Observation Matrix Design for Densifying MIMO Channel Estimation via 2D Ice Filling

In recent years, densifying multiple-input multiple-output (MIMO) has attracted much attention from the communication community. Thanks to the subwavelength antenna spacing, the strong correlations among densifying antennas provide sufficient prior knowledge about channel state information (CSI). This inspires the careful design of observation matrices (e.g., transmit precoders and receive combiners), that exploits the CSI prior knowledge, to boost channel estimation performance. Aligned with this vision, this work proposes to jointly design the combiners and precoders by maximizing the mutual information between the received pilots and densifying MIMO channels. A two-dimensional ice-filling (2DIF) algorithm is proposed to efficiently accomplish this objective. The algorithm is motivated by the fact that the eigenspace of MIMO channel covariance can be decoupled into two sub-eigenspaces, which are associated with the correlations of transmitter antennas and receiver antennas, respectively. By properly setting the precoder and the combiner as the eigenvectors from these two sub-eigenspaces, the 2DIF promises to generate near-optimal observation matrices. Moreover, we further extend the 2DIF method to the popular hybrid combining systems, where a two-stage 2DIF (TS-2DIF) algorithm is developed to handle the analog combining circuits realized by phase shifters. Simulation results demonstrate that, compared to the state-of-the-art schemes, the proposed 2DIF and TS-2DIF methods can achieve superior channel estimation accuracy.

💡 Research Summary

**
The paper addresses the problem of channel estimation in densely packed massive MIMO systems, where antenna spacing is far below the conventional half‑wavelength (e.g., λ/8, λ/10). Such ultra‑dense arrays exhibit very strong spatial correlation and electromagnetic mutual coupling, which can be captured by a structured channel covariance matrix (the “kernel”). Existing estimators—least‑squares, MMSE, compressed‑sensing, or deep‑learning based—typically ignore this rich prior information and therefore fall short of the performance of the optimal linear MMSE (LMMSE) estimator.

The authors formulate the design of the observation matrix (the Kronecker product of the transmit precoder and the receive combiner) as a mutual‑information (MI) maximization problem between the received pilot signals and the unknown channel vector. Under the Gaussian process regression (GPR) framework, maximizing MI is equivalent to minimizing the posterior covariance of the channel estimate, i.e., improving the LMMSE error.

A key theoretical contribution is the proof that the eigenspace of the channel covariance can be perfectly decoupled into two independent sub‑eigenspaces: one associated with the transmit‑side correlation (R_tx·C_tx) and the other with the receive‑side correlation (R_rx·C_rx). Consequently, the optimal observation matrix can be constructed by selecting eigenvectors from each sub‑space. This insight leads to the proposed Two‑Dimensional Ice‑Filling (2DIF) algorithm.

2DIF algorithm

1. Perform eigen‑decomposition of the full covariance Σ_h.
2. Separate the eigenvectors into transmit and receive components (U_tx and U_rx).
3. Rank the eigenvalues in each sub‑space and greedily pick the largest‑value eigenvectors, assigning them to the precoder V and the combiner W respectively.
4. Form the observation matrix X = V* ⊗ W.

The “ice‑filling” metaphor reflects the process of filling the most energetic dimensions (largest eigenvalues) first, analogous to placing ice blocks into a two‑dimensional container. The algorithm requires only a full eigen‑decomposition (O(N_T³+N_R³)) and a linear greedy search, making it computationally tractable for typical dense arrays (e.g., 64 × 64).

Hybrid implementation challenge
In practical hybrid analog‑digital architectures, the analog combiner A can often control only the phase of each antenna‑to‑RF‑chain link (phase‑only hardware). Since 2DIF assumes full complex weighting (amplitude and phase), it cannot be directly applied. To bridge this gap, the authors propose Two‑Stage 2DIF (TS‑2DIF):

Stage 1: Optimize the analog phase matrix A within a discrete phase set (e.g., {0, π/2, π, 3π/2}) while keeping the digital combiner D and the precoder V fixed.
Stage 2: With the analog matrix fixed, apply the 2DIF principle to jointly design the digital combiner D and the precoder V using the transmit/receive eigenvectors.

The two stages are iterated alternately until the posterior covariance no longer improves. This alternating optimization respects the hardware constraints yet still approaches the performance of the fully‑digital 2DIF solution.

Complexity and performance
The authors analyze computational complexity, showing that TS‑2DIF adds only a modest overhead proportional to the number of phase candidates and pilot slots. Simulations are conducted for a 64‑antenna uplink system with 8 RF chains, 16 pilot symbols, and various SNR levels (0–30 dB). Benchmarks include random precoder/combiner, a water‑filling based design, OMP compressed‑sensing, and a deep‑learning estimator. Results demonstrate that:

• 2DIF reduces NMSE by roughly 3–4 dB compared with random designs and by 2 dB over water‑filling.
• TS‑2DIF, despite the phase‑only constraint, achieves performance within 1.5 dB of the fully‑digital 2DIF.
• Both methods approach the theoretical LMMSE bound, especially at high SNR where the eigenvalue distribution of Σ_h is sharply decaying.

Contributions and limitations
The paper’s main contributions are: (i) a rigorous MI‑based formulation linking observation matrix design to LMMSE error; (ii) the decoupling of the channel covariance eigenspace, enabling a simple yet near‑optimal greedy construction (2DIF); (iii) an extension to realistic hybrid hardware (TS‑2DIF). Limitations include the need for accurate prior knowledge of Σ_h (which may vary in fast‑fading environments), the computational cost of full eigen‑decomposition for extremely large arrays, and the fact that TS‑2DIF provides only a heuristic guarantee under phase‑only constraints.

Future directions suggested by the authors involve: adaptive estimation of the covariance matrix, low‑complexity approximate SVD methods, multi‑user or multi‑cell extensions where observation matrices can be jointly optimized, and integration with reconfigurable intelligent surfaces (RIS) to further exploit spatial correlation.

In summary, the paper presents a theoretically grounded and practically viable framework for designing transmit precoders and receive combiners in densifying MIMO systems. By leveraging the structured channel covariance through the 2D Ice‑Filling concept, it achieves substantial gains over state‑of‑the‑art channel estimation techniques, and its two‑stage hybrid extension makes it applicable to modern phased‑array hardware.