{"id":"db80e12c-7923-474d-a3c3-df36b8011e5a","arxiv_id":"2506.19476","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"NCDSFL fixes output and auxiliary layer weights to neural collapse solutions to align clients in federated OFDM signal detection, converging faster than FedAvg in simulation, but the optimality theorem is invalid as written.","lead":"Federated learning for wireless signal detection can slow down when different users see different channels. This paper fixes parts of each neural network to a 'neural collapse' structure to align users' hidden features and speed up convergence, with simulations showing faster training; the supporting theorem, however, has a proof flaw.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Load-bearing concern: Theorem II.1 governs the layer-peeled problem with free features and balanced labels, but the actual DNN/FL objective is not shown to satisfy these assumptions; without a transfer argument the fixed NC weights lack theoretical support.","rationale":"Stress-test pass. First, I re-derived the construction of A. With B_1 = [-I_d, I_d] and the recurrence in (8), B_i is i d x 2^i d and satisfies B_i B_i^T = 2^i I_{i d}; hence A = [B_I,...,B_I] is dI x dK2^I with A A^T = K2^I I, and the reader's stated dimension inconsistency does not materialize. The lower-bound argument in Lemma A.2 is consistent, and an equality-attaining configuration exists, so Theorem II.1 is not obviously false. The paper's real vulnerability is the transfer from the layer-peeled problem to the deployed architecture. The theorem needs free H, feature regularization, and balanced coverage of all 2^I labels; the NCDSFL objective has a fixed ReLU backbone, softmax CE, no explicit feature norm penalty, and the 16-bit-per-DNN OFDM labels are drawn randomly rather than balanced over 65,536 sequences. The cited universal approximation property only says some wide network can approximate a target function; it does not make intermediate features unconstrained in the loss landscape. Thus fixing WNC/eWNC on the basis of Theorem II.1 is an unsupported leap. The empirical demonstration is also too weak to close the gap: theta and vartheta are not shown to reach zero, no error bars are given, and there is no ablation severing the NC structure (e.g., random frozen orthogonal head). Because the central theoretical justification for NCDSFL is not connected to the actual training problem, the reject verdict stands, although it should rest on this transfer gap rather than on the alleged dimension error.","tokens_in":14321,"tokens_out":22839,"duration_ms":233517,"concrete_test":"Run the NCDSFL simulations identically, but replace the two fixed NC matrices WNC and eWNC by random orthogonal matrices of the same shapes (e.g., drawn from the Haar distribution, then frozen), keeping all hyperparameters unchanged; report convergence epochs and final BER over at least 10 seeds. If the random-head variant matches NCDSFL, the NC-optimality theorem is not doing the causal work and the central claim fails. If it is substantially worse, report theta and vartheta at convergence for the unfixed network to check whether the actual optimizer is close to an NC solution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Re-deriving the algebra in Appendix A did not reveal the dimension inconsistency claimed by the reader: B_i has i d rows and 2^i d columns, so A is dI x dK2^I and A A^T = K2^I I_{dI}; Lemma A.1 is consistent. The proof of Theorem II.1 can be completed by explicitly constructing equality-attaining points (orthogonal equal-norm class differences plus matching feature norm), so the theorem itself is not the weak point. The load-bearing gap is the step from problem (5) to the algorithm. Theorem II.1 requires H to be free variables regularized by lambda||H||^2 and K balanced samples for every one of the 2^I label sequences. The real objective (14) has a fixed backbone, uses softmax CE, and the OFDM experiments use I=16 bits per DNN with randomly drawn bit streams, so the balanced-label assumption is not met and the 500-250-128 ReLU networks are not shown to be wide enough for the universal-approximation remark to make H effectively unconstrained. Universal approximation of functions does not imply the intermediate features are free optimization variables or that the global minimizer of the constrained problem is an NC solution. Figures 3-4 show theta and vartheta decreasing but not converging to zero, with no error bars or comparison to a random fixed orthogonal head. Hence the observed FL speed-up is not causally tied to the NC theorem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes NCDSFL, a federated learning algorithm for signal detection in OFDM systems. It defines a neural collapse (NC) solution for multi-binary classification, proves in a layer-peeled model that all global minimizers of a regularized logistic-type loss are NC solutions, and then fixes the output layer and an auxiliary deep-supervision layer to orthogonal equal-norm classifiers derived from the NC definition. The remaining backbone weights are trained in a federated setup. Experiments on WINNER II and Rician channel models report that NCDSFL converges faster than FedAvg and independent learning and achieves comparable or better BER than these baselines and than MMSE.","tokens_in":14676,"tokens_out":11674,"duration_ms":117216,"significance":"If the theoretical and empirical claims hold, the paper offers a concrete way to reduce trainable parameters and communication payload in federated receivers while mitigating client drift: freezing NC-derived classifiers in both the output and an auxiliary layer. The explicit global-optimality characterization of neural collapse for multi-binary classification with balanced labels is of independent interest, and the authors provide a self-contained proof with consistent matrix definitions. The OFDM experiments cover realistic heterogeneity in power delay profiles and include a conventional MMSE baseline. The main caveat is the gap between the layer-peeled theorem and the actual constrained DNN/FL objective, which is examined below.","major_comments":[{"comment":"Theorem II.1 is proved for the layer-peeled problem (5), where the hidden features H are free optimization variables and the data are balanced with exactly K samples for each of the 2^I label sequences. The actual training objective (14) uses a fixed 500-250-128 ReLU backbone, softmax cross-entropy, and I=16 bits with randomly drawn bit streams; the theorem is not shown to transfer. Universal approximation of functions does not imply that intermediate features are unconstrained optimization variables, and the balanced-label condition of problem (5) is not satisfied exactly in the experiments. Therefore the claim in §IV.B that faster convergence follows 'as stated in Theorem II.1' is not supported. Please either prove a transfer or approximation result showing that the constrained minimizer inherits NC structure (for example, as width grows), or rephrase the algorithmic benefit as heuristic motivation and add a control experiment with a random fixed orthogonal head.","section":"§II.C, §III.B, §IV.B"},{"comment":"The chain of inequalities (45)-(49) shows that every feasible point has loss at least Lbar(ρopt), but the sentence 'which indicates that problem (5) also admits non-trivial global minimizers' does not follow without exhibiting a point that attains the lower bound. The proof should explicitly construct equality-attaining points: choose orthogonal equal-norm class differences Δw_i, set w_i0 = -w_i1, choose h in the row space of A satisfying h = -1/(c K 2^I) A^T w, and scale c so that ||W||^2 + ||H||^2 = ρopt. For the record, the dimension concern about A does not land: A in (7) is dI × dK2^I, and Lemma A.1 is consistent with this dimension.","section":"Appendix A, proof of Theorem II.1"},{"comment":"The validation of Theorem II.1 reports two quantities θ and ϑ for a single training run with no error bars or repeated seeds. Both figures show values that decrease but remain visibly above zero at 5000 epochs, yet the text says they 'confirm the presence of the NC phenomenon' and that hidden features 'gradually converge toward the subspace'. To support this claim, report final metric values, multiple seeds with error bars, a threshold or statistical test, and ideally a comparison against a random fixed orthogonal head to show that the observed alignment is caused by the NC weights rather than by generic training dynamics.","section":"§IV.A, Figs. 3-4"},{"comment":"The FL motivation is data heterogeneity, and the paper explicitly contrasts NCDSFL with FedProx [14], MOON [16], and matched averaging [17]. However, the experiments compare only against FedAvg and independent learning, so the reader cannot judge whether the proposed fixed-NC deep supervision adds value over existing feature-alignment FL methods. Please add FedProx and MOON (and ideally matched averaging) with matched hyperparameters, and report communication cost or transmitted parameter counts if the reduced-overhead claim is retained.","section":"§IV.B"}],"minor_comments":[{"comment":"Several figure axis labels contain typos, for example 'T e ting BER' and 'Number of t aining epochs'; please correct them.","section":"§IV figures"},{"comment":"'imagery part' should be 'imaginary part'.","section":"§III.D, Eq. (27)"},{"comment":"Reference [11] contains the typo 'IIEEE J. Sel. Top. Signal Process.' and should be corrected.","section":"References"},{"comment":"The notation in 'K samples for each label s∈S I' is malformed; please define the label indexing and the relationship between s^{(j)} and the binary value of the sequence more carefully.","section":"§II.A"},{"comment":"The update rule in line 5 sets v_{i,0}=v_k inside the epoch loop, but the reset of local parameters and the relation between v_k and the aggregated global model are not explicit; please clarify the notation.","section":"Algorithm 1"},{"comment":"The formula for ρopt is valid only in the regime 0<λ<t/2; the manuscript should state this restriction explicitly when presenting the formula.","section":"Appendix A, Eq. (48)"}],"recommendation":"major_revision","confidential_remarks":"The dimension inconsistency raised in the internal review does not reproduce: the block recursion (8) gives B_I with I d rows and 2^I d columns, so A is dI × dK2^I and Lemma A.1 is consistent. The proof gap is the missing equality-attaining construction, which is easily fixable. The more serious issue is the transfer from the layer-peeled balanced theorem to the practical constrained DNN/FL objective; this is fixable by repositioning the theory as motivation and adding the suggested control experiments and baselines, but without such changes the central convergence claim is not backed by the theorem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the NC characterization for multi-binary classification, with a proof attempt for global optimality in the layer-peeled model. I checked the algebra in Appendix A: B_i has i d rows and 2^i d columns, so A is dI x dK2^I and A A^T = K2^I I_{dI}. The reader's claimed dimension inconsistency is not present. The proof is sketchy—it never explicitly constructs an equality-attaining point to show the lower bound is tight—but that gap is repairable, and the theorem itself looks like a correct extension of the existing layer-peeled NC literature. Credit where due: the algorithm idea, fixing NC classifiers at both the output and an auxiliary layer and then applying FedAvg, is sensible and reduces trainable parameters. The experiments show faster convergence than plain FedAvg and independent learning, and the NC metrics do decrease, which is consistent with the theory's spirit.\n\nThe soft spots are real, though. The load-bearing gap is the step from Theorem II.1 to the algorithm. The theorem requires H to be free optimization variables, balanced samples across all 2^I label sequences, and a specific regularized logistic loss. The actual DNN/FL objective has a fixed backbone, uses softmax cross-entropy, and the OFDM experiments use I=16 bits per DNN with randomly drawn bit streams, so the balanced-label assumption is not met. The 500-250-128 ReLU networks are not shown to be wide enough for the universal-approximation remark to make H effectively unconstrained. Without a transfer argument, fixing NC weights has no theoretical support for the actual training problem. The empirical evaluation also lacks error bars, code, and comparisons against FedProx or MOON, which are the feature-alignment baselines the paper cites as related work. And the NC validation is somewhat self-referential: they fix weights using their NC definition and then measure NC metrics derived from the same definition. That is not a circular proof, but it limits how much the observations can support the theorem-to-algorithm bridge. The figures show theta and vartheta decreasing, not converging to zero, so the 'collapse' claim is weaker than the text suggests.\n\nBottom line: this is a paper with a probably-correct theoretical contribution and a plausible but unproven algorithmic application. A serious referee should be engaged, not a desk reject. The revision should either provide a real transfer argument connecting the layer-peeled minimizers to the constrained DNN problem, or clearly reposition the theory as motivation and back the algorithm with much stronger experiments, including error bars and the missing baselines. I would read a revised version and would likely cite the multi-binary NC result if it gets properly vetted.","headline":"The paper's core theorem about neural collapse in multi-binary classification appears correct and is genuinely new, but the connection to the federated learning algorithm is not established; the reader's dimension objection does not hold up.","tokens_in":15158,"tokens_out":2342,"would_cite":true,"duration_ms":25494,"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":"For the regularized multi-binary logistic-loss problem, the paper proves that every global minimizer of its layer-peeled model is a neural-collapse solution, and exploits that fixed orthogonal structure to speed up federated OFDM signal…","keywords":["neural collapse","federated learning","signal detection","OFDM","deep supervision","multi-binary classification","feature alignment","layer-peeled model"],"falsifier":"Train the same three-hidden-layer OFDM detector used in Section IV with trainable output and auxiliary layers, then check whether the converged classifiers satisfy $w_{i,1}+w_{i,0}=0$ and $\\langle w_{i,1}, w_{j,1}\\rangle=0$ for all $i\\neq j$, and whether the training loss reaches the predicted value $L(\\rho_{\\mathrm{opt}})$; if either check fails, the minimizers of the real constrained problem are not captured by Theorem II.1.","tokens_in":14155,"feed_emoji":"📡","tokens_out":10502,"duration_ms":93681,"temperature":0.7,"pith_summary":"The paper claims that in a multi-binary classification problem with regularized logistic loss, every global minimizer of the layer-peeled model has a rigid neural-collapse structure: features from the same label sequence collapse to their common mean, the $I$ binary classifiers form an orthogonal set with $w_{i,1}+w_{i,0}=0$ and equal norms, and the features are linearly aligned with the classifiers. This is Theorem II.1, proved for regularizer $0<\\lambda<1/(2I\\sqrt{2K2^I})$. The paper then treats those optimal classifiers as fixed weights at the output layer and at an auxiliary deep-supervision layer, so federated clients share the same classifier anchors. The resulting NCDSFL algorithm is applied to OFDM signal detection, where it converges faster than standard federated averaging and reaches equal or better bit error rate. A sympathetic reader should care because the result converts a feature-alignment heuristic into a theoretically grounded design principle.","feed_headline":"Fixed neural-collapse classifiers cut OFDM training epochs 2.5x","feed_subtitle":"Federated learning with shared orthogonal classifiers converges in 60 epochs instead of 150.","key_machinery":"The load-bearing object is the layer-peeled problem (5) together with the block matrix $A$ built recursively in (7)--(8). The matrix satisfies $AA^{\\top}=K2^I I_{dI}$, so the bilinear form $P=w^{\\top}Ah$ for the $K2^I$ samples of $I$ binary labels is bounded below by $-\\sqrt{K2^{I-1}}(\\|W\\|^2+\\|H\\|^2)$, with equality only under four explicit conditions. Jensen's inequality on the logistic loss then makes every global minimizer attain equality and constant per-sample arguments, and Lemma A.3 turns these conditions into feature collapse, classifier orthogonality and duality (NC1--NC3). The algorithm's practical mechanism is the direct consequence NC2: because an orthogonal classifier set with $w_{i,1}=-w_{i,0}$ and equal norms is optimal, those weights can be fixed in advance at the output layer and at an auxiliary deep-supervision layer, reducing trainable parameters and giving all federated clients a shared geometric anchor for their features.","core_discovery":"The central discovery is Theorem II.1: if $0<\\lambda<1/(2I\\sqrt{2K2^I})$, then any global minimizer of problem (5), the layer-peeled multi-binary classification model with cross-entropy loss and $\\ell^2$ regularization on weights and features, is an NC solution satisfying Definition 1. The proof bounds the sum of the logistic arguments $P=w^{\\top}Ah$ from below using the row-orthogonality of the block matrix $A$, showing equality forces the four conditions of Lemma A.2 together with constant argument values; Lemma A.3 then converts those conditions into the three neural-collapse properties. The paper leverages NC2 to generate an orthogonal classifier set, fixes it at the output layer, adds a second fixed NC classifier as a deep-supervision auxiliary layer, and embeds the resulting model in federated learning. In OFDM simulations the monitored quantities $\\theta$ and $\\vartheta$ decrease monotonically, and the NCDSFL algorithm reaches a testing BER at about 60 epochs while the FedAvg baseline needs about 150 epochs.","pith_inferences":["The theorem is proved for the layer-peeled model, not for the finite three-hidden-layer networks actually simulated; extending the argument would require showing that the optimization over features constrained by a finite backbone has the same minimizers, or testing the NC metrics on much wider networks.","Since the paper notes the multi-binary formulation is equivalent to multi-label classification, the fixed-NC-weight recipe should transfer to other multi-label tasks; a direct experiment would apply NCDSFL to a standard multi-label image benchmark.","The shared orthogonal classifiers provide an anchor for aggregation, so NCDSFL may remain stable under partial client participation or asynchronous updates, regimes the paper does not simulate.","The convergence-speed gain suggests a quantitative test: vary the degree of power-delay-profile heterogeneity across clients and record the rounds needed to reach a target BER; the mechanism predicts the gap between NCDSFL and FedAvg should widen as heterogeneity increases."],"forward_implications":["Fixing the output and auxiliary classifiers to an orthogonal NC weight set does not sacrifice the global optimum of the layer-peeled loss, so the reduced-parameter network is in principle as expressive at the final layer as a fully trained one.","Because all clients share fixed classifiers, their hidden features are pushed toward a common low-dimensional subspace, which should reduce model divergence under heterogeneous power delay profiles.","The design removes the output layer and auxiliary layer weights from local updates and communication, lowering both computation and transmission cost per round.","In the reported OFDM experiments, NCDSFL converges in about 60 training epochs versus about 150 for FedAvg at SNR 10 dB, and it matches or improves the testing BER especially when client SNRs or Rician factors are heterogeneous.","The monotone decay of $\\theta$ and $\\vartheta$ observed during training supports the claim that the multi-binary OFDM detector actually exhibits neural collapse."],"supporting_citations":[{"why":"supplies the layer-peeled model and the idea of replacing a learnable final classifier with a fixed classifier, which Theorem II.1 extends to multi-binary classification","marker":"[22]"},{"why":"introduces the layer-peeled model as an analysis device that treats hidden features as free variables","marker":"[30]"},{"why":"defines the neural collapse phenomenon that the paper adapts to the multi-binary setting","marker":"[18]"},{"why":"justifies fixing NC classifiers by showing global minimizers with unconstrained features are equiangular tight frames and other critical points are saddles","marker":"[20]"},{"why":"provides the existing NC treatment of multi-label learning with pick-all-label loss that the paper's new NC definition is designed to replace","marker":"[23]"},{"why":"introduces deep supervision, the auxiliary-layer training mechanism that NCDSFL equips with fixed NC weights","marker":"[25]"},{"why":"reviews deep supervision and is cited for its usefulness in guiding intermediate features","marker":"[26]"},{"why":"provides the FedAvg baseline that NCDSFL is compared against and whose aggregation steps it follows","marker":"[9]"},{"why":"generates the WINNER II power delay profiles used to create heterogeneous client datasets and the test channels","marker":"[32]"}],"fun_headline_variants":["Neural collapse 2.5x speeds OFDM federated learning","Fixed NC classifiers cut OFDM epochs from 150 to 60","NC-inspired scheme trains OFDM in 60 epochs","Federated OFDM detection via neural collapse: 2.5x faster","Shared orthogonal classifiers boost OFDM training speed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof relies on the layer-peeled model, which treats hidden features as free optimization variables; the actual three-hidden-layer networks used in the OFDM simulations are not shown to be over-parameterized enough for that model to apply exactly.","fun_headline_variants_meta":{"raw":{"variants":["Neural collapse 2.5x speeds OFDM federated learning","Fixed NC classifiers cut OFDM epochs from 150 to 60","NC-inspired scheme trains OFDM in 60 epochs","Federated OFDM detection via neural collapse: 2.5x faster","Shared orthogonal classifiers boost OFDM training speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000257,"raw_usage":{"total_tokens":1522,"prompt_tokens":834,"completion_tokens":688,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":601}},"tokens_in":450,"tokens_out":688,"duration_ms":7192,"temperature":1.0,"reasoning_tokens":601,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:33:20.012007+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same three-hidden-layer OFDM detector used in Section IV with trainable output and auxiliary layers, then check whether the converged classifiers satisfy $w_{i,1}+w_{i,0}=0$ and $\\langle w_{i,1}, w_{j,1}\\rangle=0$ for all $i\\neq j$, and whether the training loss reaches the predicted value $L(\\rho_{\\mathrm{opt}})$; if either check fails, the minimizers of the real constrained problem are not captured by Theorem II.1.","supporting_citations":[{"cited_title":"Inducing neural collapse in imbalanced learning: Do we really need a learnable classifier at the end of deep neural network?","cited_arxiv_id":null,"evidence_quote":"supplies the layer-peeled model and the idea of replacing a learnable final classifier with a fixed classifier, which Theorem II.1 extends to multi-binary classification"},{"cited_title":"Exploring deep neural networks via layer-peeled model: Minority collapse in imbalanced training,","cited_arxiv_id":null,"evidence_quote":"introduces the layer-peeled model as an analysis device that treats hidden features as free variables"},{"cited_title":"Prevalence of neural collapse during the terminal phase of deep learning training,","cited_arxiv_id":null,"evidence_quote":"defines the neural collapse phenomenon that the paper adapts to the multi-binary setting"},{"cited_title":"A geometric analysis of neural collapse with unconstrained features,","cited_arxiv_id":null,"evidence_quote":"justifies fixing NC classifiers by showing global minimizers with unconstrained features are equiangular tight frames and other critical points are saddles"},{"cited_title":"Deeply-supervised nets,","cited_arxiv_id":null,"evidence_quote":"introduces deep supervision, the auxiliary-layer training mechanism that NCDSFL equips with fixed NC weights"},{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"provides the FedAvg baseline that NCDSFL is compared against and whose aggregation steps it follows"},{"cited_title":"Winner ii channel models,","cited_arxiv_id":null,"evidence_quote":"generates the WINNER II power delay profiles used to create heterogeneous client datasets and the test channels"}],"review_version":2}