Source-linked AI summary
UL-Assisted Parametric CSI Feedback and Prediction for FDD Systems
Hao Chang, Mingcheng Nie, Shuangyang Li, Chang Liu, Branka Vucetic, Yonghui Li, Wibowo Hardjawana
TL;DR
FDD systems require DL CSI feedback because UL and DL channels are not fully reciprocal, and feedback can become outdated before use. The paper encodes DL pilot observations for feedback and combines them with UL parameters to predict future channel parameters. On the Sionna RT dataset, the scheme achieves lower NMSE than evaluated baselines across DL pilot SNRs and prediction lengths.
Problem
In FDD, UL and DL channels are not fully reciprocal, so UEs must feed DL CSI to the BS, but delay can make reconstructed CSI outdated.
Source: §I · p. 1Method
UEs directly encode DL pilot observations into feedback symbols, and the BS combines them with partially reciprocal UL parameters to predict future channel parameters and synthesize the DL matrix.
Source: §I · p. 2, B. UL and DL Transmission Phases · p. 3, B. BS-Side Decoder · p. 4Results
The proposed scheme achieves lower NMSE than all evaluated baselines across DL pilot SNRs and various prediction lengths on the Sionna RT dataset.
Source: §V · p. 6Takeaways & Limitations
The parametric design avoids directly predicting the high-dimensional channel matrix and reduces model complexity.
Source: §I · p. 2
Abstract
from arXiv · showhide
In frequency-division duplex (FDD) systems, the uplink (UL) and downlink (DL) channels are not fully reciprocal. Therefore, DL channel state information (CSI) estimated by each user equipment (UE) must be fed back to the base station (BS). However, because the channel varies between CSI acquisition and utilization, the feedback CSI may become outdated before it is used at the BS. To address this challenge, this paper proposes an UL-assisted parametric CSI feedback and prediction scheme for multi-user FDD systems with time-varying channel parameters. By exploiting the partial reciprocity between the UL and DL channel parameters, the BS first estimates the UL channel parameters from the sounding reference signals (SRSs) transmitted by the UEs. The BS then broadcasts DL pilots, allowing each UE to obtain observations containing information about its current DL channel. Instead of explicitly estimating the DL channel from these observations, each UE directly encodes them into feedback symbols and transmits the symbols to the BS. The BS decoder combines the feedback symbols with the partially reciprocal UL parameters to predict the future DL channel parameters. Finally, the corresponding DL channel matrix is synthesized from the predicted parameters using an explicit physical channel model. Simulation results on Sionna ray tracing (RT) dataset show that the proposed scheme outperforms the considered baselines in terms of channel prediction performance over different signal-to-noise ratios and prediction lengths.
I. INTRODUCTION
FDD systems require UEs to feed DL CSI back to the BS, but feedback delay can make reconstructed CSI outdated. The paper proposes UL-assisted parametric feedback and prediction that uses pilot observations and partially reciprocal UL parameters to predict future DL channels.
I. INTRODUCTION
- In FDD, UL and DL channels operate at different carrier frequencies and are not fully reciprocal, so the BS cannot directly use UL CSI for DL transmission.Each UE estimates DL CSI from pilots and feeds it back to the BS for reconstruction.Source: §I · p. 1
- Traditional codebook feedback estimates and quantizes DL CSI before transmission, while neural approaches include channel-matrix compression and direct processing of pilot observations.Codebook search may incur substantial feedback overhead and complexity; CsiNet assumes the UE perfectly knows the DL channel matrix.Source: §I · p. 1
- Feedback delay can make reconstructed DL CSI outdated, while prior predictive approaches either assume stable parameters or incur high-dimensional processing complexity.The cited predictive methods exploit UL reciprocity, but one assumes near-constant structure and another processes historical channel matrices.Source: §I · pp. 1–2
- The proposed scheme predicts future DL channel parameters from pilot-observation feedback and partially reciprocal UL parameters, then synthesizes the channel matrix using a physical model.The UE encodes observations without explicit channel estimation or historical DL CSI.Source: §I · p. 2
II. SYSTEM MODEL
- The system model considers a multi-user OFDM link with a BS serving UEs over K subcarriers.The BS has M antennas; the UL and DL use carrier frequencies fUL and fDL.Source: §II · p. 2
A. Channel Model
The channel model represents a time-frequency channel using path parameters that vary slowly as UEs move, within an approximately unchanged propagation environment.
A. Channel Model
- Within a stationarity region, the propagation environment is approximately unchanged while path parameters vary slowly due to UE mobility.The channel is indexed over users, OFDM symbols, and subcarriers.Source: A. Channel Model · p. 2
- The frequency-domain channel matrix has dimensions M × K, and its channel parameters are assumed constant across subcarriers within each OFDM symbol.The channel vector is described as belonging to C^(M×1).Source: A. Channel Model · pp. 2–3
- The model's path parameters include complex path gain, delay, azimuth and zenith angles, and Doppler shift.The path gain is associated with a particular user and OFDM symbol.Source: A. Channel Model · p. 2
- The BS antenna array is described by horizontal and vertical antenna counts and spacings, with total antenna count M = Mh × Mv.The spacings dh and dv correspond to the horizontal and vertical directions.Source: A. Channel Model · p. 2
B. UL and DL Transmission Phases
The transmission procedure moves from UL sounding and DL training to UL feedback and future DL CSI prediction. The BS uses SRS-derived UL parameters and UE pilot-observation feedback to predict the future channel.
B. UL and DL Transmission Phases
- The procedure has four phases: UL sounding, DL training, UL feedback, and DL CSI prediction.The BS estimates UL parameters at t0, broadcasts DL pilots, receives UE feedback, and predicts the future DL channel.Source: B. UL and DL Transmission Phases · p. 3
- During UL sounding, the BS estimates UL parameters from UE SRSs; path gains are nonreciprocal, while delays and angles are reciprocal.UL parameters are assumed perfectly estimated in this description.Source: B. UL and DL Transmission Phases · p. 3
- During DL training, the BS broadcasts orthogonal pilots, and each UE collects noisy observations across T1 OFDM symbols and K subcarriers.The observation noise is modeled as zero-mean AWGN with variance σ².Source: B. UL and DL Transmission Phases · p. 3
- Each UE compresses its DL pilot observations into feedback symbols, which the BS decoder uses with UL information to predict future channel parameters and synthesize the DL matrix.The prediction covers the future interval t3.Source: B. UL and DL Transmission Phases · p. 3
III. PROPOSED UL-ASSISTED PARAMETRIC CSI FEEDBACK AND PREDICTION
The proposed design compresses DL pilot observations at the UE and predicts low-dimensional channel parameters at the BS before synthesizing the channel matrix.
III. PROPOSED UL-ASSISTED PARAMETRIC CSI FEEDBACK AND PREDICTION
- The UE encoder directly compresses DL pilot observations into feedback symbols without explicit channel estimation.This avoids encoding an estimated channel matrix.Source: §III · p. 3
- The BS decoder predicts low-dimensional channel parameters and synthesizes the channel matrix from them rather than directly outputting a high-dimensional matrix.Source: §III · p. 3
- The paper introduces UE-side encoder Fenc and BS-side decoder Fdec, with their detailed structures summarized in Table I.Source: §III · p. 3
A. UE-Side Encoder
The UE encoder maps received DL pilot observations directly to real-valued feedback symbols, without explicitly estimating the DL channel. A shared encoder processes all UEs’ observations.
A. UE-Side Encoder
- The shared UE encoder maps each pilot observation to a real-valued feedback symbol without explicitly estimating the DL channel.It first stacks the observation’s real and imaginary parts, then compresses the resulting representation.Source: A. UE-Side Encoder · p. 3
- The encoder uses convolutional layers, adaptive pooling, and a fully connected layer to compress the pilot observations.Source: A. UE-Side Encoder · p. 3
- Feedback-symbol normalization ensures unit-average transmit power, and the BS is assumed to receive the symbols without error.Source: A. UE-Side Encoder · p. 3
B. BS-Side Decoder
The BS decoder combines each UE’s feedback symbol with reciprocal UL channel parameters to predict future DL parameters. It predicts changes relative to the UL parameters and synthesizes a DL channel matrix from the predicted parameters.
B. BS-Side Decoder
- The decoder combines each UE’s feedback representation with its UL parameter prior to predict future DL channel parameters independently.The UL parameters are normalized, with angles represented by their sine and cosine values.Source: B. BS-Side Decoder · pp. 3–4
- The decoder predicts future path gains and normalized changes to delay, Doppler, and angles relative to the UL parameters.Source: B. BS-Side Decoder · p. 4
- The residual formulation makes the decoder learn UL–DL mismatch and temporal variations rather than absolute parameters.The changes typically have smaller dynamic ranges than the absolute parameters, which facilitates network training.Source: B. BS-Side Decoder · p. 4
- The predicted parameters are substituted into the physical channel model to produce the future DL channel matrix.Predicted azimuth is wrapped to its principal interval, and zenith is clipped to [0, π].Source: B. BS-Side Decoder · p. 4
C. End-to-End Training
The UE encoder and BS decoder are jointly optimized using channel-matrix and channel-parameter losses. The two losses balance end-to-end channel prediction accuracy with parameter-level fidelity.
C. End-to-End Training
- Joint training combines predicted channel-matrix loss with channel-parameter loss to balance channel accuracy and parameter fidelity.The channel-matrix loss optimizes the synthesized DL channel, while the parameter loss supervises each predicted path parameter.Source: C. End-to-End Training · p. 4
- The channel-matrix loss is computed over UEs and prediction-period OFDM symbols, while parameter loss is computed over paths, UEs, and symbols.Source: C. End-to-End Training · p. 4
- The encoder and decoder parameters are jointly optimized using Adam.Source: C. End-to-End Training · p. 4
A. System Configurations
The evaluation uses a Sionna RT dataset with four UEs, four BS antennas, 32 subcarriers, and moving UEs. The configuration includes a 1/64 compression ratio and training across multiple SNR values.
A. System Configurations
- The Sionna RT dataset uses 4 UEs, 4 BS antennas, 32 subcarriers, 4 paths per UE, and UE speed of 108 km/h.The UL and DL carrier frequencies are 3.4 GHz and 3.5 GHz, respectively.Source: A. System Configurations · p. 4
- The configuration uses 16 DL training symbols, 8 UL feedback symbols, and 32 DL prediction symbols, yielding a compression ratio of 1/64.Source: A. System Configurations · p. 4
- As UEs move along trajectories, propagation geometry changes the path gains, delays, angles, and Doppler shifts; UL parameters at t0 are assumed perfectly known.DL training uses all subcarriers for pilot symbols.Source: A. System Configurations · p. 4
- Table I summarizes the encoder and decoder layer settings and their input and output dimensions.Source: A. System Configurations · p. 4, Table I · p. 5
- The 50,000 samples are split into 40,000 training, 5,000 validation, and 5,000 test samples, with batch SNRs drawn from 0 to 25 dB.Models are trained for 100 epochs with batch size 256 and learning rate 5 × 10^-4.Source: A. System Configurations · p. 5
B. Baseline Definitions
The baselines differ in whether UEs estimate the DL channel and whether BS decoders use reciprocal UL parameters and a physical channel model. The proposed scheme achieves the lowest NMSE across the reported DL pilot SNRs and prediction lengths.
B. Baseline Definitions
- Baseline 1 encodes an LS-estimated DL channel matrix, while Baseline 2 directly compresses DL pilot observations without explicit channel estimation.Both decoders predict the future channel matrix without reciprocal UL parameters or the physical channel model.Source: B. Baseline Definitions · p. 5
- Baseline 3 directly compresses pilot observations and combines feedback with reciprocal UL parameters, but predicts only future path gains and reuses outdated delays, angles, and Doppler shifts.All baselines share the datasets, feedback budget, and training configuration.Source: B. Baseline Definitions · p. 5
C. Channel Prediction Performance
- The proposed scheme consistently outperforms all evaluated baselines across DL pilot SNRs and achieves the lowest NMSE at every evaluated prediction length.For prediction lengths, evaluation uses a fixed DL pilot SNR of 20 dB.Source: C. Channel Prediction Performance · pp. 5–6, Figure 3 · p. 5, Figure 4 · p. 6
- Baseline 2 performs worse than Baseline 1 because it must implicitly extract channel information from noisy pilots, whereas Baseline 1 encodes an explicitly estimated channel matrix.Baseline 3 also outperforms Baselines 1 and 2 by using UL parameters and the physical channel model, allowing it to predict low-dimensional path gains.Source: C. Channel Prediction Performance · p. 5
- For prediction lengths 8, 16, 24, and 32, NMSE decreases slightly through length 16 and then increases, while the proposed scheme remains best.Models are trained once at maximum length 32, minimizing average NMSE across the full prediction interval.Source: C. Channel Prediction Performance · pp. 5–6
D. Complexity Analysis
The proposed scheme predicts low-dimensional channel parameters and uses them to synthesize the DL channel matrix. It uses approximately 70% fewer trainable parameters than Baselines 1 and 2.
D. Complexity Analysis
- The proposed scheme uses approximately 70% fewer trainable parameters than Baselines 1 and 2 by predicting low-dimensional parameters instead of a high-dimensional channel matrix.Baseline 3 has the fewest trainable parameters because it predicts only channel path gains.Source: D. Complexity Analysis · p. 6
- Equal trainable-parameter counts do not imply equal computational cost: Baseline 1 costs more on the UE than Baseline 2 because it processes an estimated channel matrix and requires LS estimation.The schemes use the same convolutional layers, so their UE encoder parameter counts are identical.Source: D. Complexity Analysis · p. 6
V. CONCLUSION
- The proposed scheme combines UE feedback with UL parameters to predict future DL channel parameters, then synthesizes the channel matrix; it achieves lower NMSE than all evaluated baselines across SNRs and prediction lengths.The conclusion reports simulations on the Sionna RT dataset.Source: §V · p. 6