In classical control theory, we rely on a beautiful result called the Separation Principle. It tells us that we can design a system's "eyes" (the state estimator) and its "brain" (the controller) independently. You build the best possible estimator to figure out what is happening, and then you apply the best possible control law as if that estimate were the absolute truth.
However, as Ather Gattami and Mikael Johansson point out in Distributed Output-Feedback LQG Control with Delayed Information Sharing, this principle collapses the moment we move to decentralized systems with communication delays. In these networked environments, the "eyes" and "brains" of different agents become inextricably linked.
Why the Separation Principle Fails
The core reason for this failure is a coupling of uncertainty. In a centralized system, the controller knows everything the estimator knows. But in a decentralized system—like a swarm of drones or a power grid—Agent A might take an action based on local information that Agent B hasn't seen yet.
As Xiao Liang and Huanshui Zhang argue in Optimal Decentralized Control with Asymmetric Partial Information Sharing, a controller's actions in these settings directly influence the quality of future state estimations for other agents. If I move left, I change the environment you are trying to sense. Because my control law affects your estimation error, we can no longer optimize them as separate modules. We must design them jointly.
Finding Structure in Asymmetry
If the general problem is too complex to solve, researchers have found that specific patterns of information sharing—called partially nested structures—allow us to regain some mathematical tractability.
Naumaan Nayyar and Rahul Jain demonstrated in Optimal Decentralized Control with Asymmetric One-Step Delayed Information Sharing that if information flows such that any agent affected by another's action eventually learns the information that prompted that action, the optimal control laws remain linear. This allows for the derivation of Riccati-type equations—the gold standard for computable control solutions—even in the presence of asymmetric delays.
Dileep Kalathil and colleagues extended this in Optimal Decentralized Control With Asymmetric One-Step Delayed Information Sharing, showing that even lopsided information access (where one agent is more "informed" than another) can be managed if the underlying communication graph respects these nested dependencies.
From Individual Accuracy to Team Consistency
In traditional estimation, the goal is Minimum Mean-Squared Error (MMSE)—making each agent as accurate as possible. But Aditya Mahajan and Mohammad Afshari suggest a shift in perspective in Team Optimal Decentralized State Estimation.
They introduce Minimum Team Mean-Squared Error (MTMSE). The intuition is that in a team, it is often more important for agents to be consistent with one another than to be individually "correct" about the global state. If two robots are carrying a heavy object, they don't just need to know where the object is; they need to have estimates that are compatible with each other's planned movements. MTMSE filters explicitly account for the coordination weights of the control task, outperforming standard Kalman filters by prioritizing the team's collective goal over individual sensing accuracy.
Realizing the Architecture
How do we actually build these controllers? Takashi Tanaka provides a state-space "blueprint" in Optimal Output Feedback Architecture for Triangular LQG Problems. By solving coupled Riccati equations, Tanaka shows how to structure the flow of information between sub-controllers to maintain optimality in hierarchical (triangular) systems.
This integrated approach is already seeing application in critical infrastructure. For instance, Jianglin Lan and Xianxian Zhao applied these concepts to load frequency control in power systems. By designing observers and controllers simultaneously, they created a grid that recovers from disturbances faster because the local controllers "understand" how their actions propagate through the interconnected areas of the network.