COLE POLYTECHNIQUE

FDRALE DE LAUSANNE

Robust Control


Alireza Karimi
Laboratoire dAutomatique

Course Program
1. Introduction
2. Norms for Signals and Systems

10. Linear Fractional Transformation

3. Basic Concepts (stability and performance)

11.

4. Uncertainty and Robustness

12. Model and Controller Reduction

5. Stabilization (coprime factorization)

13. Robust Control by Convex Optimization

6. Design Constraints

14. LMIs in Robust Control

7. Loopshaping

15. Robust Pole Placement

8. Model Matching

16. Parametric uncertainty

H Control

9. Design for Performance

References:

Feedback Control Theory by Doyle, Francis and Tannenbaum (on the website of the course)
Essentials of Robust Control by Kemin Zhou with Doyle, Prentice-Hall, 1998
Robust Control

Introduction
Classical Control

Robust Control
-

u(t)

P0

y(t)
-

u(t)

P0

? y(t)
-

P = P0 +
P0 : nominal model

: plant uncertainty

Uncertainty sources:
Structured: parametric uncertainty, multimodel uncertainty
Unstructured: frequency-domain uncertainty, unmodeled dynamics, nonlinearity

Robust Control Objective: Design a controller satisfying stability and

performance for a set of models
Robust Control

Model Uncertainty and Feedback

Feedback control has been used for the first time to overcome the model uncertainty

r(t)-

T =
S=

CP
1+CP
1
1+CP

e(t)
-

u(t)
-

v(t)
-+ ? y(t)

Transfer function between r and y

Transfer function between v and y

For very large CP , T

1 (tracking) and S 0 (disturbance rejection) whatever the plant model is.

For an open-loop stable system:

C = 0 robust stability
Loopshaping:

- C robust performance

|C(j)P (j)| should be large in the frequencies where good performances are

desired and small where the stability is critical

Introduction

Norms for Signals and Systems

Norms for signals: Consider piecewise continuous signals mapping (, +) to R. A
norm must have the following four properties:
1.

u 0 (positivity)

2.

au = |a| u, a R (homogenity)

3.

u = 0 u(t) = 0 t (positive definiteness)

u + v u + v (triangle inequality)

1-Norm: u1 =
|u(t)|dt
4.



2-Norm:

u2 =

1/2
u2 (t)dt

u22 is the total signal energy

-Norm: u = sup |u(t)|



p-Norm:

up =

Robust Control

1/p
|u(t)|p dt

1p
5

Norms for Signals

Average power of a signal is denoted by:


pow(u) =

1
T 2T

lim

Remark: One says u(t)

1/2
u2 (t)dt

pow is not a norm (it is not positive definite)

Lp if up < where Lp is an infinite-dimensional Banach space (L 2

is an infinite-dimensional Hilbert space as well).

Recall:

A Banach space is a complete vector space with a norm.

A Hilbert space is a complete vector space with an inner product < x, y > such that the norm

defined by x = < x, x >. A Hilbert space is always a Banach space, but the converse need
not hold.

1(t) L but
/ L1 ,
/ L2
u(t) = (1 et )1(t) L but
/ L1 ,
/ L2
u(t) = et 1(t) L , L1 , L2 (Besides, u(t) is a power signal pow(u) = 0)
u(t) = sin(t) L but
/ L1 ,
/ L2 (u(t) is a power signal)
Examples:

Norms for Signals and Systems

Norms for Systems

We consider linear, time-invariant, causal and usually finite-dimensional systems.

y(t) = G(t) u(t),

y(t) =

G(t )u( )d ,

G(s)
= L[G]

Some definitions:

G(s)
is stable if it is analytic in the closed RHP (Re s 0)

G(s)
is proper if G(j)
is finite (deg den deg num)

= 0 deg den > deg num

G(s)
is strictly proper if G(j)

G(s)
is biproper if (deg den = deg num)
Norms for SISO systems:
2-Norm:

2=
G

1
2

|G(j)|
d

Parsvals theorem: (for stable systems)

2=
G

1
2

|G(j)|
d

Norms for Signals and Systems

1/2

1/2



=

= sup |G(j)|

-Norm: G

1/2
|G(t)|2 dt
7

Norms for Systems

Remarks:

L2 is a Hilbert space of scalar-valued functions on jR. The inner product for this Hilbert space is
defined as:

>= 1
< F , G
2

F (j)G(j)d

is strictly proper and has no poles on the

2 < . It is the case iff G
L2 if G
We say G(s)
imaginary axis.

is proper and has no poles on the

< . It is the case iff G
L if G
We say G(s)
imaginary axis. L is a Banach space of scalar-valued functions on jR.
. It is also the distance from the origin to
is the peak value of the Bode magnitude plot of G
G
.
the farthest point on the Nyquist plot of G

H2 and H are respectively subspaces of L 2 and L with G(s)

stable (Hp spaces are
usually called Hardy spaces).
1
Examples: s1

L2 , L but
/ H2 , H

Norms for Signals and Systems

s+1
s+2

H but
/ H2

s2 +1

/ L2 , L
8

Norms for Systems

Norms for matrices:
1-Norm: The maximum absolute column sum norm is defined as A 1
2-Norm: The spectral norm or simply the norm of A is defined as:

= max

A 2 =

-Norm: The maximum absolute row sum norm is defined as A = max
i

F-Norm: Frobenius norm is defined as A F

n


|aij |.

i=1

max (A A).
n


|aij |.

j=1

trace(A A)

Axp
Induced p-norm: The induced p-norm is defined from a vector p-norm: A p = max
x=0 xp
Remarks:

The induced 2-norm and the norm of A are the same and called also the natural norm. This norm
is also equal to the maximum singular value of A. A =
(A).
The spectral radius (A) = |max (A)| is not a norm.
Norms for Signals and Systems

Norms for Systems

Norms for MIMO systems: Given

G(s)
a multi-input multi-output system

2-Norm: This norm is defined as

2=
G

1
2

1/2

(j)G(j)

trace G
d

-Norm: The H norm is defined as

= sup G(j)

G
= sup
[G(j)]

Remark: The infinity norm has an important property (submultiplicative)

H
G
H

G

Norms for Signals and Systems

10

Computing the Norms

How to compute the 2-norm: Suppose that G

1
2

G2 =
2

1
2

|G(j)| d =
2j

L2 , we have:
j

G(s)G(s)ds =
2j
j

22 equals the sum of the residues of

Then by the residue theorem,  G

G(s)
G(s)ds

G(s)
G(s)
at its poles in the

left half-plane (LHP).

How to compute the -norm:

is:
Choose a fine grid of frequency point { 1 , . . . , N }, then an estimate for G

max |G(j
k )|

1kN

or for the MIMO systems:

max
[G(j
k )]

1kN

Matlab commands:

norm, bode, frsp, vsvd

Norms for Signals and Systems

11

Computing the Norms

State-space methods for 2-norm: Consider a SISO state-space model

x(t)

= Ax(t) + Bu(t)
y(t) = Cx(t)

G(s)
= C(sI A)1 B
From Parsevals theorem:

2 = G2 =
G
2
2

s
x(s) = A
x(s) + B u
(s)
y(s) = C x
(s)

impulse response

CetA B

H2

: G(t) = CetA B

B T etA C T dt = CLC T

where

L=

etA BB T etA dt

is the observability Gramian and can be obtained from following Lyapunov equation:

2 = (CLC T )1/2
AL + LAT + BB T = 0 and the 2-norm is G


T 1/2

For MIMO systems we have  G2 = trace(CLC )

Norms for Signals and Systems

12

Computing the Norms

State-space methods for -norm: Consider a SISO strictly proper state-space model
Theorem:

L.

< iff the Hamiltonian matrix H has no eigenvalues on the imaginary axis:
G

A
2 BB T

H=
C T C
AT

Bisection algorithm:
1. Select u and l such that l

u ;
G

2. If (u

(u + l )/2. Otherwise continue;

l )/l specified level., stop and  G

3. Set

<
= (u + l )/2 and test if G

4. If H has no eigenvalue on jR, set u

For MIMO
biproper (D

H=

= otherwise set l = ; go back to step 2.

= 0) systems:

A + BR1 D T C

BR1 B T

C T (I + DR1 D T )C

(A + BR1 D T C)T

Norms for Signals and Systems

and R = 2 I D T D
13

Input-output relationships
If we know how big the input is, how big is the output going to be?
Output Norms for Two Inputs

sin(t)

y2

(t)
2
G

y

G

|G(j)|

u(t)

Proofs:

If u(t) = (t) then y(t) =

2
G(t )( )d = G(t), so y2 = G2 = G

If u(t) = (t) then y(t) = G(t), so y = G

If u(t) = sin(t) then y(t) = |G(j)|

sin(t + ), so y2 = and y = |G(j)|

Norms for Signals and Systems

14

Input-output relationships
System Gain:

sup{y : u 1}
u(t) L

y2

u(t) L2

G

y

2
G

G1

Proofs:
Entry (1,1): We have

1
2
2
y2 = 
y 2 =
2

|G(j)|
|
u(j)|2 d

1
2
G
|
u(j)|2 d
2
2 
2 u22
= G
u22 = G

Entry(2,1): According to the Cauchy-Schwartz inequality




|y(t)| = 



G(t )u( )d 



G2 (t )d

1/2 

u2 ( )d

1/2

2 u2 y G

2 u2
= G2 u2 = G
Norms for Signals and Systems

15

Basic Concepts
r-

x1 -

u - ? x2 -

C
v

Basic Feedback Loop:

x1 = r F x3
x1
r
1
0
F

x2 = d + Cx1 = C
1
0
x2 = d

x3 = n + P x2
n
0
P 1
x3

1

x1
1
0
F
r
1
P F

1
x2 = C

C
1
0 d =
1

1 + P CF
0
P 1
n
PC
P
x3

P
x3

?
n

r
F


CF
d
n
1

Well-posedness: The system is well-posed iff the above matrix is nonsingular. A stronger notion of
well-posedness is that all the nine transfer functions be proper. A necessary and sufficient condition
for this is that 1 + P CF not be strictly proper (P CF ()
Robust Control

= 1)
16

Internal Stability
If the following nine transfer functions are stable then the feedback system is internally stable.

1
C
1 + P CF
PC

P F
1
P

CF

P =

NP
NC
NF
, C=
, F =
MP
MC
MF

Theorem: The feedback system is internally stable iff

there are no zeros in Re s 0 in the characteristic polynomial

NP NC NF + MP MC MF = 0
or

the following two conditions hold:

(a) The transfer function 1 + P CF has no zeros in Re s
(b) There is no pole-zero cancellation in Re s

Basic Concepts

0.

0 when the product P CF is formed.

17

Asymptotic Tracking
Internal Model Principle: For perfect asymptotic tracking of r(t), the loop transfer function

(with F

= P C
L

= 1) must contain the unstable poles of r(s).

Theorem: Assume that the feedback system is internally stable and n=d=0.
(a) If r(t) is a step, then

1 has at least one zero at

lim e(t) = r(t) y(t) = 0 iff S = (1 + L)

the origin.
(b) If r(t) is a ramp, then
(c) If r(t)

lim e(t) = 0 iff S has at least two zeros at the origin.

= sin(t), then lim e(t) = 0 iff S has at least one zero at s = j .

t

Final-Value Theorem: If y
(s) has no poles in Re s

0 except possibly one pole at s = 0 then:

lim y(t) = lim s

y (s)

s0

c
c
c

Proof (a): We have r

(s) = and e(s) = S(s) then lim e(t) = lim sS(s) . The limit is zero
t
s0
s
s
s
has at least one zero at origin. For this, P or C should have a pole at origin.
iff S
Basic Concepts

18

Performance
Tracking performance can be quantified in terms of a weighted norm of the sensitivity
function
1
Sensitivity Function: Transfer function from r to tracking error e : S =
1 + PC
PC
Complementary Sensitivity Function: Transfer function from r to y : T =
1 + PC
S is the relative sensitivity of T with respect to relative perturbations in P:

dT P
C(1 + P C) P C 2 P (1 + P C)
1
T /T
=
=
=
S = lim
P 0 P/P
dP T
(1 + P C)2
PC
1 + PC
Performance Specification:
1.

r(t) is any sinusoid of amplitude 1 filtered by W 1 , then the max. amp. of e is W1 S .

2. Suppose that {r

= W1 rpf , rpf 2 1}, then supr e2 = W1 S .

3. In some applications good performance is achieved if |S(j)|

< |W 1 (j)|1 , or

W1 S < 1 |W1 (j)| < |1 + L(j)|,

Basic Concepts

19

Uncertainty and Robustness

Plant Uncertainty: We cannot exactly model the physical systems so there is always the modeling
errors. The best technic is to define a model set which can be structured or unstructured.

Structured model set such as parametric uncertainty or multiple model set

P ={

1
: amin a amax } or P = {P0 , P1 , P2 , P3 }
2
s + as + 1

Unstructured model set such as unmodeled dynamics or disk uncertainty

P = {P0 + :

 }

Conservatism: Controller design for a model set greater than the real model set leads to a
conservative design.
Uncertainty Models (unstructured):
Additive uncertainty

Multiplicative uncertainty

P = P + W2

P = P (1 + W2 )

P : true model
Robust Control

P : nominal model

Feedback uncertainty

P
P
=
or P
1 + W2 P
1 + W2
: norm-bounded uncertainty W 2 : weighting filter
P =

20

Examples
Example 1:

k frequency-response models are identified. Find the multiplicative uncertainty model and

the weighting filter.

P
P = P (1 + W2 )
1 = W2
P 

 P (j)



if  1 
1 |W2 (j)|
 P (j)


Let (Mik , ik ) be the magnitude-phase at i in k -th experiment and (M i , i ) that of the nominal
model (e.g. the mean value).

Example 2: Suppose that




 Mik eik
 |W2 (ji )|

max 

1


k
Mi e i

k
P (s) = { s2
: 0.1 k 10}. Represent this model by the

multiplicative uncertainty.










k0
k
 P (j)



1 |W2 (j)| max  1 |W2 (j)|
P (s) =
 P (j)

0.1k10 k0
s2

The best value for k0 is 5.05 which gives W2 (s)

Uncertainty and Robustness

= 4.95/5.05
21

Examples
Example 3: Assume that P (s)

1
s2 and

P (s) = e s s12 where 0 0.1. Find W2 (s) for the

multiplicative uncertainty model.




 P (j)


1 |W2 (j)| |e j 1| |W2 (j)| ,


 P (j)

Using the Bode diagram we can find W 2 (s)

1
Example 4: Consider the model set { s2 +as+1

0.21s
0.1s+1 .

: 0.4 a 0.8}. Find W2 (s) for Feedback

uncertainty model.
Take:

a = 0.6 + 0.2, 1 1
So

P (s) =

1
P (s)
=
s2 + 0.6s + 0.2s + 1
1 + W2 (s)P (s)

where

P (s) =
Uncertainty and Robustness

1
, W2 (s) = 0.2s
2
s + 0.6s + 1
22

Robust Stability
Robustness: A controller is robust with respect to a closed-loop characteristic, if this characteristic
holds for every plant in P
Robust Stability: A controller is robust in stability if it provides internal stability for every plant in P
Stability margin: For a given model set with an associate size, it can be defined as the largest model
set stabilized by a controller.

Stability margin for an uncertainty model: Given P

= P (1 + W2 ) with  , the
stability margin for a controller C is the least upper bound of .
Im L

Modulus margin: The distance from -1 to the

open-loop Nyquist curve.

Mm

= inf | 1 L(j)| = inf |1 + L(j)|

1

1
= S1
=
sup

1 + L(j)

Uncertainty and Robustness

Re L

Mm
L(j)

23

Robust Stability
Small Gain Theorem: Suppose

(s)

H RH

> 0. The following feedback loop is

internally stable for all (s) RH with
and let

 1/ if and only if H <

H(s)

Remark: For a given with 

1/ the condition H < is only sufficient and very

conservative. However for all RH , it is a necessary and sufficient condition.
Robust stability condition for plants with additive uncertainty:

P = P + W2 H = W2

C
1 + CP

Closed-loop system is internally stable for

all 

iff

W2 CS < 1

Uncertainty and Robustness

- - W2
r(t)-

-+ ? y(t)

24

Robust Stability
Robust stability condition for plants with multiplicative uncertainty:

CP

P = P (1+W2 ) H = W2
1 + CP
Closed-loop system is internally stable for
all 

iff

W2 T  < 1

Proof: Assume that W2 T 

- - W2
r(t)- - P
- C

-+ ? y(t)

< 1. We show that the winding number of 1 + CP around zero is

.
equal to that of 1 + C P

1 + C P = 1 + CP (1 + W2 ) = 1 + CP + CP W2 = 1 + CP + (1 + CP )T W2
1 + C P = (1 + CP )(1 + W2 T )
)}
so Wno { (1 + C P

= Wno{(1 + CP )} + Wno{(1 + W2 T )}.

But Wno {(1 + W2 T )}

= 0 because W2 T  < 1

Uncertainty and Robustness

25

Robust Stability
Robust stability condition for plants with feedback uncertainty (1):

P =

P
1
H = W2
1 + W2
1 + CP

Closed-loop system is internally stable for

all 

iff

W2  
?
r(t)- P
- C

y(t)
-

W2 S < 1

Robust stability condition for plants with feedback uncertainty (2):

P
P

P =
H = W2
1 + W2 P
1 + CP
Closed-loop system is internally stable for
all 

iff

W2 P S < 1

Uncertainty and Robustness

W2 
r(t)-

-?
-

y(t)

26

Robust Performance
Nominal performance condition:

W 1 S < 1

Robust stability condition for multiplicative uncertainty:

Robust performance for multiplicative uncertainty:

W 2 T  < 1

< 1 where:
W 2 T  < 1 and W1 S

1
1
1
S
=
=
=

1
+
CP
(1
+
W
)
(1
+
CP
)(1
+
W
T
)
1 + W2 T
1 + CP
2
2


 W1 S 

Robust performance conditions: W2 T  < 1 and 
 1 + W2 T  < 1

S =

Theorem: A necessary and sufficient condition for robust performance is

 |W1 S| + |W2 T |  < 1

< 1 where:
W2 CS < 1 and W1 S




1
1
S
W1 S


 <1
=
S=

=
1 + CP + CW2
1 + W2 CS
1 + W2 CS 
1 + C P

Robust performance for additive uncertainty:

Or equivalently in one inequality condition:

Uncertainty and Robustness

 |W 1 S| + |W2 CS|  < 1

27

Stabilization
The main objective is to parameterize all of the controllers which provide internal stability for a
given plant
Theorem: Assume that P

Proof: (F

= 1)

1
C
1 + P CF
PC

RH (P is stable). The set of all stabilizing controllers is given by:



Q
| Q RH
C :=
1 PQ

P F
1
P

1 P Q P (1 P Q) 1(1 P Q)
F

RH

1 PQ
Q
CF

= Q
PQ
P (1 P Q)
1 PQ
1

On the other hand, suppose that C stabilizes P then define

C
Q
RH which leads to C =
Q :=
1 + CP
1 PQ
In this parameterization sensitivity and complementary sensitivity are

Robust Control

S = 1 PQ

T = PQ
28

Coprime Factorization
Objective: Given P , find M, N, X and Y

RH such that:

N
P =
M

NX + MY = 1

Remarks:

N and M are called coprime factors of G over RH

N and M can have no common zeros in Re s 0 nor at s =
N (s0 )X(s0 ) + M (s0 )Y (s0 ) = 0 = 1
If P is stable we have : M = 1, N = P, X = 0, Y = 1
It is easy to obtain N and M , for example:
P (s) =
if k

N (s)
1
=
s1
M (s)

N (s) =

1
s1
,
M
(s)
=
(s + 1)k
(s + 1)k

> 1 then M and N have a common zero at s = , so k = 1

How to compute X(s) and Y (s)?

Stabilization

29

Coprime Factorization
Euclids algorithm: Given polynomials m() and n() (deg n
and y() such that nx + my

deg m ) find polynomials x()

= 1.

Step 1: Divide m into n to get quotient q1 and remainder r1 :

n = mq1 + r1 , deg r1 < deg m

Step 2: Divide r1 into m to get quotient q2 and remainder r2 :

m = r1 q2 + r2 , deg r2 < deg r1

Step 3: Divide r2 into r1 to get quotient q3 and remainder r3 : r1

= r2 q3 + r3 , deg r3 < deg r2

Continue Stop at step k when rk is a nonzero constant.

Find r3 as a function of m, n and q i :
r2




r3 = (n mq1 ) (m (n mq1 ) q2 ) q3 = n(1 + q2 q3 ) + m(q3 q1 q1 q2 q3 )
 
 
r1

which gives:

Stabilization

r1

1
1
x = (1 + q2 q3 ) and y = (q3 q1 q1 q2 q3 )
r3
r3
30

Coprime Factorization
Procedure to find M, N, X and Y for an unstable plant G:

n()

Step 1: Transform G(s) to G() under the mapping s = (1 )/. Write G =

m()
Step 2: Using Euclids algorithm, find x() and y() such that:

nx + my = 1
= 1/(s + 1)

Step 3: Find M, N, X and Y from m, n, x and y under the mapping

State-Space Method:
Step 1: Transform G(s) to A, B, C and D (state space realization)

Step 2: Compute F and H so that A + BF and A + HC are stable (F=-place(A,B,Pf))

Step 3: Compute M, N, X and Y as follows:

M (s) :=

X(s) :=
Stabilization

A + BF

A + HC

N (s) :=

Y (s) :=

A + BF

C + DF

A + HC

B HD

31

Controller Parametrization
Theorem: The set of all C s for which the feedback system is internally stable equal:

C=

X + MQ
:
Y NQ

Q RH

Nc
Proof: For C =
, the stability condition is: (N Nc + M Mc )1 RH , but we have:
Mc
N (X + M Q) + M (Y N Q) = N X + M Y = 1 (N Nc + M Mc )1 RH
Conversely, if C stabilizes the closed-loop system we should show that it belongs to the above set.

C is stabilizing V := (N Nc + M Mc )1 RH N Nc V + M Mc V = 1
= Y N Q. From the above equation and N X + M Y = 1 we find
Nc V
that Nc V = X + M Q so the controller C =
the set of all stabilizing controller. It is easy
Mc V
to verify that Q RH
Let Q be the solution of Mc V

Remark: The sensitivity functions are:

1
= M (Y N Q)
S=
1 + CP
Stabilization

CP
T =
= N (X + M Q)
1 + CP
32

Example
Let

P (s) =

1
(s 1)(s 2)

Compute a proper controller C so that:

1. The feedback system is internally stable.
2. Perfect asymptotic tracking of step reference (d

= 0).

3. Perfect asymptotic disturbance rejection when d

= sin 10t (r = 0).

Procedure:

Parameterize all stabilizing controllers.

Reduce the asymptotic specs to interpolation constraints on the parameters.
Find (if possible) a parameter to satisfy these constraints.
Back-substitute to get the controller.

Stabilization

33

Design Constraints
Algebraic Constraints:

S + T = 1 so |S(j)| and |T (j)| cannot both be less than 1/2 at the same frequency.
A necessary condition for robust performance is that:
min{|W1 (j)|, |W2 (j)|} < 1,
So at every frequency either |W 1 | or |W2 | must be less than 1. Typically |W 1 | is monotonically
decreasing and |W 2 | is monotonically increasing.

If p is a pole and z a zero of L both in Re s 0 then:

S(p) = 0 S(z) = 1 T (p) = 1 T (z) = 0
Analytic Constraints:

Bounds on the weights W 1 and W2 :

W1 S |W1 (z)|
Proof from the Maximum Modulus Theorem:

W2 T  |W2 (p)|

F  = sup |F (s)|
Re s>0

Robust Control

34

Analytic Constraints
All-Pass and Minimum-Phase Transfer Functions:

F (s) RH is all-pass if |F (j)| = 1

G(s) RH is minimum-phase if it has no zeros in Re s > 0. It has the minimum phase
among all transfer functions with the same magnitude (F G where F is all-pass).
Every function G in RH can be presented as G = G ap Gmp
Suppose that L = CP has no poles on the imaginary axis, so S = (1 + L) 1 = Sap Smp and
1
Smp has no zeros on the imaginary axis. Thus S mp
RH .
Suppose that z and p are the only zero and pole of P in the closed RHP and C has neither poles
nor zeros there. Then:

sp
=
s+p

z+p
S(z) = 1 Smp (z) =
Sap
=
zp



z + p 

Then: W1 S = W1 Smp  |W1 (z)Smp (z)| = W1 (z)
z p



sz
p + z 

Similarly: Tap =
and T (p) = 1 W2 T  W2 (p)
s+z
p z
Design Constraints

1
Sap
(z)

35

Analytic Constraints
Example: Consider the inverse pendulum problem.

(M + m)
x + ml( cos 2 sin ) = u
m(
x cos + l g sin ) = d
Linearized model:

x
ls2 g
1
=

2
2
s [M ls (M + m)g]
s2

ls2
u

M +m 2
d
m s

Tux

ls g
= 2
s [M ls2 (M + m)g]




RHP poles and zeros:

z=

q
- m

x -

uM

%
g/l p = 0, 0,

(M + m)g
Ml

1
g
=
no RHP zero
T
uy
2
2
2
M ls (M + m)g
s [M ls (M + m)g]
For Tux if m  M W2 T   1 (|W2 (p)| is an increasing function) the system is difficult to
control. The best case is m/M and l large.
Tu =

For Tu and Tuy a larger l gives a smaller p so the system is easier to stabilize.
Design Constraints

36

Analytic Constraints
The Waterbed Effect

= 0 + j0 with 0 > 0,


1
0
log |Smp (s0 )| =
log |S(j)| 2
d
2

0 + ( 0 )

Lemma: For every point s 0

Theorem: Suppose that P has a zero at z with Re z

M1 :=

max

1 2

> 0 and:

|S(j)|

M2 := S

Then there exist positive constants c 1 and c2 , depending only on 1 , 2 and z , such that :
1
c1 log M1 + c2 log M2 log |Sap
(z)| 0

Theorem (The Area Formula): Assume that the relative degree of L is at least 2. Then

log |S(j)|d = (log e)

where {pi } denotes the set of poles of L in Re s

Design Constraints

Re pi

> 0.
37

Loopshaping
Objective: Given P, W1 and W2 find controller C providing internal stability and robust performance:

 |W1 S| + |W2 T |  < 1 or

 


 W1 (j)   W2 (j)L(j) 
+
<1
(j) := 


1 + L(j)
1 + L(j) 

Idea: Find graphically L(j) satisfying the above condition and then compute C

= L/P

Note that we assume P is minimum phase and stable.

We have:

|1 + L| = |W1 | + |W2 L| and |1 |L| | |1 + L| 1 + |L|

|W1 | + |W2 L|
|W1 | + |W2 L|

1 + |L|
|1 |L| |

So if |W1 | + |W2 L|

< |1 |L| | < 1:

In low frequencies

|L| > 1

|L| >

|W1 |
|W1 | + 1

1 |W2 |
1 |W2 |

|W1 |  1 > |W2 |

In high frequencies

|L| < 1

|L| <

1 |W1 |
1 |W1 |

1 + |W2 |
|W2 |

|W2 |  1 > |W1 |

Robust Control

38

Procedure
step 1: Plot two curves on log-log scale:
at LF (|W1 |

> 1 > |W2 |)

|W1 |
and at HF
1 |W2 |

(|W2 | > 1 > |W1 |)

1 |W1 |
|W2 |

step 2: Fit the graph of |L| on the same plot such that:

at low frequency it lies above the first curve and also  1

at high frequency it lies below the second curve and  1
at very high frequency let it roll off at least as fast as does |P | (so C is proper)
near crossover frequency do a smooth transition, keeping the slope as gentle as possible.
Because the slope of |L| determines the phase of L (Bodes integral):


||
1
d ln |L|
ln coth
d
L(j0 ) =
where = ln(/0 )
d
2
The steeper the graph of L near the crossover frequency, the smaller the value of L and
larger the phase margin
step 3: Get a stable, minimum-phase TF for L such that L(0)
Loopshaping

> 0 and compute C = L/P

39

Example
Assume that the relative degree of P equals 1. Find L for robust performance if the objective is to
track sinusoidal signals over the frequency range from 0 to 1 rad/s and the weighting function W 2 is:

s+1
W2 (s) =
20(0.01s + 1)
We can define W 1 as follows (in loopshaping design it is not necessary to have a rational TF for W 1 ):

a
|W1 (j)| =
0

01
else

LF (|W1 | > 1): < 1

Plot

The larger the value of a, the smaller the tracking error

HF (|W2 |

> 1): 20

|W1 |
1 |W1 |
in LF ( < 1) and
in HF ( > 20)
1 |W2 |
|W2 |

Choose L =

1
b
1 |W1 |
=
and find b such that in HF |L|
( |b| 20)
s+1
|W2 |
|W2 |

a
|W1 |
=
a = 13.15
Find the maximum value of a such that in LF |L|
1 |W2 |
1 |W2 |
Loopshaping

40

Model Matching
T1 (s) and T2 (s), stable
proper transfer functions, find a stable Q(s) to
minimize T1 T2 Q
Trivial case: If T1 /T2 is stable then the unique
optimal Q is T1 /T2 and

Objective: Given

r(t)

- T2

- T1

? (t)

opt = min T1 T2 Q = 0

Simplest nontrivial case:

T2 has only one RHP zero at s = s 0 . Then by the maximum modulus

theorem:

T1 T2 Q |T1 (s0 ) T2 (s0 )Q(s0 )| = |T1 (s0 )| opt |T1 (s0 )|
T1 T1 (s0 )
Note that Q =
is stable and leads to opt = |T1 (s0 )|.
T2
Example:

T1 (s) =

Robust Control

4
s+3 ,

T2 (s) =

s2
(s+1)3

Q=

T1 T1 (2)
T2

4(s+1)3
5(s+3)

41

Nevanlinna-Pick Problem
Problem: Let {a1 , . . . , an } be a set of points in the open RHP and {b 1 , . . . , bn } a set of distinct
points in complex plane. Find a stable, proper, complex-rational function G satisfying:

G 1 and G(ai ) = bi , i = 1, . . . , n

1 bi bj
Solvability: The NP problem is solvable iff the n n Pick matrix Q, whose ij th element is
ai + aj
is positive semidefinite (Q 0). Note that Q is Hermitian (Q = Q where Q is the complex
conjugate transpose of Q). Q 0 iff all its eigenvalues are 0.
Mobius Function: A Mobius function has the form:

zb
Mb (z) =
1 zb

where |b|

<1

Mb has a zero at z = b and a pole at z = 1/b so M b is analytic in open unit disk..

Mb maps the unit disk onto the unit disk and the unit circle onto the unit circle.
The inverse map Mb1 =
Model Matching

z+b
= Mb is a Mobius function too.
1 + zb
42

Nevanlinna-Pick Problem
= 1: Find a stable, proper G(s) such that G 1 and G(a1 ) = b1 where
|b1 | 1 and Re a1 > 0.
NP problem for n

Case 1 |b1 |

= 1: The unique solution is G(s) = b1 .

Case 2 |b1 |

< 1: The set of all solutions is:

{G : G(s) = Mb1 [G1 (s)Aa1 (s)], G1 CRH , G1  1]}

where the all-pass function A a (s)

Example: For a1

:=

sa
s+a

= 2 and b1 = 0.6 we have: G(s) =

G1 (s) s2
s+2 + 0.6
1 + 0.6G1 (s) s2
s+2

s 0.5
G1 (s) = 1 results in G(s) =
s + 0.5
Remark 1: If G1 is an all-pass function, so is G
Remark 2: When ai are the complex-conjugate pairs, if G

= G R + jGI is the solution of the NP

problem then GR is also a solution to the NP problem.

Model Matching

43

Nevanlinna-Pick Problem
Consider the NP problem with n points:
Case 1 |b1 |

= 1: G(s) = b1 is the unique solution (and hence b 1 = b2 = = bn ).

< 1: Pose the NP problem with n 1 data points: {a 2 , . . . an } and {b2 , . . . , bn }
where bi := Mb1 (bi )/Aa1 (ai ) i = 2, . . . , n

Case 2 |b1 |

Lemma: The set of all solutions to the NP problem is G(s)

= M b1 [G1 (s)Aa1 (s)] where G1 (s)

ranges over the solutions to the NP problem.

Example: Consider the NP problem with a

= {1, 2} and b = {1/2, 1/3}.

Solvability:
solvable,
The problem is
because

Q=

1b21

2a1
1b2 b1
a2 +a1

1b1 b2
a1 +a2
1b22
2a2

3/8

5/18

5/18
eig(Q) = [0.5867 0.0105] Q 0
2/9

b2 b1
0.2
1b
2 b1

= 0.6
NP problem: a2 = 2, b2 = a a =
2
1
1/3
a2 +a1
s8 s1
1
+
s2 3s + 8
s+8 s+1
2
NP problem: G(s) =
1 s8 s1 = s2 + 3s + 8
1 + 2 s+8 s+1

Model Matching

G1 (s) =

s2
s+2

0.6

1 0.6 s2
s+2

s8
=
s+8

44

Model Matching Problem

Find Q such that

opt = min{T1 T2 Q }

Define:

G=

1
(T1 T2 Q)

T1 G
. However, to ensure the
T2
stability of Q, T1 G should contain the RHP zeros of T 2 (i.e. zi ), that is:

We find first G such that G

1 then we compute Q =

1
G(zi ) = T1 (zi ) G(zi ) = T1 (zi )

This is a NP problem and opt is the smallest for which the problem has a solution. That is, the
associated Pick matrix is positive semidefinite.

Aij =

1
zi + zj

A 2 B 0 where :
Bij =

T1 (zi )T1 (zj )

zi + zj

Lemma: opt equals the square root of the largest eigenvalue of the matrix A 1/2 B A1/2 .

Model Matching

45

Model Matching Problem

Procedure: Given T1 and T2 find a stable Q to minimize T 1

T2 Q (T1=tf(num,den))

> 0.
zz=zero(T2);z=zz(find(real(zz)>0))

Step 1: Determine zi the zeros of T2 in Res

Step 2: Form the matrices A and B :

Aij =

1
zi + zj

Bij =

T1 (zi )T1 (zj )

zi + zj

Step 3: Compute opt as the square root of the largest eigenvalue of the matrix A 1/2 B A1/2 .

gamma=sqrt(max(eig(inv(sqrtm(A)) *B*inv(sqrtm(A)))))
Step 4: Find G, the solution of the NP problem with data:

z1

...

zn

1
1
opt
T1 (z1 ) . . . opt
T1 (zn )

Step 5: Set

Q=

Model Matching

T1 opt G
T2

Q=minreal((T1-gamma*G)/T2,0.01)
46

Model Matching Problem

State-Space Procedure:
Step 1: Factor T2 as the product of an all-pass T 2ap and a minimum phase factor T 2mp
Step 2: Define R

:=

T1
and factor R as R = R1 + R2 with R1 strictly proper with all poles in
T2ap

RHP and R2 H and find a minimum realization of R 1 (s) =

Step 3: Solve the Lyapunove equations:

ALc + Lc A

= BB 

A Lo + Lo A = C  C
Step 4: Find the maximum eigenvalue 2 of Lc Lo and a corresponding eigenvector w .

Step 5: Define:

f (s) =

Step 6: Then opt

Model Matching

g(s) =

A

1 Lo w

B

(s)
= and Q = (R fg(s)
)/T2mp

47

Design for Performance

Objective: Find a proper C for which the feedback system is internally stable and W 1 S
Lemma: If G is stable and strictly proper, then

lim G(1 J)

<1
1
= 0 where J(s) =
( s + 1)k

P and P 1 stable: In this case the set of all stabilizing controller is:
Q
C=
1 PQ

Q H and

W1 S = W1 (1 P Q)

= P 1 is stable but not proper, so lets try Q = P 1 J to make it proper. Then

W1 S = W1 (1 J) whose -norm is less than 1 for sufficiently small .
Clearly, Q

P 1 stable:
Do a coprime factorization of P = N/M, N X + M Y = 1
Set J = ( s + 1)k with k = the relative degree of P
Choose so small that W1 M Y (1 J) < 1
Set Q = Y N 1 J and C = (X + M Q)/(Y N Q)
Robust Control

48

P 1 Unstable (General Case)

P has no poles or zeros on the imaginary axis, only distinct poles and zeros in the
RHP and at least one zero in the RHP. W 1 is stable and strictly proper.

Assumptions:

Procedure:
Step 1: Do a coprime factorization of P

= N/M, N X + M Y = 1

Step 2: Find a stable improper Qim such that:

W1 S = W1 M (Y N Qim ) < 1

It is a standard model matching problem that can be solved using the NP algorithm.

1
with k = large enough that Q is proper and small enough that
Step 3: Set J =
( s + 1)k
W1 M (Y N Qim J) < 1
Step 4: Set Q

= Qim J

Step 5: Set C

= (X + M Q)/(Y N Q)

Design for Performance

49

Design Example
Flexible Beam: Consider the following simplified plant transfer function:

zeros 4.91
5.53
poles
0
0.0007

6.47s2 + 4.03s + 176

P (s) =
s(5s3 + 3.57s2 + 140s + 0.093)

0.356 5.27j

Performance Specification: Settling time 8s and overshoot 10%

Assume that the ideal T (s) is a standard second-order system:

n2
Tid (s) = 2
s + 2n s + n2

4.6
8
n

exp


1 2


= 0.1 = 0.6 n = 1

s(s + 1.2)
Then the ideal sensitivity function is S id (s) = 1 Tid (s) = 2
s + 1.2s + 1
1

We take the weighting function W 1 (s) to be Sid

(s):

s2 + 1.2s + 1 stable, strictly proper

s2 + 1.2s + 1
W1 (s) =
=
W1 (s) =
s(s + 1.2)
(s + 0.0001)(s + 1.2)(0.0001s + 1)
Design for Performance

50

Design Example
Step 1:

P (s) has a pole on the imaginary axis (s = 0) so we perturb P to fix the problem (we add

106 to the denominator)

Step 2: The model matching problem is to minimize:

W 1 S = W1 (1 P Qim )

P has only one RHP zero at 5.53, thus min W 1 (1 P Qim ) = |W1 (5.53)| = 1.02 and
the specification is not achievable.
Step 3: Let us scale W1 as W1
Step 4: Set J(s)

norm

0.1

1.12

0.05

1.01

0.04

0.988

C=

0.9
W1 0.9
W1 . Then the optimal Qim =
1.02
W1 P

1
and compute W 1 (1 P Qim J) for decreasing values of
3
( s + 1)

Step 5:

:=

take

= 0.04 and set Q = Qim J

(W1 0.9)J
Q
=
P 1
1 PQ
W1 (1 J) + 0.9J

Design for Performance

51

2-Norm Minimization
Objective: Given P and W , find a proper stabilizing controller to minimize the 2-norm of a weighted
closed-loop transfer function: e.g.

min W P S 2

Define: The subspace of functions in L 2 that are analytic in the open RHP (all poles with Res

the orthogonal complement of H 2 and is denoted by H 2 . Every function F

as F

= Fst + Fun where Fst H2 , Fun H2

Lemma: If F

0) is

L2 can be expressed

H2 and G H2 , then F + G22 = F 22 + G22

Problem: Obtain Q
Idea: Factor U

H to minimize W P S2 = W N Y W N 2 Q2

:= W N 2 = Uap Ump , then we have:

W N Y W N 2 Q22

1
= W N Y Uap Ump Q22 = Uap
W N Y Ump Q22
1
1
= (Uap
W N Y )un + (Uap
W N Y )st Ump Q22
1
1
= (Uap
W N Y )un 22 + (Uap
W N Y )st Ump Q22

which leads to:

1
1
1
Qim = Ump
(Uap
W N Y )st and the minimum of the criterion: (U ap
W N Y )un 2

To get a proper suboptimal Q, Q im should be rolled off at high frequency.

Design for Performance

52

Optimal Robust Stability

Objective: Given P

= (1 + W2 )P where  , find the controller C that stabilizes every

plant in P and maximizes the stability margin:

inf := inf W2 T 

C

sup = 1/inf

Procedure: Input P and W2

Step 1: Do a coprime factorization of P

= N/M , N X + M Y = 1

Step 2: Solve the model-matching problem:

W2 T  = W2 N (X + M Q)

and find Qim and sup
Step 3: Let 

with

T1 = W2 N X T2 = W2 N M

= 1/opt

< sup and set J(s) = ( s + 1)k where k is large enough that Q im J is proper and

small enough that:

W2 N (X + M Qim J)
Step 4: Set Q

1
<


= Qim J and C = (X + M Q)/(Y N Q)

Robust Control

53

Robust Performance Problem

Objective: Given P, W1 , W2 find a proper controller C so that the feedback system for the nominal
plant is internally stable and that:

 |W1 S| + |W2 T |  < 1

This problem cannot be solved !

Modified Problem: Consider the following inequality:

 |W1 S|2 + |W2 T |2  < 1/2

The robust performance problem with this inequality can be converted to a model matching problem
(See Feedback Control Theory chapter 12.3)
This inequality is a sufficient condition for the inequality in the exact problem.
General framework: The inequality in the modified problem can be presented also as:



 W1 S


 W2 T

Robust Control








= max max

|W1 S(j)|

< 1
2
|W2 T (j)|
54