Robust Kalman Filter

• Robust Kalman filter is a state estimator that maintains stability and accuracy by accounting for model mismatches, noise distribution errors, and outliers.
• Distributionally robust methods using Wasserstein ambiguity sets and semidefinite programming optimize the worst-case mean-square error under uncertainty.
• Recent advancements include divergence-based approaches and heavy-tailed noise handling techniques, ensuring efficient and reliable estimation in adversarial settings.

A robust Kalman filter refers to any Kalman-type state estimator specifically designed to deliver high-fidelity, stable estimation despite violations of the classical Kalman assumptions, particularly model mismatch, noise distribution misspecification, or the presence of outliers. This includes approaches based on minimax optimality, distributional robustness, divergence-based uncertainty sets, heavy-tailed modeling, risk-sensitive estimation, and adversarial settings. Robustness is achieved through minimax recursions, semidefinite relaxations, statistical robustification, or explicit risk inflation, often at low additional computational overhead relative to the Kalman filter. Contemporary robust Kalman filters play a central role in control, robotics, signal processing, and safety-critical estimation due to their ability to provide nondivergent, stable state estimates in realistic, uncertain, or adversarially perturbed settings.

1. Foundations and Problem Formulation

Classical Kalman filtering assumes a linear time-invariant discrete-time model: $x_{k+1} = A x_k + w_k, \quad y_k = C x_k + v_k$ where $w_k$ and $v_k$ are assumed zero-mean Gaussian with known covariances. The filter is optimal in the MMSE sense only under these stringent assumptions. In robust filtering, noise distributions, covariances, or even system matrices are only approximately known. Robust Kalman filters aim to guarantee stability and controlled estimation error under (sometimes substantial) model misspecification.

The most general robust formulation is minimax: for a specified set of model perturbations (ambiguity set), design a filter that minimizes the worst-case mean-square error (MSE) over all admissible perturbations. A typical robust filtering objective is: $$\min_{K} \sup_{Q_w \in \mathcal{W}_w,\, Q_v \in \mathcal{W}_v} \; \mathbb{E} \left[ \| x_k - \widehat{x}_k \|^2 \right]$$ where $\mathcal{W}_w$, $\mathcal{W}_v$ are uncertainty (ambiguity) sets around the nominal process and measurement noise distributions (Jang et al., 31 Mar 2025). Ambiguity sets may be defined using divergence balls (e.g., Kullback-Leibler, Wasserstein distance, τ-divergence), or by allowing for outliers or heavy-tailed noise.

Robustness in this sense is fundamentally different from adaptivity: it corresponds to a guarantee against worst-case deviations within the prescribed set, rather than asymptotic parameter learning (Li et al., 17 Dec 2025).

2. Distributional Robustness via Wasserstein Ambiguity Sets

Recent advances focus on distributionally robust Kalman filtering (DRKF), where ambiguity sets are specified as 2-Wasserstein balls centered on nominal Gaussians: $$\mathcal{W}_w = \{ Q_w : W_2(Q_w, \mathcal{N}(0, \hat{\Sigma}_w)) \leq \rho_w \},\ \mathcal{W}_v = \{ Q_v : W_2(Q_v, \mathcal{N}(0, \hat{\Sigma}_v)) \leq \rho_v \}$$ with $W_2(\cdot,\cdot)$ the quadratic Wasserstein distance and $\rho_w$, $\rho_v$ radii set by robustness requirements (Jang et al., 31 Mar 2025, Jang et al., 6 Dec 2025).

Key advances include the reduction of the robust steady-state filter to a single convex semidefinite program (SDP), solved offline:

• The minimax steady-state estimator is realized by solving an SDP whose variables are the steady-state posterior, prior, and worst-case measurement noise covariances.
• The optimal robust steady-state gain is constant: $w_k$0
• Explicit constraints enforce that the worst-case MSE across all admissible noise laws remains bounded; trace and Gramian conditions ensure stability and convergence of the recursion.

The steady-state robust Kalman filter then operates as:

• Offline: Solve the SDP to obtain $w_k$1.
• Online: For all $w_k$2,
  • Time update: $w_k$3
  • Measurement update: $w_k$4

Convergence of the time-varying robust gain to the steady-state $w_k$5 is guaranteed under standard controllability/observability, with explicit bounds on the allowable process-noise ambiguity radius (Jang et al., 31 Mar 2025).

Numerical evaluations show that the DRKF outperforms classical and risk-sensitive filters in both Gaussian and non-Gaussian scenarios, with the smallest mean LQR cost and MSE across a range of ambiguity radii.

3. Divergence-Based and Minimax Kalman Filtering

Beyond Wasserstein balls, other robust Kalman filters employ divergence-based ambiguity sets, notably based on Kullback-Leibler or more general τ-divergence families (Zorzi, 2015, Yi et al., 2021, Yi et al., 2020). The general approach:

• At each increment, restrict allowable transitions to lie in a divergence ball (e.g., $w_k$6).
• The minimax solution inflates the Riccati covariance update according to a risk-sensitive term derived from the dual of the divergence constraint.
• For KL-divergence, the robust posterior covariance update is $w_k$7, where $w_k$8 is a time-varying risk sensitivity parameter determined implicitly by the allowable divergence (Yi et al., 2021, Yi et al., 2020).

These filters generalize to low-rank and degenerate scenarios, with convergence to unique fixed-points under reachability and observability.

Key theoretical results include:

• Existence and uniqueness of the robust Riccati iteration under strict contraction.
• Explicit demonstration that classical filters may fail dramatically under model mismatch, whereas robust versions maintain bounded error as long as the true system remains within the specified divergence ball.

4. Robust Filters for Heavy-Tailed and Outlier Noise

A parallel stream of research addresses robust Kalman filtering in the presence of outliers or heavy-tailed disturbances, departing from Gaussian assumptions (Saha et al., 2023, Wang et al., 2018, Wang et al., 2019, Ruckdeschel, 2010). Approaches include:

• Maximum Correntropy Kalman Filters (MCKF, RMCKF):
  • Replace quadratic loss with a (possibly weighted) sum of exponential terms (correntropy criterion), leading to recursive weighted updates that downweight large residuals.
  • Robustness is enhanced by adaptively selecting kernel bandwidths and incorporating risk-sensitivity, resulting in RMCKF, which dominates MCKF and classical KF in presence of both model mismatch and non-Gaussian noise (Saha et al., 2023).
• Mixture Correntropy CKF and Laplace/Gaussian Mixtures:
  • Measurement-fitting errors are regularized by non-quadratic, heavy-tail-robust mixture-correntropy induced loss. IRLS is used for iterative adaptation within a cubature Kalman filter framework, yielding performance improvements under inlier and outlier-dominated regimes (Wang et al., 2019).
• Heavy-tailed/Student’s t-based Kalman Filters:
  • Robust cost functions or losses derived from Student’s t-distributions confine the influence of extreme noise, with variational inference or fixed-point iterations providing tractable recursions. Adaptivity is introduced through switching between robust and adaptive steps, governed by innovation or covariance drift detection (Li et al., 17 Dec 2025).

Robust filters based on these criteria demonstrate consistently lower RMSE and superior track loss statistics under contaminated, impulsive, or adversarial noise, compared to classical KFs, risk-sensitive KFs, and conventional outlier-robust procedures.

5. Adversarial and Minimax Robustness

Models with adversarial corruption or worst-case noise insertions necessitate fundamentally robust estimation strategies:

• Sum-of-Squares (SoS) Smoothing for Adversarial Corruptions:
  • Estimates the state sequence by searching for the filter output that agrees with the data on a subset of likely uncorrupted measurements, using constraints and loss functions that are robust to a fraction $w_k$9 of arbitrary outliers. The SoS approach provides polynomial time algorithms with oracle-like excess error guarantees—$v_k$0—well beyond the scope of classical KFs (Chen et al., 2021).
  • Online robustification combines SoS filtering for cleaning and a second-stage KF, maintaining performance even with substantial adversarial contamination.
• Optimally Robust Filters for Additive and Innovation Outliers:
  • Filters are constructed to achieve minimax optimality in distributional neighborhoods allowing for additive (AO), innovation (IO), or mixed outliers, typically using Huberized innovation terms and windowed decision logic (Ruckdeschel, 2010, Fisch et al., 2020).
  • Hybrid schemes can address the distinction and simultaneity of AO and IO sequences, with rigorous breakdown-point analysis underlines their resilience.

These models articulate the gap between ad hoc outlier handling and provable, application-independent robust estimation, moving robust Kalman filtering into reliability-critical and security-sensitive domains.

6. Numerical Performance, Implementation, and Applications

Comprehensive empirical evaluations confirm the efficacy of robust Kalman filtering:

• Under both Gaussian and non-Gaussian (e.g., U-quadratic) system noise, robust filters maintain tight trajectory tracking and reduced mean-square error, outperforming time-varying or steady-state Kalman, risk-sensitive, and classical DRKF alternatives (Jang et al., 31 Mar 2025, Jang et al., 6 Dec 2025).
• Time to convergence of the DRKF is typically $v_k$1 iterations; offline SDP solves impose negligible computational cost relative to even large online filtering tasks.
• Robust Kalman filters are computationally efficient, with per-step complexity nearly matching that of standard KFs, and all key algorithms (offline SDP, IRLS, fixed-point, variational Bayes) are compatible with existing embedded and real-time control stacks.
• Application domains span robust MPC, LQR (trajectory tracking), distributed sensor fusion over unreliable networks, and model-based anomaly detection.

The parameter selection (ambiguity set radius, robust loss degree-of-freedom, adaptation rates) is informed by theoretically justified explicit stability and performance bounds, practical cross-validation, or direct operational constraints.

7. Theoretical Guarantees, Design Trade-offs, and Future Directions

Modern robust Kalman filtering offers a spectrum of theoretical guarantees, from contraction-based convergence proofs to explicit MSE bounds and minimax optimality. The essential design trade-off is between robustness margin (size of ambiguity set, risk sensitivity parameter) and nominal filter efficiency—robust filters pay a modest penalty when true noise remains in family, but dominate when mismatch, contamination, or adversarial deviations are present (Zorzi, 2015, Jang et al., 6 Dec 2025).

Emerging research is focused on:

• Further integration of robust filtering with distributed, cloud-based, or networked estimation architectures (He et al., 2019).
• Joint robust and adaptive strategies that can dynamically transition between minimax robustification and online adaptation as dictated by streaming performance criteria.
• Deeper analysis of performance under nonstationary, time-varying, or unknown model parameters, including probabilistic model uncertainty and heavy-tailed process noise (Li et al., 17 Dec 2025, Kim et al., 2020).
• Expanding policy-robust filtering to nonlinear systems, high-dimensional inference, and applications in learning-based/adaptive model predictive control.

Robust Kalman filtering, both conceptually and algorithmically, is now foundational in safety-critical autonomy and modern uncertain-data-driven control (Jang et al., 31 Mar 2025, Jang et al., 6 Dec 2025, Li et al., 17 Dec 2025, Chen et al., 2021, Yi et al., 2021).