Spectral Filtering for Linear Dynamical Systems

15 concepts9 papers1 proposal

When modeling sequences or controlling physical systems, we often rely on Linear Dynamical Systems (LDS) or State Space Models (SSMs). These frameworks assume that an unobserved hidden state evolves over time, mapping past inputs to future outputs. However, learning the parameters of these systems from data is notoriously difficult. Because the hidden transitions interact multiplicatively over time, the optimization landscape is fundamentally non-convex. When a system possesses long-term memory, gradients easily vanish or explode, forcing traditional algorithms like Expectation-Maximization to take exponential time or get trapped in local minima.

A powerful line of research bypasses this non-convexity entirely through a technique known as spectral filtering. Instead of trying to learn the hidden transition matrices, spectral filtering projects the history of the system onto a fixed, mathematically grounded basis derived from a Hankel matrix. This approach transforms a seemingly intractable hidden-state estimation problem into a simple, convex linear regression task.

The Intuition: Hankel Matrix Spectral Decay

To understand why spectral filtering works, we must look at the system from an input-output perspective rather than tracking its internal states. The relationship between past inputs and future outputs in a stable linear system can be organized into a block-structured operator called a Hankel matrix, where each entry represents the system's response to an impulse at a given time lag.

The core mathematical insight driving this field is that the singular values (the spectrum) of a stable system's Hankel matrix decay exponentially fast.

As demonstrated by Elad Hazan et al. in Learning Linear Dynamical Systems via Spectral Filtering, this exponential decay has a profound implication: the infinite-dimensional history of a stable dynamical system effectively lives in a very low-dimensional subspace. Instead of needing a massive or infinitely long lookback window to capture long-term dependencies, a small, polylogarithmic number of basis functions—derived from the top eigenvectors of a universal Hankel matrix—is sufficient to approximate the system's dynamics with near-perfect precision.

Bypassing Recurrence with Fixed Filters

Instead of treating the history-compressing filters as parameters to be learned via backpropagation, spectral filtering uses a set of fixed convolutional filters. These filters are pre-computed from the eigenvectors of the Hankel matrix, which is constructed solely based on a known or assumed stability margin of the system.

Past Inputs ---> [ Fixed Hankel Filter 1 ] ---> Feature 1 \
            ---> [ Fixed Hankel Filter 2 ] ---> Feature 2 ---> [ Learned Linear Layer ] ---> Prediction
            ---> [ Fixed Hankel Filter K ] ---> Feature K /

By convolving the input history with these fixed spectral filters, the system's past is compressed into a compact feature vector. Because the filters themselves are fixed and theoretically guaranteed to span the space of stable dynamics, the learning algorithm only needs to find the optimal linear combination of these features to predict the next output. This overparameterization trick completely eliminates the non-convexity of system identification.

The authors extended this foundational mechanism in Spectral Filtering for General Linear Dynamical Systems by introducing "phased wave-filtering," providing the first polynomial-time algorithm capable of predicting general, asymmetric linear systems without restrictive assumptions. Similarly, Paria Rashidinejad et al. utilized this spectral compression in their SLIP (Spectral LDS Improper Predictor) algorithm (SLIP: Learning to predict in unknown dynamical systems with long-term memory), proving that one can achieve logarithmic regret—matching the performance of an optimal Kalman filter—without ever explicitly reconstructing the latent state.

From Prediction to Provable Control

The benefits of spectral filtering extend directly from prediction to optimal control, particularly in environments featuring unpredictable or adversarial disturbances.

Sanjeev Arora et al. first bridged this gap in Towards Provable Control for Unknown Linear Dynamical Systems, showing that the spectral basis could be used to maintain stable control guarantees even when the underlying physics of the system are entirely unknown.

A persistent challenge in classical control, however, is that algorithms slow down dramatically as a system approaches its "stability margin" (becoming highly sensitive to perturbations). To resolve this computational bottleneck, Anand Paresh Brahmbhatt et al. introduced Online Spectral Control (OSC) in A New Approach to Controlling Linear Dynamical Systems. By constructing a universal Hankel matrix tuned to the system's stability margin, OSC isolates the most critical components of past disturbances. The authors expanded this into Double Spectral Control (DSC) in Efficient Spectral Control of Partially Observed Linear Dynamical Systems, proving that stable linear controllers can be compactly represented via this Hankel basis, yielding an exponential runtime speedup over prior state-of-the-art methods while maintaining robust performance guarantees.

Scaling to Deep Sequence Models

The exact same mathematical guarantees that stabilize control loops have recently been adapted to solve the training instabilities of deep learning architectures designed for long-context sequence modeling. Modern State Space Models (such as S4 or Mamba) typically require intricate, highly sensitive initialization schemes to prevent gradients from vanishing over thousands of tokens.

In Spectral State Space Models, Naman Agarwal et al. proposed replacing learned recurrence parameters entirely with fixed convolutional filters derived from Hankel matrices. Because this fixed basis naturally captures long-horizon features independently of the system's spectral gap, the architecture remains inherently stable during training, bypassing the need for specialized normalization or initialization tricks.

To scale this concept to large-scale language and robotics tasks, Y. Isabel Liu et al. developed Flash STU (Spectral Transform Units) in Flash STU: Fast Spectral Transform Units. By pairing fixed spectral filters with a sliding window attention mechanism, Flash STU achieves near-linear scaling and a significantly smoother optimization landscape than traditional recurrent models or Transformers.

Finally, this spectral framework provides a rigorous solution to one of deep learning's most stubborn empirical challenges: length generalization. In Provable Length Generalization in Sequence Prediction via Spectral Filtering, Annie Marsden et al. proved that because spectral filtering projects sequences into a stable subspace defined by the system's underlying impulse response, models trained on short sequences possess the mathematical foundation required to naturally extrapolate to much longer contexts at test time.

Key Foundational Work

Learning Linear Dynamical Systems via Spectral Filtering (2017) — contributed 3 concept elements to this direction.
Spectral Filtering for General Linear Dynamical Systems (2018) — contributed 1 concept elements to this direction.
Towards Provable Control for Unknown Linear Dynamical Systems (2018) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 9 papers
  2. Version 1 · 2026-10-04 · 20 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_fixed_spectral_filtering_bbf27417,
  title = {fixed spectral filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fixed-spectral-filtering/rdr_bbf27417cc924535b17bffed47db0399},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_fixed_spectral_filtering_bbf27417_v2,
  title = {fixed spectral filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fixed-spectral-filtering/rdr_bbf27417cc924535b17bffed47db0399/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake