In control theory, we rarely have the luxury of directly measuring every state variable of a system. Whether managing a robotic arm, a power converter, or an aircraft, we rely on observers—mathematical algorithms that reconstruct internal, unmeasurable states from noisy and limited sensor outputs.
To achieve rapid convergence, eliminate steady-state tracking errors, and robustly reject external disturbances, the intuitive mathematical solution is simple: turn up the observer gain. As noted by Lars Imsland et al. in Output feedback stabilization with nonlinear predictive control: asymptotic properties, traditional approaches to perfect tracking in complex nonlinear systems theoretically require high or even infinite gains to override model uncertainties. High-gain architectures allow estimators to rapidly reconstruct unmeasured states or isolate system faults, a capability leveraged by Farhad Ghanipoor et al. in Linear Fault Estimators for Nonlinear Systems: An Ultra-Local Model Design to track fast-moving fault signals without complex nonlinear modeling.
However, in the physical world, turning up the gain triggers a severe penalty: high-gain noise sensitivity. This creates a fundamental performance ceiling where further error reduction is prevented by the physical limitations of sensors and discrete-time processing.
Why High Gain Amplifies Noise
Why can't we simply set the observer gain arbitrarily high? The answer lies in how observers reconstruct hidden states. To estimate a hidden state (like velocity) from a measured output (like position), an observer fundamentally acts as a generalized differentiator.
Consider a sensor measurement $y(t) = x(t) + n(t)$, where $x(t)$ is the true state and $n(t)$ is high-frequency sensor noise, modeled as a small-amplitude, high-frequency sinusoid:
$$n(t) = A \sin(\omega t)$$
When an observer attempts to compute the derivative to estimate the underlying dynamics, the derivative of the noise component becomes:
$$\dot{n}(t) = A \omega \cos(\omega t)$$
Even if the noise amplitude $A$ is tiny, a massive frequency $\omega$ acts as a multiplier, blowing up the noise derivative. When a high-gain observer multiplies this already-amplified signal by a large gain parameter, the noise completely dominates the true state estimate.
As Xiaodi Li explains in Speed Control for a PMSM Servo System Using Model Reference Adaptive Control and an Extended State Observer, these derivative-like mechanisms are highly prone to amplifying high-frequency noise from feedback hardware like incremental encoders. This creates a strict performance ceiling. In time-varying optimization, Matteo Marchi et al. demonstrate in A Framework for Time-Varying Optimization via Derivative Estimation that while increasing estimator gain improves tracking speed for moving targets under ideal conditions, it simultaneously increases vulnerability to measurement jitter and sampling rates in practice.
The Consequences: Peaking, Jitter, and Saturation
In physical systems, this mathematical sensitivity manifests in three destructive ways:
1. The Peaking Phenomenon
When an observer is initialized with an incorrect estimate, high gains cause the estimation error to decay exponentially fast, but at the cost of a massive transient spike—or "peak"—in the initial control signal. Mona Buisson-Fenet et al. show in Joint state and dynamics estimation with high-gain observers and Gaussian process models that these peaking effects can severely degrade performance or outright destabilize the control loop.
2. Actuator Saturation
These initial spikes and continuous noise amplification force actuators (like motors or valves) to hit their physical limits. Lihua Xie et al. highlight in Practical Output Consensus of Nonlinear Heterogeneous Multi-Agent Systems with Limited Data Rate that tuning observer gains too high directly risks actuator saturation due to these initial state spikes and noise vulnerability.
3. Control Signal Jitter
In steady state, high-frequency noise amplification creates a continuous "chattering" or jitter in the control action. Krzysztof Łakomy et al. document this in Active Disturbance Rejection Control Design With Suppression of Sensor Noise Effects in Application to DC-DC Buck Power Converter, noting that high-gain observers used for rapid disturbance estimation introduce severe high-frequency jitter into the control signals of DC-DC buck converters.
The Ubiquity of the Trade-off
This conflict between tracking speed and noise rejection spans across diverse control paradigms:
- Adaptive Control: Masayoshi Tomizuka established in Basic adaptive control theory and its application to robot control-I. that high adaptation gains designed to minimize tracking errors in robotics inevitably amplify sensor noise from components like tachometers and encoders.
- Disturbance Compensation: When using Linear Extended State Observers (LESOs) to compensate for dead-time effects in power electronics, Jie Shi in Analysis and Compensation Control of Dead-Time Effect on Space Vector PWM notes that increasing observer gain to improve disturbance rejection inadvertently amplifies high-frequency measurement noise, requiring a highly sensitive tuning process.
- MIMO and Chaotic Systems: Zhenfeng Chen et al. demonstrate in Output Feedback Stabilization of a Class of MIMO Uncertain Non-affine Nonlinear Systems that the gain-noise trade-off defines the practical limits of stabilization for multi-input multi-output systems in high-interference environments. Similarly, Chuansheng Tang et al. find in Robust Guaranteed Cost Control of PMSM Chaotic System with Uncertain Parameters that high-gain feedback matrices optimized offline can cause extreme sensitivity and instability when deployed on physical hardware.
- Safety and Networks: Kendra Lesser in Safety verification of output feedback controllers for nonlinear systems points out that deterministic safety buffers can fail when high-gain observers amplify stochastic sensor noise. In distributed networks, Aritra Mitra et al. show in Distributed State Estimation over Time-Varying Graphs: Exploiting the Age-of-Information that high-gain observers suffer from transient spikes and error amplification when real-world communication or sensor noise is introduced. Finally, Yanjun Shen et al. in Nonfragile high‐gain observers for nonlinear systems with output uncertainty highlight that non-constant sensor bias or nonlinear interference causes state estimations to oscillate wildly under high gains.
Architectural Solutions: Breaking the Trade-off
To bypass this fundamental ceiling, modern research focuses on structural modifications that decouple rapid estimation from noise amplification.
Hybrid Filtering
Instead of forcing a single observer to handle both disturbance rejection and noise suppression, researchers combine complementary frameworks. For instance, Bao Zhang et al. in High-Precision Control of Aviation Photoelectric-Stabilized Platform Using Extended State Observer-Based Kalman Filter integrated an Extended State Observer (ESO) with a Kalman Filter. The Kalman Filter explicitly dampens sensor noise, allowing the ESO to operate at the high gains necessary for rapid disturbance rejection without saturating the controller.
Structural Decomposition
Rather than using a single monolithic high-gain observer, the total disturbance can be broken down into multi-level stages. Krzysztof Łakomy et al. designed a Cascade Extended State Observer (CESO) that distributes the estimation burden across a cascaded structure. This architecture achieves the tracking accuracy of high-gain systems while drastically reducing high-frequency jitter in the control signal.
Key Foundational Work
Output feedback stabilization with nonlinear predictive control: asymptotic properties (2003) — contributed 2 concept elements to this direction.
Basic adaptive control theory and its application to robot control-I. (1985) — contributed 1 concept elements to this direction.