Abstract
Dynamic mode decomposition (DMD) has become synonymous with the Koopman operator, where continuous time dynamics are discretized and examined using Koopman (i.e. composition) operators. Using the newly introduced “occupation kernels,” the present manuscript develops an approach to DMD that treats continuous time dynamics directly through the Liouville operator. This manuscript outlines the technical and theoretical differences between Koopman-based DMD for discrete time systems and Liouville-based DMD for continuous time systems, which includes an examination of Koopman and Liouville operators over several reproducing kernel Hilbert spaces. While Liouville operators are modally unbounded, this manuscript introduces the concept of a scaled Liouville operator, which, for many dynamical systems, is a compact operator over the native space of the exponential dot product kernel. Compactness of scaled Liouville operators allows for norm convergence of Liouville-based DMD, which is a decided advantage over Koopman-based DMD.
Similar content being viewed by others
Notes
It should be noted that the operator \(P_{\alpha _M} A_{f,a} P_{\alpha _M}\) is simply \(P_{\alpha _M} A_{f,a}\) when restricted to \({{\,\mathrm{span}\,}}(\alpha _M)\) as \(P_{\alpha _M} g = g\) for all \(g \in {{\,\mathrm{span}\,}}(\alpha _M)\).
References
Aronszajn, N.: Theory of reproducing kernels. Trans. Am. Math. Soc. 68(3), 337–404 (1950)
Bakardjian, H., Tanaka, T., Cichocki, A.: Optimization of SSVEP brain responses with application to eight-command brain-computer interface. Neurosci. Lett. 469(1), 34–38 (2010)
Bin, G., Gao, X., Yan, Z., Hong, B., Gao, S.: An online multi-channel SSVEP-based brain-computer interface using a canonical correlation analysis method. J. Neural Eng. 6(4), 046002 (2009)
Bittracher, A., Koltai, P., Junge, O.: Pseudogenerators of spatial transfer operators. SIAM J. Appl. Dyn. Syst. 14(3), 1478–1517 (2015)
Budišić, M., Mohr, R., Mezić, I.: Applied Koopmanism. Chaos Interdiscip. J. Nonlinear Sci. 22(4), 047510 (2012)
Carswell, B., MacCluer, B.D., Schuster, A.: Composition operators on the Fock space. Acta Sci. Math. (Szeged) 69(3–4), 871–887 (2003)
Cichella, V., Kaminer, I., Dobrokhodov, V., Xargay, E., Choe, R., Hovakimyan, N., Aguiar, A.P., Pascoal, A.M.: Cooperative path following of multiple multirotors over time-varying networks. IEEE Trans. Autom. Sci. Eng. 12(3), 945–957 (2015)
Coddington, E.A., Levinson, N.: Theory of Ordinary Differential Equations. Tata McGraw-Hill Education, New York (1955)
Cowen, C.C., Jr., MacCluer, B.I.: Composition Operators on Spaces of Analytic Functions, vol. 20. CRC Press, Boca Raton (1995)
Črnjarić-Žic, N., Maćešić, S., Mezić, I.: Koopman operator spectrum for random dynamical systems. J. Nonlinear Sci. 30, 2007–2056 (2020)
Cvitanovic, P., Artuso, R., Mainieri, R., Tanner, G., Vattay, G., Whelan, N., Wirzba, A.: Chaos: Classical and Quantum. ChaosBook. org. Niels Bohr Institute, Copenhagen (2005)
Das, S., Giannakis, D.: Koopman spectra in reproducing kernel Hilbert spaces. Appl. Comput. Harm. Anal. 49(2), 573–607 (2020)
Folland, G.B.: Real Analysis: Modern Techniques and Their Applications. Wiley, Hoboken (2013)
Froyland, G., González-Tokman, C., Quas, A.: Detecting isolated spectrum of transfer and Koopman operators with Fourier analytic tools. J. Comput. Dyn. 1(2), 249–278 (2014)
Giannakis, D.: Data-driven spectral decomposition and forecasting of ergodic dynamical systems. Appl. Comput. Harm. Anal. 47(2), 338–396 (2019)
Giannakis, D., Das, S.: Extraction and prediction of coherent patterns in incompressible flows through space-time Koopman analysis. Physica D 402, 132211 (2020)
Giannakis, D., Kolchinskaya, A., Krasnov, D., Schumacher, J.: Koopman analysis of the long-term evolution in a turbulent convection cell. arXiv:1804.01944 (2018)
Gonzalez, E., Abudia, M., Jury, M., Kamalapurkar, R., Rosenfeld, J.A.: Anti-koopmanism. arXiv:2106.00106v2 (2021)
Gruss, L.F., Keil, A.: Sympathetic responding to unconditioned stimuli predicts subsequent threat expectancy, orienting, and visuocortical bias in human aversive Pavlovian conditioning. Biol. Psychol. 140, 64–74 (2019)
Haddad, W.: A Dynamical Systems Theory of Thermodynamics. Princeton Series in Applied Mathematics. Princeton University Press, Princeton (2019)
Hallam, T.G., Levin, S.A.: Mathematical Ecology: An Introduction, vol. 17. Springer Science & Business Media, Berlin (2012)
Hastie, T., Tibshirani, R., Friedman, J., Franklin, J.: The elements of statistical learning: data mining, inference and prediction. Math. Intell. 27(2), 83–85 (2005)
Jury, M.T.: C*-algebras generated by groups of composition operators. Indiana Univ. Math. J. 56(6), 3171–3192 (2007)
Khalil, H.K.: Nonlinear Systems, 3rd edn. Prentice Hall, Upper Saddle River (2002)
Klus, S., Nüske, F., Peitz, S., Niemann, J.H., Clementi, C., Schütte, C.: Data-driven approximation of the Koopman generator: model reduction, system identification, and control. Physica D 406, 132416 (2020)
Korda, M., Mezić, I.: On convergence of extended dynamic mode decomposition to the Koopman operator. J. Nonlinear Sci. 28(2), 687–710 (2018)
Kutz, J.N., Brunton, S.L., Brunton, B.W., Proctor, J.L.: Dynamic Mode Decomposition: Data-Driven Modeling of Complex Systems. SIAM, Philadelphia (2016)
Luery, K.E.: Composition Operators on Hardy Spaces of the Disk and Half-Plane. University of Florida, Gainesville (2013)
Mezić, I.: Spectral properties of dynamical systems, model reduction and decompositions. Nonlinear Dyn. 41(1–3), 309–325 (2005)
Mezić, I.: Analysis of fluid flows via spectral properties of the Koopman operator. Annu. Rev. Fluid Mech. 45, 357–378 (2013)
Middendorf, M., McMillan, G., Calhoun, G., Jones, K.S.: Brain-computer interfaces based on the steady-state visual-evoked response. IEEE Trans. Rehabil. Eng. 8(2), 211–214 (2000)
Pedersen, G.K.: Analysis Now, Graduate Texts in Mathematics, vol. 118. Springer Science & Business Media, Berlin (2012)
Petro, N.M., Gruss, L.F., Yin, S., Huang, H., Miskovic, V., Ding, M., Keil, A.: Multimodal imaging evidence for a frontoparietal modulation of visual cortex during the selective processing of conditioned threat. J. Cogn. Neurosci. 29(6), 953–967 (2017)
Regan, D.: Human brain Electrophysiology: Evoked Potentials and Evoked Magnetic Fields in Science and Medicine. Elsevier, Amsterdam (1989)
Rosenfeld, J.A.: Densely defined multiplication on several Sobolev spaces of a single variable. Complex Anal. Oper. Theory 9(6), 1303–1309 (2015a)
Rosenfeld, J.A.: Introducing the polylogarithmic hardy space. Integral Equ. Oper. Theory 83(4), 589–600 (2015b)
Rosenfeld, J.A.: The Sarason sub-symbol and the recovery of the symbol of densely defined Toeplitz operators over the Hardy space. J. Math. Anal. Appl. 440(2), 911–921 (2016)
Rosenfeld, J.A., Kamalapurkar, R.: Dynamic mode decomposition with control Liouville operators. In: IFAC-PapersOnLine, vol. 54, pp. 707–712 (2021)
Rosenfeld, J.A., Kamalapurkar, R., Gruss, L.F., Johnson, T.T.: On occupation kernels, Liouville operators, and dynamic mode decomposition. In: Proceedings of the American Control Conference, pp. 3957–3962. New Orleans, LA, USA (2021)
Rosenfeld, J.A., Kamalapurkar, R., Russo, B., Johnson, T.T.: Occupation kernels and densely defined Liouville operators for system identification. In: Szafraniec are Ramirez de Arellano, E. and Shapiro, M. V. and Tovar, L. M. and Vasilevski N. L. Proceedings of the IEEE Conference on Decision and Control, pp. 6455–6460 (2019a)
Rosenfeld, J.A., Russo, B., Kamalapurkar, R., Johnson, T.T.: The occupation kernel method for nonlinear system identification. arXiv:1909.11792 (2019b)
Steinwart, I., Christmann, A.: Support Vector Machines. Springer Science & Business Media, Berlin (2008)
Szafraniec, F.H.: The reproducing kernel Hilbert space and its multiplication operators. In: Ramirez de Arellano, E., Shapiro, M. V., Tovar, L. M., Vasilevski N. L. (eds.) Complex Analysis and Related Topics, pp. 253–263. Springer (2000)
Tóth, J., Nagy, A.L., Papp, D.: Reaction Kinetics: Exercises, Programs and Theorems. Springer, Berlin (2018)
Walters, P., Kamalapurkar, R., Voight, F., Schwartz, E.M., Dixon, W.E.: Online approximate optimal station keeping of a marine craft in the presence of an irrotational current. IEEE Trans. Robot. 34(2), 486–496 (2018)
Wendland, H.: Scattered Data Approximation, Cambridge Monographs on Applied and Computational Mathematics, vol. 17. Cambridge University Press, Cambridge (2004)
Williams, M.O., Kevrekidis, I.G., Rowley, C.W.: A data-driven approximation of the Koopman operator: extending dynamic mode decomposition. J. Nonlinear Sci. 25(6), 1307–1346 (2015a)
Williams, M.O., Rowley, C.W., Kevrekidis, I.G.: A kernel-based method for data-driven Koopman spectral analysis. J. Comput. Dyn. 2(2), 247–265 (2015b)
Acknowledgements
This research was supported by the Air Force Office of Scientific Research under Contract Numbers FA9550-20-1-0127, FA9550-18-1-0122 and FA9550-21-1-0134, the Air Force Research Laboratory under contract number FA8651-19-2-0009, and the National Science Foundation under grant numbers 2027976, 2027999, and 2028001. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the sponsoring agencies.
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Oliver Junge.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
A subset of the results in this manuscript was presented at the 2021 American Control Conference and is published in the proceedings (Rosenfeld et al. 2021). A YouTube Playlist supporting the content of this manuscript (including MATLAB code) may be found here: https://youtube.com/playlist?list=PLldiDnQu2phuIdps0DcIQJ_gF0YIb-g6y.
Proofs of Theorem 1 and Proposition 2
Proofs of Theorem 1 and Proposition 2
Theorem 1 restated: Let \(F^2({\mathbb {R}}^n)\) be the Bargmann-Fock space of real valued functions, which is the native space for the exponential dot product kernel, \(K(x,y) = \exp (x^Ty)\), \(a \in {\mathbb {R}}\) with \(|a| < 1\), and let \(A_{f,a}\) be the scaled Liouville operator with symbol \(f:{\mathbb {R}}^n \rightarrow {\mathbb {R}}^n\). There exists a collection of coefficients, \(\{ C_\alpha \}_{\alpha }\), indexed by the multi-index \(\alpha \), such that if f is representable by a multi-variate power series, \(f(x) = \sum _{\alpha } f_\alpha x^\alpha \), satisfying
then \(A_{f,a}\) is bounded and compact over \(F^2({\mathbb {R}}^n)\).
Proof
The proof for the case \(n = 1\) is presented to simplify the exposition. The case for \(n > 1\) follows with some additional bookkeeping of the multi-index.
If \(A_{x,a}\) is compact for all \(|a| < 1\), then \(A_{x^m,a} = A^{m}_{x,\root m \of {a}}\) is compact since products of compact operators are compact. If \(f(x) = \sum _{m=0}^\infty f_m x^m\) is such that \(\sum _{m=0}^\infty |f_m| \Vert A_{x^m,a}\Vert < \infty \), then \(A_{f,a} = \lim _{m\rightarrow \infty } \sum _{m=0}^M f_m A_{x^m,a},\) with respect the operator norm via the triangle inequality, and \(A_{f,a}\) is compact since it is the limit of compact operators. Thus, it is sufficient to demonstrate that \(A_{x,a}\) is compact to prove the theorem.
Let \(g \in F^{2}({\mathbb {R}})\), then \(g(x) = \sum _{m=0}^\infty g_m \frac{x^m}{\sqrt{m!}}\) with norm \(\Vert g \Vert _{F^{2}({\mathbb {R}})}^2 = \sum _{m=0}^\infty |g_m|^2 < \infty .\) Applying the scaled Liouville operator, \(A_{x,a}\), yields
Hence, \(\Vert A_{x,a} g \Vert _{F^({\mathbb {R}})}^2 = |a|^{2m} m^2 |g_{m}|^2 < \infty \) as for large enough m, \(|a|^{2m} m^2 < 1\). Hence, \(A_{x,a}\) is everywhere defined and by the closed graph theorem \(A_{x,a}\) is bounded.
As \(|a|^{m} m^2 \rightarrow 0\), there is an M such that for all \(m > M\), \(|a|^{m} m^2 < 1\). Let \(P_M\) be the projection onto \({{\,\mathrm{span}\,}}\{ 1, x, x^2, \ldots , x^M\}\). Now consider
Hence, the operator norm of \((A_{x,a} - A_{x,a} P_M)\) is bounded by \(|a|^{M/2}\), and as \(|a| < 1\), \(A_{x,a} P_m \rightarrow A_{x,a}\) in the operator norm. \(P_m\) is finite rank and therefore compact. It follows that \(A_{x,a} P_m\) is compact, since compact operators form an ideal in the ring of bounded operators. Thus, \(A_{x,a}\) is compact as it is the limit of compact operators. \(\square \)
Proposition 2restated: Let H be a RKHS of twice continuously differentiable functions over \({\mathbb {R}}^n\), f be Lipschitz continuous, and suppose that \(\varphi _{i,a}\) is an eigenfunction of \(A_{f,a}\) with eigenvalue \(\lambda _{i,a}\). Let D be a compact subset of \({\mathbb {R}}^n\) that contains x(t) for all \(0< t < T\). In this setting, if \(\lambda _{i,a} \rightarrow \lambda _{i,1}\) and \(\varphi _{i,a}(x(0)) \rightarrow \varphi _{i,1}(x(0))\) as \(a \rightarrow 1^-\), then
Proof
Suppose that x(t) remains in a compact set \(D \subset {\mathbb {R}}^n\). Since \(\phi _{m,a} \in H\) and H consists of twice continuously differentiable functions, there exists \(M_1,M_2,F > 0\) such that
First, it is necessary to demonstrate that \(M_{1,a}\) and \(M_{2,a}\) may be bounded independent of a. For each \(i,j=1,\ldots ,n\) and \(y \in {\mathbb {R}}^n\), the functionals \(g \mapsto \frac{\partial }{\partial x_i} g(y)\) and \(g \mapsto \frac{\partial ^2}{\partial x_i \partial x_j} g(y)\) are bounded (cf. Steinwart and Christmann 2008). Setting, \(k_{y} = K(\cdot ,y)\), it can be seen that the functions \(\frac{\partial }{\partial x_i} k_y\) and \(\frac{\partial ^2}{\partial x_i \partial x_j} k_y\) are the unique functions that represent these functionals through the inner product of the RKHS (cf. Steinwart and Christmann 2008). As \(\phi _{m,a}\) is a normal vector, \(\Vert \phi _{m,a} \Vert _H = 1\), and by Cauchy-Schwarz
(13) is bounded over D as \(x \mapsto \frac{\partial }{\partial x_i} k_y(x)\) is continuous. Thus, \(M_{1,a}\) is bounded independent of a. A similar argument may be carried out for \(M_{2,a}\). Let \(M_1\) and \(M_2\) be the respective bounding constants.
Note that
Then by the mean value inequality, Cauchy-Schwarz, and the bounds given above,
Setting \(\epsilon _a(t) := \frac{\partial }{\partial t} \phi _{m,a}(ax(t)) - \frac{\partial }{\partial t} \phi _{m,a}( x(t))\), it follows that \(\sup _{0 \le t \le T} \Vert \epsilon _a(t) \Vert _2 = O(|a-1|)\). Hence,
and
As the time interval is fixed to [0, T], \(e^{\mu _{m,a}t} \int _0^t e^{-\mu _{m,a} \tau } \epsilon (\tau ) d\tau = O(|a-1|),\) since \(\mu _{m,a}\) is bounded with respect to a. \(\square \)
Theorem 2restated: Let \(|a| < 1\). Suppose that \(\{ \gamma _{i}:[0,T_i] \rightarrow {\mathbb {R}}^n \}_{i=1}^\infty \) is a sequence of trajectories satisfying \({\dot{\gamma }} = f(\gamma )\) for a dynamical system f corresponding to a compact scaled Liouville operator, \(A_{f,a}\). If the collection of functions, \(\{ \varGamma _{\gamma _i} \}_{i=1}^\infty \) is dense in the Bargmann-Fock space, then the sequence of operators \(\{ P_{\alpha _M} A_{f,a} P_{\alpha _M} \}_{M=1}^\infty \) converges to \(A_{f,a}\) in the norm topology, where \(\alpha _M = \{ \varGamma _{\gamma _1}, \ldots , \varGamma _{\gamma _M} \}\).
Proof
The following proof is more general than what is indicated in the theorem statement of Theorem 2. In fact, for any compact operator, T, and any set \(\{ g_i \}_{i=1}^\infty \) such that \(\overline{{{\,\mathrm{span}\,}}(\{g_i\}_{i=1}^\infty )} = H\), the sequence of operators \(P_{\alpha _M} T P_{\alpha _M} \rightarrow T\) in norm, where \(P_{\alpha _M}\) is the projection onto \({{\,\mathrm{span}\,}}(\{g_i\}_{i=1}^M)\). Henceforth, it will be assumed that \(\{ g_i \}_{i=1}^\infty \) is an orthonormal basis for H, since given any complete basis in H, an orthonormal basis may be obtained via the Gram-Schmidt process.
First note that every compact operator has a representation as \(T = \sum _{i=1}^\infty \lambda _i \langle \cdot , v_i \rangle _H u_i\), where \(\{ v_i \}\) and \(\{ u_i \}\) are orthonormal collections of vectors (functions) in H, and \(\{ \lambda _i \}_{i=1}^\infty \subset {\mathbb {C}}\) are the singular values of T. If \(T_M := \sum _{i=1}^M \lambda _i \langle \cdot , v_i \rangle _H u_i\) then \(T_M \rightarrow T\) as \(M \rightarrow \infty \) in the operator norm.
Suppose that \(\epsilon > 0\). Select M such that \(\Vert T - T_M\Vert < \epsilon \), and select N such that for all \(n > N\),
for all \(i=1,\ldots ,M\). Let \(g \in H\) be arbitrary.
Now consider,
The second term after the inequality may be expanded as
Now the objective is to demonstrate that both \(\Vert T_M g - P_n T_M g\Vert _H\) and \(\Vert T_M g - T_M P_n g\Vert _H\) are proportional to \(\epsilon \Vert g\Vert _H\). Note that
and
Thus, for every \(\epsilon > 0\), there is an N such that for all \(n > N\), \(\Vert T g - P_n T P_n g\Vert _H \le 4\epsilon \Vert g\Vert _H.\) Hence, it follows that \(\Vert T - P_n T P_n \Vert \le 4\epsilon \). Thus, as \(n \rightarrow \infty \), \(P_n T P_n \rightarrow T\) in the operator norm. \(\square \)
Rights and permissions
About this article
Cite this article
Rosenfeld, J.A., Kamalapurkar, R., Gruss, L.F. et al. Dynamic Mode Decomposition for Continuous Time Systems with the Liouville Operator. J Nonlinear Sci 32, 5 (2022). https://doi.org/10.1007/s00332-021-09746-w
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s00332-021-09746-w
Keywords
- Dynamic mode decomposition
- Densely defined operators
- Liouville operator
- Reproducing kernel Hilbert spaces