Autoencoder-Based CSI Compression for Beyond Wi-Fi 8 Coordinated Beamforming Ibrahim Abou Shehada∗ , Boris Bellalta∗ , Giovanni Geraci∗† , Lorenzo Galati Giordano‡ ∗ Department of Engineering, Universitat Pompeu Fabra, Barcelona, Spain † Nokia Standards, Spain ‡ Radio Systems Research, Nokia Bell Labs, Stuttgart, Germany
arXiv:2604.13500v1 [cs.NI] 15 Apr 2026
Email: {[email protected]} Abstract—Coordinated beamforming (Co-BF) is a key multiaccess-point coordination (MAPC) technique for dense Wi-Fi deployments, but its performance can be hindered by the large channel state information (CSI) feedback required through channel sounding across overlapping basic service sets (OBSS). This work proposes an autoencoder (AE)-based CSI compression mechanism integrated into a standards-aligned IEEE 802.11bn MAC design. Using an event-driven simulator with realistic channels generated through Sionna RT, we evaluate the tradeoff between AE reconstruction accuracy and feedback size by measuring their impact on channel sounding overhead and data latency. Our results show that AE-based compression reduces channel sounding overhead by more than 50% compared to IEEE 802.11 CSI compression, with a compression ratio of 1/4 providing the best accuracy/feedback-size tradeoff for lowest data latency. Compared to legacy transmissions without MAPC, IEEE 802.11 CSI compression limits Co-BF due to high channel sounding overhead, causing it to underperform the legacy in some situations. However, AE-based CSI compression enables better Co-BF performance with substantial gains in throughput and data latency compared to legacy, demonstrating its promise as an enabler of efficient MAPC operation in future Wi-Fi systems.
I. I NTRODUCTION Wi-Fi usage continues to grow rapidly, driven by applications with high data rate, low latency, and high reliability requirements such as augmented reality/virtual reality (AR/VR), 4K/8K video streaming, and cloud gaming [1]. However, contention-based channel access using carrier sense multiple access with collision avoidance (CSMA/CA) in IEEE 802.11 limits the ability to serve many highly-demanding users simultaneously in dense deployments [2]. Multi-access-point coordination (MAPC) is a key feature anticipated for next-generation Wi-Fi (Wi-Fi 8 / ultra-high reliability (UHR) / IEEE 802.11bn) [3]. With MAPC, when an access point (AP) gains channel access via CSMA/CA, it can share time, frequency, or spatial resources with another AP during the same transmission opportunity (TXOP), improving channel access in overlapping basic service sets (OBSS) [4], [5]. Coordinated beamforming (Co-BF) is a promising MAPC technique due to its interference-mitigation capabilities, but it This work was supported in part by grants PID2024-155470NB-I00 (MCIU/AEI/FEDER,UE), PCI2023-145958-2 (MLDR ANR-23-CHR4-0005), PID2021-123999OB-I00, PID2021-123995NB-I00, PID2024-156488OB-I00, CEX2021-001195-M, CNS2023-145384 and AGAUR ICREA Academia 00077.
requires each AP to acquire channel state information (CSI) from both its own stations (STAs) and OBSS STAs, resulting in large CSI feedback overhead that can negatively impact performance. Prior Co-BF studies remain limited. The work in [6] demonstrates zero multiuser interference (MUI) but only for one BSS, leaving inter-BSS interference unsuppressed for the other BSS. The analysis in [7] focuses only on throughput, and [8] explores deep reinforcement learning (DRL) for Co-BF but uses simplified channel modeling with free-space path loss (FSPL) and shadowing. Most prior work assumes full-buffer traffic and relies on overly simplified channels, which can lead to exaggerated performance in multiuser (MU) multiple-input multiple-output (MIMO) settings, plus ignoring latency evaluation. Furthermore, the impact of channel sounding overhead on Co-BF performance remains unexplored. In this work, we propose the use of an autoencoder (AE) for CSI compression to mitigate the negative impact of the large channel sounding overhead in Co-BF. We simulate CoBF and legacy transmissions using a custom simulator built on SimPy [9] for event-driven simulation and Sionna Ray-Tracer (RT) [10] for realistic channel modeling, performing end-toend (E2E) performance evaluations in terms of data latency and throughput. AE-based CSI compression has been studied in several works, such as CsiNet [11] and CRNet [12] for cellular communication and LB-SciFi [13] and EFNet [14] for Wi-Fi networks. To our knowledge, however, this is the first work that evaluates AE performance in terms of latency and investigates the optimal compression ratio based on E2E performance. First, we analyze the tradeoff between AE reconstruction accuracy and feedback size across different compression ratios and show that AE-based CSI compression reduces Co-BF channel sounding overhead by more than 50% compared to the IEEE 802.11 method. In addition, an optimal tradeoff is achieved at a compression ratio of 1/4, which yields the lowest data latency. Furthermore, compared to legacy transmission without MAPC, Co-BF yields higher data latency in some scenarios using IEEE 802.11 CSI compression due to large sounding overhead, but the use of AE-based CSI compression allows Co-BF to outperform the legacy in all scenarios and offers substantial performance gains that can exceed 30% in both data latency and throughput.
Figure 1: TXOP signaling structure example with 2 STAs for Co-BF with IEEE 802.11bn joint NDP channel sounding procedure.
II. S YSTEM M ODEL In this section, details of the system model are discussed. In particular, we present the MAC layer model, channel model, formulation of the Co-BF precoder, and traffic model. A. MAC Layer Model Each AP contends for the channel using the distributed coordination function (DCF). Once an AP gains channel access, it sends a Co-BF invite frame to the neighboring AP, which replies with a Co-BF response after a short inter-frame spacing (SIFS) interval (see Fig. 1). The Co-BF invite and response frames follow the buffer state report poll (BSRP) non-trigger-based (NTB) and multi-STA BlockACK (BA) formats, respectively, as recommended by the IEEE802.11bn task group (TGbn) [3]. We will refer to the initiating AP as the coordinating AP, and the other as the coordinated AP. Scheduling decisions are performed by the coordinating AP based on the oldest packets across both BSSs and are embedded in the Co-BF trigger frame, with a maximum of 2 scheduled STAs per BSS in the same TXOP. Both APs then transmit aggregated MAC protocol data unit (A-MPDU) frames to their scheduled STAs using Co-BF to suppress intra- and inter-BSS interference. If the coordinated AP has an empty buffer, it declines the session, and the coordinating AP starts a legacy A-MPDU transmission. Outdated CSI of any scheduled STA(s) triggers channel sounding prior to AMPDU transmission, and we consider CSI to be outdated after 25 ms from the last update. This CSI age limit was specified based on the work in [15] and our own simulations to ensure that STAs can always receive data from their APs even if at the lowest modulation and coding scheme (MCS). Co-BF sounding uses the IEEE 802.11bn joint null data packet (NDP) procedure [3]. After the Co-BF response, the coordinating AP sends a sounding invite followed by the coordinated AP’s response. The coordinating AP then sends a UHR NDP announcement (NDPA), after which both APs simultaneously transmit Extremely-High Throughput (EHT) NDP frames. A beamforming report poll (BFRP) frame triggers uplink CSI feedback from the STAs. The coordinated AP repeats the NDPA/NDP/BFRP sequence for its own STA(s). If enough time remains in the TXOP, a Co-BF trigger initiates A-MPDU transmissions. Legacy transmissions without MAPC use the standard request-to-send/clear-to-send (RTS/CTS) pro-
cedure, and the MAC layer parameters are summarized in Table I. B. Channel Model Let hki,j ∈ CNt denote the channel between STA i and AP j in subcarrier k. The channels are generated using Sionna RT, and imperfect CSI is modeled by adding least-squares (LS) estimation noise to hki,j . During joint NDP sounding, STA i receives long training field (LTF) pilot symbols from both APs as follows k yLTF = hki,1 xkLTF1 + hki,2 xkLTF2 + n, i
(1)
where n ∼ CN(0, N0 ) is a complex white Gaussian noise with N0 being the thermal noise power and the pilot symbols xkLTF1 from AP 1 and xkLTF2 from AP 2 are assumed to be orthogonal. k To estimate hki,j , yLTF is multiplied by the pseudo-inverse of i k xLTFj , which is orthogonal to xkLTFi , yielding k ĥki,j = yLTF (xkLTFj )† = hki,j + nLTF , i
(2)
2 where nLTF ∼ CN(0, σLTF ) ∈ CNt is the LS estimation noise approximated as a complex Gaussian random vector with a variance of N0 · NF 2 , (3) σLTF = NLTF · E[x2LTF ]
where NF is the noise figure and NLTF is the number of pilot LTF symbols in the EHT NDP frame. Table I: MAC layer parameters Parameter Maximum TXOP duration (Ts ) [ms] Minimum contention window size (CWmin ) Maximum contention window size (CWmax ) Payload length (LD ) [B] MAC header length [b] MPDU delimiter [b] Service field length [b] Packet tail [b] EHT/UHR Preamble duration [µs] Legacy preamble duration [µs] Legacy rate [Mbps] Short inter-frame spacing (TSIFS ) [µs] Distributed inter-frame spacing (TDIFS ) [µs] Empty slot duration (Te ) [µs]
Value 5.484 16 1024 1500 240 32 16 18 88.8 20 6 16 34 9
C. Co-BF Precoder Co-BF is implemented using cell-edge-aware zero forcing (CEA-ZF) [16] to null interference between all scheduled STAs in both BSSs, assuming single-antenna receivers. For each AP–STA pair and subcarrier, we apply singular value decomposition (SVD) to the estimated channel as follows: k ĥki,j = uki,j ski,j Vi,j .
(4)
k The first column of Vi,j is compressed at the STA then decomk pressed at the AP to obtain ṽi,j (details of CSI compression are presented in Section III). For AP j, the CEA-ZF matrix k ṼCEA ∈ CNt ×|Sj ∪S̄j | is constructed as j k k k k , ṽ1,j ṼCEA = [ṽ1,j , ..., ṽ|S , ..., ṽ|kS̄j |,j ], j j |,j
(5)
where Sj is the set of scheduled in-BSS STAs of AP j and S̄j is the set of scheduled OBSS STAs. The CEA-ZF precoder is then computed as 1
k k k k ((ṼCEA )H ṼCEA )−1 }[1:|Sj |] , WCEA = (Dkj )(− 2 ) {ṼCEA j j j j (6) with {V }[1:n] indicating the first n columns of the matrix V . The signal-to-interference-plus-noise ratio (SINR) at STA i that is associated with AP j in subcarrier k is given by
SINRki =
k 2 |hki,j wj,i | XX , k k 2 k 2 |hi,j wj,s | + |hki,p wp,r | +N0 · NF
X
p̸=j r∈Sp
s̸=i∈Sj
(7) Furthermore, the effective SINR is computed as the arithmetic mean across data subcarriers, N
f SINRef = i
sc 1 X SINRki , Nsc
(8)
k=1
and the highest MCS guaranteeing a maximum packet error rate (PER) of 10% based on the effective SINR is selected. D. Traffic Model Packet arrivals at the AP buffer follow a batch Poisson process (BPP). Batches arrive every interval Tb , modeled as Tb ∼ Exp(λb ), with λPKT λb = , (9) B where λPKT is the packet arrival rate and B is the average batch size. Data latency is measured as data packet delays: dPKT = trec − tarr ,
(10)
where trec and tarr are the reception time at the STA and the arrival time at the AP buffer, respectively. This captures queuing, contention, and transmission delays. III. CSI C OMPRESSION AND F EEDBACK CSI compression and feedback are discussed in this section. We describe the IEEE 802.11 standard mechanism for CSI compression, followed by AE-based CSI compression.
A. CSI Compression and Feedback in IEEE 802.11 As noted in Section II-B, the right singular vectors vk across multiple subcarriers are compressed at the STA and reported to the AP in the compressed beamforming (CB) report frame [17]. IEEE 802.11 uses Givens rotations to decompose each vector into angle pairs (ϕ, ψ), which are quantized (see Signal Model Section in [18] for details). The subcarriers are grouped into blocks and the quantized angle pairs of the first subcarrier per group are reported. The number of angle pairs is Nss (2Nt − Nss − 1), (11) Na = 2 where Nss is the number of spatial streams and Nt is the number of Tx antennas. Then, the CB report size is Nsc Na (bϕ + bψ ). (12) Ng where Nsc is the number of data subcarriers and Ng is the subcarrier grouping. For Nsc = 980, Nt = 16, Ng = 16, and (bϕ , bψ ) = (9, 7), the feedback size is 14,880 bits per STA, which doubles under joint NDP sounding. This motivates AEbased compression to reduce channel sounding overheads. B. CSI Compression using Autoencoders The proposed CSI AE is shown in Fig. 2, which follows the EFNet design in [14] but with some enhancements on the encoder side. The unit-norm vectors vk are sampled every Ng subcarriers and stacked into an input tensor Tv ∈ R2×Nvs ×Nt , where Nvs = ⌈Nsc /Ng ⌉ is the number of valid subcarriers. At the STA, Tv is encoded into a latent vector of dimension M = ⌈η(2Nvs Nt )⌉, where η is the compression ratio, which is then quantized and entropy–coded into a bit stream using an entropy bottleneck layer [19], [20]. The AP decodes and dequantizes the latent representation, which is then expanded, reshaped and passed into two channel attention refinement (CA-Refine) blocks (see Fig. 3 for the channel attention module design) to reconstruct Tv . A final convolution layer and per-vector normalization ensure that all reconstructed vectors have unit norm, yielding T̃v ∈ R2×Nvs ×Nt . Using tanh as in EFNet [14] bounds vector elements to [−1, 1] but does not ensure unit-norm vectors. The entropy bottleneck layer follows the model of [20] and is implemented using compressai library [21]. The details of integrating the entropy bottleneck layer with AE are in Section III of [19]. Training uses ADAM [22] to minimize the meansquared error (MSE) between Tv and T̃v in the training set D: 1 X |Tv − T̃v |2 . (13) LMSE = |D| Tv ∈D
k
To obtain ṽ for all data subcarriers, T̃v is reshaped into unitnorm vectors, which are repeated Ng times each. Reconstruction accuracy is measured using the cosine correlation between the original and reconstructed vectors: N
ρ=
sc 1 X |(ṽk )H vk |. Nsc
k=1
(14)
Figure 2: Proposed autoencoder architecture for CSI compression with encoder layers (blue) decoder layers (green) and entropy bottleneck layer (checkered). The number of feature maps of the input/output of each layer is indicated on top.
Figure 3: Architecture of the channel attention module [14].
IV. P ERFORMANCE E VALUATION In this section, we discuss the simulation scene, then describe the AE training data and evaluate the tradeoff between reconstruction accuracy and feedback size by examining their effect on Co-BF sounding overhead and data latency at different compression ratios. We further compare Co-BF against legacy transmissions in terms of throughput and latency. A. Simulation Scene The scene used for AE training and event-driven simulations is presented in Fig. 4, which is an indoor office scene with 4 rooms. Each room in the scene contains a ceiling-mounted AP with a 4×4 uniform planar array (UPA) antenna with a directional radiation pattern, while STAs use single omnidirectional antennas. A received signal strength (RSS) heatmap of one of the APs over the scene is also shown in Fig. 4. The STAs are deployed at random heights with at least 50 cm horizontal separation from the walls. Furthermore, during
Figure 4: Indoor office with RSS heatmap of bottom-left AP.
event-driven simulations, the STAs move inside the scene at a fixed speed in random directions (random-walk), with the direction of movement changing every 2 seconds. The scene and channel parameters are listed in Table II. For event-driven simulations, we consider multiple deployment scenarios with different transmission modes, CSI compression methods, and traffic loads. For each scenario, the STAs are placed in two randomly selected rooms of the scene in Fig. 4, each served by the AP in that room. Per-STA traffic loads are 177/93/63 Mbps for 2/4/6 STAs per AP (high load),1 1 High loads were set at 95% of worst-case STA throughput of legacy 80.
Table II: Scene and channel parameters. Parameter Room length, width, height [m] Corridor width [m] STA height [m] STA speed [km/h] Channel bandwidth [MHz] Subcarrier spacing [kHz] Carrier frequency [GHz] Noise floor [dBm/Hz] Power spectral density [dBm/MHz] Noise figure (NF) [dB]
Value (5, 5, 3) 1 1.2 - 1.7 0.9 80 78.125 5.3 -174 5 7
Table III: Reconstruction accuracy and feedback size. Mean 1st Percentile Feedback size [bits] LOS NLOS LOS NLOS Median Min Max StDev IEEE 802.11 (conf. 1) 0.9998 0.9861 0.9994 0.8609 14880 – – – IEEE 802.11 (conf. 2) 0.9988 0.9880 0.9986 0.8886 44100 – – – Train (η = 1/2) 0.9975 0.9756 0.9951 0.7857 6240 5760 7072 205.19 Test (η = 1/2) 0.9974 0.9748 0.9938 0.7846 Train (η = 1/4) 0.9974 0.9758 0.9931 0.7715 2944 2784 3232 75.20 Test (η = 1/4) 0.9970 0.9745 0.9885 0.7957 Train (η = 1/8) 0.9931 0.9718 0.9783 0.7887 1472 1408 1600 28.78 Test (η = 1/8) 0.9914 0.9634 0.9598 0.7713 Train (η = 1/16) 0.9853 0.9662 0.9505 0.7859 768 736 864 21.36 Test (η = 1/16) 0.9821 0.9564 0.9077 0.7480 Dataset
Figure 5: Violin plots of Co-BF channel sounding overhead at high load.
with medium load set to half these values. Each setting is evaluated in 50 deployments using 10-second simulations with identical seeds for topology, mobility, and traffic generation to ensure a fair comparison between different settings. B. Co-BF Evaluation with AE CSI Compression 1) AE Training Data: The AE is trained using CSI collected from the scene in Fig. 4 via Sionna RT. We collected 12,000 CSI samples per room (four APs at 3000 random STA positions). For Table III, the training uses three rooms (36,000 samples), while the remaining room provides validation and test data to assess generalization between rooms. For the performance evaluations in this section, we use a final AE trained on the entire dataset. 2) AE Evaluation: Table III reports the mean and bottom 1st percentile reconstruction accuracy under LOS and NLOS conditions and the feedback sizes for various η. Furthermore, two IEEE 802.11 CSI feedback configurations are included for reference: configuration 1 with the baseline parameters from Section III-A, and configuration 2 with Ng = 4 and (bϕ , bψ ) = (7, 5). Reconstruction accuracy is similar for η = 1/2 and η = 1/4, but drops at lower ratios, while feedback size decreases almost linearly with η. In addition, the AE shows very good generalization between different rooms, which is evident from the almost identical accuracy between the training and testing sets. The generalization is best with η = 1/2 and η = 1/4, but decreases with lower η. Nevertheless, the 4 rooms in the scene have very similar channel distributions, and further generalization capabilities between different scenarios shall be studied in future work. Furthermore, Figures 5 and 6 illustrate the system-level impact. Fig. 5 shows that IEEE 802.11 compression can push the sounding overhead near the 5.484 ms TXOP limit, while AE compression keeps it well
Figure 6: 99th percentile and median latency with 4 STAs per AP.
below, allowing sufficient time for A-MPDU transmission after channel sounding. Fig. 6 indicates that η = 1/4 offers the lowest data latency by balancing reconstruction accuracy and feedback size: it halves the feedback relative to η = 1/2 at nearly the same reconstruction accuracy and avoids the degraded precoder–channel alignment seen at η < 1/4. C. Co-BF Comparison with Legacy Figures 7 and 8 show comparisons of Co-BF and legacy non-MAPC performance in terms of throughput (full buffer) and data latency (high load), respectively. Co-BF is evaluated with IEEE 802.11 CSI compression and AE-based CSI compression, which we will refer to as Co-BFST and Co-BFAE , respectively. In addition, legacy transmissions are evaluated with both APs using a shared 80 MHz channel or each using a separate 40 MHz channel, which we will call legacy 80 and legacy 40, respectively. With 2 STAs per AP, Co-BF increases throughput by more than 50% compared to the legacy 40
V. F UTURE W ORK
Figure 7: Average throughput per STA comparison.
The results presented show that using AE for CSI compression can significantly boost the performance of Co-BF. However, the AE in this work is trained with site-specific data from the same simulation scene. Generalization between different scenes has been studied for cellular systems [23], [24], but the evaluation is limited to the reconstruction accuracy of the channel. The impact of generalization on E2E performance has not yet been studied. In addition, different considerations for model architectures may be required for Wi-Fi channels, since most current AE architectures were built for 3GPP channels, not Wi-Fi channels. AE generalization for different channel models, bandwidths, and antenna sizes in Wi-Fi was studied in IEEE contributions [25], but they lack results of the impact on E2E performance. Wi-Fi-based studies on AE CSI compression are also limited for SU-MIMO. A universal AE model that generalizes to different channel distributions, bandwidths, and antenna configurations must be developed and tested in MUMIMO/Co-BF scenarios for practical deployment. Model size, sharing, and run-time are other crucial aspects that have to be studied for practical use of AEs for CSI compression. VI. C ONCLUSION
Figure 8: 99th percentile and median data latency comparison.
MHz case using either CSI compression technique. There are marginal gains in throughput and data latency with Co-BFAE compared to Co-BFST , but the performance is already well improved with Co-BF at this STA density. With 4 STAs per AP, Co-BF offers marginal latency reductions but with decent throughput gains compared to the legacy, which are further increased by AE CSI compression. However, Co-BFST results in higher latency than legacy 40 with 6 STAs per AP, which is caused by large sounding overheads. However, Co-BFAE reduces the 99th latency compared to legacy 40 and also produces a significant increase of 39.9% in throughput. The median latency is still slightly higher with Co-BFAE compared to legacy 40, but this is due to the absence of collisions and minimal contention overhead in legacy 40. These results show that AE CSI compression is an effective tool for reducing sounding overhead, significantly improving Co-BF performance across a wide range of STA densities and traffic conditions.
We presented an AE-based CSI compression mechanism to reduce the large channel sounding overhead of Co-BF for nextgeneration Wi-Fi. Using a standards-aligned MAC design and realistic Sionna RT channels, we showed that AE compression substantially reduces channel sounding overhead by over 50%, and provides an optimal accuracy/feedback-size tradeoff at η = 1/4 in terms of data latency performance. We further evaluated Co-BF against non-MAPC legacy transmissions and saw that IEEE 802.11 CSI compression limits Co-BF performance, leading to higher latency than the legacy at high STA densities. However, Co-BF with AEbased CSI compression outperforms the legacy in all cases with significant gains in throughput and latency. These results indicate that AE-based CSI compression is a strong candidate for enabling practical and high-performance MAPC operation in Wi-Fi 8 and beyond. R EFERENCES [1] G. Geraci, F. Meneghello, F. Wilhelmi, D. Lopez-Perez, I. Val, L. G. Giordano, C. Cordeiro, M. Ghosh, E. Knightly, and B. Bellalta, “Wi-Fi: Twenty-five years and counting,” 2025. [Online]. Available: https://arxiv.org/abs/2507.09613 [2] L. Zhang, H. Yin, S. Roy, L. Cao, X. Gao, and V. Sathya, “IEEE 802.11be network throughput optimization with multilink operation and AP controller,” IEEE Internet of Things Journal, vol. 11, pp. 23 850– 23 861, 7 2024. [3] “IEEE P802.11bn/D1.0: Draft Standard for Information technology—Telecommunications and information exchange between systems Local and metropolitan area networks—Specific requirements Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications—Amendment 6: Enhancements for ultra-high reliability (UHR),” pp. 1–502, Aug 2025. [4] L. Galati-Giordano, G. Geraci, M. Carrascosa, and B. Bellalta, “What will Wi-Fi 8 be? a primer on IEEE 802.11bn ultra high reliability,” IEEE Communications Magazine, vol. 62, no. 8, pp. 126–132, 2024.
[5] S. Verma, T. K. Rodrigues, Y. Kawamoto, M. M. Fouda, and N. Kato, “A survey on multi-AP coordination approaches over emerging WLANs: Future directions and open challenges,” IEEE Communications Surveys & Tutorials, vol. 26, no. 2, p. 858–889, 2024. [Online]. Available: http://dx.doi.org/10.1109/COMST.2023.3344167 [6] D. Ezri, S. Shilo, and R. Keren, “Zero-MUI coordinated beamforming,” November 2023, doc.: IEEE 802.11-23/1998r0. [7] R. Strobel, S. Schelstraete, I. Val, and M. Martinez, “Coordinated spatial nulling (C-SN) simulations,” January 2024, doc.: IEEE 802.1123/0012r0. [8] J. Bacha, A. Zubow, and F. Dressler, “Multi-agent reinforcement learning approach for interference optimization in Wi-Fi 8,” 2025. [9] N. Matloff, A. König, M. Müller, and The SimPy Development Team, SimPy: Discrete-event simulation for Python, 2023, version 4.1.1. [Online]. Available: https://simpy.readthedocs.io/ [10] J. Hoydis, F. A. Aoudia, S. Cammerer, M. Nimier-David, N. Binder, G. Marcus, and A. Keller, “Sionna RT: Differentiable ray tracing for radio propagation modeling,” arXiv:2303.11103, 2023. [11] J. Li, Q. Zhang, X. Xin, Y. Tao, Q. Tian, F. Tian, D. Chen, Y. Shen, G. Cao, Z. Gao, and J. Qian, “Deep learning-based massive MIMO CSI feedback,” in 2019 18th International Conference on Optical Communications and Networks (ICOCN), 2019, pp. 1–3. [12] Z. Lu, J. Wang, and J. Song, “Multi-resolution CSI feedback with deep learning in massive MIMO system,” in ICC 2020 - 2020 IEEE International Conference on Communications (ICC). IEEE, 6 2020. [13] S. Gorinsky, R. Guerin, and P. Steenkiste, “LB-SciFi: Online learningbased channel feedback for MU-MIMO in wireless LANs,” in 2020 IEEE 28th International Conference on Network Protocols (ICNP). IEEE, 10 2020. [14] F. Qi, J. Guo, Y. Cui, X. Li, C.-K. Wen, and S. Jin, “Deep learningbased CSI feedback in Wi-Fi systems,” in 2024 5th International Symposium on Computer Engineering and Intelligent Communications (ISCEIC), 7 2024. [Online]. Available: http://arxiv.org/abs/2407.05905 [15] E. Endovitskiy, S. Tutelian, K. Chemrov, V. Loginov, and E. Khorov, “Impact of explicit channel sounding period on the Wi-Fi MU-MIMO performance,” in 2024 IEEE International Black Sea Conference on Communications and Networking, BlackSeaCom 2024. Institute of Electrical and Electronics Engineers Inc., 2024, pp. 78–83. [16] H. H. Yang, G. Geraci, T. Q. S. Quek, and J. G. Andrews, “Cell-edgeaware precoding for downlink massive MIMO cellular networks,” IEEE Transactions on Signal Processing, vol. 65, no. 13, pp. 3344–3358, 2017. [17] “IEEE standard for information technology–telecommunications and information exchange between systems local and metropolitan area networks–specific requirements part 11: Wireless LAN medium access control (MAC) and physical layer (PHY) specifications,” in: IEEE Std 802.11-2024 (Revision of IEEE Std 802.11-2020), pp. 1–5956, 2025. [18] C. Yuen, S. Sun, M. M. S. Ho, and Z. Zhang, “Beamforming matrix quantization with variable feedback rate,” Eurasip Journal on Wireless Communications and Networking, vol. 2012, 2012. [19] S. Ravula and S. Jain, “Deep autoencoder-based massive MIMO CSI feedback with quantization and entropy coding,” in Proceedings - IEEE Global Communications Conference, GLOBECOM, 2021. [20] J. Ballé, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,” in International Conference on Learning Representations 2018, 5 2018. [Online]. Available: http://arxiv.org/abs/1802.01436 [21] J. Bégaint, F. Racapé, S. Feltman, and A. Pushparaja, “Compressai: a pytorch library and evaluation platform for end-to-end compression research,” arXiv preprint arXiv:2011.03029, 2020. [22] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” International Conference on Learning Representations, 12 2014. [23] Y. Song, J. Roa, R. Zhao, Z. Rong, W. Xiao, J. Jalali, and B. Sheen, “Variable code size autoencoder (VCSA) meets CSI compression in model generalization,” in 2024 International Conference on Computing, Networking and Communications, ICNC 2024. Institute of Electrical and Electronics Engineers Inc., 2 2024, pp. 209–214. [24] K. Lou and X. Wu, “Attention-infused autoencoder for massive MIMO CSI compression,” IEEE Transactions on Wireless Communications, pp. 8005–8017, 11 2025. [Online]. Available: http://arxiv.org/abs/2504. 12440 [25] Z. Guo, P. Liu, W. Zhang, J. Yu, M. Gan, and X. Yang, “Study on AI CSI compression,” September 2023, doc.: IEEE 802.11-23/0290r4.