the Kalman-Bucy filter
/ KAL-mun BYOO-see /
Filtering is the problem of estimating a hidden signal from noisy, partial observations as the observations arrive — in real time, using everything seen so far but nothing from the future. You have a signal X_t you cannot see directly, and an observation process Y_t that is a corrupted, noise-blurred function of X; the filtering problem is to compute the conditional distribution of X_t given the observation history F^Y_t, and in particular the best estimate (conditional mean) X-hat_t = E[ X_t given F^Y_t ]. The Kalman-Bucy filter is the complete, closed-form solution in the linear-Gaussian case, and it is one of the most-used algorithms in all of applied mathematics — in navigation, tracking, signal processing, and economics.
The setup is a linear-Gaussian state-space model: signal dX_t = A X_t dt + C dW_t and observation dY_t = H X_t dt + D dV_t, with W, V independent Brownian motions (Gaussian noise). Because everything is linear and Gaussian, the conditional distribution of X_t stays Gaussian for all time, so it is fully described by just two quantities: the conditional mean X-hat_t and the conditional covariance P_t. These evolve by closed-form equations. The mean obeys dX-hat_t = A X-hat_t dt + K_t ( dY_t - H X-hat_t dt ), where K_t = P_t H^T (D D^T)^{-1} is the Kalman gain and the bracket dY_t - H X-hat_t dt is the innovation — the part of the new observation not predicted by the current estimate. The covariance obeys a deterministic matrix Riccati equation, dP_t/dt = A P_t + P_t A^T + C C^T - P_t H^T (D D^T)^{-1} H P_t. Note that P_t does NOT depend on the observations — the uncertainty's evolution can be precomputed offline; only the mean is updated by data.
The filter is exact and optimal: in the linear-Gaussian world X-hat_t is genuinely the conditional expectation, the minimum-mean-square-error estimate, AND the conditional mode, all at once — there is no approximation. The structure (predict-and-correct via innovations, Riccati covariance) and the duality with the LQG control Riccati equation are deep and beautiful; together with certainty equivalence they give the complete solution to linear-Gaussian control under partial observation. Honest caveats: optimality among ALL estimators (not just linear ones) requires the Gaussian assumption — without it the Kalman-Bucy filter is still the best LINEAR estimator but no longer the conditional mean; for nonlinear or non-Gaussian models the conditional distribution is not finite-dimensional and one must turn to the Zakai/Kushner-Stratonovich equations or to particle filters. The extended Kalman filter (linearising a nonlinear model) is a heuristic, not exact, and can diverge.
A GPS receiver tracking a car: the hidden state X is position and velocity (which evolve by Newton's laws plus random acceleration), the observation Y is the noisy GPS fix. The Kalman gain K_t blends prediction and measurement: when the GPS is precise, K is large and the estimate trusts the new fix; in a tunnel (measurement noise huge), K shrinks and the estimate coasts on the motion model. The covariance P_t quantifies how sure the filter is at each instant.
The Kalman gain automatically weights model vs measurement by their relative noise — the filter's central trick.
Optimality among ALL estimators needs the Gaussian assumption; without it the Kalman-Bucy filter is only the best LINEAR estimator, not the conditional mean. For nonlinear models it is not exact — use Zakai/Kushner-Stratonovich or a particle filter, not the (heuristic, divergence-prone) extended Kalman filter.