Gaussian behaviors: representations and data-driven control
András Sasfi, Ivan Markovsky, Alberto Padoan, Florian Dörfler · 2025
Read Gaussian behaviors: representations and data-driven control by András Sasfi, Ivan Markovsky, Alberto Padoan, Florian Dörfler (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-15838.txt.
Introduction
András Sasfi and co-authors propose a data driven model for uncertain systems in 2025. Their frame, called Gaussian behaviors, treats a short run of a system as a random curve with a known spread. The model keeps uncertainty visible while staying simple enough for control design. It links neatly to classic state space models with random noise.
The model can be learned straight from observed runs by a plain sample spread. Future outputs can then be forecast given inputs and past outputs. The authors show a known control method matches this frame under one plain reading. The frame may ease control of systems that are only seen through data.
Catalog record
| CREATOR | András Sasfi, Ivan Markovsky, Alberto Padoan, Florian Dörfler |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 50 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-15838
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Gaussian behaviors: representations and data-driven control András Sasfi, Ivan Markovsky, Alberto Padoan, Florian Dörfler Abstract We propose a modeling framework for stochastic systems based on Gaussian processes. Finite-length trajectories of the system are modeled as random vectors from a Gaussian distribution, which we call a Gaussian behavior. The proposed model naturally quantifies the uncertainty in the trajectories, yet it is simple enough to allow for tractable formulations. We relate the proposed model to existing descriptions of dynamical systems including deterministic and stochastic behaviors, and linear time-invariant (LTI) state-space models with Gaussian process and measurement noise. Gaussian behaviors can be estimated directly from observed data as the empirical sample covariance under the assumption that the measured trajectories are from independent experiments. The distribution of future outputs conditioned on inputs and past outputs provides a predictive model that can be incorporated in predictive control frameworks. We show that subspace predictive control (SPC) is a certainty-equivalence control formulation with the estimated Gaussian behavior. Furthermore, the regularized data-enabled predictive control (DeePC) method is shown to be a distributionally optimistic formulation that optimistically accounts for uncertainty in the Gaussian behavior. To mitigate the excessive optimism of DeePC , we propose a novel distributionally robust control formulation, and provide a convex reformulation allowing for efficient implementation. I Introduction Recent data-driven control methods based on the behavioral approach [ 1 , 2 , 3 , 4 , 5 , 6 , 7 ] have gained significant attention. These formulations rely on behavioral systems theory that treats systems as sets of trajectories, and allows to represent linear time-invariant (LTI) systems directly with data [ 8 , 9 ] . The methods exploit this data representation and typically add regularization to the problem [ 10 , 11 ] for a posteriori robustification. The resulting formulations can achieve comparable or even superior performance compared to classical methods consisting of an identification and a control step, especially in challenging scenarios involving nonlinear effects and large amount of uncertainty [ 10 , 12 ] . Existing works explain this observation by the robustifing effect of regularization, yet they rely on the inherently deterministic behavioral description [ 13 , 2 , 3 ] . On the other hand, only a few works propose frameworks that start directly from a stochastic system model. In fact, the survey [ 12 ] points out that a major open problem in this field is precisely the extension of data representations to the stochastic domain. Stochastic extensions to behavioral systems theory have been defined bottom up in the literature [ 14 , 15 ] . However, the existing works provide general and abstract (and thus also blunt) perspectives that hinder the practical applicability of the frameworks. Stochastic behaviors have also been modeled in the literature using polynomial chaos expansions [ 16 , 17 ] . However, the complexity of the resulting methods grow with the order of the expansion, limiting scalability. Recently, a stochastic interpretation of data-driven control formulations was proposed in [ 18 ] , highlighting that regularization accounts for the uncertainty in a linear model estimated from data. In this work, we take a different approach, and propose a new bottom-up stochastic modeling framework that admits a data representation and leads to tractable control formulations which can be solved efficiently. The contributions of this work are the following. First, in Section III , we propose a stochastic modeling framework based on Gaussian processes that enables us to model trajectories as normally distributed random vectors. We define the distribution of trajectories to be a Gaussian behavior , characterized by its mean and covariance, which inherently captures uncertainty. Gaussian behaviors give rise to predictive models given by the conditional distribution, which can be calculated readily in the proposed framework. Furthermore, in a data-driven context, Gaussian behaviors can be estimated directly as the empirical sample covariance, which is also used as a system representation in [ 19 ] . We show that Gaussian behaviors are a special class of stochastic behaviors. Furthermore, the deterministic behavioral description of LTI systems as subspaces is captured by the proposed definition with zero mean and singular covariance matrix. Finally, stochastic LTI systems given by state-space equations with normally distributed initial state, process and measurement noise are also special cases of the proposed definition by imposing additional structure on the covariance. Additionally, Gaussian behaviors are closely connected to covariance steering problems [ 20 , 21 , 22 , 23 ] that usually consider linear stochastic systems with a normally distributed initial state and aim at finding a feedback controller that achieves a desired distribution for the closed-loop output. Second, we interpret existing data-driven control formulations in the proposed framework and derive a new robust method in Section IV . In particular, subspace predictive control (SPC) is a certainty-equivalence formulation in this framework that uses the sample covariance to predict the future outputs. To account for inaccuracies in the predictive model, we optimize the predictive distribution within an uncertainty set centered around the estimate. This results in a distributionally optimistic formulation, which is shown to be equivalent to the regularized data-enabled predictive control (DeePC) approach. In this method, the distribution that minimizes the control cost is selected from the uncertainty set. In contrast, we propose a distributionally robust formulation that optimizes the control input under the worst-case mismatch between the true and estimated distributions. The resulting min-max problem can be reformulated as a convex problem, enabling efficient computation. II Preliminaries II-A Notation The image and kernel of a matrix A are denoted by \mathrm{im}(A) and \mathrm{ker}(A) , respectively. Furthermore, A^{\dagger} denotes the Moore–Penrose inverse of A . The set of symmetric positive (semi) definite matrices is denoted by \mathbb{S}_{>0}~(\mathbb{S}_{\geq 0}) . The identity matrix of dimension n\times n is I_{n} . Furthermore, the (semi-) norm of x weighted by a matrix M\in\mathbb{S}_{>0}~(\mathbb{S}_{\geq 0}) is denoted by \|x\|_{M}=x^{\top}Mx . II-B Stochastic processes Classically, a stochastic process (or, briefly, process) is defined as a family of random variables \{w_{t}\}_{t\in\mathbb{T}} , where w_{t} is a random variable for each time t in the time range \mathbb{T} [ 24 ] . Given their finite-dimensional distributions, the entire process is uniquely characterized via Kolmogorov’s extension theorem [ 24 ] . A stochastic process \{w_{t}\}_{t\in\mathbb{T}} is wide-sense stationary [ 25 , Sec. 10.1] if its mean is constant, i.e., \mathbb{E}[w_{t}]=\mu,\quad\text{for all }t\in\mathbb{T}, for some vector \mu , and the covariance between w_{t_{i}} and w_{t_{j}} depends only on the time shift t_{i}-t_{j} , i.e., \operatorname{Cov}(w_{t_{i}},w_{t_{j}})=K(t_{i}-t_{j}),\quad\text{for all }t_{i},t_{j}\in\mathbb{T}, for some matrix-valued covariance function K(\cdot) . A stochastic process is said to be a Gaussian process if, for any finite set of time instants T=\{t_{1},t_{2},\dots,t_{k}\}\subset\mathbb{T}, the joint probability distribution of the corresponding random variables \{w_{t}\}_{t\in T} is Gaussian, that is, \{w_{t}\}_{t\in T}\sim\mathcal{N}(\mu_{T},K_{T}), where \mu_{T}=[\mu(t_{1})~\dots~\mu(t_{k})]^{\top} is the mean vector, and K_{T} is the k\times k covariance matrix with entries K_{T}(i,j)=\mathbb{E}[(w_{t_{i}}-\mu({t_{i}}))(w_{t_{j}}-\mu({t_{j}}))^{\top}]~\text{for}~t_{i},t_{j}\in T. II-C Behavioral systems theory for discrete-time LTI systems Behavioral systems theory abstractly defines systems as a set of trajectories, called the behavior , regardless of its representation in terms of particular equations. For LTI systems, the behavior is a shift-invariant subspace of the set of all possible trajectories [ 11 ] . In this deterministic context, with a slight abuse of notation, we denote