Imagine trying to track the remaining charge in an electric vehicle battery, predict the internal temperature of an eye during retinal laser surgery, or guide an autonomous underwater vehicle through turbulent currents. Standard filtering techniques, such as the Kalman filter, excel at tracking these hidden states over time based on noisy observations. However, these filters require a precise mathematical model of the system's underlying physics.
What happens when the model itself contains unknowns? A battery's internal resistance changes as it ages. A patient's metabolic response to glucose varies throughout the day. A mechanical structure's stiffness degrades under stress. If you plug incorrect, static parameters into a state estimator, the filter's predictions will rapidly diverge from reality.
The traditional workaround is a two-step process: first, fit the parameters offline using historical data, and second, run the state estimator online. But this strategy fails when parameters fluctuate or vary across individuals. The solution is joint state-parameter estimation: treating unknown parameters as latent, time-varying variables and estimating them simultaneously alongside the system states in a single, unified inference engine.
The Core Mechanism: State Augmentation
How do you force a state filter to learn a static or slowly drifting parameter? The most elegant and widely used mechanism is state augmentation.
Consider a discrete-time dynamical system with a state vector $x_t$ and an unknown parameter vector $\theta_t$:
$$x_{t+1} = f(x_t, \theta_t) + w_t$$
To estimate both simultaneously, we construct an expanded, augmented state vector $z_t$ that concatenates the physical states and the parameters:
$$z_t = \begin{bmatrix} x_t \\ \theta_t \end{bmatrix}$$
We then define artificial dynamics for the parameters. If a parameter is assumed to be constant or slowly drifting, we model its transition as a random walk with zero-mean dynamics:
$$\theta_{t+1} = \theta_t + \eta_t$$
where $\eta_t$ is a small artificial process noise. This noise gives the filter the "flexibility" to adjust its parameter estimates over time as new data arrives.
As highlighted by Alejandro F. Villaverde et al. in Structural Identifiability of Dynamic Systems Biology Models, this mathematical transformation effectively converts a parameter identification problem into a nonlinear observability problem within a unified state-space framework. Similarly, Matthias A. Müller et al. demonstrate in A moving horizon state and parameter estimation scheme with guaranteed robust convergence that treating constant unknown parameters as states with zero dynamics allows standard nonlinear stability tools to provide rigorous convergence guarantees within an optimization window.
The Catch: Induced Non-Linearity
Why does state augmentation require specialized algorithms? The catch is that augmentation introduces non-linearity, even if the original system equations were perfectly linear.
Suppose a system is governed by the linear equation $x_{t+1} = \theta_t x_t$. If $\theta_t$ is a known constant coefficient, this equation is trivial to filter. But if $\theta_t$ is an unknown state, the transition equation now involves the multiplication of two states ($\theta_t$ and $x_t$).
In Simultaneous estimation of parameters and the state of an optical parametric oscillator system, the authors note that this product transforms a linear problem with unknown coefficients into a fundamentally nonlinear estimation task. Consequently, standard linear Kalman filters cannot be used directly; researchers must deploy nonlinear filtering architectures to approximate the joint distribution.
Architectural Choices: Joint vs. Dual Filtering
When implementing this concept, researchers generally choose between two structural paradigms, as formalized in the foundational work by Eric A. Wan et al. in Dual Estimation and the Unscented Transformation:
- Joint (Augmented) Filtering: A single filter operates on the fully augmented state vector $z_t$. This approach is mathematically rigorous because the filter explicitly tracks the cross-covariances between the states and the parameters. If a change in a parameter causes a predictable change in the state, the joint filter exploits this correlation to update both simultaneously.
- Dual Filtering: Instead of one massive filter, two separate filters run in parallel—one estimating the states assuming the current parameters are correct, and the other estimating the parameters using the state estimates as inputs.
While joint filtering captures the full error geometry, dual filtering can be computationally cheaper and more modular. For instance, David J. Albers et al. leverage a dual filtering mechanism in Personalized glucose forecasting for type 2 diabetes using data assimilation to personalize mechanistic physiological models, adapting a generalized biological model to an individual's unique biological constants over time.
Algorithmic Strategies for Non-Linear Inference
Because the augmented state space is inherently nonlinear, the choice of the underlying filter dictates how accurately the system handles uncertainty.
Analytical Linearization (EKF)
The Extended Kalman Filter (EKF) approximates nonlinear dynamics by taking a first-order Taylor series expansion at each time step. While computationally efficient, it can fail if the parameter-state interactions are highly nonlinear or if the initial guesses are poor. Despite these limitations, it remains highly effective for real-time tracking when paired with physical models. For example, Dimitris N. Metaxas et al. utilize an EKF framework in Elastically Adaptive Deformable Models to dynamically tune the stiffness parameters of physics-based deformable models during shape tracking.
Sigma-Point Methods (UKF)
To avoid the errors introduced by analytical linearization, the Unscented Kalman Filter (UKF) uses a deterministic set of sample points (sigma points) to capture the true mean and covariance of the augmented state distribution, mapping them directly through the nonlinear dynamics. In Dual Estimation and the Unscented Transformation, the authors demonstrate that this approach significantly outperforms the EKF in tracking accuracy for dual estimation tasks. This has made the UKF a cornerstone for physical systems, such as tracking vehicle dynamics (Model updating of a nonlinear experimental vehicle using substructuring and unscented Kalman filtering) and battery management (Adaptive State of Charge Estimation for Li-Ion Batteries Based on an Unscented Kalman Filter with an Enhanced Battery Model).
Advanced Probabilistic and Optimization Methods
For highly non-Gaussian, high-dimensional, or constrained systems, researchers move beyond standard Kalman variants:
- Laplace Approximations: Rafael Anderka et al. introduce Iterated INLA in Iterated INLA for State and Parameter Estimation in Nonlinear Dynamical Systems, combining iterative linearization with Integrated Nested Laplace Approximations to produce a mixture-of-Gaussians posterior that captures complex, non-Gaussian parameter uncertainties where traditional ensemble filters fail.
- Moving Horizon Estimation (MHE): When physical constraints must be strictly enforced (e.g., ensuring a temperature or diffusion coefficient never drops below zero), MHE optimizes over a sliding window of past data. In State and parameter estimation for retinal laser treatment, the authors show that MHE offers superior stability over the EKF in multi-parameter medical scenarios by enforcing these physical boundaries.
- Particle Filters and Langevin Dynamics: For systems evolving on non-Euclidean spaces, Frank Chongwoo Park et al. develop a particle filtering framework on the Euclidean group SE(3) in Particle filtering on the Euclidean group: framework and applications to simultaneously track poses and unknown noise covariances. For complex partial differential equations (PDEs), Alex Glyn-Davies et al. combine statistical finite elements with interacting particle Langevin dynamics in Statistical Finite Elements via Interacting Particle Langevin Dynamics to jointly infer latent spatial solutions and unknown physical constants like diffusivity.
- Unified Continuous Passes: Rather than relying on traditional iterative sampling methods like Markov Chain Monte Carlo (MCMC), Jonathan Schmidt et al. treat parameters as stochastic processes within an augmented state in A Probabilistic State Space Model for Joint Inference from Differential Equations and Data, solving differential equations and inferring parameters in a single computational pass with linear complexity. Similarly, in Double-Chain Unscented Expectation Propagation for Inference in Stochastic Dynamical Models of Biological Processes, Stefan Bernhard et al. develop an iterative refinement technique over the posterior distribution to discover underlying biological rules from noisy, partial observations.
Real-World Application: Battery Management
The power of joint state-parameter estimation is most apparent in its ability to make models self-correcting in the wild. Battery management systems provide a clear case study.
To prevent an electric vehicle from stalling, the management system must track the battery's State of Charge (SOC). However, as a battery degrades over months of use, its capacity shrinks and its internal resistance climbs. A filter tracking charge based on factory-new parameters will quickly become highly inaccurate.
Augmented State Vector:
z_t = [ State of Charge (SOC) , Internal Resistance , Capacity ]^T
│ │ │
▼ ▼ ▼
(Fast-changing State) (Slowly Drifting Physical Parameters)
By treating these degradation indicators as augmented states within a recursive filtering framework, the filter tracks immediate energy availability alongside long-term structural integrity. In State-of-Charge and State-of-Health Estimation for Lithium-Ion Batteries Based on Dual Fractional-Order Extended Kalman Filter and Online Parameter Identification, Liuyi Ling leverages online parameter identification to reduce SOC estimation errors from over 7% to less than 1% across standard driving cycles. As demonstrated by Xiaoxuan Wu et al. in Robust Online Estimation of State of Health for Lithium-Ion Batteries Based on Capacities under Dynamical Operation Conditions, this filter-based co-estimation approach yields faster dynamic responses and superior accuracy under fluctuating temperatures compared to traditional static models.
The same principle applies to robotics. Pablo Barreno et al. show in An Efficient Underwater Navigation Method Using MPC with Unknown Kinematics and Non-Linear Disturbances that simultaneously estimating an underwater vehicle's kinematics and environmental disturbance parameters significantly refines the transition model matrices, leading to robust control under limited computational resources.
By unifying the physics of the system (the states) with the properties of the system (the parameters), joint estimation transforms static mathematical models into adaptive, self-correcting digital twins.
Key Foundational Work
Elastically Adaptive Deformable Models (1996) — contributed 2 concept elements to this direction.
Dual Estimation and the Unscented Transformation (1999) — contributed 1 concept elements to this direction.