{"id":"b544384c-d151-4d9d-a109-f7f2bf41754d","arxiv_id":"2411.15589","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A CNN estimates THz channel path factors from sub-6GHz uplink channels, and a dense network uses them to select near-optimal THz codebook beamformers.","lead":"This paper trains a neural network to estimate terahertz (THz) channel details from a lower-frequency, sub-6GHz wireless measurement, then uses those details to pick a good THz beam from a fixed codebook. A generalist may care because it suggests future THz systems could avoid expensive THz pilot training by reusing cheap sub-6GHz measurements.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing assumption that a 4-antenna sub-6GHz channel determines THz channel factors is asserted, not tested; the single-scene noiseless simulation cannot distinguish a genuine cross-frequency mapping from scene-specific fitting.","rationale":"After reading the full text, the central claim is conditional on a deterministic, learnable relation from the 4-antenna sub-6GHz channel to the THz path factors. The paper's support for that relation is Assumption 1 (stated without proof and only about hT, not about hS) plus reference [12] (invoked without stating or checking its conditions). The results do not independently validate the relation because train and test samples are drawn from the same scene and the input hS is noiseless; therefore the network could succeed by learning the smooth, scene-specific map from user position to THz factors. This is not an accusation of misconduct; it is a statement that the experiment as reported cannot distinguish the claimed cross-frequency mapping from an environment-specific fit. The output-dimension mismatch and the absence of estimation noise on the sub-6 input are additional correctness risks, but the mapping assumption is the most load-bearing. Because the approach may still be productive with per-scene training and noisy-input evaluation, the appropriate verdict remains conditional; the reader's weakest_assumption identifies the same central issue.","tokens_in":7612,"tokens_out":8992,"duration_ms":89242,"concrete_test":"Hold out entire scenes: generate a second, geographically disjoint outdoor ray-tracing scene with the same antenna arrays, carrier frequencies, and propagation models, and evaluate the CNN trained on the original block on this unseen scene without fine-tuning. Report Table III errors and Fig. 4 spectral-efficiency curves for the unseen scene. If errors degrade substantially (e.g., mean AoA error above the 6-degree threshold or top-3 rate falling well below the UB curve), the mapping is environment-specific and the claim must be restricted to per-scene training; if the curves match, the cross-frequency mapping concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Everything downstream depends on the existence of a stable mapping g: hS -> ST. The paper justifies g with Assumption 1 and [12], but neither is verified for the stated geometry: NS=4 at 2.4 GHz, NT=128 at 100 GHz, up to 8 sub-6 paths and up to 4 THz paths. With four sub-6 antennas the channel is a low-dimensional projection of a multipath scene, so whether the THz factors are identifiable from hS is exactly the point at issue; citing a bijectivity result for a different frequency-mapping problem is not a substitute. The empirical protocol cannot settle this: the 100,000 users come from a single downtown ray-tracing block, the test split is random within that block, and the CNN is trained and evaluated on the noiseless hS. Under these conditions, a network can fit scene-specific correlations between location/propagation and THz factors without implementing any general sub-6-to-THz mapping. The internal dimension inconsistency (Sec. VI states output ST in R^{8x7}, while Table II allows only 4 THz paths) underscores that the data-generation and evaluation pipeline is under-specified. Thus the central claim is currently supported only within one unperturbed synthetic environment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage deep learning pipeline for THz communication: a CNN estimates THz channel factors (pathloss, angles, time of arrival, phase) from the uplink sub-6GHz channel, and a dense neural network selects an analog beamformer from a codebook using those estimated factors. The method is trained and evaluated on MATLAB raytracing data from a single Chicago downtown block with 100,000 users, and is compared against a direct THz-channel-matrix deep learning baseline and an exhaustive-search upper bound. The authors report small factor-estimation errors and near-optimal top-3 spectral efficiency over SNR from -17 dB to 25 dB, concluding that the sub-6GHz-based approach is valid and efficient.","tokens_in":7846,"tokens_out":2471,"duration_ms":24560,"significance":"If the central mapping is real and learnable, the paper addresses a genuine practical problem: avoiding expensive THz pilot-based channel estimation and beam training by leveraging the more easily estimated sub-6GHz channel. The paper's main trends are internally consistent: the CNN and beam predictor are genuine functions trained on labeled samples, and the spectral efficiency evaluation uses held-out users and an upper bound. However, the practical significance is contingent on a cross-frequency identifiability assumption that is asserted rather than demonstrated, and the evaluation is confined to one noiseless synthetic scene. The paper does not provide machine-checked proofs, code, or a public dataset, so the results are not independently reproducible from the manuscript alone.","major_comments":[{"comment":"The load-bearing premise—that a 4-antenna sub-6GHz uplink channel determines the THz channel factors—is asserted, not derived. Assumption 1 states bijectivity of f from factors to channel without proof, and the invocation of [12] for g: hS -> hT is not verified for NS=4 at 2.4 GHz and NT=128 at 100 GHz. With only four sub-6GHz antennas, hS is a low-dimensional projection of a multipath scene, so the identifiability of up to 4 THz path factors from hS is exactly the point at issue. Please provide a concrete identifiability argument or numerical test for this geometry, or clearly limit the claims.","section":"Section IV, Assumption 1 and Eqs. (5)-(6)"},{"comment":"The output dimension of the CNN is inconsistent: Section IV defines the output as ST in R^{LT x 7}, but Section VI states ST in R^{8 x 7}, while Table II sets the maximum number of THz paths to 4 and Table III reports errors across 4 paths. If LT=8 but each user has at most 4 paths, it is unclear how the remaining rows are defined, how paths are ordered, and how the loss treats absent paths. This under-specification affects both the training target and the interpretation of the reported errors, and it must be resolved in the revised manuscript.","section":"Section VI vs. Section IV and Table II"},{"comment":"The deployment scenario described in the paper feeds an estimated sub-6GHz channel into the CNN, but the training and test inputs are noiseless sub-6GHz channel matrices computed directly from raytracing data. No experiment injects noise or channel-estimation error into hS, despite the system model in Eq. (1) including noise. Without a robustness test against imperfect sub-6GHz channel estimates, the claim that the pipeline works in deployment is not supported.","section":"Section IV, Algorithm Deployment Phase and Eq. (1)"},{"comment":"The evaluation uses 100,000 users drawn from a single 400 m x 30 m raytraced block and a random train/test split within that block. This setup cannot distinguish a general learnable sub-6GHz-to-THz mapping from scene-specific fitting of location-dependent propagation correlations. Please evaluate on a second environment or at least use a spatially disjoint train/test partition (e.g., different streets or a separate building block) and report the resulting performance drop, if any.","section":"Section VI, Performance Evaluation"},{"comment":"The baseline comparison is potentially unfair or unclear. Section V says the baseline THz channel matrix is computed from raytracing data, which would be a clean input, but Section VI explains the baseline's poor performance by saying 'the estimated THz channel matrix [is] highly noisy.' If the baseline actually uses noisy THz channel estimates, the noise model and SNR for that input are not specified; if it uses exact raytracing channels, the explanation is inconsistent. Please clarify the baseline input and ensure both methods are compared at comparable input fidelity.","section":"Section V, Baseline, and Section VI"}],"minor_comments":[{"comment":"There is a typo: 'prohibtive' should be 'prohibitive.'","section":"Section II"},{"comment":"The phrase 'reducing incoming noise' in the context of instance normalization is vague; please clarify what noise is reduced and how.","section":"Section IV, CNN Architecture"},{"comment":"Figure 3 plots predictions for 20,000 test samples, but the x-axis range and the dense overlap make it difficult to visually assess error magnitude; consider plotting a small random subset or using quantile bands.","section":"Section VI, Fig. 3"},{"comment":"Reference [7] is missing volume, page, and year information; please complete the bibliographic details.","section":"References"},{"comment":"The scenario description says 'S Clarke St. of Chicago downtown area,' but the exact building footprints and raytracing configuration are not provided; please give the MATLAB raytrace settings or cite the map source so the experiment can be reproduced.","section":"Section VI"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is plausible and relevant, but the main claim currently rests on an unverified identifiability assumption for a very asymmetric antenna configuration (4 vs. 128 antennas), and the single-scene noiseless evaluation cannot establish the existence of a general mapping. The dimension inconsistency between the stated 8x7 output and the 4-path data is a concrete sign that the data generation pipeline needs more rigorous specification. I would ask the authors for an identifiability test, input-noise robustness results, a spatially or environmentally disjoint generalization test, and a clarification of the baseline input before considering acceptance. The citation [12] is used as the sole support for the cross-frequency mapping; the authors should check whether that paper's conditions hold for their setup or provide their own evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something concrete: it trains a CNN to map a 4-antenna sub-6 GHz uplink channel to THz path factors, then a dense network to pick a codebook beamformer from those factors. In one Chicago-downtown ray-tracing scene, it reports small prediction errors and near-optimal top-3 spectral efficiency, beating a direct THz-matrix baseline. That is a legitimate extension of Alrabeiah and Alkhateeb's mmWave work, and the intermediate channel-factor regression is new relative to that line.\n\nWhat it does well: the architecture is clearly described, the training/test split is respectable, errors are reported with mean and standard deviation, and the beamformer results are compared against both a baseline and an exhaustive-search upper bound. The authors also state Assumption 1 explicitly as an assumption rather than burying it. The citation pattern is fine — the predecessors are appropriately credited.\n\nThe soft spots are in proportion to how much the central claim depends on them. The load-bearing premise is that a learnable mapping exists from hS to hT and then to ST. Assumption 1 is asserted without proof, and [12] is invoked for a different massive-MIMO channel-mapping setup whose conditions are not checked for NS=4 at 2.4 GHz and NT=128 at 100 GHz. With four antennas, hS is a low-dimensional projection; whether up to 4 THz paths are identifiable from that projection is exactly the point at issue. The empirical protocol cannot distinguish a genuine cross-frequency mapping from scene-specific fitting, because the 100,000 users are random positions in a single 400m x 30m block, the test split is random within that block, and the sub-6 input is noiseless. There is also a dimension inconsistency: the output is described as R^{8x7}, while Table II caps THz paths at 4 and Table III reports errors across 4 paths. That suggests the data-generation and evaluation pipeline is under-specified. These are not fatal to the idea, but they mean the general claim is currently supported only within one unperturbed synthetic environment.\n\nFor a reader working on sub-6-assisted THz beamforming, the paper is worth reading: the architecture and problem framing are useful, and the negative result about the direct THz-matrix baseline is interesting. I would not yet cite it as evidence that the sub-6-to-THz factor mapping exists. The paper deserves a serious referee — the topic is timely and the work is a genuine extension — but it would need major revision: multi-scene evaluation, noise on the sub-6 input, and at least a numerical check or discussion of identifiability.","headline":"Plausible sub-6GHz-to-THz beamforming pipeline, but the load-bearing mapping is assumed, not shown, and the single-scene noiseless evaluation can't carry the general claim.","tokens_in":8405,"tokens_out":2184,"would_cite":false,"duration_ms":20858,"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":"A convolutional network can infer THz channel factors from the sub-6GHz uplink channel and select codebook beams with near-optimal spectral efficiency, according to this paper.","keywords":["THz communication","convolutional neural network","deep learning","channel estimation","beamforming prediction","sub-6GHz channel","channel factors","codebook beamforming"],"falsifier":"A one-antenna variant of the sub-6GHz receiver is the cleanest test: with no spatial resolution, the claimed bijective map from sub-6GHz channel to THz angles cannot exist, so if a CNN trained on single-antenna inputs still predicts THz angles with small error, the network is exploiting the raytraced location prior rather than information carried by the sub-6GHz channel.","tokens_in":7381,"feed_emoji":"📡","tokens_out":12496,"duration_ms":92309,"temperature":0.7,"pith_summary":"This paper tries to establish that a convolutional neural network can estimate the essential parameters of a terahertz (THz) channel—path loss, angle of arrival and departure, time of arrival, and phase—directly from the uplink sub-6GHz channel, and that a second dense network can turn those estimated parameters into a near-optimal choice of analog beamformer from a codebook. If the claim holds, THz systems can skip the heavy pilot overhead and matrix computations of conventional channel estimation, and skip beam training, by reusing the inexpensive sub-6GHz measurements. The authors train on raytraced data from a downtown scenario and report mean angle errors of a few degrees, time-of-arrival errors near $3.8\\times10^{-18}$ seconds, and pathloss errors near $5.7\\times10^{-19}$; their top-3 beam predictor reaches spectral efficiency close to an exhaustive-search upper bound across $-17$ to $25$ dB SNR. The paper thus presents a new application of deep learning: a learned cross-frequency mapping from a lower band to THz beam selection.","feed_headline":"Sub-6GHz uplink picks THz beams with near-optimal rates","feed_subtitle":"CNN pulls THz angles, delays from sub-6GHz channel; dense net picks beams near exhaustive-search rates.","key_machinery":"The machinery is a learned composition of two maps. The first is the factor-to-channel map $f$ of the geometric channel model (Eq. 4), which builds the sub-6GHz and THz channel vectors from per-path factors $S_T = [\\lambda'_l, \\tau, \\theta, \\phi_r, \\theta_r, \\phi_t, \\theta_t]$; Assumption 1 states $f$ is bijective under suitable conditions, so channel factors can be recovered from the channel matrix. The second is the cross-frequency map $g: h_S \\to h_T$ invoked from [12], which the paper combines with $f^{-1}$ to justify estimating THz factors from sub-6GHz measurements. The CNN realizes that composite map in a regression head with two convolution-maxpool-instance-normalization-ReLU blocks, global average pooling, and a dense output; the beamformer predictor is a four-layer dense network with softmax doing classification over the 2048-entry codebook. The key load-bearing identity is the existence of the composite $f^{-1} \\circ g$: if it is not bijective in practice, the network can only memorize the training environment.","core_discovery":"The central claim is that the sub-6GHz uplink channel carries enough information to determine the THz channel's path-level parameters, and that a CNN can learn that map well enough to drive codebook beam selection. The paper argues for the existence of the map by combining Assumption 1, the bijectivity of the channel-factor-to-channel map $f$ for a fixed array and environment, with the cross-frequency result of [12], which asserts a bijective mapping $g: h_S \\to h_T$ from the user's channel at one frequency to its channel at another. The CNN consumes the $64\\times 4\\times 2$ sub-6GHz channel tensor (magnitude and phase over 64 subcarriers and 4 antennas) and regresses the first $L_T=4$ THz paths' seven factors each. A dense classifier then selects among the 2048 quantized beamformers of the codebook, with softmax probabilities providing top-1 and top-3 candidates. The authors report that the estimated factors carry mean errors below $6^\\circ$ for angles, $3.80\\times10^{-18}$ s for ToA, and $5.70\\times10^{-19}$ for pathloss, and that the top-3 beam predictor's spectral efficiency tracks the exhaustive-search upper bound across SNR from $-17$ to $25$ dB, while the top-1 predictor beats both top-1 and top-3 variants of a baseline that consumes the full THz channel matrix.","pith_inferences":["The authors do not ablate the sub-6GHz array size; a natural extension would be to retrain with 1, 2, 4, and 8 antennas and measure whether the THz factor estimation error falls monotonically, which would separate the value of spatial information from the network's location prior.","Because the dataset is a single raytraced street layout, the cross-frequency map may be overfit to that environment; testing on a second city layout or with randomized scatterer positions would reveal whether the sub-6GHz-to-THz coupling generalizes.","The factor representation opens a direct path to blockage prediction: a drop in estimated THz pathloss or the disappearance of a path estimated by the CNN could trigger beam re-selection without fresh THz pilots, though the paper does not explore this."],"forward_implications":["THz beam acquisition can be moved to the sub-6GHz band, removing the need for uplink THz pilots, matrix inversions, and exhaustive beam training at the THz array.","The estimated THz channel factors are reusable for other THz tasks such as blockage prediction and user positioning, without reconstructing the full THz channel matrix.","Because the baseline that consumes the actual THz channel matrix performs worse, the method implies that sub-6GHz-derived factors are a cleaner learning target than noisy THz channel observations.","The near-optimal top-3 performance across $-17$ to $25$ dB SNR means the method is not tuned to a single operating point; the softmax top-3 supply a low-cost fallback when the top-1 beam is wrong.","The requirement of co-located sub-6GHz and THz arrays, plus a static environment, delimits where the claimed mapping can be expected to hold."],"supporting_citations":[{"why":"Supplies the THz pathloss model with an extra exponential molecular-absorption term that the paper adopts as part of the channel factors.","marker":"[1]"},{"why":"Provides the geometric wideband channel model with steering vectors and the beamforming architecture assumptions the system model uses.","marker":"[3]"},{"why":"Motivates the CNN-based approach to channel state information prediction from proxy inputs, the template for the THz factor estimator.","marker":"[5]"},{"why":"Establishes the predecessor result that sub-6GHz channels can predict mmWave beams, which this paper extends to THz factor estimation and beamformer selection.","marker":"[11]"},{"why":"It is the cited source for the assumed bijective cross-frequency channel mapping $g: h_S \\to h_T$, the load-bearing existence claim of the whole pipeline.","marker":"[12]"}],"fun_headline_variants":["Sub-6GHz channel alone forecasts THz beams","THz beams predicted from sub-6GHz, near-optimal rates","No THz estimation needed: sub-6GHz picks THz beams","Near-optimal THz beamforming via sub-6GHz channel","Sub-6GHz uplink steers THz beamformer, beats THz-input"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire prediction chain stands on the premise that the 4-antenna 2.4 GHz uplink channel uniquely determines the THz paths' angles, delays, and losses; if two different THz scenarios can look identical at sub-6GHz, the network cannot separate them and the claimed near-optimal beams are not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Sub-6GHz channel alone forecasts THz beams","THz beams predicted from sub-6GHz, near-optimal rates","No THz estimation needed: sub-6GHz picks THz beams","Near-optimal THz beamforming via sub-6GHz channel","Sub-6GHz uplink steers THz beamformer, beats THz-input"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001032,"raw_usage":{"total_tokens":4377,"prompt_tokens":1007,"completion_tokens":3370,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":3276}},"tokens_in":623,"tokens_out":3370,"duration_ms":21341,"temperature":1.0,"reasoning_tokens":3276,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:07:49.772692+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A one-antenna variant of the sub-6GHz receiver is the cleanest test: with no spatial resolution, the claimed bijective map from sub-6GHz channel to THz angles cannot exist, so if a CNN trained on single-antenna inputs still predicts THz angles with small error, the network is exploiting the raytraced location prior rather than information carried by the sub-6GHz channel.","supporting_citations":[{"cited_title":"Millimeter-wave and Terahertz Spectrum for 6G Wireless","cited_arxiv_id":"2102.10267","evidence_quote":"Supplies the THz pathloss model with an extra exponential molecular-absorption term that the paper adopts as part of the channel factors."},{"cited_title":"Channel state information prediction for 5G wireless communications: A deep learning approach,","cited_arxiv_id":null,"evidence_quote":"Motivates the CNN-based approach to channel state information prediction from proxy inputs, the template for the THz factor estimator."},{"cited_title":"Deep learning for mmwave beam and blockage prediction using sub-6 Ghz channels,","cited_arxiv_id":null,"evidence_quote":"Establishes the predecessor result that sub-6GHz channels can predict mmWave beams, which this paper extends to THz factor estimation and beamformer selection."},{"cited_title":"Deep learning for TDD and FDD massive MIMO: Mapping channels in space and frequency,","cited_arxiv_id":null,"evidence_quote":"It is the cited source for the assumed bijective cross-frequency channel mapping $g: h_S \\to h_T$, the load-bearing existence claim of the whole pipeline."}],"review_version":1}