CHAPTER 39 / STATE ESTIMATION

Observer design: duality, Ackermann's formula and the minimum-order observer

An observer gain is a state-feedback gain for the transposed system. This chapter computes it with Ackermann's formula, builds a smaller observer when some states are measured, and states exactly when the controller and the observer can be designed separately.

THE SHORT ANSWER

Observer design chooses the gain L so that the estimation error of a Luenberger observer, ė = (A − LC)e, decays with chosen eigenvalues. By duality this is state-feedback design for (Aᵀ, Cᵀ), so the dual Ackermann formula L = φ_o(A)M_o⁻¹[0 … 0 1]ᵀ applies whenever the observability matrix M_o is invertible. A minimum-order observer estimates only the unmeasured states, and the separation principle lets controller and observer be designed separately for an exact linear model.

L=φo(A) Mo−1[0⋮01],Mo=[CCA⋮CAn−1],e˙=(A−LC)eL=\varphi_o(A)\,M_o^{-1}\begin{bmatrix}0\\\vdots\\0\\1\end{bmatrix},\qquad M_o=\begin{bmatrix}C\\CA\\\vdots\\CA^{n-1}\end{bmatrix},\qquad \dot e=(A-LC)e
WHAT YOU WILL LEARN

Compute a full-order observer gain by the dual Ackermann formula, design a minimum-order observer for the states a sensor does not measure, and give the closed-loop eigenvalues of an observer-based controller.

The Luenberger observer chapter explained why a model copy corrected by the measurement recovers hidden states, and the previous state-feedback chapter gave a procedure for placing poles. Here the two meet: the same procedure, applied to the transposed system, designs the observer.

Review the previous lesson: Luenberger observers and separation
START WITH AN EXAMPLE

First, picture it

A Luenberger observer runs a copy of the model and steers it with the measurement error y − Cx̂. Its estimation error obeys ė = (A − LC)e, so designing L means placing the eigenvalues of A − LC. That is the state-feedback problem with the matrices in the wrong places: L stands to the left of C, whereas K stood to the right of B. Transposing fixes the order, and a matrix has the same eigenvalues as its transpose. So every tool from the state-feedback chapters, Ackermann's formula included, works on the transposed problem. A second idea saves effort. If a sensor already reports some states, estimating them again is wasted work, and a smaller observer can estimate only the rest.

THE IDEA

Build the idea step by step

Write the estimation error as e = x − x̂. With an exact model and the same input applied to plant and observer, ė = (A − LC)e, whatever the input does. The pair (A, C) is observable exactly when the observability matrix M_o = [C; CA; …; CAⁿ⁻¹] is invertible, and then the eigenvalues of A − LC can be placed anywhere. Applying Ackermann's formula to (Aᵀ, Cᵀ) and transposing the result gives L = φ_o(A)·M_o⁻¹·[0 … 0 1]ᵀ. When the sensor measures part of the state directly, split the state into the measured part x_a = y and the unmeasured part x_b. The equations of x_b form a smaller system, and the measured equation ẋ_a − A_aa·y − B_a·u = A_ab·x_b acts as its measurement. An observer of order n − m estimates x_b from it. Finally, the controller u = −Kx̂ and the observer together have the eigenvalues of A − BK and those of the observer, so each part can be designed on its own. This separation principle holds for the exact linear model; it does not promise robustness.

Read the equations

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

Full-order observer and its error dynamics. The input cancels from the error equation when plant and observer receive the same u.

det⁡(sI−A+LC)=det⁡ ⁣(sI−AT+CTLT)\det(sI-A+LC)=\det\!\left(sI-A^{\mathsf T}+C^{\mathsf T}L^{\mathsf T}\right)

Duality. Observer design for (A, C) is state-feedback design for (Aᵀ, Cᵀ) with the gain Lᵀ.

Mo=[CCA⋮CAn−1],L=φo(A) Mo−1[0⋮01]M_o=\begin{bmatrix}C\\CA\\\vdots\\CA^{n-1}\end{bmatrix},\qquad L=\varphi_o(A)\,M_o^{-1}\begin{bmatrix}0\\\vdots\\0\\1\end{bmatrix}

Observability matrix and the dual Ackermann formula. φ_o is the desired observer polynomial.

Mo v=[0⋮01],L=Anv+αn−1An−1v+⋯+α1Av+α0vM_o\,v=\begin{bmatrix}0\\\vdots\\0\\1\end{bmatrix},\qquad L=A^nv+\alpha_{n-1}A^{n-1}v+\cdots+\alpha_1Av+\alpha_0v

The hand shortcut: one column v, then repeated matrix-times-column products.

[x˙ax˙b]=[AaaAabAbaAbb][xaxb]+[BaBb]u,y=xa\begin{bmatrix}\dot x_a\\\dot x_b\end{bmatrix}=\begin{bmatrix}A_{aa}&A_{ab}\\A_{ba}&A_{bb}\end{bmatrix}\begin{bmatrix}x_a\\x_b\end{bmatrix}+\begin{bmatrix}B_a\\B_b\end{bmatrix}u,\qquad y=x_a

Partition for the minimum-order observer: x_a is measured, x_b is not.

e˙b=(Abb−LrAab) eb\dot e_b=(A_{bb}-L_rA_{ab})\,e_b

Error of the minimum-order observer. Design L_r as an observer gain for the pair (A_bb, A_ab).

η˙=(Abb−LrAab)η+[(Abb−LrAab)Lr+Aba−LrAaa]y+(Bb−LrBa)ux^b=η+Lry\begin{aligned}\dot\eta&=(A_{bb}-L_rA_{ab})\eta+\big[(A_{bb}-L_rA_{ab})L_r+A_{ba}-L_rA_{aa}\big]y+(B_b-L_rB_a)u\\ \hat x_b&=\eta+L_ry\end{aligned}

The implementable minimum-order observer. The variable η = x̂_b − L_r·y avoids differentiating the measurement.

[x˙e˙]=[A−BKBK0A−LC][xe]\begin{bmatrix}\dot x\\\dot e\end{bmatrix}=\begin{bmatrix}A-BK&BK\\0&A-LC\end{bmatrix}\begin{bmatrix}x\\e\end{bmatrix}

Separation principle. The block-triangular form shows that the closed-loop eigenvalues are those of A − BK and those of A − LC.

U(s)Y(s)=−K (sI−A+BK+LC)−1L\frac{U(s)}{Y(s)}=-K\,(sI-A+BK+LC)^{-1}L

The observer-based controller seen from the measurement, with r = 0: a dynamic compensator of order n.

What the symbols mean

x^, e\hat x,\ e
state estimate and estimation error x − x̂
LL
full-order observer gain, a column with one entry per state
MoM_o
observability matrix [C; CA; …; CAⁿ⁻¹]
φo(s), αi\varphi_o(s),\ \alpha_i
desired observer polynomial and its coefficients
vv
last column of M_o⁻¹, found from M_o·v = [0 … 0 1]ᵀ
xa, xbx_a,\ x_b
measured part of the state (y = x_a) and the part to be estimated
Aaa,Aab,Aba,Abb,Ba,BbA_{aa},A_{ab},A_{ba},A_{bb},B_a,B_b
blocks of A and B in that partition
LrL_r
gain of the minimum-order observer
η\eta
minimum-order observer state, η = x̂_b − L_r·y
mm
number of measured states
KK
state-feedback gain applied to the estimate

The formula needs a single output, so that C is a row, and an observable pair (A, C). With several outputs L is not unique, and the minimum-order observer then has free gains. All error equations assume that the observer uses the plant's own matrices and the input actually applied to the plant.

WORKED EXAMPLE

Observers for the positioning drive

  1. Take the drive from the state-feedback design chapter, now with an angle sensor only. The controller K = [80, 35, 11] places −2 ± j2 and −10, so it needs speed and torque estimates.

    A=[0100−1100−2],B=[001],C=[100]A=\begin{bmatrix}0&1&0\\0&-1&1\\0&0&-2\end{bmatrix},\quad B=\begin{bmatrix}0\\0\\1\end{bmatrix},\quad C=\begin{bmatrix}1&0&0\end{bmatrix}
  2. Check observability. The determinant is 1, so every observer pole can be placed.

    Mo=[CCACA2]=[1000100−11],det⁡Mo=1M_o=\begin{bmatrix}C\\CA\\CA^2\end{bmatrix}=\begin{bmatrix}1&0&0\\0&1&0\\0&-1&1\end{bmatrix},\qquad \det M_o=1
  3. Choose all three observer poles at −6, three times the dominant real part of the controller.

    φo(s)=(s+6)3=s3+18s2+108s+216\varphi_o(s)=(s+6)^3=s^3+18s^2+108s+216
  4. Dual Ackermann, hand version. Solve M_o·v = [0, 0, 1]ᵀ, then multiply by A repeatedly.

    v=[001],Av=[01−2],A2v=[1−34],A3v=[−37−8]v=\begin{bmatrix}0\\0\\1\end{bmatrix},\quad Av=\begin{bmatrix}0\\1\\-2\end{bmatrix},\quad A^2v=\begin{bmatrix}1\\-3\\4\end{bmatrix},\quad A^3v=\begin{bmatrix}-3\\7\\-8\end{bmatrix}
  5. Combine the columns with the coefficients of φ_o.

    L=A3v+18A2v+108Av+216v=[−3+187−54+108−8+72−216+216]=[156164]L=A^3v+18A^2v+108Av+216v=\begin{bmatrix}-3+18\\7-54+108\\-8+72-216+216\end{bmatrix}=\begin{bmatrix}15\\61\\64\end{bmatrix}
  6. Check by coefficient matching. Only the first column of A − LC changes, and the three coefficients give l₁, l₂ and l₃ in turn.

    det⁡(sI−A+LC)=s3+(3+l1)s2+(2+3l1+l2)s+(2l1+2l2+l3) ⟹ l1=15, l2=61, l3=64\det(sI-A+LC)=s^3+(3+l_1)s^2+(2+3l_1+l_2)s+(2l_1+2l_2+l_3)\ \Longrightarrow\ l_1=15,\ l_2=61,\ l_3=64
  7. Minimum-order observer. The angle is measured, so x_a = x₁ and x_b = [x₂, x₃]. Read off the blocks.

    Aaa=0,Aab=[10],Aba=[00],Abb=[−110−2],Ba=0,Bb=[01]A_{aa}=0,\quad A_{ab}=\begin{bmatrix}1&0\end{bmatrix},\quad A_{ba}=\begin{bmatrix}0\\0\end{bmatrix},\quad A_{bb}=\begin{bmatrix}-1&1\\0&-2\end{bmatrix},\quad B_a=0,\quad B_b=\begin{bmatrix}0\\1\end{bmatrix}
  8. Place both poles of A_bb − L_rA_ab at −6. Dual Ackermann on (A_bb, A_ab) gives the same L_r.

    det⁡(sI−Abb+LrAab)=s2+(3+ℓ1)s+(2+2ℓ1+ℓ2)=s2+12s+36 ⟹ Lr=[916]\det(sI-A_{bb}+L_rA_{ab})=s^2+(3+\ell_1)s+(2+2\ell_1+\ell_2)=s^2+12s+36\ \Longrightarrow\ L_r=\begin{bmatrix}9\\16\end{bmatrix}
  9. Assemble the implementable observer. With A_aa = 0, A_ba = 0 and B_a = 0, only the term (A_bb − L_rA_ab)L_r remains in the y coefficient.

    η˙=[−101−16−2]η+[−74−176]y+[01]u,[x^2x^3]=η+[916]y\dot\eta=\begin{bmatrix}-10&1\\-16&-2\end{bmatrix}\eta+\begin{bmatrix}-74\\-176\end{bmatrix}y+\begin{bmatrix}0\\1\end{bmatrix}u,\qquad \begin{bmatrix}\hat x_2\\\hat x_3\end{bmatrix}=\eta+\begin{bmatrix}9\\16\end{bmatrix}y
  10. The error matrix F = A_bb − L_rA_ab has the double eigenvalue −6, and F + 6I squares to zero, so e^{Ft} = e^{−6t}(I + (F + 6I)t). Starting with a speed error of 1 and no torque error, the speed-estimate error is e^{−6t}(1 − 4t), which is −0.050 at 0.5 s and −0.0074 at 1 s, while the torque estimate errs by up to 16/(6e) ≈ 0.98 at t = 1/6 s before it settles.

    eFt=e−6t[1−4tt−16t1+4t],eb(t)=e−6t[1−4t−16t] for eb(0)=[10]e^{Ft}=e^{-6t}\begin{bmatrix}1-4t&t\\-16t&1+4t\end{bmatrix},\qquad e_b(t)=e^{-6t}\begin{bmatrix}1-4t\\-16t\end{bmatrix}\ \text{for}\ e_b(0)=\begin{bmatrix}1\\0\end{bmatrix}
  11. Separation. With u = −K[y; x̂₂; x̂₃] the five-state loop has the characteristic polynomial of A − BK times that of the observer, and its eigenvalues computed from the 5×5 matrix are −2 ± j2, −10, −6 and −6. With the full-order observer instead, the six eigenvalues are −2 ± j2, −10 and −6 three times.

    (s3+14s2+48s+80)(s+6)2=s5+26s4+252s3+1160s2+2688s+2880(s^3+14s^2+48s+80)(s+6)^2=s^5+26s^4+252s^3+1160s^2+2688s+2880
REFERENCE

Reference table

State feedback and observer design side by side
ItemState feedbackObserver
Matrix whose eigenvalues are placedA−BKA-BKA−LCA-LC
Gainrow K, one entry per statecolumn L, one entry per state
Test matrixMc=[BAB⋯An−1B]M_c=\begin{bmatrix}B&AB&\cdots&A^{n-1}B\end{bmatrix}Mo=[CCA⋮CAn−1]M_o=\begin{bmatrix}C\\CA\\\vdots\\CA^{n-1}\end{bmatrix}
Condition for placing every polecontrollable: det M_c ≠ 0observable: det M_o ≠ 0
Ackermann's formulaK=[0⋯01]Mc−1φ(A)K=\begin{bmatrix}0&\cdots&0&1\end{bmatrix}M_c^{-1}\varphi(A)L=φo(A) Mo−1[0⋯01]TL=\varphi_o(A)\,M_o^{-1}\begin{bmatrix}0&\cdots&0&1\end{bmatrix}^{\mathsf T}
Hand shortcutqMc=[0⋯01], K=∑iαi qAiqM_c=\begin{bmatrix}0&\cdots&0&1\end{bmatrix},\ K=\sum_i\alpha_i\,qA^iMov=[0⋯01]T, L=∑iαiAivM_ov=\begin{bmatrix}0&\cdots&0&1\end{bmatrix}^{\mathsf T},\ L=\sum_i\alpha_iA^iv
Positioning drive exampleK = [80, 35, 11] places −2 ± j2 and −10L = [15, 61, 64]ᵀ places all three poles at −6

A matrix and its transpose have the same eigenvalues, so the observer problem for (A, C) is the state-feedback problem for (Aᵀ, Cᵀ) with gain Lᵀ. In the sums αₙ = 1.

More detail and model assumptions
A CLOSER LOOK

Duality turns observer design into state feedback

Because a matrix and its transpose have the same characteristic polynomial, det(sI − A + LC) = det(sI − Aᵀ + CᵀLᵀ). Compare this with det(sI − A + BK): A becomes Aᵀ, B becomes Cᵀ and K becomes Lᵀ. The controllability matrix of (Aᵀ, Cᵀ) is M_oᵀ, so the observer poles can all be placed exactly when (A, C) is observable.

Applying Ackermann's formula gives Lᵀ = [0 … 0 1](M_oᵀ)⁻¹φ_o(Aᵀ), and transposing gives L = φ_o(A)M_o⁻¹[0 … 0 1]ᵀ. The hand shortcut transposes as well. Solve M_o·v = [0 … 0 1]ᵀ for one column v, then form L = Aⁿv + αₙ₋₁Aⁿ⁻¹v + … + α₀v using only matrix-times-column products.

A CLOSER LOOK

Choosing the observer poles

The estimate should converge before the controller relies on it, so observer poles are usually placed two to five times farther left than the controller's dominant poles. Faster is not free. In x̂̇ = Ax̂ + Bu + L(y − Cx̂) every bit of sensor noise is multiplied by L, and L grows quickly with the observer speed.

For the lab plant measured in position, with damping 0.8 and natural frequency ωo, the observer gain is L = [2.4, −1.92]ᵀ for ωo = 2 rad/s, [5.6, 7.52]ᵀ for ωo = 4 rad/s and [8.8, 24.96]ᵀ for ωo = 6 rad/s. The Kalman filter chapter chooses L from stated noise statistics instead of from pole positions.

A CLOSER LOOK

The minimum-order observer

Suppose y = x_a holds m states exactly, and order the state so that these come first. Partition the model: ẋ_a = A_aa·x_a + A_ab·x_b + B_a·u and ẋ_b = A_ba·x_a + A_bb·x_b + B_b·u. The second equation is a model of x_b driven by the known signals y and u. The first, rearranged as ẋ_a − A_aa·y − B_a·u = A_ab·x_b, is a measurement of x_b with output matrix A_ab.

An observer for this smaller problem is x̂̇_b = A_bb·x̂_b + A_ba·y + B_b·u + L_r(ẋ_a − A_aa·y − B_a·u − A_ab·x̂_b). Its error e_b = x_b − x̂_b obeys ė_b = (A_bb − L_rA_ab)e_b, so L_r is designed exactly like a full-order gain with A_bb in place of A and A_ab in place of C. When the measured states are the first m states, (A_bb, A_ab) is observable whenever (A, C) is.

The observer above contains ẋ_a, which would mean differentiating a measured signal. Define η = x̂_b − L_r·y instead. Differentiating η and substituting removes ẋ_a and leaves an observer with n − m states that uses only y and u; the estimate is x̂_b = η + L_r·y.

A CLOSER LOOK

Full-order or minimum-order?

The minimum-order observer is smaller, and with a perfect sensor its estimate of the measured states is exact from the first instant. With a noisy sensor that becomes a weakness, because x̂_a = y passes the noise straight through and the direct path L_r·y adds it to x̂_b as well. A full-order observer filters every state. A clean sensor such as an encoder favours the minimum-order observer; a noisy one favours a full-order observer or a Kalman filter.

A CLOSER LOOK

The separation principle and its conditions

With u = −Kx̂ = −Kx + Ke, the plant becomes ẋ = (A − BK)x + BKe, while the error still obeys ė = (A − LC)e. In the coordinates (x, e) the closed loop is block upper-triangular, so its 2n eigenvalues are those of A − BK together with those of A − LC. With a minimum-order observer the second block is A_bb − L_rA_ab, and the loop has 2n − m eigenvalues.

The argument needs four conditions: a linear time-invariant model; an observer that uses exactly the plant's matrices; an observer that receives the input actually applied, so no unmodelled saturation; and (A, B) controllable and (A, C) observable if the poles are to be placed freely. If any of these fails, the controller and observer eigenvalues mix, and the whole loop has to be checked.

The principle also leaves some things unsaid. The combined loop can have much smaller stability margins than full-state feedback had, and the transient of e still enters the plant through BKe. Separation lets you design in two steps; it does not let you skip checking the result.

A CLOSER LOOK

The observer-based controller is a dynamic compensator

From the measurement y to the input u, observer and gain together form a single transfer function: U(s) = −K(sI − A + BK + LC)⁻¹L·Y(s) when r = 0. It sits in the feedback path like a classical compensator. For a double integrator it has one zero much closer to the origin than its two poles, which is the shape of a lead network; exercise 4 computes it.

TRY IT

Connect the example to the lab

The state-feedback lab measures position only, so C = [1, 0] and M_o is the identity matrix. The dual Ackermann formula then says that L is the last column of φ_o(A). With ωn = 2 rad/s and an observer speed ratio of 3, the lab places the observer poles at ωo = 6 rad/s with damping 0.8, so φ_o(s) = s² + 9.6s + 36 and L = [8.8, 24.96]ᵀ. Compare this with the L printed in the lab's model panel.

Check the observer gain L in the lab
YOUR EXPERIMENT

Make a prediction before moving a control.

  1. PREDICTBefore opening the lab, compute the observer gain for ωn = 2 rad/s and observer speed ratios 1 and 3. The lab's observer polynomial is s² + 1.6ωo·s + ωo² with ωo = ratio × ωn. Which observer removes a 0.3 m initial estimate error sooner, and at what cost in force?
  2. CHANGE ONE THINGOpen the observer-recovery experiment, which uses observer feedback with ratio 1, and pin the run. Change only the observer speed ratio to 3.
  3. OBSERVEThe model panel shows L = [2.4, −1.92]ᵀ for ratio 1 and L = [8.8, 24.96]ᵀ for ratio 3. The position-estimate error stays above 0.01 m until about 2.2 s with ratio 1 but only until about 0.8 s with ratio 3, and the largest requested force in the first second rises from 0.8 N to about 1.6 N.
  4. EXPLAINWith C = [1, 0], M_o = I and L is the last column of φ_o(A). Tripling ωo multiplies the constant coefficient of φ_o by nine, which is why the second entry of L changes so much. The faster observer corrects the wrong estimate sooner, but while it does, the estimated velocity swings further and the controller pushes harder on that estimate.
BE CAREFUL

Observer poles placed too far left make L large, and L multiplies the sensor noise before it reaches the estimate. A minimum-order observer has its own weakness: it uses the raw measurement as the estimate of x_a and adds L_r·y to x̂_b, so it does not filter noise on the measured states at all. Finally, the separation principle needs the observer to receive the input actually applied to the plant. Feed it the clipped actuator signal, not the requested one.

Check your understanding+

A double integrator measures velocity only: A = [[0, 1], [0, 0]], C = [0, 1]. Can any observer recover the position?

No. M_o = [[0, 1], [0, 0]] has rank 1, and for every L the matrix A − LC keeps an eigenvalue at 0, so a position error never decays. A constant position offset changes nothing in the velocity, so the sensor cannot see it.
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. Design a full-order observer for the model below that places both observer poles at −10. Use the dual Ackermann formula and check by coefficient matching.

    A=[01−6−5],C=[10]A=\begin{bmatrix}0&1\\-6&-5\end{bmatrix},\qquad C=\begin{bmatrix}1&0\end{bmatrix}
    Show answer

    AnswerL = [15, 19]ᵀ.

    L=[1519]L=\begin{bmatrix}15\\19\end{bmatrix}
    Show full solution

    Solution

    1. M_o = [C; CA] is the identity, so v = [0, 1]ᵀ.

      Mo=[1001],v=[01],Av=[1−5],A2v=[−519]M_o=\begin{bmatrix}1&0\\0&1\end{bmatrix},\quad v=\begin{bmatrix}0\\1\end{bmatrix},\quad Av=\begin{bmatrix}1\\-5\end{bmatrix},\quad A^2v=\begin{bmatrix}-5\\19\end{bmatrix}
    2. The desired polynomial is (s + 10)² = s² + 20s + 100.

      L=A2v+20Av+100v=[−5+2019−100+100]=[1519]L=A^2v+20Av+100v=\begin{bmatrix}-5+20\\19-100+100\end{bmatrix}=\begin{bmatrix}15\\19\end{bmatrix}
    3. Check: A − LC changes only in its first column, and its characteristic polynomial matches.

      det⁡(sI−A+LC)=s2+(5+l1)s+(6+5l1+l2)=s2+20s+100\det(sI-A+LC)=s^2+(5+l_1)s+(6+5l_1+l_2)=s^2+20s+100
  2. The lab's mass–spring–damper is fitted with a velocity sensor instead of a position sensor. (a) Is it observable? (b) Design L for observer poles at −4 and −5. (c) Remove the spring, leaving a double integrator. Is the position still observable from velocity?

    A=[01−4−0.8],C=[01]A=\begin{bmatrix}0&1\\-4&-0.8\end{bmatrix},\qquad C=\begin{bmatrix}0&1\end{bmatrix}
    Show answer

    Answer(a) Yes, det M_o = 4. (b) L = [−4, 8.2]ᵀ. (c) No: without the spring M_o has rank 1 and a pole stays at 0 for every L.

    L=[−48.2]L=\begin{bmatrix}-4\\8.2\end{bmatrix}
    Show full solution

    Solution

    1. (a) The spring turns position into force, and force shows up in the velocity, so the second row of M_o depends on position.

      Mo=[CCA]=[01−4−0.8],det⁡Mo=4M_o=\begin{bmatrix}C\\CA\end{bmatrix}=\begin{bmatrix}0&1\\-4&-0.8\end{bmatrix},\qquad \det M_o=4
    2. (b) Solve M_o·v = [0, 1]ᵀ: the first row gives v₂ = 0, the second gives −4v₁ = 1.

      v=[−0.250],Av=[01],A2v=[1−0.8]v=\begin{bmatrix}-0.25\\0\end{bmatrix},\quad Av=\begin{bmatrix}0\\1\end{bmatrix},\quad A^2v=\begin{bmatrix}1\\-0.8\end{bmatrix}
    3. With φ_o(s) = (s + 4)(s + 5) = s² + 9s + 20, combine. The negative first entry is not an error: it is what this sensor geometry requires.

      L=A2v+9Av+20v=[1−5−0.8+9]=[−48.2],A−LC=[05−4−9]L=A^2v+9Av+20v=\begin{bmatrix}1-5\\-0.8+9\end{bmatrix}=\begin{bmatrix}-4\\8.2\end{bmatrix},\qquad A-LC=\begin{bmatrix}0&5\\-4&-9\end{bmatrix}
    4. Check: the trace of A − LC is −9 and its determinant 20, matching poles at −4 and −5.

    5. (c) With A = [[0, 1], [0, 0]], M_o = [[0, 1], [0, 0]] has rank 1, and A − LC = [[0, 1 − l₁], [0, −l₂]] has characteristic polynomial s² + l₂s, so the eigenvalue 0 stays for every L.

      det⁡ ⁣(sI−[01−l10−l2])=s(s+l2)\det\!\left(sI-\begin{bmatrix}0&1-l_1\\0&-l_2\end{bmatrix}\right)=s(s+l_2)
  3. Design a minimum-order observer for the velocity of the lab's mass–spring–damper (position measured) with its pole at −10. Write the observer equations. If the mass rests at 0.1 m and the observer starts at η = 0, what is the velocity-estimate error at t = 0 and at t = 0.5 s?

    A=[01−4−0.8],B=[01],y=x1A=\begin{bmatrix}0&1\\-4&-0.8\end{bmatrix},\qquad B=\begin{bmatrix}0\\1\end{bmatrix},\qquad y=x_1
    Show answer

    AnswerL_r = 9.2, η̇ = −10η − 96y + u and v̂ = η + 9.2y. The error is −0.92 m/s at t = 0 and −0.0062 m/s at 0.5 s.

    η˙=−10η−96y+u,v^=η+9.2y\dot\eta=-10\eta-96y+u,\qquad \hat v=\eta+9.2y
    Show full solution

    Solution

    1. Partition with x_a = position and x_b = velocity.

      Aaa=0,Aab=1,Aba=−4,Abb=−0.8,Ba=0,Bb=1A_{aa}=0,\quad A_{ab}=1,\quad A_{ba}=-4,\quad A_{bb}=-0.8,\quad B_a=0,\quad B_b=1
    2. The scalar error obeys ė_b = (−0.8 − L_r)e_b. A pole at −10 needs L_r = 9.2.

      Abb−LrAab=−0.8−9.2=−10A_{bb}-L_rA_{ab}=-0.8-9.2=-10
    3. Fill in the observer formula. The y coefficient is (−10)(9.2) + (−4) − 9.2·0 = −96, and the u coefficient is 1 − 9.2·0 = 1.

      η˙=−10η−96y+u,v^=η+9.2y\dot\eta=-10\eta-96y+u,\qquad \hat v=\eta+9.2y
    4. At t = 0 the estimate is v̂ = 0 + 9.2 × 0.1 = 0.92 m/s while the mass is at rest, so e_b(0) = −0.92 m/s. The error then decays as e^{−10t}; a simulation with a sinusoidal input gives the same value, because the input cancels from the error.

      eb(0.5)=−0.92 e−5≈−0.0062 m/se_b(0.5)=-0.92\,e^{-5}\approx-0.0062\ \mathrm{m/s}
  4. A double integrator with position measurement uses K = [4, 2.8] (poles −1.4 ± j1.43) and an observer with L = [8, 16]ᵀ (double pole at −4). Give the closed-loop eigenvalues and the compensator transfer function from y to −u.

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

    AnswerThe eigenvalues are −1.4 ± j1.43 and −4 twice. The compensator is (76.8s + 64)/(s² + 10.8s + 42.4), with its zero at −0.833 and its poles at −5.4 ± j3.64.

    Gc(s)=76.8s+64s2+10.8s+42.4G_c(s)=\frac{76.8s+64}{s^2+10.8s+42.4}
    Show full solution

    Solution

    1. By separation, the characteristic polynomial is the product of the controller and observer polynomials.

      (s2+2.8s+4)(s2+8s+16)=s4+10.8s3+42.4s2+76.8s+64(s^2+2.8s+4)(s^2+8s+16)=s^4+10.8s^3+42.4s^2+76.8s+64
    2. The compensator matrix and its 2×2 inverse.

      A−BK−LC=[−81−20−2.8],(sI−A+BK+LC)−1=1s2+10.8s+42.4[s+2.81−20s+8]A-BK-LC=\begin{bmatrix}-8&1\\-20&-2.8\end{bmatrix},\qquad (sI-A+BK+LC)^{-1}=\frac{1}{s^2+10.8s+42.4}\begin{bmatrix}s+2.8&1\\-20&s+8\end{bmatrix}
    3. Multiply by L on the right and by K on the left.

      K[8s+38.416s−32]=4(8s+38.4)+2.8(16s−32)=76.8s+64K\begin{bmatrix}8s+38.4\\16s-32\end{bmatrix}=4(8s+38.4)+2.8(16s-32)=76.8s+64
    4. Check with the loop: the plant is 1/s², and s²(s² + 10.8s + 42.4) + 76.8s + 64 reproduces the fourth-order polynomial above. The zero at −64/76.8 = −0.833 lies far to the right of the poles, the shape of a lead compensator.

  5. Exam level. The positioning drive of the worked example gets a speed encoder as well, so y = [x₁, x₂] and only the torque x₃ is unknown. Design a minimum-order observer with its pole at −10, explain which gain is free, and give the closed-loop eigenvalues with K = [80, 35, 11].

    Show answer

    AnswerL_r = [ℓ₁, 8] with ℓ₁ free; with ℓ₁ = 0, η̇ = −10η − 72x₂ + u and x̂₃ = η + 8x₂. The eigenvalues are −2 ± j2 and −10 twice.

    η˙=−10η−72x2+u,x^3=η+8x2\dot\eta=-10\eta-72x_2+u,\qquad \hat x_3=\eta+8x_2
    Show full solution

    Solution

    1. Partition with x_a = [x₁, x₂] and x_b = x₃.

      Aaa=[010−1],Aab=[01],Aba=[00],Abb=−2,Ba=[00],Bb=1A_{aa}=\begin{bmatrix}0&1\\0&-1\end{bmatrix},\quad A_{ab}=\begin{bmatrix}0\\1\end{bmatrix},\quad A_{ba}=\begin{bmatrix}0&0\end{bmatrix},\quad A_{bb}=-2,\quad B_a=\begin{bmatrix}0\\0\end{bmatrix},\quad B_b=1
    2. L_r is a row [ℓ₁, ℓ₂]. Because the first entry of A_ab is 0, ℓ₁ does not affect the error pole; that freedom is what two outputs bring. Take ℓ₁ = 0.

      Abb−LrAab=−2−ℓ2=−10 ⟹ ℓ2=8A_{bb}-L_rA_{ab}=-2-\ell_2=-10\ \Longrightarrow\ \ell_2=8
    3. The y coefficient is (−10)[0, 8] − [0, 8]A_aa = [0, −80] − [0, −8] = [0, −72].

      η˙=−10η+[0−72]y+u,x^3=η+8x2\dot\eta=-10\eta+\begin{bmatrix}0&-72\end{bmatrix}y+u,\qquad \hat x_3=\eta+8x_2
    4. Physically, ẋ₂ = −x₂ + x₃ means x₃ = ẋ₂ + x₂, and the observer computes a filtered version of that derivative of the speed.

    5. By separation the four closed-loop eigenvalues are those of A − BK and the observer pole. Computing them from the 4×4 matrix of [x; η] confirms the product polynomial.

      (s3+14s2+48s+80)(s+10)=s4+24s3+188s2+560s+800(s^3+14s^2+48s+80)(s+10)=s^4+24s^3+188s^2+560s+800

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 does duality mean for observer design?

A matrix has the same eigenvalues as its transpose, so det(sI − A + LC) = det(sI − Aᵀ + CᵀLᵀ). Placing the observer poles of (A, C) is therefore state-feedback design for (Aᵀ, Cᵀ) with gain Lᵀ, and every pole can be placed exactly when (A, C) is observable.

How fast should the observer poles be?

Usually two to five times farther left than the controller's dominant poles, so that the estimate converges before the controller relies on it. Faster poles make L larger, and L multiplies sensor noise: for the lab plant measured in position, with damping 0.8, L grows from [2.4, −1.92] at 2 rad/s to [8.8, 24.96] at 6 rad/s.

When is a minimum-order observer better than a full-order one?

It estimates only the n − m states the sensor does not measure, so it is smaller and exact for the measured states when the sensor is clean, such as an encoder. With a noisy sensor it passes the noise straight into x̂_a and, through L_r·y, into x̂_b, so a full-order observer or a Kalman filter is the better choice.

Why does the minimum-order observer use the variable η?

Written directly, the observer contains ẋ_a, the derivative of a measured signal. Defining η = x̂_b − L_r·y removes that derivative and leaves an observer of order n − m driven only by y and u, with the estimate x̂_b = η + L_r·y.

Does the separation principle make an observer-based controller as robust as full-state feedback?

No. It says that the closed-loop eigenvalues are those of A − BK together with those of A − LC, for an exact linear model whose observer receives the input actually applied. The combined loop can still have much smaller stability margins, and saturation or model error mixes the two sets, so check the whole loop.

Why the next lesson follows

Pole placement, for the controller and for the observer, asks you to choose every pole yourself. LQR instead computes the gain from a stated balance between state error and control effort.

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/observer-design · Updated 2026-11-15