State-space models decompose a system into a state equation (how the state evolves) and an observation equation (how the state maps to measurements). The Kalman filter estimates the state from noisy observations by maintaining a Gaussian posterior distribution — mean and covariance — updated at each observation.
The ensemble Kalman filter replaces the covariance matrix with a sample: N particles, each representing a possible state, evolved forward and corrected against observations. The ensemble approximates the posterior without ever forming the full covariance matrix, which is essential for high-dimensional systems where the covariance matrix is too large to store.
The paper reframes ensemble methods as a control problem. Instead of viewing the ensemble as an approximation to a probability distribution, it views the ensemble as a collection of systems to be steered simultaneously — ensemble control. The control objective: make the ensemble consistent with the observations while maintaining diversity (the particles should explore the posterior, not collapse to a point).
The reframing reveals structure. The ensemble Kalman update is the solution to a specific optimal control problem: minimize the squared distance to the observations plus a regularization term that prevents ensemble collapse. Different regularizations produce different ensemble methods — the ensemble Kalman filter, the ensemble square root filter, and localized variants all correspond to different control objectives on the same ensemble.
The unifying lens: state estimation is not filtering but control. You're not passively incorporating observations; you're actively steering an ensemble to track the truth. The filter is a controller, and the controller has a cost function that explains why it works.