Questions: Separation Principle and Output Feedback

Question 1

A control engineer designs state feedback u = -Kx with closed-loop poles at {-2, -4}, then designs a Luenberger observer with poles at {-10, -20}. She combines them as u = -Kx̂. What are the closed-loop poles of the resulting output-feedback system?

• A. {-2, -4} only — the faster observer poles vanish once estimation error converges
• B. {-10, -20} only — the observer poles dominate the combined dynamics
• C. {-2, -4, -10, -20} — the union of both sets, with no interaction between them
• D. New poles must be computed because substituting x̂ for x changes the eigenvalue problem

Answer: (C)

The separation principle guarantees that the combined output-feedback system has poles equal to the exact union of the state-feedback poles and the observer poles — with no coupling. The closed-loop is characterized by two independent subsystems: the state dynamics with feedback gain K, and the estimation error dynamics with observer gain L. Because the estimation error evolves independently of u, the two pole sets cannot mix. Recomputing poles after combination is unnecessary — this is precisely what the separation principle eliminates.

Question 2

Why can the controller gain K and observer gain L be designed independently under the separation principle?

• A. K and L appear in separate equations, so they trivially cannot interact in any dynamical system
• B. The estimation error dynamics ė = (A − LC)e depend only on L, not K — making error evolution independent of the control input
• C. Both K and L are chosen to minimize the same quadratic cost function, so they automatically decouple
• D. The separation holds approximately because observer poles are placed much faster than control poles

Answer: (B)

The mathematical key is that ė = (A − LC)e does not contain u or K. When you write the combined dynamics with state [x; e], the system matrix is block-triangular: the upper block (x dynamics) depends on K, and the lower block (e dynamics) depends only on L. Block-triangular matrices have eigenvalues equal to those of each diagonal block independently. This is not an approximation — it is exact for LTI systems regardless of how the observer and control poles are placed relative to each other.

Question 3

A practical rule of thumb is to place observer poles 2–5 times faster than the control poles because faster observer poles ensure the estimated state x̂ tracks the true state x before significant state changes occur during transients.

• T. True
• F. False

Answer: True

This is correct engineering practice. The separation principle guarantees combined stability regardless of relative pole speeds, but practical performance depends on estimation quality. If observer poles are too slow, x̂ lags behind x during transients, and the controller u = -Kx̂ acts on stale estimates. Placing observer poles 2–5× faster ensures the estimation error decays on a much shorter timescale than the controlled state, so x̂ ≈ x during most of the system's response.

Question 4

The separation principle holds for most dynamical systems — linear, nonlinear, and time-varying — as long as both the state feedback controller and the observer are individually designed to be stable.

• T. True
• F. False

Answer: False

The separation principle holds exactly only for linear time-invariant (LTI) systems. For nonlinear systems, the estimation error dynamics generally depend on the control input u, destroying the block-triangular structure that makes poles split cleanly. Combining a stable nonlinear controller with a stable nonlinear observer does not guarantee stability of the combined system — the separation into independent subproblems is a special property of linearity, not a universal truth.

Question 5

Explain why the estimation error e(t) = x(t) − x̂(t) evolves independently of the control gain K, and how this independence is what makes the separation principle work.

Model answer: The Luenberger observer evolves as x̂̇ = Ax̂ + Bu + L(y − Cx̂). Subtracting from the plant ẋ = Ax + Bu gives ė = (A − LC)e. The control input u and gain K cancel because u appears with identical coefficients in both the plant and the observer, so it drops from the error dynamics. Because ė = (A − LC)e depends only on L, the combined system [x; e] has block-triangular dynamics, and block-triangular matrices have eigenvalues equal to those of each diagonal block. This means the state-feedback poles (determined by K) and observer poles (determined by L) can be placed completely independently.

The cancellation of u in the error dynamics is the key algebraic fact. It happens because the observer uses the same input u as the plant — both x and x̂ receive the same forcing, so it cancels in the difference. Only the output injection term L(y − Cx̂) remains, correcting for initial estimation error. The resulting block-triangular structure is why 'separation' is the right word: the two design problems are mathematically separated, not just approximately decoupled.