{"id":"04d3995c-2a1c-4a8a-9a43-6a1d295730a3","arxiv_id":"2501.09946","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Client-centric federated adaptive optimization lets clients participate asynchronously with heterogeneous local work while the server runs Adam-style updates, with a proven O(sqrt(1/(A E T))) nonconvex rate.","lead":"This paper proposes a family of federated learning algorithms, CC-FedAdam, CC-FedAdagrad, and CC-FedAMS, that combine asynchronous, self-scheduled client updates with server-side adaptive optimization. The authors prove a nonconvex convergence rate and report that their methods outperform generalized FedAvg baselines by large margins on several benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1 requires uniform random participation, so the advertised arbitrary client participation is not covered by the claimed best-known rate; Remark 4.1.2 concedes an unavoidable Ω(σ_g^2) term otherwise.","rationale":"The reader identified the uniform-participation assumption as the weakest load-bearing premise, and the manuscript itself confirms this in Remark 4.1.2. I agree: the central claim of the paper is that Client-Centric Federated Adaptive Optimization simultaneously enables arbitrary client participation, asynchronous aggregation, and heterogeneous local computing while matching the best-known convergence rate O(√(1/(AET))). The theorem only proves this rate under uniform random participation, and the paper explicitly concedes that arbitrary participation introduces an unavoidable Ω(σ_g^2) term. Therefore the advertised feature is not covered by the main theoretical result, and the abstract overstates the scope of the guarantee. This is a substantive correctness-scope issue, not merely a presentation issue, because a reader relying on the abstract would reasonably believe the algorithm converges to a stationary point under self-determined client availability. The paper deserves credit for stating Remark 4.1.2 explicitly, which makes the limitation transparent, but the limitation still undercuts the central claim. Other reservations—lack of machine-checked proofs, experiments without error bars, and the asymptotic regime T≥AE^5 not matching the T=500 experiments—are secondary and addressable. Since the identified concern is real but fixable (e.g., by rephrasing claims or adding a theorem for positive participation probabilities with a bias term), the verdict remains CONDITIONAL, unchanged from the reader's assessment.","tokens_in":66971,"tokens_out":10409,"duration_ms":104305,"concrete_test":"Re-derive the key unbiasedness step E[(1/A)Σ_{i∈S_t} h_i] = (1/B)Σ_{i=1}^B h_i and the bound on E[||Δ_t||^2] under a non-uniform participation distribution, e.g., p_i = A/B for a subset of clients and p_i = 0 for an excluded client, while keeping all other assumptions unchanged. If the derivation produces an additional term of the form Σ_i (p_i - A/B)^2 ||∇f_i(x_t)||^2 that does not vanish as T→∞, then Theorem 4.1's O(√(1/(AET))) rate cannot hold for arbitrary participation, confirming that the abstract's 'arbitrary client participation' claim is unsupported by the theory.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's abstract and Section 3 advertise arbitrary client participation as a central feature of Client-Centric Federated Adaptive Optimization, but the main convergence guarantee (Theorem 4.1) assumes 'each client is included in S_t with probability A/B uniformly and independently.' Remark 4.1.2 explicitly concedes that without this uniform-participation assumption the rate degrades to O(√(1/(AET))) + O(E^2/T) + O(τ^2/T) + Ω(σ_g^2), and it constructs a two-client example showing the Ω(σ_g^2) term is unavoidable when a client never participates. Consequently, the paper's headline claim that the framework converges with the best-known rate under arbitrary client participation is not supported by the theorem; the theoretical guarantee only covers the special case of uniformly random participation, and the advertised 'arbitrary' regime has only a much weaker convergence-to-a-neighborhood result. This is the most load-bearing concern because it directly affects the paper's central contribution: the claimed unification of system heterogeneity features with best-known convergence is only partial, and the abstract overstates the scope of the theorem. A secondary but related issue is that the theorem's stated 'independently' participation is inconsistent with Algorithm 2's fixed-size A buffer; the proof uses without-replacement probabilities P{i,j∈S_t}=A(A-1)/(B(B-1)), so the formal assumption is at best imprecise. The proof itself is not machine-checked, and the clean O(√(1/(AET))) rate requires T≥AE^5 and τ≤(T/(AE))^{1/4}, conditions not met in the experiments (T=500, τ=5), but these are secondary to the participation-scope gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Client-Centric Federated Adaptive Optimization (CC-FedAdam, CC-FedAdagrad, CC-FedAMS), a class of federated adaptive optimizers that combine asynchronous server aggregation, time-varying and device-dependent local computation, and client-determined participation. The main theoretical contribution is Theorem 4.1, which under Assumptions 1–4, bounded delay, and a uniform random participation model gives a convergence rate for general nonconvex objectives of O(sqrt(1/(AET))) + O(E^2/T) + O(tau^2/T), improving to O(sqrt(1/(AET))) for large T and moderate lag. Experiments on CIFAR-10, CIFAR-100, Fashion-MNIST, and StackOverflow compare the proposed methods against CC-FedSGD and CC-FedAvg.","tokens_in":67281,"tokens_out":3088,"duration_ms":32418,"significance":"If the stated convergence result holds, the paper makes a useful step by combining server-side adaptivity with asynchronous aggregation and heterogeneous local computation in a single framework. The proof is extensive and self-contained, and the rate matches the best known rate for synchronous and asynchronous federated optimization under the uniform-participation assumption. The paper also provides an honest Remark 4.1.2 showing that arbitrary participation introduces an unavoidable Omega(sigma_g^2) term, which is a valuable observation. However, the advertised feature of arbitrary client participation is not covered by Theorem 4.1, and the empirical comparisons do not include the most relevant asynchronous or adaptive baselines. The core theoretical result is narrower than the paper's framing suggests.","major_comments":[{"comment":"The abstract and Section 1 advertise 'arbitrary client participation' as a central feature, but Theorem 4.1 requires that 'each client is included in S_t with probability A/B uniformly and independently.' This is a specific sampling model, not arbitrary availability. Remark 4.1.2 concedes that without this assumption the rate degrades to O(sqrt(1/(AET))) + O(E^2/T) + O(tau^2/T) + Omega(sigma_g^2), and it constructs a two-client example in which a never-participating client makes the Omega(sigma_g^2) term unavoidable. Consequently, the paper's headline claim of converging with the best-known rate under arbitrary participation is unsupported; the theorem covers only uniformly random participation. This mismatch between the advertised scope and the proven guarantee is load-bearing and should be reconciled, either by restricting the claims or by presenting the arbitrary-participation result as a convergence-to-neighborhood guarantee.","section":"Theorem 4.1 and Remark 4.1.2"},{"comment":"The statement of Theorem 4.1 says clients are included 'independently,' but Algorithm 2 uses a fixed-size buffer of A clients and the proof in Appendix B uses without-replacement probabilities P{i,j in S_t} = A(A-1)/(B(B-1)). These two models are inconsistent: independent Bernoulli inclusion with probability A/B would give P{i,j in S_t} = A^2/B^2. Since the proof's cancellation step relies on the without-replacement probability, the formal assumption in Theorem 4.1 should be corrected to state without-replacement sampling, or the proof and algorithm should be adjusted to match the stated independence assumption.","section":"Appendix B, Theorem 4.1 proof"},{"comment":"The empirical evaluation compares only against CC-FedSGD and CC-FedAvg, which are variants of the proposed framework rather than established asynchronous or adaptive baselines such as FedAsync, FedBuff, AFL, or FedAdam. The claim that the approaches 'consistently outperform the baseline by a large margin' is therefore not yet supported against the most relevant prior work. Additionally, no error bars or multiple-seed results are reported, so the stability of the observed margins is unknown. The experiments should be extended with the relevant baselines and repeated-seed variability to substantiate the empirical contribution.","section":"Section 5.2, Tables 1-2, Figures 1-3"}],"minor_comments":[{"comment":"Lines 11-14 in Algorithm 2 mix the update rules for AdaGrad and Adam/AMSGrad without a clear conditional structure; the pseudocode would be clearer if it explicitly showed three separate variants or labeled each line with the corresponding optimizer.","section":"Algorithm 2"},{"comment":"The condition 'T >= A E^5' and tau <= (T/(A E))^{1/4} in Corollary 4.1.1 is stated without intuition; adding a sentence explaining how these regimes relate to the O(tau^2/T) and O(E^2/T) terms would improve readability.","section":"Corollary 4.1.1"},{"comment":"The phrase 'arbitrary client participation' is used in multiple places (e.g., Section 1 bullet list and Section 3.2) before the uniform-participation assumption of Theorem 4.1 is introduced; this creates a misleading first impression of the theoretical scope. Please qualify the claim in the introduction.","section":"Section 1 and 3"},{"comment":"The proof uses the notation 1/E_t both as the reciprocal of an average and in sums over clients; the definition in footnote 5 is helpful but the notation is easy to confuse with a simple reciprocal. A distinct symbol, e.g., r_t, would be clearer.","section":"Appendix B, B.1"},{"comment":"Some references are incomplete or inconsistent (e.g., the ACM reference format on the first page shows a placeholder year and DOI). Please ensure all bibliographic entries are complete.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's theoretical proof is substantial and appears self-contained, but the central advertised feature—arbitrary client participation—is not what the main theorem proves. This is not a fatal flaw because the uniform-participation result is still valuable, but the authors should be required to rewrite the abstract and introduction to align with the actual assumptions, and to reposition the arbitrary-participation result as a convergence-to-neighborhood bound. The experimental section would also benefit from a comparison with the cited asynchronous baselines before the current empirical claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on arXiv:2501.09946. The paper combines buffered asynchronous aggregation, server-side adaptive optimization (Adam/AMSGrad/Adagrad), and normalized heterogeneous local epoch updates into one FL framework, and proves a convergence rate for nonconvex objectives. The combination is genuinely new—FedAdam/FedAMS are synchronous, FedBuff/AFL don't do adaptive server updates, FedNova is synchronous—and the proof is a self-contained derivation from standard assumptions. That's real work. The O(sqrt(1/(AET))) rate matching the best-known async rate is a credible result if the theorem's conditions hold. So the paper deserves a serious referee and would be citable in FL theory.\n\nThe soft spots are in the gap between the advertising and the theorem. The abstract and intro list 'arbitrary client participation' as a core feature, but Theorem 4.1 assumes each client is included with probability A/B uniformly and independently. Remark 4.1.2 then concedes that without this, the rate degrades to O(sqrt(1/(AET))) + O(E^2/T) + O(tau^2/T) + Omega(sigma_g^2), and the two-client example shows it's unavoidable. So the headline claim of arbitrary participation with best-known rate isn't actually established. The uniform-participation case is what the theory covers; the arbitrary case gets only convergence to a neighborhood. That's a substantial caveat, though the authors are transparent about it in the remark.\n\nA secondary issue: the assumption says 'independently,' but the proof uses the without-replacement probability A(A-1)/(B(B-1)) for pairs, which corresponds to Algorithm 2's fixed-size A buffer. So the formal statement is imprecise, even though the proof itself uses the right probability.\n\nThe experiments are the weakest part. No error bars or multiple seeds, and the baselines are CC-FedSGD and CC-FedAvg—not FedBuff, AFL, or any other async adaptive method. The reported margins are large, but without variance or the most relevant comparisons, the empirical claims are only suggestive. The hyperparameter sensitivity analysis is a plus, and the code is linked.\n\nIn sum: the theory is solid conditional on uniform random participation; the paper overclaims arbitrary participation in its framing; experiments need strengthening. This is a conditional accept, not a desk reject. I'd send it to a referee who knows the async FL and adaptive optimizer literature. For my own work, I'd cite it for the combination result but not for the arbitrary-participation claim.","headline":"Genuinely novel combination of buffered async aggregation with server-side adaptivity, but the 'arbitrary participation' claim in the abstract is not backed by the main theorem, which assumes uniform random participation.","tokens_in":67833,"tokens_out":2072,"would_cite":true,"duration_ms":21647,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A client-centric federated learning framework achieves the best-known convergence rate for nonconvex objectives while tolerating asynchronous participation and heterogeneous local compute.","keywords":["federated learning","client-centric federated learning","adaptive optimization","asynchronous aggregation","system heterogeneity","client drift","nonconvex convergence","AMSGrad"],"falsifier":"Construct the two-client example in Remark 4.1.2 (f(x) = (1/2)[(x+K)^2 + (x-K)^2]) and run the algorithm with client 1 never participating; the paper's lower-bound argument predicts any algorithm is stuck with E[||grad f||^2] = $\\Omega$($sigma_g^{2}$), so if the method converges past that floor the participation model is not the deciding factor. Alternatively, measure the average squared gradient norm on a small nonconvex problem under uniform independent participation with probability A/B and bounded delay; the rate should follow O(1/$\\sqrt$(T)) for large T.","tokens_in":66790,"feed_emoji":"📡","tokens_out":8619,"duration_ms":83129,"temperature":0.7,"pith_summary":"This paper proposes Client-Centric Federated Adaptive Optimization (CC-FedAdagrad, CC-FedAdam, CC-FedAMS), a class of federated learning algorithms in which clients decide whether to participate, choose a device-dependent number of local steps, and send updates asynchronously to a server that updates the global model as soon as a fixed number of responses arrive. The paper's central claim is that despite this asynchronous, heterogeneous schedule, the algorithms converge for general nonconvex objectives at the rate O($\\sqrt$(1/(AET))) + O($E^{2}$/T) + O($tau^{2}$/T), where A is the number of participating clients per round, E the local epochs, and tau the maximum delay; for long runs with moderate delay this reduces to O($\\sqrt$(1/(AET))), matching the best-known rate in asynchronous federated learning. If correct, this means real-world system heterogeneity—stragglers, variable local compute, outdated global-model views—need not degrade the worst-case convergence rate, and server-side adaptive optimization can regularize client drift while preserving linear speedup in A and E. The paper reports experiments on CIFAR-10, CIFAR-100, Fashion-MNIST, and StackOverflow where the proposed methods outperform FedAvg-based baselines by large margins. The guarantee rests on the assumption that each client is included in each round uniformly at random with probability A/B, which is not the fully arbitrary participation the introduction emphasizes.","feed_headline":"Tolerating stragglers without losing federated learning's best rate","feed_subtitle":"A new algorithm class lets clients pick when to join and still match the best-known convergence rate.","key_machinery":"The central object is the server-side adaptive update rule: the server maintains a momentum buffer m_t = (1-$\\beta$)Delta_t + $\\beta$ m_{t-1} and a non-decreasing second-moment estimate v_hat_t (AMSGrad-style), and updates x_{t+1} = x_t - eta m_t/$\\sqrt$(v_hat_t + epsilon), where Delta_t is the average of client model differences normalized by their local step counts. Three supporting mechanisms carry the argument: (1) the normalized model update $\\Delta$ = (x_tau - x_local)/E, which stops clients running more local epochs from biasing the aggregate; (2) a size-A client buffer, so the server updates once A responses arrive and no client waits for stragglers, with the maximum delay tau bounded; and (3) a Lyapunov sequence z_t = x_t + ($\\beta$/(1-$\\beta$))(x_t - x_{t-1}) that re-centers the momentum in the proof, letting the adaptive step be analyzed as a stochastic descent of the global loss.","core_discovery":"The core claim is that server-side adaptive optimization—treating the averaged, normalized client updates as a pseudo-gradient and updating the global model with an Adagrad/Adam/AMSGrad-style rule—is compatible with client-centric system heterogeneity. Theorem 4.1 and Corollary 4.1.1 establish that under smoothness, bounded gradients, bounded local and global variance, bounded delay tau, and uniform independent participation with probability A/B, the average squared gradient norm is at most O($\\sqrt$(1/(AET))) + O($E^{2}$/T) + O($tau^{2}$/T), with the first term dominating for large T and moderate tau. This matches the best-known convergence rate in asynchronous federated learning and gives linear speedup in both the number of participating clients A and the number of local epochs E. The proof uses a normalized model update to debias clients with more local steps, a size-A buffer to bound the delay, and a Lyapunov sequence that absorbs the server-side momentum so the adaptive step becomes a controlled descent step. The authors note that for fully arbitrary participation, an unavoidable $\\Omega$($sigma_g^{2}$) term appears, so the advertised 'clients participate whenever they want' is analyzed only under a specific uniform random participation pattern.","pith_inferences":["The uniform random participation assumption is doing real work: outside the theorem, the authors' own Remark 4.1.2 shows an unavoidable Omega(sigma_g^2) floor, so practitioners using the algorithm under participation patterns determined by battery or network conditions should expect that floor to appear unless they correct for the bias.","The normalized model update is the ingredient that makes heterogeneous local epochs benign; a natural ablation comparing CC-FedAdam with and without normalization under identical conditions would isolate how much of the reported accuracy gain comes from this debiasing rather than from adaptivity.","As the delay tau goes to zero and the buffer size A grows to the full client set, CC-FedAdam should reduce to the synchronous FedAdam of prior work; verifying that final accuracy tracks this limit would confirm that the gains come from adaptivity rather than from the asynchronous schedule alone."],"forward_implications":["If the convergence guarantee is correct, federated systems can let clients choose their own participation and local compute without sacrificing the O(sqrt(1/(AET))) rate, as long as participation is uniform random with probability A/B and delays are bounded.","The linear speedup in A and E means that recruiting more clients per round or allowing more local epochs directly shortens the number of rounds needed to reach a target accuracy, preserving the communication-efficiency advantage of FedAvg.","The O(tau^2/T) term vanishes asymptotically when the maximum delay is moderate (tau <= (T/(AE))^{1/4}), so stale global-model views do not permanently hurt convergence.","Empirical results across four benchmarks show accuracy gains of roughly 10-30% over FedAvg-based baselines under the same asynchronous and heterogeneous schedule, with gains largest on tasks with sparse features such as StackOverflow.","Because the framework treats any adaptive optimizer as a plug-in server module, the same convergence analysis is claimed to extend to other adaptive optimizers beyond the three instantiations shown."],"supporting_citations":[{"why":"Supplies the FedAvg baseline and the 'computation then aggregation' template that the client-centric framework generalizes.","marker":"[51]"},{"why":"Introduces server-side adaptive federated optimization (FedAdam and variants), which this paper extends to asynchronous and heterogeneous settings.","marker":"[58]"},{"why":"Provides the AMSGrad max-stabilization step used in the CC-FedAMS update rule.","marker":"[59]"},{"why":"Establishes the prior asynchronous nonconvex analysis whose best-known rate this work claims to match.","marker":"[92]"},{"why":"Proposes the buffered asynchronous aggregation scheme (update when A responses arrive) that the framework adopts.","marker":"[54]"},{"why":"Sources the normalized model update that debiases clients with differing numbers of local epochs.","marker":"[77]"},{"why":"Early asynchronous federated optimization algorithm whose convex-only analysis and single-client privacy concern motivate the buffer design here.","marker":"[88]"},{"why":"Analyzes arbitrary device unavailability but assumes synchronous aggregation and a Lipschitz Hessian, serving as the comparison point for the relaxation here.","marker":"[22]"}],"fun_headline_variants":["Client-centric adaptive FL matches best rate with stragglers","Adaptive server updates keep FL's best rate despite client delays","Federated adaptive optimizer tolerates random client joins","Client-centric adaptive steps nail best-known FL convergence rate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorem assumes each client is included in the round's update set uniformly at random with probability A/B, independently; the paper's motivating promise that clients self-determine participation is not what the guarantee actually covers, and the authors concede an unavoidable error floor under fully arbitrary participation in Remark 4.1.2.","fun_headline_variants_meta":{"raw":{"variants":["Client-centric adaptive FL matches best rate with stragglers","Adaptive server updates keep FL's best rate despite client delays","Federated adaptive optimizer tolerates random client joins","Client-centric adaptive steps nail best-known FL convergence rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000675,"raw_usage":{"total_tokens":3072,"prompt_tokens":947,"completion_tokens":2125,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":2059}},"tokens_in":563,"tokens_out":2125,"duration_ms":16436,"temperature":1.0,"reasoning_tokens":2059,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:30:33.655787+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct the two-client example in Remark 4.1.2 (f(x) = (1/2)[(x+K)^2 + (x-K)^2]) and run the algorithm with client 1 never participating; the paper's lower-bound argument predicts any algorithm is stuck with E[||grad f||^2] = $\\Omega$($sigma_g^{2}$), so if the method converges past that floor the participation model is not the deciding factor. Alternatively, measure the average squared gradient norm on a small nonconvex problem under uniform independent participation with probability A/B and bounded delay; the rate should follow O(1/$\\sqrt$(T)) for large T.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the FedAvg baseline and the 'computation then aggregation' template that the client-centric framework generalizes."},{"cited_title":"Rabbat, Mani Malek, and Dzmitry Huba","cited_arxiv_id":null,"evidence_quote":"Introduces server-side adaptive federated optimization (FedAdam and variants), which this paper extends to asynchronous and heterogeneous settings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the AMSGrad max-stabilization step used in the CC-FedAMS update rule."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the prior asynchronous nonconvex analysis whose best-known rate this work claims to match."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proposes the buffered asynchronous aggregation scheme (update when A responses arrive) that the framework adopts."},{"cited_title":"Vincent Poor","cited_arxiv_id":null,"evidence_quote":"Sources the normalized model update that debiases clients with differing numbers of local epochs."}],"review_version":1}