CHAPTER 38 / Model-based experiments

Luenberger observers and separation

An observer runs a model beside the plant and corrects it with the measured output, recovering states that no sensor reports.

THE SHORT ANSWER

A Luenberger observer estimates unmeasured states by running a copy of the state model with the same applied input and correcting it with the output error y − Cx̂ through a gain L. With an exact model and no noise, the estimation error obeys e′ = (A − LC)e, so L sets how fast the error decays. Faster error poles remove an initial mismatch sooner but pass more measurement noise into the estimate.

x^˙=Ax^+Bu+L(y−Cx^),e˙=(A−LC) e\dot{\hat x}=A\hat x+Bu+L(y-C\hat x),\qquad \dot e=(A-LC)\,e
WHAT YOU WILL LEARN

Explain how a measurement error corrects a state estimate and why faster correction can amplify noise.

Pole placement assumed we knew the whole state. An observer fills in unmeasured states by comparing a model's prediction with sensor data.

Review the previous lesson: State feedback design: Ackermann's formula and the integral servo
START WITH AN EXAMPLE

First, picture it

A camera reports position, but your controller also needs velocity. Instead of taking a noisy difference of pictures alone, run a model of the moving cart and correct it when its predicted position disagrees with the camera.

THE IDEA

Build the idea step by step

An observer is a copy of the state model with a correction term. It receives the known applied input and predicts the state. Compare its predicted output with the sensor output. Multiply the difference by observer gain L and use it to correct the state estimate. Observability is what makes recovery of missing states possible in the ideal model. A faster correction may reduce an initial estimation error quickly, but also transmit more measurement noise.

Read the equations

x^˙=Ax^+Bu+L(y−Cx^)\dot{\hat x}=A\hat x+Bu+L(y-C\hat x)

Predict with the same model and input as the plant, then correct the estimate using measured output minus predicted output. This version has no direct input-to-output term.

What the symbols mean

x^\hat x
estimated state; the hat means estimate
A,B,CA,B,C
state, input, and output matrices
u,yu,y
known input and measured output
LL
observer correction gain; not the loop transfer function used earlier

The scalar example assumes a correct model and no noise. The lab estimates position and velocity from noisy position. Its observer is continuous-time and uses fixed pole-placement gains, not a Kalman covariance update.

WORKED EXAMPLE

Correct a scalar estimate

  1. Take a normalized scalar plant with rate minus state plus input. Measure the state directly for this first example. Choose observer gain 3.

    A=−1,B=C=1,L=3A=-1,\quad B=C=1,\quad L=3
  2. With the correct model and no noise, subtract the observer equation from the plant equation. The same input cancels.

    x~=x−x^,x~˙=(−1−3)x~=−4x~\tilde x=x-\hat x,\qquad\dot{\tilde x}=(-1-3)\tilde x=-4\tilde x
  3. An initial estimation error of 1 falls to about 0.018 after one second. This is the error decay, not necessarily the plant motion. Noise would continually drive the correction term.

    x~(1)=e−4≈0.0183\tilde x(1)=e^{-4}\approx0.0183
More detail and model assumptions
A CLOSER LOOK

A model prediction corrected by a measurement

A position sensor cannot report velocity directly. The observer runs a copy of the model: x̂′ = Ax̂ + Bu + L(y−C x̂). The term y−C x̂, called the innovation, compares measured position with predicted position. Its correction changes both the position estimate and the velocity estimate. The lab deliberately starts the estimate at the wrong position so that recovery is visible.

With an exact model, no noise and the same applied input, subtracting observer from plant gives e′ = (A−LC)e. For A = [[0,1],[-4,-0.8]] and C = [1,0], choose L₁ = 1.6ωo−0.8 and L₂ = ωo²−4−0.8L₁. This gives observer error denominator s²+1.6ωo s+ωo², with damping ratio 0.8. Unlike feedback K, which acts through the physical input, L injects a measurement correction into the internal estimate.

A CLOSER LOOK

Fast estimation is a tradeoff

For the ideal unsaturated observer-based loop, the combined eigenvalues are those of A−BK and A−LC. This separation principle makes design manageable. It does not say that transient peaks, saturation or noise are harmless. With measurement noise n, error dynamics include −Ln; large observer gains can magnify its effect on velocity estimates and the resulting control force.

First compare recovery with no noise and a deliberately wrong initial estimate. Then reset that initial mismatch and compare observer speed using the same seeded noise. This separates two effects that a single busy plot can hide. Kalman filtering introduces statistical process and measurement assumptions to choose correction gains; choosing a fast Luenberger observer is not the same procedure.

TRY IT

Connect the example to the lab

The worked example measures the state directly. In the connected experiment only position is measured, so the observer must also infer velocity; watch the estimation error and the force during the first second.

Open the connected experiment
YOUR EXPERIMENT

Make a prediction before moving a control.

  1. PREDICTOnly position is measured. Can the observer correct an initially wrong position estimate and infer velocity?
  2. CHANGE ONE THINGPin observer feedback with speed ratio 1. Raise only the ratio to 3. Compare position estimation error, estimated velocity and force during the first second. Then compare true-state feedback.
  3. OBSERVEPin the baseline, then compare the true and estimated states and the applied force over the first second.
  4. EXPLAINThe innovation y−C x̂ corrects a model prediction. L places the eigenvalues of A−LC; faster error dynamics remove this initial mismatch sooner but can produce a transient peak in estimated velocity. The observer receives the applied, clipped force, not an unavailable requested force. Separation of controller and observer poles assumes a linear, unsaturated exact model.
BE CAREFUL

The separation principle lets you choose controller and observer poles independently only for an exact, linear, unsaturated model. It does not make estimation transients, force saturation or measurement noise harmless.

Check your understanding+

In the worked example the estimation error falls from 1 to about 0.018 in one second. Does a faster observer always give a better estimate?

No. Faster error dynamics remove an initial mismatch sooner, but a larger L also passes more measurement noise into the estimate, because the error dynamics contain −Ln.
PRACTICE PROBLEMS

Solve these before opening the solutions

Work each problem on paper first. Open the answer to check yours, then the full solution if you need the steps, and mark how it went.

  1. The scalar plant below is unstable, and its sensor reads y = 2x. Choose the observer gain L so that the estimation error decays at rate 4.5 s⁻¹, and find the error after 1 s if the initial error is 1.

    x˙=0.5x+u,y=2x,x^˙=0.5x^+u+L(y−2x^)\dot x=0.5x+u,\quad y=2x,\qquad \dot{\hat x}=0.5\hat x+u+L(y-2\hat x)
    Show answer

    AnswerL = 2.5, giving error dynamics with rate −4.5. The error falls from 1 to about 0.0111 after 1 s, even though the plant itself is unstable.

    L=2.5,x~(1)=e−4.5≈0.0111L=2.5,\qquad \tilde x(1)=e^{-4.5}\approx0.0111
    Show full solution

    Solution

    1. Subtract the observer from the plant. The input cancels.

      x~˙=(0.5−2L)x~\dot{\tilde x}=(0.5-2L)\tilde x
    2. Place the error pole.

      0.5−2L=−4.5 ⇒ L=2.50.5-2L=-4.5\ \Rightarrow\ L=2.5
    3. Error after one second.

      x~(1)=e−4.5≈0.0111\tilde x(1)=e^{-4.5}\approx0.0111
    4. The observer converges because the correction term overpowers the unstable model mode in the error equation. The estimate follows a state that may itself be growing; stabilizing the plant is the controller's job.

  2. For the lab model below with position sensing, use the lecture's design rule with observer frequency ωo = 5 rad/s. Find L₁, L₂ and the observer poles, and verify the characteristic polynomial of A − LC.

    A=[01−4−0.8],C=[10],L1=1.6ωo−0.8,L2=ωo2−4−0.8L1A=\begin{bmatrix}0&1\\-4&-0.8\end{bmatrix},\quad C=\begin{bmatrix}1&0\end{bmatrix},\qquad L_1=1.6\omega_o-0.8,\quad L_2=\omega_o^2-4-0.8L_1
    Show answer

    AnswerL₁ = 7.2 and L₂ = 15.24. The observer poles are −4 ± j3, with ωo = 5 rad/s and damping ratio 0.8.

    L=[7.215.24],s=−4±j3L=\begin{bmatrix}7.2\\15.24\end{bmatrix},\qquad s=-4\pm j3
    Show full solution

    Solution

    1. Apply the design rule.

      L1=1.6(5)−0.8=7.2,L2=25−4−0.8(7.2)=15.24L_1=1.6(5)-0.8=7.2,\qquad L_2=25-4-0.8(7.2)=15.24
    2. Form the error matrix.

      A−LC=[−7.21−19.24−0.8]A-LC=\begin{bmatrix}-7.2&1\\-19.24&-0.8\end{bmatrix}
    3. Characteristic polynomial.

      det⁡(sI−A+LC)=(s+7.2)(s+0.8)+19.24=s2+8s+25\det(sI-A+LC)=(s+7.2)(s+0.8)+19.24=s^2+8s+25
    4. Roots and damping.

      s=−4±j3,ωo=25=5,ζ=82⋅5=0.8s=-4\pm j3,\qquad \omega_o=\sqrt{25}=5,\qquad \zeta=\frac{8}{2\cdot5}=0.8
  3. A cart is modelled as a double integrator with a position sensor, as below. (a) Find L that places both observer poles at −5. (b) Find K in u = −Kx̂ that places the controller poles at −1 and −2. (c) List the eigenvalues of the combined observer-based loop.

    A=[0100],B=[01],C=[10]A=\begin{bmatrix}0&1\\0&0\end{bmatrix},\quad B=\begin{bmatrix}0\\1\end{bmatrix},\quad C=\begin{bmatrix}1&0\end{bmatrix}
    Show answer

    Answer(a) L = [10 25]ᵀ. (b) K = [2 3]. (c) −1, −2, −5 and −5: by the separation principle the combined loop has the controller poles and the observer poles.

    L=[1025],K=[23],λ=−1, −2, −5, −5L=\begin{bmatrix}10\\25\end{bmatrix},\quad K=\begin{bmatrix}2&3\end{bmatrix},\quad \lambda=-1,\,-2,\,-5,\,-5
    Show full solution

    Solution

    1. Observer gain.

      A−LC=[−L11−L20],s2+L1s+L2=(s+5)2 ⇒ L1=10, L2=25A-LC=\begin{bmatrix}-L_1&1\\-L_2&0\end{bmatrix},\quad s^2+L_1s+L_2=(s+5)^2\ \Rightarrow\ L_1=10,\ L_2=25
    2. Controller gain.

      A−BK=[01−k1−k2],s2+k2s+k1=(s+1)(s+2) ⇒ k1=2, k2=3A-BK=\begin{bmatrix}0&1\\-k_1&-k_2\end{bmatrix},\quad s^2+k_2s+k_1=(s+1)(s+2)\ \Rightarrow\ k_1=2,\ k_2=3
    3. In coordinates (x, e) with e = x − x̂ the combined matrix is block triangular, so its eigenvalues are those of the two diagonal blocks.

      ddt[xe]=[A−BKBK0A−LC][xe]\frac{d}{dt}\begin{bmatrix}x\\e\end{bmatrix}=\begin{bmatrix}A-BK&BK\\0&A-LC\end{bmatrix}\begin{bmatrix}x\\e\end{bmatrix}
    4. Combined characteristic polynomial.

      (s2+3s+2)(s2+10s+25)=s4+13s3+57s2+95s+50(s^2+3s+2)(s^2+10s+25)=s^4+13s^3+57s^2+95s+50
  4. Exam level. Use the double-integrator observer of the previous problem with no noise. The estimate starts with a 0.1 m position error and no velocity error. (a) For observer poles at −5, −5 (L = [10 25]ᵀ), find the position and velocity estimation errors as functions of time and the peak velocity-estimate error. (b) Repeat for poles at −10, −10. (c) What does doubling the observer speed do?

    Show answer

    Answer(a) e₁ = (0.1 − 0.5t)e^(−5t) and e₂ = −2.5t e^(−5t); |e₂| peaks at 0.184 m/s at t = 0.2 s. (b) L = [20 100]ᵀ gives e₂ = −10t e^(−10t), peaking at 0.368 m/s at t = 0.1 s. (c) The error disappears twice as fast, but the transient velocity-estimate error doubles; a state-feedback controller would pass that peak into the force.

    ωo=5: ∣e2∣max⁡≈0.184 m/s;ωo=10: ∣e2∣max⁡≈0.368 m/s\omega_o=5:\ |e_2|_{\max}\approx0.184\ \mathrm{m/s};\qquad \omega_o=10:\ |e_2|_{\max}\approx0.368\ \mathrm{m/s}
    Show full solution

    Solution

    1. The errors obey ė₁ = e₂ − L₁e₁ and ė₂ = −L₂e₁, so e₁ satisfies a second-order equation with a double root. Use e₁(0) = 0.1 and ė₁(0) = e₂(0) − L₁e₁(0) = −1.

      e¨1+10e˙1+25e1=0,e1=(0.1+ct)e−5t,c−0.5=−1 ⇒ c=−0.5\ddot e_1+10\dot e_1+25e_1=0,\qquad e_1=(0.1+ct)e^{-5t},\qquad c-0.5=-1\ \Rightarrow\ c=-0.5
    2. Recover e₂ = ė₁ + 10e₁ and find its extremum.

      e2(t)=−2.5t e−5t,ddt(te−5t)=0 ⇒ t=0.2 s,∣e2∣max⁡=0.5e−1≈0.184 m/se_2(t)=-2.5t\,e^{-5t},\qquad \frac{d}{dt}\big(te^{-5t}\big)=0\ \Rightarrow\ t=0.2\ \mathrm s,\quad |e_2|_{\max}=0.5e^{-1}\approx0.184\ \mathrm{m/s}
    3. With poles at −10: L₁ = 20, L₂ = 100 and ė₁(0) = −2.

      e1=(0.1−t)e−10t,e2=−10t e−10t,∣e2∣max⁡=e−1≈0.368 m/s at t=0.1 se_1=(0.1-t)e^{-10t},\qquad e_2=-10t\,e^{-10t},\qquad |e_2|_{\max}=e^{-1}\approx0.368\ \mathrm{m/s}\ \text{at}\ t=0.1\ \mathrm s
    4. (c) The faster observer halves the time scale but doubles the peak: the large gain L₂ multiplies the initial position error and drives the velocity estimate hard. With measurement noise the same large gains would also inject more noise. Observer speed is a tradeoff, not something to maximize.

Mark each exercise after you check your answer. Marks stay in this browser.

KEY TERMS

Terms this lecture teaches

All terms in the glossary →
QUESTIONS STUDENTS ASK

Frequently asked questions

What is the separation principle?

For a linear, unsaturated, exact model, the closed loop that feeds back an observer estimate has the eigenvalues of A − BK together with those of A − LC. That lets you design the feedback gain and the observer gain separately, but it says nothing about transient peaks, saturation or noise.

Why not just differentiate the position measurement to get velocity?

Differentiating a noisy measurement amplifies the noise. An observer instead propagates a model with the known input and uses the measurement only as a correction, so velocity is inferred from how the predicted position disagrees with the sensor over time.

How fast should the observer poles be?

Fast enough that the estimate converges well before the controller relies on it, but not so fast that a large gain L carries sensor noise into the estimated velocity and the control force. Compare recovery from a wrong initial estimate and the response to the same seeded noise before choosing.

Should the observer use the requested input or the applied input?

The applied input. If the actuator clips the command, feeding the requested value into the observer model makes its prediction disagree with the real plant, and the correction term then works against a modelling error the observer created itself.

What is the difference between a Luenberger observer and a Kalman filter?

Both correct a model prediction with the innovation y − Cx̂. A Luenberger observer uses a gain chosen by placing the error poles, while a Kalman filter computes its gain from stated process and measurement noise covariances.

Why the next lesson follows

The next chapter turns observer design into a procedure: Ackermann's formula by duality, a smaller observer when some states are measured, and the exact conditions of the separation principle.

YOUR PROGRESS

Finished the reading and the exercises? Mark the lecture as done to tick it on the syllabus. Progress stays in this browser.

EIGENROOM · https://eigenroom.com/theory/luenberger-observer · Updated 2026-11-15