State-Coupled Volatility in Latent Dynamical Systems: Recovery Under Partial Observation

Technical Summary: State-Coupled Volatility in Latent Dynamical Systems

Problem Statement
Latent state-space models are standard tools for analyzing partially observed dynamical systems, yet they typically assume that process variability (stochastic noise) is independent of the latent state's position. In many biological, behavioral, and physiological systems, however, variability exhibits systematic organization tied to the underlying dynamical state. Traditional models often treat this variability as nuisance noise or attempt to model it using observed-state proxies, which can fail to capture structured stochasticity when observations are noisy. The core challenge addressed is the recovery of the relationship between latent-state displacement from equilibrium and process variance under conditions of partial observation, where the variance-driving quantity is unobserved.

Methodology
The paper introduces a state-coupled stochastic volatility framework where the latent process variance is a function of the latent state's displacement from a long-run equilibrium.

• Model Formulation:

  • Latent Dynamics: The latent state $x_t$ follows an equilibrium-seeking autoregressive process: $x_t = \mu + \phi(x_{t-1} - \mu) + w_t$.
  • State-Coupled Variance: Unlike constant-variance models, the process noise $w_t$ has a time-varying variance $q_t$ defined by: $\log q_t = \alpha + \gamma|x_{t-1} - \mu|$.
    • $\mu$: Equilibrium mean.
    • $\phi$: Persistence parameter.
    • $\alpha$: Baseline volatility.
    • $\gamma$: Coupling parameter quantifying the strength of association between state displacement and variability.
  • Observation Model: $y_t = x_t + v_t$, where $v_t \sim N(0, r)$ is observation noise.

• Inference Procedure:
  Because the variance-driving quantity ($|x_{t-1} - \mu|$) is latent, a standard plug-in approach (estimating a single trajectory then fitting variance) is insufficient as it ignores uncertainty propagation. The authors develop a Particle Expectation-Maximization (EM) procedure:

  1. E-step: Uses Bootstrap Particle Filtering combined with Backward Trajectory Smoothing to approximate the posterior distribution of the complete latent trajectory $p(x_{1:T} | y_{1:T})$. This propagates latent-state uncertainty into the variance estimation.
  2. M-step: Updates the volatility parameters ($\alpha, \gamma$) by minimizing a likelihood-based objective function using the smoothed particle trajectories, while holding mean-dynamics parameters ($\mu, \phi, r$) fixed.
  3. Optimization: Uses bounded L-BFGS-B with damping to ensure numerical stability.

Key Contributions

1. Interpretable Framework: A specific heteroskedastic state-space formulation where transition variance is explicitly parameterized as a function of latent-state geometry, rather than an independent latent volatility process (as in Stochastic Volatility models) or past residuals (as in GARCH).
2. Inference Strategy: A particle EM algorithm that jointly estimates latent trajectories and state-dependent variance parameters, explicitly accounting for the coupling between state reconstruction and variance estimation.
3. Identifiability Analysis: A theoretical proposition establishing that the volatility parameters are structurally identifiable conditional on the latent trajectory and mean-dynamics parameters.
4. Benchmarking: A large-scale simulation study (4,050 datasets) comparing the proposed method against:
   • Constant-variance Latent Dynamical Systems (LDS).
   • Observed-state heteroskedastic proxies (modeling variance based on $y_t$).
   • GARCH(1,1) proxies.
   • Simplified Stochastic Volatility (SV) proxies.

Results
The simulation benchmark evaluated parameter recovery (bias, RMSE) and detection performance (AUC) across varying coupling strengths ($\gamma$), observation noise ($r$), trajectory lengths ($T$), and persistence ($\phi$).

• Bias Reduction: The proposed state-coupled framework consistently reduced recovery bias compared to the observed-state heteroskedastic proxy. The proxy systematically underestimated coupling strength (attenuation bias), particularly as observation noise increased. The proposed method maintained estimates closer to the true $\gamma$, even in high-noise regimes.
• Bias-Variance Tradeoff: While the proposed method reduced systematic bias, it exhibited higher Root Mean Squared Error (RMSE) than the proxy in low-noise, short-trajectory settings due to the uncertainty introduced by latent-state reconstruction. However, as coupling strength and trajectory length increased, the RMSE gap narrowed, and the proposed method often achieved lower RMSE in high-persistence, high-noise regimes.
• Detection Performance: The proposed framework demonstrated superior or competitive Area Under the Curve (AUC) for detecting the presence of state-coupled volatility. Its advantage became most pronounced as observation noise increased and latent persistence ($\phi$) strengthened.
• Comparison to Alternatives:
  • Constant LDS: Showed high detection AUC for strong coupling (due to model misspecification signatures) but could not estimate the coupling magnitude.
  • GARCH/SV Proxies: Performed near chance levels, indicating that generic volatility clustering or independent volatility trajectories fail to capture state-dependent variance mechanisms.

Significance and Claims
The paper claims that state-coupled volatility can be identified and estimated under partial observation when the latent-state structure is explicitly modeled.

• Methodological Feasibility: The work establishes that structured stochasticity (variability dependent on latent state) is recoverable from noisy observations, provided one accounts for the uncertainty in the latent trajectory.
• Contextual Advantage: The framework is particularly advantageous for systems characterized by substantial observation noise and strong temporal persistence (e.g., neural, physiological, or behavioral systems), where direct observation-based proxies fail to distinguish measurement error from true state-dependent variability.
• Limitations and Scope: The authors explicitly state that these results demonstrate methodological feasibility based on simulated data generated under the assumed model. The paper does not claim empirical validation in specific biological systems, nor does it assert universal superiority over all volatility specifications. Future work is identified as necessary to test the framework on real-world data and assess robustness to model misspecification.

In summary, the paper provides a practical methodological foundation for investigating whether structured variability contributes information about system dynamics beyond what is contained in mean-state trajectories alone.