Channel Estimation for Downlink Communications Based on Dynamic Metasurface Antennas

1. Problem Statement

The paper addresses the challenge of Channel Estimation (CE) in downlink Multiple-Input Single-Output (MISO) systems utilizing Dynamic Metasurface Antennas (DMAs) and Orthogonal Frequency Division Multiplexing (OFDM).

• The DMA Context: DMAs are emerging as a low-cost, energy-efficient alternative to traditional massive MIMO arrays. They consist of waveguides with tunable radiating elements.
• The Core Challenge: In DMA systems, the effective channel is a product of two distinct components:
  1. The wireless channel ($h$) between the antenna elements and the user.
  2. The inner waveguide propagation vector ($m$), representing signal propagation within the DMA microstrips.
• Limitations of Existing Work: Previous CE methods for DMAs often assumed perfect knowledge of the inner waveguide vector ($m$) or relied on complex pilot sequences. In practice, $m$ is often unknown due to manufacturing mismatches or non-homogeneous element behavior. Furthermore, separating these two components is crucial because the inner channel ($m$) is relatively static, while the wireless channel ($h$) changes rapidly.
• Goal: To develop a scheme that jointly estimates the wireless channel, the unknown inner waveguide vector, and the transmitted data symbols without requiring separate pilot training phases, thereby reducing latency and overhead.

2. Methodology

The authors propose a data-aided, semi-blind channel estimation scheme based on Tensor Modeling and PARAFAC (Parallel Factor) Decomposition.

A. System Model

• Signal Structure: The system uses $D$ waveguides with $L$ elements each ($N=DL$ total elements) and $K$ subcarriers.
• Received Signal: The received signal at subcarrier $k$, time $t$, and beamforming state $p$ is modeled as:
  $$y_{k,t,p} = h_k^H \text{diag}\{m\} f_p^{DMA} s_t + n_{k,t,p}$$
  Where $f_p^{DMA}$ is the beamforming vector, $s_t$ is the symbol, and $n$ is noise.
• Tensor Formulation: By collecting data over $T$ symbols, $K$ subcarriers, and $P$ beamforming states, the received signal is structured as a 3rd-order tensor $\mathcal{Y} \in \mathbb{C}^{K \times T \times P}$. This tensor admits a PARAFAC decomposition:
  $$\mathcal{Y} = \mathcal{I}_{3,N} \times_1 \mathbf{H} \times_2 \mathbf{X} \times_3 \mathbf{F} + \mathcal{N}$$
  Where:
  • $\mathbf{H}$: Frequency-domain MISO channel matrix.
  • $\mathbf{X} = \mathbf{s}\mathbf{m}^T$: A rank-1 matrix containing the product of transmitted symbols and the inner channel vector.
  • $\mathbf{F}$: DMA training matrix (beamforming vectors).

B. Estimation Algorithm

The proposed solution uses a Two-Stage Semi-Blind Receiver:

1. Stage 1: Joint Estimation via Alternating Least Squares (ALS)

   • The algorithm solves a multilinear optimization problem to jointly estimate $\mathbf{H}$ and the rank-1 matrix $\mathbf{X}$.
   • It utilizes the ALS algorithm, iteratively updating $\mathbf{H}$ and $\mathbf{X}$ by solving least-squares problems based on the mode-1 and mode-2 matrix unfoldings of the tensor.
   • This stage operates in a data-aided manner, meaning it uses the actual transmitted data symbols (not just pilots) to refine the estimates.

2. Stage 2: Rank-1 Approximation

   • Once $\hat{\mathbf{X}}$ is estimated, it is decomposed to separate the symbol vector $\mathbf{s}$ and the inner channel vector $\mathbf{m}$.
   • Since $\mathbf{X}$ is theoretically rank-1, the authors apply Truncated Singular Value Decomposition (SVD) to $\hat{\mathbf{X}}$ to extract $\hat{\mathbf{s}}$ and $\hat{\mathbf{m}}$.

C. Uniqueness and Conditions

The paper analyzes the uniqueness of the PARAFAC decomposition. While standard Kruskal's condition is too restrictive for rank-1 matrices, the authors rely on relaxed conditions (e.g., from [25], [26]) proving that if the factor matrices are drawn from continuous distributions (Gaussian for channels, random phases for waveguides, QAM for symbols), the decomposition is unique almost surely, even with a rank-1 factor.

3. Key Contributions

1. Decoupled Estimation: The first work to successfully separate the estimation of the wireless channel and the DMA inner waveguide vector without prior knowledge of the latter. This allows for independent beamformer design that compensates for the static inner channel regardless of the fast-fading wireless channel.
2. Data-Aided Semi-Blind Approach: The method eliminates the need for distinct pilot and data transmission phases. It jointly estimates channel parameters and data symbols, reducing latency and spectral overhead.
3. Tensor-Based Framework: It introduces a PARAFAC-based tensor model specifically for DMA systems, leveraging the multilinear structure of the received signal to solve the underdetermined estimation problem.
4. Low-Complexity Iterative Receiver: The proposed algorithm uses standard ALS and SVD operations, offering a computationally feasible solution that converges rapidly at moderate-to-high SNR.

4. Simulation Results

The authors evaluated the scheme using 10,000 Monte Carlo runs with parameters: $K=8$ subcarriers, $N=16$ elements, $P=32$ beamforming states, and $T=10$ symbols (64-QAM).

• Channel Estimation Accuracy (NMSE):
  • The proposed method achieved accurate estimation for both the wireless channel ($\mathbf{H}$) and the inner vector ($\mathbf{m}$).
  • When using a semi-unitary DMA precoder, the proposed method showed a ~5 dB improvement in Normalized Mean Square Error (NMSE) compared to benchmarks.
  • The performance gap between estimating $\mathbf{H}$ and $\mathbf{m}$ was negligible, despite $\mathbf{m}$ being estimated in a second stage (demonstrating robustness against error propagation).
• Symbol Error Rate (SER):
  • The SER curves followed the same trend as the channel estimation performance, confirming that the data-aided approach successfully decodes symbols during the estimation phase.
  • The proposed method outperformed benchmarks that relied solely on pilots or assumed simplified precoders.
• Computational Complexity:
  • The algorithm converges quickly (few iterations) in medium-to-high SNR regimes ($\ge 15$ dB).
  • At low SNR, more iterations are required, but the overall runtime remains competitive.
  • Using a semi-unitary precoder and pilot-aided benchmarks further reduced complexity by avoiding matrix inversions.

5. Significance and Future Work

• Practical Impact: This work bridges the gap between theoretical DMA models and practical implementation by addressing the unknown nature of the internal waveguide propagation. It enables more robust beamforming designs that do not need to constantly re-estimate the static internal hardware characteristics.
• Efficiency: By removing the need for dedicated pilot sequences, the system improves spectral efficiency and reduces latency, which is critical for 6G and next-generation networks.
• Future Directions: The authors suggest extending this research to:
  • Multi-user scenarios.
  • Reducing the dependency on the number of DMA measurements.
  • Handling scenarios where the DMA training matrix is imperfectly known at the receiver (e.g., due to limited feedback or hardware impairments).

In summary, this paper presents a novel, efficient, and robust channel estimation framework for DMA-based MISO-OFDM systems, utilizing tensor decomposition to jointly estimate channels and data without the overhead of traditional pilot-based training.