[
Abstract
A datadriven, kernelbased method for approximating the leading Koopman eigenvalues, eigenfunctions, and modes in problems with highdimensional state spaces is presented. This approach approximates the Koopman operator using a set of scalar observables, which are functions that map states to scalars, that is determined implicitly by the choice of a kernel function. This circumvents the computational issues that arise due to the number of basis functions required to span a “sufficiently rich” subspace of the space of scalar observables in such applications. We illustrate this method on two examples: the FitzHughNagumo PDE, a prototypical onedimensional reactiondiffusion system, and vorticity data obtained from experimentally obtained velocity data for flow over a cylinder at Reynolds number 413. In both examples, we compare our results with related methods, such as Dynamic Mode Decomposition (DMD) that have the same cost as our approach.
KernelBased Methods for Koopman Spectral Analysis] A KernelBased Method
for DataDriven Koopman Spectral Analysis
M.O. Williams and C.W. Rowley and I.G. Kevrekidis]
\subjclassPrimary: 37M10, 65P99, 47B33
Matthew O. Williams
Program in Applied and Computational Mathematics (PACM)
Princeton University
Princeton, NJ 08544, USA
Clarence W. Rowley
Department of Mechanical and Aerospace Engineering
Princeton University
Princeton, NJ 08544, USA
Ioannis G. Kevrekidis
Department of Chemical and Biological Engineering & PACM
Princeton University
Princeton, NJ 08544, USA
(Communicated by the associate editor name)
1 Introduction
In many applications, the evolution of complex spatiotemporal phenomena can be characterized using models based on the interactions between a relatively small number of modes. Due to the availability of data and computational power, algorithmic techniques for identifying such modes have become increasingly common [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]. Perhaps the best known method is the Proper Orthogonal Decomposition (POD) [1, 2, 15], which is also known as Principal Component Analysis (PCA) [15, 14]. However, other algorithms that generate different sets of modes exist. In recent years, approximations of the modes of the Koopman operator [16, 17] have become popular in fluid applications [5, 6, 18, 19, 10, 8, 20]. The Koopman modes of the full state observable, which are vectors for systems of ODEs or spatial profiles for PDEs, are intrinsic to a particular evolution law, and have temporal dynamics that are determined by their corresponding Koopman eigenvalues [5]; in effect, the Koopman modes look and act like the eigenvectors of a linear system even when the underlying evolution law is nonlinear.
This representation is possible because the Koopman operator, which defines these quantities, is a linear operator and therefore can have eigenfunctions and eigenvalues, which are later used to define the modes. However, it acts on scalar observables, which are functions defined on state space, and is therefore an infinitedimensional operator even for finitedimensional dynamical systems. As a result, methods that approximate the Koopman operator (often implicitly) select a finitedimensional subspace of the space of scalar observables to use during the computation. Currently, the most widely used method is Dynamic Mode Decomposition (DMD) [21, 3, 4, 5, 6] (along with its related modifications [10, 8, 20]), which implicitly select linear functions of the state as the basis functions [7]. This restrictive choice of basis allows DMD to be applied to large systems of ODEs or PDEs, but in many applications, this subspace is simply not “rich” enough to effectively approximate the Koopman operator [7, 11].
Other methods, such as Extended DMD [7], use an expanded set of observables, which can produce a more accurate approximation of the set of Koopman eigenvalues, eigenfunctions, and modes for nonlinear systems. For small systems of ODEs, a “sufficiently rich” subspace can often be spanned using a few hundred or thousand basis functions, which is computationally tractable even on a laptop. However, the size of the needed basis grows rapidly as the dimension of the state space increases, so the necessary computations quickly become infeasible. This explosion in the computational cost is common in machine learning applications, and one facet of the “curse of dimensionality” [15].
In this manuscript, we introduce a datadriven, kernelbased method to approximate the Koopman operator in systems with large state dimension. This approach circumvents the dimensionality issues encountered by Extended DMD by defining a kernel function that implicitly computes inner products in the highdimensional space of observables. Because we do not form this space explicitly, the computational cost of the method is determined by the number of snapshots and the dimension of state space rather than the number of basis functions used to represent the scalar observables; therefore, the cost of this approach is comparable to that of DMD. We apply this approach to a pair of sample problems: the onedimensional FitzHughNagumo PDE, a prototypical reactiondiffusion system; and experimentally obtained vorticity data from flow over a cylinder obtained at Reynolds number 413. Using the FitzHughNagumo PDE example, we will demonstrate that using a higherdimensional subspace of observables can result in more accurate and reproducible approximations of the leading Koopman modes, eigenvalues, and eigenfunctions. Finally, the cylinder example demonstrates that this approach has practical advantages in more realistic settings where the data are noisy and the “true” Koopman modes and eigenvalues are unknown.
The remainder of the manuscript is outlined as follows: in Sec. 2 we present the kernel reformulation of Extended Dynamic Mode Decomposition. In Sec. 3, we apply it to a numerical discretization of the FitzHughNagumo PDE in one spatial dimension, where a subset of the Koopman eigenvalues and modes are known. In Sec. 4, we apply our method to experimentally obtained data of flow over a cylinder, which is a more realistic example with measurement noise and other experimental realities. Finally, in Sec. 5, we present some concluding remarks and future outlook.
2 A DataDriven Approximation of the Koopman Operator
In this section, we present a datadriven approach for approximating the Koopman operator that can be applied to systems with highdimensional state spaces. The method presented here is a reformulation of the Extended DMD procedure that makes use of the socalled kernel trick [22]. In the following subsections, we will (i) briefly review the Koopman operator (ii) give a brief derivation of “standard” Extended DMD, which enables the “tuples” of Koopman eigenvalues, eigenfunctions, and modes to be approximated from data; (iii) present the kernel approach; and (iv) give a practical algorithm for computing the leading tuples of eigenvalues, eigenfunctions, and modes.
2.1 The Koopman Operator
The Koopman operator [16, 17], along with its eigenvalues, eigenfunctions, and modes, is defined by a dynamical system and not a set of data. Given the discrete time dynamical system , where is time, is the state space, and defines the dynamics, the action of the Koopman operator, , on a scalar observable, , is
(1) 
Intuitively, is a new function that gives the value of “one step in the future”. Note that the Koopman operator acts on functions of the state, and not the states themselves. Since , where is an appropriate space of scalar observables, the Koopman operator is infinite dimensional. However, it is also linear, and thus it can have eigenvalues and eigenfunctions, which we refer to as Koopman eigenvalues and eigenfunctions. Accompanying the eigenvalues and eigenfunctions are the Koopman modes for a given vector valued observable, , where . These modes are vectors in a system of ODEs (or spatial profiles in a PDE) that contain the coefficients required to construct using a Koopman eigenfunction basis [5, 7]. One particularly useful set of modes is that of the identity operator or, equivalently, the full state observable, , which we refer to as simply the Koopman modes in all that follows.
In many systems [5, 23, 18], tuples consisting of an eigenvalue, an eigenfunction, and a mode enable a simple yet powerful means of representing the system state and making predictions of future values. In particular,
(2) 
where is the Koopman mode corresponding to the eigenfunction , is the corresponding eigenvalue, and is the number of tuples required for the reconstruction, which could be infinite. The eigenfunction is a function, but the mode is a vector. From (2), the coefficient associated with the th mode, , is obtained by evaluating the th eigenfunction, , and the temporal evolution is dictated by . Ultimately, one benefit of this Koopmanbased approach is that the eigenvalues, eigenfunctions, and modes are intrinsic to the dynamical system; in the subsequent section they will be approximated using data, but unlike the Proper Orthogonal Decomposition (POD) and related methods, they still exist absent any data.
2.2 Extended Dynamic Mode Decomposition
Extended Dynamic Mode Decomposition (Extended DMD) [7] is a regression procedure whose solution produces a finitedimensional approximation of the Koopman operator. To obtain this approximation, we define a basis set that consists of scalar observables, which we denote as for , that span . We also define the vector valued observable, , where
(3) 
In this application, is the mapping from physical space to feature space. Any can be written as
(4) 
for some set of coefficients . Although is unknown, we assume access to a data set of snapshot pairs:
(5) 
where . One important special case of such a data set is a single timeseries of data, which can be written in the above form by “grouping” sequential pairs of snapshots that were obtained with a fixed sampling interval, say .
Let , where is a residual function that appears because is not necessarily invariant to the action of the Koopman operator. Using the notation in (4), the objective of the Extended DMD procedure [7] is to define a mapping from some given to a new vector that minimizes this residual. Because the Koopman operator is linear, this mapping can be represented by a matrix . To determine the entries of , the Extended DMD approach takes ideas from collocation methods typically used to solve PDEs, but uses the as the collocation points rather than a predetermined grid [24, 25]. As a result, the finite dimensional approximation is
(6a)  
where  
(6b) 
are in , and denotes the pseudoinverse.
In Williams et al. [7], the relationship was used to rewrite (6) as
(7a)  
(7b) 
which is advantageous when the number of snapshots is much larger than the number of basis functions, i.e., , because each term in the sum can be evaluated individually. As a result, one need only store the matrices rather than , which would be much larger in this regime.
Expressions (6) and (7) are mathematically equivalent, and produce the same matrix . Therefore, regardless of how it was computed, the properties of of interest here are unchanged:

The th eigenvalue of , , is an approximation of an eigenvalue of . When the data are generated by sampling a continuous time dynamical system with a fixed sampling interval , we also define the approximation of the continuous time eigenvalue as .

Using (4), the corresponding eigenvector, , approximates an eigenfunction of the Koopman operator via
(8) 
The left eigenvector, , can be used to approximate the Koopman mode, .
For a derivation of these relationships, see Williams et al. [7].
2.3 The Kernel Method
In Sec. 2.2, we considered the case where the number of snapshots was large compared to the dimension of the desired subspace of functions. Now we will consider the opposite and more commonly encountered regime [4, 3, 21, 5, 6, 18, 19, 10, 8, 20] where the number of snapshots is small compared to the dimension of our subspace of scalar observables (i.e., ).
The difficulty here is that the Extended DMD procedure, as formulated in Sec. 2.2, requires a matrix to be formed and decomposed, which requires and time respectively, and the value of for a “rich” set of basis functions grows rapidly as the dimension of state space increases. For instance, consider the case where is the space of all (multivariate) polynomials on with degree up to 20, as it will be in our first example discussed in Section 3. In this case, , which is far too large for practical computations [15]. This explosion in the size of the problem is common, and is one facet of the curse of dimensionality.
Because the matrix is the solution to a regression problem, the nonzero eigenvalues and their associated left and right eigenvectors can also be obtained by solving the dual form of this problem [15]. To show this, note that , i.e., the range of contains the range of . If we could compute the SVD of ,
(9) 
where and , then an eigenvector of with could be written as for some . With some simple algebraic manipulations, the eigenvalue problem can be written as:
Therefore, an alternative method for computing an eigenvector of is to form the matrix
(10) 
where , compute an eigenvector of , say , and set . Here , so the computational cost of the decomposition is determined by the number of snapshots rather than the dimension of the system state or “feature” space.
The benefit of the expression in is that all the required matrices can be obtained by computing inner products in feature space. In addition to , we define the matrix . Despite appearing “flipped,” the th elements of and are
(11) 
However , using the definition of and in (9). Therefore, given we can obtain and via its eigendecomposition, which is simply the method of snapshots approach for computing the POD modes from snapshot data [26]. As a result, we could compute by forming and using (11) in time. This is a large improvement over “standard” Extended DMD, but still impractical as can be extremely large.
Rather than explicitly defining the function and computing the entries of directly, the kernel trick is a common technique for implicitly computing inner products [27, 28, 22, 29], which can be used to assemble in time. Instead of defining , we define a kernel function that computes inner products in feature space given pairs of data points; that is, [27]. In effect, the choice of defines , which is equivalent to choosing the basis in Extended DMD. It is, however, crucial to note that does not compute these inner products directly. The simplest example is the polynomial kernel
(12) 
with , which, when expanded, is
if . In general, a polynomial kernel of the form
(13) 
is equivalent to a basis that can represent all polynomials up to and including terms of degree , and takes only time to evaluate.
In the example that follows, we use a polynomial kernel with . We selected a polynomial kernel because the Koopman eigenfunctions are often analytic in a disk about a fixed point [30], and polynomial kernels mimic an order powerseries expansion. The specific choice of is more ad hoc. In general, large values of use a “richer” set of basis functions, but also deleteriously impact the condition number of . Other choices of kernels, such as Gaussian kernels (i.e., ), are also used in machine learning applications [15, 31, 22], and may result in better performance (both in terms of numerical conditioning and in approximating the Koopman operator) [32]. The optimal choice of kernel, which is equivalent to the ideal choice of the basis set, remains an open question.
Ultimately, the procedure for obtaining an approximation of the Koopman operator is as follows:

Using the data set of snapshot pairs and the kernel function, , compute the elements of and using:
(14) 
Compute the eigendecomposition of the Gramian to obtain and .

Construct using (10).
The difference between (11) and (14) is that the former defines explicitly, and as a result, computes the set of the needed inner products in time, while the latter defines implicitly, which allows the inner products to be computed in time. If a linear kernel, is chosen, the kernel approach outlined here is identical to DMD [11], where the subspace of observables used to approximate the Koopman eigenfunctions is chosen using the Proper Orthogonal Decomposition. The overarching idea here is that the kernel approach chooses this subspace differently by using what is in effect Kernel Principal Component Analysis [15], which exploits the kernel trick to make the computation efficient when .
2.4 Computing the Koopman Eigenvalues, Modes, and Eigenfunctions
In this subsection, we show how to approximate the Koopman eigenvalues, modes, and eigenfunctions given . Let be the matrix whose columns are the eigenvectors of . Then using (8), we define the matrix of eigenfunctions values:
(15)  
where is the diagonal matrix of singular values with any entry neglected in the pseudoinverse set to zero. The th row of contains the numerically computed eigenfunctions evaluated at . The th numerically approximated Koopman eigenfunction can also be evaluated at a new data point via:
(16)  
using the same arguments as in (15) though without the simplifications and cancellations that occur in that case.
To compute the Koopman modes, we use (2), which when evaluated at each of the data points, results in the matrix equation
(17) 
where
(18) 
provided that is full rank. In this case,
(19) 
where is a left eigenvector of scaled so that . This implies that the th Koopman mode, , is
(20) 
and therefore approximate Koopman tuples can be obtained via the left and right eigenvectors of , and do not require to be computed in its entirety. For the problems considered here, the complete decomposition of is computed, but for problems with larger numbers of snapshots, Krylov methods could be used to compute a leading subset of its eigenvalues and vectors [33].
3 Example: The FitzHughNagumo PDE
In order to highlight the effectiveness of the kernel method, we first apply it and, as a benchmark, DMD to the FitzHughNagumo PDE [34] in one spatial dimension. This example is particularly useful, because a subset of the true Koopman eigenvalues and modes can be deduced from the system linearization. The governing equations are:
(21a)  
(21b) 
where is the activator field, is the inhibitor field, , , , , for with Neumann boundary conditions. Both and are approximated using a discrete cosine transformbased spectral method with 128 basis functions each.
With these parameter values, (21) has a stable equilibrium point that resembles a standing wavefront in both the activator and inhibitor fields. In order to explore a subset of state space and to mimic the presence of impulsive actuation, every 25 time units we perturb the system state by setting
(22) 
where , , and . Note that remains unchanged. Starting from the equilibrium point, we generate five separate trajectories consisting of 2500 snapshots (or 2499 snapshot pairs) with a sampling interval of starting with five different random seeds. Figure 1 shows the and data for a prototypical trajectory: the perturbations applied even 25 units in time excite “fast” directions that decay almost immediately (and thus are not visible in the figure) as well as “slow” directions that take the form of “shifts” in the position of the wave front.
To compute the Koopman eigenvalues, eigenfunctions, and modes, we use a polynomial kernel with . Furthermore, we scale each data set so that the mean value of , which is equivalent to introducing a scaling parameter in the kernel . For these data sets, both (for DMD) and (for the kernel method) have small singular values, so when we compute the pseudoinverse of (for DMD) or (for the kernel method), we retain the 150 largest singular values and set the others to zero. This is equivalent to projecting and onto the 150 leading POD modes of or and onto the leading 150 kernel principal components of before carrying out the computation. In our experience, this often helps to avoid the spurious unstable eigenvalues that both DMD and Extended DMD can produce.
Figure 2 shows the approximation of the continuoustime eigenvalues obtained for the five different data sets using this methodology. Because we used truncated SVDs for computing pseudoinverses, both DMD and the kernel method produce eigenvalues that are, up to numerical errors, contained in the left half plane. Furthermore, both DMD and the kernel method consistently identify the eigenvalues with and though, as shown in the figure, the variance in this pair is smaller when the kernel method is used. Although DMD identifies other eigenvalues with more negative real parts, no other eigenvalues lie in the subset of the complex plane shown in the figure. The kernel method, which uses a “richer” set of basis functions to approximate the Koopman eigenfunctions, also identifies two additional layers of the “pyramid” of eigenvalues that the Koopman operator should possess [36]. As a result, although the numerically computed eigenvalues are truly data dependent, the kernel approach: (i) appears to approximate the leading eigenvalues of the Koopman operator with less variance than the DMD benchmark, and (ii) identifies additional Koopman eigenvalues the DMD benchmark appears to neglect.
The part of the modes corresponding to four of the leading eigenvalues are shown in Fig. 3. The first mode, which has , is a scaled version of the solution at the fixed point where the appropriate scaling factor is determined by the corresponding Koopman eigenfunction. The error, i.e., with the normalization , was less than 0.01 for the five data sets used in the computation (the mean value of the error was ). The next pair of modes (modes 2 and 3), which have , can be shown to be the “slowest” pair of eigenvectors of the system linearization, and provide a useful description of the evolution near the fixed point. The error in these modes was less than 0.02 for the five data sets (with a mean value of ). As a result, not only does the kernel method consistently identify the leading Koopman modes in this problem, but it is accurate as well.
The first mode associated with nonlinear effects, which has the eigenvalue , is shown in the third image of Fig. 3. As shown in the figure, the kernel method consistently identifies this mode, though there are differences in the mode shape that are visible to the eye if one “zooms in” near in the figure. In general, both DMD and the kernel method become less accurate the further “down” (i.e., into the left halfplane) one goes, as is demonstrated by the slight yet visible differences in the 7th mode with shown in the bottom panel of Fig. 3. Intuitively, this is because the corresponding eigenfunctions tend to become more complicated functions; in this example, the eigenfunctions corresponding to each subsequent layer of the pyramid are higher powers of the leading two eigenfunctions. As a result, they are less likely to lie in or near the subspace spanned by our implicitly chosen basis set, and therefore less accurately computed.
In this section, we applied the kernel approach to the FitzHughNagumo PDE, and found that it consistently identified a subset of the Koopman modes and eigenvalues including those that are not associated with the system linearization. The first three modes, where the true solutions happen to be known, are identified accurately, with a relative error of less than 2%. Given a sufficiently large amount of data and the proper choice of kernel, the this approach can also identify additional “layers” in the pyramid of Koopman eigenvalues, but here the variance in the numerically obtained eigenvalues and modes is larger, which is likely due to the fact the eigenfunctions of interest to not lie entirely within the span of the basis functions used to represent them. As a result, at a certain point the method transitions from yielding quantitatively accurate modes and eigenvalues to a more qualitative description of the underlying dynamics.
4 Example: Experimental Data of Flow Over a Cylinder
In this example, we apply the kernel approach to experimentally obtained PIV data for flow over a cylinder taken at Reynolds number 413, used in previous work by Tu et al. [37] and Hemati et al. [20]. The experiment was conducted in a water channel designed to minimize threedimensional effects, and the computed velocity field sampled at 20 Hz at a resolution of pixels [37]. From these snapshots of the velocity field, we compute the vorticity via standard finite difference methods. In the original manuscripts, a total of 8000 snapshots of the velocity field were obtained, but here we use 3000 of them in our computation.
This example is intended to highlight how we foresee the kernel approach being used in practice. Due to the experimental nature of the data and the presence of measurement and process noise, the true Koopman eigenvalues and modes are not known. However as we will show shortly, the numerically computed Koopman modes and eigenvalues have more of the properties that the true eigenvalues should have, which implies the kernel method is producing a more accurate approximation than DMD. As a result, our objective here is to demonstrate the practical advantages of the kernel method, rather than the accuracy of the resulting modes and eigenvalues.
As before, we use a polynomial kernel with , scale the data such that the mean value of , and use (10) to approximate the Koopman eigenvalues and modes. From previous efforts including Tu et al. [37] and Hemati et al. [20], an effective set of modes and eigenvalues can be obtained from this data using DMD and similar methods. Figure 4 shows the leading eigenvalues obtained using the kernel method and, for reference, those computed using DMD. In both cases, all 2999 singular values were above machine epsilon, and were retained when computing pseudoinverses. The practical advantage of the kernel approach is that the eigenvalues and modes of interest can be identified visually; in particular, they are the “most slowly decaying” eigenvalues, which could be computed efficiently even in large systems using Krylov methods. On the other hand, DMD has a “cloud” of both stable and unstable eigenvalues, and the modes known to be most important are not necessarily those that decay the slowest. We should note that a “useful” set of eigenvalues are contained in this cloud, but energy [37] or sparsitypromoting [38] methods must be used in order to select this subset of the eigenvalues. To do this, however, both of these methods require all of eigenvalues and modes to be computed, and then truncate this “complete” set, which can become computationally expensive in systems with larger numbers of snapshots and, therefore, modes to choose from.
The other eigenvalues computed by the kernel method decay more rapidly than those pictured in Fig. 4, and no eigenvalues have a positive real part (up to roundoff errors). Conceptually, the underlying system possesses a limit cycle with the addition of both process and measurement noise. In the absence of noise, the Koopman eigenvalues would lie on the imaginary axis and in regular intervals in the left halfplane [36]. The addition of noise causes many of the eigenvalues to decay (or decay more rapidly) [23, 39, 7], as the distribution of trajectories approaches the invariant distribution that is concentrated near the limit cycle. However, this decay rate depends upon the amount and type of noise added to the system, so while some decay is expected, it is unclear whether or not these eigenvalues are accurate.
The real part of three of the leading Koopman modes are shown in Fig. 5. Both the shape and angular frequency (i.e., ) of these modes are similar to those obtained using the full set of 8000 snapshots in Tu et al. [37] or Hemati et al. [20]. In particular, corresponds to an oscillation frequency of 0.90 Hz, to a frequency of 1.79 Hz, and to 2.71 Hz. In Hemati et al. [20], these frequencies were reported as 0.89 Hz, 1.77 Hz, and 2.73 Hz respectively. The modes shown in Fig. 5 were obtained from the left eigenvectors of the matrix using (20), which in this example were obtained by completely decomposing , but could also have been obtained by computing the leading lefteigenvectors using the implicitly restarted Arnoldi method [33], which would be more efficient if a larger number of snapshots were available.
In this section, we applied the kernel method to experimental data and computed the leading Koopman modes. Unlike the previous example, where the kernel approach produced modes that DMD could not, in this example both methods produce similar sets of modes. The difference, however, is that the “important” modes identified by Tu et al. [37] and Hemati et al. [20] using an energybased method, are the clearlyvisible leading (i.e., most slowlydecaying) modes of the kernel method. As a result, rather than computing all of the Koopman modes and then truncate, one could compute only the leading Koopman modes via (20). Furthermore, the eigenvalues identified by the kernel method possess more of the properties that the Koopman eigenvalues should have; in particular, they lie on the imaginary axis or the left halfplane. Therefore, even though we cannot definitively say that the kernel method has accurately identified the Koopman eigenvalues, there are conceptual and practical advantages to using the kernel method in lieu of standard DMD even in experimental applications like the one shown here.
5 Conclusions
We have presented a datadriven method for approximating the Koopman operator in problems with large state spaces, which commonly occurs in physical applications such as fluid dynamics. The kernel method we have developed defines a subspace of scalar observables implicitly through a kernel function, which allows the method to approximate the Koopman operator with the same asymptotic cost as DMD, but with a far larger set of observables. In this manuscript, we used a polynomial kernel, whose associated set of basis functions can represent th order polynomials. However, many other kernels are available, associated with other basis sets; as a result, the choice of kernel, like the choice of a basis set in Extended DMD, is important but user determined.
To highlight the effectiveness of the kernel method, we applied it to a pair of illustrative examples. The first is the FitzHughNagumo PDE in one spatial dimension, where a subset of the Koopman eigenvalues and modes could be deduced by linearizing the system about an equilibrium point. For this subset, the kernel method consistently and accurately identified the leading Koopman eigenvalues and modes. However, it also accurately computed additional Koopman eigenvalues that are not associated with the system linearization and, although we do not have an analytical expression for the corresponding modes, consistently identified a mode shape.
In practical applications, the data often come from systems whose true Koopman eigenvalues (or related information such as the location and linearization about fixed points) are unknown. Our second example, which used experimental data from flow over a cylinder, was a more realistic example of how we envision the kernel approach being applied. In that example, the numerically obtained eigenvalues had more of the properties the Koopman eigenvalues should have, but it was unclear whether or not they were quantitatively accurate because the distribution and nature of the noise in this example was unknown. The primary advantage of the kernel method here was that the Koopman modes of interest were associated with the eigenvalues that decay the slowest. As a result, they could be computed using the leading lefteigenvectors of the finite dimensional approximation, rather than requiring the complete set of eigenvectors and eigenvalue be computed and then truncated using energybased or sparsity promoting methods.
In the end, the Koopman operator is an appealing mathematical framework for defining a set of spatial modes because these modes are intrinsic to the dynamics rather than associated with a set of data. However, obtaining effective approximations of this operator is nontrivial, particularly when all that is available are data. One method for obtaining such an approximation is Extended Dynamic Mode Decomposition [7], which is, for certain choices of basis functions, only computationally feasible in problems with small state spaces. The kernel method presented here is conceptually identical to Extended Dynamic Mode Decomposition, but computationally feasible even in large systems; in particular, the asymptotic cost of the kernel method is identical to that of “standard” Dynamic Mode Decomposition, and therefore, it can be used anywhere DMD is currently being used. Like most existing datadriven methods, there is no guarantee that the kernel approach will produce accurate approximations of even the leading eigenvalues and modes, but it often appears to produce useful sets of modes in practice if the kernel and truncation level are chosen properly. As a result, approaches like the kernel method presented here are a first step towards the ultimate goal of enabling the conceptual tools provided by Koopman spectral analysis to be used in practical applications.
Acknowledgments
The authors acknowledge support from NSF (awards DMS1204783 and CDSE1310173) and AFOSR.
References
 Holmes et al. [2012] P. Holmes, J. L. Lumley, G. Berkooz, and C. W. Rowley, Turbulence, Coherent Structures, Dynamical Systems and Symmetry, 2nd ed. Cambridge University Press, 2012.
 Chatterjee [2000] A. Chatterjee, “An introduction to the proper orthogonal decomposition,” Current Science, vol. 78, no. 7, pp. 808–817, 2000.
 Schmid et al. [2011] P. Schmid, L. Li, M. Juniper, and O. Pust, “Applications of the dynamic mode decomposition,” Theoretical and Computational Fluid Dynamics, vol. 25, no. 14, pp. 249–259, 2011.
 Schmid et al. [2012] P. J. Schmid, D. Violato, and F. Scarano, “Decomposition of timeresolved tomographic PIV,” Experiments in Fluids, vol. 52, no. 6, pp. 1567–1579, 2012.
 Rowley et al. [2009] C. W. Rowley, I. Mezić, S. Bagheri, P. Schlatter, and D. S. Henningson, “Spectral analysis of nonlinear flows,” Journal of Fluid Mechanics, vol. 641, pp. 115–127, 2009.
 Budišić et al. [2012] M. Budišić, R. Mohr, and I. Mezić, “Applied Koopmanism,” Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 22, no. 4, p. 047510, 2012.
 Williams et al. [2014] M. O. Williams, I. G. Kevrekidis, and C. W. Rowley, “A datadriven approximation of the Koopman operator: Extending dynamic mode decomposition,” arXiv:1408.4408, 2014.
 Wynn et al. [2013] A. Wynn, D. Pearson, B. Ganapathisubramani, and P. Goulart, “Optimal mode decomposition for unsteady flows,” Journal of Fluid Mechanics, vol. 733, pp. 473–503, 2013.
 Juang [1994] J.N. Juang, Applied system identification. Prentice Hall, 1994.
 Chen et al. [2012] K. K. Chen, J. H. Tu, and C. W. Rowley, “Variants of dynamic mode decomposition: boundary condition, Koopman, and Fourier analyses,” Journal of Nonlinear Science, vol. 22, no. 6, pp. 887–915, 2012.
 Tu et al. [2013] J. H. Tu, C. W. Rowley, D. M. Luchtenburg, S. L. Brunton, and J. N. Kutz, “On dynamic mode decomposition: theory and applications,” arXiv preprint arXiv:1312.0041, 2013.
 Tissot et al. [2014] G. Tissot, L. Cordier, N. Benard, and B. R. Noack, “Model reduction using dynamic mode decomposition,” Comptes Rendus Mécanique, vol. 342, no. 6, pp. 410–416, 2014.
 Coifman and Lafon [2006] R. R. Coifman and S. Lafon, “Diffusion maps,” Applied and Computational Harmonic Analysis, vol. 21, no. 1, pp. 5–30, 2006.
 Lee and Verleysen [2007] J. A. Lee and M. Verleysen, Nonlinear Dimensionality Reduction. Springer, 2007.
 Bishop et al. [2006] C. M. Bishop et al., Pattern Recognition and Machine Learning. Springer, 2006, vol. 1.
 Koopman [1931] B. O. Koopman, “Hamiltonian systems and transformation in Hilbert space,” Proceedings of the National Academy of Sciences of the United States of America, vol. 17, no. 5, p. 315, 1931.
 Koopman and von Neumann [1932] B. O. Koopman and J. von Neumann, “Dynamical systems of continuous spectra,” Proceedings of the National Academy of Sciences of the United States of America, vol. 18, no. 3, p. 255, 1932.
 Mezic [2013] I. Mezic, “Analysis of fluid flows via spectral properties of the Koopman operator,” Annual Review of Fluid Mechanics, vol. 45, pp. 357–378, 2013.
 Bagheri [2013] S. Bagheri, “Koopmanmode decomposition of the cylinder wake,” Journal of Fluid Mechanics, vol. 726, pp. 596–623, 2013.
 Hemati et al. [2014] M. S. Hemati, M. O. Williams, and C. W. Rowley, “Dynamic mode decomposition for large and streaming datasets,” Physics of Fluids, vol. 26, no. 11, p. 111701, 2014.
 Schmid [2010] P. J. Schmid, “Dynamic mode decomposition of numerical and experimental data,” Journal of Fluid Mechanics, vol. 656, pp. 5–28, 2010.
 Scholkopf [2001] B. Scholkopf, “The kernel trick for distances,” Advances in Neural Information Processing Systems, pp. 301–307, 2001.
 Mezić [2005] I. Mezić, “Spectral properties of dynamical systems, model reduction and decompositions,” Nonlinear Dynamics, vol. 41, no. 13, pp. 309–325, 2005.
 Boyd [2013] J. P. Boyd, Chebyshev and Fourier Spectral Methods. Courier Dover Publications, 2013.
 Trefethen [2000] L. N. Trefethen, Spectral Methods in MATLAB. SIAM, 2000, vol. 10.
 Sirovich [1987] L. Sirovich, “Turbulence and the dynamics of coherent structures. part i: Coherent structures,” Quarterly of applied mathematics, vol. 45, no. 3, pp. 561–571, 1987.
 Burges [1998] C. J. Burges, “A tutorial on support vector machines for pattern recognition,” Data Mining and Knowledge Discovery, vol. 2, no. 2, pp. 121–167, 1998.
 Baudat and Anouar [2001] G. Baudat and F. Anouar, “Kernelbased methods and function approximation,” in Proceedings of the International Joint Conference on Neural Networks, vol. 2. IEEE, 2001, pp. 1244–1249.
 Rasmussen [2006] C. E. Rasmussen, “Gaussian processes for machine learning.” MIT Press, 2006.
 Mauroy and Mezic [2013] A. Mauroy and I. Mezic, “A spectral operatortheoretic framework for global stability,” in 52nd IEEE Conference on Decision and Control, Dec 2013, pp. 5234–5239.
 Cristianini and ShaweTaylor [2000] N. Cristianini and J. ShaweTaylor, An Introduction to Support Vector Machines and Other KernelBased Learning Methods. Cambridge University Press, 2000.
 Ham et al. [2004] J. Ham, D. D. Lee, S. Mika, and B. Schölkopf, “A kernel view of the dimensionality reduction of manifolds,” in Proceedings of the 21st International Conference on Machine Learning. ACM, 2004, p. 47.
 Lehoucq et al. [1998] R. B. Lehoucq, D. C. Sorensen, and C. Yang, ARPACK Users’ Guide: Solution of LargeScale Eigenvalue Problems with Implicitly Restarted Arnoldi Methods. SIAM, 1998, vol. 6.
 Elmer and Van Vleck [2005] C. E. Elmer and E. S. Van Vleck, “Spatially discrete FitzHugh–Nagumo equations,” SIAM Journal on Applied Mathematics, vol. 65, no. 4, pp. 1153–1174, 2005.
 Gaspard et al. [1995] P. Gaspard, G. Nicolis, A. Provata, and S. Tasaki, “Spectral signature of the pitchfork bifurcation: Liouville equation approach,” Physical Review E, vol. 51, no. 1, p. 74, 1995.
 Gaspard and Tasaki [2001] P. Gaspard and S. Tasaki, “Liouvillian dynamics of the hopf bifurcation,” Physical Review E, vol. 64, no. 5, p. 056232, 2001.
 Tu et al. [2014] J. H. Tu, C. W. Rowley, J. N. Kutz, and J. K. Shang, “Toward compressed dmd: spectral analysis of fluid flows using subnyquistrate piv data,” arXiv preprint arXiv:1401.7047, 2014.
 Jovanović et al. [2013] M. R. Jovanović, P. J. Schmid, and J. W. Nichols, “Sparsitypromoting dynamic mode decomposition,” arXiv preprint arXiv:1309.4165, 2013.
 Bagheri [2014] S. Bagheri, “Effects of weak noise on oscillating flows: Linking quality factor, Floquet modes, and Koopman spectrum,” Physics of Fluids, vol. 26, no. 9, p. 094104, 2014.