Module 2-1: Pole Placement

Modern Control - Theory and Design (2023)

Ding Zhao

Associate Professor
College of Engineering
School of Computer Science

Carnegie Mellon University

Ding Zhao (CMU) M2-1: Pole Placement 1 / 46

Table of Contents

1 PID Feedback Control

2 Design (State) Feedback Control

3 Design (Luenberger) observer

4 Separation Principle

5 MIMO Systems

6 (Optional) Reduced Order Observer

Ding Zhao (CMU) M2-1: Pole Placement 2 / 46

Table of Contents

1 PID Feedback Control

2 Design (State) Feedback Control

3 Design (Luenberger) observer

4 Separation Principle

5 MIMO Systems

6 (Optional) Reduced Order Observer

Ding Zhao (CMU) M2-1: Pole Placement 3 / 46

Proportional–Integral–Derivative (PID) Control
Z t
de(t)
Algorithms: u(t) = Kp e(t) + Ki e(τ ) dτ + Kd
0 dt
Transfer function: L(s) = Kp + Ki /s + Kd s
Minorsky, 1885-1970

image Courtesy @ Arturo Urquizo

Ding Zhao (CMU) M2-1: Pole Placement 4 / 46
Intuition of PID

Take the example of lateral control of a car. Define Cross Track Error (CTE) as the distance
of the car from trajectory.

Ding Zhao (CMU) M2-1: Pole Placement 5 / 46

Recap: Linear CONTROL Systems - A Brief History
Control:continuously operating dynamical systems
Frequency response methods
made it possible to
design linear closed-loop Norbert Wiener (Cybernetics)
State space methods
Minorsky worked on linear model-based (MIMO)
automatic controllers (PID) optimal/stochastic/ modern control theory in
for steering ships adaptive control Rudolf Kalman (Apollo) engr,bio,med,econ,social etc.

1920s 1930s 1940s 1950s 1960s 1980s Present

Root-locus method
due to Evans
was fully developed
Nyquist/Bode (Bell Lab) developed Machine learning controls
methods for analyzing nolinear model-free
the stability of genetic algorithm, neural network
controlled systems reinforcement learning control

Ding Zhao (CMU) M2-1: Pole Placement 6 / 46

A Brief History
In 1922, Minorsky helped in the installation and testing of automatic steering on board
the battleship USS New Mexico. In relation to this work Minorsky authored a paper
introducing the concept of Integral and Derivative Control. This paper is one of the
earliest formal discussions on control theory in the English language.
1924–1934 Nicolas Minorsky was a Professor of Electronics and Applied Physics at Penn.
Navy requests to investigate anti-rolling devices on ships. The ability to stabilize a ship
such as an aircraft carrier would be extremely useful during the landing of airplanes.
1934 to 1940, Minorsky worked on roll stabilization of ships for the navy, designing an
activated-tank stabilization system into a 5-ton model ship, later on dubbed as ”USS
Minorsky”.
A full-scale version of the system was tested in the USS Hamilton but exhibited control
stability problems. Very promising results were beginning to appear when the outbreak of
the Second World War interrupted further development as the Hamilton was called to
active duty and the 5 ton model was put into storage.
In 1942, Ziegler and Nichols introduced tuning rules.
Ding Zhao (CMU) M2-1: Pole Placement 7 / 46
Tuning method: Ziegler–Nichols Method
The ultimate gain Ku : set KI and KD to 0, increase KP until the output of the control loop
has stable and consistent oscillations to get Ku and the oscillation period Tu . Ki = Kp /Ti .
Kd = Kp Td . Z-N yields an aggressive gain and overshoot – some applications wish to instead
minimize or eliminate overshoot, and for these this method is inappropriate. In this case, use
the last two rows.

Ding Zhao (CMU) M2-1: Pole Placement 8 / 46

Tuning method: Twiddle Algorithm

# Choose an initialization parameter vector

else: # There was no improvement
p = [0, 0, 0]
p[i] -= 2 * dp[i]
# Define potential changes
# Go into the other direction
dp = [1, 1, 1]
err = robot(p)
# Calculate the error
best_err = robot(p)
if err < best_err:
threshold = 0.001
# There was an improvement
while sum(dp) > threshold:
best_err = err
for i in range(len(p)):
dp[i] *= 1.05
p[i] += dp[i]
else: # There was no improvement
err = robot(p)
p[i] += dp[i]
# As there was no improvement,
if err < best_err:
# the step size in either direction,
# There was some improvement
# the step size might simply be too big.
best_err = err
dp[i] *= 0.95
dp[i] *= 1.1

Courtesy at Martin Thoma

Ding Zhao (CMU) M2-1: Pole Placement 9 / 46
Table of Contents

1 PID Feedback Control

2 Design (State) Feedback Control

3 Design (Luenberger) observer

4 Separation Principle

5 MIMO Systems

6 (Optional) Reduced Order Observer

Ding Zhao (CMU) M2-1: Pole Placement 10 / 46

Feedback Control

Benefits of feedback
1 Stabilize unstable systems
2 Improve transient response (speed)
3 Reject disturbances
4 Increase robustness to modeling errors
We will consider “state feedback” ⇒ u = Kx
For systems that do not output all the states, we will estimate unmeasured states with an
observer
Without loss of generalization, we will assume the desired state is at the origin.

Ding Zhao (CMU) M2-1: Pole Placement 11 / 46

SISO State Feedback

Consider the LTI SISO system

ẋ = Ax + bu
y = cx + du

and u = Kx + Ev (Feed forward. We will see how to design it later.)

⇒ ẋ = (A + bK)x + bEv, y = (c + dK)x + dEv
where, A + bK = Af b

We know that the stability of the system depends on the eigenvalues of Af b

Goal: Given a set of desired eigenvalues {λd }, design K s.t. A + bK has eigenvalues λd .

Ding Zhao (CMU) M2-1: Pole Placement 12 / 46

Recap: Controllable Canonical Forms for the SISO System

bn sn + · · · + b0
Given a SISO system: G(s) =
sn + an−1 sn−1 + · · · + a0
   
0 1 0 ··· 0 0
 0
 0 1 ··· 0  
0
 
ẋ =  ... .. .. . .  .. 
x + . u
 
 . . . 
  
 0 0 0 ··· 1  0
−a0 −a1 · · · · · · −an−1 1
 
y = b0 − bn a0 b1 − bn a1 · · · bn−1 − bn an−1 x + bn u

We can also arrive at this form via a similarity

Mc = P Pc−1
transformation using Mc .

Ding Zhao (CMU) M2-1: Pole Placement 13 / 46

Recap: Similarity Transformations for Ctrl/Obs. Canonical Forms

We can get here via a similarity transformation and an observable state space model.
Let Mc = P Pc−1 , with Pc the controllability matrix for the system in controllable
canonical form, calculated by Ac and Bc
Let Mo = Q−1 Qo , with Qo the observability matrix for the system in observable
canonical form, calculated by Ao and Co
Q−1
o and
−1
 Pc have special forms: 
a1 a2 · · · an−1 1
 
1 0 ··· 0 0
 a2 a3 · · · 1 0
  an−1 1 ··· 0 0
 .. .. . . ..  −1  . .. .. ..

.. 
Pc−1 =  . . . . 0  ..
, Qo =  . . . .
.. 
 
 . .

a a ··· 1 0

an−1 1 · · · . .  2 3
1 0 ··· ··· 0 a 1 a 2 · · · an−1 1
These can be written by inspection!

Ding Zhao (CMU) M2-1: Pole Placement 14 / 46

Feedback with Controllable Canonical Form

The easiest way to design SISO feedback controllers is to start with controllable canonical
form (of course, given the system is controllable).
Let
   
0 1 0 ··· 0 0
0
 0 1 ··· 0 

0
 
..
x̃˙ = Ãx̃ + B̃u =  .. .. .. .. .
 x̃ +  ..  u
 
.
 . . . .   
 0 0 0 ··· 1  0
−a0 −a1 −a2 · · · −an−1 1
 
Let u = K̃ x̃ = k̃0 k̃1 · · · k̃n−1 x̃
 T  
⇒ B̃ K̃ = 0 0 · · · 0 1 k̃0 k̃1 · · · k̃n−1

Ding Zhao (CMU) M2-1: Pole Placement 15 / 46

Pole Placement with Controllable Canonical Form

 
0 1 0 ··· 0
 0 0 1 ···  0
.. .. .. ..
 
⇒Ã + B̃ K̃ = 
 .. 
. . . . .
 
 0 0 0 ··· 1 
−a0 + k̃0 −a1 + k̃1 −a2 + k̃2 · · · −an−1 + k̃n−1

From Controllable Canonical Form, the characteristic equation for à + B̃ K̃ is

∆(s) = sn + (an−1 − k̃n−1 )sn−1 + · · · + (a1 − k̃1 )s + (a0 − k̃0 )

⇒ By choosing K̃, we can get any ∆(s) we want ⇒ any eigenvalues!

Ding Zhao (CMU) M2-1: Pole Placement 16 / 46

Controllable Canonical Form

We still need to convert K̃ to K to work with the original system form (similarity
transformation)
Let the original system be x and the Controllable Canonical Form system be x̃. Then
x = M x̃ with M = P Pc−1 . Given K̃ designed for Ã, we can calculate K.

u = Kx = K̃ x̃ = KM x̃
K = K̃M −1

Ding Zhao (CMU) M2-1: Pole Placement 17 / 46

Example
Place poles of
   
1 3 1
ẋ = x+ u, at -5 and -6
4 2 1
Solution:    
0 1 0
∆(s) = (s − 1)(s − 2) − 12 = s2 − 3s − 10 x̃˙ = x̃ + u
10 3 1
         
1 4 0 1 −1 −3 1 −1 1 1 −1 1 −1 1
P = , Pc = , Pc = ⇒ M = P Pc = ,M =
1 6 1 3 1 0 3 1 2 3 −1
(λ + 5)(λ + 6) = λ2 + 11λ + 30 = λ2 + (−3 − k1 )λ + (−10 − k0 )
 
⇒ k̃1 = −14, k̃0 = −40 ⇒ K̃ = −40 −14
⇒ K = −40 −14 M −1 ⇒ K = −1 −13
   

Ding Zhao (CMU) M2-1: Pole Placement 18 / 46

Example

It’s not hard to verify that the closed loop system has the desired eigenvalues.
Clearly if the system is controllable, we can place the poles... what if it’s not?

Ding Zhao (CMU) M2-1: Pole Placement 19 / 46

Example

   
−2 0 1
ẋ = x+ u
0 −2 3
Controllable? No - by observing the Jordan blocks.
Solution:  
  −2 + k0 3k1
Let K = k0 k1 and u = Kx ⇒ A + bK =
k0 −2 + 3k1
 
λ + 2 − k0 −3k1
⇒ |λI − (A + bK)| = | |
−k0 λ + 2 − 3k1
= (λ + 2 − k0 )(λ + 2 − 3k1 ) − 3k0 k1 = (λ + 2)2 − (k0 + 3k1 )(λ + 2)
= (λ + 2)(λ + (2 − k0 − 3k1 ))
In general, one or more eigenvalues of an uncontrollable system will be unaffected by
state feedback.
Ding Zhao (CMU) M2-1: Pole Placement 20 / 46
Table of Contents

1 PID Feedback Control

2 Design (State) Feedback Control

3 Design (Luenberger) observer

4 Separation Principle

5 MIMO Systems

6 (Optional) Reduced Order Observer

Ding Zhao (CMU) M2-1: Pole Placement 21 / 46

SISO State Observers

We have used the full state vector in feedback, but really only measure y.
An observer takes as inputs u and y returns an estimator of x.
Why not just use x̃˙ = Ax̃ + Bu, ỹ = C x̃ + Du, x̃(t0 ) = x(t0 )?

Ding Zhao (CMU) M2-1: Pole Placement 22 / 46

SISO State Observers

We have used the full state vector in feedback, but really only measure y.
An observer takes as inputs u and y returns an estimator of x.
Why not just use x̃˙ = Ax̃ + Bu, ỹ = C x̃ + Du, x̃(t0 ) = x(t0 )?
This only relies on model-based prediction, it can be used in a short period of time, but we
also need to correction to deal with modeling errors/disturbance.
Only replying on prediction is akin to running with eyes closed.

Ding Zhao (CMU) M2-1: Pole Placement 22 / 46

Dynamics of SISO State Observers

Let’s try the following to get y involved

x̃˙ = Ax̃ + Bu + L(y − ỹ)

= Ax̃ + Bu + L(y − (C x̃ + Du))
= (A − LC)x̃ + (B − LD)u + Ly

Define the error e = x − x̃

⇒ ė = Ax + Bu − (A − LC)x̃ − (B − LD)u − L(Cx + Du)

= Ae − LCe = (A − LC)e
⇒ ė = (A − LC)e

Based on stability theory, e → 0 as t → ∞ ⇔ A − LC has poles in the LHP

Ding Zhao (CMU) M2-1: Pole Placement 23 / 46

Recap: Observable Cononical Form for the SISO System

 
−an−1 1 0 ··· 0  
 ..  bn−1 − bn an−1
 . 0 1 · · · 0
  .. 

ẋ =  . .. .. .. x +  . u

 .. . . .  
   b1 − bn a1 
 −a1 0 0 · · · 1 b0 − bn a0
−a0 0 ··· ··· 0
 
y= 1 0 · · · 0 x + bn u
Similarly: we could arrive at the observable canonical via similarity transformation:
Mo = Q−1 Qo

Ding Zhao (CMU) M2-1: Pole Placement 24 / 46

Pole Placemnet for SISO Observers in Observable Canonical Form
We can arbitrarily place the poles ⇔ (A,C) is observable.
 
−ãn−1 1 0 ··· 0
˜ln−1
 
−ãn−2 0 1 · · · 0
 
à − L̃C̃ =  ... .. .. ..   ..   
−  .  1 0 ··· 0

. . . 0
˜l0
 
 −ã1 0 0 · · · 1
−ã0 0 0 ··· 0
−ãn−1 − ˜ln−1
 
1 0 ··· 0
−ãn−2 − ˜ln−2 0 1 · · · 0
 
=
 .. .. .. . . 
. . . . 0
 −ã1 − ˜l1
 
0 0 · · · 1
−ã0 − ˜l0 0 0 ··· 0

Can achieve arbitrary pole placement via choice of L̃! To solve original eigenvalue problem,
convert via L = M L̃ with M = Q−1 Qo .
Ding Zhao (CMU) M2-1: Pole Placement 25 / 46
Example
 
1 3  
Design an observer for A = , C = 1 0 with poles at -10, -20
4 2
Solution:
∆(s) = (s − 1)(s − 2) − 12 = s2 − 3s − 10
∆f b (s) = (s + 10)(s + 20) = s2 + 30s + 200
= s2 + (−3 + ˜l1 )s + (−10 + ˜l0 )
 
33
⇒L̃ =
210
Now convert back to original variables
   −1    
3 1 −1 1 0 1 0 1 0
à = , M = Q Qo = = 2 1
10 0 1 3 3 1 3 3
 
33
L = M L̃ =
92
Ding Zhao (CMU) M2-1: Pole Placement 26 / 46
Observer Structure

For SISO, the observer can be expressed as a 2-input ([u, y − ỹ]) and (n+1)-output system
([x̃, ỹ])

x̃˙ = Ax̃ + Bu + L(y − ỹ)

  T
= Ax̃ + B L u y − ỹ
     
x̃ In 0
yo = = x̃ + n u
ỹ C D

Ding Zhao (CMU) M2-1: Pole Placement 27 / 46

Table of Contents

1 PID Feedback Control

2 Design (State) Feedback Control

3 Design (Luenberger) observer

4 Separation Principle

5 MIMO Systems

6 (Optional) Reduced Order Observer

Ding Zhao (CMU) M2-1: Pole Placement 28 / 46

 SISO Feedback Control with Observer

By using an observer and performing state feedback on x̃, we can build an output feedback
controller.

x̃˙ = Ax̃ + Bu + L(y − ỹ)

u = K x̃ ỹ = C x̃ + DK x̃
⇒ẋ = Ax + BK x̃, y = Cx + Du ⇒x̃˙ = Ax̃ + BK x̃ + L(Cx + DK x̃ − (C x̃ + DK x̃))
= (A − LC + BK)x̃ + LCx
    
ẋ A BK x
Now stack ˙ =
x̃ LC A − LC + BK x̃

Ding Zhao (CMU) M2-1: Pole Placement 29 / 46

SISO Feedback Control

 
I 0
Transfom via Me = . Me−1 = Me
I −I
    
x(t) I 0 x
=
e(t) I −I x̃
      
ẋ I O A BK I O x
⇒ =
ė I −I LC A − LC + BK I −I e
  
A + BK −BK x
=
0 A − LC e

Ding Zhao (CMU) M2-1: Pole Placement 30 / 46

Separation Principle

Can show through properties of the determinant that the characteristic equations is:

∆(s) = |sI − (A − LC)| · |sI − (A + BK)|

⇒ The eigenvalues of the augmented system are the same as the separate controller/observer.
-This is called the separation principle. Observers & controllers can be designed separately.
-However, we should be careful to make the observer faster (3-5 times to the left on complex
plane) than the controller.

Ding Zhao (CMU) M2-1: Pole Placement 31 / 46

Structure of the Full Controller

The full controller takes the form

+ ey
y
 
u x̃
x̃˙ = Ax̃ + [B L] u
- ey ỹ u = K x̃
      
x̃ I 0 0 u
= x̃ +
ỹ C D 0 ey

Ding Zhao (CMU) M2-1: Pole Placement 32 / 46

Controller Transfer Functions

-Let’s now look at the controller TF to compare with classical methods,

x̃˙ = Ax̃ + BK x̃ + L(y − (C x̃ + DK x̃))

= (A + BK − LC − LDK)x̃ + Ly
u = K x̃

⇒ C(s) = YU (s)
(s) = K(sI − (A + BK − LC − LDK)) L
−1

-For a full state observer, the controller has order n

Ding Zhao (CMU) M2-1: Pole Placement 33 / 46

Table of Contents

1 PID Feedback Control

2 Design (State) Feedback Control

3 Design (Luenberger) observer

4 Separation Principle

5 MIMO Systems

6 (Optional) Reduced Order Observer

Ding Zhao (CMU) M2-1: Pole Placement 34 / 46

MIMO Systems

MIMO systems
For multi-input controllers and multi-output observers, the pole placement problem does
not have a unique solution.
Fundamentally, need to solve the eigenvalue problem to set eigenvalue of A + BK
Note that the observer design problem can be written in the same form by placing
eigenvalue of AT − C T LT

Ding Zhao (CMU) M2-1: Pole Placement 35 / 46

MIMO Systems

MIMO systems
For multi-input controllers and multi-output observers, the pole placement problem does
not have a unique solution.
Fundamentally, need to solve the eigenvalue problem to set eigenvalue of A + BK
Note that the observer design problem can be written in the same form by placing
eigenvalue of AT − C T LT
(Dual Problem)
ẋ = AT x + C T u
 
ẋ = Ax + Bu
The observability of ⇔ Controllability of
y = Cx + Du y = B T x + Du
The python scipy.signal.place poles command solves the MIMO problem as well

Ding Zhao (CMU) M2-1: Pole Placement 35 / 46

Design Controllers for MIMO Systems

By hand, use a similarity transformation and solve a Lyapunov equation.

Sylvester, 1814-1897
1 Create an n × n matrix F with desired eigenvalues with no
overlapping eigenvalues of A (restriction on this method)


2 Select an arbitrary p × n matrix K such that F, K is observable
3 Solve for T in AT − T F = −BK (Sylvester equation), if T is singular, go back to 2
4 Use K = KT −1 as feedback controller
Proof: Similarity transformation! AT − T F = −BKT , (A + BK)T = T F
⇒ A + BK = T F T −1 ⇒ We wantA + BK is similar to F

Ding Zhao (CMU) M2-1: Pole Placement 36 / 46

Example
   
0 1 0 0 0 0
 0 0 1 0 0 0
A= −3 1 2 3, B = 1 2
  

2 1 0 0 0 2
λd = −4 ± 3j, −5 ± 4j
Write F in Controllable Canonical Form

∆(s) = (s + 4 + 3j)(s + 4 − 3j)(s + 5 + 4j)(s + 5 − 4j)

= s4 + 18s3 + 146s2 + 578s + 1025
 
0 1 0 0
 0 0 1 0 
⇒F =
 0

0 0 1 
−1025 −578 −146 −18

Ding Zhao (CMU) M2-1: Pole Placement 37 / 46

Example

 
  K
1 0 0 0  KF 
Let K = (can check rank KF 2  = 4 Now use Python
 
0 1 0 0
KF 3
(scipy.linalg.solve
 sylvester) to solve the Sylvester equation

−1 −0.2256 0.0641 0.008 0.5027
⇒ K = KT = Can check ...
−146.73 −40.66 −9.61 −0.39
numpy.linalg.eig(A + BK) = −4 ± 3j, −5 ± 4j

Ding Zhao (CMU) M2-1: Pole Placement 38 / 46

Example

 
1 0 0 1
Control is not unique! with K = get
0 1 1 0
 
−267.98 −237.25 −22.50 44.44
K= , same eigenvalues!
−13.73 −5.12 0.91 0.33

Ding Zhao (CMU) M2-1: Pole Placement 39 / 46

Design Observer for MIMO Systems

An analogous procedure exists for observers:

1 Create an n × n matrix F with desired eigenvalues ̸= eigenvalues of A (restriction on this
method)


2 Select an arbitrary n × q matrix L such that F, L is controllable
3 Solve Sylvester equation T A − F T = LC, if T is singular, go to (2)
4 Use L = T −1 L as observer gains

Ding Zhao (CMU) M2-1: Pole Placement 40 / 46

Table of Contents

1 PID Feedback Control

2 Design (State) Feedback Control

3 Design (Luenberger) observer

4 Separation Principle

5 MIMO Systems

6 (Optional) Reduced Order Observer

Ding Zhao (CMU) M2-1: Pole Placement 41 / 46

Reduced Order Observer

-In some cases, the output contains direct measurement of a subset of the state variables - no
need to estimate those.  
−1 q linearly independent rows of C
-The transformation M = transforms any system into
n − q additional ind. rows
the form
      
ẋ1 A11 A12 x1 B1
= + u
ẋ2 A21 A22 x2 B2
 
  x1
y = Iq×q 0 + Du = x1 + Du
x2

Ding Zhao (CMU) M2-1: Pole Placement 42 / 46

Reduced Order Observer
Note that x1 is measured by y & we can rewrite as

A12 x2 = ẋ1 − A11 x1 − B1 u

ẋ2 = A22 x2 + B2 u + A21 x1
Define: u ≜ A21 x1 + B2 u
y ≜ ẋ1 − A11 x1 − B1 u
⇒ ẋ2 = A22 x2 + u
y = A12 x2

We can now observe x2 via

predictioncorrection
z }| { z }| {
x̃˙ 2 = A22 x̃2 + u + L(y − ỹ) = (A22 − LA12 )x̃2 + u + Ly

where ȳ˜ = A12 x̃2 and ȳ can be obtained from direct measurement ȳ = ẋ1 − A11 x1 − B1 u
Ding Zhao (CMU) M2-1: Pole Placement 43 / 46
Reduced Order Observer

-Now look at the error

ė = ẋ2 − x̃˙ 2
= A22 x2 + u − (A22 x̃2 + u + L(A12 x2 − A12 x̃2 ))
= A22 (x2 − x̃2 ) − LA12 (x2 − x̃2 )
= (A22 − LA12 )e

Check e-values of A22 − LA12 for convergence. If (A, C) observable, we could always find L
to stablize the system.

Ding Zhao (CMU) M2-1: Pole Placement 44 / 46

Reduced Order Observer

However, what if we don’t have access to ẋ1 ?

Define: z = x̃2 − Lx1

ż = x̃˙ 2 − Lẋ1
= (A22 − LA12 )x̃2 + u + Ly − Lẋ1
= (A22 − LA12 )x̃2 + u − L(A11 x1 + B1 u)
ż = (A22 − LA12 )z + [(A22 − LA12 )L + A21 − LA11 ]x1 + (B2 − LB1 )u

We could then use it to estimate z and then get x̃2 = z + Lx1

Ding Zhao (CMU) M2-1: Pole Placement 45 / 46

Reduced Order Observer

Design L to place the observer eigenvalues.

Design state feedback as before with
   
x1 y − Du
=
x̃2 z + Lx1
   
x y − Du
u=k 1 =K
x̃2 z + Lx1

Ding Zhao (CMU) M2-1: Pole Placement 46 / 46