COPRIME FACTOR REDUCTION OF H∞ CONTROLLERS

A. Varga
German Aerospace Center, DLR - Oberpfaffenhofen, Institute of Robotics and System Dynamics, D-82234 Wessling, Germany,
[Link]@[Link]

Keywords: Controller reduction, frequency-weighted model frequency-weighted coprime factor controller reduction meth-
reduction, balanced realization, numerical methods. ods.
In this paper, we address the efficient solution of frequency-
Abstract weighted balancing-related coprime factor controller reduction
problems for the special stability and performance preserv-
We consider the efficient solution of the coprime factoriza- ing frequency-weights proposed in [4, 5]. We show that for
tion based H∞ controller approximation problems by us- the reduction of the H∞ central controller, the computation
ing frequency-weighted balancing related model reduction ap- of frequency-weighted grammians for the coprime factor con-
proaches. It is shown that for a class of frequency-weighted troller reduction can be done efficiently by solving Lyapunov
performance preserving coprime factor reduction as well as for equations of the order of the controller. The Lyapunov equa-
a relative error coprime factor reduction method, the compu- tions can be solved directly for the Cholesky factors of the
tation of the frequency-weighted controllability and observ- grammians, thus allowing the application of the balancing-free
ability grammians can be done by solving Lyapunov equa- square-root accuracy enhancing method for coprime factor re-
tions of the order of the controller. The new approach can be duction [13].
used in conjunction with accuracy enhancing square-root and
balancing-free techniques developed for the balancing related Notation. Throughout the paper, the following notational con-
coprime factors based model reduction. vention is used. The bold-notation G is used to denote a state-
space system G := (A, B, C, D) with the transfer-function
matrix (TFM)
1 Introduction
Using the H∞ controller synthesis methodology (see for exam- · ¸
A B
ple [19]) often leads to controllers whose orders are too large G(λ) = C(λI − A)−1 B + D := .
C D
for practical use. Therefore, in such cases it is necessary to per-
form controller reduction by determining a lower order approx-
imation of the original controller. Controller reduction prob- According to the system type, λ is either the complex variable s
lems are often formulated as special frequency-weighted model appearing in the Laplace transform in the case of a continuous-
reduction (FWMR) problems, where the frequency-weights are time system or the variable z appearing in the Z-transform in
chosen to enforce closed-loop stability and an acceptable per- the case of a discrete-time system. Throughout the paper we
formance degradation when the low order controller is used denote G(λ) simply as G, when the system type is not relevant.
instead the original high order one [1]. The bold-notation is used consistently to denote systems corre-
sponding to particular TFMs: G1 G2 denotes the series cou-
The idea to apply frequency-weighted balancing techniques to
pling of two systems having the TFM G1 (λ)G2 (λ), G1 + G2
reduce the stable coprime factors of the controller has been dis-
represents the (additive) parallel coupling of two systems with
cussed in several papers [1, 8, 18]. For the reduction of con-
TFM G1 (λ)+G2 (λ), G−1 represents the inverse systems with
trollers originating from H∞ synthesis several methods have
TFM G−1 , [ G1 G2 ] represents the realization of the com-
been proposed [4, 5, 16, 2]. While the frequency-weights in
pound TFM with [ G1 G2 ], etc.
[1, 8] have been primarily chosen to guarantee closed-loop sta-
bility, the H∞ controller reduction mainly focusses on pre-
serving the performance bounds achieved by the original con- 2 Coprime factor controller reduction
trollers. Interestingly, many stability/performance preserv-
Consider G := (A, B, C, D), an n-th order state-space
ing controller reduction problems have very special structure
model and let K be a stabilizing controller with a stabiliz-
which can be exploited when developing efficient numerical
able and detectable nc -th order state space realization K :=
algorithms for controller reduction. For example, it has been
(Ac , Bc , Cc , Dc ). The solution of the a frequency-weighted
shown in [15] that for the frequency-weighted balancing re-
coprime factor controller reduction problem (see for example
lated approaches applied to several controller reduction prob-
[1, 8]) consists in computing an approximation of the coprime
lems with the special stability/performance enforcing weights
factors of the controller. Specifically, the Frequency-Weighted
proposed in [1], the computation of grammians can be done
Left Coprime Factor Reduction (FWLCFR) Problem is: given a
by solving reduced order Lyapunov equations. Similarly, it
left coprime factorization (LCF) K = Ve −1 U e of the controller,
was shown recently in [14] that this is also true for a class of
find Kr , an rc -th order approximation of K, in a LCF form
 er , such that the weighted approximation error
Kr = Ver−1 U H∞ controller reduction problems formulated in [4]. Let
· ¸
fo [ U
e −U
er Ve − Ver ]W
fi k∞ , M11 M12
kW (1) M= (3)
M21 M22
is minimized. Similarly, the Frequency-Weighted Right Co- be the TFM used to parameterize all admissible γ-suboptimal
prime Factor Reduction (FWRCFR) Problem is: given a right controllers [19]. It follows that K can be expressed in terms of
coprime factorization (RCF) K = U V −1 of the controller, a lower linear fractional transformation (LFT) in the form
find Kr , an rc -th order approximation of K, in the RCF form
Kr = Ur Vr−1 , such that the weighted approximation error K = Fl (M, Q) := M11 + M12 Q(I − M22 Q)−1 M21 ,
· ¸ where Q is a stable and proper rational matrix satisfying
U − Ur
kWo Wi k∞ , (2) kQk∞ < γ. Since for standard H∞ problems both M12 and
V − Vr
M21 are invertible and minimum-phase [19], a ”natural” LCF
fo , W
fi , Wo and Wi are stable of the central controller (Q = 0) as K0 = Ve −1 U
e can be ob-
is minimized. In (1) and (2), W
tained with
weighting TFMs, which are specially chosen to enforce closed- e = M −1 M11 , Ve = M −1
U 12 12
loop stability and performance.
while a ”natural” RCF of the central controller as K0 = U V −1
Balancing related FWMR techniques which attempt to min- can be obtained with
imize (1) or (2) can be used to determine reduced order
−1 −1
controllers. The following procedure to solve the FWLCFR U = M11 M21 , V = M21
Problem is based on the FWMR approach proposed by Enns
in [3]: These coprime factorizations can be used to perform un-
weighted coprime factor controller reduction using accuracy
enhancing model reduction algorithms [13].
FWLCFR Procedure.
The frequency weighted left coprime factor reduction formu-
1. Compute the controllability grammian of [ U e V
e ]Wf i and
lated in [4] is one sided with
f e e
the observability grammian of Wo [ U V ] and define ac- · −1 ¸
cording to [3], appropriate nc order frequency-weighted f f e −1 γ I 0
Wo = I, Wi = Θ (4)
controllability and observability grammians PE and QE , re- 0 I
spectively.
where
2. Using PE and QE in place of standard grammians of " # · ¸
[Ue V
e ], determine a reduced order approximation [ U
er Ver ] e
Θ e 12
Θ −1
M21 − M22 M12 −1
M11 −M22 M12
e
Θ = e 11 := −1 −1
by applying, for example, the balanced truncation (BT) Θ21 e 22
Θ M12 M11 M12
method [9] or the singular perturbation approximation
(SPA) [7]. Note that Θe is stable, invertible and minimum-phase. With the
3. Form Kr = V e −1 U
e r. help of the submatrices of Θ e it is possible to express K also as
r

A completely similar procedure can be used to solve the e 22 + QΘ

K = (Θ e 12 )−1 (Θ
e 21 + QΘ
e 11 )
FWRCFR Problem:
e −1 Θ
and thus the central controller is factorized as K0 = Θ e 21 .
22

FWRCFR Procedure. Similarly, a frequency-weighted right coprime factor reduction

can be formulated with the one sided weights
1. Compute
· ¸ the controllability
· ¸and observability grammians of · −1 ¸
U U γ I 0
Wi and Wo , respectively, and define ac- Wo = Θ−1 , Wi = I (5)
V V 0 I
cording to [3], appropriate nc order frequency-weighted
controllability and observability grammians PE and QE , re- where
spectively. · ¸ · −1 −1
¸
Θ11 Θ12 M12 − M11 M21 M22 M11 M21
Θ= := −1 −1
2. Using
· ¸PE and QE · in place
¸ of standard grammians of Θ21 Θ22 −M21 M22 M21
U Ur
, determine , a reduced order approximation, Note that this time we have
V Vr
by applying either the BT method [9] or the SPA [7].
K = (Θ12 + Θ11 Q)(Θ22 + Θ21 Q)−1
3. Form Kr = Ur Vr −1 .
and the central controller is factorized as K0 = Θ12 Θ−1
22 .
In this paper we focus on the efficient and numerically accurate
computation of low order controllers by using these procedures The importance of the above frequency-weighted coprime fac-
to solve the frequency-weighted coprime factorization based tor reduction can be seen from the following result [4].
Theorem 2.1 Let K0 be a stabilizing continuous-time γ- 3 Efficient solution of frequency-weighted H∞
suboptimal H∞ central controller, and let Kr be an approx- controller reduction problems
imation of K0 computed by applying either the FWLCFR
Procedure or FWRCFR Procedure. Then Kr stabilizes the 3.1 LCF controller reduction
closed-loop system and preserves the γ-suboptimal perfor-
We consider the efficient computation of the frequency-
mance, provided√ the weighted approximation error (1) or (2) weighted controllability grammian at Step 1 of the FWLCFR
is less than 1/ 2.
Procedure for the weights defined in (4). Let consider a real-
ization of the parameterization TFM M (3) in the form
We conjecture that this result holds also in the discrete-time,
 
and can be proved along the lines of the proof provided in [19]. Ab B b1 Bb2
 b b 12 
b 11 D
An alternative approach to H∞ controller reduction uses the M= C 1 D 
relative error method as suggested in [17]. Using this ap- b b
C2 D21 D22 b
proach in conjunction with the LCF reduction we can define
the weights as Note that for the central controller we have (Ac , Bc , Cc , Dc ) =
b B
(A, b1 , C
b1 , D
b 11 ). Since M12 and M21 are stable, minimum-
fo = [ U
W e Ve ]+ , fi = I
W (6) b 12 and D b 21 are
phase and invertible TFMs, it follows that D
e Ve ]+ denotes a stable right inverse of [ U e Ve ]. A vari- invertible, A,b A b−B b2 D
b −1 C
b1 and A b−B b1 Db −1 Cb2 are all sta-
where [ U 12 21
ant of this approach (see [19]) is to perform a relative error co- ble matrices, i.e., have eigenvalues in the open left half plane
prime factor reduction on an invertible augmented minimum- for a continuous-time controller and in the interior of the unit
ea Vea ] instead of [ U
e Ve ]. In our case, Θe can be circle for a discrete-time controller. The realizations Θ e =
phase system [ U e −1 = (A −1 , B −1 , C −1 , D −1 )
taken as the augmented system. Thus this method essentially (AΘ ,
e Θ B ,
e Θ C ,
e ΘD e ) and Θ e
Θ e
Θ e
Θ e
Θ
consists of determining an approximation Θ e r of Θ
e by mini- can be computed as [19]
e e −1 e e  
mizing the relative error ∆r = Θ (Θ − Θr ). The reduced Ab−B b2 Db −1 Cb Bb1 − B b2 Db −1 Db 11 −B b2 Db −1
12 1 12 12
controller is recovered from the sub-blocks (2,1) and (2,2) of
Θe = C b2 − D b 22 D
b −1 Cb b b b −1 b b b −1 
e r as Kr = Θ
Θ e −1 Θe 12 1 D21 − D22 D12 D11 −D22 D12 
r,22 r,21 . b −1 C b b −1 Db 11 b −1
D 12 1 D 12 D 12
A relative error RCF reduction can be formulated with the  
Ab−B b1 Db C
−1 b b b −1 b2 − B b1 D
b D−1 b 22
weights 21 2 −B1 D21 B 21
· ¸+ Θe −1 =  Db −1 Cb Db −1 Db −1 Db 22 

U 21 2 21 21
Wo = I, Wi = (7) Cb1 − D b 11 D
b −1 Cb2 −Db 11 Db −1 D b 12 − D b 11 D
b −1 Db 22
V 21 21 21
· ¸+ · ¸
U U e V e ]W
f i has order 2nc , it fol-
where denotes a stable left inverse of . Alter- Since the realization of [ U
V V
lows that the solution of the controller reduction problem for
natively, an augmented relative error problem can be solved by
the special weights defined in (4) involves the solution of a
approximating Θ by a reduced order system Θr which mini-
Lyapunov equation of order 2nc to determine the frequency-
mizes the relative error ∆r = (Θ − Θr )Θ−1 . The reduced
weighted controllability grammian PE and a Lyapunov equa-
controller is recovered from the sub-blocks (1,2) and (2,2) of
e r as Kr = Θr,12 Θ−1 . This method has been also considered tion of order nc to compute the observability grammian QE .
Θ r,22 The following result shows that it is always possible to solve
in [2] for the case of normalized coprime factor H∞ controller
two Lyapunov equations of order nc to compute the frequency-
reduction.
weighted grammians for the special weights in (4).
The main computational burden in applying to these problems
either the FWLCFR or FWRCFR procedure is the compu- Lemma 3.1 The frequency-weighted controllability grammian
tation of the grammians at Step 1. Apparently, the computa- PE and observability grammian QE according to Enn’s choice
tion of grammians involves the solutions of at least one Lya- [3] satisfy, according to the system type: continuous-time (c) or
punov equation of order 2nc . In this paper we show that for the discrete-time (d), the corresponding Lyapunov equations
method of [4] as well as for the augmentation based relative er- (
AΘ T e −1 BeT
ror methods, the frequency-weighted grammians can be com-
(c) e−1 PE + PE AΘ e−1
+B e
Θ e−1
Θ
= 0
puted by solving Lyapunov equations each of order nc . Com- T
A QE + QE AΘ e T e
e e + CΘ Ce
e Θ
= 0
plete formulas for both continuous- and discrete-time systems ( Θ
AΘ e −1 BeT
are given for both LCF and RCF based approaches.
(d) e−1 PE AΘe−1 + BΘ e e−1
Θ
= PE
T
A QE AΘ e C
T e
In a separate section, we discuss shortly the direct computation e
Θ e + CΘ e Θ e = QE
of the Cholesky factors of the frequency-weighted grammians. e −1 = B −1 diag (γ −1 I, I) and C
e =Db −1 C
b
where BΘe e
Θ e
Θ 12 1 .
This is a prerequisite for the applicability of the balancing-free
square-root accuracy-enhancing techniques to coprime factor
controller reduction of [13], along the lines of the model re- Proof: We can construct immediately the realization of
e V
[U e ]W
f i := (Ai , B i , C i , Di ) with
duction methods developed for the BT in [12] and SPA in [11].
 Lemma 3.2 The frequency-weighted controllability grammian
" # PE and observability grammian QE for Enns’ method [3] sat-
b−B
A b2 D
b −1 C
b B b1 D
b −1 C
b b b −1 b
Ai = 12 1 21 2 − B2 D12 C1 (8) isfy, according to the system type, the corresponding Lyapunov
0 b−B
A b1 D
b −1 Cb equations
21 2
" # (
γ −1 B b −1 −B
b1 D b2 + Bb1 D b 22
b −1 D AΘ T T
Bi = b1 D
21
b −1 b2 − Bb1 D
21
b −1 D
b 22 (c) e PE + PE AΘ e
+ BΘ e BΘ
e
= 0
−γ −1 B B T
A −1 QE + QE AΘ T
21 21
e e−1 + CΘ Ce−1 = 0
e−1 Θ
( Θ
and let P i and Q be the controllability grammian of [ U e V
e ]W
fi AΘ T
(d) e PE AΘe + BΘ e BΘe
= PE
e e
and the observability grammian of [ U V ], respectively. Ac- T
A −1 QE AΘ + C T
C = QE
e
Θ e−1 e−1 Θ
Θ e−1
cording to the system type, P i and Q satisfy the corresponding
Lyapunov equations
( T T
Remark. For the relative error method with the weights given
Ai P i + P i Ai + B i B i = 0 in (6), a right inverse can be immediately constructed as
(c) b −1 Cb1 )T D
b −1 Cb1 = 0
AT Q + QAΘ e + ( D · ¸
( Θ e 12 12 −1
M21 M22
T T e e
[U V ] =+
−1
Ai P i Ai + B i B i = Pi M12 − M11 M21 M22
(d) b −1 Cb1 )T D
b −1 Cb1 = Q
AT QAΘ e + (D
e
Θ 12 12
e Ve ]+ is given by
A realization of the output weight Wo = [ U
Partition P i in accordance with the structure of Ai in (8) " #
AΘ e −1
B
" # e −1 e
Θ
P 11 P 12 Wo =
CΘ e −1
Pi = T (9) e−1 DΘ e
P 12 P 22
where B e −1 = B −1 [ 0 I ]T and D e −1 = D −1 [ 0 I ]T .
e
Θ e
Θ e
Θ e
Θ
such that P 11 is an nc × nc matrix. Enns defines in [3] PE = Thus, the grammians PE and QE used in Lemma 3.2 are the
P 11 and QE = Q as the frequency-weighted controllability e Ve ] and the observability gram-
controllability grammian of [ U
and observability grammians, respectively. e e +
mian of [ U V ] , respectively. 2
· matrix T ¸
Consider the transformation
Inc −Inc 3.3 RCF controller reduction
T =
0 Inc
We consider the efficient computation of the frequency-
It is easy to see that the controllability grammian Pei for the weighted controllability grammian at Step 1 of the FWRCFR
transformed pair (A ei , B
ei ) := (T −1 Ai T, T −1 B i ) has the form Procedure for the weights defined in (5). The realizations Θ =
e
Pi = diag (0, Pi ), where Pi satisfies the appropriate Lyapunov (AΘ , BΘ , CΘ , DΘ ) and Θ−1 = (AΘ−1 , BΘ−1 , CΘ−1 , DΘ−1 )
equation can be computed as [19]
T e −1 BeT  
(c) AΘ e−1 Pi + Pi AΘ +B e =0 b−B
A b1 D
b −1 Cb2 b2 − B
B b1 D b −1 Db 22 B b1 Db −1
e−1 Θ e−1
Θ (10) 21 21 21
(d) AΘ e e T  b b b −1 
e−1 Pi AΘe−1 + BΘ e−1 BΘe−1
= Pi Θ=C b b −1 b b b b −1 b
1 − D11 D21 C2 D12 − D11 D21 D22 D11 D21 
−Db −1 Cb −D b −1 Db 22 Db −1
The grammian in the original coordinate basis results as 21 2 21 21
 
· ¸ Ab−B b2 Db C
−1 b b b −1 b1 − B b2 D b D
−1 b 11
Pi −Pi 12 1 B2 D12 B 12
P i = T Pei T T =  
−Pi Pi Θ−1 =  −Db −1 C
12 1
b Db −1
12 −D b −1 D
12
b 11 
Cb2 − Db 22 D
b −1 C
b1 D
b 22 Db −1 D b 21 − D b 22 Db −1 Db 11
12 12 12
Thus, the frequency-weighted controllability grammian ac-
cording to Enns’ method is PE = Pi , the leading nc × nc · ¸
U
block of P i . 2 Since the realization of Wo has order 2nc , it follows
V
e Ve ]W
Remark. It is easy to see that [ U fi = [ 0 I ], thus com- that the solution of the controller reduction problem for the spe-
plete pole-zero cancellation takes place between the system to cial weights defined in (5) involves the solution of a Lyapunov
be reduced and the input weight. This situation is typical for equation of order 2nc to determine the frequency-weighted
several frequency-weighted controller reduction problems (see controllability grammian PE and a Lyapunov equation of or-
for instance [1, 15, 19]) and can be addressed by using Enns’ der nc to compute the observability grammian QE . The fol-
choice of frequency-weighted grammians. lowing result shows that it is always possible to solve two
Lyapunov equations of order nc to compute the frequency-
weighted grammians for the special weights in (5).
3.2 Relative error LCF reduction
e is in fact a FWMR prob-
The relative error approximation of Θ Lemma 3.3 The frequency-weighted controllability grammian
e −1
lem with the weights Wo = Θ and Wi = I. We have the PE and observability grammian QE for Enns’ method [3] sat-
following straightforward result [19, Theorem 7.5]: isfy, according to the system type, the corresponding Lyapunov
equations The grammian in original coordinates results as
( · ¸
eΘ B
AΘ PE + PE ATΘ + B eT = 0 e o T −1 = Qo −Qo
(c) Θ Qo = T −T Q
T e T −1 C
eΘ−1 = 0 −Qo Qo
AΘ−1 QE + QE AΘ−1 + C Θ
( According to Enns’ method, the frequency-weighted observ-
eΘ B
AΘ PE ATΘ + B eT = PE
(d) Θ ability grammian is QE = Qo , the trailing nc × nc block of Qo
T
AΘ−1 QE AΘ−1 + C e T −1 C
eΘ−1 = QE
Θ in (12). 2
· ¸
where B eΘ = BΘ 0 = B b1 D
b −1 and CΘ−1 = 3.4 Relative error RCF reduction
I 21
−1
diag (γ I, I)CΘ−1 . The relative error approximation of Θ is a FWMR problem
· ¸ with the weights Wo = I and Wi = Θ−1 . We have the fol-
U lowing straightforward result [19, Theorem 7.5]:
Proof: We can construct the realization of Wo :=
V
(Ao , B o , C o , Do ) with the matrices Lemma 3.4 The frequency-weighted controllability grammian
" # PE and observability grammian QE for Enns’ method [3] sat-
b−B
A b2 D
b −1 C
b1 −B b1 D
b −1 C
b2 + Bb2 Db −1 C
b1 isfy, according to the system type, the corresponding Lyapunov
Ao = 12 21 12 (11)
0 Ab−B b1 D
b −1 C
b2 equations
21
" # ½
b −1 C
b b −1 Cb AΘ−1 PE + PE ATΘ−1 + BΘ−1 BΘ T
−1 = 0
−γ −1 D 12 1 γ −1 D 12 1 (c)
Co = ATΘ QE + QE AΘ + CΘ T
CΘ = 0
b2 − D
C b 22 D
b C
−1 b b2 + D
−C b 22 D
b −1 C
b ½
12 1 12 1 T
AΘ−1 PE AΘ−1 + BΘ−1 BΘ−1 = PE
(d)
ATΘ QE AΘ + CΘ T
CΘ = QE
Let P and· Qo be¸ the controllability
· ¸ and observability gram-
U U
mians of and Wo , respectively. According to Remark. For the relative error method with the weights given
V V
in (6), a left inverse can be immediately constructed as
the system type, P and Qo satisfy the corresponding Lyapunov
· ¸+
equations U −1 −1
( = [ M22 M12 M12 − M22 M12 M11 ]
V
AΘ P + P ATΘ + B eΘ BeT = 0
(c) T T
Θ · ¸+
Ao Qo + Qo Ao + C o C o = 0 U
( A realization of the input weight Wi = is given by
V
AΘ P ATΘ + B eΘ BeT = P
(d) Θ " #
T T AΘ−1 BΘ−1
Ao Qo Ao + C o C o = Qo
Wi =
eΘ−1 D
C e Θ−1
Partition Qo in accordance with the structure of the matrix Ao
in (11) " # where C eΘ−1 = [ 0 I ]CΘ−1 and D e Θ−1 = [ 0 I ]DΘ−1 . Thus,
Q11 Q12 the grammians PE and QE used in Lemma 3.4 are the control-
Qo = T (12) · ¸+
Q12 Q22 U
lability grammian of and the observability grammian
V
where Q22 is an nc × nc matrix. The approach proposed by [3] · ¸
U
defines of , respectively. 2
V
PE = P, QE = Q22 (13)
as the frequency-weighted observability grammian. 4 Square-root techniques
Consider the transformation matrix T
· ¸ Accuracy enhancing balancing-free square-root techniques for
Inc Inc coprime factor model reduction have been proposed in [13]
T =
0 Inc along the lines of similar methods developed for the BT in
[12] and SPA in [11]. The key computation in the proposed
It is easy to see that the observability grammian Q e o for the
procedures is the determination of the Cholesky factors of the
e e −1
transformed pair (Ao , Co ) := (T Ao T, C o T ) has the form T T
grammians such that PE = SE SE and Q = RE RE . The
· ¸ method of Hammarling [6] can be generally employed to solve
e o = Qo 0
Q the Lyapunov equations in question directly for the Cholesky
0 0
factors. Having these factors, the reduction of coprime factors
where Qo satisfies the appropriate Lyapunov equation can be performed by computing two truncation matrices L and
T using the singular value decomposition
(c) ATΘ−1 Qo + Qo AΘ−1 + C eΘ−1
e T −1 C =0
Θ (14) £ ¤ £ ¤T
T e T e
(d) AΘ−1 Qo AΘ−1 + CΘ−1 CΘ−1 = Qo RE SE = U1 U2 diag(Σ1 , Σ2 ) V1 V2
with Σ1 = diag(σ1 , . . . , σrc ), Σ2 = diag(σrc +1 , . . . , σnc ) and [6] S. J. Hammarling. Numerical solution of the stable,
σ1 ≥ . . . ≥ σrc > σrc +1 ≥ . . . ≥ σnc ≥ 0. The square-root non-negative definite Lyapunov equation. IMA J. Numer.
method determines L and T as [10] Anal., 2:303–323, 1982.
−1/2 −1/2
L = Σ1 U1T RE , T = S E V1 Σ 1 . [7] Y. Liu and B. D. O. Anderson. Singular perturbation
approximation of balanced systems. Int. J. Control,
If the original system is highly unbalanced, potential accuracy 50:1379–1405, 1989.
losses can be induced in the reduced model if either of the trun-
[8] Y. Liu, B. D. O. Anderson, and U. L. Ly. Coprime factor-
cation matrices L or T is ill-conditioned (i.e., nearly rank defi-
ization controller reduction with Bezout identity induced
cient). To avoid ill-conditioned truncation matrices, balancing-
frequency weighting. Automatica, 26:233–249, 1990.
free approaches can be used, as for example, the balancing-free
square-root algorithm for the BT introduced by [12]. Similar [9] B. C. Moore. Principal component analysis in linear sys-
formulas have been developed for the SPA approach in [11]. tem: controllability, observability and model reduction.
IEEE Trans. Autom. Control, 26:17–32, 1981.
5 Conclusions
[10] M. S. Tombs and I. Postlethwaite. Truncated balanced re-
Efficient and numerically reliable balancing related compu- alization of a stable non-minimal state-space system. Int.
tational approaches have been proposed for the frequency- J. Control, 46:1319–1330, 1987.
weighted coprime factors H∞ controller reduction with spe-
cial frequency weights enforcing closed-loop stability and per- [11] A. Varga. Balancing-free square-root algorithm for com-
formance. To compute lower order approximations of the co- puting singular perturbation approximations. Proc. of
prime factors, ”natural” coprime factorizations of the central 30th IEEE CDC, Brighton, UK, pp. 1062–1065, 1991.
H∞ controller are used, which result from the parameteriza-
tion of all suboptimal H∞ controllers. We developed com- [12] A. Varga. Efficient minimal realization procedure based
plete formulas to compute the frequency-weighted grammians on balancing. Proc. of IMACS/IFAC Symp. on Modelling
for both LCF and RCF based reductions, which are generally and Control of Technological Systems, Lille, France,
applicable for the reduction of all types of H∞ controllers, pro- (A. El Moudni, P. Borne, S. G. Tzafestas, Eds.), vol. 2,
vided the associated parameterization of all controllers is also pp. 42–47, 1991.
available. To compute the grammians, in all cases it is suffi- [13] A. Varga. Coprime factors model reduction based on
cient to solve two Lyapunov equations of the order of the con- square-root balancing-free techniques. In Computational
troller. Therefore, the new procedures are sensibly more ef- System Analysis 1992, Proc. 4-th Int. Symp. Systems
ficient than the standard frequency-weighted balancing based Analysis and Simulation, Berlin, Germany, (A. Sydow,
reduction approach. The frequency weighted grammians can Ed.), pp. 91–96, Elsevier, Amsterdam, 1992.
be determined directly in Cholesky factored forms to facilitate
the application of square-root and balancing-free accuracy en- [14] A. Varga. On frequency-weighted coprime factorization
hancing techniques. based controller reduction. Proc. of ACC’20003, Denver,
CO, USA, 2003.
References [15] A. Varga and B. D. O. Anderson. Frequency-
[1] B. D. O. Anderson and Y. Liu. Controller reduction: weighted balancing related controller reduction. Proc. of
concepts and approaches. IEEE Trans. Autom. Control, IFAC’2002 Congress, Barcelona, Spain, 2002.
34:802–812, 1989. [16] G. Wang, V. Sreeram, and W. Q. Liu. Performance pre-
[2] H. M. H. El-Zobaidi, I. M. Jaimoukha, and D. J. N. Lime- serving controller reduction via additive perturbation of
beer. Normalized H∞ controller reduction with a pri- the closed-loop transfer function. IEEE Trans. Autom.
ori error bounds. IEEE Trans. Autom. Control, 46:1477– Control, 46:771–775, 2001.
1483, 2001. [17] K. Zhou. Frequency-weighted L∞ norm and optimal
[3] D. Enns. Model Reduction for Control Systems Design. Hankel norm model reduction. IEEE Trans. Autom. Con-
PhD thesis, Dept. Aeronaut. Astronaut., Stanford Univ., trol, 40:1687–1699, 1995.
Stanford, CA, 1984. [18] K. Zhou and J. Chen. Performance bounds for coprime
[4] P. J. Goddard and K. Glover. Controller approximation: factor controller reductions. Systems & Control Lett.,
approaches for preserving H∞ performance. IEEE Trans. 26:119–127, 1995.
Autom. Control, 43:858–871, 1999. [19] K. Zhou, J. C. Doyle, and K. Glover. Robust and Optimal
[5] G. Gu. Model reduction with relative/multiplicative error Control. Prentice Hall, 1996.
bounds and relations to controller reduction. IEEE Trans.
Autom. Control, 40:1478–1485, 1995.