Dynamic Output-Feedback Law

• Dynamic output-feedback law is a control paradigm that uses internal state dynamics based on measured outputs to achieve stabilization and optimal performance.
• It leverages observer-controller formulations and Riccati-Lyapunov frameworks to address control challenges in linear, nonlinear, and distributed systems.
• Recent developments integrate data-driven and learning-based techniques to synthesize robust controllers under uncertainties and complex dynamics.

A dynamic output-feedback law is a controller that, through the use of its own dynamic internal state, computes control actions as functions of real-time measured outputs of a system, without explicit access to the full plant state. This paradigm is fundamental in modern control theory, covering both linear and nonlinear, finite- and infinite-dimensional, deterministic and stochastic, and centralized or distributed systems. Dynamic output-feedback allows for stabilization, optimality, tracking, decoupling, dissipativity, and other closed-loop properties under varying information and actuation constraints.

1. Core Mathematical Structure

A dynamic output-feedback controller is characterized by a state-space realization: $$\begin{aligned} x_c(k+1) &= A_c\,x_c(k) + B_c\,y(k) \ u(k) &= C_c\,x_c(k) + D_c\,y(k) \end{aligned}$$ where $x_c$ is the controller’s internal state, $y$ the measured output, and $u$ the control input applied to the plant. In continuous-time, the standard realization is similar, and the plant is typically described by: $$\dot x = A x + B u, \qquad y = C x, \qquad x(0) = x_0$$

The design goal is to select $(A_c,B_c,C_c,D_c)$ such that the interconnected closed-loop system achieves prescribed objectives, e.g., stabilization, regulation, tracking, performance optimization, or robust stability in the presence of disturbances or uncertainties.

2. Optimal Dynamic Output Feedback: State-Space and Structural Results

The most widely studied setting is Linear Quadratic Regulation (LQR) under dynamic output feedback, where the goal is to minimize an infinite-horizon cost: $$J = \mathbb{E}\left[\sum_{t=0}^\infty \big(x_t^\top Q x_t + u_t^\top R u_t\big)\right]$$ for given $Q \succeq 0$, $R \succ 0$, subject to $x_{t+1} = A x_t + B u_t$, $y_t = C x_t$.

Recent work rigorously characterizes the dynamic output-feedback solution as follows (Duan et al., 2022, Duan et al., 2022):

• Any stabilizing dynamic output-feedback law can be expressed (up to similarity) in "observer–controller" (separation) form:

$$\begin{aligned} \xi_{t+1} &= (A - B K^* - L^* C)\, \xi_t + L^* y_t \ u_t &= -K^* \xi_t \end{aligned}$$

where $K^*$ and $L^*$ are uniquely determined by, respectively, the state-feedback and observer Riccati equations.

• The corresponding closed-loop cost is uniquely minimized at this controller (up to a coordinate transformation of the controller's state).
• The optimization landscape of the dynamic output-feedback cost functional has a unique observable stationary point, simplifying global convergence analyses for policy-gradient methods.

Key to this result is the interaction between the controller dynamics and plant output, the structure of the coupled Lyapunov and Riccati equations, and the explicit similarity transformation invariance of dynamic controller realizations (Duan et al., 2022). These findings carry over to stochastic settings (LQG) under additional statistical structural assumptions.

3. Output Feedback in Broader Problem Classes

Dynamic output feedback is central in a wide range of advanced control problems:

• Distributed optimal control: For multi-agent or networked systems, dynamic output-feedback controllers are synthesized by solving coupled Riccati equations and constructing distributed observer-controller structures, ensuring optimality under partially nested information patterns (Gattami et al., 2012).
• Robust and dissipative control: Methods for robust stabilization and dissipativity certification employ dynamic output-feedback laws constrained via Linear Matrix Inequalities (LMIs), both in model-based and data-driven contexts (Kristović et al., 9 Jul 2025, Strong et al., 2022).
• Output-feedback for nonlinear and noninvertible systems: In nonlinear or non-minimum phase cases, extended observer architectures, high-gain observer designs, and feedback-linearizing output-feedback laws are constructed to achieve semiglobal or global stabilization (Peralez et al., 2016, Goyal et al., 2021).
• Positive systems and monotonic tracking: Output-feedback controllers that enforce external positivity or monotonicity are derived using behavioral system representations, with explicit LMIs ensuring closed-loop positivity and tracking (Makdah et al., 2023).
• Disturbance decoupling and geometric control: For biproper systems (with feedthrough), geometric methods yield necessary and sufficient solvability conditions, maximizing assignable closed-loop poles via dynamic output-feedback and analyzing the fixed spectrum induced by the plant's subspace structure (Padula et al., 2019).

4. Data-driven and Learning-based Dynamic Output Feedback

Recent developments enable the synthesis of dynamic output-feedback laws directly from data, bypassing explicit model identification (Lin et al., 23 Sep 2025, Xie et al., 8 Mar 2025):

• Model-free observer construction: "Virtual observer" or "information-state" filters are constructed from input–output trajectory data, with state-estimation dynamics embedded into the controller's parameterization. This sidesteps conventional requirements on observer convergence.
• Data-based policy iteration/ADP: By solving value iteration (VI) or policy iteration (PI) recursions using recorded data, optimal dynamic output-feedback gains can be recovered, with stability and convergence guaranteed under generic (mild) rank conditions.
• Robust, data-driven dissipativity: LMIs are formulated to synthesize a dynamic output-feedback law guaranteed to render all models consistent with noisy data strictly dissipative, employing matrix S-lemma machinery for necessary and sufficient conditions (Kristović et al., 9 Jul 2025).

The encoding of all state and observer error dynamics into the controller’s extended state, and the exploitation of linear parameterizations, is central to these data-driven approaches (Xie et al., 8 Mar 2025).

5. Advanced Structures: Event-Triggered, Infinite-Dimensional, and Nonlinear Laws

Dynamic output-feedback generalizes classically to:

• Event-triggered and high-gain scaling: Event-based dynamic output-feedback laws employ internal high-gain and adaptation variables to modulate sampling and observer gains dynamically, optimizing resource use while retaining closed-loop stability (Peralez et al., 2016).
• Infinite-dimensional and PDE control: In parabolic PDEs (e.g., heat equations with memory), output-feedback is achieved via finite-dimensional (sampled or distributed) observer-controller realizations, discretized with IMEX schemes and leveraging system-theoretic projections for stability (Khan et al., 28 Apr 2025, Liu et al., 2018).
• Nonlinear and information-state-based feedback: For nonlinear systems, information-state dynamic output-feedback employs sliding-window or ARMA models, with feedback synthesized via trajectory optimization methods such as iterative LQR on the lifted "information state" (Goyal et al., 2021).

6. Theoretical Guarantees and Design/Implementation Considerations

Major analytical and design properties for dynamic output-feedback controllers include:

• Separation principles: In LTI/LQG settings, output-feedback controllers decouple estimation and control, with the optimal dynamic law consisting of a Kalman filter and state-feedback controller (Duan et al., 2022, Duan et al., 2022).
• Robustness and dissipativity: For uncertain or poorly modelled plants, dissipativity-enforcing output-feedback designs guarantee closed-loop stability for all plants in a given uncertainty set (Strong et al., 2022, Kristović et al., 9 Jul 2025).
• Certifiable stability: Data-driven and learning-based algorithms use Lyapunov-residual criteria and switched iteration schemes to verify and guarantee Schur/contractive closed-loop behavior during online updates (Xie et al., 8 Mar 2025).
• Geometric assignability and algebraic constraints: In disturbance decoupling and fixed-pole assignment, the assignable spectrum and well-posedness are characterized by self-bounded/self-hidden subspace pairs, leading to explicit dynamic compensator parameterizations (Padula et al., 2019).

The realization and order of the dynamic compensator, structural properties (such as observer detectability), and computational tractability (for LMI/SDP or DP-based synthesis) must be judiciously addressed for each application domain.

7. Applications, Impact, and Ongoing Directions

Dynamic output-feedback is fundamental to modern and emerging fields:

• Large-scale and distributed systems: System Level Synthesis (SLS) leverages dynamic programming for scalable output-feedback design, enabling internally stabilizing controllers for high-dimensional systems (Conger et al., 2021).
• Reinforcement learning for partially observed systems: The structure of the output-feedback optimization landscape clarifies the convergence and optimality of policy-gradient algorithms in RL under partial observability, permitting direct search over full-order dynamic controllers (Duan et al., 2022, Duan et al., 2022).
• Biological, epidemiological, and network systems: Observer-based output-feedback is effectively used for disease spread mitigation (age-structured SIRD models), ensuring practical constraints and robust performance (Sonveaux et al., 2024).
• Nonlinear and hybrid systems: Closed-loop identification, event-triggered sampling, and power series (Chen–Fliess) approaches extend the dynamic output-feedback paradigm to nonlinear and hybrid system classes (Venkatesh, 2022, Peralez et al., 2016, Goyal et al., 2021).

Dynamic output-feedback remains an active area of research, increasingly leveraging data-driven, learning-based, and distributed paradigms to achieve robust, optimal, and scalable closed-loop control in complex systems.