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

1. Problem Statement

The paper addresses the challenge of downlink channel estimation in Frequency-Division Duplex (FDD) massive Multiple-Input Multiple-Output (MIMO) systems under a 5G New Radio (NR) type limited-feedback architecture.

• Context: In FDD systems, channel reciprocity cannot be exploited. The Base Station (BS) has $d$ antennas, and the user has a single antenna (initially). The downlink channel vector is $h \in \mathbb{C}^d$.
• Feedback Mechanism: The system uses a shared codebook $\mathcal{V} = [v_1, \dots, v_N] \in \mathbb{C}^{p \times N}$. At each communication round $t$, the BS sends a dimensionality-reduction matrix $Q_t$. The user computes the effective channel and selects the codeword index $i_t$ that maximizes the channel gain:
  $$i_t = \arg \max_{i} |v_i^H Q_t^H h|^2$$
  The user feeds back only this index, known as the Precoding Matrix Indicator (PMI).
• The Challenge: The BS must reconstruct the high-dimensional complex channel vector $h$ solely from the sequence of discrete PMI indices $\{i_t\}_{t=1}^T$. Crucially, the BS does not have access to the Channel Quality Indicator (CQI) values (the actual gain magnitudes), which are often available in other works but omitted here to reflect a stricter, more challenging scenario. The observations are highly quantized, nonlinear, and ordinal (comparative).

2. Methodology

A. Probabilistic Perturbation Model

To enable statistical inference, the authors replace the deterministic "hard" arg-max decision rule with a stochastic perturbation model using the Gumbel-max trick.

• The observed index $I_t$ is modeled as a random variable following a multinomial distribution derived from a softmax function:
  $$P(I_t = i) = \frac{\exp(|a_{t,i}^H h|^2 / \tau)}{\sum_{j=1}^N \exp(|a_{t,j}^H h|^2 / \tau)}$$
  where $a_{t,i} = Q_t v_i$ and $\tau > 0$ is a temperature parameter controlling the sharpness of the approximation (as $\tau \to 0$, it approaches the hard decision).

B. Constrained Maximum Likelihood Estimator (MLE)

Based on the probabilistic model, the authors formulate the estimation as a constrained optimization problem.

• Objective: Minimize the negative log-likelihood (NLL):
  $$L_T(x) = \frac{1}{T} \sum_{t=1}^T \log \left( \sum_{j=1}^N \exp\left( \frac{|a_{t,j}^H x|^2 - |a_{t,i_t}^H x|^2}{\tau} \right) \right)$$
• Constraint: The estimator $\hat{h}_T$ is found by minimizing $L_T(x)$ subject to $\|x\|_2 \leq R$, where $R$ is a bound on the channel energy. This ensures the problem is well-posed and parameter space is compact.
• Interpretation: The authors demonstrate that this likelihood objective acts as a two-stage relaxation of the hard empirical decision error (0-1 loss):
  1. Replacing the discontinuous indicator with a margin loss.
  2. Smoothing the non-smooth max operator using the log-sum-exp (Moreau envelope) function.

C. Cramér-Rao Bound (CRB)

The paper derives the theoretical lower bound on the estimation error variance.

• Phase Ambiguity: Since PMI feedback is invariant to global phase rotation ($h$ and $h e^{j\phi}$ yield the same PMI), the Fisher Information Matrix (FIM) is singular.
• Gauge Fixing: The authors handle this by restricting the parameter space to a subspace orthogonal to the phase rotation direction (gauge fixing). They derive the CRB for the real-valued representation of the complex channel, showing the bound scales as $O(1/T)$.

3. Key Contributions

1. Statistical Formulation: The first principled statistical inference framework for direct channel estimation using PMI-only feedback (without CQI), treating the problem as a constrained MLE under a Gumbel-max perturbation model.
2. Sharp Excess Risk Bounds:
   • Global Bound: Proved a global excess-risk bound of order $O(1/\sqrt{T})$.
   • Local Sharp Bound: Under suitable identifiability conditions (Assumptions 1-3 regarding measurement matrices and codebook properties), the authors proved a sharp local excess-risk bound of order $O(1/T)$. This implies the Mean Squared Error (MSE) decays at the optimal parametric rate.
   • The proof utilizes local strong convexity of the population risk and concentration inequalities (McDiarmid's inequality, Matrix Bernstein).
3. Theoretical-CR Alignment: The analysis establishes that the MLE asymptotically attains the derived Cramér-Rao bound.
4. Extension to Multi-Stream: While theoretical proofs focus on the single-antenna user case, the methodology is extended to multi-antenna users and multi-stream scenarios in the numerical experiments.

4. Numerical Results

The authors validated their method using both synthetic data and realistic channels generated by QuaDRiGa (3GPP 3D Urban Macro-cell scenario).

• CRB Validation: Experiments confirmed that as the number of feedback rounds $T$ increases, the empirical MSE of the MLE converges to the theoretical CRB, following the $c/T$ decay rate.
• Comparison with Baselines: The MLE was compared against:
  • Spectral Method: A heuristic eigenvector-based approach.
  • Alternating Minimization (AM): An optimization method often used with CQI.
  • Subspace Phase Retrieval (Subspace-PR): A method leveraging uplink-downlink spatial correlation.
• Performance:
  • The proposed MLE outperformed all baselines in both single-stream and two-stream scenarios.
  • It showed robustness even when the dimensionality-reduction matrices were fully random.
  • Notably, the MLE achieved superior performance without using CQI, whereas baselines like AM and Subspace-PR typically rely on CQI values for better performance. This highlights the efficiency of the likelihood-based formulation in extracting information from ordinal PMI data.
• Ablation Studies:
  • Temperature ($\tau$): The method is robust to $\tau$ within a moderate range (e.g., $[5, 10]$).
  • Initialization: The MLE is largely insensitive to initialization (identity, random, or spectral), unlike some non-convex baselines.

5. Significance and Future Work

• Significance: This work bridges the gap between theoretical statistical inference and practical 5G/6G limited-feedback systems. It proves that high-precision channel estimation is possible using only discrete PMI feedback, which is the standard in current cellular networks, without requiring the transmission of additional CQI overhead. The derivation of the sharp $O(1/T)$ risk bound provides a rigorous theoretical guarantee for the efficiency of this approach.
• Future Directions:
  • Extending the sharp risk theory to complex-valued channels and multi-antenna user settings (currently proven for real-valued single-antenna).
  • Investigating lifted formulations (using $X = hh^H$) to analyze the relationship between convex relaxations and the non-convex MLE.

In summary, the paper provides a mathematically rigorous and empirically validated solution for downlink channel estimation in FDD massive MIMO systems, demonstrating that Maximum Likelihood Estimation on PMI-only feedback is statistically optimal and practically superior to existing heuristic methods.