Data-driven discovery of Koopman eigenfunctions for control

E. Kaiser a,⋆, J. N. Kutz b , S. L. Brunton a

a
Department of Mechanical Engineering, University of Washington, Seattle, WA, 98195
b
Department of Applied Mathematics, University of Washington, Seattle, WA, 98195

Abstract
arXiv:1707.01146v2 [math.OC] 12 Apr 2018

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 eigenfunctions establish intrinsic coordinates along
which the dynamics behave linearly. In this work, we demonstrate a data-driven control architecture, termed Koopman Reduced
Order Nonlinear Identification and Control (KRONIC), that utilizes Koopman eigenfunctions to manipulate nonlinear systems
using linear systems theory. We approximate these eigenfunctions with data-driven regression and power series expansions,
based on the partial differential equation governing the infinitesimal generator of the Koopman operator. Although previous
regression-based methods may identify spurious dynamics, we show that lightly damped eigenfunctions may be faithfully
extracted using sparse regression. 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. We derive the form
of control in these intrinsic eigenfunction coordinates and design nonlinear controllers using standard linear control theory.
KRONIC is then demonstrated 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.

Key words: Dynamical systems, nonlinear control, Koopman theory, embedding, system identification, machine learning.

1 Introduction ing control [22,9]. Although considerable progress has been

In contrast to linear systems, a generally applicable made in the control of nonlinear systems [34,24,56], methods
and scalable framework for the control of nonlinear systems are generally tailored to a specific class of problems, require
remains an engineering grand challenge. Improved nonlin- considerable mathematical and computational resources, or
ear control has the potential to transform our ability to don’t readily generalize to new applications. Currently there
interact with and manipulate complex systems across broad is no overarching framework for nonlinear control as exists
scientific, technological, and industrial domains. From tur- for linear systems [19,59]. Fortunately, the rise of big data,
bulence control to brain-machine interfaces, emerging tech- advances in machine learning, and new approaches in dy-
nologies are characterized by high-dimensional, strongly namical systems are changing how we approach these canon-
nonlinear, and multiscale phenomena that lack simple mod- ically challenging nonlinear control problems.
els suitable for control design. This lack of simple equations Koopman operator theory has recently emerged as a
motivates data-driven control methodologies, which include leading framework to obtain linear representations of non-
system identification for model discovery [37,4,48,6,10]. linear dynamical systems from data [42]. This operator-
Alternatively, one can seek transformations that embed theoretic perspective complements the more standard geo-
nonlinear dynamics in a global linear representation, as in metric [21] and probabilistic [17,18] perspectives. The abil-
the Koopman framework [32,42]. The goal of the Koopman ity to embed nonlinear dynamics in a linear framework is
control developed in this paper is to reformulate nonlinear particularly promising for the prediction, estimation, and
dynamics in a linear framework to enable the use of powerful control of nonlinear systems.
linear optimal and robust control techniques [59,19,62]. The In 1931, Koopman showed that a nonlinear dynamical
result is an innovative, data-driven mathematical frame- system may be represented by an infinite-dimensional linear
work for nonlinear control via our KRONIC architecture. operator acting on the space of measurement functions of
A wide range of data-driven and nonlinear control ap- the state of the system [32]. Formulating dynamics in terms
proaches exist in the literature, including model-free adap- of measurements is appealing in the era of big data. Since
tive control [34], extremum-seeking [3], gain scheduling [55], the seminal work of Mezić and Banaszuk [46] and Mezić [42],
feedback linearization [15], describing functions [68], slid- Koopman theory has been the focus of efforts to charac-
ing mode control [20], singular perturbation [31], geomet- terize nonlinear systems. Many classical results have been
ric control [7], back-stepping [30], model predictive con- extended to the Koopman formalism [13,43]. For example,
trol [14,41], reinforcement learning [66], and machine learn- level sets of Koopman eigenfunctions form invariant parti-
tions [12] and may be used to analyze mixing. The Hartman-
Grobman theorem has also been generalized to provide a
⋆ This paper was not presented at any IFAC meeting. Corre- linearizing transform in the entire basin of attraction of a
sponding author E. Kaiser. Tel. +1-206-5500793. stable or unstable equilibrium or periodic orbit [36].
Email addresses: [email protected] (E. Kaiser), [email protected] Recently, Koopman theory has been applied for sys-
(J. N. Kutz), [email protected] (S. L. Brunton). tem identification [44,52,69,39], estimation [63,64] and

Preprint submitted to Automatica 16 April 2018

 System Discover Incorporate control Control design
Unknown dynamics Data Eigenfunctions Reduced-order dynamics Nonlinear feedback

ẋ = f (x) + Bu (∇x ϕ)T · ẋ = λϕ ϕ̇ = λϕ + (∇x ϕ)T · Bu u = Cϕ(x)

Fig. 1. Nonlinear, data-driven control becomes linear in Koopman eigenfunction directions.

control [72,11,33,61,1] of nonlinear systems. The Koop- of the dynamics, Kt g = g ◦ F. Thus, the Koopman operator
man operator is infinite-dimensional, and control laws are advances measurements linearly: g(xk+1 ) = Kt g(xk ). For
typically based on a finite-dimensional approximation. Dy- smooth dynamics, there is a continuous system
namic mode decomposition (DMD) [57,54,67,35] approxi-
mates the Koopman operator with a best-fit linear model. d
However, DMD is based on linear measurements, which do g = Kg, (2)
dt
not span a Koopman invariant subspace for many nonlinear
systems [70,11,35]. Current data-driven methods to ap- where K is the infinitesimal generator of the Koopman op-
proximate the Koopman operator include extended DMD erator. A Koopman eigenfunction ϕ(x) corresponding to
(EDMD) [70,71,29] and the variational approach of confor- eigenvalue λ satisfies λϕ(x) = ϕ(F(x)). In continuous-time,
mation dynamics (VAC) [49,50]. EDMD was recently used a Koopman eigenfunction ϕ(x) satisfies
for model predictive control with promising results [33].
However, EDMD models may suffer from closure issues for d
systems with multiple fixed points or attractors, as a linear ϕ(x) = λϕ(x). (3)
dt
model only has a single fixed point [11]. This may lead to
spurious or corrupted dynamics [11,35]. For chaotic systems, Obtaining Koopman eigenfunctions from data or analyti-
delay coordinates provides a promising embedding [65,8,2]. cally is a central challenge in modern dynamical systems.
Obtaining useful data-driven coordinate transformations Instead of capturing the evolution of all measurement func-
that approximate Koopman eigenfunctions remains an tions in a Hilbert space, applied Koopman analysis approx-
open challenge in data-driven dynamical systems [35,11]. imates the evolution on a subspace spanned by a finite set
In the present work, we develop a strategy to iden- of measurement functions. A Koopman invariant subspace
tify lightly damped Koopman eigenfunctions from data and is spanned by any set of eigenfunctions of the Koopman
use these eigenfunctions for control. We show that Koop- operator. Discovering these eigenfunctions enables globally
man eigenfunctions provide a principled linear embedding linear representations of strongly nonlinear systems.
of nonlinear dynamics, resulting in an intrinsic coordinate Applying the chain rule to (3) yields
system to design controllers that manipulate coherent struc-
tures using linear control theory. Importantly, we show that d
even if VAC/EDMD models contain spurious dynamics, the ϕ(x) = ∇ϕ(x) · ẋ = ∇ϕ(x) · f (x). (4)
dt
lightly damped eigenfunctions are often not corrupted. For
example, the Hamiltonian energy is a Koopman eigenfunc- Combined with (3), this results in a linear partial differential
tion, and we are able to manipulate this function with lin- equation (PDE) for the eigenfunction ϕ(x):
ear control. These nonlinear control techniques generalize
to any lightly damped eigenfunction. As a more sophisti- ∇ϕ(x) · f (x) = λϕ(x). (5)
cated example, we consider the double gyre flow, which is a
model for ocean mixing. The discovery of intrinsic coordi-
This formulation assumes that the eigenfunctions are
nates for optimized nonlinear control establishes our data-
smooth [45]. It is possible to approximate eigenfunctions
driven KRONIC framework 1 , shown in Fig. 1.
with this PDE by solving for the Laurent series [26] or
2 Identifying Koopman eigenfunctions from data by regression. This assumes continuous and differentiable
dynamics.
The classical geometric theory of dynamical systems
considers a set of coupled ordinary differential equations

d
x(t) = f (x) (1) 2.1 Data-driven discovery of continuous-time eigenfunc-
dt tions
in terms of the state of the system x ∈ M, where M is a Sparse identification of nonlinear dynamics (SINDy) [10]
differentiable manifold, often given by M = Rn . In discrete is used to identify Koopman eigenfunctions for a particular
time, the dynamics are given by xk+1 = F(xk ), where F value of λ. This formalism assumes that the system has
may be the flow map of the dynamics in (1). a point or mixed spectrum, for which eigenfunctions with
In 1931, B. O. Koopman introduced the operator the- distinct eigenvalues exist. A schematic is displayed in Fig. 2.
oretic perspective, showing that there exists an infinite- First, we build a library of candidate functions:
dimensional linear operator, given by Kt , that acts to ad- h i
vance all measurement functions g of the state with the flow Θ(x) = θ1 (x) θ2 (x) · · · θp (x) . (6)

1
Code at https://github.com/eurika-kaiser/KRONIC. We choose Θ large enough so that the Koopman eigenfunc-

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

Time
λ − =0 Truth:
−1/3

1 1 1
H = − x 2 + y 2 + x4
2 2 4
Θ(X) Γ(X, Ẋ) ξ
Fig. 2. Identification of Koopman eigenfunctions from data using implicit-SINDy.

tion may be well approximated in this library: be sampled from full trajectories, but can be obtained using
more sophisticated strategies such as latin hypercube sam-
p
X pling or sampling from a distribution over the phase space.
ϕ(x) ≈ θk (x)ξk = Θ(x)ξ. (7) Moreover, reproducing kernel Hilbert spaces (RKHS) can be
k=1 employed to describe ϕ(x) locally in patches of M. It may
also be possible to directly identify a recursion relationship
Given data X = [x1 x2 · · · xm ], the time derivative Ẋ = to obtain a power series expansion as in [26,40].
[ẋ1 ẋ2 · · · ẋm ] can be approximated numerically from x(t) 2.2 Data-driven discovery of discrete-time eigenfunctions
if not measured directly [10]. The total variation deriva- In discrete-time, an eigenfunction evaluated at a num-
tive [16] is recommended for noise-corrupted measurements. ber of data points in X will satisfy:
It is then possible to build a data matrix Θ(X):
h iT h iT
h i λϕ(x1 ) . . . λϕ(xm ) = ϕ(x2 ) . . . ϕ(xm+1 ) . (11)
Θ(X) = θ1 (XT ) θ2 (XT ) · · · θp (XT ) . (8)

Again, searching for such an eigenfunction ϕ(x) in a library

Moreover, we can define a library of directional derivatives, Θ(x) yields the matrix system:
representing the possible terms in ∇ϕ(x) · f (x) from (5):
Γ(x, ẋ) = [∇θ1 (x) · ẋ ∇θ2 (x) · ẋ · · · ∇θp (x) · ẋ]. It is then
(λΘ(X) − Θ(X′ )) ξ = 0, (12)
possible to construct Γ from data:
h i h i
Γ(X, Ẋ)= ∇θ1 (XT ) · Ẋ . (9) where X′ = x2 x3 · · · xm+1 is a time-shifted matrix. This
∇θ2 (XT ) · Ẋ · · · ∇θp (XT ) · Ẋ
formalism directly identifies the functional representation
of a Koopman eigenfunction with eigenvalue λ.
For a specific eigenvalue λ, the Koopman PDE in (5) If we seek the best least-squares fit to (12), this reduces
may be evaluated on data, yielding: to the extended DMD [70] formulation:
 
λΘ(X) − Γ(X, Ẋ) ξ = 0. (10) λξ = Θ† Θ′ ξ. (13)

The formulation in (10) is implicit, so that ξ will be in Again, it is necessary to confirm that predicted eigenfunc-
the null-space of λΘ(X) − Γ(X, Ẋ). The right null-space tions actually behave linearly on trajectories.
of (10) for a given λ is spanned by the right singular vec- 3 Koopman operator control in eigenfunctions
tors of λΘ(X) − Γ(X, Ẋ) = UΣV∗ (i.e., columns of V) We now propose a general, data-driven control archi-
corresponding to zero-valued singular values. It may be tecture in Koopman eigenfunction coordinates, referred to
possible to identify the few active terms in an eigenfunc- as Koopman Reduced Order Nonlinear Identification and
tion by finding the sparsest vector in the null-space [53], Control (KRONIC), to enable the use of powerful opti-
as in the implicit-SINDy algorithm [38]. In this formula- mal and robust control techniques available for linear sys-
tion, the eigenvalues λ are not known a priori, and must tems [59,19,62]. First, dominant eigenfunctions are discov-
be learned online along with the approximate eigenfunc- ered in an unsupervised, data-driven manner (see Sec. 2)
tion. Koopman eigenfuntions and eigenvalues can also providing a reduced-order representation of the system. A
be determined as the solution to the eigenvalue problem feedback controller is then developed in these coordinates,
Kξ α = λα ξ α , where K = Θ† Γ is obtained via least- yielding a possibly nonlinear control law in the original
squares (LS) regression. While many eigenfunctions are state variables. Control in eigenfunction coordinates is quite
spurious, i.e. these eigenfunctions do not behave linearly, general, encompassing the stabilization of fixed points and
those corresponding to lightly damped eigenvalues can be periodic orbits, e.g. via the Hamiltonian eigenfunction, or
well approximated [11,35]. In either case, it is critical to the manipulation of more general spatial-temporal coherent
test any candidate eigenfunctions to make sure that a can- structures given by level sets of other eigenfunctions.
didate eigenfunction ϕ(x(t)) actually behaves linearly as 3.1 Control–affine systems
the predicted eλt on trajectories x(t). We first examine how adding control to a dynamical
From a practical standpoint, data in X does not need to system (1) affects a single Koopman eigenfunction. This

3
formulation then readily generalizes to multiple eigenfunc- For a control-affine system (14), the dynamics are
tions. Consider a control-affine system
d
d ϕ(x) = Λϕ(x) + ∇x ϕ(x) · Bu (21)
x(t) = f (x) + Bu, (14) dt
dt
with Λ = diag(λ1 , . . . , λr ). Depending on the structure of
with a multi-channel input u ∈ Rq and continuously dif- ϕ(x) and B, the actuation vector Bϕ = ∇x ϕ(x) · B may
ferentiable dynamics f (x, u) : Rn × Rq → Rn . The linear be a function of x. A state-dependent control term may be
system in Koopman eigenfunction coordinates (4) becomes interpreted as a spatially distributed actuation.
A feedback controller is now sought in the Koopman
d representation, leading to
ϕ(x) = ∇ϕ(x) · (f (x) + Bu) (15a)
dt
= λϕ(x) + ∇ϕ(x) · Bu. (15b) u = −Cϕ ϕ(x), (22)

ϕ(x) is a Koopman eigenfunction associated with the un- which is linear in the eigenfunctions, but generally nonlin-
forced dynamics. The control input enters the dynamics of ear in the state x. The optimal gain Cϕ can be determined
ϕ via an additional term leading to a control-affine system, by solving the associated algebraic Riccati equation leading
which is linear in ϕ and possibly nonlinear in the control. to a linear quadratic regulator (LQR) formulated in Koop-
3.2 Nonaffine control systems man eigenfunctions. We may also
consider reference track-
More generally, a nonlinear system of the following form ing, u = −Cϕ ϕ(x) − ϕREF , with a modified cost func-
is considered: tional (20). The resulting feedback law is optimal for manip-
d ulating Koopman eigenfunctions. The controller (22) aims
x(t) = f (x, u). (16) to minimize the value of J(ϕ, u). However, the controller
dt
We may define a Koopman eigenfunction ϕ(x, u) so that may be directly applied to (14), for which the control prob-
lem may be suboptimal and can be interpreted as solving
d the state-dependent Riccati equation [51] (SDRE).
ϕ(x, u) = λϕ(x, u). (17) For the general case, it is possible to augment the state
dt with the control input and include the derivative of the
Applying the chain rule, we find that the dynamics of the control as new input û := u̇:
Koopman eigenfunction depends on u̇, which is arbitrary: " # " #" # " #
d
dt ϕ(x, u) = ∇x ϕ(x, u) · f (x, u) + ∇u ϕ(x, u) · u̇. Instead, d ϕ Λ Bϕ ϕ 0
we may specify that ϕ(x, u) reduces to the eigenfunction = + û (23)
dt u 0 0 u Iq
ϕ(x, ū) of ẋ = f (x, ū) for all locked ū ∈ N , as in [52]. In
this case, the eigenfunction is parameterized by the input ū
with q × q identity matrix Iq . This may be interpreted as
∇x ϕ(x, ū) · f (x, ū) = λϕ(x, ū), ∀ ū ∈ N . (18) integral control. The cost functional is then given by

Z∞ h " #" #
If we augment the eigenfunction vector with the input u, 1 i Q 0 ϕ
ϕ
we obtain J= ϕ T uT + ûT R̂û dt. (24)
2 0 R u
0
d
ϕ(x, u) = λϕ(x, u) + ∇u ϕ(x, u) · u̇, (19)
dt with some restrictions on R̂. Modifying the system struc-
ture, by moving the nonlinearity in the control term into the
where we view u̇ as the input to the Koopman linear system, state dynamics, improves the tractability of the problem [5].
and the ∇u ϕ(x, u) matrix varies based on x and u. Thus, If a closed linear regression model K is found in the
we may enact a gain-scheduled control law. space of observables, LQR applied to this system is equiv-
3.3 Formulation of the optimal control problem alent to LQR in a suitably chosen subspace of eigenfunc-
We now formulate the infinite-horizon, optimal control tions. However, the matrix K is usually large, as the state is
problem [62] for a reduced set of Koopman eigenfunctions. lifted to a high-dimensional space, and the resulting model
The control objective is a quadratic cost functional: rarely closes. Model reduction, such as balanced truncation,
Z∞ may be used for real-time applications. Moreover, general-
1 izations such as LQG balancing [25] yield both a reduction
J(ϕ, u) = ϕT (x(t))Qϕ ϕ(x(t)) + uT (t)Ru(t) dt, (20)
2 for a possibly unstable system along with the compensator.
0 As elaborated above, finding such a closed linear regres-
where ϕ = [ϕλ1 ϕλ2 . . . ϕλr ]T comprises r eigenfunctions sion model, K, is extremely challenging in practice; instead
with ϕλj associated with eigenvalue λj . For this cost to lightly damped Koopman eigenfunctions can be robustly
be equivalent to the cost in the original state x, a modi- identified, providing a reduced-order model. Thus, control
fied weight matrix may be considered such that ϕT Qϕ ϕ ≈ of a few dominant controllable Koopman eigenfunctions is
xT Qx. However, this may only be achieved for special in- realizable.
vertible eigenfunctions. More generally, the matrix Qϕ al-
lows one to weight particular eigenfunction directions.

4
 x1
t

Jx xk
x2
6 x1 5
× 105
4 Example: System with a slow manifold 4 10 6 Linear. LQR
2 KRONIC in y
We now demonstrate optimal control in intrinsic Koop- 5
KRONIC

xk
20 4
man coordinates for a system with quadratic nonlinearity 0 Feedback lin.
40 x2 Jx
that gives rise to a slow manifold [67]: −5 Linear. LQR
Numer. TPBV
2
" # " # −10 KRONIC in y
d x1 µx1 t −15 KRONIC 0
= + Bu (25)

Jx xk
dt x2 −4 −2Feedback0lin. 0 20 40
λ(x2 − x21 ) (a) (b)
xNumer.
1 TPBV t

This system can be represented as a finite-dimensional, lin- Fig. 3. LQR for B = [0 1]T , µ = −0.1 and λ = 1 using stan-
ear system in a special choice of observable functions, mak- dard linearization, truncated Koopman in y and ϕ (KRONIC)
ing it amenable to optimal control [11]. KRONIC in intrin- compared with the solution of the nonlinear control problem
sic coordinates provides a powerful alternative if the system (TPBV) and feedback linearization: (a) phase plot and (b) cost
does not allow for a fully controllable, linear representation. evolution. KRONIC is outperforming all other approaches.
The system exhibits slow and fast dynamics for |λ| ≪ ×104 KRONIC Num. TPBV Linear. LQR
|µ| and has a single fixed point at the origin. This nonlin-
ear system can be embedded in a higher-dimensional space Jx 1.660 1.337 1.337
(y1 , y2 , y3 ) = (x1 , x2 , x21 ) where the dynamics are linear: Jϕ 1.232 1.335 1.335
     Table 1
y1 µ 0 0 y1 Control for B = [1 0]T , µ = 0.1 and λ = −1 comparing KRO-
d      NIC (28) solving SDRE, Num. TPBV, and LQR on the lin-
 y2  =  0 λ −λ y2  + By u. (26) earized dynamics over 50 time units.
 
dt   
y3 0 0 2µ y3 [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 un-
| {z }
K
controllable, but also stable. We then consider the system
The unforced dynamics form a closed linear system in a where the stable and unstable directions are reversed.
Koopman-invariant subspace. However, By may be a func- Standard LQR results are compared (see Fig. 3) for the
tion of state x depending on the specific choice of B. Koop- linearized dynamics, truncated Koopman system in y, trun-
man eigenfunctions of the unforced system, i.e. B ≡ [0 0]T , cated Koopman system in ϕ (KRONIC), as well as with
are ϕµ = x1 and ϕλ = x2 − bx21 with b = λ−2µ λ
with eigen- feedback linearization [28] (uF L = λx21 − CF L x) and with
numerically solving the nonlinear control problem as a two-
values λ and µ, respectively. These eigenfunctions remain
point boundary value problem (TPBV), with performance
invariant under the Koopman operator K and can be in-
evaluated in terms of Jx (t). Both controllers, in observable
terpreted as intrinsic coordinates. Note that ϕpα := ϕpα are
functions and intrinsic coordinates, achieve the same perfor-
also Koopman eigenfunctions with eigenvalue pα for p ∈ N
mance and outperform all other controllers. The results for
(and p ∈ Z for non-vanishing ϕpα ).
the truncated Koopman system in observables correspond
The dynamics of the Koopman eigenfunctions are af-
to those presented in [11]. There is no difference between
fected by the additional control term B6=[0 0]T according to
those results and the control results of the system in intrin-
    sic coordinates, as the systems are connected via an invert-
µ 0 0 1 0 ible linear transformation. One advantage of a formulation
d     in intrinsic coordinates will become apparent in the next
 0 λ 0  ϕ + −2bx1
ϕ=  · B u.
1 (27)
 
dt case, where the stable and unstable directions are reversed.
0 0 2µ 2x1 0 We now consider (25) where the stable and unstable di-
rections are reversed, i.e. µ = 0.1 and λ = −1. The control
where the first term represents the uncoupled, linear dy- input now affects the first state x1 with B = [1 0]T , oth-
namics of the eigenfunctions and the second term a possibly erwise this state is uncontrollable. As elaborated in [11], in
state-dependent control term ∇ϕ · B. this case the linear system in observables (26) will become
The controller shall stabilize the unstable fixed point nonlinear in the control term and, more importantly, will
at the origin if either µ or λ are unstable. The control ob- become unstabilizeable as the third state y3 has a positive
jective isR to minimize the quadratic cost function Jx = eigenvalue 2µ. Analogously, the Koopman system in intrin-
t
limt→∞ 0 xT Qx + R u2 dτ with Q = [ 10 01 ] and R = 1, sic coordinates has an uncontrollable, unstable direction in
weighing state and control expenditures equally. We can also ϕ2µ . However, the dynamics of the Koopman eigenfunctions
define a cost function in observable functions with Qy = are uncoupled, thus the third direction can be discarded and
Q 0 h 0
i the controller is developed in the controllable subspace:
0 0
and in intrinsic coordinates with Qϕ = Q b2 . Con-
0bb " # " #" # " #
trol penalization R is kept the same. Here, Qy and Qϕ are d ϕµ µ 0 ϕµ 1
chosen to yield the same cost in x. Linear optimal control = + u. (28)
dt ϕλ 0 λ ϕλ −2bx1
is then directly applied to (27). The controller is linear in y
and ϕ and yields a nonlinear controller in the state x.
First, we consider the system with µ = −0.1 and λ = 1 Note that the third direction φ2µ (x) is a harmonic of φµ (x),
with an unstable x2 direction. The control vector is B = i.e., φ2µ = φ2µ . Thus, these two directions may not be inde-

5
 4
pendently controllable with a single input. Here, the con- Pendulum ε=1 ε=0
troller for x is determined by solving the SDRE to account ! " ! " 2
x2 1
for the nonlinear control term. In a truncated Koopman PSfrag replacements
ẋ =
−a2 sin(x1 )
+ u
0
eigenfunction system, the weights in Jϕ can generally not be 1 2
0

modified to directly replicate the cost Jx . Here, the weights PSfrag replacements
H = x2 − cos(x1 )
2 -2
for Jx are Q = eye(2) and R = 1, and for Jϕ := J(ϕ, u) Ḣ = sin(x1 ) u ε = 2 ε = −1
are Qϕ = Q and Rϕ = 4R. The choice for Rϕ ensures a fair # Measurements -4

comparison by enforcing the same applied energy input for |Θξ − H|

Duffing Oscillator
all methods. Performance results are summarized in Tab. 1. # Measurements
! " ! "
x 0
It is also possible to combine KRONIC PSfrag replacements
with MPC allow- = − H|
ẋ|Θξ 2
x1 − x1 3 +
1
u
ing for more general objective functions. PSfrag
Then, the control
replacements 1 2 1 2 1 4
could be formulated in terms of Jx by computing the inverse H =Time x −[s] x + x
2 2 2 1 4 1
−1 r n
ϕ : R → R if it exists or estimating x from ϕ using, Ḣ = x2 u 10
e.g., multidimensional scaling [27].
6
Control in Koopman intrinsic coordinates allows |Θξone −toH|
2
discard uncontrollable, unstable directions, for which stan- 10Fig. 4. KRONIC demonstrated for several Hamiltonian-3
systems.
-3 × 10

Time [s]
dard control toolboxes such as Matlab’s lqr fail. Note that 2

|Θξ − H|
6

max of
feedback linearization fails in this case too: The control law 1 2
10
is of the form uFL = x−11 CFL x. As the system approaches 1 (a) 2 (c)
the origin, the control input becomes unboundedly large. 10
0

Eq.(10)
Error of
10 6

Jt=10
5 Example: Hamiltonian energy control 10 # Measurements
# Measurements
-6 (b) (d)
Conserved quantities, such as the Hamiltonian, are 10 10 4
3
10 3
5 · 10 4
10 103 5 · 103 104
Koopman eigenfunctions associated with the eigenvalue
λ = 0. Hamiltonian systems represent a special class of # Measurements # Measurements
systems for which we can easily discover a Koopman eigen-
Fig. 5. Dependency in increasing number of measurements for
function directly from data. the Duffing system: (a) error in the Koopman eigenfunction, (b)
The dynamics of a Hamiltonian system are governed by error in (10) with estimated ξ, (c) computational time to solve
for ξ in (10), and (d) control performance over 10 time units.
d ∂H d ∂H
q= , p=− , (29) the desired state, and vice versa. For the specific case Q = R,
dt ∂p dt ∂q
the control law reduces further to u = −sign(BH )H(x).
Note that this feedback control law is linear in the Hamilto-
where q and p are the generalized state and momenta vec-
nian function, but nonlinear in the state x. In the following,
tors, respectively. The Hamiltonian H = H(q, p) considered
we demonstrate the control approach for several Hamilto-
here is time-independent, representing the conserved energy
nian systems by solving the SDRE; an overview is provided
in the system. Trajectories of the system evolve on constant
in Fig. 4, where colored curves represent Koopman con-
energy hypersurfaces {(q, p) : H(q, p) = E}, which may be
trolled trajectories. We assume Q = R = 1 for all examples.
interpreted as oscillatory modes. Thus, energy level stabi-
lization is a form of oscillation control and corresponds to
stabilizing invariant manifolds in phase space. Also nonlin- 5.1 Frictionless pendulum
ear fixed point stabilization may correspond to stabilizing A frictionless pendulum given by a negative cosine po-
a particular value of the Hamiltonian energy. tential V (x) = − cos(x1 ) is steered towards different energy
Consider the nonlinear, control-affine Hamiltonian sys- levels H(x) = E ∀x, E ∈ {−1, 0, 1, 2}, with feedback law
tem dtd
x = f (x) + Bu where f = [∂H/∂p − ∂H/∂q]T , with u = −C(H(x) − E). Minimizing the Hamiltonian drives the
state vector x = [q p]T ∈ Rn , multi-channel input vector system to the equilibrium point at the origin with E = −1
u ∈ Rq , and constant control matrix B ∈ Rn×q . We may (green curve in Fig. 4). Koopman control has improved per-
develop the control directly for the eigenfunction equation formance over an LQR controller based on linearized dy-
namics near the center.
d 5.2 Duffing system
H = 0 · H + ∇x H · Bu := BH u, (30) Certain damped or forced oscillators are described by
dt
the Duffing equation (Fig. 4). The system is steered towards
where ϕ = H. This equation also represents the energy con- the energy level E = 0, which corresponds to the separatrix
servation law of the system: A change in the energy, i.e. cycle (yellow dashed lines) yielding a periodic solution. The
the Hamiltonian, corresponds to the external supplied work origin is a saddle
p point leading to a homoclinic
√ orbit defined
via u. The infinite-horizon
Rt cost function to be minimized is by x∗2 = ±x∗1 1 − 1(x∗1 )2 /2 for x∗1 ≤ ± 2. It is possible
J = limt→∞ 21 0 Q (H(x(t)))2 + uT (t)Ru(t) dt with scalar to stabilize the center fixed points by commanding a lower
Q penalizing energy deviations and R penalizing the cost reference energy; however, because of symmetry in the sys-
expenditure. Assuming apsingle input u, the control law is tem, both fixed points are indistinguishable in eigenfunc-
given by u = −sign(BH ) Q/RH(x) feeding back the cur- tion coordinates. This illustrates a fundamental uncertainty
rent energy level H(x). The ratio Q/R determines how ag- associated with Koopman eigenfunction control.
gressive the controller is. A more aggressive controller, with The error of the regression problem, the computational
Q > R, leads to a faster but also more costly convergence to time for identifying the eigenfunction, and the control

6
 H
J

H J
PSfrag replacements
0
0

V
0.5
g replacements 1
1.5
0.5
0.4 1

J
V H J
1.5 basins of different heights associated with different fixed
−0.4 0.4 1.5 points. In particular, the following control strategy steers
V
Unforced −0.41 0.41
}KRONIC 0
J−0.4 particles from the left to the right basin:
0.5
Unforced
(a) 0 1.5 0.5
}KRONIC 
1 −C(x)(H(x) − H([a, 0])) if H(x) < H([a, 1]),
0.5
1
-2 -1 0 1 2 (c)
u= 0 if H(x) = H([a, 1])and x1 ≤ a,
1.5 20
−C(x)(H(x) − H([1, 0])) if x > a.
Unforced Unforced 1
0.4 }KRONIC 0.4
40
x2 0
−0.4 t }KRONIC
H A particle on an energy level lower than H([a, 1]), associ-
Unforced 0
}KRONIC ated with the saddle point, is first steered onto a trajec-
−1 tory with slightly higher energy than the homoclinic orbit
(b) (d)−0.4
−2 −1 0 1 2 20t 40 connecting the two basins. On this orbit, control is turned
x1 off and the particle travels to the right well exploiting the
Fig. 6. Switching control strategy based on KRONIC and the
homoclinic orbit to jump between wells: (a) potential function intrinsic system dynamics. As soon as it passes the saddle
V (x1 ) showing initial conditions (blue and cyan) and extrema point, control is turned on again directing it to the lowest
(red), (b) phase plot with unforced (yellow) and controlled tra- energy level H([1, 0]) at the desired fixed point. The con-
jectories (blue and cyan), (c) cost function and (d) total energy. troller is demonstrated for two initial conditions, as shown
performance (steering towards E = 0) for an increasing in Fig. 6(b-d), driving both to the desired energy level.
number of measurements in the estimation step are dis- This controller can be fully derived from data: First, rel-
played in Fig. 5. For the eigenvalue λ = 0, (10) becomes evant Koopman eigenfunctions can be identified from data,
−Γ(X, Ẋ)ξ = 0, and hence a sparse ξ is sought in the as shown in Sec. 2. By analyzing roots and extrema of the
eigenfunction corresponding to λ = 0, equilibrium and sad-
null-space of −Γ(X, Ẋ). Polynomials up to fourth order
dle points can be identified. The homoclinic and heteroclinic
are employed to construct a library of candidate functions
orbits associated with the saddles can be used as natural
in (8)-(9). A single time series for t ∈ [0, 10] with time step
transit paths between basins. In each basin, eigenfunction
∆t = 0.001 starting at the initial state x0 = [0, −2.8]T
control drives the system to the desired state. Future appli-
is collected. Thus, each row in (8) and (9) corresponds
cations for this control strategy include space mission design
to a time instant of the trajectory. The identified Koop-
and multi-stable systems such as proteins.
man eigenfunction with λ = 0 from 1792 measurements
T
(kink in Fig. 5(b)) is ϕ(x) = [ x21 x22 x41 ] [ − 23 23 31 ] with error 7 Example: Double Gyre flow
−8
O(10 ). This eigenfunction represents a perfect recovery
of the Hamiltonian up to a scaling, as a Hamiltonian mul- We now consider a high-dimensional, spatially evolving,
tiplied by a constant scalar is also a conserved quantity. non-autonomous dynamical system, with time-dependent
Using a larger time step of ∆t = 0.05, 56 measurements are Koopman eigenfunctions. The periodically driven double
sufficient to learn the eigenfunction with error O(10−6 ). gyre flow models the transport between convection cells in
the Rayleigh-Bénard flow due to lateral oscillations, yielding
6 Example: Basin hopping in a double well a simple model for the gulf stream ocean front [60]. We
A Koopman eigenfunction represents a topography over employ here the same parameters as in Shadden et al.’s
the state; e.g., the Hamiltonian function depicts the energy seminal work on Lagrangian coherent structures [58].
landscape of the system. Trajectory control of a set of par- The time-dependent stream function is
ticles based on a single Koopman eigenfunction is driven by
the difference between the current and desired value in this Ψ(x, y, t) = A sin(πf (x, t)) sin(πy) (32a)
topography. While the Koopman eigenfunction is a global, with f (x, t) = ε sin(ωt)x2 + (1 − 2ε sin(ωt))x (32b)
linear representation of the system, the control of the par-
ticle is local, e.g. by solving a suboptimal SDRE. This is with A = 0.25, ω = 2π, and ε = 0.25 on a periodic domain
illustrated for a particle in an asymmetric double potential [0, 2] × [0, 1]. The velocity field v = [vx vy ]T is given by
V (x1 ) = 41 x41 − 12 x21 − a3 x31 + ax1 with a = −0.25 (Fig. 6(a)). vx = − ∂Ψ ∂Ψ
∂y and vy = ∂x .
The Hamiltonian is H = x22 /2+V (x1 ) and the dynamics are The control objective is to steer an ensemble of trajec-
" # " # " # tories to a level set of the stream function. This can be in-
d x1 x2 1 0 terpreted as the control of an ensemble of active drifters or
= + u. (31) autonomous gliders in the ocean, which drift due to hydro-
dt x2 3 2
−x1 + ax1 + x1 − a 0 1 dynamic forces associated with h vx and vyi. The dynamics of
v +γ sin(θ )
[ yi ] = vyx+γii cos(θii ) = [ vvxy ] + [ 10 01 ] u.
d xi
the ith drifter are dt
For an initial condition in the left well (blue dots in Fig. 6(a))
the controller will fail to steer the state to the fixed point For the autonomous and unforced flow with ε = 0, the
x∗ = [1 0]T in the center of the right well as the trajectory stream function is a Koopman eigenfunction associated with
will become trapped in the bottom of the left well. Instead, the eigenvalue λ = 0. The forced system becomes:
the controller must first increase the energy level to the " #
saddle transition, and after the trajectory passes to the right d h i − ∂Ψ h i
∂y
basin, the energy can be decreased further. Ψ(x, y) = ∂Ψ
∂x
∂Ψ
∂y
· + ∂Ψ ∂Ψ
∂x ∂y
· B u. (33)
dt ∂Ψ
We propose a switching control strategy that exploits ∂x | {z }
| {z } BΨ (x,y)
the Koopman eigenfunctions to transport particles between =0

7
 1
K-Sys ID EDMDc KRONIC
0.2 PSfrag replacements
PSfrag replacements
y Unforced PSfrag replacements Control
PSfragreplacements MPC EDMDc-MPC LQR/SDRE
KRONIC-MPC
KRONIC PSfrag replacements
EDMDc-MPC KRONIC-SDRE
Cost J(ϕ, u)

J
KRONIC-MPC EDMDc-MPC
J(x, u) 0
0 KRONIC-SDRE KRONIC-MPC 4 4
0.05 0 × 10 × 10
KRONIC-SDRE
Fig. 7. Controlled autonomous double gyre flow with ε = 0 5
steering an ensemble of drifters (initial 0.15
condition depicted by 10 2 J(x, u) J(ϕ, u)
PSfrag
red replacements
dots) to the level Ψ = 0.2.
t EDMDc-MPC
KRONIC-MPC 1 1
1 Unforced 0.2 4 KRONIC-SDRE 2
0.2 KRONIC 1
0.1
2 0
y
ag replacements Ψ 0 0 5 10 5 10
−0.1 t t
Unforced Fig. 9. Comparison of Koopman system identification methods
KRONIC 0 −0.2 for control demonstrated for the Duffing system. Here, EDMD-
0 x 1 2
(a) (b) 0 20 t 40
c-MPC requires knowledge of the Koopman eigenfunction. The
Fig. 8. A single drifter trajectory (unforced and controlled) in employed weights are QH = 1 and R = 1 for Jϕ , and Q = eye(2)
the non-autonomous double gyre flow with ε = 0.25. and R = 1 for Jx . The prediction and control horizon for MPC
is in both cases N = 5.
Without control, the stream function Ψ is conserved, as and can be precomputed. Control drives the amplitude of
it is the negative of the Hamiltonian. The particles follow the stream function at each point to the desired value. In
streamlines, which are isolines of the stream function. contrast, the control system for a single drifter includes a
In the non-integrable case, with ε > 0, the total deriva- nonlinear control term, resulting in a suboptimal controller.
tive of the stream function is given by While the time-dependency of the stream function poses
challenges for control design, it is nevertheless promising.
d
Ψ(x, y, t) = AΨ (x, y, t)Ψ + BΨ (x, y, t) u (34) 8 Discussion and conclusions
dt
In summary, we have presented a data-driven frame-
where the vanishing term in (33) is not displayed. work to identify leading eigenfunctions of the Koopman op-
The first term in (34) arises from the time derivative erator and use these eigenfunctions to control nonlinear sys-
∂ tems. We find that lightly damped or undamped eigenfunc-
∂t Ψ(x, y, t) and is reformulated into a linear-like struc-
ture in Ψ: ∂Ψ/∂t = Aπ cos(πf (x, t)) sin(πy)(∂f /∂t) = tions may be accurately approximated from data via regres-
π tan−1 (πf (x, t))(∂f /∂t)Ψ = AΨ Ψ. The second term is the sion, as these eigenfunctions correspond to persistent phe-
time-dependent analogue of the corresponding term in (33). nomena, such as conserved quantities. Moreover, these are
For both cases, ε = 0 and ε > 0, a controller is devel- often the structures that we seek to control, since they affect
oped for the stream function. The control is then applied to long-time behavior. Next, we extend the Koopman opera-
an ensemble of drifters to steer them towards the level set tor formalism to include actuation, and demonstrate how a
Ψ = 0.2. As in the previous examples a quadratic cost func- nonlinear control problem may be converted into a control-
tion with Q = 1 and R = ( 10 01 ) is considered. In both cases, affine linear problem in eigenfunction coordinates. We have
BΨ and AΨ depend on the state, and for ε > 0 also on time. demonstrated the efficacy of this new data-driven control
Thus, the state-dependent Riccati equation is solved at each architecture on a number of nonlinear systems, including
point in space and time. The controller successfully drives Hamiltonian systems and a challenging high-dimensional
an ensemble of drifters distributed over the domain to the ocean mixing model. These results suggest that identifying
desired level, as shown in Fig. 7. Trajectories are integrated and controlling Koopman eigenfunctions may enable signif-
using a 4th-order Runge-Kutta scheme from t ∈ [0, 10]. Ex- icant progress towards the ultimate goal of a universal data-
ample trajectories of the non-autonomous system, with and driven nonlinear control strategy.
without control, are presented in Fig. 8. Note that in the EDMD with control (EDMDc) has recently been com-
non-autonomous case, the reference isocurve ΨREF = 0.2 bined with model predictive control with promising re-
(white dashed in Fig. 8(a)) oscillates from left to right while sults [33]. Both KRONIC and EDMDc provide Koopman-
being periodically compressed and expanded in x-direction. based linear system identification that can be leveraged
The particles follow the moving isocurve resulting in a small for model-based control, such as MPC or LQR, as shown
oscillation around the desired value (see Fig. 8(b)). in Fig. 9. However, there are a number of key differences:
Koopman eigenfunction control can be interpreted in (1) KRONIC directly identifies Koopman eigenfunctions,
two ways: (1) applying control to swimmers or particles in while EDMDc approximates the Koopman operator re-
an external field such as a fluid flow or magnetic field; or (2) stricted to a high-dimensional span of observables. (2)
changing the external field in which the swimmers or parti- EDMD augments the state vector with nonlinear measure-
cles drift. For the autonomous double gyre, the second case ments, increasing the dimension of the system. In contrast,
(not shown) is optimal for the stream function, as (33) is KRONIC yields a reduced-order model in terms of a few
truly linear in Ψ, where BΨ (x, y) represents a spatially dis- Koopman eigenfunctions. (3) Control is incorporated in
tributed actuator. The spatially distributed gain is constant EDMDc as an approximated affine linear term. KRONIC

8
derives an expression for how eigenfunctions are affected by ics. It may also be possible to use Koopman linear embed-
control through the generator equation. However, this may dings to optimize sensor and actuator placement for nonlin-
render the control term nonlinear in the state. (4) The cost ear systems.
function for EDMDc is defined in the state or measurement Finally, as undamped or lightly damped eigenfunctions
space, while KRONIC defines the cost in eigenfunctions; correspond to conserved or nearly conserved quantities,
note that these cost functions are not always transferable. there are many potential applications of the proposed con-
(5) Finally, KRONIC readily admits more complicated trol strategy. For example, symmetries give rise to other
solutions, such as limit cycle stabilization, as these cor- conserved quantities, which will likewise yield new Koop-
respond to level sets of the eigenfunctions, which is more man eigenfunctions. In many physical systems, simultane-
challenging to incorporate with EDMDc. ously controlling the system energy and angular momen-
As with previous studies, this work further cements the tum may be an important goal. Much of the present work
importance of accurate identification and representation was formulated with the problem of space mission design
of Koopman eigenfunctions. Future work will continue to in mind. Energy efficient transport throughout the solar
develop algorithms to extract approximate eigenfunctions system has long driven advances in theoretical and compu-
from data, and it is likely that these efforts will benefit from tational dynamical systems, and may stand to benefit from
advances in machine learning. In addition, there is a fun- control based on Koopman eigenfunctions. More generally,
damental uncertainty principle in representing Koopman there is a broad range of applications that stand to ben-
eigenfunctions, as these eigenfunctions may themselves be efit from improved nonlinear control, include self-driving
irrepresentable, as are the long-time flow maps for chaotic cars, the control of turbulence, suppressing the spread of
systems. Instead of seeking perfect Koopman eigenfunc- disease, stabilizing financial markets, human machine in-
tions, which may not be attainable, it will be important to terfaces, prosthetics and rehabilitation, and the treatment
incorporate uncertainty quantification into the data-driven of neurological disorders, to name only a few.
Koopman framework. Model uncertainties may then be
managed with robust control. Acknowledgements
The present work also highlights an important choice of EK acknowledges funding by the Moore/Sloan foun-
perspective when working with Koopman approximations. dation, the Washington Research Foundation, and the
Generally, Koopman eigenfunctions are global objects, such eScience Institute. SLB and JNK acknowledge support from
as the Hamiltonian energy function. Although a global, lin- the Defense Advanced Research Projects Agency (DARPA
ear representation of the dynamics is appealing, there is also contract HR011-16-C-0016). SLB acknowledges support
information that is stripped from these representations. For from the Army Research Office (W911NF-17-1-0306) and
example, in the case of the Hamiltonian eigenfunction, in- the Air Force Office of Scientific Research (FA9550-16-1-
formation about specific fixed points and spatial locations 0650). JNK acknowledges support from the Air Force Office
are folded into a single scalar energy. If the Hamiltonian of Scientific Research (FA9550-15-1-0385). The authors
is viewed as a topography over the phase space, then this gratefully acknowledge many valuable discussions with
eigenfunction only carries information about the altitude, Josh Proctor about Koopman theory and extensions to con-
and not the location. In contrast, Lan and Mezić [36] show trol. We would also like to acknowledge Igor Mezić, Maria
that it is possible to extend the Hartman-Grobman theo- Fonoberova, Bernd Noack, Clancy Rowley, and Sam Taira.
rem to the entire basin of attraction of certain fixed points
References
and periodic orbits, providing a local linear embedding of [1] I. Abraham, G. De La Torre, and T. D. Murphey. Model-based
the dynamics. Connecting these perspectives will continue control using Koopman operators. arXiv:1709.01568, 2017.
to yield interesting and important advances in Koopman [2] H. Arbabi and I. Mezić. Ergodic theory, dynamic mode
decomposition and computation of spectral properties of the
theory. In addition, there are known connections between Koopman operator. arXiv preprint arXiv:1611.06664, 2016.
the eigenvalues of geometric structures in phase space and [3] K. B. Ariyur and M. Krstić. Real-Time Optimization by
the spectrum of the Koopman operator [47]. This knowledge Extremum-Seeking Control. Wiley, 2003.
[4] B. Bamieh and L. Giarré. Identification of linear parameter
may guide the accurate identification of Koopman eigen- varying models. Int. J. Robust and Nonlinear Control, 12:841–
functions using sparsity-promoting regression techniques. 853, 2002.
Formulating control in terms of Koopman eigenfunc- [5] S. C. Beeler and D. E. Cox. State-dependent Riccati equation
regulation of systems with state and control nonlinearities.
tions motivates additional work to understand how control- NASA/CR-2004-213245, 2004.
lability and observability in these coordinates relate to prop- [6] S. A. Billings. Nonlinear system identification: NARMAX
erties of the nonlinear system. The degree of observability methods in the time, frequency, and spatio-temporal domains.
John Wiley & Sons, 2013.
and controllability will generally vary with different eigen- [7] R. W. Brockett. Volterra series and geometric control theory.
functions, so that it may be possible to obtain balanced re- Automatica, 12(2):167–176, 1976.
alizations. Moreover, classic results, such as the PBH test, [8] S. L. Brunton, B. W. Brunton, J. L. Proctor, E. Kaiser, and J. N.
Kutz. Chaos as an intermittently forced linear system. Nature
indicate that multi-channel actuation may be necessary to Commun., 8(19), 2017.
simultaneously control different eigenfunctions correspond- [9] S. L. Brunton and B. R. Noack. Closed-loop turbulence control:
ing to the same eigenvalue, such as the Hamiltonian en- Progress and challenges. Appl. Mech. Rev., 67:050801–1–050801–
48, 2015.
ergy and conserved angular momentum. The additional de- [10] S. L. Brunton, J. L. Proctor, and J. N. Kutz. Discovering
grees of freedom arising from multi-channel inputs can also governing equations from data by sparse identification of
be used for eigenstructure assignment to shape Koopman nonlinear dynamical systems. Proc. Natl. Acad. Sci,
113(15):3932–3937, 2016.
eigenfunctions [23]. Thus, actuation may modify both the [11] S. L. Brunton, B. W. Brunton, J. L. Proctor, and J.N. Kutz.
shape of coherent structures (i.e., Koopman modes associ- Koopman invariant subspaces and finite linear representations
ated with a particular eigenfunction) and their time dynam- of nonlinear dynamical systems for control. PLoS ONE,
11(2):e0150171, 2016.

9
[12] M. Budišić and I. Mezić. Geometry of the ergodic quotient reveals [43] I. Mezić. Analysis of fluid flows via spectral properties of the
coherent structures in flows. Physica D, 241(15):1255–1269, 2012. Koopman operator. Ann. Rev. Fluid Mech., 45:357–378, 2013.
[13] M. Budišić, R. Mohr, and I. Mezić. Applied Koopmanism. Chaos, [44] I. Mezić. On applications of the spectral theory of the Koopman
22(4):047510, 2012. operator in dynamical systems and control theory. In IEEE Conf.
Decision and Control (CDC), pages 7034–7041. IEEE, 2015.
[14] E. F. Camacho and C. B. Alba. Model predictive control.
Springer, 2013. [45] I. Mezić. Spectral operator methods in dynamical systems:
[15] B. Charlet, J. Lévine, and R. Marino. On dynamic feedback Theory and applications. Springer, 2017.
[46] I. Mezić and A. Banaszuk. Comparison of systems with complex
linearization. Systems & Control Letters, 13(2):143–151, 1989. behavior. Physica D, 197(1):101–133, 2004.
[16] R. Chartrand. Numerical differentiation of noisy, nonsmooth [47] Igor Mezic. Koopman operator spectrum and data analysis.
data. ISRN Appl. Math., 2011, 2011. arXiv:1702.07597, 2017.
[17] M. Dellnitz, G. Froyland, and O. Junge. The algorithms behind [48] O. Nelles. Nonlinear system identification: from classical
GAIO—set oriented numerical methods for dynamical systems. approaches to neural networks and fuzzy models. Springer, 2013.
In Ergodic theory, analysis, and efficient simulation of dynamical [49] F. Noé and F. Nuske. A variational approach to modeling slow
systems, pages 145–174. Springer, 2001. processes in stochastic dynamical systems. Multiscale Modeling
[18] M. Dellnitz and O. Junge. Set oriented numerical methods for & Simulation, 11(2):635–655, 2013.
dynamical systems. Handbook of dynamical systems, 2:221–264,
2002. [50] F. Nüske, R. Schneider, F. Vitalini, and F. Noé. Variational
[19] G. E. Dullerud and F. Paganini. A course in robust control tensor approach for approximating the rare-event kinetics of
theory: A convex approach. Texts in Applied Mathematics. macromolecular systems. J. Chemical Physics, 144(5):054105,
Springer, Berlin, Heidelberg, 2000. 2016.
[20] C. Edwards and S. Spurgeon. Sliding mode control: theory and [51] J. D. Pearson. Approximation methods in optimal control i. Sub-
applications. Crc Press, 1998. optimal control. I. J. Electronics, 13(5):453–469, 1962.
[21] J. Guckenheimer and P. Holmes. Nonlinear oscillations, [52] J. L.
dynamical systems, and bifurcations of vector fields, volume 42 Proctor, S. L. Brunton, and J. N. Kutz. Generalizing Koopman
of Applied Mathematical Sciences. Springer, 1983. theory to allow for inputs and control. arXiv:1602.07647, 2016.
[22] N. Hansen, A. S. P. Niederberger, L. Guzzella, and [53] Q. Qu, J. Sun, and J. Wright. Finding a sparse vector in a
P. Koumoutsakos. A method for handling uncertainty in subspace: Linear sparsity using alternating directions. In Adv.
evolutionary optimization with an application to feedback control Neural Inform. Process. Syst. 27, pages 3401–3409, 2014.
of combustion. IEEE Trans. Evol. Comput., 13(1):180–197, 2009. [54] C. W. Rowley, I. Mezić, S. Bagheri, P. Schlatter, and D.S.
[23] M. Hemati and H. Yao. Dynamic mode shaping for fluid flow Henningson. Spectral analysis of nonlinear flows. J. Fluid Mech.,
control: New strategies for transient growth suppression. In 8th 645:115–127, 2009.
AIAA Theoretical Fluid Mechanics Conference, page 3160, 2017. [55] W. J. Rugh and J. S. Shamma. Research on gain scheduling.
[24] A. Isidori. Nonlinear control systems. Springer, 2013. Automatica, 36(10):1401–1425, 2000.
[25] E. Jonckheere and L. Silverman. A new set of invariants for linear [56] S. S. Sastry. Nonlinear systems: analysis, stability, and control,
systems–Application to reduced order compensator design. IEEE volume 10. Springer, 2013.
Trans Autom Control, 28(10):953–964, 1983. [57] P. J. Schmid. Dynamic mode decomposition of numerical and
experimental data. J. Fluid Mech., 656:5–28, 2010.
[26] E. Kaiser, J. N. Kutz, and S. L. Brunton. Data-driven [58] S. Shadden, F. Lekien, and J. Marsden. Definition and properties
discovery of Koopman eigenfunctions for control. arXiv Preprint of Lagrangian coherent structures from finite-time Lyapunov
arXiv:1707.01146, 2017. exponents in two-dimensional aperiodic flows. Physica D,
[27] Y. Kawahara. Dynamic mode decomposition with reproducing 212:271–304, 2005.
kernels for Koopman spectral analysis. In Advances in Neural [59] S. Skogestad and I. Postlethwaite. Multivariable feedback control:
Information Processing Systems, pages 911–919, 2016. analysis and design. John Wiley & Sons, Inc., 2 edition, 2005.
[28] H. K. Khalil. Noninear Systems. Prentice-Hall, 1996. [60] T. H. Solomon and J. P. Gollub. Chaotic particle transport
[29] S. Klus, P. Koltai, and C. Schütte. On the numerical in time-dependent Rayleigh-Bénard convection. Phys. Rev. A,
approximation of the Perron-Frobenius and Koopman operator. 38:6280–6286, 1988.
J. Comp. Dyn., 3(1):51–79, 2016. [61] A. Sootla, A. Mauroy, and D. Ernst. An optimal control
[30] P. V. Kokotovic. The joy of feedback: nonlinear and adaptive. formulation of pulse-based control using Koopman operator.
IEEE Control systems, 12(3):7–17, 1992. arXiv:1707.08462, 2017.
[31] P. V. Kokotovic, R. E. O’malley, and P. Sannuti. Singular [62] R. F. Stengel. Optimal control and estimation. Courier Corp.,
perturbations and order reduction in control theory—an 2012.
overview. Automatica, 12(2):123–132, 1976. [63] A. Surana. Koopman operator based observer synthesis for
[32] B. O. Koopman. Hamiltonian systems and transformation in control-affine nonlinear systems. In 55th IEEE Conf. Decision
and Control (CDC, pages 6492–6499, 2016.
Hilbert space. Proc. Natl. Acad. Sci, 17(5):315–318, 1931.
[64] A. Surana and A. Banaszuk. Linear observer synthesis for
[33] M. Korda and I. Mezić. Linear predictors for nonlinear dynamical nonlinear systems using Koopman operator framework. IFAC-
systems: Koopman operator meets model predictive control. PapersOnLine, 49(18):716–723, 2016.
arXiv:1611.03537, 2016.
[34] M. Krstić, I. Kanellakopoulos, and P. V. Kokotović. Nonlinear [65] Y. Susuki and I. Mezić. A prony approximation of Koopman
and adaptive control design. Wiley, 1995. mode decomposition. IEEE Conf. Decision and Control (CDC),
[35] J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L. pages 7022–7027, 2015.
Proctor. Dynamic Mode Decomposition: Data-Driven Modeling [66] R. S. Sutton and A. G. Barto. Reinforcement Learning: An
of Complex Systems. SIAM, 2016. Introduction. The MIT Press, 1998.
[36] Y. Lan and I. Mezić. Linearization in the large of nonlinear [67] J. H. Tu, C. W. Rowley, D. M. Luchtenburg, S. L. Brunton,
and J. N. Kutz. On dynamic mode decomposition: theory and
systems and Koopman operator spectrum. Physica D, 242(1):42– applications. J. Comp. Dyn., 1(2):391–421, 2014.
53, 2013.
[37] L. Ljung. System Identification: Theory for the User. Prentice [68] Wallace E Vander Velde. Multiple-input describing functions and
Hall, 1999. nonlinear system design. McGraw-Hill, New York, 1968.
[38] N. M. Mangan, S. L. Brunton, J. L. Proctor, and J. N. Kutz. [69] M. O. Williams, M. S. Hemati, S. T. M. Dawson, I. G. Kevrekidis,
Inferring biological networks by sparse identification of nonlinear and C. W. Rowley. Extending data-driven Koopman analysis to
dynamics. IEEE Trans. Molecular, Biological, and Multi-Scale actuated systems. IFAC-PapersOnLine, 49(18):704–709, 2016.
Commun., 2(1):52–63, 2016. [70] M. O. Williams, I. G. Kevrekidis, and C. W. Rowley. A
[39] A Mauroy and J Goncalves. Linear identification of nonlinear data-driven approximation of the Koopman operator: extending
systems: A lifting technique based on the Koopman operator. dynamic mode decomposition. J. Nonlin. Sci., 2015.
IEEE Conf. Decision and Control (CDC), pages 6500–6505, [71] M. O. Williams, C. W. Rowley, and I. G. Kevrekidis. A kernel
2016. approach to data-driven Koopman spectral analysis. J. Comp.
[40] A. Mauroy and I. Mezić. Global stability analysis using the Dyn., 2(2):247–265, 2015.
eigenfunctions of the Koopman operator. IEEE Trans Autom [72] D. Wilson and J. Moehlis. An energy-optimal methodology for
Control, 61:3356–3369, 2016. synchronization of excitable media. SIAM J. Appl. Dyn. Sys.,
[41] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert. 13(2):944–957, 2014.
Constrained model predictive control: Stability and optimality.
Automatica, 36(6):789–814, 2000.
[42] I. Mezić. Spectral properties of dynamical systems, model
reduction and decompositions. Nonlinear Dynamics, 41(1-
3):309–325, 2005.

10