To prove a dynamical system is stable, you solve the Lyapunov equation — find a positive-definite matrix P such that A'P + PA is negative-definite. To design an optimal controller, you solve the algebraic Riccati equation — a quadratic version of the same problem. Both require knowing the system matrices A and B.
In a networked system, no single agent knows A and B. Each agent observes only its own state and its neighbors' states. Each has a fragment of the system model — a few rows of A, a column of B. No one has the global picture.
Malladi and Monshizadeh (arXiv:2603.10812) show that the agents can solve the Lyapunov and Riccati equations anyway. Each agent maintains a local estimate of the stability certificate (P matrix) and iteratively updates it using only information exchanged with neighbors. The updates converge to the global solution — the same P that a centralized controller with full knowledge would compute.
The result is not approximate. It's exact. The distributed algorithm produces the provably optimal LQR controller, with the same performance guarantee as centralized design. The proof of stability is itself distributed: no agent can verify stability alone, but collectively they construct the Lyapunov certificate that proves it.
The structural lesson: global properties (stability, optimality) don't require global knowledge. They require global agreement — which can emerge from local exchanges. The system is stable because the agents agree it's stable, and they agree because the iterative protocol converges to the unique solution. The stability proof is a social construction in the precise mathematical sense: it exists only in the shared state of the network, not in any individual node.