State Observers for Linear Systems with Prescribed Residual Bounds

Technical Summary: State Observers for Linear Systems with Prescribed Residual Bounds

Problem Formulation
The paper addresses a specific limitation in classical continuous-time state observers for Linear Time-Invariant (LTI) systems subject to unknown bounded disturbances. While standard Luenberger observers guarantee asymptotic convergence and uniform boundedness of the estimation error under suitable assumptions, they do not explicitly regulate the transient behavior of the observer residual, defined as $r(t) = y(t) - C\hat{x}(t)$. In applications such as fault detection and safety supervision, large transient deviations in the residual can trigger false alarms or degrade performance. Existing robust designs provide bounds on the residual as a consequence of system dynamics and gain selection, but these bounds are not directly assignable by the designer. The problem is formulated as a constraint enforcement task: design an observer such that $\|r(t)\| \leq \delta$ for all $t \geq t_0$, where $\delta > 0$ is a pre-specified bound chosen by the designer, regardless of the disturbance magnitude (within known bounds).

Methodology
The proposed solution is a hybrid observer architecture that augments a standard continuous-time Luenberger observer with discrete state resets. The system dynamics are modeled as:
$$\dot{x} = Ax + Bu + d(t), \quad y = Cx$$
where $d(t)$ is an unknown bounded disturbance. The observer state $\hat{x}$ evolves continuously according to the standard Luenberger equation but undergoes instantaneous jumps when the residual norm reaches the prescribed threshold $\delta$.

The hybrid dynamics are defined by flow and jump sets:

• Flow Set ($C$): $\|r\| \leq \delta$. The observer evolves continuously.
• Jump Set ($D$): $\|r\| \geq \delta$. A reset is triggered.

At the jump instant $t_k$, the observer state is updated via a linear reset map:
$$\hat{x}^+ = \hat{x}^- + J r^-$$
where $J$ is a constant reset gain matrix. The paper derives a closed-form expression for $J$ based on the Lyapunov matrix $P$ (associated with the continuous error dynamics) and the contraction factor $\rho \in [0, 1)$:
$$J = (1 - \rho) P^{-1} C^\top (C P^{-1} C^\top)^{-1}$$
This specific choice of $J$ ensures that the residual contracts by a factor of $\rho$ at every jump ($\|r^+\| = \rho \|r^-\|$) while maintaining the non-expansiveness of the estimation error in the Lyapunov metric $V = \tilde{x}^\top P \tilde{x}$.

Key Contributions and Theoretical Guarantees
The paper establishes four main theoretical results under the assumptions of detectability, bounded disturbances, and full row rank of $C$:

1. Residual Contraction: At every reset instant, the residual norm is strictly reduced by the factor $\rho$, ensuring $\|r^+\| = \rho \|r^-\|$.
2. Forward Invariance: If the initial residual satisfies $\|r(t_0)\| \leq \delta$, the residual envelope $\{r : \|r\| \leq \delta\}$ is forward invariant. The residual never exceeds the prescribed bound $\delta$.
3. Non-Expansiveness of Estimation Error: The Lyapunov function $V(\tilde{x})$ does not increase at jump instants ($V^+ \leq V^-$). This ensures that the resetting mechanism does not destabilize the estimation error.
4. Uniform Boundedness: The hybrid observer preserves the uniform boundedness properties of the underlying continuous-time Luenberger observer. The estimation error remains bounded by the same expression derived for the standard observer, ensuring Input-to-State Stability (ISS) with respect to the disturbance.
5. Absence of Zeno Behavior: The paper proves a strictly positive minimum dwell time between consecutive resets, preventing infinite switching in finite time.

Simulation Results
The authors validate the design through simulations on a 4-state LTI system with a time-varying sinusoidal disturbance active on the interval $t \in [2, 5]$.

• Residual Performance: A standard Luenberger observer with identical gains violates the prescribed residual bound $\delta$ when the disturbance is active. In contrast, the proposed hybrid observer strictly enforces $\|r(t)\| \leq \delta$. Whenever the residual reaches $\delta$, a reset occurs, instantly contracting the residual to $\rho\delta$ (where $\rho=0.1$ in the simulation).
• Estimation Error: The state estimation errors for the hybrid observer exhibit jumps at reset instants but remain bounded and comparable to the continuous-time observer. The Lyapunov function $V(t)$ remains bounded, confirming the theoretical non-expansiveness property.
• Initialization: Simulations also demonstrate that if the system starts outside the residual envelope ($\|r(0)\| > \delta$), a single reset at $t=0$ (with $\rho=0$) can immediately bring the residual into the invariant set.

Significance and Claims
The paper claims to offer a novel mechanism for regulating transient residual behavior without relying on high-gain continuous feedback, which is prone to peaking phenomena and noise sensitivity. Unlike previous works that use resets to improve transient performance or reduce $L_2$ gains, this work treats the residual bound as a hard, invariant performance specification enforced by design.

The authors position this work as a solution to a "constraint enforcement problem" rather than a "robustness characterization problem." By guaranteeing that the residual stays within a designer-specified envelope, the approach provides a more structured characterization of nominal behavior. This is particularly relevant for fault detection systems where thresholds are typically tuned conservatively; the proposed observer mitigates disturbance-induced spikes that might otherwise trigger false alarms. The paper concludes by noting that while the framework is established for LTI systems, the hybrid approach is not restricted to linear dynamics, suggesting potential for future extensions to nonlinear and sampled-data systems.