DiSKO: Distribution Steering for Nonlinear Systems using the Koopman Operator
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: "DiSKO: Distribution Steering for Nonlinear Systems using the Koopman Operator".
Dev: Here's the gist of the paper. This paper proposes DiSKO, a robust framework for data-driven distribution steering of stochastic nonlinear dynamical systems using Koopman operator theory.
Rosa: The method learns an approximate linear representation from data and employs a distributionally robust approach to handle approximation errors and disturbance ambiguity, which is significant because it allows control over the mean and covariance of the underlying nonlinear system Framework Overview The core idea involves approximating nonlinear dynamics with a linear model in a lifted state space using Koopman operator theory and then designing controllers for the mean and covariance using linear distribution steering algorithms.
Dev: To manage errors from the learned dynamics and disturbance ambiguity, it formulates the problem as a distributionally robust covariance steering problem.
Rosa: A deep neural network (DNN) is leveraged to simultaneously learn Koopman observables and linear model parameters, which eliminates manual observable selection and promotes desirable properties like stability during training.
Title and authors: Rosa: Now, looking at this paper's title, "DiSKO: Distribution Steering for Nonlinear Systems using the Koopman Operator," it really tells you what they're trying to achieve. It’s about steering the distribution—the way the system spreads out in its possible states—using that Koopman operator idea.
Dev: And from what I can tell, they are moving beyond just predicting where the state will go next; they are designing controllers for both the average behavior and how spread out those possible states are.
Rosa: That means for a stochastic nonlinear system, you get control over both the mean and the covariance of that distribution. The paper says this is done by using linear distribution steering algorithms on that learned linear model.
Taro: I always wonder if steering the covariance is as hard as steering just the average state, especially when you're dealing with nonlinearities where small errors can get really big.
Dev: That’s where their distributional robust approach comes in; they formulate it as a distributionally robust covariance steering problem to handle those approximation errors and disturbance ambiguity.
Rosa: And they use that DNN to learn the observables and the model parameters simultaneously, which is supposed to eliminate the need for manual observable selection while also shrinking the size of that lifted representation.
Taro: Shrinking it down sounds good for computation, but I wonder if you lose crucial information about what’s really happening in those nonlinear dynamics.
Dev: They try to prevent that by adding a stability penalty during training, which pushes the linear models they learn to have a spectral radius away from one, promoting desirable model properties like stability while the DNN is learning.
The paper's summary: Rosa: So, when you read their summary of "DiSKO: Distribution Steering for Nonlinear Systems using the Koopman Operator," it boils down to this: they use Koopman theory to get a linear approximation from data in a lifted space. Then, they use that linear model to design controllers for the mean and covariance using distributionally robust steering methods.
Dev: They’re essentially trying to bridge the gap between nonlinear physics and standard linear control by learning that linear map directly from observed data instead of assuming it beforehand.
Rosa: And they use a deep neural network to do this mapping—it learns both the observables and the model parameters together, which means they don't have to pick those important features manually.
Taro: So, they're automating the discovery of what observables are useful for steering these systems? That’s a powerful move if it works reliably in practice.
Dev: It should work well when you think about the goal: being able to control the system's mean and covariance even when there are errors in your model or unknown disturbances affecting your system.
Rosa: Because they address those errors by framing it as a distributionally robust problem, which allows them to steer the actual state distribution into a safe set defined by risk constraints.
Taro: So the paper claims this method gives you control over the mean and covariance of a nonlinear stochastic system, accounting for real-world model imperfections.
The paper's improvements: Rosa: The authors point out several improvements they’ve made to this framework, starting with how they handle the learning process itself. They introduce specific loss terms to make sure the DNN observables don't just collapse into constant maps by minimizing the expected trace of the covariance of that lifted state.
Dev: That loss term is important because it stops you from having a representation where everything is essentially stuck, which is a common issue when learning these kinds of maps.
Rosa: Plus, they add another thing to make sure the learned linear models are actually stable by including a stability penalty during training that penalizes the spectral radius of the model matrix away from one.
Taro: So they’re not just focusing on accuracy in prediction, but on making sure the underlying linear model itself is well-behaved and stable during that learning process.
Dev: That makes sense from an engineering standpoint; a system that learns a wildly unstable representation won't give you reliable control inputs later on.
Rosa: Then there’s the aspect of state constraints, where they use distributionally robust Conditional Value at Risk or DR-CVaR constraints to ensure the probability of violating certain limits stays below a risk level you specify, delta.
Taro: So they can enforce safety boundaries on these stochastic systems while steering them toward a desired distribution.
Dev: And finally, they reformulate the whole distribution steering problem as a convex semi-definite program or SDP, which means it’s solvable by standard off-the-shelf solvers, which is essential when you have those high-dimensional lifted states.
Conclusion: Rosa: So to wrap up on "DiSKO: Distribution Steering for Nonlinear Systems using the Koopman Operator," the authors show a way to use data to get a linear representation, then steer both the mean and covariance of a stochastic nonlinear system robustly.
Dev: The implication is that you can design controllers that explicitly account for errors in your learned dynamics and uncertainty in your disturbances by solving it as a distributionally robust problem.
Rosa: It also means you can automate the selection of those Koopman observables using a deep neural network, which simplifies the setup significantly while reducing the dimensionality of the lifted space.
Taro: The stability penalty and state constraint handling through DR-CVaR show they’ve thought about what happens when this system meets real-world operational limits and doesn't behave exactly as modeled.
Dev: And because they reformulate the whole thing as an SDP, it moves the problem from something computationally intractable to something solvable with standard tools, which is a big step for practical application.
Rosa: It’s a lot of machinery—Koopman theory, DNNs, distributionally robust optimization—but the goal is clear: control over mean and covariance in complex nonlinear systems using data-driven methods.
Taro: It’s promising because it tackles the difficulty of getting precise control when you don't have perfect knowledge of the system dynamics or the noise.
Dev: Yeah, it gives engineers a way to design robust feedback policies that work even when things are messy and nonlinear.
Rosa: That’s what we had with this paper, "DiSKO: Distribution Steering for Nonlinear Systems using the Koopman Operator." It shows how leveraging learned linear models can give us finer control over complex stochastic systems in the real world.
Dimitrios Moustroufis, Panagiotis Tsiotras
eess.SY, cs.SY
Submitted: 2026-11-12
Updated: 2026-11-12
License: http://creativecommons.org/licenses/by-sa/4.0/
Importance score: 78/100
The gist: This paper proposes DiSKO, a robust framework for data-driven distribution steering of stochastic nonlinear dynamical systems using Koopman operator theory.
Key concepts
- Koopman Operator Theory
- This theory allows complex nonlinear dynamics to be approximated by linear dynamics in a higher-dimensional space. It finds 'observables'—functions of the state—that make nonlinear evolution look like simple linear evolution, which is key for learning the system's underlying structure.
- Distributionally Robust Approach
- This method handles two types of uncertainty: approximation errors from using a finite-dimensional linear model and ambiguity in the actual disturbance. It ensures that the resulting control works well even when these uncertainties are at their worst possible values within defined sets.
- Deep Neural Network (DNN) Learning
- A DNN is used to simultaneously learn the Koopman observables and the parameters of the linear model. This automation eliminates the need for manually selecting which state variables to use as observables, leading to a more efficient and potentially more stable learned representation.
Terminology
Summary
This paper proposes DiSKO, a robust framework for data-driven distribution steering of stochastic nonlinear dynamical systems using Koopman operator theory. p. 1 Recently, data-driven algorithms for covariance steering
have been developed in [50], [51], which, however, are only
applicable to linear systems. The main contribution of this paper is to propose a robust
framework, henceforth referred to as DiSKO (Distribution
Steering using the Koopman Operator), for data-driven distribution steering of stochastic nonlinear dynamical systems.Where the paper says it
Framework Overview
The core idea involves approximating nonlinear dynamics with a linear model in a lifted state space using Koopman operator theory and then designing controllers for the mean and covariance using linear distribution steering algorithms. To manage errors from the learned dynamics and disturbance ambiguity, it formulates the problem as a distributionally robust covariance steering problem. p. 0 By exploiting this linear model, we design controllers for the
mean and covariance of the underlying nonlinear system by
utilizing linear distribution steering algorithms. To account for
approximation errors in the learned dynamics and the ambiguity in the disturbance distribution induced in the lifted space, we
formulate the problem as a distributionally robust covariance
steering problem. Furthermore, we leverage a deep neural network (DNN) to simultaneously learn the Koopman observables
and the linear model parameters, thus eliminating the need
for manual observables basis selection while also reducing the
dimensionality of the lifted representation. Additionally, we can
promote desirable model properties, such as stability, during
DNN training.Where the paper says it
Where the paper says it
Model Learning via DNNs
The framework utilizes a DNN to learn the Koopman observables and the linear model parameters jointly. The loss function for this optimization minimizes both the single-step prediction error and the total conditional variance of the lifted state, which is crucial for stochastic systems. p. 9 (72)
For a deterministic system, Cov(Y |X) = 0, and thus, the
goal of the optimization reduces to just minimizing the
single-step prediction error. For a stochastic system, we
have that EY [Y |X] = Ewk
[Ψθ
k+1|Ψθ
k
, uk], and thus the loss
function optimizes the single-step prediction error along with
the total conditional variance of the lifted state. In particular, to minimize the term EX[tr(Cov(Y |X))] in
(72), it is possible that the DNN observables “collapse” into
constant maps. To prevent this, we introduce the following
term into the loss function
Lvar = λvarReLU(κ
2
min − Var[Φd
(θ, xk)])2
, (73 Furthermore, we leverage a deep neural network (DNN) to simultaneously learn the Koopman observables
and the linear model parameters, thus eliminating the need
for manual observables basis selection while also reducing the
dimensionality of the lifted representation. Additionally, we can
promote desirable model properties, such as stability, during
DNN training.Where the paper says it
Where the paper says it
Where the paper says it
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 covariance steering for stochastic nonlinear systems: The system can steer mean and covariance using a linear model learned directly from data via Koopman operator theory, accounting for approximation errors and disturbance ambiguity through a distributionally robust approach.
-
Automated observable selection: A deep neural network can jointly learn the Koopman observables and the linear model parameters, eliminating the need for manual basis selection while simultaneously reducing the dimensionality of the lifted representation.
-
Robust control synthesis: The framework allows for designing controllers that explicitly account for both approximation errors in the learned dynamics and ambiguity in disturbance distribution by formulating a distributionally robust covariance steering problem.
-
Handling complex system dynamics: The system can model nonlinear, control-affine systems and stochastic nonlinear systems using lifted state spaces and EDMD with control (EDMDc) to approximate the Koopman generator.
-
Incorporating stability guarantees: The training objective function can be augmented with a stability penalty term, Lstab = lambda stab ReLU(rho(A) - (1 - epsilon)) squared, promoting desirable model properties like stability during DNN training.
-
State constraint satisfaction: The system can enforce state constraints using distributionally robust Conditional Value at Risk (DR-CVaR) constraints, ensuring that the probability of constraint violation remains below a specified risk level delta j,k.
-
Tractable optimization via Semi-Definite Programming (SDP): The entire distribution steering problem is reformulated as a convex Semi-Definite Program (SDP), allowing for solution by off-the-shelf solvers, which is essential for high-dimensional lifted states.
-
Guaranteed terminal performance: The framework provides sufficient conditions to guarantee that the final state distribution belongs to the target ambiguity set by controlling the terminal covariance and radius constraints through LMI reformulations.
Abstract
This paper studies the problem of data-driven covariance steering for stochastic, nonlinear dynamical systems. We utilize Koopman operator theory to learn, directly from data, an approximate linear representation of the nonlinear dynamics in a lifted space of observable functions of the state. By exploiting this linear model, we design controllers for the mean and covariance of the underlying nonlinear system by utilizing linear distribution steering algorithms. To account for approximation errors in the learned dynamics and the ambiguity in the disturbance distribution induced in the lifted space, we formulate the problem as a distributionally robust covariance steering problem. Furthermore, we leverage a deep neural network (DNN) to simultaneously learn the Koopman observables and the linear model parameters, thus eliminating the need for manual observables basis selection while also reducing the dimensionality of the lifted representation. Additionally, we can promote desirable model properties, such as stability, during DNN training. We demonstrate the effectiveness of the proposed framework via numerical simulations.
Sources
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