Pith. sign in

REVIEW 4 major objections 5 minor 71 references

CSI-4CAST: A Hybrid Deep Learning Model for CSI Prediction with Comprehensive Robustness and Generalization Testing

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A hybrid deep network predicts future wireless channel states better than larger LLM-based models at a fifth of the compute.

desk verdict The CSI-RRG benchmark is the real contribution; the TDD improvement is plausible, but the FDD claim rests on point estimates without error bars and the paper needs statistical and quality fixes before its headline numbers can be trusted. read the letter →

arxiv 2510.12996 v2 pith:63ACY4KF submitted 2025-10-14 cs.LG

Sikai Cheng , Reza Zandehshahvar , Haoruo Zhao , Daniel A. Garcia-Ulloa , Alejandro Villena-Rodriguez , Carles Navarro Manchón , Pascal Van Hentenryck This is my paper · ORCID
classification cs.LG
keywords CSIpredictionmassiveMIMOchannelagingdeeplearninghybridarchitecturerobustnessgeneralizationbenchmark
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

CSI-4CAST claims that a deliberately lightweight hybrid network—CNN residual blocks, adaptive correction layers, ShuffleNet blocks, and a Transformer encoder—can predict future downlink channel state information more accurately than much larger deep-learning models while using a fraction of the compute. The paper backs this with a new benchmark, CSI-RRG, containing over 300,000 samples across 3,060 TDD and FDD scenarios that mix standardized channel models, delay spreads, velocities, and three non-Gaussian noise types: phase, burst, and packet-drop. On that benchmark, the model reports the lowest NMSE in 88.9% of TDD scenarios and 43.8% of FDD scenarios—the best overall—while reducing FLOPs by roughly 5x and 3x versus the strongest LLM-based baseline. A sympathetic reader would take away that efficient hybrid designs can match or beat much larger sequence models on TDD CSI prediction, and that the hard open problem is FDD cross-band generalization, where all models degrade.

What carries the argument

CSI-4CAST itself: a per-antenna pipeline that stacks (1) a CNN residual block for denoising, (2) adaptive correction layers (MLPs with additive or multiplicative skip connections) over time and subcarrier/delay dimensions, (3) ShuffleNet blocks for efficient feature extraction, and (4) a Transformer encoder with sinusoidal position embeddings. A parallel IDFT branch converts the CSI into the delay domain, and the two representations are summed before prediction. The companion CSI-RRG benchmark supplies the stress-test protocol: 27 training configurations, three evaluation tracks (regular, robustness, generalization), rank-based scoring to avoid averaging across incomparable NMSE scales, and

What would settle it

Run CSI-4CAST on measured mMIMO channel traces from a deployed base station with real oscillator phase noise and network packet losses, comparing against the same baselines under identical compute budgets; if it does not rank first in a majority of TDD scenarios, or if its NMSE advantage over a no-prediction persistent baseline disappears, the robustness and generalization claims do not transfer.

Watch

Extended reading notes

Core claim

On its own terms, the paper's core discovery is that the channel-aging problem in massive MIMO can be attacked by a four-component architecture. The CNN residual refines noisy observations; the adaptive correction layers adjust learned representations along time and, for FDD, along subcarrier dimensions; ShuffleNet compresses features; and the Transformer captures long-range temporal dependencies. A frequency-to-delay IDFT branch adds a sparse, physically structured view of multipath. Evaluated on CSI-RRG, this combination achieves the best NMSE ranking in 88.9% of TDD scenarios and leads in 43.8% of FDD scenarios, with FLOPs reduced by factors of five (TDD) and three (FDD) relative to the s

Load-bearing premise

The synthetic 3GPP CDL channels and the calibrated noise models—phase noise only at 10–25 dB SNR, burst noise capped at one pulse per input, and packet drops filled with the last sample—faithfully represent real-world channels and interference; if actual deployment noise or propagation has different structure or intensity, the claimed robustness and generalization margins may not transfer.

Editorial extensions

If this is right

  • If the TDD results hold, accurate CSI prediction does not require large pretrained language models; a roughly 20-million-parameter network can run in single-digit milliseconds at a base station.
  • The CSI-RRG protocol gives the field a common yardstick: future predictors can be compared under identical phase, burst, and packet-drop noise rather than under AWGN only.
  • Robustness depends on noise structure as much as intensity: burst noise degrades prediction more than phase noise at equal SNR, and packet drops hurt TDD more than FDD.
  • FDD inter-band prediction remains a bottleneck; no evaluated model generalizes to unseen FDD scenarios, pointing toward active learning or adaptive retraining as the next step.
  • Rank-based scoring avoids the bias of averaging NMSE across scenarios with very different difficulty, so performance claims should be read scenario-by-scenario rather than as a single number.

Reading between the lines

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

  • A natural ablation—dropping the Transformer or the delay-domain branch—would reveal which component actually carries the robustness; the paper does not report this, but the modular design makes it directly testable.
  • The same four components are generic enough to transfer to other time-frequency correlated wireless tasks such as beam prediction or mmWave sensing, provided the synthetic-to-real gap is closed.
  • Because burst noise is capped at one pulse per input and phase noise is tested only at 10–25 dB SNR, real-world interference with repeated pulses or deep fades could erase the reported margins; that is a testable limitation rather than a contradiction.
  • Training on additional channel models or LOS/NLOS mixes from the start would likely shrink the generalization gap, consistent with the paper's own finding that channel-model shifts are the hardest factor.
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

4 major / 5 minor

Summary. The paper proposes CSI-4CAST, a hybrid deep learning architecture combining CNN residuals, adaptive correction layers, ShuffleNet blocks, and Transformer encoders for CSI prediction in massive MIMO systems. The authors introduce CSI-RRG, a large-scale benchmark synthesized with Sionna, covering 3GPP CDL channel models, multiple delay spreads, user velocities, AWGN plus phase, burst, and packet-drop noise, and TDD/FDD duplexing. The central claim is that CSI-4CAST achieves the lowest NMSE in 88.9% of TDD and 43.8% of FDD Regular-track scenarios, ranks first among all evaluated models, and reduces FLOPs by 5× (TDD) and 3× (FDD) versus the strongest baseline LLM4CP. The paper also reports rank-based and spectral-efficiency metrics, robustness and generalization analyses, and publicly releases the dataset and evaluation code.

Significance. If the reported gains are real, the paper makes a valuable contribution to CSI prediction: it introduces a new architecture with strong efficiency advantages (5× FLOP reduction vs. LLM4CP in TDD) and offers a large, publicly available benchmark that goes beyond standard AWGN and in-distribution testing. The generalization tracks across unseen channel models, delay spreads, and velocities, and the robustness tracks with realistic noise, are useful for the community. The release of dataset and evaluation protocols is a concrete reproducibility asset. However, the headline superiority claim is built entirely on point estimates of NMSE per scenario, with no uncertainty quantification, paired significance testing, or seed variance. Given that several FDD margins are small, the quantitative precision of the central claim is not yet supported. That said, the qualitative direction of the results is plausible, and the issue is fixable with additional statistical analysis.

major comments (4)
  1. [Sections IV-A.c and V-A, Eqs. (7), (16)–(19)] The headline claims that CSI-4CAST 'achieves the lowest NMSE in 88.9% of TDD scenarios and 43.8% of FDD scenarios' are computed from a single NMSE point estimate per scenario, with each scenario containing only 100 samples. No confidence intervals, paired significance tests, or multiple-seed variance are reported. A rank is a nonlinear function of a noisy estimate: with 100 samples, models with similar true NMSE can easily flip ranks, and the model with lower estimate variance will be over-counted as rank-1. The FDD margins are particularly tight (MeanRank 1.62 in the Regular track; Prank1 43.8% versus what the next-best model would achieve under random sampling). The manuscript should provide bootstrap confidence intervals for Prank1 and MeanRank, paired tests across scenarios (e.g., Wilcoxon signed-rank), or repeated-seed training/evaluation to quantify the stability of the ranking. At
  2. [Sections I, V-A, VI] The abstract and conclusion state '88.9% of TDD scenarios and 43.8% of FDD scenarios' without specifying that these are Regular-track (in-distribution) numbers. Section V-A explicitly shows that in FDD Generalization, 'none of the methods achieve reliable generalization' and the rank distributions are wide. Presenting the 43.8% FDD figure as a general statement overstates the model's performance on unseen conditions. Please qualify these percentages as 'on the in-distribution Regular track' or report the full breakdown across Regular, Robustness, and Generalization. Additionally, the abstract's '3,060 scenarios' refers only to the Generalization track; the total test set has 3,708 scenarios (162+486+3,060). The current wording conflates these numbers.
  3. [Section IV-C] The Regular track is not an independent test of generalization because the models are selected by lowest validation NMSE on data drawn from the same configuration grid (Table II) and the Regular track uses the same configurations. Thus the Regular results largely reflect model-selection performance, not out-of-distribution generalization. The paper should clearly state this distinction and ensure that the Generalization track is the primary evidence for cross-scenario claims. This does not invalidate the benchmark, but it tempers the 'Comprehensive Robustness and Generalization Testing' claim in the title and abstract.
  4. [Appendix D-C.a] The robustness track for phase and burst noise is limited to SNR 10–25 dB, while the Regular track evaluates AWGN down to 0 dB. The authors justify excluding 0–5 dB as 'uncommon in practice,' but this means the robustness stress test is not as demanding as the AWGN test and makes cross-noise comparisons at low SNR impossible. State this limitation prominently in Section V-C and consider adding low-SNR robustness scenarios (e.g., 5 and 0 dB) to match the Regular-track difficulty. The current robustness claims ('CSI-4CAST consistently achieves the lowest NMSE' across all noise types and levels, Section V-C) are based on this reduced SNR range.
minor comments (5)
  1. [Abstract] The abstract in the submission contains a numerical inconsistency: it states '81.5% of TDD scenarios and 44.4% of FDD scenario' in one place, while the full text abstract and Section V-A report 88.9% and 43.8%. Please reconcile all occurrences.
  2. [Figs. 5 and 6] The left axis is labeled 'RankScore' but the values range 0–100. RankScore as defined in Eq. (18) lies in [0, |Π|−1] = [0,5] for six models. It appears the violin plots actually show the rank distribution on a 1–6 scale, and the red line shows Prank1 (%). Please correct the axis label or clarify the definition in the caption.
  3. [Appendix E] The text says 'delay spreads (Table VIII); channel models (Table VIII)' but the first cross-reference should be Table VII. Also, Table X is referenced as Table IX in the appendix? Please check all table numbering.
  4. [Appendix F, Table IX] Typo: 'STEGMNN' should be 'STEMGNN'. Also, the search space table lists 'STEGMNN' and 'STEGMNN' variants; please standardize names.
  5. [Eq. (13)] The positional encoding formula is nonstandard; the denominators mix |T| and γ in an unusual way. If intentional, a brief explanation would help; otherwise, consider using the standard Transformer encoding or clarifying why this form was chosen.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: CSI-4CAST's accuracy claims rest on independent test tracks; the only mild self-referential point is validation/test distribution overlap, which is selection bias, not circularity.

full rationale

The paper is an empirical architecture/benchmark contribution; it does not derive predictions from fitted parameters. The training objective (Eq. 6) and evaluation metric (Eq. 7) are both NMSE, but the test scenarios are generated from Sionna independently of the training/validation samples, and the Robustness/Generalization tracks use held-out channel models, velocities, delay spreads, and noise types. No claim reduces to its input by construction: per-scenario ranks and Prank1 values are computed from separate test samples, and no fitted parameter is renamed as a prediction. The only mild self-referential aspect is that the Regular test distribution coincides with the validation distribution used for Optuna-based checkpoint selection (Sec. IV-C), so in-distribution rank margins may be optimistically selected; this is a benchmark-design/statistical-support caveat, not circularity. There is no load-bearing self-citation: architecture choices cite external works (e.g., ShuffleNet, Transformer), baseline implementations are attributed to their original authors, and no uniqueness theorem or prior result by the same authors is used to force the conclusion. Lack of confidence intervals on the 100-sample per-scenario NMSE ranks is a statistical robustness concern, outside the circularity definition.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

There is no first-principles derivation; the central claim is an empirical model comparison. The ledger records the model-selection and data-generation assumptions on which the benchmark conclusions rest.

free parameters (2)
  • Optuna-tuned hyperparameters for each (model, duplexing) pair = Not enumerated fully; search ranges in Table IX
    Architecture dimensions, optimizer settings, ACL operations are chosen by validation NMSE on a split of the Regular distribution, making the Regular track partly in-sample after model selection.
  • Noise calibration parameters (σ for phase noise; A_burst, P_burst for burst noise) = Derived from empirical SNR calibration; exact values not listed
    These determine the strength of phase/burst noise at the evaluated SNR targets 10-25 dB; they constrain the robustness analysis.
assumptions (4)
  • domain assumption Sionna-generated 3GPP TR 38.901 CDL channel models adequately represent real-world CSI dynamics.
    All training and testing data is synthetic; no real-world channel measurements are used (Section IV-A).
  • domain assumption Per-antenna independence: each transmitter-receiver pair is modeled separately.
    Architecture input Eq. (8) operates on a single antenna pair; inter-antenna correlations are ignored (Section III-a).
  • domain assumption Noise models: phase (Gaussian phase perturbation), burst (at most one bell-shaped pulse per input), packet drop (Bernoulli erasures with last-sample imputation).
    These synthetic structures are assumed to cover practical non-Gaussian noise; e.g., at most one burst per sequence and only SNR 10-25 dB for phase/burst (Appendix D-B, D-C).
  • standard math NMSE and SE correctly quantify prediction quality for precoding, as defined in Eq. (7) and Eq. (35).
    Standard definitions widely used in wireless communications; accepted as background.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CSI-4CAST: A Hybrid Deep Learning Model for CSI Prediction with Comprehensive Robustness and Generalization Testing." pith.science (2026). https://pith.science/paper/63ACY4KF

@misc{pith2026251012996,
  author       = {Pith},
  title        = {Pith review of: CSI-4CAST: A Hybrid Deep Learning Model for CSI Prediction with Comprehensive Robustness and Generalization Testing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/63ACY4KF}},
  note         = {Machine review of arXiv:2510.12996}
}
read the original abstract

Channel state information (CSI) prediction is a promising strategy for ensuring reliable and efficient operation of massive multiple-input multiple-output (mMIMO) systems by providing timely downlink (DL) CSI. While deep learning-based methods have advanced beyond conventional model-driven and statistical approaches, they remain limited in robustness to practical non-Gaussian noise, generalization across diverse channel conditions, and computational efficiency. This paper introduces CSI-4CAST, a hybrid deep learning architecture that integrates 4 key components, i.e., Convolutional neural network residuals, Adaptive correction layers, ShuffleNet blocks, and Transformers, to efficiently capture both local and long-range dependencies in CSI prediction. To enable rigorous evaluation, this work further presents a comprehensive benchmark, CSI-RRG for Regular, Robustness and Generalization testing, which includes more than 300,000 samples across 3,060 realistic scenarios for both TDD and FDD systems. The dataset spans multiple channel models, a wide range of delay spreads and user velocities, and diverse noise types and intensity degrees. Experimental results show that CSI-4CAST achieves superior prediction accuracy with substantially lower computational cost, outperforming baselines in 81.5% of TDD scenarios and 44.4% of FDD scenario, the best performance among all evaluated models, while reducing FLOPs by 5x and 3x compared to LLM4CP, the strongest baseline. In addition, evaluation over CSI-RRG provides valuable insights into how different channel factors affect the performance and generalization capability of deep learning models. Both the dataset (https://huggingface.co/CSI-4CAST) and evaluation protocols (https://github.com/AI4OPT/CSI-4CAST) are publicly released to establish a standardized benchmark and to encourage further research on robust and efficient CSI prediction.

Figures

Figures reproduced from arXiv: 2510.12996 by the authors.

Figure 2
Figure 2. An illustration of the DL CSI acquisition schemas. (a) TDD (b) FDD [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An illustration of the CSI Prediction in the time-frequency domain. problem: by the time DL transmission occurs, the reported CSI is already outdated and no longer reflects current DL conditions. The bottom part of the figure illustrates detailed time-interval breakdowns for TDD and FDD scenarios, highlighting the oc￾currence of channel aging. According to 3GPP procedures [40], in TDD systems, where UL and DL share … view at source ↗
Figure 4
Figure 4. The proposed CSI-4CAST. Historical CSI is first processed by a CNN residual block, followed by an inverse DFT (IDFT) to obtain the delay-domain representation. Both frequency- and delay-domain features are then refined by ACL layers and passed through a ShuffleNet block. Finally, the Transformer block maps the embedded features to predict future CSI. The ACL2 layer (in gray) applies only to the FDD. Here, X C f ∈ C … view at source ↗
Figures from the paper (17 more)
Figure 5
Figure 5. Figure 5: TDD: NMSE rank distribution of Regular, Robustness, and Generalization. Within each panel, models are ordered left to right by their mean rank, MeanRank in (17) (lower is better). Rank distributions are shown as violin plots, while top-1 percentages, Prank1 in (19), ar…
Figure 6
Figure 6. Figure 6: FDD: NMSE rank distribution of Regular, Robustness, and Generalization. The plotting conventions follow those in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: NMSE under varying SNR (dB). Model performance under (a) TDD and (b) FDD is evaluated across SNR from 0 to 25 dB using AWGN. 40.7% in Robustness testing. Under Generalization testing, however, all models exhibit significant degradation compared to TDD. Due to the inher…
Figure 8
Figure 8. Figure 8: NMSE across user velocities. Red dashed vertical lines mark the velocities included in the regular set; all other velocities belong to the generalization set. Light green shading denotes the interpolation region (velocities within the regular range), whereas light red …
Figure 9
Figure 9. Figure 9: NMSE under varying realistic additional noises: TDD and FDD [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: CSI-4CAST: NMSE under Phase and Burst Noise. TDD than in FDD, consistent with the expectation that temporal correlation is more critical in TDD. Moreover, [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: illustrates the overall performance of CSI-4CAST compared to baseline models, jointly evaluating prediction accu￾racy on CSI-RRG and computational overhead. The prediction performance axes are presented based on the RankScore (18). Unlike earlier sections that focused…
Figure 12
Figure 12. Figure 12: ACF across different user velocities (FDD | CDL-A | 30ns) Furthermore, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: ACF across different user velocities (TDD | CDL-A | 30ns) 0 50 100 150 200 250 300 Lag (Subcarriers) 0.0 0.2 0.4 0.6 0.8 1.0 Frequency Autocorrelation (a) 30ns 0 50 100 150 200 250 300 Lag (Subcarriers) 0.0 0.2 0.4 0.6 0.8 1.0 Frequency Autocorrelation (b) 50ns 0 50 1…
Figure 14
Figure 14. Figure 14: Frequency ACF across different delay spreads. (FDD | CDL-C | 1m/s) where 𝜎 2 is the variance of the noise. The relationship between the SNR and the 𝜎 2 is given by: SNR = 10 log10 ∥H∥ 2 𝐹 𝜎2 ! 𝜎 2 = ∥H∥ 2 𝐹 · 10−SNR/10 (23) 0 30 60 90 120 150 180 210 240 270 Frequency…
Figure 16
Figure 16. Figure 16: 2D Frequency ACF across different delay spreads. (TDD | CDL-D | 1m/s) 2) Burst Noise: To better mimic practical channel conditions, burst noise (Fig. 17b) is introduced to simulate sudden spike￾like disturbances that may result from abrupt environmental changes or une…
Figure 17
Figure 17. Figure 17: Visualization of realistic additive noises. For the first dataset sample, we plot the real and imaginary parts of the injected noise on antenna index 0 and subcarrier index 1 for each additional noise type (phase, burst, and packet drop) [PITH_FULL_IMAGE:figures/full…
Figure 18
Figure 18. Figure 18: TDD: SE rank distribution of Regular, Robustness, and Generalization. Within each panel, models are ordered left to right by their mean rank, MeanRank in (17) (lower is better). Rank distributions are shown as violin plots, while top-1 percentages, Prank1 in (19), are…
Figure 19
Figure 19. Figure 19: FDD: SE rank distribution of Regular, Robustness, and Generalization. 0 5 10 15 20 25 SNR 7.0 7.2 7.4 7.6 7.8 8.0 SE CNN LLM4CP CSI-4CAST NP RNN STEMGNN (a) TDD 0 5 10 15 20 25 SNR 6.6 6.8 7.0 7.2 7.4 7.6 7.8 8.0 SE CNN LLM4CP CSI-4CAST NP RNN STEMGNN (b) FDD [PITH_F…
Figure 20
Figure 20. Figure 20: SE under varying SNR of noises. Appendix F Training Configurations Table IX outlines the defined hyperparameter search space and the trainer settings used with the Optuna framework for automated tuning. The optimizer, scheduler, and training settings are shared across…
Figure 21
Figure 21. Figure 21: SE across user velocities. Red dashed vertical lines mark the velocities included in the regular set; all other velocities belong to the generalization set. Light green shading denotes the interpolation region (velocities within the regular range), whereas light red d…
Figure 22
Figure 22. Figure 22: SE under varying realistic additional noises: TDD and FDD. Baseline models follow the official implementations provided by their authors [28,29]. Although a flexible search space is ap￾plied, some constraints remain due to the limited configurability of the original c…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

71 extracted references · 2 canonical work pages

  1. [1]

    Massive mimo systems for 5g and beyond networks—overview, recent trends, challenges, and future research direction,

    R. Chataut and R. Akl, “Massive mimo systems for 5g and beyond networks—overview, recent trends, challenges, and future research direction,”Sensors, vol. 20, no. 10, 2020. [Online]. Available: https://www.mdpi.com/1424-8220/20/10/2753

    2020

  2. [2]

    T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo,Fundamentals of Massive MIMO. Cambridge University Press, 2016

    2016

  3. [3]

    Björnson, J

    E. Björnson, J. Hoydis, and L. Sanguinetti, 2017

    2017

  4. [4]

    Energy and spectral efficiency of very large multiuser mimo systems,

    H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral efficiency of very large multiuser mimo systems,”IEEE Transactions on Communications, vol. 61, no. 4, pp. 1436–1449, 2013

    2013

  5. [5]

    Massive mimo has unlimited capacity,

    E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive mimo has unlimited capacity,”IEEE Transactions on Wireless Communications, vol. 17, no. 1, pp. 574–590, 2018

    2018

  6. [6]

    Optimal design of energy-efficient multi-user mimo systems: Is massive mimo the answer?

    E. Björnson, L. Sanguinetti, J. Hoydis, and M. Debbah, “Optimal design of energy-efficient multi-user mimo systems: Is massive mimo the answer?”IEEE Transactions on Wireless Communications, vol. 14, no. 6, pp. 3059–3075, 2015

    2015

  7. [7]

    Limited feedback beamforming over temporally-correlated channels,

    K. Huang, R. W. Heath, and J. G. Andrews, “Limited feedback beamforming over temporally-correlated channels,”Trans. Sig. Proc., vol. 57, no. 5, pp. 1959–1975, May 2009. [Online]. Available: https://doi.org/10.1109/TSP.2009.2014272

    arXiv 1959

  8. [8]

    Pilot con- tamination and precoding in multi-cell tdd systems,

    J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot con- tamination and precoding in multi-cell tdd systems,”IEEE Transactions on Wireless Communications, vol. 10, no. 8, pp. 2640–2651, 2011

    2011

Show all 71 references
  1. [9]

    Capacity and performance of MIMO systems under the impact of feedback delay

    H. Nguyen, J. Andersen, and G. Pedersen, “Capacity and performance of MIMO systems under the impact of feedback delay.” 2004 IEEE 15th International Symposium on Personal, Indoor and Mobile Radio Communications (IEEE Cat. No.04TH8754), 1 2005. [Online]. Available: https://doi....

    2004 arXiv

  2. [10]

    Effects of channel aging in massive MIMO systems,

    K. T. Truong and R. W. Heath, “Effects of channel aging in massive MIMO systems,”Journal of communications and networks, vol. 15, no. 4, pp. 338–351, 8 2013. [Online]. Available: https://doi.org/10.1109/jcn.2013.000065

    2013

  3. [11]

    Estimating doubly-selective channels for hybrid mmwave massive mimo systems: A doubly-sparse approach,

    S. Gao, X. Cheng, and L. Yang, “Estimating doubly-selective channels for hybrid mmwave massive mimo systems: A doubly-sparse approach,” IEEE Transactions on Wireless Communications, vol. 19, no. 9, pp. 5703–5715, 2020

    2020

  4. [12]

    Performance analysis of MIMO MRC systems with feedback delay and channel estimation error,

    M. Li, M. Lin, W.-P. Zhu, Y. Huang, A. Nallanathan, and Q. Yu, “Performance analysis of MIMO MRC systems with feedback delay and channel estimation error,”IEEE transactions on vehicular technology, vol. 65, no. 2, pp. 707–717, 2 2016. [Online]. Available: https://doi.org/10.11...

    2016

  5. [13]

    Deterministic channel modeling and long range prediction of fast fading mobile radio channels,

    T. Eyceoz, A. Duel-Hallen, and H. Hallen, “Deterministic channel modeling and long range prediction of fast fading mobile radio channels,” IEEE communications letters, vol. 2, no. 9, pp. 254–256, 9 1998. [Online]. Available: https://doi.org/10.1109/4234.718494

    1998

  6. [14]

    Short range wireless channel prediction using local information,

    Z. Shen, J. Andrews, and B. Evans, “Short range wireless channel prediction using local information,” inThe Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, vol. 1, 2003, pp. 1147–1151 Vol.1

    2003

  7. [15]

    Channel prediction in time-varying massive mimo environments,

    W. Peng, M. Zou, and T. Jiang, “Channel prediction in time-varying massive mimo environments,”IEEE Access, vol. 5, pp. 23938–23946, 2017

    2017

  8. [16]

    Performance analysis of (tdd) massive mimo with kalman channel prediction,

    S. Kashyap, C. Mollén, E. Björnson, and E. G. Larsson, “Performance analysis of (tdd) massive mimo with kalman channel prediction,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 3554–3558

    2017

  9. [17]

    Massive MIMO Channel Prediction: Kalman Filtering vs. Machine Learning,

    H. Kim, S. Kim, H. Lee, C. Jang, Y. Choi, and J. Choi, “Massive MIMO Channel Prediction: Kalman Filtering vs. Machine Learning,” IEEE transactions on communications, vol. 69, no. 1, pp. 518–528, 1

  10. [18]

    Addressing the curse of mobility in massive mimo with prony-based angular-delay domain channel predictions,

    H. Yin, H. Wang, Y. Liu, and D. Gesbert, “Addressing the curse of mobility in massive mimo with prony-based angular-delay domain channel predictions,”IEEE Journal on Selected Areas in Communications, vol. 38, no. 12, pp. 2903–2917, 2020. 15

    2020

  11. [19]

    Machine learning for future wireless communications: Channel prediction perspectives,

    H. Kim, J. Choi, and D. J. Love, “Machine learning for future wireless communications: Channel prediction perspectives,” 2025. [Online]. Available: https://arxiv.org/abs/2502.18196

    2025 arXiv

  12. [20]

    A comparison of neural networks for wireless channel prediction,

    O. Stenhammar, G. Fodor, and C. Fischione, “A comparison of neural networks for wireless channel prediction,” 2023. [Online]. Available: https://arxiv.org/abs/2308.14020

    2023 arXiv

  13. [21]

    Simplification of MUSIC and ESPRIT by exploitation of cyclostationarity,

    W. Gardner, “Simplification of MUSIC and ESPRIT by exploitation of cyclostationarity,”Proceedings of the IEEE, vol. 76, no. 7, pp. 845–847, 7 1988. [Online]. Available: https://doi.org/10.1109/5.7152

    1988 doi

  14. [22]

    Ai for csi prediction in 5g-advanced and beyond,

    C. Jiang, J. Guo, X. Li, S. Jin, and J. Zhang, “Ai for csi prediction in 5g-advanced and beyond,” 2025. [Online]. Available: https://arxiv.org/abs/2504.12571

    2025 arXiv

  15. [23]

    Toward intelligent fading channel prediction: A compre- hensive survey,

    R. Adeogun, “Toward intelligent fading channel prediction: A compre- hensive survey,”IEEE Access, vol. 13, pp. 111260–111281, 2025

    2025

  16. [24]

    Lstm-gru model-based channel prediction for one-bit massive mimo system,

    I. Helmy, P. Tarafder, and W. Choi, “Lstm-gru model-based channel prediction for one-bit massive mimo system,”IEEE Transactions on Vehicular Technology, vol. 72, no. 8, pp. 11053–11057, 2023

    2023

  17. [25]

    CV-3DCNN: Complex-Valued Deep Learning for CSI Prediction in FDD Massive MIMO Systems,

    Y. Zhang, J. Wang, J. Sun, B. Adebisi, H. Gacanin, G. Gui, and F. Adachi, “CV-3DCNN: Complex-Valued Deep Learning for CSI Prediction in FDD Massive MIMO Systems,”IEEE wireless communications letters, vol. 10, no. 2, pp. 266–270, 2 2021. [Online]. Available: https://doi.org/10....

    2021

  18. [26]

    Deep UL2DL: Data-Driven Channel Knowledge transfer from uplink to Downlink,

    M. S. Safari, V. Pourahmadi, and S. Sodagari, “Deep UL2DL: Data-Driven Channel Knowledge transfer from uplink to Downlink,” IEEE open journal of vehicular technology, vol. 1, pp. 29–44, 1 2020. [Online]. Available: https://doi.org/10.1109/ojvt.2019.2962631

    2020

  19. [27]

    Accurate channel prediction based on transformer: making mobility negligible,

    H. Jiang, M. Cui, D. W. K. Ng, and L. Dai, “Accurate channel prediction based on transformer: making mobility negligible,”IEEE journal on selected areas in communications, vol. 40, no. 9, pp. 2717–2732, 9

  20. [28]

    Spectral Temporal Graph neural network for massive MIMO CSI prediction,

    S. Mourya, P. Reddy, S. Amuru, and K. K. Kuchi, “Spectral Temporal Graph neural network for massive MIMO CSI prediction,”IEEE wireless communications letters, p. 1, 1 2024. [Online]. Available: https://doi.org/10.1109/lwc.2024.3372148

    2024

  21. [29]

    LLM4CP: Adapting Large Language Models for channel Prediction,

    B. Liu, X. Liu, S. Gao, X. Cheng, and L. Yang, “LLM4CP: Adapting Large Language Models for channel Prediction,” 6 2024. [Online]. Available: https://ieeexplore.ieee.org/document/10582829

    2024

  22. [30]

    Mining limited data sufficiently: A bert-inspired approach for csi time series application in wireless communication and sensing,

    Z. Zhao, F. Meng, H. Li, X. Li, and G. Zhu, “Mining limited data sufficiently: A bert-inspired approach for csi time series application in wireless communication and sensing,”arXiv preprint arXiv:2412.06861, 2024, submitted on 9 Dec 2024. [Online]. Available: https://doi.org/1...

    2024 doi

  23. [31]

    Uplink perfor- mance of time-reversal mrc in massive mimo systems subject to phase noise,

    A. Pitarokoilis, S. K. Mohammed, and E. G. Larsson, “Uplink perfor- mance of time-reversal mrc in massive mimo systems subject to phase noise,”IEEE Transactions on Wireless Communications, vol. 14, no. 2, pp. 711–723, 2015

    2015

  24. [32]

    Effect of realistic oscillator phase noise on the performance of cell-free massive mimo systems,

    I. Zhilin, E. Vinogradov, and I. Akyildiz, “Effect of realistic oscillator phase noise on the performance of cell-free massive mimo systems,”

  25. [33]

    Performance evaluation of mimo systems in a mixture of gaussian noise and impulsive noise,

    A. Li, Y. Wang, W. Xu, and Z. Zhou, “Performance evaluation of mimo systems in a mixture of gaussian noise and impulsive noise,” in APCC/MDMC ’04. The 2004 Joint Conference of the 10th Asia-Pacific Conference on Communications and the 5th International Symposium on Multi-Dimen...

    2004

  26. [34]

    Uplink channel estimation for massive mimo systems with impulsive noise,

    L. Zhou, J. Dai, W. Xu, and C. Chang, “Uplink channel estimation for massive mimo systems with impulsive noise,”IEEE Communications Letters, vol. 25, no. 5, pp. 1534–1538, 2021

    2021

  27. [35]

    Packet losses distributions in 5g networks for pmu-based monitoring systems,

    A. Morato, G. Frigo, and F. Tramarin, “Packet losses distributions in 5g networks for pmu-based monitoring systems,” inProceedings of the 2024 IEEE International Instrumentation and Measurement Technology Conference (I2MTC). Glasgow, United Kingdom: IEEE, May 2024

    2024

  28. [36]

    Are language models actually useful for time series forecasting?

    M. Tan, M. A. Merrill, V. Gupta, T. Althoff, and T. Hartvigsen, “Are language models actually useful for time series forecasting?”arXiv preprint arXiv:2406.16964, 2024, accepted to NeurIPS 2024 (Spotlight). [Online]. Available: https://doi.org/10.48550/arXiv.2406.16964

    2024 doi arXiv

  29. [37]

    Construction and validation of analytical wireless MIMO channel models based on channel measurement data,

    A. A. Kalachikov and N. S. Shelkunov, “Construction and validation of analytical wireless MIMO channel models based on channel measurement data,”2018 XIV International Scientific-Technical Conference on Actual Problems of Electronics Instrument Engineering (APEIE), vol. 2007, ...

    2018

  30. [38]

    Static CSI extraction and application in the tomographic channel model,

    H. Li, Y. Li, S. Zhou, and J. Wang, “Static CSI extraction and application in the tomographic channel model,”China Communications, vol. 16, no. 12, pp. 132–144, 12 2019. [Online]. Available: https://doi.org/10.23919/jcc.2019.12.010

    2019 doi

  31. [39]

    A geometry-based non-stationary MIMO channel model for vehicular communications,

    Y. Ma, L. Yang, and X. Zheng, “A geometry-based non-stationary MIMO channel model for vehicular communications,”China Communications, vol. 15, no. 7, pp. 30–38, 7 2018. [Online]. Available: https: //doi.org/10.1109/cc.2018.8424580

    2018

  32. [40]

    NR; Radio Resource Control (RRC); Protocol specification,

    “NR; Radio Resource Control (RRC); Protocol specification,” 3rd Generation Partnership Project (3GPP), Technical Specification TS 38.331, 2022, release 17, Version V17.2.0. [On- line]. Available: https://portal.3gpp.org/desktopmodules/Specifications/ SpecificationDetails.aspx?...

    2022

  33. [41]

    FDD-RT: A simple CSI acquisition technique via channel reciprocity for FDD massive MIMO downlink,

    H.-W. Liang, W.-H. Chung, and S.-Y. Kuo, “FDD-RT: A simple CSI acquisition technique via channel reciprocity for FDD massive MIMO downlink,”IEEE systems journal, vol. 12, no. 1, pp. 714–724, 3 2018. [Online]. Available: https://doi.org/10.1109/jsyst.2016.2556222

    2018

  34. [42]

    FDD Massive MIMO Channel Spatial Covariance Conversion Using Projection Methods

    L. Miretti, R. L. Cavalcante, and S. Stanczak, “FDD Massive MIMO Channel Spatial Covariance Conversion Using Projection Methods.” 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 4 2018. [Online]. Available: https://doi.org/10.1109/icassp...

    2018

  35. [43]

    Complex cnn csi enhancer for integrated sensing and communications,

    X. Chen, Z. Feng, J. A. Zhang, F. Gao, X. Yuan, Z. Yang, and P. Zhang, “Complex cnn csi enhancer for integrated sensing and communications,” IEEE Journal of Selected Topics in Signal Processing, vol. 18, no. 5, pp. 752–765, 2024

    2024

  36. [44]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,

    K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, “Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,”IEEE Transactions on Image Processing, vol. 26, no. 7, pp. 3142–3155, 2017

    2017

  37. [45]

    T. S. Rappaport,Wireless Communications: Principles and Practice, 2nd ed. Cambridge University Press, 2024

    2024

  38. [46]

    Tse and P

    D. Tse and P. Viswanath,Fundamentals of wireless Communication, 5

  39. [47]

    Multi-resolution csi feedback with deep learning inmassivemimo system,

    Z. Lu, J. Wang, and J. Song, “Multi-resolution csi feedback with deep learning inmassivemimo system,”inICC2020 -2020 IEEEInternational Conference on Communications (ICC), 2020, pp. 1–6

    2020

  40. [48]

    Enhancing deep learning performance of massive mimo csi feedback,

    S. Ji and M. Li, “Enhancing deep learning performance of massive mimo csi feedback,” 2023. [Online]. Available: https://arxiv.org/abs/2208.11333

    2023 arXiv

  41. [49]

    Modelling the 5g energy consumption using real-world data: Energy fingerprint is all you need,

    T. Chen, Y. Wang, H. Chen, Z. Zhao, X. Li, N. Piovesan, G. Zhu, and Q. Shi, “Modelling the 5g energy consumption using real-world data: Energy fingerprint is all you need,” 2024. [Online]. Available: https://arxiv.org/abs/2406.16929

    2024 arXiv

  42. [50]

    Shufflenet: An extremely efficient convolutional neural network for mobile devices,

    X. Zhang, X. Zhou, M. Lin, and J. Sun, “Shufflenet: An extremely efficient convolutional neural network for mobile devices,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    2018

  43. [51]

    Shufflenet v2: Practical guidelines for efficient cnn architecture design,

    N. Ma, X. Zhang, H.-T. Zheng, and J. Sun, “Shufflenet v2: Practical guidelines for efficient cnn architecture design,” inProceedings of the European Conference on Computer Vision (ECCV), September 2018

    2018

  44. [52]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett...

    2017

  45. [53]

    Hoydis, S

    J. Hoydis, S. Cammerer, F. Ait Aoudia, M. Nimier-David, L. Maggi, G. Marcus, A. Vem, and A. Keller, “Sionna,” 2022, https://nvlabs.github.io/sionna/

    2022

  46. [54]

    NR; Physical channels and modulation (3GPP TS 38.211 version 15.3.0 Release 15),

    J. M. Meredith, “NR; Physical channels and modulation (3GPP TS 38.211 version 15.3.0 Release 15),” 3rd Generation Partnership Project (3GPP), Tech. Rep. TS 38.211, October 2018, eTSI TS 138 211 V15.3.0 (2018-10). [Online]. Available: https://portal.3gpp.org/desktopmodules/ Spe...

    2018

  47. [55]

    Study on channel model for frequencies from 0.5 to 100 GHz (3GPP TR 38.901 version 16.1.0 Release 16),

    “Study on channel model for frequencies from 0.5 to 100 GHz (3GPP TR 38.901 version 16.1.0 Release 16),” 3rd Generation Partnership Project (3GPP), Tech. Rep. TR 38.901, Nov. 2020, version 16.1.0. [Online]. Available: https://cdn. standards.iteh.ai/samples/59772/4d8229a2e4c149...

    2020

  48. [56]

    A spatially separable attention mechanism for massive mimo csi feedback,

    S. Mourya, S. Amuru, and K. K. Kuchi, “A spatially separable attention mechanism for massive mimo csi feedback,”IEEE Wireless Communications Letters, vol. 12, no. 1, p. 40–44, Jan. 2023. [Online]. Available: http://dx.doi.org/10.1109/LWC.2022.3216352

    2023

  49. [57]

    Neural Network-Based Fading Channel Prediction: A Comprehensive Overview,

    W. Jiang and H. D. Schotten, “Neural Network-Based Fading Channel Prediction: A Comprehensive Overview,”IEEE access, vol. 7, pp. 118112–118124, 1 2019. [Online]. Available: https: //doi.org/10.1109/access.2019.2937588

    2019

  50. [58]

    A comparison of Wireless Channel Predictors: Artificial Intelligence versus Kalman Filter

    ——, “A comparison of Wireless Channel Predictors: Artificial Intelligence versus Kalman Filter.” 2019 IEEE International 16 Conference on Communications (ICC), 5 2019. [Online]. Available: https://doi.org/10.1109/icc.2019.8761308

    2019

  51. [59]

    Optuna: A next-generation hyperparameter optimization framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next-generation hyperparameter optimization framework,” inProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2019

    2019

  52. [60]

    Enhancing reliability in ai-based csi prediction: A proxy-based performance monitoring approach,

    C. Jiang, J. Guo, C.-K. Wen, and S. Jin, “Enhancing reliability in ai-based csi prediction: A proxy-based performance monitoring approach,”IEEE Transactions on Communications, vol. 73, no. 4, pp. 2602–2615, 2025

    2025

  53. [61]

    Deep learning-based downlink channel prediction for fdd massive mimo system,

    Y. Yang, F. Gao, G. Y. Li, and M. Jian, “Deep learning-based downlink channel prediction for fdd massive mimo system,”IEEE Communications Letters, vol. 23, no. 11, pp. 1994–1998, 2019

    1994

  54. [62]

    Phase noise in modular millimeter wave massive MIMO,

    M. E. Rasekh, M. Abdelghany, U. Madhow, and M. Rodwell, “Phase noise in modular millimeter wave massive MIMO,”IEEE Transactions on Wireless Communications, vol. 20, no. 10, pp. 6522–6535, 4 2021. [Online]. Available: https://doi.org/10.1109/twc.2021.3074911

    2021

  55. [63]

    Joint estimation of channel and oscillator phase noise in MIMO systems,

    H. Mehrpouyan, A. A. Nasir, S. D. Blostein, T. Eriksson, G. K. Karagiannidis, and T. Svensson, “Joint estimation of channel and oscillator phase noise in MIMO systems,”IEEE Transactions on Signal Processing, vol. 60, no. 9, pp. 4790–4807, 6 2012. [Online]. Available: https://d...

    2012

  56. [64]

    Communication in the presence of noise,

    C. Shannon, “Communication in the presence of noise,”Proceedings of the IRE, vol. 37, no. 1, pp. 10–21, 1 1949. [Online]. Available: https://ieeexplore.ieee.org/document/1697831 17 Appendix A Computational Overhead Table VI presents the computational overhead of the proposed m...

    1949

  57. [69]

    Phase Noise:Phase noise (Fig. 17a) is pervasive in practical communication systems and is a crucial factor limiting the performance of high-speed communications, thus requiring model robustness to phase fluctuations [62,63]. The complex element CSI can be represented by gain a...

  58. [70]

    17b) is introduced to simulate sudden spike- like disturbances that may result from abrupt environmental changes or unexpected obstacles between the BS and UE [33,34]

    Burst Noise:To better mimic practical channel conditions, burst noise (Fig. 17b) is introduced to simulate sudden spike- like disturbances that may result from abrupt environmental changes or unexpected obstacles between the BS and UE [33,34]. Currently, burst noise is modeled...

  59. [71]

    17c) refers to the random omission of CSI packets [35]

    Packet drop Noise:Packet drop noise (Fig. 17c) refers to the random omission of CSI packets [35]. For each time step𝑡, whether a packet is dropped is modeled as a realization of a Bernoulli random variable with parameter𝑝𝑑, i.e.,𝑑𝑡∼ Bernoulli(𝑝𝑑). The packet drop noise is then...

  60. [2005]

    Available: https://doi.org/10.1017/cbo9780511807213

    [Online]. Available: https://doi.org/10.1017/cbo9780511807213

    doi

  61. [2021]

    Available: https://doi.org/10.1109/tcomm.2020.3027882

    [Online]. Available: https://doi.org/10.1109/tcomm.2020.3027882

    2020

  62. [2022]

    Available: https://doi.org/10.1109/jsac.2022.3191334

    [Online]. Available: https://doi.org/10.1109/jsac.2022.3191334

    2022

  63. [2025]

    Available: https://arxiv.org/abs/2405.04099

    [Online]. Available: https://arxiv.org/abs/2405.04099

    arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.