Attention-Aided MMSE for OFDM Channel Estimation: Learning Linear Filters with Attention

In orthogonal frequency division multiplexing (OFDM), accurate channel estimation is crucial. Classical signal processing-based approaches, such as linear minimum mean-squared error (LMMSE) estimation, often require second-order statistics that are difficult to obtain in practice. Recent deep neural network (DNN)-based methods have been introduced to address this; yet they often suffer from high inference complexity. This paper proposes an Attention-aided MMSE (A-MMSE), a model-based DNN framework that learns the linear MMSE filter via the Attention Transformer. Once trained, the A-MMSE performs channel estimation through a single linear operation, eliminating nonlinear activations during inference and thus reducing computational complexity. To improve the learning efficiency of the A-MMSE, we develop a two-stage Attention encoder that captures the frequency and temporal correlation structure of OFDM channels. We also introduce a rank-adaptive extension that enables a flexible performance-complexity trade-off. Numerical simulations show that the proposed A-MMSE consistently outperforms other baseline methods in terms of normalized MSE across a wide range of signal-to-noise ratio (SNR) conditions. In particular, the A-MMSE and its rank-adaptive extension provide an improved performance-complexity trade-off, providing a powerful and highly efficient solution for practical channel estimation.

💡 Research Summary

This paper addresses the long‑standing challenge of accurate yet low‑complexity channel estimation for orthogonal frequency‑division multiplexing (OFDM) systems, which are central to modern 5G and upcoming 6G networks. Classical signal‑processing (SP) techniques such as least‑squares (LS) and linear minimum mean‑squared error (LMMSE) either ignore prior channel statistics (LS) or require precise knowledge of second‑order statistics that are difficult to obtain in practice (LMMSE). Recent deep‑neural‑network (DNN) approaches treat the pilot‑based channel estimation problem as an end‑to‑end image‑to‑image mapping. While these data‑driven methods improve accuracy, they typically involve hundreds of thousands of parameters, multiple nonlinear activations, and consequently high inference latency, memory consumption, and power draw. Moreover, they lack interpretability and flexibility for runtime complexity scaling.

To bridge this gap, the authors propose a model‑based DNN framework called Attention‑aided MMSE (A‑MMSE). The central idea is to use a Transformer‑style multi‑head self‑attention (MHSA) encoder to learn the coefficients of the linear MMSE filter directly, rather than learning a nonlinear mapping from pilots to channel estimates. During training, a two‑stage attention encoder extracts latent features that capture the intrinsic frequency‑domain and temporal‑domain correlations of the OFDM channel. The first stage, the Frequency Encoder, operates on the subcarrier dimension (N) with MHSA and a linear embedding sized to N, thereby learning frequency‑selective fading patterns. The second stage, the Temporal Encoder, processes the output of the Frequency Encoder across the OFDM symbol dimension (M), again using MHSA with an embedding dimension matching the total number of channel elements (N × M). The concatenated features are fed into a residual fully‑connected network that outputs the linear filter matrix.

Once training is complete, inference requires only a single matrix‑vector multiplication: the learned linear filter is applied to the noisy pilot observations, and the remaining data subcarriers are obtained by linear interpolation. No nonlinear activation functions, pooling, or recurrent operations are needed, dramatically reducing computational load and enabling efficient hardware acceleration.

The authors further extend A‑MMSE with a rank‑adaptive variant (RA‑A‑MMSE). By performing a low‑rank approximation of the learned filter (e.g., retaining only the top K singular values), the method can trade off performance against complexity at runtime. Experiments show that reducing the rank to as low as 10 % of the full dimension incurs only modest NMSE degradation while cutting the number of floating‑point operations by more than 98 %.

Extensive simulations are conducted on the 3GPP COST2100 channel model under various scenarios (urban microcell, high mobility, different Doppler spreads) and a wide SNR range (0–30 dB). The performance metric is normalized mean‑squared error (NMSE). Results demonstrate that A‑MMSE achieves roughly 56 % lower NMSE than conventional LMMSE and about 72 % lower NMSE than the recent Channelformer transformer‑based estimator. RA‑A‑MMSE retains over 74 % of the full‑rank performance while using only 1.5 % of the computational budget of Channelformer or LMMSE. The proposed methods also exhibit strong robustness to SNR mismatch, maintaining near‑optimal NMSE even when the test SNR deviates from the training distribution.

Key contributions of the paper are:

1. Hybrid Design – Integration of domain knowledge (linear MMSE structure) with the expressive power of Transformers to learn optimal linear filters.
2. Two‑Stage Attention Encoder – A novel decomposition of channel correlation into frequency and temporal components, enabling efficient capture of non‑stationary channel statistics.
3. Inference‑Only Linear Operation – After training, the estimator reduces to a single linear multiplication, eliminating the need for costly nonlinear layers and facilitating real‑time deployment on resource‑constrained devices.
4. Rank‑Adaptive Capability – Dynamic adjustment of filter rank provides a flexible complexity‑performance trade‑off, crucial for heterogeneous hardware platforms.

The paper concludes that A‑MMSE and its rank‑adaptive extension constitute a practical, high‑performance solution for OFDM channel estimation in next‑generation wireless systems. Future work is suggested in extending the framework to multi‑antenna (MIMO) scenarios, multi‑user settings, and online adaptation to time‑varying channel statistics.