{"id":"29333e6e-9029-4026-8d15-8f8acba243d2","arxiv_id":"2506.00452","paper_version":5,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An attention network learns a linear MMSE filter for OFDM channel estimation, claiming major complexity savings, but the fixed-filter versus input-dependent-filter descriptions are contradictory.","lead":"An attention transformer is trained to output a linear channel-estimation filter for OFDM wireless links, so that deployment is just one matrix multiplication. The authors report lower estimation error than classical and transformer baselines, but the paper's own equations and complexity table disagree about whether the filter is really fixed at deployment.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (19) makes W_A-MMSE a function of the current pilot observation, yet Sec. IV-B.3 fixes it after training; until reconciled, the linear-inference claim is either false or reduces to an unbenchmarked fixed linear filter.","rationale":"Reading the paper in good faith, the intended contribution is a model-based DNN that learns a linear MMSE filter, so that at inference only a single matrix-vector product is needed. The strongest support would require (1) a consistent deployment-time description, (2) an accurate complexity count, and (3) an empirical comparison showing the learned linear filter beats mismatched LMMSE and DNN baselines. The weakest point is (1): Eq. (19) and Section IV-B.3 contradict each other. If Section IV-B.3 is authoritative, Eq. (19) should be rewritten as a training-time meta-learning step, but then the attention architecture is absent at inference and must be benchmarked against a directly optimized linear filter; the reported gains over mismatched LMMSE do not isolate the attention contribution. If Eq. (19) is authoritative, the complexity claim and Table II are wrong. Either way, the central value proposition is not established as written. The reader's REJECT verdict is therefore appropriate. My emphasis differs slightly from the reader's weakest-assumption statement: the reader focuses on whether a fixed filter remains near-optimal under nonstationary drift, whereas I place the primary weight on the unresolved deployment-time definition plus the missing direct linear baseline; both point to the same conclusion.","tokens_in":22758,"tokens_out":7125,"duration_ms":69471,"concrete_test":"From the released code [37], instrument the inference graph for 100 random test pilot vectors and save W_A-MMSE for each. If the saved filter differs across inputs, the attention encoder is executed per frame; recompute Table II including its FLOPs, and the linear-inference claim fails. If the filter is identical, compute the fixed linear filter W* = argmin_W sum_i ||H_i - W Y_p,i||_2^2 on the same training frames (optionally with the same Huber loss) and compare NMSE on the same 4,000 test frames; if W* matches or beats A-MMSE, the attention architecture contributes nothing beyond direct linear regression.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central value proposition is that, once trained, A-MMSE estimates channels with one matrix-vector product because nonlinearities appear only in the training-time attention encoder. Section IV-A Eq. (19), however, states W_A-MMSE = F_A-MMSE([Re(Y), Im(Y)]_P | theta); if this same function is computed on the current pilot observation, the attention encoder's softmax, FFNs, and layer norms are executed at inference, and the complexity claim collapses. Section IV-B.3 asserts the opposite: 'one final A-MMSE filter is produced from the entire training dataset. This A-MMSE filter remains fixed and is consistently applied to all test data without further modification.' These two descriptions are mutually incompatible. The paper also provides no baseline consisting of the optimal fixed linear filter estimated from the same training data (e.g., sample-covariance LMMSE or ridge regression). Thus, if the filter is fixed, the two-stage attention machinery is training-only, and it is not shown that it beats simply solving the linear least-squares problem; if it is input-dependent, the claimed linear inference is unsupported. Table II adds a second inconsistency: with N=72, M=14, and L=72, the full-rank cost 8NML is roughly 5.8e5 real FLOPs, not the listed 5.8K. Both issues bear directly on the paper's core complexity and performance claims.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes A-MMSE, a model-based OFDM channel estimator in which a two-stage attention transformer is trained to produce a linear filter W_A-MMSE, so that channel estimates are obtained at inference as vec(H_hat) = W_A-MMSE Y_p. A rank-adaptive extension, RA-A-MMSE, factorizes the filter to reduce complexity. The authors report large NMSE reductions relative to LMMSE and Channelformer on COST2100 Semi-Urban and High-Speed Rail scenarios, and claim very low per-inference FLOPs. The core value proposition is that nonlinear operations are confined to the training phase and inference is a single matrix-vector product.","tokens_in":22976,"tokens_out":9141,"duration_ms":91512,"significance":"If the central claims were fully supported, the paper would provide a practical channel estimator that combines the accuracy of a learned model with the inference cost of a linear filter, and the rank-adaptive extension would offer a useful deployment-time complexity knob. The manuscript has strengths: it makes source code publicly available, evaluates on standard COST2100 scenarios with a temporal train/test split, and explicitly targets the interpretability and complexity limitations of end-to-end neural estimators. However, the current text contains a load-bearing inconsistency between the input-dependent definition of W_A-MMSE in Eq. (19) and the fixed-filter statement in Sec. IV-B.3, an internally inconsistent FLOP table, and no comparison against a directly optimized fixed linear filter. Until these issues are resolved, the significance of the contribution cannot be properly assessed.","major_comments":[{"comment":"Eq. (19) defines W_A-MMSE as a function of the current pilot observation, F_A-MMSE([Re(Y); Im(Y)]_P | theta). Under that definition, the estimator in Eq. (18) is input-dependent and nonlinear, because the attention encoder's softmax, FFN, and LayerNorm must be executed at inference, contradicting the paper's core claim of a single linear operation. Section IV-B.3 instead states that one final filter is produced from the entire training dataset and applied fixed to all test data. Please reconcile these statements: if the filter is fixed, rewrite Eq. (19) so that W_A-MMSE is produced from the training data only, and add the missing baseline of the optimal fixed linear filter estimated from the same data (e.g., sample-covariance LMMSE or ridge regression). Without that baseline, the claimed gains of the attention-trained filter over a directly optimized linear filter are not demonstrated.","section":"IV-A, Eq. (19), and IV-B.3"},{"comment":"The FLOP count for A-MMSE is internally inconsistent. With N=72, M=14, and L=72, the full-rank cost 8NML equals roughly 5.8e5 real FLOPs, not the approximately 5.8K listed in Table II. The RA-A-MMSE row (approximately 8608 r) and Eq. (35)'s 18% reduction for r=12 are consistent with a full-rank cost of 8NML about 5.8e5, so the A-MMSE row appears to be low by a factor of 100. In addition, clarify whether the LMMSE row is the per-inference cost with a precomputed filter or includes filter computation and matrix inversion; if the filter is precomputed, LMMSE inference also costs approximately 8NML, and the complexity comparison must be stated on the same basis.","section":"Table II and Sec. V, Eq. (35)"},{"comment":"The nonstationarity claim is tested only on 4,000 test frames that immediately follow the 40,000 training/validation frames. Because the A-MMSE filter is fixed after training, this experiment does not establish robustness to long-term statistical drift. Please either extend the evaluation horizon or temper the claim that the method effectively captures nonstationary channel statistics over time.","section":"VI-A and IV-B.3"},{"comment":"The two-stage attention encoder is motivated by the separable covariance model in Eq. (21), but no ablation is provided to isolate its contribution. Under the fixed-filter interpretation, the encoder and decoder run only during training, so the reader cannot tell whether the reported gains come from the two-stage structure, from the Huber loss, or simply from learning any linear filter on this data. Please add ablations, for example replacing the attention encoder with an unstructured MLP or with direct empirical-risk minimization of W, and a sensitivity analysis for the Huber threshold delta.","section":"IV-B.1 and VI"}],"minor_comments":[{"comment":"The parameter count is inconsistent: the text says the filter requires 2(NM x L) real-valued parameters, which for N=72, M=14, L=72 is 145,152, but the next sentence says approximately 70,000 (= NM x L) parameters. Please correct this.","section":"IV-C.2"},{"comment":"The reported performance gains are inconsistent across the manuscript: the abstract says 56% lower NMSE than LMMSE and 72% lower than Channelformer, the introduction says 61% and 67%, while Section VI-B reports 73.8% and 82.2% in the SU scenario at 35 dB. Please harmonize the numbers and state the operating point for each claim.","section":"Abstract, Sec. I, Sec. VI-B"},{"comment":"The table should state explicitly whether FLOPs are counted per OFDM slot, per frame, or per resource block, and should specify the complex-arithmetic convention (e.g., one complex multiplication equals six real FLOPs) in the caption or text.","section":"Table II"},{"comment":"The notation [Re(Y); Im(Y)]_P uses vertical stacking, while the notation section defines [.,.] as horizontal stacking and [.;.] as vertical stacking; please clarify this usage in Eq. (19).","section":"Notations and Eq. (19)"},{"comment":"The NMSE curves in Figs. 6 and 7 are presented without error bars or multiple-seed variability; please indicate the variance across independent training runs or state that a single run was used.","section":"VI-B"}],"recommendation":"major_revision","confidential_remarks":"The core idea is salvageable: a fixed linear filter learned by an attention architecture could be a legitimate contribution if it is shown to outperform a directly optimized linear filter on the same training data. The current manuscript overclaims because Eq. (19) and Sec. IV-B.3 describe different inference procedures, and the FLOP table has a clear arithmetic error. I recommend major revision rather than rejection, provided the authors can resolve the fixed-filter question, add the direct linear baseline, and correct the complexity accounting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a genuinely new way to train a linear channel-estimation filter with a transformer, and it ships code and data, which is more than many papers in this area do. The reported NMSE gains over mismatched LMMSE and Channelformer are large enough to matter. But the deployment story does not hang together as written, and the FLOPs table has a 100x error. These are fixable, but they are not cosmetic.\n\nWhat is new: learning W in the linear estimator via a two-stage frequency-then-temporal attention encoder, then freezing it for inference, is not in the cited prior work. The rank-adaptive factorization is a reasonable extension. The empirical setup is non-trivial: COST2100 with a temporal split that at least tests some drift, and they compare against a sensing-aided baseline plus the usual suspects. The code link is a real plus.\n\nSoft spots, in order of severity:\n\n1. Eq. (19) defines W_A-MMSE as a function of the current pilot observation. Section IV-B.3 says one final filter is produced from the training set and stays fixed. Those are different estimators. If W is fixed, the transformer is a training-time device, and the paper needs a baseline: the optimal fixed linear filter estimated from the same training data (sample LMMSE, ridge regression). Without that baseline, we cannot tell whether the attention machinery buys anything over just solving a least-squares problem. If W is input-dependent, the \"single matrix-vector product\" claim is false because you have to run the encoder at inference. The paper cannot have it both ways.\n\n2. Table II lists A-MMSE at ~5.8K FLOPs. The paper's own formula, 8 N M L with N=72, M=14, L=72, gives ~5.8e5 FLOPs. That is a 100x discrepancy in the central complexity claim. The rank-adaptive formula lines up with the table only for the RA rows, which suggests the A-MMSE row is a typo, but a typo in the headline number is exactly what a referee should catch.\n\n3. Missing direct linear-filter baseline and error bars. The performance gains over mismatched LMMSE are plausible, but without a learned-linear baseline the specific contribution of the two-stage attention is not isolated. Single-seed results without error bars are common in this subfield but still worth noting.\n\n4. Minor: the temporal split uses the 4,000 frames immediately after training frames; that tests short-term drift, not long-term non-stationarity.\n\nWho this is for: people working on model-based deep learning for the physical layer, OFDM receivers, and low-complexity inference. A serious referee should engage with it. I would not desk-reject; the empirical claims are checkable and the code is available. But I would send it back for major revision with the two consistency issues front and center.\n\nRecommendation: accept for peer review, with the expectation of major revision. If the authors fix Eq. (19) vs. the fixed-filter description, correct Table II, and add a plain learned-linear baseline, this could be a solid contribution.","headline":"A useful idea and strong reported numbers, but the paper's central linear-inference claim is internally inconsistent and the complexity table is off by two orders of magnitude; needs revision before it can be trusted.","tokens_in":23592,"tokens_out":3490,"would_cite":false,"duration_ms":32352,"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":"Attention learns a single linear filter that beats LMMSE on OFDM.","keywords":["OFDM channel estimation","linear MMSE filter","Attention Transformer","two-stage encoder","rank-adaptive estimation","model-based deep learning","computational complexity","COST2100 channel model"],"falsifier":"Compare A-MMSE's fixed filter against a directly optimized linear filter built from the empirical covariance of the same training frames, evaluated on the same test frames; if the directly optimized filter matches or beats A-MMSE, the attention machinery is not the source of the gain. Alternatively, take the final filter and test it on frames separated from the training set by an increasing time gap; if NMSE grows with the gap, the fixed-filter assumption fails.","tokens_in":22488,"feed_emoji":"📡","tokens_out":4985,"duration_ms":43480,"temperature":0.7,"pith_summary":"The paper proposes A-MMSE, an OFDM channel estimator that uses an Attention Transformer only during training to learn the coefficients of a linear minimum mean-squared-error filter. At inference, channel estimation becomes a single matrix-vector multiplication with no nonlinear activations, removing the main computational burden of neural channel estimators. The authors report roughly 56% lower normalized MSE than classical LMMSE and roughly 72% lower NMSE than a recent Transformer-based baseline on COST2100 channels, and a rank-adaptive variant trades a small accuracy loss for large complexity cuts. If correct, the contribution is a model-based network whose training captures nonstationary channel statistics while deployment stays linear and cheap.","feed_headline":"Attention learns a single linear filter that beats LMMSE on OFDM","feed_subtitle":"Once trained, estimation is one matrix-vector product, cutting inference cost to about 5.8k FLOPs.","key_machinery":"The load-bearing object is the two-stage Attention encoder. It consists of a Frequency Encoder and a Temporal Encoder, each using multi-head self-attention with embedding dimensions matched to the number of subcarriers and to the total number of channel elements, respectively. This structure is motivated by the separable correlation model $\\mathbf{R}_{\\text{full}} \\approx \\mathbf{R}_f \\otimes \\mathbf{R}_t$ for WSSUS channels: the first stage learns frequency correlations among pilots and subcarriers, and the second learns temporal correlations across OFDM symbols. A residual fully-connected decoder then maps the encoded features to the real and imaginary parts of the complex filter matrix $\\mathbf{W}_{\\text{A-MMSE}} \\in \\mathbb{C}^{NM \\times L}$. At inference the encoder and decoder are discarded: estimation is the single multiplication of the fixed filter with the pilot vector.","core_discovery":"The central claim is that a linear MMSE filter can be learned end-to-end by an attention-based encoder-decoder, and that the resulting fixed filter is better matched to real, nonstationary OFDM channels than covariance-based LMMSE filters estimated from training samples. The filter is applied as $\\mathrm{vec}(\\hat{\\mathbf{H}}_{\\text{A-MMSE}}) = \\mathbf{W}_{\\text{A-MMSE}} \\mathbf{y}_p$, where $\\mathbf{W}_{\\text{A-MMSE}}$ is produced by the network from the pilot observations during training. After training, one final filter is fixed and applied to all test frames. The paper reports that this fixed filter outperforms both mismatched LMMSE and Channelformer across SNR levels in semi-urban and high-speed-rail COST2100 scenarios, with larger gains in the semi-urban case.","pith_inferences":["A key test not fully addressed in the paper is whether the single fixed filter remains near-optimal as channel statistics drift beyond the 4,000 consecutive test frames; one could quantify this by increasing the temporal gap between training and test frames.","If the A-MMSE filter is truly input-independent after training, then the attention encoder is a training-time device, and the method's advantage over directly optimized linear filters, such as an empirical-covariance LMMSE, should be verified on the same data; the paper's baseline LMMSE is mismatched, which is the relevant comparison but should be made explicit.","The same recipe of learning a linear operator and then deploying it linearly could carry over to other structured estimation problems, such as MIMO detection or channel prediction, where the covariance has a similar separable structure.","The paper mentions an online adaptation framework only in supplementary material; if the online version updates the filter per frame, the linear-inference complexity claim would need to be restated to include the update cost."],"forward_implications":["Once trained, A-MMSE runs channel estimation as one matrix-vector product, so per-inference FLOPs drop to about 5.8k for the configured frame, roughly 95% less than LMMSE or Channelformer.","The reported NMSE gains imply that a data-driven linear filter can beat both mismatched classical LMMSE and a nonlinear Transformer baseline under nonstationary COST2100 channels.","The rank-adaptive extension means the same trained filter can be deployed at reduced rank, with complexity scaling as $\\mathcal{O}(NMr + Lr)$ and still retaining most of the full-rank accuracy.","Robustness results suggest that training at a moderate SNR of 20 dB gives one filter that works over a wide SNR range, avoiding per-SNR retraining."],"supporting_citations":[{"why":"Supplies the multi-head self-attention mechanism used in both encoder stages.","marker":"[11]"},{"why":"Channelformer, the Transformer-based baseline whose NMSE the A-MMSE is measured against.","marker":"[13]"},{"why":"ChannelNet, the CNN baseline for comparison.","marker":"[7]"},{"why":"COST2100 channel model, providing the semi-urban and high-speed-rail test data.","marker":"[31]"},{"why":"WSSUS separable scattering model, motivating the frequency-temporal factorization behind the two-stage encoder.","marker":"[33]"},{"why":"3GPP TS 38.211, defining the DM-RS pilot configuration used in the simulations.","marker":"[36]"},{"why":"Classical SVD-based OFDM channel estimation, providing background for the LMMSE filter structure.","marker":"[4]"}],"fun_headline_variants":["Attention learns one linear filter, beats LMMSE on OFDM","Single learned linear filter outperforms LMMSE in OFDM","Attention-trained linear filter cuts OFDM estimation cost","One matrix-vector product replaces LMMSE in OFDM","Attention transformer yields linear MMSE filter for OFDM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single fixed linear filter, trained on the first 40,000 frames, stays near-optimal for the 4,000 subsequent nonstationary test frames and across SNR levels, rather than needing to be recomputed from each pilot observation.","fun_headline_variants_meta":{"raw":{"variants":["Attention learns one linear filter, beats LMMSE on OFDM","Single learned linear filter outperforms LMMSE in OFDM","Attention-trained linear filter cuts OFDM estimation cost","One matrix-vector product replaces LMMSE in OFDM","Attention transformer yields linear MMSE filter for OFDM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000543,"raw_usage":{"total_tokens":2596,"prompt_tokens":934,"completion_tokens":1662,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":1580}},"tokens_in":550,"tokens_out":1662,"duration_ms":12015,"temperature":1.0,"reasoning_tokens":1580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:05:26.326350+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare A-MMSE's fixed filter against a directly optimized linear filter built from the empirical covariance of the same training frames, evaluated on the same test frames; if the directly optimized filter matches or beats A-MMSE, the attention machinery is not the source of the gain. Alternatively, take the final filter and test it on frames separated from the training set by an increasing time gap; if NMSE grows with the gap, the fixed-filter assumption fails.","supporting_citations":[{"cited_title":"Channelformer: Attention based neural solution for wireless channel estimation and effective online training,","cited_arxiv_id":null,"evidence_quote":"Channelformer, the Transformer-based baseline whose NMSE the A-MMSE is measured against."},{"cited_title":"Deep learning-based channel estimation,","cited_arxiv_id":null,"evidence_quote":"ChannelNet, the CNN baseline for comparison."},{"cited_title":"The COST 2100 MIMO channel model,","cited_arxiv_id":null,"evidence_quote":"COST2100 channel model, providing the semi-urban and high-speed-rail test data."},{"cited_title":"Low-complexity 2D LMMSE channel estimation for ofdm systems,","cited_arxiv_id":null,"evidence_quote":"WSSUS separable scattering model, motivating the frequency-temporal factorization behind the two-stage encoder."},{"cited_title":"NR; physical channels and modulation,","cited_arxiv_id":null,"evidence_quote":"3GPP TS 38.211, defining the DM-RS pilot configuration used in the simulations."},{"cited_title":"OFDM channel estimation by singular value decomposition,","cited_arxiv_id":null,"evidence_quote":"Classical SVD-based OFDM channel estimation, providing background for the LMMSE filter structure."}],"review_version":1}