betaFeedback

WorksEngineering

Bayesian Learning of Uncertain Linear System Behaviours

Abstract

We present a Bayesian approach to learning uncertain behavioural representations of linear dynamical systems from noisy input-output data. In behavioural systems theory, the finite-horizon behaviour of a linear system is represented by a subspace. We quantify uncertainty in this subspace through probability distributions on the Grassmann manifold. This leads to a geometric formulation of Bayesian inference over the Grassmannian. We derive the posterior distribution under Gaussian measurement noise and discuss how uncertainty in the inferred behaviour can be propagated to trajectory predictions, with a view towards data-driven control.

Introduction

Behavioural systems theory [1] characterizes a linear time-invariant (LTI) system through its subspace of admissible trajectories. A representation of this subspace can be constructed from sufficiently informative input–output data [2], without first identifying a parametric model. Over finite horizons, the restricted behaviour is represented by the column space of a suitably constructed data matrix. This perspective supports methods for system identification [3], [4], [5] and data-enabled control [6], [7], [8], with demonstrated practical effectiveness [8], [9].

In practice, noise corrupts both the offline training data and the online measurements used for feedback and prediction, and can severely degrade controller performance if its effects are not accounted for [10]. Quantifying and compensating for these effects is therefore a fundamental challenge in both system identification and data-driven control. This has motivated formulations that extend beyond the deterministic, noise-free setting [8], including regularized optimization [6], [10], robust formulations [11], and distributionally robust formulations [7], which can improve robustness but often introduce hyperparameters that require tuning. Other strategies preprocess data before constructing the behavioural representation, for example through low-rank approximation [12], including truncated singular value decomposition (SVD) [13]. Maximum a posteriori estimation offers another approach to trajectory inference under uncertainty [14].

We propose an uncertainty quantification framework for restricted behaviours of linear dynamical systems. The central idea is to represent uncertainty about the restricted behaviour by a probability distribution on the Grassmann manifold and learn this distribution from data using Bayesian inference. Motivated by [15], [16], this formulation respects the invariance of the subspace to the choice of basis and quantifies uncertainty arising from noisy observations. It also allows uncertainty to be propagated to finite-horizon trajectory predictions using tools from geometric probability and computation on manifolds [17], [18], [19], [20].

The statistical analysis of estimation error and model uncertainty has long been central to system identification [21]. Uncertainty quantification for identified LTI models has been developed for state-space [22], [23], [24], frequency-response [25], [26], and ARMA representations [27]. For predictors constructed from behavioural representations, uncertainty in the predicted outputs is quantified in [28], [29], and a deterministic analysis of prediction error is given in [30]. Related work has also investigated uncertainty quantification and geometric distances for restricted behaviours [15], [16].

Closely related research provides complementary perspectives on probabilistic descriptions of dynamical systems. Faulwasser et al. [31] use polynomial chaos expansions (PCE) to obtain a deterministic representation of stochastic dynamics in a higher-dimensional coefficient space. Sasfi et al. [32] model finite-length trajectories as Gaussian random vectors, characterizing trajectory variability through their covariance structure. Finally, Preciado [33] develops a measure-theoretic formulation based on probability measures supported on admissible trajectories, leading to convex formulations of optimal control problems. Our focus is Bayesian inference over restricted behaviours themselves, with uncertainty in the inferred behaviour propagated to trajectory predictions.

Preliminaries

Let S≥0q\mathbb{S}_{\geq0}^{q} and S>0q\mathbb{S}_{>0}^{q} be the sets of q×qq \times q symmetric positive semidefinite and positive definite matrices, respectively. For a matrix AA, we write ATA^{\mathsf{T}} for its transpose, A†A^{\dagger} for its Moore–Penrose pseudoinverse, ∣A∣|A| for its determinant, and σi(A)\sigma_i(A) for its iith largest singular value. The k×kk \times k identity matrix is IkI_k. We use the abbreviations

etr⁡(A):=exp⁡(tr⁡(A)),∥x∥M2:=xTMx,\operatorname{etr}(A) := \exp(\operatorname{tr}(A)), \qquad\lVert x \rVert_M^2 := x^{\mathsf{T}} M x,

where MM is symmetric positive definite. The Gaussian distribution with mean μ∈Rn\mu\in\mathbb{R}^n and covariance Σ∈S>0n\Sigma\in\mathbb{S}_{>0}^{n} is written N(μ,Σ)\mathcal{N}(\mu,\Sigma), with density

N(y;μ,Σ):=(2π)−n/2∣Σ∣−1/2exp⁡(−12∥y−μ∥Σ−12).\mathcal{N}(y;\mu,\Sigma) := (2\pi)^{-n/2}|\Sigma|^{-1/2}\exp\left(-\frac{1}{2}\lVert y-\mu\rVert_{\Sigma^{-1}}^2\right).

We will make use of two types of smooth manifolds in this paper: the compact Stiefel manifold and the Grassmann manifold [34] (see also [35], [36]). Subspaces are denoted by bold uppercase letters and their orthonormal basis matrices by the corresponding uppercase letters. For integers $n > 0 and k>0k > 0, with k≤nk \le n, the Grassmann manifold Gr⁡(n,k)\operatorname{Gr}(n,k) is the set of all kk-dimensional linear subspaces of Rn\mathbb{R}^n

Gr⁡(n,k)={U⊆Rn:dim⁡(U)=k}.\operatorname{Gr}(n,k)=\{\mathrm{U}\subseteq\mathbb{R}^n:\dim(\mathrm{U})=k\}.

The compact Stiefel manifold St⁡(n,k)\operatorname{St}(n,k) is the set of all matrices in Rn×k\mathbb{R}^{n\times k} with orthonormal columns

St⁡(n,k)={U∈Rn×k:UTU=Ik}.\operatorname{St}(n,k)=\{\mathrm{U}\in\mathbb{R}^{n\times k}:\mathrm{U}^{\mathsf{T}}\mathrm{U}=\mathrm{I}_k\}.

The compact Stiefel manifold is an embedded submanifold of Rn×k\mathbb{R}^{n\times k}, and the Grassmann manifold is its quotient under orthogonal changes of basis Gr⁡(n,k)≅St⁡(n,k)/O(k)\operatorname{Gr}(n,k)\cong\operatorname{St}(n,k)/\mathrm{O}(k) [34, 36], where O(k)\mathrm{O}(k) is the orthogonal group of k×kk\times k matrices. Thus, a subspace U∈Gr⁡(n,k)\mathrm{U}\in\operatorname{Gr}(n,k) is represented by any U∈St⁡(n,k)\mathrm{U}\in\operatorname{St}(n,k) whose columns span U\mathrm{U}. Two representatives U\mathrm{U} and V\mathrm{V} describe the same subspace if and only if V=UQ\mathrm{V}=\mathrm{UQ} for some Q∈O(k)\mathrm{Q}\in\mathrm{O}(k).

We can also identify U∈Gr⁡(n,k)\mathrm{U}\in\operatorname{Gr}(n,k) with its orthogonal projector PU=UUT\mathrm{P}_{\mathrm{U}}=\mathrm{UU}^{\mathsf{T}}, which is independent of the representative U\mathrm{U} [37, 38]. For V=im⁡V\mathrm{V}=\operatorname{im}\mathrm{V} with V∈St⁡(n,k)\mathrm{V}\in\operatorname{St}(n,k), the chordal distance is defined by

dc(U,V):=12∥PU−PV∥F,d_{\mathrm{c}}(\mathrm{U},\mathrm{V}):=\frac{1}{\sqrt{2}}\lVert\mathrm{P}_{\mathrm{U}}-\mathrm{P}_{\mathrm{V}}\rVert_{\mathrm{F}},

where ∥⋅∥F\lVert\cdot\rVert_{\mathrm{F}} is the Frobenius norm. Equivalently,

dc(U,V)2=∑i=1ksin⁡2θi,d_{\mathrm{c}}(\mathrm{U},\mathrm{V})^2=\sum_{i=1}^{k}\sin^2\theta_i,

where θi∈[0,π2]\theta_i\in[0,\frac{\pi}{2}], with θ1≤⋯≤θk\theta_1\le\cdots\le\theta_k, are the principal angles between the subspaces U\mathrm{U} and V\mathrm{V}, and satisfy cos⁡θi=σi(UTV)\cos\theta_i=\sigma_i(\mathrm{U}^{\mathsf{T}}\mathrm{V}) [39, 40].

The compact Grassmann manifold Gr⁡(n,k)\operatorname{Gr}(n,k) carries a uniform probability measure ν\nu [37]. Throughout the paper, densities on Gr⁡(n,k)\operatorname{Gr}(n,k) are taken with respect to ν\nu. For a symmetric matrix A∈Rn×n\mathrm{A}\in\mathbb{R}^{n\times n}, the matrix Bingham distribution on Gr⁡(n,k)\operatorname{Gr}(n,k) has density [37]

Bing⁡n,k(U;A):=etr⁡(APU)1F1(k2;n2;A).(1)\operatorname{Bing}_{n,k}(\mathrm{U};\mathrm{A}):=\frac{\operatorname{etr}(\mathrm{AP}_{\mathrm{U}})}{{}_1F_1\left(\frac{k}{2};\frac{n}{2};\mathrm{A}\right)}. \tag*{(1)}

where 1F1{}_1F_1 is the confluent hypergeometric function of matrix argument [41], which can be evaluated numerically using the algorithms in [42].

Problem Setup

We consider a discrete-time LTI system

xt+1=Axt+But,(2)x_{t+1}=Ax_t+Bu_t, \tag*{(2)}
yt=Cxt+Dut,(3)y_t=Cx_t+Du_t, \tag*{(3)}

where xt∈Rnx_t\in\mathbb{R}^n, ut∈Rmu_t\in\mathbb{R}^m, and yt∈Rpy_t\in\mathbb{R}^p denote the state, input, and output, respectively. The system matrices AA, BB, CC, and DD are unknown, and the state is unobserved.

For integers t≥1t\ge1 and L≥1L\ge1, define the stacked input and output vectors

u[t,L]:=[utT ⋯ ut+L−1T]T,u_{[t,L]}:=\left[u_t^{\mathsf{T}}\ \cdots\ u_{t+L-1}^{\mathsf{T}}\right]^{\mathsf{T}},
y[t,L]:=[ytT ⋯ yt+L−1T]T.y_{[t,L]}:=\left[y_t^{\mathsf{T}}\ \cdots\ y_{t+L-1}^{\mathsf{T}}\right]^{\mathsf{T}}.

The stacked vector

w[t,L]:=[u[t,L]y[t,L]]w_{[t,L]}:=\begin{bmatrix}u_{[t,L]}\\y_{[t,L]}\end{bmatrix}

is an admissible length-LL input–output trajectory of (2) if there exists a state sequence xt,…,xt+Lx_t,\ldots,x_{t+L} satisfying (2) at each time τ=t,…,t+L−1\tau=t,\ldots,t+L-1. The restricted behaviour BL\mathcal{B}_L of (2) consists of all admissible input–output trajectories w[1,L]w_{[1,L]} of (2). The restricted behaviour BL\mathcal{B}_L of (2) consists of all admissible input–output trajectories w[1,L]w_{[1,L]} of (2). The restricted behaviour BL\mathcal{B}_L is a linear subspace of Rd\mathbb{R}^d, where q=m+pq=m+p and d=qLd=qL. We denote its dimension by rr. For an observable realization (2), r=mL+nr=mL+n for L≥ℓL\ge\ell, where ℓ\ell is the observability index of (C,A)(C,A) [2].

The stacked vector

w[t,L]:=[u[t,L]y[t,L]]w_{[t,L]}:=\begin{bmatrix}u_{[t,L]}\\y_{[t,L]}\end{bmatrix}

is an admissible length-LL input–output trajectory of (2) if there exists a state sequence xt,…,xt+Lx_t,\ldots,x_{t+L} satisfying (2) at each time τ=t,…,t+L−1\tau=t,\ldots,t+L-1. The restricted behaviour BL\mathcal{B}_L of (2) consists of all admissible input–output trajectories w[1,L]w_{[1,L]} of (2). The restricted behaviour BL\mathcal{B}_L of (2) consists of all admissible input–output trajectories w[1,L]w_{[1,L]} of (2). The restricted behaviour BL\mathcal{B}_L is a linear subspace of Rd\mathbb{R}^d, where q=m+pq=m+p and d=qLd=qL. We denote its dimension by rr. For an observable realization (2), r=mL+nr=mL+n for L≥ℓL\ge\ell, where ℓ\ell is the observability index of (C,A)(C,A) [2].

Our objective is to infer the restricted behaviour of (2) from noisy observations and provide a principled quantification of uncertainty in the inferred subspace. Inputs and outputs are assumed to be measured with additive Gaussian noise, giving an errors-in-variables model [43]. Specifically, we consider the probabilistic model

wi=wˉi+ϵi,wˉi∈BL,i=1,…,N.(4)w_i=\bar{w}_i+\epsilon_i,\qquad\bar{w}_i\in\mathcal{B}_L,\qquad i=1,\ldots,N. \tag*{(4)}

where ϵi∈Rd\epsilon_i\in\mathbb{R}^d is independent of the input-output trajectory wˉi\bar{w}_i and such that ϵi∼i.i.d.N(0,σϵ2Id)\epsilon_i\overset{\mathrm{i.i.d.}}{\sim}\mathcal{N}(0,\sigma_\epsilon^2\mathrm{I}_d).

Subspace Representation from Data

We first recall how a restricted behaviour can be represented in the absence of noise. Consider a measured trajectory

wd(t)=[ud(t)yd(t)]∈Rq,t=1,…,T.w_d(t)=\begin{bmatrix}u_d(t)\\y_d(t)\end{bmatrix}\in\mathbb{R}^q,\qquad t=1,\ldots,T.

For a positive integer L≤TL\le T, define the depth-LL Hankel matrix HL(wd)∈RqL×(T−L+1)\mathcal{H}_L(w_d)\in\mathbb{R}^{qL\times(T-L+1)}, by

HL(wd)=[wd(1)wd(2)⋯wd(T−L+1)wd(2)wd(3)⋯wd(T−L+2)⋮⋮⋱⋮wd(L)wd(L+1)⋯wd(T)].\mathcal{H}_L(w_d)= \begin{bmatrix} w_d(1)&w_d(2)&\cdots&w_d(T-L+1)\\ w_d(2)&w_d(3)&\cdots&w_d(T-L+2)\\ \vdots&\vdots&\ddots&\vdots\\ w_d(L)&w_d(L+1)&\cdots&w_d(T) \end{bmatrix}.

If (A,B)(A,B) is controllable and udu_d is persistently exciting of order L+nL+n, the fundamental lemma [2, 44] gives

BL=im⁡HL.(5)\mathcal{B}_L=\operatorname{im}\mathcal{H}_L. \tag*{(5)}

Thus, every admissible trajectory w∈BLw\in\mathcal{B}_L can be written as

w=HLgw=\mathcal{H}_Lg

for some g∈RT−L+1g\in\mathbb{R}^{T-L+1}. The construction extends to multiple records by concatenating their Hankel matrices under a collective excitation condition [44]. For length-LL input-output trajectories, the corresponding noise-free data matrix represents BL\mathcal{B}_L whenever it has rank rr.

An orthonormal representation follows from the compact SVD [8],

HL=UrΣrVrT,\mathcal{H}_L=\mathrm{U}_r\Sigma_r\mathrm{V}_r^{\mathsf{T}},

where Ur∈Rd×rU_r \in\mathbb{R}^{d \times r} satisfies UrTUr=IrU_r^{\mathsf{T}}U_r=I_r, and BL=im⁡Ur\mathcal{B}_L=\operatorname{im}U_r. Thus, the restricted behaviour is a point on Gr⁡(d,r)\operatorname{Gr}(d,r) [15], [16], and represented by any of its orthonormal basis matrices. For noisy measurements, the rank of a Hankel matrix may be larger than the dimension of underlying noise-free restricted behaviour. In this case, standard practice is to truncate the singular value decomposition by retaining rr dominant singular directions to obtain an estimate of the subspace [8].

Probabilistic Beliefs over Behaviours

The core idea in this paper is to model finite-horizon behaviours as elements of the Grassmann manifold and to learn the belief

p(BL∣W),(6)p(\mathcal{B}_L \mid W), \tag*{(6)}

where W∈Rd×NW \in\mathbb{R}^{d \times N} is a data matrix of observed trajectories

W=[w1 ⋯ wN]∈Rd×N.(7)W=[w_1\ \cdots\ w_N]\in\mathbb{R}^{d\times N}. \tag*{(7)}

Uncertainty about the restricted behaviour BL\mathcal{B}_L is modelled by a probability distribution3 on Gr⁡(d,r)\operatorname{Gr}(d,r). Thus, our framework replaces a deterministic restricted behaviour with a probability measure over admissible behaviours. A distribution can subsequently be propagated into probabilistic behavioural predictive control and simulation problems, providing a principled mechanism for representing epistemic uncertainty in data-driven predictions.

Inference

Let S∈Gr⁡(d,r)S \in\operatorname{Gr}(d,r) denote the unknown restricted behaviour and let Γ∈St⁡(d,r)\Gamma\in\operatorname{St}(d,r) be any representative, so that S=im⁡ΓS=\operatorname{im}\Gamma and PS=ΓΓTP_S=\Gamma\Gamma^{\mathsf{T}}. We use the observation model

wi=Γzi+εi,i=1,…,N,(8)w_i=\Gamma z_i+\varepsilon_i,\qquad i=1,\ldots,N, \tag*{(8)}

where all zi,εiz_i,\varepsilon_i are independent, zi∼N(0,σz2Ir)z_i\sim\mathcal{N}(0,\sigma_z^2I_r) and εi∼N(0,σε2Id)\varepsilon_i\sim\mathcal{N}(0,\sigma_\varepsilon^2I_d), with σz2>0\sigma_z^2>0 being a hyperparameter (usually estimated using expectation maximization) and σε2>0\sigma_\varepsilon^2>0 being a known noise covariance. This model (7) generates noisy samples

wi∼N(0,σz2PS+σε2Id)(9)w_i\sim\mathcal{N}(0,\sigma_z^2P_S+\sigma_\varepsilon^2I_d) \tag*{(9)}

from SS. Noise-free samples wi=Γzi∼N(0,σz2PS)w_i=\Gamma z_i\sim\mathcal{N}(0,\sigma_z^2P_S) satisfy wi∈Sw_i\in S with a distribution that does not depend on the specific representative Γ\Gamma but only the rotation-invariant projection PSP_S. The latent vector ziz_i can be thought of as a prior on the initial state of the system and serves as a hyperparameter. The choice of μz∈Rr\mu_z\in\mathbb{R}^r is not critical (as a prior its influence will be forgotten as the dataset grows), so we take μz=0\mu_z=0 throughout for simplicity. The variance of ziz_i must be isotropic however, if instead zi∼N(0,Λ)z_i\sim\mathcal{N}(0,\Lambda) with a fixed Λ∝̸I\Lambda\not\propto I, then the noise-free trajectories depend on the representation Γ\Gamma rather than the subspace SS.

By Bayes’ rule, the posterior density of SS with respect to ν\nu is

p(S∣W)=p(W∣S)p(S)p(W),(10)p(S\mid W)=\frac{p(W\mid S)p(S)}{p(W)}, \tag*{(10)}

with

p(W)=∫Gr⁡(d,r)p(W∣T)p(T) dν(T).p(W)=\int_{\operatorname{Gr}(d,r)}p(W\mid T)p(T)\,\mathrm{d}\nu(T).

Next, we examine the likelihood p(W∣Γ)p(W\mid\Gamma), prior p(Γ)p(\Gamma), and normalizing constant (or evidence) p(W)p(W).

  1. The likelihood: We first introduce the sample covariance matrix and aggregated precision, defined as

SW:=1N−1WWT,α:=N−12σz2σε2(σε2+σz2),(11)S_W:=\frac{1}{N-1}WW^{\mathsf{T}},\qquad\alpha:=\frac{N-1}{2}\frac{\sigma_z^2}{\sigma_\varepsilon^2(\sigma_\varepsilon^2+\sigma_z^2)}, \tag*{(11)}

respectively. The next result characterizes the likelihood.

Proposition 1 (Likelihood): Consider the observation model (7) and the data matrix WW defined in (6). Assume that σz2>0\sigma_z^2>0 and σε2>0\sigma_\varepsilon^2>0 are known. Then, p(W∣S)∝Bing⁡d,r(αSW)p(W\mid S)\propto\operatorname{Bing}_{d,r}(\alpha S_W), i.e., for every S∈Gr⁡(d,r)S\in\operatorname{Gr}(d,r),

p(W∣S)=C1etr⁡(αSWPS)(12)p(W\mid S)=C_1\operatorname{etr}(\alpha S_WP_S) \tag*{(12)}

where SWS_W and α\alpha are defined in (10) and

C1=((2π)d(σε2)d−r(σε2+σz2)r)−N2exp⁡(−N−12σε2tr⁡(SW)).C_1=\left((2\pi)^d(\sigma_\varepsilon^2)^{d-r}(\sigma_\varepsilon^2+\sigma_z^2)^r\right)^{-\frac{N}{2}}\exp\left(-\frac{N-1}{2\sigma_\varepsilon^2}\operatorname{tr}(S_W)\right).

Proof: Fix a representative Γ\Gamma of SS. By (7), the wiw_i are i.i.d. with wi∼N(0,Σw)w_i\sim\mathcal{N}(0,\Sigma_w), Σw=σz2PS+σε2Id\Sigma_w=\sigma_z^2P_S+\sigma_\varepsilon^2I_d. Since PSP_S is an orthogonal projector of rank rr, ∣Σw∣=(σε2+σz2)r(σε2)d−r\lvert\Sigma_w\rvert=(\sigma_\varepsilon^2+\sigma_z^2)^r(\sigma_\varepsilon^2)^{d-r} and Σw−1=σε−2Id−αN−1PS\Sigma_w^{-1}=\sigma_\varepsilon^{-2}I_d-\frac{\alpha}{N-1}P_S [45]. Hence

p(W∣S)=(2π)−Nd2∣Σw∣−N2etr⁡(−N−12Σw−1SW),p(W\mid S)=(2\pi)^{-\frac{Nd}{2}}\lvert\Sigma_w\rvert^{-\frac{N}{2}}\operatorname{etr}\left(-\frac{N-1}{2}\Sigma_w^{-1}S_W\right),

which is (11), and depends on Γ\Gamma only through PSP_S.

Proposition 1 shows that the log-likelihood is an affine function of PSP_S and that the data enter only through SWS_W and, as a function of SS, the likelihood is proportional to a matrix Bingham density. The likelihood can also be naturally expressed using projectors. Some algebraic manipulation yields

p(W∣S)=C1etr⁡(αSW)exp⁡(−αN−1M(S)),p(W\mid S)=C_1\operatorname{etr}(\alpha S_W)\exp\left(-\frac{\alpha}{N-1}M(S)\right),

where M(S)=∑i=1N∥(Id−PS)wi∥2M(S)=\sum_{i=1}^{N}\lVert(I_d-P_S)w_i\rVert^2. M(S)M(S) is known as the “misfit” and the terms ∥(I−PS)wi∥\lVert(I-P_S)w_i\rVert are a measure of distance between the data and the subspace SS [46].

Remark 1: (Maximum Likelihood Estimation) The Maximum Likelihood Estimate (MLE)

S^MLE∈arg⁡max⁡S∈Gr⁡(d,r) p(W∣S)(13)\widehat{S}_{\mathrm{MLE}}\in\underset{S\in\operatorname{Gr}(d,r)}{\arg\max}\ p(W\mid S) \tag*{(13)}

provides some insights about the likelihood. By Ky Fan’s maximum principle [47] S^MLE=Im⁡Γ^r\widehat{S}_{\mathrm{MLE}}=\operatorname{Im}\widehat{\Gamma}_r, where Γ^r\widehat{\Gamma}_r is a matrix of the rr dominant Eigenvectors of SWS_W. This equivalent to the PCA estimate of [45] and the column space of the truncated SVD of the data matrix used in subspace identification [3] and in subspace predictive control [48]. The subspace S^\widehat{S} is unique if and only if λr>λr+1\lambda_r>\lambda_{r+1}, where λ1≥⋯≥λd\lambda_1\geq\cdots\geq\lambda_d are the eigenvalues of SWS_W. This condition holds almost surely under Gaussian noise whenever N>rN>r, i.e., we have enough data relative to the rank of the system.

Selection of the prior

We use priors of the form

p(S)=Bing⁡d,r(S;A0),(14)p(\mathbf{S})=\operatorname{Bing}_{d,r}(\mathbf{S};A_0), \tag*{(14)}

in the matrix Bingham family (1). We consider two specific choices of the prior. The choice A0=0A_0=0 gives the uninformative uniform prior p(S0)=1p(S_0)=1, that is, the invariant distribution ν\nu. Alternatively, let S0∈Gr⁡(d,r)S_0\in\operatorname{Gr}(d,r) be a reference subspace with projector P0=Γ0Γ0TP_0=\Gamma_0\Gamma_0^{\mathsf{T}} obtained for example by a SVD of an independent data matrix WdW_d. Then the informative prior

p(S)=Bing⁡d,r(S;κP0)=eκrexp⁡(−κdc(S,S0)2)(15)p(\mathbf{S})=\operatorname{Bing}_{d,r}(\mathbf{S};\kappa P_0)=e^{\kappa r}\exp\left(-\kappa d_c(\mathbf{S},S_0)^2\right) \tag*{(15)}

is a decreasing function of the chordal distance to S0S_0 and has mode S0S_0. The hyperparameter κ≥0\kappa\geq0 thus quantifies trust in S0S_0 and (14) reduces to the uniform prior for κ=0\kappa=0.

The evidence

For both choices of prior, the evidence p(W)p(W) in (9) cannot be written in closed-form but can be expressed in terms of hypergeometric functions. With the uniform prior,

p(W)=C1 1F1(r2;d2;αSW)p(W)=C_1\,{}_1F_1\left(\frac{r}{2};\frac{d}{2};\alpha S_W\right)

by Proposition 1 and (1). With the prior (14),

p(W)=C1 1F1(r2;d2;αSW+κP0)/1F1(r2;d2;κP0).p(W)=C_1\,{}_1F_1\left(\frac{r}{2};\frac{d}{2};\alpha S_W+\kappa P_0\right)\big/{}_1F_1\left(\frac{r}{2};\frac{d}{2};\kappa P_0\right).

Both functions can be evaluated with the algorithm of [42] or by Monte Carlo integration with exact samples from ν\nu [49].

The posterior distribution

With the likelihood and prior in hand, we can derive expressions for the posterior.

Theorem 1 (Posterior). Consider the observation model (7), the data matrix WW, and SWS_W and α\alpha as in (10). Assume that σz2>0\sigma_z^2>0 and σε2>0\sigma_\varepsilon^2>0 are known and that the prior is Bing⁡d,r(A0)\operatorname{Bing}_{d,r}(A_0) for a symmetric A0∈Rd×dA_0\in\mathbb{R}^{d\times d}. Then the posterior is distributed according to

p(S∣W)∼Bing⁡d,r(A0+αSW)(16)p(S\mid W)\sim\operatorname{Bing}_{d,r}(A_0+\alpha S_W) \tag*{(16)}

In particular, the uniform prior gives Bing⁡d,r(αSW)\operatorname{Bing}_{d,r}(\alpha S_W) and the prior (14) gives Bing⁡d,r(αSW+κP0)\operatorname{Bing}_{d,r}(\alpha S_W+\kappa P_0).

Proof. By Proposition 1 and (1), p(W∣S)p(S)p(W\mid S)p(S) equals exp⁡(tr⁡((A0+αSW)PS))\exp(\operatorname{tr}((A_0+\alpha S_W)P_S)) times a factor that does not depend on SS. Normalizing with respect to ν\nu by (1) gives (15).

Theorem 1 shows that the matrix Bingham family is conjugate to the likelihood of Proposition 1. Conjugacy follows because the log-likelihood is affine in PSP_S, the sufficient statistic of the exponential family Bing⁡d,r\operatorname{Bing}_{d,r} on Gr⁡(d,r)\operatorname{Gr}(d,r). The data enter through the additive update A=A0+αSWA=A_0+\alpha S_W. Since α\alpha grows linearly with NN, a fixed prior contribution A0A_0 becomes negligible relative to the data term as N→∞N\to\infty.

Remark 2: (The MAP Estimate) The maximum a posteriori (MAP) estimate is

S^MAP∈arg⁡max⁡S∈Gr⁡(d,r) p(S∣W).\widehat{S}_{\mathrm{MAP}}\in\underset{S\in\operatorname{Gr}(d,r)}{\arg\max}\ p(S\mid W).

By Ky Fan’s principle, it can be computed as S^MAP=im⁡Vr\widehat{S}_{\mathrm{MAP}}=\operatorname{im}V_r, where the columns of VrV_r are eigenvectors associated with the rr largest eigenvalues of AA. For the prior A0=κP0A_0=\kappa P_0, the MAP estimate coincides with the MLE when κ=0\kappa=0 and converges to S0S_0 as κ→∞\kappa\to\infty. The concentration of the posterior around the MAP estimate is governed by the gaps between the rr largest eigenvalues of AA and the remaining eigenvalues.

The condition (4) for noise-free data has a natural statistical counterpart. One may require the distribution of the underlying trajectories to span the restricted behaviour, i.e.,

im⁡Σwˉ=BL,Σwˉ:=E[wˉiwˉiT].(17)\operatorname{im}\Sigma_{\bar{w}}=\mathcal{B}_L,\qquad\Sigma_{\bar{w}}:=\mathbb{E}[\bar{w}_i\bar{w}_i^{\mathsf{T}}]. \tag*{(17)}

Under i.i.d. sampling with finite second moments and isotropic measurement noise, the MLE then converges almost surely to BL\mathcal{B}_L as N→∞N\to\infty. The MAP estimate has the same limit for a fixed Bingham prior. The observation model (7) satisfies this condition automatically.

Prediction

Let Tini+Tf=LT_{\mathrm{ini}}+T_f=L and let

w=Γz+ε(18)w=\Gamma z+\varepsilon \tag*{(18)}

be a new trajectory, where (z,ε)(z,\varepsilon) is independent of WW and distributed as (7). We permute and partition the trajectory as w=(wa,yf)w=(w_a,y_f), where wa=(uini,yini,uf)∈Rdaw_a=(u_{\mathrm{ini}},y_{\mathrm{ini}},u_f)\in\mathbb{R}^{d_a} collects the initial inputs uini∈RmTiniu_{\mathrm{ini}}\in\mathbb{R}^{mT_{\mathrm{ini}}}, the initial outputs yini∈RpTiniy_{\mathrm{ini}}\in\mathbb{R}^{pT_{\mathrm{ini}}} and the future inputs uf∈RmTfu_f\in\mathbb{R}^{mT_f}, and yf∈RpTfy_f\in\mathbb{R}^{pT_f} collects the future outputs we wish to predict. We also partition a representative and the projector,

Γ=[Γ1Γ2],PS=[P11P12P21P22]=[Γ1Γ1TΓ1Γ2TΓ2Γ1TΓ2Γ2T].\Gamma=\begin{bmatrix}\Gamma_1\\\Gamma_2\end{bmatrix},\qquad P_S=\begin{bmatrix}P_{11}&P_{12}\\P_{21}&P_{22}\end{bmatrix} =\begin{bmatrix}\Gamma_1\Gamma_1^{\mathsf{T}}&\Gamma_1\Gamma_2^{\mathsf{T}}\\\Gamma_2\Gamma_1^{\mathsf{T}}&\Gamma_2\Gamma_2^{\mathsf{T}}\end{bmatrix}.

in the same way. The following theorem characterizes the predictive distribution.

Theorem 2 (Predictive distribution). Consider the observation model (7), the data matrix WW, and the new trajectory w=(wa,yf)w=(w_a,y_f). Assume that σz2>0\sigma_z^2>0 and σε2>0\sigma_\varepsilon^2>0 are known and let S∣W∼Bing⁡d,r(A)S\mid W\sim\operatorname{Bing}_{d,r}(A) be our belief about SS. Then:

i) For any fixed SS, (yf∣wa,S)∼N(m(S),Σ(S))(y_f\mid w_a,S)\sim\mathcal{N}(m(S),\Sigma(S)) with

m(S)=σz2P21(σz2P11+σε2I)−1wa,m(S)=\sigma_z^2P_{21}\left(\sigma_z^2P_{11}+\sigma_\varepsilon^2I\right)^{-1}w_a,
Σ(S)=σz2P22+σε2I−σz4P21(σz2P11+σε2I)−1P12,\Sigma(S)=\sigma_z^2P_{22}+\sigma_\varepsilon^2I-\sigma_z^4P_{21}\left(\sigma_z^2P_{11}+\sigma_\varepsilon^2I\right)^{-1}P_{12},

ii) the predictive density is

p(yf∣wa,W)=∫Gr⁡(d,r)N(yf;m(S),Σ(S)) p(S∣wa,W) dν(S),(19)p(y_f\mid w_a,W)=\int_{\operatorname{Gr}(d,r)}\mathcal{N}(y_f;m(S),\Sigma(S))\,p(S\mid w_a,W)\,d\nu(S), \tag*{(19)}

where p(S∣wa,W)∝N(wa;0,σz2P11+σε2I)p(S∣W)p(S\mid w_a,W)\propto\mathcal{N}(w_a;0,\sigma_z^2P_{11}+\sigma_\varepsilon^2I)p(S\mid W). Replacing p(S∣wa,W)p(S\mid w_a,W) by p(S∣W)p(S\mid W) in (18) omits the additional information about SS supplied by waw_a. The approximation is accurate when p(wa∣S,W)p(w_a\mid S,W) is approximately constant over the region where p(S∣W)p(S\mid W) concentrates, as may occur for sufficiently large NN.

ii) Since ww and WW are conditionally independent given S\mathbf{S}, p(yf∣wa,S,W)=p(yf∣wa,S)p(y_f \mid w_a,\mathbf{S},W)=p(y_f \mid w_a,\mathbf{S}) and p(S∣wa,W)∝p(wa∣S)p(S∣W)p(\mathbf{S}\mid w_a,W)\propto p(w_a\mid\mathbf{S})p(\mathbf{S}\mid W). The claim follows from the law of total probability and Theorem 1.

Theorem 2 expresses the predictive density as a continuous mixture of Gaussian densities weighted by the posterior p(S∣W,wa)p(\mathbf{S}\mid W,w_a). The resulting density need not be Gaussian and may be multimodal when substantial uncertainty about the restricted behaviour remains. The integral over the manifold can be evaluated by quadrature in low dimensions or approximated using Riemannian Monte Carlo methods in higher dimensions.

When the posterior is concentrated near a point estimate S^\hat{\mathbf{S}}, such as the MLE or MAP estimate, one may use the approximation

p(yf∣wa,W)≈p(yf∣wa,S^),p(y_f\mid w_a,W)\approx p(y_f\mid w_a,\hat{\mathbf{S}}),

where the right-hand side is the Gaussian density characterized in Theorem 2. For fixed S\mathbf{S}, the noise-free limit σε→0\sigma_\varepsilon\to0 gives

m(S)→Γ2Γ1†wa=P21P11†wa,m(\mathbf{S})\to\Gamma_2\Gamma_1^\dagger w_a=P_{21}P_{11}^\dagger w_a,
Σ(S)→Σˉ=σz2Γ2(Ir−Γ1†Γ1)Γ2⊤.\Sigma(\mathbf{S})\to\bar{\Sigma}=\sigma_z^2\Gamma_2(I_r-\Gamma_1^\dagger\Gamma_1)\Gamma_2^\top.

If Γ1\Gamma_1 has full column rank, then Σˉ=0\bar{\Sigma}=0 and, for compatible wa∈im⁡Γ1w_a\in\operatorname{im}\Gamma_1, the limiting prediction is

yf=Γ2Γ1†wa.y_f=\Gamma_2\Gamma_1^\dagger w_a.

This recovers the noise-free predictor of [50, 48], evaluated at S\mathbf{S}. For S=BL\mathbf{S}=\mathcal{B}_L, the full column rank condition holds if Tini≥l(S)T_{\mathrm{ini}}\geq l(\mathbf{S}), where l(S)l(\mathbf{S}) is the system lag.

Numerical Case Study

Consider a first-order autonomous system

xt+1=axtx_{t+1}=ax_t
yt=xt,y_t=x_t,

where a=0.8a=0.8, with additive measurement noise on the output according to ε∼N(0,σε2)\varepsilon\sim\mathcal{N}(0,\sigma_\varepsilon^2) and σε=0.15\sigma_\varepsilon=0.15. We choose this system because its restricted behaviour for L=2L=2 is a line in R2\mathbb{R}^2, allowing direct visualization of uncertainty in the inferred subspace. We generate NN trajectories of length L=2L=2. We rewrite the system as

w=[y0y1]=[x0ax0]+ε=Γz+εw=\begin{bmatrix}y_0\\y_1\end{bmatrix} =\begin{bmatrix}x_0\\ax_0\end{bmatrix} +\varepsilon=\Gamma z+\varepsilon

A basis for the noise-free trajectory subspace is deduced from this representation as

Γtrue=11+a2[1a]=[cos⁡θsin⁡θ]\Gamma_{\mathrm{true}}=\frac{1}{\sqrt{1+a^2}}\begin{bmatrix}1\\a\end{bmatrix} =\begin{bmatrix}\cos\theta\\\sin\theta\end{bmatrix}

yielding the noise-free subspace as a line through the origin angled at θ0≈38.66∘\theta_0\approx38.66^\circ.

  1. Inference: Assuming an uninformative uniform prior, we compute the posterior density over a subset of candidate subspaces Γ∈St⁡(2,1)\Gamma\in\operatorname{St}(2,1). Figure 1 shows the posterior density with true subspace and MAP estimates. Figure 2 illustrates that the angular error scales as we gather more data. Figure 3 demonstrates the evolution of the subspace angle MAP estimate as a function of the signal-to-noise ratio (SNR), SNR=y0/σε\mathrm{SNR}=y_0/\sigma_\varepsilon.

Posterior density as a function of subspace angle

Figure 1. Posterior density with true subspace θ0=38.66∘\theta_0=38.66^\circ and MAP estimates θ^=37.92∘\hat{\theta}=37.92^\circ, 38.51∘38.51^\circ and 38.62∘38.62^\circ for N=100N=100, 500500 and 10001000 samples, respectively.

Principal angle error as a function of the number of samples

Figure 2. Principal angle error between MAP subspace estimate and true subspace as a function of NN for σε=0.15\sigma_\varepsilon=0.15 with corresponding trendline. The error decreases as N−1/2N^{-1/2}. Each data point represents the average over multiple independent noise realizations to obtain a more representative estimate of the underlying trend.

MAP subspace estimate and confidence interval as a function of signal-to-noise ratio

Figure 3. MAP subspace estimate and corresponding 95% confidence interval as a function of the signal-to-noise ratio, using N=500N=500 samples. The posterior converges to a Dirac measure at the true subspace as σε→0\sigma_\varepsilon\to0

  1. Prediction: Next, we perform prediction as in § 3-B. We focus on the MAP conditioned distribution p(yf∣wa,S^)p(y_f \mid w_a,\hat{\mathbf{S}}) for computational feasibility as evaluating the full predictive posterior will require specialized sampling algorithms which we defer to an extension.

Here da=r=1d_a=r=1 and, writing S=span⁡{(cos⁡θ,sin⁡θ)}\mathbf{S}=\operatorname{span}\{(\cos\theta,\sin\theta)\} with θ∈(−π/2,π/2]\theta\in(-\pi/2,\pi/2], the noise-free predictor is y1=y0tan⁡θy_1=y_0\tan\theta. Interestingly, the posterior density of θ\theta is bounded below by a positive constant and ∫−π/2π/2∣tan⁡θ∣ dθ=∞\int_{-\pi/2}^{\pi/2}|\tan\theta|\,d\theta=\infty, so y1y_1 has no posterior mean. Since tan⁡\tan is increasing on (−π/2,π/2)(-\pi/2,\pi/2), the posterior quantiles of y1y_1 are y0tan⁡y_0\tan of the posterior quantiles of θ\theta. Figure 4 compares the output distribution obtained using the true subspace S\mathbf{S} with the predictive distribution obtained by propagating subspace uncertainty according to Theorem 1. Finally, Figure 5 illustrates the dependence of the MAP-based output prediction on the SNR.

Posterior densities

Figure 4. Posterior densities for N=500N=500 with true noisy yfy_f and posterior yfy_f with σε=0.15\sigma_\varepsilon=0.15 and σz=1\sigma_z=1, using a fixed MAP subspace estimate rather than the full posterior distribution. The true and predicted distributions are close but offset, illustrating how errors in the estimated subspace propagate to the output prediction and introduce systematic bias. For σε=0.15\sigma_\varepsilon=0.15 and y0=1y_0=1, the posterior median of y1y_1 is 0.768 and its 95% credible interval is [0.695, 0.840].

MAP output estimate and confidence intervals

Figure 5. MAP output estimate and corresponding 95% confidence interval as a function of signal-to-noise ratio, calculated using fixed MAP subspace estimate with N=500N=500. This predictive distribution is conditioned on a single MAP subspace estimate and neglects uncertainty in the underlying subspace, producing overconfident predictions. The full predictive distribution (18) should suffer less from this issue. However, the true prediction does reliably lie in the 95% confidence interval indicating an acceptable level of approximation error.

  1. Summary: At a high level, the numerical example indicates that: 1) the framework is consistent in the sense that the posterior distribution and predictive distribution approach that of the data generating process (almost surely) as N→∞N\to\infty or σε→0\sigma_\varepsilon\to0 and 2) the MAP conditioned distribution p(yf∣wa,S^)p(y_f\mid w_a,\hat{\mathbf{S}}) is overconfident but still provides a useful prediction and degree of uncertainty quantification.

Conclusions and Future Work

The paper introduced a Bayesian framework for learning restricted behaviours from noisy input–output data. Uncertainty about the restricted behaviour is represented by a probability distribution on the Grassmann manifold. Under the Gaussian observation model considered, we established conjugacy of the matrix Bingham family and derived predictive distributions that account for uncertainty in the inferred behaviour. A numerical case study illustrated the inference and prediction procedures.

Future work includes extending the observation model beyond isotropic noise, developing scalable algorithms for probabilistic prediction in larger systems, and exploring non-parametric representations of uncertainty about the restricted behaviour. Applications to model fusion and model predictive control could combine information from first-principles models with operational data while accounting for uncertainty in the inferred behaviour.

References

  1. [1]J. C. Willems, “From time series to linear system—Part I. Finite dimensional linear time invariant systems,” Automatica, vol. 22, no. 5, pp. 561–580, Sep. 1986.DOI
  2. [2]J. C. Willems, P. Rapisarda, I. Markovsky, and B. L. De Moor, “A note on persistency of excitation,” Syst. Control Lett., vol. 54, no. 4, pp. 325–329, Apr. 2005.
  3. [3]P. van Overschee and B. L. R. De Moor, Subspace Identification for Linear Systems: Theory - Implementation - Applications. Boston, MA, USA: Kluwer Acad. Publ, 1996.
  4. [4]I. Markovsky, J. C. Willems, P. Rapisarda, and B. L. De Moor, “Algorithms for deterministic balanced subspace identification,” Automatica, vol. 41, no. 5, pp. 755–766, May 2005.DOI
  5. [5]M. Verhaegen and V. Verdult, Filtering and System Identification: A Least Squares Approach, 1st ed. Cambridge, U.K.: Cambridge University Press, Apr. 2007.
  6. [6]J. Coulson, J. Lygeros, and F. Dörfler, “Data-enabled predictive control: In the shallows of the DeePC,” in 2019 18th European Control Conference (ECC). Naples, Italy: IEEE, Jun. 2019, pp. 307–312.arxiv.org/abs/1811.05890
  7. [7]J. Coulson, J. Lygeros, and F. Dörfler, “Distributionally robust chance constrained data-enabled predictive control,” IEEE Trans. Automat. Contr., vol. 67, no. 7, pp. 3289–3304, Jul. 2022.DOI
  8. [8]I. Markovsky and F. Dörfler, “Behavioral systems theory in data-driven analysis, signal processing, and control,” Annu. Rev. Control, vol. 52, pp. 42–64, 2021.DOI
  9. [9]I. Markovsky, L. Huang, and F. Dörfler, “Data-Driven Control Based on the Behavioral Approach: From Theory to Applications in Power Systems,” IEEE Control Syst., vol. 43, no. 5, pp. 28–68, Oct. 2023.DOI
  10. [10]F. Dörfler, J. Coulson, and I. Markovsky, “Bridging direct and indirect data-driven control formulations via regularizations and relaxations,” IEEE Trans. Automat. Contr., vol. 68, no. 2, pp. 883–897, Feb. 2023.
  11. [11]L. Huang, J. Zhen, J. Lygeros, and F. Dörfler, “Robust data-enabled predictive control: Tractable formulations and performance guarantees,” IEEE Trans. Automat. Contr., vol. 68, no. 5, pp. 3163–3170, May 2023.
  12. [12]I. Markovsky, Low Rank Approximation: Algorithms, Implementation, Applications. London, U.K.: Springer, 2012.
  13. [13]I. Markovsky and F. Dörfler, “Data-driven dynamic interpolation and approximation,” Automatica, vol. 135, Jan. 2022.DOI
  14. [14]M. Yin, A. Ianelli, S. A. Nazari, and M. A. Müller, “A unified Bayesian framework for data-driven smoothing, prediction, and control,” Apr. 2026, https://arxiv.org/abs/2512.01475.
  15. [15]A. Padoan, J. Coulson, H. J. Van Waarde, D. Lygeros, and F. Dörfler, “Behavioral uncertainty quantification for data-driven control,” in 2022 IEEE 61st Conference on Decision and Control (CDC). Cancun, Mexico: IEEE, Dec. 2022, pp. 4726–4731.
  16. [16]A. Padoan and J. Coulson, “Distances between finite-horizon linear behaviors,” IEEE Control Syst. Lett., vol. 9, pp. 1772–1777, 2025.arxiv.org/abs/2503.22849
  17. [17]M. Girolami and B. Calderhead, “Riemann Manifold Langevin and Hamiltonian Monte Carlo Methods,” Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 73, no. 2, pp. 123–214, Mar. 2011.DOI
  18. [18]M.-N. Tran, D. H. Nguyen, and D. Nguyen, “Variational Bayes on manifolds,” Stat Comput, vol. 31, no. 6, p. 71, Nov. 2021.arxiv.org/abs/1908.03097
  19. [19]L. Lin, V. Rao, and D. Dunson, “Bayesian nonparametric inference on the Stiefel manifold,” Statist. Sinica, vol. 27, no. 2, pp. 535–553, 2017.arxiv.org/abs/1311.0907
  20. [20]P. Turaga, A. Veeraraghavan, and R. Chellappa, “Statistical analysis on Stiefel and Grassmann manifolds with applications in computer vision,” in 2008 IEEE Conference on Computer Vision and Pattern Recognition. Anchorage, AK, USA: IEEE, Jun. 2008, pp. 1–8.DOI
  21. [21]L. Ljung, System Identification: Theory for the User, 2nd ed. Upper Saddle River, NJ, USA: Prentice Hall, 1999.
  22. [22]A. Chiuso and G. Picci, “Numerical conditioning and asymptotic variance of subspace estimates,” Automatica, vol. 40, no. 4, pp. 677–683, Apr. 2004.DOI
  23. [23]A. Tsiamis, I. Ziemann, N. Matni, and G. J. Pappas, “Statistical Learning Theory for Control: A Finite-Sample Perspective,” IEEE Control Syst., vol. 43, no. 6, pp. 67–97, Dec. 2023.arxiv.org/abs/2209.05423
  24. [24]A. R. Mesquita, “Robust and Bayesian subspace identification,” IEEE Trans. Automat. Contr., vol. 70, no. 2, pp. 1395–1401, Feb. 2025.DOI
  25. [25]S. Damjanovic, J. W. Simpson-Porco, and K. van Heusden, “Distributionally robust frequency response estimation using non-steady-state frequency-domain data,” Oct. 2026, https://arxiv.org/abs/2610.09395.
  26. [26]Y. Pedari, F. Kaviani, and H. R. Ossareh, “Error bounds for direct data-driven frequency-response estimation,” International Journal of Control, vol. 99, no. 4, pp. 901–911, Apr. 2026.DOI
  27. [27]T. Söderström and P. Stoica, System Identification. New York, NY, USA: Prentice Hall, 1989.
  28. [28]M. Yin, A. Ianelli, and R. S. Smith, “Data-driven prediction with stochastic data: Confidence regions and minimum mean-squared error estimates,” in 2022 European Control Conference (ECC). London, United Kingdom: IEEE, Jul. 2022, pp. 853–858.DOI
  29. [29]F. Kaviani, I. Markovsky, and H. R. Ossareh, “Uncertainty quantification of data-driven output predictors in the output error setting,” IEEE Trans. Automat. Contr., vol. 70, no. 11, pp. 7588–7595, Nov. 2025.arxiv.org/abs/2404.15098
  30. [30]D. Jin and J. Coulson, “On the sensitivity of the subspace predictor to behavioral perturbations,” IEEE Control Syst. Lett., vol. 10, pp. 1609–1614, 2026.
  31. [31]T. Faulwasser, R. Ou, G. Pan, P. Schmitz, and K. Worthmann, “Behavioral theory for stochastic systems? A data-driven journey from Willems to Wiener and back again,” Annu. Rev. Control, vol. 55, pp. 92–117, 2023.
  32. [32]A. Sasfi, I. Markovskyy, A. Padoan, and F. Dörfler, “Gaussian behaviors: representations and data-driven control,” in 2025 IEEE 64th Conference on Decision and Control (CDC). Rio de Janeiro, Brazil: IEEE, Dec. 2025, pp. 1993–1998.arxiv.org/abs/2504.15838
  33. [33]V. M. Preciado, “A measure-theoretic formulation of behavioral systems,” May 2026, http://arxiv.org/abs/2605.01558.
  34. [34]P.-A. Absil, R. Mahony, and R. Sepulchre, “Riemannian geometry of Grassmann manifolds with a view on algorithmic computation,” Acta Appl. Math., vol. 80, no. 2, pp. 199–220, Jan. 2004.DOI
  35. [35]N. Boumal, An Introduction to Optimization on Smooth Manifolds, 1st ed. Cambridge, U.K.: Cambridge University Press, Mar. 2023.DOI
  36. [36]P.-A. Absil, R. Mahony, and R. Sepulchre, Optimization Algorithms on Matrix Manifolds. Princeton, NJ, USA: Princeton Univ. Press, Dec. 2008.DOI
  37. [37]Y. Chikuse, Statistics on Special Manifolds. New York, NY, USA: Springer, 2003.
  38. [38]T. Bendokat, R. Zimmermann, and P.-A. Absil, “A Grassmann manifold handbook: Basic geometry and computational aspects,” Adv Comput Math, vol. 50, no. 1, p. 6, Feb. 2024.DOI
  39. [39]J. H. Conway, R. H. Hardin, and N. J. A. Sloane, “Packing lines, planes, etc.: Packings in Grassmannian spaces,” Exp. Math., vol. 5, no. 2, pp. 139–159, Jan. 1996.DOI
  40. [40]A. Edelman, T. A. Arias, and S. T. Smith, “The geometry of algorithms with orthogonality constraints,” SIAM J. Matrix Anal. Appl., vol. 20, no. 2, pp. 303–353, Jun. 1998.arxiv.org/abs/physics/9806030
  41. [41]R. J. Muirhead, Aspects of Multivariate Statistical Theory. Hoboken, NJ, USA: Wiley-Interscience, 2005.
  42. [42]P. Koev and A. Edelman, “The efficient evaluation of the hypergeometric function of a matrix argument,” Math. Comp., vol. 75, no. 254, pp. 833–846, Jan. 2006.arxiv.org/abs/math/0505344
  43. [43]T. Söderström, Errors-in-Variables Methods in System Identification. Cham, Switzerland: Springer, 2018.
  44. [44]H. J. Van Waarde, C. De Persis, M. K. Camlibel, and P. Tesi, “Willems’ fundamental lemma for state-space systems and its extension to multiple datasets,” IEEE Control Syst. Lett., vol. 4, no. 3, pp. 602–607, Jul. 2020.arxiv.org/abs/2002.01023
  45. [45]M. E. Tipping and C. M. Bishop, “Probabilistic principal component analysis,” J. R. Statist. Soc. Ser. B (Statist. Methodol.), vol. 61, no. 3, pp. 611–622, Sep. 1999.DOI
  46. [46]I. Markovsky, J. C. Willems, S. Van Huffel, and B. De Moor, Exact and Approximate Modeling of Linear Systems: A Behavioral Approach. Philadelphia, PA, USA: SIAM, Jan. 2006.
  47. [47]K. Fan, “On a theorem of Weyl concerning eigenvalues of linear transformations I,” Proc. Natl. Acad. Sci. U.S.A., vol. 35, no. 11, pp. 652–655, Nov. 1949.
  48. [48]W. Favoreel, B. D. Moor, and M. Gevers, “SPC: Subspace predictive control,” IFAC Proc. Vol., vol. 32, no. 2, pp. 4004–4009, Jul. 1999.DOI
  49. [49]F. Mezzadri, “How to generate random matrices from the classical compact groups,” Feb. 2007, https://arxiv.org/abs/math-ph/0609050.
  50. [50]I. Markovsky and P. Rapisarda, “Data-driven simulation and control,” Int. J. Control, vol. 81, no. 12, pp. 1946–1959, Dec. 2008.DOI

Paper details

vlnc:W_ecdgmeh2
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents