acceptodds
Under review as a conference paper at ICLR 2027

Equivalent on Accuracy, Ordered in Fidelity: Aggregation under Partial Reception in Decentralized Federated Learning

Abstract

In decentralized federated learning over unreliable links without retransmission, a receiver obtains only a subset of each neighbor's model coordinates with reception probability , which are unobservable in advance and channel-imposed rather than designed or selected. We show that final accuracy, the standard comparison metric, fails to distinguish aggregation rules under standard annealing schedules. We propose a direct measure of aggregation fidelity, the distance between the aggregate formed under coordinate loss and its lossless counterpart, and use it to give a two-part characterization. First, fidelity separates aggregators into two families. Convex per-coordinate weights cannot amplify disagreement in any realization, whereas naive loss compensation incurs variance that is coupled to parameter magnitude rather than to the step size. This property is path-wise, so learning-rate schedules cannot change family membership, as confirmed under constant and cosine schedules. Second, within this stable family the difference between rules persists, but their visibility depends on the schedule. Final accuracy separates aggregation rules at a constant rate, but converges, or even inverts at cosine annealing. Fidelity, consensus, and convergence cost nevertheless preserve their ordering, with the fidelity gap widening monotonically from to against the strongest baseline as availability falls. Benchmarking on final accuracy under a single schedule therefore identifies a rule’s family but never the ordering within it.

Primary area:
optimization
Keywords:
decentralized learning, federated learning, partial model reception, unreliable communication, evaluation methodology, consensus

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

The paper studies aggregation under partial model reception in decentralized federated learning. It compares a realized-coordinate average, MFG, with imputation and loss-compensation rules, analyzing per-step convexity and inverse-probability variance. Experiments on 20 clients examine how constant versus cosine learning-rate schedules affect accuracy, aggregation fidelity, consensus, and rounds-to-target. The main datasets are CIFAR-10 and EMNIST, with a CIFAR-100/ResNet-18 extension. The central claim is that annealing can conceal aggregation differences in final accuracy.

Critical strengths
  1. The evaluation question is useful: final accuracy can conceal differences in training trajectories and the cost of reaching a target. Comparing two schedules with fidelity, consensus, and rounds-to-target provides a more informative assessment than endpoint accuracy alone.

  2. Proposition 5.3 correctly establishes coordinate-diameter non-expansion for convex aggregation in each realization. Together with the explicit expected mixing coefficient in Proposition 5.4, this explains the rule's self-retention behavior as reception decreases.

  3. The reported design pairs data partitions, initializations, and reception masks across methods, and uses GroupNorm to avoid averaging running statistics. Learning-rate and topology ablations, plus the larger ResNet-18 model, address several plausible confounds within the tested setting.

Critical weaknesses
  1. Dataset and network scale are limited. Evaluation uses relatively small, low-resolution benchmarks: CIFAR-10/100 (32x32) and EMNIST (28x28), all with only 20 clients (Appendix C.1). The CIFAR-100/ResNet-18 extension increases model size but provides limited evidence beyond small-image tasks. Please add a larger-scale benchmark such as ImageNet-1K with higher-resolution inputs to test whether the accuracy-fidelity separation and convergence-cost ordering persist on a more demanding task. Also extend the client count to at least 50, preferably 100, with comparable average degree and per-client work. Applicability to larger datasets and decentralized networks therefore remains untested.

  2. The accuracy-equivalence claim is too broad. Under cosine annealing at 90% loss, Table 1 reports EMNIST accuracy of 83.9+/-0.1 for MFG versus 81.9+/-0.3 for Soft-DSGD, so final accuracy still separates these methods. Conversely, overlapping standard deviations on CIFAR-10 do not establish equivalence. With only three seeds and no stated equivalence margin or paired confidence intervals, the headline conclusion needs narrower scope and a suitable statistical analysis.

  3. Fidelity comparisons mix operator effects with reference and trajectory differences. Section 7.2 measures distance to an unweighted neighbor mean, while the lossless D-PSGD reference uses Metropolis weights (Appendix C.5). These targets differ on irregular graphs. Raw Q(t) also depends on each method's current disagreement, so comparisons along separate training trajectories do not isolate the aggregation operator. A common-state replay and errors relative to each rule's own lossless operator would clarify whether the reported ratios reflect loss robustness or nominal mixing and trajectory differences.

  4. Corollary 4.3 does not establish collapse during training. Appendix A.3 iterates the aggregation-only contraction while omitting ongoing SGD. For identical scalar objectives f(x)=(x-a)^2/2 on a regular graph, one SGD step followed by ZeroFill gives the expected-iterate recursion z_next=rho*((1-eta)z+etaa). For constant 0<eta<1 and 0<rho<1, its fixed point is rhoetaa/[1-rho*(1-eta)], which is nonzero for a nonzero a. Thus the stated geometric decay to zero and consequent chance accuracy require additional conditions; the observed failures alone do not prove this general claim.

  5. The released implementation has a reproducibility blocker. In the Zenodo archive, node.py calls the aggregator with four positional arguments, but the registered DFL-AA implementation, IPWGossip.aggregate in aggregators.py, accepts only local and received. Static inspection therefore identifies an interface mismatch in the provided experiment path. Please release the runnable version and exact configurations that generated the reported DFL-AA results; these comparisons are central to the claimed within-family ordering.

Questions for authors
  1. Do the main rankings persist with 50 and 100 clients at comparable average degree and per-client work? Can one larger-scale setting also test heterogeneous or temporally correlated reception?

  2. Which dataset/loss settings support accuracy equivalence under a prespecified practical margin, and how do you reconcile that claim with the EMNIST differences in Table 1?

  3. Does the fidelity ordering persist when all aggregators are evaluated on identical model states and masks, using both a common target and their respective lossless operators?

  4. Can Corollary 4.3 be restated for aggregation alone, or extended to include gradient forcing and identify conditions under which training actually collapses?

  5. Which runnable source version generated the DFL-AA results, and how is the aggregator-call signature mismatch resolved?

2 comments

Thank you for a very thorough review, including checking our code and working out a counterexample. Each point led to a concrete change, shown in blue in the revised PDF.

Corollary 4.3 (W4, Q4). You are right. The old proof (now App. B.3) iterated the aggregation step alone and left out the gradient steps. The corrected corollary includes them and says when collapse actually happens. (a) For f(x)=12(x−a)2f(x)=\frac12(x-a)^2 with kk gradient steps per round, the expected iterate settles at s as\,a with s=ρ(1−β)/(1−ρβ)<1s=\rho(1-\beta)/(1-\rho\beta)<1 and β=(1−η)k\beta=(1-\eta)^k. With k=1k=1 this is exactly your fixed point. So at a constant step size, zero-fill shrinks a single-layer model rather than wiping it out, and s→0s\to0 as the step size goes to zero. (b) For a product of LL weights, f(w)=12(∏lwl−a)2f(\mathbf w)=\frac12(\prod_l w_l-a)^2, the all-zero model is a fixed point. With two layers it becomes stable once loss passes a threshold, and with three or more layers it is stable at every loss rate, because the gradient that would restore one layer is proportional to the other layers' weights. In our setting, (a) gives s=0.50s=0.50 at 10% loss and 0.030.03 at 90%, the two-layer threshold is 10.6% loss, and a simulation matches these values to three decimals. We state clearly that local stability alone does not prove training falls into this state. That it does, at every loss level we tested, is shown in Table 1.

Fidelity comparisons (W3, Q3). We agree with both concerns and ran the experiment you suggested (new App. D.3). We saved node states from two lossless D-PSGD runs, one with a constant and one with a cosine step size, so the states belong to none of the compared rules. Every rule then received the same states and the same loss patterns, and we measured its error both against the plain neighborhood average and against its own output with nothing lost. MFG has the lowest error of every rule at every loss level, measured either way, so its lead comes from the rule itself and not from its training path. DFL-AA's output with nothing lost equals the plain average, so that comparison does not depend on the reference, and MFG's error is 1.4 to 2.4 times lower. The larger gaps in Section 7.2 also include the extra disagreement DFL-AA builds up during training. Your concern about the reference is correct for Soft-DSGD. Against the plain average it looks 1.4 to 4.5 times worse than MFG, but against its own weighted output only 1.1 to 2.1 times, which puts it closer to MFG than DFL-AA. We now claim only MFG's lead at this level. The cosine replay gives the same ratios for the stable rules, while the unnormalized rules' errors become 10710^7 to 10810^8 times MFG's, because annealing brings nodes together but does not shrink the parameters.

Accuracy equivalence (W2, Q2). We agree, and the abstract, Section 7.2 and Section 8 no longer claim equivalence. Your EMNIST example is correct. App. D.4 now gives, for every cosine cell, a 95% interval for the accuracy difference and an equivalence test with a ±1 point margin chosen in advance. On EMNIST, MFG is within ±1 point of both baselines at 10 to 20% loss and significantly better at 50 to 90% loss. On CIFAR-10 every cell is inconclusive, since three seeds can show neither a difference nor equivalence, and we now say exactly that.

Code (W5, Q5). You are right. While cleaning the code for release, we removed two unused arguments from the DFL-AA aggregator, which broke the call in node.py. The experiments ran with the original version, and DFL-AA does not use those arguments, so the results are unaffected. The new archive (https://zenodo.org/records/24299491) restores the original signature.

Scale (W1, Q1). We agree this is the main limitation of the current evaluation, and Section 8 now says so. Runs with 50 and 100 nodes and on Tiny-ImageNet are planned for the extended version.

Thank you again. Your corrections made the paper more accurate. We are happy to answer any further questions.

Thank you for the detailed response and revisions. The common-state replay clarifies the fidelity comparison, and the revised Corollary 4.3 addresses my concern about ZeroFill. I am maintaining my original rating because the empirical evaluation remains limited to 20-client networks with homogeneous, independent packet loss. It therefore remains unclear whether the rankings in aggregation fidelity and rounds to target would hold in larger networks or under heterogeneous or temporally correlated packet loss.

Reviewer_Kbw5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper aims to delve deeper into the aggregation rules of decentralized federated learning and study them in settings when the model parameters are exchanged in an imperfect environment (e.g., subject to packet drops). The study different aggregation rules and propose an aggregation rule a theoretically investigate different properties of the aggregation rules.

Critical strengths

The studied problem of the paper is timely.

The theoretical results seem to be new.

Critical weaknesses

The paper suffers from major limitations that makes it unsuitable for publication:

  1. The presentation of the paper is extremely difficult to follow. In particular, the same concepts are explained in different wording and it seems that sometimes elementary concepts are explained in such a complicated manner that makes the paper nearly impossible to follow.

  2. Besides the tone of the writing, the presentation of results do not seem to follow a cohesive order and the text seems scattered. For example, the reader is exposed to "Remark 3.1 (Scope)" in which a result that will show up way later in the paper is explained.

  3. Following the comments above, there are multiple sentences in the paper that are very hard to understand as the context and the wordings are not scoped well in the paper. Below, I provide some examples:

"A fixed-weight rule that substitutes a mask-independent surrogate"

"For u = 0 the missing weight mass is multiplied by zero rather than redistribute"

"An error term proportional to the state is a decay rate, not an offset."

"The main dividing line is convexity, not correction vs. faithfulness."

"Zero-fill applies no correction at all yet is unnormalized, since the weight on each missing coordinate is discarded rather than redistributed"

"It coincides with the H´ajek estimator (Datta & Polson, 2026) when reception is homogeneous, since the inverse weights then cancel, and under heterogeneous qji it takes the unweighted mean over the realized arrival set rather than weighting by inclusion probabilities, which Def. 3.2 excludes. It is faithful rather than imputing, and therefore belongs to a convex family."

These are just a subset of sentences which are extremely vague and will not be clear to the readers.

  1. Definition 5.1 formally defines the meaning behind the convexity of an aggregator, although that concept is informally used multiple times before that definition. Until that definition, the reader will have no idea about the clear definition of convexity of an aggregator.

  2. The main technique of the paper seems to be the simple update rule in Eq. (2). The reader will wonder what is unique about this simple aggregation technique as it is rather intuitive.

  3. The simulation results of the paper also follow the same scattered approach. In particular, I was not able to find one clear set of plots that shows the effect of increase on the network size and/or effect of increase in data heterogeneity on the performance.

Questions for authors

Please refer to the weaknesses mentioned above (I do not have clear questions as my main concerns center around how the paper is formatted and written). The paper seems to need a clear revision, through which the order of presentation is improved and results are made more accessible to the readers. The current heavy presentation of the paper hinders its understanding.

3 comments

Thank you for the detailed reading, and especially for quoting specific sentences. That made the revision concrete. We agree the presentation needed work, and all changes are shown in blue in the revised PDF.

Order of presentation.

  • The definition of the convex and unnormalized families now appears in Section 3 (Def 3.2), before it is first used. A short note after it fixes our terminology, and we now use one name per idea, for example "MFG" instead of several different phrases.
  • Remark 3.1 no longer appears early. Its content now sits right after the result it describes (Remark 5.3, after Prop 5.2).
  • Section 4 now opens with a short paragraph that connects its results into one argument.

The sentences you quoted, and what they now say.

  1. "A fixed-weight rule that substitutes a mask-independent surrogate…" now describes a rule with fixed weights that replaces each missing value by a value chosen without seeing it, such as zero or the node's own value.
  2. "For u = 0 the missing weight mass is multiplied by zero rather than redistributed" now explains that with zero-filling, the weight a missing value would have received is dropped instead of being passed to the values that arrived, so each round keeps only part of the total weight.
  3. "An error term proportional to the state is a decay rate, not an offset" is replaced by "zero-fill multiplies the model by a factor below one in every round." The corollary itself was also corrected, as described in our reply to Reviewer AKQp.
  4. "The main dividing line is convexity, not correction vs. faithfulness" now says that the families are separated by whether a rule's weights add up to one, not by whether it corrects for loss.
  5. "Zero-fill applies no correction at all yet is unnormalized…" now explains that ZeroFill makes no correction, yet its weights add up to less than one because the weight of each missing value is dropped.
  6. The Hájek sentence now says that the simplest convex rule averages each coordinate over the values that arrived, and that we call it MFG. It needs no knowledge of reception probabilities and fills in nothing. When all links share the same reception probability it equals the Hájek estimator, and when the probabilities differ, MFG gives every arrived value the same weight.

"The main technique is the simple update rule in Eq. (2)." We agree the rule is simple, and that is on purpose. We do not propose MFG as a new method, and similar normalization appears in earlier work, which we cite in Section 2. We use MFG as a reference point, the stable rule that adds nothing beyond averaging what arrived. Comparing it with rules that add estimation, reweighting or filled-in values shows what that extra machinery actually contributes. The contributions are what this comparison reveals. No rule that works without knowing the reception probabilities can be unbiased in every case, and the two usual fixes fail in different ways (Section 4). Rules split into two families depending on whether their weights add up to one, and this split holds for every loss pattern, so no learning-rate schedule changes it (Prop 5.2). Within the stable family, differences remain in fidelity, consensus and rounds-to-target even when final accuracy hides them (Section 7). Section 1 now says this directly.

Network size and data heterogeneity. For data heterogeneity, App. D.2 compares strongly non-IID data with near-IID data at 70% loss, and the family split and the order of the stable rules stay the same. For connectivity, the topology study in App. D.2 varies the degree from 2 (a ring) to 19 (fully connected), and MFG's advantage grows with degree, as Remark 5.5 predicts. Runs with 50 and 100 nodes are planned for the extended version, where we will show size, degree and data heterogeneity together in one figure, as you suggest.

Thank you again for the specific examples. We hope the revised version is much easier to follow, and we are happy to answer any further questions.

I would like to thank the authors for their responses. While some of my comments are addressed, still I do not think the paper is in the publication stage. As I mentioned in my comments, I have provided only a subset of very vague sentences in the paper, which the authors have corrected. Providing all the set of vague sentences is not practical as the paper has a lot of such instances. As a result, the text remains to be quite vague even after such corrections and I highly encourage the authors to conduct a major revamp on the paper. Due to the quality of the presentation and writing of the paper, I maintain my original rating.

Thank you for reading our response and for your honest advice. We agree the writing needs a full rewrite, and we will do that in the next version.

Reviewer_djqTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript addresses the challenge of aggregating model updates in non-reliable decentralized FL without retransmission, where clients receive a channel-imposed and unobservable subset of model coordinates. Instead of comparing aggregation rules by training with a decaying learning rate and checking the final accuracy, the authors propose to measure per-round fidelity, since a vanishing step size suppresses each rule's aggregation error by the end of training. Through theoretical results and experimental validation, they show that convex averaging progresses towards consensus, whereas imputation and inverse-probability correction, and that fidelity separates rules where annealed final accuracy doesn't.

Critical strengths

I appreciated the mathematical clarity of the manuscript, in particular, the authors, which relied on linear algebra, graph theory tools and probability tools. In particular, the notation and the statements are clear and read reasonably

Introduction and Related Work sections are well written and properly frames the method within Decentralized Federated Learning literature, clearly stating the state-of-the-art limitations that the authors aim to address

Critical weaknesses

Besides the theoretical contribution, the experimental advantages are hard to grasp. Even though state-of-the-art limitations are properly stated in the introduction, the novelty convincingly demonstrated.

Section 4 on the Imputation Trap is too dense and the theoretical contributions are not linked between each other. They should be more carefully and thoroughly discussed. Even if the statements are reasonable, they are not properly contextualized within their contribution. I advise the authors to better incorporate Definitions, Theorems, Lemmas and Propositions in the discussion.

Questions for authors

Main concerns are listed above.

Further Questions.

Could the authors clarify some variables in Eq.1, eg, aij(⋅)a_{ij}(\cdot)?

Could the authors discuss how independence across links impacts their results (ll 135-137)?

Could the authors detail and explain better contributions iv and v?

What do the authors refer to as heterogeneous qjiq_{ji}, ll 216-217?

Evaluation is limited to small networks, 20 clients, could the authors show how the results change when increasing the number of clients to larger scale?

Minor concerns. Consistently with what observed in the Weaknesses above, I would advise the authors to order the Appendix differently, moving the Background before the Proofs, as it could help a little to better visualize to non-DFL readers the impact of the results. (Typo on title of Appendix B)

3 comments

Thank you for the careful reading. We are glad the statements and the framing of the introduction and related work came across clearly. All changes are in the revised PDF, shown in blue.

W1. The practical advantage and the novelty are hard to see. We agree the main message was not stated clearly enough. We do not propose MFG as a new method. It is the simplest rule in the stable family, and we use it as a reference point. The contribution is about how aggregation rules behave under packet loss and how they should be compared. Final accuracy under one learning-rate schedule tells you which family a rule belongs to, but it does not reliably rank rules within the stable family. Aggregation fidelity, consensus distance and rounds-to-target do rank them, under both schedules. One example from the same runs makes this concrete. On CIFAR-10 with cosine annealing at 50% loss, Soft-DSGD ends with a higher mean accuracy than MFG (75.5 vs 74.0), yet MFG's consensus distance is about 4.8× lower and its aggregation error about 6.2× lower. The two measures rank the same runs in opposite orders. Contributions (iv) and (v) now say this directly.

W2. Section 4 is dense and its results are not linked. We agree and rewrote the opening of Section 4 so the results read as one argument. No rule that works without knowing the reception probabilities can be unbiased in every case (Thm 4.1). So a rule must either fill in the missing values or accept some bias. Filling them in leaves a bias that no choice of weights removes (Lem 4.2), and filling in zero shrinks the model (Cor 4.3). Reweighting by inverse probabilities removes the bias but makes the error grow with the size of the parameters (Prop 4.4), because the weights no longer add up to one (Prop 4.5). Rules whose weights stay non-negative and add up to one avoid both problems (Section 5). The definition of this family now appears in Section 3, before Section 4 uses it.

Q1. Variables in Eq. (1). Every symbol is now defined right below the equation. For node ii and coordinate pp, xj(p)x_j^{(p)} is neighbor jj's value, mji(p)m_{ji}^{(p)} is 1 if that value arrived and 0 otherwise, and aiia_{ii} and aija_{ij} are the weights on the node's own value and on each arrived value. The weights may depend on which values arrived, but never on the missing ones. We also give each method's weights as examples and fixed two notation slips nearby.

Q2. How independence across links affects the results. It depends on the result, and a new paragraph at the end of Section 3 explains this. Lem 5.1 and Prop 5.2 hold for every loss pattern, so they need no independence and still hold under correlated or bursty loss. The split between the two families rests on these results. Lem 4.2, Cor 4.3 and Prop 4.4 only use the average reception rate of each link. Thm 4.1 still holds with correlated loss, because a rule would then have to be unbiased in even more cases. Only Prop 5.4 and the coverage formula need independence. When losses are positively correlated, a coordinate is missed by all neighbors more often than (1−q)di(1-q)^{d_i} predicts.

Q3. Contributions (iv) and (v). Both are rewritten. (iv) Aggregation fidelity measures how far a node's aggregate is from the one it would form if nothing were lost. Along training, the stable rules keep the same order at every loss level, under both schedules and on both datasets, and the gap to DFL-AA grows from 1.56× to 7.62× on CIFAR-10. When every rule is given the same model states and the same losses, MFG is still the most faithful (new App. D.3). (v) Annealing shrinks the accuracy differences inside the stable family but not these other differences, so comparing final accuracy under one schedule can rank a less faithful rule higher. We suggest reporting fidelity, consensus distance and rounds-to-target under at least two schedules.

Q4. Meaning of "heterogeneous". It means that different links have different reception probabilities. It is not about data heterogeneity. When all links share the same probability, MFG equals the Hájek estimator. When they differ, the Hájek estimator would need each link's probability, which MFG does not use. The sentence is rewritten. All our experiments use equal probabilities, and we plan to test unequal ones in the extended version.

Q5. Larger networks. We agree this matters. Runs with 50 and 100 nodes at the same average degree are planned for the extended version, and Section 8 now states the 20-node scope plainly.

Minor. The background appendix now comes before the proofs, and the typo is fixed.

Thank you again. We are happy to answer any further questions.

I thank the authors for their response and for clarifying some points. However, my concerns remain unsolved and I believe that the paper is not mature and sound enough for the venue. I keep my initial assessment.

Thank you for taking the time to read our response. We understand your view, and we will keep working to improve the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General addressing for reviews

We thank all reviewers for their careful reading and concrete suggestions. We have uploaded a revised PDF with all changes are shown in blue, and the main text remains within 10 pages.

Presentation.

  • The convex/unnormalized definition now appears in Sec.3 (Def. 3.2), before it is first used, followed by a one-line terminology note. We use one name per concept throughout.
  • Remark 3.1 (Scope) was removed from Sec. 3 and now follows the result it qualifies (Remark 5.3, after Prop. 5.2).
  • Sec. 4 opens with a paragraph that links its results into one argument (Thm. 4.1 --> Lem 4.2 --> Cor 4.3 --> Props 4.4–4.5 --> convex family).
  • Every symbol of Eq. (1) is defined, with each method's coefficients as examples, and a paragraph states which results depend on independence across links.
  • All sentences quoted by Reviewer Kbw5 were rewritten. Background now precedes Proofs in the appendix (typo fixed).

Correctness and scope.

  • Corollary 4.3 is corrected to include the gradient steps between aggregation rounds (new proof in App. B.3).
  • The accuracy-equivalence claim is narrowed in the abstract, Sec. 7.2 and Sec. 8, and supported by a statistical comparison with a ± 1-point equivalence margin fixed in advance (new App. D.4).
  • A common-state replay applies every aggregator to identical model states and masks, measuring error against both a common target and each rule's own lossless operator (new App. D.3). MFG remains the most faithful rule at every loss level. We have narrowed the within-family claim accordingly.

Reproducibility.

Detailed answers follow under each review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.