{"id":"1978a38f-e5bb-4b2e-b788-4dc7f7e2e4e3","arxiv_id":"2412.09629","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A CNN with a feature-masking module and online batch-normalization updates is claimed to deliver millisecond beamforming for cell-free networks with changing channels and varying AP and user counts.","lead":"This paper proposes a deep learning system, HGNet with an online update algorithm, to compute beamforming in cell-free wireless networks when channels and the number of access points and users change over time. It claims millisecond-level beamforming that outperforms a standard optimization baseline, which would matter for real-time wireless systems if the results hold.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Magnitude-only CSI input in Section III-A discards the relative phase information required for beamforming; the central empirical claim is unsupported unless phase information is actually used.","rationale":"The reader and I converge on the same weakest point: Section III-A. The stated input tensor has exactly half the real degrees of freedom of the full complex CSI, so phase is discarded. This is not a stylistic issue; it changes the information available to beamforming. Sum-rate-maximizing beamforming requires aligning desired signal phases and managing interference, both of which depend on the relative phase structure across the M AP antennas and across APs. A magnitude-only CNN can learn some average policy, but it cannot adapt to the phase realization of the current channel. Proposition 3 does not repair this: it bounds a generalization error on output features G_l but never shows that discarding 'sensitive features' preserves enough information for beamforming, nor does it connect the bound to sum rate. The OAU entropy loss is a separate weakness, but the input representation is more fundamental: if taken literally, the reported WMMSE-beating numbers cannot be explained by the described method. The concrete test directly compares two phase-distinct channels with identical magnitude tensors; if HGNet returns the same beamformer while WMMSE rates differ, the central claim fails for the architecture as described. Therefore the reader's REJECT verdict is appropriate, and no change is needed.","tokens_in":19708,"tokens_out":6354,"duration_ms":67501,"concrete_test":"Construct two channel realizations with identical Q_t x I_t x MN magnitude tensors but different relative phases, for example by applying independent unit-modulus phase rotations that preserve all magnitudes while changing the relative phases across AP antennas within each user row. Feed both realizations to the trained HGNet and to WMMSE. If the WMMSE sum rates differ while HGNet outputs the same beamforming tensor for both, the magnitude-only representation cannot achieve per-instance WMMSE-level sum rates. To verify the fix, retrain the same architecture with full real/imaginary CSI represented as 2MN real input channels under the identical protocol; if full-CSI HGNet matches or exceeds WMMSE while magnitude-only HGNet does not, the phase discard is confirmed as the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that HGNet can map the real tensor H3D_t,mod in R^(Q_t x I_t x MN), formed in Section III-A 'by computing modulus values', to complex beamforming V3D_t,com in C^(Q_t x I_t x M) that achieves a higher sum rate than WMMSE. A complex AP-user channel matrix N x M contains 2MN real degrees of freedom (real/imaginary, or magnitude/phase). The specified input keeps only MN values, so the relative phases across AP antennas and across APs, which determine coherent combining gain and interference suppression, are not available to the network. Global row phases do not affect sum rate, but relative phases inside each user's channel row generally do: two matrices with identical magnitudes but different relative phases can have different optimal beamformers and different WMMSE rates. Since the unsupervised training loss is the negative sum rate measured on the true full CSI, a magnitude-only network can only minimize an average over phase realizations; it cannot track the per-channel optimum. No theorem or ablation in the paper establishes that phase-free CSI is sufficient, and no experiment uses a full-CSI input to HGNet. If the implementation actually feeds phase information, Section III-A and the stated tensor dimensions are inconsistent. Either way, the central empirical claim is not supported by the described method.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers downlink beamforming design in cell-free systems under dynamic wireless environments in which channel distributions and the numbers of APs and users change between periods. The authors formulate a sum-rate maximization problem and propose HGNet, a CNN-based architecture whose output dimensions adapt to variable AP/user counts, together with a high-generalization module that masks features deemed sensitive to period identity. They also propose an OAU algorithm that online updates about 3% of the network parameters, namely the affine parameters of batch-normalization layers, using a computationally cheap information-entropy-like loss. The paper claims a theoretical proof that the masking module decreases an upper bound on generalization error, and numerical results compare HGNet with OAU against WMMSE, Edge-GNN, and SUNet in terms of sum rate and computation time.","tokens_in":19977,"tokens_out":6475,"duration_ms":67591,"significance":"The problem is timely, and the paper contains useful engineering pieces: Propositions 1 and 2 give explicit architectural conditions that let a CNN output layer track variable AP/user counts; the partial-update strategy of updating only batch-normalization affine parameters is computationally plausible; and the experiments compare against WMMSE and two recent learning-based baselines rather than only toy settings. If the method worked as described, it would be a practical real-time beamforming solution with a generalization story. However, several load-bearing points are not supported: the input representation appears to discard channel phase, the claimed bound reduction from feature masking is not proven, and the surrogate loss used for online adaptation is not a valid information entropy nor is it linked to sum rate. These issues undermine the central claims of the paper.","major_comments":[{"comment":"The input module converts the complex channel H_t into a real tensor H3D_t,mod in R^{Q_t x I_t x MN} 'by computing modulus values'. This keeps only the per-entry magnitudes of the complex channel and discards all phase information. For beamforming in a MIMO cell-free system, relative phases across the antennas of each AP and across APs determine coherent combining and interference suppression; two channels with identical magnitudes but different relative phases can have different optimal beamformers and different WMMSE sum rates. Since the unsupervised training loss in Section III-D is computed from the true complex CSI, a network fed only magnitudes can at best minimize an average over phase realizations and cannot track the per-channel optimum. The paper provides no theorem or ablation showing that phase-free CSI is sufficient, and no experiment uses a full-CSI input to HGNet. If the implementation actually feeds phase information, the description of Section III-A and the stated tensor dimension are inconsistent. Either way, the central empirical claim is not supported by the described method.","section":"Section III-A"},{"comment":"Proposition 3 and its proof derive a generic MMD-type upper bound on the generalization error for the output of G(·, θ_l). The proof does not compare the masking operation with a baseline without masking, and it does not bound the G-MMD distance of the masked output relative to the unmasked output. The sentences following Proposition 3 assert that discarding sensitive features 'promotes' learning non-sensitive features and 'would become more similar' to the target, but this is a heuristic statement, not a consequence of the proposition. The advertised claim that the high-generalization beamforming module 'theoretically' decreases the upper bound of the generalization error is therefore not established; the proposition as stated is also compatible with masking increasing the bound.","section":"Section III-E and Appendix A"},{"comment":"The loss L_ie is called information entropy, but it is not an entropy. The entries of V3D_t,HGNet are complex-valued beamforming coefficients, not probabilities; the logarithm of a complex quantity is not the standard entropy operation, and the absolute-value weighting does not convert the expression into a meaningful entropy. No theoretical or experimental evidence is given that minimizing L_ie improves achievable sum rate. Since OAU is claimed to improve sum rate specifically through this surrogate loss, the mechanism is unvalidated. The empirical improvement may be an artifact of updating batch-normalization parameters with any gradient signal, but the paper does not test or explain this.","section":"Section IV, Eq. (21)"},{"comment":"The residual identity mapping adds V_IM to C_L in AL(C_L + V_IM). The text states that V_IM is the output of the identity mapping with H3D_t,mod as input, whose third dimension is MN, while C_L has third dimension 2M by Remark 4. For arbitrary M and N, MN and 2M are not equal, so the addition is undefined unless N=2. The experiments use N=2, which hides this inconsistency, but the paper claims general MIMO cell-free systems. The architecture needs a projection for the residual path or an explicit restriction to N=2.","section":"Section III-D, Eq. (15)"}],"minor_comments":[{"comment":"The sentence 'The learning rate and batch size are set to 64 and 0.1, respectively' appears to have the two values swapped; please clarify the learning rate and the batch size.","section":"Section V-A"},{"comment":"The weighted random selection mask is described as being applied during the training stage, but no description states how the mask is chosen at inference time. Specify whether a fixed mask, an averaged mask, or the expected value is used during inference; otherwise the network's output is stochastic.","section":"Section III-C"},{"comment":"There are typographical errors, including 'inclueds' in Section III-C, 'traing' in the caption of Fig. 3, and 'Y ongming' in the author affiliation. These should be corrected.","section":"Throughout"},{"comment":"The figures labeled as average generalization sum rate are plotted against 'Data stream' with 2000 points; clarify whether these are averages over 2000 test samples and over how many random channel realizations, and provide standard deviations or confidence intervals for the main comparisons.","section":"Section V-B"}],"recommendation":"reject","confidential_remarks":"The manuscript is within the scope of the journal, but the phase-information issue is decisive: as written, the method cannot perform the task it claims. In the absence of code or a full-CSI ablation, there is no way to verify whether the actual implementation uses complex CSI despite the text. The theoretical bound-reduction claim and the entropy-based OAU objective are also not supported. These are not presentation issues; they affect the central claims. I would not recommend publication in the present form, and the authors would need substantially new analysis and experiments to address them."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a sensible engineering paper with a novel architecture, but Section III-A's magnitude-only input is a load-bearing flaw that makes the headline result implausible as stated. If the implementation actually uses full complex CSI, the text is wrong; if it doesn't, the network cannot beat WMMSE on Rayleigh channels. The referee should demand an ablation with full CSI input.\n\nThe genuinely new part is the combination: a fully convolutional network that keeps the spatial dimensions Qt x It and lets the channel count vary, plus an adversarial feature-masking module that tries to discard domain-sensitive features, plus online adaptation of batch-norm affine parameters. That is a reasonable toolbox for dynamic cell-free beamforming, and the paper's motivation (fixed-configuration deep learning fails under distribution shift) is well made.\n\nThe soft spots. First and most serious: the input tensor H3D_t,mod has only MN real values per AP-user pair, exactly half the real degrees of freedom of the complex N x M channel. 'Computing modulus values' means phase is gone. Since both coherent combining and interference suppression depend on relative phases, a mapping from |H| to V cannot track the per-realization optimal beamformer. The unsupervised training loss uses the true complex H, so the network learns a phase-averaged mapping; it may do okay statistically but cannot beat WMMSE on each realization. No ablation compares magnitude-only against full-CSI input, so the empirical claim in Figs.4-7 is unsupported as described.\n\nSecond: Proposition 3 is a standard MMD domain-adaptation bound (recycled from Redko et al. via the appendix), and the proof uses only triangle inequality and known lemmas. It holds for any G; it does not show the masking module decreases the bound. The verbal argument that removing 'sensitive' features reduces MMD is plausible but unproven.\n\nThird: the online entropy loss in Eq.(21) is not connected to sum rate; it's a heuristic. The OAU experiments show monotone improvement, but without a gradient analysis or ties to rate it could be an artifact. Fourth: no error bars, seeds, or code for the main comparisons, which matters more when the head result is surprising.\n\nBottom line: the architecture idea may be worth pursuing, and the paper is written clearly enough to be worth a referee's time, but as it stands the central claim cannot be assessed. A reviewer should ask for a full-CSI version and a proper statistical validation. I would not cite this in its current form, but I would bring it to a reading group as a case study in input representation.","headline":"A well-motivated architecture with a likely fatal input-representation flaw that undermines the headline result.","tokens_in":20441,"tokens_out":4284,"would_cite":false,"duration_ms":41487,"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 claims a CNN can deliver millisecond beamforming for cell-free systems as channels, access points, and user counts change, and that an online update of about 3% of its parameters further improves the achieved sum rate.","keywords":["cell-free systems","beamforming design","real-time processing","deep learning","domain generalization","online adaptation","convolutional neural network","sum-rate maximization"],"falsifier":"Take a channel with a fixed set of magnitudes $|H_{ij}|$ and vary only the phases of the complex channel entries; feed the two phase-different versions through the trained HGNet and compare the achieved sum rates, since the network's stated input cannot distinguish them, any difference in output rate would reveal hidden phase use, while identical rates would show that phase-sensitive beamforming gains are not being achieved.","tokens_in":19490,"feed_emoji":"📡","tokens_out":5514,"duration_ms":52424,"temperature":0.7,"pith_summary":"This paper tries to solve real-time beamforming design in cell-free wireless systems where channels, user counts, and access-point counts change between periods. It proposes a convolutional network, HGNet, that maps CSI directly to beamforming vectors, plus a module that discards channel-sensitive features to improve generalization across changing channel distributions. A theoretical bound is given showing the module reduces an upper bound on the generalization error. To track a changing environment online, the OAU algorithm updates only the batch-normalization affine parameters (about 3% of the weights) using an information-entropy loss, keeping computation in the millisecond range. Numerical comparisons report higher sum rates and lower compute than WMMSE, Edge-GNN, and SUNet.","feed_headline":"Beamforming network adapts in milliseconds as users and APs move","feed_subtitle":"A CNN with an online update of about 3% of its weights claims higher sum rate than WMMSE and prior learning baselines.","key_machinery":"The load-bearing object is the high-generalization beamforming module embedded in the convolutional stack. It uses a gradient reversal layer to make the feature extractor adversarially ignore channel identity, a sensitive-feature discriminator that scores each feature channel by its contribution to predicting the channel distribution, and weighted random selection to build a binary mask that zeroes out the most sensitive channels during training. The masked features then flow on to produce the beamforming tensor. Online, OAU updates only the affine parameters $\\gamma_l$ and $\\beta_l$ of each batch-normalization layer, about 3% of all parameters, using an information-entropy loss in place of the expensive sum-rate loss.","core_discovery":"On the paper's own terms, the discovery is that a fixed CNN can serve as a real-time beamformer for dynamic cell-free environments if (i) its convolution architecture is chosen so output dimensions automatically track the varying AP and user counts, (ii) an adversarial feature-discarding module removes the feature-map components most predictive of which channel distribution a sample came from, and (iii) at inference time only the scale and shift parameters of the batch-normalization layers are updated with a cheap entropy objective. The paper proves that the feature-discarding step lowers a G-MMD-based upper bound on the generalization error of the feature extractor, and it reports that the combined system achieves higher average sum rate than WMMSE, Edge-GNN, and SUNet while running in the $10^{-3}$ to $10^{-2}$ second range.","pith_inferences":["Editorial extension: a natural ablation the paper does not run is replacing the modulus-only input with a magnitude-plus-phase or complex two-channel input; such an ablation would show whether the reported gains come from the generalization module or from phase information the current input discards.","Editorial reading: the G-MMD bound in Proposition 3 is a bound on the feature extractor, not on end-to-end achievable sum rate; the sum-rate claim is carried by the numerical comparisons, with the theorem supporting the feature-transfer component.","Editorial extension: the OAU recipe of freezing the network and tuning only batch-normalization affine parameters with a cheap surrogate loss could transfer to other wireless optimization tasks with differentiable objectives and varying system sizes."],"forward_implications":["If the central claim is right, cell-free systems can re-optimize beamformers between periods without retraining, because the CNN architecture is designed so output dimensions follow the input dimensions as AP and user counts change.","Operators can keep one trained network deployed across environments and use the OAU update to recover sum rate after a channel-distribution shift, at a reported computational cost below $10^{-2}$ seconds.","The feature-discarding module directly targets distribution shift: by removing features that best predict which channel model produced the data, the network transfers across Rayleigh, Rician, and geometric multi-path channels better than fixed-configuration GNN or CNN baselines.","The generalization benefit is presented as structural rather than incidental, because the high-generalization module is proven to lower the G-MMD upper bound on the feature extractor's generalization error."],"supporting_citations":[{"why":"Provides the iteratively weighted MMSE algorithm used as the classical optimization baseline whose sum rate and computation time HGNet must beat.","marker":"[7]"},{"why":"Provides the Edge-GNN benchmark, a graph-neural-network beamforming method that assumes a fixed configuration and degrades under distribution shift.","marker":"[10]"},{"why":"Provides the SUNet benchmark, a CNN beamforming method with high computational efficiency that also assumes fixed channel distributions.","marker":"[11]"},{"why":"Supplies the gradient reversal layer mechanism that enables the adversarial training between the beamforming task and the sensitive-feature discriminator.","marker":"[21]"},{"why":"Supplies the weighted random sampling algorithm used to build the binary mask that discards the most channel-sensitive features with O(C) complexity.","marker":"[23]"},{"why":"Supplies the projection function used to enforce the per-AP power constraint on the network output beamforming.","marker":"[25]"},{"why":"Provides the MMD definition and the domain-adaptation lemma (Lemma 5.3) on which the G-MMD generalization bound in Proposition 3 is built.","marker":"[28]"},{"why":"Provides Theorem 29, used in the proof to relate the empirical G-MMD distance to the population G-MMD distance with high probability.","marker":"[32]"}],"fun_headline_variants":["Beamforming adapts in milliseconds with just 3% weight updates","Real-time beamforming for dynamic cell-free networks via HGNet","Online adaptive beamforming: 3% updates for real-time cell-free"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that beamforming can be learned from channel magnitudes alone, because the input module converts complex CSI into a real tensor of modulus values and discards phase, yet optimal beamforming depends on phase alignment and interference structure.","fun_headline_variants_meta":{"raw":{"variants":["Beamforming adapts in milliseconds with just 3% weight updates","Real-time beamforming for dynamic cell-free networks via HGNet","Online adaptive beamforming: 3% updates for real-time cell-free"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000834,"raw_usage":{"total_tokens":3626,"prompt_tokens":917,"completion_tokens":2709,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":2657}},"tokens_in":533,"tokens_out":2709,"duration_ms":18333,"temperature":1.0,"reasoning_tokens":2657,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:35:09.367210+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a channel with a fixed set of magnitudes $|H_{ij}|$ and vary only the phases of the complex channel entries; feed the two phase-different versions through the trained HGNet and compare the achieved sum rates, since the network's stated input cannot distinguish them, any difference in output rate would reveal hidden phase use, while identical rates would show that phase-sensitive beamforming gains are not being achieved.","supporting_citations":[{"cited_title":"An iterative ly weighted mmse approach to distributed sum-utility maximization for a mimo interfering broadcast channel,","cited_arxiv_id":null,"evidence_quote":"Provides the iteratively weighted MMSE algorithm used as the classical optimization baseline whose sum rate and computation time HGNet must beat."},{"cited_title":"Learning cooperati ve beam- forming with edge-update empowered graph neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the Edge-GNN benchmark, a graph-neural-network beamforming method that assumes a fixed configuration and degrades under distribution shift."},{"cited_title":"An efficie nt architec- ture search for scalable beamforming design in cell-free sy stems,","cited_arxiv_id":null,"evidence_quote":"Provides the SUNet benchmark, a CNN beamforming method with high computational efficiency that also assumes fixed channel distributions."},{"cited_title":"Unsupervised domain adapta tion by back- propagation,","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient reversal layer mechanism that enables the adversarial training between the beamforming task and the sensitive-feature discriminator."},{"cited_title":"Weighted random sa mpling with a reservoir,","cited_arxiv_id":null,"evidence_quote":"Supplies the weighted random sampling algorithm used to build the binary mask that discards the most channel-sensitive features with O(C) complexity."},{"cited_title":"Deep weighte d mmse downlink beamforming,","cited_arxiv_id":null,"evidence_quote":"Supplies the projection function used to enforce the per-AP power constraint on the network output beamforming."},{"cited_title":"Redko, E","cited_arxiv_id":null,"evidence_quote":"Provides the MMD definition and the domain-adaptation lemma (Lemma 5.3) on which the G-MMD generalization bound in Proposition 3 is built."},{"cited_title":"Learning via hilbert space embedding of distr ibutions,","cited_arxiv_id":null,"evidence_quote":"Provides Theorem 29, used in the proof to relate the empirical G-MMD distance to the population G-MMD distance with high probability."}],"review_version":1}