Lecture 2

FRTN10 Multivariable Control

Automatic Control LTH, 2019
 Course Outline

L1–L5 Specifications, models and loop-shaping by hand

1 Introduction
2 Stability and robustness
3 Specifications and disturbance models
4 Control synthesis in frequency domain
5 Case study: DVD player
L6–L8 Limitations on achievable performance
L9–L11 Controller optimization: analytic approach
L12–L14 Controller optimization: numerical approach
L15 Course review

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 L2: Stability and robustness

1 Stability

2 Sensitivity and robustness

3 The Small Gain Theorem

4 Singular values

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 L2: Stability and robustness

1 Stability

2 Sensitivity and robustness

3 The Small Gain Theorem

4 Singular values

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Stability is crucial

Examples:

bicycle
JAS 39 Gripen
Mercedes A-class
ABS brakes

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Input–output stability

u y = S(u)
S

A general system S is called input–output stable (or “ L2 stable”

or “BIBO stable” or just “stable”) if its L2 gain is finite:

kS(u)k
kSk = sup <∞
u kuk

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Input–output stability of LTI systems

For an LTI system S with impulse response g(t) and transfer

function G(s), the following stability conditions are equivalent:

kSk is bounded
g(t) decays exponentially
All poles of G(s) are in the left half-plane (LHP), i.e., all poles
have negative real part

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Internal stability

The LTI system

dx
= Ax + Bu
dt
y = C x + Du

is called internally stable if the following equivalent conditions

hold:

The state x decays exponentially when u = 0

All eigenvalues of A are in the LHP

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Internal vs input–output stability

If
xÛ = Ax + Bu
y = C x + Du
is internally stable then

G(s) = C(sI − A)−1 B + D

is input–output stable.

Warning
The opposite is not always true! There may be unstable pole-zero
cancellations (that also render the system uncontrollable and/or
unobservable), and these may not be seen in the transfer function!

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Stability of feedback loops

Assume scalar open-loop system G0 (s)

✲ Σ♥ ✲ G0 (s) ✲
✻

−1 ✛

The closed-loop system is stable if and only if all solutions to the

characteristic equation

1 + G0 (s) = 0

are in the left half-plane.

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Simplified Nyquist criterion

If G0 (s) is stable, then the closed-loop system [1 + G0 (s)]−1 is stable

if and only if the Nyquist curve of G 0 (s) does not encircle −1.

1
Imaginary Axis

−1

−1 −0.5 0 0.5

Real Axis

(Note: Matlab gives a Nyquist plot for both positive and negative frequencies)

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 General Nyquist criterion

Let

P = number of unstable (RHP) poles in G0 (s)

N = number of clockwise encirclements of −1 by the Nyquist
plot of G0 (s)

Then the closed-loop system [1 + G0 (s)]−1 has P + N unstable poles

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 L2: Stability and robustness

1 Stability

2 Sensitivity and robustness

3 The Small Gain Theorem

4 Singular values

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Sensitivity and robustness

How sensitive is the closed-loop system to model errors and

disturbances?

How do we measure the “distance to instability”?

Is it possible to guarantee stability for all systems within some

distance from the ideal model?

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Amplitude and phase margins

Amplitude margin Am :
arg G0 (iω0 ) = −180◦, |G0 (iω0 )| = 1/Am
Phase margin ϕm :
|G0 (iωc )| = 1, arg G0 (iωc ) = ϕm − 180◦
1
10
1

0.8 0
ωc
10

Gain
0.6 1/A m 1
Am −1
0.4 10

0.2
−2

0
G0 (ω o ) 10
−1
10 10
0

ϕm
−0.2
−50

−0.4
−100
Phase

−0.6

−0.80 (ω c )
G −150
ϕm
−1 −200 ωo
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
−250
−1 0
10 10

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Mini-problem

✲ ❡ ✲ k(s + 1) ✲ e−sL ✲
✻ s2 + cs + 1

−1 ✛

Nominally k = 1, c = 1 and L = 0. How much margin is there in each

parameter before the closed-loop system becomes unstable?
Gm = Inf , Pm = 109 deg (at 1.41 rad/s)

Magnitude (abs)
10 0
0

−0.5
10 -1
Im

0
Phase (deg)

−1
-30

-60
−1.5
0 0.5 1 1.5
Re -90
10 -1 10 0 10 1
Frequency (rad/s)
Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control
 Mini-problem

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Sensitivity functions

r y
✲ ❡ ✲ C(s) ✲ P(s) ✲
✻

−1 ✛

1
S(s) = sensitivity function
1 + P(s)C(s)
P(s)C(s)
T(s) = complementary sensitivity function
1 + P(s)C(s)

Note that we always have

S(s) + T(s) = 1

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Sensitivity towards changes in plant

r y
✲ ❡ ✲ C(s) ✲ P(s) ✲
✻

−1 ✛

How sensitive is the closed loop to a (small) change in P ?

dT C T
= 2
=
dP (1 + PC) P(1 + PC)
Relative change in T compared to relative change in P :

dT/T 1
= =S
dP/P 1 + PC

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Sensitivity towards disturbances

d v
r e u y
+ C + P +

n
−1 +

Open-loop response (C = 0) to process disturbances d , v :

Yol = V + PD

Closed-loop response:

1 P
Ycl = V+ D = S Yol
1 + PC 1 + PC

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Interpretation as stability margin

The maximum gain of the sensitivity function measures the

inverse of the distance between the Nyquist plot and the point −1:

1
R−1 = sup = Ms
ω 1 + P(iω)C(iω)

Im

−1

Re
R

P(iω)C(iω)

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 L2: Stability and robustness

1 Stability

2 Sensitivity and robustness

3 The Small Gain Theorem

4 Singular values

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Robustness analysis

How large plant uncertainty ∆ can be tolerated without risking

instability?

Example (multiplicative uncertainty):

v w
✲ ∆(s)

✲❢ ✲ ❄
✲❢
P(s)
✻

−C(s) ✛

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 The Small Gain Theorem

r1 e1
S1

e2 r2
S2

Assume that S1 and S2 are stable. If kS1 k · kS2 k < 1, then the
closed-loop system (from (r1, r2 ) to (e1, e2 )) is stable.

Note 1: The theorem applies also to nonlinear, time-varying, and

multivariable systems

Note 2: The stability condition is sufficient but not necessary, so the

results may be conservative

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 The Small Gain Theorem

r1 e1
S1

e2 r2
S2

Assume that S1 and S2 are stable. If kS1 k · kS2 k < 1, then the
closed-loop system (from (r1, r2 ) to (e1, e2 )) is stable.

Note 1: The theorem applies also to nonlinear, time-varying, and

multivariable systems

Note 2: The stability condition is sufficient but not necessary, so the

results may be conservative

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Proof

e1 = r1 + S2 (r2 + S1 (e1 ))
 
ke1 k ≤ kr1 k + kS2 k kr2 k + kS1 k · ke1 k
kr1 k + kS2 k · kr2 k
ke1 k ≤
1 − kS1 k · kS2 k
This shows bounded gain from (r1, r2 ) to e1 .

The gain to e2 is bounded in the same way.

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Application to robustness analysis

v w
✲ ∆(s)

✲❢ ✲ ❄
✲❢
P(s)
✻

−C(s) ✛

The diagram can be redrawn as

v w
✲❞ ✲ ∆ ✲
✻

−PC ✛
1+PC

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Application to robustness analysis

v w
✲❡ ✲ ∆ ✲
✻

−PC ✛
1+PC

PC
Assuming that T = 1+PC is stable, the Small Gain Theorem
guarantees stability if

k∆k · kT k < 1

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 L2: Stability and robustness

1 Stability

2 Sensitivity and robustness

3 The Small Gain Theorem

4 Singular values

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Gain of multivariable systems

Recall from Lecture 1 that

kSk = sup |G(iω)| = kGk∞

ω

for a stable LTI system S .

How to calculate |G(iω)| for a multivariable system?

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Vector norm and matrix gain

For a vector x ∈ Cn , we use the 2-norm

√ p
|x| = x∗ x = |x1 | 2 + · · · + |xn | 2

( A∗ denotes the conjugate transpose of A)

For a matrix A ∈ Cn×m , we use the L2 -induced norm

r q
| Ax| x ∗ A∗ Ax
k Ak := sup = sup = λ(A∗ A)
x |x| x x∗ x

λ(A∗ A) denotes the largest eigenvalue of A∗ A. The ratio | Ax|/|x| is

maximized when x is a corresponding eigenvector.

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

     
y1 2 4 u1
Example: Different gains in different directions: =
y2 0 3 u2

Input u= [0.309 0.951]T, |u|= 1

1.5

0.5
u2

−0.5

−1

−1.5
−4 −3 −2 −1 0 1 2 3 4
u1

y=Gu = [4.42 2.85]T, |y|= 5.26

5
y2

−5
−15 −10 −5 0 5 10 15
(red):eigenvectors ; (blue): V ; (green): U A=U*S*VT

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Singular Values

For a matrix A, its singular values σi are defined as

p
σi = λi

where λi are the eigenvalues of A∗ A.

Let σ(A) denote the largest singular value and σ(A) the smallest
singular value.

For a linear map y = Ax , it holds that

|y|
σ(A) ≤ ≤ σ(A)
|x|

The singular values are typically computed using singular value

decomposition (SVD)
Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control
 Singular value decomposition (SVD)

Let A be an m × n complex matrix. It can be factored as

A = UΣV ∗
where

U is an m × m unitary complex matrix, whose columns

represent different output directions
Σ is an m × n matrix with non-negative real numbers (the
singular values) on the diagonal, representing different gains
V is an n × n unitary complex matrix, whose columns
represent different input directions

With A = G(iω), the complex directions reveal both the relative

magnitude and phase of the input/output signals with frequency ω.

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Example: Gain of multivariable system

 2 4 

 s+1 2s + 1 
G(s) =  s 3 

 2 
 s + 0.1s + 1 s+1 

>> s = zpk('s');
>> G = [2/(s+1) 4/(2*s+1); s/(s^2+0.1*s+1) 3/(s+1)];
>> sigma(G) % plot sigma values of G wrt freq
>> [gain,w] = norm(G,inf) % infinity norm = system gain
gain =
10.3698
w =
1.0000

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Singular Values
System: G
101 Frequency (rad/s): 0.01
Singular value (abs): 5.26 System: G
Frequency (rad/s): 0.997
Singular value (abs): 10.4
Singular Values (abs)

0
10

10-1

10-2
10-2 10-1 100 101 102
Frequency (rad/s)

The singular values of the transfer function matrix (prev slide). Note that
G(0) = [2 4 ; 0 3] (prev example). kGk∞ = kG(1i)k = 10.3698.

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

>> A = evalfr(G,i*1)
A =
1.0000 - 1.0000i 0.8000 - 1.6000i
10.0000 - 0.0000i 1.5000 - 1.5000i

>> [U,S,V] = svd(A)

U =
-0.1307 + 0.1082i 0.9472 - 0.2720i
-0.9855 + 0.0023i -0.1557 - 0.0675i
S =
10.3698 0
0 1.4720
V =
-0.9734 + 0.0000i -0.2292 + 0.0000i
-0.1697 - 0.1541i 0.7206 + 0.6544i

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control

 Summary of Lecture 2

Input–output stability: kSk < ∞

1
Sensitivity function: S(s) := 1+P(s)C(s)
Three different interpretations
Small Gain Theorem: The feedback interconnection of S1 and
S2 is stable if kS1 k · kS2 k < 1
Conservative compared to the Nyquist criterion
Useful for robustness analysis
The gain of a multivariable system G(s) is given by
supω σ(G(iω)), where σ is the largest singular value
Singular values by SVD (on computer): G(iω) = UΣV ∗

Automatic Control LTH, 2019 Lecture 2 FRTN10 Multivariable Control