acceptodds
Under review as a conference paper at ICLR 2027

A Decentralized Lens on Centralized Federated Learning: Unifying FL via Inexact Douglas–Rachford Splitting

Abstract

Federated learning enables collaborative training without pooling client data, but data heterogeneity remains a major challenge, especially when combined with partial participation and inexact local computation, whose joint effects are difficult to capture in a unified analysis of existing methods. To analyze these interactions, we develop an operator-splitting framework on the full client product space with a consensus constraint. Within this framework, a target field defines the residual and energy, while a conditional mean-square error criterion connects the target dynamics to actual client computations. This connection enables convex convergence guarantees under an additional summable conditional-mean condition, as well as nonconvex stationarity guarantees through expected forward–backward-envelope descent. Under local Kurdyka-Łojasiewicz (KL) geometry, almost-sure boundedness, and explicit process and error-tail conditions, the nonconvex analysis further yields finite length and convergence of the full sequence. Applying the abstract algorithm to federated optimization yields the FedExSplit (FES) family, with variants defined by the target potential. We then use an exact heterogeneous quadratic construction to explain how its mean-anchored variant (FES-MA) can suppress the effect of client disagreement on the mean update in a specified parameter regime. Controlled one-step experiments support this mechanism, while training experiments show that FES-MA improves validation accuracy over FedAvg by approximately 1.7 percentage points on FEMNIST and 2.6 percentage points on CIFAR-100 under high heterogeneity ().

Primary area:
optimization
Keywords:
Optimization, Federated Learning, Splitting Methods, Learning Theory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3TTJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes FedExSplit (FES), a Douglas–Rachford splitting (DRS)-based framework for federated learning that explicitly models persistent client states, partial participation, and inexact local updates. Its key idea is to control the implicit error of the full update, providing a unified basis for convergence analysis. The paper establishes convergence guarantees for convex and nonconvex objectives, including stationarity and full-sequence convergence under additional assumptions, and instantiates the framework through two variants.

Critical strengths

One strength of this paper is its unified convergence analysis for inexact DRS updates. By considering the error in the full update rather than only the local solver error, the framework provides a useful way to analyze federated algorithms with partial participation and persistent client states. The paper also goes beyond the general theory by deriving an error bound for FES-MA and examining how client disagreement affects its updates. While the theoretical advantages are established under specific assumptions, the analysis offers some useful insights into the behavior of the proposed method.

Critical weaknesses

The main concern is that the theoretical contribution beyond existing DRS-based federated optimization methods, such as FedDR and iFedDR, is not entirely clear. While the full-update error formulation is useful, the nonconvex convergence rate is comparable to existing results, and the paper could better clarify what important cases can be handled by the proposed framework that were not covered before. In addition, some of the stronger convergence guarantees rely on restrictive assumptions. These issues do not invalidate the technical results, but they make it harder to assess the significance and practical scope of the theoretical advances.

Although the paper is primarily theoretical, the empirical evaluation could be a bit more comprehensive for an ICLR submission. I see this mainly as a suggestion to strengthen the paper rather than a major concern.

Questions for authors

Could the authors clarify more precisely how the proposed full-update error framework extends existing analyses such as FedDR and iFedDR? In particular, what important settings or convergence guarantees can be handled by this framework but are not directly covered by prior results?

1 comment

We thank the reviewer for recognizing the complete-update analysis and for asking which concrete settings it covers. The framework retains the O(1/k)O(1/k) general stationarity-rate order and analyzes sampling and local-solve errors through a complete-update budget. The revision details these error estimates, separates the assumptions by conclusion, and expands empirical reporting. The italicized questions summarize the review.


Q1. Which concrete cases go beyond a shared reformulation of existing DRS analyses?

A1. The Introduction now motivates a common convergence argument by comparing the error and participation conditions in existing analyses. Appendix D verifies complete-update errors for FES-DR and FES-MA under partial participation and controlled local inaccuracies. For FES-MA, the shared gradient is evaluated at the aggregate mean. An unselected client's state can remain fixed while its gradient component changes with that mean. The error estimate tracks this dependence together with the local solver residuals.

Propositions D.5–D.6 establish expected descent and mean-square step control using the sampling structure. With exact correction and the local stopping rule (3.4), Corollary D.10 gives sufficient structured-descent regimes for every positive participation rate. Proposition C.4 supplies the common convergence argument. These implementation-specific estimates connect the two update designs to the same O(1/k)O(1/k) best expected squared-residual bound under the stated conditions.


Q2. What is the relation to FedDR, DANE, and iFedDR, and what remains to be verified?

A2. Section 3.1 states the correspondence conditions. MA recovers DANE's mean trajectory at γ=τ=1\gamma=\tau=1 with full participation, equal weights, exact correction and solves, gradient scale one, and penalty 1/β1/\beta. DR at τ=1\tau=1 recovers FedDR without server regularization through the stated transformed states, with exact correction and solves, consistent initialization, and identical selections.

The full-update framework separates an implementation-specific error calculation from the common convergence argument. For an iFedDR update, this calculation would track the complete error induced by its relative-error correction and adaptive steps.

Persistent states originate in existing methods such as FedDR. MA combines them with mean-anchored gradient-corrected local models. FedDANE-related sampled corrections can enter the local model subject to verification of their induced error. Algorithm 2 aggregates persistent-state increments, while FedDANE averages sampled responses.

Exact multicenter aggregation preserves the result when it computes the same projection and other required global quantities. Approximate gossip changes the update and needs a complete-error bound, including its effect on MA's gradient evaluation point.


Q3. How restrictive are the stronger convergence assumptions?

A3. Section 2.2 separates the conclusions. Basic nonconvex stationarity requires smoothness, a finite constrained lower bound, initial-energy integrability, a positive descent margin, and square-summable additive errors. Finite-length and whole-sequence convergence additionally require boundedness, local KL geometry, an energy-process condition, and stronger error-tail summability.

The PL condition discussed in Section 2.2 supplies one route to the required limiting-energy property. The weaker energy-process condition (2.10) permits random limiting energies and is assumed separately from local KL geometry. Appendix C extends the conditional-expectation argument to additive errors. These extra assumptions are used for finite-length and whole-sequence convergence. The basic O(1/k)O(1/k) expected squared-residual order uses the stationarity assumptions listed above.


Q4. What empirical evidence has been added, and what does it support?

A4. Section 4 includes a controlled MLP study over five new seeds and (β,γ)(\beta,\gamma) heatmaps across inner-solver step sizes, including γ=1\gamma=1. All four methods use five inner iterations. At LDA .075, the sample SD of A50 is 1.99 for MA and 7.70 for DR, showing greater consistency for MA across these five runs. In the MLP study, both FES variants perform better than Avg and Prox on all five seeds at the strongly heterogeneous LDA levels .15 and .2, and below these baselines at the more homogeneous LDA level 1.

The revised CIFAR-100 task trains a 746,852-parameter nonlinear head on frozen features, with participation .5 and ten paired seeds per LDA level .05, .03, and .01. MA's mean accuracy-AULC gains over DR are 1.900, 2.651, and 4.395 points; gains over Avg are .580, .705, and 1.083. The curves include pointwise 95% Student-t intervals. Appendix H reports the complete development search and selected settings.

DR trails Avg in CIFAR-100 accuracy-AULC. MA's per-iteration accuracy-AULC advantage does not persist at 100 matched communication round trips.

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

This paper aims to develop a so called product space framework for federated learning.

Critical strengths

None

Critical weaknesses

To be honest, the reviewer had a very difficult time understanding the paper, even if the reviewer is fairly familiar with the distributed optimization/FL literature. The paper, in its current form, is unreadable and I am afraid it would require a significant amount of effort for revision before it could be considered for publication.

  • One major concern is that I do not understand why one needs a decentralized perspective on centralized FL. Can the authors explain what new insights this perspective brings? Are there new technical challenges emerging and if so, how were they solved?

  • The math in this paper is very hard to follow. First, many equations are not numbered, and for the numbered ones, they are numbered inconsistently. Second, there are too many definitions but no discussion of their physical meanings.

  • There are missing discussions for all important math results. For example, Theorem 2.3 talks about convergence, but this convergence is almost trivial, and there is no discussion of what this means. A convergence bound that talks about the rate, and its correlation to different system parameters, is generally expected.

  • The comparison to existing studies is very uninformative. For example, in Table 1, the reviewer expects to see a list of concrete assumptions, convergence rate etc, while the table misses such most important information. Appendix A has the same problem.

  • The experiment results look problematic. For example, results on CIFAR-100 with \alpha=0.1 achieve around 0.35 across all methods, which is far below expectation. The authors, please look into the implementation in detail.

Questions for authors

See above.

2 comments

We thank the reviewer for identifying the obstacles to understanding and assessing the paper. The revision clarifies the motivation, separates convergence claims by assumptions, and replaces the empirical evaluation with explicitly specified tasks and statistical reporting. The italicized questions summarize the five concerns.


Q1. What does a decentralized perspective add to centralized federated learning?

A1. The revised Introduction separates two motivations: reusing convergence analysis across sampling, local accuracy, and aggregation choices, and studying how persistent client states affect subsequent updates. The product-space view makes each client's optimization state explicit and separates local computation from the consensus projection. Building on FedDR's persistent-state design, it allows gradient evaluation, sampling, local errors, and aggregation to be analyzed together through the complete update.

For MA, the shared gradient depends on the aggregate mean. An unselected client's state can remain fixed while its gradient component changes with that mean, so the complete-update error analysis accounts for changes in the shared gradients alongside local solver residuals. Appendix D derives this error and establishes structured sampling estimates. Appendix F also gives a same-parameter example in which retaining states produces a greater objective decrease than resetting them to their mean.

For multicenter implementations, the iteration is preserved when aggregation computes the same projection and required global quantities while preserving sampling and local updates. Finite-round gossip requires a separate bound on the resulting complete error, including MA's gradient evaluation at an approximate mean.


Q2. Can the mathematical objects and equation references be made easier to follow?

A2. The revision explains the roles of the main objects: PP is the objective, g=∇Pg=\nabla P its gradient, BB contains the client gradients actually evaluated, and C=ΠBC=\Pi B is their shared correction. The residual measures disagreement and stationarity; Θk\Theta_k is the complete-update error; the forward–backward envelope is the auxiliary descent energy.

Section 2.2 displays the central descent inequality and separates the assumptions for stationarity, convex convergence, and finite-length convergence. Appendix D connects the local proximal stopping rule to the outer error estimates. Referenced update, error, and descent expressions use consistent section-based numbering. Expressions not subsequently referenced remain unnumbered.


Q3. What does convergence mean, and how does the bound depend on system parameters?

A3. The basic nonconvex result gives vanishing residuals and stationarity of existing cluster points. Theorem 2.2 establishes almost-sure finite path length and convergence of all client states to one consensus stationary point under boundedness, KL geometry, an energy-process condition, and stronger error-tail control. The basic stationarity analysis gives an O(1/n)O(1/n) best expected squared-residual bound.

Proposition C.4, Eq. (C.8), gives

\frac{\mathbb E\mathcal E_\beta(x^0)-P_{\inf}+c_\varepsilon\sum_{k\ge0}\varepsilon_k^2}{a(n+1)}.$$ Propositions D.5–D.6 specify $a_{\mathrm{DR}}=p\nu_{\mathrm{DR}}/(2\beta)$ and $a_{\mathrm{MA}}=pc_\tau/(4\beta)$ with explicit positive margins. Corollary D.10 gives sufficient regimes for every $p>0$: $\tau=1$, $\gamma=1.1$, $\beta L\le1/16$, or $\tau=1/p$, $\gamma=1.1p$, $\beta L\le p/16$. Participation enters the descent coefficient, the normalized regime scales additive local errors by $1/p$, and the residual depends on $\beta$. Theorem 2.1 treats convex residual and objective-gap rates under its separate conditional-mean requirement.

Q4. How do the assumptions and rates compare with existing analyses?

A4. Lines 43–50 compare the assumptions in the relevant analyses. Our stationarity result uses the standard smooth nonconvex setting: a finite constrained lower bound, integrable initial energy, and a conditional mean-square budget on the complete-update error, with square-summable additive tolerances and parameters ensuring a positive descent margin. KL geometry and bounded trajectories enter only the full-sequence result. The cited inexact FedSplit and fixed-penalty FedADMM-InSa guarantees use strongly convex objectives and full participation.

For full-sequence convergence, Section 2.2 (lines 172–221) states the conditions needed by our stochastic KL argument: almost-sure boundedness, ordinary local power-KL geometry with a common deterministic exponent, control of accumulated error tails, and the conditional energy condition (2.10). The uniformized KL constants and entry times may depend on the sample path, and stationary points may be nonisolated. Using (2.10) in place of global PL geometry also allows random limiting energies.

The nonconvex analyses of INEXACTDANE, FedDR, FedADMM, FedADMM-MM, and FedCDR, together with FES-DR/FES-MA, give the same O(1/K)O(1/K) order for their respective squared stationarity measures under their stated assumptions and iteration counts. FedADMM-MM's O(K−1/2)O(K^{-1/2}) norm bound is consistent with this order; EF-Feddr's bound includes an additional error term that persists at fixed compression accuracy.


Q5. Why was the original CIFAR-100 accuracy low, and what does the revised experiment establish?

A5. The original CIFAR-100 experiment was designed as an illustrative comparison of the algorithms under common training and evaluation conditions. It used a basic end-to-end pipeline with limited task-specific refinement of hyperparameters, data preprocessing, and training procedures, yielding approximately 35% accuracy. The revised evaluation uses a frozen pretrained encoder to provide a stronger shared image representation and compares the methods by training the same nonlinear head.

Section 4.3 uses a frozen PE-Core-G14-448 encoder and a nonlinear head with 746,852 trainable parameters, 100 clients, participation .5, and five full-local-gradient inner iterations for every method. Parameters are tuned on development runs and frozen for ten new paired seeds at each LDA level .05, .03, and .01. Appendix H reports all evaluated development configurations. Figure 4 gives means and pointwise 95% Student-t intervals.

AULC is accuracy integrated over the iteration horizon and divided by that horizon. MA's mean validation-accuracy AULC gains over Avg are .580, .705, and 1.083 percentage points; its gains over DR are 1.900, 2.651, and 4.395. MA exceeds Avg in all thirty seed–LDA pairs. On the frozen-feature task, the comparison shows these gains per outer iteration and no accuracy-AULC advantage over Avg at 100 matched communication round trips.

Reviewer_G2MaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a framework that casts centralized federated learning as relaxed, inexact Douglas–Rachford splitting on the client product space with a consensus constraint. A single conditional mean-square error budget captures errors from partial participation, correction estimates and inexact local solves. They prove nonconvex stationarity and convex almost-sure convergence. Under local KL geometry and additional boundedness and error-tail conditions, they also prove nonconvex whole-sequence convergence. The framework yields the FedExSplit family: FES-DR recovers FedDR, and the mean-anchored FES-MA generalizes DANE. An exact quadratic analysis identifies a regime where FES-MA reduces the effect of client disagreement on the mean update. Experiments show FES-MA improving over FedAvg by about 1.7 points on FEMNIST and 2.6 points on CIFAR-100 under high heterogeneity.

Critical strengths
  • Rigorous, carefully scoped analysis under a general inexactness model. The work controls the error of the full client update, covering sampling, correction estimates and local-solve errors together. It does so through a conditional mean-square budget rather than a pathwise or proximal-only error. This is a natural fit for stochastic federated learning, and it yields stationarity, convex convergence and KL-based sequence convergence from one abstract scheme. The constants are explicit and the assumptions are stated precisely, and the derivations I checked hold (the descent constants, the convex distance recursion, the feasibility example). This matters to my assessment because it is the paper's core contribution: a technically correct and reusable framework, not a loose analogy.

  • A clean structural explanation connecting FedDR and DANE. The Douglas–Rachford view places FedDR and DANE in a single persistent-state template that differs only in where gradients are evaluated. The exact quadratic decomposition (Proposition 4.1) shows precisely how client disagreement couples to curvature heterogeneity in each method's mean update, and why mean anchoring removes that coupling. This matters because it is the paper's most informative insight for practitioners. It gives a principled account of when DANE-style anchoring helps under heterogeneity, independent of the empirical results.

Critical weaknesses
  • The theory does not cover the setting the paper motivates or the one it evaluates. The paper presents the interaction of partial participation and inexact computation as its central challenge. However, Prop. E.2 gives the exact FES-DR sampling error, ρ = γ²p(1−p), and this identity also holds for nonconvex f_i with βL < 1. Substituting it into condition (2.5) shows that the nonconvex guarantees require p ≳ 0.996. This is effectively full participation, and it is weaker than FedDR's own analysis under sampling. The experiments also fall outside every certified regime: they use μ ≫ 1, γ = 1.1, local SGD, and a sampling ratio of 0.5. In addition, they average only participating clients, which differs from the aggregation analyzed in Algorithm 2. I consider this the most consequential weakness. The theory and the experiments end up as separate contributions, and neither delivers the unified treatment of participation and inexactness that the paper claims.

  • The evidence does not establish the value of FES-MA's distinctive design. In Example 4.2, FES-MA's disagreement-induced perturbation vanishes at γ = 1 (the DANE case) and grows as γ moves away from 1. The mechanism therefore suggests that retaining client history is a cost relative to DANE rather than a benefit. The experiments do not resolve this question. There is no γ ablation including γ = 1, and the comparison omits DANE, FedDR, SCAFFOLD, and FedDyn. Reported accuracies are read from plots, CIFAR-100 variance is not reported, and the FedProx baseline appears poorly tuned. As a result, I cannot attribute the reported 1.4–2.7 point gains to the proposed design rather than to DANE-style anchoring, tuning differences, or run-to-run variation.

Questions for authors

Clarification

  1. Partial participation in the nonconvex theory. For FES-DR with exact solves, Prop. E.2 gives ρ = γ²p(1−p), and as far as I can tell this also holds for nonconvex f_i when βL < 1. Substituting into (2.5) appears to require p ≳ 0.996. Is this reading correct? If it is, do you have a nonconvex guarantee that covers realistic participation rates? If you do not, would you state this limitation explicitly in the abstract and in Table 1?

  2. Aggregation used in experiments. The training implementation averages only participating clients, while Algorithm 2 maintains a full-state mean that retains unselected clients. Could you report results using Algorithm 2's aggregation at sampling ratio 0.5? Alternatively, could you explain how the analysis applies to the aggregation you implemented?

  3. The role of γ. Example 4.2 suggests FES-MA's advantage is greatest at γ = 1, the DANE case. Could you report training results over a range of γ that includes γ = 1, together with DANE (or FedDANE) and FedDR baselines? This would show whether retaining client history (γ ≠ 1) improves on DANE-style anchoring.

  4. Experimental reporting and tuning parity. Please provide the following:

    • the number of clients for each dataset;
    • the number of runs and the variance for CIFAR-100;
    • numerical tables in place of values estimated from plots;
    • the tuning grids for every method, given that the FedAvg learning rate is fixed and FedProx uses μ_prox of 100–200.
  5. Communication cost. Does forming FES-MA's correction require an additional client–server round trip in each round? If so, how do the methods compare at equal communication?

Suggestions

  • Include SCAFFOLD and FedDyn, the standard drift-correction baselines, in the comparison.
  • Measure the disagreement-induced mean perturbation during training, so the quadratic mechanism can be tested beyond one-step examples.
  • Extend Proposition 3.1 to μ ≠ 1 and to local SGD, which would bring the experiments closer to the certified regime.
4 comments

We thank the reviewer for the detailed assessment and for recognizing the full-update analysis and the DR/MA connection. The revision adds structured sampling certificates, aligns aggregation with the persistent-state iteration, and reports paired multi-seed experiments. We distinguish the certified conditions from the experimental settings below. The italicized questions summarize the review.


Q1. Does the original nonconvex sampling certificate effectively require full participation?

A1. Your reading of the original scalar-error certificate is correct: it imposed a near-full-participation restriction. The revised proof incorporates the sampling structure directly into the descent estimate.

Propositions D.5–D.6 establish expected descent and mean-square step control for both FES-DR and FES-MA. Proposition C.4 supplies the common convergence argument. With exact correction and local stopping rule (3.4), Corollary D.10 gives two sufficient regimes for every positive participation rate pp:

τ=1,γ=1.1,0<βL≤1/16,\tau=1,\quad \gamma=1.1,\quad 0<\beta L\le 1/16,

or

τ=1/p,γ=1.1p,0<βL≤p/16.\tau=1/p,\quad \gamma=1.1p,\quad 0<\beta L\le p/16.

The additive tolerance is τεkprox\tau\varepsilon_k^{\mathrm{prox}} and must satisfy the stated summability requirements. The descent coefficients retain a factor pp; the normalized regime also amplifies local tolerances by 1/p1/p. Under these parameter and error conditions, the certificates apply to every p>0p>0. Section 3.2 and Appendices C.4 and D state the conditions. Whole-sequence convergence additionally requires boundedness, KL geometry, and the energy-process and error-tail conditions.


Q2. Does the implementation use Algorithm 2's full-state aggregation?

A2. The revised implementation retains unselected client states and maintains the full weighted mean:

xik+1=xik+τ(uik+1−xik),i∈Sk,x_i^{k+1}=x_i^k+\tau(u_i^{k+1}-x_i^k),\quad i\in S_k,

xˉk+1=xˉk+τ∑i∈Skai(uik+1−xik).\bar x^{k+1}=\bar x^k+\tau\sum_{i\in S_k}a_i(u_i^{k+1}-x_i^k).

Equal weights recover Algorithm 2; fixed sample weights give the corresponding weighted projection. CIFAR uses 100 clients, 50 selections per round, and τ=1\tau=1. Across 1,800 update checks, the largest relative discrepancy from explicitly recomputing the weighted mean was below 2.47×10−72.47\times10^{-7}.

The local objective is fi(u)+∥u−vik∥2/(2β)f_i(u)+\|u-v_i^k\|^2/(2\beta), solved using (4.1). We removed the independent scaling parameter μ\mu. The revised experiments use five full-local-gradient inner iterations for every method. Sections 3.1–3.2 state the local-residual and correction-error conditions required for convergence.


Q3. Does retaining history improve on the DANE case?

A3. The original quadratic example starts at an optimal global mean, where removing the history perturbation is desirable. Appendix F adds a same-parameter example away from optimality: from objective 1, the next values are approximately .4472 for persistent MA, .4613 after resetting its states, and .5256 for persistent DR.

Section 4.2 scans (β,γ)(\beta,\gamma), including γ=1\gamma=1, across inner step sizes. Exact DANE correspondence additionally requires full participation, exact correction and solves, and τ=1\tau=1. In the MLP experiment, both FES variants perform better than FedAvg and FedProx on all five seeds at the strongly heterogeneous LDA levels .15 and .2, and below these baselines at the more homogeneous LDA level 1.

Section 4.1 and Appendix F also measure the mean-update change caused by resetting all client states to their current mean. DR and MA share development-selected parameters within each diagnostic block. At LDA .03, γ=1.5,1.9\gamma=1.5,1.9, MA's mean perturbation is 1.32% and 3.59% of DR's, respectively; at LDA 1, γ=1.9\gamma=1.9, MA instead has larger perturbation. These diagnostics measure mean-update perturbations along each method's own training trajectory.

Q4. What are the client counts, variability, numerical results, and tuning protocol?

A4. MLP uses 12 fully participating clients; CIFAR uses 100 clients with participation .5. MLP transfer parameters are selected on seeds 4000–4001 at LDA .1 and evaluated on seeds 4010–4014. The separate perturbation diagnostic selects shared DR/MA parameters per LDA/relaxation block.

CIFAR parameters are selected per method and LDA on seed 6000, then frozen for seeds 6001–6010. Partitions, initialization, and sampling are paired. Appendix H lists all 139 development configurations, scores, selected settings, search ranges, boundary expansion, and unequal per-method candidate counts.

Over 100 CIFAR outer iterations, validation-accuracy AULC is the trapezoidal area divided by the horizon. Values below are percentages, reported as mean ± sample SD across ten seeds:

LDAFES-DRFES-MAFedAvgFedProx
.0589.815 ± .22991.715 ± .10191.135 ± .09291.133 ± .090
.0388.639 ± .18491.289 ± .12790.584 ± .16590.584 ± .165
.0185.636 ± .49290.031 ± .25188.948 ± .17088.948 ± .170

Paired MA-minus-Avg differences are .580 ± .096, .705 ± .182, and 1.083 ± .189 points. Figure 4 reports pointwise 95% Student-t intervals for the mean; these differ from the sample SDs above. The validation image pool is fixed.


Q5. Is the advantage retained at equal communication cost?

A5. The direct FES implementation uses two round trips per outer update, versus one for FedAvg and FedProx. MA's accuracy-AULC advantage per outer iteration does not persist at 100 matched communication round trips. The correction stage adds computation beyond the matched inner-iteration counts.

Original FedDR can communicate transformed states directly. Remark 3.1 and Appendix E analyze a pipelined alternative using previous corrections and refreshed gradient differences. Ordinary iterations use one exchange under its stated conditions, with refreshes and vector payloads counted separately. The experiments reported here use the direct FES implementation.

Thank you for the clarifications. It addresses some of concerns.

Thank you for the follow-up. Please let us know if there are any remaining concerns we can address. If the revisions and clarifications have improved your overall assessment, would you consider updating your score accordingly?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.