On Data-Driven Koopman Representations of Nonlinear Delay Differential Equations
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "On Data-Driven Koopman Representations of Nonlinear Delay Differential Equations".
Dev: Here's the gist of the paper. On Data-Driven Koopman Representations of Nonlinear Delay Differential Equations this work establishes a rigorous bridge between infinite-dimensional delay dynamics and finite-dimensional Koopman learning with explicit and interpretable error guarantees > The Core Problem Addressed Delay differential equations (DDEs) evolve on an infinite-dimensional phase space consisting of history segments, which fundamentally complicates modeling, prediction, and analysis Existing approaches face a fundamental trade-off between infinite-dimensional fidelity, computational tractability, and theoretical reliability While Koopman analysis is well-developed for ordinary differential equations (ODEs) and partially for partial differential equations (PDEs), its extension to delay differential equations (DDEs) remains limited due to the infinite-dimensional phase space of DDEs To the best of our knowledge, however, a quantitative bridge from infinitedimensional DDE dynamics to finite-dimensional data-driven Koopman representations with explicit approximation error bounds has not yet been established > Finite-Dimensional Realization Framework The key idea is to construct a finite-dimensional realization of the infinite-dimensional delay dynamics via history discretization and a suitable reconstruction operator, enabling a principled connection between the Koopman operator on the space of history functions and a finite-dimensional datadriven representation 1.
Rosa: Koopman realization for DDEs establishes a rigorous bridge between infinite-dimensional delay dynamics and finite-dimensional Koopman representations through a sampling–reconstruction framework, yielding a tractable surrogate model 2.
Dev: A finite-dimensional realization of W as the compact operational domain Z is defined using uniformly spaced discretization points and a sampling operator Q: W → Z 3.
Rosa: The reconstruction operator R: Z → C is defined such that R(ζ)(θ) = ζj + θ−θj δ(M) (3), and Lemma 1 proves that R maps Z into W, i.e., R(Z) ⊆ W Data-Driven Surrogate Construction Using kernelbased extended dynamic mode decomposition (kEDMD) with Wendland radial basis function (RBF) kernels, the authors construct a finite-dimensional linear surrogate model that rigorously approximates the infinite-dimensional Koopman operator defined on the Banach space of history functions through a data-driven matrix representation 1.
Title and authors: Rosa: So we're looking at this paper now called "On Data-Driven Koopman Representations of Nonlinear Delay Differential Equations". It seems like the authors are trying to connect these infinite-dimensional delay dynamics—which are super complicated because they involve the whole history—to something finite that we can actually learn with data.
Dev: Right, it's tackling that fundamental problem where delay differential equations live in this infinite space of past states, which makes modeling and predicting them really hard. This paper proposes a way to bridge that gap using data-driven Koopman learning with explicit error guarantees.
Taro: So it’s not just trying to guess the future, it’s building a rigorous mathematical link between those continuous history functions and a finite set of data points we can actually use for prediction. I wonder how stable this connection is when things get messy in the real world.
Rosa: Well, what they're doing is setting up this finite-dimensional realization by discretizing the history and using a reconstruction operator to link it back to a manageable space Z. They’ve done that sampling and reconstruction framework to create a tractable surrogate model for that infinite-dimensional Koopman operator.
Dev: That’s the first part, creating this finite representation Z using uniformly spaced discretization points and this sampling operator Q, which maps history functions W into Z. Then they define this reconstruction operator R that takes data from Z back into the history space C, and Lemma one proves that this reconstruction actually stays within the original space W.
Taro: So it’s not just a theoretical trick; they’ve shown mathematically that you can go from your real infinite dynamics down to something finite with a controlled mapping. That sounds promising for testing things out where the physics is complex.
Rosa: Exactly, and then they use this setup to construct a data-driven surrogate model using kernel based extended dynamic mode decomposition, or kEDMD, with Wendland radial basis functions as the kernels. This lets them build a linear model that approximates the Koopman operator defined on the space of history functions through a matrix representation derived from their dataset T.
Dev: Their dataset T is built from transition pairs, which are these true history state evolution pairs, and they approximate the lifted state evolution using this surrogate model Ψ of z plus being close to A times Ψ of z. The matrix A itself is constructed as K transpose F hat X K minus one X.
Title and authors: Taro: Wait, so the prediction isn't just a simple lookup; it’s approximating this lifted state evolution using this specific data-driven model, and they claim theorem one gives them an explicit error bound for the kEDMD surrogate. That’s a big deal for reliability.
Rosa: It is, because that deterministic error bound breaks down the total approximation error into three pieces: one from the history discretization, Rse delta M; another from the kernel interpolation, Rih X; and a third one from the data-driven regression, Rd rho squared.
Dev: And they show that this residual norm is bounded by those three terms combined. Specifically, it says that the residual norm is less than or equal to Rs plus Ri plus Rd squared, where Rs and Ri are defined by norms involving the history space C.
Taro: So when you look at it for someone who just wants to run a simulation, what does this mean practically? Are these error terms something you can tune or control?
Rosa: It means the prediction accuracy is governed jointly by how well the kernel reconstruction works and how tight that residual bound is. They also developed a state reconstruction mechanism using kernels that lets them recover the finite dimensional discretized state from those lifted Koopman coordinates, and they establish deterministic bounds on the error in predicting that reconstructed state.
Dev: That part of it is key for control applications because it supports reliable simulation and potential control applications, which means we can trust what the model predicts when we try to use it to steer a system.
Taro: It sounds like they've done more than just mapping the dynamics; they’ve built a complete loop from infinite space down to a bounded prediction error that you can calculate. But what’s the catch? The paper mentions these error terms, and I want to know what those limitations are in practice.
Rosa: They do mention the dependency of the regression error on assumption three remaining satisfied, which means their performance is tied to how well that specific assumption holds in reality. And they also note that this whole framework relies on the smoothness of the time delta map F being twice continuously Fréchet differentiable on a neighborhood of W.
Title and authors: Dev: So, to put it plainly for someone driving or running a loop, it’s a powerful tool because they give you the numbers for how much error you can expect from discretization and kernel choices, but you still have to make sure the underlying assumptions about the delay dynamics hold true.
Taro: So if we take this idea—taking infinite delay dynamics and mapping them to a finite system with provable error bounds—what does that actually change for how we approach modeling really complex, time-delayed systems in robotics or control?
Rosa: It changes it by providing a concrete path. Instead of just saying "Koopman learning might work," they show exactly how to build the necessary framework, which is based on history discretization and kernel methods. This gives us a principled way to tackle DDEs without just throwing random data at a black box.
Dev: It’s about moving beyond the simple ODE case where Koopman analysis is already solid, and finally giving us an explicit roadmap for applying that theory to delay dynamics where the phase space explodes.
Taro: I’m interested in how this connects to other work we're seeing, like those papers on real-time execution or agent planning; can this framework be used when we need fast reactivity loops?
Rosa: It could certainly feed into that. If you can build a fast, low-error surrogate model for the dynamics of a delayed system, you have a better starting point for training those real-time autoregressive policies you're working on.
Dev: That makes sense. If the underlying system model is more accurate and has known error bounds, then the real-time policy doesn't have to work as hard to compensate for model uncertainty in its feedback loop.
Taro: So it’s about making the world model itself more reliable before you even try to make an autonomous decision on top of it. That seems like a solid layer of safety for complex tasks.
Rosa: Yeah, that's the implication I see for deployment—you get a more trustworthy predictor because you know exactly where the error is coming from, whether it’s in the sampling or in the kernel interpolation.
Dev: Ultimately, this paper on "On Data-Driven Koopman Representations of Nonlinear Delay Differential Equations" gives us a concrete way to handle these infinite-dimensional dynamics by turning them into finite, verifiable models with explicit error control.
Taro: It’s a solid piece of theory that brings rigor to a very messy area, and I think having these error guarantees is what makes it really useful for moving from simulation to real-world testing.
The paper's summary: Rosa: So, we just talked about how this paper connects infinite dimensional delay dynamics to something finite we can learn from data. Now, let's look at what they actually claim they did with that bridge and those error guarantees.
Dev: They built a finite-dimensional realization of the history dynamics using a sampling and reconstruction framework. They defined a space Z where the data lives, and then proved mathematically that you can reconstruct the actual history information from that finite set, which is really neat because it gives us a controlled way to move from theory to data.
Rosa: And they didn't just stop there; they used this setup to build a linear surrogate model using kernel based extended dynamic mode decomposition, or kEDMD. This lets them approximate that infinite dimensional Koopman operator with a matrix representation derived straight from their dataset T.
Dev: The real meat is the error analysis. They provide an explicit and interpretable error bound for the learned Koopman predictor. They break down the total approximation error into three distinct parts: one from how they discretize the history, another from how well they interpolate with those kernels, and a third from the actual data driven regression.
Rosa: That decomposition is what makes it useful. It tells you exactly where your prediction error is coming from so you can try to fix it. For example, if your sampling interval is too big, that first error term goes up directly.
Dev: And they show the residual norm—the total leftover prediction error—is bounded by the sum of those three terms. This gives us a concrete measure of accuracy you can trust when you're deploying this model for something real-time, like in a control loop.
Rosa: What this means for people using this, whether it's field robotics or autonomous driving, is that they can finally stop guessing how accurate the model is and start calculating the actual risk associated with their prediction. You get a deterministic bound on how far off your prediction might be.
Dev: It also includes a state reconstruction mechanism using those same kernels. This lets you recover the finite dimensional state from those lifted Koopman coordinates, and they give you error bounds on that reconstruction too, which is vital for keeping your system stable when the world gets weird.
Rosa: So it’s not just a new way to learn dynamics; it’s a rigorous framework that turns messy infinite delay systems into finite models with provable safety margins. It moves us past just having a good guess toward having a verifiable predictor.
Dev: And they flag the limitations pretty clearly, though. They say the regression error bound depends on an assumption about how smooth the time map F is—if that assumption doesn't hold perfectly in reality, that specific part of your error calculation might be off.
Rosa: So while they give us a very solid roadmap for building these finite models and quantifying their errors, you still have to ensure the underlying physical assumptions about the delay dynamics are actually true for what you're modeling.
Dev: Exactly. It’s a powerful tool because it gives you the numbers for how much error you can expect from discretization and kernel choices, but you still have to make sure those assumptions hold true when your system starts misbehaving in the real world.
Rosa: It’s about moving beyond just saying "Koopman learning might work" to showing exactly how to build the necessary framework, which is based on history discretization and kernel methods. This gives us a principled way to tackle delay differential equations without just throwing random data at a black box.
The paper's improvements: Rosa: So we just looked at the core idea of how they map infinite delay dynamics to a finite representation using kernels and sampling, but now let’s talk about what they suggest changing things about this field.
Dev: They aren't just presenting a model; they’re proposing a way to make the entire system more robust for real-time use by focusing on that error analysis we talked about earlier.
Rosa: Right, they focus on making those three error terms—discretization, kernel interpolation, and regression—more manageable. They suggest ways to tune the sampling interval or choose better kernel functions based on those explicit bounds.
Dev: That’s important because it gives a practical path for someone writing a control loop. It means you can explicitly decide how much error you are willing to tolerate in the prediction based on your hardware limits and your required safety case.
Rosa: And they also hint at future work that involves integrating this with other learning methods, like using this surrogate model as a better starting point for training those real-time autoregressive policies we’ve been discussing.
Dev: That makes sense. If you have a world model that has known error bounds, the policy doesn't have to work as hard to compensate for uncertainty when it tries to act on that prediction. It smooths out the learning process significantly.
Rosa: The paper also points toward testing this framework in more extreme scenarios, like when the system starts misbehaving in ways that aren't just small noise—really testing those failure modes Taro and I always talk about.
Dev: And for that, they’re looking at how to make the reconstruction operator itself more adaptive. Instead of a fixed mapping, maybe it can adjust based on what kind of weird behavior the system is actually exhibiting in real time.
Rosa: So the idea is to build a system where you don't just run one fixed model; you have a framework that can adapt its level of detail or its reconstruction method depending on whether the dynamics are behaving normally or if they’re entering a chaotic regime.
Dev: That sounds like it could lead to much faster recovery mechanisms when things go wrong, which is exactly what we need for reliable control systems. We want graceful failure, not just a sudden crash.
Rosa: It really pushes us toward building these world models that are more than just accurate snapshots; they become active participants in managing the uncertainty of the system over time.
Dev: So it moves us toward creating a dynamic safety case where the model's reliability isn't static, but is continuously assessed based on real-time performance metrics.
Conclusion: Tom: So we’ve gone through how they build these finite models for delay differential equations using kernel methods and how they break down the error into three parts—discretization, interpolation, and regression—but now we’re wrapping up with the big picture implications of "On Data-Driven Koopman Representations of Nonlinear Delay Differential Equations."
Rosa: Basically, this paper gives us a way to turn messy infinite delay dynamics into finite models that we can actually train data on while knowing exactly how much error to expect in our predictions.
Dev: It’s a significant step because it finally provides an explicit safety case for using Koopman methods on delay systems, moving it beyond just theoretical neatness into something you can put on a loop.
Taro: I think the biggest implication for autonomy is that we can start building world models that are more reliable when they have inherent time lags, which is a huge part of interacting with the physical world.
Rosa: Exactly, it means we stop guessing how stable our prediction is and start calculating the actual risk based on those known error bounds. It’s about building a trustworthy predictor before you even try to use it for control or planning.
Dev: For real-time systems, knowing that your latency will not explode unpredictably because of bad sampling or poor kernel choice is crucial for designing a robust safety case that works in practice.
Taro: And when the world misbehaves, this framework suggests we can recover more gracefully because we have better knowledge of where the error is coming from and how to bound it. It supports more resilient recovery strategies.
Rosa: So, to sum up, "On Data-Driven Koopman Representations of Nonlinear Delay Differential Equations" provides the mathematical toolkit to bridge the gap between complex history dynamics and data driven learning with provable performance metrics.
Dev: It’s a solid piece of theory that brings necessary rigor to modeling time delays where we usually just have to hope things work out.
Taro: It sets a new standard for how we should approach modeling these kinds of systems in robotics and control, making the uncertainty quantifiable instead of just abstract.
Rosa: Next week, we’re looking at how this idea connects directly to building those fast autoregressive policies for real-time execution and what that means for immediate action on the field.
Santosh Mohan Rajkumar, Dibyasri Barman, Kumar Vikram Singh, Debdipta Goswami
The Ohio State University · Miami University
eess.SY, cs.LG, cs.SY, math.DS
Submitted: 2026-04-03
Updated: 2026-08-12
Comments: Github: https://github.com/santoshrajkumar/koopman-dde-kEDMD
Journal ref: 2026 65th IEEE Conference on Decision and Control
Code: https://github.com/santoshrajkumar/koopman-dde-kEDMD
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 89/100
The gist: On Data-Driven Koopman Representations of Nonlinear Delay Differential Equations this work establishes a rigorous bridge between infinite-dimensional delay dynamics and finite-dimensional Koopman…
Key concepts
- Delay Differential Equations (DDEs)
- These are mathematical equations that describe how a system's state changes over time, but they depend not just on the current state but also on its past history. Because of this dependence on an infinite sequence of past states, they exist in an infinite-dimensional space, making them very hard to analyze directly.
- Koopman Operator
- The Koopman operator is a mathematical tool that allows you to transform a complex nonlinear system (like DDEs) into an equivalent linear system in a higher-dimensional space. If you can find this linear representation, you can use well-established linear methods to predict the future state of the original nonlinear system.
- kEDMD
- Kernel Extended Dynamic Mode Decomposition (kEDMD) is a data-driven technique used here to build a finite-dimensional linear model. It uses radial basis functions (kernels) to map the infinite-dimensional dynamics onto a manageable, finite set of coordinates, effectively approximating the true Koopman operator using observed data.
Terminology
Summary
On Data-Driven Koopman Representations of Nonlinear Delay Differential Equations this work establishes a rigorous bridge between infinite-dimensional delay dynamics and finite-dimensional Koopman learning with explicit and interpretable error guarantees >
The Core Problem Addressed
Delay differential equations (DDEs) evolve on an infinite-dimensional phase space consisting of history segments, which fundamentally complicates modeling, prediction, and analysis Existing approaches face a fundamental trade-off between infinite-dimensional fidelity, computational tractability, and theoretical reliability While Koopman analysis is well-developed for ordinary differential equations (ODEs) and partially for partial differential equations (PDEs), its extension to delay differential equations (DDEs) remains limited due to the infinite-dimensional phase space of DDEs To the best of our knowledge, however, a quantitative bridge from infinitedimensional DDE dynamics to finite-dimensional data-driven Koopman representations with explicit approximation error bounds has not yet been established >
Finite-Dimensional Realization Framework
The key idea is to construct a finite-dimensional realization of the infinite-dimensional delay dynamics via history discretization and a suitable reconstruction operator, enabling a principled connection between the Koopman operator on the space of history functions and a finite-dimensional datadriven representation
-
Koopman realization for DDEs establishes a rigorous bridge between infinite-dimensional delay dynamics and finite-dimensional Koopman representations through a sampling–reconstruction framework, yielding a tractable surrogate model
-
A finite-dimensional realization of W as the compact operational domain Z is defined using uniformly spaced discretization points and a sampling operator Q: W → Z
-
The reconstruction operator R: Z → C is defined such that R(ζ)(θ) = ζj + θ−θj δ(M) (3), and Lemma 1 proves that R maps Z into W, i. p. 1
e., R(Z) ⊆ WWhere the paper says it
(2) Then Q(W) ⊆ Z, since every η ∈ W satisfies the defining constraints of Z. Now, define the reconstruction operator R : Z → C for any ζ ∈ Z, R is defined as R(ζ)(θ) := ζj + θ−θj δ(M) (ζj+1 − ζj ), ∀θ ∈ [θj , θj+1], (3) and j ∈ [1 : M − 1].
Data-Driven Surrogate Construction
Using kernelbased extended dynamic mode decomposition (kEDMD) with Wendland radial basis function (RBF) kernels, the authors construct a finite-dimensional linear surrogate model that rigorously approximates the infinite-dimensional Koopman operator defined on the Banach space of history functions through a data-driven matrix representation
-
The dataset T is constructed from transition pairs (zk, zk+1) corresponding to true history state evolution pairs (ϕk, ϕk+1)
-
The lifted state evolution is approximated by the data-driven surrogate model Ψ(z+) ≈ AΨ(z), where A = K⊤ Fˆ(X)K−1X
-
Theorem 1 provides a deterministic error bound for the kEDMD-based surrogate (20), comprising the history discretization error Rseδ(M), kernel interpolation error RihX, and local regression error Rdρ2
Error Analysis and Guarantees
The deterministic error bound (21) decomposes the total approximation error into three components:
((
Rseδ(M), kernel interpolation error RihX, and data-driven regression error Rdρ2 The regression error is quadratic in ρ, provided Assumption 3 remains satisfied
The residual satisfies∥rϕ∥ ≤ Rseδ(M) + RihX + Rdρ2, with Rse = Rs + Re where Rs = C1LF Lθ∥Ψ∥H, Ri = BX∥Ψ∥H, and Rd = C0√d p LkZ¯ DF2/2∥Ψ∥HK−1X
The prediction accuracy is governed jointly by kernel reconstruction and the residual bound in Theorem 1
Numerical Validation
Numerical results demonstrate convergence with respect to discretization resolution and data density, showing consistently lower step-wise mean prediction error (MPE) µzk as M increases (hX ≈ 0.
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
-
Data-Driven Linear Surrogate Model Construction: The system can construct a finite-dimensional linear surrogate model using kernel based extended dynamic mode decomposition (kEDMD) with Wendland radial basis function kernels to rigorously approximate the infinite dimensional Koopman operator on the space of history functions.
-
Deterministic Error Guarantees: The improved system provides explicit and interpretable error bounds for the learned Koopman predictor, decomposing total approximation error into contributions from history discretization, kernel interpolation, and data driven regression.
-
State Reconstruction with Guarantees: The system develops a kernel based mechanism to recover the finitedimensional discretized state from lifted Koopman coordinates and establishes deterministic bounds on the resulting state prediction error, supporting reliable simulation and potential control applications.
Abstract
This work establishes a rigorous bridge between infinite-dimensional delay dynamics and finite-dimensional Koopman learning, with explicit and interpretable error guarantees. While Koopman analysis is well-developed for ordinary differential equations (ODEs) and partially for partial differential equations (PDEs), its extension to delay differential equations (DDEs) remains limited due to the infinite-dimensional phase space of DDEs. We propose a finite-dimensional Koopman approximation framework based on history discretization and a suitable reconstruction operator, enabling a tractable representation of the Koopman operator via kernel-based extended dynamic mode decomposition (kEDMD). Deterministic error bounds are derived for the learned predictor, decomposing the total error into contributions from history discretization, kernel interpolation, and data-driven regression. Additionally, we develop a kernel-based reconstruction method to recover discretized states from lifted Koopman coordinates, with provable guarantees. Numerical results demonstrate reliable prediction of nonlinear delay systems, with potential relevance to future control applications.
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