pith. sign in

arxiv: 2511.06416 · v2 · submitted 2025-11-09 · 🧮 math.OC

Online Subspace Learning on Flag Manifolds for System Identification

Dian Jin , Jeremy Coulson This is my paper

Pith reviewed 2026-05-17 23:33 UTC · model grok-4.3

classification 🧮 math.OC
keywords online subspace learningflag manifoldssystem identificationtime-varying systemsdata-driven simulationnested subspacesrecursive tracking
0
0 comments X

The pith

An online algorithm on flag manifolds tracks nested subspaces from streaming data to identify time-varying systems of unknown and changing order.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a subspace learning method that works directly on flag manifolds to process incoming measurements recursively. It maintains an ensemble of nested subspaces at multiple dimensions simultaneously, so the representation can grow or shrink as the underlying dynamics change without any preset model order. The same learned subspaces feed into a data-driven predictor that simulates future behavior. In the reported case study the predictor adapts to sudden shifts in system behavior and exceeds several standard baselines.

Core claim

By lifting the recursive update from the Grassmann manifold to the flag manifold, the algorithm can track an entire chain of nested subspaces at once; each new data vector updates the whole chain in a single step, automatically selecting the most suitable dimension for the current regime and supplying the subspaces needed for subsequent simulation.

What carries the argument

The flag manifold of nested subspaces together with its recursive gradient update that generalizes the Grassmannian recursive tracking algorithm.

If this is right

  • Data-driven simulation becomes possible without a separate model-order selection step.
  • The same subspace chain supports prediction at every dimension the algorithm currently tracks.
  • Abrupt changes in dynamics are handled by the manifold geometry rather than by restarting the estimator.
  • The learned subspaces can be plugged directly into existing subspace-based control routines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same flag-manifold tracker could be tested on systems whose order changes continuously rather than in jumps.
  • Embedding the method inside a receding-horizon controller would test whether the multi-order representation improves closed-loop performance.
  • The recursive update might be combined with forgetting factors to emphasize recent data more strongly.

Load-bearing premise

The unknown time-varying system can be represented well enough by an evolving ensemble of nested subspaces that a recursive update on the flag manifold produces useful predictions.

What would settle it

A controlled experiment in which the true system order jumps outside the tracked range and the data-driven predictor's error remains higher than a fixed-order baseline after the jump.

Figures

Figures reproduced from arXiv: 2511.06416 by Dian Jin, Jeremy Coulson.

Figure 1
Figure 1. Figure 1: Average chordal distance for differ￾ent window sizes across 100 experiments [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Median cumulative prediction errors with interquartile whiskers: a single flag eval￾uated at different ranks (orange) versus indi￾vidually learned Grassmann subspaces (teal) [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: True and predicted trajectories. Left: no learning baseline. Right: the ensemble- [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Median cumulative prediction error (y-axis) versus varying NSR (x-axis) [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes an online subspace learning algorithm on flag manifolds that recursively tracks nested subspaces from streaming data for time-varying systems with unknown or changing model order, generalizing the Grassmannian recursive tracking method and integrating the models into a data-driven simulation framework for prediction, with a case study showing outperformance on abrupt changes.

Significance. If the generalization and recursive updates are rigorously shown to recover the Grassmannian case while preserving nested structure, this could enable adaptive data-driven control for systems with varying complexity without prior model order knowledge, addressing a practical limitation in subspace-based methods.

major comments (2)
  1. §3, Algorithm 1 and the generalization claim: the recursive update on the flag manifold must be shown to reduce exactly to the Grassmannian Recursive Algorithm for Tracking when restricted to a single subspace; without an explicit reduction check or proof, the central generalization claim lacks verification.
  2. §4, case study: the abstract and results claim outperformance on abrupt changes, but no derivation details, error analysis, baseline specifications, or quantitative metrics (e.g., prediction error norms) are provided to support the claim.
minor comments (2)
  1. Abstract: the final sentence appears truncated and should be completed to fully summarize the contributions and results.
  2. §2: the transition from the Grassmannian background to the flag manifold extension would benefit from a brief recap of the prior method's assumptions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We address each major comment point by point below, indicating the revisions we will make to strengthen the presentation.

read point-by-point responses

  1. Referee: §3, Algorithm 1 and the generalization claim: the recursive update on the flag manifold must be shown to reduce exactly to the Grassmannian Recursive Algorithm for Tracking when restricted to a single subspace; without an explicit reduction check or proof, the central generalization claim lacks verification.

    Authors: We agree that the manuscript would benefit from an explicit verification of the reduction. While the text states that the algorithm generalizes the Grassmannian recursive tracking method, a detailed algebraic check confirming exact recovery when the flag manifold is restricted to a single subspace is not included. In the revised manuscript we will add this verification, either as a dedicated paragraph in Section 3 or as an appendix, by substituting the appropriate flag parameters (k=1) and showing that the update equations and projection steps coincide with the Grassmannian case. revision: yes

  2. Referee: §4, case study: the abstract and results claim outperformance on abrupt changes, but no derivation details, error analysis, baseline specifications, or quantitative metrics (e.g., prediction error norms) are provided to support the claim.

    Authors: We acknowledge that the case-study section requires additional supporting material. The current version reports qualitative success on abrupt changes and states that the method outperforms several baselines, yet it does not supply the requested derivation details, error analysis, baseline parameter settings, or quantitative metrics such as prediction-error norms. In the revision we will expand Section 4 with these elements, including explicit baseline descriptions, time-series plots of prediction error norms, and a brief error analysis to substantiate the performance claims. revision: yes

Circularity Check

0 steps flagged

No circularity: generalization claim is independent extension of prior Grassmannian recursion

full rationale

The paper presents an online subspace tracking algorithm on flag manifolds as a direct generalization of the existing Grassmannian Recursive Algorithm for Tracking. The abstract and description state that the recursive update on nested subspaces adapts to unknown/varying model order and integrates into data-driven simulation. No equations or claims reduce by construction to fitted parameters, self-definitions, or load-bearing self-citations; the derivation chain remains self-contained against external benchmarks for subspace identification. The reader's assessment of score 2.0 aligns with a minor self-citation that is not load-bearing for the central result.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. The central claim rests on the domain assumption that flag manifolds provide a suitable geometry for representing and updating ensembles of nested subspaces that capture time-varying linear dynamics.

axioms (1)
  • domain assumption Flag manifolds can represent ensembles of nested subspaces suitable for modeling systems with varying dimensions.
    This assumption underpins the choice of manifold and the recursive tracking procedure described in the abstract.

pith-pipeline@v0.9.0 · 5435 in / 1199 out tokens · 40864 ms · 2026-05-17T23:33:35.462629+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

  1. [1]

    Gene H Golub and Charles F Van Loan.Matrix computations

    doi: 10.1109/LRA.2022.3184007. Gene H Golub and Charles F Van Loan.Matrix computations. JHU press,

    work page doi:10.1109/lra.2022.3184007 2022

  2. [2]

    Data-driven simulation and control.International Journal of Control, 81(12):1946–1959,

    Ivan Markovsky and Paolo Rapisarda. Data-driven simulation and control.International Journal of Control, 81(12):1946–1959,

    work page 1946

  3. [3]

    org/abs/2412.09052

    URLhttps://arxiv. org/abs/2412.09052. 12 ONLINESUBSPACELEARNING ONFLAGMANIFOLDS FORSYSTEMIDENTIFICATION Tom Szwagier and Xavier Pennec. Nested subspace learning with flags,

    work page arXiv

  4. [4]

    James Townsend, Niklas Koep, and Sebastian Weichwald

    URLhttps: //arxiv.org/abs/2502.06022. James Townsend, Niklas Koep, and Sebastian Weichwald. Pymanopt: A python toolbox for opti- mization on manifolds using automatic differentiation.Journal of Machine Learning Research, 17(137):1–5,

    work page arXiv