FDD CSI Feedback under Finite Downlink Training: A Rate-Distortion Perspective

This paper establishes the theoretical limits of channel state information (CSI) feedback in frequency-division duplexing (FDD) multi-antenna orthogonal frequency-division multiplexing (OFDM) systems under finite-length training with Gaussian pilots. The user employs minimum mean-squared error (MMSE) channel estimation followed by asymptotically optimal uplink feedback. Specifically, we derive a general rate-distortion function (RDF) of the overall CSI feedback system. We then provide both non-asymptotic bounds and asymptotic scaling for the RDF under arbitrary downlink signal-to-noise ratio (SNR) when the number of training symbols exceeds the antenna dimension. A key observation is that, with sufficient training, the overall RDF converges to the direct RDF corresponding to the case where the user has full access to the downlink CSI. More importantly, we demonstrate that even at a fixed downlink SNR, the convergence rate is inversely proportional to the training length. The simulation results show that our bounds are tight, and under very limited training, the deviation between the overall RDF and the direct RDF is substantial.

💡 Research Summary

This paper investigates the fundamental information‑theoretic limits of channel state information (CSI) feedback in frequency‑division duplex (FDD) massive‑MIMO OFDM systems when the downlink training phase is of finite length and uses Gaussian pilots. The user equipment (UE) first performs minimum‑mean‑square‑error (MMSE) channel estimation based on the received pilot symbols and then compresses the estimate for uplink feedback. By modeling the whole process as a remote source‑coding problem (H → S → Z), the authors derive a closed‑form overall rate‑distortion function (RDF) that captures the trade‑off between the number of feedback bits and the mean‑square error distortion on the reconstructed channel at the base station (BS).

The main result (Theorem 1) expresses the overall RDF as the sum of three terms: (i) the direct RDF R_H(d) for compressing the true channel H under distortion d, (ii) a reduction term ΔR_{P,S} that quantifies how much the MMSE estimate S reduces source uncertainty compared with H, and (iii) an increase term ΔR_{P,d} that accounts for the fact that, after MMSE estimation, the effective distortion budget becomes d − D_{mmse|P}. The reduction term is simply half the difference of differential entropies h(S|P) − h(H) and can be rewritten as the entropy gap between the noisy and noiseless received pilot observations. The increase term depends on the rank mismatch between H and S and on the residual distortion after estimation.

Proposition 1 shows that compressing the MMSE estimate always requires fewer bits than compressing the raw channel, because the estimate discards the component of H that is uncorrelated with the pilot observations. Theorem 2 provides non‑asymptotic upper and lower bounds on the expected value of ΔR_{P,S} when the pilot matrix is drawn from a Gaussian ensemble and the number of training symbols n_t is at least the number of transmit antennas m_t. The bounds involve the downlink SNR, the eigenstructure of the channel covariance C_H, and the digamma function ψ(·). Importantly, the expected RDF reduction decays as O(1/n_t); thus, even at a fixed downlink SNR, increasing the training length makes the noisy channel behave increasingly like a noiseless one. In the large‑n_t regime, ΔR_{P,S} converges to a constant C_S with a 1/n_t correction term, where C_S is sandwiched between two expressions that depend on tr(C_H^{-1}) and SNR_{dl}.

Theorem 3 addresses the opposite effect: because the MMSE estimator consumes part of the distortion budget, the effective distortion for compression becomes stricter, leading to an increase ΔR_{P,d}. This term grows with the rank deficiency r_H − r_S and with the logarithm of the ratio d/(d − D_{mmse|P}). Consequently, when training is scarce (small n_t), ΔR_{P,d} can dominate, causing the overall RDF to be substantially larger than the direct RDF.

Simulation results validate the tightness of the derived bounds. With ample training (large n_t), the overall RDF practically coincides with the direct RDF, confirming the theoretical convergence. Conversely, with very limited training, the gap between the two RDFs is pronounced, illustrating the practical impact of training overhead on feedback efficiency.

The authors also discuss practical considerations. The derived RDF assumes joint encoding over many pilot observations, which implies a long feedback delay; real‑time “one‑shot” feedback remains an open problem. Nonetheless, the paper provides a unified analytical framework that links downlink training length, pilot power, channel statistics, and feedback rate, offering valuable design guidelines for future 6G FDD massive‑MIMO systems.