pith. sign in

arxiv: 2604.08458 · v2 · submitted 2026-04-09 · 💻 cs.NI

LITE: Lightweight Channel Gain Estimation with Reduced X-Haul CSI Signaling in O-RAN

David Goez , Marco Piazzola , Giulia Costa , Achiel Colpaert , Rodney Martinez Alonso , Esra Aycan Beyazit , Nina Slamnik-Krijestorac , Johann M. Marquez-Barja
show 1 more author
Miguel Camelo Botero
This is my paper

Pith reviewed 2026-05-10 17:29 UTC · model grok-4.3

classification 💻 cs.NI
keywords O-RANchannel state informationcell-free massive MIMOchannel gain predictionCSI compressionBiLSTMautoencoderX-haul signaling
0
0 comments X

The pith

LITE combines 50% CSI compression with an asymmetric SE-BiLSTM to cut X-haul signaling and model complexity while preserving short-horizon channel gain accuracy in O-RAN.

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

LITE addresses the bandwidth strain from frequent CSI exchanges in cell-free massive MIMO O-RAN setups that otherwise overload X-haul links. The system places a one-dimensional convolutional autoencoder at the distributed unit to compress channel data by half and routes the compact information to an enhanced bidirectional LSTM at the intelligent controller for short-term gain prediction. This combination cuts the predictor's complexity by more than eighty percent relative to a plain bidirectional LSTM while actually raising accuracy by five percent. When the model is trained to account for the compression step, the accuracy drop stays at only six percent, beating both separate compression and full end-to-end alternatives. An optimized version of the pipeline processes queries at 147 thousand per second, four times faster than before.

Core claim

LITE introduces a pipeline that combines a 1-D convolutional autoencoder at the O-DU for CSI compression with a squeeze-and-excitation bidirectional LSTM at the Near-RT-RIC for channel gain forecasting. This setup supports short-horizon trajectory-unaware predictions under tight transport and processing limits in O-RAN. The design achieves 50% compression, 83.39% complexity reduction, 5% accuracy gain over baseline, and only 6% loss with aware training, plus high throughput after optimization, making it compatible with O-RAN splits.

What carries the argument

The LITE pipeline consisting of a 1-D convolutional Autoencoder for 50% CSI compression at the O-DU paired with an asymmetric SE-enhanced BiLSTM predictor at the Near-RT-RIC that performs trajectory-unaware short-horizon channel gain forecasting on the compressed input.

If this is right

  • O-RAN cell-free massive MIMO deployments can operate with half the X-haul CSI traffic without major performance loss.
  • The asymmetric SE-BiLSTM yields both lower complexity and higher accuracy than a standard BiLSTM baseline.
  • Compression-aware training retains nearly all accuracy while independent or end-to-end training does not.
  • TensorRT optimization raises throughput to 147k QPS, enabling real-time operation within O-RAN latency budgets.
  • The resulting predictor integrates directly into existing O-RAN functional splits for practical deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the accuracy holds for longer prediction horizons, LITE could support proactive beamforming or scheduling decisions in faster-changing environments.
  • The same compression-plus-predictor structure might be reused to forecast other O-RAN quantities such as interference levels or traffic load.
  • Real hardware trials that include RF impairments could quantify how much extra training data is needed to keep the 6% accuracy margin.
  • Adoption in multi-vendor O-RAN networks could reduce overall fronthaul energy use by lowering the volume of CSI packets exchanged.

Load-bearing premise

The short-horizon trajectory-unaware forecasts produced under the paper's test conditions will stay accurate enough for O-RAN control loops once real channels include mobility patterns, interference, or hardware impairments absent from the training data.

What would settle it

Running LITE on channel measurement traces that add higher user mobility speeds or extra-cell interference and checking whether prediction error rises substantially above the levels reported for the original test set.

Figures

Figures reproduced from arXiv: 2604.08458 by Achiel Colpaert, David Goez, Esra Aycan Beyazit, Giulia Costa, Johann M. Marquez-Barja, Marco Piazzola, Miguel Camelo Botero, Nina Slamnik-Krijestorac, Rodney Martinez Alonso.

Figure 1
Figure 1. Figure 1: Overview of the LITE system architecture, illustrating the four processing layers [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Detailed LITE DL architecture showing the symmetric 1-D convolu [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Step-wise Pearson correlation of CSI trajectories for increasing dataset [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
read the original abstract

Cell-Free Massive Multiple-Input Multiple-Output (CF-MaMIMO) in Open Radio Access Network (O-RAN) promises high spectral efficiency but is limited by frequent Channel State Information (CSI) exchanges, which strain fronthaul/midhaul/backhaul (X-haul) bandwidth and exceed the capabilities of existing approaches relying on uncompressed CSI or heavy predictors. To overcome these constraints, we propose LITE, a lightweight pipeline combining a 1-D convolutional Autoencoder (AE) at the O-RAN Distributed Unit (O-DU) with a Squeeze-and-Excitation (SE)-enhanced Bidirectional Long Short-Term Memory (BiLSTM) predictor at the Near-Real-Time RAN Intelligent Controller (Near-RT-RIC), enabling short-horizon trajectory-unaware forecasting under strict transport and processing budgets. LITE applies 50% CSI compression and an asymmetric SE-BiLSTM, reducing model complexity by 83.39% while improving accuracy by 5% relative to a baseline BiLSTM. With compression-aware training, the Lightweight Intelligent Trajectory Estimator (LITE) incurs only 6% accuracy loss versus the BiLSTM baseline, outperforming independent and end-to-end strategies. A TensorRT-optimized implementation achieves 147k Queries per Second (QPS), a 4.6x throughput gain. These results demonstrate that LITE delivers X-haul-efficient, low-latency, and deployment-ready channel-gain prediction compatible with O-RAN splits.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes LITE, a lightweight pipeline for channel gain estimation in cell-free massive MIMO O-RAN systems. It deploys a 1-D convolutional autoencoder at the O-DU for 50% CSI compression and an asymmetric SE-enhanced BiLSTM predictor at the Near-RT-RIC for short-horizon trajectory-unaware forecasting. The central claims are a 83.39% reduction in model complexity, 5% accuracy improvement over a baseline BiLSTM, only 6% accuracy loss under compression-aware training (outperforming independent and end-to-end strategies), and 147k QPS throughput (4.6x gain) after TensorRT optimization, enabling reduced X-haul signaling while remaining deployment-ready.

Significance. If the quantitative results prove reproducible and the short-horizon predictions generalize, LITE could meaningfully alleviate X-haul bandwidth pressure in O-RAN CF-MaMIMO deployments while supporting low-latency control loops. The asymmetric SE-BiLSTM design and compression-aware training offer a concrete complexity-accuracy trade-off, and the reported QPS figure indicates practical inference speed. These elements address a recognized practical bottleneck; however, the current lack of methodological transparency limits assessment of whether the gains are robust enough to support the deployment-ready assertion.

major comments (3)
  1. Abstract: quantitative performance claims (50% CSI compression, 83.39% complexity reduction, 5% accuracy gain, 6% loss under aware training, 147k QPS) are presented without any dataset description, baseline implementation details, statistical tests, or ablation results on the SE blocks, rendering the numbers unverifiable from the provided text.
  2. Evaluation (implied by abstract claims): the 5% accuracy improvement and 6% loss figures are reported against an external BiLSTM baseline with no indication that the comparison uses identically fitted quantities or the same training/validation splits, undermining the cross-strategy superiority statements.
  3. Abstract / forecasting description: the short-horizon trajectory-unaware forecasting is asserted to be deployment-ready, yet no evaluation incorporates 3GPP mobility traces, multi-user interference, or RF impairments (phase noise, I/Q imbalance); if these increase NMSE beyond the 6% margin, the X-haul savings and control-loop compatibility do not follow.
minor comments (1)
  1. Abstract: the parenthetical expansion 'Lightweight Intelligent Trajectory Estimator (LITE)' appears after the title's 'Lightweight Channel Gain Estimation'; a single consistent expansion would reduce minor confusion.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point-by-point below, providing clarifications based on the full paper content and indicating where revisions will be made to enhance transparency.

read point-by-point responses

  1. Referee: Abstract: quantitative performance claims (50% CSI compression, 83.39% complexity reduction, 5% accuracy gain, 6% loss under aware training, 147k QPS) are presented without any dataset description, baseline implementation details, statistical tests, or ablation results on the SE blocks, rendering the numbers unverifiable from the provided text.

    Authors: The abstract is length-constrained by design, but the full manuscript details the dataset and channel model in Section III-A, baseline BiLSTM implementation and hyperparameters in Section IV-B, statistical significance via 10 independent runs with mean/std in Section V-A, and SE-block ablations in Section V-C. We will revise the abstract to include a brief evaluation-setup sentence and explicit section cross-references so that the quantitative claims are immediately traceable. revision: yes

  2. Referee: Evaluation (implied by abstract claims): the 5% accuracy improvement and 6% loss figures are reported against an external BiLSTM baseline with no indication that the comparison uses identically fitted quantities or the same training/validation splits, undermining the cross-strategy superiority statements.

    Authors: All strategies (baseline BiLSTM, independent training, end-to-end, and compression-aware) were trained and tested on identical data splits and with the same random seeds, as stated in Section IV-C. The reported 5% gain and 6% loss are computed on the same held-out test set. We will add an explicit sentence in the revised abstract and evaluation section confirming identical fitting procedures and data partitions. revision: yes

  3. Referee: Abstract / forecasting description: the short-horizon trajectory-unaware forecasting is asserted to be deployment-ready, yet no evaluation incorporates 3GPP mobility traces, multi-user interference, or RF impairments (phase noise, I/Q imbalance); if these increase NMSE beyond the 6% margin, the X-haul savings and control-loop compatibility do not follow.

    Authors: The current results are obtained on the synthetic CF-MaMIMO dataset and channel model described in Section III, which does not include 3GPP mobility traces or explicit RF impairments. The deployment-ready claim rests on the measured complexity reduction, 147k QPS throughput, and accuracy under the evaluated conditions. We will add a dedicated limitations paragraph discussing the potential impact of these unmodeled factors on NMSE and X-haul savings, while preserving the core contribution on the compression-prediction trade-off. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical ML results against external baseline

full rationale

The paper reports experimental outcomes for a 1-D convolutional AE + asymmetric SE-BiLSTM pipeline on CSI data, with accuracy, complexity, and throughput metrics measured relative to a separate baseline BiLSTM model. No derivation chain reduces a claimed prediction to its own fitted inputs by construction, nor does any load-bearing step rely on self-citation of an unverified uniqueness result. Performance numbers (5% accuracy gain, 6% loss under compression, 83.39% complexity reduction, 4.6x QPS) are obtained via standard train/test evaluation and TensorRT benchmarking, remaining independent of the quantities being predicted.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on standard supervised learning assumptions (i.i.d. training/validation splits, gradient-based optimization) and O-RAN functional splits; no explicit free parameters, new axioms, or invented physical entities are introduced beyond the neural network architecture itself.

pith-pipeline@v0.9.0 · 5612 in / 1293 out tokens · 60693 ms · 2026-05-10T17:29:27.336568+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

  1. [1]

    Ericsson mobility report: November 2025,

    Ericsson, “Ericsson mobility report: November 2025,” https://www.ericsson.com/en/reports-and-papers/mobility-report/ reports/november-2025, November 2025, accessed 10 January 2026

    work page 2025

  2. [2]

    Next- generation multiple access with cell-free massive mimo,

    M. Mohammadi, Z. Mobini, H. Quoc Ngo, and M. Matthaiou, “Next- generation multiple access with cell-free massive mimo,”Proceedings of the IEEE, vol. 112, no. 9, pp. 1372–1420, 2024

    work page 2024

  3. [3]

    Spectral efficiency analysis of cell-free massive mimo systems with zero-forcing detector,

    P. Liu, K. Luo, D. Chen, and T. Jiang, “Spectral efficiency analysis of cell-free massive mimo systems with zero-forcing detector,”IEEE Transactions on Wireless Communications, vol. 19, no. 2, pp. 795–807, 2020

    work page 2020

  4. [4]

    Scalable cell-free massive mimo systems,

    E. Björnson and L. Sanguinetti, “Scalable cell-free massive mimo systems,”IEEE Transactions on Communications, vol. 68, no. 7, pp. 4247–4261, 2020

    work page 2020

  5. [5]

    A survey on user- centric cell-free massive mimo systems,

    S. Chen, J. Zhang, J. Zhang, E. Björnson, and B. Ai, “A survey on user- centric cell-free massive mimo systems,”Digital Communications and Networks, vol. 8, no. 5, pp. 695–719, 2022

    work page 2022

  6. [6]

    Structured massive access for scalable cell-free massive mimo systems,

    S. Chen, J. Zhang, E. Björnson, J. Zhang, and B. Ai, “Structured massive access for scalable cell-free massive mimo systems,”IEEE Journal on Selected Areas in Communications, vol. 39, no. 4, pp. 1086–1100, 2021

    work page 2021

  7. [7]

    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, 2022

    work page 2022

  8. [8]

    Trajectory-unaware channel gain forecast in a distributed massive mimo system based on a multivariate bilstm model,

    R. Martinez Alonso, R. Beerten, A. Colpaert, A. P. Guevara, and S. Pollin, “Trajectory-unaware channel gain forecast in a distributed massive mimo system based on a multivariate bilstm model,”IEEE Open Journal of the Communications Society, vol. 5, pp. 5348–5363, 2024

    work page 2024

  9. [9]

    Mobile cell-free massive mimo: A practical o-ran-based approach,

    R. Beerten, V . Ranjbar, K. A. P. Guevara, and S. Pollin, “Mobile cell-free massive mimo: A practical o-ran-based approach,”IEEE Open Journal of the Communications Society, vol. 6, pp. 593–610, 2025

    work page 2025

  10. [10]

    Beamforming and functional split selection for scalable cell-free mmimo networks,

    A. Girycki, M. A. Rahman, A. P. Guevara, and S. Pollin, “Beamforming and functional split selection for scalable cell-free mmimo networks,” in2024 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit), 2024, pp. 1–6

    work page 2024

  11. [11]

    Model-driven deep learning based precoding for fdd cell-free massive mimo with imperfect csi,

    S. Liu, Z. Gao, C. Hu, S. Tan, L. Fang, and L. Qiao, “Model-driven deep learning based precoding for fdd cell-free massive mimo with imperfect csi,” in2022 International Wireless Communications and Mobile Computing (IWCMC), 2022, pp. 696–701

    work page 2022

  12. [12]

    Adaptive compression of massive mimo channel state information with deep learning,

    F. B. Mismar and A. O. Kaya, “Adaptive compression of massive mimo channel state information with deep learning,”IEEE Networking Letters, vol. 6, no. 4, pp. 267–271, 2024

    work page 2024

  13. [13]

    A semi-supervised learning approach towards automatic wireless technology recognition,

    M. Camelo, A. Shahid, J. Fontaine, F. A. P. de Figueiredo, E. De Poorter, I. Moerman, and S. Latre, “A semi-supervised learning approach towards automatic wireless technology recognition,” in2019 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), 2019, pp. 1–10

    work page 2019

  14. [14]

    An ai- based incumbent protection system for collaborative intelligent radio networks,

    M. Camelo, R. Mennes, A. Shahid, J. Struye, C. Donato, I. Jabandzic, S. Giannoulis, F. Mahfoudhi, P. Maddala, I. Seskaret al., “An ai- based incumbent protection system for collaborative intelligent radio networks,”IEEE Wireless Communications, vol. 27, no. 5, pp. 16–23, 2020

    work page 2020

  15. [15]

    Edge inference for uwb ranging error correction using autoencoders,

    J. Fontaine, M. Ridolfi, B. Van Herbruggen, A. Shahid, and E. De Poorter, “Edge inference for uwb ranging error correction using autoencoders,”IEEE Access, vol. 8, pp. 139 143–139 155, 2020

    work page 2020

  16. [16]

    Bidirectional recurrent neural networks,

    M. Schuster and K. Paliwal, “Bidirectional recurrent neural networks,” IEEE Transactions on Signal Processing, vol. 45, no. 11, pp. 2673–2681, 1997

    work page 1997

  17. [17]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 7132–7141

    work page 2018

  18. [18]

    Ultra dense indoor mamimo csi dataset,

    S. D. Bast and S. Pollin, “Ultra dense indoor mamimo csi dataset,”

    work page

  19. [19]

    Available: https://dx.doi.org/10.21227/nr6k-8r78

    [Online]. Available: https://dx.doi.org/10.21227/nr6k-8r78

    work page doi:10.21227/nr6k-8r78

  20. [20]

    On the diversity of synthetic data and its impact on training large language models,

    H. Chen, A. Waheed, X. Li, Y . Wang, J. Wang, B. Raj, and M. I. Abdin, “On the diversity of synthetic data and its impact on training large language models,” 2024

    work page 2024