Pith. sign in

REVIEW 5 major objections 5 minor 12 references

Deep Learning for THz Channel Estimation and Beamforming Prediction via Sub-6GHz Channel

T0 review · 5 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2411.15589 v1 pith:TYZLL2YX submitted 2024-11-23 eess.SP cs.AI

Sagnik Bhattacharya , Abhishek K. Gupta This is my paper · ORCID
classification eess.SPcs.AI
keywords THzcommunicationconvolutionalneuralnetworkdeeplearningchannelestimationbeamformingpredictionsub-6GHzfactorscodebook
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

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.

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 (5)
  1. [Section IV, Assumption 1 and Eqs. (5)-(6)] 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.
  2. [Section VI vs. Section IV and Table II] 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.
  3. [Section IV, Algorithm Deployment Phase and Eq. (1)] 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.
  4. [Section VI, Performance Evaluation] 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.
  5. [Section V, Baseline, and Section VI] 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.
minor comments (5)
  1. [Section II] There is a typo: 'prohibtive' should be 'prohibitive.'
  2. [Section IV, CNN Architecture] The phrase 'reducing incoming noise' in the context of instance normalization is vague; please clarify what noise is reduced and how.
  3. [Section VI, Fig. 3] 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.
  4. [References] Reference [7] is missing volume, page, and year information; please complete the bibliographic details.
  5. [Section VI] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the sub-6-to-THz mapping is explicitly assumed, not constructed from the paper's own outputs.

full rationale

The paper's derivation chain is a supervised-learning pipeline, not a closed loop. The CNN is trained on labeled pairs (h_S, S_T) and evaluated on held-out users; the beam predictor is trained on exact raytraced S_T with labels from exhaustive codebook search, then applied to CNN-estimated factors. No fitted parameter is relabeled as a prediction, and no equation reduces to its own input by construction. The load-bearing premise that h_S determines h_T and hence S_T is stated explicitly as Assumption 1 and supported by a citation [12]; it is an unverified identifiability assumption for the specific 4-antenna sub-6GHz and 128-antenna THz geometry, but an unsupported premise is a correctness or generalization risk, not circularity. Similarly, the fact that training labels and the upper bound come from the same single-scene ray-tracing simulator limits external validity but does not make the prediction statistically forced: the test split is disjoint and the network outputs are genuine functions of held-out inputs. The only self-citation, [1], is background on THz propagation and is not load-bearing. No circular step can be exhibited from the paper's equations or citations.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central method rests on five modeling assumptions rather than on a derivation: the bijectivity of the channel-factor mapping, the transferability of [12] to THz, the fidelity of raytracing as ground truth, the availability of an accurate sub-6GHz channel estimate, and the geometric channel model. The first two are load-bearing and are asserted, not proved.

free parameters (2)
  • Trained weights of the CNN THz-factor estimator and the dense beamformer predictor = Not released
    The central performance claims are carried by these fitted weights; without the trained models or training code the fitted function cannot be independently reproduced or audited.
  • LT, the number of THz channel paths output by the CNN = Inconsistent: Section VI reports R^{8x7}; Table II limits THz paths to 4; Table III reports errors for 4 paths
    The output dimensionality of the estimator is a key design choice and is presented inconsistently, affecting the meaning of the reported estimation errors.
assumptions (5)
  • domain assumption Assumption 1: the mapping f from channel factors to the channel matrix is bijective under suitable conditions.
    Stated in Section IV without proof; it is the basis for defining f^-1 and for the composed mapping hS to ST.
  • domain assumption Existence of g: hS to hT from [12] holds when extended to the 4-antenna sub-6GHz and 128-antenna THz arrays used here.
    The paper invokes [12] but does not verify its conditions for THz frequencies, the antenna counts, or the simulated environment.
  • domain assumption Raytracing channels (sbr+gas+cloud) faithfully represent real sub-6GHz and THz propagation in the deployment.
    All training and test labels are generated from MATLAB raytrace; no measurement data are used.
  • domain assumption The uplink sub-6GHz channel matrix is available with high accuracy and low overhead.
    Section II and IV feed hS directly into the network; estimation noise on the sub-6GHz input is not modeled in training or testing.
  • domain assumption Geometric channel model with negligible Doppler applies.
    Adopted from [3] in Section III; reasonable for a static outdoor scenario but not verified for mobile users.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Learning for THz Channel Estimation and Beamforming Prediction via Sub-6GHz Channel." pith.science (2026). https://pith.science/paper/TYZLL2YX

@misc{pith2026241115589,
  author       = {Pith},
  title        = {Pith review of: Deep Learning for THz Channel Estimation and Beamforming Prediction via Sub-6GHz Channel},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TYZLL2YX}},
  note         = {Machine review of arXiv:2411.15589}
}
read the original abstract

An efficient channel estimation is of vital importance to help THz communication systems achieve their full potential. Conventional uplink channel estimation methods, such as least square estimation, are practically inefficient for THz systems because of their large computation overhead. In this paper, we propose an efficient convolutional neural network (CNN) based THz channel estimator that estimates the THz channel factors using uplink sub-6GHz channel. Further, we use the estimated THz channel factors to predict the optimal beamformer from a pre-given codebook, using a dense neural network. We not only get rid of the overhead associated with the conventional methods, but also achieve near-optimal spectral efficiency rates using the proposed beamformer predictor. The proposed method also outperforms deep learning based beamformer predictors accepting THz channel matrices as input, thus proving the validity and efficiency of our sub-6GHz based approach.

Figures

Figures reproduced from arXiv: 2411.15589 by the authors.

Figure 1
Figure 1. Illustration showing a dual-band network system with sub-6GHz and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of (a) CNN based THz factors estimator (top) (b) Optimal [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Prediction performance on the first THz channel path for various [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Variation of the spectral efficiency using the proposed beamformer [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 7 canonical work pages

  1. [12]

    Deep learning for TDD and FDD massive MIMO: Mapping channels in space and frequency,

    ——, “Deep learning for TDD and FDD massive MIMO: Mapping channels in space and frequency,” in Proc. Asilomar Conference on Signals, Systems, and Computers , 2019, pp. 1465–1470

    work page 2019

  2. [1]

    Millimeter-wave and Terahertz Spectrum for 6G Wireless

    S. Tripathi, N. V . Sabu, A. K. Gupta, and H. S. Dhillon, “Millimeter- wave and terahertz spectrum for 6G wireless,” 2021. [Online]. Available: https://arxiv.org/abs/2102.10267

    work page Pith review arXiv 2021

  3. [2]

    An overview of signal processing techniques for millimeter wave MIMO systems,

    R. W. Heath, N. Gonz ´alez-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed, “An overview of signal processing techniques for millimeter wave MIMO systems,” IEEE Journal of Selected Topics in Signal Processing, vol. 10, no. 3, pp. 436–453, 2016

    2016

  4. [3]

    An overview of signal processing techniques for terahertz communications,

    H. Sarieddeen, M.-S. Alouini, and T. Y . Al-Naffouri, “An overview of signal processing techniques for terahertz communications,”Proceedings of the IEEE , vol. 109, no. 10, pp. 1628–1665, 2021

    2021

  5. [4]

    Deep-learning- based millimeter-wave massive MIMO for hybrid precoding,

    H. Huang, Y . Song, J. Yang, G. Gui, and F. Adachi, “Deep-learning- based millimeter-wave massive MIMO for hybrid precoding,” IEEE Transactions on V ehicular Technology, vol. 68, no. 3, pp. 3027–3032, 2019

    work page 2019

  6. [5]

    Channel state information prediction for 5G wireless communications: A deep learning approach,

    C. Luo, J. Ji, Q. Wang, X. Chen, and P. Li, “Channel state information prediction for 5G wireless communications: A deep learning approach,” IEEE Transactions on Network Science and Engineering , vol. 7, no. 1, pp. 227–236, 2020

    work page 2020

  7. [6]

    Compressive sensing for indoor THz channel estimation,

    V . Schram, A. Moldovan, and W. H. Gerstacker, “Compressive sensing for indoor THz channel estimation,” in 2018 52nd Asilomar Conference on Signals, Systems, and Computers , 2018, pp. 1539–1546

    work page 2018

  8. [7]

    Estimation of wideband dynamic mmwave and THz channels for 5G systems and beyond,

    A. Brighente, M. Cerutti, M. Nicoli, S. Tomasin, and U. Spagnolini, “Estimation of wideband dynamic mmwave and THz channels for 5G systems and beyond,” IEEE Journal on Selected Areas in Communica- tions

    work page

Show all 12 references
  1. [8]

    DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,

    A. Alkhateeb, “DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,” 2019. [Online]. Available: https://arxiv.org/abs/1902.06435

    2019 arXiv

  2. [9]

    Deep learning coordinated beamforming for highly-mobile millimeter wave systems,

    A. Alkhateeb, S. Alex, P. Varkey, Y . Li, Q. Qu, and D. Tujkovic, “Deep learning coordinated beamforming for highly-mobile millimeter wave systems,” IEEE Access , vol. 6, pp. 37 328–37 348, 2018

    2018

  3. [10]

    ViWi vision-aided mmwave beam tracking: Dataset, task, and baseline solutions,

    M. Alrabeiah, J. Booth, A. Hredzak, and A. Alkhateeb, “ViWi vision-aided mmwave beam tracking: Dataset, task, and baseline solutions,” 2020. [Online]. Available: https://arxiv.org/abs/2002.02445

    2020 arXiv

  4. [11]

    Deep learning for mmwave beam and blockage prediction using sub-6 Ghz channels,

    M. Alrabeiah and A. Alkhateeb, “Deep learning for mmwave beam and blockage prediction using sub-6 Ghz channels,” IEEE Transactions on Communications, vol. 68, no. 9, pp. 5504–5518, 2020

    2020

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.