An IQC-Based Stability Criterion for Systems with Slowly Varying Parameters
Anders Helmerssonn
Department of Electrical Engineering Linkoping University, S-581 83 Linkoping, Sweden
www:
http://www.control.isy.liu.seemail:
andersh@isy.liu.seLiTH-ISY-R-1979 September 10, 1997
REGLERTEKNIK
AUTOMATIC CONTROL
LINKÖPING
Technical reports from the Automatic Control group in Linkoping are available as UNIX-compressed Postscript les by anonymous ftp at the address
130.236.20.24(ftp.control.isy.liu.se)
.
An IQC-Based Stability Criterion for Systems with Slowly Varying Parameters
Anders Helmersson
Department of Electrical Engineering Linkoping University, S-581 83 Linkoping, Sweden
www:
http://www.control.isy.liu .seemail:
andersh@isy.liu.seSubmitted to ACC 1998 September 10, 1997
Abstract
An integral quadratic constraints (IQC) is introduced for stability analysis of linear systems with slowly varying parameters. The param- eters are assumed to be bounded and with bounded derivatives. Other types of uncertainties can be included in the problem. The new criterion yields signicantly less conservative bounds than previously proposed cri- teria.
Keywords:
Stability analysis, robustness, linear time-varying, inte- gral quadratic constraints.
1 Introduction
Integral Quadratic Constraints (IQC) has recently been proposed by Megretski and Rantzer as a unied approach to robustness analysis 5].
We will here elaborate on IQCs for verifying stability of linear systems with slowly varying parameters. We assume that the parameters are bounded as well as are their derivatives. An IQC-based stability criterion for slowly-varying parameters (uncertainties) has been proposed by Jonsson and Rantzer 3, 2].
The IQC proposed in this paper exploits more of the structure of the problem than the previous criterion. Since higher exibility in the IQC multiplier can be allowed the new criterion gives less conservative bounds. This is illustrated by two examples.
The paper is organized as follows. Integral Quadratic Constraints are briey introduced in section 2. In section 3, the main result is elaborated including the swapping lemma and choice of multipliers. The new algorithm is applied to two examples in section 4. The conclusions are given in section 5.
This work was supported by the Swedish National Board for Industrial and Technical
Development (NUTEK), which is gratefully acknowledged
G
(
s)
6
-
+
i -
?
+
i fv
e w
Figure 1: Basic feedback conguration.
1.1 Notation
For matrices
Adenotes the complex conjugate transpose.
RH
1
denotes denotes stable real-rational transfer functions
L2denotes the Hilbert space of measurable functions
R!Rnsatisfying
kfk22
=
Z 1;1
jf
(
t)
j22dt<1This is a subspace of
L2e, whose members only need to be square integrable on
nite intervals.
2 Integral Quadratic Constraints
The integral quadratic constraints (IQCs) has been proposed for robustness analysis 5]. The IQC states a stability criterion for the interconnection of a stable system
M2RH1and a bounded causal operator , see gure 1.
(
v
=
Mw+
fw
=
v+
e:(1)
We say that the interconnection of
Mand is well-posed if the map (
vw)
!(
ef) dened by (1) has a causal inverse on
L2e. The interconnection is stable if, in addition, the inverse is bounded, that is if there exists a constant
Csuch that
Z
T
0
(
jv(
t)
j2+
jw(
t)
j2)
dtCZ
T
0
(
jf(
t)
j2+
je(
t)
j2)
dtfor any
T0 and for any solution of (1).
Depending on the particular application, various versions of IQCs are avail- able. Two signals
w2 L20
1) and
v 2 L20
1) are said to satisfy the IQC dened by , if
Z
1
1
v
^ (
j!)
^
w
(
j!)
(
j!)
v
^ (
j!)
^
w
(
j!)
d!
0 (2)
where absolute integrability is assumed. Here ^
v(
j!) and ^
v(
j!) represent the harmonic spectrum of the signals
vand
wat the frequency
!. In principle
:
jR !Ccan be any measurable Hermitian-valued function. In most appli-
cations, however, it is sucient to use rational functions that are bounded on
the imaginary axis.
A time-domain form of (2) is
Z
1
0
(
x(
t)
v(
t)
w(
t))
dt0 (3) where
is a quadratic form, and
xis dened by
_
x
(
t) =
Ax(
t) +
Bvv(
t) +
Bww(
t)
x(0) = 0 where
Ais a Hurwitz matrix.
The main theorem from 5] goes as follows
Theorem 1 (5]) Let
G2RH1and let be a bounded causal operator. As- sume that:
i) for every
20
1], the interconnection of
Gand
is well-posed
ii) for every
20
1], the IQC dened by is satised by
iii) there exists
>0 such that
G
(
j!)
I
(
j!)
G
(
j!)
I
;I 8!2R:
(4) Then the feedback interconnection of
Gand is stable.
Note that if the upper left corner,
11, of is positive semidenite then = 0 satises (2). If further the lower right corner,
22, is negative semidenite then any convex combination of 's satisfying (2) also satises the IQC. Thus,
110 and
220 imply that
satises (2) for
20
1] if and only if does so. This simplies assumption ii).
This follows by the fact that
I
1+ (1
;)
211 12 12 22
I
1+ (1
;)
2
=
I1
11 12 12 22
I1
+ (1
;)
I2
11 12 12 22
I2
;
(1
;)
0
1;2
0 0 0
221;
0
2
I1
11 12 12 22
I1
+ (1
;)
I2
11 12 12 22
I2
if
20
1] and
220. Also,
110 assures that the inequality is satised for = 0.
The search for multipliers, , can be carried out as a convex optimization problem by parametrizing
(
j!) =
Xi x
i
i(
j!)
where
xiare positive real parameters and
iis a set of basis multipliers. By
applying the Kalman-Yakubovich-Popov lemma 7, 8, 6], the search for
xi, can
be implemented using linear matrix inequalities (LMIs).
3 Slowly Varying Uncertainties
We will here propose an IQC for verifying stability of linear system with slowly varying uncertainties. We dene slowly varying uncertainties such that their rate of change _ is norm-bounded by
d, that is
k_
kd.
The main tool for this analysis is the swapping lemma that tells how a linear operator on state-space form commutes with a time-varying uncertainty.
3.1 Swapping Lemma
The following swapping lemma is essentially from 2].
Lemma 1 (Swapping Lemma) Suppose that ^ has the derivative _^ and
u2L2
. Further let
T 2RH1:
T
(
s) = ^
D+ ^
C(
sI;A^ )
;1B^
U
(
s) = ^
C(
sI;A^ )
;1V
(
s) = (
sI;A^ )
;1B^ such that
A
^
B^
^
C D
^
^ 0 0
=
^ 0 0
^
A B
^
^
C D
^
:
(5)
Then
T
(
s) =
T(
s)
u+
U(
s) _^
V(
s)
u:(6)
Proof: Let
s=
dtd. We note that
(
sI;A^ )^
V(
s)
u= _^
V(
s)
u+ ^
Bu= _^
V(
s)
u+
Bu:(7) Let
U(
s) operate from the left on (7). After addition of ^
Du= ^
Du, we get
^
D+ ^
C^
V(
s)
u=
U(
s) _^
V(
s) + ^
D+
U(
s)
Buwhich is equivalent to (6).
2Note that we can generalize the swapping lemma to also apply to a more general set of linear operators ^ for which we dene _^ by
d
dt
(^
v) = _^
v+ ^_
vfor all
vv_
2L2.
Also note that the block-diagonal structure of
T,
Vand
Uis implicitly dened by the commuting equation (5).
The set of uncertainties, , can be assumed to have a blockdiagonal struc- ture, possibly with repeated sub-blocks:
=
f= diag
In1 1:::Inf f]
gwhere
denotes the Kronecker product. The uncertainties consist of
fdiagonal blocks, which could be either dynamic (complex) or parametric (real). For parametric blocks also
i=
iapplies. It is easy to show that for any
i 2C
kiki
and
Di2Cninithe following commutative equation holds:
(
Ini i)(
Di Iki) = (
Di Iki)(
Ini i) (8) In the paper the structure of is implicit, and instead the structure is dened by commuting properties, such as (5) and (8).
For instance, if we assume that = diag
1In12] where
1is a slowly varying parameter and
2contains remaining uncertainties, a natural choice of
T
would be to use
A
^
B^
^
C D
^
=
2
6
6
4
^
A B
^ 0
I
0 0 0
I0 0 0
I3
7
7
5
:
(9)
Then ^ =
1Ik1, = diag
1Ik1] = diag
1In1+k12], where
k1is the dimension of ^
A. In general ^ contains the slowly varying parameters in possibly (di!erently) repeated.
3.2 IQC Formulation
We will study a stability criterion for the system
x=
Gxwhere and _^ are norm-bounded. Consider the following IQC-like matrix inequality.
2
6
6
4
T
(
j!)
G(
j!) 0
V
(
j!)
G(
j!) 0
T
(
j!)
U(
j!)
0
I3
7
7
5
~
2
6
6
4
T
(
j!)
G(
j!) 0
V
(
j!)
G(
j!) 0
T
(
j!)
U(
j!)
0
I3
7
7
5
<
0 (10) where ~ = ~ is a static matrix. We can rewrite (10) as a proper IQC for the augmented system
G0
:
2
4
G
(
j!) 0
I
0
0
I3
5
(
j!)
2
4
G
(
j!) 0
I
0
0
I3
5
<
0
!2Rf1g(11) where
=
11 12 12 22
=
2
6
6
4
T
0 0
V
0 0 0
T U0 0
I3
7
7
5
~
2
6
6
4
T
0 0
V
0 0 0
T U0 0
I3
7
7
5
:
(12)
The multiplier dened by (12) satises the IQC for the augmented uncertainty given by
"
_^
V#
. To show this we rst observe that the augmented uncertainty
is bounded since
T 2 RH1and, consequently, _^
Vis bounded if _^ is such.
Using the swapping lemma, we obtain
w
=
2
6
6
4
T
0 0
V
0 0 0
T U0 0
I3
7
7
5 2
4
I_^
V3
5
v
=
2
6
6
6
4 T
V
T
+
U_^
V_^
V3
7
7
7
5 v
=
2
6
6
4 T
VT
_^
V3
7
7
5 v:
Thus (3) becomes
Z
1
0 w
(
t) ~
w(
t)
dt=
Z
1
0
w
T
(
t)
w
V
(
t)
2
6
6
6
4
I
0
0
I(
t) 0 0 _^(
t)
3
7
7
7
5
~
2
6
6
6
4
I
0
0
I(
t) 0 0 _^(
t)
3
7
7
7
5
w
T
(
t)
w
V
(
t)
dt
0 (13) where
wT=
Tvand
wV=
Vvare dened by
_^
x
(
t) = ^
Ax^ (
t) + ^
Bv(
t)
x^ (0) = 0 (14)
w
T
(
t)
w
V
(
t)
=
C
^
D^
I
0
^
x
(
t)
v
(
t)
:
Using (5) we have
"
(
t) 0 0 _^(
t)
#
C
^
D^
I
0
=
C
^
D^ 0 0 0
I2
6
4
^(
t) 0 0 (
t) _^(
t) 0
3
7
5 :
Consequently, (13) is equivalent to
Z
1
0
x
^ (
t)
v
(
t)
2
6
6
6
6
6
4
I
0
0
I^(
t) 0 0 (
t) _^(
t) 0
3
7
7
7
7
7
5
~~
2
6
6
6
6
6
4
I
0
0
I^(
t) 0 0 (
t) _^(
t) 0
3
7
7
7
7
7
5
x
^ (
t)
v
(
t)
dt
0 (15) for any
v2L2and ^
xas dened by (14) and where
~~ =
2
4
~~
11~~
12~~
12~~
223
5
=
2
6
6
4
^
C D
^ 0 0 0
I
0 0 0 0 0 0 ^
C D^ 0 0 0 0 0
I3
7
7
5
~
11~
12~
12~
222
6
6
4
^
C D
^ 0 0 0
I
0 0 0 0 0 0 ^
C D^ 0 0 0 0 0
I3
7
7
5 :
(16)
If ~~
110 and ~~
220, then the inequality (15) is satised for any time-varying
convex combinations of ^'s that satisfy the same inequality. See the remark on
theorem 1.
We will now show that there exists such a multiplier, ~~, if
11(
j!)
0 and
22(
j!)
0 for all
! 2 R. We assume that ^
Ais Hurwitz and that the state-space representation of
Tis minimal so that ( ^
AB^ ) is controllable.
The positive semideniteness of
11(
j!) is equivalent to
T
(
j!)
V
(
j!)
~
11
T
(
j!)
V
(
j!)
=
(
j!I;A^ )
;1B^
I
~~
11(
j!I;A^ )
;1B^
I
0 for all
! 2 R. Using the Kalman-Yakubovich-Popov lemma 7, 8, 6], this in- equality holds if and only if there exists
Y=
Ysuch that
~~
11+
YA
^ + ^
A Y YB^
^
B Y
0
0
:Thus, if
11(
j!)
0 for all
! 2 Rwe can modify ~~
11to become positive semidenite by adding
~~
11Y=
YA
^ + ^
A Y YB^
^
B Y
0
:
This does not a!ect neither (11) nor (15), since
(
j!I;A^ )
;1B^
I
~~
11Y(
j!I;A^ )
;1B^
I
= 0
:Using the same sequence of arguments
22(
j!) =
T
(
j!)
U(
j!)
0
I
~
22
T
(
j!)
U(
j!)
0
I
(17)
=
2
4
(
j!;A^ )
;1B^ (
j!;A^ )
;1I
0
0
I3
5
~~
222
4
(
j!;A^ )
;1B^ (
j!;A^ )
;1I
0
0
I3
5
0 holds for all
!2R, if and only if there exists
X=
Xsuch that
~~
22+
2
4
XA
^ + ^
A X XB^
X^
B X
0 0
X
0 0
3
5
0
:(18)
Thus we can modify ~~
22to become negative semidenite if and only if
22(
j!)
0 for all
!2R. Further, we can show that also (15) is una!ected by the same modication of ~~
22, since
Z
1
0
^
x(
t)
v
(
t)
2
6
4
^(
t) 0 0 (
t) _^(
t) 0
3
7
5 2
4
XA
^ + ^
A X XB^
X^
B X
0 0
X
0 0
3
5 2
6
4
^(
t) 0 0 (
t) _^(
t) 0
3
7
5
x
^ (
t)
v
(
t)
dt
=
Z
1
0
^
x(
t) ^(
t)
X( ^
A^(
t)^
x(
t) + ^
B(
t)
v(
t) + _^(
t)^
x(
t)) + ()
dt=
Z
1
0
^
x(
t) ^(
t)
X(^(
t) ^
Ax^ (
t) + ^ ^
Bv(
t) + _^(
t)^
x(
t)) + ()
dt=
Z
1
0
d
dt
(^
x(
t) ^(
t)
X^(
t)^
x(
t))
dt= lim
t!1
^
x
(
t) ^(
t)
X^(
t)^
x(
t)
;x^ (0) ^(0)
X^(0)^
x(0) = 0
since ^
x x_^
2L2and ^
x(0) = 0. Here we have used () to denote the transpose of the previous terms.
We have shown that if
11(
j!)
0 and
22(
j!)
0 for
!2R, then we can
nd an equivalent IQC with ~~ such that ~~
110 and ~~
220. Thus (13) and (15) both hold for any time-varying convex-cone combination such that
^(
t) =
Xi
i
(
t) ^
i(
t) (19a) _^(
t) =
Xi
i
(
t) _^
i(
t) (19b) for
i(
t)
20
1] and
Pii(
t)
1. Note that (19) constrains
idynamically since ^ and _^ are related. Specically, it follows that
Pi_
i(
t) ^
i(
t) = 0. For instance, if _^
i= 0 then only time-invariant ^ are allowed.
If we have a multiplier ~ such that
2
6
6
4
I
0
0
Ii
(
t) 0 0 ^
i(
t)
3
7
7
5
~
2
6
6
4
I
0
0
Ii
(
t) 0 0 ^
i(
t)
3
7
7
5
0
8it(20) that satises (11) and if
11(
j!)
0 and
22(
j!)
0 for all
!2R, then the system is veried to be stable for any time-varying convex-cone combination of the
iand _^
i.
For instance, if =
Iand _^ = _
Isuch that
jj1 and
jj_
<d, then we may use
~ =
2
6
6
4
P1
0 ;
10
0
d2P20 ;
2;
10
;P1;"1I0
0 ;
20
;P23
7
7
5
(21)
where
P1=
P1,
P2=
P20, ;
1=
;;
1and ;
2=
;;
2. We do not require
P1to be positive semidenite, while
P2must be so in order not to violate
11(
j!)
0 and
22(
j!)
0. Since, (21) satises (20) for
=
1
;"21 +
"2] and
jj_
dwe can allow any time-varying convex-cone combination if the IQC holds. The variable
"1is an arbitrarily small positive constant that is chosen in order to allow a variation of
within small bands of width 2
"2around
1.
To summarize, we have arrived at the following theorem:
Theorem 2 Assume that
iand _^
isatisfy (13) and that the interconnection of and
G(
s) is well-posed for every in the convex cone dened by
i. Then the system
u=
G(
s)
uis asymptotically stable for any time-varying convex- cone combination of
iand _^
iif there exists a as dened by (12) such that
2
4
G
(
j!) 0
I
0
0
I3
5
(
j!)
2
4
G
(
j!) 0
I
0
0
I3
5
<
0 (22a)
11(
j!)
0 (22b)
22(
j!)
0 (22c)
hold for all
!2Rf1g.
3.3 Comparison with Previous Criterion
Jonsson and Rantzer proposed an IQC for the stability analysis of linear system with slowly varying parameters 3, 2]:
M
I
R R
+
VRVR+
d2VSVS+
USUS S;SS ;S ;R
(
I;URUR)
R
M
I
<
0
(23) where
R
=
DR+
CR(
sI;AR)
;1BR S=
DS+
CS(
sI;AS)
;1BSU
R
=
CR(
sI;AR)
;1 US=
CS(
sI;AS)
;1V
R
= (
sI;AR)
;1BR VS= (
sI;AS)
;1BSsuch that
Rand
Sare both stable. We will now show that (23) is more conser- vative than (10). Specically, we will show that it is slightly more conservative than theorem 2 even if we assume that ~
110 and ~
220. Let
~ =
2
6
6
6
6
6
6
6
6
6
6
6
6
6
6
4
I
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0
;I0 0
0 0 0 0 0 0
I0 0 0
0 0 0
I0 0 0 0 0 0
0 0 0 0
d2I0 0 0 0 0
0 0 0 0 0
;I0 0 0 0
0 0
I0 0 0 0 0 0 0
0
;I0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
;I0
0 0 0 0 0 0 0 0 0
;I3
7
7
7
7
7
7
7
7
7
7
7
7
7
7
5
and
T
=
2
4 R
S
I 3
5
U
=
2
4 U
R
0 0
US0 0
3
5
V
=
V
R
V
S
:
This yields
2
6
6
4
TM
0
VM
0
T U
0
I3
7
7
5
~
2
6
6
4
TM
0
VM
0
T U
0
I3
7
7
5
=
2
6
6
4 3
7
7
5 2
6
6
4
R R
+
VRVR+
d2VSVS S;S0
USS ;S ;R R ;R U
R
0
0
;URR ;(
I+
URUR) 0
U
S
0 0
;I3
7
7
5 2
6
6
4 M
0 0
I
0 0 0
I0 0 0
I3
7
7
5
<
0
:Using Schur complements, we rewrite it as
M
I
R R
+
VRVR+
d2VSVS+
USUS S;SS ;S ;R
(
I+
URUR)
;1R
M
I
<
0
:(24)
Note that (24) is very close to the IQC (23) proposed by Jonsson and Rantzer
3, 2] for slowly varying uncertainties, which is slightly more conservative, since
I;U
R U
R
(
I+
URUR)
;1:for all
UR. Note that main restriction in (23) is that it assumes semidenite
~
22and ~
11.
4 Examples
We will show the stability criterion on two examples. In both examples the improvements compared to previous IQC-based bounds are signicant.
4.1 Example I
This example is from 3]. A simple SISO system is considered:
_
x
(
t) = (
A+
C(
t)
B)
x(
t) =
;
0
:21
;1 + 0
:8
(
t)
1 0
x
(
t)
where it is assumed that
(
t)
2;]. In the case the uncertainty parameter is constant, the system is asymptotically stable if and only if
<1
:25, that is, when
<1
:25.
If _
is unbounded, then stability is assured if
jj <0
:261. This bound is obtained as the inverse of the
H1-norm of
G
(
s) =
D+
C(
sI;A)
;1Bwhere
A B
C D
=
2
4
;
0
:21
;1 0
:8
1 0 0
0 1 0
3
5
:
We now apply theorem 2 using
T
(
s) =
"
1
s+1
1
#
:
(25)
This multiplier has the same pole as that in 3]. There the analysis using a frequency dependent IQC yields a stability bound of
jj_
<d= 0
:1 for
= 1.
Using (25) yields
d= 0
:1036, which is in close agreement with 3].
If we instead choose
T
(
s) =
"
1
s+10
1
#
(26)
we obtain
d= 0
:1456 for
= 1, assuming ~
110 and ~
0. Relaxing the
requirements on ~ and instead assuming
11(
j!)
0 and
22(
j!)
0, we
obtain
d= 0
:3111 for
= 1 with (26). Figure 2 shows the stability bounds,
d,
as a function of
. Instability has been veried for
= 0
:41 without any bounds
0 0.2 0.4 0.6 0.8 1 1.2 0
0.1 0.2 0.3 0.4 0.5
d
stable
Figure 2: Example I: Stability bounds
das a function of
. The system is stable of
(
t)
2 ;] and
j_ (
t)
j d. The lower, dashed line uses
Tas dened in (25) while the middle, dashed-dotted line uses (26). Both assume semidenite
~
ii. The upper, solid line uses the relaxed constraints on ~
iiand the multiplier
dened by (26). The cross (
) marks the bound given in 3]. The circle (
)
at
= 1
:25, denotes the bound obtained with constant
. Arbitrarily fast
variations in
can be allowed if
jj<0
:261, marked by the vertical line. The
criteria show stability in the area below and to the left of the lines. The asterisk
(
) at
= 1 and
d= 0
:46, shows a case for which instability has been veried.
on _
and
d= 0
:46 for
= 1 these bounds are (probably) not the true bounds though.
The lower bound could probably be improved by using a basis multiplier,
T
, of high order. However, this leads to an increased computational load, as well as that the condition number of the required ~ usually increases. This may cause numerical problems, which could result in worse bounds even if this is not theoretically true.
4.2 Example II
This example is has been analyzed in 4] using IQCs. We consider the ship steering dynamics as in Example 9.6 in 1]. The dynamics can be approximated by the Nomoto model
_
x
(
t) =
v(
t)(
;ax(
t) +
bv(
t)
(
t)) _
(
t) =
x(
t)
where
denotes the heading of the ship,
denotes the rudder angle and
vis the speed of the ship. It is assumed that
v(
t)
0. We will, as in 1], study the stability of the ship dynamics for an unstable tanker with
a=
;0
:3 and
b= 0
:8, which is controlled by a PD regulator
v
=
;Kwhere
K(
s) =
k(1 +
Tds) with
k= 2
:5 and
Td= 0
:86.
The closed loop system can be described by
x
(
t) = (
v(
t)
I2)
G0(
s)
x(
t) with
G0
(
s) =
;a=s b
;K
(
s)
=s20
:
This system is not strictly stable and instead we replace it for the purpose of analysis by
G
(
s) = (
I+
G0(
s))(
I ;G0(
s))
;1=
D+
C(
sI;A)
;1Bwhere
A B
C D
=
2
6
6
4
;
(
a+
bkTd)
;bk ;2
;2
b1 0 0 0
a
+
bkTd bk1 2
bkT
d
k
0 1
3
7
7
5
and
(
t) =
v(
t)
;vnomv
(
t) +
vnom 2;1
1]
where
vnomis the nominal speed of the ship. For the transformed system,
dis solved as a function of
such that the system is veried to be stable when
jj