An Uncertainty Augmentation Approach to Varying Analysis
34th CDC
Anders Helmersson
Department of Electrical Engineering Linkoping University
S-581 83 Linkoping, Sweden tel: +46 13 281622 fax: +46 13 282622 email:
andersh@isy.liu.seFebruary 27, 1995
Abstract
We present a concept for analysis and synthesis with varying un- certainties. We assume that the uncertainties are unknown by bounded and with bounded rate of variation. The approach taken here is to aug- ment the uncertainty block with its derivative. This can be achieved using appropriate dynamic multipliers. The approach ts nicely in the framework and
D-
K-iteration-like methods can be used for synthesis.
Keywords:
Structured singular values, parametric dependent sys- tems, time-varying systems.
1 Introduction
Systematic theories and tools for analysis and design of linear systems with uncertainties have been developed during the last decade and are still being rened and extended. With the advent of the two-Riccati-equation method for solving the
H1control problem in 1988 2], the tools became more easy to use and the spread to new applications took a stride. Doyle 1] introduced the structured uncertainties and the
-formalism, where linear uncertainties can be of two types: real and complex. Real uncertainties are parametric, while complex uncertainties include linear dynamic e ects.
The
-analysis involves scaling matrices with a structure corresponding to the structure of the uncertainties. The scaling matrices are varied in order to optimize a criterion: either a matrix inequality or the maximum singular value.
The minimization of the maximum singular value can be reformulated into a
This work was supported by the Swedish National Board for Industrial and Technical
Development (NUTEK), which is gratefully acknowledged
linear matrix inequality (LMI). If the uncertainties are constants the scaling matrices can be frequency dependent. If the uncertainties are varying without any bound on the rate of change the scaling matrices are constants. Recently methods for bridging the gap between constant and fast-changing uncertainties have been presented 6, 10, 11, 12].
In this paper an approach based on uncertainty augmentation is presented.
The main idea behind this approach is to introduce an uncertainty block that in addition to the uncertainty itself contains its derivative. With appropriate left and right dynamic multipliers the original uncertainty block can be recovered.
Using these multipliers we can reformulate the original problem into a standard
one.
The paper is organized as follows. In section 2 we review the
analysis concepts and linear fractional transformations (LFTs). Section 3 presents an approach for handling time varying uncertainties when these have bounded rate of variation. An example illustrating the method is given in section 4. A discussion on uncertainty structures and their classication follows in section 5.
How to include the approach in controller synthesis is discussed in section 6.
The conclusions are given in section 7.
1.1 Notations
The notations used are fairly standard. We use
In to denote a unit matrix of dimension
n nXdenotes the complex conjugate transpose of
XX>(
) 0 a hermitian (
X=
X) positive denite (semidenite) matrix
X;= (
X)
;1ker
Xdenotes the null space of
Xand range
Xits range
X?denotes a matrix such that ker
X?= range
Xand
X?X? >0 note that
X?only exists if
Xhas linearly dependent rows and that
X?is not unique but in this paper any choice is acceptable
Xyis the Moore-Penrose pseudo inverse of
Xdiag
X1X2] a block-diagonal matrix composed of
X1and
X2rank
Xdenotes the rank of the matrix
Xherm
X=
12(
X+
X)
S(
::) denotes the Redhe er star product
AB
denotes the Kronecker product of the matrices
Aand
B(
X) the maximal singular value of
Xand
kGk1is the
H1norm of the linear system
G.
2 -analysis and LFTs
This section gives a short review on structured singular values and linear frac- tional transformations (LFTs), see also e.g. 3].
2.1 Denitions
The denition of
depends upon the underlying block structure of the un- certainties , which could be either real or complex, see gure 1. See also
15, 16]. For notational convenience we assume that all uncertainty blocks are square. This can be done without loss of generality by adding dummy inputs or outputs.
Given a matrix
M 2Cn
n and two non-negative integers
fR , and
fC , with
f
=
fR +
fC
nthe block structure is an
f-tuple of pairs of positive integers
N K
= (
n1k1)
:::(
nf
Rkf
R)
(
nf
R+1kf
R+1)
:::(
nf
R+f
Ckf
R+f
C)]
(1)
M
-
Figure 1: System with uncertainties.
where
Pf i
=1ni
ki =
nfor dimensional compatibility. The block repetition struc- ture is dened by
Nand the basic blocks by
K. The set of allowable perturba- tions is dened by a set of block diagonal matrices
X 2Cn
n dened by
X
=
f= diag
In
1R
1:::In
fRRf
RI
n
fR+1Cf
R+1:::In
fR+fCCf
R+f
C] :
Ri
2Rk
ik
iCf
R+i
2Ck
fR+ik
fR+ig(2) where
denotes the Kronecker product:
AB
=
2
6
4 a
11 B a
12
B ::: a
1
n
B... ... ...
a
m
1B am
2B ::: amn
B3
7
5 :
The Kronecker product has the following property, which is straight-forward to show,
(
AB)(
CD) = (
AC)
(
BD)
:(3) Using this equation we can now show the following commutative equation
(
In
)(
DIk ) = (
In
D)
(
Ik ) = (
DIn )
(
Ik ) = (
DIk )(
In
) where
2Ck
k and
D2Cn
n . That is,
In
and
DIk commute.
The uncertainty structure used here is slightly more general than in e.g.
15, 16], since also repeated full blocks are allowed both for real and complex uncertainties.
Assuming the uncertainty structure
N K, the structured singular value
of a matrix
M2Cn
n is dened by
=
min
2X
f
() : det(
I;M) = 0
g
;1
(4)
and if no
2Xsatises det(
I;M) = 0 then
(
M) = 0.
2.2 Upper Bounds
Generally the structured singular value cannot be exactly computed, and instead
we have to resort to upper and lower bounds, which are usually sucient for
most practical applications. A tutorial review of the complex structured singular value is given in 9].
An upper bound can be determined using convex methods, either involving minimization of singular values with respect to a scaling matrix or by solving a linear matrix inequality (LMI) problem. The upper bound is conservative in the general case, but can be improved by branch and bound schemes.
A lower bound can be found by maximizing the real eigenvalue of a scaled matrix. This bound is nonconservative in the sense that if the true global maximum is found it is equal to
. However, since the problem is not convex, we cannot guarantee that we nd the global maximum.
We will here focus on the computation of the upper bound, which we here denote
, in order to distinguish it from the true
function. The upper bound
can be computed as a convex optimization problem. For complex uncertainties it is dened by
(
M) = inf D
2D
(
DMD;1)
(
M) (5) where
Dis the set of block diagonal Hermitian matrices that commute with
X, that is
D
=
f0
<D=
D2Cn
n :
D=
D82Xg:(6) This problem is equivalent to an LMI problem
(
M) = inf >
P
2D0f
:
MPM<2Pg:(7) Real uncertainties can be included in the LMI problem for computing the upper bound (see e.g. 4, 15, 16]). We dene
(
M) = inf >
P G
2D2G0f
:
MPM+
j(
GM;MG)
<2Pg(8) where
G
=
fG=
G2Cn
n :
G=
G82Xg:(9) Every
G2Gis block diagonal with zero blocks for complex uncertainties. If we let
G=
f0
gin (8) we recover the complex upper bound (7).
We can reformulated (8) as a positive real property
(
M) = inf >
W
2W0
: herm
;(
I+
1M)
W(
I;1M
)
>0
(10) where herm
X=
12(
X+
X) and
W=
fW=
P+
jG:
P 2DG2Gg. Note that herm
W >0 always and that
W=
Wfor complex uncertainties. Another equivalent reformulation of (8) is
(
M) = inf >
D G
2D2G0 n:
;1DMD;1;jG
(
I+
G2)
;12<1
o:(11)
2.3 Linear Fractional Transformations (LFTs)
Suppose
Mis a complex matrix partitioned as
M
=
M
11 M
12
M
21 M
22
2C
(
p
1+p
2)(m
1+m
2)(12) and let u
2Cm
1p
1and l
2Cm
2p
2. The upper and lower linear fractional transformations (LFTs) are dened by
F
u (
Mu ) =
M22+
M21u (
I;M11u )
;1M12(13) and
F
l (
Ml ) =
M11+
M12l (
I;M22l )
;1M21(14) respectively. Clearly, the existence of the LFTs depends on the invertibility of
I;M
11
u and
I;M22l respectively.
The Redhe er star product 13] is a generalization of the LFTs. Assume that
Qis partitioned similarly to
M. Then the star product is dened by
S
(
QM) =
F
l (
QM11)
Q12(
I;M11Q22)
;1M12M
21
(
I;Q22M11)
;1Q21 Fu (
MQ22)
:
(15) Note that the denition above is dependent on the partitioning of the ma- trices
Qand
M. The LFTs can be dened by the
Snotation, as
F
u (
Mu ) =
S( u
M) and
Fl (
Ml ) =
S(
Ml )
:The star product is associative, that is
S(
AS(
BC)) =
S(
S(
AB)
C)
:2.4 Frequency transformation
The bilinear transformation between the
z-domain and the
s-domain is
s= z z
;1+1, which is given by
Fu (
Nz;1I) where
N
=
I
p
2
I;
p
2
I ;I
:
Using this transformation we can map the continuous time problem to a static
problem 3].
2.5 on Linear Systems
If
G(
s) =
D+
C(
sI;A)
;1Bis a stable linear system, we dene
kGk
= max !
2R
f
(
G(
j!))
g:(16) Note that this notation is somewhat misleading since
k:kis not a norm.
Using the frequency transformation
s= z z
+1;1together with LFTs we can recast the
k:kfunction into a pure
problem. We can then obtain the following LMI test for giving an upper bound on
kGk. If there exist
>0,
P2D, ;
2jGand
X >0, such that
A
T
X+
XA XBB
T
X0
+
C D
0
IT
P
; T
;
;2P
C D
0
I
<
0
then
kGk.
3 Time Varying Uncertainties
3.1 A Unied Approach
We will here adopt a unied approach for including time-varying uncertainties in the
formalism. We have previously based the computation of the upper
bound on two commutative sets
Dand
G. In the case of time-invariant (constant) uncertainties, either parametric (real) or dynamic (complex), we may have multipliers
Dand
Gthat are frequency dependent (dynamic). In the case of time-varying uncertainties, including nonlinear elements with bounded
L2gain and parametric varying parameters,
Dand
Gare generally restricted to constants (with respect to time or frequency).
Both these structures, constant and nonlinear, are extremes. A \constant"
parameter in practice is usually not constant but slowly varying. A dynamic nonlinear uncertainty is normally too conservative. Thus, it would be of great importance to include other structures in between these two extremes. We will here look at uncertainties that are varying with bounds on the rate of change.
The approach presented allows us to include these uncertainties in the standard
framework.
We denote uncertainties that have bounds on
_
:::(m
), but not on
(m
+1)or higher derivatives, to belong to the class
Vm . Any time-varying uncertainty (without bounds on _) belongs to
V0. We can include these time- varying structures into the framework presented here.
3.2 Uncertainty Block Augmentation
The main idea in this approach to handle slowly time-varying uncertainties is the concept of uncertainty block augmentation. This idea can be used also for constant uncertainties. An uncertainty block can be augmented by either adding copies of the original uncertainty or its derivative other choices may also be possible.
We have previously used the commutative property as an important tool for
nding the upper bound on
. The commutative property
D=
Dcan also be stated as
DD;1= , since
Dby denition is nonsingular.
Generally, we need no commutative set, but it is enough to nd an augmented uncertainty block ~ together with left and right (dynamic) multipliers,
Y(
s) and
Z
(
s) respectively, such that
Y
(
s)~
Z(
s) =
:If this is possible, then the system
Mcan be shown to be stable if
(~)
<1
=kZMYk.
For a constant uncertainty
we can augment the uncertainty block by copies of itself: ~ =
In . Then we can choose any dynamic multipliers
Y(
s) and
Z(
s) such that
Y(
s)
Z(
s) = 1. Thus,
Y(
s)~
Z(
s) =
.
Example 3.1 Let ~ = diag
],
Y(
s) =
12 1+21+s s
1+21+s s
and
Z(
s) =
"
1+2
s
1+
s
1+
s
1+2
s
#
, then
Y(
s)
Z(
s) = 1 and consequently
Y(
s)~
Z(
s) =
.
2For slowly time-varying uncertainties we include _ = ddt into the aug- mented uncertainty block ~. We use
sto denote the di erential operator ddt interchangeably with the Laplace argument.
Example 3.2 Let ~ =
_
,
Y(
s) =
1 +
as ;aand
Z(
s) = 1 1 +
as. Let
y= 1 1 +
asx. Then
Y
(
s) ~
Z(
s)
x= (1 +
as)
y;ay_
=
y+
ay_ +
ay_
;ay_
=
(
y+
ay_ ) =
(1 +
as)
y=
x:Thus,
Y(
s)~
Z(
s) =
.
2We can generalize this to the multivariable case, for which we provide the following lemma.
Lemma 3.1 Let =
Iwhere
is a real or complex scalar. Assume that
Ahas all its eigenvalues with negative real part and let the system be initialized to zero state at
t=
;1. Then we have =
Y(
s)~
Z(
s), if either
(i) ~ =
_
,
Y(
s) =
sI;A ;Iand
Z(
s) = (
sI;A)
;1or (ii) ~ =
_
,
Y(
s) = (
sI ;A)
;1and
Z(
s) =
sI;A
I
.
Proof:
(i) Let
y= (
sI;A)
;1x, then
Y
(
s)~
Z(
s)
x= (
sI;A)
y;y_
= _
y+
y_
;Ay;y_
=
( _
y;Ay) =
(
sI;A)
y=
x:(ii)
Y
(
s) ~
Z(
s)
x= (
sI;A)
;1(
(
sI;A)
x+ _
x)
= (
sI;A)
;1(
x_
;Ax+ _
x)
= (
sI;A)
;1(
sI;A)(
x) =
x:2
We can now combine the two results in lemma 3.1 into the following more general lemma.
Lemma 3.2 Let =
Iwhere
is a real or complex scalar. Assume that
A
1
and
A2have all their eigenvalues with negative real part and let the system
be initialized to zero state at
t=
;1. Then with ~ = diag
_] we have
=
Y(
s)~
Z(
s) if
Y
(
s) =
Y0(
s)
Y1(
s)
=
(
sI;A1)
;1(
sI;A2) (
sI;A1)
;1(17) and
Z
(
s) =
Z
0
(
s)
Z
1
(
s)
=
(
sI;A2)
;1(
sI;A1) (
sI;A2)
;1(
A1;A2)
:
(18)
Proof: We expand the -block by using lemma 3.1 (ii) and (i) in two consec- utive steps.
= (
sI;A1)
;1_
sI;A
1
I
= (
sI;A1)
;1
sI;A
2
;I
_
(
sI;A2)
;1_
sI;A1I
= (
sI;A1)
;1(
sI;A2)(
sI;A2)
;1(
sI;A1) + (
sI;A1)
;1;_(
sI;A2)
;1(
sI;A1) + _
= (
sI;A1)
;1(
sI;A2)(
sI;A2)
;1(
sI;A1) + (
sI;A1)
;1_(
sI;A2)
;1(
A1;A2)
=
Y(
s)
0 0 _
Z
(
s)
:2
Remark 3.1 ~ If =
Ik we can generalize the set of scaling matrices by letting
Y
(
s) =
CY(
s) and ~
Z(
s) =
Z(
s)
Bwhere
Yand
Zare dened by (17) and (18) respectively and
CB=
Ik .
Remark 3.2 We can also include !
and higher order derivatives in the un- certainty structure. The bounds on these higher order derivatives of
can be included easily in the same formalism, since the augmented block has the same properties as the original uncertainties. Hence, we can augment the _ -block with ! similarly to the augmentation of with _.
Remark 3.3 We have assumed that both
and _
are weighted equally in each block. Dierent weighting can be introduced by a scaling. Let ~ = diag
=d_ ] and scale the last block in
Yor
Zby
d.
Remark 3.4 The approach taken here is related to \the swapping lemma" 7]
used in adaptive control for stability analysis. The swapping lemma states in principle that the dierence
D(
s)
;D(
s) is bounded. In our approach we explicitly solve the dierence and compensates for it exactly:
Z
0
(
s)
;Z0(
s) =
Z0(
s)
Y1(
s)_
Z1(
s)
where we have used the fact that
Z0=
Y0;1.
4 An Example
To illustrate the technique we give a simple example. The problem is described by
M
(
s) =
0
1+44+s s
1+4
s
4+
s 0
=
1
0 0
2
1
2
2C
(19) for which
kMk= 1 with time-invariant uncertainties. To obtain this we can use the dynamic scaling
D
(
s) = diag 1
4 +
s1 + 4
s]
since
kMk
=
kDMD;1k1=
0 1 1 0
= 1
:Assuming dynamic uncertainties then
kMk= 4 with a constant
D=
I. In order to make the system unstable using a nonlinear uncertainty already at a low gain
> 14, we need uncertainty elements that are able to transform energy from one frequency to another. If the uncertainty elements are slowly varying we may expect a lower value on
, which means that we can tolerate larger uncertainties while maintaining stability.
We now assume that the uncertainty
22V1is slowly time-varying (
12V0may change without bound on rate). We will use the following multipliers
Y
(
s) =
"
1 0 0
0
1+44+s s
p4+15d= s
4#
and
Z(
s) =
2
6
6
6
4
1 0 0
1+44+s s 0
p1+460d s
3
7
7
7
5 :
that satises the structure given in lemma 3.2. Then
=
Y(
s) ~
Z(
s) =
Y(
s)
2
6
4
1
0 0
0
20 0 0 _
2=d3
7
5 Z
(
s)
where ~ is the augmented uncertainty. The scaled system now becomes
Z
(
s)
M(
s)
Y(
s) =
2
6
6
4
0 1
p1+415d= s
41 0 0
p
60
d
4+
s 0 0
3
7
7
5 :
The following upper bound is then obtained
kMk
kZMYk0
B
@ 2
6
4
0 1
p152d
1 0 0
p
15
d
2
0 0
3
7
5 1
C
A
=
q1 +
154d :We can now choose
din order to cope with varying uncertainties. If we choose
d
to be small, the
value will be approximately 1. For instance, let
d= 0
:1
10−2 10−1 100 101 102 0
0.2 0.4 0.6 0.8 1
d
1/mu
Varying mu analysis
Figure 2: Varying
analysis. The horizontal line through 1
== 0
:25 is the bound obtained with a constant scaling (no bound on the rate of
). The bell- shaped line starting close to 1 and then decreasing is the obtain with the scaling described in the example. The crosses mark improved bounds on 1
=obtained using modied scalings.
then
1
:1726. Thus the system is stable if
j1j<0
:8528,
j2j<0
:8528 and
j
_
2j<0
:0853, since
(~) = max
fj1jj2jj_
2j=dg.
If we choose
dto be large, the
value increases. For instance, with
d= 1 we obtain
2
:1794 and thus the system is stable if
j1j<0
:4588,
j2j<0
:4588 and
j_
2j0
:4588. The bound can be improved further, especially for large values of
dby modifying
Yand
Z.
The result obtained from this analysis is given in gure 2. The two lines gives the bounds obtained from constant scaling and the augmented uncertainty technique. The bounds can be further improved by modifying the scalings (
Yand
Z). The crosses in the gure indicate such improved bounds.
5 Uncertainty Structures
In this section, we will discuss the various structures of uncertainties that can be included in the
framework. The denition of the uncertainty structure is closely related to its commuting properties. We will here make a more general classication of uncertainties than in e.g. 15, 16].
5.1 Classication
The uncertainties can be described by the four characteristics given in the table
below.
1 eld (
F) complex (
C) real (
R)
dynamic parametric
2 structure full matrix repeated scalar
N K
without structure or full blocks 3 variability time-varying constant
V V
0