Variational Robust Kalman Filter

• Variational Robust Kalman Filter is a state estimation method that merges variational Bayesian inference with robust statistics to handle heavy-tailed and unknown noise.
• It employs moving-horizon estimation and coordinate ascent variational Bayes for joint inference over state trajectories and noise parameters.
• The approach offers enhanced stability and accuracy under outlier conditions, making it suitable for applications like target tracking and robotics.

A Variational Robust Kalman Filter (VRKF) is a class of state estimation methods that merge variational Bayesian inference with design principles from robust statistics and adaptive filtering within the Kalman filtering framework. These approaches address non-stationary, unknown, or heavy-tailed noise environments, aiming to preserve estimation accuracy and stability guarantees even when standard Gaussian assumptions are violated or noise statistics are mis-specified. VRKFs feature joint inference over state trajectories and noise parameters, robustification against outliers, principled adaptive updates, and (where relevant) finite-memory implementations suited for online, real-time scenarios (Dong et al., 2021, Li et al., 17 Dec 2025, Das et al., 2021, Zorzi, 2015).

1. Core Principles and Problem Setting

Variational robust Kalman filtering is rooted in Bayesian state-space modeling. The canonical problem is the discrete-time linear system: $x_{k+1} = F x_k + v_k,\quad y_k = H x_k + w_k$ where $v_k\sim \mathcal{N}(0,Q)$ (process noise) and $w_k\sim\mathcal{N}(0,R)$ (measurement noise). In robust/adaptive scenarios, the covariances $Q$ and $R$ may be unknown, non-constant, or non-Gaussian. The VRKF extends the joint inference to include $Q$, $R$, or even latent variables capturing scale/heavy-tailed effects.

The VRKF family uses a mean-field variational inference approach, approximating the posterior $p(x_{1:k}, Q, R\,|\,y_{1:k})$ by a tractable distribution $q(x_{1:k}) q(Q) q(R)$ or more general factorizations, and minimizing the KL divergence $\mathrm{KL}(q || p)$. The variational objective can be defined globally over full data or locally in receding-horizon or recursive forms (Dong et al., 2021).

Key robustification mechanisms include the explicit modeling of non-Gaussian, heavy-tailed noise via Student’s t, alpha-stable, or scale-mixture representations (Li et al., 17 Dec 2025, Hao et al., 2023), alternative divergence measures (such as those induced by the $\tau$-divergence minimax principle (Zorzi, 2015)), and de-weighting or inflation of the effective likelihood for detected outlier measurements (Li et al., 17 Dec 2025, Das et al., 2021).

2. Methodological Framework and Algorithms

2.1. Variational Bayes Coordinate Ascent

The principal inference routine employs coordinate ascent variational Bayes (VB) to iteratively update the distributions over state trajectories, process, and measurement noise. For a moving horizon of length $T$, the state distribution $q_X(X)$ remains Gaussian, while the noise covariances $q_Q(Q)$ and $q_R(R)$ are modeled as inverse-Wishart:

• $q_X(X) = \mathcal{N}(X; \hat{X}, P)$, where $X = x_{k-T}, \ldots, x_k$.
• $q_Q(Q) = \mathrm{IW}(Q; M, m)$, $q_R(R)=\mathrm{IW}(R; S, s)$, with closed-form updates for the sufficient statistics $M$ and $S$.

At each VB iteration, expectations $$\Phi=\mathbb{E}_{q_Q}[Q^{-1}], \Psi=\mathbb{E}_{q_R}[R^{-1}]$$ are computed and the mean and covariance of $X$ are updated via precision matrices. Monte Carlo integration with importance sampling is optionally used to propagate covariances for prediction steps (Dong et al., 2021).

2.2. Moving Horizon and Adaptive Filtering

A hallmark of advanced VRKF formulations is the use of moving-horizon estimation (MHE), processing data over a fixed window rather than the entire history. This ensures finite algorithmic memory and allows for efficient online operation. Priors for $x_{k-T}, Q, R$ are propagated forward, and forgetting factors can be used to favor recent information. After each MHE/VB cycle, estimates for $Q$ and $R$ are obtained by weighted averaging/sampling, and these are used in time updates for the next horizon (Dong et al., 2021).

2.3. Outlier and Heavy-Tailed Noise Modeling

To handle outliers and non-Gaussianity, VRKFs leverage:

• Marginalization over latent scale variables, leading to Student-t or alpha-stable induced losses; e.g., introducing auxiliary $\lambda_{i,k}$ per observed residual and placing inverse-Gamma priors so that marginalization recovers the Student-t likelihood (Li et al., 17 Dec 2025).
• Robust divergence constraints, such as the $\tau$-divergence minimax approach, which solves for a worst-case model increment in an allowed divergence ball around the nominal dynamics, and replaces the Riccati step with a nonlinear distortion (Zorzi, 2015).
• Variational outlier rejection: dynamically inflating the measurement covariance or down-weighting the impact of extreme residuals within the VB update or through adaptive loss-modifying mechanisms (Das et al., 2021).

2.4. Summary of Key Algorithmic Variants

Variant	Robustness Mechanism	Adaptivity Style
MHE-VB (Dong et al., 2021)	Online $Q, R$ estimation; windowed smoothing	Propagate priors, sample-based covariance update
ORKF (Das et al., 2021)	Latent scale/posterior of $R_k$ (Student-t)	Iterative VB update in EKF framework
$\tau$-VRKF (Zorzi, 2015)	Divergence-constrained minimax, time-varying gain	Risk-sensitive style, parameterized robustness
Unified VRKF (Li et al., 17 Dec 2025)	Student-t loss, fixed-point weight updates	Switches between robust and adaptive modes

3. Stability, Robustness, and Performance Guarantees

VRKFs are constructed to provide explicit stability guarantees despite the presence of unknown, time-varying, or adversarial noise statistics. In MHE-VB, Lyapunov-type induction arguments show that the posterior error covariance is uniformly bounded, and mean-square boundedness of the estimation error holds for any window length, any number of VB iterations, and any number of Monte Carlo samples (Dong et al., 2021). Similarly, in the $\tau$-divergence minimax formulation, robust Riccati updates guarantee that the posterior remains within an upper and lower bound explicitly controlled by the divergence budget $c_t$ and parameter $\tau$ (Zorzi, 2015).

Robustness to outliers is achieved mathematically by (i) tempering the influence function of the likelihood (e.g., Student-t loss grows only logarithmically for large errors), (ii) inflating the innovation covariance in the face of outliers, and (iii) using outlier-de-weighting rules within the VB fixed-point loops (Li et al., 17 Dec 2025, Das et al., 2021).

4. Application Domains and Experimental Outcomes

VRKF methodologies are applicable across high-integrity state estimation problems, especially where classical Kalman filtering fails due to unknown or non-Gaussian noise environments:

• Target tracking with unknown, time-varying process and measurement noise covariance, outperforming standard and EM-based adaptive filters in RMSE and convergence (e.g., simulated tracking benchmarks (Dong et al., 2021)).
• Mobile robotics and localization under wheel/inertial sensor outliers, where VB-based ORKFs down-weight odometry outliers and maintain localization accuracy, surpassing classical adaptive and robust (Huber, covariance-scaling) filters in planetary-analogue environments (Das et al., 2021).
• Robust estimation under model perturbations, demonstrated to outperform both standard and classical risk-sensitive Kalman filters under adversarial noise conditions (Zorzi, 2015).
• The unified VRKF framework can tune or switch between pure robust, pure adaptive, and standard KF modes, adapting automatically to both outlier and smoothly varying noise covariances in synthetic tracking and fault-diagnosis tasks (Li et al., 17 Dec 2025).

5. Connections and Theoretical Extensions

The variational robust Kalman filtering paradigm unifies several threads in probabilistic estimation:

• Bayesian adaptive KFs: online learning of noise covariances via conjugate priors and mean-field factorization, as opposed to point estimation in EM-based methods.
• Robust M-estimation: deploying heavy-tailed loss functions (e.g., Student-t, $\alpha$-stable) and their variational equivalents.
• Minimax filtering: direct optimization under worst-case divergence constraints (e.g., $\tau$-VRKF).
• Memory-efficient estimation: moving-horizon (MHE) and windowed smoothing ensure that model complexity is bounded regardless of data horizon, with explicit guarantees on stability.
• Outlier-detection and rejection: VB-outlier weights, dynamic covariance inflation, and switching rules for adaptivity-vs-robustness trade-offs.

6. Implementation Notes and Computational Aspects

VRKFs are algorithmically tractable, typically reducing to a finite sequence of linear-algebraic updates per time step. Key computational elements include:

• Solving block-tridiagonal (windowed) precision systems for trajectory smoothing.
• Closed-form or importance-sampled inverse-Wishart update formulas for $Q$, $R$ posterior approximations.
• Low-dimensional fixed-point or Newton solves for gain parameters in minimax robustification.
• Monte Carlo or analytic evaluations for outlier latent variable moments.
• Existence of provably convergent fixed-point updates and guarantees on bounded iteration count for practical convergence (Dong et al., 2021, Zorzi, 2015, Das et al., 2021).

For real-time or embedded applications, the moving-horizon and windowed-VB schemes reduce memory demands compared to full-smoothing (batch) approaches, and forgetting factors enable responsive adaptation to changing noise statistics.

7. Comparative Perspective and Impact

VRKFs surpass standard Kalman filtering under uncertainty and non-Gaussianity by combining (i) real-time full posterior inference for both state and noise statistics, (ii) robustification via heavier-tailed models or divergence constraints, and (iii) formal stability. Unlike pure EM or sliding-window VB adaptive KFs, moving-horizon VRKFs guarantee bounded memory, faster transient convergence, and resilience to drastic noise changes (Dong et al., 2021). Compared to point-estimate adaptive schemes, full VB filters naturally quantify posterior uncertainty. Against robust but nonadaptive methods, the adaptive VRKF remains efficient in normal operation while not succumbing to degraded estimator performance under mis-specified conditions (Li et al., 17 Dec 2025).

Simulation studies consistently evidence lower RMSE, improved outage performance, and faithful covariance tracking under both synthetic and real datasets, supporting their adoption in advanced estimation and sensor fusion applications (Dong et al., 2021, Das et al., 2021, Li et al., 17 Dec 2025, Zorzi, 2015).

---

References

• "A Variational Bayes Moving Horizon Estimation Adaptive Filter with Guaranteed Stability" (Dong et al., 2021)
• "A Comparison of Robust Kalman Filters for Improving Wheel-Inertial Odometry in Planetary Rovers" (Das et al., 2021)
• "Robust Kalman Filtering under Model Perturbations" (Zorzi, 2015)
• "Variational Robust Kalman Filters: A Unified Framework" (Li et al., 17 Dec 2025)