TL;DR
This paper analyzes the spectral efficiency of Large Intelligent Surfaces (LISs) in practical wireless environments, focusing on the impact of pilot contamination and optimizing pilot training length. The key takeaway is that LISs can achieve high performance even with pilot contamination, especially when using optimized training strategies, and their performance is fundamentally different from traditional massive MIMO systems.
What Problem Does This Paper Address?
The paper tackles the challenge of maximizing spectral efficiency (SSE) in Large Intelligent Surfaces (LISs), a novel wireless communication paradigm that leverages massive antenna arrays integrated into man-made structures. While LISs promise reliable, space-intensive communication through line-of-sight (LOS) channels, a critical issue arises from pilot contamination. In practical uplink communication, devices transmit pilot signals to help the LIS estimate the channel. When multiple LISs share the same frequency band and reuse pilot sequences, these signals become non-orthogonal, leading to interference that degrades performance. This paper aims to analyze and mitigate the impact of this pilot contamination on SSE, particularly in multi-LIS scenarios, and to determine optimal system parameters like pilot training length and device scheduling.
Methods
The authors employ asymptotic analysis to study the system spectral efficiency (SSE) as the number of antennas on the LIS ($M$) approaches infinity. This approach allows for deriving theoretical performance bounds and understanding the fundamental behavior of LIS systems.
Key methodological contributions include:
- Asymptotic SSE Analysis: Deriving the SSE under a practical uplink frame structure that includes pilot training and data transmission periods.
- Pilot Contamination Modeling: Specifically addressing the non-orthogonality of pilot sequences in multi-LIS systems and its impact on channel estimation.
- Performance Bound Derivation: Establishing theoretical upper bounds on SSE for LIS systems, considering pilot contamination and interference.
- Optimal Pilot Training Length Characterization: Developing a method to determine the optimal pilot training duration that maximizes SSE, even in the presence of pilot contamination.
- Device Scheduling Optimization: Deriving the optimal number of devices to schedule per LIS to maximize overall performance.
- Channel Modeling: Utilizing deterministic LOS channels for desired signals and spatially correlated Rician fading for interference channels, reflecting a realistic LIS environment.
Datasets
Not discussed in the paper. The evaluation relies on Monte Carlo simulations with parameters based on LTE specifications.
Key Insights and Ideas
- LISs are fundamentally different from Massive MIMO: Unlike massive MIMO where performance is heavily limited by pilot contamination, LISs show a remarkable resilience. As the number of antennas ($M$) increases, the impact of pilot contamination and interference through NLOS paths becomes negligible.
- Channel hardening is a key LIS characteristic: As $M$ grows, the SSE converges to a deterministic value, indicating improved reliability and reduced latency. This "channel hardening" effect is a significant advantage.
- Optimal pilot training is surprisingly simple: The optimal pilot training length ($t$) to maximize SSE converges to $K$ (the number of devices per LIS), meaning one pilot symbol per device is sufficient, regardless of pilot contamination. This is a crucial finding for practical system design.
- Performance convergence in multi-LIS: In multi-LIS systems, as $M$ increases, the performance gap between single-LIS and multi-LIS systems narrows, especially when interference is primarily through the LOS path. This suggests that with sufficient antennas, multi-LIS systems can approach the performance of a single-LIS with perfect CSI.
- Trade-offs in device scheduling: While more scheduled devices ($K$) can increase overall network capacity, it also leads to increased pilot training overhead and interference, creating a trade-off that needs careful optimization.
Key Results
- The asymptotic analysis accurately predicts the SSE, closely matching simulation results.
- The derived optimal pilot training length ($t=K$) achieves near-brute-force optimal performance for SSE.
- The SSE of a multi-LIS system is bounded by pilot contamination, intra-LIS, and inter-LIS interference through the LOS path.
- As $M$ increases, pilot contamination and NLOS interference become negligible, leading to performance convergence.
- Simulation results demonstrate that the proposed pilot training length and scheduling strategy achieve performance close to optimal solutions found through brute-force search. For instance, Figure 10 shows the proposed method achieving performance very close to brute-force search across a wide range of $M$.
Limitations and Future Work
- MMSE vs. LS Estimator: While the paper uses a simpler Least Squares (LS) estimator for channel estimation due to complexity concerns in LIS systems, it acknowledges that a Minimum Mean Square Error (MMSE) estimator could potentially offer better performance by reducing pilot contamination, though at a higher computational cost.
- Practical Implementation: The paper focuses on theoretical analysis and simulation. Future work could explore the practical implementation challenges and performance of these strategies in real-world LIS deployments.
- Network LIS Cooperation: The paper touches upon network LIS cooperation for joint scheduling, suggesting this as a promising avenue for further research to maximize network SE.
- Standardization: The authors emphasize the need to consider adequate frame structures, including pilot training length and scheduled devices, for future LIS standardization.
Summary updated