Uncertainty Propagation

The primary mathematical object for representing uncertainty. A covariance matrix quantifies the spread and correlation of errors in a multidimensional state estimate.

• Diagonal elements represent the variance (uncertainty) of each state variable independently.
• Off-diagonal elements represent the correlation between errors in different variables, which is critical for correctly propagating uncertainty through dynamic models.
• Maintaining a valid, positive semi-definite covariance matrix is a strict requirement for any consistent estimator.

Uncertainty propagation follows a strict two-step recursive loop, most famously formalized in the Kalman filter.

• Prediction Step: The state estimate and its covariance are projected forward in time using a process model. This step always increases uncertainty due to the injection of process noise.
• Correction Step: A new sensor measurement is incorporated. The Kalman Gain optimally weights the prediction against the measurement based on their relative covariances, reducing the overall uncertainty.
• This cycle ensures the system's confidence dynamically tracks the quality of available information.

The method for moving uncertainty through a system model depends entirely on the model's mathematical nature.

• Linear Systems: A state transition matrix A is applied directly to the covariance: P' = A P Aᵀ + Q. This is exact and computationally cheap.
• Nonlinear Systems: The model f(x) cannot be applied directly to a covariance. The Extended Kalman Filter (EKF) uses a Jacobian matrix for local linearization, while the Unscented Kalman Filter (UKF) propagates sigma points through the true nonlinear function for higher accuracy.
• Choosing the wrong propagation method leads to inconsistent estimates and filter divergence.

These two distinct covariance matrices are the tuning knobs that define the filter's trust in the model versus the sensors.

• Process Noise Covariance (Q): Represents the uncertainty injected by the system's own dynamics—how much the model's prediction is trusted. A high Q makes the filter more responsive to new measurements.
• Measurement Noise Covariance (R): Represents the uncertainty inherent in the sensor itself. A high R makes the filter trust the prediction more, smoothing out noisy data.
• The ratio of Q to R, not their absolute values, fundamentally determines the filter's steady-state gain and bandwidth.

A state estimator is consistent if its calculated covariance matrix accurately reflects the true distribution of the estimation errors.

• A filter is overconfident (inconsistent) if the true errors are larger than the covariance predicts, leading to filter divergence and a brittle system.
• A filter is conservative if the covariance is intentionally inflated to over-bound the true errors, a common safety strategy in safety-critical systems.
• The Normalized Innovation Squared (NIS) test is a standard statistical gate used online to monitor filter consistency and detect faults.

A hidden challenge that arises when fusing estimates from multiple local filters that have shared common process noise or prior information.

• Local track estimates become statistically correlated, but this cross-covariance is often unknown to the central fuser.
• Naively fusing them as independent leads to an overconfident, inconsistent global estimate—a phenomenon known as double-counting of information.
• Solutions include Covariance Intersection, which computes a consistent fused covariance without knowing the cross-correlation, or maintaining the full cross-covariance matrix in a centralized architecture.

A data fusion algorithm designed for decentralized sensor networks where the cross-correlation between estimates from different nodes is unknown or computationally intractable to maintain. Standard fusion would produce an overconfident, inconsistent result. Covariance intersection solves this by computing a weighted convex combination of the individual estimates, guaranteeing a consistent fused covariance that never underestimates the true error, even with unknown correlations.

A modern batch estimation framework that represents the uncertainty propagation problem as a bipartite graph. Variable nodes represent the unknown states (e.g., robot poses, landmark positions), and factor nodes represent probabilistic constraints derived from sensor measurements. Solving the graph via nonlinear least squares finds the maximum a posteriori (MAP) estimate by minimizing the error across all constraints simultaneously, naturally handling the full correlation structure.

A sequential Monte Carlo method that represents uncertainty non-parametrically using a set of weighted random samples called particles. Each particle represents a hypothesis of the system's true state. Unlike Kalman filters, particle filters can represent arbitrary, multi-modal probability distributions, making them ideal for highly nonlinear systems or when measurements are ambiguous. Uncertainty is visualized directly by the spread and density of the particle cloud.

A critical pre-deployment mathematical assessment that determines if the system's internal states can be uniquely inferred from the available sensor suite. A system is observable if the observability matrix has full rank. If a state is unobservable, its uncertainty will grow without bound regardless of sensor quality. This analysis directly informs sensor placement and identifies fundamental limitations in the fusion architecture.

A systematic framework that prevents corrupted measurements from contaminating the fused state estimate. FDI uses analytical redundancy—comparing sensor readings against model-based predictions—to generate residuals. Statistical tests on these residuals detect when a fault occurs and isolate the specific failed sensor. This ensures that uncertainty propagation remains valid by excluding erroneous data before it corrupts the covariance estimate.