A Kalman filter estimates an evolving, partly hidden state as observations arrive. It performs Bayesian updating repeatedly rather than waiting for all data, maintaining a probability distribution over the current state. section 9.0 section 9.7
The filter follows a repeating predict–update cycle:
- Predict the next state and its uncertainty from the previous state. section 9.0
- Update that prediction when a noisy observation arrives, combining the prior prediction with the measurement through Bayes’ theorem. section 9.13
- Repeat the process so beliefs and uncertainty change as information arrives. section 9.0
The name applies when the underlying dynamics are linear and the observational noise is Gaussian. Under these assumptions, Gaussian distributions remain convenient for fast, closed-form updates. With several latent variables, means become vectors, covariances become matrices, and the Kalman gain becomes a vector. section 9.7 section 9.13
Tracking atmospheric CO₂
A univariate model can treat CO₂ concentration as a random walk, but a multivariate filter tracks both concentration and its yearly slope. This produces forecasts that increase linearly, avoids underestimating concentration, and better follows the observed trend. Adding another state for changes in the slope can provide flexibility for quadratic trends. section 9.12 section 9.13
Use the observation equation to see how a noisy measurement relates to the hidden state. section 9.7
