Iterative Linearization Refinement for Non-Linear State Estimation

40 concepts28 papers

State estimation is the challenge of guessing the hidden trajectory of a system from a sequence of noisy, indirect observations. When the physics governing the system and the sensors measuring it are linear, the classical Kalman filter provides an exact, optimal solution. But real-world systems—such as a vehicle projecting onto a camera lens or a smartphone tracking orientation—are rarely linear.

To handle non-linearities, traditional approaches like the Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF) rely on a single-pass approximation. The EKF takes a first-order Taylor expansion around its current prediction, while the UKF uses a set of deterministically chosen "sigma points" to probe the non-linear function. However, if the initial prediction is inaccurate or the non-linearity is severe, this single-pass approximation introduces massive errors. As G. Sibley et al. observe, non-iterative filters typically fail to reach the optimal Maximum Likelihood Estimate because they lock in the errors of a suboptimal linearization point.

Turning the Measurement Update into Local Optimization

The core intuition behind iterative state estimation is simple: turn the measurement update into a local optimization problem. Instead of accepting the first-pass linearization, the filter uses its newly computed estimate to re-evaluate and re-linearize the system at the exact same time step.

Mathematically, this process embeds a Gauss-Newton or Levenberg-Marquardt optimization loop directly inside the filter's update step (G. Sibley et al.; Zhaoming Li et al.). By repeatedly adjusting the state estimate using the latest measurement until convergence, the filter drives the estimate toward the true peak of the posterior distribution. In geometric tracking scenarios, such as projecting 3D models onto 2D image planes, Daniela Koller et al. demonstrated that this iterative refinement prevents the filter from becoming overly optimistic about its predictions, significantly improving feature-matching accuracy.

Standard EKF:  [State Prediction] ---> [Single Linearization] ---> [One-Shot Update]
                                                                          |
Iterated EKF:  [State Prediction] ---> [Linearize around Estimate] ---> [Update State]
                                             ^                                 |
                                             |__________(Iterate)______________|

This principle applies beautifully to derivative-free filters as well. The Iterated Sigma Point Kalman Filter (ISPKF) replaces rigid Jacobian matrices with sample-based statistical linearization within an iterative loop (G. Sibley et al.; Xiangmo Zhao et al.). By using feedback from previous iterations to dynamically reposition the sigma points, the estimator adaptively captures complex dynamics that one-shot methods miss entirely (Stefan Bernhard et al.). To ensure stability when models are ill-conditioned or noise is non-Gaussian, researchers have augmented these loops with Levenberg-Marquardt regularization (Rui Gao et al.) or robust maximum correntropy criteria (Badong Chen et al.).

Global Multi-Pass Smoothing

If iterating within a single time step improves local accuracy, what happens when we iterate across an entire trajectory? This is the domain of multi-pass smoothing.

In traditional smoothing, a filter processes data forward in time, and a smoother refines it backward. However, a single forward-backward pass still suffers from local linearization errors. By re-processing the entire motion sequence through multiple global forward-backward sweeps, the system can use global trajectory context to correct local misinterpretations. For instance, Santiago Cortés Reina et al. demonstrated that multi-pass smoothing drastically reduces drift in smartphone inertial navigation.

Similarly, for continuous-time systems governed by stochastic differential equations, Filip Tronarp et al. showed that repeatedly re-linearizing the system around the entire smoothed path rather than a single guess minimizes estimation errors, a technique highly effective for reconstructing complex flight paths (Jong‐Seok Lee). Even in particle-based architectures, structured multi-sweep sequences (such as forward-backward-forward architectures) ensure that linear and non-linear components are optimally balanced across the entire data sequence (Fredrik Lindsten et al.).

The Unified View: Expectation Propagation as Message Passing

Why does this multi-pass approach work so fundamentally well? A powerful theoretical connection reveals that classical engineering smoothers are actually special cases of a broader machine learning framework: Expectation Propagation (EP).

Sanket Kamthe et al. proved that standard non-linear smoothers (like the Extended or Unscented Kalman Smoothers) are mathematically equivalent to the very first iteration of an EP message-passing scheme. This means that traditional smoothers are inherently incomplete. By transforming state estimation into an explicit iterative message-passing process, forward, backward, and measurement messages are repeatedly updated across a time series until the posterior distribution fully converges (Marc Peter Deisenroth et al.; Sanket Kamthe et al.).

This iterative refinement allows the model to re-evaluate past states in light of future observations, providing tighter posterior estimates in high-uncertainty regimes. Furthermore, this closed-loop structure allows state estimation to mutually refine other latent variables, such as data association decisions in multi-target tracking (Hua Lan et al.).

Advanced Constraints and Optimization Frameworks

Iterative linearization also unlocks the ability to handle complex structural constraints and non-Euclidean geometries that break standard filters.

For example, incorporating physical constraints or $L_1$-regularization (to enforce sparsity or reject outliers) typically requires expensive batch optimization. To circumvent this, Rui Gao and Rui Gao et al. combined variable splitting methods like ADMM with the Iterated Extended Kalman Smoother. By repeatedly linearizing the non-linear dynamics around the current estimate within the splitting framework, they solve the primal update recursively. This achieves a computational complexity that scales linearly with time ($O(T)$), bringing batch-level optimization accuracy to massive, high-dimensional datasets.

When the state space is a non-Euclidean manifold, a single-pass filter struggles to transport uncertainty accurately. Pieter van Goor et al. showed that an iterative update strategy refines the parallel transport of covariance matrices across a manifold. By using successive approximations of the true state, the filter aligns measurement information accurately, accelerating convergence during the initial transient phase.

Remarkably, this principle extends even to non-convex optimization problems outside tracking, such as phase retrieval. Andrew S. Lan et al. demonstrated that iteratively applying a simple linear estimator—where each output serves as the linearization point for the next step—bridges the performance gap between fast linear methods and slow, high-precision semidefinite programming. Even in deep learning, architectures like the Mnemonic Descent Method mimic this behavior by formulating facial feature refinement as a non-linear dynamical system where a recurrent hidden state controls a trajectory of successive updates (George Trigeorgis et al.).

Key Foundational Work

Model-Based Object Tracking in Traffic Scenes (1992) — contributed 2 concept elements to this direction.
Bias Reduction and Filter Convergence for Long Range Stereo (2005) — 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_iterative_measurement_update_9c531123,
  title = {iterative measurement update},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/iterative-measurement-update/rdr_9c531123b19b43fd832f9dfebf1fbc0b},
  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_iterative_measurement_update_9c531123_v1,
  title = {iterative measurement update},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/iterative-measurement-update/rdr_9c531123b19b43fd832f9dfebf1fbc0b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake