Most of modern machine learning theory is built on a convenient fiction: the assumption that data points are independent and identically distributed (i.i.d.). In this world, every sample is a fresh roll of the dice, unrelated to the one before it.
But for a robot navigating a room, a sensor monitoring a power grid, or a model forecasting the climate, this assumption is fundamentally broken. In these dynamical systems, the state at time $t+1$ is inextricably linked to the state at time $t$. This temporal correlation creates a "theoretical gap" where standard generalization proofs—which rely on the law of large numbers for independent variables—simply collapse.
Researchers are now bridging this gap, developing a rigorous statistical foundation for learning from dependent, non-stationary trajectories.
The Problem of Temporal Correlation
In a static dataset, like a collection of cat photos, the order of the images doesn't matter. In a dynamical system, the order is everything. As Anastasios Tsiamis, Nikolai Matni, and George J. Pappas point out in Statistical Learning Theory for Control: A Finite Sample Perspective, this feedback loop means that errors can compound. If a model's prediction is slightly off, the next state it observes is also "off," potentially leading to a divergent spiral where the model encounters data it never saw during training.
To provide guarantees for these systems, we cannot treat a trajectory of 1,000 steps as 1,000 independent data points. Instead, we must treat it as a single, highly correlated observation.
Learning from a Single Trajectory
If we only have one continuous path (a "single trajectory"), how can we ever hope to generalize? The key intuition lies in mixing coefficients and ergodicity.
If a system "mixes" well, it means that the influence of the past eventually fades. Two states observed far apart in time become effectively independent. Vladimir Kostic, Lorenzo Rosasco, and colleagues use this to provide statistical guarantees for Learning Dynamical Systems via Koopman Operator Regression in Reproducing Kernel Hilbert Spaces. By framing the Koopman operator—which lifts nonlinear dynamics into an infinite-dimensional linear space—as a regression task in an RKHS, they show that we can learn the underlying physics of a system from a single trajectory, provided we account for these temporal dependencies.
Similarly, Ingvar Ziemann and Nikolai Matni demonstrate in Single Trajectory Nonparametric Learning of Nonlinear Dynamics that we can achieve "optimal" learning speeds even in these correlated environments by combining information theory with dynamical stability.
Stability as a Generalization Bound
In deep learning, we often use "regularization" to prevent overfitting. In dynamical systems, stability plays a similar role.
A stable system is one where small perturbations in the input don't lead to wild swings in the output. Deividas Eringis, Mihály Petreczky, and their co-authors have pioneered the use of PAC-Bayesian bounds to analyze this. In PAC-Bayes Generalisation Bounds for Dynamical Systems Including Stable RNNs, they show that if a Recurrent Neural Network (RNN) is constrained to be stable, its generalization error doesn't explode as the sequence length increases.
By replacing the standard Kullback-Leibler (KL) divergence with Rényi divergence, as seen in PAC-Bayesian Error Bound, via Rényi Divergence, for a Class of Linear Time-Invariant State-Space Models, they can handle the "unbounded noise" typical of physical systems, providing a convergence rate of $O(1/\sqrt{N})$—the same rate we expect in the i.i.d. setting.
Beyond Linear Models: Transformers and Reservoirs
This theoretical push isn't limited to simple linear models. It is being extended to the most complex architectures in use today:
- Reservoir Computing: Lukas Gonon, Juan-Pablo Ortega, and Lyudmila Grigoryeva have provided the first risk bounds for reservoir computing, proving that these randomized recurrent systems are mathematically sound learners for complex time-series.
- Transformers: In Transformers and Their Roles as Time Series Foundation Models, Dennis Wu and Han Liu explain why pretraining on diverse time-series works. They argue that Transformers implicitly perform gradient descent "in-context" to fit autoregressive models, and that this process generalizes even with non-i.i.d. data if the system satisfies stability conditions like Dobrushin’s uniqueness.
The Double Descent of Dynamics
Finally, this research direction is re-examining classic engineering wisdom. Traditional system identification (SI) relies on "Occam’s Razor"—keeping models as simple as possible. However, Antonio H. Ribeiro shows in Beyond Occam's Razor in System Identification that the double-descent phenomenon exists in dynamics too.
Even when a model has enough parameters to perfectly interpolate the training trajectory, adding more parameters can actually improve generalization. This suggests that the massive overparameterization of deep learning isn't just a hack—it's a valid way to identify the complex, hidden patterns of the physical world, provided we have the right theoretical tools to understand why it works.