Robust Kalman Filter
Also known as: RKF, heavy-tailed Kalman filter, outlier-robust Kalman filter, robust state estimation
The Robust Kalman Filter is an extension of the classical Kalman filter designed to maintain reliable state estimation when observations or process noise depart from the Gaussian assumption — particularly when data contain outliers, heavy-tailed distributions, or gross errors. By replacing or downweighting the standard least-squares update with influence-limited or M-estimation-based corrections, it prevents a single anomalous measurement from distorting the entire state estimate.
Key highlights
- Maintains accurate state estimates even when a fraction of observations are outliers or gross errors.
- Retains the recursive, computationally efficient structure of the standard Kalman filter.
- Applicable to real-time sequential data streams without storing the full history.
- Multiple robustification strategies (Huber, Student-t, bisquare) allow tuning the trade-off between efficiency and breakdown point.
- Can be extended to nonlinear systems via robust extended or unscented variants.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use the Robust Kalman Filter when tracking or estimating dynamic system states from time-series data that is likely contaminated by outliers, sensor malfunctions, or impulsive noise — such as GPS positioning, financial state-space models, or industrial process monitoring. It is also appropriate when the observation noise is known to be heavy-tailed rather than Gaussian. Do not apply it when the data are genuinely Gaussian and clean: the classical Kalman filter is optimal in that setting and computationally simpler. Avoid robust variants if interpretability of the weighting scheme is critical and the robustification mechanism cannot be clearly justified to stakeholders.
Strengths & limitations
- Maintains accurate state estimates even when a fraction of observations are outliers or gross errors.
- Retains the recursive, computationally efficient structure of the standard Kalman filter.
- Applicable to real-time sequential data streams without storing the full history.
- Multiple robustification strategies (Huber, Student-t, bisquare) allow tuning the trade-off between efficiency and breakdown point.
- Can be extended to nonlinear systems via robust extended or unscented variants.
- Robustification adds a tuning parameter (e.g., the Huber threshold c) that must be chosen, often heuristically.
- In iterative reweighted or variational implementations, each time step requires internal iterations, increasing computational cost.
- Performance degrades if the fraction of outliers exceeds the assumed breakdown point of the influence function.
- The state-space model structure (linearity, known F, H, Q, R) is still required; model misspecification is a separate issue not addressed by robustness alone.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does the Robust Kalman Filter differ from the standard Kalman filter?
The standard Kalman filter minimises mean squared error under Gaussian noise and can be heavily distorted by even a single outlier. The robust version applies an influence function to the innovation, capping the effect of any single anomalous measurement so the state estimate stays on track through corrupted data.
What is the Huber threshold and how do I choose it?
The Huber threshold c defines the boundary between quadratic (Gaussian-like) and linear (outlier-resistant) treatment of the innovation residual. Values around 1.345 times the expected innovation standard deviation are a common default that preserves about 95% asymptotic efficiency under Gaussian data while providing robustness. Cross-validation or knowledge of the expected outlier magnitude can guide the choice.
Can the Robust Kalman Filter handle nonlinear systems?
Yes. The robustification principle can be grafted onto the Extended Kalman Filter (linearised around the current estimate) or the Unscented Kalman Filter (sigma-point propagation), yielding Robust EKF and Robust UKF variants. Particle filters offer a fully nonlinear alternative with built-in non-Gaussian noise handling.
Is a Student-t observation model the same as a robust Kalman filter?
Effectively yes: modelling the observation noise as Student-t instead of Gaussian leads to an update step equivalent to iteratively reweighted least squares with Huber-like weights. Variational Bayes implementations of the Student-t Kalman filter make this connection explicit and also estimate the degrees-of-freedom parameter from the data.
What should I check after running a robust Kalman filter?
Inspect the innovation sequence and the sequence of weights assigned to each observation. Most weights near 1 with occasional values close to 0 indicate the filter is correctly identifying and downweighting genuine outliers. Persistently low weights or diverging covariance matrices suggest model misspecification rather than mere outliers.
Sources
- 1.Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82(1), 35-45.
- 2.Huber, P. J. & Ronchetti, E. M. (2011). Robust Statistics (2nd ed.). Wiley.ISBN 978-0470129906
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Robust Kalman Filter. ScholarGate. https://scholargate.app/bayesian/robust-kalman-filter