Downlink Channel Matrix Estimation from PMI-Only Feedback in FDD Systems: Maximum Likelihood and Sharp Excess Risk Bound

Jinchi Chen, Mingxi Hu, Peigang Jiang, Xin Meng, Ke Wei, Xianyin ZhangarXiv 2026 · DOI · arXiv

TL;DR

This paper tackles the challenge of estimating the wireless channel in FDD (Frequency Division Duplex) massive MIMO systems using only limited feedback from the user. They propose a novel Maximum Likelihood Estimator (MLE) that achieves a sharp theoretical performance bound and demonstrates superior practical performance compared to existing methods, even with highly quantized feedback.

What Problem Does This Paper Address?

In modern wireless communication systems like 5G, massive MIMO (Multiple-Input Multiple-Output) technology uses a large number of antennas at the base station to significantly boost data rates and reliability. However, in FDD systems, the base station and user operate on different frequency bands, meaning they cannot directly use channel reciprocity (where the uplink and downlink channels are assumed to be the same). This makes estimating the downlink channel at the base station a critical but difficult task. To reduce the overhead of sending channel information, users provide limited feedback, often in the form of a Precoding Matrix Indicator (PMI) – essentially an index pointing to a pre-defined codebook that best represents the channel. The problem is that this PMI feedback is highly quantized and nonlinear, making accurate channel estimation a significant challenge. Existing methods often rely on additional information like Channel Quality Indicators (CQI) or focus on estimating channel statistics rather than the full channel matrix. This paper aims to directly estimate the downlink channel matrix using only PMI feedback, a more challenging but practical scenario.

Methods

The core of this paper is a Maximum Likelihood Estimator (MLE) for the downlink channel matrix. The authors first model the PMI feedback process using a probabilistic perturbation model, which replaces the hard "argmax" decision with a "softmax" function. This makes the likelihood function continuous and differentiable, allowing for optimization. They then derive a constrained MLE to estimate the channel vector, with the constraint ensuring a compact parameter space.

Key technical contributions include:

  • Probabilistic PMI Model: Replacing the hard PMI selection with a softmax distribution, controlled by a "temperature" parameter $\tau$, to approximate the true feedback mechanism.
  • Constrained Maximum Likelihood Estimator (MLE): Formulating the estimation as minimizing a negative log-likelihood function subject to a norm constraint on the channel vector.
  • Cramér-Rao Bound (CRB) Derivation: Establishing a theoretical lower bound on the achievable estimation error, carefully handling the global phase ambiguity inherent in channel estimation.
  • Sharp Excess Risk Bounds: Proving both a global excess risk bound of order $O(1/\sqrt{T})$ and a sharper local excess risk bound of order $O(1/T)$ for the MLE, where $T$ is the number of feedback rounds. This $O(1/T)$ rate is shown to be tight.

Datasets

The paper evaluates its proposed method using two types of datasets:

  • Synthetic Data: Generated to verify the theoretical performance and compare the MLE against the CRB.
  • QuaDRiGa-generated FDD Channels: Realistic channel data generated using the QuaDRiGa simulator, simulating a 3D urban macro-cell (3D UMa) scenario. This dataset is used to evaluate the method's performance in practical FDD massive MIMO settings.

Key Insights and Ideas

  • PMI-only feedback is surprisingly informative: The paper demonstrates that even highly quantized PMI feedback, when processed through a principled statistical inference framework like MLE, can yield accurate channel estimates.
  • The MLE offers a principled approach: By formulating the problem probabilistically and using MLE, the authors provide a statistically sound method that goes beyond heuristic approaches.
  • $O(1/T)$ rate is achievable and tight: The theoretical analysis showing a sharp $O(1/T)$ excess risk bound for the MLE is a significant result, indicating that the estimation accuracy improves quadratically with the number of feedback rounds.
  • Robustness across settings: The MLE shows strong performance in both single-stream and multi-stream scenarios, and is relatively insensitive to the choice of the temperature parameter $\tau$.

Key Results

  • Theoretical Guarantees: The paper establishes a sharp local excess-risk bound of order $O(1/T)$ for the MLE in the real-valued single-antenna user setting.
  • Numerical Validation: Experiments show that the MLE asymptotically achieves the Cramér-Rao bound.
  • Performance Comparison: On both synthetic and QuaDRiGa-generated FDD channels, the proposed MLE method consistently outperforms baseline methods such as the spectral method, alternating minimization, and subspace phase retrieval. For instance, in the single-stream case with structured dimensionality reduction, the MLE shows a clear performance advantage.

Limitations and Future Work

  • Extension to Complex-Valued Channels and Multi-Antenna Users: While the MLE formulation is general, the sharp risk analysis is currently limited to the real-valued single-antenna user case. The authors suggest extending this analysis to more general settings, potentially leveraging the gauge-fixing technique used for CRB analysis.
  • Lifted Formulation: Investigating a "lifted" formulation based on $X=xx^T$ and clarifying its relationship with the non-convex maximum likelihood formulation is proposed as future work. This could offer further insights into the estimation problem.

Summary updated

BibTeX
@misc{chen2026downlink,
  author = {Jinchi Chen and Mingxi Hu and Peigang Jiang and Xin Meng and Ke Wei and Xianyin Zhang},
  title = {{Downlink Channel Matrix Estimation from PMI-Only Feedback in FDD Systems: Maximum Likelihood and Sharp Excess Risk Bound}},
  year = {2026},
  eprint = {2604.21271},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.21271},
}
Figure