Pith. sign in

REVIEW 4 major objections 5 minor 21 references

On the performance of some new Multiuser FSO-MIMO Communication Systems

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A neural receiver with no channel knowledge matches ideal maximum likelihood detection in FSO-MIMO links

desk verdict The single-user DNN detector is plausible and worth a look, but the multiuser interference model is physically wrong, so the headline claims don't hold. read the letter →

arxiv 1909.05147 v1 pith:W7APYTYR submitted 2019-08-25 eess.SP

M. A. Amirabadi This is my paper · ORCID
classification eess.SP
keywords blinddetectiondeepneuralnetworkfree-spaceopticalcommunicationFSO-MIMOGamma-Gammaturbulencemultiuserinterferenceconstellationshapingequalgaincombining
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 claims that a deep neural network can act as a fully blind detector in free-space optical MIMO links, with no channel estimation at all, and still come close to—and in the 4-QAM case match—maximum likelihood detection that assumes perfect channel state information. The claim covers single-user links, multiuser links with channel allocation, and multiuser links with interference, across weak to strong atmospheric turbulence and SISO/MIMO structures with equal-gain and selection combining. A second proposed structure adds a transmitter-side neural network that jointly shapes the constellation, and the paper argues this end-to-end design roughly doubles the improvement over the receiver-only detector. The reason a sympathetic reader would care is that the deep-learning detector removes the equalizer, pilots, and channel-estimation cost, so if the claim holds, practical FSO terminals could cut cost and latency while preserving near-optimal error rates.

What carries the argument

The central object is a feed-forward deep neural network used as a blind detector: two input neurons take the real and imaginary parts of the combined received signal, hidden layers (four layers, forty neurons per layer, ReLU activations) transform it, and $M$ output neurons with softmax cross-entropy produce a one-hot estimate of the transmitted $M$-ary symbol. Training uses labeled one-hot vectors as ground truth, the Adam optimizer with learning rate $0.005$, and a batch size of $264$ over $1000$ iterations. In the joint structure a second DNN sits at the transmitter with $M$ input and two output neurons, and its complex output defines a learned constellation point, shaping the constellation so that turbulence has less effect; the receiver DNN then detects the shaped symbol. The paper's argument is carried by the ability of this pair to learn the nonlinear Gamma-Gamma channel behavior without explicit channel estimates.

What would settle it

Run the same 4-QAM and 16-QAM simulations with the physical superposition model—all users' turbulence-faded optical signals added at each receive aperture before detection—and compare the symbol error rate with the paper's one-target-aperture multiuser results; a material increase in SER would falsify the multiuser-interference claim as stated.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the symbol error rate of a receiver that is given no channel state information—and that does not even estimate the channel blindly—is close to the symbol error rate of a maximum likelihood receiver that knows the turbulence fading exactly, provided the detector is a trained feed-forward deep network trained on labeled one-hot symbols. In the 16-QAM simulations the blind network tracks ML across weak, moderate, and strong Gamma-Gamma turbulence, and with 4-QAM the end-to-end network with transmitter-side constellation shaping matches the QAM-ML result for both SISO and MIMO. Because EGC and SC combining in FSO require no channel information, the comparison is held in the detection stage, and the paper attributes the residual gap to hyperparameters tuned once on the SISO single-user case and reused everywhere else.

Load-bearing premise

The load-bearing premise is the multiuser-interference model, in which one receive aperture is treated as carrying the desired user's signal and the other apertures as carrying interference; in physical FSO reception all users' optical signals add at every aperture, and if that physical summing replaces the assumed separation the multiuser claim may collapse.

Editorial extensions

If this is right

  • For 4-QAM, the end-to-end blind system reaches the symbol error rate of ideal QAM-ML detection in both SISO and MIMO, which implies the pilot overhead and channel-estimation stage can be removed without losing performance.
  • For 16-QAM, the blind detector stays close to QAM-ML over weak, moderate, and strong turbulence, so the claimed saving in complexity, cost, and latency does not require a benign channel.
  • Adding the transmitter-side constellation-shaping DNN improves performance substantially beyond the receiver-only DNN, with the paper attributing roughly half the end-to-end gain to shaping and half to neural detection.
  • Multiuser scheduling on the best channel can beat the single-user case, because the probability that all users simultaneously see bad turbulence falls as the number of users grows.
  • The gap between blind detection and perfect-CSI detection widens as the number of apertures or users grows, which the paper ties to hyperparameters tuned once for the SISO single-user scenario rather than per scenario.

Reading between the lines

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

  • The paper's multiuser-interference model assumes one receive aperture carries the target user while the other apertures carry interferers; in a real FSO link every user's optical field adds at every aperture, so the multiuser-interference results are conditional on a separation that physics does not provide.
  • A fair energy comparison would credit the blind structures with the power saved by not sending pilots for channel estimation; on a fixed total-energy budget the reported advantage over ML with perfect CSI could be larger than the figures show.
  • Because the authors state the hyperparameters were tuned once for the SISO single-user case, re-tuning per number of apertures and users is a direct testable extension that could close the MIMO and multiuser gaps.
  • The learned constellation shaping at the transmitter is a general idea: if it works under Gamma-Gamma turbulence, the same end-to-end training principle should transfer to other fading channels, though the paper only demonstrates FSO.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 deep neural network (DNN) based 'blind' detectors and a joint DNN detector-constellation shaping structure for free-space optical (FSO) MIMO links with intensity modulation and direct detection. It considers single-user and multiuser scenarios (with and without resource allocation), SISO and MIMO configurations, EGC and SC combining, and Gamma-Gamma turbulence from weak to strong. The main claim is that the DNN-based detectors, without channel state information, achieve symbol error rates close to, and for 4-QAM matching, maximum likelihood detection with perfect channel estimation, while allegedly having low complexity, cost, and latency. The results are presented as SER versus Es/N0 curves in Figs. 2 through 6.

Significance. If the results held, the paper would contribute a practical alternative to ML detection with perfect CSI in FSO-MIMO systems, particularly by avoiding explicit channel estimation. The claimed universality across turbulence regimes, combining schemes, and multiuser configurations is attractive. However, the significance is conditional on correcting a physically inconsistent multiuser signal model and on a fair treatment of the training labels used by the proposed 'blind' detector. The paper also ships no code and gives no statistical confidence measures, so the empirical claims are not yet verifiable. The combination of a wrong multiuser model and the unfair comparison with the ML baseline substantially weakens the contribution as it stands.

major comments (4)
  1. [Section II, Eqs. (1)-(5)] The multiuser signal model is physically inconsistent. In an FSO system with multiple users transmitting simultaneously on the same wavelength, the optical intensities from all users are collected by every receive aperture and the photodetector current is a superposition of all users' signals plus noise. Equation (1), however, contains only the target user's symbol x_l and has no summation over users; Eqs. (3)-(5) similarly describe only one user's signal. The description in Section II that 'one of the received N_r signals is the target signal and the others are interference signals' assumes spatial separation of users across receive apertures, which is neither a stated design condition nor the standard FSO multiuser channel. Because the title, abstract, and conclusion foreground multiuser interference, this modeling error is load-bearing: the curves labeled 'multiuser without resource allocation' in Figs. 2-6 simulate a spatially separated reception model, not multiuser FSO-MIMO interference. The physical scenario and the implemented equations are internally inconsistent.
  2. [Section III.A] The claim that the proposed receiver is 'completely blind' is contradicted by the training procedure. Section III.A states that the one-hot vector is 'the label vector required for training the DNN,' meaning the DNN is trained with the true transmitted symbols. The conventional ML receiver is penalized for needing perfect or blind channel estimation, while the DNN is given ground-truth labels without accounting for how those labels are obtained in practice. If the labels require pilot transmission or a prior training phase, the comparison is unfair in terms of power, latency, and spectral efficiency. The paper should either explicitly state that the DNN is a supervised detector trained offline with known labels and compare against a conventional receiver with the same training overhead, or it should use a genuinely blind/unsupervised training objective. As written, the term 'blind' is misleading.
  3. [Section IV] The simulation results are reported without any statistical confidence measures. No error bars, no train/test split, and no number of test symbols are given, so it is impossible to know whether the gap between the DNN curves and the ML curves in Figs. 2-6 is significant or merely a consequence of a particular random seed. Moreover, hyperparameters are tuned once on the SISO single-user case (Section IV) and reused for all other scenarios; Fig. 5 shows the gap grows with the number of apertures, which the authors attribute to this fixed tuning, but no retuning experiments are shown. The paper should provide confidence intervals and describe the data split, the number of Monte Carlo trials, and the test-set size.
  4. [Abstract and Section IV] The complexity claim is asserted but never quantified. The abstract and conclusion state that the proposed structure has 'very few complexity' and 'greatly reduce[s] cost, and processing latency,' yet no analysis of the inference complexity (number of multiply-accumulate operations, number of parameters) or a comparison with the complexity of maximum likelihood detection is provided. The paper also ignores the cost of DNN training and the pilot overhead for obtaining labels. Without a concrete complexity model, the central practical motivation remains unsupported.
minor comments (5)
  1. [Section III.B] The text says 'other DNN hyperparameters are exactly the same as descriptions of section III.B,' but it should refer to Section III.A; as written, it is self-referential and ambiguous.
  2. [Section I, contribution list] Contribution 4 says 'SISO/MIMO (SC/ EEG) structures'; 'EEG' is a typo and should be 'EGC'.
  3. [Section II, Eq. (2)] The Meijer-G function notation in Eq. (2) is incomplete; the arguments of G are not fully specified, which makes the distribution definition difficult to parse.
  4. [Section V] The conclusion says 'a few deferring number of apertures would greatly change the performance'; 'deferring' should be 'differing'.
  5. [References] The reference list contains an unusually high fraction of self-citations ([1]-[5], [17]-[18], and [22]), and [22] is an arXiv preprint whose title includes a nonstandard phrase 'under the shelf of Optical Communication'; the novelty claims in the Introduction should be supported by a broader independent literature review.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the DNN-based detector is trained and tested under a stated Gamma-Gamma simulation, with an independent maximum-likelihood perfect-CSI baseline.

full rationale

The paper does not derive its performance predictions from the claims being made. The DNN detector in Section III is trained on simulated Gamma-Gamma labeled data (one-hot vectors) and evaluated on held-out simulated data; this is standard supervised learning, not fitting an input then re-naming it as a prediction. The comparison baseline, ML with perfect channel estimation (Section II, Eqs. (3)-(5)), is specified independently and implemented from the same channel model, so it provides an external reference. Self-citations appear for standard facts (e.g., Gamma-Gamma distribution [4], Rytov variance [17], hyperparameter tuning [22]), but the paper states the channel model explicitly in Eq. (2) and the training procedure explicitly in Eqs. (6)-(7), so none of these citations carries the argument. The multiuser interference model, where one of the received N_r signals is the target signal and the others are interference signals, is physically questionable for FSO because optical signals superpose at every aperture; this is a modeling and correctness limitation, not a circular reduction. No element of the claimed derivation reduces by construction to its own input.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central performance claims rest on manually tuned DNN hyperparameters, a simplified Gamma-Gamma AWGN channel, an assumed availability of true labels for blind training, and an incorrect multiuser separation-by-aperture model. The free parameters are listed in Table 1; the axioms are not experimentally justified.

free parameters (6)
  • Number of DNN layers = 4
    Manually tuned in Table 1 and reused for all scenarios.
  • Hidden neurons per layer = 40
    Manually tuned in Table 1; affects capacity to learn the detector.
  • Batch size = 264
    Manually tuned in Table 1; training dynamics depend on it.
  • Learning rate = 0.005
    Manually tuned in Table 1; convergence and final accuracy depend on it.
  • Number of iterations = 1000
    Manually tuned in Table 1; stopping criterion for training.
  • Sample size to batch size ratio = 4
    Reported as better in tuning, effectively a tuned training-data quantity.
assumptions (6)
  • ad hoc to paper Multiuser received signal can be separated by receive aperture: each aperture carries one user's signal.
    Section II: 'one of the received N_r signals is the target signal and the others are interference signals.' Physically, users' signals superpose at every aperture, so this is an invented simplification.
  • ad hoc to paper Training labels, i.e., true transmitted one-hot vectors, are available to the 'blind' receiver during training.
    Section III.A: the one-hot vector is the label vector required for training; the paper does not explain how a blind receiver obtains these labels.
  • domain assumption Gamma-Gamma distribution models weak to strong turbulence without pointing errors.
    Section II, Eq. (2); standard FSO model but assumed throughout and not experimentally validated here.
  • domain assumption AWGN with zero mean and variance sigma^2 models photodetection noise.
    Section II, after Eq. (1); neglects scintillation correlation and signal-dependent shot noise.
  • domain assumption EGC and SC do not require channel information for real-valued FSO channels.
    Section II; used to justify the DNN input construction, but no derivation of optimality is given.
  • domain assumption A 4-layer 40-neuron ReLU DNN can approximate the optimal detector for these channels.
    Section IV, Table 1; no approximation error analysis or architecture search beyond manual tuning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the performance of some new Multiuser FSO-MIMO Communication Systems." pith.science (2026). https://pith.science/paper/W7APYTYR

@misc{pith2026190905147,
  author       = {Pith},
  title        = {Pith review of: On the performance of some new Multiuser FSO-MIMO Communication Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W7APYTYR}},
  note         = {Machine review of arXiv:1909.05147}
}
read the original abstract

The practical implementation of maximum likelihood detection is limited by its high complexity as well as requiring perfect channel state information. Although conventional blind detection techniques reduce complexity, they degrade performance and require blind channel state information. In this paper (for the first time), a deep learning based blind detection and a joint blind detection-constellation shaping structure are presented (to solve this problem). This paper (deeply) goes through the problem and discusses several (practical) scenarios, including single user, multiuser with resource (channel) allocation, and multiuser without resource allocation (multiuser interference). In order to show the universality of the proposed systems, wide atmospheric turbulence regimes, from weak to strong are considered, and single input single output, as well as multi-input multi-output structures are considered. Results indicate that without channel estimation, a deep learning based (blind) detector (despite its very few complexity, and despite conventional systems require it), could have a very favorable performance at all around. So, it is expected that practical implementations of the proposed structures greatly reduce cost, and processing latency, while maintaining performance close enough to the outstanding conventional systems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 18 canonical work pages

  1. [1]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2018). Performance comparison of two novel relay-assisted hybrid fso/rf communication systems. IET. COMM

    work page 2018

  2. [2]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2019). A novel hybrid FSO/RF communication system with receive diversity. Optik

    2019

  3. [3]

    Amirabadi, M. A. (2018). Performance analysis of hybrid FSO/RF communication systems with Alamouti Coding or Antenna Selection. IET JOE

    work page 2018

  4. [4]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2019). On the performance of a multi-user multi -hop hybrid FSO/RF communication system. Optics Communications

    work page 2019

  5. [5]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2018). A new optimization problem in FSO communication system. IEEE Communications Letters, 22(7), 1442-1445

    2018

  6. [7]

    Giacoumidis, E., Lin, Y., Wei, J., Aldaya, I., Tsokanos, A., & Barry, L. (2019). Harnessing machine learning for fiber-induced nonlinearity mitigation in long-haul coherent optical OFDM. Future Internet, 11(1), 2

    work page 2019

  7. [8]

    N., Zhong, K., Zhou, X., Al-Arashi, W

    Khan, F. N., Zhong, K., Zhou, X., Al-Arashi, W. H., Yu, C., Lu, C., & Lau, A. P. T. (2017). Joint OSNR monitoring and modulation format identification in digital coherent receivers using deep neural networks. Optics express, 25(15), 17767-17776

    work page 2017

  8. [9]

    C., Jones, R., & Zibar, D

    Thrane, J., Wass, J., Piels, M., Diniz, J. C., Jones, R., & Zibar, D. (2017). Machine learning techniques for optical performance monitoring f rom directly detected PDM -QAM signals. Journal of Lightwave Technology, 35(4), 868- 875

    work page 2017

Show all 21 references
  1. [10]

    L., Samadi, P., Cho, P

    Huang, Y., Gutterman, C. L., Samadi, P., Cho, P. B., Samoud, W., Ware, C., ... & Bergman, K. (2017). Dynamic mitigation of EDFA power excursions with machine learning. Optics express, 25(3), 2245-2258

    2017

  2. [11]

    Rottondi, C., Barletta, L., Giusti, A., & Tornatore, M. (2018). Machine-learning method for quality of transmission prediction of unestablished lightpaths. IEEE/OSA Journal of Optical Communications and Networking, 10(2), A286-A297

    2018

  3. [12]

    (2015, May)

    Zheng, C., Yu, S., & Gu, W. (2015, May). A SVM-based processor for free-space optical communication. In 2015 IEEE 5th International Conference on Electronics Information and Emergency Communication (pp. 30-33). IEEE

    2015

  4. [13]

    Li, Z., & Zhao, X. (2017). BP artificial neural network based wave front correction for sensor-less free space optics communication. Optics Communications, 385, 219-228

    2017

  5. [14]

    & Xin, X

    Tian, Q., Li, Z., Hu, K., Zhu, L., Pan, X., Zhang, Q., ... & Xin, X. (2018). Turbo-coded 16-ary OAM shift keying FSO communication system combining the CNN-based adaptive demodulator. Optics express, 26(21), 27849-27864

    2018

  6. [15]

    Free-space optics optimization models for building sway and atmospheric interference using variable wavelength,

    X. Liu, “Free-space optics optimization models for building sway and atmospheric interference using variable wavelength,” IEEE Transactions on Communications, vol. 57, no. 2, pp. 492–498, 2009

    2009

  7. [16]

    Wolfram, The Wolfram functions site, Available: http://functions.wolfram.com

  8. [17]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2019). Performance of a relay-assisted hybrid FSO/RF communication system. Physical Communication, 100729. 12

    2019

  9. [18]

    A., & Vakili, V

    Amirabadi, M. A., & Vakili, V. T. (2018). Performance evaluation of a novel relay assisted hybrid FSO/RF communication system with receive diversity. IET OPT

    2018

  10. [19]

    R., & Ghassemlooy, Z

    Bhatnagar, M. R., & Ghassemlooy, Z. (2016). Performance analysis of gamma–gamma fading FSO MIMO links with pointing errors. Journal of Lightwave technology, 34(9), 2158-2169

    2016

  11. [20]

    T., Eriksson, T

    Jones, R. T., Eriksson, T. A., Yankov, M. P., & Zibar, D. (2018, September). Deep learning of geometric constellation shaping including fiber nonlinearities. In 2018 European Conference on Optical Communication (ECOC) (pp. 1-3). IEEE

    2018

  12. [21]

    A., Skoog, R

    Wu, X., Jargon, J. A., Skoog, R. A., Paraschis, L., & Willner, A. E. (2009). Applications of artificial neural networks in optical performance monitoring. Journal of Lightwave Technology, 27(16), 3580-3589

    2009

  13. [22]

    Amirabadi, M. A. (2019). Novel Suboptimal approaches for Hyperparameter Tuning of Deep Neural Network [under the shelf of Optical Communication]. arXiv preprint arXiv:1907.00036

    2019 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.