In classical control theory and signal processing, tracking the internal state of a dynamic system requires an observer or filter (such as a Kalman filter or a Luenberger observer). Traditionally, these systems rely on a centralized architecture: a single central processing unit aggregates every sensor measurement across the entire network, updates a global state vector, and computes the corresponding error covariance matrices.
While mathematically straightforward, this centralized approach hits a hard scalability wall when applied to modern, massive cyber-physical systems like continental power grids, large robotic swarms, or dense sensor networks.
The Centralized Bottleneck
To understand why centralized estimation fails to scale, consider a network with $N$ distinct states. A centralized filter must continuously manipulate matrices of size $N \times N$. As the network expands, the computational complexity of matrix inversions and multiplications scales cubically ($\mathcal{O}(N^3)$).
Beyond computation, the communication overhead becomes prohibitive. Transmitting raw, high-frequency sensor readings from thousands of geographically dispersed nodes to a single central hub chokes available network bandwidth. Furthermore, a centralized architecture introduces a single point of failure; if the central hub experiences a hardware malfunction or a targeted cyber-attack, the entire system loses visibility.
To bypass these limitations, a prominent research direction focuses on Distributed State Estimation. Instead of forcing a single processor to understand the entire world, the system is broken down into localized pieces that collaborate to reconstruct either the global state or relevant local subsets of it.
Partitioning and Subsystem-Level Estimation
The most direct way to distribute the computational load is to partition a massive system into smaller, interconnected subsystems. Each subsystem runs its own local estimator to track its internal states, treating the influence of neighboring subsystems as external disturbances or coupled inputs.
For instance, Riverso et al. proposed a modular "Plug-and-Play" distributed estimation framework in Plug-and-play distributed state estimation for linear systems and extended it to bounded-error settings in Distributed bounded-error state estimation for partitioned systems based on practical robust positive invariance. This architecture ensures that when a new subsystem joins or leaves the network, only its immediate neighbors need to adjust their local estimators, preventing a global computational recalculation.
Similarly, Mclaughlin et al. leverage hierarchical learning in Fast and Robust State Estimation and Tracking via Hierarchical Learning to divide large networks into manageable subnetworks. By utilizing a parameter server to coordinate inter-subnetwork dynamics, their approach accelerates convergence and isolates communication failures, ensuring that a localized disruption does not compromise the entire system.
Functional Observation and Clustering
A key insight in scaling up estimation is realizing that not every node needs to know the entire global state. If a local controller only requires a specific linear combination or a regional average of the state vector to execute its control law, tracking the full high-dimensional state is wasteful.
As Mitra et al. demonstrate in Distributed Observers for LTI Systems, shifting from full-state estimation to distributed functional observers eliminates the need for "state augmentation"—a costly practice where every node maintains a redundant copy of the entire global state vector. Instead, nodes focus strictly on estimating specific linear functions of the state.
Another way to compress the state space is through spatial aggregation. Cheng et al. in Clustering-based average state observer design for large-scale network systems group network nodes into clusters and design observers to estimate only the average state of each cluster. This compresses the mathematical dimension of the observer from the total number of nodes down to the number of clusters. Deplano et al. formalize this concept in Scale-Free Estimation of the Average State in Large-Scale Systems, proving that under certain structural conditions, the estimation error bound of an average-state observer can remain completely independent of the overall system size.
Consensus and Information Fusion
When a system is distributed, local estimators cannot operate in total isolation; they must reconcile the physical coupling along the boundaries of their subsystems. Without communication, local "blind spots" (unobservable dynamics) would cause local filters to diverge. Distributed estimation algorithms solve this by incorporating consensus protocols and iterative optimization techniques.
In Cooperative ℋ∞-estimation for large-scale interconnected linear systems, Wu et al. show how local estimators can track a small subset of relevant states while exchanging information with immediate neighbors to fill in unobservable dynamics. To minimize the resulting communication overhead, Sahu et al. introduce a subspace consensus method in $\mathcal {CIRFE}$: A Distributed Random Fields Estimator, which allows nodes to ignore irrelevant dimensions of a high-dimensional field and only exchange parameters that are mutually necessary.
For optimization-based estimation, mathematical frameworks like the Alternating Direction Method of Multipliers (ADMM) or iterative projections are used to fuse local data optimally. Pasqualetti et al. utilize iterative projections in Distributed Estimation via Iterative Projections with Application to Power Network Monitoring to achieve decentralized monitoring without a central bottleneck. Furthermore, Luo et al. prove in Globally Optimal Multisensor Distributed Random Parameter Matrices Kalman Filtering Fusion with Applications that distributed fusion can achieve the exact same mathematical optimality as a centralized Kalman filter, even under complex conditions like correlated sensor noise and uncertain system parameters.
Application to Critical Infrastructure and Robotics
The theoretical necessity of distributed estimation is vividly illustrated in modern smart grids and mobile robotics.
Large-scale power grids are highly vulnerable to cyber-attacks, such as False Data Injection (FDI) attacks, where malicious actors corrupt sensor readings to mislead operators. In a centralized setup, an attacker only needs to compromise the central data stream to blind the utility provider. Özay et al. explore this vulnerability in Sparse Attack Construction and State Estimation in the Smart Grid: Centralized and Distributed Models, demonstrating how distributed estimation using ADMM can robustly detect stealthy, sparse attacks by cross-verifying localized estimations. To protect these network measurements from becoming bottlenecks, Aziz et al. propose a two-tier secure architecture in T2S2G: A Novel Two-Tier Secure Smart Grid Architecture to Protect Network Measurements, mitigating the performance overhead of traditional centralized monitoring.
Real-time grid operations also require integrating diverse sensor types, such as traditional SCADA systems and modern Phasor Measurement Units (PMUs). Du et al. address this in Distributed Hybrid Power State Estimation Under PMU Sampling Phase Errors, showing that distributed estimation can handle sampling phase errors across wide areas without overloading a single control center. Xia et al. reinforce this in Distributed State Estimation of Multi-region Power System based on Consensus Theory, demonstrating how multi-region power systems can maintain security and stability by communicating exclusively with adjacent regions.
Beyond power grids, this scalability is crucial for mobile robotics operating in GPS-denied environments. Jung et al. demonstrate in Modular Meshed Ultra-Wideband Aided Inertial Navigation with Robust Anchor Calibration that updating global error matrices for a large number of reference nodes quickly exceeds the capacity of small, energy-efficient onboard computers. By distributing the estimation filter across a modular, meshed network, robots can dynamically expand their localization network on the fly without hitting a computational wall.
Key Foundational Work
Distributed Estimation via Iterative Projections with Application to Power Network Monitoring (2011) — contributed 1 concept elements to this direction.
Globally Optimal Multisensor Distributed Random Parameter Matrices Kalman Filtering Fusion with Applications (2008) — contributed 1 concept elements to this direction.