Abstract
The millimeter wave is a promising technique for the next generation of mobile communication. The large antenna array is able to provide sufficient precoding gain to overcome the high pathloss at millimeter wave band. However, the accurate channel state information is the key for the precoding design. Unfortunately, the channel use overhead and complexity are two major challenges when estimating the channel with high-dimensional array. In this paper, we propose a two-stage approach which reduces the channel use overhead and the computational complexity. Specifically, in the first stage, we estimate the column subspace of the channel matrix. Based on the estimated column subspace, we design the training sounders to acquire the remaining coefficient matrix of the column subspace. By dividing the estimation task into two stages, the training sounders for the second stages are only targeted for the column subspace, which will save the channel uses and the computational complexity as well.
Introduction
Millimeter-wave (mmWave) communication is a cornerstone of 5G and beyond, offering massive bandwidth. However, mmWave signals suffer from high pathloss, requiring large antenna arrays to "beamform" signals effectively. To design these beams, the system needs accurate Channel State Information (CSI).
As the number of antennas grows, the cost of estimating the channel in terms of time and radio resources usually explodes. Traditional methods often require more observations than are practical. While mmWave channels are "sparse," existing sparse recovery algorithms like Orthogonal Matching Pursuit (OMP) can still be computationally expensive. This paper introduces a two-stage approach to estimate mmWave channels in massive MIMO systems by identifying the channel's column subspace first, significantly reducing both signaling overhead and computational burden.
Proposed Two-Stage Method
The proposed method exploits the low-rank nature of the mmWave channel matrix by dividing the estimation process into two distinct phases.
- Stage 1: Column Subspace Sampling: Instead of trying to estimate the whole channel at once, the system samples a small number of columns ($m$) of the channel matrix. Since the number of physical paths ($L$) is small, the entire channel's spatial structure is contained within a low-dimensional subspace. By performing a Singular Value Decomposition (SVD) on these few sampled columns, the receiver identifies the "dominant" directions (the column subspace).
- Stage 2: Learning the Remaining Matrix: Once the subspace is known, the receiver designs specific "training sounders" (measurement filters) that are mathematically aligned with that subspace. This allows the system to estimate the remaining parts of the channel matrix much more efficiently, as it no longer wastes resources looking in directions where no signal exists.
The authors use a DFT matrix for initial sampling and employ the OMP algorithm specifically for designing the analog/digital sounders in the second stage to meet hardware constraints like the use of phase shifters.
Experimental Evaluation
The performance of the two-stage method was evaluated through numerical simulations using a standard physical channel model for mmWave MIMO ($N_r = 32$, $N_t = 128$, $L = 4$ paths).
As shown in the performance scaling analysis, the total number of channel uses is significantly reduced. Increasing the number of sampled columns ($m$) in the first stage from $1L$ to $2L$ directly improves the accuracy of the subspace estimation. The Mean Square Error (MSE) drops significantly as $m$ increases, approaching the performance of a "full observation" scenario, which would be prohibitively expensive in practice.
Conclusion and Future Work
A key insight of this work is "subspace sufficiency": sampling just $m \ge L$ columns is theoretically sufficient to capture the full column subspace of the channel. By projecting second-stage measurements onto the estimated subspace, the system avoids wasting observations on the orthogonal (empty) subspace.
While effective, a performance gap remains between this two-stage method and a theoretical "full observation" model due to error propagation from the first stage. Furthermore, the current model assumes a static channel, which may not hold in high-mobility scenarios. Future work will focus on further optimizing receiver sounder design and exploring performance in multi-user scenarios.
Summary updated