Abstract
We present a method for estimating conditionally Gaussian random vectors with random covariance matrices, which uses techniques from the field of machine learning. Such models are typical in communication systems, where the covariance matrix of the channel vector depends on random parameters, e.g., angles of propagation paths. If the covariance matrices exhibit certain Toeplitz and shift-invariance structures, the complexity of the MMSE channel estimator can be reduced to $O(M \log M)$ floating point operations, where $M$ is the channel dimension. While in the absence of structure the complexity is much higher, we obtain a similarly efficient (but suboptimal) estimator by using the MMSE estimator of the structured model as a blueprint for the architecture of a neural network. This network learns the MMSE estimator for the unstructured model, but only within the given class of estimators that contains the MMSE estimator for the structured model. Numerical simulations with typical spatial channel models demonstrate the generalization properties of the chosen class of estimators to realistic channel models.
Introduction
In massive MIMO (Multiple-Input Multiple-Output) systems, base stations utilize large antenna arrays to beamform signals toward users. Achieving high spectral efficiency requires accurate Channel State Information (CSI). The Minimum Mean Squared Error (MMSE) estimator is the optimal choice for accuracy; however, for realistic non-Gaussian channels, it is often mathematically intractable or computationally prohibitive, requiring $O(M^3)$ operations. This complexity is a significant barrier for real-time implementation as the number of antennas $M$ increases. Current practical solutions often rely on linear approximations or sparsity-based methods, which may not fully capture the complexities of multi-path environments.
Methodology
The proposed approach avoids treating the neural network as a "black box" by using the mathematical derivation of a "gridded" MMSE estimator as a structural blueprint for the network architecture.
CNN-MMSE Architecture
By assuming a Toeplitz structure for the channel covariance, the MMSE estimator can be mapped to a Convolutional Neural Network (CNN) architecture. This structure involves circular convolutions and specific non-linearities.
The use of circular convolutions allows the network to be implemented efficiently using Fast Fourier Transforms (FFTs), reducing complexity from $O(M^2N)$ to $O(M \log M)$. While theoretical derivations suggest a Softmax activation, the authors found that a Rectified Linear Unit (ReLU) provides better training stability and robustness to model mismatches.
Hierarchical Learning
Training networks for large antenna arrays (e.g., $M \ge 128$) is challenging due to local optima. To address this, a hierarchical "warm-start" strategy is employed: a small network is trained first, and its weights are interpolated to initialize a larger network, which is then refined. This process ensures convergence for high-dimensional channel vectors.
Results and Discussion
The ReLU-CNN architecture demonstrates superior accuracy compared to traditional methods like Orthogonal Matching Pursuit (OMP) and Maximum Likelihood estimators in multi-path simulations using 3GPP spatial channel models.
As shown in the figure above, the hierarchical training approach is critical for maintaining low Mean Squared Error (MSE) as the system scales. Without this strategy, "Basic" training often fails to converge to a good solution for larger $M$. The CNN-based approach maintains $O(M \log M)$ complexity, offering a high-performance alternative to computationally intensive optimal estimators.
Conclusion
This work highlights that CNNs are a natural functional form for structured MMSE estimators, providing a theoretical bridge between signal processing and deep learning. The shift-invariance of CNNs aligns with the physical properties of wireless channels. Future research will focus on extending these models to irregular antenna geometries and validating performance against real-world experimental data to account for hardware imperfections.
Summary updated