⚡ electrical engineering

Event-Triggered Newton Extremum Seeking for Multivariable Optimization

This paper proposes a static event-triggered Newton-based extremum seeking strategy for multivariable optimization that utilizes a Riccati-based Hessian inverse estimator to achieve assignable exponential convergence rates while significantly reducing control update frequency compared to traditional gradient-based and continuously actuated methods.

Original authors: Victor Hugo Pereira Rodrigues, Tiago Roux Oliveira, Miroslav Krstic, Paulo Tabuada

Published 2026-01-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Victor Hugo Pereira Rodrigues, Tiago Roux Oliveira, Miroslav Krstic, Paulo Tabuada

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find the highest peak in a vast, foggy mountain range, but you can't see the landscape. You only have a barometer that tells you your current altitude. This is the essence of Extremum Seeking: a method for finding the best possible setting (the "extremum") for a system when you don't know the map of the terrain.

Here is how the paper explains the new method they developed, broken down into simple concepts:

1. The Old Way: Feeling Your Way Up (Gradient-Based)

Traditionally, to find the peak, you take a small step in a random direction. If the barometer says you went higher, you keep going that way. If you went lower, you turn around.

The Problem: This is like walking up a very long, narrow valley. If the valley is steep on the sides but flat at the bottom, you end up zig-zagging wildly. You might take thousands of tiny, inefficient steps to reach the top because you don't know how "curved" the mountain is.
The Update Issue: In the digital world, a computer has to check your position and tell the system to move constantly. Doing this every single second wastes a lot of energy and computer power, even when you are just standing still or moving slowly.

2. The New Method: The "Smart" Hiker (Newton-Based)

The authors propose a smarter way to climb. Instead of just feeling the slope (gradient), they use a special tool to estimate the curvature of the mountain (the Hessian).

The Analogy: Imagine you are blindfolded, but you have a magical compass that not only tells you which way is up, but also tells you if the ground is flat, steep, or curved like a bowl.
The Benefit: With this "curvature" information, the hiker can take long, direct strides straight to the peak, rather than zig-zagging. The paper shows that this method allows the user to decide exactly how fast they want to reach the top, regardless of how weird the mountain's shape is.

3. The "Event-Triggered" Trick: Only Moving When Necessary

Even with the smart compass, a traditional computer would still shout out new instructions every millisecond. The authors added a "traffic light" system called Event-Triggered Control.

The Metaphor: Imagine a guard at a gate. In the old system, the guard checks the hiker's position every second and yells, "Move! Move! Move!" even if the hiker hasn't moved an inch.
The New System: The guard only checks if the hiker has drifted far enough from the planned path. If the hiker is doing fine, the guard stays silent. The guard only shouts a new instruction when the error gets too big.
The Result: This saves a massive amount of "shouting" (control updates). In the paper's simulation, the new method needed only 16 updates to find the peak, while the old method needed 68 updates.

4. How It Works Together

The paper combines these two ideas:

Newton's Method: Uses a dynamic "Riccati filter" (a mathematical engine) to guess the shape of the mountain in real-time, allowing for fast, direct movement.
Event-Triggering: Only sends a new command to the system when the error between where the system should be and where it is becomes too large.

5. The Proof

The authors didn't just guess this would work; they used advanced math (called "averaging theory") to prove that:

The system will always find the peak (stability).
It won't get stuck in an infinite loop of checking itself (avoiding "Zeno behavior").
It converges (reaches the goal) much faster than the old method.

Summary

Think of this paper as introducing a smart, energy-saving GPS for optimization.

Old GPS: Tells you to turn left, then right, then left again every second, regardless of traffic. It gets you there, but it's exhausting and slow on winding roads.
New GPS: Knows the shape of the road ahead. It tells you to drive straight for a long time, only giving you a new instruction if you drift off course. It gets you to the destination faster and uses much less battery power.

The paper demonstrates this with computer simulations, showing that this new "Smart GPS" reaches the optimal point faster and requires far fewer "instructions" than the traditional method.

Technical Summary: Event-Triggered Newton Extremum Seeking for Multivariable Optimization

Problem Statement
Extremum seeking (ES) is a model-free optimization technique used to find the extremum of an unknown static map by employing periodic probing signals. While gradient-based ES is widely used, it suffers from a fundamental limitation: its convergence rate is inherently tied to the unknown Hessian matrix (second derivative) of the cost function. In multivariable settings, if the Hessian is ill-conditioned (e.g., highly elongated level sets), gradient-based methods exhibit slow convergence along specific directions. Furthermore, traditional continuous-time implementations require constant control updates, which can be inefficient in networked systems where bandwidth is limited and communication congestion is a concern. While event-triggered control (ETC) has been applied to scalar and multivariable gradient-based ES to reduce update frequency, these methods still retain the dependency on the unknown Hessian for convergence speed.

Methodology
The paper proposes a static event-triggered control strategy for multivariable Newton-based extremum seeking. The methodology integrates three core components:

Newton-Based Framework: Unlike gradient-based approaches, the proposed method estimates the inverse of the Hessian matrix ( $H^{*-1}$ ) dynamically. This is achieved using a Riccati differential equation (Eq. 39) driven by a positive frequency $\omega_r$ . The Riccati filter estimates the inverse Hessian, allowing the control law to decouple the convergence rate from the unknown curvature of the cost function. The control input is formulated as $u(t) = -K\Gamma(t)\hat{G}(t)$ , where $\Gamma(t)$ is the inverse Hessian estimate and $\hat{G}(t)$ is the gradient estimate.
Event-Triggered Mechanism: To reduce actuation updates, the continuous control law is emulated using a Zero-Order Hold (ZOH). Updates occur only at discrete time instants $t_\kappa$ determined by a static triggering condition. The condition monitors the deviation between the current product of the inverse Hessian estimate and the gradient estimate, $\Gamma(t)\hat{G}(t)$ , and its last sampled value. An update is triggered when:
$\sigma \|\Gamma(t)\hat{G}(t)\| - \alpha \|e(t)\| < 0$
where $e(t)$ is the error vector, and $\sigma, \alpha$ are design parameters.
Stability Analysis: The authors employ averaging theory for discontinuous systems to analyze the closed-loop stability. By rescaling time and defining an average system, they establish that the equilibrium is locally exponentially stable. The analysis utilizes Lyapunov functions to prove that the system converges to a neighborhood of the extremum point with user-assignable rates, independent of the unknown Hessian.

Key Contributions

Decoupling Convergence from Curvature: The primary contribution is the integration of a dynamic Riccati-based estimator for the inverse Hessian into an event-triggered framework. This allows the designer to assign convergence rates via the gain matrix $K$ and excitation frequency $\omega_r$ , overcoming the slow convergence issues associated with ill-conditioned Hessians in gradient-based methods.
Reduced Actuation Frequency: The static event-triggered mechanism significantly reduces the number of control updates required to reach the extremum compared to both continuous-time implementations and event-triggered gradient-based schemes.
Theoretical Guarantees: The paper provides rigorous local exponential stability results and proves the avoidance of Zeno behavior (infinite updates in finite time) for the proposed multivariable Newton-based system.
Handling Multivariable Maps: The approach extends previous scalar event-triggered ES works to the multivariable case, addressing the complexities of matrix-valued Hessian estimation and error dynamics.

Results
Numerical simulations were conducted on a two-dimensional static map with a known but unknown-to-the-controller positive definite Hessian. The results compare the proposed Event-Triggered Newton-based ES (ET-Newton-ES) against an Event-Triggered Gradient-based ES (ET-Gradient-ES):

Convergence Speed: The Newton-based strategy demonstrated faster and more direct convergence to the optimal point $\theta^*$ compared to the gradient-based approach.
Control Updates: The ET-Newton-ES required only 16 control updates to converge, whereas the ET-Gradient-ES required 68 updates for the same scenario.
Performance: Both schemes successfully drove the output to the optimal value, but the Newton-based method achieved this with sharper control actions and significantly reduced actuation frequency.
Estimator Behavior: The dynamic inverse Hessian estimate $\Gamma(t)$ converged to the true inverse Hessian after a brief transient, validating the Riccati filter's effectiveness.

Significance and Claims
The paper claims that the proposed strategy offers a resource-efficient solution for real-time optimization in networked or embedded systems. By combining the curvature-compensation capabilities of Newton methods with the communication efficiency of event-triggered control, the approach achieves:

Faster Convergence: Independent of the unknown Hessian's condition number.
Efficiency: A substantial reduction in control update frequency, leading to lower computational and communication burdens.
Robustness: Theoretical guarantees of local exponential stability and practical convergence to the extremum neighborhood.

The authors position this work as an advancement in extremum seeking that bridges second-order optimization techniques with sparse, feedback-dependent control policies, making it suitable for scenarios where precise models are unavailable and communication resources are constrained. Future work is suggested to explore dynamic event-triggering conditions, robustness to noise, and extensions to time-varying or distributed optimization problems.