LMI Methods in Optimal and Robust Control

Matthew M. Peet
Arizona State University

Lecture 4: LMIs for State-Space Internal Stability

Solving the Equations
Find the output given the input

State-Space:
ẋ = Ax(t) + Bu(t) Input Output
u State-Space y
y(t) = Cx(t) + Du(t) x(0) = 0 System

Basic Question: Given an input function, u(t), what is the output?

Definition 1.
Define the Matrix Exponential:
1 1 1
e A = I + A + A2 + A3 + · · · + Ak + · · ·
2 6 k!

M. Peet Lecture 4: State-Space Theory 2 / 23

Properties of the Matrix Exponential
The matrix exponential is similar to the scalar exponential with important
differences.
• e0 = I
1 1
e0 = I + 0 + 02 + 03 + · · · = I
2 6
M T
M

T
• e = e
T 1 1 1
eM = I + M T + (M T )2 + (M T )3 + · · · + (M T )k + · · ·
2 6 k!

d At
• dt e = AeAt
 
d At d 1 2 1 k
e = I + (At) + (At) + · · · + (At) + · · ·
dt dt 2 k!
 
1 2 1 k−1
= A I + (At) + (At) + · · · + (At) + ···
2 (k − 1)!

• However, eM +N 6= eM eN unless, M N = N M .
M. Peet Lecture 4: State-Space Theory 3 / 23
State-Space Solutions and the Jordan Decomposition
The equation ẋ(t) = Ax(t), x(0) = x0
has solution
x(t) = eAt x0
Proof.
Let x(t) = eAt x0 , then
• ẋ(t) = AeAt x0 = Ax(t).
• x(0) = e0 x0 = x0

Question: So what does eAt look like? Let λi be eigenvalues of A

Definition 2.
A Jordan Block, Ji is a matrix which looks like
 
  0 1
λi 
 0 1 

Ji = λi I + N = 
 .. +
  .. .. 
.  
 . . 

λi  0 1
| {z } 0
λi I | {z }
N
M. Peet Lecture 4: State-Space Theory 4 / 23
State-Space Solutions and the Jordan Decomposition
Theorem 3.
For and A ∈ Rn×n , there exists an invertible T such that
 
J1
A = T JT −1 , J = 
 .. 
where Ji are Jordan Blocks
. 
Jn
• Ak = T J k T −1 . Hence  Jt 
e 1
eAt = T eJt T −1 =T
 ..  −1
T
.
eJn t
• λi I and N commute, hence eλi I+N = eλi I eN . Further N d = 0 for some d.
 λt 
e i  
Ji t .. 1 2 1 k−1 k−1
e =  1 + N + N + ··· + N t
 
. 2 (k − 1)!
eλi t

• limt→∞ ti eλt = 0 if and only if Re λ < 0

M. Peet Lecture 4: State-Space Theory 5 / 23
Stability of Continuous and Discrete-Time Systems

Definition 4.
A is Hurwitz if Re λi (A) < 0 for all i.

Theorem 5.
ẋ(t) = Ax(t) is stable if and only if A is Hurwitz.

For Discrete-Time Systems: xk+1 = Axk ,

x k = Ak x 0

Definition 6.
A is Schur if |λi (A)| < 1 for all i.

Theorem 7.
xk+1 = Axk is stable if and only if A is Schur.
M. Peet Lecture 4: State-Space Theory 6 / 23
Lyapunov Functions

ẋ(t) = f (x(t))
Theorem 8 (Lyapunov).
V is a Lyapunov Function if V (0) = 0 and V (x) > 0 for x 6= 0 and
limkxk→∞ V (x) = ∞. If

d
V (x(t)) < 0 for ẋ(t) = f (x(t)).
dt
Then for any x(0) ∈ R the system ẋ(t) = f (x(t)) has a unique solution which is
stable in the sense of Lyapunov.
M. Peet Lecture 4: State-Space Theory 7 / 23
Lyapunov Functions: Local Stability

ẋ(t) = f (x(t))

Theorem 9 (Lyapunov Stability).

Suppose there exists a continuous V and α, β, γ > 0 where

βkxk2 ≤ V (x) ≤ αkxk2

V̇ (x) = ∇V (x)T f (x) ≤ −γkxk2

for all x ∈ X. Then any sub-level set of V in X is a Domain of Attraction.

M. Peet Lecture 4: State-Space Theory 8 / 23

Lyapunov Functions

Mass-Spring: Pendulum:

c k g
ẍ = − ẋ − x ẋ2 = − sin x1 ẋ1 = x2
m m l
1 1 1
V (x) = mẋ2 + kx2 V (x) = (1 − cos x1 )gl + l2 x22
2 2 2

V̇ (x) = ẋ(−cẋ − kx) + kxẋ V̇ (x) = glx2 sin x1 − glx2 sin x1

2
= −cẋ − k ẋx + kxẋ =0

= −cẋ2 ≤ 0
M. Peet Lecture 4: State-Space Theory 9 / 23
An Example of Global Stability Analysis
A controlled model of a jet engine (Derived from
Moore-Greitzer).
3 1
ẋ = −y − x2 − x3
2 2
ẏ = 3x − y
This is feasible with
V (x) = 4.5819x2 − 1.5786xy + 1.7834y 2 − 0.12739x3 + 2.5189x2 y − 0.34069xy 2
+ 0.61188y 3 + 0.47537x4 − 0.052424x3 y + 0.44289x2 y 2 + 0.090723y 4

M. Peet Lecture 4: State-Space Theory 10 / 23

The Lyapunov Inequality (Our First LMI)
Lemma 10 (An LMI for Hurwitz Stability).
A is Hurwitz if and only if there exists a P > 0 such that

AT P + P A < 0

Proof.
Suppose there exists a P > 0 such that AT P + P A < 0.
• Define the Lyapunov function V (x) = xT P x.
• Then V (x) > 0 for x 6= 0 and V (0) = 0.
• Furthermore,
V̇ (x(t)) = ẋ(t)T P x(t) + x(t)T P ẋ(t)
= x(t)T AT P x(t) + x(t)T P Ax(t)
= x(t)T AT P + P A x(t)

• Hence V̇ (x(t)) < 0 for all x 6= 0. Thus the system is globally stable.
• Global stability implies A is Hurwitz.
M. Peet Lecture 4: State-Space Theory 11 / 23
The Lyapunov Inequality
Proof.
For the other direction, if A is Hurwitz, for any Q > 0, let
Z ∞
T
P = eA s QeAs ds
0

• Converges because A is Hurwitz.

• Furthermore
Z ∞
T
PA = eA s QeAs Ads
Z0 ∞ Z ∞
T T d As

= eA s QAeAs ds = eA s Q e ds
0 0 ds
 ∞ Z ∞
T d AT s As
= eA s QeAs − e Qe
0 0 ds
Z ∞
T
= −Q − AT eA s QeAs = −Q − AT P
0

• Thus P A + AT P = −Q < 0.

M. Peet Lecture 4: State-Space Theory 12 / 23

Discrete-Time Lyapunov Functions

xk+1 = f (xk )

Theorem 11 (Lyapunov).
V is a Lyapunov Function if V (0) = 0 and V (x) > 0 for x 6= 0 and
limkxk→∞ V (x) = ∞. If

V (xk+1 ) < V (xk ) where xk+1 = f (xk ).

Then for any x0 ∈ Rn the system xk+1 = f (xk ) is stable in the sense of
Lyapunov.

M. Peet Lecture 4: State-Space Theory 13 / 23

Discrete-Time Lyapunov Functions
Lemma 12 (An LMI for Schur Stability).
A is Schur if and only if there exists a P > 0 such that

AT P A − P < 0

Proof.
Suppose there exists a P > 0 such that AT P A − P < 0.
• Define the Lyapunov function V (x) = xT P x.
• Then V (x) > 0 for x 6= 0 and V (0) = 0.
• Furthermore,
V (xk+1 ) = xTk+1 P xk+1
= xTk AT P Axk
< xTk P xk = V (xk )

• Hence V (xk+1 ) < V (xk ) for all k ≥ 0. Thus the system is Stable.
• Stability implies A is Schur.
M. Peet Lecture 4: State-Space Theory 14 / 23
Lyapunov Functions
Proof.
For the other direction, if A is Hurwitz, for any Q > 0, let
X∞
P = (AT )k QAk
k=0

∞
X ∞
X
AT P A − P = (AT )k QAk − (AT )k QAk
k=1 k=0
= −(AT )0 QA0 = −Q < 0

• Thus AT P A − P < 0.

YALMIP Code:
> P = sdpvar(n); eta=.1;
> F=[P>=eta*eye(n)];
> F=[F; A’ P A - P<=0];
> optimize(F);
M. Peet Lecture 4: State-Space Theory 15 / 23
Pole Locations AKA D-stability
Some people still care about pole locations.
• For these people, we have D-stability.
To begin, you have to define the acceptable region of the complex plane using
inequality constraints.
1.8
• Rise Time: ωn ≤ tr
• Settling Time: σ ≤ − 4.6
ts
ln Mp
• Percent Overshoot: σ≤− π |ωd |
Recall that if z is the complex pole location:
2
• ωn = kzk2 = z ∗ z
• ωd = Im z = (z − z ∗ )/2
• σ = Re z = (z + z ∗ )/2
Which yields
• Rise Time: z ∗ z − 1.82
t2r ≤0
• Settling Time: (z + z ∗ ) + 4.6
ts ≤ 0
• Percent Overshoot: z − z ∗ + ln πM |z + z ∗ | ≤ 0
p

M. Peet Lecture 4: State-Space Theory 16 / 23

An LMI for Pole Locations
Gutman proposed a nice LMI for D-stability with a single constraint

Theorem 13 (Gutman).
The pole locations, z ∈ C of A satisfy
X
z ∈ {z ∈ C : ckl z k (z ∗ )l < 0}
k,l

if and only if there exists some P > 0 such that

X
ckl Ak P (AT )l < 0
k,l

But this has some disadvantages

• There can only be one constraint.
• The LMI is not linear in A.
I So controller synthesis is not an LMI.

M. Peet Lecture 4: State-Space Theory 17 / 23

An LMI for Convex Regions of the Complex Plane

To get around the limitations of Gutman’s result, we introduce the concept of

LMI regions.
• These are regions which can be represented using LMIs in the z and z ∗
variables

Definition 14.
An LMI Region of the complex plane has the form

{z ∈ C : F0 + zF1 + z ∗ F2 < 0}

Such regions are hard to visualize, but

• Are convex
I e.g. Minimum rise time is not allowed!
• Can combine multiple convex regions

M. Peet Lecture 4: State-Space Theory 18 / 23

An LMI for Convex Regions of the Complex Plane
Examples

1.82
Rise Time: z ∗ z − ≤0
t2r
|{z}
 r 2      
−r z −r 0 0 1 0 0 ∗
= + z+ z <0
z∗ −r 0 −r 0 0 1 0
| {z } | {z } | {z }
F0 F1 F2
Which by the Schur complement is equivalent to r − z ∗ r−1 z > 0.

Settling Time: 4.6

ts + z + z∗ ≤ 0

Percent Overshoot: z − z ∗ + π
ln Mp |(z + z ∗ )| ≤ 0

ln Mp (z + z ∗ ) π(z − z ∗ )
       
0 0 ln Mp π ln Mp −π
= + z+ z∗ < 0
π(z − z ∗ )∗ ln Mp (z + z ∗ ) 0 0 −π ln Mp π ln Mp
| {z } | {z } | {z }
F0 F1 F2
Which by the Schur complement is equivalent to z − z ∗ < 0 and
(z − z ∗ )2 > ( ln πMp )2 |z + z ∗ |2 ≤ 0.
M. Peet Lecture 4: State-Space Theory 19 / 23
An LMI for Convex Regions of the Complex Plane

Theorem 15 (Chilali + Gahinet).

The pole locations, z ∈ C of A satisfy

z ∈ {z ∈ C : F0 + zF1 + z ∗ F2 < 0}

if and only if there exists some P > 0 such that

F0 ⊗ P + F1 ⊗ (AP ) + F2 ⊗ (AP )T < 0

The notation F ⊗ P is Kronecker notation and means for each element of F z,

replace the scalar z with the matrix P . So, e.g.
   
f11 f12 f P f12 P
⊗ P := 11
f12 f22 f12 P f22 P

M. Peet Lecture 4: State-Space Theory 20 / 23

An LMI for Sector Regions of the Complex Plane
Rise Time:        
−r z −r 0 0 1 0 0 ∗
= + z+ z <0
z∗ −r 0 −r 0 0 1 0
| {z } | {z } | {z }
F0 F1 F2
becomes
Lemma 16.
The pole locations, z ∈ C of A satisfy z ∗ z ≤ r2 if and only if there exists some
P > 0 such that  
−rP AP
<0
(AP )T −rP

Settling Time: 4.6

ts + z + z∗ ≤ 0
becomes
Lemma 17.
The pole locations, z ∈ C of A satisfy 2 Re x ≤ −α if and only if there exists
some P > 0 such that
AP + (AP )T + αP < 0

M. Peet Lecture 4: State-Space Theory 21 / 23

An LMI for Sector Regions of the Complex Plane

ln Mp
Percent Overshoot: z + z ∗ ≤ − π |z − z∗|

Lemma 18.
ln M
The pole locations, z ∈ C of A satisfyz + z ∗ ≤ − π p |z − z ∗ | if and only if
there exists some P > 0 such that
ln Mp (AP + (AP )T ) π(AP − (AP )T )
 
<0
π(AP − (AP )T )T ln Mp (AP + (AP )T )

M. Peet Lecture 4: State-Space Theory 22 / 23

Discrete-Time Lyapunov Functions

Next Time: LMIs for Controllability and Observability

M. Peet Lecture 4: State-Space Theory 23 / 23