The Separation Principle
For linear systems, the optimal controller and optimal estimator can be designed independently and combined without loss of optimality.
Design in Two Independent Parts
The separation principle is one of the most useful results in control. It states that for a linear system, the problem of designing a state-feedback controller and the problem of designing a state estimator can be solved separately, and the two designs, joined together, still work correctly. The controller acts on the estimate as if it were the true state.
What it lets you do
- Design the feedback gain K assuming the full state is known, ignoring the estimation problem entirely.
- Design the observer or Kalman filter gain L assuming the control is fixed, ignoring the feedback problem.
- Combine them: use u = -K*x-hat, feeding the estimate into the same gain designed for the true state.
- The closed-loop poles are exactly the union of the controller poles and the observer poles.
Why the poles separate
When the controller and observer are combined, the closed-loop dynamics can be written in coordinates of state and estimation error. In these coordinates the system matrix becomes block-triangular, so its eigenvalues are simply the eigenvalues of A - B*K (the controller) together with those of A - L*C (the observer). The two sets do not interact, which is the precise statement of separation.
In the stochastic case
For linear systems with Gaussian noise, the separation principle takes a stronger form: the LQR gain and the Kalman filter, each optimal on its own, combine into the LQG controller that is optimal for the full output-feedback problem. Estimation and control decouple exactly.
Limits
Separation is a property of linear systems. For nonlinear systems it generally fails: the estimator and controller interact, and designing them independently gives no guarantee of a working combination. Nonlinear designs must often be treated jointly, though the principle still guides intuition and approximate designs.
The separation principle is what makes modern output-feedback design tractable. It reduces a coupled, high-dimensional problem into two familiar subproblems, each with mature solution methods.