Neural Models for Non-Markovian Dynamics

This direction is split into other directions: memoryless state transitions. Its last version stays readable and citable here.

Version 1 of 1173 papers180 conceptsfrozen 2026-10-04

Most of our tools for modeling the world rely on a convenient lie: the Markov property. This assumption posits that the future is independent of the past, given the present. In a Markovian world, if you know the current position and velocity of a particle, you know everything necessary to predict its next step. The history of how it got there is irrelevant.

However, as researchers like Yifeng Tian, Marian Anghel, and Daniel Livescu point out in Data-driven learning for the Mori-Zwanzig formalism: a generalization of the Koopman learning framework, this assumption often fails in practice. When we observe a complex system through a "reduced-order" lens—meaning we can't see every single variable—the ignored variables manifest as memory effects. The system appears non-Markovian because the "hidden" parts of the state are still influencing the parts we can see, tied to the system's history.

A new wave of research is moving beyond this "memoryless" constraint, developing neural architectures that explicitly learn and integrate historical context to predict future behavior.

The Mathematical Origin of Memory

To understand why memory is necessary, we can look to the Mori-Zwanzig (MZ) formalism from statistical mechanics. MZ proves that when you project a high-dimensional dynamical system onto a lower-dimensional space, the resulting equations must include a "memory kernel"—an integral over the system's past states.

Traditional Neural Ordinary Differential Equations (Neural ODEs) are strictly Markovian: the rate of change $dx/dt$ depends only on the current state $x(t)$. To fix this, Emanuele Zappala and Antonio Henrique de Oliveira Fonseca introduced Neural Integro-Differential Equations (NIDE). Instead of just looking at the present, NIDEs use an integral operator to "sweep" across the system's history, allowing the model to account for cumulative effects that a simple derivative would miss. This is crucial for systems like brain activity or disease spread, where the current state is a product of a long, continuous trajectory.

Learning What to Remember

If the future depends on the past, which parts of the past matter most? Not all history is equally relevant.

One approach is to use attention mechanisms to weight historical states. In Neural Fractional Attention Differential Equations, Qiyu Kang and colleagues replace the static weights found in classical fractional calculus with learnable attention. This allows the model to adaptively "focus" on specific past events that are predictive of the future, such as a sudden spike in a fluid's velocity or a specific historical trend in traffic patterns.

Similarly, in the realm of physical simulations, Liming Wu and Wenbing Huang developed Equivariant Spatio-Temporal Attentive Graph Networks (ESTAG). By modeling entire trajectories rather than frame-to-frame transitions, ESTAG can capture "hidden" dynamics—like the drag of an unobserved solvent on a protein—that a Markovian model would simply treat as random noise.

Transforming the Domain of Time

Sometimes, the best way to handle history isn't to look back in time, but to change the coordinate system entirely. Samuel Holt, Zhaozhi Qian, and Mihaela van der Schaar proposed Neural Laplace, which learns dynamics in the Laplace domain.

In the time domain, a delay equation (where the current change depends on the state five seconds ago) is difficult to solve because it requires keeping a buffer of the past. In the Laplace domain, these complex temporal dependencies and delays become simple algebraic relationships. By approximating the Laplace transform of a system's solution, Neural Laplace can model long-range history and abrupt discontinuities that would cause a standard Neural ODE to fail.

From White Noise to Colored Context

In stochastic systems, the Markov assumption often appears as "white noise"—random fluctuations that are completely uncorrelated in time. But real-world noise is often "colored," meaning the randomness itself has a memory.

Morgane Borreguero and Marco Ellero demonstrate in Investigation of dissipative particle dynamics with colored noise that accounting for these correlations is essential for simulating complex fluids. When modeling financial markets, Patryk Gierjatowicz and Lukasz Szpruch argue in Robust pricing and hedging via neural SDEs that path-dependent dynamics are necessary to capture the "sticky" nature of market volatility.

Why This Matters

The shift toward non-Markovian modeling represents a move from curve-fitting to system-understanding. Whether it is Matthijs T. J. Spaan using weighted historical scenarios to plan in uncertain markets (Planning Under Uncertainty with Weighted State Scenarios) or Ahmed Alaa using attention to track disease progression (Attentive State-Space Modeling of Disease Progression), the goal is the same: to acknowledge that the past is not just a sequence of events we've already passed, but a fundamental component of where we are going next.

Papers in this version

Showing 60 of 173 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_neural_models_for_non_markovian_dynamics_de953e25,
  title = {Neural Models for Non-Markovian Dynamics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-models-for-non-markovian-dynamics/rdr_de953e250598493fa73df3a4b771df4d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_neural_models_for_non_markovian_dynamics_de953e25_v1,
  title = {Neural Models for Non-Markovian Dynamics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-models-for-non-markovian-dynamics/rdr_de953e250598493fa73df3a4b771df4d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}