In classical system identification, researchers heavily rely on a comforting assumption: strict stability. When a dynamical system is strictly stable, its eigenvalues lie strictly inside the unit circle. If you perturb it, the system eventually forgets its past and returns to equilibrium. Mathematically, this "fading memory" property creates a mixing environment where data collected along a single trajectory quickly begins to look like a collection of independent samples.
But many real-world systems—ranging from robotics and power grids to orbital mechanics—are explicitly designed to operate outside this safe regime. They may be unstable (states grow exponentially without control) or marginally stable (energy is preserved, and states neither decay nor explode, like a frictionless pendulum or a pure integrator).
Learning the parameters of these systems from data forces us to throw out traditional mixing-time arguments. It introduces a fascinating tension: while instability makes long-term simulation incredibly fragile, it counter-intuitively makes certain parts of the learning process much easier.
The Paradox of Explosive Growth
Intuitively, you might expect an unstable system to be impossible to learn from a single trajectory because its states rapidly blow up. However, Simchowitz et al. and Sarkar et al. uncovered a striking paradox: fully observed linear systems are often easier to identify when they are unstable.
To understand why, consider a simple autoregressive process $x_{t+1} = Ax_t + w_t$, where $w_t$ is bounded noise. If the matrix $A$ is unstable (spectral radius greater than one), the state $x_t$ grows exponentially over time. Because the noise $w_t$ remains bounded while the state $x_t$ explodes, the signal-to-noise ratio of the system skyrockets.
When we run Ordinary Least Squares (OLS) to estimate $A$, the estimator relies on the empirical covariance matrix $\sum x_t x_t^T$. In an unstable system, this matrix grows exponentially, acting as a massive statistical amplifier that rapidly drowns out the noise. As noted by Jadbabaie et al., this high signal-to-noise ratio allows algorithms to recover the underlying transition dynamics with sharp, non-asymptotic guarantees without needing the system to "mix" or forget its past.
The Edge of Stability: The Nightmare of Infinite Memory
If explosive systems amplify the signal, what happens at the exact boundary of stability? Marginally stable systems—such as ideal oscillators or integrators where eigenvalues lie exactly on the unit circle—present a unique statistical challenge.
In a marginally stable system, the state does not explode exponentially, but it never dampens either. The system retains a perfect memory of its initial conditions and past noise injections forever. As Bakshi et al. demonstrate, this lack of decay causes the variance of standard statistical estimators to grow unbounded as the observation trajectory increases. Because the data never becomes independent, standard regression techniques struggle to disentangle long-term temporal correlations from the true underlying parameters.
Strictly Stable Marginally Stable Unstable
(Fading Memory) (Infinite Memory) (Explosive Signal)
State State State
▲ ▲ ▲
│ ▄ │ ▄ ▄ │ ▄▀
│ █ █ │ █ █ █ █ │ ▄▀
│█ █▄▄ │ █ █ █ │ ▄█▀
└─────────► Time └───────────► Time └───────────► Time
Errors Decay Errors Persist Signal Amplified
The Hidden State and Non-Linearity Bottleneck
The statistical benefits of instability completely vanish when we lose the luxury of full observability or linearity.
- Exponential Uncertainty Growth: In a partially observed system, we only see a noisy projection of the hidden state ($y_t = Cx_t + v_t$). If the underlying dynamics are unstable, any minor error in our estimate of the current hidden state compounds exponentially at the next time step. Berkenkamp et al. show that if a learning algorithm relies purely on forward simulation without real-time measurement feedback, the simulated trajectory will rapidly diverge from reality, causing the learning process to collapse.
- The Exploration Risk: In reinforcement learning and adaptive control, trying to learn an unstable system from scratch is inherently dangerous. As Lale et al. point out, without a pre-existing stabilizing controller to keep the system bounded during the initial data-collection phase, the system can fail catastrophically before enough data is gathered to learn its parameters.
Algorithmic Solutions
To bypass these limitations, modern research focuses on designing estimators that can handle or explicitly constrain non-contractive dynamics.
1. Spectral and Prefiltered Filtering
Instead of running regression directly on raw, highly correlated trajectories, researchers use filtering to tame the data. Simchowitz et al. introduced Prefiltered Least Squares (PF-LS), which applies a coarse linear filter to the data before estimation. This filter effectively pre-whitens the trajectory, mitigating long-term correlations and allowing consistent identification of marginally stable Newtonian systems without suffering from exploding variance.
Similarly, Hazan et al. bypass the rapid-dissipation assumption entirely by using wave-filtering over the spectral properties of Hankel matrices, transforming a non-convex system identification problem into a well-behaved convex optimization task that remains robust even as the system approaches the limit of stability.
2. Inductive Stability Constraints
When the goal is to learn a model that will be used for long-term forecasting, we must ensure the learned parameters do not accidentally introduce artificial instability. Siddiqi et al. address this by formulating parameter estimation as a quadratic program with an iterative constraint generation loop. If a candidate model exhibits unstable behavior during simulation, the algorithm automatically generates and applies a new linear constraint to force the parameters back into a stable regime, ensuring realistic long-term simulations.
3. Subspace Robustness
Surprisingly, some classical algorithms possess inherent mathematical resilience to these boundary conditions. Tsiamis et al. provided a non-asymptotic analysis proving that stochastic subspace identification methods remain highly effective even when eigenvalues sit exactly on the unit circle. They showed that these algorithms recover the system's true dynamics and Kalman filter gains at the optimal $1/\sqrt{N}$ convergence rate, validating their use for real-world oscillators and integrators.
Key Foundational Work
A Constraint Generation Approach to Learning Stable Linear Dynamical Systems (2007) — contributed 1 concept elements to this direction.
Symbolic models for nonlinear control systems affected by disturbances (2008) — contributed 1 concept elements to this direction.