Nudging state-space models for Bayesian filtering under misspecified dynamics

Fernando Rodríguez González, Ömer Deniz Akyıldız, Dan Crisan, Joaquı́n Mı́guezStat. Comput 2024

Abstract

Nudging is a popular algorithmic strategy in numerical filtering to deal with the problem of inference in high-dimensional dynamical systems. We demonstrate in this paper that general nudging techniques can also tackle another crucial statistical problem in filtering, namely the misspecification of the transition kernel. Specifically, we rely on the formulation of nudging as a general operation increasing the likelihood and prove analytically that, when applied carefully, nudging techniques implicitly define state-space models that have higher marginal likelihoods for a given (fixed) sequence of observations. This provides a theoretical justification of nudging techniques as data-informed algorithmic modifications of state-space models to obtain robust models under misspecified dynamics. To demonstrate the use of nudging, we provide numerical experiments on linear Gaussian state-space models and a stochastic Lorenz 63 model with misspecified dynamics and show that nudging offers a robust filtering strategy for these cases.

Introduction

In Bayesian filtering, the objective is to estimate the hidden state of a dynamical system—such as atmospheric conditions or the trajectory of an autonomous vehicle—based on a sequence of noisy observations. This estimation relies on two fundamental components: a likelihood function, which relates states to observations, and a transition kernel, which models the system's evolution over time.

A significant challenge in practical applications is model misspecification. When the assumed transition kernel deviates from the true physical process—due to ignored forces, incorrect parameters, or structural simplifications—the filter's estimates often drift significantly from reality. While traditional methods like outlier detection or parameter estimation exist, they often lack the flexibility to correct fundamentally flawed model structures. This paper investigates "nudging" as a robust solution to bridge the gap between misspecified models and observed reality.

Methodology: Nudging as Model Modification

Nudging is traditionally viewed as an algorithmic heuristic to steer a model toward data. This research reframes nudging as a formal modification of the state-space model itself. By introducing nudging maps, the authors define a new class of models, denoted as $M^\alpha$, which are adaptively informed by incoming data.

Gradient-Ascent Nudging

A primary implementation explored is Gradient-Ascent Nudging. In this approach, the state is shifted at each time step in the direction of the gradient of the log-likelihood. This ensures that the model's predicted state moves toward regions of the state space that better explain the current observation. For systems with physical boundaries (e.g., non-negative concentrations), the authors introduce Projected Nudging, which utilizes a projected gradient-ascent version to maintain the state within valid constraints.

Theoretical Foundations

The core contribution of this work is the analytical proof that nudging is not merely a heuristic but a method for Bayesian evidence maximization. Under mild regularity conditions, the authors prove that there exists a nudging sequence ensuring the modified model $M^\alpha$ possesses a higher marginal likelihood than the original misspecified model.

This improvement is quantified through the "nudging gain" ($\Delta^\alpha_T$), representing the boost in marginal likelihood achieved. This theoretical link provides a rigorous justification for using nudging as a data-driven way to approximate the true maximum likelihood estimator in the presence of model error.

Experimental Evaluation

The methodology was validated using two synthetic benchmarks designed to simulate varying degrees of model-reality gaps.

1. 4D Controlled Linear Dynamical System

In this experiment, a linear Gaussian state-space model was used where the filter was intentionally kept "blind" to specific control terms. The results indicated that the nudged Kalman Filter could effectively compensate for the missing control information, achieving a higher marginal likelihood than even a standard filter that lacked the nudging mechanism.

2. Stochastic Lorenz 63 Model

To test robustness in chaotic environments, the authors applied the method to the Lorenz 63 model with extreme parameter misspecification (e.g., doubling standard parameter values) and partial observations.

  • Tracking Accuracy: The standard Particle Filter (PF) failed to track the state, yielding a Normalized Mean Square Error (NMSE) of approximately 1.7484. In contrast, the Nudged PF maintained tracking with an NMSE of ~0.1190.
  • Model Fit: The log-Bayesian evidence saw a massive improvement, moving from $-1.3366 \times 10^5$ in the standard model to $-1.2961 \times 10^3$ in the nudged version, demonstrating a significantly stronger alignment with observed data.

Discussion and Limitations

While nudging provides a powerful tool for robust filtering, several considerations remain for practical implementation:

  1. Step-Size Selection: While the existence of an optimal nudging step size ($\gamma$) is proven, identifying it in practice requires empirical tuning or cross-validation.
  2. Risk of Degeneracy: Aggressive nudging (excessively large step sizes) can lead to kernel degeneracy, where the transition kernel collapses to a single point. This may increase estimation errors despite achieving a high likelihood.
  3. Computational Complexity: Gradient-based methods are generally efficient, but calculating gradients for extremely high-dimensional or non-differentiable models remains a computational hurdle.

Conclusion

This work provides a formal theoretical foundation for nudging in Bayesian filtering. By demonstrating that nudging systematically increases marginal likelihood, the authors move the technique from a numerical heuristic to a statistically grounded method for handling misspecified dynamics. Future research will focus on non-gradient-based nudging schemes and applications to large-scale, real-world geophysical models.

Summary updated

Abstract

Nudging is a popular algorithmic strategy in numerical filtering to deal with the problem of inference in high-dimensional dynamical systems. We demonstrate in this paper that general nudging techniques can also tackle another crucial statistical problem in filtering, namely the misspecification of the transition kernel. Specifically, we rely on the formulation of nudging as a general operation increasing the likelihood and prove analytically that, when applied carefully, nudging techniques implicitly define state-space models that have higher marginal likelihoods for a given (fixed) sequence of observations. This provides a theoretical justification of nudging techniques as data-informed algorithmic modifications of state-space models to obtain robust models under misspecified dynamics. To demonstrate the use of nudging, we provide numerical experiments on linear Gaussian state-space models and a stochastic Lorenz 63 model with misspecified dynamics and show that nudging offers a robust filtering strategy for these cases.
BibTeX
@article{gonzalez2024nudging,
  author = {Fernando Rodríguez González and Ömer Deniz Akyıldız and Dan Crisan and Joaquı́n Mı́guez},
  title = {{Nudging state-space models for Bayesian filtering under misspecified dynamics}},
  journal = {Stat. Comput.},
  year = {2024},
}
Figure