Abstract
In the era of Massive Machine-Type Communication (mMTC), base stations must detect sporadic activity from thousands of IoT devices using short signature sequences. While covariance-based detection theoretically allows for a quadratic scaling law ($K \approx L^2$) in single-cell systems, its applicability to multi-cell networks with inter-cell interference has remained a challenge. This paper proves that this quadratic scaling law holds in cooperative multi-cell massive MIMO systems. Furthermore, we introduce "Active Set" and "Inexact" Coordinate Descent algorithms that significantly accelerate the Maximum Likelihood Estimation (MLE) process, making high-capacity activity detection computationally feasible for practical deployments.
Introduction
The proliferation of IoT devices necessitates efficient Massive Machine-Type Communication (mMTC). A key characteristic of mMTC is sporadic activity: out of a vast pool of devices, only a small fraction is active at any given time. To minimize overhead, these devices transmit short, unique signature sequences for identification.
Traditional Compressed Sensing (CS) techniques typically limit the number of detectable devices ($K$) to the length of the signature sequence ($L$). However, covariance-based detection methods have the potential to break this barrier, theoretically detecting up to $L^2$ devices. The primary obstacles to implementing this in real-world networks are the interference from neighboring cells and the high computational complexity of the resulting optimization problems.
Theoretical Framework and Scaling Laws
This research establishes a rigorous theoretical foundation for multi-cell activity detection. By employing the Stable Null Space Property (NSP), the authors prove that the maximum number of detectable devices per cell increases quadratically with the sequence length $L$, even under inter-cell interference. Specifically, the scaling law is established as $K = O(L^2 / \log^2(BN/L^2))$, where $B$ is the number of cells and $N$ is the total number of devices.
As shown in the comparison table above, this work extends the state-of-the-art by providing scaling laws and estimation error bounds for multi-cell scenarios using both continuous (Type II) and discrete (Type I) signature sequences. A significant finding is that inter-cell interference only degrades detection capacity by a logarithmic factor ($\log B$), suggesting that multi-cell cooperation effectively manages interference.
Algorithmic Innovations
To solve the MLE problem efficiently, the paper moves beyond standard Coordinate Descent (CD), which is computationally expensive due to the need for exact polynomial root solving for every device. The authors propose three key enhancements:
- Inexact CD: This approach utilizes a quadratic approximation for inter-cell interference, reducing the update complexity to $O(1)$ per iteration without sacrificing the convergence properties of the MLE.
- Active Set CD: Leveraging the inherent sparsity of the problem (most devices are inactive), this strategy selectively updates only a subset of coordinates, bypassing thousands of redundant calculations for silent devices.
- Hybrid Active Set Inexact CD: By combining these methods, the authors achieve a massive computational speedup, reaching error tolerances significantly faster than vanilla CD or existing clustering-based benchmarks.
Performance Evaluation
The proposed methods were validated through extensive Monte Carlo simulations using a hexagonal multi-cell layout (7 to 9 cells) and a standard path-loss model (exponent $\gamma > 2$). The results confirm that:
- Discrete Sequences are Effective: Type I discrete sequences (easier for hardware implementation) follow the same quadratic scaling law as continuous sequences.
- Efficiency and Accuracy: The Active Set Inexact CD algorithm achieves the same low probability of error as exact MLE but in a fraction of the time, making it suitable for real-time IoT applications.
Conclusion and Future Directions
This work confirms that the quadratic scaling law for device activity detection is robust to inter-cell interference in cooperative massive MIMO systems. While the current model assumes known large-scale fading coefficients, future research will focus on scenarios where these coefficients are unknown or where devices operate under varying power constraints. Additionally, while the proposed CD algorithms perform excellently in practice, exploring global optimality guarantees for the non-convex MLE problem remains an open area for investigation.
Summary updated