Data-driven discovery of Koopman

eigenfunctions for control

Eurika Kaiser1∗, J. Nathan Kutz2 , and Steven L. Brunton1
arXiv:1707.01146v4 [math.OC] 10 Feb 2021

1
Department of Mechanical Engineering, University of Washington, Seattle, WA 98195, United States
2
Department of Applied Mathematics, University of Washington, Seattle, WA 98195, United States

February 11, 2021

Abstract
Data-driven transformations that reformulate nonlinear systems in a linear framework have the potential
to enable the prediction, estimation, and control of strongly nonlinear dynamics using linear systems theory.
The Koopman operator has emerged as a principled linear embedding of nonlinear dynamics, and its eigen-
functions establish intrinsic coordinates along which the dynamics behave linearly. Previous studies have used
finite-dimensional approximations of the Koopman operator for model-predictive control approaches. In this
work, we illustrate a fundamental closure issue of this approach and argue that it is beneficial to first validate
eigenfunctions and then construct reduced-order models in these validated eigenfunctions. These coordinates
form a Koopman-invariant subspace by design and, thus, have improved predictive power. We show then how
the control can be formulated directly in these intrinsic coordinates and discuss potential benefits and caveats
of this perspective. The resulting control architecture is termed Koopman Reduced Order Nonlinear Identifica-
tion and Control (KRONIC). It is further demonstrated that these eigenfunctions can be approximated with
data-driven regression and power series expansions, based on the partial differential equation governing the
infinitesimal generator of the Koopman operator. Validating discovered eigenfunctions is crucial and we show
that lightly damped eigenfunctions may be faithfully extracted from EDMD or an implicit formulation. These
lightly damped eigenfunctions are particularly relevant for control, as they correspond to nearly conserved
quantities that are associated with persistent dynamics, such as the Hamiltonian. KRONIC is then demon-
strated on a number of relevant examples, including 1) a nonlinear system with a known linear embedding, 2)
a variety of Hamiltonian systems, and 3) a high-dimensional double-gyre model for ocean mixing.

Keywords: Dynamical systems, nonlinear control, Koopman theory, system identification, machine learning.

1 Introduction
In contrast to linear systems, a generally applicable and scalable framework for the control of nonlinear systems
remains an engineering grand challenge. Improved nonlinear control has the potential to transform our ability
to interact with and manipulate complex systems across broad scientific, technological, and industrial domains.
From turbulence control to brain-machine interfaces, emerging technologies are characterized by high-dimensional,
strongly nonlinear, and multiscale phenomena that lack simple models suitable for control design. This lack
of simple equations motivates data-driven control methodologies, which include system identification for model
discovery [73, 9, 86, 16, 22]. Alternatively, one can seek transformations that embed nonlinear dynamics in a global
linear representation, as in the Koopman framework [64, 80]. The goal of this work is to reformulate controlled
nonlinear dynamics in a Koopman-eigenfunction framework, referred to as Koopman Reduced Order Nonlinear
Identification and Control (KRONIC), that shows improved predictive power and is amenable to powerful linear
optimal and robust control techniques [104, 37, 107].
∗ E-mail: [email protected]

1
 M g M
xm ym
y1
x1 y2 y3
x3 K
F x2

Figure 1: Koopman embedding to linearize nonlinear dynamics.

A wide range of data-driven and nonlinear control approaches exist in the literature, including model-free
adaptive control [68], extremum-seeking [6], gain scheduling [98], feedback linearization [27], describing func-
tions [116], sliding mode control [38], singular perturbation [63], geometric control [17], back-stepping [62], model
predictive control [26, 79], reinforcement learning [111], and machine learning control [48, 21]. Although consid-
erable progress has been made in the control of nonlinear systems [68, 54, 100], methods are generally tailored
to a specific class of problems, require considerable mathematical and computational resources, or don’t readily
generalize to new applications. Currently there is no overarching framework for nonlinear control, that is gener-
ically applicable to a wide class of potentially high-dimensional systems, as exists for linear systems [37, 104].
Generally applicable frameworks, such as dynamic programming [15], Pontryagin’s maximum principle [92], and
model-predictive control (MPC) [2, 26], often suffer from the curse of dimensionality and require considerably
computational effort, e.g. solving adjoint equations, when applied to nonlinear systems, which can be mitigated
to some degree by combining these with low-dimensional, linear representations. Fortunately, the rise of big data,
advances in machine learning, and new approaches in dynamical systems are changing how we approach these
canonically challenging nonlinear control problems. For instance, recently deep reinforcement learning has been
combined with MPC [91, 121], yielding impressive results in the large-data limit.
Koopman operator theory has recently emerged as a leading framework to obtain linear representations of
nonlinear dynamical systems from data [80]. This operator-theoretic perspective complements the more standard
geometric [46] and probabilistic [35, 36] perspectives. The ability to embed nonlinear dynamics in a linear
framework (see Fig. 1) is particularly promising for the prediction, estimation, and control of nonlinear systems.
In 1931, Koopman showed that a nonlinear dynamical system may be represented by an infinite-dimensional linear
operator acting on the space of measurement functions of the state of the system [64]. Formulating dynamics in
terms of measurements is appealing in the era of big data. Since the seminal work of Mezić and Banaszuk [85]
and Mezić [80], Koopman theory has been the focus of efforts to characterize nonlinear systems. Many classical
results have been extended to the Koopman formalism [25, 81]. For example, level sets of Koopman eigenfunctions
form invariant partitions [24] and may be used to analyze mixing. The Hartman-Grobman theorem has also been
generalized to provide a linearizing transform in the entire basin of attraction of a stable or unstable equilibrium
or periodic orbit [72].
Recently, Koopman theory has been applied for system identification [82, 94, 117, 76], estimation [108, 109] and
control [120, 20, 65, 106, 1] of nonlinear systems. The Koopman operator is infinite-dimensional, and control laws
are typically based on a finite-dimensional approximation. Dynamic mode decomposition (DMD) [101, 97, 115, 69]
approximates the Koopman operator with a best-fit linear model. However, DMD is based on linear measurements,
which do not span a Koopman invariant subspace for many nonlinear systems [118, 20, 69]. Current data-
driven methods to approximate the Koopman operator include extended DMD (EDMD) [118, 119, 60] and the
variational approach of conformation dynamics (VAC) [87, 88]. EDMD was recently used for model predictive
control with promising results [65]. However, EDMD models may suffer from closure issues for systems with
multiple fixed points or attractors, as a linear model only has a single fixed point [20], which may lead to
corrupted dynamics [20, 69] and emphasizes the importance of model validation. For instance, some eigenfunctions
may be distorted when projected onto a finite-dimensional measurement subspace, and it may be advantageous
to construct a reduced-order description from validated eigenfunctions, that exhibit behavior as predicted by
their associated eigenvalue. For chaotic systems, delay coordinates provides a promising embedding [110, 19, 5].
Obtaining useful data-driven coordinate transformations that approximate Koopman eigenfunctions remains an
open challenge in data-driven dynamical systems [69, 20].
In the present work, we build on the existing EDMD and reformulate the Koopman-based control problem in
eigenfunction coordinates and provide strategies to identify lightly damped eigenfunctions from data, that can be

2
 Section 5 Section 4

Figure 2: Control of nonlinear systems via reduced Koopman-invariant representations in eigenfunction coordi-
nates.

subsequently used for control. In particular:

• Koopman eigenfunctions provide a principled linear embedding of nonlinear dynamics resulting in an
intrinsic coordinate system, in which the system is closed under the action of the Koopman operator. We
formulate the incorporation of control in these intrinsic coordinates and discuss benefits and caveats of
this approach. Further, Koopman eigenfunctions can be associated with geometric system properties and
coherent structures. Thus, the designed controller may be employed to manipulate particular coherent
structures. For example, the Hamiltonian energy is a Koopman eigenfunction, and we are able to control
the system by manipulating this function.
• Smooth eigenfunctions in the point spectrum of the Koopman operator can be discovered from given
data using sparse regression providing interpretable representations. We propose sparsity-promoting
algorithms to regularize EDMD or to discover eigenfunctions directly in an implicit formulation.
• We further demonstrate the importance of model validation to distinguish accurately identified eigen-
functions from spurious ones. Lightly damped eigenfunctions are often not corrupted and can be used to
construct reduced-order Koopman models.

These nonlinear control techniques generalize to any lightly damped eigenfunction. As a more sophisticated
example, we consider the double gyre flow, which is a model for ocean mixing. The discovery of intrinsic coordi-
nates for optimized nonlinear control establishes our data-driven KRONIC framework1 , shown in Fig. 2.
The present work is outlined as follows: In Sec. 2, we demonstrate the importance of eigenfunction validation
and motivate the use of sparse regression for their discovery. In Sec. 3, key results in Koopman spectral theory and
corresponding data-driven approaches are summarized, and a brief background on optimal control is provided.
The approach for identifying of Koopman eigenfunctions from data using sparse regression is outlined in Sec. 5.
In Sec. 4, it is shown how control can be incorporated in the eigenfunction formulation. An analytical example
is examined in Sec. 6 to illustrate the control problem in terms of Koopman eigenfunction coordinates. The
KRONIC framework is then demonstrated on several Hamiltonian systems (Sec. 7), for basin-hopping in an
asymmetric double potential well (Sec. 8), and the autonomous and non-autonomous double gyre flow (Sec. 9).
A discussion and outlook on future directions is provided in Sec. 10.

2 Motivation
Finite-dimensional approximations of the Koopman operator are typically obtained as its projection onto a spec-
ified basis or dictionary. Extended dynamic mode decomposition (EDMD) [117] has emerged as the leading
numerical approach by solving a least-squares problem. A well-known issue arises when trying to identify the full
operator in a finite set of basis functions, that sometimes the finite-dimensional approximation is not closed and
spurious eigenfunctions may appear. For this reason, it can be important to perform consistency checks, such
as validating the linearity property of eigenfunctions. The present work builds on EDMD addressing this limita-
tion by re-formulating the regression problem for the direct identification of Koopman eigenfunctions. Further,
we demonstrate how EDMD may be regularized to obtain more accurate eigenfunction representations. In the
following, we illustrate the closure problem using a polynomial basis. As a motivating example (examined in
1 Code at https://github.com/eurika-kaiser/KRONIC.

3
Figure 3: EDMD models give rise to inaccurate eigenvectors, demonstrated for the slow manifold system: (a)
EDMD regression model on up to third degree monomials, (b) fitting onto higher-degree monomials, (c) evolution
of eigenvectors identified from the model in (a), which are evaluated on a single test trajectory, and as predicted
using the corresponding eigenvalue.

detail in Sec. 6), we consider a system with quadratic nonlinearity that gives rise to a slow manifold [115]:
   
d x1 µx1
= (1)
dt x2 λ(x2 − x21 )

with µ = −0.1 and λ = −1. By a clever choice of observable functions y, the nonlinear system (1) may be
represented as a linear system:
d
y = Ky, (2)
dt
where K represents a finite-dimensional approximation of the Koopman operator. The system (1) is one of few
analytical examples, for which a closed, finite-dimensional, linear Koopman approximation exists.
The EDMD model fit on the first 9 monomials (up to third degree) is shown in Fig. 3(a). Some of the
eigenvectors are spurious (see Fig. 3(c)), i.e. the evolution of the eigenfunction ϕ(x(t)) obtained by evaluating the
eigenfunction on a trajectory x(t) does not correspond to the linear prediction using the eigenvalue, eλt ϕ(x(t0 )).
The reason is that the model is not closed in the particular choice of basis functions, because the derivatives of
third degree monomials give rise to higher-degree monomials (see Fig. 3(b)) resulting in a nonlinear model:
      
y1 µ 0 0 0 0 0 0 0 0 y1 0
y2   0 λ −λ 0 0 0 0 0 0    
    y2   0 
y3   0 0 2µ 0 0 0 0 0   
0  y3    0 
   
y4   0 0 0 µ + λ −λ 0 0 0 0  y4   0 

d         
y5 =  0 0 0 0 2λ 0 −2λ 0 0  y  
5 +  0 , (3)

dt      
y 0 0 0 0 0 3µ 0 0 0  y   0 
 6    6  
y7   0 0 0 0    4 
   0 0 0 λ + 2µ 0  y7   −λx1 
y8   0 0 0 0 0 0 0 2λ + µ 0  y8  −2λx31 x2 
y9 0 0 0 0 0 0 0 0 3λ y9 −3λx21 x22

where (y1 , y2 , y3 , y4 , y5 , y6 , y7 , y8 , y9 ) = (x1 , x2 , x21 , x1 x2 , x22 , x31 , x21 x2 , x1 x22 , x33 ). These additional terms, that are
not in the span of {y1 , · · · , y9 }, will be aliased in the corresponding row equations corrupting the system matrix
K. Thus, some of the eigenfunctions will be spurious, affecting the prediction accuracy of the model based
on K. However, it may be possible to identify a subset of the eigenfunctions that are not corrupted, e.g.
those eigenpairs that show good agreement in Fig. 3(c), and use these to construct a reduced-order model with
improved prediction. Alternatively, EDMD may be regularized using sparsity-promoting techniques to regress a
(approximate) closed model on a subset of basis functions. For instance, choosing the five observable functions

4
Figure 4: Sparsified EDMD models increase robustness and prevent overfitting on noisy data: (a) EDMD model,
(b) validation of eigenpairs, and (c) phase plot with L2 error for different models. For details see Fig. 3 and text.

(y1 , y2 , y3 , y4 , y5 ) = (x1 , x2 , x21 , x1 x2 , x31 ), which are a subset of the nine monomials used above, yields:
    
y1 µ 0 0 0 0 y1
y2   0 λ −λ 0 0  y2 
d   
y3  =  0 0 2µ
 
0 0   
 y3  , (4)

dt   
y4   0 0 0 µ + λ −λ y4 
y5 0 0 0 0 3µ y5

which is a 5-dim. linear system, that remains closed under the action of the Koopman operator.
Measurements from real-world systems are generally corrupted by noise, which can be more challenging for
model identification procedures. Figure 4(c) shows the poor prediction performance of the EDMD model trained
on noisy data (displayed in Fig. 4(a)). More eigenfunctions are inaccurate as a result of an increasing number of
non-vanishing coefficients in the state matrix compared with the noise-free situation in Fig. 3. The least-squares
solution overfits resulting in a full matrix with small, but non-vanishing coefficients. By sparsifying the EDMD
state matrix or constructing a reduced-order model based on accurate eigenfunctions (validated from Fig. 4(b))
higher prediction accuracy and robustness to noise can be achieved (compare models in Figure 4(c)). Ideally,
it would be possible to learn good eigenfunctions directly, which are designed to behave linearly and evolve as
predicted by their associated eigenvalue, which would potentially significantly increase prediction accuracy and
reduce the dimension of the model.

3 Background
In this section, we provide a brief background on Koopman spectral theory in Sec. 3.1 and a numerical algorithm,
dynamic mode decomposition, to approximate the Koopman operator in Sec. 3.2. Key results in optimal control
are then summarized in Sec. 3.3.

3.1 Koopman spectral theory

The classical geometric theory of dynamical systems considers a set of coupled ordinary differential equations
d
x(t) = f (x) (5)
dt
in terms of the state of the system x ∈ M, where M is a differentiable manifold, often given by M = Rn . In
discrete time, the dynamics are given by
xk+1 = F(xk ), (6)
where F may be the flow map of the dynamics in (5):
Z t0 +t
F(x(t0 )) = x(t0 ) + f (x(τ )) dτ. (7)
t0

5
Discrete-time systems are more general and form a superset, containing those induced by continuous-time dynam-
ics. Moreover, discrete-time dynamics are often more consistent with experimental measurements, and may be
preferred for numerical analysis. The geometric perspective then considers fixed points and invariant structures
of the dynamics.
In 1931, B. O. Koopman introduced the operator theoretic perspective, showing that there exists an infinite-
dimensional linear operator, given by K, that acts to advance all measurement functions g : M → R of the state
with the flow of the dynamics:
Kg = g ◦ F. (8)
Thus, the Koopman operator advances measurements linearly:

g(xk+1 ) = Kg(xk ). (9)

For smooth dynamics, there is a continuous system

d
g(x) = Kg(x) = ∇g(x) · f (x), (10)
dt
where K is the infinitesimal generator of the one-parameter family of Koopman operators K.
The Koopman operator is linear, which is appealing, but is infinite dimensional, posing issues for representation
and computation. Instead of capturing the evolution of all measurement functions in a Hilbert space, applied
Koopman analysis approximates the evolution on a subspace spanned by a finite set of measurement functions.
It is possible to obtain a finite-dimensional matrix representation of the Koopman operator by restricting it to
an invariant subspace. A Koopman invariant subspace is spanned by any set of eigenfunctions of the Koopman
operator. A Koopman eigenfunction ϕ(x) corresponding to eigenvalue λ satisfies

λϕ(x) = ϕ(F(x)). (11)

In continuous-time, a Koopman eigenfunction ϕ(x) : M → C satisfies

d
ϕ(x) = λϕ(x). (12)
dt
Obtaining Koopman eigenfunctions from data or analytically is a central applied challenge in modern dynamical
systems. Discovering these eigenfunctions enables globally linear representations of strongly nonlinear systems in
terms of these intrinsic observables. The evolution equation (12) describes the unactuated behavior, which will
be extended to incorporate the effect of control in our KRONIC framework (compare the third box in Fig. 2 and
for further details we refer to Sec. 4 and 5).

3.2 Dynamic mode decomposition

Dynamic mode decomposition (DMD) is a simple numerical algorithm that approximates the Koopman opera-
tor with a best-fit linear model that advances measurements from one time step to the next [97, 101, 115, 69].
DMD was originally introduced in the fluid dynamics community [101, 102, 44, 8, 81, 21] to decompose large
data sets into dominant spatial-temporal coherent structures, and a connection with Koopman theory was soon
established [97]. In addition to fluid dynamics, DMD has been widely applied to a range of problems in neu-
roscience [18], robotics [14], epidemiology [95], and video processing [45, 71, 41]. The DMD algorithm has also
been extended to include sparsity and compressed sensing [114, 56, 23, 47, 41], actuation and control [93], multi-
resolution analysis [70], de-noising [34, 51], and streaming variants [52].
In DMD, the infinite-dimensional Koopman operator in (9) is approximated with a finite-dimensional matrix
A that advances the system state x:
xk+1 ≈ Axk . (13)
Given data from a nonlinear system in (6), it is possible to stack snapshots into a matrix X = [x1 x2 . . . xm−1 ]
0
and a time-shifted matrix X = [x2 x3 . . . xm ]. In terms of these data matrices, (13) becomes
0
X ≈ AX. (14)

6
Various DMD algorithms then compute the leading eigendecomposition of the best-fit linear operator A, given
by 0
A = arg min ||X − ÃX||F , (15)
Ã
0
where k · kF is the Frobenius norm. The best-fit A is given by A = X X† , where † is the pseudo-inverse, which
is computed via singular value decomposition.
DMD has proven to be an extremely useful technique for the analysis of high-dimensional dynamical systems
data. However, DMD is based on linear measurements of the system, which do not typically span a Koopman-
invariant subspace of a general nonlinear system. For example, a linear DMD model may perfectly capture
the periodic attractor dynamics of a system on a limit cycle, but will fail to capture the nonlinear transients
if the system is perturbed off the attractor. DMD has since been augmented with nonlinear measurements
to enrich the model in EDMD [118, 119, 60] and VAC [87, 88]. EDMD models have been used with success
for estimation [108, 109] and model predictive control [65]. However, EDMD models are based on a large set of
nonlinear measurements of the state, and there is no guarantee that these measurements form a Koopman invariant
subspace. In fact, EDMD measurement subspaces will generally not be closed [20, 69]. For example, there is no
finite-dimensional Koopman invariant subspace that includes the state of the system x for any dynamical system
that has multiple attractors (e.g., fixed points, periodic orbits, etc.), since the resulting finite-dimensional linear
model cannot be topologically conjugate to the original dynamics. This is closely related to the representation of
the Koopman operator in a polynomial basis, similar to Carleman linearization [10, 67, 112]. Thus, EDMD as well
as other models are often plagued with spurious eigenfunctions that do not behave linearly as predicted by the
associated eigenvalue. Fortunately, although these models may have corrupted eigenvalues and eigenfunctions,
eigenfunctions corresponding to lightly damped eigenvalues may be faithfully extracted.

3.3 Optimal control

The overarching goal of Koopman control is to reformulate strongly nonlinear dynamics in a linear framework
to enable the use of powerful optimal and robust control techniques available for linear systems [104, 37, 107].
Here, we summarize key results in optimal control theory that will be used for control in Koopman eigenfunction
coordinates. We consider the nonlinear system (5) affected by an external input

d
x(t) = f (x, u), x(0) = x0 (16)
dt
with multi-channel control input u ∈ Rq and continuously differentiable dynamics f (x, u) : Rn × Rq → Rn .
Without loss of generality, the origin is an equilibrium: f (0, 0) = 0.
Infinite-horizon optimal control minimizes the following quadratic cost functional
Z∞
1
J(x, u) = xT (t)Qx(t) + uT (t)Ru(t) dt (17)
2 | {z }
0 L(x,u)

with state and input weight matrices Q ∈ Rn×n and R ∈ Rq×q . Both matrices are symmetric and fulfill Q > 0
and R ≥ 0. A full-state feedback control law

u(x) = −C(x)x = c(x), c(0) = 0 (18)

with gain C : Rn → Rq×n is sought that minimizes the cost function (17) subject to the state dynamics (16) to
drive the system to the origin, i.e. lim x(t) = 0, ∀ x.
t→∞

4 Koopman operator control in eigenfunctions

We now propose a general control architecture in Koopman eigenfunction coordinates, referred to as Koopman
Reduced Order Nonlinear Identification and Control (KRONIC)(see Fig. 2). This eigenfunction perspective relies
on a model constructed from validated Koopman eigenfunctions, which is closed and linear by design. First, we
derive how the effect of actuation affects the dynamics of these eigenfunction coordinates. Second, the optimal

7
control problem is formulated in these coordinates and a corresponding feedback controller is then developed,
yielding a possibly nonlinear control law in the original state variables. Control in eigenfunction coordinates
is quite general, encompassing the stabilization of fixed points and periodic orbits, e.g. via the Hamiltonian
eigenfunction, or the manipulation of more general spatial-temporal coherent structures given by level sets of
other eigenfunctions.

4.1 Control–affine systems

We first examine how adding control to a dynamical system (5) affects a single Koopman eigenfunction. This
formulation then readily generalizes to multiple eigenfunctions. Consider a control-affine system
Xq
d
x(t) = f (x) + bi (x)ui , (19)
dt i=1

with a multi-channel input u ∈ Rq , continuously differentiable dynamics f (x) : Rn → Rn associated with the
unforced dynamics, and each bi (x) is a vector field acting on the state space.
Starting with the Koopman operator associated with the uncontrolled, autonomous system (see Sec. 3.1), we
examine how the control terms in (19) affect the dynamics of its eigenfunctions. By applying the chain rule, we
obtain
q
!
d X
ϕ(x) = ∇ϕ(x) · f (x) + bi (x)ui (20a)
dt i=1
q
X
= λϕ(x) + ∇ϕ(x) · bi (x)ui . (20b)
i=1

This equation differs from Eq. (12) in the additional second term associated with the control terms. The ϕ(x)
is a Koopman eigenfunction associated with the autonomous Koopman operator for the unforced dynamics f (x).
For instance, a Hilbert space of the Lebesque square-integrable functions may be considered as function space.
The control enters the dynamics of ϕ via the additional term leading to a control-affine system, which is linear
in ϕ and possibly nonlinear in the control.
Without loss of generality, we assume in the examples presented in later sections a linear control term:
d
x(t) = f (x) + Bu, (21)
dt
with control matrix B ∈ Rn×q , so that the dynamics of the eigenfunctions become
d
ϕ(x) = λϕ(x) + ∇ϕ(x) · Bu. (22a)
dt

4.2 Nonaffine control systems

More generally, we may be interested in the control of a nonlinear, non-affine system:
d
x(t) = f (x, u), (23)
dt
with continuously differentiable dynamics f (x, u) : Rn × Rq → Rn . We may now consider an observable g(x, u)
as a function of the extended state space M × U, where x ∈ M and u ∈ U, and the non-autonomous Koopman
operator acting on these observables. For discrete-time dynamics with discrete map F(x, u), the Koopman op-
erator propagates scalar measurements according to Kg(xk , uk ) = g(F(xk , uk ), uk+1 ) by assuming the Koopman
operator acts on the extended state space in the same manner as the Koopman operator associated with the
autonomous, unforced system. This assumption has been previously [65] considered, where the extended space
was defined as the product of the original state-space and the space of all control sequences. Since the first ap-
pearance of this article, they have been further studies [94, 4, 57] examining the approximation of the Koopman
operator for the non-affine control system. In [94], it is in more detail discussed how the Koopman operator

8
formulation can be modified based on the dynamics of u itself, e.g. open-loop versus closed-loop control. As
associated function space, in which observables are defined on the extended state space, a Hilbert space of the
Lebesque square-integrable functions or polynomial functions defined on a compact set may be considered. If
the dynamics on u are governed by a specific feedback law of the form u = C(x), that is only a function of the
state x, then choices about the function space for the controlled system are equivalent to those applying to the
autonomous system.
For smooth dynamics (23), the continuous-time dynamics are given by
d
g(x, u) = Kg(x, u). (24)
dt
An associated Koopman eigenfunction ϕ(x, u) satisfies
d
ϕ(x, u) = λϕ(x, u). (25)
dt
Applying the chain rule, we find that the dynamics of the Koopman eigenfunction depends on u̇, which is generally
d
arbitrary: dt ϕ(x, u) = ∇x ϕ(x, u) · f (x, u) + ∇u ϕ(x, u) · u̇. Instead, we may specify that ϕ(x, u) reduces to the
eigenfunction ϕ(x, ū) of ẋ = f (x, ū) for all locked ū ∈ U, as in [94]. In this case, the eigenfunction is parametrized
by the input ū
∇x ϕ(x, ū) · f (x, ū) = λϕ(x, ū), ∀ ū ∈ U. (26)
These are eigenfunctions of the parametrized Koopman generator, which is autonomous for each locked ū and
can be defined on the commonly used function spaces as stated above. This perspective is also assumed within
the gEDMD framework and its extension for control [61]. The Koopman generator K is approximated as a finite-
rank matrix K parametrized by the discrete control input using EDMD. The control problem is then solved by
optimizing switching times among the finite set of discrete control inputs and associated models.
If we augment the eigenfunction vector with the input u, we obtain
d
ϕ(x, u) = λϕ(x, u) + ∇u ϕ(x, u) · u̇, (27)
dt
where we view u̇ as the input to the Koopman linear system, and the ∇u ϕ(x, u) matrix varies based on x and u.
Thus, we may enact a gain-scheduled control law.
Summarizing, if the original dynamics are nonlinear and control-affine, the dynamics in Koopman eigenfunction
coordinates are control-affine and split into a linear part associated with the nonlinear unforced dynamics and
a bilinear part associated with the control term. If the original dynamics are nonlinear and non-affine in the
control, the eigenfunction dynamics can become linear if the Koopman operator is defined on the extended state.
Considering practical implications, it is also possible to modify these dynamics so that these consist of a linear
part and a bilinear control term as in Eq. (27). We also point out, that while the control is generally nonlinear
in the state x, it may become linear in the eigenfunction coordinates for special cases, such as ∇ϕ(x) = const.

4.3 Formulation of the optimal control problem

We now formulate the infinite-horizon, optimal control problem [107] for a reduced set of Koopman eigenfunctions.
The control objective is a quadratic cost functional:
Z∞
1
J(ϕ, u) = ϕT (x(t))Qϕ ϕ(x(t)) + uT (t)Ru(t) dt, (28)
2
0
T
where ϕ = [ϕ1 ϕ2 . . . ϕr ] comprises r eigenfunctions with ϕj associated with eigenvalue λj . For this cost
function to be equivalent to the cost in the original state x, a modified weight matrix may be considered such
that ϕT Qϕ ϕ ≈ xT Qx. This can only be achieved exactly if the state itself is an eigenfunction of the Koopman
operator; however, it can be sufficient requiring this only for a subset of states that enter the cost function.
Alternatively, it is possible to estimate x via the inverse mapping ϕ−1 . Eigenfunctions may generally not be
invertible exactly. Nevertheless, the inverse mapping may be approximated using multidimensional scaling as
in [58] or learned jointly with ϕ itself using autoencoders [74]. More generally, the matrix Qϕ allows one to
weight particular eigenfunction directions, which are related to properties of the underlying system and coherent

9
structures. The selection of a specific set of eigenfunctions, in which a model is constructed and which are used to
formulate the cost functional, is problem specific. However, given a target state xREF in the original state space,
the associated target value of the eigenfunctions may be directly determined by evaluating the eigenfunctions on
the target state, i.e. ϕREF := ϕ(xREF ).
For the general case, it is possible to augment the state with the control input and include the derivative of
the control as new input û := u̇:       
d ϕ Λ Bϕ ϕ 0
= + û (29)
dt u 0 0 u Iq
with q × q identity matrix Iq . This may be interpreted as integral control. The cost functional is then given by
Z∞   
1   Qϕ 0 ϕ
J= ϕT uT + ûT R̂û dt. (30)
2 0 R u
0

with some restrictions on R̂. Modifying the system structure, by moving the nonlinearity in the control term into
the state dynamics, improves the tractability of the problem [11].
In the following, we will focus on multiple-input, control-affine systems (21), for which the dynamics in intrinsic
coordinates becomes
d
ϕ(x) = Λϕ(x) + ∇x ϕ(x) · Bu (31)
dt
with Λ = diag(λ1 , . . . , λr ). Depending on the structure of ϕ(x) and B, the actuation matrix Bϕ = ∇x ϕ(x) · B
may be a function of x. A state-dependent control term may be interpreted as a gain-scheduled control. A
feedback controller is now sought in the Koopman representation of the form

u = −cϕ [ϕ(x)] = −Cϕ (x)ϕ(x). (32)

 
We may also consider reference tracking, u = −Cϕ (x) ϕ(x) − ϕREF , with a modified cost functional (28).
The objective is then to determine the gain function Cϕ by minimizing J(ϕ, u) and the resulting control u may
be directly applied to the original system (21), for which the control problem may be suboptimal. Generally,
Koopman eigenfunction control can be used in two ways depending on which means of control one has access
to: (1) internally driven swimmers or particles, which state is given by x and which dynamics are subject to an
external field, such as a fluid flow or magnetic field; or (2) driving an external field, represented in terms of these
eigenfunctions, in which these swimmers or particles drift.

4.4 Solving the optimal control problem

The model (31) may be combined with any model-based control strategy. In the following, we discuss ways to
solve the optimal control problem formulated above using standard techniques.
In the simplest case, for which Eq. (31) becomes fully linear, the optimal gain matrix Cϕ can be determined
by solving the associated algebraic Riccati equation leading to a linear quadratic regulator (LQR) formulated in
Koopman eigenfunctions. For the more general case, where the control term becomes nonlinear, other techniques
are required. A common extension of LQR for nonlinear systems considers state-dependent state and control
matrices as outlined in Sec. A.3 and solves the state-dependent Riccati equation (72). Here, the state dynamics
would be constant and linear, i.e. A = Λ, and only the control matrix Bϕ := ∇x ϕ · B depends on the state x:

Q + PΛ + ΛT P − PBϕ (x)R−1 BTϕ (x)P = 0. (33)

We note that parametrizations for the dynamics (71) in x are unique for scalar systems, but generally nonunique
for multivariable systems [31]. In contrast, the state-dependent dynamics in eigenfunction coordinates (31) are
unique with respect to x and Λ is constant. These are not a result from the factorization. The (non)uniqueness
of the factorization is generally related to global optimal control and global optimal stability. However, further
studies are required to connect these properties for solutions in eigenfunction coordinates to the state dynamics
in x.
We examine the effect of an error εψ(x) in the representation of a Koopman eigenfunction, ϕ̂(x) := ϕ(x) +
εψ(x), on its closed-loop dynamics based on (33) and provide an upper bound for the error (for details see

10
Appendix B). We assume control-affine dynamics (21) of the underlying system,
Pq access to full-state measurements
x, and control vector fields are known. Further, we reformulate B(x)u := i=1 bi (x)ui for simplicity. The closed-
loop dynamics of ϕ(x) are then given by
q
√
ϕ̇(x) = − λ2 + QC(x)R−1 CT (x)ϕ(x) = − µϕ(x), (34)

where µ := λ2 + QC(x)R−1 CT (x). The upper bound for the error in µ due to the misrepresentation of ϕ(x) due
to εψ(x)is
    
 ∇ψ · f (∇ψ · f )2 
|µ − µ̂| ≤  − ε QCR−1 DT + QDR−1 CT + 2λ − ε2 QDR−1 DT + . (35)
ϕ(x) ϕ2 (x)
where C(x) := ∇ϕ(x) · B(x) and D(x) := ∇ψ(x) · B(x) are the control terms associated with ϕ(x) and ψ(x),
respectively. For small ϕ(x) the contribution of εψ(x) becomes important/may be dominant.
More generally, Koopman control in eigenfunction coordinates may be combined with any model-based control
approach. Under certain conditions, it may also be possible to feedback linearize the dynamics [59]. The data-
driven identification of Koopman eigenfunctions can be challenging and they may only be approximated accurately
in a certain domain. Further, dynamics may also drift away from the situations captured in the training data
due to external disturbances. Especially in these cases it is advantageous to couple the resulting model with a
receding horizon estimator and controller to adapt quickly to changing conditions. In particular, model predictive
control has gained increasing popularity over the last decade due to its success in a wide range of applications
and its ability to incorporate customized cost functions and constraints [26, 3, 40].

5 Identifying Koopman eigenfunctions from data

The overarching goal of this work is to control nonlinear systems in intrinsic Koopman eigenfunction coordinates.
A leading method for approximating the Koopman operator is EDMD relying on a set of basis or dictionary
functions. Here, we aim to address the well-known closure issue, i.e. the model may not be closed in the set
of basis functions and can therefore have spurious eigenfunctions, by assuming an eigenfunction perspective.
Approximating these eigenfunctions from data is an ongoing challenge. Building on EDMD, we propose a strategy
to directly identify dominant Koopman eigenfunctions associated with lightly damped eigenvalues, that can then
be utilized to construct low-dimensional, closed models. Further, we show that it is possible to sparsify the
state-transition matrix K of EDMD to improve the representation of eigenfunctions within the selected basis.
It has long been recognized that numerical regularization via the truncation of the SVD in the computation
of the pseudoinverse, as it is done for DMD/EDMD, is crucial to reduce noise corruption. However, this does
not produce a sparser approximation of K, which is crucial for systems that are inherently sparse in their
representation. An additional L1 regularization is here advantageous as it is able to improve the predictive power
of the model and eigenfunctions become more accurately represented in the chosen function library preventing
overfitting. Interestingly, the infinitesimal generator of the Koopman operator may be sparse even when the
Koopman operator itself is not [61]. However, even in these cases, the approach may benefit from a sparsity
constraint to counteract spurious non-zero entries arising from noise and numerical approximation. More
generally, non-compactness or continuous spectra of the Koopman operator can pose issues for the numerical
analysis requiring some form of regularization (see [43] for a detailed discussion for the Koopman operator and [42]
regarding the Perron-Frobenius operator).
In the following we formulate a framework using sparsity-promoting techniques to identify Koopman eigen-
functions directly building on the partial differential equation (PDE) governing the evolution of an eigenfunction.
Applying the chain rule to (12) yields
d
ϕ(x) = ∇ϕ(x) · ẋ = ∇ϕ(x) · f (x). (36)
dt
Combined with (12), this results in a linear PDE for the eigenfunction ϕ(x):

∇ϕ(x) · f (x) = λϕ(x). (37)

This formulation assumes continuous and differentiable dynamics and that the eigenfunctions are smooth [84].

11
 System PDE for Koopman eigenfunction Eigenfunction
Unknown dynamics ! " Sparse coefficients
ẋ = f (x) λΘ(X) − Γ(X, Ẋ) ξ = 0 ϕ(x) ≈ Θ(x)ξ
Example: y ẋ+xẏ 4y ẏ 3 Example:
ẋ = y
x y x2xy y 2 x4y 4 ẋ ẏ 2xẋ 2y ẏ 4x3 ẋ Sparse
ẏ = x − x3 For λ = 0 :  
Data regression ! 2 " 2/3
ϕ(x) = x y2 x4 −2/3
ξ −1/3
Time λ − =0
Truth:
1 1 1
H = − x2 + y 2 + x4
Θ(X) Γ(X, Ẋ) 2 2 4
ξ

Figure 5: Identification of Koopman eigenfunctions from data using implicit-SINDy.

Spectral properties of the Koopman operator have been shown to relate to intrinsic time scales, geometrical
properties, and the long-term behavior of dynamical systems [80, 85, 77, 69]. It has been shown [83], that the
evolution of observables can be described by a linear expansion in Koopman eigenfunctions for systems which
consist only of the point spectrum, i.e. smooth dynamical systems exhibiting, e.g., hyperbolic fixed points, limit
cycles and tori. If systems with a mixed spectrum are considered, it may be possible to restrict the following
analysis to the point spectrum as in [33].

5.1 Data-driven discovery of continuous-time eigenfunctions

Sparse identification of nonlinear dynamics (SINDy) [22] is used to identify Koopman eigenfunctions for a partic-
ular value of λ. This formalism assumes that the system has a point or mixed spectrum, for which eigenfunctions
with distinct eigenvalues exist. A schematic is displayed in Fig. 5.
First, we build a library of candidate functions:
 
Θ(x) = θ1 (x) θ2 (x) · · · θp (x) . (38)

We choose Θ large enough so that the Koopman eigenfunction may be well approximated in this library:
p
X
ϕ(x) ≈ θk (x)ξk = Θ(x)ξ. (39)
k=1

Given data X = [x1 x2 · · · xm ], the time derivative Ẋ = [ẋ1 ẋ2 · · · ẋm ] can be approximated numerically from x(t)
if not measured directly [22]. The total variation derivative [28] is recommended for noise-corrupted measurements.
It is then possible to build a data matrix Θ(X):
 
Θ(X) = θ1 (XT ) θ2 (XT ) · · · θp (XT ) . (40)

Moreover, we can define a library of directional derivatives, representing the possible terms in ∇ϕ(x) · f (x)
from (37): Γ(x, ẋ) = [∇θ1 (x) · ẋ ∇θ2 (x) · ẋ · · · ∇θp (x) · ẋ]. It is then possible to construct Γ from data:
 
Γ(X, Ẋ) = ∇θ1 (XT ) · Ẋ ∇θ2 (XT ) · Ẋ · · · ∇θp (XT ) · Ẋ . (41)

For a specific eigenvalue λ, the Koopman PDE in (37) may be evaluated on data, yielding:
 
λΘ(X) − Γ(X, Ẋ) ξ = 0. (42)

The formulation in (42) is implicit, so that ξ will be in the null-space of the matrix λΘ(X) − Γ(X, Ẋ). The
right null-space of (42) for a given λ is spanned by the right singular vectors of λΘ(X) − Γ(X, Ẋ) = UΣV∗
(i.e., columns of V) corresponding to zero-valued singular values. It is possible to identify the few active terms

12
Algorithm 1 Direct discovery of eigenfunctions based on the implicit formulation (42).
1: Initialize: Λguess = eigs(Θ(X)† Γ(X, Ẋ))}
2: for each eigenvalue λi in Λguess do
3: Initialize: λ = λguess
i
4: while λ not converged do
5: M := Γ(X, Ẋ) − λΘ(X)
6: N := null(M)
7: for each row l in N do
8: q0 := (N(l, :)/norm(N(l, :)))T
9: qmtx (:, l) := adm(N, q0 , α, M axIter, tolerance);
10: Ξ(:, l) = Nqm tx(:, l)
11: end for
12: for each column l in Ξ do
13: Ξ(:, l) := soft thresholding(Ξ(:, l), α)
14: Ξ(:, l) := normalize(Ξ(:, l))
test
15: Error(l) := norm(Θ(Xtest )Ξ(:, l) − eλt Θ(xtest 0 )Ξ(:, l))
16: end for
17: best := min(Error)
18: ξ := Ξ(:, best)
19: Update λ := (ξ T Θ(X)† Γ(X, Ẋ)ξ)/(ξ T ξ)
20: end while
21: T(:, i) = ξ
22: Λ(i, i) = λ
23: end for
24: return Λ, T

in an eigenfunction by finding the sparsest vector in the null-space [96], which is used, e.g., in the implicit-SINDy
algorithm [75]. This is a nonconvex approach based on alternating directions (adm) with linear scaling [96],
which can be adapted to our problem. In this formulation, the eigenvalues λ are not known a priori, and must
be learned online along with the approximate eigenfunction. In Alg. 1, we propose an implicit formulation, which
starts with an initial guess of the eigenvalues given by the least-squares solution, and subsequently alternates
between an searching for the sparsest vector in the null-space [96] and updating of the eigenvalue. As the
approach in [96] depends on the initial condition, we evaluate all initial conditions given by each row in the
nullspace N := null(Γ(X, Ẋ) − λΘ(X)). An additional soft-thresholding with parameter α and validation on
a test dataset is applied to select the best eigenvector for each initial condition. While the approach has been
observed to converge to accurate eigenvalues in clean data, the found eigenvector in the solution set may not
be unique. This approach solves for each eigenpair separately; however, it may also be possible to extend it to
solve for eigenpairs jointly which is part of ongoing research. From a practical standpoint, data in X does not
need to be sampled from full trajectories, but can be obtained using more sophisticated strategies such as latin
hypercube sampling or sampling from a distribution over the phase space. It may also be possible to directly
identify a recursion relationship to obtain a power series expansion as shown in Appendix C and in [78].
Koopman eigenfuntions and eigenvalues can also be determined as the solution to the eigenvalue problem
ξ α K = λα ξ α , where K = Θ† Γ is obtained via least-squares (LS) regression. While many eigenfunctions are
spurious, i.e. these eigenfunctions do not behave linearly as predicted by the corresponding eigenvalue, those
corresponding to lightly damped eigenvalues can be well approximated [20, 69], and a reduced-order Koopman
model may developed on these coordinates (see also Sec. 2). The accuracy of eigenfunctions of K can be improved,
by improving the recovery of K itself, which can be achieved by sparsifying K directly. The minimization problem,

min ||∇θi (XT ) · Ẋ − Θ(X)kTi ||22 + ρ(ki ), ∀ki , i = 1, . . . , p, (43)

ki

where ρ is a regularization term that promotes the sparsity of ki , i.e. the number of non-zero coefficients, is
solved separately for each row in K = [kT1 , . . . , kTp ]. For instance, an L1 constraint on the coefficients in ki may
be chosen and ki may equivalently be determined as in SINDy [22]. In general, the problem (43) can be solved

13
using standard techniques, such as LASSO [99, 113], Least Angle Regression [39], or an iterated least-squares
thresholding method [22].
This formulation is closely related to the gEDMD framework [61], a generalization of the EDMD method to
approximate the infinitesimal generator of the Koopman operator. The gEDMD least-squares formulation is solved
here row-wise with an additional sparsity constraint. There are also similarities to sparsity-promoting DMD [56]
and variants [53, 55], which aims to reconstruct a signal through a sparse combination of modes/eigenfunctions,
which have been computed from a least-squares solution. In contrast to these works, we argue that the sparse
representation of eigenfunctions themselves should be promoted. Alternatively, it is also possible to use the
dominant terms in accurately identified eigenfunctions, generally associated with lightly damped eigenvalues, as
guidance to select observables to regress on. Since the first appearance of this article, further promising methods
have been proposed to identify Koopman eigenfunctions directly [66, 49, 89].

5.2 Data-driven discovery of discrete-time eigenfunctions

In discrete-time, an eigenfunction evaluated at a number of data points in X will satisfy:
 T  T
λϕ(x1 ) . . . λϕ(xm ) = ϕ(x2 ) . . . ϕ(xm+1 ) . (44)

Again, searching for such an eigenfunction ϕ(x) in a library Θ(x) yields the matrix system:

(λΘ(X) − Θ(X0 )) ξ = 0, (45)

 
where X0 = x2 x3 · · · xm+1 is a time-shifted matrix. This formalism directly identifies the functional
representation of a Koopman eigenfunction with eigenvalue λ.
If we seek the best least-squares fit to (45), this reduces to the extended DMD [118] formulation:

λξ = Θ† Θ0 ξ. (46)

Again, it is necessary to confirm that predicted eigenfunctions actually behave linearly on trajectories.

5.3 Koopman model reduction and validation

For any data-driven modeling framework cross-validation is critical to ensure predictive and generalization capa-
bilities. Approximations of the Koopman operator allow a systematic evaluation based on its linearity property.
We can learn a finite-rank approximation using any EDMD-like method or learn eigenfunctions directly. In either
case, it is critical to validate that any candidate eigenfunction ϕ(x(t)) actually behaves linearly on trajectories
x(t) as the eigenvalue λ predicts. The error of an eigenfunction can be defined as

E = ||ϕ(x(t)) − eλt ϕ(x(0))||2 (47)

evaluated on a test trajectory x(t) and identified eigenfunctions can be ranked according to the error E. All
eigenfunctions with error below a threshold may then be used to construct a reduced-order model:

ϕ(x) = Λϕ + ∇ϕ(x) · Bu (48)

analogous to Eq. (31). This model is closed and behaves linearly by design and we demonstrate in the following
its increased predictive power.

6 Example: System with a slow manifold

6.1 Identification of eigenfunctions and models
We consider a system with quadratic nonlinearity that gives rise to a slow manifold [115]:
   
d x1 µx1
= (49)
dt x2 λ(x2 − x21 )

14
Figure 6: Model comparison for the slow manifold system with noise level η = 0.1: (a) training and test data,
(b) state prediction error on test data for each identified model, (c) state prediction error on test data for each
identified reduced-order model (ROM), (d) transition matrix on set of measurement functions, (e) eigenpairs
(β, ϕ(x)) validated on test trajectory (marked blue if below threshold (dashed line)), (f) prediction error for each
eigenpair (β, ϕ(x)) (error between trajectory as predicted by eigenvalue and by evaluating eigenfunction on test
trajectory), (g) state prediction using model K and ROM constructed from validated (or good) eigenfunctions
(marked in blue/yellow in (f)). For details see text.

with µ = −0.1 and λ = −1. In the following, we apply the methods outlined in Sec. 5 to the dynamical
system (49). For all models, a library of the first 9 monomials (up to the third degree) is considered. Noise
corruption in measurement data can be particularly problematic. Here, we examine the recovery of eigenfunctions
and prediction performance for different noise magnitudes η = 0.1 and η = 0.9. Further, reduced-order models are
constructed based on eigenpairs (β, ϕ(x)) that are deemed accurate, i.e. have a small L2 error when compared with
the prediction using the associated eigenvalue, here denoted by β. The threshold for the selection of eigenfunctions
is 1 and 3 for noise magnitudes η = 0.1 and η = 0.9, respectively. The L1 regularized problem is solved using the
iterative least-squares thresholding algorithm (ITLS) as in SINDy [22] and least angle regression (LARS) [39];
however, detailed results are only shown for LARS as these have equivalent performance. LARS has the advantage
that the number of iterations scales with the number of candidate functions; only 10 iterations are required here.
The sparse solution is then selected when there is minimal improvement in the absolute value of correlation with

15
the evolving residual. The soft thresholding parameter in the implicit formulation is set to α = 0.1 and α = 0.2
for η = 0.1 and η = 0.9, respectively.

Figure 7: Model comparison for the slow manifold system as in Fig. 6 but with higher measurement noise level
η = 0.9. The ROM model error displayed in (c) is cut at 100. For details see Fig. 6.

The results are displayed in Figs. 6 and 7 for the two noise cases. We observe that with increasing noise level,
the EDMD state transition matrix becomes denser due to overfitting (see (d)). In contrast, the implicit approach
and L1 −regularized EDMD (termed ‘LARS’ or ‘ITLS’ in the following) yield sparse matrices, which is more
apparent for large noise magnitude. Note that the state matrix for the implicit formulation is reconstructed using
the identified eigenfunctions on the set of candidate functions for visualization purposes, i.e. there are many zero
entries and candidate functions that do not contribute to the actual dynamics. Lightly damped eigenfunctions
can be recovered for all approaches. Eigenfunctions, or specifically eigenpairs (β, ϕ(x)), are deemed as recovered,
if the prediction error falls below a threshold (marked as dashed line in (f)). The prediction error is computed
by evaluating the eigenfunction on a test trajectory and comparing it with the evolution as predicted by the
associated eigenvalue; the evolution is displayed in (e). These good eigenpairs (colored as blue bars in (f)) are
then used to construct reduced-order models (ROMs), that are by design linear and closed (subject to small
errors). Note that solution from the implicit formulation may not be unique, i.e. there are several identical
eigenpairs. However, only unique eigenpairs are used to construct the reduce-order model. This is the reason
why, e.g. in Fig. (6)(f), the error for all eigenpairs falls below the threshold, but only unique ones (marked in

16
blue) are selected to construct the ROM.
For small noise magnitude (Fig. 6), the identified models and ROMs perform all well, except the EDMD-based
ROM, as one crucial eigenfunction associated with β = −0.19 falls above the threshold and is not selected. The
eigenvalue is correct, but not the associated eigenfunction. If the eigenfunction is included in the model (yellow
bar in (f) and yellow dashed line in (g)), i.e. the model is constructed from blue and yellow marked eigenpairs,
it yields a similar performance as the full-state EDMD model. It can also be observed that even in the low
noise setting sparsification yields improvements. We note that the full-state model and ROM obtained from the
implicit formulation are identical, as all discovered eigenpairs fall below the threshold (in (f)) and only unique
pairs are selected to construct either model. We note that the 4D-ROM from the implicit formulation can achieve
better accuracy with one dimension lower compared with EDMD and the sparse EDMD model identified using
LARS. Moreover, while all eigenpairs from the implicit formulation are accurate, both EDMD and LARS yield
also non-physical eigenfunctions. The overall L2 prediction error on the test trajectory in the original state is
summarized for the full-state models and ROMs in (b) and (c), respectively.
For higher noise levels (Fig. 7), the performance differences become more apparent. LARS still yields several
more accurate eigenfunctions than EDMD, so that LARS and the LARS-ROM significantly outperform EDMD
and the EDMD-based ROM. The performance of EDMD does not improve here by truncating the SVD when
computing the pseudo-inverse. The implicit formulation is more noise sensitive, as it is searching for eigenfunctions
in the nullspace of a matrix 42, and is only able to discover one accurate eigenfunction. Interestingly, despite this
caveat, the 2D model is able to outperform EDMD and the EDMD-based ROM. The implicit model is unable
to predict the exact transient behavior; however, it accurately predicts the convergence to the steady state. The
implicit ROM model, however, is insufficient with just one eigenfunction ϕ(x) = x1 and is unable to predict
the evolution of the second state x2 . When selecting one or two additional eigenfunctions for the EDMD-based
ROM (in yellow shown in (g) for two additional eigenfunctions, i.e. in total six eigenfunctions marked by blue
and yellow in (f)), the prediction performance improves considerably, although it is still unable to predict the
steady-state behavior.
Summarizing, EDMD is suffering from overfitting and sparsity-promoting formulations can yield significant
performance enhancements. Model validation in terms of eigenfunctions is a crucial step and can be used to
construct better performing and lower-dimensional reduced-order models. More generally, these results also
demonstrate the importance of not just selecting good, but also the right eigenfunctions.

6.2 Control design

We now demonstrate control in intrinsic Koopman coordinates for the controlled system:
   
d x1 µx1
= + Bu (50)
dt x2 λ(x2 − x21 )

where the control vector is B ∈ R2 . This system can be represented as a finite-dimensional, linear system in a spe-
cial choice of observable functions, making it amenable to optimal control [20]. KRONIC in intrinsic coordinates
provides a powerful alternative if the system does not allow for a fully controllable, linear representation.
The system exhibits slow and fast dynamics for |λ|  |µ| and has a single fixed point at the origin. This
nonlinear system can be embedded in a higher-dimensional space (y1 , y2 , y3 ) = (x1 , x2 , x21 ) where the unforced
dynamics form a closed linear system in a Koopman-invariant subspace:
      
y1 µ 0 0 y1 1 0
d   
y2 = 0 λ −λ y2  +  0 1 B u. (51)
dt
y3 0 0 2µ y3 2y1 0
| {z } | {z }
K =By

However, By may be a function of y, and hence of state x, depending on the specific choice of B. Koopman
λ
eigenfunctions of the unforced system, i.e. B ≡ [0 0]T , are ϕµ = x1 and ϕλ = x2 − bx21 with b = λ−2µ with
eigenvalues λ and µ, respectively. These eigenfunctions remain invariant under the Koopman operator K and can
be interpreted as intrinsic coordinates. Note that ϕpβ := ϕpβ are also Koopman eigenfunctions with eigenvalue pβ
for p ∈ N (and p ∈ Z for non-vanishing ϕpβ ).

17
Figure 8: LQR for B = [0 1]T , µ = −0.1 and λ = 1 using standard linearization, truncated Koopman in y and
ϕ (KRONIC) compared with the solution of the nonlinear control problem (TPBV) and feedback linearization:
(a) phase plot and (b) cost evolution. KRONIC is outperforming all other approaches.

The dynamics of the Koopman eigenfunctions are affected by the additional control term B6=[0 0]T according
to (see also Eq. (31))
   
µ 0 0 1 0
d
ϕ = 0 λ 0  ϕ + −2bx1 1 · B u. (52)
dt
0 0 2µ 2x1 0

Here, the first term represents the unforced, uncoupled, linear dynamics of the eigenfunctions and the second
term a possibly state-dependent control term ∇ϕ·B, that incorporates the effect of control on each eigenfunction.
The controller shall stabilize the unstable fixed point at the origin if either µ or λ are unstable. The control
objective is to minimize the quadratic cost function
Z ∞
Jx = xT Qx + R u2 dt (53)
0

with Q = [ 10 01 ] and R = 1, weighing state and control expenditures equally. Analogously, we can define a cost
function in observable functions,
Z ∞ hQ 0i
Jy = yT Qy y + R u2 dt, Qy = 0 , (54)
0 0 00

and in intrinsic coordinates,

Z ∞  
Q 0
T 2
Jϕ = ϕ Qϕ ϕ + R u dt, Qϕ = b . (55)
0 0 b b2

Here, Qy and Qϕ are chosen to yield the same cost in x. Linear optimal control is then directly applied to (52)
to derive the control law, which is then incorporated in (50). The controller is linear in y and ϕ and yields a
nonlinear controller in the state x:
 
  x1
uy = − Cy y = − Cy,1 Cy,2 − Cy,3 x21
x2
 
  x1
uϕ = − Cϕ ϕ = − Cϕ,1 Cϕ,2 − (Cϕ,3 − bCϕ,1 )x21 ,
x2

where Cy ∈ R1×3 and Cϕ ∈ R1×3 are the control gain vectors in observable or intrinsic coordinates, respectively.

6.3 Stabilization with unstable fast dynamics

First, we consider the system with µ = −0.1 and λ = 1 with an unstable x2 direction. The control vector is
B = [0 1]T , resulting in a constant vector in y or ϕ coordinates, By = Bϕ = [0 1 0]T . Note that the first direction
is uncontrollable, but also stable.

18
Figure 9: Control for B = [1 0]T , µ = 0.1 and λ = −1 comparing KRONIC (56) by solving the SDRE, Num.
TPBV, and LQR on the linearized dynamics: (a) phase plot, and (b) cumulative control performance over T = 50
time units. Note that the cost for KRONIC can not be directly compared with TPBV and linear. LQR, as the
controller in the latter two is optimized with respect to Jx , while the former is optimized for Jϕ .

Standard LQR results are compared (see Fig. 8) for the linearized dynamics, truncated Koopman system in
y, truncated Koopman system in ϕ (KRONIC), as well as with feedback linearization [59] (uF L = λx21 − CF L x)
and with numerically solving the nonlinear control problem as a two-point
Pt boundary value problem (TPBV),
with performance evaluated in terms of the cumulative cost Jˆxt = τ =0 Jx (τ ). Both controllers, in observable
functions and intrinsic coordinates, achieve the same performance and outperform linearized LQR and feedback
linearization. The results for the truncated Koopman system in observables correspond to those presented in [20].
There is no difference between those results and the control results of the system in intrinsic coordinates, as
these systems are connected via an invertible linear transformation. One advantage of a formulation in intrinsic
coordinates will become apparent in the next case, where the stable and unstable directions are reversed.

6.4 Stabilization with unstable slow dynamics

We now consider (50) where the stable and unstable directions are reversed, i.e. µ = 0.1 and λ = −1. The control
input now affects the first state x1 with B = [1 0]T , otherwise this state is uncontrollable. As elaborated in [20],
in this case the linear system in observables (51) will become nonlinear in the control term and, more importantly,
will become unstabilizeable as the third state y3 has a positive eigenvalue 2µ. Analogously, the Koopman system
in intrinsic coordinates has an uncontrollable, unstable direction in ϕ2µ . However, the dynamics of the Koopman
eigenfunctions are uncoupled, thus the third direction can be discarded and the controller is developed in the
controllable subspace:       
d ϕµ µ 0 ϕµ 1
= + u. (56)
dt ϕλ 0 λ ϕλ −2bx1
Note that the third direction φ2µ (x) is a harmonic of φµ (x), i.e., φ2µ = φ2µ . Thus, these two directions may not be
independently controllable with a single input. Here, the controller for x is determined by solving the SDRE (see
Sec. 4.3) at each time instant to account for the nonlinear control term. In a truncated Koopman eigenfunction
system like in this example, the weights in Jϕ can generally not be modified to directly replicate the cost Jx .
Here, the weights for Jx are Q = eye(2) and R = 1, and for Jϕ := J(ϕ, u) are Qϕ = Q and Rϕ = 4R. The
choice for Rϕ ensures a fair comparison by enforcing the same applied energy input for all methods.
Performance results are summarized in Fig. 9. It is also possible to combine KRONIC with MPC allowing
for more general objective functions. Then, the control could be formulated in terms of Jx by computing the
inverse ϕ−1 : Rr → Rn if it exists or estimating x from ϕ using, e.g., multidimensional scaling as in [58]. Control
in Koopman intrinsic coordinates allows one to discard uncontrollable, unstable directions, for which standard
control toolboxes such as Matlab’s lqr fail. Note that feedback linearization fails in this case too: The control
law is of the form uFL = x−1
1 CFL x. As the system approaches the origin, the control input becomes unboundedly
large.

19
 Figure 10: KRONIC demonstrated for several Hamiltonian systems.

7 Example: Hamiltonian energy control

Conserved quantities, such as the Hamiltonian, are Koopman eigenfunctions associated with the eigenvalue λ = 0.
Hamiltonian systems represent a special class of systems for which we can easily discover a Koopman eigenfunction
directly from data.
The dynamics of a Hamiltonian system are governed by
d ∂H d ∂H
q= , p=− , (57)
dt ∂p dt ∂q

where q and p are the generalized state and momenta vectors, respectively. The Hamiltonian H = H(q, p)
considered here is time-independent, representing the conserved energy in the system. Trajectories of the system
evolve on constant energy hypersurfaces {(q, p) : H(q, p) = E}, which may be interpreted as oscillatory modes.
Thus, energy level stabilization is a form of oscillation control and corresponds to stabilizing invariant manifolds
in phase space. Also nonlinear fixed point stabilization may correspond to stabilizing a particular value of the
Hamiltonian energy.
d
Consider the nonlinear, control-affine Hamiltonian system dt x = f (x) + Bu where f = [∂H/∂p − ∂H/∂q]T ,
with state vector x = [q p] ∈ R , multi-channel input vector u ∈ Rq , and constant control matrix B ∈ Rn×q .
T n

We may develop the control directly for the eigenfunction equation

d
ϕ = 0 · ϕ + ∇x ϕ · Bu := Bϕ u, (58)
dt
where ϕ = H. This equation also represents the energy conservation law of the system: A change in the
energy, i.e. the Hamiltonian, corresponds to Rthe external supplied work via u. The infinite-time horizon cost
t 2
function to be minimized is J = limt→∞ 12 0 Q (H(x(t))) + uT (t)Ru(t) dt with scalar Q penalizing energy
deviations and p R penalizing the cost expenditure. Assuming a single input u, the control law is given by
u = −sign(BH ) Q/RH(x) feeding back the current energy level H(x). The ratio Q/R determines how aggressive
the controller is. A more aggressive controller, with Q > R, leads to a faster but also more costly convergence to the
desired state, and vice versa. For the specific case Q = R, the control law reduces further to u = −sign(BH )H(x).
Note that this feedback control law is linear in the Hamiltonian function, but nonlinear in the state x. In the
following, we demonstrate the control approach for several Hamiltonian systems by solving the SDRE; an overview
is provided in Fig. 10, where colored curves represent Koopman controlled trajectories. We assume Q = R = 1
for all examples.

20
Figure 11: Comparison of the energy control control strategy [7] and SDRE for the pendulum with ε = 0.1. The
color coding in the phase plots depicts the respective state-dependent control

7.1 Frictionless pendulum

A frictionless pendulum given by a negative cosine potential V (x) = − cos(x1 ) is steered towards different energy
levels HREF (x) = E ∀x, E ∈ {−1, 0, 1, 2}, with feedback law u = −C(H(x) − HREF ) = −C(H(x) − E).
Minimizing the Hamiltonian drives the system to the equilibrium point at the origin with E = −1 (green curve
in Fig. 10). Koopman control has improved performance over an LQR controller based on linearized dynamics
near the center.
We further compare the controller resulting from solving the state-dependent Riccati equation with an energy-
based control (EC) [7], which is developed from physical considerations:
 
0
SDRE : B = , Ḣ = x2 u, u = −sign(x2 )(H(x) − E) (59a)
1
 
0
EC : B = , Ḣ = −x2 cos(x1 )u, u = −sign(−x2 cos(x1 ))(H(x) − E) (59b)
− cos(x1 )
If the functional representation of the Koopman eigenfunction, or specifically here the Hamiltonian energy func-
tion, is identified from data as in KRONIC, the feedback gain can be precomputed and does not need to be
computed on-line. In Fig. 11, the state-dependent feedback law is shown for both controllers as color-coded phase
plots, where negative values are depicted by blue and positive values by red. The extrema are identical. The
control input from the EC controller is here slightly modified compared with (59b) by moving the term − cos(x1 )
into u itself. Then the effective control and performance are displayed in terms of û = − cos(x1 )u for a fair
comparison; otherwise, the EC controller would appear less performing. The pendulum dynamics are locally
controllable as long as u does not vanish, which is achieved if x1 6=π/2 and x2 6=0 for the EC controller and if x2 6=0
for the SDRE controller. Both controllers yield a bang-bang strategy and successfully steer the system to the
desired energy level, though the SDRE controller is performing slightly better.

7.2 Duffing system

Certain damped or forced oscillators are described by the Duffing equation (Fig. 10). The system is steered
towards the energy level E = 0, which corresponds to the separatrix cycle (yellow dashed lines) pyielding a periodic
solution.√ The origin is a saddle point leading to a homoclinic orbit defined by x∗2 = ±x∗1 1 − 1(x∗1 )2 /2 for
x∗1 ≤ ± 2. It is possible to stabilize the center fixed points by commanding a lower reference energy; however,
because of symmetry in the system, both fixed points are indistinguishable in eigenfunction coordinates. This
illustrates a fundamental uncertainty associated with Koopman eigenfunction control.
For the eigenvalue λ = 0, (42) becomes −Γ(X, Ẋ)ξ = 0, and hence a sparse ξ is sought in the null-space of
−Γ(X, Ẋ). Polynomials up to fourth order are employed to construct a library of candidate functions in (40)-
(41). A single time series for t ∈ [0, 10] with time step ∆t = 0.001 starting at the initial state x0 = [0, −2.8]T
is collected. Thus, each row in (40) and (41) corresponds to a time instant of the trajectory. The prediction
and associated error of the identified eigenfunction are displayed in Fig. 12. Here, the energy is predicted over
the full state space using the identified eigenfunction, the Hamiltonian energy function, which is trained from a
single trajectory. The magnitude of the error is very small of O(10−17 ). However, the eigenfunction evaluate on
a trajectory would oscillate with a tiny amplitude around the true energy level.

21
Figure 12: Prediction over the state space and error using the discovered Hamiltonian for the Duffing system
trained on a single trajectory.

Figure 13: Dependency in increasing number of measurements for the Duffing system: (a) error in the Koopman
eigenfunction, (b) error in (42) with estimated ξ, (c) computational time to solve for ξ in (42), and (d) cumulative
control performance Jˆ over T = 10 time units.

The error of the regression problem, the computational time for identifying the eigenfunction, and the control
performance (steering towards E = 0) for an increasing number of measurements in the estimation step are
displayed in Fig. 13. The identified Koopman eigenfunction with λ = 0 from 1792 measurements (kink in
T
Fig. 13(b)) is ϕ(x) = [ x21 x22 x41 ] [ − 32 32 31 ] with error O(10−8 ). This eigenfunction represents a perfect recovery of
the Hamiltonian up to a scaling, as a Hamiltonian multiplied by a constant scalar is also a conserved quantity.
Using a larger time step of ∆t = 0.05, 56 measurements are sufficient to learn the eigenfunction with error
O(10−6 ).

8 Example: Basin hopping in a double well

A Koopman eigenfunction represents a topography over the state; e.g., the Hamiltonian function depicts the energy
landscape of the system. Trajectory control of a set of particles based on a single Koopman eigenfunction is driven
by the difference between the current and desired value in this topography. While the Koopman eigenfunction
is a global, linear representation of the system, the control of the particle is local, e.g. by solving a suboptimal
SDRE. This is illustrated for a particle in an asymmetric double potential V (x1 ) = 41 x41 − 12 x21 − a3 x31 + ax1 with
a = −0.25 (Fig. 14(a)). The Hamiltonian is H = x22 /2 + V (x1 ) and the dynamics are
     
d x1 x2 1 0
= + u. (60)
dt x2 −x31 + ax21 + x1 − a 0 1

22
Figure 14: Switching control strategy based on KRONIC and the homoclinic orbit to jump between wells: (a)
potential function V (x1 ) showing initial conditions (blue and cyan) and extrema (red), (b) phase plot with
unforced (yellow) and controlled trajectories (blue and cyan), (c) instantaneous cost function and (d) total energy.

For an initial condition in the left well (blue dots in Fig. 14(a)) the controller will fail to steer the state to the
fixed point x∗ = [1 0]T in the center of the right well as the trajectory will become trapped in the bottom of
the left well. Instead, the controller must first increase the energy level to the saddle transition, and after the
trajectory passes to the right basin, the energy can be decreased further.
We propose a switching control strategy that exploits the Koopman eigenfunctions to transport particles
between basins of different heights associated with different fixed points. In particular, the following control
strategy steers particles from the left to the right basin:


−C(x)(H(x) − H([a, 0])) if H(x) < H([a, 1]),
u= 0 if H(x) = H([a, 1])and x1 ≤ a,


−C(x)(H(x) − H([1, 0])) if x1 > a.
A particle on an energy level lower than H([a, 1]), associated with the saddle point, is first steered onto a trajectory
with slightly higher energy than the homoclinic orbit connecting the two basins. On this orbit, control is turned
off and the particle travels to the right well exploiting the intrinsic system dynamics. As soon as it passes the
saddle point, control is turned on again directing it to the lowest energy level H([1, 0]) at the desired fixed point.
The controller is demonstrated for two initial conditions, as shown in Fig. 14(b-d), driving both to the desired
energy level.
This controller can be fully derived from data: First, relevant Koopman eigenfunctions can be identified from
data, as shown in Sec. 5. By analyzing roots and extrema of the eigenfunction corresponding to λ = 0, equilibrium
and saddle points can be identified. The homoclinic and heteroclinic orbits associated with the saddles can be
used as natural transit paths between basins. In each basin, eigenfunction control drives the system to the desired
state. Future applications for this control strategy include space mission design and multi-stable systems such as
proteins.

9 Example: Double Gyre flow

We now consider a high-dimensional, spatially evolving, non-autonomous dynamical system, with time-dependent
Koopman eigenfunctions. The periodically driven double gyre flow models the transport between convection cells
in the Rayleigh-Bénard flow due to lateral oscillations, yielding a simple model for the gulf stream ocean front [105].
We employ here the same parameters as in Shadden et al.’s seminal work on Lagrangian coherent structures [103].
The time-dependent stream function is
Ψ(x, y, t) = A sin(πf (x, t)) sin(πy) (61a)
2
with f (x, t) = ε sin(ωt)x + (1 − 2ε sin(ωt))x (61b)
with A = 0.25, ω = 2π, and ε = 0.25 on a periodic domain [0, 2] × [0, 1]. The velocity field v = [vx vy ]T is given
by vx = − ∂Ψ ∂Ψ
∂y and vy = ∂x .

23
Figure 15: Controlled double gyre flow: (a) autonomous with ε = 0 steering an ensemble of drifters (initial
condition depicted by red dots) to the level Ψ = 0.2, (b) a single drifter trajectory (unforced and controlled) in
the non-autonomous double gyre flow with ε = 0.25, and (c) corresponding stream function values.

The control objective is to steer an ensemble of trajectories to a level set of the stream function. This can be
interpreted as the control of an ensemble of active drifters or autonomous gliders in the ocean, which
h drift duei to
d xi vx +γi sin(θi )
hydrodynamic forces associated with vx and vy . The dynamics of the ith drifter are dt [ yi ] = vy +γi cos(θi ) =
[ vvxy ] + [ 10 01 ] u.
For the autonomous and unforced flow with ε = 0, the stream function is a Koopman eigenfunction associated
with the eigenvalue λ = 0. The forced system becomes:

d h i − ∂Ψ  h i
∂Ψ ∂Ψ ∂y + ∂Ψ ∂Ψ
Ψ(x, y) = ∂x ∂y · ∂Ψ ∂x ∂y · B u. (62)
dt ∂x | {z }
| {z }
=0 BΨ (x,y)

Without control, the stream function Ψ is conserved, as it is the negative of the Hamiltonian. The particles follow
streamlines, which are isolines of the stream function.
In the non-integrable case, with ε > 0, the total derivative of the stream function is given by
d
Ψ(x, y, t) = AΨ (x, y, t)Ψ + BΨ (x, y, t) u (63)
dt
where the vanishing term in (62) is not displayed. The first term in (63) arises from the time derivative
∂
∂t Ψ(x, y, t) and is reformulated into a linear-like structure in Ψ: ∂Ψ/∂t = Aπ cos(πf (x, t)) sin(πy)(∂f /∂t) =
π tan−1 (πf (x, t))(∂f /∂t)Ψ = AΨ Ψ. The second term is the time-dependent analogue of the corresponding term
in (62).
For both cases, ε = 0 and ε > 0, a controller is developed for the stream function. The control is then applied
to an ensemble of drifters to steer them towards the level set Ψ = 0.2. As in the previous examples a quadratic
cost function with Q = 1 and R = ( 10 01 ) is considered. In both cases, BΨ and AΨ depend on the state, and
for ε > 0 also on time. Thus, the state-dependent Riccati equation is solved at each point in space and time.
The controller successfully drives an ensemble of drifters distributed over the domain to the desired level, as
shown in Fig. 15(a). Trajectories are integrated using a 4th-order Runge-Kutta scheme from t ∈ [0, 10]. Example
trajectories of the non-autonomous system, with and without control, are presented in Fig. 15(b-c).
Note that in the non-autonomous case, the reference isocurve ΨREF = 0.2 (white dashed in Fig. 15(b))
oscillates from left to right while being periodically compressed and expanded in x-direction. The particles follow
the moving isocurve resulting in a small oscillation around the desired value (see Fig. 15(c)).
Koopman eigenfunction control can be interpreted in two ways: (1) applying local control to internally driven
swimmers or particles in an external field such as a fluid flow or magnetic field; or (2) driving the external field
in which the swimmers or particles drift. In the latter case, control drives the amplitude of the stream function
at each point to the desired value. For the double gyre flow with a constrained spatial domain, the spatially
distributed gain may be precomputed.

24
Figure 16: Comparison of Koopman system identification methods for control demonstrated for the Duffing
system. Here, EDMDc-MPC requires knowledge of the Koopman eigenfunction. The employed weights are
QH = 1 and R = 1 for Jϕ , and Q = eye(2) and R = 1 for Jx (instantaneous cumulative costs are shown here).
The prediction and control horizon for MPC is in both cases N = 5.

10 Discussion and conclusions

In summary, we extend the Koopman operator formalism to include actuation, and demonstrate how a nonlinear
control problem may be converted into a bilinear control problem in eigenfunction coordinates. Next, we have
presented a data-driven framework to identify leading eigenfunctions of the Koopman operator and sparsity-
promoting extensions to EDMD. We find that lightly damped or undamped eigenfunctions may be accurately
approximated from data via regression, as these eigenfunctions correspond to persistent phenomena, such as
conserved quantities. Moreover, these are often the structures that we seek to control, since they affect long-
time behavior. We have demonstrated the efficacy of this new data-driven control architecture on a number of
nonlinear systems, including Hamiltonian systems and a challenging high-dimensional ocean mixing model. These
results suggest that identifying and controlling Koopman eigenfunctions may enable significant progress towards
the ultimate goal of a universal data-driven nonlinear control strategy.
Finite-dimensional approximations of the Koopman operator are typically obtained as its projection onto a
specified basis, which may suffer from a well-known closure issue. The matrix K is usually large, as the state is
lifted to a high-dimensional space, and the resulting model rarely closes and spurious eigenfunctions may appear.
Building on EDMD, we construct a reduced-order model using validated Koopman eigenfunctions, so that the
resulting model will be closed by design. Both KRONIC and EDMD with control [65] provide Koopman-based
system identification that can be leveraged for model-based control, such as MPC or LQR, as shown in Fig. 16.
However, there are a number of key differences: (1) KRONIC directly identifies Koopman eigenfunctions, while
EDMDc approximates the Koopman operator restricted to a high-dimensional span of observables. (2) EDMD
augments the state vector with nonlinear measurements, increasing the dimension of the system. In contrast,
KRONIC yields a reduced-order model in terms of a few Koopman eigenfunctions. (3) Control is incorporated in
EDMDc as an approximated affine linear term. KRONIC derives an expression for how eigenfunctions are affected
by control through the generator equation. However, this may render the control term bilinear (nonlinear in the
state). (4) The cost function for EDMDc is defined in the state or measurement space, while KRONIC defines
the cost in eigenfunctions; note that these cost functions are not always transferable. (5) Finally, KRONIC
readily admits more complicated solutions, such as limit cycle stabilization, as these correspond to level sets of
the eigenfunctions.
As with previous studies, this work further cements the importance of accurate identification and represen-
tation of Koopman eigenfunctions. Future work will continue to develop algorithms to extract approximate
eigenfunctions from data, and it is likely that these efforts will benefit from advances in machine learning. In
addition, there is a fundamental uncertainty principle in representing Koopman eigenfunctions, as these eigenfunc-
tions may themselves be irrepresentable, as are the long-time flow maps for chaotic systems. Instead of seeking
perfect Koopman eigenfunctions, which may not be attainable, it will be important to incorporate uncertainty
quantification into the data-driven Koopman framework. Model uncertainties may then be managed with robust
control.
The present work also highlights an important choice of perspective when working with Koopman approxima-
tions. Generally, Koopman eigenfunctions are global objects, such as the Hamiltonian energy function. Although
a global, linear representation of the dynamics is appealing, there is also information that is stripped from these

25
representations. For example, in the case of the Hamiltonian eigenfunction, information about specific fixed
points and spatial locations are folded into a single scalar energy. If the Hamiltonian is viewed as a topography
over the phase space, then this eigenfunction only carries information about the altitude, and not the location.
In contrast, Lan and Mezić [72] show that it is possible to extend the Hartman-Grobman theorem to the entire
basin of attraction of certain fixed points and periodic orbits, providing a local linear embedding of the dynamics.
Connecting these perspectives will continue to yield interesting and important advances in Koopman theory. In
addition, there are known connections between the eigenvalues of geometric structures in phase space and the
spectrum of the Koopman operator [83]. This knowledge may guide the accurate identification of Koopman
eigenfunctions.
Formulating control in terms of Koopman eigenfunctions requires a change of perspective as the control
objective may now be defined in eigenfunctions coordinates. Eigenfunctions characterize, e.g., geometric properties
of the system, such as fixed points and limit cycles, as particular level sets, and the control objective can be
equivalently formulated to steer the system towards these objects. Further, particular eigenfunctions represent
coherent structures, i.e. persistent features of the system, which have been targeted in control for a long time.
However, the specific selection of eigenfunctions to control and their interpretation regarding specific control
goals remains an open problem. Nevertheless, it may still be possible to formulate the control in the original
state space, e.g. by incorporating the state as observables, modifying the state weight matrix appropriately,
or by learning an approximation of the inverse mapping [58], which can be more easily incorporated in the
context of model predictive control. This also motivates additional work to understand how controllability and
observability in these coordinates relate to properties of the nonlinear system. The degree of observability and
controllability will generally vary with different eigenfunctions, so that it may be possible to obtain balanced
realizations. Moreover, classic results, such as the PBH test, indicate that multi-channel actuation may be
necessary to simultaneously control different eigenfunctions corresponding to the same eigenvalue, such as the
Hamiltonian energy and conserved angular momentum. The additional degrees of freedom arising from multi-
channel inputs can also be used for eigenstructure assignment to shape Koopman eigenfunctions [50]. Thus,
actuation may modify both the shape of coherent structures (i.e., Koopman modes associated with a particular
eigenfunction) and their time dynamics. It may also be possible to use Koopman linear embeddings to optimize
sensor and actuator placement for nonlinear systems.
Finally, as undamped or lightly damped eigenfunctions correspond to conserved or nearly conserved quantities,
there are many potential applications of the proposed control strategy. For example, symmetries give rise to
other conserved quantities, which will likewise yield new Koopman eigenfunctions. In many physical systems,
simultaneously controlling the system energy and angular momentum may be an important goal. Much of the
present work was formulated with the problem of space mission design in mind. Energy efficient transport
throughout the solar system has long driven advances in theoretical and computational dynamical systems, and
may stand to benefit from control based on Koopman eigenfunctions. More generally, there is a broad range
of applications that stand to benefit from improved nonlinear control, include self-driving cars, the control of
turbulence, suppressing the spread of disease, stabilizing financial markets, human machine interfaces, prosthetics
and rehabilitation, and the treatment of neurological disorders, to name only a few.

Acknowledgments
EK acknowledges funding by the Moore/Sloan foundation, the Washington Research Foundation, and the eScience
Institute. SLB and JNK acknowledge support from the Defense Advanced Research Projects Agency (DARPA
contract HR011-16-C-0016) and the UW Engineering Data Science Institute, NSF HDR award #1934292. SLB
acknowledges support from the Army Research Office (W911NF-17-1-0306) and the Air Force Office of Scien-
tific Research (FA9550-16-1-0650). JNK acknowledges support from the Air Force Office of Scientific Research
(FA9550-15-1-0385). The authors gratefully acknowledge many valuable discussions with Josh Proctor about
Koopman theory and extensions to control. We would also like to acknowledge Igor Mezić, Maria Fonoberova,
Bernd Noack, Clancy Rowley, Sam Taira, and Lionel Mathelin.

26
A Optimal control
We begin by formulating optimal control for the full nonlinear system in (16). This procedure then simplifies
considerably for linearly factorizable or linear dynamics, motivating Koopman embeddings.

A.1 Hamilton-Jacobi-Bellman equation

A nonlinear optimal control formulation for the system in (16) can be established using dynamic programming [15].
This relies on Bellman’s principle of optimality [13], stating that an optimal trajectory remains optimal when the
problem is initialized at intermediate steps, and leads to the Hamilton-Jacobi-Bellman (HJB) equation [12]:
0 = inf {L(x, u) + ∇x J ∗ · f (x, u)} = inf H (x, ∇x J ∗ , u) . (64)
u u

If the HJB equation (64) has a continuously differentiable, positive definite solution J ∗ (x) [15], also referred to
as an optimal value function, then the optimal control law is
 ∗

∗ ∗ ∂J
u = arg min H x , ,u . (65)
u ∂x
The solution is a state-feedback control law, i.e. a closed-loop controller, which is optimal for any initial condition
and solved for all states at once. Solving this nonlinear PDE is computationally challenging and only feasible for
low-dimensional problems. If solved, however, the HJB equation provides a global solution to the optimal control
problem.

A.2 Euler-Lagrange equations

An alternative approach searches for the optimal trajectory under certain constraints using a variational argument
and Pontryagin’s minimum (or maximum) principle [92]. Here, a small variation around an optimal trajectory
is considered, for which the change in cost shall vanish. The sequence of control inputs that solves this problem,
constrained by the state dynamics (16), constitutes the optimal control. However, the Euler-Lagrange equations
provide only necessary, but not sufficient, conditions, so that the solution may only converge to a local minimum
and the control is generally not optimal in practice.
Similar to (64), a Lagrange-type expression is formulated for the control Hamiltonian H(x, u, z) := L(x, u) +
zT f (x, u), where z is referred to as Lagrange multiplier, co-state or adjoint variable. A necessary condition is
the minimization of the control Hamiltonian (i.e., Pontryagin’s minimum principle), leading to the following set
of coupled ordinary differential equations:
∂H T
ẋ = (x, u∗ , z) = f (x, u) (66a)
∂z
∂H T
ż = − (x, u∗ , z) (66b)
∂x
T
with stationarity condition ∂H ∗
∂u (x, u , z) = 0 and transversality boundary conditions x(0) = x0 and limT →∞ z(T ) =
0. Then, the optimal control is given by
u∗ = arg min H(x, u, z). (67)
u

If L(·) and f (x(·), u(·)) are concave, then the necessary condition is also sufficient and any path that satisfies
these conditions also solves the optimal control problem.
The Euler-Lagrange equations (66) lead to an optimal open-loop control: The control u∗ point-wise minimizes
the control Hamiltonian H(x∗ (t), u, z∗ (t)). This two-point boundary value (TPBV) problem is solvable for higher-
dimensional problems in contrast to the HJB equation. However, in practice, only a local minimum is found. For
systems with a high number of degrees of freedom, such as fluid flows, expensive direct and adjoint simulations
render this approach infeasible, instead motivating the use of reduced-order models.

27
A.3 Linear or factorized systems
The control problem above simplifies considerably for linear systems of the form
d
x(t) = Ax + Bu, x(0) = x0 . (68)
dt
Using ansatz z = Px for the co-state, the optimal control is given by

u = −R−1 BT z = −R−1 BT Px, (69)

with constant gain C = R−1 BT P and where the positive semi-definite matrix P ∈ Rn×n is the solution to the
algebraic Riccati equation (ARE) [107]:

Q + PA + AT P − PBR−1 BT P = 0. (70)

For this special case, referred to as the linear quadratic regulator (LQR), the Euler-Lagrange equations and the
HJB equation lead to the same solution, using a quadratic ansatz for co-state and value function. The ARE (70)
can be solved upfront and yields a global state-feedback control law. This simplicity motivates efforts to find linear
representations for nonlinear systems and explains why nonlinear embeddings via Koopman operator theory are
so appealing.
For later reference, we point out an extension of LQR to control-affine, nonlinear systems by factoring the
governing equations into a linear-like structure, as in (68), where the state transition matrix and actuation matrix
become state-dependent:
d
x(t) = A(x)x + B(x)u, x(0) = x0 . (71)
dt
For these systems, it is common to solve the ARE point-wise, i.e. at each point x in time, leading to the state-
dependent Ricatti (SDRE) equation [90]

Q + PA(x) + AT (x)P − PB(x)R−1 BT (x)P = 0 (72)

where the weight matrices Q and R may also be state dependent. In contrast to the ARE, which can be pre-
computed off-line, the SDRE is solved on-line at each x. The required computational power and memory during
the on-line phase may render it infeasible for high-dimensional systems. The SDRE generalizes LQR for nonlinear
systems, retaining a simple implementation and often yielding near optimal controllers [30]. We refer to [29, 32]
for a review on the SDRE and related discussions, e.g. on controllability and stability.

B Effect of misrepresentation of eigenfunctions

We examine the effect of an error in the representation of a Koopman eigenfunction on the closed-loop dynamics
based on (33) and provide an upper error bound. We consider the following cost function
Z
J = Qϕ2 (x) + Ru2 dt (73)

where a single eigenfunction ϕ(x) is to be controlled and with positive weights Q and R. Assuming a control-affine
system ẋ = f (x) + B(x)u, the eigenfunction satisfies

ϕ̇(x) = λϕ(x) + C(x)u (74)

with C(x) := ∇ϕ(x) · B(x). Here, ϕ is the Koopman eigenfunction associated with the autonomous dynamics
satisfying ϕ̇ = ∇ϕ(x) · f (x) = λϕ(x). We assume full-state measurements x are available and knowledge of how
the control affects these measurements B(x). Thus, we do not have to represent the control term C(x) in terms
of eigenfunctions but can evaluate it exactly given the current measurement x and knowledge of ϕ(x). Solving
the scalar Riccati equation (33) analytically yields
 q 
1 2 + QC(x)R−1 CT (x) .
P = λ + λ (75)
C(x)R−1 CT (x)

28
The feedback control is then given by
 q 
R−1 CT (x)
u = −R−1 CT P ϕ = − λ+ λ2 + QC(x)R−1 CT (x) ϕ(x) (76)
C(x)R−1 CT (x)
and the resulting closed-loop dynamics are
q
√
ϕ̇(x) = − λ2 + QC(x)R−1 CT (x)ϕ(x) = − µϕ(x), (77)

where µ := λ2 + QC(x)R−1 CT (x). We examine the effect of a misrepresentation of the eigenfunction:

ϕ̂(x) := ϕ(x) + εψ(x) (78)
where εψ(x) is the discrepancy to the true eigenfunction ϕ(x) with small ε. The dynamics of ϕ̂(x) are given in a
factorized representation by
˙
ϕ̂(x) = ϕ̇(x) + εψ̇(x) = λϕ(x) + C(x)u + ε∇ψ · (f (x) + B(x)u)
λϕ(x) + ε∇ψ(x) · f (x)
= ϕ̂(x) + (C(x) + εD(x)) u (79)
ϕ(x) + εψ(x)
with D(x) := ∇ψ(x) · B(x). The upper bound for the error in µ due to the misrepresentation of ϕ(x) is
  2 
 2 −1 T λϕ(x) + ε∇ψ(x) · f (x) 
|µ − µ̂| = λ + QC(x)R C (x) − − Q(C(x) + εD(x))R−1 (C(x) + εD(x))T 
ϕ(x) + εψ(x)
 λ2 ϕ2 + ε2λϕ∇ψ · f + ε2 (∇ψ · f )2 

= λ2 − εQ(CR−1 DT + DR−1 CT ) − ε2 QDR−1 DT − 
(ϕ + εψ)2
 λ2 ϕ2 + ε2λϕ∇ψ · f + ε2 (∇ψ · f )2 

≤ λ2 − εQ(CR−1 DT + DR−1 CT ) − ε2 QDR−1 DT − 
ϕ2 (x)
 ε2λϕ∇ψ · f + ε2 (∇ψ · f )2 

=  − εQ(CR−1 DT + DR−1 CT ) − ε2 QDR−1 DT − 
ϕ2
    
 ∇ψ · f (∇ψ · f )2 
=  − ε QCR−1 DT + QDR−1 CT + 2λ − ε2 QDR−1 DT + . (80)
ϕ(x) ϕ2 (x)

C Series solutions for eigenfunctions

It is possible to solve the PDE in (37) for eigenfunctions using standard techniques, such as Taylor or Laurent
series.

C.1 Linear dynamics

Consider linear dynamics
d
x = x. (81)
dt
Assuming a Taylor series expansion for ϕ(x):
ϕ(x) = c0 + c1 x + c2 x2 + c3 x3 + · · ·
then the gradient and directional derivatives are given by:
∇ϕ = c1 + 2c2 x + 3c3 x2 + 4c4 x3 + · · ·
∇ϕ · f = c1 x + 2c2 x2 + 3c3 x3 + 4c4 x4 + · · ·
Solving for terms in (37), c1 = 0 must hold. For any positive integer λ, only one coefficient may be nonzero.
Specifically, for λ = k ∈ Z+ , then ϕ(x) = cxk is an eigenfunction for any constant c. For instance, if λ = 1, then
ϕ(x) = x.

29
C.2 Quadratic nonlinear dynamics
Consider a nonlinear dynamical system
d
= x2 . (82)
dt
There is no Taylor series that satisfies (37), except the trivial solution ϕ = 0 for λ = 0. Instead, we use a Laurent
series:

ϕ(x) = · · · + c−3 x−3 + c−2 x−2 + c−1 x−1 + c0

+ c1 x + c2 x2 + c3 x3 + · · · .

The gradient and directional derivatives are given by:

∇ϕ = · · · − 3c−3 x−4 − 2c−2 x−3 − c−1 x−2 + c1 + 2c2 x

+ 3c3 x2 + 4c4 x3 + · · ·
∇ϕ · f = · · · − 3c−3 x−2 − 2c−2 x−1 − c−1 + c1 x2 + 2c2 x3
+ 3c3 x4 + 4c4 x5 + · · · .

All coefficients with positive index are zero, i.e. ck = 0 for k ≥ 1. The nonpositive index coefficients are given by
the recursion λck+1 = kck , for k ≤ −1. The Laurent series is
 
λ2 λ3
ϕ(x) = c0 1 − λx−1 + x−2 − x−3 + · · · = c0 e−λ/x .
2 3!

This holds for all λ ∈ C. There are also other Koopman eigenfunctions that can be identified from the series.

C.3 Polynomial nonlinear dynamics

For a more general nonlinear dynamical system
d
= axn , (83)
dt
λ 1−n
ϕ(x) = e (1−n)a x is an eigenfunction for all λ ∈ C.

References
[1] I. Abraham, G. De La Torre, and T. D. Murphey. Model-based control using Koopman operators.
arXiv:1709.01568, 2017.
[2] F. Allgöwer, T. A. Badgwell, J. S. Qin, J. B. Rawlings, and S. J. Wright. Nonlinear predictive control and
moving horizon estimation: an introductory overview. In Advances in control, pages 391–449. Springer,
1999.
[3] F. Allgöwer, R. Findeisen, and Z. K. Nagy. Nonlinear model predictive control: From theory to application.
J. Chin. Inst. Chem. Engrs, 35(3):299–315, 2004.
[4] H. Arbabi, M. Korda, and I. Mezić. A data-driven koopman model predictive control framework for nonlinear
partial differential equations. In Proceedings of the IEEE Conference on Decision and Control, pages 6409–
6414. IEEE, 2019.
[5] H. Arbabi and I. Mezić. Ergodic theory, dynamic mode decomposition and computation of spectral prop-
erties of the Koopman operator. arXiv preprint arXiv:1611.06664, 2016.
[6] K. B. Ariyur and M. Krstić. Real-Time Optimization by Extremum-Seeking Control. Wiley, 2003.

30
 [7] K. J. Aström and K. Furuta. Swinging up a pendulum by energy control. Automatica, 36(2):287–295, 2000.
[8] S. Bagheri. Koopman-mode decomposition of the cylinder wake. J. Fluid Mech., 726:596–623, 2013.
[9] B. Bamieh and L. Giarré. Identification of linear parameter varying models. Int. J. Robust and Nonlinear
Control, 12:841–853, 2002.

[10] S. P. Banks. Infinite-dimensional Carleman linearization, the Lie series and optimal control of non-linear
partial differential equations. International journal of systems science, 23(5):663–675, 1992.
[11] S. C. Beeler and D. E. Cox. State-dependent Riccati equation regulation of systems with state and control
nonlinearities. NASA/CR-2004-213245, 2004.

[12] R. Bellman and R. Kalaba. Selected papers on mathematical trends in control theory. Dover, 1964.
[13] R. E. Bellman. Dynamic programming. Princeton University Press, Princeton, N.J., 1957.
[14] E. Berger, M. Sastuba, D. Vogt, B. Jung, and H. B. Amor. Estimation of perturbations in robotic behavior
using dynamic mode decomposition. J. Adv. Robotics, 29(5):331–343, 2015.

[15] D. P. Bertsekas. Dynamic Programming and Optimal Control. Athena Scientific, 2005.
[16] S. A. Billings. Nonlinear system identification: NARMAX methods in the time, frequency, and spatio-
temporal domains. John Wiley & Sons, 2013.
[17] R. W. Brockett. Volterra series and geometric control theory. Automatica, 12(2):167–176, 1976.

[18] B. W. Brunton, L. A. Johnson, J. G. Ojemann, and J. N. Kutz. Extracting spatial–temporal coherent

patterns in large-scale neural recordings using dynamic mode decomposition. J. Neuroscience Methods,
258:1–15, 2016.
[19] S. L. Brunton, B. W. Brunton, J. L. Proctor, E. Kaiser, and J. N. Kutz. Chaos as an intermittently forced
linear system. Nature Commun., 8(19), 2017.

[20] S. L. Brunton, B. W. Brunton, J. L. Proctor, and J. Kutz. Koopman invariant subspaces and finite linear
representations of nonlinear dynamical systems for control. PLoS ONE, 11(2):e0150171, 2016.
[21] S. L. Brunton and B. R. Noack. Closed-loop turbulence control: Progress and challenges. Appl. Mech. Rev.,
67:050801–1–050801–48, 2015.

[22] S. L. Brunton, J. L. Proctor, and J. N. Kutz. Discovering governing equations from data by sparse identi-
fication of nonlinear dynamical systems. Proc. Natl. Acad. Sci, 113(15):3932–3937, 2016.
[23] S. L. Brunton, J. L. Proctor, J. H. Tu, and J. N. Kutz. Compressed sensing and dynamic mode decompo-
sition. J. Comp. Dyn., 2(2):165–191, 2015.

[24] M. Budišić and I. Mezić. Geometry of the ergodic quotient reveals coherent structures in flows. Physica D,
241(15):1255–1269, 2012.
[25] M. Budišić, R. Mohr, and I. Mezić. Applied Koopmanism. Chaos, 22(4):047510, 2012.
[26] E. F. Camacho and C. B. Alba. Model predictive control. Springer Science & Business Media, 2013.

[27] B. Charlet, J. Lévine, and R. Marino. On dynamic feedback linearization. Systems & Control Letters,
13(2):143–151, 1989.
[28] R. Chartrand. Numerical differentiation of noisy, nonsmooth data. ISRN Appl. Math., 2011, 2011.
[29] T. Cimen. State-dependent Riccati equation (SDRE) control: A survey. IFAC Proc. Volumes, 41(2):3761–
3775, 2008.

31
[30] J. R. Clautier, N. D’Souza, and C. P. Mracek. Nonlinear regulation and nonlinear H control via the state-
dependent Riccati equation technique: Part 1, theory. In Proc. Int. Conf. Nonlinear problems in aviation
and aerospace, pages 117–131, 1996.
[31] J. Cloutier, C. D’Souza, and C. Mracek. Nonlinear regulation and nonlinear h∞ control via the state-
dependent riccati equation technique: Part 1, theory; part 2, examples. In Proc. of the First International
Conference on Nonlinear Problems in Aviation and Aerospace, Daytona Beach, FL, pages 117–141, 1996.
[32] J. R. Cloutier. State-dependent Riccati equation techniques: an overview. In Proc. American Control
Conference, volume 2, pages 932–936. IEEE, 1997.
[33] S. Das and D. Giannakis. Delay-coordinate maps and the spectra of koopman operators. Journal of
Statistical Physics, 175(6):1107–1145, 2019.
[34] S. T. M. Dawson, M. S. Hemati, M. O. Williams, and C. W. Rowley. Characterizing and correcting for the
effect of sensor noise in the dynamic mode decomposition. Exp. Fluids, 57(3):1–19, 2016.
[35] M. Dellnitz, G. Froyland, and O. Junge. The algorithms behind GAIO—set oriented numerical methods
for dynamical systems. In Ergodic theory, analysis, and efficient simulation of dynamical systems, pages
145–174. Springer, 2001.
[36] M. Dellnitz and O. Junge. Set oriented numerical methods for dynamical systems. Handbook of dynamical
systems, 2:221–264, 2002.
[37] G. E. Dullerud and F. Paganini. A course in robust control theory: A convex approach. Texts in Applied
Mathematics. Springer, Berlin, Heidelberg, 2000.
[38] C. Edwards and S. Spurgeon. Sliding mode control: theory and applications. Crc Press, 1998.
[39] B. Efron, T. Hastie, I. Johnstone, and R. Tibshirani. Least angle regression. The Annals of Statistics,
32(2):407–499, 2004.

[40] U. Eren, A. Prach, B. B. Koçer, S. V. Raković, E. Kayacan, and B. Açıkmeşe. Model predictive control in
aerospace systems: Current state and opportunities. Journal of Guidance, Control, and Dynamics, 2017.
[41] N. B. Erichson, S. L. Brunton, and J. N. Kutz. Compressed dynamic mode decomposition for real-time
object detection. J Real-Time Imag. Proc., 2016.
[42] G. Froyland. An analytic framework for identifying finite-time coherent sets in time-dependent dynamical
systems. Physica D, 250:1–19, 2013.
[43] D. Giannakis. Data-driven spectral decomposition and forecasting of ergodic dynamical systems. Appl.
Comput. Harmon. Anal., 47(2):338–396, 2019.
[44] M. Grilli, P. J. Schmid, S. Hickel, and N. A. Adams. Analysis of unsteady behaviour in shockwave turbulent
boundary layer interaction. J. Fluid Mech., 700:16–28, 2012.
[45] J. Grosek and J. N. Kutz. Dynamic mode decomposition for real-time background/foreground separation
in video. arXiv:1404.7592, 2014.
[46] J. Guckenheimer and P. Holmes. Nonlinear oscillations, dynamical systems, and bifurcations of vector fields,
volume 42 of Applied Mathematical Sciences. Springer, 1983.

[47] F. Gueniat, L. Mathelin, and L. Pastur. A dynamic mode decomposition approach for large and arbitrarily
sampled systems. Phys. Fluids, 27(2):025113, 2015.
[48] N. Hansen, A. S. P. Niederberger, L. Guzzella, and P. Koumoutsakos. A method for handling uncertainty
in evolutionary optimization with an application to feedback control of combustion. IEEE Trans. Evol.
Comput., 13(1):180–197, 2009.

32
[49] M. Haseli and J. Cortés. Learning koopman eigenfunctions and invariant subspaces from data: Symmetric
subspace decomposition. arXiv Preprint arXiv:1909:01419, 2019.
[50] M. Hemati and H. Yao. Dynamic mode shaping for fluid flow control: New strategies for transient growth
suppression. In 8th AIAA Theoretical Fluid Mechanics Conference, page 3160, 2017.
[51] M. S. Hemati and C. W. Rowley. De-biasing the dynamic mode decomposition for applied Koopman spectral
analysis. arXiv preprint arXiv:1502.03854, 2015.
[52] M. S. Hemati, M. O. Williams, and C. W. Rowley. Dynamic mode decomposition for large and streaming
datasets. Phys. Fluids, 26(11):111701, 2014.
[53] J.-C. Hua, F. Noorian, D. Moss, P. H. Leong, and G. H. Gunaratne. High-dimensional time series prediction
using kernel-based koopman mode regression. Nonlinear Dynamics, 90:1785–1806, 2017.
[54] A. Isidori. Nonlinear control systems. Springer, 2013.
[55] F. D. L. J. Graff, X. Xu and T. Singh. Reduced-order modeling using dynamic mode decomposition and
least angle regression. In AIAA Aviation Forum, page 3072, 2019.
[56] M. Jovanović, P. Schmid, and J. Nichols. Sparsity-promoting dynamic mode decomposition. Physics of
Fluids, 26:024103, 2014.
[57] E. Kaiser, J. Kutz, and S. Brunton. Data-driven approximations of dynamical systems operators for control.
In invited chapter for The Koopman Operator in Systems and control, pages 197–234. Springer, 2020.
[58] Y. Kawahara. Dynamic mode decomposition with reproducing kernels for Koopman spectral analysis. In
Advances in Neural Information Processing Systems, pages 911–919, 2016.
[59] H. K. Khalil. Noninear Systems. Prentice-Hall, 1996.
[60] S. Klus, P. Koltai, and C. Schütte. On the numerical approximation of the Perron-Frobenius and Koopman
operator. J. Comp. Dyn., 3(1):51–79, 2016.
[61] S. Klus, F. Nüske, S. Peitz, J.-H. Niemann, C. Clementi, and C. Schütte. Data-driven approxima-
tion of the koopman generator: Model reduction, system identification, and control. arXiv Preprint
arXiv:1909:10638v2, 2019.
[62] P. V. Kokotovic. The joy of feedback: nonlinear and adaptive. IEEE Control systems, 12(3):7–17, 1992.
[63] P. V. Kokotovic, R. E. O’malley, and P. Sannuti. Singular perturbations and order reduction in control
theory—an overview. Automatica, 12(2):123–132, 1976.
[64] B. O. Koopman. Hamiltonian systems and transformation in Hilbert space. Proc. Natl. Acad. Sci, 17(5):315–
318, 1931.
[65] M. Korda and I. Mezić. Linear predictors for nonlinear dynamical systems: Koopman operator meets model
predictive control. arXiv:1611.03537, 2016.
[66] M. Korda and I. Mezić. Optimal construction of koopman eigenfunctions for prediction and control. IEEE
Transactions on Automatic Control, 1:1, 2020.
[67] K. Kowalski, W.-H. Steeb, and K. Kowalksi. Nonlinear dynamical systems and Carleman linearization.
World Scientific, 1991.
[68] M. Krstić, I. Kanellakopoulos, and P. V. Kokotović. Nonlinear and adaptive control design. Wiley, 1995.
[69] J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L. Proctor. Dynamic Mode Decomposition: Data-Driven
Modeling of Complex Systems. SIAM, 2016.
[70] J. N. Kutz, X. Fu, and S. L. Brunton. Multi-resolution dynamic mode decomposition. SIAM Journal on
Applied Dynamical Systems, 15(2):713–735, 2016. Preprint. Available: arXiv:1506.00564.

33
[71] J. N. Kutz, J. Grosek, and S. L. Brunton. Dynamic mode decomposition for robust PCA with applications
to foreground/background subtraction in video streams and multi-resolution analysis. In T. Bouwmans,
editor, CRC Handbook on Robust Low-Rank and Sparse Matrix Decomposition: Applications in Image and
Video Processing, 2015.
[72] Y. Lan and I. Mezić. Linearization in the large of nonlinear systems and Koopman operator spectrum.
Physica D, 242(1):42–53, 2013.
[73] L. Ljung. System Identification: Theory for the User. Prentice Hall, 1999.
[74] B. Lusch, J. N. Kutz, and S. L. Brunton. Deep learning for universal linear embeddings of nonlinear
dynamics. Nature Communications, 9:4950, 2018.
[75] N. M. Mangan, S. L. Brunton, J. L. Proctor, and J. N. Kutz. Inferring biological networks by sparse
identification of nonlinear dynamics. IEEE Trans. Molecular, Biological, and Multi-Scale Commun., 2(1):52–
63, 2016.
[76] A. Mauroy and J. Goncalves. Linear identification of nonlinear systems: A lifting technique based on the
Koopman operator. IEEE Conf. Decision and Control (CDC), pages 6500–6505, 2016.
[77] A. Mauroy and I. Mezic. A spectral operator-theoretic framework for global stability. In Decision and
Control (CDC), 2013 IEEE 52nd Annual Conference on, pages 5234–5239. IEEE, 2013.
[78] A. Mauroy and I. Mezić. Global stability analysis using the eigenfunctions of the Koopman operator. IEEE
Trans Autom Control, 61:3356–3369, 2016.
[79] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert. Constrained model predictive control:
Stability and optimality. Automatica, 36(6):789–814, 2000.
[80] I. Mezić. Spectral properties of dynamical systems, model reduction and decompositions. Nonlinear Dy-
namics, 41(1-3):309–325, 2005.
[81] I. Mezić. Analysis of fluid flows via spectral properties of the Koopman operator. Ann. Rev. Fluid Mech.,
45:357–378, 2013.
[82] I. Mezić. On applications of the spectral theory of the Koopman operator in dynamical systems and control
theory. In IEEE Conf. Decision and Control (CDC), pages 7034–7041. IEEE, 2015.
[83] I. Mezic. Koopman operator spectrum and data analysis. arXiv:1702.07597, 2017.
[84] I. Mezić. Spectral operator methods in dynamical systems: Theory and applications. Springer, 2017.
[85] I. Mezić and A. Banaszuk. Comparison of systems with complex behavior. Physica D, 197(1):101–133,
2004.
[86] O. Nelles. Nonlinear system identification: from classical approaches to neural networks and fuzzy models.
Springer, 2013.
[87] F. Noé and F. Nuske. A variational approach to modeling slow processes in stochastic dynamical systems.
Multiscale Modeling & Simulation, 11(2):635–655, 2013.
[88] F. Nüske, R. Schneider, F. Vitalini, and F. Noé. Variational tensor approach for approximating the rare-
event kinetics of macromolecular systems. J. Chemical Physics, 144(5):054105, 2016.
[89] S. Pan, N. Arnold-Medabalimi, and K. Duraisamy. Sparsity-promoting algorithms for the discovery of
informative koopman invariant subspaces. arXiv Preprint arXiv:2002:10637, 2020.
[90] J. D. Pearson. Approximation methods in optimal control i. Sub-optimal control. I. J. Electronics,
13(5):453–469, 1962.
[91] H. Peng, J. Wu, G. Inoussa, Q. Deng, and K. Nakano. Nonlinear system modeling and predictive control
using the RBF nets-based quasi-linear ARX model. Control Eng Pract, 17(1):59–66, 2009.

34
 [92] L. Pontryagin, V. Boltyanskii, R. Gamkrelidze, and E. Mishchenko. The mathematical theory of optimal
processes. Interscience, 1962.
[93] J. L. Proctor, S. L. Brunton, and J. N. Kutz. Dynamic mode decomposition with control. SIAM J. Appl.
Dyn. Sys., 15(1):142–161, 2016.

[94] J. L. Proctor, S. L. Brunton, and J. N. Kutz. Generalizing Koopman theory to allow for inputs and control.
SIAM J. Applied Dynamical Systems, 17(1):909–930, 2018.
[95] J. L. Proctor and P. A. Eckhoff. Discovering dynamic patterns from infectious disease data using dynamic
mode decomposition. International Health, 7(2):139–145, 2015.
[96] Q. Qu, J. Sun, and J. Wright. Finding a sparse vector in a subspace: Linear sparsity using alternating
directions. In Adv. Neural Inform. Process. Syst. 27, pages 3401–3409, 2014.
[97] C. W. Rowley, I. Mezić, S. Bagheri, P. Schlatter, and D. Henningson. Spectral analysis of nonlinear flows.
J. Fluid Mech., 645:115–127, 2009.
[98] W. J. Rugh and J. S. Shamma. Research on gain scheduling. Automatica, 36(10):1401–1425, 2000.

[99] F. Santosa and W. W. Symes. Linear inversion of band-limited reflection seismograms. SIAM Journal on
Scientific and Statistical Computing, 7(4):1307–1330, 1986.
[100] S. S. Sastry. Nonlinear systems: analysis, stability, and control, volume 10. Springer, 2013.
[101] P. J. Schmid. Dynamic mode decomposition of numerical and experimental data. J. Fluid Mech., 656:5–28,
2010.
[102] P. J. Schmid. Application of the dynamic mode decomposition to experimental data. Exp. Fluids, 50:1123–
1130, 2011.
[103] S. Shadden, F. Lekien, and J. Marsden. Definition and properties of Lagrangian coherent structures from
finite-time Lyapunov exponents in two-dimensional aperiodic flows. Physica D, 212:271–304, 2005.

[104] S. Skogestad and I. Postlethwaite. Multivariable feedback control: analysis and design. John Wiley & Sons,
Inc., 2 edition, 2005.
[105] T. H. Solomon and J. P. Gollub. Chaotic particle transport in time-dependent Rayleigh-Bénard convection.
Phys. Rev. A, 38:6280–6286, 1988.

[106] A. Sootla, A. Mauroy, and D. Ernst. An optimal control formulation of pulse-based control using Koopman
operator. arXiv:1707.08462, 2017.
[107] R. F. Stengel. Optimal control and estimation. Courier Corp., 2012.
[108] A. Surana. Koopman operator based observer synthesis for control-affine nonlinear systems. In 55th IEEE
Conf. Decision and Control (CDC, pages 6492–6499, 2016.
[109] A. Surana and A. Banaszuk. Linear observer synthesis for nonlinear systems using Koopman operator
framework. IFAC-PapersOnLine, 49(18):716–723, 2016.
[110] Y. Susuki and I. Mezić. A prony approximation of Koopman mode decomposition. IEEE Conf. Decision
and Control (CDC), pages 7022–7027, 2015.

[111] R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction. The MIT Press, 1998.
[112] S. A. Svoronos, D. Papageorgiou, and C. Tsiligiannis. Discretization of nonlinear control systems via the
Carleman linearization. Chem. Eng. Science, 49(19):3263–3267, 1994.
[113] R. Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society.
Series B (methodological), 58(1):267–288, 1996.

35
[114] J. H. Tu, C. W. Rowley, J. N. Kutz, and J. K. Shang. Spectral analysis of fluid flows using sub-Nyquist-rate
PIV data. Exp. Fluids, 55(9):1–13, 2014.
[115] J. H. Tu, C. W. Rowley, D. M. Luchtenburg, S. L. Brunton, and J. N. Kutz. On dynamic mode decompo-
sition: theory and applications. J. Comp. Dyn., 1(2):391–421, 2014.

[116] W. E. Vander Velde. Multiple-input describing functions and nonlinear system design. McGraw-Hill, New
York, 1968.
[117] M. O. Williams, M. S. Hemati, S. T. M. Dawson, I. G. Kevrekidis, and C. W. Rowley. Extending data-driven
Koopman analysis to actuated systems. IFAC-PapersOnLine, 49(18):704–709, 2016.
[118] M. O. Williams, I. G. Kevrekidis, and C. W. Rowley. A data-driven approximation of the Koopman operator:
extending dynamic mode decomposition. J. Nonlin. Sci., 2015.
[119] M. O. Williams, C. W. Rowley, and I. G. Kevrekidis. A kernel approach to data-driven Koopman spectral
analysis. J. Comp. Dyn., 2(2):247–265, 2015.
[120] D. Wilson and J. Moehlis. An energy-optimal methodology for synchronization of excitable media. SIAM
J. Appl. Dyn. Sys., 13(2):944–957, 2014.
[121] T. Zhang, G. Kahn, S. Levine, and P. Abbeel. Learning deep control policies for autonomous aerial vehicles
with MPC-guided policy search. In IEEE Int Conf Robotics Autom, pages 528–535, 2016.

36