Residual-Corrected Physics-Informed Neural Network KKL Observers with Input-to-State Stability Guarantees
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "Residual-Corrected Physics-Informed Neural Network KKL Observers with Input-to-State Stability Guarantees".
Paper summary: Rosa: So, we're looking at this paper, "Residual-Corrected Physics-Informed Neural Network KKL Observers with Input-to-State Stability Guarantees." Basically, they're tackling the problem that these physics informed neural network observers usually have a bit of an error.
Dev: They say that to use a physics informed neural network observer for something called the Kazantzis–Kravaris–Luenberger or KKL observer, you need two things: an approximate answer to this partial differential equation, and then an approximate inverse of that solution. And they point out that either of those approximations can leave some estimation error even when there isn't any measurement noise at all.
Taro: That approximation gap is the core issue here, right? If you get that wrong, you get a nonzero estimation bias right away before even considering what happens when the world throws us noisy data.
Rosa: Exactly. So what this paper claims is that they remove both of those errors by building something new from scratch during construction. They build an encoder that has an explicitly invertible neural map.
Dev: An explicitly invertible map means they have a closed-form inverse for it, and they use that inverse to reconstruct the internal state of the system. This reconstruction is then used to check the learned residual and fix the latent dynamics.
Taro: And because they do this construction, they make sure that the synchronization manifold—which is basically where everything should be perfectly aligned—is exactly invariant no matter what you set for those network weights. That sounds like a big deal for stability.
Rosa: It really does sound like it. And based on how they analyze the dynamics after correcting for that residual, they prove something called input-to-state stability or ISS for the error dynamics when there is measurement noise involved.
Dev: ISS means if you have some disturbance from the measurements, you can guarantee that your estimation error will converge exponentially to zero over time. That's a very strong claim for a real-time system.
Taro: Exponential convergence in the noise-free setting is even stronger, which means if there's no noise at all, the error just vanishes super fast without any external disturbances pushing it around.
Paper summary: Rosa: They back that up with two numerical examples where they test this out: a van der pol oscillator and the chaotic lorenz system. And in both cases, these residual-corrected observers showed exact noise-free convergence and better error response than the standard physics informed neural network KKL baselines.
Dev: So, to put it plainly, they've fixed the known approximation errors by building a specific structure into their encoder, and they proved that this fix leads to guaranteed stable behavior even when you have noisy measurements.
Taro: It means for autonomous systems or anything running in real-time loops where you need reliability under imperfect conditions, this construction offers a way to ensure the error doesn't just get small, but disappears exponentially fast when things are quiet.
Rosa: It suggests that the way they train the system can be structured so that it inherently knows how to correct its own approximation errors during operation, rather than just hoping it gets good enough through training alone.
Dev: So, if you're designing a control loop and you know the dynamics are governed by this kind of equation, this paper suggests structuring your neural network components in a specific way—with that invertible map—to get those stability guarantees.
Taro: It means the hardware or software design itself can be leveraged to enforce the required mathematical properties, instead of just tuning parameters until they look good on a test set.
Rosa: And the training part is interesting too. They use a two-stage process where they first train some components using a loss function that penalizes things like the KKL residual and contraction terms. Then, in the second stage, they freeze those main parts and only train an output map to handle noise reconstruction while simulating noisy trajectories.
Dev: That two-stage training sounds like a practical way to separate the core physics learning from the noise handling part of the system. It makes sense for real-time deployment because you can get the stable dynamics right first, and then fine-tune it against real world noise later.
Paper summary: Taro: But what about those two stages? If you have to do two separate training runs, how do we know if that separation actually helps or if it just complicates the process unnecessarily for a complex system?
Rosa: Well, the paper seems to argue that separating them is what lets them get these distinct stability proofs. They are proving ISS on the corrected observer dynamics separately from the output reconstruction map Q.
Dev: That's good because it means you can analyze and verify the stability of each part independently before putting it all together into one final system running in real time. It addresses a key safety concern for control engineers, which is how much uncertainty you actually have in your estimation process.
Taro: When we test these things outside the lab, the world misbehaves constantly. If the system is designed to handle that misbehavior gracefully—if it has this inherent stability property—it means when the sensor data gets weird, the system doesn't just blow up; it converges back to a safe state quickly.
Rosa: That’s where I see it. It moves the problem from "how do we tune this until it works?" to "how do we build this structure so that no matter what weird input comes in, the error stays bounded and shrinks predictably."
Dev: So, for those of us who deal with real-time loops and safety cases, the implication is that if you use a system designed with these residual corrections, your safety case doesn't rely on assuming noise will be small; it relies on proven stability properties.
Taro: It’s about moving beyond just getting decent performance in simulation to having a mathematical guarantee about how the error behaves when you actually run it in the real world.
Rosa: So, looking at this "Residual-Corrected Physics-Informed Neural Network KKL Observers with Input-to-State Stability Guarantees," it’s about making these physics informed observers fundamentally more robust by fixing errors in their construction and proving stability for the error dynamics under noise.
Dev: It’s a way to build more trustworthy observers, especially when you need that convergence guarantee in noisy conditions.
Taro: It gives us a blueprint for designing systems where the mathematical structure itself provides the safety margin against real-world disturbances.
Conclusion: Rosa: So, we're wrapping up on this paper about these residual-corrected physics informed neural network Kazantzis Kravaris Luenberger observers.
Dev: Yeah, it’s about fixing those approximation gaps that usually show up in the solution and its inverse.
Rosa: Exactly. They build an encoder with a map that you can actually invert, which lets them reconstruct the internal state to correct the residual dynamics.
Dev: And they proved input-to-state stability for the error dynamics when there is measurement noise involved, meaning it converges exponentially to zero in the quiet settings.
Rosa: So they're essentially removing both errors by construction—the approximation error and that inversion error—and showing that this new structure guarantees stable behavior under real-world conditions.
Dev: It moves the focus from just getting good simulation results to having a mathematical guarantee about how the error behaves when you actually run the system.
Rosa: That's right. They use examples like the van der pol oscillator and the chaotic lorenz system to show that these corrected observers converge exactly in noise-free settings and handle noisy ones better than standard physics informed neural network approaches.
Dev: The implication is that for real-time loops, this construction offers a way to build more trustworthy observers because they can rely on proven stability properties instead of just tuning parameters until it looks okay on paper.
Rosa: It’s about building the system structure itself to enforce the required mathematical properties, which is something we need when things go wrong outside the lab.
Dev: So, if you're designing a control loop and you know your dynamics are governed by this kind of equation, this paper suggests structuring your neural network components with that invertible map to get those stability guarantees.
Rosa: It means for autonomous systems or anything running in real-time where you need reliability under imperfect conditions, the design can inherently handle the uncertainty.
Dev: We'll take a look at how they actually built these components in their training procedure next.
Hannah Berin-Costain, Jun Liu, Kirsten Morris
University of Waterloo
eess.SY, cs.SY
Submitted: 2026-11-10
Updated: 2026-11-10
Comments: 7 pages, 1 figure, ACC 2027 conference submission
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The gist: Physics-informed neural network (PINN) Kazantzis–Kravaris–Luenberger (KKL) observers require an approximate solution to a KKL partial differential equation (PDE) and an approximate inverse of…
Key concepts
- KKL Observer
- A type of observer used for systems described by a KKL partial differential equation (PDE). These observers estimate the internal state of a system based on measurements and the PDE, but they often suffer from errors because they rely on approximate solutions to the underlying PDE.
- Explicitly Invertible Neural Map
- A specific neural network component designed to have a closed-form inverse. This map is used as an encoder that allows for the exact reconstruction of the internal state from latent coordinates, eliminating approximation errors in state recovery.
- Input-to-State Stability (ISS)
- A mathematical guarantee that describes how the error in a system's estimation behaves when subjected to external inputs, such as measurement noise. Proving ISS shows that the estimation error will be bounded and converge exponentially to zero if the input (noise) is zero.
- Residual-Corrected Observer Dynamics
- The mathematical model describing how the observer's state evolves after correcting for errors in the learned KKL residual. This corrected dynamic is shown to contract, leading to stable and accurate estimation, even in the presence of measurement noise.
Summary
Physics-informed neural network (PINN) Kazantzis–Kravaris–Luenberger (KKL) observers require an approximate solution to a KKL partial differential equation (PDE) and an approximate inverse of this solution <ref:2611. p. 013411#pg2>. In this paper, we remove both errors from the implemented observer by construction <ref:2611. p. 013411#pg6>. Based on contraction of the residual-corrected observer dynamics, we prove an input-to-state stability (ISS) result for the implemented error dynamics with respect to measurement noise. p. 0
Pruning redundant reasoning tokens
The paper introduces a construction that eliminates approximation gaps in learned KKL observers by building an encoder containing an explicitly invertible neural map <ref:2611.13411#pg6>. This closed-form inverse provides the internal state reconstruction used to evaluate the learned KKL residual and correct the latent dynamics <ref:2611. p. 013411#pg6>. This construction makes the synchronization manifold exactly invariant for every value of the network weights <ref:2611. p. 013411#pg6>. The authors also provide an input-to-state stability (ISS) theorem for the error dynamics, which implies exact exponential convergence of the estimation error in the noise-free setting <ref:2611. p. 013411#pg6>.
How the training signal is built
The training procedure involves a two-stage process for building the neural components <ref:2611.13411#pg5>. In the first stage, the parameters of Tˆθ and φθ are trained jointly using a loss function Lobs that penalizes the KKL residual, contraction term, hard constraint on symmetric derivatives, linear inversion error, and measurement gain <ref:2611. p. 413411#pg5>. In the second stage, Tˆθ, φθ, and Q0 are frozen while only Qψ is trained by simulating noisy observer trajectories to minimize the output reconstruction error under noise <ref:2611. p. 413411#pg5>.
Where the gains come from
The residual-corrected observer dynamics are defined as F(z, m):= φ(z, m) + r(Q0(z)) <ref:2611.13411#pg6>. The internal reconstruction map Q0 is defined explicitly by Q0(z) = T−10 (z1:nx) <ref:2611. p. 013411#pg6>. The output reconstruction map Qψ uses all latent coordinates and is trained to reduce the reconstruction error under measurement noise <ref:2611. p. 213411#pg6>.
Numerical assessment of the stability conditions
The authors evaluate the ISS conditions by numerical sampling rather than uniform verification over the continuous domain <ref:2611.13411#pg7>. They fit a constant metric P≻0 using corrected-field Jacobians on the encoder image and retain a tube only when sampled contraction and tube-invariance conditions of Theorem 2 are satisfied. p. 5 The screen gives contraction rates 0. p. 5173 on the VDP tube R = 0. p. 5003 and 0.228 on the Lorenz tube R = 0. p. 5015, and both tubes pass the sampled invariance test at the principal noise levels. p. 5
Conclusion
In this work we have developed a neural KKL observer that removes the nominal error caused by approximate encoder dynamics and state reconstruction <ref:2611.13411#pg7>. An invertible component of the encoder provides the internal reconstruction needed to evaluate the KKL residual correction, making the synchronization manifold exactly invariant <ref:2611. p. 513411#pg7>. A separate output map uses all latent coordinates while remaining an exact left inverse of the encoder for every value of its trainable parameters <ref:2611. p. 513411#pg6>. We also established an ISS result for the corrected observer. p. 5 The corrected observers converged to the integration accuracy limit without noise and generally reduced noisy estimation errors relative to the released PINN–KKL implementations. p. 5
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Precise estimation error convergence in noise-free settings: The residual-corrected observer achieves
exact exponential convergence of the estimation error in the noise-free setting
as proven by Theorem 2, leading to superior performance compared to released PINN–KKL baselines which retain anonzero estimation bias even without measurement noise.
-
Robustness against bounded measurement noise: The system provides an explicit ultimate bound on the estimation error with respect to noise, stated in equation 16 as "lim sup t->inf x hat(t) - x(t) 2 <= LQ ∆(¯v) α," ensuring reliable performance even when measurement disturbances are present.
-
Exact synchronization manifold invariance: The construction ensures that the
synchronization manifold is exactly invariant for every value of the network weights,
which guarantees thatthe encoded system trajectory an exact solution of the observer dynamics, regardless of the residual magnitude.
-
Improved state estimation accuracy on complex systems: Numerical studies demonstrate substantial improvements in error reduction, such as achieving approximately 8.19 times smaller overall RMSE for the Van der Pol oscillator and 93.8 times smaller overall RMSE for the chaotic Lorenz system compared to released PINN–KKL checkpoints under similar noise conditions.
-
Separation of stability and output training: The architecture separates the stability condition from output training by utilizing two distinct reconstruction maps, Q0 and Q, allowing the observer dynamics to be governed by a contraction condition involving a fixed map Q0 while using a trainable map Q for noise-aware state estimation.
Sources
- KKL Observer Synthesis for Nonlinear Systems via Physics-Informed Learning
- Learning a Contracting KKL-observer with Local Optimal Guarantees
Related papers
- Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems
- Exact and Approximate Convex Reformulation of Linear Stochastic Optimal Control with Chance Constraints
- Certified Set Convergence for Piecewise Affine Systems
- Symbolic Control for Stochastic Systems via Finite Parity Games
- Multiobjective optimization-based design and dispatch of islanded, hybrid microgrids for remote, off-grid communities in sub-Saharan Africa
- Physiology-informed layered sensing for intelligent human-exoskeleton interaction