Joint State and Parameter Estimation in Nonlinear Systems

62 concepts53 papers1 proposal

In many real-world applications, we want to track the internal, unobservable state of a system over time. However, doing so accurately requires knowing the underlying physical constants or rules that govern the system's transitions. When these parameters are also unknown, we face a challenging problem: we must simultaneously track the hidden states and identify the system parameters from noisy observations.

The Circular Dilemma of Dual Estimation

The fundamental difficulty of this joint estimation task lies in its inherent circular dependency. To accurately estimate the time-varying latent states of a system, you need to know its physical parameters. Conversely, to identify those fixed parameters, you typically require precise measurements of the states.

As Jean-Sébastien Brouillon et al. explain in Robust online joint state/input/parameter estimation of linear systems, this interdependency creates a bilinear coupling where the system's states and parameters multiply each other in the underlying mathematical models. Consequently, errors in the state estimate propagate directly into the parameter estimate and vice versa, complicating the initialization and convergence of tracking algorithms.

In Dual Estimation and the Unscented Transformation, Eric A. Wan et al. note that this coupling is particularly problematic when the dynamics are nonlinear, making accurate inference highly complex. Andreas Svensson et al. echo this framing in Identification of jump Markov linear models using particle filters, emphasizing that this circularity prevents direct analytical solutions and necessitates iterative estimation techniques. When the internal states are completely hidden and the transitions are nonlinear, Zoubin Ghahramani et al. show in Learning Nonlinear Dynamical Systems Using an EM Algorithm that standard linear tracking methods break down entirely, making it exceptionally difficult to map noisy observations back to internal state transitions.

The Mathematical Bottleneck of Intractable Likelihoods

Why can we not simply solve this joint estimation problem using standard optimization tools? The answer lies in the geometry of the problem's marginal likelihood function.

To evaluate how well a set of parameters fits the observed data, we must integrate out all possible trajectories of the hidden states. As Arno Solin et al. point out in Sigma-Point Filtering and Smoothing Based Parameter Estimation in Nonlinear Dynamic Systems, calculating this marginal likelihood requires solving complex integrals that lack closed-form solutions. Fredrik Lindsten et al. note in Second-order Particle MCMC for Bayesian Parameter Inference that this analytical intractability forces researchers to rely on simulation-based approximations, which introduce significant noise and computational overhead.

This bottleneck becomes even more pronounced when trying to optimize parameters online or in real-time. Online parameter estimation typically relies on calculating the gradient of the likelihood function. However, Vladislav Z. B. Tadic et al. demonstrate in Bias of Particle Approximations to Optimal Filter Derivative that this gradient calculation depends directly on the optimal filter and its derivative. For nonlinear state-space models, these quantities cannot be computed analytically, posing a severe challenge for reliable estimation. Even when moving to non-parametric representations, the math remains unyielding; Jakob Lindinger et al. observe in Laplace approximated Gaussian process state-space models that calculating the marginal likelihood for non-parametric transitions is mathematically intractable, preventing direct optimization of model parameters and requiring complex approximation schemes to handle the evidence.

Algorithmic Strategies for Joint Inference

To bypass these mathematical hurdles, researchers have developed several distinct strategies to approximate the intractable integrals and decouple the state-parameter relationship.

Simulation and Particle Methods

One dominant paradigm relies on Monte Carlo simulations to approximate the state distributions. For example, Johan Dahlin et al. demonstrate in Particle Metropolis-Hastings using gradient and Hessian information that gradient and Hessian information can be estimated "on the fly" using a fixed-lag particle smoother, allowing for scale-invariant proposals. However, these simulation methods come with a steep trade-off. Daniel Durstewitz presents in A state space approach for piecewise-linear recurrent neural networks for identifying computational dynamics from neural measurements that numerical methods like particle filters are computationally expensive and scale poorly with the number of variables, necessitating more efficient frameworks when both process and observation noise are present.

Variational Inference

To achieve better scalability, researchers frequently turn to variational approximations. Instead of simulating particles, variational methods approximate the true posterior distribution with a simpler, tractable distribution and optimize its parameters. In Scalable Bayesian Learning for State Space Models using Variational Inference with SMC Samplers, Marcel Hirt et al. combine Sequential Monte Carlo with Variational Inference to handle the significant computational burden imposed by the highly interdependent latent states and fixed parameters. Similarly, Wil O. C. Ward et al. apply variational techniques in Variational Bridge Constructs for Grey Box Modelling with Gaussian Processes. to balance the learning of static physical model parameters with the dynamic inference of time-varying latent variables.

Iterative Linearization and Observers

Another approach relies on deterministic approximations and control-theoretic observer designs. Rafael Anderka et al. introduce a framework in Iterated INLA for State and Parameter Estimation in Nonlinear Dynamical Systems that pairs iterative linearization with the Integrated Nested Laplace Approximation (INLA), providing robust, non-Gaussian uncertainty measures even when data is sparse or noisy.

From a control perspective, Jong-Kun Yang et al. show in Indirect Adaptive Fuzzy Observer Design that integrating adaptive laws directly into observer design allows the system to simultaneously learn the governing dynamics while tracking unobservable states in real-time. To ensure these methods do not diverge, Matthias A. Müller et al. develop a Moving Horizon Estimation (MHE) scheme in A moving horizon state and parameter estimation scheme with guaranteed robust convergence, leveraging incremental stability to mathematically guarantee that the joint estimation error converges exponentially to a small neighborhood around the truth.

Key Foundational Work

Dual Estimation and the Unscented Transformation (1999) — contributed 1 concept elements to this direction.
Fisher Scoring and a Mixture of Modes Approach for Approximate Inference and Learning in Nonlinear State Space Models (1998) — contributed 1 concept elements to this direction.
Learning Nonlinear Dynamical Systems Using an EM Algorithm (1998) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_simultaneous_state_and_parameter_estimat_e31b691a,
  title = {simultaneous state and parameter estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/simultaneous-state-and-parameter-estimation/rdr_e31b691a738f4d6f946ca7a5db985727},
  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_simultaneous_state_and_parameter_estimat_e31b691a_v1,
  title = {simultaneous state and parameter estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/simultaneous-state-and-parameter-estimation/rdr_e31b691a738f4d6f946ca7a5db985727/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake