State estimation is the mathematical engine behind tracking systems, robotics, and navigation. At its core, it asks a simple question: given a noisy history of sensor readings and an imperfect physical model, what is the true state of the system right now? For decades, the standard answer was the Extended Kalman Filter (EKF). The EKF operates on a simple premise: if a system's dynamics or measurement models are non-linear, we can approximate them using a first-order Taylor series expansion—essentially drawing a tangent line at the current state estimate.
However, this local flat-world assumption introduces linearization-induced estimation errors. When a system exhibits high curvature, rapid dynamics, or severe uncertainty, treating a curved space as flat causes filters to underestimate uncertainty, introduce systematic bias, or diverge entirely.
To understand why linearization fails and how modern estimation theory overcomes it, we must look at the geometry of the failure itself.
The Geometry of Linearization Failure
Why does first-order linearization break down? The fundamental issue is that a tangent line only matches a curve at a single, infinitesimal point. As you move away from that point, the true function and its linear approximation drift apart. In state estimation, this manifests as three distinct structural failures.
The Jensen's Gap and Mean Shift
When a probability distribution is passed through a non-linear function, the mean of the transformed distribution is rarely equal to the function of the original mean ($E[f(x)] \neq f(E[x])$). As Dan MacKinlay et al. point out, first-order linearization completely ignores this "Jensen gap." By evaluating the Jacobian at the mean value alone, traditional filters drop the higher-order terms that describe how the variance of the input distribution interacts with the curvature of the function. This omission introduces a persistent mathematical bias into the state estimate.
Noisy Evaluation Points and Compounding Drift
A Taylor series expansion requires an expansion point. In recursive filtering, this point is the current state estimate. This creates a dangerous feedback loop: if the current state estimate is even slightly inaccurate, the Jacobian is calculated at the wrong location.
As Shiyu Bai et al. and Sebastian Curi et al. observe, calculating derivatives at a noisy estimate introduces mathematically inconsistent updates. The filter misinterprets the geometry of the state space, leading to "linearization drift" where small errors compound over time, ultimately causing the filter to lose track of the system entirely. This is especially severe during aggressive maneuvers (Angela P. Schoellig et al.), sharp state transitions (Arash Mehrjou), or complex coordinate transformations (Hua Lan et al.).
The Inability to Re-linearize
Standard recursive filters operate in a single forward pass. Once a measurement is processed, the linearization point is locked in. Patrick Geneva et al. emphasize that this inability to re-linearize past states means that errors introduced during initial, highly uncertain feature estimations cannot be corrected later, drawing a sharp contrast between traditional filters and iterative optimization methods.
Linearization (EKF):
True Curve: /‾‾‾\ Approximation: /‾‾‾\
/ \ / \
/ * \ *------- (Tangent line misses curvature)
^ Mean
Sigma-Point (UKF):
True Curve: /‾‾‾\ Approximation: /‾‾‾\
/ \ * * * (Points capture the true spread)
* * *
Paradigm 1: Deterministic Sampling (Sigma-Point Filters)
Rather than approximating a non-linear function analytically, what if we approximated the probability distribution instead? This is the core intuition behind the Unscented Transformation (UT), introduced by Eric A. Wan et al..
Instead of calculating complex Jacobians, an Unscented Kalman Filter (UKF) selects a small, deterministic set of points—called sigma points—that perfectly capture the mean and covariance of the current state distribution. These points are then propagated directly through the true, unaltered non-linear function.
This approach works because it is far easier to approximate a probability distribution than it is to approximate an arbitrary non-linear function. By passing discrete points through the actual system dynamics, the filter naturally captures how the curvature bends the distribution, effectively recovering higher-order Taylor series terms without ever calculating a derivative. This makes it highly effective for complex geometries like AGV navigation (Chuanqian Yuan et al.) and firefighter tracking (Jinbo Hu et al.).
However, sigma-point methods are not infallible. As Jakub Prüher et al. note, standard sigma-point filters rely on fixed quadrature rules that implicitly assume the non-linear functions behave like polynomials. When the underlying physics deviates from this assumption, uncalibrated numerical integration errors can still cause the filter to become overconfident and fail.
Paradigm 2: Iterative Refinement
Another way to combat linearization error is to turn the filtering step into a local optimization problem. If the primary failure mode of the EKF is evaluating the Jacobian at an inaccurate state estimate, we can mitigate this by iteratively updating the expansion point.
Algorithms like the Global Iterated Extended Kalman Filter (GIEKF) and the Iterated Sigma Point Kalman Filter (ISPKF) re-process the same measurement multiple times within a single time step. As Santiago Cortés Reina et al. demonstrate in smartphone navigation, re-linearizing the measurement model around the newly updated estimate allows the filter to gradually converge toward the true maximum a posteriori (MAP) state.
By bridging the gap between recursive filtering and Gauss-Newton optimization, G. Sibley et al. show that iterative filters can handle highly curved measurement functions and high state uncertainty where standard filters diverge.
Paradigm 3: Set-Theoretic and Bounded Uncertainty
When safety or absolute guarantees are required, stochastic approximations (like Gaussians) can be dangerous. An alternative approach abandons point-mass distributions entirely in favor of set-theoretic bounds.
Instead of tracking a mean and a variance, Box Particle Filtering uses multi-dimensional rectangular regions known as "box particles." As explained by Amadou Gning et al., these intervals are propagated through non-linear functions using inclusion functions.
Because interval analysis computes rigorous mathematical bounds rather than local point approximations, it completely eliminates linearization error. The resulting bounds are guaranteed to contain the true state, providing a highly reliable framework for handling non-Gaussian noise and data association ambiguity (Amadou Gning et al.), as well as minimizing the conservatism found in traditional set-membership filters (Dual Set Membership Filter With Minimizing Nonlinear Transformation of Ellipsoid).
Paradigm 4: Global and Non-Parametric Models
The most radical departure from local linearization is to model the system globally or non-parametrically, bypassing step-by-step Taylor expansions altogether.
- Continuous-Time Global Noise: Simo Särkkä et al. argue that continuous-time non-linear filters fail because they frequently re-project the state distribution back into a Gaussian shape at every infinitesimal step, discarding vital structural information. By approximating noise globally over an entire time interval using series expansions, filters can maintain tracking integrity in highly non-linear regimes.
- Gaussian Processes: Rather than linearizing a known but complex function, the system dynamics can be modeled non-parametrically. Marc Peter Deisenroth et al. demonstrate that using Gaussian Processes (GPs) allows smoothers to compute exact analytic predictions that account for full model uncertainty, preventing the divergence common to local linearizations.
- Data-Driven and Geometric Alignment: Modern approaches also leverage deep learning and differential geometry to bypass traditional linearization entirely. This includes using neural networks to learn complex plasma geometries (Christopher M. Bishop), utilizing variational sequential Monte Carlo for chaotic systems (Wessel L. van Nierop et al.), or morphing unobserved fields using differential geometry to preserve intrinsic physical structures (Yicun Zhen et al.).
Summary
Linearization error is not a minor numerical nuisance; it is a fundamental geometric limitation of first-order Taylor series expansions. Whether by sampling the distribution deterministically (Sigma-Point filters), optimizing the expansion point iteratively, bounding the state space via interval analysis, or modeling the system globally, modern state estimation focuses on a singular goal: respecting the true curvature of the world.