{"id":"e849dbe7-ee5d-4c66-950e-352790427cde","arxiv_id":"2507.15601","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Wireless federated learning can cut end-to-end training time by choosing per-device batch sizes with a closed-form rule that balances convergence rounds against per-round latency.","lead":"This paper derives a rule for setting the mini-batch size of every device in a wireless federated learning system so that training finishes as fast as possible. The rule balances cleaner gradients from larger batches against the longer per-round time those batches cause.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Closed-form B* is optimal for the fitted surrogate, but the surrogate's functional form comes from an upper-bound rearrangement and is validated only in-sample; with the optimum sitting near the law's pole (beta-hat/epsilon), unquantified surrogate error at B* can silently shift the claimed…","rationale":"I re-derived the key steps before settling on the concern. The Theorem 1 bound depends on the global batch size B only (the per-device split cancels in the variance and drift terms), so the two-stage decomposition—latency equilibrium per round, then global-B optimization—is internally consistent; the weak link is not the decomposition but the surrogate law N_epsilon(B) = alpha/(epsilon - beta/B) that carries the fitted parameters into Theorem 2. This is exactly the reader's weakest assumption, and I agree with it. I sharpened it in three ways: (1) the functional form is not validated as a law but inherited from an upper bound, so the two-parameter fit can be good inside the sampled range while the shape is wrong elsewhere; (2) the predicted optimum B_epsilon is proportional to the pole beta-hat/epsilon, placing B* in the region where the surrogate is most sensitive to parameter error—for the reported MNIST values the pole is at B=46.4, and the steep part of the curve extends well beyond it; (3) the reported validation (Figs. 2 and 7) is in-sample and qualitative, with no residuals or leave-one-out checks at B*. I also verified the boundary behavior: psi-tilde is continuous at B-tilde_th (the numerator identity HW*B-tilde_th + f-hat-sigma = f_sigma*tau_1b holds by construction), and the rounding of (34) can legitimately violate (20c); the paper's negligible claim has no bound. The appendix contains a W-versus-HW typo in the B_epsilon derivation, but my re-derivation confirms the HW version in the theorem statement, so this is cosmetic. I give the paper credit where it has independent support: the predicted optimum does align with the empirical minimum in Fig. 7 for the two tested settings, the O(K) closed-form controller is a genuine practical contribution, and the convergence analysis is standard and coherent. Those supports are real but narrow. They do not quantify how far B* can move under surrogate error, which is the load-bearing condition for the headline claim of E2E-optimality. Because the paper frames Theorem 2 as solving the surrogate-based problem and the experiments support the surrogate at the tested operating points, the honest verdict remains CONDITIONAL: the claim stands if a hold-out check at B* confirms the surrogate, and fails if it does not. My concern therefore does not change the reader's verdict, and I mark it UNCHANGED.","tokens_in":21153,"tokens_out":21520,"duration_ms":243193,"concrete_test":"Hold-out validation at the predicted optimum: for each dataset (MNIST, CIFAR-10) and for three system configurations drawn from Figs. 10-11 (e.g., BW = 1, 5, 10 MHz at fixed computation scale), refit alpha-hat and beta-hat using only interior sampled B values (exclude the largest and smallest from the Fig. 2-type data), then run FL to the target accuracy at B in {0.7B*, 0.85B*, B*, 1.15B*, 1.3B*} using the allocation rule (34), recording true N_epsilon(B) and E2E latency. Accept the central claim only if (i) B* is an empirical argmin of E2E latency within measurement uncertainty, and (ii) the surrogate-predicted N_epsilon(B) matches the measured value at B* within 20%.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The practical headline claim—that the batch sizes in Theorem 2 minimize E2E latency—is a statement about the surrogate-based problem, not about the true system. The proof in Appendix C establishes unimodality and the minimizer only for the continuous relaxation psi-tilde (Eq. 31) of psi (Eq. 30), and only because N_epsilon(B) is assumed to be exactly alpha-hat/(epsilon - beta-hat/B). The functional form of that law is not derived empirically: it is the rearrangement of the Theorem 1 upper bound (Eqs. 16-17). Least-squares fitting of two parameters (Eq. 18) can appear accurate inside the sampled B range (Fig. 2) even when the true rounds-to-target curve has a different shape, and the paper reports no residuals, confidence intervals, or hold-out checks.\n\nThe risk concentrates at the point where the optimum is computed. The surrogate has a pole at B = beta-hat/epsilon; measured N(B) is steepest there, and the closed-form optimum B_epsilon = (beta-hat/epsilon)(1 + sqrt(1 + f-hat-sigma*epsilon/(HW*beta-hat))) is proportional to that pole. For the reported MNIST parameters (alpha-hat=34.5, beta-hat=23.2, epsilon=0.5), the pole is at B=46.4, so B* lies in the steepest part of the fitted curve, where small errors in beta-hat or in the assumed 1/(epsilon-beta/B) shape translate directly into a wrong optimum and a wrong allocation (34). Fig. 7 compares the predicted optimum against an empirical E2E curve that shares the same fitted data, at one threshold per dataset; it is a consistency check, not an extrapolation test.\n\nTwo secondary gaps reinforce the conditionality. The rounded allocation in (34) can violate constraint (20c) (sum of round(x_k) need not equal B*), and negligible is asserted rather than bounded; this perturbs both per-round latency and the global batch size entering the convergence condition.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a communication-and-computation (C2)-aware batch-size control framework for synchronous federated learning over heterogeneous devices. The authors derive a non-convex convergence upper bound (Theorem 1), rearrange it into a round-batch scaling law N_epsilon(B) = alpha/(epsilon - beta/B) (Corollary 1), and fit alpha and beta by least squares to empirical convergence-round data on MNIST and CIFAR-10. They then formulate an E2E latency minimization problem, relax integrality, prove a latency-equilibrium lemma for per-device allocation, and derive a closed-form optimal global batch size (Theorem 2) for slow-fading channels. For fast-fading channels they present an adaptive protocol that recomputes a batch-size threshold from instantaneous communication latencies. Experiments compare the proposed scheme with fixed-batch baselines, DBFL, and a no-local-adaptation variant.","tokens_in":21561,"tokens_out":4948,"duration_ms":61376,"significance":"If the surrogate and the relaxations were fully justified, the paper would offer a practically attractive result: an O(K) closed-form batch-size control policy with an explicit C2 tradeoff and consistent experimental latency reductions. The convergence analysis follows a standard non-convex SGD template, and the latency-equilibrium lemma is sound for the continuous relaxation. The main contribution, however, is a surrogate-based optimization procedure rather than a proven optimal policy for the original MINLP: the optimality of Theorem 2 depends on an in-sample fitted functional form, and the integer-rounding gap is unquantified. These issues are load-bearing for the central claim, but they are addressable within the manuscript's scope.","major_comments":[{"comment":"The central optimality claim is established only for the surrogate psi-tilde(B) in (31), whose parameters hat-alpha and hat-beta are least-squares fitted to the same tasks used later for validation. The functional form N_epsilon(B) = alpha/(epsilon - beta/B) is a rearrangement of the Theorem 1 upper bound, but the actual parameter values are free, and Fig. 2 reports only an in-sample fit: no residuals, confidence intervals, or hold-out checks are given. Because B_epsilon in (55) is proportional to the pole hat-beta/epsilon, errors in the fit or in the assumed functional form translate directly into a biased optimum, and the pole lies in the steep region of the fitted curve for the reported MNIST parameters. The manuscript should either provide a robustness bound on B* under a bounded surrogate error or validate the scaling law out-of-sample over a grid that contains the predicted B*, with error bars.","section":"§IV-A, Eq. (18); §IV-C, Theorem 2"},{"comment":"Rounding each b*_{k,n} to the nearest integer can violate the global batch-sum constraint (20c), because sum_k round(x_k) need not equal B*. The claim that the induced performance gap is negligible rests on experimental observation rather than a mathematical bound, and Theorem 2's proof does not address the integer MINLP formulation of Problem (20). A feasibility-correction step or an explicit approximation bound is needed before the closed-form allocation can be called optimal for the original problem.","section":"§IV-C, Eq. (34)"},{"comment":"The adaptive protocol changes the global batch size B_n in every round, whereas Theorem 1 and Corollary 1 assume a fixed global batch size B_n = B for all rounds. Consequently, Eqs. (21), (31), and the optimality of B*_n = max(bar-B*, Bth_n) in (39) are not justified by the convergence analysis. The fast-fading extension may be a reasonable heuristic, but the introduction and conclusion present it as part of the optimal-control result; the paper should clearly label it as heuristic or provide a convergence analysis that covers time-varying batch sizes.","section":"§V-B, Algorithm 1"},{"comment":"Theorem 2 optimizes the continuous relaxation psi-tilde(B), not the exact integer objective psi(B) in (30) with its ceiling operation. Fig. 4 shows a close match for two displayed scenarios, but no general bound is given for the gap between the relaxed minimizer and the true integer minimizer of Problem (20). Since the original problem is MINLP and claimed to be NP-hard, the closed form requires either an exactness proof or a quantified integrality gap before it can be stated as the solution of Problem (20).","section":"§IV-C, Eqs. (30)-(33)"}],"minor_comments":[{"comment":"The definition of Bth uses a floor operator in (24), while the continuous threshold tilde-Bth in (31) is real-valued; the text should state explicitly whether the piecewise switch in Fig. 3 and the condition B > Bth refer to the integer threshold or its continuous relaxation, otherwise the two are easily conflated.","section":"§IV-C, Eq. (24) and surrounding text"},{"comment":"In the contradiction argument, the perturbed allocation tilde-b'_ki = tilde-b*_ki +/- delta may leave the feasible interval [1, B-K+1]; adding a short feasibility check would make the proof complete.","section":"Appendix B, proof of Lemma 1"},{"comment":"The convergence tolerance epsilon = 0.5 is set alongside validation-accuracy thresholds epsilon_A = 95% and 58%, but the relationship between the gradient-norm threshold epsilon in the analysis and the validation-accuracy proxy epsilon_A is never stated; this connection should be made explicit.","section":"§VI-A, experimental settings"},{"comment":"The symbol W denotes per-sample computational workload, which is natural, but the same letter is reused for the model dimension in the text around Eq. (2); using distinct symbols for workload and model dimension would avoid confusion.","section":"§II-B, Eq. (5)"},{"comment":"The DBFL baseline is described as having a 'stiff incremental factor', but the paper does not specify how the DBFL parameters are chosen in the experiments; a sentence describing the configuration would improve reproducibility.","section":"§VI-C, Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"The core issue is not the convergence proof but the gap between the fitted surrogate and the claimed optimality for the original integer problem. The paper would be much stronger if it were re-framed as a surrogate-based design method with explicit validation of the surrogate at and around the predicted optimum, and with a feasibility/rounding analysis for the device-level allocations. This is fixable within the current scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper gives a closed-form batch-size controller for slow-fading federated learning with heterogeneous devices, balancing convergence rounds against per-round latency, and an adaptive heuristic for fast fading. The convergence bound in Theorem 1 is the standard non-convex SGD template, the latency-equilibrium lemma is clean, and the closed-form expressions in Theorem 2 are genuinely new: prior work fixed the global batch size or ignored convergence rounds. The O(K) complexity replacing an NP-hard MINLP is a real practical win, and the experiments show consistent latency reductions on MNIST and CIFAR-10 against reasonable baselines.\n\nThe soft spot is where the stress-test points: the optimality of B* is optimality of the surrogate problem, not the true system. The scaling law N_epsilon(B) = alpha/(epsilon - beta/B) is a rearrangement of the Theorem 1 upper bound, and alpha and beta are least-squares fitted in-sample. The predicted optimum sits near the pole of that law, so small errors in beta-hat can shift B* noticeably. There are no residuals, confidence intervals, or hold-out checks, and the fast-fading variant is explicitly a heuristic. Integer rounding of the per-device allocation can also violate the sum constraint; the paper asserts \"negligible\" without a bound. None of this destroys the engineering value, but the word \"optimal\" needs to be read as \"optimal for the fitted surrogate.\"\n\nI would send this to a serious referee. The core idea is solid, the theory is mostly sound, and the limitations are addressable in revision. Push the authors for error bars, code/data release, an out-of-sample test of the surrogate, and some bound on the rounding gap. For my own work, I'd cite it as a representative low-complexity batch-control scheme, but not as the final word on optimality. Reading group: maybe; it would spark a useful discussion on surrogate-based optimization in systems papers.","headline":"A solid systems paper that turns an intractable MINLP into a closed-form batch-size rule, with the caveat that 'optimal' holds for the fitted surrogate, not the true learning curve.","tokens_in":22100,"tokens_out":3524,"would_cite":true,"duration_ms":36097,"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":"The paper derives a closed-form rule that chooses global and per-device batch sizes to minimize end-to-end federated learning latency while preserving convergence accuracy.","keywords":["federated learning","batch-size control","end-to-end latency","communication-computation tradeoff","device heterogeneity","convergence analysis","slow fading","fast fading"],"falsifier":"Run a dense sweep of global batch sizes on a fresh task, fit the two surrogate parameters using only a subset of low batch sizes, compute the closed-form $B^{*}$, and compare measured end-to-end latency at $B^{*}$ with latencies at neighboring values; if any neighbor is faster, or if the fitted law misses measured convergence rounds by more than the data noise near $B^{*}$, the optimality claim fails.","tokens_in":20955,"feed_emoji":"⚙️","tokens_out":10921,"duration_ms":112229,"temperature":0.7,"pith_summary":"This paper tries to show that the wall-clock time needed to train a federated model over heterogeneous wireless devices can be minimized by a closed-form batch-size rule, without solving the NP-hard mixed-integer program that the problem first appears to be. The argument rests on a communication-and-computation tradeoff: larger global batches sharpen gradient estimates and reduce the number of communication rounds, but they lengthen every round. The authors derive a round-batch scaling law from convergence analysis, fit its two parameters to real training data, and then prove that the optimal global batch size is the larger of a threshold and a simple closed-form point, with per-device batches given by an explicit linear allocation. Experiments on MNIST and CIFAR-10 report that this controller reaches the same accuracy thresholds faster than fixed, uniform, and exponentially growing batch-size schemes, with reported latency reductions of 26.7% and 37.5% over the no-local-adaptation baseline.","feed_headline":"A closed-form formula finds the optimal batch size in federated learning","feed_subtitle":"Per-device batch allocation balances communication and computation to cut end-to-end training time.","key_machinery":"The load-bearing object is the round-batch scaling law $N_{\\epsilon}(B) = \\hat{\\alpha}/(\\epsilon - \\hat{\\beta}/B)$, a two-parameter surrogate for the number of communication rounds required to reach a target accuracy as a function of the global batch size. It is fit by least squares to measured convergence-round data, and it converts the intractable mixed-integer problem into a unimodal one-dimensional minimization. The second mechanism is the latency equilibrium principle: for a fixed global batch size, per-round latency is minimized when every device's computation-plus-communication time is equal, so the optimal local batch size is the linear allocation that equalizes device latencies. Together these generate the continuous surrogate $\\tilde{\\psi}(B)$ whose unique minimizer is $B_{\\epsilon}$, and the threshold $B_{\\mathrm{th}}$ below which no equalizing allocation exists.","core_discovery":"Under slow-fading channels, the paper's Theorem 2 states that the optimal global batch size is $B^{*} = \\max\\{B_{\\mathrm{th}}, \\lfloor B_{\\epsilon}\\rceil_{\\tilde{\\psi}(\\cdot)}\\}$, where $B_{\\epsilon} = \\frac{\\hat{\\beta}}{\\epsilon}\\left(1+\\sqrt{1+\\frac{\\hat{f}_{\\Sigma}\\epsilon}{H W \\hat{\\beta}}}\\right)$ and the ceiling-floor notation means the integer neighbor of $B_{\\epsilon}$ that gives the smaller surrogate latency. The matching device-level allocation is $b^{*}_{k,n} = \\mathrm{round}\\!\\left(\\frac{f_k}{H W}\\left(\\frac{H W B^{*} + \\hat{f}_{\\Sigma}}{f_{\\Sigma}} - T^{\\mathrm{cmm}}_k\\right)\\right)$, which drives every device to finish its round at the same time---the latency equilibrium principle that Lemma 1 proves optimal. The global formula balances the round saving from larger batches against the per-round computation they add, using the fitted scaling law $N_{\\epsilon}(B) = \\hat{\\alpha}/(\\epsilon - \\hat{\\beta}/B)$ as the surrogate for convergence speed. For fast fading, the same construction is evaluated at expected channel statistics and then raised to the instantaneous channel-aware threshold each round. The paper's central claim is that this two-part control is the latency minimizer under the fitted surrogate, and that the surrogate matches measured convergence behavior closely enough that the closed-form point lands on the empirical latency minimum.","pith_inferences":["The same surrogate-plus-equilibrium template should transfer to other round-level controls, such as gradient compression ratio, number of local epochs, or quantization bits, wherever a larger control value improves gradient quality but lengthens per-round time; the paper's communication-computation tradeoff is one instance of that family.","Because $\\hat{\\alpha}$ and $\\hat{\\beta}$ are fitted per task, a practical deployment would need a cheap online estimator of these two parameters from early training rounds rather than a full offline sweep; the paper does not develop this, but its own data-fitting step makes it a natural next step.","A natural extension is to treat the fast-fading rule as a stochastic online control problem and seek a regret guarantee relative to the clairvoyant channel-sequence optimum, which the paper does not provide."],"forward_implications":["With slow-fading channels, the closed-form $B^{*}$ and $b^{*}_{k,n}$ compute in $O(K)$ arithmetic, replacing an NP-hard MINLP with a rule that can run in real time.","With fast-fading channels, setting $B^{*}_n = \\max\\{\\bar{B}^{*}, B^{\\mathrm{th}}_n\\}$ each round keeps per-round latency at the straggler floor while using the extra samples to reduce the number of rounds, lowering E2E latency.","If bandwidth rises, the optimal global batch size falls; if device computation speed rises, the optimal global batch size grows, giving operators a direct knob for tuning the communication-computation tradeoff.","Reported results compare with fixed, uniform, and DBFL baselines, with the proposed adaptive control reaching the accuracy threshold with 26.7% (MNIST) and 37.5% (CIFAR-10) less E2E latency than the no-local-adaptation scheme."],"supporting_citations":[{"why":"Defines the synchronous FedAvg protocol whose weighted aggregation and local-update structure the paper builds on.","marker":"[4]"},{"why":"Provides the DBFL baseline, the bounded-gradient-variance assumption, and the convergence-bound template that Theorem 1 and Corollary 1 extend.","marker":"[21]"},{"why":"Supplies the convergence-rate metric (expected average gradient norm) and the wireless-FL straggler latency model used throughout.","marker":"[31]"},{"why":"Prior optimal batch allocation for wireless FL with a fixed global batch; the per-round latency minimization in Lemma 1 extends it to heterogeneous communication and computation.","marker":"[34]"},{"why":"The least-squares method used to fit the surrogate parameters $\\hat{\\alpha}$ and $\\hat{\\beta}$ in Eq. (18).","marker":"[41]"},{"why":"MNIST dataset used to validate the round-batch scaling law and compare latency.","marker":"[46]"},{"why":"CIFAR-10 dataset used for the larger ResNet-18 experiments.","marker":"[47]"},{"why":"ResNet-18 architecture used in the CIFAR-10 experiments.","marker":"[48]"}],"fun_headline_variants":["Optimal batch size formula minimizes FL latency","Per-device batch allocation cuts end-to-end FL latency","Closed-form batch size control for low-latency FL","Device-aware batch sizing balances comm and compute in FL","Slow and fast fading share one optimal batch formula"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire optimality result depends on the fitted curve that predicts how many training rounds are needed for each batch size staying correct for all feasible batch sizes, especially the one the formula picks; if the curve bends differently outside the sampled points, the claimed optimum is not guaranteed for the real system.","fun_headline_variants_meta":{"raw":{"variants":["Optimal batch size formula minimizes FL latency","Per-device batch allocation cuts end-to-end FL latency","Closed-form batch size control for low-latency FL","Device-aware batch sizing balances comm and compute in FL","Slow and fast fading share one optimal batch formula"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001097,"raw_usage":{"total_tokens":4684,"prompt_tokens":1155,"completion_tokens":3529,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":771,"completion_tokens_details":{"reasoning_tokens":3454}},"tokens_in":771,"tokens_out":3529,"duration_ms":30487,"temperature":1.0,"reasoning_tokens":3454,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:28:02.896316+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a dense sweep of global batch sizes on a fresh task, fit the two surrogate parameters using only a subset of low batch sizes, compute the closed-form $B^{*}$, and compare measured end-to-end latency at $B^{*}$ with latencies at neighboring values; if any neighbor is faster, or if the fitted law misses measured convergence rounds by more than the data noise near $B^{*}$, the optimality claim fails.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Defines the synchronous FedAvg protocol whose weighted aggregation and local-update structure the paper builds on."},{"cited_title":"To talk or to work: Dynamic batch sizes assisted time efficient federated learning over future mobile edge devices,","cited_arxiv_id":null,"evidence_quote":"Provides the DBFL baseline, the bounded-gradient-variance assumption, and the convergence-bound template that Theorem 1 and Corollary 1 extend."},{"cited_title":"Wirelessly powered federated edge learning: Optimal tradeoffs between convergence and power transfer,","cited_arxiv_id":null,"evidence_quote":"Supplies the convergence-rate metric (expected average gradient norm) and the wireless-FL straggler latency model used throughout."},{"cited_title":"Optimal batch allocation for wireless federated learning,","cited_arxiv_id":null,"evidence_quote":"Prior optimal batch allocation for wireless FL with a fixed global batch; the per-round latency minimization in Lemma 1 extends it to heterogeneous communication and computation."},{"cited_title":"A method for the solution of certain non-linear problems in least squares,","cited_arxiv_id":null,"evidence_quote":"The least-squares method used to fit the surrogate parameters $\\hat{\\alpha}$ and $\\hat{\\beta}$ in Eq. (18)."},{"cited_title":"Gradient-based learning applied to document recognition,","cited_arxiv_id":null,"evidence_quote":"MNIST dataset used to validate the round-batch scaling law and compare latency."},{"cited_title":"Deep residual learning for image recognition,","cited_arxiv_id":null,"evidence_quote":"ResNet-18 architecture used in the CIFAR-10 experiments."}],"review_version":1}