Autoregressive Processes on Stiefel and Grassmann Manifolds Journal: JOURNALTOFIND
Jordi-Lluís Figueras, Aron Persson · 2025
Read Autoregressive Processes on Stiefel and Grassmann Manifolds Journal: JOURNALTOFIND by Jordi-Lluís Figueras, Aron Persson (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-2509-24767.txt.
Introduction
Jordi-Lluís Figueras and Aron Persson extend a classic time series tool to data that live on curved spaces. Their 2025 paper treats autoregressive processes, where each new value builds on past ones, on Stiefel and Grassmann manifolds. These manifolds collect sets of perpendicular frames and flat subspaces, shapes that arise in vision and signal work.
The authors model the system's moving parts as members of the rotation group. They show the system can be estimated by averaging over observations, a simple rule with firm footing. A conjugate gradient algorithm, reshaped to respect the curved geometry, computes the estimates. The work gives system identification a route for data whose state is a frame or a subspace, not a plain vector.
Catalog record
| CREATOR | Jordi-Lluís Figueras, Aron Persson |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-24767
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Autoregressive Processes on Stiefel and Grassmann Manifolds Journal: JOURNALTOFIND Jordi-Lluís Figueras, Aron Persson Abstract System identification of autoregressive processes on Stiefel and Grassmann manifolds are presented and studied. We define the system parameters as elements in the orthogonal group and we show that the system can be estimated by averaging over observations. Then we propose an algorithm on how to compute these system parameters using conjugate gradient descent on Stiefel and Grassmann manifolds, respectively. 2020 MSC 62R30 , 53Z50 1 Introduction In modern signal processing and MIMO systems, a state is described by a general n by k matrix. But it is only the orthogonal part of the state that is measured, i.e. one measures the orthogonal part of the polar decomposition of a n by k matrix. In other words the measurements are points on Grassmann or Stiefel manifolds, depending on application and context, see for example [ 1 , 2 , 3 , 4 , 5 , 6 ] . Now consider a sequence of measurements over time, a time-series that takes values on a Grassmann or Stiefel manifold. If one now wants to make a prediction of future states or analyse the dynamics of such a time series, we have to make a suitable model for the time series. In \mathbb{R} -valued statistics, the most basic model used in time series analysis are the autoregressive processes. This then begs the question: What is a natural definition of an autoregressive process on Stiefel and Grassmann manifolds? Moreover, how can one estimate the parameters of the model in a reasonable manner? In [ 7 ] , Xavier and Manton introduced some AR-processes on Riemannian manifolds, a treatise we draw inspiration from and aim to specialise to the class of homogeneous spaces. The goal is to enable us to estimate the system parameters on Grassmann and Stiefel manifolds. Since, by definition, a homogeneous space has a transitive Lie-group action, this naturally guides us to consider system parameters as elements in the Lie group. In Proposition 3.1 , we show that one may model an AR( N )-processes on Stiefel and Grassmann manifolds by considering the system parameter as an element in the orthogonal group, see Equation ( 2 ). Moreover, by Theorem 3.6 we show that the system parameter can be estimated, and numerically computed, by averaging out over the observations. Using the definition for the expectation on manifolds, we propose a conjugate gradient descent method in Algorithm 1 , which utilise an approximation of the distance function for the Grassmann and Stiefel manifolds. This paper is relevant to anyone doing directional statistics, the most common form is that of directional statistics on the spheres \mathbb{S}^{n} . The spheres are special cases of the Stiefel manifolds, more precisely \mathbb{S}^{n}\cong\St_{n+1,1} . Ever since Fischer in [ 8 ] popularized the idea of modelling data on spheres in order to ascertain geological information on rocks, see also [ 9 ] , there has been an abundance of applications of directional statistics. Directional statistics has found its way into meteorology [ 10 , 11 , 12 , 13 ] , into medicine [ 14 , 15 ] and into machine learning [ 16 ] , to name a few. We start, in Section 2 , by giving an overview on what has been done previously in the area of autoregressive processes on Riemannian manifolds, and general statistics on manifolds. Then, in Section 3 , we specialise to homogeneous spaces and we give some new results. Thereafter, in Section 4 , we give some algorithms that estimate and approximate the system parameters on the Stiefel and Grassmann manifold using the conjugate gradient method. A fundamental introduction to Riemannian geometry is given in Appendix A that introduces concepts needed in order to understand the reasoning in this paper. In the second part of the appendix, Appendix B , one can find some figures showing simulations for a select few Stiefel and Grassmann manifolds illustrating the algorithms given in Section 4 . 2 Statistics on Riemannian Manifolds Throughout, let (\Omega,\mathcal{A},\mathbb{P}) be a probability space. Let M be a Riemannian manifold. We shall recall a slightly modified, but still equivalent, version of the "wrapped process" definition of an AR-process found in [ 7 ] . Denote \exp_{p} and \log_{p} to be the Riemannian exponential map centred at p and its inverse, respectively, see Definition A.2 . Moreover, the parallel transport from p\in M to q\in M along a geodesic from p to q is written as P_{p,q}:T_{p}M\rightarrow T_{q}M . Definition 2.1 . Let M be a connected and geodesically complete Riemannian manifold, and let Z_{\ell}\in M , \ell\in\mathbb{Z} be a time-series. Let N\in\mathbb{N} , and suppose there are N (1,1)-tensor fields A_{j}:TM\rightarrow TM , i.e. for every p\in M it holds that A_{j}\mid_{T_{p}M}:T_{p}M\rightarrow T_{p}M is linear. Then, an AR( N )-process on M is defined as Z_{\ell}=\exp_{Z_{\ell-1}}\left(\varepsilon_{\ell}+\sum_{j=1}^{N}A_{j}P_{Z_{\ell-j},Z_{\ell-1}}(\log_{Z_{\ell-j}}(Z_{\ell-j-1}))\right), where \varepsilon_{\ell}:\Omega\rightarrow T_{Z_{\ell-1}}M is a random variable with \mathbb{E}[\varepsilon_{\ell}]=0 , \Var(\varepsilon_{\ell})=\id_{T_{Z_{\ell-1}}M} . For our purposes we need the notion of expected value of a Riemann manifold-valued random variable. This is precisely the well known least square miminiser on \mathbb{R}^{n} adapted and generalized to Riemannian manifolds. Suppose \dist is the Riemannian distance function on M and \Vol is the Riemannian volume measure on M . Following [ 17 , 18 ] , the expectation of an absolutely continuous random variable X:(\Omega,\mathcal{F},\mathbb{P})\rightarrow M is then any point y\in M satisfying \mathbb{E}[X]=\arginf_{y\in M}\int_{M}\dist^{2}(y,x)p_{X}(x)\dVol(x). Here p_{X}=\frac{\differential X_{*}\mathbb{P}}{\dVol} is the Radon-Nikodyn derivative of the push forward measure, defined by X_{*}\mathbb{P}(A)=\mathbb{P}(X^{-1}(A)) for any Borel subset A\subseteq M , with respect to the Riemannian volume measure. Note that in contrast to random variables on \mathbb{R}^{n} , the expected value need not be uniquely defined. This is clearly illustrated by considering a uniformly distributed random variable on \mathbb{S}^{2} , i.e. constant probability density function, then any, and thus all, points on \mathbb{S}^{2} miminise the integral above. 3 AR-Processes on Homogeneous Spaces by Lie Group Action Let G be a compact Lie group with bi-invariant Riemannian metric, and let M be a geodesically complete G -homogeneous space with metric induced by G . See Appendix A for the definition. Proposition 3.1 . Suppose Z_{\ell} is an AR( N )-process on M as in Definition 2.1 with system parameters A_{j} . Then, for every 1\leq j\leq N and any pair of fixed points Z_{\ell-j},Z_{\ell-j-1}\in M there is a \Phi_{j}\in G such that \Phi_{j}Z_{\ell-j}=\exp_{Z_{\ell-j}}(A_{j}\log_{Z_{\ell-j}}(Z_{\ell-j-1})). (1) Proof. Denote q=\exp_{Z_{\ell-j}}(A_{j}\log_{Z_{\ell-j}}(Z_{\ell-j-1})), and \pi:G\rightarrow M=G/H as the quotient projection from G to M . Note that since \pi is defining the Lie-group action on M , we have \Phi_{j}Z_{\ell-j}=\Phi_{j}\pi(\Psi)=\pi(\Phi_{j}\Psi) where \Phi_{j},\Psi\in G and \Psi is a representative for Z_{\ell-j} . Suppose \xi\in G is such that \pi(\xi)=q then taking \Phi_{j}=\Psi\xi^{-1} is a Lie group element such that the equality in ( 1 ) holds true. ∎ Note that the above may also be concluded by observing that the Lie group action on homogeneous spaces is transitive. Remark . One important reason one may want to consider system parameters as Lie group elements is that for Stiefel and Grassmann manifolds the degrees of freedom for finding parameters \Phi\in O(n) is much smaller than the degrees of freedom for finding tensor fields A_{j}\in\mathcal{T}^{1}_{1}(M) so they will in