Pathwise Learning of Stochastic Dynamical Systems with Partial Observations

This paper introduces a neural path estimation framework based on variational inference and stochastic control to reconstruct and infer stochastic dynamical systems from noisy, nonlinear, and sparse observations by learning a generative model that maps prior path measures to filtering posterior measures.

The Gist

Imagine you are trying to figure out exactly where a lost hiker is walking through a dense, foggy forest. You can't see them directly. Instead, you only have a shaky, noisy radio signal that tells you roughly where they might be, but the signal is full of static and sometimes cuts out completely.

This is the core problem of stochastic dynamical systems with partial observations. Scientists and engineers face this every day, whether they are tracking a virus spreading through a city, predicting stock market crashes, or guiding a robot through a chaotic environment.

Here is a simple breakdown of what this paper does, using everyday analogies.

1. The Old Way: The "Crowd of Guessers"

Traditionally, to solve this problem, scientists used methods called Particle Filters.

• The Analogy: Imagine you release 10,000 tiny, invisible drones into the forest. Each drone makes a random guess about where the hiker is. Every time you get a new radio signal, you check which drones are closest to the signal. You throw away the drones that are way off and "clone" the ones that are close.
• The Problem:
  • Too many drones: If the forest is huge (high dimensions), you need millions of drones to get an accurate picture, which is computationally expensive.
  • Clumping: Sometimes, all the drones accidentally clump together in one spot, and you lose the ability to see other possibilities (like if the hiker is actually in two different places at once).
  • Fragility: If the radio signal is missing for a while, the drones get lost and the whole system breaks.

2. The New Way: The "Smart GPS App"

This paper proposes a new method called Pathwise Learning with Neural SDEs. Instead of using a crowd of drones, they build a single, incredibly smart "GPS App" that learns the rules of the forest.

• The Analogy: Instead of guessing with thousands of drones, you train a neural network (a type of AI) to act like a generative map.
  • Learning the Rules: The AI looks at thousands of examples of "Hiker + Radio Signal" pairs. It learns not just where the hiker is at a specific moment, but how the hiker moves over time. It learns the "flow" of the forest.
  • The "Amortized" Magic: In the old way, if you wanted to track a new hiker, you had to start the drone simulation from scratch. In this new way, once the AI is trained, it has a "shortcut." When you feed it a new noisy radio signal, it instantly generates the most likely path the hiker took. It doesn't need to re-learn; it just applies what it already knows.

3. How It Works: The "Controlled Diffusion"

The paper uses some heavy math (Stochastic Differential Equations and Variational Inference), but here is the simple version:

• The Problem: The radio signal is noisy. The AI needs to figure out the "true path" hidden behind the noise.
• The Solution: The AI creates a "controlled" version of the hiker's movement. Imagine the hiker is walking on a slippery slope (random noise). The AI acts like a guide who gently nudges the hiker in the right direction based on the radio signal.
• The "Pathwise" Trick: Most methods only look at the hiker's location at this exact second. This paper looks at the entire journey. It understands that if the hiker was at Point A at 9:00 AM and Point B at 9:05 AM, the path between them matters. This helps the AI handle weird situations, like the hiker getting stuck in a "double-well" (a valley with two peaks) or moving chaotically like a butterfly.

4. Why This Is a Big Deal

The authors tested their "Smart GPS" on three difficult scenarios:

1. The Double-Well (The Valley with Two Peaks): Imagine a ball that can roll into either the left valley or the right valley. It's hard to predict which one it will pick. The old methods struggled to see both possibilities at once. The new method easily handled this "multimodal" uncertainty.
2. The Lorenz System (The Chaotic Butterfly): This is a famous weather model that is extremely chaotic. Small changes lead to huge differences. The new method tracked the path much better than the old "drone" methods, even when the data was sparse.
3. Missing Data (The Broken Radio): What if the radio signal stops for 20% of the time? The old methods often crashed or gave bad guesses. The new method kept working, filling in the gaps by remembering the "flow" of the movement it learned during training.

The Bottom Line

This paper introduces a learning-based approach to tracking things in the dark.

Instead of running a massive, slow simulation every time you get new data, they trained a neural network to understand the entire story of how a system moves. Once trained, this network can instantly reconstruct the hidden path of a system, even if the data is noisy, missing, or chaotic.

It's like upgrading from a team of 10,000 people guessing where a car is driving based on a broken GPS, to a self-driving car that has memorized the entire map and can instantly predict the route, even if the GPS signal flickers.

Technical Summary

1. Problem Statement

The paper addresses the fundamental challenge of reconstructing and inferring stochastic dynamical systems from noisy, partial, and potentially sparse observations.

• Context: In many scientific fields (geosciences, robotics, finance), the true state $X_t$ of a system evolves according to a Stochastic Differential Equation (SDE) but is only observable through a noisy, nonlinear measurement process $Y_t$.
• The Gap: Traditional methods like Sequential Monte Carlo (Particle Filters) suffer from particle degeneracy in high dimensions and require resampling, which introduces discontinuities. Ensemble Kalman Filters (EnKFs) are computationally efficient but inconsistent with Bayes' theorem for non-Gaussian models. Furthermore, standard data-driven surrogate models often assume high-fidelity training data and struggle when the underlying dynamics are unknown or partially specified.
• Goal: To develop a data-driven, amortized method that simultaneously learns the SDE dynamics and infers the posterior path measure (the distribution of the entire trajectory $X_{0:T}$ given observations $Y_{0:T}$) without requiring explicit knowledge of the governing equations or large particle ensembles.

2. Methodology

The proposed approach combines stochastic control theory, variational inference, and neural differential equations.

A. Theoretical Foundation: Pathwise Filtering and Control

The authors reformulate the nonlinear filtering problem as a stochastic optimal control problem on path space.

1. Pathwise Zakai Equation: They derive a parabolic PDE governing the unnormalized conditional density $q^y_t(x)$ for a deterministic observation path $y$. This equation accounts for time-dependent diffusion coefficients and nonlinear observations.
2. Variational Representation: Using the Gibbs variational principle and the Kallianpur-Striebel formula, they show that the posterior path measure is the solution to a stochastic control problem. The optimal feedback control $u^*$ is explicitly related to the gradient of the log-posterior density:
   $$u^*(t, x) = -\sigma(t, x)^\top \nabla_x \log q^y_t(x)$$
3. HJB Formulation: They establish a connection between the pathwise Zakai equation and the Hamilton-Jacobi-Bellman (HJB) equation, proving that the optimal control minimizes a specific cost functional involving the drift and diffusion terms.

B. Neural Architecture: Conditional Latent SDE

To solve this control problem efficiently, the authors propose a Conditional Latent SDE framework:

1. Generative Model: Instead of learning a static map, they learn a conditional generative model that maps a prior path measure to the posterior path measure.
2. Latent Space: The system operates in a latent space $Z$. An encoder network $\mathcal{E}_\phi$ processes the observation history $y_{0:t}$ to parameterize the initial latent state distribution.
3. Controlled Dynamics: The latent process $Z_t$ evolves via a controlled SDE:
   $$dZ_t = \left( w_\theta(t, Z_t; y_{0:t}) + g_\theta(t, Z_t) u_\phi(t, Z_t; y_{0:t}) \right) dt + g_\theta(t, Z_t) dW_t$$
   Here, $u_\phi$ is a neural network acting as the amortized feedback control, approximating the optimal control derived in the theoretical section.
4. Decoder: A decoder $D_\theta$ maps the latent states back to the physical state space $X_t$.

C. Training Objective: Pathwise ELBO

The model is trained by maximizing a Pathwise Evidence Lower Bound (ELBO).

• The objective function consists of a reconstruction term (log-likelihood of observations given latent states) and a regularization term (KL divergence between the variational path measure and the model path measure).
• Using Girsanov's theorem, the KL divergence is computed analytically as the integral of the squared difference between the drifts of the variational and model processes, normalized by the diffusion coefficient.
• This allows for end-to-end training without explicit particle resampling.

3. Key Contributions

1. Variational Control Formulation: The paper bridges nonlinear filtering and stochastic control by deriving a pathwise Zakai equation and showing its equivalence to an optimal control problem where the control is the gradient of the log-density.
2. Amortized Inference: Unlike recursive filters that update state estimates step-by-step, this method learns a conditional generative map. Once trained, it can generate posterior sample paths for any new observation trajectory instantly without re-running a filter or retraining.
3. Model-Free Learning: The method does not require prior knowledge of the system's drift or diffusion coefficients. It learns the dynamics directly from noisy data.
4. Robustness to Sparsity and Noise: The continuous-time formulation naturally handles irregularly sampled data and missing observations, a common issue in real-world applications.
5. Path Functional Estimation: The method provides uncertainty quantification not just for marginal distributions but for path functionals (e.g., hitting times, occupation times, dwell times), which are critical in chaotic and metastable systems.

4. Experimental Results

The authors validate the method on three distinct stochastic systems:

• Stochastic Double-Well Equation:

  • Task: Estimate trajectories and compute "dwell times" (time spent in a specific potential well).
  • Result: The model successfully captures the bimodal stationary distribution and metastable behavior, accurately estimating path functionals with high coverage probability (90% confidence intervals).

• Stochastic Lorenz-63 (Chaotic System):

  • Task: State estimation in a chaotic system with nonlinear observations ($\arctan$).
  • Comparison: Outperforms Particle Filters (PF) and Particle Gibbs (PG) in terms of RMSE and Wasserstein distance.
  • Key Finding: Particle methods required thousands of particles to achieve accuracy comparable to the learned SDE, which used only 64 samples. The learned model handled mode transitions better in sparse observation regimes.

• Stochastic Lorenz-96 (High-Dimensional & Missing Data):

  • Task: 15-dimensional system with randomly missing observations (up to 40% missing rate).
  • Result: The method demonstrated superior robustness compared to PF and PG baselines. The continuous-time encoder effectively handled masked inputs, maintaining accurate trajectory reconstruction even with significant data gaps.

• Real-World Application (MuJoCo Hopper):

  • Task: Denoising and trajectory inference for a physics simulation with multimodal dynamics.
  • Result: The conditional latent SDE outperformed a standard GRU autoregressive model, particularly in capturing multimodal distributions and maintaining temporal coherence in irregularly sampled data.

5. Significance and Future Directions

• Scalability: By avoiding particle resampling and large ensembles, the method offers a scalable alternative for high-dimensional filtering problems where traditional Monte Carlo methods fail.
• Flexibility: The framework unifies parameter estimation and state estimation, making it suitable for "inverse problems" where the system dynamics are unknown.
• Future Work: The authors suggest extending the framework to online settings with stability guarantees, exploring Mean-Field limits (McKean-Vlasov), and integrating Transformer architectures to handle longer sequences more efficiently, addressing the current quadratic complexity of the attention-based encoder.

In summary, this paper presents a novel neural path estimation approach that leverages the mathematical structure of stochastic control to solve the filtering problem, offering a robust, efficient, and model-free solution for inferring complex stochastic dynamics from noisy, partial data.