High-Gain Observer Theory

• High-gain observer theory is a framework for designing observers that achieve exponential convergence of state and disturbance estimates in uncertain systems.
• It employs tunable gain parameters to balance fast error decay against noise amplification and peaking phenomena in both linear and nonlinear models.
• Extensions like cascade and block observer structures enhance robustness and practical implementation by limiting gain growth in complex multi-dimensional systems.

High-gain observer theory addresses the design and rigorous analysis of asymptotically convergent state and disturbance estimation schemes for both linear and nonlinear systems, with particular focus on their robustness, convergence rates, and implementation trade-offs. The theory provides explicit observer structures whose gains, via a tunable parameter, can be made arbitrarily large—leading to fast estimation error decay but introducing noise-sensitivity and “peaking” phenomena, thus necessitating careful synthesis of stability, robustness, and practical utility. High-gain observers (HGOs), extended state observers (ESOs), cascade observers, and their variants are foundational for disturbance rejection, output-feedback stabilization, and robust trajectory tracking in the presence of model uncertainty and exogenous inputs.

1. Theoretical Foundations and System Classes

High-gain observer schemes are built on the principle that for certain system representations, an observer with sufficiently large gains achieves exponential estimation error convergence. For SISO plants in observable canonical or companion form, or general $n^{\text{th}}$-order nonlinear systems with output $y=x$, the basic paradigm involves augmenting the system state to include uncertainties or disturbances as additional estimable variables. A canonical example is the extended observer for a nonlinear uncertain system

$$x^{(n)}(t) = f(x, \dot x, \dots, x^{(n-1)}, t) + b\,u(t), \qquad y(t) = x(t)$$

where all model uncertainties and external disturbances are lumped into $f(\cdot)$, which is assumed continuously differentiable and with bounded total derivative $g(\cdot) = \frac{d}{dt}f(\cdot)$ (Wang et al., 2023).

The extended state vector becomes $X = [x, \dot x, \dots, x^{(n-1)}, f(\cdot)]^T$, so that the observer must estimate both system states and disturbances. In the linear case, the so-called extended dynamics observer (EDO) and classical high-gain observer structures exploit observability canonical forms and disturbance representation in an augmented space, allowing simultaneous disturbance and state estimation (Feng et al., 2020).

When system nonlinearities are non-Lipschitz—e.g., Hölder continuous or merely continuous—convergence can be established with modified structures (cascade, homogeneous observers), provided certain exponents or growth bounds are met (Bernard et al., 2017). Cascade observer architectures and multi-block structures are key for minimizing gain exponent and for treatability of systems with lower regularity.

2. Observer Structure, Gain Selection, and Error Dynamics

The high-gain observer is defined via recursive injection of the measurement innovation through a series of differentiated coordinates, each scaled by increasing powers of $\varepsilon^{-1}$ or a gain parameter $L \gg 1$. For a general $(n+1)$-order observer targeting both state and lumped disturbance estimation: $$\begin{aligned} \dot{\hat x}_1 &= \hat x_2 + \frac{h_1}{\varepsilon}(y - \hat x_1) \ \dot{\hat x}_2 &= \hat x_3 + \frac{h_2}{\varepsilon^2}(y - \hat x_1) \ &\,\,\vdots \ \dot{\hat x}_n &= \hat x_{n+1} + \frac{h_n}{\varepsilon^n}(y - \hat x_1) + b\,u \ \dot{\hat x}_{n+1} &= \frac{h_{n+1}}{\varepsilon^{n+1}}(y - \hat x_1) \end{aligned}$$ where $h_i$ are selected so that the polynomial $s^{n+1} + h_1 s^n + \cdots + h_{n+1}$ is Hurwitz (Wang et al., 2023). The scalar $\varepsilon \in (0,1)$ (or $L^{-1}$ in alternate notation) is the high-gain parameter: as $\varepsilon \to 0$, convergence accelerates.

The estimation error $\tilde X$ satisfies singularly perturbed (fast-slow) dynamics: $$\dot{\tilde X} = \frac{1}{\varepsilon} A_{cl} \tilde X + E\Delta(t)$$ where $A_{cl}$ is the closed-loop observer matrix, and $\Delta(t)$ collects the bounded time derivatives of uncertainties (e.g., $\Delta=g(x,\dots, x_{n}, t)-y_a^{(n+1)}(t)$). Lyapunov analysis yields exponential decay of $\|\tilde X(t)\|$, with the ultimate bound $$\limsup_{t\to\infty} \|\tilde X(t)\| \leq K_2 \varepsilon M$$; thus, the steady-state error is proportional to both the disturbance bound and the high-gain parameter.

When an explicit disturbance model is available (e.g., a known finite-dimensional subspace), the observer can incorporate exact cancellation for those components, as in the EDO, while unknown modes are suppressed by the high-gain mechanism. This generalized framework yields a performance envelope interpolating between perfect model-based estimation and pure high-gain state disturbance rejection (Feng et al., 2020).

3. Extensions, Cascade, and Block Structures

Standard high-gain observation relies on the full gain parameter rising rapidly with system order (up to $\epsilon^{-n}$ in the $n$-state system). Novel observer structures, such as the 2D-block cascade observer, limit gain growth to quadratic order by augmenting the observer dimension (from $n$ to $2n-2$) and coupling blocks through hierarchical error injection. This design paradigm ensures the convergence rate remains tunable via the scalar $\ell$, but the maximum observer gain grows only like $\ell^2$, improving both numerical conditioning and noise robustness for large-scale systems (Astolfi et al., 2015).

For non-Lipschitz or merely continuous nonlinearities, cascaded high-gain observers (with growing block size or number of blocks) or homogeneous observer constructions are introduced. Homogeneous high-gain structures employ nonlinear error injection (fractional or sign-power law) and Lyapunov functions in the bi-limit, allowing finite-time or asymptotic convergence under Hölder regularity, with much weaker smoothness requirements than the classical linear case (Bernard et al., 2017).

Cascade extended state observers (cascade-ESOs) further decompose the overall observer into a sequence of sub-observers, each handling a distinct bandwidth regime. This enables the overall system to achieve arbitrarily small disturbance estimation error while controlling noise amplification. With $p$ cascade stages, the effective disturbance gain decays as $\prod_{i=1}^p 1/\omega_{oi}$ (where $\omega_{oi}$ is the gain at the $i$th stage), while noise amplification is determined by the output-stage (lowest bandwidth) only (Łakomy et al., 2020).

4. Stability Analysis and Robustness Guarantees

High-gain observer stability analysis is primarily based on singular perturbations, Lyapunov functions (typically quadratic or homogeneous), and input-to-state stability (ISS) concepts. For linear and nonlinear systems in canonical form, the error coordinates are scaled by the gain parameter so that the fast subsystem dominates. This setting yields:

• Exponential convergence rates $\sim \exp(-\alpha t/\varepsilon)$ for any initial error.
• Ultimate bounds linear in $\varepsilon$ and the size of lumped disturbances or their derivatives (e.g., $$\limsup_{t\to\infty} \|e(t)\| \leq K\varepsilon\sup |g(x,\dots, x_n, t)|$$) (Wang et al., 2023).
• In the presence of bounded measurement noise, the error steady-state is governed by a trade-off $$\sim (\text{noise})/\epsilon^{n-1} + \epsilon (\text{uncertainty})$$ (Esfandiari et al., 2019).
• In multi-observer or bank-of-observers architectures, convex adaptation and recursive least-squares can improve both transient convergence and noise immunity, allowing explicit tuning of the steady-state noise tube radius via additional adaptation parameters (Esfandiari et al., 2019).

In observer designs incorporating disturbance models, exact cancellation is possible for known exogenous signals, and the steady-state error is determined only by components residing outside the modeled subspace. Practical implementation constrains high-gain selection due to noise, actuator bandwidth, and peaking effects, necessitating careful gain selection and sometimes saturation or filtering mechanisms to avoid instability or excessive noise amplification (Boss et al., 2021, Al-Nadawi et al., 2020).

5. Practical Applications and Implementation Guidelines

High-gain observer theory underpins output-feedback stabilization, robust output tracking, and active disturbance rejection in diverse domains, such as multi-rotor UAV trajectory tracking (Boss et al., 2021), vehicle path-following under modeling uncertainty and road perturbations (Al-Nadawi et al., 2020), or general SISO/MIMO disturbance rejection.

Extended high-gain observers (EHGO) allow simultaneous estimation of states, unmeasured disturbance terms, and even derivatives of reference trajectories; this supports robust feedback linearizing control strategies that guarantee tracking error (and its derivatives) convergence under exogenous disturbances and parameter variation. The observer gains are typically chosen via pole-placement for the companion polynomial, with the high-gain parameter adjusted for the desired bandwidth.

Cascade and multi-stage observers are indicated for situations where measurement noise is significant and fast convergence is required without overwhelming the system with noise-induced error (Łakomy et al., 2020). Multi-high-gain observer (MHGO) banks and convex-combination techniques with adaptation can break the classical trade-off between high transient speed and noise resilience (Esfandiari et al., 2019).

In all high-gain observer applications, explicit trade-offs must be balanced:

• Lowering the high-gain parameter accelerates convergence but amplifies noise and introduces peaking.
• Increasing observer dimension (as in 2D-block cascades) reduces the maximal required gain at the cost of increased computational complexity.
• Augmenting with adaptive, model-based, or homogeneous/hybrid corrections enables robustness to lower-regularity nonlinearities or more complex uncertainty structures (Bernard et al., 2017, Astolfi et al., 2015).

6. Limitations and Advanced Developments

High-gain observer theory is not universally applicable, and several limitations require advanced solutions:

• Measurement noise places a hard bound on how high observer gains can be tuned, which otherwise is unbounded in the limit $\varepsilon \to 0$ (Łakomy et al., 2020).
• Non-Lipschitz systems (e.g., with only Hölder continuous nonlinearities) may experience only ultimate boundedness to an arbitrarily small residue, not true asymptotic convergence, unless addressed with homogeneous observer logic (Bernard et al., 2017).
• For high-dimensional systems, classical high-gain designs suffer from rapidly growing gain exponents, leading to poor numerical conditioning and implementability, addressed by block or limited gain-power observer designs (Astolfi et al., 2015).
• Integration with actuator and sensor dynamics, and the requirement for output-feedback only (i.e., no unmeasured state derivatives), further complicates observer design, requiring two-time-scale singular perturbation analysis and tailored observer-controller architectures (Boss et al., 2021).

Further extensions of the theory incorporate exact-differentiator gains (homogeneous observers), multi-output generalizations, adaptive noise attenuation strategies, and explicit Lyapunov function constructions for non-Lipschitz and switched/hybrid systems.

---

Key references:

• High-gain extended observer for nonlinear uncertain systems (Wang et al., 2023)
• Extended dynamics observer for linear disturbances (Feng et al., 2020)
• Cascade extended state observer and noise trade-offs (Łakomy et al., 2020)
• Homogeneous/cascade observers for non-Lipschitz systems (Bernard et al., 2017)
• Output-feedback and MHGO for noise resilience (Esfandiari et al., 2019)
• Observer design with limited gain power (Astolfi et al., 2015)
• Applications to UAVs (Boss et al., 2021), vehicle control (Al-Nadawi et al., 2020), and trajectory tracking under disturbances (Izadi et al., 2023).