⚡ electrical engineering

Approximate Feedback Linearization for a Nonlinear Hyperbolic PDE Class -- Part I: Volterra Truncation

This paper demonstrates that low-order truncations of infinite Volterra series feedback controllers, while no longer providing exact linearization, effectively stabilize nonlinear hyperbolic PDEs by leveraging the diminishing influence of higher-order terms near the origin to achieve finite-time practical and asymptotic stability within a region of attraction that expands with the truncation order.

Original authors: Miroslav Krstic

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Miroslav Krstic

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 steer a very complex, wild river (the nonlinear PDE) to flow perfectly straight and calm. In the world of math, there is a "perfect map" (an exact feedback law) that tells you exactly how to steer the river at every single point to make it behave like a simple, predictable stream.

However, this perfect map is like a library containing an infinite number of books. To use it, you would need to read and calculate an infinite number of pages in real-time. A computer cannot do that; it would crash trying to process the infinite details.

The Problem:
The author, Miroslav Krstic, previously found this "infinite library" solution. But since we can't use an infinite solution in the real world, engineers usually have to throw away the "tail" of the library—using only the first few books (a finite truncation).

The big fear was: If we throw away the infinite books, will the river go out of control? Will the simplified map fail to keep the water calm?

The Solution (The "Good Enough" Map):
This paper proves that you don't need the whole infinite library to keep the river safe. You can use a very short summary (a low-order truncation) and it will still work, provided the river isn't starting out too wild.

Here is the core idea broken down with simple analogies:

1. The "Infinite Library" vs. The "Short Summary"

Think of the perfect controller as a recipe that requires an infinite number of ingredients to be perfect.

The Exact Method: You try to cook the dish using every single ingredient from the infinite list. The result is perfect, but you can't actually cook it because the list is endless.
The Truncated Method: You decide to use only the first 3 or 5 ingredients. The paper shows that even with just these few, the dish still tastes "safe" and won't poison you (the system remains stable).

2. The "Small Start" Rule

The paper introduces a crucial condition: The river must start out relatively calm.
If the river is already a massive, chaotic flood (a large initial disturbance), a short summary of the map might not be enough to tame it. But if the river starts as a gentle stream (a small initial condition), the short summary works perfectly.

The Insight: The "missing" ingredients (the infinite tail we threw away) only matter when the river is huge. When the river is small, those missing ingredients are so tiny they don't matter. As the river gets smaller and calmer, the missing parts become even more negligible.

3. The "Safety Net" (Practical Stability)

The paper proves two things about this "short summary" controller:

It keeps you in the safe zone: No matter what, the water level will never rise above a certain safe height. It might not go all the way down to zero immediately, but it won't flood the city.
It eventually settles down: Over time, the water will calm down and approach a perfectly still state. The "missing" infinite parts create a tiny, tiny error, but that error shrinks rapidly as you add more ingredients to your summary (increase the truncation order).

4. The "Math Bridge" (How they proved it)

Usually, to prove these things, mathematicians need to know the exact value of every single point in the river at every moment. But for this specific type of complex river, those exact point-by-point values are impossible to calculate.

The Trick: The author built a "bridge" using a mathematical tool called the Cauchy-Schwarz inequality. Instead of looking at every single drop of water individually, they looked at the "total energy" of the water in chunks. This allowed them to prove the system is safe without needing to know every single drop's exact location.

5. The "Two Lenses" (Sup-norm vs. L2)

The paper looks at the river through two different lenses:

The "Sup-norm" lens: This looks at the highest wave in the river. The paper proves that even the highest wave will stay under control. This is the "sharper" and more practical proof.
The "L2" lens: This looks at the total energy of the water. The paper also proves the system is stable here, but the "highest wave" proof is considered more precise for this specific problem.

The Bottom Line

You don't need a perfect, infinite computer brain to control this complex system. You can use a simplified, finite version of the controller.

If the system starts small: The simplified controller works great.
The more terms you add: The "error" from leaving out the infinite tail gets smaller and smaller, like a geometric progression.
The result: You get a controller that a real computer can actually run, which keeps the system stable and safe, effectively capturing the "essence" of the perfect infinite solution without the impossible math.

In short: You can approximate the perfect solution with a simple one, and as long as you don't start with a disaster, the system will stay safe and eventually calm down.

Technical Summary: Approximate Feedback Linearization for a Nonlinear Hyperbolic PDE Class — Part I: Volterra Truncation

Problem Formulation
This paper addresses the boundary control of a class of one-dimensional nonlinear hyperbolic partial differential equations (PDEs) featuring spatial Volterra nonlinearities. The plant is defined by a transport equation $u_t = u_x + F[u]$ , where the nonlinearity $F[u]$ is an infinite Volterra series in the state $u$ . Previous work (Krstic, 2026, referred to as [6]) established that exact feedback linearization for this system is achievable via an infinite Volterra series backstepping transformation. This exact controller maps the nonlinear plant to a stable target transport equation. However, the exact controller requires the evaluation of an infinite cascade of nested integrals and the solution of transport PDEs on simplices of unbounded dimension, rendering it impossible to implement on digital hardware.

The core problem is to determine whether a finite-order truncation of this exact infinite Volterra controller can retain the stabilizing properties of the exact law. Specifically, the paper investigates if the "residual" error introduced by discarding higher-order terms (the Volterra tail) can be managed to ensure closed-loop stability, and how the region of attraction scales with the truncation order.

Methodology
The paper employs a quantitative analysis in the sup-norm ( $L^\infty$ ) to establish closed-loop stability for the truncated feedback law. The methodology diverges from the $L^2$ -based analysis of the exact controller in [6] due to the lack of pointwise bounds on the Volterra kernels.

Truncation and Residual Analysis: The exact controller $K[u]$ is replaced by a finite sum $K_N[u]$ of order $N$ . This truncation leaves a non-zero boundary residual $b(t) = K_N[u] - K[u]$ at the target boundary. The paper characterizes this residual as a higher-order term that decays geometrically with the truncation order $N$ within a specific state ball.
Kernel Bounds via Cauchy–Schwarz: A primary technical challenge is that the available kernel bounds from [6] are integral bounds on growing simplices, while the truncation residual is a pointwise quantity. The authors bridge this gap by applying the Cauchy–Schwarz inequality on the simplices. This converts the $L^2$ kernel bounds into $L^1$ bounds, which are then used to derive sup-norm bounds for the operator $K$ and its truncation tail without requiring pointwise kernel estimates.
Fixed-Point Well-Posedness: Unlike the exact controller, which maps the plant to a trivially well-posed target, the truncated controller results in a genuinely coupled fixed-point problem due to the boundary residual. The paper establishes local well-posedness and forward invariance using a contraction mapping argument on a closed ball in $L^\infty([0, \tau]; L^\infty(0, 1))$ .
Stability Proofs: The stability analysis proceeds in two stages:
- Practical Stability: The system is shown to be practically stable, with the state converging to a residual ball whose size depends on the truncation order.
- Asymptotic Stability: By exploiting the fact that the residual is a higher-order term (vanishing as the state approaches zero), the authors prove local asymptotic stability characterized by a class- $\mathcal{KL}$ estimate. The decay is shown to be superexponential near the origin.

Key Contributions

Quantitative Stability of Truncated Controllers: The paper proves that truncating the exact infinite Volterra linearizer to any finite order $N \ge 2$ preserves closed-loop stability, provided the initial condition is sufficiently small.
Sup-Norm Analysis: It establishes the first $L^\infty$ stability results for this class of nonlinear hyperbolic PDEs under Volterra backstepping. This is significant because $L^\infty$ bounds are often more relevant for physical constraints than $L^2$ bounds.
Region of Attraction Characterization: The paper provides explicit estimates for the region of attraction. It demonstrates that the size of this region grows with the truncation order $N$ and the analyticity radius of the plant nonlinearity, shrinking as the nonlinearity's growth rate increases.
Resolution of Well-Posedness: The analysis resolves the well-posedness of the nonlinear closed-loop system under truncation, showing that the coupling introduced by the residual does not prevent the existence of a unique mild solution.
$L^2$ Alternative: A secondary contribution is a proof of stability in the $L^2$ norm. The authors show that while $L^2$ stability holds, the sup-norm result is mathematically sharper and practically more relevant, as the $L^2$ analysis requires more conservative constants due to partial symmetrization of the Lipschitz estimates.

Results
The main result, Theorem 1, establishes four properties for the closed-loop system under the truncated feedback $U(t) = K_N[u](1, t)$ :

Forward Invariance: The state remains within a fixed sup-norm ball for all time.
Practical Stability: The state is bounded by a function of the initial condition and a residual term proportional to the truncation error $\varepsilon_N(r_0)$ .
Finite-Time Attractivity: After a transport crossing time ( $t=1$ ), the state enters and remains within a residual ball determined by the truncation order.
Asymptotic Stability: The system converges to the origin with a class- $\mathcal{KL}$ estimate. The residual term decays geometrically with $N$ (as $(C_K r_0)^N$ ), meaning the exact stabilization behavior is recovered as $N \to \infty$ .

The paper also presents Theorem 2, which mirrors these results in the $L^2$ norm, noting that the $L^2$ constants are more conservative (e.g., the Lipschitz constants involve a factor of $\sqrt{2}$ ) and the residual bound is additive rather than a maximum.

Significance and Claims
The paper claims that surprisingly simple finite-order approximations capture the essence of nonlinear PDE feedback linearization. It argues that while exact feedback linearization is theoretically elegant, it is practically infeasible. The truncation approach offers a viable implementation strategy where the loss of exact linearization is compensated by a smaller, yet quantifiable, region of attraction.

The authors emphasize that the "higher-order vanishing" of the truncation residual is the key mechanism that reconciles practical stability with asymptotic stability. As the state shrinks, the residual becomes negligible, allowing the system to converge to zero despite the controller not being an exact linearizer.

The paper positions itself as Part I of a two-part study. It establishes the analytical guarantees for the truncated controller, while Part II (cited as [5]) is tasked with replacing the computationally expensive kernel evaluations with a learned neural-operator surrogate, thereby eliminating the need for online PDE solving and nested integration. The current work does not propose experimental validation or specific engineering applications beyond the theoretical framework of boundary control for hyperbolic PDEs.