dual identification methods in koopman operator theory
TRANSCRIPT
![Page 1: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/1.jpg)
Dual identification methods in Koopmanoperator theory: from ODEs to PDEs
Alexandre Mauroy (University of Namur, Belgium)
![Page 2: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/2.jpg)
We aim at identifying a vector field from (possible low-sampled) data points generated by the dynamics
Find such that
No direct estimation of time derivatives(e.g. low sampling rate, noise)
Constraint
![Page 3: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/3.jpg)
Flowacting on the state space
We can identify the linear Koopman operatorinstead of identifying the nonlinear system
Linear identification(infinite-dimensional)
LIFTING
Nonlinear identification (finite-dimensional)
Nonlinear system
Linear Koopman operator
Operatoracting on a functional space
![Page 4: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/4.jpg)
Outline
The Koopman operator in a nutshell
Two dual lifting methods
(joint work with J. Goncalves, U. Luxembourg)
Extension to PDEs
(preliminary work)
![Page 5: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/5.jpg)
The Koopman operator in a nutshell
is an observable
Semigroup of Koopman operators
Strongly continuous semigroup
Infinitesimal generator of the Koopman operator:
[Koopman 1931, Budisic, Mohr and Mezic 2012]
![Page 6: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/6.jpg)
We estimate the vector field by identifyingthe infinitesimal generator in the lifted space
2. Linearidentification
3. “Lifting back”1. Lifting of the data
![Page 7: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/7.jpg)
Outline
The Koopman operator in a nutshell
Two dual lifting methods
(joint work with J. Goncalves, U. Luxembourg)
Extension to PDEs
(preliminary work)
![Page 8: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/8.jpg)
Our identification method relies on the Trotter-Kato approximation theorem
Trotter-Kato approximation theorem:
strongly continuous semigroups with generators
: orthogonal projection onto
If , then
matrix logarithm
matrix representation
dense in
Nyquist
![Page 9: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/9.jpg)
The method yields the expansion of the vector field in the basis of functions
EDMD, Williams et al., JNLS, 2015]
Step 1
Step 2
Step 3
[AM and Goncalves, CDC 2016]
Assume:
![Page 10: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/10.jpg)
The method is efficient to recover the vector fieldwith low sampling rate
Van der Pol(limit cycle)
Unstableequilibrium
Lorenz (chaotic)
Lifting method vs central differences
(Van der Pol)Extensions
Inputs, stochastic
![Page 11: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/11.jpg)
Theorem
Assumptions:
• polynomial (basis of monomials)• invertible non-singular flow• sample points uniformly randomly distributed in compact invariant set
Then the estimated vector field is given by with
Under some conditions, we have convergence guarantees
Exponential convergence rate
[Kurdila and Bobade, Koopman theoryand nonlinear approximation spaces, 2018]
(with probability one).
![Page 12: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/12.jpg)
Theorem
Assumptions:
• polynomial (basis of monomials)• invertible non-singular flow• sample points uniformly randomly distributed in compact invariant set
Then the estimated vector field is given by with
Under some conditions, we have convergence guarantees
Exponential convergence rate
[Kurdila and Bobade, Koopman theoryand nonlinear approximation spaces, 2018]
(with probability one).
![Page 13: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/13.jpg)
A dual method can also be obtained
: projection onto (piecewise-constant functions)
(Trotter-Kato theorem)
Main method Dual method
Function space
« Sample space »
![Page 14: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/14.jpg)
The dual method yieldsthe value of the vector field at the sample points
(+ regression, e.g. Lasso)
Step 1
Step 2
Construct
Step 3
![Page 15: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/15.jpg)
coefficients
The method is efficient to reconstructlarge networks with nonlinear dynamics
states (nodes)
data points
Sampling period:
states (nodes)
data points
coefficients
![Page 16: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/16.jpg)
coefficients
The method is efficient to reconstructlarge networks with nonlinear dynamics
states (nodes)
data points
Sampling period:
states (nodes)
data points
coefficients
![Page 17: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/17.jpg)
Theorem
• invertible non-singular flow• sample points uniformly randomly distributed over compact invariant set• specific properties for the test functions
(e.g. Gaussian radial basis functions)
Then, the estimated vector field satisfies
with probability one.
Under some conditions, we have convergence guarantees
Convergence rate
![Page 18: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/18.jpg)
Outline
The Koopman operator in a nutshell
Two dual lifting methods
(joint work with J. Goncalves, U. Luxembourg)
Extension to PDEs
(preliminary work)
![Page 19: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/19.jpg)
Can we use a similar method to identify PDEs?
[Rudy, Brunton, Proctor and Kutz, Data-driven discovery of PDEs, Science Advances, 2017]
![Page 20: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/20.jpg)
Can we use a similar method to identify PDEs?
![Page 21: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/21.jpg)
A semigroup of Koopman operatorsis defined in the case of nonlinear PDEs
Nonlinear semigroupNonlinear PDE
(Nonlinear) observable functionals Semigroup of Koopman operators
Infinitesimal generator
(Gâteaux derivative)Strong continuity?
[Mezic, Spectral Koopman operator methods in dynamical systems][Nakao, Proceedings of SICE, 2018] - Hiroya Nakao’s talk
![Page 22: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/22.jpg)
Basis of nonlinear functionals
The infinitesimal generator applied to a specificfunctional provides information on the PDE
for some
Action of the generator
![Page 23: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/23.jpg)
We obtain a lifting methodfor identifying nonlinear PDEs
Step 1
Step 2
Step 3
Assume
![Page 24: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/24.jpg)
Preliminary numerical results suggest thatthe method is efficient to identify nonlinear PDEs
……
![Page 25: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/25.jpg)
A not so smart idea: What about identifyingthe « Koopman operator of the Koopman operator »?
sample functionsbasis functionals
basis/test functionssample points
linear functional
dual method projection of the dual operator (PF)
![Page 26: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/26.jpg)
We have proposed a lifting methodfor system identification / parameter estimation
Two dual methods
Convergence results
Extension to PDEs
Perpsectives
Classic system input-output system identification (unobserved states)
Extension to PDEs (to be continued)
![Page 27: Dual identification methods in Koopman operator theory](https://reader033.vdocuments.site/reader033/viewer/2022051618/6280cf49c9b631583b246023/html5/thumbnails/27.jpg)
Dual identification methods in Koopmanoperator theory: from ODEs to PDEs
Alexandre Mauroy (University of Namur, Belgium)
[AM and Goncalves, Koopman-based lifting techniques for nonlinear systems identification, arxiv]
Codes (Matlab, Julia) available online soon