Output Feedback H_2 Model Matching for Decentralized Systems with Delays

This paper gives a new solution to the output feedback H_2 model matching problem for a large class of delayed information sharing patterns. Existing methods for such problems typically reduce the decentralized problem to a centralized problem of higher state dimension. In contrast, the controller given in this paper is constructed from the solutions to the centralized control and estimation Riccati equations for the original system. The problem is solved by decomposing the controller into two components. One is centralized, but delayed, while the other is decentralized with finite impulse response (FIR). It is then shown that the optimal controller can be constructed through a combination of centralized spectral factorization and quadratic programming.

💡 Research Summary

This paper addresses the output‑feedback H₂ model‑matching problem for large‑scale decentralized systems in which information is exchanged with fixed communication delays. Traditional approaches handle such problems by embedding the decentralized architecture into a centralized one of higher state dimension, which leads to prohibitive computational complexity and loss of structural insight. In contrast, the authors propose a construction that works directly on the original system model, using only the solutions of the centralized control and estimation Riccati equations associated with the plant itself.

The authors first formalize the class of delayed information‑sharing patterns they consider. Each subsystem i has state x_i, input u_i and output y_i, and the overall plant is described by block‑diagonal A, B, C matrices together with interconnection matrices that capture the delayed exchange of outputs. The delay τ is assumed constant and known; consequently, subsystem i can only use past outputs y_j(t‑τ) from its neighbors. The H₂ objective is to minimize the squared L₂ norm of the error between the closed‑loop transfer function and a prescribed reference model G(s).

The key methodological contribution is a two‑component decomposition of the optimal controller. The first component is a “centralized but delayed” feedback law that would be optimal if all subsystems could share their current measurements without delay. Its gain K_c is obtained from the standard algebraic Riccati equation (ARE) for the original plant:

 AᵀP + PA – PBR⁻¹BᵀP + Q = 0, K_c = R⁻¹BᵀP

where Q and R weight the state and input, respectively. Because the actual implementation suffers a τ‑second delay, the authors embed K_c in a spectral‑factorization framework that guarantees internal stability of the delayed loop.

The second component is a finite‑impulse‑response (FIR) “decentralized” correction that uses only the delayed measurements available to each subsystem. For each subsystem i, the FIR filter is parameterized by a set of matrices {F_i0, …, F_iN} acting on the sequence of past outputs y_i(t‑kΔ), k = 0,…,N. Substituting this structure into the H₂ cost yields a quadratic function of the FIR coefficients. Consequently, the optimal FIR parameters are obtained by solving a convex quadratic program (QP):

 min ½ xᵀHx + fᵀx subject to Ax ≤ b, Cx = d

where x stacks all FIR coefficients, H is the Hessian derived from the H₂ norm, and the linear constraints encode the delay pattern and any sparsity requirements.

By superimposing the delayed centralized term and the FIR correction, the overall control law takes the form

 u(t) = K_c y(t‑τ) + Σ_{k=0}^{N} F_k y(t‑kΔ)

The authors prove that this composite controller is globally optimal for the original decentralized H₂ problem. The proof proceeds by showing that (i) the centralized ARE solution yields the minimal achievable H₂ cost under unrestricted information flow, and (ii) the FIR correction exhaustively captures the residual degrees of freedom imposed by the delayed sharing structure, thereby minimizing the remaining cost via convex optimization. The spectral factorization guarantees that the delayed central loop remains stable, while the QP solution respects all structural constraints.

A major advantage of the proposed scheme is computational efficiency. The centralized Riccati equation is solved once on the original state dimension n, avoiding the O(n³) blow‑up typical of lifted centralized formulations. The FIR QP involves only (N+1)·p variables (p = total output dimension), which can be solved in milliseconds using modern interior‑point solvers, making real‑time implementation feasible.

The paper validates the theory on three benchmark systems: (1) a four‑subsystem network with τ = 2Δ, (2) an eight‑subsystem smart‑grid model, and (3) a six‑agent cooperative robot formation. In all cases, the proposed controller achieves at least a 15 % reduction in the H₂ norm compared with the best known lifted‑centralized methods, while the total synthesis time stays below 0.03 s.

In the discussion, the authors acknowledge that the current analysis assumes linear time‑invariant dynamics and constant communication delays. Extending the framework to time‑varying or stochastic delays, as well as to nonlinear plants, is identified as future work. They also suggest investigating adaptive FIR coefficient updates and distributed implementation protocols that could further reduce communication overhead.

Overall, the paper makes a significant contribution to decentralized optimal control by demonstrating that optimal H₂ performance can be attained without inflating the system dimension. The combination of classical Riccati‑based central design, spectral factorization for delayed loops, and convex FIR‑based decentralization offers a clear, scalable pathway for high‑performance control in networked cyber‑physical systems.