Abstract
We investigate the stabilization of unstable multidimensional partially observed single-sensor and multi-sensor linear systems driven by unbounded noise and controlled over discrete noiseless channels under fixed-rate information constraints. Stability is achieved under fixed-rate communication requirements that are asymptotically tight in the limit of large sampling periods. Through the use of similarity transforms, sampling and random-time drift conditions we obtain a coding and control policy leading to the existence of a unique invariant distribution and finite second moment for the sampled state. We use a vector stabilization scheme in which all modes of the linear system visit a compact set together infinitely often. We prove tight necessary and sufficient conditions for the general multi-sensor case under an assumption related to the Jordan form structure of such systems. In the absence of this assumption, we give sufficient conditions for stabilization.
Introduction
In Networked Control Systems (NCS), the "Data Rate Theorem" establishes that an unstable system can only be stabilized if the communication link between the sensor and the controller carries enough information. Specifically, the bit rate must exceed the sum of the logarithms of the system's unstable eigenvalues. While this limit is well-understood for systems with no noise or bounded noise, the presence of unbounded Gaussian noise introduces significant complexity. Because noise can be arbitrarily large, the system state can "jump" outside the range of a fixed-rate quantizer, leading to an "overflow" event.
This paper addresses the challenge of achieving statistical stability (a finite second moment) for complex, multi-dimensional, and multi-sensor systems using a fixed-rate information constraint, even under unbounded noise.
System Architecture
The study considers both single-sensor and multi-sensor configurations. In the single-sensor case, a single observer monitors the plant and communicates with the controller over a limited-bandwidth channel.
For more complex environments, the multi-sensor setup involves multiple sensors, each potentially observing only a part of the system state. These sensors must coordinate their transmissions to a central controller to maintain stability.
Methodology
The proposed coding and control policy relies on several mathematical innovations:
- Sampling and Similarity Transforms: The system is sampled periodically. Similarity transforms are applied to convert the system matrix into Jordan Normal Form, decoupling complex dynamics into manageable blocks.
- Dynamic Quantization: A uniform quantizer with a "zoom" feature is employed. The bin size ($\Delta_t$) dynamically expands during "overflow" events to capture escaping states and shrinks when the state is within the granular region.
- Random-Time Drift Criteria: The system state and quantizer bin size are modeled as a Markov chain. Stability is proven using "random-time drift" conditions, analyzing the system at specific stopping times when the quantizer is successfully "zoomed."
- Decentralized Decomposition: For multi-sensor setups, a decentralized version of the observability canonical decomposition determines the information each sensor must transmit.
Key Insights and Results
The research yields several critical findings:
- Asymptotic Optimality: As the sampling period ($T$) increases, the required bit rate for stability converges to the theoretical minimum ($R_{min} = \sum_{|\lambda_i|>1} \log_2(|\lambda_i|)$).
- Multi-Sensor "Safety Net": Individual sensors do not need to observe the entire state. Stability is achievable as long as every unstable mode is observed by at least one sensor.
- Stability Guarantees: The policy ensures the existence of a unique invariant distribution and a finite second moment: $\lim_{t \to \infty} E[\|x_t\|^2] < \infty$.
- Vector Stabilization: The authors utilize a "vector" approach where all system modes visit a compact set simultaneously, preserving the Markovian properties necessary for proving strong stability.
Conclusion and Future Work
While the framework provides tight bounds and efficient rates, it faces practical limitations such as sampling latency—where high efficiency requires longer waiting periods. Future research will focus on extending these results to noisy channels (e.g., those with packet drops) and relaxing assumptions regarding the joint observability of eigenspaces in multi-sensor configurations.
Summary updated