Subspace Tracking with Dynamical Models on the Grassmannian

Alex Saad-Falcon, Brighton Ancelin, Justin RombergIEEE Sensor Array and Multichannel Signal Processing (SAM) 2026 · OpenReview

TL;DR

This paper introduces a new method for tracking how linear subspaces change over time, particularly useful for dynamic signals like those from antenna arrays. By modeling subspace motion on the "Grassmannian" manifold with dynamical systems and using efficient "chordal" approximations, they achieve accurate tracking with significantly reduced computational cost compared to previous geodesic-based methods.

What Problem Does This Paper Address?

In many real-world scenarios, like signal processing with antenna arrays, the underlying structure of data isn't static but evolves over time. This structure can often be represented by a low-dimensional linear subspace within a high-dimensional space. "Subspace tracking" aims to estimate this evolving subspace from a stream of data. Traditional methods often assume the subspace is constant or change abruptly. More advanced techniques leverage the mathematical structure of the "Grassmannian manifold" (the space of all possible linear subspaces) to model subspace motion. However, these methods can be computationally expensive, especially when dealing with complex dynamics. This paper addresses the challenge of efficiently and accurately tracking these dynamic subspaces by introducing a framework that combines dynamical models on the Grassmannian with computationally efficient approximations.

Methods

The core of the paper is a "regularized least squares" (RLS) framework for subspace tracking. The problem is formulated as minimizing a loss function that balances two objectives:

  1. Data Fidelity: How well the estimated subspace fits the current batch of data.
  2. Model Regularization: How well the estimated subspace trajectory adheres to a chosen dynamical model of motion on the Grassmannian.

The authors propose two types of dynamical models:

  • Position Regularization: Assumes the subspace remains relatively static between batches.
  • Velocity Regularization: Models the subspace as having a "tangential velocity" on the manifold, allowing for smoother, continuous motion.

Crucially, they leverage approximations for operations on the Grassmannian. Instead of using the computationally intensive "geodesic" distance and logarithmic maps (which require Singular Value Decomposition - SVD), they employ "chordal" distances and projector differences. These approximations only require matrix multiplications, leading to significant speedups. They derive the necessary gradients for these regularizers to perform gradient descent on the manifold.

Datasets

The paper does not introduce new datasets. Instead, it uses simulated data to evaluate its proposed methods. The simulations involve:

  • Generating subspaces that follow a random geodesic on the Grassmannian with small tangential velocity.
  • Simulating a narrowband beamforming scenario with multiple emitters moving in front of an antenna array, where the underlying subspace dynamics are complex.

Key Insights and Ideas

  • Grassmannian as a Natural Framework: The Grassmannian provides a mathematically sound way to represent and reason about subspaces, naturally handling the invariance to basis choices.
  • Dynamical Models for Smoothness: Explicitly modeling subspace motion (position or velocity) as a dynamical system on the manifold is key to achieving smooth and accurate tracking, especially in the presence of noise.
  • Chordal Approximations are Powerful: The insight that chordal distance and projector differences can effectively approximate geodesic operations for regularization is a major contribution. This dramatically reduces computational cost without sacrificing tracking accuracy.
  • Trade-off Control: The regularization parameter ($\lambda$) allows for a tunable trade-off between fitting the current data and adhering to the learned dynamical model, enabling flexibility in different scenarios.

Key Results

  • Computational Speedup: The chordal approximation for position regularization achieved a 4.74x speedup in runtime compared to the geodesic version, while maintaining similar tracking performance.
  • Accurate Tracking: Both chordal and geodesic position RLS methods accurately track subspaces generated by a geodesic motion on the Grassmannian.
  • Capturing Complex Dynamics: Chordal velocity RLS successfully captured the complex, dynamic motion of multiple emitters in an antenna array scenario, outperforming simpler methods like batch 2D-SVD and fitting a single geodesic.
  • Performance at Batch Edges: The paper notes that tracking performance can degrade at the edges of data batches due to limited data.

Limitations and Future Work

  • Limitations:
    • Performance can suffer at the edges of data batches due to insufficient data.
  • The paper focuses on first-order velocity models; more complex acceleration-based models are mentioned as a potential extension but not fully explored due to computational intensity.
  • Future Work:
    • Extending the framework to handle more complex, higher-order dynamical models on the Grassmannian.
  • Investigating the impact of different manifold metrics and regularization strategies.
  • Exploring applications beyond beamforming, such as in computer vision or other signal processing domains where dynamic subspace estimation is relevant.

Summary updated

Abstract

Tracking signals in dynamic environments presents difficulties in both analysis and implementation. In this work, we expand on a class of subspace tracking algorithms which utilize the Grassmann manifold – the set of linear subspaces of a high-dimensional vector space. We design regularized least squares algorithms based on common manifold operations and intuitive dynamical models. We demonstrate the efficacy of the approach for a narrowband beamforming scenario, where the dynamics of multiple signals of interest are captured by motion on the Grassmannian.
BibTeX
@inproceedings{saadfalcon2026subspace,
  author = {Alex Saad-Falcon and Brighton Ancelin and Justin Romberg},
  title = {{Subspace Tracking with Dynamical Models on the Grassmannian}},
  booktitle = {IEEE Sensor Array and Multichannel Signal Processing (SAM)},
  year = {2026},
  url = {https://openreview.net/forum?id=6hUsuEgEpp},
}
Figure