{"id":"5aade5c4-e2bd-4cd2-ae5f-95bc15b774bb","arxiv_id":"2412.20785","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An adaptive element-wise quantization scheme and power allocation method for federated learning over cell-free massive MIMO improves test accuracy by up to 7-19% under equal energy and latency budgets.","lead":"This paper proposes an energy-saving way to train federated learning models over cell-free wireless networks by compressing each client's model update with an adaptive quantization scheme and optimizing transmit powers. A smart generalist might read it because it quantifies how much accuracy can be gained under fixed energy and latency budgets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convergence proof for the adaptive local-iteration rule uses gradient monotonicity without convexity (Eq. 41), so Proposition 1 and Remark 1 are unsupported; Eq. (33) also does not establish convergence.","rationale":"The reader's weakest-assumption identification is correct and is the most load-bearing issue. The central claim is that EMQ with adaptive local iterations and power allocation improves test accuracy within fixed energy and latency budgets. Unlike the missing hyperparameters or the somewhat opaque headline numbers in Tables II and III, the invalid monotonicity step in Appendix C directly undermines the adaptive local-iteration rule, which is one of the paper's stated contributions and is used in Algorithm 1 and all simulations. Without Proposition 1, Remark 1's stopping rule is heuristic rather than optimal, and the claimed vanishing quantization error and the computational-savings statements are unsupported. Proposition 2's Eq. (33) is also not a convergence bound, and the step-size dependence on the final unknown K compounds the problem. The concrete counterexample test settles whether the proof concern lands: L-smoothness alone does not give gradient monotonicity, so the proof is invalid as written unless a missing assumption is supplied. This does not prove the empirical results are wrong, but it confirms that the theoretical support for the framework's adaptive mechanism is missing, which is exactly why a CONDITIONAL verdict is appropriate. Since the reader already assigned CONDITIONAL, no verdict change is needed.","tokens_in":28854,"tokens_out":13452,"duration_ms":134455,"concrete_test":"Independently re-derive Eq. (41) of Appendix C from Assumptions 1-5 without adding convexity. A decisive check is to evaluate (f'(y)-f'(x))(y-x) for the L-smooth nonconvex function f(x) = -x^2 + x^4 on [-1,1] at x = 0, y = 0.2, which is negative, showing that Eq. (41) does not follow from L-smoothness alone. If this holds, the proof of Proposition 1 must be revised or an explicit convexity/monotonicity assumption must be added; the numerical claims should then be re-run under the corrected stopping rule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The framework's claimed acceleration depends on the adaptive local-iteration rule of Remark 1 and Algorithm 1, which is justified by Proposition 1. The proof of Proposition 1 in Appendix C invokes, at Eq. (41), the gradient monotonicity inequality (g_l - g_{l-1})^T (w_l - w_{l-1}) >= 0 for every client and local iteration. For an L-smooth but nonconvex loss, this inequality is not implied by Assumption 2; without convexity (or an explicit monotone-gradient assumption) it can fail. The CIFAR-10 CNN used in Section V is nonconvex, and no convexity assumption is stated. If Eq. (41) fails, the chain of inequalities through Eq. (45) collapses, and Proposition 1, the claimed non-increasing quantization exponent u_k, and the 'optimal' stopping rule in Remark 1 are unsupported. Separately, Proposition 2's final bound, Eq. (33), is not a convergence statement: it only says E f(w_{k+1}) <= f(w_k) + O(1/(k+1)), which upper-bounds the per-step expected increase and permits non-convergence; no summable gradient-norm inequality is derived. The step-size conditions in (32) also depend on the total number of global iterations K, which is itself an output of the energy/latency budgets, not a known constant when alpha is chosen. These gaps do not falsify the numerical comparisons, but they remove the theoretical basis for the adaptive iteration-count mechanism that is part of the headline framework.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EFCAQ, a federated learning framework over cell-free massive MIMO networks in which clients use local AdaDelta updates with an adaptive number of local iterations, quantize local updates with the proposed exponent-mantissa quantization (EMQ) scheme, and transmit under uplink power coefficients obtained from an SQP-based solution of a weighted energy-latency optimization problem. The authors provide a quantization error analysis, a convergence analysis of FedAvg with local AdaDelta updates, and CIFAR-10 simulations comparing against Dinkelbach and max-sum rate power allocation and AQUILA/LAQ quantization baselines under equal energy and latency budgets.","tokens_in":29253,"tokens_out":7915,"duration_ms":75311,"significance":"If the numerical results are taken at face value, the paper demonstrates a practically useful combination of techniques: adaptive element-wise EMQ quantization together with per-iteration power allocation increases both test accuracy and the number of completed global iterations under identical energy and latency budgets. The system model is detailed, the comparisons in Tables II and III are extensive across IID/non-IID settings and multiple budgets, and the reported gains are consistent with the proposed mechanisms. The paper also contributes an explicit quantization error analysis and a first analysis of FedAvg with local AdaDelta updates, though the theoretical parts currently contain gaps that need to be fixed before the convergence claims can be accepted.","major_comments":[{"comment":"Proposition 2's bound (33) is not a convergence statement. The inequality E_k{f(w_{k+1})} ≤ f(w_k) + [1+G^2+σ_k^2+(1-ρ^{L-1})^{-1}]/(k+1) only bounds the expected one-step increase by O(1/k); a sequence such as f(w_k)=log k satisfies this bound while diverging. The sentence after (33), stating that E{f(w_{k+1})}-f(w_k)→0, only shows that the per-step increments vanish, not that f(w_k) converges or that the gradient norms are summable. No descent lemma or gradient-norm bound is derived, so the claimed convergence rate of FedAvg with AdaDelta is unsupported.","section":"§IV, Eq. (33)"},{"comment":"The step-size condition in Eq. (32) depends on the total number of global iterations K, both through α1 and α2. However, K is not known when training begins; it is computed at run time from the energy and latency budgets via Eqs. (29)-(31). Thus the theorem cannot be applied a priori, and the step-size constraint is not verifiable before the training trajectory is known. This is a load-bearing gap because the convergence argument relies on choosing α according to this condition.","section":"§IV, Eq. (32)"},{"comment":"The proof of Proposition 1 invokes the monotone mapping inequality (g_l - g_{l-1})^T (w_l - w_{l-1}) ≥ 0 at Eq. (41), citing reference [36], a convex optimization textbook. This inequality holds for convex functions or monotone gradients, but the paper only assumes L-smoothness (Assumption 2) and uses a nonconvex CNN in Section V. For L-smooth nonconvex losses the inequality is generally false, so the chain of inequalities (41)-(45) collapses. This invalidates Proposition 1 and consequently Remark 1 and the stopping rule in Algorithm 1, which are part of the claimed acceleration mechanism.","section":"Appendix C, Eq. (41)"},{"comment":"Lemma 1's error bound ∥ε_j^k∥∞ ≤ 5·10^{u_j^k-1} assumes round-to-nearest with error at most 0.5. But the EMQ scheme encodes only mantissa values in {0,1,...,9} (Fig. 2), while the mantissa interval is [0,10). A value such as 9.7 has nearest integer 10, which is not representable, so the actual rounding rule must saturate or truncate, and the error can exceed 0.5 near the upper boundary. The quantization error analysis needs an explicit boundary condition or a modified rounding rule for this gap to be closed.","section":"Appendix A, Eq. (34)"},{"comment":"Remark 2 asserts without proof that solving sub-problems P1 and P2 at each iteration yields a sub-optimal solution to the original problem (13). No argument is provided that the per-iteration decomposition preserves feasibility of the global energy and latency budgets, or that the online policy approximates the solution of the offline joint problem. Moreover, Proposition 2 analyzes unquantized FedAvg with AdaDelta, not the full EFCAQ algorithm with EMQ; Lemma 2's error bound is not integrated into the convergence analysis. The theoretical link between the proposed framework and the convergence result is therefore missing.","section":"Remark 2"}],"minor_comments":[{"comment":"The text refers to 'Figs. 5(c), and 5(c)' and later 'Figs. 6(c), and 6(c)', but the intended references are to panels (c) and (d) of the respective figures; please correct the duplicated labels.","section":"Section V-B"},{"comment":"Remark 1 states that Proposition 1 demonstrates 'a non-increasing sequence of l_j^k w.r.t. k' leads to a non-increasing sequence of u_j^k, but Proposition 1 and the rule in (15) use a non-decreasing sequence of l_j^k. This is a typo that should be fixed because the direction of monotonicity is central to the adaptive stopping rule.","section":"Remark 1"},{"comment":"Algorithm 1 uses l_j^{k-1} in line 11, but for k=1 no initial value l_j^0 is defined in the inputs. Please specify the initialization of l_j^0 for the first global iteration.","section":"Algorithm 1"},{"comment":"The text says that rows corresponding to Algorithm 2 are 'highlighted', but the tables as typeset contain no highlighting; please add a visual marker or remove the reference to highlighting.","section":"Tables II and III"},{"comment":"The numerical tables report single accuracy values without standard deviations or the number of random seeds; since the CNN training is stochastic, please add error bars or at least state the number of independent runs used for the reported numbers.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the journal and the numerical comparisons are concrete and useful. I recommend major revision rather than rejection because the central engineering claims are defensible and the theoretical gaps appear fixable with a corrected convergence analysis and an explicit convexity or monotone-gradient assumption for Proposition 1. The main risk is that the adaptive local-iteration mechanism, which is part of the headline framework, currently rests on an invalid proof step; please require the authors to either repair the proof or clearly label the rule as a heuristic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the real contribution is EMQ, an element-wise exponent-mantissa quantization with a common exponent and adaptive mantissa bit lengths, plus the integration with AdaDelta local updates in a cell-free MIMO FL setup. The numerical comparison is solid enough that I would want to see this go to peer review. The convergence theory, however, is the soft spot and it is a real one.\n\nWhat the paper does well: the quantization scheme is concrete, easy to implement, and the bit budget formula is transparent. The tables show consistent accuracy gains over AQUILA and LAQ and over the Dinkelbach and max-sum-rate power allocation baselines, within the same energy and latency budgets. I checked the headline numbers: the up to 7%, 19%, and 36% / 35% claims match the tables. The authors also report the main simulation parameters and the CNN architecture, which makes reproduction feasible in principle.\n\nThe soft spots are in Appendix C and D, and they are not cosmetic. Proposition 1's proof uses the monotone mapping inequality at Eq. (41), which requires convexity or monotone gradients, but only L-smoothness is assumed. The CIFAR-10 CNN is nonconvex, so that inequality is not justified. Everything downstream - the non-increasing exponent claim and the stopping rule in Remark 1, which is called 'optimal' but is a heuristic - rests on that step. Separately, Proposition 2's Eq. (33) bounds a one-step expected increase by a term that decays as O(1/k), but that does not imply convergence of the iterates or of gradients; no summable gradient-norm bound is derived. The step-size condition also depends on the total number of iterations K, which is not known when training starts. These gaps do not invalidate the numerical comparisons, but they do mean the theoretical claims need to be substantially revised or removed.\n\nWho is it for: researchers working on communication-efficient FL and on joint communication-computation optimization in cell-free MIMO. They will get a useful new compression scheme and a clean system model, and they should read the convergence section with care.\n\nRecommendation: send to peer review, but the referee should insist that either the convergence analysis is fixed (with explicit convexity or monotone gradient assumptions) or the claims are downgraded to empirical observations. Also ask for seeds/hyperparameters or code. The paper is not junk; it is a competent engineering paper with an overreaching theory section.","headline":"EMQ is a concrete, useful compression scheme with solid numerical support, but the convergence proofs are not there yet; the paper deserves review on the engineering for a major revision.","tokens_in":29743,"tokens_out":2222,"would_cite":true,"duration_ms":23006,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68W15","90C30","90C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"An element-wise exponent-mantissa quantization scheme for cell-free massive MIMO federated learning, combined with adaptive local iterations and uplink power allocation, is claimed to raise test accuracy by up to 36% over…","keywords":["Federated learning","Cell-free massive MIMO networks","Adaptive quantization","Exponent-mantissa quantization","Power allocation","Energy efficiency","AdaDelta","Straggler effect"],"falsifier":"Track, on a non-convex CNN such as the paper's CIFAR-10 setup, the per-client exponent u_{l,k}^j = floor(log10 ||Δw_{l,k}^j||_infinity) through local AdaDelta iterations: if any client's exponent increases between consecutive checked local iterations, or exceeds the previous global iteration's exponent while the adaptive stopping rule is active, then the monotonicity claim that carries Proposition 1 and Remark 1 is empirically false.","tokens_in":28622,"feed_emoji":"📶","tokens_out":7176,"duration_ms":71178,"temperature":0.7,"pith_summary":"This paper tries to establish that federated learning over cell-free massive MIMO can be made substantially cheaper in communication without losing accuracy by exploiting the sparsity of local gradient updates. The proposed framework combines exponent-mantissa quantization (EMQ), which spends more bits on large gradient entries and very few bits on near-zero entries, with an adaptive number of local iterations per client and an uplink power-allocation scheme. The central quantitative claim is that, within the same energy and latency budgets, the power-allocation scheme improves test accuracy over Dinkelbach and max-sum-rate allocation by up to 7% and 19%, respectively, and that EMQ improves over AQUILA and LAQ by up to 36% and 35% under all three power-allocation methods. A sympathetic reader would care because communication, not computation, is usually the bottleneck in wireless federated learning; if the claim holds, many more useful global iterations fit into the same wireless budget, which is what drives the reported accuracy gains.","feed_headline":"Adaptive bit scheme lifts FL accuracy by up to 36 percent","feed_subtitle":"Exponent-mantissa encoding plus power control packs more global model rounds into the same energy and latency budget.","key_machinery":"The central object is EMQ (Exponent-Mantissa Quantization), an element-wise quantization scheme that writes every entry of an update vector as sign times mantissa times 10^exponent, with a single vector-wide exponent u_j^k = floor(log10 ||δw_j^k||_infinity) and mantissas rounded to the nearest integer in {0,...,9} encoded by a prefix-free bit table. This encoding concentrates bits on the few large gradient entries while spending almost nothing on the many near-zero entries, and Lemma 1 bounds the induced quantization error by roughly 5*$10^{{u_j^k-1}}$ per element. The second load-bearing piece is the adaptive local-iteration rule from Proposition 1 and Remark 1: clients run local AdaDelta iterations until the update exponent stops increasing (u_j^k <= $u_j^{{k-1}}$), which jointly controls local effort and quantization error. The third piece is the per-iteration power-allocation subproblem P2, solved by SQP with BFGS Hessian approximation and Armijo line search, which converts the bit counts and channel conditions into uplink powers that minimize a weighted sum of straggler latency and energy consumption.","core_discovery":"The paper's central claim is that each element of a client's local model update can be quantized according to its own magnitude — one shared exponent for the whole vector, one sign bit per entry, and a small prefix-free code for the rounded mantissa — yet the quantized training still converges to accuracy close to that of full-precision FedAvg with the maximum number of local iterations, while saving at least 49–56% of computation. The mechanism is that local AdaDelta updates make the order of magnitude of each update, u_j^k = floor(log10 ||δw_j^k||_infinity), non-increasing as training progresses, so an adaptive stopping rule can reduce the number of local iterations while also shrinking the quantization error. On top of this, a per-iteration sequential quadratic programming (SQP) power-allocation scheme minimizes a weighted sum of straggler latency and total uplink energy, and the remaining energy and latency budgets determine the number of global iterations K. Because EMQ shrinks the bit counts b_j^k, the same energy and latency budget buys more global iterations, which is the mechanism behind the reported test-accuracy gains of up to 7% and 19% over Dinkelbach and max-sum-rate power allocation, and up to 36% and 35% over AQUILA and LAQ quantization baselines.","pith_inferences":["The same exponent-mantissa encoding applies naturally to the downlink broadcast of the global model, not just uplink gradient updates, so the scheme could be extended to shrink both directions of communication with only minor changes to the bit-count accounting.","Because the gain mechanism is \"fewer bits per update buys more global iterations,\" the accuracy advantage over uniform-quantization baselines should grow with the model dimension d, since longer gradient vectors make the sparsity that EMQ exploits more pronounced.","If the prefix-free code in Fig. 2 were re-estimated online from the empirical distribution of rounded mantissas, the average bit count per update would likely drop further without changing the convergence analysis, since Lemma 1's error bound depends only on rounding to the nearest integer, not on the code's exact lengths.","The paper compares methods at the number of global iterations allowed by the same budget, so a reader should treat the reported percentage gains as combined gains in both accuracy and iteration count, not as accuracy gains at a fixed number of iterations."],"forward_implications":["Under equal energy and latency budgets, the proposed power-allocation scheme is claimed to beat Dinkelbach by up to 7% test accuracy and max-sum-rate allocation by up to 19%, purely by converting saved resources into more global iterations.","Across all three power-allocation methods, EMQ is claimed to raise test accuracy by up to 36% over AQUILA and up to 35% over LAQ, showing that the element-wise adaptive bit allocation itself, not the power control, is the main source of the quantization gain.","FedAvg+EMQ with adaptive local iterations reaches test accuracy comparable to full-precision FedAvg with the maximum number of local iterations, while saving at least 49% of computation in the IID case and 56% in the non-IID case.","FedAvg with local AdaDelta updates reaches about 78% test accuracy in fewer than 39 iterations in the paper's CIFAR-10 setup, roughly three times faster than local SGD, which underpins the benefit of pairing AdaDelta with adaptive quantization.","The convergence analysis claims that, for a sufficiently small step size alpha, the expected loss increment satisfies E{f(w_{k+1})} - f(w_k) <= O(1/(k+1)), so the quantized adaptive scheme converges in the sense that the expected per-round improvement vanishes as k grows."],"supporting_citations":[{"why":"Supplies AdaDelta, the per-dimension adaptive learning-rate local optimizer whose magnitude alignment with gradient entries drives the adaptive quantization story.","marker":"[27]"},{"why":"Provides the FedAvg algorithm whose convergence with local AdaDelta updates is analyzed in Proposition 2.","marker":"[28]"},{"why":"Supplies the user-centric cell-free massive MIMO channel model, pilot-based MMSE estimation, and the max-sum-rate power-allocation baseline.","marker":"[2]"},{"why":"Provides the achievable uplink rate expression (Th. 2) from which the SINR, latency, and energy formulas are constructed.","marker":"[30]"},{"why":"Provides AQUILA, the adaptive-quantization baseline that EMQ is compared against under equal energy and latency budgets.","marker":"[8]"},{"why":"Provides LAQ, the lazily aggregated quantized gradient baseline used in the same budget-constrained comparisons.","marker":"[10]"},{"why":"Provides the Dinkelbach-based energy-efficiency power-control method used as a power-allocation baseline.","marker":"[35]"},{"why":"Provides the SQP, BFGS, and Armijo line-search machinery used to solve the per-iteration power-allocation problem.","marker":"[31]"},{"why":"Supplies the CNN architecture and CIFAR-10 experimental setup on which all reported test-accuracy numbers are based.","marker":"[34]"},{"why":"Supplies the per-iteration decomposition strategy, solving the resource-allocation problem online without data from future iterations.","marker":"[17]"}],"fun_headline_variants":["Adaptive quantization and power control lift FL accuracy 36%","Exponent-mantissa encoding speeds up energy-efficient federated learning","Cell-free FL: SQP power and adaptive bits gain 36% over baselines","Smart bit allocation in federated learning cuts energy, boosts accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The adaptive stopping rule and the convergence bound rely on the assumption that consecutive local gradients point in the same general direction as training proceeds, a monotonicity property that is guaranteed for convex losses but not established for the general smooth, possibly non-convex loss functions the paper says it is analyzing.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive quantization and power control lift FL accuracy 36%","Exponent-mantissa encoding speeds up energy-efficient federated learning","Cell-free FL: SQP power and adaptive bits gain 36% over baselines","Smart bit allocation in federated learning cuts energy, boosts accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001022,"raw_usage":{"total_tokens":4376,"prompt_tokens":1079,"completion_tokens":3297,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":3229}},"tokens_in":695,"tokens_out":3297,"duration_ms":25337,"temperature":1.0,"reasoning_tokens":3229,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:11:43.725099+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track, on a non-convex CNN such as the paper's CIFAR-10 setup, the per-client exponent u_{l,k}^j = floor(log10 ||Δw_{l,k}^j||_infinity) through local AdaDelta iterations: if any client's exponent increases between consecutive checked local iterations, or exceeds the previous global iteration's exponent while the adaptive stopping rule is active, then the monotonicity claim that carries Proposition 1 and Remark 1 is empirically false.","supporting_citations":[{"cited_title":"Foundations of user-centric cell-free Massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Supplies the user-centric cell-free massive MIMO channel model, pilot-based MMSE estimation, and the max-sum-rate power-allocation baseline."},{"cited_title":"Cell-free Massive MIMO versus small cells,","cited_arxiv_id":null,"evidence_quote":"Provides the achievable uplink rate expression (Th. 2) from which the SINR, latency, and energy formulas are constructed."},{"cited_title":"AQUILA: Communication efficient Federated Learning with adaptive quantization in device selection strategy,","cited_arxiv_id":null,"evidence_quote":"Provides AQUILA, the adaptive-quantization baseline that EMQ is compared against under equal energy and latency budgets."},{"cited_title":"Lazily Aggregated Quantized Gradient (LAQ) innovation for communication-efficient Federated Learning,","cited_arxiv_id":null,"evidence_quote":"Provides LAQ, the lazily aggregated quantized gradient baseline used in the same budget-constrained comparisons."},{"cited_title":"Energy-efficient power control: A look at 5G wireless technologies,","cited_arxiv_id":null,"evidence_quote":"Provides the Dinkelbach-based energy-efficiency power-control method used as a power-allocation baseline."},{"cited_title":"Nocedal et al., Numerical optimization","cited_arxiv_id":null,"evidence_quote":"Provides the SQP, BFGS, and Armijo line-search machinery used to solve the per-iteration power-allocation problem."},{"cited_title":"Over-the-air federated learning over scalable cell- free massive mimo,","cited_arxiv_id":null,"evidence_quote":"Supplies the CNN architecture and CIFAR-10 experimental setup on which all reported test-accuracy numbers are based."},{"cited_title":"Fedcau: A proactive stop policy for communica- tion and computation efficient Federated Learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the per-iteration decomposition strategy, solving the resource-allocation problem online without data from future iterations."}],"review_version":1}