Abstract
Sequence models have demonstrated the ability to perform tasks like channel equalization and symbol detection by automatically adapting to current channel conditions. This is done without requiring any explicit optimization and by leveraging not only short pilot sequences but also contextual information such as long-term channel statistics. The operating principle underlying automatic adaptation is in-context learning (ICL), an emerging property of sequence models. Prior art adopted transformer-based sequence models, which, however, have a computational complexity scaling quadratically with the context length due to batch processing. Recently, state-space models (SSMs) have emerged as a more efficient alternative, affording a linear inference complexity in the context size. This work explores the potential of SSMs for ICL-based equalization in cell-free massive MIMO systems. Results show that selective SSMs achieve comparable performance to transformer-based models while requiring approximately eight times fewer parameters and five times fewer floating-point operations.
Introduction
In the pursuit of next-generation (6G) Radio Access Network (RAN) architectures, researchers have developed a novel approach to channel equalization in cell-free massive MIMO systems. By utilizing State-Space Models (SSMs) and the principle of "in-context learning" (ICL), this method matches the performance of state-of-the-art Transformer models while being significantly more efficient. Specifically, the proposed SSM-ICL requires 8.5x fewer parameters and 5.5x fewer floating-point operations (FLOPs), making it a prime candidate for real-time deployment in resource-constrained wireless hardware.
The Challenge of Cell-Free Massive MIMO
Cell-free massive MIMO is a key architecture for future networks where numerous distributed Access Points (APs) coordinate to serve users. However, two major hurdles persist:
- Limited Fronthaul Capacity: Links between APs and the central processor (CPU) have finite bandwidth. This necessitates signal quantization, which introduces noise and non-linear distortions.
- Pilot Contamination: Interference occurs when multiple users reuse the same pilot sequences for channel estimation, degrading signal quality.
Traditional linear estimators, such as LMMSE, often struggle with these non-linearities and interference patterns. While Transformer-based sequence models can learn to mitigate these issues through ICL, their quadratic computational complexity relative to sequence length poses a significant barrier for edge deployment.
SSM-ICL: Efficient In-Context Learning
The authors introduce SSM-ICL, an equalizer based on the Mamba architecture—a type of selective State-Space Model. Unlike Transformers, which process sequences with quadratic complexity, SSMs maintain a hidden "state" or memory, allowing for inference complexity that scales linearly with the context length.
Key Innovations
- In-Context Learning (ICL): Instead of per-channel fine-tuning, the model is provided with a "prompt" containing pilot signals and long-term channel statistics. It uses this context to automatically adapt its equalization strategy.
- Contextual Prompt Design: The prompt includes quantized received signals, known pilots, and fading statistics for both target and interfering users. This "contextual awareness" allows the model to effectively "de-contaminate" pilots.
- Tokenization: Complex-valued wireless signals are converted into real-valued tokens and augmented with metadata about the modulation scheme (e.g., QAM-16) for processing by the SSM.
Experimental Setup
The study's performance was evaluated in a simulated environment based on 3GPP Urban Microcell standards at a 2 GHz carrier frequency. The setup involved a $1 \text{ km} \times 1 \text{ km}$ area with 4 APs and 4 User Equipments (UEs). Pilot sequences were length-8 Walsh-Hadamard sequences, with intentional reuse to simulate contamination. A variety of modulation schemes were tested, including BPSK, 8-PSK, and QAM sizes up to 64.
Performance and Efficiency
The results demonstrate that SSM-ICL achieves Mean Squared Error (MSE) levels comparable to Transformer-based models (T-ICL) across various Signal-to-Noise Ratios (SNR). Crucially, SSMs maintain high accuracy even with significantly reduced model sizes.
Key findings include:
- Efficiency Gains: To reach a target MSE of 0.23, SSM-ICL required 8.5x fewer parameters and 5.5x fewer FLOPs than the Transformer equivalent.
- Robustness in Low-Resource Scenarios: In a "tiny model" comparison (2 layers), SSM-ICL achieved an MSE of 0.33, significantly outperforming the Transformer model's 0.48.
- Quantization Handling: The ICL approach naturally learns to compensate for distortions from low-bitrate fronthaul links (1-4 bits per symbol), outperforming traditional linear estimators in these constrained environments.
Conclusion and Future Directions
While these results are promising, the study currently relies on 3GPP statistical models. Future work will need to address real-world hardware impairments like power amplifier non-linearities and phase noise. Additionally, scaling the approach to larger numbers of APs and users is a key area for investigation. Beyond equalization, SSMs hold potential for other in-context wireless tasks such as beamforming and resource allocation in O-RAN systems.
Summary updated