Dual Estimation for Nonlinear State Space Models

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 174 papers75 conceptsfrozen 2026-10-04

In classical control theory and signal processing, we usually assume a division of labor: a scientist provides a mathematical model (the parameters), and an engineer builds a filter to track the system’s status (the state). However, in complex real-world systems—from the erratic flight of a spinning ping-pong ball to the hidden firing patterns of neurons—we rarely have the luxury of a perfect model.

This creates a "chicken-and-egg" problem known as dual estimation. To estimate the hidden state of a system, you need an accurate model of its dynamics. But to identify the parameters of that model, you need clean, noise-free state data. This circular dependency is the central challenge of modern system identification.

The Circular Dependency of Inference

The fundamental difficulty in dual estimation is that states and parameters are often nonlinearly coupled. As noted in Dual Estimation and the Unscented Transformation by Eric A. Wan and colleagues, the unknown parameters govern the transition of the states, while the states provide the only evidence for the parameters.

If your model parameters are slightly off, your state estimate will drift. If you then use that drifting state estimate to "correct" your parameters, you may end up reinforcing the error rather than fixing it. This feedback loop can lead to divergence or "path degeneracy," where the estimator becomes overconfident in a wrong solution.

Strategies for Breaking the Loop

Researchers have developed several distinct architectural patterns to decouple these variables:

1. State Augmentation

The simplest conceptual approach is to treat the unknown parameters as "extra" hidden states that simply don't change over time (or change very slowly). By concatenating the state vector $x$ and the parameter vector $\theta$ into a single augmented vector, a standard nonlinear filter like the Unscented Kalman Filter (UKF) can be applied. Eric A. Wan demonstrated that the Unscented Transformation is particularly effective here because it captures the higher-order interactions between states and parameters better than the linear approximations used in older filters.

2. Dual Filtering

Instead of one giant filter, researchers often use two filters running in parallel. One filter estimates the state assuming the current parameters are correct; the other updates the parameters based on the state filter’s residual errors. This separation allows for different mathematical treatments of the two components. For instance, David Angwenyi proposed a hybrid approach using an Ensemble Kalman-Bucy Filter (EnKBF) for high-dimensional states and an Ensemble Transform Particle Filter (ETPF) for non-Gaussian parameters.

3. Iterative Smoothing (EM and Variational Inference)

When data can be processed in batches, we can use the Expectation-Maximization (EM) algorithm. In the "E-step," we use a smoother to find the best state trajectory given the current model. In the "M-step," we update the model parameters to maximize the likelihood of that trajectory.

Recent work has pushed this into the "online" or streaming domain. Yuan Zhao and Il Memming Park introduced Variational Joint Filtering (VJF), which uses variational inference to approximate the joint distribution of states and dynamics in real-time. This bypasses the need for expensive sampling by optimizing a "filtering ELBO" (Evidence Lower Bound), making it possible to uncover the rules of neural dynamics as they happen.

Leveraging Physical Constraints

A recurring theme in recent literature is that "pure" machine learning often struggles with dual estimation because it lacks a sense of physical reality. To solve this, researchers are embedding physical laws directly into the estimation architecture.

In Learning Dynamics of a Ball with Differentiable Factor Graph and Roto-Translational Invariant Representations, Qingyu Xiao and Matthew C. Gombolay use a differentiable factor graph. This allows the system to backpropagate errors through the physics engine itself. By enforcing symmetries—like the fact that physics shouldn't change just because you rotate the camera—the model constrains the search space, preventing the "circular dependency" from spiraling into unphysical solutions.

Similarly, Priyabrata Saha and colleagues developed PhICNet, which embeds partial differential equations (PDEs) into recurrent neural networks. This allows the model to distinguish between the internal dynamics of a system and external "source terms" (unobserved forces) that are driving it.

The Problem of Convergence

Even with sophisticated models, dual estimation is sensitive to how it starts. Zhankun Luo and Abolfazl Hashemi identified an "initialization gap" in EM-based estimates. They found that if the initial guess for how different components are mixed is too balanced, the algorithm converges much more slowly and with lower accuracy.

This highlights the "curse" of joint estimation: the landscape of the likelihood function is often riddled with local optima. As Fredrik Lindsten and Thomas B. Schön have shown, techniques like Rao-Blackwellization—which analytically solves the linear parts of a model while using particles for the nonlinear parts—are essential for reducing the variance of these estimates and ensuring the solver doesn't get stuck in a "degenerate" path where it stops exploring the parameter space entirely.

Papers in this version

Showing 60 of 74 papers.

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dual_estimation_for_nonlinear_state_spac_22559205,
  title = {Dual Estimation for Nonlinear State Space Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-estimation-for-nonlinear-state-space-models/rdr_22559205fdb448aab7218ac4ceace19a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_dual_estimation_for_nonlinear_state_spac_22559205_v1,
  title = {Dual Estimation for Nonlinear State Space Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-estimation-for-nonlinear-state-space-models/rdr_22559205fdb448aab7218ac4ceace19a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}