Adaptive Kalman Filter with Measurement Trust

• The approach adapts the measurement noise covariance (R_k) online to modulate sensor trust and ensure optimal filtering performance.
• It employs Riemannian trust-region and autocovariance methods to update noise statistics, guaranteeing stable and statistically consistent estimation.
• Empirical studies confirm improved convergence, reduced estimation errors, and robust performance compared to fixed-covariance filters.

A Kalman Filter with adapting measurement trust is a class of adaptive state estimation algorithms that dynamically adjust the measurement noise covariance matrix $R_k$ online, thereby modulating the filter’s reliance (“trust”) on new sensor data. This mechanism is central to achieving optimal filtering performance in scenarios with unknown, varying, or nonstationary measurement noise statistics. Adaptive measurement trust plays a critical role in linear and nonlinear filtering, robustness to noise model mismatch, and online system identification. Below is a comprehensive exposition of core methodologies, mathematical frameworks, and convergence properties, drawing from recent advances in the literature.

1. Problem Formulation and Measurement Trust Principle

The canonical setting is a discrete-time state-space model:

$$x_{k+1} = F_k x_k + G_k u_k + w_k,\quad y_k = H_k x_k + v_k,$$

where process noise $w_k\sim\mathcal{N}(0, Q)$ and measurement noise $v_k\sim\mathcal{N}(0, R)$ are mutually uncorrelated. The key innovation in adaptive Kalman filtering is treating $Q$ and, in particular, $R$ as unknown and updating them recursively using measurement data. The measurement noise covariance $R_k$ directly modulates measurement trust: lower $R_k$ values increase reliance on sensor data, higher values down-weight the measurement channel.

Adaptation of $R_k$ is crucial when sensor characteristics are unknown, time-varying, or when the noise is non-Gaussian, censored, or contaminated by outliers. Successful estimation of $R_k$ permits the filter to maintain statistical consistency and minimize state estimation error, even in adverse or nonstationary environments (Moghe et al., 2021).

2. Autocovariance-Based and Riemannian Trust-Region Approaches

The Riemannian Trust-Region (RTR) based Adaptive Kalman Filter provides a theoretically principled methodology for online estimation of the process and measurement noise covariances in observable linear time-varying systems (Moghe et al., 2021). This approach employs the measurement-difference autocovariance method: by stacking $m+1$ past outputs, eliminating the state, and constructing suitable lagged autocovariances, a linear least squares cost function is derived,

$$J_k(\theta) = \frac{1}{2}(D_k \theta - b_k)^T R_W^{-1} (D_k\theta - b_k) + \frac{1}{2}(\theta - \Theta_{k-1})^T \Psi_{k-1}^{-1} (\theta - \Theta_{k-1}),$$

with $\theta$ vectorizing vech($Q$), vech($R$).

The adaptive update is cast as a trust-region optimization on the manifold of symmetric positive definite (SPD) matrices, utilizing the affine-invariant Riemannian metric to preserve the symmetry and positive definiteness of $Q$ and $R$. The local quadratic model is minimized via truncated conjugate gradient within a geodesic ball, and the step is accepted or rejected based on the ratio of actual to predicted reduction. This ensures every $R_k$ (and $Q_k$) estimate is valid, sidestepping the instability of simpler Euclidean-based iterative methods.

Under uniform observability and controllability, and with sufficient excitation, the recursive estimates converge in probability to the true covariance matrices, and the corresponding adaptive Kalman filter is provably exponentially stable and asymptotically equivalent to the optimal filter with known $Q$ and $R$ (Moghe et al., 2021).

3. Alternative Adaptation Mechanisms: Heuristic Recipes, Nonparametric Methods, and Robust Filters

Multiple frameworks complement the Riemannian manifold approach:

• Reference Recursive Recipe (RRR): Uses multiple forward-backward passes through the data (with Extended Kalman Filter and Rauch–Tung–Striebel smoothing), alternately refining $R$ (and optionally $Q$) by empirical covariance formulas applied to innovations, filtered residuals, or smoothed residuals:

$$R \gets \frac{1}{N}\sum_k [ (z_k - h(x_{k|N})) (z_k - h(x_{k|N}))^T + H_k P_{k|N} H_k^T ].$$

This is iterated until convergence in $R$, with generalized cost functions (e.g., innovation cost $J_1$) evaluated to ensure statistical consistency (Ananthasayanam et al., 2015, M et al., 2015).

• Nonparametric Jackknife Ensemble Estimation: Constructs an ensemble of state estimates via jackknife LSQ on subsampled measurements, then estimates $R_k$ as the difference between the out-of-sample residual covariance and the prediction-error covariance. This approach is theoretically consistent and asymptotically normal, with convergence rate $o(1/k)$ (Busch et al., 2014).
• Robustification via M-Estimation: For measurement noise with heavy tails or outliers, robust Cubature Kalman Filters utilize Huber's M-estimation to adapt measurement trust on each channel, down-weighting large residuals by scaling the innovation covariance (Li et al., 2019).
• Nonlinear and Non-Gaussian Settings: Adaptive Unscented Kalman Filters integrate error-entropy or correntropy criteria into measurement updates, and use online Sage–Husa recursions or Gaussian mixture models to adapt $R_k$, improving robustness to impulsive, multimodal, or heavy-tailed noise (Tian et al., 2023, Gong et al., 2023).

4. Algorithms and Implementation Strategies

Several algorithmic strategies have been formalized for practical realization of adapting measurement trust:

• Exponential Forgetting (ROSE-Filter, Adaptive Covariance): Updates $R_k$ via exponential windows:

$$R_k = \gamma\,\alpha_R\, (E[y_k] - y_k) (E[y_k] - y_k)^T + (1-\alpha_R) R_{k-1}$$

with tuning parameters controlling reactivity. Auxiliary KFs can provide $E[y_k]$ in nonlinear settings (Marchthaler, 2021).

• Fading-Gain and Strong-Tracking Filters: Apply "fading factors" or scale $R_k$ downward when residuals are small, thereby increasing the Kalman gain and filter agility in response to high-quality measurements (Narasimhappa, 2021).
• Online Maximum Likelihood Estimation: Recursive updates for $R_k$ are derived from one-step or sliding-window innovation statistics, in both Gaussian and censored measurement frameworks (e.g., Adaptive Tobit KF) (Chiariotti, 2019).
• Variational Bayes Moving Horizon Estimation: Models $R$ as a random matrix with an inverse-Wishart posterior on a finite window for enhanced tracking and stability guarantees, extracting point estimates via Monte Carlo importance sampling (Dong et al., 2021).

All procedures require maintaining positive definiteness of $R_k$ at each step, commonly by eigenvalue flooring, Cholesky re-factorization, or Riemannian retraction.

5. Theoretical Properties: Convergence and Stability

Adaptive Kalman filters with measurement trust adaptation exhibit rigorous convergence and stability properties under broad conditions:

• Convergence: Under persistent excitation, uniform observability, and boundedness assumptions, the adaptive measurement-noise estimate $\hat{R}_k$ converges almost surely (a.s.) to the true $R$ for stationary cases (Moghe et al., 2021, Busch et al., 2014, Ananthasayanam et al., 2015). For non-stationary or time-varying noise, windowed or forgetting-factor versions ensure tracking of slowly varying $R_k$.
• Stability: Proper adaptation guarantees the error covariance sequence remains uniformly bounded and the filter is exponentially stable, as established via Lyapunov arguments and error propagation in both standard and moving-horizon VB frameworks (Moghe et al., 2021, Dong et al., 2021).
• Robustness: In the presence of outliers and non-Gaussian stochasticity, robust adaptation methods such as Huberization or MEEF-based updates recuperate filter consistency and maintain bounded estimation error, outperforming non-adaptive or parametric alternatives (Li et al., 2019, Tian et al., 2023).

6. Practical Applications and Numerical Evidence

Adaptive measurement trust mechanisms have been validated across distinct application domains:

• Time-Varying and Nonlinear Systems: Adaptive strategies maintain low error and consistency in 2D/3D oscillators and vehicle tracking with time-varying system matrices and measurement channels, attaining performance equivalent to filters with oracle knowledge of $Q$ and $R$ (Moghe et al., 2021, Marchthaler, 2021).
• Robust Target Tracking: In impulsive noise or censored measurement regimes (e.g., IoT navigation with sensor saturation), adaptive schemes (ATKF, TGKF) avoid divergence observed in non-adaptive KFs and exhibit superior RMSE scores (Chiariotti, 2019, Gong et al., 2023).
• Real-World Data: In flight test data, adaptive R-estimation ensures that residual cost functions settle to the expected measurement dimension, producing definitive parameter estimates and correct confidence bounds, even under coupled state-parameter dynamics and non-ideal noise (M et al., 2015).

Empirical studies consistently demonstrate that filters with adapting measurement trust outperform fixed-covariance KFs and classic adaptive-R approaches, attaining rapid convergence and resilience to model misspecification.

7. Comparative Summary of Adaptive Measurement Trust Algorithms

Method	Measurement Trust Mechanism	Application Scope
Riemannian Trust-Region (RTR)	Manifold-constrained R update	LTV/LTI, strict SPD preservation
Reference Recursive Recipe	Full-data, smoothed residuals	EKF, joint parameter and covariance estimation
Jackknife Ensemble	Nonparametric, cross-validated	Nonlinear, streaming, no prior distribution required
Fading/Strong-Tracking	Innovation/residual scaling	Nonlinear, unknown and time-varying noise
Robust M-Estimation	Huberized gain, per-channel	Non-Gaussian/outlier-rich environments
GMM/Entropy-Based (TGKF)	Mixture-based, per-regime	Heavy-tailed, impulsive, or multimodal measurement
Variational Bayes MHE	Inverse-Wishart/moving window	Linear/nonlinear, guaranteed mean-square boundedness

Each methodology provides a framework for adapting measurement trust, tailored to underlying system structure, computational resources, and the statistical nature of noise and measurement data.

---

All claims, algorithmic procedures, and theoretical guarantees are taken directly from (Moghe et al., 2021, Narasimhappa, 2021, Busch et al., 2014, Chiariotti, 2019, Ananthasayanam et al., 2015, Li et al., 2019, Marchthaler, 2021, Tian et al., 2023, Gong et al., 2023, Dong et al., 2021), and (M et al., 2015). These works collectively establish adapting measurement trust as an essential and mature component of modern Kalman filtering.