pqSEDMD: Subspace Methods for Extended Dynamic Mode Decomposition Identification
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Mathematics. The day in mathematics: new theorems, new proofs, and what they open.
Ada: We're looking at the paper: "pqSEDMD: Subspace Methods for Extended Dynamic Mode Decomposition Identification".
Title and authors: Ada: So we're looking at this paper today called "pqSEDMD: Subspace Methods for Extended Dynamic Mode Decomposition Identification". It looks like they're trying to mix two different ideas for finding nonlinear dynamics, which are subspace identification methods and extended DMD.
Emmy: That's right. They call it pqSEDMD because it combines the strengths of these two approaches, trying to get a more stable way to approximate nonlinear systems in a linear function space than what we have now.
Kolya: So how does this actually help us when we're dealing with real data? Is it just about making the math look prettier, or does it really handle that noise we always see?
Ada: It's definitely about handling that noise more robustly. The paper introduces a p-q quasi-norm reduction on an orthogonal polynomial basis to do this, which lets them identify nonlinear dynamics using the extended DMD algorithm within this lifted function space.
Emmy: That lifting of the data into that function space is key because it allows them to treat the nonlinear dynamics as if they were linear ones, which subspace methods are really good at doing.
Kolya: So what does that mean for the practical application? Does it just give us a better fit to our data, or does it open up a new way to look at these kinds of systems?
Ada: It increases the set of nonlinear dynamical systems that you can actually identify just by looking at your data, which is pretty significant. They're aiming for better identification with less training data.
Emmy: And they also boost the numerical stability of the extended DMD methods themselves and how those methods approximate what we call the Koopman operator, which is a big deal in this field.
Kolya: That sounds promising, but I always wonder about the practical limits. What are these conditions they set up to make sure this whole lifting process actually works without breaking?
Ada: They lay out three necessary conditions for deterministic lifted subspace identification of autonomous systems to guarantee that the column spaces end up being equal. First, they need the number of Hankel blocks to be greater than the dimension of the function space of the state.
Emmy: And second, they have a condition where the number of columns in each Hankel matrix is greater than some quantity related to those blocks. That's about ensuring enough data points are involved for each mode you're trying to find.
Kolya: And third, they need the sequences of lifted states to have full row rank equal to that dimension of the function space. What does that condition tell us about the quality of our input signals?
Title and authors: Ada: Lemma seven talks about this, and if you have a controllable lifted system and your inputs are persistently exciting enough, then the rank of those lifted state sequences will be equal to that dimension plus something related to the number of blocks.
Emmy: So it's basically saying that if we feed the system enough different kinds of input data, we can guarantee that our identification is accurate for those systems. That moves us toward a more reliable way to model things than just relying on one method or another.
Kolya: Okay, so the paper suggests an alternative way to handle multiple experiments by using extended observability matrices to get these lifted state sequences. How does that specifically help when we have several different runs?
Ada: They propose defining future and past lifted state sequences based on the application of some extended observability matrix to each of the output matrices. These are defined as the sequence of future lifted states, which is calculated using some function phi applied to the initial state.
Emmy: And then they have a corresponding past lifted state sequence, which involves applying that same function to zero and some input data up to a certain point. It's like they’re creating two linked sequences for each experiment.
Kolya: So this leads into their prediction part, right? If you only have one initial output measurement, how do we actually get a long-term prediction from this setup?
Ada: If you only have the initial output measurement, the mapping from that to the initial condition of the lifted state is defined by some equation involving a matrix C and S and then psi applied to the measured output.
Emmy: And then they evolve that lifted state one step ahead using a specific transition mapping, which is related to something called function phi. It's like propagating the state forward based on what we know about the system.
Kolya: Then they project that evolved state back down to the original output space using another mapping and recover the output. That whole loop has to repeat for every single step you want a long-term prediction.
Ada: Exactly, it’s a step by step procedure: lift the initial output, map it to the lifted state, evolve one step forward, project back to the original output space, and then repeat that process for as many steps as you need.
Title and authors: Emmy: It's a way of ensuring that even though they only have an approximation for a finite horizon because of using finite observables, they can correct the lifted output at every single step to keep it accurate over time.
Kolya: So what's the big picture here for someone who just wants to use this? Is it going to be plug-and-play for any nonlinear system we encounter?
Ada: The paper suggests that by combining these subspace methods with the extended DMD algorithm using a p-q quasi-norm reduction on an orthogonal polynomial basis, you get a computational stable way to model nonlinear systems from data.
Emmy: The main implication is that this pqSEDMD method provides robust approximations of nonlinear systems in a linear function space, which significantly increases the numerical robustness of extended DMD methods and their approximations of the Koopman operator.
Kolya: It sounds like it’s a solid tool for complex systems where noise and nonlinearity are both present, but I always wonder about its limitations with respect to system complexity. Where does it start failing?
Ada: The paper mentions that because neither standard extended DMD nor this subspace extension spans an invariant subspace of the dynamical system, they don't actually span the Koopman operator in general. So their approximations are only valid for a finite horizon because of that constraint.
Emmy: That means for truly long term predictions, you have to use that specific prediction procedure we just talked about—lifting, mapping to state, evolving one step at a time, projecting back down—and doing it repeatedly.
Kolya: So the main thing is that it makes the identification process more stable and capable of handling noise better than before, but it still has that finite horizon problem for prediction without constant correction. That’s a real caveat.
Ada: Exactly. The pqSEDMD algorithm combines these advantages in a computationally stable algorithm for data-driven modeling of dynamical systems, offering a way to get reliable results when you're dealing with noisy and limited data.
Emmy: It's about getting that robust approximation of nonlinear dynamics into a linear function space so we can use the strengths of subspace methods without losing the nonlinearity.
Kolya: So yeah, it’s a lot of machinery, but it seems like they’ve built a method that handles uncertainty and nonlinearity in a way that’s computationally stable for data-driven modeling.
Ada: It certainly seems like they've put together something solid here to make those complex systems more tractable through these subspace methods.
Emmy: And it opens up new avenues for how we can use these tools to study things that are inherently nonlinear but might be noisy in the real world.
The paper's summary: Ada: So, we’re looking at how this pqSEDMD algorithm combines subspace methods with extended DMD to find nonlinear dynamics in a linear function space for better data handling and longer predictions.
Emmy: It basically takes the strengths of both worlds. Subspace methods are good at handling noise and finding linear patterns, while extended DMD lets you look at nonlinear systems through observables.
Ada: Exactly. The core idea is lifting the data into this higher-dimensional function space so that we can identify those nonlinear dynamics using a linear operator. That’s the central trick here.
Emmy: And they use something called a p-q quasi norm reduction on an orthogonal polynomial basis to make sure this approximation of the nonlinear system is really robust and stable.
Ada: It means when you’re dealing with noisy, limited data, this method gives you a much more reliable way to approximate what the system is doing. The results show it increases the set of nonlinear systems that can be identified just from data.
Emmy: And it improves how well extended DMD approximates the Koopman operator, which is that bridge between nonlinear dynamics and linear operators. That’s a big structural improvement for modeling these things.
Ada: But there are still caveats. The authors laid out some necessary conditions—like the number of Hankel blocks having to be bigger than the dimension of the state space—to make sure this identification actually works mathematically, especially for deterministic autonomous systems.
Emmy: And they address how to handle multiple experiments too. Instead of just treating each run as a separate past and future sequence, they use an extended observability matrix to define linked sequences for each experiment. That helps train the AI on multiple data points from the same system.
Ada: The prediction part is clever because it has a specific six-step procedure if you only have one initial output measurement. You lift that output, map it to the state sequence, evolve one step forward, project back to the output space, and repeat that whole loop.
Emmy: That repeated projection step is what allows them to correct the lifted output at every single time step so the long-term prediction stays accurate. It’s a continuous correction process for those finite horizon approximations they have.
Ada: So, while it’s computationally stable and handles noise better than some other methods, it still has that limitation that because the approximation isn't exactly spanning the Koopman operator, you only get accurate predictions for a finite horizon without that step-by-step correction.
Emmy: It means for really long term analysis, you have to stick to their prediction procedure. But overall, it’s a solid algorithm combining subspace identification with extended DMD for more robust nonlinear system modeling from data.
Ada: We’ll be talking about how this lifting works in more detail next.
The paper's improvements: Ada: So, we're looking at how this paper suggests improvements to the pqSEDMD algorithm itself. It’s not just about the initial setup, but about making the whole process more powerful.
Emmy: They are focusing on how to make this identification method even more flexible when dealing with different types of nonlinear systems and data structures.
Ada: The main suggestion is really about extending the applicability beyond just deterministic autonomous systems to cover a wider range of dynamics, which is a big win for real-world modeling.
Emmy: They also look at how to handle the data matrices differently so that you don't have to rely on that strict concatenation method we talked about earlier, instead carrying individual matrices and joining them when needed.
Ada: That’s important because it makes the training process much more adaptable for different experimental setups. It moves away from a single monolithic data matrix approach to something more modular.
Emmy: And they are pushing the numerical robustness even further by exploring how the p-q quasi norm reduction can be used in different ways to get better approximations of that Koopman operator. That means fewer artifacts in the final result.
Ada: It seems like they are trying to build a framework that is less dependent on having perfect, perfectly structured input data, which is where most of these identification methods struggle.
Emmy: And they also look at how to formalize the lifted state sequences even more cleanly using that extended observability matrix approach we mentioned before. It makes the connection between the output and those hidden state variables clearer.
Ada: This whole direction points toward a more generalized method for data-driven modeling, one that can handle noise and nonlinearity in a way that’s computationally stable across different system types.
Emmy: And they are setting up the groundwork for how this approach might be used to build even more complex models where you have multiple interacting subsystems.
Ada: It suggests that as we push AI into systems science, we need methods like this that don't just work for one specific type of system but can scale up.
Emmy: So what they’re hinting at is a more generalized toolbox for identifying nonlinear dynamics from data, rather than just a single algorithm with some tweaks.
Ada: We'll be diving into the practical implications of these improvements next, especially how this might affect control engineering applications.
Conclusion: Ada: So we're wrapping up this look at pqSEDMD: Subspace Methods for Extended Dynamic Mode Decomposition Identification, summarizing how this algorithm combines subspace methods with extended DMD for robust nonlinear dynamics modeling from noisy data.
Emmy: It boils down to using a p-q quasi norm reduction on an orthogonal polynomial basis to lift the data into a linear space where we can identify nonlinear systems more reliably.
Ada: The main thing is that it provides a computationally stable way to approximate those complex nonlinear systems, and it significantly boosts the numerical robustness of extended DMD methods and their approximations of the Koopman operator.
Emmy: It means we get better results when our data is messy, which is what most real-world science deals with. The authors show that they can increase the set of nonlinear dynamical systems you can identify just from your raw data.
Ada: Kolya, what’s your take on the practical application? Does this method really survive being applied to systems where you have a lot of experimental noise and limited training time?
Kolya: Well, the paper sets up some necessary conditions for deterministic autonomous systems to work, like making sure the number of Hankel blocks is bigger than the dimension of your state space. That gives us a concrete bound on what’s required for this identification to actually be possible.
Emmy: And they also showed that they can handle multiple experiments by using extended observability matrices to define linked past and future state sequences for each run, which helps when you have several data sets.
Ada: The prediction procedure is a bit of a workout, but it’s structured: lift the initial output, map it to the state sequence, evolve one step ahead using a specific transition mapping, project back to the output space, and repeat that loop for every step you want.
Kolya: That repeated projection corrects for those finite horizon approximations we discussed; it keeps the prediction accurate over time by constantly adjusting based on what it knows. The constants in their analysis seem to survive the application provided those rank conditions are met.
Emmy: It's a robust framework because it addresses both the nonlinearity and the noise through that function space lifting, which is a really elegant structural solution to the problem of modeling nonlinear dynamics linearly.
Ada: So, while there’s this finite horizon limitation for prediction without constant correction, pqSEDMD gives us a stable starting point for data-driven models of complex systems.
Kolya: It means we have a more reliable tool to use when we want to extract insights from noisy experimental data without having to clean everything up perfectly first.
Ada: Exactly. So that’s the summary of pqSEDMD: Subspace Methods for Extended Dynamic Mode Decomposition Identification, combining subspace methods with extended DMD for robust nonlinear system modeling.
Emmy: We’ll be looking at how these ideas connect to control engineering applications next.
Camilo Garcia-Tenorio, Alan Vande Wouwer
Electrical Power Engineering Unit, University of Mons · Systems, Estimation, Control and Optimization, University of Mons
math.OC
Submitted: 2026-09-10
Updated: 2026-11-17
Comments: 14 pages, 4 figures
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 73/100
The gist: The pqSEDMD algorithm introduces an algorithm that uses subspace identification methods to identify nonlinear dynamics using the Extended Dynamic Mode Decomposition (EDMD) algorithm with a p-q-quasi…
Key concepts
- Subspace Identification Methods
- These methods are robust techniques used to identify linear dynamics from data by projecting Hankel matrices onto subspaces. They are effective at handling process and measurement noise, providing a stable foundation for approximating nonlinear systems.
- Extended Dynamic Mode Decomposition (EDMD)
- EDMD is a broader method capable of handling nonlinear systems by using observable functions. It allows for the identification of dynamics that standard linear methods cannot capture, extending the scope beyond simple linear models.
- Function Space Lifting
- This involves transforming raw data into a higher-dimensional function space where nonlinear dynamics can be approximated by a linear operator. This transformation is key to using subspace methods effectively for nonlinear system identification.
- p-q-quasi norm reduction on an orthogonal polynomial basis
- This technique uses a specific mathematical framework based on orthogonal polynomials to reduce the complexity of the analysis. It helps create a more robust approximation of nonlinear systems within the linear function space.
Terminology
Summary
The pqSEDMD algorithm introduces an algorithm that uses subspace identification methods to identify nonlinear dynamics using the Extended Dynamic Mode Decomposition (EDMD) algorithm with a p-q-quasi norm reduction on an orthogonal polynomial basis, resulting in a robust approximation of nonlinear systems in a linear function space. p. 1 This paper addresses the former
question by introducing the p-q quasi-norm Subspace
Extended Dynamic Mode Decomposition (pqSEDMD). This method explores the use of subspace methods to
identify nonlinear dynamics using the EDMD algorithm
with a p-q-quasi norm reduction of an orthogonal polynomial basis, referred to as pqEDMD (Garcia-Tenorio
and Vande Wouwer, 2022).Where the paper says it
Pruning redundant reasoning tokens
The paper introduces the pqSEDMD, an algorithm that combines the complementary strengths of two identification paradigms. Adapting the subspace identification methods to operate in a lifted function space achieves a robust approximation of nonlinear dynamics from noisy and limited data. p. 6 Lifting the data into the function space enables the identification of nonlinear dynamics with a linear operator. p. 6 3 The pqSEDMD
This section shows how the shortcomings of hte two previous algorithms are alleviated via their combination. Lifting the data into the function space enables the identification of nonlinear dynamics with a linear operator. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. Subspace identification methods, are robust tools able to provide accurate state-space models for multi-variable linear
systems directly from input-output data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. Subspace identification methods, are robust tools able to provide accurate state-space models for multi-variable linear
systems directly from input-output data.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
How the training signal is built
For the pqSEDMD algorithm, there will be no concatenation of the data matrices before the lifting or Hankel transformation, instead, we carry the individual matrices and concatenate when necessary. Different from the pqEDMD, there will be no concatenation of the data matrices before the lifting or Hankel transformation, instead, we carry the individual matrices and concatenate when necessary. p. 6 As with the pqEDMD, we will
not apply any lifting to the input. Different from the
pqEDMD, there will be no concatenation of the data
matrices before the lifting or Hankel transformation, instead, we carry the individual matrices and concatenate
when necessary.Where the paper says it
**)&Ψi0bf +bp−1 ∈ R(bf +bp)d×jiE i=1, where the Ψi0bH−1 denotes the Hankel matrix with bH = bp + bf blocks of the lifted matrix Y i, i.e., Ψ(Y i). The set of input signals follows a similar approach to (6), where instead of 2i Hankel blocks, there are bp+bf in each matrix of the set. p. 6 Ψi
f
. (32)
For the set of input signals, we follow a similar approach
to (6), where instead of 2i Hankel blocks, there are bp+bf
in each matrix of the set,
{
U
i
0|bp+bf −1 ∈ R
(bp+bf )m×ji
}E
i=1
.Where the paper says it
WiΨp ≜ [U i0bp−1 Ψi0bp−1] ∈ R bp(d+m)×ji E i=1.
Where the gains come from
The benefits of combining the two methods are: increasing the set of nonlinear dynamical system that can be identified just with data, and increasing the numerical robustness of EDMD methods and their approximations of the Koopman operator. The method explores the use of subspace methods to identify nonlinear dynamics using the EDMD algorithm with a p-q-quasi norm reduction of an orthogonal polynomial basis, referred to as pqEDMD. p. 1 This paper addresses the former
question by introducing the p-q quasi-norm Subspace
Extended Dynamic Mode Decomposition (pqSEDMD). This method explores the use of subspace methods to
identify nonlinear dynamics using the EDMD algorithm
with a p-q-quasi norm reduction of an orthogonal polynomial basis, referred to as pqEDMD (Garcia-Tenorio
and Vande Wouwer, 2022).Where the paper says it
Lifted rank conditions and persistence of excitation
For the case of deterministic lifted-subspace identification of autonomous systems, there are three necessary conditions to guarantee that the columns spaces are equal. These conditions are:
-
The number of Hankel Blocks is greater than the dimension of the function space of the state, i. p. 9
e., bbH > ds. p. 9Where the paper says it
To guarantee that the columns spaces are equal, i. , col(ΨY ) = col(ΓbH ), the following conditions must be met: (i) The number of Hankel Blocks is greater than the dimension of the function space of the state, i.
Where the paper says it
, col(ΨY ) = col(ΓbH ), the following conditions must be met: (i) The number of Hankel Blocks is greater than the dimension of the function space of the state, i. , bbH > ds.
-
The number of columns in each of the Hankel matrix of outputs is greater then the number of blocks, i. p. 4
e., ji > HbH i = 0E. p. 7Where the paper says it
To conclude that in fact there is an equality, the following conditions must be met: (1) The number of of Hankel blocks is greater than the order of the system, i. (2) The number of columns in the Hankel matrix of outputs is greater than the number of Hankel blocks, i.
Where the paper says it
(35) For each available experiment, the input-output relationship in the function space follows from the system solution (35), the lifted output equation (34b), and the lifted Hankel matrices (32) with bH block rows. Define the extended observability matrix as ΓbH ∈ R bHd×ns and the Markov parameter matrix as HbH ∈ R bHd×bHm similar to (8) and (9) respectively, but with respect to AS, BS, and CS.
-
The sequences of lifted states Ξ i E i=1 has full row rank ds. p. 9
Where the paper says it
, {ji > HbH }i = 0E. Finally, (iii) The sequences of lifted states {Ξ i} E i=1 has full row rank ds.
Lemma 7 (Lifted rank condition) Consider a controllable lifted system (34) and a set of input-to-lifted-state trajectories, from zero initial conditions. p. 5 For a state sequence X0 : T −2i and the corresponding
Hankel matrix of inputs U0|2i−1, the rank condition to
satisfy is:
Lemma 4 (Willems et al. , 2004; Verhaegen and Verdult, 2007) Consider a controllable system (4), and an
input-state trajectory (x, u) = {xk; uk}
T −1
k=0 ∈ R
(n+m)×T
from zero initial conditions. If the input sequence u0|T −1 is is persistently exciting of order
n + 2im. Then,
(i) Any 2i-length trajectory (u0|2i−1, y0|2i−1) of the linear system (4) can be written as a linear combination of of the columns of the combined Hankel matrix of inputs and outputs,
[
U0|2i−1
Y0|2i−1
]
(27)
(ii) Any linear combination of (27) is a 2i length trajectory of the system.Where the paper says it
Where the paper says it
Lifted state sequence and solution
For the sequences of lifted states, we propose an alternative algorithm to (21-24) that handles multiple experiments. The set of lifted state sequences is the application of (21) to each of the Obf matrices using the extended observability matrix. p. 8 In addition, we replace equations (22 and 23) for the calculation of the shifted-lifted
states for a past and future division of the lifted state
sequence like the pqEDMD formulation (29). The set of lifted state sequences is the application
of (21) to each of the Obf matrices using the extended
observability matrix (8), {Ξˆ
0|Ti−bH = Γ†
bf
Oi
bf
}
E
i=1,
where each of the projection matrices {Oi
bf
}
E
i=1 is the
application of (42 or 45) depending on the system
(stochastic/deterministic and forced/autonomous).Where the paper says it
**)&Ξˆi+ = [Φˆi(x(1)) · · · Φˆi(x(Ti − bf))]E i=1. **
**)&Ξˆi− = [Φˆi(x(0)) · · · ˆξ i(Ti − bf − 1)]E i=1. p. 8 The set of lifted state sequences is the application
of (21) to each of the Obf matrices using the extended
observability matrix (8), {Ξˆ
0|Ti−bH = Γ†
bf
Oi
bf
}
E
i=1,
where each of the projection matrices {Oi
bf
}
E
i=1 is the
application of (42 or 45) depending on the system
(stochastic/deterministic and forced/autonomous). From the approximation of the lifted state sequences,
define the future and past lifted state sequences and the
corresponding input and lifted output matrices as,
{
Ξˆi
+ =
[
Φˆi
(x(1)) · · · Φˆi
(x(Ti − bf ))]}E
i=1
(50a)
{
Ξˆi
− =
[
Φˆi
(x(0)) · · · ˆξ
i
(Ti − bf − 1)]}E
i=1
, (50b)
{
U
i
− =
[
u
i
bp|bp
(0) · · · u
i
bp|bp
(Ti − bf − 1)]}E
i=1
(50c)
{
Ψ
i
− =
[
Ψi
bp|bp
(0) · · · Ψi
bp|bp
(Ti − bf − 1)]}E
i=1
(50d)
where each U
i
− is the first block of the future Hankel
matrix of inputs sliced from the first element until the
9Where the paper says it
The formulation of the least squares problem follows from the concatenation of the past and future matrices. p. 9 Then, the fomulation of the leasosquares
problem follow from the concatenation of the past and
future matrices,
Ξˆ+ =
[
Ξˆ1
+ · · · ΞˆE
+
]
(51a)
Ξˆ− =
[
Ξˆ1
− · · · ΞˆE
−
]
(51b)
U− =
[
U
1
− · · · U
E
−
]
(51c)
Ψ
Y
− =
[
Ψ
Y1
− · · · Ψ
YE
−
]
. (51d)
With the two sequences of lifted states, and the first
block of the data and input Hankel matrices properly
sliced, the least squares problem for the calculation of
the AS operator, and the system matrices BS and CS is,
Ξˆ+ =
[
AS BS
]
[
Ξˆ−
U−
]
+ r[AS BS ] (52)
Ψ
Y
− = CSΞˆ+ + rCS
, (53)
where r[AS BS ] and rCS
are the residual term to minimize.Where the paper says it
Prediction
Assuming that only the initial output y(0) is available, the required mapping between an initial measured output, and the corresponding initial condition of the lifted state ξ(0) is, xi(k) = C†SΨ(y(k)). It is necessary to have a relationship between the lifted output and the lifted state because in general, neither the standard EDMD nor this subspace extension span an invariant subspace of the dynamical system; they do not span the Koopman operator. p. 9 Assuming that only the initial output y(0) is
available, the required mapping between an initial measured output, and the corresponding initial condition of
the lifted state ξ(0) is,
ξ(k) = C
†
SΨ(y(k)), (54)
for an arbitrary time instant k. It is necessary to have
a relationship between the lifted output and the lifted
state because in general, neither the standard EDMD
nor this subspace extension span an invariant subspace
of the dynamical system; they do not span the Koopman operator. In practice, this is generally
not the case (none of the examples in this paper span a
Koopman sybspace, despite the accuracy of the approximations). With a finite set of observables, the approximations are only valid for a finite horizon. With a finite set of observables, the approximations are only valid for a finite horizon. To have accurate long term predictions, it is necessary to correct the
lifted output at every step of the trajectory (Junker et
al. To have accurate long term predictions, it is necessary to correct the
lifted output at every step of the trajectory (Junker et
al. Accordingly, the prediction for the pqSEDMD
form an arbitrary initial output proceeds as follows: 1)
lift the initial output with Ψ. Accordingly, the prediction for the pqSEDMD
form an arbitrary initial output proceeds as follows: 1)
lift the initial output with Ψ. 2) map it to the lifted state
via (54). 2) map it to the lifted state
via (54). 3) evolve one step ahead via (34a). 3) evolve one step ahead via (34a). 4) project
back to the output space via (34b). 4) project
back to the output space via (34b). 5) recover the output
via (34c). 5) recover the output
via (34c). 6) repeat from step 1 for the required number
of steps.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The pqSEDMD combines these advantages in a computationally stable algorithm for data-driven modeling of dynamical systems. p. 0 Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al.Where the paper says it
References
Baddoo, Peter J, Benjamin Herrmann, Beverley J McKeon, J Nathan Kutz and Steven L Brunton (2023). Physics-informed dynamic mode decomposition. p. 0 pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise.Where the paper says it
Bohon, M. D., A. Orchini, R. Bluemner, C. O. Paschereit and E. J. Gutmark (2020). p. 0 , 2018; Bohon et al. ,
2020) and control (Proctor et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al.Where the paper says it
Where the paper says it
Brunton, Steven L, Joshua L Proctor and J Nathan Kutz (2016). p. 0 The
SINDy framework by Brunton et al. (2016) has similarly
demonstrated the potential of dictionary-based methods for discovering governing equations from data. (2016) has similarly
demonstrated the potential of dictionary-based methods for discovering governing equations from data. For
the subject of non-autonomous system identification,
Haseli and Cortés (2026) provide a universal form showing the equivalence between considering the operator
dynamics as a linear parameter varying system parameterized by the input (Williams et al.Where the paper says it
Where the paper says it
De Persis, Claudio and Pietro Tesi (2020). p. 1 Their success in linear system identification raises a interesting question: Is it possible to
combine the robustness of subspace methods with the
EDMD’s ability to capture nonlinear dynamics in a lifted
function space? De Persis and Tesi (2020) formulate a similar question
when dealing with data-driven control methods. Their success in linear system identification raises a interesting question: Is it possible to
combine the robustness of subspace methods with the
EDMD’s ability to capture nonlinear dynamics in a lifted
function space? De Persis and Tesi (2020) formulate a similar question
when dealing with data-driven control methods. Their success in linear system identification raises a interesting question: Is it possible to
combine the robustness of subspace methods with the
EDMD’s ability to capture nonlinear dynamics in a lifted
function space? De Persis and Tesi (2020) formulate a similar question
when dealing with data-driven control methods.Where the paper says it
Where the paper says it
Where the paper says it
Garcia-Tenorio, Camilo and Alain Vande Wouwer (2022). p. 0 pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. Combining these two methods, we introduce the p-q quasi-norm Subspace Extended
Dynamic Mode Decomposition (pqSEDMD). An algorithm that uses our previous improvements to the EDMD by the use of
a p-q-quasi-norm reduction on an orthogonal polynomial basis, the pqEDMD algorithm, along with subspace identification
methods.Where the paper says it
Where the paper says it
Garcia-Tenorio, Camilo and Alain Vande Wouwer (2025). p. 0 pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. 1, the pqEDMD algorithm in 2. 2, and the concepts from data driven control
that support our claims. 1, the pqEDMD algorithm in 2. 2, and the concepts from data driven control
that support our claims.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Giesl, Peter and Sigurdur Hafstein (2015). p. 2 ,
linearized versions of (2) that take the form:
x(k + 1) = Ax(k) + Bu(k) + w(k) (4a)
y(k) = Cx(k) + v(k) (4b)
where A ∈ R
n×n is the state transition matrix that governs the evolution of the unknown state, B ∈ R
m×n
is the input-to-state transition matrix, and C ∈ R
n×l
is the output matrix. The linear approximation of the
dynamics assumes that the input-output data lies in
the vicinity of a hyperbolic fixed point of the nonlinear
system (2) that satisfies the Hartman-Grobman theorem (Giesl and Hafstein, 2015). ,
linearized versions of (2) that take the form:
x(k + 1) = Ax(k) + Bu(k) + w(k) (4a)
y(k) = Cx(k) + v(k) (4b)
where A ∈ R
n×n is the state transition matrix that governs the evolution of the unknown state, B ∈ R
m×n
is the input-to-state transition matrix, and C ∈ R
n×l
is the output matrix. The linear approximation of the
dynamics assumes that the input-output data lies in
the vicinity of a hyperbolic fixed point of the nonlinear
system (2) that satisfies the Hartman-Grobman theorem (Giesl and Hafstein, 2015). The expectation of the process and
output noise is,
E
[(w(i)
v(i)
)
(
w
T
(j) v
T
(j)
)
]
=
(
Q S
S
T R
)
δij ≥ 0. The solution of (2) produces the sequences of states
x = {xk}
T −1
k=0 ∈ R
n×T and outputs y = {yk}
T −1
k=0 ∈
R
l×T
that come from the successive application of the
state transition mapping from an initial condition x(0)
and the application of a known sequence of inputs u =
{uk}
T −1
k=0 ∈ R
m×T according to the input and output
functions. The expectation of the process and
output noise is,
E
[(w(i)
v(i)
)
(
w
T
(j) v
T
(j)
)
]
=
(
Q S
S
T R
)
δij ≥ 0. The solution of (2) produces the sequences of states
x = {xk}
T −1
k=0 ∈ R
n×T and outputs y = {yk}
T −1
k=0 ∈
R
l×T
that come from the successive application of the
state transition mapping from an initial condition x(0)
and the application of a known sequence of inputs u =
{uk}
T −1
k=0 ∈ R
m×T according to the input and output
functions.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Golub, Gene H and Charles F Van Loan (2013). p. 1 Mainly, when there is uncertainty
in the system dynamics and measurement noise. Subspace identification methods (Van Overschee and
De Moor, 1996; Verhaegen and Verdult, 2007) provide
a robust framework for the identification of uncertain,
linear, and time-invariant systems. Mainly, when there is uncertainty
in the system dynamics and measurement noise. Subspace identification methods (Van Overschee and
De Moor, 1996; Verhaegen and Verdult, 2007) provide
a robust framework for the identification of uncertain,
linear, and time-invariant systems. , the available measurements of
the system). The system evolves according to the nonlinear state transition mapping f : R
n → R
n, an inputto-state transition mappig g : R
m → R
n, and the process noise w(k) ∼ N (0, Q), a zero-mean white noise sequence with covariance matrix Q ∈ R
n×n. , the available measurements of
the system). The system evolves according to the nonlinear state transition mapping f : R
n → R
n, an inputto-state transition mappig g : R
m → R
n, and the process noise w(k) ∼ N (0, Q), a zero-mean white noise sequence with covariance matrix Q ∈ R
n×n.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
H. Tu, Jonathan, Clarence W. Rowley, Dirk M. p. 0 Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Despite its limitations to capture nonlinear behavior, it has found widespread applications for analysis (H. , 2014; Towne et al. Despite its limitations to capture nonlinear behavior, it has found widespread applications for analysis (H. , 2014; Towne et al. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Despite its limitations to capture nonlinear behavior, it has found widespread applications for analysis (H. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Despite its limitations to capture nonlinear behavior, it has found widespread applications for analysis (H.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Haseli, Masih and Jorge Cortés (2026). p. 0 (2016) has similarly
demonstrated the potential of dictionary-based methods for discovering governing equations from data. For
the subject of non-autonomous system identification,
Haseli and Cortés (2026) provide a universal form showing the equivalence between considering the operator
dynamics as a linear parameter varying system parameterized by the input (Williams et al. (2016) has similarly
demonstrated the potential of dictionary-based methods for discovering governing equations from data. For
the subject of non-autonomous system identification,
Haseli and Cortés (2026) provide a universal form showing the equivalence between considering the operator
dynamics as a linear parameter varying system parameterized by the input (Williams et al. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Junker, Annika, Julia Timmermann and Ansgar Trächtler (2022). p. 9 With a finite set of observables, the approximations are only valid for a finite horizon. To have accurate long term predictions, it is necessary to correct the
lifted output at every step of the trajectory (Junker et
al. With a finite set of observables, the approximations are only valid for a finite horizon. To have accurate long term predictions, it is necessary to correct the
lifted output at every step of the trajectory (Junker et
al. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Klus, Stefan, Feliks Nuske, Peter Koltai, Hao Wu, Ioannis Kevrekidis, Christof Schütte and Frank Noé (2018). p. 0 The Koopman framework has since attracted substantial research interest: Klus et al. (2018) extended the
framework to stochastic systems; and Baddoo et al. This method explores the use of subspace methods to
identify nonlinear dynamics using the EDMD algorithm
with a p-q-quasi norm reduction of an orthogonal polynomial basis, referred to as pqEDMD (Garcia-Tenorio
and Vande Wouwer, 2022). The benefits of combining
the two methods are: increasing the set of nonlinear dynamical system that can be identified just with data, and
increasing the numerical robustness of EDMD methods
and their approximations of the Koopman operator 1
.Where the paper says it
Where the paper says it
Koopman, Bernard O (1931). p. 0 To directly address the linear limitation of DMD Williams et al. (2015),
introduced the extended DMD (EDMD), an approximation of nonlinear system dynamics in a linear function space of observables, the so-called, Koopman operator (Koopman, 1931). To directly address the linear limitation of DMD Williams et al. (2015),
introduced the extended DMD (EDMD), an approximation of nonlinear system dynamics in a linear function space of observables, the so-called, Koopman operator (Koopman, 1931). Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. Subspace identification methods, are robust tools able to provide accurate state-space models for multi-variable linear
systems directly from input-output data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. Subspace identification methods, are robust tools able to provide accurate state-space models for multi-variable linear
systems directly from input-output data.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Korda, Milan and Igor Mezić (2018a). p. 0 , 2016), and as a
bilinear form (Proctor et al. , 2016; Korda and Mezić,
2018a). , 2016), and as a
bilinear form (Proctor et al. , 2016; Korda and Mezić,
2018a). pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Korda, Milan and Igor Mezić (2018b). p. 0 (2015),
introduced the extended DMD (EDMD), an approximation of nonlinear system dynamics in a linear function space of observables, the so-called, Koopman operator (Koopman, 1931). Korda and Mezić provide the
convergence guarantees of the EDMD to the Koopman
operator at the limit of data and observables (Korda
and Mezić, 2018b). (2015),
introduced the extended DMD (EDMD), an approximation of nonlinear system dynamics in a linear function space of observables, the so-called, Koopman operator (Koopman, 1931). Korda and Mezić provide the
convergence guarantees of the EDMD to the Koopman
operator at the limit of data and observables (Korda
and Mezić, 2018b). pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Markovsky, I. p. 5 If u is persistently exciting
of order n + 2i, then
rank ([X0 : T −2i
U0|2i−1
]) = n + 2im,
then,
rank(Y0|2i−1ΠU⊥ ) = n
and,
col(Y0|2i−1ΠU⊥ ) = col(Γ2i),
Lemma 4 proves that the observability matrix Γn of a
system is identifiable, and the remaining process to get
the system matrices feasible. Finally, the fundamental
lemma is,
Lemma 5 (Markovsky et al. If u is persistently exciting
of order n + 2i, then
rank ([X0 : T −2i
U0|2i−1
]) = n + 2im,
then,
rank(Y0|2i−1ΠU⊥ ) = n
and,
col(Y0|2i−1ΠU⊥ ) = col(Γ2i),
Lemma 4 proves that the observability matrix Γn of a
system is identifiable, and the remaining process to get
the system matrices feasible. Finally, the fundamental
lemma is,
Lemma 5 (Markovsky et al. Also,
recent work by Shang et al. (2024) use the Willems et
al. Also,
recent work by Shang et al. (2024) use the Willems et
al. Finally, the fundamental
lemma is,
Lemma 5 (Markovsky et al. , 2005) Consider a controllable system (4), and an input-output trajectory
(u0|T −1, y0|T −1) = {u0|T −1; } ∈ R
(m+l)×T
. Finally, the fundamental
lemma is,
Lemma 5 (Markovsky et al. , 2005) Consider a controllable system (4), and an input-output trajectory
(u0|T −1, y0|T −1) = {u0|T −1; } ∈ R
(m+l)×T
. In practice, this is generally
not the case (none of the examples in this paper span a
Koopman sybspace, despite the accuracy of the approximations). With a finite set of observables, the approximations are only valid for a finite horizon. In practice, this is generally
not the case (none of the examples in this paper span a
Koopman sybspace, despite the accuracy of the approximations). With a finite set of observables, the approximations are only valid for a finite horizon.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Mauroy, Alexandre, Mezić, Igor and Susuki, Yoshihiko, Eds. p. 0 mate the theoretically robust Koopman operator for the
analysis of nonlinear dynamics. For example, Mauroy et
al. mate the theoretically robust Koopman operator for the
analysis of nonlinear dynamics. For example, Mauroy et
al. 1, the pqEDMD algorithm in 2. 2, and the concepts from data driven control
that support our claims. 1, the pqEDMD algorithm in 2. 2, and the concepts from data driven control
that support our claims.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Otto, Samuel E and Clarence W Rowley (2021). p. 0 (2018) extended the
framework to stochastic systems; and Baddoo et al. (2023) and Otto and Rowley (2021) explored connections to physics-informed and manifold learning. (2018) extended the
framework to stochastic systems; and Baddoo et al. (2023) and Otto and Rowley (2021) explored connections to physics-informed and manifold learning. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Proctor, Joshua L, Steven L Brunton and J Nathan Kutz (2016). p. 0 The
SINDy framework by Brunton et al. (2016) has similarly
demonstrated the potential of dictionary-based methods for discovering governing equations from data. The
SINDy framework by Brunton et al. (2016) has similarly
demonstrated the potential of dictionary-based methods for discovering governing equations from data. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Rowley, Clarence W, Igor Mezić, Shervin Bagheri, Philipp Schlatter and Dan S Henningson (2009). Spectral analysis of nonlinear flows. p. 0 Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al.Where the paper says it
Where the paper says it
Salas, Alvaro and Lorenzo Castillo, Jairo and (2021). p. 1 1 Motivating Example: The Duffing Oscillator
The Duffing equation is the canonical benchmark to
test the accuracy of algorithms and methods related to
the EDMD algorithm and the Koopman operator. It is
a nonlinear second order differential equation that can
describe many types of dynamic behavior in nonlinear
systems (Salas and Castillo, 2021). 1 Motivating Example: The Duffing Oscillator
The Duffing equation is the canonical benchmark to
test the accuracy of algorithms and methods related to
the EDMD algorithm and the Koopman operator. It is
a nonlinear second order differential equation that can
describe many types of dynamic behavior in nonlinear
systems (Salas and Castillo, 2021). The benefits of combining
the two methods are: increasing the set of nonlinear dynamical system that can be identified just with data, and
increasing the numerical robustness of EDMD methods
and their approximations of the Koopman operator 1
. 1 Motivating Example: The Duffing Oscillator
The Duffing equation is the canonical benchmark to
test the accuracy of algorithms and methods related to
the EDMD algorithm and the Koopman operator. The benefits of combining
the two methods are: increasing the set of nonlinear dynamical system that can be identified just with data, and
increasing the numerical robustness of EDMD methods
and their approximations of the Koopman operator 1
. 1 Motivating Example: The Duffing Oscillator
The Duffing equation is the canonical benchmark to
test the accuracy of algorithms and methods related to
the EDMD algorithm and the Koopman operator.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Schmid, Peter J (2010). p. 0 Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise. pqSEDMD: Subspace Methods for Extended Dynamic Mode
Decomposition Identification
Camilo Garcia-Tenorio a
, Alan Vande Wouwer b
aElectrical Power Engineering Unit, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
bSystems, Estimation, Control and Optimization, Université de Mons, 31 Bd Dolez, 7000 Mons, Belgium
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are
powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition
based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement
noise.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Shang, Xu, Jorge Cortés and Yang Zheng (2024). p. 1 Also,
recent work by Shang et al. (2024) use the Willems et
al. Also,
recent work by Shang et al. (2024) use the Willems et
al. (2024) use the Willems et
al. (2004) fundamental lemma to propose guarantees on
the available data for identification under the assumption of the Koopman operator. (2024) use the Willems et
al. (2004) fundamental lemma to propose guarantees on
the available data for identification under the assumption of the Koopman operator.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Towne, Aaron, Oliver T. p. 0 Despite its limitations to capture nonlinear behavior, it has found widespread applications for analysis (H. , 2014; Towne et al. Despite its limitations to capture nonlinear behavior, it has found widespread applications for analysis (H. , 2014; Towne et al. , 2014; Towne et al. , 2018; Bohon et al. , 2014; Towne et al. , 2018; Bohon et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
van Waarde, Henk J. p. 1 (2004) fundamental lemma to propose guarantees on
the available data for identification under the assumption of the Koopman operator. Finally, van Waarde et
al. (2004) fundamental lemma to propose guarantees on
the available data for identification under the assumption of the Koopman operator. Finally, van Waarde et
al. Their success in linear system identification raises a interesting question: Is it possible to
combine the robustness of subspace methods with the
EDMD’s ability to capture nonlinear dynamics in a lifted
function space? De Persis and Tesi (2020) formulate a similar question
when dealing with data-driven control methods. Their success in linear system identification raises a interesting question: Is it possible to
combine the robustness of subspace methods with the
EDMD’s ability to capture nonlinear dynamics in a lifted
function space? De Persis and Tesi (2020) formulate a similar question
when dealing with data-driven control methods. Their success in linear system identification raises a interesting question: Is it possible to
combine the robustness of subspace methods with the
EDMD’s ability to capture nonlinear dynamics in a lifted
function space? De Persis and Tesi (2020) formulate a similar question
when dealing with data-driven control methods. Their success in linear system identification raises a interesting question: Is it possible to
combine the robustness of subspace methods with the
EDMD’s ability to capture nonlinear dynamics in a lifted
function space? De Persis and Tesi (2020) formulate a similar question
when dealing with data-driven control methods. (2024) use the Willems et
al. (2004) fundamental lemma to propose guarantees on
the available data for identification under the assumption of the Koopman operator. (2024) use the Willems et
al. (2004) fundamental lemma to propose guarantees on
the available data for identification under the assumption of the Koopman operator.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Verhaegen, Michel and Vincent Verdult (2007). p. 1 Mainly, when there is uncertainty
in the system dynamics and measurement noise. Subspace identification methods (Van Overschee and
De Moor, 1996; Verhaegen and Verdult, 2007) provide
a robust framework for the identification of uncertain,
linear, and time-invariant systems. Mainly, when there is uncertainty
in the system dynamics and measurement noise. Subspace identification methods (Van Overschee and
De Moor, 1996; Verhaegen and Verdult, 2007) provide
a robust framework for the identification of uncertain,
linear, and time-invariant systems. 4 Lifted system matrices
The method to get the lifted systems matrices is analogous to the linear case from equations ((21)-(24)), where
the objective is to get a sequence of lifted states from the
extended observability matrix and set up a least squares
problem. 1 Order of the lifted subspace
For the subspace identification of linear systems, the
order n of the systems corresponds to the number of
nonzero singular values from the SVD (19). 4 Lifted system matrices
The method to get the lifted systems matrices is analogous to the linear case from equations ((21)-(24)), where
the objective is to get a sequence of lifted states from the
extended observability matrix and set up a least squares
problem. 1 Order of the lifted subspace
For the subspace identification of linear systems, the
order n of the systems corresponds to the number of
nonzero singular values from the SVD (19).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Viberg, Mats (1995). Subspace-based methods for the identification of linear time-invariant systems. p. 1 Subspace identification methods (Van Overschee and
De Moor, 1996; Verhaegen and Verdult, 2007) provide
a robust framework for the identification of uncertain,
linear, and time-invariant systems. These methods construct low-dimensional subspaces capturing dominant
system modes through numerically stable algorithms
based on singular value decomposition (SVD) and QR
factorization. Subspace identification methods (Van Overschee and
De Moor, 1996; Verhaegen and Verdult, 2007) provide
a robust framework for the identification of uncertain,
linear, and time-invariant systems. These methods construct low-dimensional subspaces capturing dominant
system modes through numerically stable algorithms
based on singular value decomposition (SVD) and QR
factorization.Where the paper says it
Where the paper says it
Willems, J. p. 1 Also,
recent work by Shang et al. (2024) use the Willems et
al. Also,
recent work by Shang et al. (2024) use the Willems et
al. If u is persistently exciting
of order n + 2i, then
rank ([X0 : T −2i
U0|2i−1
]) = n + 2im,
then,
rank(Y0|2i−1ΠU⊥ ) = n
and,
col(Y0|2i−1ΠU⊥ ) = col(Γ2i),
Lemma 4 proves that the observability matrix Γn of a
system is identifiable, and the remaining process to get
the system matrices feasible. Finally, the fundamental
lemma is,
Lemma 5 (Markovsky et al. If u is persistently exciting
of order n + 2i, then
rank ([X0 : T −2i
U0|2i−1
]) = n + 2im,
then,
rank(Y0|2i−1ΠU⊥ ) = n
and,
col(Y0|2i−1ΠU⊥ ) = col(Γ2i),
Lemma 4 proves that the observability matrix Γn of a
system is identifiable, and the remaining process to get
the system matrices feasible. Finally, the fundamental
lemma is,
Lemma 5 (Markovsky et al. (2024) use the Willems et
al. (2004) fundamental lemma to propose guarantees on
the available data for identification under the assumption of the Koopman operator. (2024) use the Willems et
al. (2004) fundamental lemma to propose guarantees on
the available data for identification under the assumption of the Koopman operator. Then, an
approximation of the extended observability matrix is
Γˆ
i = W−1
1 U1Σ
1/2
1
, (20)
and a sequence of states that explains the dynamics of
the system is,
Xˆ
i = Γ†
iOi
, (21)
where † denotes the Moore-Penrose pseudo inverse. Note
that these approximations of the extended observability matrix and the approximation of the state sequence
are not equivalent to their definitions in (8) and (11),
i. Then, an
approximation of the extended observability matrix is
Γˆ
i = W−1
1 U1Σ
1/2
1
, (20)
and a sequence of states that explains the dynamics of
the system is,
Xˆ
i = Γ†
iOi
, (21)
where † denotes the Moore-Penrose pseudo inverse. Note
that these approximations of the extended observability matrix and the approximation of the state sequence
are not equivalent to their definitions in (8) and (11),
i.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Williams, Matthew O. p. 0 ,
2020) and control (Proctor et al. To directly address the linear limitation of DMD Williams et al. ,
2020) and control (Proctor et al. To directly address the linear limitation of DMD Williams et al. ,
2020) and control (Proctor et al. To directly address the linear limitation of DMD Williams et al. ,
2020) and control (Proctor et al. To directly address the linear limitation of DMD Williams et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. (2020) also provide an interesting extension to the
linear data-driven formulation when there are multiple
data-trajectories available for the approximation of the
subspace dynamics. This paper addresses the former
question by introducing the p-q quasi-norm Subspace
Extended Dynamic Mode Decomposition (pqSEDMD). (2020) also provide an interesting extension to the
linear data-driven formulation when there are multiple
data-trajectories available for the approximation of the
subspace dynamics. This paper addresses the former
question by introducing the p-q quasi-norm Subspace
Extended Dynamic Mode Decomposition (pqSEDMD).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Williams, Matthew O., Maziar S. Hemati, Scott T.M. Dawson, Ioannis G. Kevrekidis and Clarence W Rowley (2016). p. 0 Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness
of the algorithm and illustrate many important aspects related to the development. Key words: Nonlinear system identification; Extended dynamic mode decomposition; Subspace identification
1 Introduction
The Dynamic Mode Decomposition (DMD), introduced
by Schmid (2010), provides a data-driven framework for
the analysis and prediction of linear dynamical systems
and even some types of nonlinear flows (Rowley et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
pqSEDMD: Subspace Methods for Extended Dynamic Mode Decomposition Identification
De Persis, Claudio and Pietro Tesi (2020). Formulas for data-driven control: Stabilization, optimality, and robustness.
Garcia-Tenorio, Camilo and Alain Vande Wouwer (2025). Application of pqedmd to modelling and control of bioprocesses.
Giesl, Peter and Sigurdur Hafstein (2015). Review on computational methods for Lyapunov functions.
Golub, Gene H and Charles F Van Loan (2013). Matrix Computations.
H. Tu, Jonathan, Clarence W. Rowley, Dirk M. Luchtenburg, Steven L. Brunton and J Nathan Kutz (2014). On dynamic mode decomposition: Theory and applications.
Haseli, Masih and Jorge Cortés (2026). Modeling nonlinear control systems via koopman control family: Universal forms and subspace invariance proximity.
Junker, Annika, Julia Timmermann and Ansgar Trächtler (2022). Data-driven models for control engineering applications using the koopman operator.
Klus, Stefan, Feliks Nuske, Peter Koltai, Hao Wu, Ioannis Kevrekidis, Christof Schütte and Frank Noé (2609.11292p8>.
Koopman, Bernard O (1931). Hamiltonian systems and transformation in hilbert space. Proceedings of the National Academy of Sciences of the United States of America 17(5), 315–318.
Korda, Milan and Igor Mezić (2018a). linear predictors for nonlinear dynamical systems: koopman operator meets model predictive control.
Korda, Milan and Igor Mezić (2018b). On convergence of extended dynamic mode decomposition to the koopman operator.
Markovsky, I., J.C. Willems and B. De Moor (2005). State representations from finite time series.
Mauroy, Alexandre, Mezić, Igor and Susuki, Yoshihiko, Eds. (2020). The Koopman Operator in Systems and Control: Concepts, Methodologies, and Applications.
Otto, Samuel E and Clarence W Rowley (2021). Koopman operators for estimation and control of dynamical systems.
Proctor, Joshua L, Steven L Brunton and J Nathan Kutz (2016). Dynamic mode decomposition with control.
Rowley, Clarence W, Igor Mezić, Shervin Bagheri, Philipp Schlatter and Dan S Henningson (2009). Spectral analysis of nonlinear flows.
Salas, Alvaro and Lorenzo Castillo, Jairo and (2021). The duffing oscillator equation and its applications in physics.
Schmid, Peter J (2010). Dynamic mode decomposition of numerical and experimental data.
Shang, Xu, Jorge Cortés and Yang Zheng (2024). Willems’ fundamental lemma for nonlinear systems with koopman linear embedding.
Towne, Aaron, Oliver T. Schmidt and Tim Colonius (2018). Spectral proper orthogonal decomposition and its relationship to dynamic mode decomposition and resolvent analysis.
van Waarde, Henk J., Claudio De Persis, M. Kanat Camlibel and Pietro Tesi (2020). Willems’ fundamental lemma for state-space systems and its extension to multiple datasets.
Verhaegen, Michel and Vincent Verdult (2007). Filtering and System Identification: A Least Squares Approach.
Viberg, Mats (1995). Subspace-based methods for the identification of linear time-invariant systems.
Willems, J.C., I. Markovsky, P. Rapisarda and B.L.M. De Moor (2004). A note on persistency of excitation.
Williams, Matthew O., Ioannis G Kevrekidis and Clarence W Rowley (2015). A Data-Driven Approximation of the Koopman Operator: Extending Dynamic Mode Decomposition.
Williams, Matthew O., Maziar S. Hemati, Scott T.M. Dawson, Ioannis G. Kevrekidis and Clarence W Rowley (2016). Extending data-driven koopman analysis to actuated systems.
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
-
Robust identification of nonlinear dynamics in a linear function space for long term prediction and analysis enables AI systems to accurately reconstruct system dynamics even when real-world data contains process and measurement noise.
-
The pqSEDMD algorithm increases the set of nonlinear dynamical systems that can be identified just with data by combining subspace identification methods with the extended DMD algorithm using a p-q quasi-norm reduction on an orthogonal polynomial basis.
-
The improved system can provide robust approximations of nonlinear systems in a linear function space, increasing the numerical robustness of EDMD methods and their approximations of the Koopman operator.
-
The system can perform accurate long term predictions by correcting the lifted output at every step of the trajectory through a defined prediction procedure:
lift the initial output with Ψ. map it to the lifted state via (54). evolve one step ahead via (34a). project back to the output space via (34b). recover the output via (34c).
-
The AI system can be trained on multiple experiments, where each trajectory has Ti data-points, by using a data matrix concatenation approach rather than dividing experiments into distinct past and future sequences.
Abstract
The dynamic mode decomposition (DMD), along with its variant for nonlinear systems, the extended DMD (EDMD), are powerful tools for the extraction of meaningful spatio-temporal characteristics of (non)linear dynamical systems from measurement data. Despite some efforts to handle the identification task when dealing with real-world data, the decomposition based methods face a critical challenge: real-world data has two inherent sources of uncertainty, the process and measurement noise. Subspace identification methods, are robust tools able to provide accurate state-space models for multi-variable linear systems directly from input-output data. Combining these two methods, we introduce the p-q quasi-norm Subspace Extended Dynamic Mode Decomposition (pqSEDMD). An algorithm that uses our previous improvements to the EDMD by the use of a p-q-quasi-norm reduction on an orthogonal polynomial basis, the pqEDMD algorithm, along with subspace identification methods. The result is a robust approximation of nonlinear systems in a linear function space, combining the strengths of the two methodologies. Throughout the paper we will use the Duffing oscillator as a benchmark problem to show the effectiveness of the algorithm and illustrate many important aspects related to the development.
Related papers
- An Accelerated Proximal Gradient Method for Bures-Wasserstein Space
- FOMTools: Development and AI-Assisted Code Generation for First-Order Methods
- Lions and Muons: Optimization via Stochastic Frank-Wolfe under Heavy-Tailed Noise
- Dynamic Service Recommendation with Congestion-Dependent Joining: Near-Optimal and Constant-Factor Approximation Algorithms
- Integrating Learning and Optimization for Open Transport Capacity Pooling under Disruption
- Zeroth-Order Gradient Estimators under Decision-Dependent Noise