Pith. sign in

REVIEW 5 major objections 5 minor 52 references

Cooperative 3D positioning can reach 0.48 m mean error with only 6.25% of the CSI fronthaul payload.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 05:55 UTC pith:DDY2VDZR

load-bearing objection A solid, honest simulation study on fronthaul-efficient cooperative positioning, but the headline accuracy is only demonstrated in-distribution with no error bars, so treat the numbers as indicative rather than validated. the 5 major comments →

arxiv 2602.00664 v2 pith:DDY2VDZR submitted 2026-01-31 eess.SP

Fronthaul-Efficient Distributed Cooperative 3D Positioning with Quantized Latent CSI Embeddings

Tong An , Jiwei Zhao , Jiayang Shi , Bin Zheng , Kai Yu , Maged Elkashlan , George K. Karagiannidis , Hongsheng Chen This is my paper
classification eess.SP MSC 68T0794A12
keywords 3D positioningcooperative localizationCSI compressionquantized embeddingsfronthauldeep learningNLOSMIMO-OFDM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Cooperative positioning across multiple base stations sharpens 3D localization in dense urban non-line-of-sight environments, but forwarding raw channel-state information (CSI) to a central unit quickly overwhelms fronthaul links. This paper claims that each base station can instead learn a compact, quantized latent embedding of its local CSI and transmit only that fixed-length bitstream; the central unit then infers the user's 3D position by jointly processing the bitstreams from all base stations. In a 5G-NR-compliant ray-traced urban scenario with six base stations, the proposed edge-cloud cooperative (ECC) framework achieves a mean 3D error of 0.48 m and a 90th-percentile error of 0.83 m while using only 6.25% of the payload of lossless CSI forwarding—close to the 0.42 m / 0.75 m of full-CSI cooperative positioning. The significance, if the result holds outside simulation, is that high-accuracy cooperative localization need not scale fronthaul cost with CSI dimensionality, which is the main obstacle to practical deployment.

Core claim

The central discovery is that a task-oriented, quantized CSI embedding can carry almost all the location-relevant information across a fronthaul link. Each base station gain-normalizes and phase-stabilizes its estimated CSI, encodes it with a residual SegNet-style autoencoder into a frequency-ordered latent vector, and uniformly quantizes it (Q bits per coefficient). At Q=10 the central unit's fusion of the resulting bitstreams yields a mean error of 0.48 m, within 0.06 m of the lossless-CSI reference of 0.42 m, while the payload drops to 6.25%. The authors also show that a two-stage training procedure—self-supervised reconstruction at the base stations, then end-to-end joint training throug

What carries the argument

The load-bearing component is the quantized latent CSI embedding: a frequency-ordered latent vector per base station, produced by a residual SegNet-style encoder and mapped by uniform midrise quantization to a DQ-bit bitstream. The embedding preserves the multipath phase structure across antennas and subcarriers—the information the central unit exploits—while discarding dynamic range and redundancy. At the central unit, channel-masked attention weights each base station's tokens by its Frobenius-norm gain indicator, suppressing weak NLOS links, and an LSTM accumulates evidence across subcarriers before a regression head estimates the 3D position.

Load-bearing premise

The reported accuracy is produced and evaluated entirely within a ray-tracing simulation whose training, validation, and test data are generated on the fly from the same channel statistics; if the simulated 3.5 GHz urban NLOS geometry and propagation do not match a real deployment, or if the environment changes after training, the 0.48 m mean error and the 6.25%-payload trade-off are not guaranteed in the field.

What would settle it

Take the trained ECC model and run it on measured 3.5 GHz uplink CSI from a real six-base-station urban deployment (or from an independent ray-tracer with different building geometry). If the mean 3D error exceeds roughly 1.5 m—three times the reported 0.48 m—then the headline accuracy is an artifact of the simulation distribution rather than a property of the learned embedding.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Fronthaul cost per base station becomes a fixed DQ bits per snapshot, independent of the number of antennas, subcarriers, or time slots, so cooperative positioning can scale to many more base stations without linear fronthaul growth.
  • Reconstruction-oriented CSI feedback—designed for beamforming—is not required for positioning; a task-oriented embedding can carry the needed information, suggesting an integrated design where one compressed representation serves both communication and sensing.
  • The 0.48 m mean error over a 220 m x 300 m region with user heights up to 30 m indicates the method is viable for low-altitude UAV tracking and street-level localization in dense urban canyons.
  • The observed monotonic gain from accumulating more subcarriers implies that positioning accuracy in the compressed regime improves with bandwidth, giving operators a direct knob to trade spectrum for accuracy.

Where Pith is reading between the lines

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

  • The payload comparison assumes lossless CSI is sent as 64-bit complex values; if a practical fronthaul already employs a high-efficiency CSI codec, the relative advantage of the learned embedding could shrink—a head-to-head comparison against a reconstruction-oriented codec at equal bit rate would clarify the real gain.
  • Because training and test snapshots come from the same ray-traced distribution, the 0.48 m figure does not reflect environmental drift; a calibration or fine-tuning step on a small amount of real-world data would likely be needed before deployment, and the method's adaptation cost is untested.
  • The method estimates a single user position; extending the same bitstreams to multi-user scenarios or to joint position-and-velocity inference is a natural next step, since the embedding retains per-subcarrier and per-slot structure that may encode Doppler and spatial separation.
  • The channel-masked attention only uses a scalar gain per base station; passing richer per-subcarrier reliability side information over the same bit budget might close the remaining 0.06 m gap to lossless CSI at even lower payloads.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes ECC, a learning-based edge-cloud cooperative 3D positioning framework for fronthaul-constrained C-RAN. Each BS compresses locally estimated MIMO-OFDM CSI via a residual SegNet encoder into a fixed-length latent vector, which is uniformly quantized into a bitstream and transmitted to a CU; the CU fuses the bitstreams from multiple BSs using channel-masked attention and an LSTM-based frequency accumulation network, and regresses the UE's 3D position. Training is two-stage: Stage I self-supervised reconstruction at each BS (cosine loss), Stage II end-to-end joint training with straight-through estimation through the quantizer. Simulations use a Sionna ray-traced 3.5 GHz urban scene (Yuquan campus, six BSs, 20 MHz, effective 24 subcarriers, 10 time slots). With Q=10 bits per latent coefficient, the payload is 6.25% of lossless CSI forwarding, achieving 0.48 m mean and 0.83 m 90th-percentile 3D error versus 0.42 m / 0.75 m for lossless CSI. Ablations show benefit of two-stage training, CMA masking, and frequency accumulation; comparisons against MUSIC, ADCPM, and MFCNet are presented.

Significance. If the reported accuracy holds in deployment, the paper offers a useful systems contribution: a task-oriented CSI embedding that preserves positioning information at a small fraction of raw-CSI fronthaul, with ablations (Tables II, Figs. 10-12) supporting the architectural choices. The quantitative claim (0.48 m at 6.25% payload, close to 0.42 m lossless) is a credible simulation-based benchmark. The paper is internally consistent: the lossless reference, quantization trade-off, and ablation comparisons are aligned, and the two-stage training rationale is clearly presented. Strengths include a well-specified system model, explicit bit-budget accounting (Eq. 39), a reproducible-style simulation protocol (on-the-fly sample generation), and comparison with standard baselines. However, the evaluation is entirely in-distribution on one simulated scene, and several load-bearing details (quantizer step calibration, lack of error bars, no cross-scene validation) limit the strength of the headline claim until generalized.

major comments (5)
  1. [Table II, Figs. 9-11] The headline trade-off (0.48 m at 6.25%) is demonstrated only under in-distribution evaluation in a single Sionna ray-traced scene. Section IV-A(c) states that training, validation, and test samples are generated on the fly from the same simulator distribution. Because the learned embeddings and the quantizer step Δ are calibrated on that distribution (Table I: 'calibrated on training set'), the result presupposes that the simulated urban NLOS statistics match deployment. The paper provides no cross-scene, cross-frequency, or real-world validation, and no code/data release. This is the central generalization risk for the paper's practical claim; at minimum it should be stated clearly as a limitation, and ideally supplemented by a second scene or an out-of-distribution test. As is, the paper's wording that the protocol 'reduces the risk of memorization' addresses overfitting to a finite d
  2. [IV-A(c), Table II] All reported accuracy figures (mean and 90th percentile) are given without error bars, confidence intervals, or repeated-seed statistics. Since training and test samples are generated on the fly with random UE positions and channel realizations, the difference between ECCNet (0.48 m) and the lossless-CSI reference (0.42 m) could be within run-to-run variability. Please report standard deviations or confidence intervals across at least a few random seeds / test draws, and state the test sample count. This is load-bearing for the claim that the degradation from lossless CSI is small.
  3. [II-C, Table I] The quantization step size Δ is described as 'calibrated on training set,' but the calibration procedure is not specified: how is Δ chosen for each Q and for the fixed latent dimension? If Δ is chosen by a grid search on validation error, then it is a hyperparameter and should be reported along with its sensitivity. More importantly, since Δ is fixed at inference, a shift in the latent distribution (e.g., different SNR, UE height range, or environment) can cause the uniform midrise quantizer in Eq. (19) to clip or under-use the codebook. The paper should either include an adaptive/scaling mechanism or provide a robustness experiment varying the channel statistics (e.g., noise level, UE height distribution) to show that the chosen Δ is not brittle.
  4. [IV-B, Table II, Algorithm 2] The payload ratio η in Eq. (39) excludes the gain side-information Bg (32 bits per snapshot per BS). While the gain is a small fraction of the 6.25% payload, Table II reports η excluding Bg, and the abstract says '6.25% of lossless CSI forwarding.' For a fair system-level accounting, the total ratio ηtot = (DQ+Bg)/(64TNrNsc) should be reported. With Bg=32, T=10, Nr=8, Nsc=24, BCSI=122,880 bits, so Bg is 0.026% of BCSI; this is negligible absolute, but the text should be precise about what is included. More importantly, the gain (Eq. 7) is not quantized in the default simulation (Step 4 of Alg. 2: 'set g_hat <- g'), and the paper does not study the effect of Bg quantization; since g is used for CMA weighting, this should be stated.
  5. [IV-E] The comparison with baselines is a strength, but the fairness of the MUSIC/AoA baseline depends on assumed array calibration and geometric mapping that are not detailed. Also, the Transformer baseline (standard self-attention) is said to follow [51] but no description of its depth/training is given. The comparison would be stronger if the paper reported the hyperparameters used for all baselines (e.g., number of layers, learning rate, epochs) and if it reported accuracy for the 'Transformer' baseline at the same payload. Currently the reader cannot assess whether differences are due to architecture or tuning.
minor comments (5)
  1. [III-B] The CMA notation is confusing. Eq. (24) defines Bl with rows for b(0) and b_n^(l), but in Eq. (27) Attn is applied without clear indexing over subcarrier tokens n; the text says 'the last-row output' but does not explain how the L+1 rows are formed for a given n. Clarify whether attention is applied per-subcarrier across BSs or jointly across BSs and subcarriers. Also, Eq. (26) uses diag([1, m1, ..., mL]) which has dimension L+1, but b(0) weights are not gated; fine, but should be stated.
  2. [III-B, IV-A] Figures 4 and 5 are schematic but lack numeric dimensions or layer counts (e.g., number of layers n, LSTM hidden size). Since the paper emphasizes the design, please provide the layer configuration in the text, table, or caption.
  3. [IV-C] In Fig. 8, the validation mean error converges to ~0.5 m for Q=10 after ~210 epochs. Specify the number of epochs used for the final models reported in Table II. Also, Stage I says convergence after ~380 epochs; confirm that the final Stage I encoder is used for Stage II.
  4. [IV-G] The qualitative reconstruction visualization (Fig. 13) is useful but the cosine-similarity loss in (22) is scale-invariant, so the claim 'amplitude scale may differ' is expected; this is more of a presentation note than a flaw. The trajectory tracking results (0.40 m vs 1.37 m) are reported without indicating whether they are from a single trajectory run; add error bars or repeated-run values.
  5. [References] Some references are to journals with odd pagination (e.g., [35] 'pg. 3579-3579'); please verify. Also, [49] is an arXiv paper, fine, but the version number should be cited consistently.

Circularity Check

0 steps flagged

No significant circularity: the reported accuracy is an empirical simulation result benchmarked against an external lossless-CSI reference, not a quantity forced by construction.

full rationale

The central claims (0.48 m mean 3D error at 6.25% fronthaul payload, close to 0.42 m with lossless CSI) are empirical results from simulating the proposed pipeline in Sionna and comparing against a full-CSI baseline (Table II). No equation defines the predicted position in terms of the reported error, and no fitted parameter is renamed as a prediction. The quantization step size is listed as "calibrated on training set" (Table I), which is a standard hyperparameter choice, while the test statistics are computed over an independently generated test set: "all CDFs and summary statistics are computed over an independently generated test set." The frequency-domain motivation in Eq. (31) is a standard Fourier phase progression used only to motivate an LSTM sequence model, not to derive the localization result. Self-citations, e.g., [4] and [7], appear only as background on cloud-RAN architectures and are not load-bearing for the positioning claim. The main limitation is that training, validation, and test snapshots are generated on the fly from the same Sionna ray-tracing distribution, so simulator-to-deployment generalization is not demonstrated; this is a correctness/generalization risk, not circularity. No uniqueness theorem is imported, no ansatz is smuggled in via a self-citation, and no known result is re-presented as a derivation from its own conclusion.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central performance claim rests on standard channel estimation, a simulator-faithfulness assumption, fixed covariance knowledge, and several hand-set hyperparameters (Δ, β, A, d_z, subcarrier subsampling). No new physical entities are introduced; the evaluation is in-distribution simulation, so generalization to real deployments is the key unvalidated premise.

free parameters (5)
  • Quantization step size Δ
    Table I: 'Quantization step size (Δ) calibrated on training set.' It sets the distortion of the uniform midrise quantizer in Eq. (19) and directly controls the accuracy-fronthaul trade-off reported in Table II; no sensitivity analysis is given.
  • Temperature β in softmax gain weighting
    Eq. (23) defines m = softmax(β ĝ); β controls how aggressively weak links are downweighted. Its value is never reported, yet it shapes the CMA fusion that the paper credits for NLOS robustness.
  • STE clipping threshold A
    Eq. (35) defines the straight-through estimator with gradient 1 for |y| ≤ A and 0 otherwise; the heuristic threshold A is never given, and it affects how the quantization bottleneck is trained.
  • Latent dimension per subcarrier d_z = 32
    Table I sets d_z=32 (D=768). Together with Q, this determines the payload; the choice is not derived from rate-distortion analysis and is a hand-set capacity knob.
  • Subcarrier subsampling factor (1 subcarrier per RB) = 12 (Nsc=24 from N_full_sc=288)
    Section IV-A-a chooses to retain one subcarrier per RB; this reduces CSI dimensionality and the lossless baseline is computed on the subsampled grid. The paper does not analyze how this choice affects positioning accuracy versus full-grid operation.
axioms (5)
  • standard math LMMSE estimation under a zero-mean circularly symmetric complex Gaussian prior (Eq. 4)
    Section II-B assumes h ~ CN(0,R) and uses the standard LMMSE estimator; the paper acknowledges it is a 'tractable second-order surrogate' and does not require physical Gaussianity.
  • domain assumption Sionna ray-tracing channels with up to 10 specular reflections accurately model dense urban 3.5 GHz NLOS propagation
    Section IV-A uses this to generate all data; external validity of the headline errors rests on this assumption.
  • domain assumption R_{h,n} is known at each BS via long-term averaging and remains fixed during training and testing
    Section II-B and IV-A-b state R_h,n is estimated by long-term averaging and kept fixed. Stale covariance in deployment would shift the input distribution.
  • domain assumption UE position is quasi-static over T=10 slots and per-(t,n) phase stabilization removes unknown oscillator/hardware phase offsets
    Section II-A and Eq. (9)-(11); residual phase errors after the reference-antenna rotation are assumed negligible for fingerprinting.
  • ad hoc to paper The two-stage training strategy with straight-through estimation produces a stable, near-optimal solution
    Stage I + Stage II with STE is a heuristic introduced by the paper (Section III-C, Eq. 35); there is no convergence guarantee or optimality certificate.

pith-pipeline@v1.3.0-alltime-deepseek · 18710 in / 13493 out tokens · 141513 ms · 2026-08-03T05:55:19.608477+00:00 · methodology

0 comments
read the original abstract

High-precision three-dimensional (3D) positioning in dense urban non-line-of-sight (NLOS) environments benefits significantly from cooperation among multiple distributed base stations (BSs). However, forwarding raw CSI from multiple BSs to a central unit (CU) incurs prohibitive fronthaul overhead, which limits scalable cooperative positioning in practice. This paper proposes a learning-based edge-cloud cooperative positioning framework under limited-capacity fronthaul constraints. In the proposed architecture, a neural network is deployed at each BS to compress the locally estimated CSI into a quantized representation subject to a fixed fronthaul payload. The quantized CSI is transmitted to the CU, which performs cooperative 3D positioning by jointly processing the compressed CSI received from multiple BSs. The proposed framework adopts a two-stage training strategy consisting of self-supervised local training at the BSs and end-to-end joint training for positioning at the CU. Simulation results based on a 3.5~GHz 5G NR compliant urban ray-tracing scenario with six BSs and 20~MHz bandwidth show that the proposed method achieves a mean 3D positioning error of 0.48~m and a 90th-percentile error of 0.83~m, while reducing the fronthaul payload to 6.25% of lossless CSI forwarding. The achieved performance is close to that of cooperative positioning with full CSI exchange.

Figures

Figures reproduced from arXiv: 2602.00664 by Bin Zheng, George K. Karagiannidis, Hongsheng Chen, Jiayang Shi, Jiwei Zhao, Kai Yu, Maged Elkashlan, Tong An.

Figure 1
Figure 1. Figure 1: Cloud-RAN architecture for cooperative 3D position [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The proposed ECC positioning framework. At the edges [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The proposed embedding network (Res-SegNet). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The proposed CMA mechanism. 1) Channel-Masked Attention (CMA): As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The proposed frequency evidence accumulation mech [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ray-traced 3D urban scene used for CSI generation. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Stage I reconstruction objective (negative cosine [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: CDF of the 3D localization error under Q = 10 for different training strategies [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: CDF of the 3D localization error for different [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Visualization of original and reconstructed CSI pe [PITH_FULL_IMAGE:figures/full_fig_p011_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: 3D trajectory tracking simulation. Left: ground-t [PITH_FULL_IMAGE:figures/full_fig_p012_14.png] view at source ↗

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

52 extracted references · 1 linked inside Pith

  1. [1]

    A Survey on Fundamental Limits of Integrat ed Sensing and Communication,

    A. Liu et al., “A Survey on Fundamental Limits of Integrat ed Sensing and Communication,” IEEE Communications Surveys & Tutorials , vol. 24, no. 2, pp. 994-1034, 2022

    2022

  2. [2]

    Positioning Using Wireless Networks: Ap plications, Recent Progress, and Future Challenges,

    Y . Y ang et al., “Positioning Using Wireless Networks: Ap plications, Recent Progress, and Future Challenges,” IEEE Journal on Selected Areas in Communications , vol. 42, no. 9, pp. 2149-2178, 2024

    2024

  3. [3]

    Geometrical mod eling and experimental measurements of indoor mmWave communication systems including finite-area reflective surfaces,

    Ivanov, V ., V olkov, A. and Peters, D. R. “Geometrical mod eling and experimental measurements of indoor mmWave communication systems including finite-area reflective surfaces,” Progress In Electromagnetics Research, vol. 182, pp. 107-119, 2025

    2025

  4. [4]

    Cost - Effective Deployment for Fully-Decoupled RAN: A Techno-Ec onomic Approach,

    J. Zhao, J. Chen, B. Cheng, B. Qian, Y . Xu and H. Zhou, “Cost - Effective Deployment for Fully-Decoupled RAN: A Techno-Ec onomic Approach,” IEEE Transactions on V ehicular Technology, vol. 73, no. 11, pp. 17007-17023, 2024

    2024

  5. [5]

    A Social-Aw are Virtual MAC Protocol for Energy-Efficient D2D Communications Under lying Heterogeneous Cellular Networks,

    B. Fan, H. Tian, L. Jiang and A. V . V asilakos, “A Social-Aw are Virtual MAC Protocol for Energy-Efficient D2D Communications Under lying Heterogeneous Cellular Networks,” IEEE Transactions on V ehicular Technology, vol. 67, no. 9, pp. 8372-8385, 2018

    2018

  6. [6]

    Y u et al., ”A Fully-Decoupled RAN Architecture for 6G I nspired by Neurotransmission,” Journal of Communications and Information Networks, vol

    Q. Y u et al., ”A Fully-Decoupled RAN Architecture for 6G I nspired by Neurotransmission,” Journal of Communications and Information Networks, vol. 4, no. 4, pp. 15-23, Dec. 2019

    2019

  7. [7]

    Fully-decou pled radio access networks: A resilient uplink base stations cooperat ive reception framework,

    J. Zhao, Q. Y u, B. Qian, K. Y u, Y . Xu, H. Zhou, “Fully-decou pled radio access networks: A resilient uplink base stations cooperat ive reception framework,” IEEE Transactions on Wireless Communications , vol. 22, no. 8, pp. 5096-5110, 2023

    2023

  8. [8]

    Q. Peng, H. Ren, C. Pan, N. Liu and M. Elkashlan, ”Resource Allocation for Uplink Cell-Free Massive MIMO Enabled URLLC in a Smart Factory,” IEEE Transactions on Communications , vol. 71, no. 1, pp. 553-568, 2023

    2023

  9. [9]

    Q. Peng, H. Ren, M. Dong, M. Elkashlan, K. -K. Wong and L. Ha nzo, ”Resource Allocation for Cell-Free Massive MIMO-Aided URL LC Systems Relying on Pilot Sharing,” IEEE Journal on Selected Areas in Communications , vol. 41, no. 7, pp. 2193-2207, 2023

    2023

  10. [10]

    Algorithm s and Bounds for Distributed TDOA-Based Positioning Using OFDM S ig- nals,

    R. K. Martin, C. Y an, H. H. Fan and C. Rondeau, “Algorithm s and Bounds for Distributed TDOA-Based Positioning Using OFDM S ig- nals,” IEEE Transactions on Signal Processing , vol. 59, no. 3, pp. 1255- 1268, 2011

    2011

  11. [11]

    Multiple emitter location and signal para meter estimation,

    R. Schmidt, “Multiple emitter location and signal para meter estimation,” IEEE Transactions on Antennas and Propagation , vol. 34, no. 3, pp. 276-280, 1986. 13

    1986

  12. [12]

    Joint Direction-of-Arrival and Time-of-Arrival Estimation With Ultra-Wideband Elliptical Arrays,

    A. Ram´ ırez-Arroyo, A. Alex-Amor, P . Padilla and J. F. V alenzuela- V ald´ es, “Joint Direction-of-Arrival and Time-of-Arrival Estimation With Ultra-Wideband Elliptical Arrays,” IEEE Transactions on Wireless Com- munications, vol. 22, no. 12, pp. 9187-9200, 2023

    2023

  13. [13]

    Linear Least Squares Approach for Ac curate Re- ceived Signal Strength Based Source Localization,

    H. C. So and L. Lin, “Linear Least Squares Approach for Ac curate Re- ceived Signal Strength Based Source Localization,” IEEE Transactions on Signal Processing , vol. 59, no. 8, pp. 4035-4040, 2011

    2011

  14. [14]

    Passively Track WiFi Users With an En hanced Particle Filter Using Power-Based Ranging,

    Z. Li and T. Braun, “Passively Track WiFi Users With an En hanced Particle Filter Using Power-Based Ranging,” IEEE Transactions on Wireless Communications, vol. 16, no. 11, pp. 7305-7318, 2017

    2017

  15. [15]

    3-D Target Localization in Wireless Sensor Networks Using RSS and AoA Measurements,

    S. Tomic, M. Beko and R. Dinis, “3-D Target Localization in Wireless Sensor Networks Using RSS and AoA Measurements,” IEEE Transac- tions on V ehicular Technology, vol. 66, no. 4, pp. 3197-3210, 2017

    2017

  16. [16]

    Three-Dimen sional Localiza- tion Algorithm for Mixed Near-Field and Far-Field Sources B ased on ESPRIT and MUSIC Method,

    Jia-Jia Jiang, Fa-Jie Duan, and Jin Chen, “Three-Dimen sional Localiza- tion Algorithm for Mixed Near-Field and Far-Field Sources B ased on ESPRIT and MUSIC Method,” Progress In Electromagnetics Research , vol. 136, pp. 435-456, 2013

    2013

  17. [17]

    DOA Estimation with Sub- Array Divided Technique and Interporlated ESPRIT Algorith m on a Cylindrical Conformal Array Antenna,

    Peng Y ang, Feng Y ang, and Zai-Ping Nie, “DOA Estimation with Sub- Array Divided Technique and Interporlated ESPRIT Algorith m on a Cylindrical Conformal Array Antenna,” Progress In Electromagnetics Research, vol. 103, pp. 201-216, 2010

    2010

  18. [18]

    CSI-based Posit ioning in Massive MIMO systems using Convolutional Neural Network s,

    S. D. Bast, A. P . Guevara and S. Pollin, “CSI-based Posit ioning in Massive MIMO systems using Convolutional Neural Network s,” 2020 IEEE 91st V ehicular Technology Conference (VTC2020-S pring), Antwerp, Belgium, pp. 1-5, 2020

    2020

  19. [19]

    CSI Phase Fingerprinting for Indoor Localization With a Deep Learning Approach,

    X. Wang, L. Gao and S. Mao, “CSI Phase Fingerprinting for Indoor Localization With a Deep Learning Approach,” IEEE Internet of Things Journal, vol. 3, no. 6, pp. 1113-1123, 2016

    2016

  20. [20]

    CSI-Fingerprinting In door Localiza- tion via Attention-Augmented Residual Convolutional Neur al Network,

    B. Zhang, H. Sifaou and G. Y . Li, “CSI-Fingerprinting In door Localiza- tion via Attention-Augmented Residual Convolutional Neur al Network,” IEEE Transactions on Wireless Communications , vol. 22, no. 8, pp. 5583-5597, 2023

    2023

  21. [21]

    F ast Estimation of Frequency and 2-d DOAs for Cylindrical Conformal Array An tenna Using State-Space and Propagator Method,

    Weijian Si, Liangtian Wan, Lutao Liu, and Zuoxi Tian, “F ast Estimation of Frequency and 2-d DOAs for Cylindrical Conformal Array An tenna Using State-Space and Propagator Method,” Progress In Electromagnet- ics Research, vol. 137, pp. 51-71, 2013

    2013

  22. [22]

    Optimal Power Control a nd CSI Acquisition for Over-the-Air Computation in OFDM System,

    X. Xie, C. Hua, J. Hong and W. Xu, “Optimal Power Control a nd CSI Acquisition for Over-the-Air Computation in OFDM System,” IEEE Transactions on Wireless Communications , vol. 23, no. 6, pp. 6533- 6545, 2024

    2024

  23. [23]

    Nonorthogo nal Time-Frequency Training-Sequence-Based CSI Acquisition for MIMO Systems,

    W. Ding, F. Y ang, S. Liu, X. Wang and J. Song, “Nonorthogo nal Time-Frequency Training-Sequence-Based CSI Acquisition for MIMO Systems,” IEEE Transactions on V ehicular Technology , vol. 65, no. 7, pp. 5714-5719, 2016

    2016

  24. [24]

    Massive MIMO-OFDM Channel Acquisition With Time- Frequency Phase-Shifted Pilots,

    J. Tang et al., “Massive MIMO-OFDM Channel Acquisition With Time- Frequency Phase-Shifted Pilots,” IEEE Transactions on Communica- tions, vol. 73, no. 6, pp. 4520-4535, 2025

    2025

  25. [25]

    Accurate Location Tra cking From CSI-Based Passive Device-Free Probabilistic Fingerprint ing,

    S. Shi, S. Sigg, L. Chen and Y . Ji, “Accurate Location Tra cking From CSI-Based Passive Device-Free Probabilistic Fingerprint ing,” IEEE Transactions on V ehicular Technology , vol. 67, no. 6, pp. 5217-5230, 2018

    2018

  26. [26]

    Ef ficient Neural Network Approach for 2D DOA Estimation Based on Anten na Array Measurements,

    Marija Agatonovic, Zoran Stankovic, Ivan Milovanovic , Nebojsa Don- cov, Leen Sit, Thomas Zwick, and Bratislav Milovanovic, “Ef ficient Neural Network Approach for 2D DOA Estimation Based on Anten na Array Measurements,” Progress In Electromagnetics Research, vol. 137, pp. 741-758, 2013

    2013

  27. [27]

    Cooperative Lidar Sensing for Pedestrian Detecti on: Data Asso- ciation Based on Message Passing Neural Networks,

    B. Camajori Tedeschini, M. Brambilla, L. Barbieri, G. B alducci and M. Nicoli, “Cooperative Lidar Sensing for Pedestrian Detecti on: Data Asso- ciation Based on Message Passing Neural Networks,” IEEE Transactions on Signal Processing , vol. 71, pp. 3028-3042, 2023

    2023

  28. [28]

    An Indoor Localization Technique Utilizing Passi ve Tags and 3-D Microwave Passive Radar Imaging,

    Quanfeng Wang, Alexander H. Paulus, Mei Song Tong, and T homas F. Eibert, “An Indoor Localization Technique Utilizing Passi ve Tags and 3-D Microwave Passive Radar Imaging,” Progress In Electromagnetics Research, vol. 181, pp. 89-98, 2024

    2024

  29. [29]

    Indoor V ehi cle Positioning for MIMO-OFDM WIFI Systems via Rearranged Spar se Bayesian Learning,

    J. Du, J. Cao, L. Jin, S. Li, J. Liu and F. Gao, “Indoor V ehi cle Positioning for MIMO-OFDM WIFI Systems via Rearranged Spar se Bayesian Learning,” IEEE Transactions on Wireless Communications , vol. 23, no. 7, pp. 7849-7864, 2024

    2024

  30. [30]

    Integrated Sensin g and Com- munications (ISAC) for V ehicular Communication Networks ( VCN),

    X. Cheng, D. Duan, S. Gao and L. Y ang, “Integrated Sensin g and Com- munications (ISAC) for V ehicular Communication Networks ( VCN),” IEEE Internet of Things Journal , vol. 9, no. 23, pp. 23441-23451, 2022

    2022

  31. [31]

    Machine Learning-Assisted Sensing Techniques for I ntegrated Communications and Sensing in WLANs: Current Status and Fut ure Directions (Invited),

    Siyuan Shao, Min Fan, Cheng Y u, Y an Li, Xiaodong Xu, and H aiming Wang, “Machine Learning-Assisted Sensing Techniques for I ntegrated Communications and Sensing in WLANs: Current Status and Fut ure Directions (Invited),” Progress In Electromagnetics Research , vol. 175, pp. 45-79, 2022

    2022

  32. [32]

    WinIPS: WiFi- Based Non-Intrusive Indoor Positioning System With Online Radio Map Construction and Adaptation,

    H. Zou, M. Jin, H. Jiang, L. Xie and C. J. Spanos, “WinIPS: WiFi- Based Non-Intrusive Indoor Positioning System With Online Radio Map Construction and Adaptation,” IEEE Transactions on Wireless Communications, vol. 16, no. 12, pp. 8118-8130, 2017

    2017

  33. [33]

    Toward 5G NR High-Preci sion Indoor Positioning via Channel Frequency Response: A New Pa radigm and Dataset Generation Method,

    K. Gao, H. Wang, H. Lv and W. Liu, “Toward 5G NR High-Preci sion Indoor Positioning via Channel Frequency Response: A New Pa radigm and Dataset Generation Method,” IEEE Journal on Selected Areas in Communications, vol. 40, no. 7, pp. 2233-2247, 2022

    2022

  34. [34]

    CSI-Based Multi-Antenna and Multi-Point Indoor Positioning Using Pr obability Fusion,

    E. G¨ on¨ ultas ¸, E. Lei, J. Langerman, H. Huang and C. Studer, “CSI-Based Multi-Antenna and Multi-Point Indoor Positioning Using Pr obability Fusion,” IEEE Transactions on Wireless Communications , vol. 21, no. 4, pp. 2162-2176, 2022

    2022

  35. [35]

    High-Resoluti on Localization Using Distributed MIMO FMCW Radars,

    H. Park, S. Chung, J. Park, and Y . Huang, “High-Resoluti on Localization Using Distributed MIMO FMCW Radars,” Sensors, vol. 25, no. 12, pp.3579-3579, 2025

    2025

  36. [36]

    Cooperative Locali zation in Wireless Sensor Networks With AOA Measurements,

    S. Wang, X. Jiang and H. Wymeersch, “Cooperative Locali zation in Wireless Sensor Networks With AOA Measurements,” IEEE Transac- tions on Wireless Communications , vol. 21, no. 8, pp. 6760-6773, 2022

    2022

  37. [37]

    Cooperative Deep-Lear ning Position- ing in mmWave 5G-Advanced Networks,

    B. C. Tedeschini and M. Nicoli, “Cooperative Deep-Lear ning Position- ing in mmWave 5G-Advanced Networks,” IEEE Journal on Selected Areas in Communications , vol. 41, no. 12, pp. 3799-3815, 2023

    2023

  38. [38]

    Learning to Localize: A 3D CNN Approach to U ser Positioning in Massive MIMO-OFDM Systems,

    C. Wu et al., “Learning to Localize: A 3D CNN Approach to U ser Positioning in Massive MIMO-OFDM Systems,” IEEE Transactions on Wireless Communications, vol. 20, no. 7, pp. 4556-4570, 2021

    2021

  39. [39]

    Overview of Deep Le arning- Based CSI Feedback in Massive MIMO Systems,

    J. Guo, C. -K. Wen, S. Jin and G. Y . Li, “Overview of Deep Le arning- Based CSI Feedback in Massive MIMO Systems,” IEEE Transactions on Communications , vol. 70, no. 12, pp. 8017-8045, 2022

    2022

  40. [40]

    Deep Learning for Mass ive MIMO CSI Feedback,

    C. -K. Wen, W. -T. Shih and S. Jin, “Deep Learning for Mass ive MIMO CSI Feedback,” IEEE Wireless Communications Letters , vol. 7, no. 5, pp. 748-751, 2018

    2018

  41. [41]

    Ra te- Adaptive Feedback With Bayesian Compressive Sensing in Mul tiuser MIMO Beamforming Systems,

    X. -L. Huang, J. Wu, Y . Wen, F. Hu, Y . Wang and T. Jiang, “Ra te- Adaptive Feedback With Bayesian Compressive Sensing in Mul tiuser MIMO Beamforming Systems,” IEEE Transactions on Wireless Com- munications, vol. 15, no. 7, pp. 4839-4851, 2016

    2016

  42. [42]

    Deep Learni ng and Compressive Sensing-Based CSI Feedback in FDD Massive MIMO Systems,

    P . Liang, J. Fan, W. Shen, Z. Qin and G. Y . Li, “Deep Learni ng and Compressive Sensing-Based CSI Feedback in FDD Massive MIMO Systems,” IEEE Transactions on V ehicular Technology , vol. 69, no. 8, pp. 9217-9222, 2020

    2020

  43. [43]

    Convolutional Neu ral Network- Based Multiple-Rate Compressive Sensing for Massive MIMO C SI Feedback: Design, Simulation, and Analysis,

    J. Guo, C. -K. Wen, S. Jin and G. Y . Li, “Convolutional Neu ral Network- Based Multiple-Rate Compressive Sensing for Massive MIMO C SI Feedback: Design, Simulation, and Analysis,” IEEE Transactions on Wireless Communications, vol. 19, no. 4, pp. 2827-2840, 2020

    2020

  44. [44]

    Two-Stage Adaptive and Com- pressed CSI Feedback for FDD Massive MIMO,

    G. Huang, A. Liu and M. -J. Zhao, “Two-Stage Adaptive and Com- pressed CSI Feedback for FDD Massive MIMO,” IEEE Transactions on V ehicular Technology, vol. 70, no. 9, pp. 9602-9606, 2021

    2021

  45. [45]

    Changeable Rate and Novel Quantization for CSI Feedback Based on Deep Learni ng,

    X. Liang, H. Chang, H. Li, X. Gu and L. Zhang, “Changeable Rate and Novel Quantization for CSI Feedback Based on Deep Learni ng,” IEEE Transactions on Wireless Communications , vol. 21, no. 12, pp. 10100-10114, 2022

    2022

  46. [46]

    Quantization Adapt or for Bit-Level Deep Learning-Based Massive MIMO CSI Feedback,

    X. Zhang, Z. Lu, R. Zeng and J. Wang, “Quantization Adapt or for Bit-Level Deep Learning-Based Massive MIMO CSI Feedback,” IEEE Transactions on V ehicular Technology , vol. 73, no. 4, pp. 5443-5453, 2024

    2024

  47. [47]

    Low-Com plexity CSI Feedback for FDD Massive MIMO Systems via Learning to Optimize,

    Y . Ma, H. He, S. Song, J. Zhang and K. B. Letaief, “Low-Com plexity CSI Feedback for FDD Massive MIMO Systems via Learning to Optimize,” IEEE Transactions on Wireless Communications , vol. 24, no. 4, pp. 3483-3498, 2025

    2025

  48. [48]

    SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Seg mentation,

    V . Badrinarayanan, A. Kendall and R. Cipolla, “SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Seg mentation,” IEEE Transactions on Pattern Analysis and Machine Intellig ence, vol. 39, no. 12, pp. 2481-2495, 2017

    2017

  49. [49]

    Sionna: An open-source library for ne xt-generation physical layer research,

    J. Hoydis et al., “Sionna: An open-source library for ne xt-generation physical layer research,” Mar. 2022, arXiv:2203.11854

    Pith/arXiv arXiv 2022

  50. [50]

    OpenCellid

    Unwiredlabs. OpenCellid. Accessed: Sep. 2024. [Onlin e]. Available: https://www.opencellid.org

    2024

  51. [51]

    Attention is all you need,

    A. V aswani et al., “Attention is all you need,” Neural Information Processing Systems, pp. 5998-6008, Jun. 2017

    2017

  52. [52]

    Deep Learn ing-Based Multi-User Positioning in Wireless FDMA Cellular Networks ,

    Z. Chen, Z. Zhang, Z. Xiao, Z. Y ang and R. Jin, “Deep Learn ing-Based Multi-User Positioning in Wireless FDMA Cellular Networks ,” IEEE Journal on Selected Areas in Communications , vol. 41, no. 12, pp. 3848- 3862, 2023

    2023