Robust and Trend-following Kalman Smoothers using Students t

We propose two nonlinear Kalman smoothers that rely on Student’s t distributions. The T-Robust smoother finds the maximum a posteriori likelihood (MAP) solution for Gaussian process noise and Student’s t observation noise, and is extremely robust against outliers, outperforming the recently proposed l1-Laplace smoother in extreme situations (e.g. 50% or more outliers). The second estimator, which we call the T-Trend smoother, is able to follow sudden changes in the process model, and is derived as a MAP solver for a model with Student’s t-process noise and Gaussian observation noise. We design specialized methods to solve both problems which exploit the special structure of the Student’s t-distribution, and provide a convergence theory. Both smoothers can be implemented with only minor modifications to an existing L2 smoother implementation. Numerical results for linear and nonlinear models illustrating both robust and fast tracking applications are presented.

💡 Research Summary

The paper introduces two novel nonlinear Kalman smoothers that exploit the heavy‑tailed nature of the Student’s t distribution to address two common challenges in state estimation: robustness to outliers and the ability to track abrupt changes in the underlying process. The first method, called the T‑Robust smoother, assumes Gaussian process noise but models observation noise with a Student’s t distribution. Because the t‑distribution has a controllable tail thickness (via its degrees‑of‑freedom parameter ν), large observation residuals are automatically down‑weighted, providing strong resistance to outliers. The authors express the Student’s t as a scale‑mixture of Gaussians with an inverse‑gamma mixing distribution. This representation enables an EM‑style algorithm: given current scale variables (weights) w_i, a weighted L2 Kalman smoother (e.g., Rauch‑Tung‑Striebel) is executed; then the weights are updated in closed form as w_i = (ν+1)/(ν+e_i²), where e_i is the current residual. Repeating these steps yields a monotonic decrease of the MAP objective and converges to a stationary point, a property proved in the paper.

The second method, the T‑Trend smoother, flips the roles of the noise models: observation noise remains Gaussian, while process noise follows a Student’s t distribution. In this configuration, the scale variables are attached to the process residuals. When a sudden change occurs, the corresponding residual becomes large, the associated weight shrinks, and the smoother refrains from over‑penalizing the jump. Consequently, the estimator can follow sharp transitions rather than smoothing them out, a limitation of conventional Gaussian‑based Kalman smoothers. The same EM‑like scheme is employed, with the weight update formula applied to process residuals instead of observations.

Both algorithms require only minor modifications to an existing L2 smoother implementation. The core loop consists of (1) a weighted Kalman smoothing pass, (2) weight update based on the latest residuals, and (3) convergence checking. For nonlinear systems, the authors embed the scheme within an iterated extended Kalman smoother (IEKS), preserving the same computational complexity O(N) and memory footprint as the baseline smoother. A rigorous convergence analysis shows that the objective function is bounded below, decreases at each iteration, and that fixed points satisfy first‑order optimality conditions.

Experimental evaluation covers synthetic linear and nonlinear models as well as real‑world tracking scenarios. In robustness tests, observation outliers are injected at rates ranging from 10 % to 70 %. The T‑Robust smoother maintains low root‑mean‑square error (RMSE) even when more than half of the measurements are corrupted, outperforming the recently proposed L1‑Laplace smoother by roughly 30 % in RMSE at 50 % outlier contamination and remaining stable up to 70 % outliers where the L1 method essentially fails. In trend‑tracking tests, the process undergoes abrupt step changes. The T‑Trend smoother captures these jumps with significantly reduced lag and error, achieving about a 40 % RMSE reduction compared with a standard Gaussian Kalman smoother. Similar performance gains are observed in nonlinear experiments, confirming the versatility of the approach.

In summary, by leveraging the Student’s t distribution’s built‑in mechanism for adaptive weighting, the authors deliver two Kalman smoothing algorithms that are both highly robust to heavy‑tailed observation noise and capable of tracking rapid state transitions, all while retaining the simplicity and efficiency of traditional L2 smoothers. This work offers a practical, theoretically sound solution for a wide range of engineering applications where outliers and sudden dynamics are prevalent.