The Koopman operator allows for handling nonlinear systems through a globally linear representation. In general, the operator is infinite-dimensional – necessitating finite approximations – for which there is no overarching framework.
Traditionally, systems are represented in the immediate state space, concerned with “dynamics of states”. Although such representations enjoy incredible success, they reach limits when it comes to efficient prediction, analysis and optimization-based control.
Lower/upper case bold symbols x / X denote vectors/matrices. Symbols N / R / ℂ denote sets of natural/real/complex numbers while N 0 denotes all natural numbers with zero and R +, 0 / R + all positive reals with/without zero.
Let us commence with the basic assumptions and definitions required for the introduction of the Koopman operator paradigm.
Consider the Koopman operator K F and two observables g 1, g 2 ∈ F and a scalar α ∈ R. Using (5) it follows that (6) K F ( α g 1 + g 2) = ( α g 1 + g 2) ∘ F = α g 1 ∘ F + g 2 ∘ F = α K F g 1 + K F g 2, showing the linearity of the operator.
As the Koopman operator acts on a function space, it is infinite-dimensional in general. For a finite-dimensional nonlinear system, infinitely many dimensions might be needed to render it linear. Thus, finding a suitable finite-dimensional representation of the operator is required.
Consider the domain X ⊆ M ⊆ R n, x ∈ X and x ̇ = f ( x) such that f: X ↦ M, then the corresponding Koopman operator has eigenfunctions ϕ ( x) that are solutions of a linear partial differential equation (PDE) ∂ ϕ ∂ x f ( x) = s ϕ ( x) if X is compact and ϕ ( x) ∈ C 1 ( X), or alternatively, if ϕ ( x) ∈ C 2 ( X).