Stable Linear System Identification by Design

This direction is merged into another direction: generalization to nonlinear systems. Its last version stays readable and citable here.

Version 1 of 19 papers16 conceptsfrozen 2026-10-04

When we train a model to predict the future of a physical system—like a robotic arm or a chemical reactor—we usually care about accuracy first. But in the world of Linear Dynamical Systems (LDS), accuracy isn't enough. If a model is 99% accurate but has even one eigenvalue slightly outside the "stable" range, its predictions will eventually explode to infinity.

Traditional machine learning often ignores this. When we fit a linear model to noisy data, the optimizer frequently "overfits" the noise by creating an unstable model. Recent research has shifted from trying to fix these unstable models after the fact to building stability directly into the architecture of the learner.

The Problem with Post-Hoc Fixes

In classical system identification, if you end up with an unstable model, you might try to "squash" the eigenvalues back into the stable region. However, Loris Di Natale, Giancarlo Ferrari‐Trecate, and their colleagues point out in Stable Linear Subspace Identification: A Machine Learning Approach that these post-hoc corrections often degrade the model's predictive power. It’s like trying to fix a blurry photo by stretching it; you might meet the technical requirements, but you lose the underlying truth.

Furthermore, in high-stakes environments like fusion energy research, an unstable model isn't just a mathematical nuisance—it's a safety risk. As Zachary L. Daniel and Steven L. Brunton demonstrate in their work on helicity injectors, noisy experimental data naturally pushes standard regression techniques toward non-physical, divergent behavior.

Stability as a Geometric Constraint

The core insight of this research direction is to treat stability as a geometric constraint on the search space. Instead of searching through the space of all possible matrices and hoping to find a stable one, researchers define a "manifold" or a specific mathematical structure that only contains stable matrices.

Giorgos Mamakoukas and Todd D. Murphey introduced an algorithm called SOC in Memory-Efficient Learning of Stable Linear Dynamical Systems for Prediction and Control. SOC ensures stability at every single iteration of the optimization. By restricting the search to a manifold of stable matrices, the algorithm can be stopped at any time and still yield a physically valid model. Crucially, they reduced the memory cost of this process from $O(n^4)$ to $O(n^2)$, making it feasible for large-scale robotics.

Direct Parametrization: The "Sigmoid" for Stability

How do you actually force a matrix to be stable during gradient descent? The solution is direct parametrization.

Think of how a Sigmoid function maps any real number into the range $[0, 1]$. Researchers are developing similar "mappings" for matrices. In the SIMBa framework, Loris Di Natale and his team use Linear Matrix Inequalities (LMIs) to create a bridge. They map unconstrained "free" parameters—which a standard optimizer like PyTorch can easily handle—to a structured state-transition matrix that is guaranteed to be Schur stable (meaning all its eigenvalues stay within the unit circle).

This approach allows the use of standard backpropagation. The optimizer "thinks" it is doing unconstrained gradient descent, but the mathematical structure of the model ensures that every step it takes remains within the bounds of physical stability.

Beyond Deterministic Models

This "stability by design" philosophy is also moving into more complex domains:

  1. Uncertainty Modeling: Farhad Ghanipoor and Nathan van de Wouw show in Uncertainty Learning for LTI Systems with Stability Guarantees how to learn the "missing pieces" of a model (uncertainties) using Semi-Definite Programming (SDP). They use Lyapunov-based criteria to ensure that even when the model is learning what it doesn't know, it doesn't become unstable.
  2. Stochastic Systems: For systems influenced by random noise, John Leth and Mohamad Al Ahdab developed parametrizations based on the Stochastic Bounded-Real Lemma. By defining system matrices as functions of positive-definite and skew-symmetric matrices, they enable robust Bayesian inference that remains stable by construction.

Why It Works: The Intuition

The reason this direction is so successful is that it aligns the optimization objective (minimizing error) with the physical constraints (stability).

In traditional methods, these two forces are in conflict: the noise in the data pulls the model toward instability, and the researcher has to push back with regularization or post-processing. By using direct parametrization, the conflict disappears. The optimizer is free to chase the data as hard as it wants, but because the "map" it's walking on only exists in the territory of stable systems, it can never accidentally fall off the cliff into divergence.

The result, as seen in the SIMBa experiments, is often a massive leap in performance—sometimes reducing prediction error by over 25% simply because the model is no longer wasting its "capacity" trying to fit non-physical, unstable solutions.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_stable_linear_system_identification_by_d_e092af33,
  title = {Stable Linear System Identification by Design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stable-linear-system-identification-by-design/rdr_e092af33e447402abb2c3697f34ea02f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_stable_linear_system_identification_by_d_e092af33_v1,
  title = {Stable Linear System Identification by Design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stable-linear-system-identification-by-design/rdr_e092af33e447402abb2c3697f34ea02f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}