betaFeedback

WorksEngineering

On Lie-Algebraic Conditions for Finite-Dimensional Koopman Bilinear Embeddings of Control-Affine Systems

Abstract

Koopman bilinear representations for control-affine systems have attracted considerable attention. In this paper, we investigate, from a Lie-algebraic perspective, conditions under which a control-affine system admits a finite-dimensional Koopman bilinear embedding. Specifically, we focus on the Lie algebra generated by the drift and control vector fields. First, we show that the finite-dimensionality of the generated Lie algebra is a necessary condition for the existence of a Koopman bilinear embedding. Next, when the generated Lie algebra contains a nonzero nilpotent ideal, we derive a sufficient condition for the existence of such an embedding. Finally, through an example of a quadratic polynomial system, we demonstrate that the finite-dimensionality of the generated Lie algebra imposes strong restrictions on the vector fields, suggesting that exact finite-dimensional Koopman bilinear embeddings may be nongeneric.

Introduction

The Koopman operator has recently attracted considerable attention in the analysis and control of nonlinear systems: see, e.g., [1], [2]. Originally introduced for nonlinear autonomous systems, it is a linear operator describing the temporal evolution of functions of the state, called observables. Because the Koopman operator can be approximated from observed data, its data-driven approximation methods, most notably extended dynamic mode decomposition (EDMD), have been widely studied [3], [4].

Koopman-based research on control systems often considers control-affine systems of the form

x˙=X0(x)+∑i=1muiXi(x),x∈M.(1)\dot{x} = X_0(x) + \sum_{i=1}^{m} u_i X_i(x), \qquad x \in M. \tag*{(1)}

Early studies investigated Koopman linear representations z˙=Az+Bu\dot{z} = Az + Bu obtained through a lifting z=Ψ(x)z = \Psi(x) [5], [6]. It was subsequently pointed out by [7] that such a representation need not exist for general control-affine systems. Motivated by this, Koopman bilinear representations of the form

z˙=A0z+∑i=1muiAiz.(2)\dot{z} = A_0 z + \sum_{i=1}^{m} u_i A_i z. \tag*{(2)}

have been studied [7], [8]. Beyond control-affine systems, more general representations have also been investigated: see [9], [10], the recent review [11], and references therein.

The objective of this paper is to establish conditions for the existence of a finite-dimensional embedding z=Ψ(x)z = \Psi(x) that transforms (1) into the bilinear form (2). We call such an embedding a Koopman bilinear embedding. Equivalently, we seek conditions under which the vector fields X0,X1,…,XmX_0, X_1, \ldots, X_m admit a common finite-dimensional Koopman invariant subspace (KIS), whose term is coined by [12]. Establishing these existence conditions has clear benefits. For instance, it provides a theoretical basis for assessing whether a finite-dimensional model approximated by EDMD is relevant to an exact finite-dimensional Koopman bilinear embedding. More fundamentally, it can clarify the nature of Koopman representations for control systems, including not only bilinear representations but also infinite-dimensional representations [13], [14], or finite-dimensional models that explicitly account for approximation errors [15], [16].

The existence of Koopman embeddings has been investigated in several previous studies. For nonlinear continuous-time autonomous systems, [17], [18] studied the existence of embeddings based on topological properties of the underlying dynamical systems. For continuous-time systems with a control input uu, [19] characterized a condition under which a Koopman linearization of the form z˙=Az+Bu\dot{z} = Az + Bu exists. Koopman linearization for discrete-time systems with a control input has also been studied in [20]. Koopman bilinear embeddings for continuous-time systems have been investigated in [7], [8]. Their existing conditions for bilinear embeddings require the existence of basis for KIS. While finding such a basis is generally challenging [2], applying these conditions in practice still requires heuristically selecting candidate observables and numerically assessing whether their span is invariant. Another condition that does not require finding KIS has been expected.

In this paper, we derive a collection of such conditions for the existence of Koopman bilinear embeddings without finding a basis for KIS. Rather than focusing on local differential-geometric properties of the state equations such as [21], we exploit the algebraic structure of vector fields defined globally on MM. Specifically, we use the Lie algebra formed by the vector fields and their Lie brackets. The main contributions are the following:

  1. A Lie-algebraic interpretation for the existence problem of Koopman bilinear embeddings (Proposition 2).

  1. A necessary condition for the existence of Koopman bilinear embeddings: finite-dimensionality of the Lie algebra generated by X0,…,XmX_0, \ldots, X_m (Theorem 1).

  1. A sufficient condition under which the generated Lie algebra contains a nilpotent ideal (Theorem 2), together with a comparison with the differential-geometric conditions for Koopman linearizations in [19].

  1. A quadratic polynomial example demonstrating that the finite-dimensionality of the generated Lie algebra

strongly restricts a class of vector fields with Koopman bilinear embedding.

The fourth contribution suggests that the existence of finite-dimensional Koopman bilinear embeddings may be nongeneric, highlighting a strong need to quantify errors in finite-dimensional approximation models.

The remainder of this paper is organized as follows. Section II reviews manifolds, Koopman embeddings, Lie algebras, and their representations. Section III derives Lie-algebraic conditions for the existence of Koopman bilinear embeddings. Section IV examines the resulting condition through an example of a quadratic polynomial system. Section V concludes this paper with a brief summary and future directions. The proofs of Proposition 2 and Theorem 2 are presented in the appendices.

Preliminaries

Manifold theory

This subsection introduces several basic concepts from manifold theory. Let MM be an nn-dimensional C∞C^\infty manifold. For simplicity of coordinate representation, we first consider the case where MM is an open subset of Rn\mathbb{R}^n. Let C∞(M;C)C^\infty(M;\mathbb{C}) denote the space of C∞C^\infty complex-valued functions on MM, and X(M)\mathcal{X}(M) be the space of C∞C^\infty real vector fields on MM. For X∈X(M)X \in\mathcal{X}(M), the Lie derivative along XX is the linear operator LX:C∞(M;C)→C∞(M;C)L_X:C^\infty(M;\mathbb{C}) \to C^\infty(M;\mathbb{C}) defined by

LXf(x):=df(x)X(x),f∈C∞(M;C), x∈M,L_X f(x) := df(x)X(x), \qquad f \in C^\infty(M;\mathbb{C}),\ x \in M,

where df(x)df(x) represents the differential of ff at x∈Mx \in M. In addition, for X,Y∈X(M)X,Y \in\mathcal{X}(M), their Lie bracket is the vector field [X,Y]∈X(M)[X,Y] \in\mathcal{X}(M) defined by

[X,Y](x):=DY(x)X(x)−DX(x)Y(x),x∈M(3)[X,Y](x) := DY(x)X(x) - DX(x)Y(x), \qquad x \in M \tag*{(3)}

where DXDX and DYDY represent the Jacobian of XX and YY, respectively. It then follows that L[X,Y]=LXLY−LYLXL_{[X,Y]} = L_XL_Y - L_YL_X. For a general smooth manifold MM, the Lie derivative and the Lie bracket can be defined by using local coordinates, and the resulting definitions are independent of the choice of local coordinates [22].

Next, let Ψ:M→CN\Psi:M \to\mathbb{C}^N be a C∞C^\infty map and dΨ(x)d\Psi(x) denote its differential at x∈Mx \in M, where CN\mathbb{C}^N is regarded as a 2N2N-dimensional real manifold. The C∞C^\infty map Ψ\Psi is called a C∞C^\infty embedding if dΨ(x):TxM→TΨ(x)CNd\Psi(x):T_xM \to T_{\Psi(x)}\mathbb{C}^N is injective for every x∈Mx \in M and Ψ:M→Ψ(M)⊂CN\Psi:M \to\Psi(M) \subset\mathbb{C}^N is a homeomorphism.

Koopman embeddings for autonomous systems

This subsection introduces finite-dimensional Koopman linear embeddings for autonomous systems. Consider the following autonomous system:

x˙=X(x),x∈M,(4)\dot{x} = X(x), \qquad x \in M, \tag*{(4)}

where MM is the state space and X∈X(M)X \in\mathcal{X}(M) is a vector field. We simply represent the system (4) by (M,X)(M,X). Throughout this paper, we call a complex-valued function f∈C∞(M;C)f \in C^\infty(M;\mathbb{C}) an observable. The Koopman linear embedding problem asks whether a finite collection of observables can be used to embed nonlinear system (4) into a finite-dimensional linear system.

Definition 1. We say that the autonomous system (M,X)(M,X) admits a finite-dimensional Koopman linear embedding if there exist a positive integer NN, linearly independent functions ψ1,…,ψN∈C∞(M;C)\psi_1,\ldots,\psi_N \in C^\infty(M;\mathbb{C}), and a matrix A∈CN×NA \in\mathbb{C}^{N \times N} such that Ψ:=(ψ1,…,ψN)T:M→CN\Psi:= (\psi_1,\ldots,\psi_N)^\mathsf{T}:M \to\mathbb{C}^N is a C∞C^\infty embedding and satisfies

LXΨ:=(LXψ1⋮LXψN)=A(ψ1⋮ψN)=AΨ.L_X\Psi:= \begin{pmatrix} L_X\psi_1\\ \vdots\\ L_X\psi_N \end{pmatrix} = A \begin{pmatrix} \psi_1\\ \vdots\\ \psi_N \end{pmatrix} = A\Psi.

In this case, the pair (Ψ,A)(\Psi,A) is called a finite-dimensional Koopman linear embedding of (M,X)(M,X), or simply a Koopman linear embedding.

Indeed, if the autonomous system (M,X)(M,X) admits a Koopman linear embedding (Ψ,A)(\Psi,A) and x(t)x(t) is a trajectory of (4), differentiating z(t)=Ψ(x(t))z(t)=\Psi(x(t)) in tt yields z˙(t)=Az(t)\dot{z}(t)=A z(t). Moreover, because Ψ\Psi is an embedding, the state xx can be uniquely reconstructed from z∈Ψ(M)z \in\Psi(M). This formulation thus recasts the dynamics from a state-based description, in which one tracks trajectories on MM, into an observable-based description formulated entirely in terms of observables on MM and their Lie derivatives along XX.

Remark 1. Definition 1 does not require the vector field XX to generate a flow globally, although a local flow (trajectory) is always defined. If the autonomous system (M,X)(M,X) generates a semi-flow Φt:M→M\Phi^t:M \to M (t≥0t \ge0), then the Koopman semi-group can be defined by Utf:=f∘ΦtU^t f := f \circ\Phi^t.

Remark 2. The existence of the Koopman linear embedding for autonomous systems has been studied in the literature [17, 18]. These studies exploit topological properties of ω\omega-limit sets and attractors to establish the existence of Koopman linear embeddings without explicitly constructing the linear embedding Ψ\Psi.

In Koopman operator theory, the linear operator LX:C∞(M;C)→C∞(M;C)L_X:C^\infty(M;\mathbb{C}) \to C^\infty(M;\mathbb{C}) is called the Koopman generator. Note that if the Koopman semi-group exists for t≥0t \ge0, then LXL_X is identical to its infinitesimal generator. By Definition 1, a finite-dimensional subspace F:=span⁡C{ψ1,…,ψN}⊂C∞(M;C)\mathcal{F}:=\operatorname{span}_{\mathbb{C}}\{\psi_1,\ldots,\psi_N\} \subset C^\infty(M;\mathbb{C}) satisfies LXF⊆FL_X\mathcal{F}\subseteq\mathcal{F}. Such a subspace F\mathcal{F} is called a Koopman invariant subspace (KIS). In particular, we have a nice spectral property as σ(LX∣F)=σ(A)\sigma(L_X|_{\mathcal{F}})=\sigma(A), where σ(LX∣F)\sigma(L_X|_{\mathcal{F}}) and σ(A)\sigma(A) represent the set of eigenvalues of LX∣FL_X|_{\mathcal{F}} and AA, respectively.

Lie algebras and representations

This subsection introduces the basic concepts of Lie algebras and their representations [23, 24]. In what follows, Lie algebras are not assumed to be finite-dimensional.

Definition 2. A real vector space g\mathfrak{g}, equipped with a bilinear map [⋅,⋅]:g×g→g[\cdot,\cdot]:\mathfrak{g}\times\mathfrak{g}\to\mathfrak{g}, is called a real Lie algebra (or simply a Lie algebra) if the following conditions hold:

  1. [X,X]=0[X,X]=0 for every X∈gX \in\mathfrak{g};

  2. [X,[Y,Z]]+[Y,[Z,X]]+[Z,[X,Y]]=0[X,[Y,Z]]+[Y,[Z,X]]+[Z,[X,Y]]=0 for every X,Y,Z∈gX,Y,Z \in\mathfrak{g} (Jacobi identity).

The bilinear map [⋅,⋅][\cdot,\cdot] is called the Lie bracket on g\mathfrak{g}.

Example 1. The space X(M)\mathcal{X}(M) of smooth vector fields on MM, equipped with the Lie bracket (3), is an infinite-dimensional real Lie algebra.

Example 2. Let VV be a complex vector space, and let End⁡(V)\operatorname{End}(V) be the space of all complex-linear endomorphisms of VV. Regarding End⁡(V)\operatorname{End}(V) as a real vector space, we define

[A,B]End⁡V:=A(Bv)−B(Av),A,B∈End⁡(V), v∈V.[A,B]_{\operatorname{End}V}:=A(Bv)-B(Av),\qquad A,B\in\operatorname{End}(V),\ v\in V.

Then, End⁡(V)\operatorname{End}(V), equipped with this bracket, is a real Lie algebra.

To develop the main results of this paper, we introduce the following notions:

Definition 3. A vector subspace h⊆g\mathfrak{h}\subseteq\mathfrak{g} is called a Lie subalgebra if h1,h2∈h⇒[h1,h2]∈hh_1,h_2\in\mathfrak{h}\Rightarrow[h_1,h_2]\in\mathfrak{h}.

Definition 4. Let h1,…,hm∈gh_1,\ldots,h_m\in\mathfrak{g}. The smallest Lie subalgebra of g\mathfrak{g} containing h1,…,hmh_1,\ldots,h_m is called Lie algebra generated by h1,…,hmh_1,\ldots,h_m and denoted by Lie⁡(h1,…,hm)\operatorname{Lie}(h_1,\ldots,h_m). Equivalently, we have

Lie⁡(h1,…,hm)=span⁡R{hi,[hi,hj],[hi,[hj,hk]],[[hi,hj],hk],[[hi,hj],[hk,hℓ]],…},\operatorname{Lie}(h_1,\ldots,h_m)=\operatorname{span}_{\mathbb{R}}\{h_i,[h_i,h_j],[h_i,[h_j,h_k]],[[h_i,h_j],h_k],[[h_i,h_j],[h_k,h_\ell]],\ldots\},

where the indices range over {1,…,m}\{1,\ldots,m\}, and all possible finite iterated Lie brackets are included. In particular, if g\mathfrak{g} is infinite-dimensional, Lie⁡(h1,…,hm)\operatorname{Lie}(h_1,\ldots,h_m) may be infinite-dimensional even when mm is finite.

Next, we introduce a representation of a Lie algebra.

Definition 5. Let g\mathfrak{g} be a real Lie algebra and VV be a complex vector space. A representation of g\mathfrak{g} on VV is a real linear map π:g→End⁡(V)\pi:\mathfrak{g}\to\operatorname{End}(V) such that

π([X,Y])=[π(X),π(Y)]End⁡,∀X,Y∈g,(5)\pi([X,Y])=[\pi(X),\pi(Y)]_{\operatorname{End}},\qquad\forall X,Y\in\mathfrak{g}, \tag*{(5)}

where End⁡(V)\operatorname{End}(V) is regarded as a real Lie algebra.

The Lie derivative provides a representation of the Lie algebra X(M)\mathcal{X}(M):

Proposition 1. For Definition 5, set g=X(M)\mathfrak{g}=\mathcal{X}(M) and V=C∞(M;C)V=C^\infty(M;\mathbb{C}). Then,

π:X(M)→End⁡(C∞(M;C)),π(X):=LX\pi:\mathcal{X}(M)\to\operatorname{End}(C^\infty(M;\mathbb{C})),\qquad\pi(X):=L_X

is a representation of X(M)\mathcal{X}(M).

Proof. The map π\pi is real linear because π(aX+bY)=LaX+bY=aLX+bLY\pi(aX+bY)=L_{aX+bY}=aL_X+bL_Y for every a,b∈Ra,b\in\mathbb{R} and every X,Y∈X(M)X,Y\in\mathcal{X}(M). Equation (5) follows from π([X,Y])=L[X,Y]=LXLY−LYLX=[π(X),π(Y)]End⁡\pi([X,Y])=L_{[X,Y]}=L_XL_Y-L_YL_X=[\pi(X),\pi(Y)]_{\operatorname{End}}.

Main results: Lie-algebraic conditions for Koopman bilinear embeddings

This section presents our main results on Lie-algebraic conditions for Koopman bilinear embeddings. Section III-A formulates the problem; Section III-B gives its Lie-algebraic interpretation; Section III-C derives a necessary condition; and Section III-D establishes a sufficient condition involving a nilpotent ideal and compares it with the existing conditions of Koopman linearization in [19].

Problem statement

This subsection formulates the existence problem for finite-dimensional Koopman bilinear embeddings of control-affine systems. Consider the following control-affine system:

x˙=X0(x)+∑j=1muj(t)Xj(x),x∈M,(6)\dot{x}=X_0(x)+\sum_{j=1}^{m}u_j(t)X_j(x),\qquad x\in M, \tag*{(6)}

where MM is an nn-dimensional C∞C^\infty manifold, X0,…,Xm∈X(M)X_0,\ldots,X_m\in\mathcal{X}(M) are C∞C^\infty vector fields, and u=(u1,…,um)T∈Rmu=(u_1,\ldots,u_m)^\mathsf{T}\in\mathbb{R}^m is the control input. We simply represent the system (6) by (M,Xj)0≤j≤m(M,X_j)_{0\leq j\leq m}. The existence problem for a Koopman bilinear embedding is to pose the question of whether a finite collection of observables can be used to embed the control-affine system (6) into a finite-dimensional bilinear system of the form

z˙=A0z+∑j=1muj(t)Ajz,z∈CN.(7)\dot{z}=A_0z+\sum_{j=1}^{m}u_j(t)A_jz,\qquad z\in\mathbb{C}^N. \tag*{(7)}

Definition 6. We say that the control-affine system (M,Xj)0≤j≤m(M,X_j)_{0\leq j\leq m} admits a finite-dimensional Koopman bilinear embedding if there exist a positive integer NN, linearly independent functions Ψ1,…,ΨN∈C∞(M;C)\Psi_1,\ldots,\Psi_N\in C^\infty(M;\mathbb{C}), and matrices A0,…,Am∈CN×NA_0,\ldots,A_m\in\mathbb{C}^{N\times N} such that Ψ=(Ψ1,…,ΨN)T:M↪CN\Psi=(\Psi_1,\ldots,\Psi_N)^\mathsf{T}:M\hookrightarrow\mathbb{C}^N is a C∞C^\infty embedding and satisfies

LXjΨ=AjΨ,j=0,…,m.(8)L_{X_j}\Psi=A_j\Psi,\qquad j=0,\ldots,m. \tag*{(8)}

In this case, the pair (Ψ,Aj)0≤j≤m(\Psi,A_j)_{0\leq j\leq m} is called a finite-dimensional Koopman bilinear embedding of (M,Xj)0≤j≤m(M,X_j)_{0\leq j\leq m}, or simply a Koopman bilinear embedding.

Indeed, if there exists a Koopman bilinear embedding (Ψ,Aj)0≤j≤m(\Psi,A_j)_{0\leq j\leq m}, then differentiating z(t)=Ψ(x(t))z(t)=\Psi(x(t)) in tt yields (7). The existence of matrices A0,…,AmA_0,\ldots,A_m satisfying (8) is equivalent to the invariance of the finite-dimensional subspace F:=span⁡{Ψ1,…,ΨN}\mathcal{F}:=\operatorname{span}\{\Psi_1,\ldots,\Psi_N\} under all the Lie derivatives LXjL_{X_j}. Namely, F\mathcal{F} is a common KIS:

LXjF⊆F,j=0,…,m.(9)L_{X_j}\mathcal{F}\subseteq\mathcal{F},\qquad j=0,\ldots,m. \tag*{(9)}

Definition 1 is applicable to each XjX_j and requires the existence of a KIS for each XjX_j, while Definition 6 requires the existence of a common KIS for all the vector fields, that is, a common embedding.

Remark 3. The constant function 11 may be included as a component of Ψ\Psi. Thus, systems of the form

z˙=A0z+b0+∑j=1muj(t)(Ajz+bj),(10)\dot{z}=A_0z+b_0+\sum_{j=1}^{m}u_j(t)(A_jz+b_j), \tag*{(10)}

including linear control systems, can be represented as bilinear systems: see [7].

The main difficulty in establishing the existence of a finite-dimensional Koopman bilinear embedding is that Definition 6 requires a single embedding map Ψ\Psi that is common to all the vector fields X0,…,XmX_0,\ldots,X_m. We will resolve this difficulty by investigating an algebraic structure of the vector fields.

Lie-algebraic interpretation of Koopman bilinear embeddings

This subsection provides a Lie-algebraic interpretation of the existence problem for Koopman bilinear embeddings. Recall from Proposition 1 that X(M)\mathcal{X}(M) is a real Lie algebra and that π(X):=LX\pi(X):=L_X is a representation. Let

Y:=Lie⁡(X0,…,Xm)\mathfrak{Y}:=\operatorname{Lie}(X_0,\ldots,X_m)

denote the Lie subalgebra1 of X(M)\mathcal{X}(M) generated by X0,…,XmX_0,\ldots,X_m as defined in Definition 4. This Lie algebra Y\mathfrak{Y} may be infinite-dimensional even though it is generated by finitely many vector fields. The following proposition provides a Lie-algebraic interpretation of the existence problem for Koopman bilinear embedding.

Proposition 2. The control-affine system (M,Xj)0≤j≤m(M,X_j)_{0\le j\le m} admits a Koopman bilinear embedding if and only if there exists a finite-dimensional subspace F⊂C∞(M;C)\mathcal{F}\subset C^\infty(M;\mathbb{C}) with the following two properties:

(i) For some basis ψ1,…,ψN\psi_1,\ldots,\psi_N of F\mathcal{F}, the map Ψ=(ψ1,…,ψN)⊤:M→CN\Psi=(\psi_1,\ldots,\psi_N)^\top:M\to\mathbb{C}^N is a C∞C^\infty embedding;

(ii) The representation

πY:Y→End⁡(F),πY(Y):=LY∣F(11)\pi_{\mathfrak{Y}}:\mathfrak{Y}\to\operatorname{End}(\mathcal{F}),\qquad\pi_{\mathfrak{Y}}(Y):=L_Y|_{\mathcal{F}} \tag*{(11)}

is well-defined and injective.

Proof. See Appendix A.

Property (i) ensures that a basis of F\mathcal{F} defines an embedding of MM. The well-definedness of πY\pi_{\mathfrak{Y}} implies

LYF⊆F,∀Y∈Y.(12)L_Y\mathcal{F}\subseteq\mathcal{F},\qquad\forall Y\in\mathfrak{Y}. \tag*{(12)}

Property (ii) thus reformulates the common-invariance condition (9) as an invariance condition under LYL_Y for every Y∈YY\in\mathfrak{Y}, including not only X0,…,XmX_0,\ldots,X_m but also their Lie brackets [Xi,Xj][X_i,X_j] and iterated Lie brackets [[Xi,Xj],Xk][[X_i,X_j],X_k], [[Xi,Xj],[Xk,Xℓ]],…[[X_i,X_j],[X_k,X_\ell]],\ldots These additional vector fields enable us to unravel a problem, for instance, to describe a structure that is never clarified with the original vector fields only. The Lie-algebraic interpretation in Proposition 2 leads to the necessary condition in the next subsection.

A necessary condition

We now derive a necessary condition for the existence of a Koopman bilinear embedding based on Proposition 2. Property (ii) yields the following necessary condition.

Theorem 1. If the control-affine system (M,Xj)0≤j≤m(M,X_j)_{0\le j\le m} admits a Koopman bilinear embedding (Ψ,Aj)0≤j≤m(\Psi,A_j)_{0\le j\le m}, then the following condition holds:

(I) The generated Lie algebra Y=Lie⁡(X0,…,Xm)\mathfrak{Y}=\operatorname{Lie}(X_0,\ldots,X_m) is finite-dimensional.

Proof. Since πY:Y→End⁡(F)\pi_{\mathfrak{Y}}:\mathfrak{Y}\to\operatorname{End}(\mathcal{F}) is injective, dim⁡Y≤dim⁡End⁡(F)\dim\mathfrak{Y}\le\dim\operatorname{End}(\mathcal{F}). Thus dim⁡F<∞\dim\mathcal{F}<\infty implies dim⁡Y≤dim⁡End⁡(F)<∞\dim\mathfrak{Y}\le\dim\operatorname{End}(\mathcal{F})<\infty.

Thus, Condition (I) requires the Lie algebra Y\mathfrak{Y} generated by X0,…,XmX_0,\ldots,X_m to be finite-dimensional. Since Y\mathfrak{Y} contains not only X0,…,XmX_0,\ldots,X_m but also their Lie brackets and iterated Lie brackets, the finite-dimensionality in Condition (I) imposes a structural restriction on the class of systems that admit a common embedding Ψ\Psi.

It should be stated that Condition (I) can be investigated directly from the vector fields, without explicitly constructing Ψ\Psi. More precisely, define a sequence of real vector subspaces of X(M)\mathcal{X}(M) by

Y(1)=span⁡R{X0,…,Xm}\mathfrak{Y}^{(1)}=\operatorname{span}_{\mathbb{R}}\{X_0,\ldots,X_m\}
Y(k+1)=Y(k)+span⁡R{[Xj,Y]∣0≤j≤m, Y∈Y(k)},(13)\mathfrak{Y}^{(k+1)}=\mathfrak{Y}^{(k)}+\operatorname{span}_{\mathbb{R}}\{[X_j,Y]\mid0\le j\le m,\ Y\in\mathfrak{Y}^{(k)}\}, \tag*{(13)}

where k≥1k\ge1. Each Y(k)\mathfrak{Y}^{(k)} is clearly finite-dimensional, and we have2

Y(1)⊆Y(2)⊆⋯ ,Y=⋃k=1∞Y(k).\mathfrak{Y}^{(1)}\subseteq\mathfrak{Y}^{(2)}\subseteq\cdots,\qquad\mathfrak{Y}=\bigcup_{k=1}^{\infty}\mathfrak{Y}^{(k)}.

If there exists an integer k≥1k\ge1 such that Y(k+1)=Y(k)\mathfrak{Y}^{(k+1)}=\mathfrak{Y}^{(k)}, then we see that a finite-dimensional generated Lie algebra Y=Y(k)\mathfrak{Y}=\mathfrak{Y}^{(k)} exists. If no such kk exists, then the dimensions of Y(k)\mathfrak{Y}^{(k)} increase without bound, that is to say, Y\mathfrak{Y} is infinite-dimensional. In such a case, by Theorem 1, we conclude that no Koopman bilinear embedding exists for the control-affine system.

A sufficient condition via nilpotent ideals

To derive a sufficient condition for Koopman bilinear embedding, we investigate the structure of the Lie algebra Y=Lie⁡(X0,…,Xm)\mathfrak{Y}=\operatorname{Lie}(X_0,\ldots,X_m). Specifically, in this subsection, we derive a sufficient condition based on the existence of a nilpotent ideal in Y\mathfrak{Y}.

We first introduce the notion of a finite-algebraic-multiplicity Koopman linear embedding. Let LX:C∞(M;C)→C∞(M;C)L_X:C^\infty(M;\mathbb{C})\to C^\infty(M;\mathbb{C}) be the Koopman generator associated with a vector field X∈X(M)X\in\mathcal{X}(M). For λ∈C\lambda\in\mathbb{C}, using the kernel space of the operator LX−λIL_X-\lambda I (II is the identity), define

Rλ(LX):=⋃k=1∞Ker⁡(LX−λI)k\mathcal{R}_{\lambda}(L_X):=\bigcup_{k=1}^{\infty}\operatorname{Ker}(L_X-\lambda I)^k

as a subspace of C∞(M;C)C^\infty(M;\mathbb{C}). This is called the algebraic root space3 of LXL_X associated with λ\lambda. Then, the finite-algebraic-multiplicity Koopman linear embedding is defined as follows.

Definition 7. A Koopman linear embedding (Ψ,A)(\Psi,\mathcal{A}) of the autonomous system (M,X)(M,X) is said to have finite algebraic multiplicity if dim⁡Rλ(LX)<∞\dim\mathcal{R}_{\lambda}(L_X)<\infty for every λ∈σ(A)\lambda\in\sigma(\mathcal{A}). We also say that (M,X)(M,X) admits a finite-algebraic-multiplicity Koopman linear embedding.

As a simple example, consider the system x˙=diag⁡(1,a)x\dot{x}=\operatorname{diag}(1,a)x with x=(x1,x2)⊤x=(x_1,x_2)^\top. For a>0a>0, the identity map itself is a finite-algebraic-multiplicity Koopman linear embedding. If a=0a=0, however, x2k∈R0(LX)x_2^k\in\mathcal{R}_0(L_X) for every k≥0k\ge0, so dim⁡R0(LX)=∞\dim\mathcal{R}_0(L_X)=\infty and a finite-algebraic-multiplicity Koopman linear embedding does not exist.

Next, we recall the notions of an ideal and a nilpotent Lie algebra from [23].

Definition 8. A Lie subalgebra h⊂g\mathfrak{h}\subset\mathfrak{g} is called an ideal if [X,Z]∈h[X,Z]\in\mathfrak{h} for every X∈gX\in\mathfrak{g} and Z∈hZ\in\mathfrak{h}.

Definition 9. A Lie algebra g\mathfrak{g} is called nilpotent if there exists a positive integer rr such that [Y1,[Y2,…,[Yr,Yr+1]…]]=0[Y_1,[Y_2,\ldots,[Y_r,Y_{r+1}]\ldots]]=0 for all Y1,…,Yr+1∈gY_1,\ldots,Y_{r+1}\in\mathfrak{g}.

With these definitions, we have the following sufficient condition.

Theorem 2. The control-affine system (M,Xj)0≤j≤m(M,X_j)_{0\leq j\leq m} admits a Koopman bilinear embedding (Ψ,Aj)0≤j≤m(\Psi,A_j)_{0\leq j\leq m} if the following conditions hold:

(A) The generated Lie algebra Y=Lie⁡(X0,…,Xm)\mathfrak{Y}=\operatorname{Lie}(X_0,\ldots,X_m) is finite-dimensional and contains a nonzero nilpotent ideal J\mathfrak{J}.

(B) There exists a vector field Z∈JZ\in\mathfrak{J} such that the autonomous system (M,Z)(M,Z) admits a finite-algebraic-multiplicity Koopman linear embedding.

Proof. See Appendix B.

As a simple example, suppose that X0,…,XmX_0,\ldots,X_m commute pairwise. Then, J=span⁡R{X0,…,Xm}\mathfrak{J}=\operatorname{span}_{\mathbb{R}}\{X_0,\ldots,X_m\} is a commutative Lie subalgebra and hence is nilpotent, and Condition (A) holds with J=Y\mathfrak{J}=\mathfrak{Y}. In this case, the control-affine system (M,Xj)0≤j≤m(M,X_j)_{0\leq j\leq m} admits a Koopman bilinear embedding if one of the autonomous systems (M,X0),…,(M,Xm)(M,X_0),\ldots,(M,X_m) admits a finite-algebraic-multiplicity Koopman linear embedding in Definition 7.

It should be noted that not every finite-dimensional Lie algebra contains a nonzero nilpotent ideal. The conditions in Theorem 2 are therefore restrictive. Establishing a sufficient condition applicable to more general classes of generated Lie algebras is left for future work.

Remark 4. We compare Theorem 2 with the differential-geometric conditions for Koopman linear embedding reported in [19]. That work considers a local map z=Φ(x)z=\Phi(x), defined on a neighborhood of an equilibrium point x0x_0 of the drift vector field X0X_0, that transforms a control-affine system into a linear system of the form z˙=Az+Bu\dot{z}=Az+Bu. By taking MM to be a sufficiently small neighborhood of x0x_0, [19] shows that such a transformation into a controllable linear system exists if and only if the following conditions hold:

(a1) Δ(x):={ad⁡X0kXi(x)∣1≤i≤m, k≥0}\Delta(x):=\{\operatorname{ad}_{X_0}^kX_i(x)\mid1\leq i\leq m,\ k\geq0\} satisfies

Δ(x)=TxMfor every x∈M.\Delta(x)=T_xM\quad\text{for every }x\in M.

(a2) [ad⁡X0kXi,ad⁡X0ℓXj]=0[\operatorname{ad}_{X_0}^kX_i,\operatorname{ad}_{X_0}^{\ell}X_j]=0 for every 1≤i,j≤m1\leq i,j\leq m and k,ℓ≥0k,\ell\geq0.

Here, the iterated adjoint actions are defined recursively by ad⁡X0Y:=Y\operatorname{ad}_X^0Y:=Y and ad⁡XkY:=[X,ad⁡Xk−1Y]\operatorname{ad}_X^kY:=[X,\operatorname{ad}_X^{k-1}Y] for k≥0k\geq0. Although these conditions are formulated locally, they become necessary conditions for a global Koopman linearization of the form z˙=Az+Bu\dot{z}=Az+Bu.

To compare Conditions (a1) and (a2) with Theorem 2, define

JD:=span⁡R{ad⁡X0kXi∣1≤i≤m, k≥0}.\mathfrak{J}_{\mathrm{D}}:=\operatorname{span}_{\mathbb{R}}\{\operatorname{ad}_{X_0}^kX_i\mid1\leq i\leq m,\ k\geq0\}.

Condition (a2) implies that JD\mathfrak{J}_{\mathrm{D}} is a commutative Lie algebra containing X1,…,XmX_1,\ldots,X_m. Condition (a2) also implies that [X0,Y]=ad⁡X0Y∈JD[X_0,Y]=\operatorname{ad}_{X_0}Y\in\mathfrak{J}_{\mathrm{D}} for every Y∈JDY\in\mathfrak{J}_{\mathrm{D}}. Thus, JD\mathfrak{J}_{\mathrm{D}} is an ideal of Y=Lie⁡(X0,…,Xm)\mathfrak{Y}=\operatorname{Lie}(X_0,\ldots,X_m) and

Y=span⁡R{X0}+JD.(14)\mathfrak{Y}=\operatorname{span}_{\mathbb{R}}\{X_0\}+\mathfrak{J}_{\mathrm{D}}. \tag*{(14)}

Furthermore, Condition (a1) implies that JD\mathfrak{J}_{\mathrm{D}} is transitive,4 which, together with commutativity of JD\mathfrak{J}_{\mathrm{D}}, yields dim⁡JD=dim⁡M\dim\mathfrak{J}_{\mathrm{D}}=\dim M; see [22 pp. 231–236]. Consequently, we have dim⁡Y≤n+1<∞\dim\mathfrak{Y}\leq n+1<\infty. Therefore, Conditions (a1) and (a2) imply Condition (A) of Theorem 2.

From this Lie-algebraic viewpoint, we can state that the structure arising from the two conditions in [19] is a special case of Condition (A) in which the nilpotent ideal J\mathfrak{J} is commutative and the generated Lie algebra takes the form (14). In this sense, the Lie-algebraic requirement in Condition (A) is more general than the structure imposed by Conditions (a1) and (a2). Since a linear system of the form z˙=Az+Bu\dot{z}=Az+Bu can be represented as a bilinear system as stated in Remark 3, the linear systems considered in [19] are included in our bilinear formulation.

However, Theorem 2 guarantees only the existence of a Koopman bilinear embedding and does not characterize when such an embedding can be reduced to a linear system of the form z˙=Az+Bu\dot{z}=Az+Bu. It also remains unclear whether Conditions (a1) and (a2) imply Condition (B). Clarifying the relationship between these conditions and Condition (B), particularly from a differential-geometric viewpoint, is left for future work.

Example and discussion

This subsection illustrates how Condition (I) in Theorem 1 can be checked for a two-dimensional, single-input polynomial system. Let M=R2M=\mathbb{R}^2, m=1m=1, and consider

\begin{pmatrix}\dot{x}\\\dot{y}\end{pmatrix} = \begin{pmatrix}y\\f(x,y)\end{pmatrix} +u\begin{pmatrix}0\\1\end{pmatrix}, \tag*{(15)}\end{pmatrix}

where f:R2→Rf:\mathbb{R}^2\to\mathbb{R} is a real polynomial of degree at most two. Equivalently, consider the vector fields

X0(x,y):=y∂∂x+f(x,y)∂∂y,X1(x,y):=∂∂y.X_0(x,y):=y\frac{\partial}{\partial x}+f(x,y)\frac{\partial}{\partial y},\qquad X_1(x,y):=\frac{\partial}{\partial y}.

This system can be viewed as a quadratic Taylor approximation at the origin of a one-degree-of-freedom equation of motion x¨=F(x,x˙)+u\ddot{x}=F(x,\dot{x})+u with F(x,x˙)=f(x,y)+O(∥(x,y)∥3)F(x,\dot{x})=f(x,y)+O(\|(x,y)\|^3) where OO represents Landau’s big-OO notation.

The following proposition characterizes the polynomials ff for which the generated Lie algebra is finite-dimensional.

Proposition 3. For the system (15), the generated Lie algebra Lie⁡(X0,X1)\operatorname{Lie}(X_0,X_1) is finite-dimensional if and only if

∂2f∂x2(0)=∂2f∂x∂y(0)=∂2f∂y2(0)∂f∂x(0)=0.(16)\frac{\partial^2 f}{\partial x^2}(0)=\frac{\partial^2 f}{\partial x\partial y}(0)=\frac{\partial^2 f}{\partial y^2}(0)\frac{\partial f}{\partial x}(0)=0. \tag*{(16)}

Equivalently, ff has one of the forms of

f(x,y)=f02y2+f01y+f00orf(x,y)=f10x+f01y+f00f(x,y)=f_{02}y^2+f_{01}y+f_{00}\quad\text{or}\quad f(x,y)=f_{10}x+f_{01}y+f_{00}

where f02,f10,f01,f00∈Rf_{02}, f_{10}, f_{01}, f_{00} \in\mathbb{R}.

A complete proof of Proposition 3 covering all possible coefficients is omitted. Instead, Table 1 compares two representative cases: f(x,y)=y2+y+1f(x,y)=y^2+y+1, which satisfies (16), and f(x,y)=xyf(x,y)=xy, which violates it. For f(x,y)=y2+y+1f(x,y)=y^2+y+1, Table 1 shows that Y(4)=Y(3)\mathfrak{Y}^{(4)}=\mathfrak{Y}^{(3)}. As demonstrated in Section III-C, this implies that Lie⁡(X0,X1)=Y(3)\operatorname{Lie}(X_0,X_1)=\mathfrak{Y}^{(3)}, which is four-dimensional. For f(x,y)=xyf(x,y)=xy, by contrast, the table shows polynomial vector fields of progressively higher degree. Proposition 3 establishes that Lie⁡(X0,X1)\operatorname{Lie}(X_0,X_1) is infinite-dimensional in this case. Similar degree growth occurs for other polynomials that violate (16). These calculations illustrate that the sequence (13) provides a useful diagnostic for the finite-dimensionality of the generated Lie algebra.

f(x,y)=y2+y+1f(x,y)=y^2+y+1f(x,y)=xyf(x,y)=xy
Y(1)\mathfrak{Y}^{(1)}⟨(10),(yy2+y+1)⟩\left\langle\binom{1}{0},\binom{y}{y^2+y+1}\right\rangle⟨(10),(yxy)⟩\left\langle\binom{1}{0},\binom{y}{xy}\right\rangle
Y(2)\mathfrak{Y}^{(2)}Y(1)+⟨(12y+1)⟩\mathfrak{Y}^{(1)}+\left\langle\binom{1}{2y+1}\right\rangleY(1)+⟨(1x)⟩\mathfrak{Y}^{(1)}+\left\langle\binom{1}{x}\right\rangle
Y(3)\mathfrak{Y}^{(3)}Y(2)+⟨(2y+12y2+2y−1)⟩\mathfrak{Y}^{(2)}+\left\langle\binom{2y+1}{2y^2+2y-1}\right\rangleY(2)+⟨(xx2)⟩\mathfrak{Y}^{(2)}+\left\langle\binom{x}{x^2}\right\rangle
Y(4)\mathfrak{Y}^{(4)}Y(3)\mathfrak{Y}^{(3)}Y(3)+⟨(x2−yx3−2xy)⟩\mathfrak{Y}^{(3)}+\left\langle\binom{x^2-y}{x^3-2xy}\right\rangle
Y(5)\mathfrak{Y}^{(5)}Y(4)\mathfrak{Y}^{(4)}Y(4)+⟨(x3−2xyx4−2x2y)⟩\mathfrak{Y}^{(4)}+\left\langle\binom{x^3-2xy}{x^4-2x^2y}\right\rangle

Table 1. Sequence (13) for two representative polynomials. Here, the notation ⟨v1,…,vr⟩:=span⁡{v1,…,vr}\langle v_1,\ldots,v_r\rangle:= \operatorname{span}\{v_1,\ldots,v_r\} is used.

Proposition 3 can be interpreted as follows: Let P≤2(R2;R)\mathcal{P}^{\leq2}(\mathbb{R}^2;\mathbb{R}) represent the vector space of real polynomials in (x,y)(x,y) of degree at most two. Because ff can be written as f(x,y)=f20x2+f11xy+f02y2+f10x+f01y+f00f(x,y)=f_{20}x^2+f_{11}xy+f_{02}y^2+f_{10}x+f_{01}y+f_{00}, P≤2(R2;R)\mathcal{P}^{\leq2}(\mathbb{R}^2;\mathbb{R}) can be identified with R6\mathbb{R}^6. Define

Pfinite≤2:={f∈P≤2(R2;R)∣dim⁡Lie⁡(X0,X1)<∞}.\mathcal{P}^{\leq2}_{\mathrm{finite}}:=\{f\in\mathcal{P}^{\leq2}(\mathbb{R}^2;\mathbb{R})\mid\dim\operatorname{Lie}(X_0,X_1)<\infty\}.

By Proposition 3, we derive

Pfinite≤2={f∈P≤2(R2;R)∣f20=f11=f02=0}∪{f∈P≤2(R2;R)∣f20=f11=f10=0}.\mathcal{P}^{\leq2}_{\mathrm{finite}}=\{f\in\mathcal{P}^{\leq2}(\mathbb{R}^2;\mathbb{R})\mid f_{20}=f_{11}=f_{02}=0\} \cup\{f\in\mathcal{P}^{\leq2}(\mathbb{R}^2;\mathbb{R})\mid f_{20}=f_{11}=f_{10}=0\}.

Thus, Pfinite≤2\mathcal{P}^{\leq2}_{\mathrm{finite}} is nowhere dense and has Lebesgue measure zero with respect to P≤2(R2;R)\mathcal{P}^{\leq2}(\mathbb{R}^2;\mathbb{R}). Hence, Theorem 1 implies that the set of polynomials ff for which a finite-dimensional Koopman bilinear embedding can exist is nowhere dense and of Lebesgue measure zero. In this sense, an exact finite-dimensional Koopman bilinear embedding is nongeneric for f∈P≤2(R2;R)f\in\mathcal{P}^{\leq2}(\mathbb{R}^2;\mathbb{R}).

This conclusion is restricted to two-dimensional, single-input systems with quadratic polynomial dynamics of the form (15) and does not directly extend to general control-affine systems. Nevertheless, this example demonstrates that the finite-dimensionality of the generated Lie algebra imposes a strong algebraic restriction. It therefore suggests that one should be cautious about assuming the existence of an exact finite-dimensional Koopman bilinear embedding without additional structure.

This observation also has implications for data-driven Koopman modeling. In general, a learned finite-dimensional model should be regarded as an approximation of the form

z˙=A0z+∑j=1mujAjz+rΨ(z,u),\dot{z}=A_0z+\sum_{j=1}^{m}u_jA_jz+r_{\Psi}(z,u),

where rΨ(z,u)r_{\Psi}(z,u) accounts for the part of the lifted dynamics that cannot be captured within the finite-dimensional span of the components of Ψ\Psi. Thus, even when a large dictionary of observables is employed, it is important to assess the residual rΨ(z,u)r_{\Psi}(z,u) explicitly; see [15], [16] and [11 Section 4].

Conclusion

This paper investigated the existence of finite-dimensional Koopman bilinear embeddings for control-affine systems from the perspective of the Lie algebra generated by the underlying vector fields and their representations. We showed that the finite-dimensionality of the generated Lie algebra is a necessary condition for such an embedding. We also derived a sufficient condition applicable when the generated Lie algebra contains a nilpotent ideal. Our analysis of a quadratic polynomial system indicates that the finite-dimensionality imposes strong restrictions on a class of vector fields, suggesting that finite-dimensional Koopman bilinear embeddings may be nongeneric. Future work will introduce appropriate norms on the generated Lie algebra and develop a quantitative theory of finite-dimensional bilinear approximations and their approximation errors.

Acknowledgement

The authors used ChatGPT (OpenAI) to assist with the derivations of Theorem 2 and Proposition 3, the comparison of Theorem 2 with the result reported in [19], the identification and correction of issues in the proofs of Proposition 2 and Theorem 2, the calculations in Section IV, searches for relevant literature, and language editing to improve the clarity and readability of the manuscript. All AI-generated suggestions were reviewed and edited by the authors to ensure accuracy and integrity.

Proof of Proposition 2

Proof. (⟹\Longrightarrow) Assume that there exist a Koopman bilinear embedding (Ψ,Aj)0≤j≤m(\Psi,A_j)_{0\leq j\leq m}. Set F=span⁡C{ψ1,…,ψN}\mathcal{F}=\operatorname{span}_{\mathbb{C}}\{\psi_1,\ldots,\psi_N\}. Property (i) follows directly from the definition of a Koopman bilinear embedding. To prove Property (ii), we first show that

LYF⊆F,∀Y∈g.(17)L_Y\mathcal{F}\subseteq\mathcal{F},\qquad\forall Y\in\mathfrak{g}. \tag*{(17)}

Define

XF:={Y∈X(M)∣LYF⊆F}.(18)\mathfrak{X}_{\mathcal{F}}:=\{Y\in\mathfrak{X}(M)\mid L_Y\mathcal{F}\subseteq\mathcal{F}\}. \tag*{(18)}

This is a Lie subalgebra. Indeed, for every Z,Y∈XFZ,Y\in\mathfrak{X}_{\mathcal{F}}, α,β∈R\alpha,\beta\in\mathbb{R}, and f∈Ff\in\mathcal{F}, we have

LαY+βZf=αLYf+βLZf∈F,L_{\alpha Y+\beta Z}f=\alpha L_Yf+\beta L_Zf\in\mathcal{F},
L[Y,Z]f=(LYLZ−LZLY)f∈F,L_{[Y,Z]}f=(L_YL_Z-L_ZL_Y)f\in\mathcal{F},

which yields αY+βZ∈XF\alpha Y+\beta Z\in\mathfrak{X}_{\mathcal{F}} and [Y,Z]∈XF[Y,Z]\in\mathfrak{X}_{\mathcal{F}}. Since we assumed the existence of a Koopman bilinear embedding, (9) gives X0,…,Xm∈XFX_0,\ldots,X_m \in\mathfrak{X}_{\mathcal{F}}. Moreover, since Lie⁡(X0,…,Xm)\operatorname{Lie}(X_0,\ldots,X_m) is the smallest subalgebra containing X0,…,XmX_0,\ldots,X_m (recall Definition 4), X0,…,Xm∈XFX_0,\ldots,X_m \in\mathfrak{X}_{\mathcal{F}} yields Y=Lie⁡(X0,…,Xm)⊆XF\mathfrak{Y}=\operatorname{Lie}(X_0,\ldots,X_m)\subseteq\mathfrak{X}_{\mathcal{F}}. This, together with (18), gives (17). Consequently, πY(Y)=LY∈End⁡(F)\pi_{\mathfrak{Y}}(Y)=L_Y\in\operatorname{End}(\mathcal{F}) for every Y∈YY\in\mathfrak{Y}, so πY\pi_{\mathfrak{Y}} is well-defined.

It remains to show that πY\pi_{\mathfrak{Y}} is injective. If Y∈Ker⁡πYY\in\operatorname{Ker}\pi_{\mathfrak{Y}},

LYψk(x)=dψk(x)Y(x)=0,x∈M,k=1,…,N.L_Y\psi_k(x)=d\psi_k(x)Y(x)=0,\qquad x\in M,\quad k=1,\ldots,N.

Equivalently, dΨ(x)Y(x)=0d\Psi(x)Y(x)=0 for every x∈Mx\in M. Since Ψ\Psi is an embedding, dΨ(x):TxM→TΨ(x)CNd\Psi(x):T_xM\to T_{\Psi(x)}\mathbb{C}^N is injective. Thus, dΨ(x)Y(x)=0d\Psi(x)Y(x)=0 (∀x∈M)(\forall x\in M) implies Y=0Y=0, and hence πY\pi_{\mathfrak{Y}} is injective.

(⟸)(\Longleftarrow) Assume that there exists a finite-dimensional subspace F\mathcal{F} satisfying Properties (i) and (ii). Set N=dim⁡FN=\dim\mathcal{F} and let ψ1,…,ψN\psi_1,\ldots,\psi_N be the basis of F\mathcal{F} satisfying Property (i). Since πY\pi_{\mathfrak{Y}} is well-defined, LYF⊆FL_Y\mathcal{F}\subseteq\mathcal{F} for every Y∈YY\in\mathfrak{Y}, including X0,…,XmX_0,\ldots,X_m. Thus, we have LXjF⊆FL_{X_j}\mathcal{F}\subseteq\mathcal{F}. Therefore, for every j=0,…,mj=0,\ldots,m, there exists a matrix Aj=(akℓ(j))1≤k,ℓ≤N∈CN×NA_j=(a_{k\ell}^{(j)})_{1\leq k,\ell\leq N}\in\mathbb{C}^{N\times N} such that

LXjψk=∑ℓ=1Nakℓ(j)ψℓ,k=1,…,N.L_{X_j}\psi_k=\sum_{\ell=1}^{N}a_{k\ell}^{(j)}\psi_\ell,\qquad k=1,\ldots,N.

Thus, Ψ:=(ψ1,…,ψN)⊤\Psi:=(\psi_1,\ldots,\psi_N)^\top gives LXjΨ=AjΨL_{X_j}\Psi=A_j\Psi, and hence (Ψ,Aj)0≤j≤m(\Psi,A_j)_{0\leq j\leq m} is a Koopman bilinear embedding. □

Appendix B: Proof of Theorem 2

Proof. By Condition (B), there exists Z∈JZ\in\mathfrak{J} such that the autonomous system (M,Z)(M,Z) admits a finite-algebraic-multiplicity Koopman linear embedding (ΨZ,AZ)(\Psi_Z,A_Z). Define

FZ:=span⁡C{ϕ∈C∞(M;C)∣ϕ is a component of ΨZ}.\mathcal{F}_Z:=\operatorname{span}_{\mathbb{C}}\{\phi\in C^\infty(M;\mathbb{C})\mid\phi\text{ is a component of }\Psi_Z\}.

Then σ(AZ)=σ(LZ∣FZ)\sigma(A_Z)=\sigma(L_Z|_{\mathcal{F}_Z}). Moreover, by the finite-algebraic-multiplicity property, the algebraic root space Rλ(LZ)\mathcal{R}_\lambda(L_Z) is finite-dimensional for every λ∈σ(AZ)\lambda\in\sigma(A_Z), which implies that

F:=⨁λ∈σ(AZ)Rλ(LZ)(19)\mathcal{F}:=\bigoplus_{\lambda\in\sigma(A_Z)}\mathcal{R}_\lambda(L_Z) \tag*{(19)}

is finite-dimensional. It also follows that FZ⊆F\mathcal{F}_Z\subseteq\mathcal{F}. Indeed, if χZ(t)=∏λ∈σ(AZ)(t−λ)mλ\chi_Z(t)=\prod_{\lambda\in\sigma(A_Z)}(t-\lambda)^{m_\lambda} is the characteristic polynomial of LZ∣FZL_Z|_{\mathcal{F}_Z}, then Cayley–Hamilton theorem gives χZ(LZ)FZ={0}\chi_Z(L_Z)\mathcal{F}_Z=\{0\} or, equivalently, FZ⊆Ker⁡χZ(LZ)\mathcal{F}_Z\subseteq\operatorname{Ker}\chi_Z(L_Z). Lemma 3 with V=C∞(M;C)V=C^\infty(M;\mathbb{C}), T=LZT=L_Z, and p=χZp=\chi_Z then gives

FZ⊆Ker⁡χZ(LZ)=⨁λ∈σ(AZ)Ker⁡(LZ−λI)mλ⊆F.\mathcal{F}_Z\subseteq\operatorname{Ker}\chi_Z(L_Z)=\bigoplus_{\lambda\in\sigma(A_Z)}\operatorname{Ker}(L_Z-\lambda I)^{m_\lambda}\subseteq\mathcal{F}.

In particular, we have σ(LZ∣F)=σ(LZ∣FZ)=σ(AZ)\sigma(L_Z|_{\mathcal{F}})=\sigma(L_Z|_{\mathcal{F}_Z})=\sigma(A_Z).

The goal of the proof is to show that

LXF⊆F,∀X∈Y.(20)L_X\mathcal{F}\subseteq\mathcal{F},\qquad\forall X\in\mathfrak{Y}. \tag*{(20)}

Once this has been established, the invariance of F\mathcal{F} allows us to construct a C∞C^\infty map Ψ\Psi whose components form a basis of F\mathcal{F}, together with a matrix-valued map5{}^{5} A:Y→CN×NA:\mathfrak{Y}\to\mathbb{C}^{N\times N} such that

LXΨ=A(X)Ψ,∀X∈Y.L_X\Psi=A(X)\Psi,\qquad\forall X\in\mathfrak{Y}.

Furthermore, since FZ⊆F\mathcal{F}_Z\subseteq\mathcal{F}, there exists a matrix CC such that ΨZ=CΨ\Psi_Z=C\Psi. Because ΨZ\Psi_Z is an embedding, Ψ\Psi is also an embedding. Hence (Ψ,A(Xj))0≤j≤m(\Psi,A(X_j))_{0\leq j\leq m} is a Koopman bilinear embedding and the proof is completed.

We first establish the following invariance property for the nilpotent ideal J\mathfrak{J}.

Lemma 1. LYF⊆FL_Y\mathcal{F}\subseteq\mathcal{F} for every Y∈JY\in\mathfrak{J}.

Proof of Lemma 1. Since Rλ(LZ)\mathcal{R}_\lambda(L_Z) is finite-dimensional for every λ∈σ(AZ)\lambda\in\sigma(A_Z), there exists a positive integer qλq_\lambda such that

Ker⁡(LZ−λI)⊆Ker⁡(LZ−λI)2⊆⋯\operatorname{Ker}(L_Z-\lambda I)\subseteq\operatorname{Ker}(L_Z-\lambda I)^2\subseteq\cdots
⋯⊆Ker⁡(LZ−λI)qλ=Ker⁡(LZ−λI)qλ+1=⋯(21)\cdots\subseteq\operatorname{Ker}(L_Z-\lambda I)^{q_\lambda}=\operatorname{Ker}(L_Z-\lambda I)^{q_\lambda+1}=\cdots \tag*{(21)}
⋯=Rλ(LZ).\cdots=\mathcal{R}_\lambda(L_Z).

Choose a positive integer rr such that every (r+1)(r+1)-fold iterated Lie bracket of elements of J\mathfrak{J} vanishes, as in Definition 9. We claim that, for every f∈Rλ(LZ)f\in\mathcal{R}_\lambda(L_Z) and every Y∈JY\in\mathfrak{J},

(LZ−λI)qλ+r−1LYf=0.(22)(L_Z-\lambda I)^{q_\lambda+r-1}L_Yf=0. \tag*{(22)}

Let ad⁡Zk\operatorname{ad}_Z^k and ad⁡LZk\operatorname{ad}_{L_Z}^k denote the kk-fold adjoint actions of ZZ and LZL_Z, respectively. Since [LZ,LY]=L[Z,Y][L_Z,L_Y]=L_{[Z,Y]}, it follows that ad⁡LZk(LY)=Lad⁡ZkY\operatorname{ad}_{L_Z}^k(L_Y)=L_{\operatorname{ad}_Z^kY} for every k≥0k\geq0. The nilpotency of J\mathfrak{J} therefore implies that ad⁡LZkLY=0\operatorname{ad}_{L_Z}^kL_Y=0 for k≥rk\geq r. Applying Lemma 4 with V=C∞(M;C)V=C^\infty(M;\mathbb{C}), A=LZ−λIA=L_Z-\lambda I, B=LYB=L_Y, and n=qλ+r−1n=q_\lambda+r-1, we obtain

(LZ−λI)qλ+r−1LY=∑k=0r−1(qλ+r−1k)π(ad⁡LZkLY)(LZ−λI)qλ+r−1−k,(23)(L_Z-\lambda I)^{q_\lambda+r-1}L_Y =\sum_{k=0}^{r-1}\binom{q_\lambda+r-1}{k}\pi(\operatorname{ad}_{L_Z}^kL_Y)(L_Z-\lambda I)^{q_\lambda+r-1-k}, \tag*{(23)}

where we used ad⁡LZ−λI=ad⁡LZ\operatorname{ad}_{L_Z-\lambda I}=\operatorname{ad}_{L_Z} and ad⁡LZkLY=0\operatorname{ad}_{L_Z}^kL_Y=0 for k≥rk\geq r. Then, for every f∈Rλ(LZ)=Ker⁡(LZ−λI)qλf\in\mathcal{R}_\lambda(L_Z)=\operatorname{Ker}(L_Z-\lambda I)^{q_\lambda}, every term on the right-hand side of (23) annihilates ff. Thus, applying (23) to ff yields (22).

Equations (22) and (21) imply that LYf∈Rλ(LZ)L_Yf\in\mathcal{R}_\lambda(L_Z) for every f∈Rλ(LZ)f\in\mathcal{R}_\lambda(L_Z). This holds for every λ∈σ(AZ)\lambda\in\sigma(A_Z). Consequently, the space F\mathcal{F} defined in (19) satisfies LYF⊆FL_Y\mathcal{F}\subseteq\mathcal{F}, which proves the lemma. □

Next, consider the characteristic polynomial

χ(t):=det⁡(tIF−LZ∣F)\chi(t):=\det(tI_{\mathcal{F}}-L_Z|_{\mathcal{F}})

of the linear operator LZ∣FL_Z|_{\mathcal{F}}. Since σ(LZ∣F)=σ(AZ)\sigma(L_Z|_{\mathcal{F}})=\sigma(A_Z), there exist positive integers dλd_\lambda with λ∈σ(AZ)\lambda\in\sigma(A_Z) such that

χ(t)=∏λ∈σ(AZ)(t−λ)dλ.(24)\chi(t)=\prod_{\lambda\in\sigma(A_Z)}(t-\lambda)^{d_\lambda}. \tag*{(24)}

Applying the Cayley–Hamilton theorem to LZ∣FL_Z|_{\mathcal{F}} gives

χ(LZ)f=0,∀f∈F.(25)\chi(L_Z)f=0,\qquad\forall f\in\mathcal{F}. \tag*{(25)}

Since Z∈JZ\in\mathfrak{J}, where J\mathfrak{J} is an ideal, Lemma 1 yields the following result.

Lemma 2. [χ(LZ),LX]End⁡F⊆F[\chi(L_Z),L_X]_{\operatorname{End}\mathcal{F}}\subseteq\mathcal{F} for every X∈YX\in\mathfrak{Y}. Proof of Lemma 2. Since I\mathcal{I} is an ideal of Y\mathcal{Y}, Z∈IZ \in\mathcal{I} implies [Z,X]∈I[Z,X] \in\mathcal{I}. It follows from Lemma 1 that [LZ,LX]EndF=L[Z,X]F⊆F[L_Z,L_X]_{\mathrm{End}}\mathcal{F}=L_{[Z,X]}\mathcal{F}\subseteq\mathcal{F}. Moreover, for every positive integer kk,

[LZk+1,LX]End=LZk+1LX−LXLZk+1=∑ℓ=0kLZℓ(LZLX−LXLZ)LZk−ℓ=∑ℓ=0kLZℓL[Z,X]LZk−ℓ.[L_Z^{k+1},L_X]_{\mathrm{End}}=L_Z^{k+1}L_X-L_XL_Z^{k+1} =\sum_{\ell=0}^{k}L_Z^\ell(L_ZL_X-L_XL_Z)L_Z^{k-\ell} =\sum_{\ell=0}^{k}L_Z^\ell L_{[Z,X]}L_Z^{k-\ell}.

Since both LZL_Z and L[Z,X]L_{[Z,X]} preserve F\mathcal{F}, we have [LZk+1,LX]EndF⊆F[L_Z^{k+1},L_X]_{\mathrm{End}}\mathcal{F}\subseteq\mathcal{F}. Since χ\chi is a polynomial, we conclude that [χ(LZ),LX]EndF⊆F[\chi(L_Z),L_X]_{\mathrm{End}}\mathcal{F}\subseteq\mathcal{F}, which proves the lemma.

We now prove (20). Let f∈Ff\in\mathcal{F} and X∈YX\in\mathcal{Y}. The definition of [⋅,⋅]End[\cdot,\cdot]_{\mathrm{End}}, Equation (25), and Lemma 2 give

χ(LZ)LXf=LXχ(LZ)f+[χ(LZ),LX]Endf=[χ(LZ),LX]Endf∈F.\chi(L_Z)L_Xf=L_X\chi(L_Z)f+[\chi(L_Z),L_X]_{\mathrm{End}}f =[\chi(L_Z),L_X]_{\mathrm{End}}f\in\mathcal{F}.

Applying χ(LZ)\chi(L_Z) again to this identity and using (25) yields χ(LZ)2LXf=0\chi(L_Z)^2L_Xf=0, and hence LXf∈Ker⁡χ(LZ)2L_Xf\in\operatorname{Ker}\chi(L_Z)^2. Then, Lemma 3 with V=C∞(M;C)V=C^\infty(M;\mathbb{C}), T=LZT=L_Z, and p=χ2p=\chi^2 gives

Ker⁡χ(LZ)2=⨁λ∈σ(AZ)Ker⁡(LZ−λI)2dλ⊆⨁λ∈σ(AZ)Rλ(LZ)=F,\operatorname{Ker}\chi(L_Z)^2 =\bigoplus_{\lambda\in\sigma(A_Z)}\operatorname{Ker}(L_Z-\lambda I)^{2d_\lambda} \subseteq\bigoplus_{\lambda\in\sigma(A_Z)}\mathcal{R}_\lambda(L_Z)=\mathcal{F},

where the final equality follows from (19). We thus obtain LXf∈Ker⁡χ(LZ)2⊆FL_Xf\in\operatorname{Ker}\chi(L_Z)^2\subseteq\mathcal{F}. Therefore, (20) follows and the proof is completed.

Appendix C: Collection of Lemmas

Lemma 3. Let VV be a possibly infinite-dimensional complex vector space and let T:V→VT:V\to V be a linear operator. Suppose that p(t)=∏j=1spj(t)p(t)=\prod_{j=1}^{s}p_j(t) where p1,…,psp_1,\ldots,p_s are complex polynomials that are pairwise relatively prime. Then

Ker⁡p(T)=⨁j=1sKer⁡pj(T).\operatorname{Ker}p(T)=\bigoplus_{j=1}^{s}\operatorname{Ker}p_j(T).

Proof. See [25 Theorem 2.2.14].

Lemma 4. Let VV be a vector space and let A,B∈End⁡(V)A,B\in\operatorname{End}(V). Then, for any integer n≥0n\geq0,

AnB=∑k=0n(nk)ad⁡Ak(B)An−k.A^nB=\sum_{k=0}^{n}\binom{n}{k}\operatorname{ad}_A^k(B)A^{n-k}.

Proof. Clearly, AiC=(Ai−1C)A+Ai−1ad⁡ACA^iC=(A^{i-1}C)A+A^{i-1}\operatorname{ad}_A C for every C∈End⁡(V)C\in\operatorname{End}(V) and i≥1i\geq1. Applying this formula to AnBA^nB by replacing C=B,ad⁡AB,ad⁡A2B,…C=B,\operatorname{ad}_A B,\operatorname{ad}_A^2B,\ldots gives the result.

References

References

  1. [1]A. Mauroy, I. Mezić, and Y. Susuki, Eds., The Koopman Operator in Systems and Control: Concepts, Methodologies, and Applications. Springer, 2020.
  2. [2]S. L. Brunton, M. Budišić, E. Kaiser, and J. N. Kutz, “Modern koopman theory for dynamical systems,” SIAM Review, vol. 64, no. 2, 2022.arxiv.org/abs/2102.12086
  3. [3]M. O. Williams, I. G. Kevrekidis, and C. W. Rowley, “A data-driven approximation of the koopman operator: Extending dynamic mode decomposition,” Journal of Nonlinear Science, vol. 25, no. 6, pp. 1307–1346, 2015.
  4. [4]J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L. Proctor, Dynamic Mode Decomposition: Data-Driven Modeling of Complex Systems. SIAM, 2016.
  5. [5]M. Korda and I. Mezić, “Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control,” Automatica, vol. 93, pp. 149–160, 2018.DOI
  6. [6]J. L. Proctor, S. L. Brunton, and J. N. Kutz, “Generalizing Koopman theory to allow for inputs and control,” SIAM Journal on Applied Dynamical Systems, vol. 17, no. 1, pp. 909–930, 2018.DOI
  7. [7]D. Bruder, X. Fu, and R. Vasudevan, “Advantages of bilinear Koopman realizations for the modeling and control of systems with unknown dynamics,” IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 4369–4376, 2021.arxiv.org/abs/2010.09961
  8. [8]D. Goswami and D. A. Paley, “Bilinearization, reachability, and optimal control of control-affine nonlinear systems: A Koopman spectral approach,” IEEE Transactions on Automatic Control, vol. 67, no. 6, pp. 2715–2728, 2021.DOI
  9. [9]M. Haseli and J. Cortés, “Modeling nonlinear control systems via Koopman control family: Universal forms and subspace invariance proximity,” arXiv preprint arXiv:2307.15368, 2023.arxiv.org/abs/2307.15368
  10. [10]L. C. Jacob, R. Tóth, and M. Schoukens, “Koopman form of nonlinear systems with inputs,” Automatica, vol. 162, p. 111525, 2024.
  11. [11]R. Strässer, K. Worthmann, I. Mezić, J. Berberich, M. Schaller, and F. Allgöwer, “An overview of koopman-based control: From error bounds to closed-loop guarantees,” Annual Reviews in Control, vol. 61, p. 101035, 2026.DOI
  12. [12]S. L. Brunton, B. W. Brunton, J. L. Proctor, and J. N. Kutz, “Koopman invariant subspaces and finite linear representations of nonlinear dynamical systems for control,” PLoS one, vol. 11, no. 2, p. e0150171, 2016.DOI
  13. [13]P. Bevanda, B. Driessen, L. C. Jacob, S. Sosnowski, R. Tóth, and S. Hirche, “Nonparametric control koopman operators,” IEEE Transactions on Automatic Control, 2026.DOI
  14. [14]M. Lazar, “From product hilbert spaces to the generalized koopman operator and the nonlinear fundamental lemma,” IEEE Transactions on Automatic Control, 2026.arxiv.org/abs/2508.07494
  15. [15]M. Schaller, K. Worthmann, F. Philipp, S. Peitz, and F. Nüske, “Towards reliable data-based optimal and predictive control using extended dmd,” IFAC-PapersOnLine, vol. 56, no. 1, pp. 169–174, 2023.arxiv.org/abs/2202.09084
  16. [16]R. Strässer, M. Schaller, K. Worthmann, J. Berberich, and F. Allgöwer, “Safedmd: A koopman-based data-driven controller design framework for nonlinear dynamical systems,” Automatica, vol. 185, p. 112732, 2026.DOI
  17. [17]Z. Liu, N. Ozay, and E. D. Sontag, “Properties of immersions for systems with multiple limit sets with implications to learning Koopman embeddings,” Automatica, vol. 176, p. 112226, 2025.arxiv.org/abs/2312.17045
  18. [18]M. D. Kalvheim and P. Arathoon, “Linearizability of flows by embeddings,” Selecta Mathematica, vol. 32, no. 2, p. 38, 2026.
  19. [19]S. A. Deka, “Differential geometric conditions for koopman linearizability of control-affine systems,” arXiv preprint arXiv:2606.13577, 2026.arxiv.org/abs/2606.13577
  20. [20]X. Shang, M. Haseli, J. Cortés, and Y. Zheng, “On the existence of koopman linear embeddings for controlled nonlinear systems,” arXiv preprint arXiv:2602.14537, 2026.arxiv.org/abs/2602.14537
  21. [21]S. Sastry, Nonlinear systems: analysis, stability, and control. Springer, 1999, vol. 10.DOI
  22. [22]J. M. Lee, Introduction to Smooth Manifolds. Springer, 2012.
  23. [23]J. E. Humphreys, Introduction to Lie Algebras and Representation Theory, ser. Graduate Texts in Mathematics. Springer-Verlag, 1972, vol. 9.DOI
  24. [24]D. L. Elliott, Bilinear Control Systems: Matrices in Action. Springer, 2009, vol. 169.DOI
  25. [25]N. Brady, “Abstract linear algebra,” 2000, lecture notes. [Online]. Available: https://math.ou.edu/~nbrady/teaching/100-4373/detail.pdf

Paper details

vlnc:W_ktwptkcq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

ChatGPT (OpenAI)

View statement
Looking for related papers…
Contents