Pith. sign in

REVIEW 3 major objections 4 minor 21 references

Fluid Democracy in Federated Data Aggregation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper introduces viscous-retained democracy, a delegation protocol for federated learning that provably never performs asymptotically worse than direct one-person-one-vote averaging (FedAvg) under the same assumptions that make…

desk verdict The viscous-retained variant is a genuinely sensible idea, but Theorem 2 is false as stated and the abstract's 'always better than 1p1v' claim does not survive the paper's own model. read the letter →

arxiv 2507.02710 v1 pith:343R5GDX submitted 2025-07-03 cs.LG

classification cs.LG
keywords federatedlearningfluiddemocracyviscousliquiddonoharmpropertyclientselectionadversarialrobustnessFedVRD
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

Federated learning typically asks every client to send its model weights to a central server every round, which wastes communication on clients whose data contributes little. This paper proposes to instead let clients delegate their 'vote' to a more competent neighbor for a few rounds, so the server only aggregates a subset of clients' weights. The central theoretical claim is that a new rule, viscous-retained democracy, never asymptotically underperforms direct one-person-one-vote voting (equated with FedAvg) under the same mild conditions that make liquid democracy safe, while still damping influence accumulation. The key is that each delegator keeps a fraction of their vote instead of losing it to decay, which prevents high-competence voters in long delegation chains from being drowned out by low-competence voters in star-shaped clusters. On top of this, the paper's FedVRD algorithm tunes a threshold to trade communication cost against resistance to adversarial clients.

What carries the argument

The central object is the vote-weight rule of viscous-retained democracy: a voter who delegates to a more competent neighbor passes a fraction α of their weight along each hop but retains the complementary fraction 1−α, so the total vote mass in the system stays constant even though guru weights are dampened by α per hop as in viscous democracy. The retained fraction is what carries the argument: it guarantees that every delegation strictly increases the competence of the weight that travels, with a minimum per-delegation competence increase of c_min · α^m (where m is the bounded longest path), which is exactly what Lemma 4.1's condition (2) needs to prove the Do No Harm property.

What would settle it

Run the Theorem 1 construction—n/2 chains of length 10 with competences in (0.65,0.7) and n/2 stars of size 10 with center competence 0.4 and leaves 0.39—using viscous-retained weights with α=0.5, and compute the weighted-mean competence after delegation; if it falls below the direct-voting mean of 0.5 for some finite n, the do-no-harm claim is false.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 2: in an upward-delegation process where each delegator passes a fraction α of their weight to a more competent neighbor and retains the complementary fraction 1−α, the resulting 'viscous-retained democracy' satisfies the do-no-harm property—its accuracy is asymptotically within ε of direct one-person-one-vote voting—provided no guru accumulates O(n) delegators and Assumption 4.2 holds (the delegation graph has at least C(n) ∈ O(n) edges and no two voters share a competence). This is contrasted with Theorem 1, which shows standard viscous democracy fails do-no-harm: in a network of length-10 chains of high-competence voters and size-10 stars of low-competence voters, the chain gurus' vote mass decays to at most 1/(1−α) while star gurus keep 1+9α, so with α=0.5 the weighted-mean competence falls from above 0.5 to at most 0.48 and the Condorcet Jury Theorem drives the collective decision to the wrong alternative. Viscous-retained democracy preserves the constant total vote mass and therefore a strict competence increase per delegation, which is exactly what Lemma 4.1 requires for do-no-harm.

Load-bearing premise

The whole guarantee rests on treating federated learning as a two-alternative binary-choice problem in which each client has a fixed probability of producing a 'correct' update, and clients can reliably identify a more competent neighbor from update-similarity scores; if that mapping does not hold, the Do No Harm theorem says nothing about whether the selected updates improve the global model versus FedAvg.

Editorial extensions

If this is right

  • Federated clients can delegate to a more competent neighbor via update similarity and the server can aggregate only the elected clients' weights, cutting per-round transfer cost without asymptotically sacrificing accuracy relative to FedAvg.
  • Viscous-retained democracy matches liquid democracy's do-no-harm guarantee while keeping guru weights identical to viscous democracy, so influence cannot accumulate in a few misaligned clients.
  • The threshold τ in FedVRD lets the server trade communication cost against adversarial robustness at each round; the paper's star-chain experiments show an optimal τ exists for a fixed budget.
  • Standard viscous democracy is unsuitable as an FL aggregation rule because its weight decay makes high-competence chains lose influence to low-competence stars, which is exactly the failure mode Theorem 1 constructs.

Reading between the lines

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

  • Because the retained-vote rule keeps total vote mass invariant, the do-no-harm proof could likely be recast as a martingale argument, where each delegation is a mean-increasing contraction; that reformulation might relax the bounded-delegator assumption.
  • The paper's adversarial analysis assumes the adversary can place agents optimally but only in star/chain topologies; a testable extension is to derive the minimal adversary budget for arbitrary graphs using the Laplacian spectrum of the delegation graph.
  • If the binary-competence assumption fails in practice, viscous-retained democracy may still be a better heuristic than viscous democracy, because it cannot suffer the pure weight-loss failure; comparing FedVRD against FedAvg and trimmed-mean on non-IID benchmarks would test this.
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 / 4 minor

Summary. This paper frames federated learning aggregation as an epistemic two-alternative social-choice problem in which each client has a binary competence qi and can delegate weight to more competent neighbors. It compares liquid democracy, viscous democracy, and a new 'viscous-retained democracy' against direct majority voting (identified with FedAvg), defines the Do No Harm (DNH) property, and claims that viscous-retained democracy satisfies DNH under the same assumptions as liquid democracy while preventing influence accumulation. It also proposes a threshold-based algorithm, FedVRD, and gives a small star/chain example to illustrate a cost-robustness trade-off. The central theoretical result, Theorem 2, is false as stated and the proof contains a false inference; the abstract's stronger claim that the protocol 'always does better than 1p1v' is not supported by the formal definitions.

Significance. The paper's high-level idea of using delegation rules as a client-selection mechanism for FL is interesting, and the viscous-retained rule is a natural variant that conserves total voting mass while bounding guru weight. The adversarial cost-budget discussion in Section 5 is also a useful concrete direction. However, the paper's main contribution is a DNH theorem, and that theorem is invalid under the stated hypotheses. With the theorem removed, the manuscript offers an algorithm with no convergence or robustness guarantees and a small illustrative experiment whose figure is absent from the text. I therefore do not see a publishable core in the current submission.

major comments (3)
  1. [Section 4.4 (Theorem 2)] The theorem is refuted by the following construction, which satisfies the stated hypotheses. Let n=m^2 and partition the voters into m disjoint chains of length m. In each chain, with voters indexed from the tail to the guru, set competence p_j = 1/2 + j/m^2 and let every non-guru delegate to the next voter; each guru then receives exactly m-1 delegators, which is o(n). Take α=1-1/m; the statement does not forbid α from depending on the graph size. For direct voting, the total mean excess over n/2 is Θ(m) and the standard deviation is Θ(m), so direct accuracy tends to Φ(1)≈0.841. Under viscous-retained democracy, the guru in each chain has weight (1-α^m)/(1-α)=Θ(m), each non-guru has weight at most 1, and total vote mass is n; the weighted sum therefore has mean excess Θ(m) and standard deviation Θ(m^{3/2}). The z-score tends to 0, so the delegated accuracy tends to 1/2. Hence gain(P_VRD,P_Direct) tends to -(Φ(1)-1/2)<0, violating DNH. The proof's claim that 'Since no guru has O(n) delegators, m is a constant' is false: this construction has max guru indegree m-1=o(n) and longest path length m=√n. If the authors intended α to be fixed independent of n, they must state that restriction explicitly and supply a different argument; as written, the theorem and proof are incorrect.
  2. [Abstract and Section 4.1] The abstract's claim that viscous-retained democracy 'always does better than 1p1v' is not a consequence of the Do No Harm property defined in Section 4.1. DNH only requires gain(M,Gn) ≥ -ε for all sufficiently large n, i.e., asymptotic no-harm, not strict improvement. Section 4.1 itself states this and then immediately draws the converse conclusion ('always performing better than direct democracy would imply performing better than FedAvg'); the implication is not established. Moreover, the formal results live entirely in the two-alternative epistemic model of Section 3.1, where 'correct' means drawing an update from data subset a+. No theorem connects binary vote correctness to the actual FedAvg objective or to the cosine-similarity heuristic in Algorithm 1, line 12, so the title's claim about improving FL aggregation is unsupported even if the social-choice theorem were correct.
  3. [Section 4.2 and Section 4.4 (proofs of Lemma 4.3 and Theorem 2)] Both proofs rely on a nonexistent uniform competence gap. The assumption that no two voters have equal competencies (Assumption 4.2) does not imply the existence of a constant cmin>0 that is independent of n; the minimum gap can tend to zero along a sequence of graphs. Similarly, in Theorem 2, from 'no guru has O(n) delegators' the proof concludes that the longest path length m is constant; even in a version where α is fixed, m can grow and α^m does not provide a uniform lower bound. Without a uniform positive lower bound on the per-delegation competence improvement, the claimed total competence increase of order n in Equation (2) is not established.
minor comments (4)
  1. [Section 5] The text refers to Figure 1 for the cost-robustness trade-off, but no figure appears in the manuscript, so the claimed plot cannot be checked.
  2. [Section 5] The threshold τ is introduced with τ ∈ (0,1), but the example uses τ=1 and calls it standard viscous democracy; the relationship between τ and the vote-share threshold is not formalized.
  3. [Section 4.1 and Assumption 4.2] The letter C(n) is used both for an o(n) bound on maximum weight in Lemma 4.1 and for an O(n) bound on the number of edges in Assumption 4.2, which is confusing.
  4. [Section 4.4] The proof of Theorem 2 writes α^m > 0 without treating the case α=0; the case of no delegation should be handled separately.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity; the derivation relies on the external Lemma 4.1 and the Condorcet Jury Theorem, with no self-citation chain or fitted-input prediction.

full rationale

The paper's central theoretical claims are proven from stated assumptions using an external lemma (Lemma 4.1, attributed to Berinsky et al. 2024) and the standard Condorcet Jury Theorem. No parameter is fitted to data and then reported as a prediction; no prior result of the same authors is invoked as a load-bearing input; and no equation is defined in terms of the conclusion it is used to derive. The proof of Theorem 2 attempts to establish the sufficient conditions of Lemma 4.1, and whether that proof is valid is a correctness question, not a circularity question. The abstract's phrase 'always does better than 1p1v' is a rhetorical inflation of the formal do-no-harm property, which only requires asymptotic non-underperformance; this is a framing issue rather than a circular reduction. The paper itself, in Section 6, states that empirical results on real-world datasets and convergence guarantees are future work, which is a scope limitation but not a circular dependency. Therefore no specific circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced; viscous-retained democracy is a mechanism variant, not a postulated entity. The free parameters are mechanism constants and counterexample values, not fitted parameters in the data-fitting sense.

free parameters (3)
  • Dampening factor α = 0.5 in examples
    Viscosity parameter of the delegation mechanism, set by hand in the paper's examples; not fitted to data.
  • Participation threshold τ = unspecified, varies in example
    Tunable threshold in FedVRD that controls which clients transmit; the paper states the server computes an optimal τ but does not specify how.
  • Counterexample competencies in Theorem 1 = α=0.5, q=0.39, a=0.4, b=0.65, c=0.7
    Chosen by hand to construct a topology where viscous democracy's weighted mean competence drops below 0.5.
assumptions (6)
  • domain assumption Upward delegation: each voter delegates only to a strictly more competent neighbor.
    Used in Lemma 4.3 and Theorems 1 and 2 to ensure acyclic paths and competence improvement; introduced in Section 3.2.
  • domain assumption No two voters have equal competence.
    Assumption 4.2, needed for the cmin > 0 argument in the proofs.
  • domain assumption The delegation graph has at least C(n) edges with C(n) ∈ O(n).
    Assumption 4.2, though the notation is ambiguous (a linear lower bound is needed for Equation (2) but a sublinear bound may not suffice).
  • domain assumption Existence of an objectively correct alternative a+ and per-client competence qi.
    Core epistemic social-choice setup in Section 3.1; the FL mapping treats data relevance as a fixed probability of correct voting.
  • standard math Lemma 4.1 from Berinsky et al. gives sufficient conditions for DNH.
    Cited without proof; used as the bridge to accuracy in Theorems 1 and 2.
  • standard math Condorcet Jury Theorem: majority accuracy tends to 1 if mean competence exceeds 0.5 and to 0 if below.
    Used in Theorem 1 to argue convergence to the wrong outcome for viscous democracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fluid Democracy in Federated Data Aggregation." pith.science (2026). https://pith.science/paper/343R5GDX

@misc{pith2026250702710,
  author       = {Pith},
  title        = {Pith review of: Fluid Democracy in Federated Data Aggregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/343R5GDX}},
  note         = {Machine review of arXiv:2507.02710}
}
read the original abstract

Federated learning (FL) mechanisms typically require each client to transfer their weights to a central server, irrespective of how useful they are. In order to avoid wasteful data transfer costs from clients to the central server, we propose the use of consensus based protocols to identify a subset of clients with most useful model weights at each data transfer step. First, we explore the application of existing fluid democracy protocols to FL from a performance standpoint, comparing them with traditional one-person-one-vote (also known as 1p1v or FedAvg). We propose a new fluid democracy protocol named viscous-retained democracy that always does better than 1p1v under the same assumptions as existing fluid democracy protocols while also not allowing for influence accumulation. Secondly, we identify weaknesses of fluid democracy protocols from an adversarial lens in terms of their dependence on topology and/ or number of adversaries required to negatively impact the global model weights. To this effect, we propose an algorithm (FedVRD) that dynamically limits the effect of adversaries while minimizing cost by leveraging the delegation topology.

Figures

Figures reproduced from arXiv: 2507.02710 by the authors.

Figure 1
Figure 1. Plot depicting the optimal τ for adversarial robustness at a fixed total data transfer cost From [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 16 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Optimizing viscous democracy

    Armstrong, B., Alouf-Heffetz, S., and Talmon, N. Optimizing viscous democracy. In Larson, K. (ed.), Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24 , pp.\ 2643--2650. International Joint Conferences on Artificial Intelligence Organization, 8 2024. doi:10.24963/ijcai.2024/292. URL https://doi.org/10.24963...

  3. [3]

    Tracking Truth with Liquid Democracy

    Berinsky, A., Halpern, D., Halpern, J. Y., Jadbabaie, A., Mossel, E., Procaccia, A. D., and Revel, M. Tracking truth with liquid democracy, 2024. URL https://arxiv.org/abs/2107.11868

  4. [4]

    Bersetche, F. M. Generalizing liquid democracy to multi-agent delegation: A voting power measure and equilibrium analysis, 2024. URL https://arxiv.org/abs/2209.14128

  5. [5]

    M., Guerraoui, R., and Stainer, J

    Blanchard, P., El Mhamdi, E. M., Guerraoui, R., and Stainer, J. Machine learning with adversaries: Byzantine tolerant gradient descent. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (eds.), Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL https://proceed...

  6. [6]

    Viscous democracy for social networks

    Boldi, P., Bonchi, F., Castillo, C., and Vigna, S. Viscous democracy for social networks. Commun. ACM, 54 0 (6): 0 129–137, June 2011. ISSN 0001-0782. doi:10.1145/1953122.1953154. URL https://doi.org/10.1145/1953122.1953154

  7. [7]

    Cao, X., Fang, M., Liu, J., and Gong, N. Z. Fltrust: Byzantine-robust federated learning via trust bootstrapping. CoRR, abs/2012.13995, 2020. URL https://arxiv.org/abs/2012.13995

  8. [8]

    DRACO : B yzantine-resilient distributed training via redundant gradients

    Chen, L., Wang, H., Charles, Z., and Papailiopoulos, D. DRACO : B yzantine-resilient distributed training via redundant gradients. In Dy, J. and Krause, A. (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp.\ 903--912. PMLR, 10--15 Jul 2018. URL https://proceedings.mlr.pr...

Show all 21 references
  1. [9]

    G\" O lz, P., Kahng, A., Mackenzie, S., and Procaccia, A. D. The fluid mechanics of liquid democracy. ACM Trans. Econ. Comput., 9 0 (4), October 2021. ISSN 2167-8375. doi:10.1145/3485012. URL https://doi.org/10.1145/3485012

  2. [10]

    Liquid democracy: An algorithmic perspective

    Kahng, A., Mackenzie, S., and Procaccia, A. Liquid democracy: An algorithmic perspective. Proceedings of the AAAI Conference on Artificial Intelligence, 32 0 (1), Apr. 2018. doi:10.1609/aaai.v32i1.11468. URL https://ojs.aaai.org/index.php/AAAI/article/view/11468

  3. [11]

    and Ramamoorthy, A

    Konstantinidis, K. and Ramamoorthy, A. Byzshield: An efficient and robust system for distributed training. In Smola, A., Dimakis, A., and Stoica, I. (eds.), Proceedings of Machine Learning and Systems, volume 3, pp.\ 812--828, 2021. URL https://proceedings.mlsys.org/paper_file...

  4. [12]

    An experimental study of byzantine-robust aggregation schemes in federated learning

    Li, S., Wu, Z., Song, W., Ma, X., Khan, N., Chen, K., and Song, X. An experimental study of byzantine-robust aggregation schemes in federated learning. IEEE Transactions on Big Data, 2023

  5. [13]

    Fedvoting: A cross-silo boosting tree construction method for privacy-preserving long-term human mobility prediction

    Liu, Y., Fan, Z., Song, X., and Shibasaki, R. Fedvoting: A cross-silo boosting tree construction method for privacy-preserving long-term human mobility prediction. Sensors, 21 0 (24), 2021. ISSN 1424-8220. doi:10.3390/s21248282. URL https://www.mdpi.com/1424-8220/21/24/8282

  6. [14]

    May, K. O. A set of independent necessary and sufficient conditions for simple majority decision. Econometrica, 20 0 (4): 0 680--684, 1952. URL https://onlinelibrary.wiley.com/doi/abs/0012-9682(195210)20:4&lt;680:ASOINA&gt;2.0.CO;2-S

  7. [15]

    McMahan, B., Moore, E., Ramage, D., Hampson, S., and Arcas, B. A. y. Communication-Efficient Learning of Deep Networks from Decentralized Data . In Singh, A. and Zhu, J. (eds.), Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, volume ...

  8. [16]

    M., and Harchaoui, Z

    Pillutla, K., Kakade, S. M., and Harchaoui, Z. Robust aggregation for federated learning. IEEE Transactions on Signal Processing, 70: 0 1142--1154, 2022. doi:10.1109/TSP.2022.3153135

  9. [17]

    Detox: A redundancy-based framework for faster and more robust gradient aggregation

    Rajput, S., Wang, H., Charles, Z., and Papailiopoulos, D. Detox: A redundancy-based framework for faster and more robust gradient aggregation. In Wallach, H., Larochelle, H., Beygelzimer, A., d Alch\' e -Buc, F., Fox, E., and Garnett, R. (eds.), Advances in Neural Information ...

  10. [18]

    Election coding for distributed learning: Protecting signsgd against byzantine attacks

    Sohn, J.-y., Han, D.-J., Choi, B., and Moon, J. Election coding for distributed learning: Protecting signsgd against byzantine attacks. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp...

  11. [19]

    Zeno: Robust and communication-efficient federated learning with stragglers

    Xie, C., Liu, M., Gong, Z., Li, L., Li, B., and Liang, P. Zeno: Robust and communication-efficient federated learning with stragglers. In International Conference on Machine Learning, pp.\ 10617--10628, 2020

  12. [20]

    B yzantine-robust distributed learning: Towards optimal statistical rates

    Yin, D., Chen, Y., Kannan, R., and Bartlett, P. B yzantine-robust distributed learning: Towards optimal statistical rates. In Dy, J. and Krause, A. (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Resear...

  13. [21]

    Fedvote: A voting-based federated learning aggregation method

    Yue, H., Li, X., Jiang, X., Zhang, X., and Yu, X. Fedvote: A voting-based federated learning aggregation method. In Proceedings of the 2022 IEEE International Conference on Big Data (BigData), pp.\ 2432--2441. IEEE, 2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.