A recursive Bayesian estimator for linear dynamical systems under Gaussian noise assumptions that produces minimum-variance estimates of the hidden state by alternating prediction (time update) and correction (measurement update) steps.

Principle

Propagate a Gaussian state estimate through linear dynamics and update that estimate by conditioning on new linear-Gaussian measurements; the Kalman update minimizes posterior covariance among linear unbiased estimators under the model assumptions.

Demonstration

Track an aircraft's position and velocity: use a linear motion model to predict state and covariance forward in time, then incorporate radar measurements with known measurement noise to compute the Kalman gain and correct the predicted state for an improved estimate.

Misapplication

Applying the standard Kalman filter to strongly nonlinear dynamics or non-Gaussian noise without modification, which yields biased or inconsistent estimates because the linear-Gaussian optimality conditions are violated.

Consequence

Under model assumptions the Kalman filter yields optimal sequential state estimates and covariances, enabling efficient real-time filtering, smoothing, and likelihood evaluation for parameter estimation.

Reversal

Nonrecursive batch estimators or particle filters operate differently: batch smoothing uses all data jointly for a posterior, while particle filters approximate arbitrary posteriors at greater computational cost and without Gaussian linearity constraints.

Boundary

Valid for linear state-space models with Gaussian additive noise or for locally linearized approximations; it excludes exact optimality when dynamics or observation models are nonlinear or noise significantly non-Gaussian without adaptations like EKF, UKF, or particle methods.

Semantic Tension

Tension exists between Kalman filtering and particle filtering or variational filters: Kalman methods prioritize tractability and closed-form covariance updates under Gaussianity, while particle methods trade analytic optimality for applicability to nonlinearity and non-Gaussianity.

Synthesis

The Kalman filter is a closed-form recursive algorithm that transforms prior Gaussian estimates through linear models and measurements into posterior Gaussians via a gain that minimizes uncertainty; it is computationally efficient and optimal under its linear-Gaussian assumptions but requires modification or replacement when those assumptions fail.