betaFeedback

WorksEngineering

Data-driven control of polynomial systems using implicit methods

Abstract

While conventional behavioral data-driven control represents linear time-invariant (LTI) systems directly from data, extending these data representations to nonlinear systems remains an open challenge. We address this gap for general polynomial systems, characterizing their finite-length trajectories as a real algebraic variety defined by implicit polynomial relations rather than explicit equations. First, we provide conditions on measured data under which these data can be used to represent the system. Building on this representation, we formulate data-driven simulation and predictive control problems that are equivalent to their model-based counterparts, and solve the resulting polynomial systems globally using homotopy continuation. Numerical case studies confirm the effectiveness of this approach by successfully simulating a stiff Van der Pol oscillator in situations where explicit models fail. Furthermore, these data-driven simulation and control methods are illustrated on a Duffing oscillator, demonstrating that our method outperforms Koopman-based approaches in predicting long trajectories.

Introduction

Data-driven predictive control based on behavioral systems theory has become popular for linear systems [10, 6, 13]. The behavioral approach describes a system by the set of its trajectories rather than by a particular representation through equations. For linear time-invariant (LTI) systems, the set of trajectories is a subspace, and this geometric viewpoint allows for a direct data representation of the behavior by the Fundamental Lemma [44, 28]. For nonlinear systems, the set of trajectories is not a linear subspace, and extending such direct data representations is an open problem, as also highlighted in the reviews by Markovsky and Dörfler [27]; Berberich and Allgöwer [5]. In this work, we focus on polynomial systems, for which this set is a real algebraic variety, and we show that it can be represented directly by data, enabling data-driven simulation and control.

Related work: in the data-driven control literature, nonlinear systems are often described by explicit models in discrete-time, such as nonlinear state-space or nonlinear ARX (NARX) equations [32, 12, 8]. While powerful, these explicit descriptions treat variables asymmetrically by isolating the future state or output, which often fails to capture the nonlinear geometry of the system. On the other hand, the behavioral approach preserves the geometric structure by treating all variables symmetrically, enabling implicit descriptions, see § 2.1 for an illustration. Data-driven methods inspired by the Fundamental Lemma have also been extended to nonlinear systems [5]. Many of these extensions are restricted to specific system classes, such as Volterra [38], Hammerstein and Wiener (Berberich and Allgöwer, 2020), or generalized bilinear systems [26], while others address general nonlinear systems through NARX models [34, 15], basis functions [24], kernel methods [20], or local linearization [7]. Most importantly, all these methods rely on explicit equations, which forgoes the main advantage of the geometric, representation-free viewpoint.

Koopman operator methods lift the state by a dictionary of observables and fit a linear model in the lifted space [45, 23, 33]. This general framework is suitable for describing implicit relations in the data. In practice, however, the lifted state is propagated by the linear model and mapped back to the state by a linear decoder [33], so nonlinear relations between the observables are typically lost along the prediction. Reprojection onto the set of lifted states restores this consistency [42, 33], but the predictions remain approximate unless the observables span a Koopman-invariant subspace, which can be infinite-dimensional even for polynomial dynamics.

Beyond data-driven control, relevant approaches also exist in the learning and system identification literature. For instance, the SINDy framework identifies nonlinear dynamics by regressing the time derivative of the state on a library of candidate functions with sparse coefficients [9]. While Mangan et al. [25]; Kaheman et al. [21] propose implicit variants of the SINDy algorithm, they aim at identifying continuous-time dynamics, and therefore require measuring or estimating time derivatives, which can be restrictive. This same limitation applies to Guo et al. [17]; Huang et al. [19] that design controllers for continuous-time polynomial systems via sum-of-squares programs. Finally, manifold learning methods also exploit that the data lie on a low-dimensional manifold, learned together with the dynamics on it [16] or via a diffusion model [22], but these approximations lack conditions on the data that guarantee an exact representation.

Contributions: our approach overcomes the limitations of the existing works by characterizing implicit relations between the variables. We represent these relations directly by data, and solve the resulting polynomial equations instead of relying on linear methods. Specifically, our contributions are as follows.

  • We motivate implicit data-driven methods through the example of simulating continuous-time systems from sampled data, and propose a pragmatic implicit behavioral description for polynomial systems, defined by the zero set of multiple, multivariate polynomials.

  • We provide conditions on the data (akin to the generalized persistency of excitation for LTI systems) enabling a direct data representation of polynomial behaviors with bounded degree. This presents a distinct challenge compared to the LTI case, as not only the defining equations but also their polynomial multiples vanish on the data, so the relevant object is a polynomial ideal rather than a subspace.

  • We propose computable data-driven simulation and predictive control methods based on the data representation, which are equivalent to their counterparts with known behavior under these conditions. These problems lead to polynomial systems of equations, which we solve globally with homotopy continuation (as in Rostalski et al. [37] for the model-based case).

  • We illustrate the data-driven simulation and control methods on a Duffing oscillator, showing superior prediction accuracy compared to Koopman based schemes. Further, we confirm the benefits of implicit descriptions over explicit models numerically on a Van der Pol oscillator.

Outline: Section 2 motivates implicit models and defines polynomial behaviors. Section 3 derives the direct data representation, which Section 4 uses for data-driven simulation and control. Section 5 contains the numerical case studies, and Section 6 concludes the paper.

Notation: let R\mathbb{R}, Z\mathbb{Z}, and N\mathbb{N} denote the sets of real numbers, integers, and positive integers, respectively. For a matrix M∈Rm×nM \in\mathbb{R}^{m \times n}, we denote its rank, image and null space by rank⁡(M)\operatorname{rank}(M), im⁡(M)\operatorname{im}(M) and null⁡(M)\operatorname{null}(M), respectively. Let R[x]\mathbb{R}[x] denote the polynomial ring in n≥1n \ge1 variables xx, and R[x]≤d⊆R[x]\mathbb{R}[x]_{\le d} \subseteq\mathbb{R}[x] the polynomials of degree at most dd. An ideal I⊂R[x]I \subset\mathbb{R}[x] is a set of polynomials that satisfies $0 \in I, a+b∈Ia+b\in I, −a∈I-a\in I, and a⋅c∈Ia\cdot c\in I for all a,b∈Ia,b\in I and all c∈R[x]c\in\mathbb{R}[x]. For an ideal II of R[x]\mathbb{R}[x], let I≤d=I∩R[x]≤dI_{\le d}=I\cap\mathbb{R}[x]_{\le d}.

Given polynomials pi∈R[x]p_i\in\mathbb{R}[x], i∈{1,…,r}i\in\{1,\ldots,r\}, we denote the ideal generated by them as

⟨p1,…,pr⟩={∑i=1rcipi∣ci∈R[x]}.\langle p_1,\ldots,p_r\rangle=\left\{\sum_{i=1}^{r}c_i p_i\mid c_i\in\mathbb{R}[x]\right\}.

For any set VV in Rn\mathbb{R}^n, I(V)\mathcal{I}(V) denotes the vanishing ideal of R[x]\mathbb{R}[x] consisting of all polynomials that vanish on VV, i.e., I(V)={p∈R[x]∣p(x)=0, ∀x∈V}\mathcal{I}(V)=\{p\in\mathbb{R}[x]\mid p(x)=0,\ \forall x\in V\}. Conversely, for any set of polynomials PP in R[x]\mathbb{R}[x], Z(P)\mathcal{Z}(P) denotes the common zeros of PP, i.e., Z(P)={x∈Rn∣p(x)=0, ∀p∈P}\mathcal{Z}(P)=\{x\in\mathbb{R}^n\mid p(x)=0,\ \forall p\in P\}.

Polynomial behaviors

In this paper, we consider simulation and control of discrete-time nonlinear systems. To do this, we will first define the class of systems we consider in a precise manner. We approach this problem geometrically, which will allow us to capture important effects by encompassing implicit models, naturally generalize the linear case, and will enable a range of computationally feasible simulation and control methods.

On implicit models

Physical systems are typically described by continuous-time dynamics, for example

x˙=f(x),(1)\dot{x}=f(x), \tag*{(1)}

for autonomous systems, where xx is the state and ff is a nonlinear function. The function ff often takes a simple, e.g., polynomial or trigonometric form in practical applications.

Data-driven methods, on the other hand, work with discrete data sampled every TsT_s seconds, i.e., xk:=x(kTs)x_k:=x(kT_s). Data-driven simulation aims at identifying the discrete relations xk+1=STs(xk)x_{k+1}=S_{T_s}(x_k), where STsS_{T_s} is the solution map of (1) over one sampling period. However, STsS_{T_s} rarely has a closed form and is typically more complex than ff itself. Indeed, approximating STsS_{T_s} is non-trivial even when ff is known, which is the subject of numerical integration. Data-driven simulation and numerical integration (with a fixed step size) thus address the same question in a data-based and a model-based setting, respectively, and we draw on the latter to guide the design of the former.

Numerical integration methods are typically categorized as explicit or implicit, taking the form

xk+1=FTs(xk)orGTs(xk,xk+1)=0,x_{k+1}=F_{T_s}(x_k) \qquad\text{or} \qquad G_{T_s}(x_k,x_{k+1})=0,

respectively. Although implicit methods are computationally more expensive per step because finding xk+1x_{k+1} requires solving a nonlinear equation, their superior stability often permits much larger step sizes. This is crucial for stiff systems, where explicit methods require prohibitively small steps for stability, while implicit methods allow step sizes based purely on accuracy ([18]). The reason for this is that the implicit model class is strictly larger than the explicit (one can always pick GTs(xk,xk+1)=xk+1−FTs(xk)G_{T_s}(x_k,x_{k+1})=x_{k+1}-F_{T_s}(x_k)). For example, for polynomial ff, implicit integration schemes like the implicit Euler or trapezoidal rules induce the richer, algebraic explicit map required for stability, whereas explicit methods merely produce a polynomial map FTsF_{T_s}.

Motivated by these observations, instead of explicit methods based on, e.g., nonlinear ARX models, we adopt the behavioral point of view that treats all variables symmetrically and defines a system by implicit relations between these variables. The numerical case study in § 7.1 validates this choice, demonstrating that an implicit data-driven behavioral method successfully simulates a stiff system where an explicit model trained on the same data fails.

Generalizing linear systems

In contrast to starting from a nonlinear set of defining difference equations, the behavioral approach views systems as subsets of the space of all trajectories. More specifically, the vast majority of methods consider only linear systems, which correspond to subspaces, or equivalently, to the zero set of linear equations. From the perspective of real algebraic geometry, the natural generalization of this is to consider systems described by zero sets of higher-order equations: real algebraic varieties.

In mathematical terms, we consider signals w:Z→Rqw:\mathbb{Z}\to\mathbb{R}^{q}, which are qq-variate signals collecting all variables of interest such as inputs, outputs, or states of the system. The behavioral setting imposes no a priori input/output partition and treats all variables symmetrically, which makes it well suited for the modeling of implicit relations in the data, see § 2.1. In the behavioral setting a discrete-time system is identified with a set of possible trajectories B⊆{w∣w:Z→Rq}\mathcal{B}\subseteq\{w\mid w:\mathbb{Z}\to\mathbb{R}^{q}\}. In this paper, we consider complete systems with lag ℓ\ell (cf. [43]), which are also time-invariant, that is, for any valid trajectory of B\mathcal{B}, any time-shift of the trajectory is also in B\mathcal{B}.

To enable data-based methods, we need to make further assumptions on the set of allowed systems, or model class. For this, we first define the restricted behavior as

BL:={w∈RqL∣w=[w1⊤w2⊤⋯wL⊤]⊤ for some w∈B}.(2)\mathcal{B}_{L}:=\left\{w\in\mathbb{R}^{qL}\mid w=\begin{bmatrix}w_{1}^{\top}&w_{2}^{\top}&\cdots&w_{L}^{\top}\end{bmatrix}^{\top}\text{ for some }w\in\mathcal{B}\right\}. \tag*{(2)}

For time-invariant systems, this is the set of all length-LL windows of trajectories w∈Bw\in\mathcal{B}. As our main goal is to propose computable data-driven simulation and control methods, we now define polynomial systems through their restricted behaviors (cf. [43 Proposition II.6]). Instead of assuming that the system is linear, i.e., that B\mathcal{B} is a subspace, we will consider systems defined by the zero set of a set of polynomials, i.e., an algebraic variety, as formalized below.

Assumption 1 For any L≥ℓ+1L\geq\ell+1, there exists a number np∈Nn_p\in\mathbb{N} and a set of multivariate finite degree polynomials pi∈R[w]p_i\in\mathbb{R}[w], such that BL\mathcal{B}_L is the zero set

BL={w∈RqL∣pi(w)=0,∀i∈{1,…,np}}.(3)\mathcal{B}_{L}=\left\{w\in\mathbb{R}^{qL}\mid p_i(w)=0,\quad\forall i\in\{1,\ldots,n_p\}\right\}. \tag*{(3)}

We then call B\mathcal{B} a polynomial behavior. Note that, while an arbitrary set of polynomials can define a behavior B\mathcal{B} (see Appendix § 9), not every such set constitutes a valid restricted behavior as in Assumption 1. Characterizing whether a given set yields a well-posed system, such as having feasible initial conditions, unique solutions, causality, or even admitting real trajectories, is beyond the scope of the present paper.

The behavioral description (3) includes linear time-invariant systems, bilinear systems, as well as nonlinear input/output models with polynomial nonlinearities. Moreover, since ww may also contain the state of the system, any state-space system with polynomial dynamics and output equation falls into the class of systems defined in this way. In technical terms, the generating set (3) is a real algebraic variety. Next, we focus on deriving a direct data representation for BL\mathcal{B}_L from Assumption 1 and applying it to data-driven simulation and control. To keep the presentation self-contained, we introduce only the algebraic geometry concepts necessary for this task, referring the reader to Appendix § 10 for additional details.

Direct data representation

Inspired by behavioral systems theory for LTI systems, we now investigate how the restricted polynomial behavior in (3) can be represented directly by data. We choose a dictionary ϕ(w)=[ϕ1(w) ⋯ ϕnθ(w)]\phi(w)=\left[\phi_1(w)\ \cdots\ \phi_{n_\theta}(w)\right] consisting of polynomials ϕi∈R[w]≤ddict\phi_i\in\mathbb{R}[w]_{\leq d_{\mathrm{dict}}} up to degree ddictd_{\mathrm{dict}}. Assume we measure DD data trajectories wi∈BLw^i\in\mathcal{B}_L, i∈{1,…,D}i\in\{1,\ldots,D\} (either windows of a single trajectory, or separate trajectories) and construct the lifted data matrix

Φ:=[ϕ(w1)ϕ(w2)⋯ϕ(wD)]∈Rnθ×D.\Phi:=\left[\phi(w^1)\quad\phi(w^2)\quad\cdots\quad\phi(w^D)\right]\in\mathbb{R}^{n_\theta\times D}.

For LTI systems, the dictionary contains only linear functions, so that Φ\Phi is a trajectory (or Hankel) matrix. For sufficiently rich data, its image is the restricted LTI behavior, while a suitable basis of null⁡(Φ)\operatorname{null}(\Phi) provides the coefficients of an ARX model (see Example 3). We now focus on the conditions guaranteeing that Φ\Phi provides a direct data representation of the polynomial behavior BL\mathcal{B}_L.

Problem 1 Given data {wi}i∈{1,…,D}\{\mathbf{w}^{i}\}_{i\in\{1,\ldots,D\}} from BL\mathcal{B}_{L} defined as in Assumption 1, find conditions on the dictionary and the data such that any w∈BL\mathbf{w}\in\mathcal{B}_{L} can be expressed in terms of Φ\Phi.

The solution for this problem is well established for LTI behaviors (Willems et al. [44]). However, a new difficulty arises when polynomial behaviors are considered. Namely, in addition to the defining polynomials pip_i in (3), all polynomials in the form ∑icipi\sum_i c_i p_i with ci∈R[w]c_i\in\mathbb{R}[\mathbf{w}] vanish on the data wi∈BL\mathbf{w}^{i}\in\mathcal{B}_{L}, cf. Example 2 below. Therefore, we provide a direct data representation for BL\mathcal{B}_{L} through the vanishing ideal I(BL)\mathcal{I}(\mathcal{B}_{L}) instead of the defining polynomials pip_i. In case the defining polynomials pip_i are given, the vanishing ideal can be expressed using the real radical, see Appendix § 10 for details.

Conditions on the dictionary

One could naively think that the dictionary degree ddictd_{\mathrm{dict}} has to be at least as large as the maximal degree of pip_i defining BL\mathcal{B}_{L} in (3). However, this is not necessary, as we illustrate on an example below. Instead, we define the minimal degree of a polynomial behavior as

δ(BL)=min⁡{d∈N∣⟨I(BL)≤d⟩=I(BL)}.\delta(\mathcal{B}_{L})=\min\{d\in\mathbb{N}\mid\langle\mathcal{I}(\mathcal{B}_{L})_{\leq d}\rangle=\mathcal{I}(\mathcal{B}_{L})\}.

In words, δ(BL)\delta(\mathcal{B}_{L}) is the minimum degree, such that the polynomials with degree at most δ(BL)\delta(\mathcal{B}_{L}) vanishing on BL\mathcal{B}_{L} generate all polynomials vanishing on BL\mathcal{B}_{L}. We illustrate the concept on an example.

Example 1. Consider the Duffing oscillator also used in Mezić et al. [33]; Rostalski et al. [37]

x˙1=x2,x˙2=−γx2−x1(α+βx12)+u,\dot{x}_{1}=x_{2},\qquad\dot{x}_{2}=-\gamma x_{2}-x_{1}(\alpha+\beta x_{1}^{2})+u,

with state x=[x1x2]T\mathbf{x}=[x_{1}\quad x_{2}]^{\mathsf{T}} and input uu. The implicit trapezoidal rule with step ΔT\Delta T yields the discrete-time dynamics defined by two polynomials p1(w)p_{1}(\mathbf{w}) and p2(w)p_{2}(\mathbf{w}) of degree 1 and 3 in the length-2 trajectory w\mathbf{w} with wt=[utx1,tx2,t]T\mathbf{w}_{t}=[u_{t}\quad x_{1,t}\quad x_{2,t}]^{\mathsf{T}}. Thus, p1p_{1} and p2p_{2} is a representation of B2\mathcal{B}_{2} in (3). In this case, ⟨p1,p2⟩=I(B2)\langle p_{1},p_{2}\rangle=\mathcal{I}(\mathcal{B}_{2}) as the ideal ⟨p1,p2⟩\langle p_{1},p_{2}\rangle is real radical (cf. Appendix § 10). As no polynomials of smaller degree can generate ⟨p1,p2⟩\langle p_{1},p_{2}\rangle, we have that δ(B2)=3\delta(\mathcal{B}_{2})=3. In fact, δ(BL)=3\delta(\mathcal{B}_{L})=3 for any L≥2L\geq2, as each time shift can be described by p1p_{1} and p2p_{2}. Alternative representations of the same B2\mathcal{B}_{2} include {p1,p2,p1⋅p2}\{p_{1},p_{2},p_{1}\cdot p_{2}\} or p12+p22p_{1}^{2}+p_{2}^{2}, which are of larger degrees, 4 and 6, respectively. The real zero set of these polynomials coincide with that of {p1,p2}\{p_{1},p_{2}\}, but their degree is not minimal.

To formulate the condition on the dictionary, note that R[w]≤ddict\mathbb{R}[\mathbf{w}]_{\leq d_{\mathrm{dict}}} is a linear subspace of R[w]\mathbb{R}[\mathbf{w}] and the dimension is the number of monomials: dim⁡(R[w]≤ddict)=(qL+ddictddict)\dim(\mathbb{R}[\mathbf{w}]_{\leq d_{\mathrm{dict}}})=\binom{qL+d_{\mathrm{dict}}}{d_{\mathrm{dict}}}. Further, any ideal I⊆R[w]\mathcal{I}\subseteq\mathbb{R}[\mathbf{w}] and set I≤ddict\mathcal{I}_{\leq d_{\mathrm{dict}}} are linear subspaces, as they are closed under addition and multiplication by constants.

Assumption 2. Suppose ddict≥δ(BL)d_{\mathrm{dict}}\geq\delta(\mathcal{B}_{L}), nΘ=dim⁡(R[w]≤ddict)n_{\Theta}=\dim(\mathbb{R}[\mathbf{w}]_{\leq d_{\mathrm{dict}}}), and let ϕ\phi be a basis for R[w]≤ddict\mathbb{R}[\mathbf{w}]_{\leq d_{\mathrm{dict}}}.

Assumption 2 ensures that the considered dictionary is sufficiently rich to explain BL\mathcal{B}_{L}. Then, the lifted data matrix Φ\Phi describes I({wi}i∈{1,…,D})≤ddict\mathcal{I}(\{\mathbf{w}^{i}\}_{i\in\{1,\ldots,D\}})_{\leq d_{\mathrm{dict}}}, i.e., the set of all polynomials in R[w]≤ddict\mathbb{R}[\mathbf{w}]_{\leq d_{\mathrm{dict}}} that vanish on the data, as formalized below.

Lemma 1. Let Assumptions 1 and 2 hold. Then, {θTϕ∣θ∈null⁡(ΦT)}=I({wi}i∈{1,…,D})≤ddict\{\theta^{\mathsf{T}}\phi\mid\theta\in\operatorname{null}(\Phi^{\mathsf{T}})\}=\mathcal{I}(\{\mathbf{w}^{i}\}_{i\in\{1,\ldots,D\}})_{\leq d_{\mathrm{dict}}}, and this subspace has dimension nΘ−rank⁡(Φ)n_{\Theta}-\operatorname{rank}(\Phi). Furthermore, I({wi}i∈{1,…,D})≤ddict⊇I(BL)≤ddict\mathcal{I}(\{\mathbf{w}^{i}\}_{i\in\{1,\ldots,D\}})_{\leq d_{\mathrm{dict}}}\supseteq\mathcal{I}(\mathcal{B}_{L})_{\leq d_{\mathrm{dict}}}.

Proof. Consider any θ∈null⁡(ΦT)\theta\in\operatorname{null}(\Phi^{\mathsf{T}}). By definition, θTϕ(wi)=0\theta^{\mathsf{T}}\phi(\mathbf{w}^{i})=0 for any i∈{1,…,D}i\in\{1,\ldots,D\}. Since ϕ\phi is a basis for R[w]≤ddict\mathbb{R}[\mathbf{w}]_{\leq d_{\mathrm{dict}}}, θ\theta can describe any polynomial in R[w]≤ddict\mathbb{R}[\mathbf{w}]_{\leq d_{\mathrm{dict}}}, and thus, the claim holds. Moreover, the set inclusion holds since wi∈BL\mathbf{w}^{i}\in\mathcal{B}_{L} implies p(wi)=0p(\mathbf{w}^{i})=0 for all p∈I(BL)p\in\mathcal{I}(\mathcal{B}_{L}).

Next, we investigate the condition on the data formulated in terms of subspace dimensions.

Conditions on the data

As we established, the nullspace of ΦT\Phi^{\mathsf{T}} represents all polynomials in R[w]≤ddict\mathbb{R}[\mathbf{w}]_{\le d_{\mathrm{dict}}} vanishing on the data, and (for noise-free data) these include any polynomial vanishing on BL\mathcal{B}_L. However, there might be more polynomials that vanish on the data than on BL\mathcal{B}_L when the data are not rich enough. Therefore, the condition on the data is that the set I(BL)≤ddict\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}} equals the set represented by null⁡(ΦT)\operatorname{null}(\Phi^{\mathsf{T}}). As the sets are linear subspaces of R[w]≤ddict\mathbb{R}[\mathbf{w}]_{\le d_{\mathrm{dict}}}, this condition can be expressed by their dimensions.

Assumption 3 rank⁡(Φ)=nθ−dim⁡(I(BL)≤ddict)\operatorname{rank}(\Phi)=n_\theta-\dim(\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}}).

Assumption 3 can be written equivalently as dim⁡(null⁡(ΦT))=dim⁡(I(BL)≤ddict)\dim(\operatorname{null}(\Phi^{\mathsf{T}}))=\dim(\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}}), and implies that at least D≥nθ−dim⁡(I(BL)≤ddict)D\ge n_\theta-\dim(\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}}) trajectories are required. The number nθ−dim⁡(I(BL)≤ddict)n_\theta-\dim(\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}}) in Assumption 3 grows with ddictd_{\mathrm{dict}}, so a richer dictionary requires more data. In case the generators p1,…,pnpp_1,\ldots,p_{n_p} in (3) are known, dim⁡(I(BL)≤ddict)\dim(\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}}) can be calculated by first computing the real radical of ⟨p1,…,pnp⟩\langle p_1,\ldots,p_{n_p}\rangle, e.g., with the algorithm of Becker and Neuhaus [3], and then a Gröbner basis of it, see Appendix § 10 and Cox et al. [11]. Under Assumptions 2 and 3, the behavior can be fully described by the lifted data matrix as formalized next.

Lemma 2 Let Assumptions 1, 2, and 3 hold, and consider the lifted data matrix Φ\Phi. Then, a trajectory w∈BL\mathbf{w}\in\mathcal{B}_L if and only if φ(w)∈im⁡(Φ)\varphi(\mathbf{w})\in\operatorname{im}(\Phi).

Proof. By Assumption 3, the inclusion in Lemma 1 is tight: I(BL)≤ddict=I({wi}i∈{1,…,D})≤ddict\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}}=\mathcal{I}(\{\mathbf{w}^i\}_{i\in\{1,\ldots,D\}})_{\le d_{\mathrm{dict}}}. If φ(w)∈im⁡(Φ)\varphi(\mathbf{w})\in\operatorname{im}(\Phi), then p(w)=0p(\mathbf{w})=0 for all p∈I(BL)≤ddictp\in\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}} by Lemma 1. Thus, w∈Z(I(BL)≤ddict)=Z(⟨I(BL)≤ddict⟩)=Z(I(BL))⊆Z(⟨p1,…,pnp⟩)=BL\mathbf{w}\in\mathcal{Z}(\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}})=\mathcal{Z}(\langle\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}}\rangle)=\mathcal{Z}(\mathcal{I}(\mathcal{B}_L))\subseteq\mathcal{Z}(\langle p_1,\ldots,p_{n_p}\rangle)=\mathcal{B}_L, where the second equality holds as ddict≥δ(BL)d_{\mathrm{dict}}\ge\delta(\mathcal{B}_L), and the set inclusion holds as ⟨p1,…,pnp⟩⊆I(BL)\langle p_1,\ldots,p_{n_p}\rangle\subseteq\mathcal{I}(\mathcal{B}_L). Conversely, if w∈BL\mathbf{w}\in\mathcal{B}_L, then p(w)=0p(\mathbf{w})=0 for all p∈I(BL)≤ddictp\in\mathcal{I}(\mathcal{B}_L)_{\le d_{\mathrm{dict}}}, which holds iff φ(w)∈im⁡(Φ)\varphi(\mathbf{w})\in\operatorname{im}(\Phi) by Lemma 1.

Example 2

Example 2 Consider again the Duffing oscillator from Example 1 over a two-step horizon. For Assumption 2 to hold, we need ddict≥δ(B2)=3d_{\mathrm{dict}}\ge\delta(\mathcal{B}_2)=3. Let us pick the minimal degree ddict=3d_{\mathrm{dict}}=3. Then, the polynomials in the set I(B2)≤3=⟨p1,p2⟩≤3\mathcal{I}(\mathcal{B}_2)_{\le3}=\langle p_1,p_2\rangle_{\le3} are p2p_2 and c⋅p1c\cdot p_1 for any c∈R[w]≤2c\in\mathbb{R}[\mathbf{w}]_{\le2} (including the constant). As dim⁡(R[w]≤2)=(6+22)=28\dim(\mathbb{R}[\mathbf{w}]_{\le2})=\binom{6+2}{2}=28, we have that dim⁡(I(B2)≤ddict)=29\dim(\mathcal{I}(\mathcal{B}_2)_{\le d_{\mathrm{dict}}})=29, and Assumption 3 requires Φ\Phi to be of rank nθ−29=55n_\theta-29=55, which is the maximal rank Φ\Phi can have. Importantly, the direct data representation of B2\mathcal{B}_2 contains 29 polynomials, instead of p1p_1 and p2p_2 only.

Example 3

Example 3 (Linear and affine systems) The considered setting with ddict=1d_{\mathrm{dict}}=1 recovers affine systems, whose restricted behavior is an affine subspace (Padoan et al. [35, 36]). LTI behaviors can be recovered by removing the constant term from the dictionary, i.e., φ(w)=w\varphi(\mathbf{w})=\mathbf{w}. Then, Φ\Phi becomes a trajectory or Hankel matrix. A linear form θTw\theta^{\mathsf{T}}\mathbf{w} vanishes on BL\mathcal{B}_L if and only if θ∈BL⊥\theta\in\mathcal{B}_L^\perp, so the linear forms in I(BL)\mathcal{I}(\mathcal{B}_L) form a subspace of dimension dim⁡(BL⊥)=qL−mL−nx\dim(\mathcal{B}_L^\perp)=qL-mL-n_x for L>ℓL>\ell, where mm is the number of inputs and nxn_x is the order. For instance, θ\theta can be chosen as the coefficients of an ARX model. Assumption 3 then reads rank⁡(Φ)=mL+nx\operatorname{rank}(\Phi)=mL+n_x, which is the generalized persistency of excitation condition (Markovsky and Dörfler [28]), and Lemma 2 states that w∈BL\mathbf{w}\in\mathcal{B}_L if and only if w∈im⁡(Φ)\mathbf{w}\in\operatorname{im}(\Phi), which is the Fundamental Lemma (Willems et al. [44]).

Remark 1 (Noisy data) If the data are noisy, Φ\Phi typically has full rank, i.e., im⁡(Φ)=Rnθ\operatorname{im}(\Phi)=\mathbb{R}^{n_\theta}. We then estimate im⁡(Φ)\operatorname{im}(\Phi) by computing the SVD of Φ\Phi and keeping the singular vectors of the hh largest singular values. The orthogonal complement of this truncated space then gives nθ−hn_\theta-h polynomials that approximately vanish on the data. Since the true rank in Assumption 3 is unknown in practice, hh is a tuning criterion. parameter. The truncation is optimal if the noise is additive on Φ\Phi with i.i.d. zero-mean Gaussian entries, as it yields the best rank-hh approximation of Φ\Phi ([14]). However, noise on the trajectories wi\mathbf{w}^{i} enters Φ\Phi through the monomials, making the noise on Φ\Phi structured and not zero-mean. The SVD ignores this structure, which biases the estimate. The adjusted least squares method ([30, 41]) corrects this bias and yields a consistent estimate.

Data-driven simulation and control

Now we use the direct data representation of the restricted behavior in Lemma 2 to formulate the data-driven simulation and control problems without identifying a parametric model first.

Data-driven simulation

Simulation amounts to computing the response of a system to a given input and initial condition ([29]). To formalize this, we partition wt\mathbf{w}_t into inputs ut\mathbf{u}_t and outputs yt\mathbf{y}_t, and w\mathbf{w} into an initial part wini\mathbf{w}_{\mathrm{ini}} of length TiniT_{\mathrm{ini}} specifying the initial condition, and a future part of length Tfut=L−TiniT_{\mathrm{fut}}=L-T_{\mathrm{ini}} with inputs ufut\mathbf{u}_{\mathrm{fut}} and outputs yfut\mathbf{y}_{\mathrm{fut}}. With this partitioning, we structure the entire trajectory as w=[wini ufut yfut]⊤\mathbf{w}=[\mathbf{w}_{\mathrm{ini}}\ \mathbf{u}_{\mathrm{fut}}\ \mathbf{y}_{\mathrm{fut}}]^{\top}, where a suitable row permutation is omitted. Note that yt\mathbf{y}_t may also be the state of the system, in which case a single sample suffices to specify the initial condition, i.e., Tini=1T_{\mathrm{ini}}=1.

Problem 2 (Simulation) Given BL\mathcal{B}_L, wini\mathbf{w}_{\mathrm{ini}}, and ufut\mathbf{u}_{\mathrm{fut}}, find yfut\mathbf{y}_{\mathrm{fut}} such that w∈BL\mathbf{w}\in\mathcal{B}_L.

As BL\mathcal{B}_L is unknown, we apply Lemma 2 to express the condition w∈BL\mathbf{w}\in\mathcal{B}_L by a system of polynomial equations through a basis Θ\Theta of null⁡(Φ⊤)\operatorname{null}(\Phi^{\top}), leading to an equivalent data-driven simulation problem.

Theorem 3 (Data-driven simulation) Let Assumptions 1, 2, and 3 hold. Then, for any wini\mathbf{w}_{\mathrm{ini}} and ufut\mathbf{u}_{\mathrm{fut}}, Problem 2 has the same solutions as the data-driven simulation problem

find yfuts.t.Θϕ(w)=0.(4)\text{find }\mathbf{y}_{\mathrm{fut}}\quad\text{s.t.}\quad\Theta\phi(\mathbf{w})=0. \tag*{(4)}

Proof Θϕ(w)=0\Theta\phi(\mathbf{w})=0 if and only if ϕ(w)∈im⁡(Φ)\phi(\mathbf{w})\in\operatorname{im}(\Phi), which is equivalent to w∈BL\mathbf{w}\in\mathcal{B}_L by Lemma 2.

Since BL\mathcal{B}_L provides an implicit description of the dynamics, Problem 2, and hence (4), may have multiple solutions, similar to implicit numerical integrators in § 2.1. The appropriate solution then has to be selected, e.g., based on physical meaning or consistency with the past trajectory.

Remark 2 (Image formulation with regularization) Another common method to deal with noise in data-driven control is using the image of the lifted data matrix (i.e. writing Φ(wd)g=ϕ(w)\Phi(\mathbf{w}^{d})g=\phi(\mathbf{w}) for some gg in this setting) and adding regularization. With a 2-norm regularizer the problem is still polynomial but now both in w\mathbf{w} and gg. Thus, the number of unknowns increases limiting practical applicability. However, since gg enters quadratically, it can be eliminated in closed form as in Sasfi et al. [39, 40], which leads to a formulation with a cost JJ that is quadratic in ϕ(w)\phi(\mathbf{w}). The problem can be solved through its KKT conditions, but these form a polynomial system of degree 2d−12d-1, which is more expensive to solve than the degree-dd kernel representation.

Data-driven predictive control

In predictive control, a finite-horizon optimal control problem is solved at every time step, and only the first optimal input is applied to the system. With the partitioning of § 6.1, wini\mathbf{w}_{\mathrm{ini}} contains the TiniT_{\mathrm{ini}} most recent measurements, and the future inputs and outputs wfut=[ufutTyfutT]T\mathbf{w}_{\mathrm{fut}} = \begin{bmatrix}\mathbf{u}_{\mathrm{fut}}^{\mathsf{T}} & \mathbf{y}_{\mathrm{fut}}^{\mathsf{T}}\end{bmatrix}^{\mathsf{T}} are the decision variables. Following Rostalski et al. [37], we consider a polynomial cost J(wfut)J(\mathbf{w}_{\mathrm{fut}}) and polynomial inequality constraints c(wfut)≤0c(\mathbf{w}_{\mathrm{fut}}) \le0. A typical choice is the quadratic tracking cost J(wfut)=∥yfut−yref∥Q2+∥ufut∥R2J(\mathbf{w}_{\mathrm{fut}})=\lVert\mathbf{y}_{\mathrm{fut}}-\mathbf{y}_{\mathrm{ref}}\rVert_{Q}^{2}+\lVert\mathbf{u}_{\mathrm{fut}}\rVert_{R}^{2} with reference yref\mathbf{y}_{\mathrm{ref}} and weights Q⪰0Q\succeq0, R≻0R\succ0.

Problem 3 (Predictive control) Given BL\mathcal{B}_L and wini\mathbf{w}_{\mathrm{ini}}, solve

min⁡wfutJ(wfut)s.t.w∈BL,c(wfut)≤0.\min_{\mathbf{w}_{\mathrm{fut}}} J(\mathbf{w}_{\mathrm{fut}}) \quad\text{s.t.}\quad\mathbf{w}\in\mathcal{B}_L,\quad c(\mathbf{w}_{\mathrm{fut}})\le0.

As in § 6.1, applying Lemma 2 to Problem 3 leads to an equivalent data-driven formulation.

Theorem 4 (Data-driven predictive control) Let Assumptions 1, 2, and 3 hold. Then, for any wini\mathbf{w}_{\mathrm{ini}}, JJ, and cc, Problem 3 has the same solutions as the data-driven predictive control problem

min⁡wfutJ(wfut)s.t.Θϕ(w)=0,c(wfut)≤0.(5)\min_{\mathbf{w}_{\mathrm{fut}}} J(\mathbf{w}_{\mathrm{fut}}) \quad\text{s.t.}\quad\Theta\phi(\mathbf{w})=0,\quad c(\mathbf{w}_{\mathrm{fut}})\le0. \tag*{(5)}

Proof Similarly to Theorem 3, the statement is a direct consequence of Lemma 2.

Solving the simulation and control problems

The data-driven simulation problem (4) is a system of polynomial equations in yfut\mathbf{y}_{\mathrm{fut}}, which can be solved approximately with Newton-type methods or, for smaller systems, to global optimality with homotopy continuation [2]. The data-driven control problem (5) is a non-convex polynomial optimization problem, which can be solved locally, e.g., by sequential quadratic programming. Instead, we reformulate (5) as a system of polynomial equations via its KKT conditions, as done by Rostalski et al. [37] in the model-based case. Under suitable constraint qualifications, every local minimizer satisfies

∇J(wfut)+∂ϕ(w)TΘTλ+∂c(wfut)Tν=0,\nabla J(\mathbf{w}_{\mathrm{fut}})+\partial\phi(\mathbf{w})^{\mathsf{T}}\Theta^{\mathsf{T}}\lambda+\partial c(\mathbf{w}_{\mathrm{fut}})^{\mathsf{T}}\nu=0,
Θϕ(w)=0,νici(wfut)=0 ∀i,ν≥0,c(wfut)≤0,(6)\Theta\phi(\mathbf{w})=0,\quad\nu_i c_i(\mathbf{w}_{\mathrm{fut}})=0\ \forall i,\quad\nu\ge0,\quad c(\mathbf{w}_{\mathrm{fut}})\le0, \tag*{(6)}

for some multipliers λ\lambda and ν\nu, where ∇\nabla and ∂\partial denote the gradient and the Jacobian with respect to wfut\mathbf{w}_{\mathrm{fut}}. Among the real solutions of (6), the one with the lowest cost is the global minimizer. Since wini\mathbf{w}_{\mathrm{ini}} and yref\mathbf{y}_{\mathrm{ref}} enter (6) only as parameters, the system can be solved offline for generic parameter values, and only the paths starting at these solutions need to be tracked online [37]. Note, however, that the computational cost of homotopy continuation grows exponentially with the number of unknowns, making it intractable for large horizons. As a remedy, one can apply the data representation recursively, with the minimal L=ℓ+1L=\ell+1, so that the computational cost grows linearly with the horizon.

Remark 3 (Homotopy continuation with redundant constraints) Homotopy continuation computes all isolated solutions of a square polynomial system, i.e., one with as many equations as unknowns [2]. Since the constraints Θϕ(w)=0\Theta\phi(\mathbf{w})=0 are highly redundant, problem (4) is overdetermined, and the solutions of (6) are not isolated, as the multipliers λ\lambda are not unique. Homotopy continuation methods typically resolve this issue by replacing Θ\Theta by AΘA\Theta, where AA is a randomly chosen matrix that makes the system square. Every solution of Θϕ(w)=0\Theta\phi(\mathbf{w})=0 also solves AΘϕ(w)=0A\Theta\phi(\mathbf{w})=0, but one needs to check the feasibility of the computed solutions on the original constraint Θϕ(w)=0\Theta\phi(\mathbf{w})=0. Instead of a random matrix, it is often preferable to minimize ∥Θϕ(w)∥2\lVert\Theta\phi(\mathbf{w})\rVert^2 rather than seeking an exact solution (see Remarks 1 and 2). A practical choice for AA is then the Jacobian ∂(Θϕ(w))T\partial(\Theta\phi(\mathbf{w}))^{\mathsf{T}}, as 2AΘϕ(w)=02A\Theta\phi(\mathbf{w})=0 for the minimizer. The Jacobian can be evaluated at a relevant reference point, such as the most recent measurement of w\mathbf{w} or the control target.

Numerical case studies

We illustrate the proposed methods on two benchmark nonlinear systems below. All computations were carried out using Matlab, on a Lenovo ThinkPad X1 Carbon Gen 10 laptop with 12th Gen Intel i7-1260P ×\times 16 CPU running Ubuntu 22.04.1 LTS. The polynomial systems are solved with Bertini (Bates et al.). As outlined in Remark 3, we square the system using the Jacobian evaluated at the current state or the origin for control, and retain the solution with the smallest residual. The code reproducing the results is available at https://gitlab.ethz.ch/asasfi/polynomial_behaviors.

Van der Pol oscillator

We consider the Van der Pol oscillator

x˙1=x2,x˙2=μ(1−x12)x2−x1.\dot{x}_1=x_2,\qquad\dot{x}_2=\mu\left(1-x_1^2\right)x_2-x_1.

where increasing the parameter μ\mu increases the stiffness of the system. We assume that the system is unknown and that its state wk=[x1(kTs)x2(kTs)]Tw_k=\begin{bmatrix}x_1(kT_s)&x_2(kT_s)\end{bmatrix}^{\mathsf{T}} is measured with sampling time Ts=0.1T_s=0.1. We use windows of length L=2L=2 and nθ=35n_\theta=35 monomials up to degree ddict=3d_{\mathrm{dict}}=3. The data are 6 noise-free trajectories of 1000 samples from initial states (−3,−10),(0,−10),(3,−10),(−3,10),(0,10),(3,10)(-3,-10),(0,-10),(3,-10),(-3,10),(0,10),(3,10) simulated with the stiff solver, ode15s, which is considered to be the true continuous-time trajectory.

Since the true trajectory does not satisfy polynomial relations up to degree 3 exactly, we compute Θ\Theta from the truncated SVD of Φ\Phi (Remark 1). The rank is selected manually as h=32h=32, since there is a visible drop in the singular values of Φ\Phi at that value. The prediction is then computed recursively by solving (4) starting always from the last predicted state. As a benchmark, we identify an explicit model wk+1≈Ξφ(wk)w_{k+1}\approx\Xi\varphi(w_k) (cf. § 2.1), where Ξ\Xi minimizes the sum of ∥wk+1−Ξφ(wk)∥22\lVert w_{k+1}-\Xi\varphi(w_k)\rVert_2^2 over the DD state pairs (wk,wk+1)(w_k,w_{k+1}) used to build Φ\Phi. The behavioral method takes about 20 ms per prediction step, whereas the explicit model takes less than 0.1 ms.

We test the methods on a simulation task for 300 steps starting from the initial condition w0=[31]Tw_0=\begin{bmatrix}3&1\end{bmatrix}^{\mathsf{T}}. As shown in Figure 1, the behavioral method stays close to the true trajectory throughout the prediction for both μ=3\mu=3 and μ=4\mu=4. In contrast, the explicit model only captures the qualitative shape of the limit cycle for μ=3\mu=3 with a significant offset, and does not produce a meaningful prediction for the stiffer system with μ=4\mu=4. Note that the explicit Euler and Runge–Kutta 4 numerical integration methods on the same task become unstable for μ>2\mu>2 and μ>3\mu>3, respectively.

Figure 1. Van der Pol oscillator trajectories for μ=3\mu=3 (left) and μ=4\mu=4 (right), showing predictions over 300 steps starting from the initial state w0w_0 (yellow square). The true trajectory is the black line, the behavioral method is in orange, and the explicit benchmark is in yellow.

Duffing oscillator

Next, we consider the forced Duffing oscillator from Mezić et al. [33]

x˙1=x2,x˙2=−γx2−αx1−βx13+u,\dot{x}_1=x_2,\qquad\dot{x}_2=-\gamma x_2-\alpha x_1-\beta x_1^3+u,

with the parameters γ=0.5\gamma=0.5, α=−1\alpha=-1 and β=1\beta=1, see Figure 3 for the phase portrait. The input and the state are measured with sampling time Ts=0.05T_s=0.05, and we use zero-order hold between the samples for uu. We adopt the code accompanying Mezić et al. [33] and use the data from their example to construct Φ\Phi. As for the Van der Pol oscillator, no polynomial relation holds exactly between the samples, and we compute Θ\Theta from the truncated SVD of Φ\Phi (Remark 1) in all experiments, also for noise-free data.

Figure 3. Duffing oscillator under the proposed controller starting from x0x_0 (yellow square). Unlike the free response (dotted), the closed-loop trajectory (orange) crosses the dashed separatrix from the basin of (−1,0)(-1,0) to stabilize at the saddle at (0,0)(0,0). The background shows the unforced phase portrait and the basins of attraction for the stable equilibria (filled circles).

Simulation

We set L=3L=3 with Tini=1T_{\mathrm{ini}}=1 and degree ddict=4d_{\mathrm{dict}}=4 (matching the polynomial lift of Mezić et al. [33]) giving nθ=715n_\theta=715. The singular values of Φ\Phi decay without a visible gap, but the predictions are not sensitive to the choice of hh for a large interval in the noise-free case, and we pick h=400h=400 from this interval. The 150 steps long test trajectory from Mezić et al. [33] is predicted by solving (4) recursively, starting from the last predicted state (75 solves). The predictions of various Koopman

models from Mezić et al. [33], trained on the same data, are shown in Figure 2 (left). Our prediction overlaps with the true trajectory, whereas the Koopman predictions drift away. This is likely because the Koopman models propagate the lifted state zz linearly, without enforcing that zz remains in the range of the lifting. In contrast, our method solves the identified implicit polynomial relations directly, at the price of increased computational load.

Figure 2. Predictions for the Duffing oscillator starting from initial state x0x_0 (yellow square). Black lines denote the true trajectory, while markers connected by dotted lines represent the behavioral method. Left: Exact data test case from Mezić et al. [33], comparing the behavioral method (orange) against different Koopman models with polynomial lift trained on the same data (linear – blue, bilinear – green, GeKo – red, and KCF – magenta). Right: Predictions under noise using the behavioral method with different truncations h=105h=105 (orange circles), h=29h=29 (yellow squares), and h=78h=78 (blue diamonds).

Next, we add zero-mean Gaussian noise with standard deviation 10−310^{-3} to the state data and to the initial state. We also use a more demanding test case, in which the input u(t)=1.25sin⁡(1.6t)u(t)=1.25\sin(1.6t) drives the state from the stable equilibrium x0=(−1,0)x_0=(-1,0) to the other basin of attraction. Figure 2 (right) shows the predictions of our method for three different truncated ranks. The best truncation is at h=105h=105 providing virtually zero error, whereas h=78h=78 tracks the true trajectory with larger error. The trajectory produced by truncation at h=29h=29 attains a very large error, although the shape resembles that of the true trajectory.

Control

We regulate the oscillator to the saddle at the origin with the data-driven predictive controller (5), using Q=diag⁡(100,1)Q=\operatorname{diag}(100,1) and R=0.1R=0.1. To keep the offline computation tractable, we reduce the degree to ddict=3d_{\mathrm{dict}}=3 (nθ=220n_\theta=220) and choose h=150h=150. The offline solve takes about 10 minutes, and the online computations about 0.07 s per time step. Figure 3 shows the closed loop from x0=(−1.5,0.5)x_0=(-1.5,0.5). Without control, the state converges to (−1,0)(-1,0), but the controller steers the state across the separatrix to the other basin of attraction and then stabilizes at the origin.

Conclusion

We extended direct data representations from LTI to polynomial behaviors by describing their finite-length trajectories as a real algebraic variety. We provided conditions on the data so that the behavior has a direct data representation. This implicit description leads to data-driven simulation and control problems that can be solved globally with homotopy continuation. The proposed methods and their benefits over existing schemes were illustrated on benchmark nonlinear case studies. Future work includes provable robustness to noisy data and solution methods that scale to long horizons.

Appendix A

Here we detail the remark after Assumption 1, namely that an arbitrary set of polynomials defines a behavior, but need not describe its restricted behavior. Let L≥ℓ+1L \ge\ell+1 and consider a generating set

S:={w∈RqL∣pi(w)=0, ∀i∈{1,…,np}}.\mathcal{S} := \left\{ \mathbf{w} \in\mathbb{R}^{qL} \mid p_i(\mathbf{w}) = 0,\ \forall i \in\{1,\ldots,n_p\} \right\}.

where pi∈R[w]p_i \in\mathbb{R}[\mathbf{w}] are multivariate finite degree polynomials. Any such set defines a behavior as

B={w∣[wtT wt+1T ⋯ wt+L−1T]T∈S, ∀t∈Z}.\mathcal{B} = \left\{ \mathbf{w} \mid\left[\mathbf{w}_t^{\mathsf{T}}\ \mathbf{w}_{t+1}^{\mathsf{T}}\ \cdots\ \mathbf{w}_{t+L-1}^{\mathsf{T}}\right]^{\mathsf{T}} \in\mathcal{S},\ \forall t \in\mathbb{Z} \right\}.

Most notably, the generating set S\mathcal{S} is not unique, as different sets S\mathcal{S} can give rise to the same system. By definition, the restricted behavior BL\mathcal{B}_L in (2) of this system satisfies BL⊆S\mathcal{B}_L \subseteq\mathcal{S} for any generating set, but the inclusion can be strict, since S\mathcal{S} may contain windows that do not extend to a trajectory of B\mathcal{B}. If BL\mathcal{B}_L is itself the zero set of polynomials, as required by Assumption 1, then BL\mathcal{B}_L is the smallest generating set of B\mathcal{B} and thus a canonical choice.

Appendix B

Let II be an ideal of R[x]\mathbb{R}[\mathbf{x}]. The real radical of II is defined as

IR={p∈R[x]∣∃m∈N, ∃b1,…,br∈R[x], p2m+b12+⋯+br2∈I},\sqrt[\mathbb{R}]{I} = \left\{p \in\mathbb{R}[\mathbf{x}] \mid\exists m \in\mathbb{N},\ \exists b_1,\ldots,b_r \in\mathbb{R}[\mathbf{x}],\ p^{2m}+b_1^2+\cdots+b_r^2 \in I \right\},

and II is called real radical if IR=I\sqrt[\mathbb{R}]{I}=I. Intuitively, IR\sqrt[\mathbb{R}]{I} collects every polynomial that is implied by the equations in II on real points. For example, consider a system described by two polynomials p1p_1 and p2p_2. The three ideals I1=⟨p1,p2⟩I_1=\langle p_1,p_2\rangle, I2=⟨p12,p2⟩I_2=\langle p_1^2,p_2\rangle, and I3=⟨p12+p22⟩I_3=\langle p_1^2+p_2^2\rangle are different, as p1∈I1p_1 \in I_1, but p1∉I2,I3p_1 \notin I_2,I_3, and p2∈I1,I2p_2 \in I_1,I_2, but p2∉I3p_2 \notin I_3. On the other hand, they all have the same real radical, i.e., I1R=I2R=I3R=I1\sqrt[\mathbb{R}]{I_1}=\sqrt[\mathbb{R}]{I_2}=\sqrt[\mathbb{R}]{I_3}=I_1. Thus, I1I_1 is a real radical ideal, but I2I_2 and I3I_3 are not.

The real Nullstellensatz (see, e.g., Marshall [31]) relates the real radical to the vanishing ideal.

Lemma 5 (Real Nullstellensatz) Let II be an ideal of R[x]\mathbb{R}[\mathbf{x}]. Then, I(Z(I))=IR\mathcal{I}(\mathcal{Z}(I))=\sqrt[\mathbb{R}]{I}.

Applied to I=⟨p1,…,pnp⟩I=\langle p_1,\ldots,p_{n_p}\rangle from (3), the lemma gives

I(BL)=⟨p1,…,pnp⟩R.\mathcal{I}(\mathcal{B}_L)=\sqrt[\mathbb{R}]{\langle p_1,\ldots,p_{n_p}\rangle}.

This result allows us to obtain the vanishing ideal from any set of polynomials representing BL\mathcal{B}_L in (3) by computing the real radical.

References

  1. [2]Daniel J. Bates, Jonathan D. Hauenstein, Andrew J. Sommese, and Charles W. Wampler. Numerically Solving Polynomial Systems with Bertini, volume 25 of Software, Environments, and Tools. SIAM, 2013. doi: 10.1137/1.9781611972702.DOI
  2. [3]Eberhard Becker and Rolf Neuhaus. Computation of real radicals of polynomial ideals. In Frédéric Eyssette and André Galligo, editors, Computational Algebraic Geometry, Progress in Mathematics, pages 1–20. Birkhäuser, 1993.DOI
  3. [5]Julian Berberich and Frank Allgöwer. An overview of systems-theoretic guarantees in data-driven model predictive control. Annual Review of Control, Robotics, and Autonomous Systems, 8, 2025.DOI
  4. [6]Julian Berberich, Johannes Köhler, Matthias A. Müller, and Frank Allgöwer. Data-driven model predictive control with stability and robustness guarantees. IEEE Trans. Automatic Control, 66 (4):1702–1717, 2020.DOI
  5. [7]Julian Berberich, Johannes Köhler, Matthias A. Müller, and Frank Allgöwer. Linear tracking MPC for nonlinear systems—Part II: The data-driven case. IEEE Transactions on Automatic Control, 67(9):4406–4421, 2022.DOI
  6. [8]Fabio Bonassi, Jing Xie, Marcello Farina, and Riccardo Scattolini. An offset-free nonlinear MPC scheme for systems learned by neural NARX models. In Proc. 61st IEEE Conference on Decision and Control, pages 2123–2128, 2022. doi: 10.1109/CDC51059.2022.9992362.DOI
  7. [9]Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113(15):3932–3937, 2016. doi: 10.1073/pnas.1517384113.DOI
  8. [10]Jeremy Coulson, John Lygeros, and Florian Dörfler. Data-enabled predictive control: In the shallows of the DeePC. In Proc. 18th European Control Conference, pages 307–312. IEEE, 2019.arxiv.org/abs/1811.05890
  9. [11]David A. Cox, John Little, and Donal O’Shea. Ideals, Varieties, and Algorithms: An Introduction to Computational Algebraic Geometry and Commutative Algebra. Undergraduate Texts in Mathematics. Springer, 4th edition, 2015.
  10. [12]Claudio De Persis, Monica Rotulo, and Pietro Tesi. Learning controllers from data via approximate nonlinearity cancellation. IEEE Transactions on Automatic Control, 2023.
  11. [13]Florian Dörfler, Jeremy Coulson, and Ivan Markovsky. Bridging direct and indirect data-driven control formulations via regularizations and relaxations. IEEE Trans. Automatic Control, 68(2):883–897, 2022.
  12. [14]Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211–218, 1936.DOI
  13. [15]Antonio Fazzi and Alessandro Chiuso. Simulation of nonlinear systems trajectories: Between models and behaviors. In 2024 10th International Conference on Control, Decision and Information Technologies (CoDIT), pages 2049–2054. IEEE, 2024.arxiv.org/abs/2304.02930
  14. [16]Daniel Floryan and Michael D. Graham. Data-driven discovery of intrinsic dynamics. Nature Machine Intelligence, 4(12):1113–1120, 2022. doi: 10.1038/s42256-022-00575-4.DOI
  15. [17]Meichen Guo, Claudio De Persis, and Pietro Tesi. Data-driven stabilization of nonlinear polynomial systems with noisy data. IEEE Transactions on Automatic Control, 67(8):4210–4217, 2022. doi: 10.1109/TAC.2021.3115436.DOI
  16. [18]Ernst Haier and Gerhard Wanner. Solving Ordinary Differential Equations II: Stiff and Differential-Algebraic Problems. Springer, 2 edition, 1996.
  17. [19]Huayuan Huang, M. Kanat Camilbel, Raffaella Carloni, and Henk J. van Waarde. Data-driven stabilization of polynomial systems using density functions. arXiv preprint arXiv:2503.07092, 2025.arxiv.org/abs/2503.07092
  18. [20]Linbin Huang, John Lygeros, and Florian Dörfler. Robust and kernelized data-enabled predictive control for nonlinear systems. IEEE Transactions on Control Systems Technology, 32(2):611–624, 2023.arxiv.org/abs/2206.01866
  19. [21]Kadierdan Kaheman, J. Nathan Kutz, and Steven L. Brunton. SINDy-PI: a robust algorithm for parallel implicit sparse identification of nonlinear dynamics. Proceedings of the Royal Society A, 476(2242):20200279, 2020. doi: 10.1098/rspa.2020.0279.DOI
  20. [22]Andrey Kharitenko, Zebang Shen, Riccardo De Santi, Niao He, and Florian Dörfler. Landing with the score: Riemannian optimization through denoising. In International Conference on Learning Representations (ICLR), 2026.arxiv.org/abs/2509.23357
  21. [23]Milan Korda and Igor Mezić. Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control. Automatica, 93:149–160, 2018. doi: 10.1016/j.automatica.2018.03.046.DOI
  22. [24]Mircea Lazar. Basis-functions nonlinear data-enabled predictive control: Consistent and computationally efficient formulations. In Proc. European Control Conference, pages 888–893, 2024. doi: 10.23919/ECC64448.2024.10591192.DOI
  23. [25]Niall M. Mangan, Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Inferring biological networks by sparse identification of nonlinear dynamics. IEEE Transactions on Molecular, Biological and Multi-Scale Communications, 2(1):52–63, 2016. doi: 10.1109/TMBMC.2016.2633265.DOI
  24. [26]Ivan Markovsky. Data-driven simulation of generalized bilinear systems via linear time-invariant embedding. IEEE Transactions on Automatic Control, 68(2):1101–1106, 2022.DOI
  25. [27]Ivan Markovsky and Florian Dörfler. Behavioral systems theory in data-driven analysis, signal processing, and control. Annual Reviews in Control, 52:42–64, 2021.DOI
  26. [28]Ivan Markovsky and Florian Dörfler. Identifiability in the behavioral setting. IEEE Transactions on Automatic Control, 68(3):1667–1677, 2023. doi: 10.1109/TAC.2022.3209954.DOI
  27. [29]Ivan Markovsky and Paolo Rapisarda. Data-driven simulation and control. International Journal of Control, 81:1946–1959, 2008.DOI
  28. [30]Ivan Markovsky, Alexander Kukush, and Sabine Van Huffel. Consistent least squares fitting of ellipsoids. Numerische Mathematik, 98(1):177–194, 2004.DOI
  29. [31]Murray Marshall. Positive polynomials and sums of squares. Number 146. American Mathematical Soc., 2008.
  30. [32]Tim Martin, Thomas B. Schön, and Frank Allgöwer. Guarantees for data-driven control of nonlinear systems using semidefinite programming: A survey. Annual Reviews in Control, 56:100911, 2023.arxiv.org/abs/2306.16042
  31. [33]Igor Mezić, Jorge Cortés, Karl Worthmann, Mircea Lazar, and Armin Lederer. Koopman operator theory: fundamentals, control, and applications. arXiv preprint arXiv:2607.01819, 2026. URL https://arxiv.org/abs/2607.01819.
  32. [34]Vikas K. Mishra, Ivan Markovsky, Antonio Fazzi, and Philippe Dreesen. Data-driven simulation for narx systems. In 2021 29th European Signal Processing Conference (EUSIPCO), pages 1055–1059, 2021. doi: 10.23919/EUSIPCO54536.2021.9616226.DOI
  33. [35]Alberto Padoan, Florian Dörfler, and John Lygeros. Data-driven representations of conical, convex, and affine behaviors. In Proc. 62nd Conference on Decision and Control, pages 596–601. IEEE, 2023.
  34. [36]Alberto Padoan, Jaap Eising, and Ivan Markovsky. From time series to affine systems. arXiv preprint arXiv:2510.22089, 2025.arxiv.org/abs/2510.22089
  35. [37]Philipp Rostalski, Ioannis A. Fotiou, Daniel J. Bates, A. Giovanni Beccuti, and Manfred Morari. Numerical algebraic geometry for optimal control applications. SIAM Journal on Optimization, 21(2):417–437, 2011.DOI
  36. [38]Juan G. Rueda-Escobedo and Johannes Schiffer. Data-driven internal model control of second-order discrete Volterra systems. In Proc. 59th IEEE Conference on Decision and Control, pages 4572–4579, 2020. doi: 10.1109/CDC42340.2020.9304122.
  37. [39]András Sasfi, Ivan Markovsky, Alberto Padoan, and Florian Dörfler. Gaussian behaviors: representations and data-driven control. In Proc. 64th Conference on Decision and Control, pages 1993–1998. IEEE, 2025.arxiv.org/abs/2504.15838
  38. [40]András Sasfi, Jaap Eising, and Florian Dörfler. Soft projections for robust data-driven control. arXiv preprint arXiv:2604.00900, 2026.arxiv.org/abs/2604.00900
  39. [41]Konstantin Usevich and Ivan Markovsky. Adjusted least squares fitting of algebraic hypersurfaces. Linear Algebra and its Applications, 502:243–274, 2016.DOI
  40. [42]Pieter van Goor, Robert Mahony, Manuel Schaller, and Karl Worthmann. Reprojection methods for Koopman-based modelling and prediction. In Proc. 62nd IEEE Conference on Decision and Control, pages 315–321, 2023. doi: 10.1109/CDC49753.2023.10383796.DOI
  41. [43]Jan C. Willems. Paradigms and puzzles in the theory of dynamical systems. IEEE Transactions on Automatic Control, 36(3):259–294, 1991. doi: 10.1109/9.73561.DOI
  42. [44]Jan C. Willems, Paolo Rapisarda, Ivan Markovsky, and Bart L. M. De Moor. A note on persistency of excitation. Systems & Control Letters, 54(4):325–329, 2005.
  43. [45]Matthew O. Williams, Ioannis G. Kevrekidis, and Clarence W. Rowley. A data-driven approximation of the Koopman operator: Extending dynamic mode decomposition. Journal of Nonlinear Science, 25(6):1307–1346, 2015. doi: 10.1007/s00332-015-9258-5.DOI

Paper details

vlnc:W_egmtsjw7
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

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

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents