EE363 Winter 2008-09

Lecture 16
Analysis of systems with sector nonlinearities
Sector nonlinearities
Lure system
Analysis via quadratic Lyapunov functions
Extension to multiple nonlinearities
161
Sector nonlinearities
a function : R R is said to be in sector [l, u] if for all q R, p = (q)
lies between lq and uq
q
p
p = (q)
p = lq
p = uq
can be expressed as quadratic inequality
(p uq)(p lq) 0 for all q, p = (q)
Analysis of systems with sector nonlinearities 162
examples:
sector [1, 1] means |(q)| |q|
sector [0, ] means (q) and q always have same sign
(graph in rst & third quadrants)
some equivalent statements:
is in sector [l, u] i for all q,

(q)
u + l
2
q

u l
2
|q|
is in sector [l, u] i for each q there is (q) [l, u] with (q) = (q)q
Analysis of systems with sector nonlinearities 163
Nonlinear feedback representation
linear dynamical system with nonlinear feedback
p q
(t, )
x = Ax + Bp
q = Cx
closed-loop system: x = Ax + B(t, Cx)
a common representation that separates linear and nonlinear
time-varying parts
often p, q are scalar signals
Analysis of systems with sector nonlinearities 164
Lure system
a (single nonlinearity) Lure system has the form
x = Ax + Bp, q = Cx, p = (t, q)
where (t, ) : R R is in sector [l, u] for each t
here A, B, C, l, and u are given; is otherwise not specied
a common method for describing time-varying nonlinearity and/or
uncertainty
goal is to prove stability, or derive a bound, using only the sector
information about
if we succeed, the result is strong, since it applies to a large family of
nonlinear time-varying systems
Analysis of systems with sector nonlinearities 165
Stability analysis via quadratic Lyapunov functions
lets try to establish global asymptotic stability of Lure system, using
quadratic Lyapunov function V (z) = z
T
Pz
well require P > 0 and

V (z) V (z), where > 0 is given
second condition is:

V (z) + V (z) = 2z
T
P (Az + B(t, Cz)) + z
T
Pz 0
for all z and all sector [l, u] functions (t, )
same as:
2z
T
P (Az + Bp) + z
T
Pz 0
for all z, and all p satisfying (p uq)(p lq) 0, where q = Cz
Analysis of systems with sector nonlinearities 166
we can express this last condition as a quadratic inequality in (z, p):

z
p

T

C
T
C C
T
C 1

z
p

0
where = lu, = (l + u)/2
so

V + V 0 is equivalent to:

z
p

T

A
T
P + PA + P PB
B
T
P 0

z
p

0
whenever

z
p

T

C
T
C C
T
C 1

z
p

0
Analysis of systems with sector nonlinearities 167
by (lossless) S-procedure this is equivalent to: there is a 0 with

A
T
P + PA + P PB
B
T
P 0

C
T
C C
T
C 1

or

A
T
P + PA + P C
T
C PB + C
T
B
T
P + C

0
an LMI in P and (2, 2 block automatically gives 0)
by homogeneity, we can replace condition P > 0 with P I
our nal LMI is

A
T
P + PA + P C
T
C PB + C
T
B
T
P + C

0, P I
with variables P and
Analysis of systems with sector nonlinearities 168
hence, can eciently determine if there exists a quadratic Lyapunov
function that proves stability of Lure system
this LMI can also be solved via an ARE-like equation, or by a graphical
method that has been known since the 1960s
this method is more sophisticated and powerful than the 1895 approach:
replace nonlinearity with (t, q) = q
choose Q > 0 (e.g., Q = I) and solve Lyapunov equation
(A + BC)
T
P + P(A + BC) + Q = 0
for P
hope P works for nonlinear system
Analysis of systems with sector nonlinearities 169
Multiple nonlinearities
we consider system
x = Ax + Bp, q = Cx, p
i
=
i
(t, q
i
), i = 1, . . . , m
where
i
(t, ) : R R is sector [l
i
, u
i
] for each t
we seek V (z) = z
T
Pz, with P > 0, so that

V + V 0
last condition equivalent to:

z
p

T

A
T
P + PA + P PB
B
T
P 0

z
p

0
whenever
(p
i
u
i
q
i
)(p
i
l
i
q
i
) 0, i = 1, . . . , m
Analysis of systems with sector nonlinearities 1610
we can express this last condition as

z
p

T

c
i
c
T
i

i
c
i
e
T
i

i
e
i
c
T
i
e
i
e
T
i

z
p

0, i = 1, . . . , m
where c
T
i
is the ith row of C, e
i
is the ith unit vector,
i
= l
i
u
i
, and

i
= (l
i
+ u
i
)/2
now we use (lossy) S-procedure to get a sucient condition: there exists

1
, . . . ,
m
0 such that

A
T
P + PA + P

m
i=1

i
c
i
c
T
i
PB +

m
i=1

i
c
i
e
T
i
B
T
P +

m
i=1

i
e
i
c
T
i

m
i=1

i
e
i
e
T
i

0
Analysis of systems with sector nonlinearities 1611
we can write this as:

A
T
P + PA + P C
T
DFC PB + C
T
DG
B
T
P + DGC D

0
where
D = diag(
1
, . . . ,
m
), F = diag(
1
, . . . ,
m
), G = diag(
1
, . . . ,
m
)
this is an LMI in variables P and D
2, 2 block automatically gives us
i
0
by homogeneity, we can add P I to ensure P > 0
solving these LMIs allows us to (sometimes) nd quadratic Lyapunov
functions for Lure system with multiple nonlinearities
(which was impossible until recently)
Analysis of systems with sector nonlinearities 1612
Example
x
1
x
2
x
3
1/s 1/s 1/s
1
()
2
()
3
()
2
we consider system
x
2
=
1
(t, x
1
), x
3
=
2
(t, x
2
), x
1
=
3
(t, 2(x
1
+ x
2
+ x
3
))
where
1
(t, ),
2
(t, ),
3
(t, ) are sector [1 , 1 + ]
gives the percentage nonlinearity
for = 0, we have (stable) linear system x =

2 2 2
1 0 0
0 1 0

x
Analysis of systems with sector nonlinearities 1613
lets put system in Lure form:
x = Ax + Bp, q = Cx, p
i
=
i
(q
i
)
where
A = 0, B =

0 0 1
1 0 0
0 1 0

, C =

1 0 0
0 1 0
2 2 2

the sector limits are l

i
= 1 , u
i
= 1 +
dene = l
i
u
i
= 1
2
, and note that (l
i
+ u
i
)/2 = 1
Analysis of systems with sector nonlinearities 1614
we take x(0) = (1, 0, 0), and seek to bound J =

0
x(t)
2
dt
(for = 0 we can calculate J exactly by solving a Lyapunov equation)
well use quadratic Lyapunov function V (z) = z
T
Pz, with P 0
Lyapunov conditions for bounding J: if

V (z) z
T
z whenever the sector
conditions are satised, then J x(0)
T
Px(0) = P
11
use S-procedure as above to get sucient condition:

A
T
P + PA + I C
T
DC PB + C
T
D
B
T
P + DC D

0
which is an LMI in variables P and D = diag(
1
,
2
,
3
)
note that LMI gives
i
0 automatically
Analysis of systems with sector nonlinearities 1615
to get best bound on J for given , we solve SDP
minimize P
11
subject to

A
T
P + PA + I C
T
DC PB + C
T
D
B
T
P + DC D

0
P 0
with variables P and D (which is diagonal)
optimal value gives best bound on J that can be obtained from a
quadratic Lyapunov function, using S-procedure
Analysis of systems with sector nonlinearities 1616
Upper bound on J
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
10
0
10
1

P
1
1
(
u
p
p
e
r
b
o
u
n
d
o
n
J
)
bound is tight for = 0; for 0.15, LMI is infeasible
Analysis of systems with sector nonlinearities 1617
Approximate worst-case simulation
heuristic method for nding bad
i
s, i.e., ones that lead to large J
nd V from worst-case analysis as above
at time t, choose p
i
s to maximize

V (x(t)) subject to sector constraints
|p
i
q
i
| |q
i
|
using

V (x(t)) = 2x
T
P(Ax + Bp), we get
p = q + diag(sign(B
T
Px))|q|
simulate
x = Ax + Bp, p = q + diag(sign(B
T
Px))|q|
starting from x(0) = (1, 0, 0)
Analysis of systems with sector nonlinearities 1618
Approximate worst-case simulation
AWC simulation with = 0.05: J
awc
= 1.49; J
ub
= 1.65
for comparison, linear case ( = 0): J
lin
= 1.00
0 5 10 15 20 25 30 35 40 45 50
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
t
x
2
(
t
)
Analysis of systems with sector nonlinearities 1619
Upper and lower bounds on worst-case J
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
10
0
10
1

b
o
u
n
d
s
o
n
J
lower curve gives J obtained from approximate worst-case simulation
Analysis of systems with sector nonlinearities 1620