Online Subspace Learning on Flag Manifolds for System Identification
Dian Jin, Jeremy Coulson · 2025
Read Online Subspace Learning on Flag Manifolds for System Identification by Dian Jin, Jeremy Coulson (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-2511-06416.txt.
Introduction
Dian Jin and Jeremy Coulson learn system models online from streaming data when the true complexity is unknown or drifting. Subspace methods summarize a dynamical system by the directions its behavior spans, but classical versions assume fixed linear systems with known order. Their framework tracks a nested family of subspaces on a flag manifold instead.
New data recursively updates the whole ensemble of subspaces, so the model adapts as dimensions change with no advance knowledge of order. The method generalizes an earlier Grassmannian tracking algorithm. Plugged into a data-driven simulator, the adaptive predictor handled sudden changes in system behavior and outperformed several baselines in the case study.
Catalog record
| CREATOR | Dian Jin, Jeremy Coulson |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-06416
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Online Subspace Learning on Flag Manifolds for System Identification Dian Jin, Jeremy Coulson Abstract Data-driven control methods based on subspace representations are powerful but are often limited to linear time-invariant systems where the model order is known. A key challenge is developing online data-driven control algorithms for time-varying systems, especially when the system’s complexity is unknown or changes over time. To address this, we propose a novel online subspace learning framework that operates on flag manifolds. Our algorithm leverages streaming data to recursively track an ensemble of nested subspaces, allowing it to adapt to varying system dimensions without prior knowledge of the true model order. We show that our algorithm is a generalization of the Grassmannian Recursive Algorithm for Tracking. The learned subspace models are then integrated into a data-driven simulation framework to perform prediction for unknown dynamical systems. The effectiveness of this approach is demonstrated through a case study where the proposed adaptive predictor successfully handles abrupt changes in system dynamics and outperforms several baselines. keywords System identification; Subspace learning; Optimization on manifolds 1 Introduction Subspace representations are becoming increasingly important in the field of systems and control, especially with the recent revival of behavioral systems theory ( 37 ) in data-driven control. A key concept in this area is Willems’ Fundamental Lemma, which provides a direct data-driven non-parametric subspace representation for finite-length trajectories of linear time-invariant systems. This has spurred the development of effective subspace identification techniques ( 16 ; 33 ) , novel data-driven simulation and control methods ( 26 ; 8 ; 9 ; 5 ; 36 ; 12 ) , which have been shown to perform remarkably well across a wide range of application domains ( 15 ; 21 ; 20 ; 22 ; 17 ) . See ( 24 ) and references therein. Despite these advances, many existing identification and control methods are limited to Linear Time-Invariant systems. A significant open challenge is the design of online adaptive approaches and recursive algorithms that can handle systems that change over time. While some subspace tracking methods have been developed for online state-space identification ( 35 ; 32 ; 27 ; 41 ; 29 ) , they typically require the system order to be known and constant. In many real-world scenarios, however, the true system order is unknown and may vary, for instance, due to component failures. In this work, we propose a novel framework for adaptive, data-driven identification of unknown, time-varying dynamical systems based on subspace learning techniques. Subspace learning (sometimes termed tracking), has long been studied in signal processing ( 38 ; 10 ; 3 ; 4 ; 19 ) . Broadly speaking, subspace learning algorithms can be classified into algebraic and geometric methods ( 2 ) . The geometric methods optimize a certain loss function via gradient descent on a matrix manifold such as the Grassmannian, a set consisting of all subspaces of a particular dimension. Such methods have a solid theoretical background rooted in optimization on manifolds ( 14 ; 1 ; 6 ) . Running these algorithms on Grassmannians requires choosing the dimensionality a priori . This poses a significant challenge in practical applications where the true underlying dimension is unknown or may change over time. Contribution. We propose a novel framework for online system identification and data-driven simulation for unknown time-varying dynamical systems. The proposed framework consists of the Flag Recursive Online Tracking algorithm on flag manifolds that leverages streaming data and yields a nested hierarchy of subspaces. To the best of our knowledge, this is the first online algorithm that optimizes over flag manifolds with streaming data. We formally show that Grassmannian Recursive Algorithm for Tracking ( 29 ) is a special case of our algorithm when the true subspace dimension is known. When the dimension is unknown, we propose a dimensionality ensemble strategy inspired by ensemble learning ( 11 ) . Finally, the Flag Recursive Online Tracking algorithm is leveraged as a subspace model in data-driven simulation ( 26 ) to perform trajectory prediction for a time and complexity-varying unknown system. The framework is showcased on an AutoRegressive model with eXogenous inputs system simulation and is shown to outperform non-adaptive subspace prediction ( 16 ) and other adaptive benchmarks. The rest of the paper is organized as follows. Section 2 contains preliminaries on subspace geometry. We formulate the problem in Section 3 . Section 4 introduces the online subspace learning algorithm. Section 5 contains numerical studies. Section 6 concludes the paper. 2 Preliminaries We use \mathbb{Z}_{\geq 0} to denote non-negative integers. Given i,j\in\mathbb{Z}_{\geq 0} , with i\leq j , we use [i,j] to denote the discrete interval [i,j]\cap\mathbb{Z}_{\geq 0} . We use I_{p\times m} to denote the p\times m matrix whose entries are 1 on the main diagonal (i.e., (i,i) -entry for i=1,\dots,\min(p,m) ), and all other entries are 0 . The diagonal matrix with diagonal elements \gamma_{1},\cdots,\gamma_{d} is denoted \diag(\gamma_{1},\cdots,\gamma_{d}) . Given a function v\colon\mathbb{Z}_{\geq 0}\to\mathbb{R}^{q} and i,j\in\mathbb{Z}_{\geq 0} , with i\leq j , we define v_{[i,j]}=(v(i),v(i+1),\dots,v(j))\in\mathbb{R}^{q(j-i+1)} . We use boldface \mathbf{U} to denote a subspace of \mathbb{R}^{p} and plain U to denote a matrix such that \mathrm{im~}U=\mathbf{U} . The orthogonal projector onto the subspace \mathbf{U} is denoted by \Pi_{\mathbf{U}} , and is used interchangeably with \Pi_{U}=UU^{\top} , where U is an orthonormal matrix whose columns span \mathbf{U} . The Moore–Penrose inverse of a matrix A is denoted A^{\dagger} . The Gaussian distribution with mean 0 and covariance \Sigma\in\mathbb{R}^{p\times p} is denoted \mathcal{N}(0,\Sigma) . 2.1 Subspace geometry A Grassmannian is the set of all subspaces in \mathbb{R}^{p} with dimension d denoted by \mathrm{Gr}(p,d)\coloneqq\{\mathbf{U} a linear subspace of \mathbb{R}^{p}:\dim\mathbf{U}=d\}. Flag manifolds generalize the Grassmannians by considering sequences of nested subspaces of increasing dimension. Let p\geq 2 be an integer and q_{1:d}\coloneqq(q_{1},\dots,q_{d}) with 0<q_{1}<\cdots<q_{d}<p. A flag of signature (p,q_{1:d}) is a nested sequence of linear subspaces \{\mathbf{U}_{i}\}_{i=1}^{d} of \mathbb{R}^{p} satisfying \{0\}\subset\mathbf{U}_{1}\subset\cdots\subset\mathbf{U}_{d}\subset\mathbb{R}^{p} with \dim\mathbf{U}_{j}=q_{j},j=1,\dots,d . We denote this flag by \mathbf{U}_{1:d} . The set of all such flags forms a smooth manifold \mathrm{Flag}(p,q_{1:d}) ( 40 ) . We also denote q_{0}=0,q=q_{d},q_{d+1}=p . We can represent flags as points on the Stiefel manifold \mathrm{St}(p,q)=\{U\in\mathbb{R}^{p\times q}:U^{\top}U=I\} consisting of orthonormal matrices. Define for each k=1,\dots,d , U_{k}\in\mathrm{St}(p,q_{k}-q_{k-1}) such that [U_{1}\cdots U_{k}] is an orthonormal basis of \mathbf{U}_{k} . We must note that under these notations, \mathbf{U}_{k} is not the subspace spanned by U_{k} . Then U_{1:d}\coloneqq[U_{1}\cdots U_{d}] is a representative of \mathbf{U}_{1:d} , called Stiefel representative. Throughout this paper we will abuse the notation to write U_{1:d}\in\mathrm{Flag}(p,q_{1:d}) . If the signature is (p,q_{1}) , then \mathrm{Flag}(p,(q_{1}))=\mathrm{Gr}(p,q_{1}) . The chordal distance of two subspaces \mathbf{A}\in\mathrm{Flag}(p,(m)),\mathbf{B}\in\mathrm{Flag}(p,(n)) is defined by \mathrm{d}(\mathbf{A},\mathbf{B})=(\sum_{i=1}^{\min(m,n)}\sin^{2}\theta_{i})^{1/2} ( 39 , Theorem 12) , where \theta_{i} is the i^{\text{th}} principal angle ( 18 ) between \mathbf{A} and \mathbf{B} . We present a simple example to illustrate the difference between flags and subspaces. Example 1 . Consider the canonical basis e_{1}=(1,0,0,0),e_{2}=(0,1,0,0),e_{3}=(0,0,1,0),e_{4}=(0,0,0,1) of \mathbb{R}^{4} . Let \mathbf{U}_{1}=\mathrm{span}\{e_{1}\},\mathbf{U}_{2}=\mathrm{span}\{e_{1},e_{2}\},\mathbf{U}_{3}=\mathrm{span}\{e_{1},e_{2},e_{3}\} , then \mathbf{U}_{1:3}\in\mathrm{Flag}(4,(1,2,3)) , and a Stiefel representation of this flag is the matrix U=\begin{bmatrix}e_{1}&e_{2}&e_{3}\end{bmatrix} . Note that even if two bases span the same subspace, the hierarchical structures of their flags may be different. Let V=\begin{bmatrix}e_{1}&e_{3}&e_{2}\end{bmatrix} , then clearly \mathrm{im~}V=\mathrm{im~}U . However, \mathbf{V}_{1}=\mathrm{span}\{e_{1}\},\mathbf{V}_{2}=\mathrm{span}\{e_{1},e_{3}\},\mathbf{V}_{3}=\mathrm{span}\{e_{1},e_{2},e_{3}\} , which implies \mathbf{V}_{2}\neq\mathbf{U}_{2} , hence \mathbf{V}_{1:d}\neq\mathbf{U}_{1:d} . 3 Problem Setup & Motivation The