L^2, can be 0 or even negative when the only condition applied is \lambda> L in (7). This issue brings the entire results under question as the value of the fraction is presumed to be strictly positive in the convergence analysis (the left hand side is denoted by \sigma^2 to denote its positive nature).

  6. In line 329, q_i^r is not clearly defined.

  7. The bounds in Theorem 3 are questionable. Specifically, they are "upper bounds" on T. That means that for any value of T holding in those bounds, the claim of the Theorem, which is "reach exact agreement on the quantized average" should hold. Yet, T=0, 1, .... all hold in the bound. That means that the algorithm reaches perfect agreement even for T=0, 1, .... which does not seem to be correct.

  8. The experimental section seems to be quite weak: 8-1) A variety of baselines on decentralized quantized FL, decentralized FL with sparsification, and event-triggered decentralized FL are missed. 8-2) The comparison to centralized FL can be moved to appendix to have more space for comparison to other decentralized FL methods. 8-3) The superiority of the algorithm should be compared with all the above methods under a variety of graph topologies (e.g., Scale-Free Graphs, Caveman Graphs, etc.) with varying connectivity level for the underlying graph topology. 8-4) The experiments above should be conducted under a wider level of data heterogeneity ranges (Dirichelet parameter ranging from 10^-3 to 10^3). This will show whether the advantage of the method is under extreme data heterogeneity or under homogeneous data across clients.

Questions for authors

Please refer to the comments above. In particular,

  1. How the claims of the paper (mentioned in the first weakness comment) can be justified?

  2. How the extreme values of conditions under varying values of L can be justified?

  3. How to avoid having negative values of the condition mentioned in line 278 and how the proof changes under negative values?

  4. How to justify the small values of T that hold in Theorem 3?

  5. What is the reason behind limited simulations?

Reviewer_ZFKxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Q-pFedMe, which combines Moreau-envelope personalization with finite-time quantized averaging on a strongly connected directed network. It claims optimization guarantees up to quantization-dependent error floors and a problem-dependent bound on consensus time. Experiments compare accuracy with centralized pFedMe and evaluate personalized accuracy, per-message payload, time, and cumulative communication against decentralized baselines.

Critical strengths

The combination addresses a practical problem: preserving full-model personalization while exchanging finite-precision messages without a central server. Separating agreement on a common quantized aggregate from recovery of the real-valued average is appropriate, and the manuscript acknowledges that reduced per-message payload need not imply reduced cumulative traffic.

The CIFAR-100 study includes two directed topologies, three-seed variability, and worst-decile client accuracy. The reported personalized-accuracy improvements are substantial and potentially useful. These results motivate taking the method seriously, although the communication and theory issues below prevent accepting the present evidence as sufficient.

Critical weaknesses

The published protocol and communication results are inconsistent in ways that affect a central claim. Algorithm 2 can terminate only when the communication step is a multiple of D. Appendix C.1 gives D=99 for the directed ring, but Table 3 reports an average of only 10 consensus steps. Moreover, Algorithm 2, lines 18–20, propagates the max/min vectors at every step, whereas Appendix B.2 charges for only T/D such transmissions. Under the displayed pseudocode, the monitoring term is 2nQ|E|T, rather than 2nQ|E|T/D. Thus the ring’s lowest C@50 in Table 2 cannot currently be reconciled with the algorithm. An alternative implementation or unit definition might explain the discrepancy, but it must be specified and the measurements recomputed consistently.

Two advertised guarantees also require correction or a materially clearer argument. Theorem 3 presents an upper bound on random consensus time using a constant dependent on an unspecified convergence probability; its proof substitutes an expected initialization-dispersion quantity without stating whether the result is in expectation or holds with a specified probability. The step from (50) to (51) additionally drops componentwise absolute values before summing; the signed deviations sum to zero, so the displayed inequality is not valid in general. This local step may be repairable, but the guarantee needs to be restated and proved. Separately, Theorem 1 imposes a positive lower bound on the effective step size, while its proposed optimum tends to zero as quantization is refined. The optimization behind the claimed Delta^(4/3) error scaling must respect that lower bound or remove it with a valid proof. These issues concern the main theoretical contribution.

Questions for authors

What exactly does a consensus “step” count in Table 3, and how can the ring terminate with an average of 10 when D=99? Can the authors provide per-round executed-step and transmission counts, including all max/min propagation, and corrected C@50 and timing results consistent with Algorithm 2?

Is Theorem 3 an expectation bound or a high-probability bound, and what is its explicit probability parameter? Can the authors repair (50)–(51) and show the admissible step-size minimization for Theorem 1, including the regime in which the unconstrained optimum falls below its stated lower bound?

Reviewer_RBxJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates decentralized personalized federated learning under data heterogeneity and communication constraints. The authors propose Q-pFedMe, which combines Moreau-envelope-based personalization with finite-time quantized coordination. The framework separates local personalized model optimization from global reference learning and employs DQuFTA to achieve finite-time agreement on a common quantized aggregate over strongly connected directed networks. The authors establish linear convergence to a quantization-dependent neighborhood for strongly convex objectives and an O(1/R) optimization term with an explicit error floor for smooth non-convex objectives. They also derive an upper bound on the communication steps required for quantized consensus. Experiments on Synthetic, MNIST, CIFAR-10, and CIFAR-100 evaluate predictive accuracy, quantization effects, and communication efficiency. Results demonstrate accuracy comparable to centralized unquantized pFedMe and improvements over several decentralized baselines, alongside reduced per-transmission communication payload.

Critical strengths

1. Systematic Theoretical Analysis of Quantized Personalized Decentralized Learning.

The paper provides an extensive theoretical treatment of personalized decentralized learning under quantized communication. Theorem 1 establishes linear convergence to a quantization-dependent neighborhood for strongly convex objectives, while Theorem 2 derives an O(1/R) optimization term with an explicit error floor for smooth non-convex objectives. The analysis explicitly incorporates quantization precision, local optimization steps, data heterogeneity, and stochastic gradient errors into the convergence bounds. Furthermore, Theorem 3 characterizes the finite-time communication steps required for quantized consensus in relation to network properties and optimization states. The detailed derivations in Appendix D demonstrate considerable theoretical effort and provide a structured characterization of how quantization affects convergence and coordination complexity.

2. Practically Relevant Problem Setting and Encouraging Personalized Learning Results.

The paper considers a meaningful problem involving personalized learning over decentralized networks with heterogeneous data and finite-precision communication. The proposed framework combines personalized optimization with quantized coordination over strongly connected directed graphs, providing a clear approach to addressing these constraints within a unified learning procedure. The empirical results are encouraging in terms of personalized model accuracy. Table 1 shows performance close to centralized unquantized pFedMe across several benchmark tasks, while Table 2 reports noticeable improvements over representative decentralized baselines on CIFAR-100, including both mean personalized accuracy and the accuracy of the worst-performing 10% of clients. These findings provide positive empirical evidence for the feasibility of decentralized personalized learning with quantized model exchanges.

Critical weaknesses

1. Limited Methodological Novelty and Unclear Theoretical Advances.

Although the paper presents an extensive theoretical analysis, the independent methodological contribution of Q-pFedMe appears relatively limited. Its framework largely builds upon existing personalized optimization and quantized coordination techniques, raising concerns about whether their integration constitutes a sufficiently substantial advance.

  • Incremental Algorithmic Design. The personalized optimization objective and approximate proximal updates are largely inherited from pFedMe (T Dinh et al., 2020), while the finite-time quantized averaging mechanism in DQuFTA builds upon Rikos et al. (2022), as acknowledged in Appendix B.1. Consequently, the proposed framework appears to integrate two established components rather than introduce a fundamentally new personalization or quantization mechanism. The paper does not sufficiently clarify which limitations prevent these existing approaches from being combined directly, or which new algorithmic challenges are resolved by their integration. This makes the independent methodological contribution difficult to distinguish from prior work.

  • Questionable Theoretical Distinctiveness and Applicability. Although Theorems 1 and 2 establish convergence guarantees under quantized communication, their derivations closely follow the original pFedMe analysis, with finite-time consensus introducing a bounded quantization error into the global update (Eqs. (27)–(30), Appendix D.4). Since the consensus mechanism eliminates inter-client reference-model disagreement, the remaining optimization analysis largely relies on established techniques. Similarly, the additional analytical contribution of Theorem 3 is not sufficiently distinguished from existing finite-time quantized consensus results. Moreover, the applicability of the theoretical guarantees requires further clarification. In particular, Theorem 1 imposes both lower and restrictive upper bounds on the effective step size without establishing a nonempty feasible parameter regime. For instance, setting LF=μF=1L_F=\mu_F=1 and β=1\beta=1 yields incompatible bounds of 2/32/3 and 1/7741/774. Theorem 3 also leaves the probabilistic interpretation of its communication-step upper bound insufficiently explicit given the randomized coordination protocol. Together, these concerns make it difficult to assess the independent theoretical significance and practical scope of the reported guarantees.

2. Insufficient Experimental Validation of the Core Design and Communication Benefits.

Despite encouraging personalized accuracy results, the empirical evaluation does not sufficiently isolate the contribution of the proposed integration or establish its overall communication advantages. This is particularly important because efficient quantized coordination is one of the central motivations of the paper.

  • Missing Component Ablations and Comparisons with Simpler Alternatives. The experiments primarily compare Q-pFedMe with existing centralized and decentralized learning methods, together with sensitivity analyses of the quantization level and global update parameter. However, they do not adequately isolate the benefits of finite-time quantized coordination from those of the underlying pFedMe personalization framework. For example, comparisons against decentralized pFedMe with conventional gossip averaging, standard quantized gossip, or alternative quantized coordination protocols would help determine whether DQuFTA is essential to the reported improvements. Without such controlled comparisons, it remains unclear how much of the observed performance comes from the proposed communication mechanism rather than the established personalization objective.

  • Unconvincing End-to-End Communication Efficiency. Tables 4 and 5 demonstrate reductions in per-transmission payload, but the repeated coordination and termination-detection exchanges introduce substantial cumulative communication costs. In Table 2, Q-pFedMe requires 39.86 TB to reach the target accuracy on the directed exponential topology, compared with 10.29 TB for D-PSGD, despite requiring fewer global rounds. Similarly, Table 6 shows substantially higher cumulative communication than centralized pFedMe at the same target accuracy. Although the authors acknowledge that Q-pFedMe primarily targets per-transmission payload reduction, these results substantially limit the practical interpretation of its communication efficiency. A more systematic characterization of when quantized finite-time coordination provides a favorable overall communication trade-off is needed.

  • Insufficient Evaluation Across Network Scales and Communication Conditions. The decentralized benchmark comparisons in Table 2 focus on 100 nodes under two static directed topologies. However, the coordination protocol's communication complexity depends on network size, topology, diameter, and quantization precision, as reflected in Theorem 3 and Appendix B. Additional experiments varying the number of nodes, graph density, and network conditions would help assess the scalability and practical robustness of the approach. Moreover, Table 2 reports identical accuracy, convergence rounds, and wall-clock time for Q-pFedMe across both topologies despite different consensus-step requirements (Table 3, Appendix C.7). Clarifying the timing methodology and how communication latency is incorporated would strengthen the credibility of the efficiency evaluation.

Questions for authors
  1. What are the fundamental technical challenges in integrating pFedMe with existing finite-time quantized consensus protocols? Since the personalization objective and coordination mechanism largely originate from prior work, how does Q-pFedMe advance beyond a direct combination of these established components?

  2. Theorems 1–3 provide extensive convergence and communication analyses, but their novel theoretical contributions remain unclear relative to prior results. Which specific analytical challenges cannot be addressed directly using existing pFedMe and quantized consensus theory? Additionally, how are the lower and upper step-size constraints in Theorem 1 simultaneously satisfied, and what probabilistic guarantee does the communication-step bound in Theorem 3 provide?

  3. How much of Q-pFedMe's performance improvement can be attributed specifically to DQuFTA rather than the underlying pFedMe personalization framework? Could the authors provide controlled comparisons with decentralized pFedMe using conventional gossip averaging, standard quantized gossip, or alternative coordination protocols under comparable training and communication budgets?

  4. Although Q-pFedMe reduces per-transmission payload, Table 2 shows substantially higher cumulative communication than D-PSGD on the directed exponential topology. Under what network and communication conditions does the proposed method provide a genuine end-to-end efficiency advantage? A more detailed comparison of cumulative traffic, consensus overhead, and communication latency would help clarify this trade-off.

  5. How does Q-pFedMe scale with increasing client numbers, different graph densities, and communication constraints? In particular, Table 2 reports identical wall-clock times for Q-pFedMe across two topologies despite different average consensus steps in Table 3. What timing methodology was used, and how were topology-dependent communication and synchronization costs incorporated?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.