Beyond Deterministic Dynamics in Continuous Time Machine Learning

34 concepts33 papers1 proposal

Imagine tracking a single particle drifting through a fluid. If the fluid is perfectly still, the particle's trajectory follows a smooth, predictable path dictated by an Ordinary Differential Equation (ODE), $\frac{\mathrm{d}x}{\mathrm{d}t} = f(x)$. If your camera shakes while recording, the resulting data contains measurement noise. This type of noise is external and superficial; the underlying physical path remains smooth and deterministic, and standard filtering techniques can easily recover it.

Now imagine the fluid is turbulent. The chaotic eddies do not just disrupt your measurements—they actively kick the particle in random directions at every fraction of a second. This is process noise. The underlying physical path itself becomes jagged, probabilistic, and non-differentiable.

For years, continuous-time machine learning frameworks—such as Neural ODEs, continuous-time reinforcement learning, and physics-informed neural networks—have operated under the deterministic dynamics assumption. They treat the real world as a collection of smooth ODEs, assuming all real-world uncertainty can be flattened into simple measurement noise. However, when applied to chaotic or inherently random environments, this assumption forces models to invent highly complex, "wiggly" trajectories to fit data that is fundamentally stochastic.

To build robust models of the physical world, researchers are shifting from deterministic differential equations to stochastic ones.

The Failure Modes of Smooth Assumptions

When a model assumes a deterministic universe, it struggles to handle the aleatoric uncertainty of real-world systems. This limitation manifests across several domains:

  • Fluid Dynamics and Climate Modeling: Natural systems are inherently chaotic. Emmanuel de Bezenac et al. demonstrate that deterministic architectures struggle to capture the true distribution of fluid flows. Similarly, JeeHyun Hwang et al. note that relying on deterministic neural networks to model climate residuals fails to capture the true probabilistic nature of weather patterns, necessitating a transition to stochastic differential equations for reliable forecasting.
  • Continuous-Time Reinforcement Learning: In environments governed by continuous physical laws, an agent must plan its actions across continuous time. Algorithms like OCoRL, introduced by Lenart Treven et al., optimize when to measure a system based on deterministic assumptions. However, as Klemens Iten et al. point out, if the environment contains true process noise rather than just sensor noise, these deterministic exploration strategies fail to account for environmental volatility.
  • Time-Series and Latent Trajectories: Models that compress sequential data into continuous latent spaces often force all variations into a single initial state or a rigid, deterministic transition function. Ilze Amanda Auzina et al. and Qiyu Kang et al. argue that this deterministic constraint prevents networks from properly representing real-world data fluctuations. When modeling the noisy residual components of sequential data, Seonkyu Lim et al. observe that standard linear ODEs are insufficient because these residuals are dominated by random noise rather than deterministic trends.

As J. Nathan Kutz et al. summarize, traditional system identification frameworks are built to filter out sensor errors, but they remain blind to internal process noise—the environmental fluctuations that alter the system's underlying dynamics.

Injecting Randomness into the Derivative

To overcome this limitation, the underlying mathematics must change. Instead of modeling a system as an ODE, researchers are leveraging Stochastic Differential Equations (SDEs) and Stochastic Partial Differential Equations (SPDEs).

Mathematically, this shifts the formulation from a standard derivative to a relation involving a random driving process (like Brownian motion $W_t$):

$$\mathrm{d}X_t = f(X_t, t)\mathrm{d}t + g(X_t, t)\mathrm{d}W_t$$

Here, $f(X_t, t)$ represents the deterministic drift (the expected trend), while $g(X_t, t)$ represents the diffusion (the magnitude of the process noise). By learning both functions simultaneously, a neural network can model both the predictable physics and the underlying randomness.

For example, to capture complex spatiotemporal dynamics under external random forces, Cristopher Salvi et al. introduced Neural Stochastic PDEs (NSPDEs), which allow neural operators to handle time-dependent random noise. In mechanical systems, Tapas Tripura et al. developed a data-driven framework to discover interpretable physical Lagrangians specifically from stochastically excited environments, separating structural mechanics from random environmental forcing.

Deterministic ODE:  [State X_t] ---> ( Smooth Drift f(X) ) ---> [Next State X_t+dt]

Stochastic SDE:     [State X_t] ---> ( Smooth Drift f(X) ) ---\
                                                               ---> [Probability Distribution of X_t+dt]
                                     ( Random Noise g(X) ) ---/

This architectural shift fundamentally changes how models evaluate future states. Instead of predicting a single trajectory, the model tracks a evolving probability distribution. In reinforcement learning, this transition requires moving away from simple value updates toward solving complex second-order partial differential equations—such as the Hamilton-Jacobi-Bellman-Fokker-Planck equation—to compute optimal policies under explicit uncertainty, as detailed by Xuefeng Wang et al..

Rebuilding Theoretical Foundations

Abandoning the deterministic dynamics assumption requires rewriting the theoretical guarantees that underpin continuous-time machine learning.

Generalization and Optimization

How do we know a continuous-time model will perform well on unseen data? For deterministic systems, Pierre Marion (Pierre Marion) proved that the generalization error of Neural ODEs depends on the smoothness of their weight transitions over time. Similarly, Linus Bleistein et al. established statistical bounds for Neural Controlled Differential Equations based on sampling coarseness. However, both works note that these guarantees are restricted to deterministic paths; extending them to the jagged, non-differentiable paths of SDEs remains an open challenge.

Furthermore, optimization techniques that make deterministic models practical do not seamlessly translate to stochastic settings. For instance, Jacob Jin Kelly et al. developed higher-order regularization terms to prevent Neural ODE trajectories from becoming too complex and computationally expensive to solve, but adapting these constraints to handle the inherent noise of SDEs requires entirely new optimization paradigms.

Causality and System Identification

If a system's dynamics are driven by random fluctuations, identifying the true underlying equations or causal structures becomes significantly more difficult. Automated model discovery tools like PySINDy, developed by Alan A. Kaptanoglu et al., rely on sparse regression to find the simplest ODEs or PDEs matching a dataset, but must be adapted to isolate stochastic volatility functions.

From a causal perspective, Paul K. Rubenstein et al. established the mathematical links required to transform deterministic ODEs into structural causal models, but noted that real-world robustness requires accounting for random differential equations. When hidden variables confound the system, establishing conditions under which the true dynamics can be uniquely identified becomes a major hurdle. Xi Geng et al. solved this for linear ODE systems with hidden confounders, but extending these identifiability conditions to stochastic transitions is vital for reliable causal inference. This sentiment is echoed by Krzysztof Kacprzyk et al. in the context of longitudinal treatment effects, where failing to model process noise leads to degraded clinical predictions.

Even the formal verification of safety critical systems is impacted. André Platzer (André Platzer) proved that verifying continuous physical systems requires deductive principles like differential cuts, but notes that existing verification hierarchies assume perfect, noise-free differential equations.

The Path Forward

Moving beyond the deterministic dynamics assumption is not simply about making models more complex; it is about aligning model priors with physical reality. Whether transforming integral equations into stochastic systems (Zhiyuan Ren et al.) or embedding Gaussian Processes into Neural ODE architectures to handle scarce, noisy data (Mohamed Aziz Bhouri et al.), treating process noise as a fundamental component of the environment allows machine learning models to transition from brittle curve-fitting to true probabilistic reasoning in continuous time.

Key Foundational Work

Fast approximate Bayesian computation for estimating parameters in differential equations (2017) — contributed 1 concept elements to this direction.
The Structure of Differential Invariants and Differential Cut Elimination (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_deterministic_dynamics_assumption_049fefa9,
  title = {deterministic dynamics assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deterministic-dynamics-assumption/rdr_049fefa9ec934c939354cfc39659254a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_deterministic_dynamics_assumption_049fefa9_v1,
  title = {deterministic dynamics assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deterministic-dynamics-assumption/rdr_049fefa9ec934c939354cfc39659254a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake