A linear operator-theoretic approach to nonlinear systems Alexandre - - PowerPoint PPT Presentation

a linear operator theoretic approach to nonlinear systems
SMART_READER_LITE
LIVE PREVIEW

A linear operator-theoretic approach to nonlinear systems Alexandre - - PowerPoint PPT Presentation

A linear operator-theoretic approach to nonlinear systems Alexandre Mauroy University of Namur You have probably already used an operator-theoretic approach to nonlinear systems You have probably already used an operator-theoretic approach to


slide-1
SLIDE 1

A linear operator-theoretic approach to nonlinear systems

Alexandre Mauroy University of Namur

slide-2
SLIDE 2

You have probably already used an operator-theoretic approach to nonlinear systems

slide-3
SLIDE 3

Globally stable equilibrium?

You have probably already used an operator-theoretic approach to nonlinear systems

Positive Lyapunov function:

slide-4
SLIDE 4

Globally stable equilibrium? Koopman operator acting on the « observable »

You have probably already used an operator-theoretic approach to nonlinear systems

Positive Lyapunov function: Operator-theoretic approach:

slide-5
SLIDE 5

Lyapunov function

  • c. 1890

Lyapunov density [Rantzer, 2001]

Stability analysis

>100 years

However, this operator-theoretic approach has been overlooked in nonlinear systems theory

It is surprising to find that Lyapunov's theorem has a close relative (…) that has been neglected until present date.

  • A. Rantzer, A dual to Lyapunov stability theorem, Systems & Control Letters, 42 (2001)
slide-6
SLIDE 6

Koopman operator [Koopman, 1930]

Operator theory

Lyapunov function

  • c. 1890

Lyapunov density [Rantzer, 2001]

Stability analysis

>100 years It is surprising to find that Lyapunov's theorem has a close relative (…) that has been neglected until present date.

  • A. Rantzer, A dual to Lyapunov stability theorem, Systems & Control Letters, 42 (2001)

However, this operator-theoretic approach has been overlooked in nonlinear systems theory

slide-7
SLIDE 7

Koopman operator [Koopman, 1930] Perron-Frobenius operator < 1960 [Ulam]

Operator theory

duality known for decades! Lyapunov function

  • c. 1890

Lyapunov density [Rantzer, 2001]

Stability analysis

>100 years [Vaidya et al., 2008] It is surprising to find that Lyapunov's theorem has a close relative (…) that has been neglected until present date.

  • A. Rantzer, A dual to Lyapunov stability theorem, Systems & Control Letters, 42 (2001)

adjoint operators

However, this operator-theoretic approach has been overlooked in nonlinear systems theory

slide-8
SLIDE 8

Koopman operator- based description Operator acting on a functional space Flow acting on the state space

The operator-theoretic approach provides general and systematic linear methods for nonlinear systems

Trajectory-oriented description

LIFTING

slide-9
SLIDE 9

Systematic, general linear methods Koopman operator- based description Operator acting on a functional space Flow acting on the state space

 finite-dimensional  infinite-dimensional  nonlinear  linear

LIFTING

The operator-theoretic approach provides general and systematic linear methods for nonlinear systems

Trajectory-oriented description

slide-10
SLIDE 10

Outline

Stability analysis: a systematic method

Joint work with I. Mezic, University of California Santa Barbara

Nonlinear identification: a lifting method

Joint work with J. Goncalves, University of Luxembourg

Control: recent works and perspectives

slide-11
SLIDE 11

Koopman eigenvalue Koopman eigenfunction

Global stability is characterized in terms of spectral properties of the Koopman operator

Continuous-time nonlinear system

slide-12
SLIDE 12

Koopman eigenvalue Koopman eigenfunction

Global stability is characterized in terms of spectral properties of the Koopman operator

Continuous-time nonlinear system Theorem: If there exist eigenfunctions with eigenvalues such that , then the set is globally asymptotically stable in .

[AM and Mezic, IEEE Trans. on Aut. Control 2016]

slide-13
SLIDE 13

We obtain a systematic approach to global stability, which mirrors linear stability analysis

[AM and Mezic, IEEE Trans. on Aut. Control 2016]

Assume that is a forward invariant connected set. The equilibrium is globally asymptotically stable in iff (i) the eigenvalues are such that (local stability) (ii) there exist eigenfunctions with

Hyperbolic equilibrium Jacobian matrix has eigenvalues Example: approximation of the basin of attraction

slide-14
SLIDE 14

The spectral approach is related to classic and (new) concepts in control theory

Differential positivity (contracting cone field) [AM, Forni and Sepulchre, CDC 2015] Eventual monotonicity [Sootla and AM, arXiv 1510.01149] Lyapunov function Contracting metric

slide-15
SLIDE 15

Outline

Stability analysis: a systematic method

Joint work with I. Mezic, University of California Santa Barbara

Nonlinear identification: a lifting method

Joint work with J. Goncalves, University of Luxembourg

Control: recent works and perspectives

slide-16
SLIDE 16

We propose to “identify” the Koopman operator

Nonlinear identification /parameter estimation Find such that

[AM and Goncalves, arXiv 1709.02003] [AM and Goncalves, CDC2016]

slide-17
SLIDE 17

We propose to “identify” the Koopman operator

  • 2. Linear

identification

  • 3. “Lifting back”
  • 1. Lifting of

the data Find such that

[AM and Goncalves, arXiv 1709.02003] [AM and Goncalves, CDC2016]

slide-18
SLIDE 18

Step 1: Data are lifted to a higher dimensional space

Choose basis functions

Data Lifted data

slide-19
SLIDE 19

Realization

  • f in the basis

Step 2: The Koopman operator is « identified » in the lifted space

Realization

  • f the infinitesimal generator

linear least squares

Lifted data

Remark: Dual method for high-dimensional systems matrix logarithm

slide-20
SLIDE 20

Step 3: The nonlinear system is finally identified

Realization

  • f the

infinitesimal generator

slide-21
SLIDE 21

Theoretical and numerical results suggest that the method is efficient

Theoretical convergence results The error tends to as (in “optimal” conditions) Van der Pol oscillator Unstable system Chaotic Lorenz system Numerical results [AM and Goncalves, arXiv 1709.02003]

slide-22
SLIDE 22

coefficients

The lifting method is efficient to reconstruct networks with low-sampled data

states (nodes) data points Sampling period: states (nodes) data points coefficients

slide-23
SLIDE 23

Outline

Stability analysis: a systematic method

Joint work with I. Mezic, University of California Santa Barbara

Nonlinear identification: a lifting method

Joint work with J. Goncalves, University of Luxembourg

Control: recent works and perspectives

slide-24
SLIDE 24

The Koopman operator-theoretic framework has been recently applied to control

Observer synthesis [Surana, CDC 2016] Model predictive control [Korda and Mezic 2016, arXiv 1611.03537] Optimal control [Kaiser et al. 2016 , arXiv 1707.01146] Controllability [Goswami and Paley, CDC 2017]

lifting

linear controller/observer design Only numerical results No theoretical framework

slide-25
SLIDE 25

Koopman operator- based description Operator acting on a functional space Flow acting on the state space Trajectory-oriented approach

 infinite-dimensional  linear

LIFTING

  • Global stability
  • Identification
  • Control

The operator-theoretic approach provides general and systematic linear methods for nonlinear systems

Systematic, general linear methods

slide-26
SLIDE 26

What you do with linear systems can (technically) be done with nonlinear systems

analysis identification control…

slide-27
SLIDE 27

A linear operator-theoretic approach to nonlinear systems

Alexandre Mauroy (alexandre.mauroy@unamur.be) University of Namur