ELEC4631 Chap.9 Observer Design, Separation Principle and Output Feedback
Observer Design, Separation Principle and Output Feedback
A Luenberger observer reproduces the plant model and corrects its estimate with the output innovation. With estimate xi and error e=x-xi, a consistent observer using y-Cxi-Du yields e-dot=(A-LC)e. Observability is therefore the exact permission for arbitrary observer-error pole placement, and duality lets controller algorithms be applied to (A^T,C^T) before transposing the gain. When state feedback must use the estimate, u=-Kxi, the augmented nominal dynamics become block triangular in state/error coordinates. The separation principle makes the combined eigenvalue multiset the union of the controller poles from A-BK and observer poles from A-LC, so K and L may be designed independently. This spectral result does not say infinitely fast observers are desirable: measurement noise, model mismatch and large correction gains remain practical boundaries.
What this chapter covers
- 01Full-order Luenberger observer
- 02Innovation y-Cxi-Du
- 03Error dynamics A-LC
- 04Observability permission and dual pole placement
- 05Estimated-state feedback u=-Kxi
- 06Block-triangular augmented dynamics
- 07Separation principle and practical pole choice
Design and verify an observer gain
- permission𝒪=[[1,0],[0,1]]=I, so rank two and (C,A) is observable.
- targetThe desired observer polynomial is (s+6)(s+7)=s^2+13s+42.
- error matrixA-LC=[[-l1,1],[-2-l2,-3]].
- polynomialIts characteristic polynomial is s^2+(l1+3)s+(3l1+2+l2).
- gainMatching gives l1=10 and l2=10.
- verificationA-LC=[[-10,1],[-12,-3]] has the requested roots -6 and -7, so estimation error decays.
Key terms
- Observer
- A dynamic system that combines the plant model, known input and output innovation to estimate an unmeasured state.
- Innovation
- Measured output minus predicted output, typically y-Cxi-Du; it drives correction through L.
- Estimation error
- The difference e=x-xi under the convention here. Its autonomous nominal dynamics are A-LC.
- Observer gain
- The n-by-p matrix L controlling how output mismatch corrects the state estimate.
- Output feedback
- A controller built from available output and an internal estimate rather than direct access to all state components.
- Separation principle
- The controller and observer nominal spectra can be designed independently; the augmented spectrum is their union.
Observer Design, Separation Principle and Output Feedback FAQ
Why does the input disappear from error dynamics?
Plant and observer use the same known Bu term and the same direct feedthrough in the predicted output, so subtraction cancels them. If the observer model or D term is inconsistent, rederive the error equation; cancellation is not automatic.
How does dual pole placement work?
Treat (A^T,C^T) as a controllable pair, design the transposed feedback gain for the desired observer polynomial, then transpose back to L. Check dimensions: L is n-by-p and the final matrix is A-LC.
Should observer poles always be much faster?
They are often chosen faster so estimation error decays before the controlled state, but extremely negative poles require large gains and can amplify measurement noise and model mismatch. Separation proves nominal eigenvalues, not an unrestricted speed rule.
What exactly does separation guarantee?
For the nominal linear model with the stated controller and observer, a coordinate change makes the augmented matrix block triangular, so its characteristic polynomial factors. It does not automatically guarantee robustness, noise attenuation, saturation avoidance or a particular eigenvector transient.
What is a complete observer-design answer?
Write the observer with the same known input and direct-feedthrough term as the plant. Define e=x-xi, subtract the equations and derive e-dot=(A-LC)e under the stated convention. Build 𝒪 and prove rank n. Expand the desired observer polynomial, design L through direct coefficient matching or the dual controllable pair, and verify the characteristic polynomial of A-LC. If output feedback is requested, add u=-Kxi, transform the augmented dynamics to [x;e] coordinates and state the union of controller and observer poles. End with a nominal convergence statement and one practical boundary. This sequence keeps plant, estimate and error roles distinct and makes a transposed gain or lost sign visible before it contaminates a larger augmented calculation.
How do D and noise change the observer discussion?
When y=Cx+Du, the predicted output must contain Cxi+Du so the known direct input contribution cancels from innovation. Omitting it injects a deterministic input term into the error equation. Measurement noise enters through the innovation and is multiplied by L, which explains why very fast observer poles can produce noisy estimates. Process/model error also prevents exact autonomous e-dot=(A-LC)e. The course’s basic separation result is nominal; a strong answer may state these boundaries without attempting a stochastic Kalman filter or robust-observer theory that the prompt did not request. In design, choose observer speed relative to controller dynamics and sensing quality, then verify dimensions L n-by-p and A-LC n-by-n.
What integrated observer practice is most valuable?
Choose one controllable and observable two-state plant. Design K for a slow controller pair and L for a faster observer pair, verify each polynomial separately and then write the four combined eigenvalues without expanding a fourth-order determinant. Simulate mentally or computationally two initial errors: e(0)=0 and e(0) nonzero. Then add D and show why the same Du term must appear in the predicted output. Ask an AI tutor to change C so observability fails and identify which step becomes impossible. This contrasts permission, design and separation more effectively than repeating isolated gain formulas, while leaving the analytical derivation as your own work.
Exam move
Keep three symbols distinct on every page: x for plant state, xi for estimate and e=x-xi. Write the observer equation including Du when direct feedthrough is present, then subtract to derive error dynamics. Prove observability before computing L and verify A-LC afterwards. Next, combine a previously verified K through u=-Kxi and derive the augmented dynamics in [x;e] coordinates; use block triangularity rather than expanding a large determinant. Practise one SISO and one multi-output dimension check. Finish by explaining the scope of separation and why faster observer poles trade convergence speed against noise sensitivity.
Working through Observer Design, Separation Principle and Output Feedback in ELEC4631? Sia is AskSia’s AI Engineering tutor — ask any ELEC4631 Observer Design, Separation Principle and Output Feedback question and get a clear, step-by-step explanation grounded in how ELEC4631 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
