Double projection for reconstructing dynamical systems: between stochastic and deterministic regimes

This paper introduces a "double projection" method within dynamical variational autoencoders that simultaneously estimates system state trajectories and noise time series from data, enabling effective multi-step reconstruction and learning of low-dimensional stochastic models across various benchmark problems.

The Gist

Imagine you are trying to teach a robot to predict the future behavior of a complex system, like the weather, a stock market, or even the firing of a single neuron in a brain. You have a video of what happened in the past, and you want the robot to learn the "rules" so it can keep the story going on its own.

This paper introduces a new way to teach that robot, called Double Projection Dynamical System Reconstruction (DPDSR).

Here is the breakdown using simple analogies:

1. The Problem: Deterministic vs. Chaotic

Most old methods try to find a single, perfect set of rules (a Deterministic model). They assume that if you know the starting point, you can predict the future exactly.

• The Flaw: Real life is messy. Sometimes, tiny, random things (like a butterfly flapping its wings) change the outcome completely. If you try to force a messy, noisy system into a rigid, rule-based box, the robot gets confused. It tries to explain random noise as if it were a complex, hidden rule, leading to wild errors.

2. The Solution: The "Double Projection" Strategy

The authors propose a smarter approach. Instead of just guessing the rules, they teach the robot to guess two things at the same time:

1. The State: Where the system is right now (e.g., the temperature, the speed).
2. The Noise: The random "jitters" or surprises happening at that moment.

The Analogy: The Detective and the Alibi
Imagine a detective trying to reconstruct a crime scene from a blurry video.

• Old Method (Single Projection): The detective tries to guess exactly what the suspect did at every second, assuming the suspect followed a perfect, logical plan. If the suspect slipped on a banana peel (random noise), the detective thinks, "Ah, the suspect must have a secret plan to slip!" and invents a complex conspiracy that doesn't exist.
• New Method (Double Projection): The detective splits the job.
  • Detective A figures out where the suspect was (the State).
  • Detective B figures out what random things happened around them (the Noise: "Oh, someone dropped a banana peel").
  • By separating the "plan" from the "accident," the detective can understand the suspect's true behavior much better.

3. How It Works: The "Teacher" and the "Student"

To train this robot, the authors use a technique called Teacher Forcing.

• The Concept: Imagine a student learning to drive. If they drive perfectly for 10 seconds, the teacher lets them keep going. But if they start to drift, the teacher grabs the wheel, corrects the car, and says, "Okay, start from here."
• The Interval (τ): The paper asks: How often should the teacher grab the wheel?
  • Too often: The student never learns to drive on their own; they just follow the teacher.
  • Too rarely: The student crashes before the teacher can help.
  • The Discovery: The authors found that the "sweet spot" depends on the system.
    • For predictable systems (like a clock), the teacher should step in often. The system learns to be deterministic.
    • For chaotic/noisy systems (like the weather), the teacher should step in less often. This forces the robot to rely on its "Noise" detector to handle the randomness, leading to a more realistic model.

4. The Results: Why It Matters

The team tested this on six different "puzzles":

• The Chaos Puzzles: (Like the famous Lorenz weather model). The new method worked just as well as the old ones.
• The Noisy Puzzles: (Like a neuron firing or a double-well energy system). Here, the old methods failed miserably because they tried to force order onto chaos. The new method (DPDSR) excelled because it admitted, "Hey, this part is random," and modeled the randomness correctly.
• Real Life Data: (Heartbeats/ECG). The new method could reproduce the tiny, natural variations in heartbeats that the old methods smoothed over and lost.

The Big Takeaway

This paper is about knowing the difference between a rule and a random event.

By using a "Double Projection," the method stops trying to force a square peg into a round hole. It acknowledges that some systems are driven by hidden rules, while others are driven by a mix of rules and random noise. By learning to separate the two, we can build better, more accurate models of the real world, from brain activity to climate change.

In short: It's a new way to teach computers to understand that sometimes, things happen just because of a random "jitter," and that's okay.

Technical Summary

Here is a detailed technical summary of the paper "Double Projection for Reconstructing Dynamical Systems: Between Stochastic and Deterministic Regimes."

1. Problem Statement

The paper addresses the challenge of Dynamical System Reconstruction (DSR) from observed time-series data. While many scientific fields rely on building mathematical models of underlying dynamics, existing approaches face significant limitations:

• Deterministic Assumptions: Many influential methods assume the underlying system is deterministic. However, real-world data often contains intrinsic noise or is generated by stochastic processes. Forcing a deterministic model onto stochastic data often leads to poor long-term predictions or requires the model to approximate randomness via chaotic dynamics, which is unstable.
• Stochastic Modeling Challenges: Existing stochastic methods, particularly Dynamical Variational Autoencoders (DVAE), struggle with multi-step prediction. Standard approaches often sample noise from a prior distribution during rollout, which fails to capture the specific noise trajectory required to match the observed data accurately.
• Training Instability: Multi-step prediction in recurrent networks often suffers from exploding/vanishing gradients. While "teacher forcing" (injecting ground truth states periodically) helps, the optimal interval for this forcing is difficult to determine, especially when balancing deterministic and stochastic regimes.

2. Methodology: Double Projection Dynamical System Reconstruction (DPDSR)

The authors propose DPDSR, a novel variant of the DVAE framework that utilizes a double projection strategy.

Core Architecture

The method learns a latent dynamical system defined by:
$$z_t = F(z_{t-1}, \epsilon_t)$$
$$x_t = g(z_t) + \Sigma_\eta \eta_t$$
Where $z_t$ is the system state, $\epsilon_t$ is the system noise (latent variable), and $x_t$ is the observation.

Key Innovations:

1. Dual Encoders: Unlike standard VAEs that map observations to a single latent state trajectory, DPDSR uses two separate encoders:
   • State Encoder: Maps observations $x_{1:T}$ to an estimated state trajectory $\hat{z}_{1:T}$. This uses a WaveNet architecture (dilated 1D convolutions) to capture temporal dependencies.
   • Noise Encoder: Maps observations and estimated states to a posterior distribution of the noise time series, $q(\epsilon_{1:T} | x, \hat{z})$. This uses an autoregressive Gaussian posterior (WaveNet + LSTM) to model the noise sequence.
2. Double Projection Training:
   • The model estimates the full state and noise trajectories from the data.
   • It then evolves the system using the generative model and the estimated noise.
   • Teacher Forcing: Every $\tau$ steps, the simulated state is replaced by the estimated state $\hat{z}_t$ to prevent error accumulation.
3. Loss Function: The objective combines:
   • Reconstruction Loss: Minimizing the difference between generated and observed data ($x$) and the estimated states ($\hat{z}$).
   • KL Divergence: Regularizing the estimated noise posterior $q(\epsilon)$ against a standard normal prior $p(\epsilon)$.
   • Regularization: L1 regularization on projection weights and scale/position constraints on states.

Key Features

• Low-Dimensional State Space: By separating the computational expressivity (handled by the encoder) from the latent state size, DPDSR allows for low-dimensional state spaces that are easier to analyze than high-dimensional RNNs.
• Posterior Sampling for Rollout: Crucially, for multi-step prediction, the model samples noise from the posterior distribution (learned from data) rather than the prior. This allows the model to "remember" the specific noise sequence that drove the observed dynamics.
• Deterministic Special Case: By setting the noise injection matrix $B=0$, the method reduces to a deterministic model (SPDSR), allowing for direct comparison between stochastic and deterministic regimes within the same architecture.

3. Key Contributions

1. Novel Framework: Introduction of the DPDSR method, which uniquely estimates both state trajectories and driving noise time series simultaneously.
2. Bridging Regimes: The ability to tune the teacher forcing interval ($\tau$) to shift the model's behavior between a deterministic chaotic regime (low $\tau$) and a noise-driven stochastic regime (high $\tau$).
3. Comprehensive Benchmarking: Evaluation on six diverse datasets, including:
   • Deterministic chaos (Lorenz, Cell Cycle).
   • Stochastic/Noise-driven systems (Double Well, Chaotic RNN, Neuron voltage).
   • Real-world experimental data (ECG).
4. Analysis of Teacher Forcing: A detailed investigation into how $\tau$ affects the nature of the learned attractors (fixed points, limit cycles, or chaos) and model performance.

4. Results

The authors evaluated DPDSR against state-of-the-art methods: SPDSR (deterministic variant), GTF-TD (Generalized Teacher Forcing with time delay), Deep Kalman Filter (DKF), AR-LSTM, and RSSM (Recurrent State Space Model).

• Stochastic Domains (Double Well, RNN, Neuron): DPDSR significantly outperformed deterministic methods (SPDSR, GTF-TD). Deterministic models failed to reproduce the bimodal distributions or spike variations, often resorting to chaotic dynamics to approximate randomness. DPDSR, DKF, and RSSM performed best, with DPDSR showing robustness in capturing noise-driven transitions.
• Deterministic Domains (Lorenz, Cell Cycle): DPDSR was competitive with deterministic methods, successfully learning chaotic attractors, though deterministic models (SPDSR, GTF-TD) were slightly superior or comparable.
• Experimental Data (ECG): DPDSR and RSSM-OO were the only methods capable of reproducing the variations in interspike intervals (ISI), a critical feature of cardiac dynamics that deterministic models failed to capture.
• Teacher Forcing Analysis:
  • Low $\tau$: Models tend toward deterministic chaos (high Lyapunov exponents).
  • High $\tau$: Models transition to noise-driven regimes (stable fixed points or limit cycles driven by noise).
  • Optimal $\tau$: The optimal interval varies by dataset. For stochastic systems, a larger $\tau$ is often required to allow the noise to drive the dynamics.
  • Adaptive $\tau$: The authors tested an adaptive scheme for $\tau$ but found it prone to converging to suboptimal fixed points, suggesting that parameter sweeps remain necessary for robust results.

5. Significance and Conclusion

This work provides a robust framework for reconstructing dynamical systems where the boundary between deterministic and stochastic behavior is unclear.

• Scientific Impact: It demonstrates that assuming a purely deterministic structure for noisy data can lead to fundamentally incorrect models (e.g., mistaking noise for chaos). DPDSR offers a principled way to learn the "true" stochastic nature of the system.
• Methodological Advancement: By utilizing posterior sampling for multi-step rollout, DPDSR overcomes a key limitation of previous DVAE approaches, enabling accurate long-term prediction in stochastic environments.
• Interpretability: The method's ability to operate in low-dimensional latent spaces and its tunable nature (via $\tau$) make it a powerful tool for analyzing the internal dynamics of complex systems, from neural activity to cardiac signals.

The paper concludes that while finding the optimal teacher forcing interval remains a challenge, the double projection approach offers a flexible and powerful alternative for modeling complex dynamical systems across the spectrum of stochasticity.